feat: integrate Opus codec into lib/libopus + voice messages in chatgui

- Add lib/libopus/ (opus v1.6.1 embedded source: CELT 18 .c, SILK 77 .c, SILK float 28 .c, OPUS core 14 .c) - Add lib/opus_codec.h/.c (wrapper: encoder, decoder, raw PCM int16) - Add tests/test_opus_codec.c (7 tests: silence, sine, compression, PLC, null, defaults, multi-frame) - chatgui: AudioRecorder (miniaudio capture, 48kHz mono) - chatgui: VoiceEncoder (PCM→Opus file with custom container) - chatgui: VoicePlayer (Opus→PCM playback via SoundManager) - chatgui: PTT button in InputBar, voice message send/receive in MessageList - chatgui: AudioDeviceSettingsPage (device selection, persisted in ui_state) - chatgui: MessageDelegate voice message rendering (play button + duration) - Update autotools: configure.ac, Makefile.am, src/Makefile.am, tests/Makefile.am - Update CMake: chatgui CMakeLists.txt, libutun CMakeLists.txt (link libopus_internal.a)
2 months ago · 4fde0b0325
260 changed files with 72665 additions and 16 deletions
--- a/Makefile.am
+++ b/Makefile.am
@ -1,7 +1,7 @@
 # Main Makefile.am for utun - root directory configuration

 # Order matters: lib -> src -> tests (tests depend on src objects)
-SUBDIRS = lib src tests
+SUBDIRS = lib lib/libopus src tests

 # Clean build directory
 clean-local:
--- a/configure.ac
+++ b/configure.ac
@ -123,7 +123,7 @@ AC_COMPILE_IFELSE([AC_LANG_PROGRAM([[#include <linux/if_tun.h>]], [[int fd = ope
    [AC_MSG_RESULT([no])])

 # Output
-AC_CONFIG_FILES([Makefile src/Makefile tests/Makefile lib/Makefile])
+AC_CONFIG_FILES([Makefile src/Makefile tests/Makefile lib/Makefile lib/libopus/Makefile])
 AC_OUTPUT

 # Summary
--- a/lib/Makefile.am
+++ b/lib/Makefile.am
@ -32,7 +32,9 @@ libuasync_a_SOURCES = \
 	tcp_io.c \
 	tcp_io.h \
 	sqlite3.c \
-	sqlite3.h
+	sqlite3.h \
+	opus_codec.c \
+	opus_codec.h

 libuasync_a_CFLAGS = \
 	-D_ISOC99_SOURCE \
@ -40,7 +42,8 @@ libuasync_a_CFLAGS = \
 	-DSQLITE_THREADSAFE=1 \
 	-g \
 	-I$(top_srcdir)/src \
-	-I$(top_srcdir)/lib
+	-I$(top_srcdir)/lib \
+	-I$(top_srcdir)/lib/libopus/include

 # Clean build directory
 clean-local:
--- a/lib/libopus/Makefile.am
+++ b/lib/libopus/Makefile.am
@ -0,0 +1,154 @@
+noinst_LIBRARIES = libopus_internal.a
+
+libopus_internal_a_SOURCES = \
+	celt/bands.c \
+	celt/celt.c \
+	celt/celt_encoder.c \
+	celt/celt_decoder.c \
+	celt/celt_lpc.c \
+	celt/cwrs.c \
+	celt/entcode.c \
+	celt/entdec.c \
+	celt/entenc.c \
+	celt/kiss_fft.c \
+	celt/laplace.c \
+	celt/mathops.c \
+	celt/mdct.c \
+	celt/modes.c \
+	celt/pitch.c \
+	celt/quant_bands.c \
+	celt/rate.c \
+	celt/vq.c \
+	silk/CNG.c \
+	silk/code_signs.c \
+	silk/init_decoder.c \
+	silk/decode_core.c \
+	silk/decode_frame.c \
+	silk/decode_parameters.c \
+	silk/decode_indices.c \
+	silk/decode_pulses.c \
+	silk/decoder_set_fs.c \
+	silk/dec_API.c \
+	silk/enc_API.c \
+	silk/encode_indices.c \
+	silk/encode_pulses.c \
+	silk/gain_quant.c \
+	silk/interpolate.c \
+	silk/LP_variable_cutoff.c \
+	silk/NLSF_decode.c \
+	silk/NSQ.c \
+	silk/NSQ_del_dec.c \
+	silk/PLC.c \
+	silk/shell_coder.c \
+	silk/tables_gain.c \
+	silk/tables_LTP.c \
+	silk/tables_NLSF_CB_NB_MB.c \
+	silk/tables_NLSF_CB_WB.c \
+	silk/tables_other.c \
+	silk/tables_pitch_lag.c \
+	silk/tables_pulses_per_block.c \
+	silk/VAD.c \
+	silk/control_audio_bandwidth.c \
+	silk/quant_LTP_gains.c \
+	silk/VQ_WMat_EC.c \
+	silk/HP_variable_cutoff.c \
+	silk/NLSF_encode.c \
+	silk/NLSF_VQ.c \
+	silk/NLSF_unpack.c \
+	silk/NLSF_del_dec_quant.c \
+	silk/process_NLSFs.c \
+	silk/stereo_LR_to_MS.c \
+	silk/stereo_MS_to_LR.c \
+	silk/check_control_input.c \
+	silk/control_SNR.c \
+	silk/init_encoder.c \
+	silk/control_codec.c \
+	silk/A2NLSF.c \
+	silk/ana_filt_bank_1.c \
+	silk/biquad_alt.c \
+	silk/bwexpander_32.c \
+	silk/bwexpander.c \
+	silk/debug.c \
+	silk/decode_pitch.c \
+	silk/inner_prod_aligned.c \
+	silk/lin2log.c \
+	silk/log2lin.c \
+	silk/LPC_analysis_filter.c \
+	silk/LPC_inv_pred_gain.c \
+	silk/table_LSF_cos.c \
+	silk/NLSF2A.c \
+	silk/NLSF_stabilize.c \
+	silk/NLSF_VQ_weights_laroia.c \
+	silk/pitch_est_tables.c \
+	silk/resampler.c \
+	silk/resampler_down2_3.c \
+	silk/resampler_down2.c \
+	silk/resampler_private_AR2.c \
+	silk/resampler_private_down_FIR.c \
+	silk/resampler_private_IIR_FIR.c \
+	silk/resampler_private_up2_HQ.c \
+	silk/resampler_rom.c \
+	silk/sigm_Q15.c \
+	silk/sort.c \
+	silk/sum_sqr_shift.c \
+	silk/stereo_decode_pred.c \
+	silk/stereo_encode_pred.c \
+	silk/stereo_find_predictor.c \
+	silk/stereo_quant_pred.c \
+	silk/LPC_fit.c \
+	silk/float/apply_sine_window_FLP.c \
+	silk/float/corrMatrix_FLP.c \
+	silk/float/encode_frame_FLP.c \
+	silk/float/find_LPC_FLP.c \
+	silk/float/find_LTP_FLP.c \
+	silk/float/find_pitch_lags_FLP.c \
+	silk/float/find_pred_coefs_FLP.c \
+	silk/float/LPC_analysis_filter_FLP.c \
+	silk/float/LTP_analysis_filter_FLP.c \
+	silk/float/LTP_scale_ctrl_FLP.c \
+	silk/float/noise_shape_analysis_FLP.c \
+	silk/float/process_gains_FLP.c \
+	silk/float/regularize_correlations_FLP.c \
+	silk/float/residual_energy_FLP.c \
+	silk/float/warped_autocorrelation_FLP.c \
+	silk/float/wrappers_FLP.c \
+	silk/float/autocorrelation_FLP.c \
+	silk/float/burg_modified_FLP.c \
+	silk/float/bwexpander_FLP.c \
+	silk/float/energy_FLP.c \
+	silk/float/inner_product_FLP.c \
+	silk/float/k2a_FLP.c \
+	silk/float/LPC_inv_pred_gain_FLP.c \
+	silk/float/pitch_analysis_core_FLP.c \
+	silk/float/scale_copy_vector_FLP.c \
+	silk/float/scale_vector_FLP.c \
+	silk/float/schur_FLP.c \
+	silk/float/sort_FLP.c \
+	src/opus.c \
+	src/opus_decoder.c \
+	src/opus_encoder.c \
+	src/extensions.c \
+	src/opus_multistream.c \
+	src/opus_multistream_encoder.c \
+	src/opus_multistream_decoder.c \
+	src/repacketizer.c \
+	src/opus_projection_encoder.c \
+	src/opus_projection_decoder.c \
+	src/mapping_matrix.c \
+	src/analysis.c \
+	src/mlp.c \
+	src/mlp_data.c
+
+libopus_internal_a_CFLAGS = \
+	-DOPUS_BUILD \
+	-DVAR_ARRAYS \
+	-DUSE_ALLOCA \
+	-DHAVE_ALLOCA_H \
+	-g \
+	-I$(srcdir)/include \
+	-I$(srcdir)/celt \
+	-I$(srcdir)/silk \
+	-I$(srcdir)/silk/float \
+	-I$(srcdir)/silk/fixed \
+	-I$(srcdir)/src \
+	-I$(srcdir)
--- a/lib/libopus/celt/_kiss_fft_guts.h
+++ b/lib/libopus/celt/_kiss_fft_guts.h
@ -0,0 +1,188 @@
+/*Copyright (c) 2003-2004, Mark Borgerding
+
+  All rights reserved.
+
+  Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are met:
+
+    * Redistributions of source code must retain the above copyright notice,
+       this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above copyright notice,
+       this list of conditions and the following disclaimer in the
+       documentation and/or other materials provided with the distribution.
+
+  THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+  AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+  IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+  ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+  LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+  CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+  SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+  INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+  CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+  ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+  POSSIBILITY OF SUCH DAMAGE.*/
+
+#ifndef KISS_FFT_GUTS_H
+#define KISS_FFT_GUTS_H
+
+#define MIN(a,b) ((a)<(b) ? (a):(b))
+#define MAX(a,b) ((a)>(b) ? (a):(b))
+
+/* kiss_fft.h
+   defines kiss_fft_scalar as either short or a float type
+   and defines
+   typedef struct { kiss_fft_scalar r; kiss_fft_scalar i; }kiss_fft_cpx; */
+#include "kiss_fft.h"
+
+/*
+  Explanation of macros dealing with complex math:
+
+   C_MUL(m,a,b)         : m = a*b
+   C_FIXDIV( c , div )  : if a fixed point impl., c /= div. noop otherwise
+   C_SUB( res, a,b)     : res = a - b
+   C_SUBFROM( res , a)  : res -= a
+   C_ADDTO( res , a)    : res += a
+ * */
+#ifdef FIXED_POINT
+#include "arch.h"
+
+
+#define SAMP_MAX 2147483647
+#define TWID_MAX 32767
+#define TRIG_UPSCALE 1
+
+#define SAMP_MIN -SAMP_MAX
+
+#ifdef ENABLE_QEXT
+#   define S_MUL(a,b) MULT32_32_P31_ovflw(b, a)
+#   define S_MUL2(a,b) MULT32_32_P31_ovflw(b, a)
+#else
+#   define S_MUL(a,b) MULT16_32_Q15(b, a)
+#   define S_MUL2(a,b) MULT16_32_Q16(b, a)
+#endif
+
+#   define C_MUL(m,a,b) \
+      do{ (m).r = SUB32_ovflw(S_MUL((a).r,(b).r) , S_MUL((a).i,(b).i)); \
+          (m).i = ADD32_ovflw(S_MUL((a).r,(b).i) , S_MUL((a).i,(b).r)); }while(0)
+
+#   define C_MULC(m,a,b) \
+      do{ (m).r = ADD32_ovflw(S_MUL((a).r,(b).r) , S_MUL((a).i,(b).i)); \
+          (m).i = SUB32_ovflw(S_MUL((a).i,(b).r) , S_MUL((a).r,(b).i)); }while(0)
+
+#   define C_MULBYSCALAR( c, s ) \
+      do{ (c).r =  S_MUL( (c).r , s ) ;\
+          (c).i =  S_MUL( (c).i , s ) ; }while(0)
+
+#   define DIVSCALAR(x,k) \
+        (x) = S_MUL(  x, (TWID_MAX-((k)>>1))/(k)+1 )
+
+#   define C_FIXDIV(c,div) \
+        do {    DIVSCALAR( (c).r , div);  \
+                DIVSCALAR( (c).i  , div); }while (0)
+
+#define  C_ADD( res, a,b)\
+    do {(res).r=ADD32_ovflw((a).r,(b).r);  (res).i=ADD32_ovflw((a).i,(b).i); \
+    }while(0)
+#define  C_SUB( res, a,b)\
+    do {(res).r=SUB32_ovflw((a).r,(b).r);  (res).i=SUB32_ovflw((a).i,(b).i); \
+    }while(0)
+#define C_ADDTO( res , a)\
+    do {(res).r = ADD32_ovflw((res).r, (a).r);  (res).i = ADD32_ovflw((res).i,(a).i);\
+    }while(0)
+
+#define C_SUBFROM( res , a)\
+    do {(res).r = ADD32_ovflw((res).r,(a).r);  (res).i = SUB32_ovflw((res).i,(a).i); \
+    }while(0)
+
+#if defined(OPUS_ARM_INLINE_ASM)
+#include "arm/kiss_fft_armv4.h"
+#endif
+
+#if defined(OPUS_ARM_INLINE_EDSP)
+#include "arm/kiss_fft_armv5e.h"
+#endif
+#if defined(__mips)
+#include "mips/kiss_fft_mipsr1.h"
+#endif
+
+#else  /* not FIXED_POINT*/
+
+#   define S_MUL(a,b) ( (a)*(b) )
+#   define S_MUL2(a,b) ( (a)*(b) )
+#define C_MUL(m,a,b) \
+    do{ (m).r = (a).r*(b).r - (a).i*(b).i;\
+        (m).i = (a).r*(b).i + (a).i*(b).r; }while(0)
+#define C_MULC(m,a,b) \
+    do{ (m).r = (a).r*(b).r + (a).i*(b).i;\
+        (m).i = (a).i*(b).r - (a).r*(b).i; }while(0)
+
+#define C_MUL4(m,a,b) C_MUL(m,a,b)
+
+#   define C_FIXDIV(c,div) /* NOOP */
+#   define C_MULBYSCALAR( c, s ) \
+    do{ (c).r *= (s);\
+        (c).i *= (s); }while(0)
+#endif
+
+#ifndef CHECK_OVERFLOW_OP
+#  define CHECK_OVERFLOW_OP(a,op,b) /* noop */
+#endif
+
+#ifndef C_ADD
+#define  C_ADD( res, a,b)\
+    do { \
+            CHECK_OVERFLOW_OP((a).r,+,(b).r)\
+            CHECK_OVERFLOW_OP((a).i,+,(b).i)\
+            (res).r=(a).r+(b).r;  (res).i=(a).i+(b).i; \
+    }while(0)
+#define  C_SUB( res, a,b)\
+    do { \
+            CHECK_OVERFLOW_OP((a).r,-,(b).r)\
+            CHECK_OVERFLOW_OP((a).i,-,(b).i)\
+            (res).r=(a).r-(b).r;  (res).i=(a).i-(b).i; \
+    }while(0)
+#define C_ADDTO( res , a)\
+    do { \
+            CHECK_OVERFLOW_OP((res).r,+,(a).r)\
+            CHECK_OVERFLOW_OP((res).i,+,(a).i)\
+            (res).r += (a).r;  (res).i += (a).i;\
+    }while(0)
+
+#define C_SUBFROM( res , a)\
+    do {\
+            CHECK_OVERFLOW_OP((res).r,-,(a).r)\
+            CHECK_OVERFLOW_OP((res).i,-,(a).i)\
+            (res).r -= (a).r;  (res).i -= (a).i; \
+    }while(0)
+#endif /* C_ADD defined */
+
+#ifdef FIXED_POINT
+/*#  define KISS_FFT_COS(phase)  TRIG_UPSCALE*floor(MIN(32767,MAX(-32767,.5+32768 * cos (phase))))
+#  define KISS_FFT_SIN(phase)  TRIG_UPSCALE*floor(MIN(32767,MAX(-32767,.5+32768 * sin (phase))))*/
+#  define KISS_FFT_COS(phase)  floor(.5+TWID_MAX*cos (phase))
+#  define KISS_FFT_SIN(phase)  floor(.5+TWID_MAX*sin (phase))
+#  define HALF_OF(x) ((x)>>1)
+#elif defined(USE_SIMD)
+#  define KISS_FFT_COS(phase) _mm_set1_ps( cos(phase) )
+#  define KISS_FFT_SIN(phase) _mm_set1_ps( sin(phase) )
+#  define HALF_OF(x) ((x)*_mm_set1_ps(.5f))
+#else
+#  define KISS_FFT_COS(phase) (kiss_fft_scalar) cos(phase)
+#  define KISS_FFT_SIN(phase) (kiss_fft_scalar) sin(phase)
+#  define HALF_OF(x) ((x)*.5f)
+#endif
+
+#define  kf_cexp(x,phase) \
+        do{ \
+                (x)->r = KISS_FFT_COS(phase);\
+                (x)->i = KISS_FFT_SIN(phase);\
+        }while(0)
+
+#define  kf_cexp2(x,phase) \
+   do{ \
+      (x)->r = TRIG_UPSCALE*celt_cos_norm((phase));\
+      (x)->i = TRIG_UPSCALE*celt_cos_norm((phase)-32768);\
+}while(0)
+
+#endif /* KISS_FFT_GUTS_H */
--- a/lib/libopus/celt/arch.h
+++ b/lib/libopus/celt/arch.h
@ -0,0 +1,400 @@
+/* Copyright (c) 2003-2008 Jean-Marc Valin
+   Copyright (c) 2007-2008 CSIRO
+   Copyright (c) 2007-2009 Xiph.Org Foundation
+   Written by Jean-Marc Valin */
+/**
+   @file arch.h
+   @brief Various architecture definitions for CELT
+*/
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef ARCH_H
+#define ARCH_H
+
+#include "opus_types.h"
+#include "opus_defines.h"
+
+# if !defined(__GNUC_PREREQ)
+#  if defined(__GNUC__)&&defined(__GNUC_MINOR__)
+#   define __GNUC_PREREQ(_maj,_min) \
+ ((__GNUC__<<16)+__GNUC_MINOR__>=((_maj)<<16)+(_min))
+#  else
+#   define __GNUC_PREREQ(_maj,_min) 0
+#  endif
+# endif
+
+#if OPUS_GNUC_PREREQ(3, 0)
+#define opus_likely(x)       (__builtin_expect(!!(x), 1))
+#define opus_unlikely(x)     (__builtin_expect(!!(x), 0))
+#else
+#define opus_likely(x)       (!!(x))
+#define opus_unlikely(x)     (!!(x))
+#endif
+
+#define CELT_SIG_SCALE 32768.f
+
+#define CELT_FATAL(str) celt_fatal(str, __FILE__, __LINE__)
+
+#if defined(ENABLE_ASSERTIONS) || defined(ENABLE_HARDENING)
+#ifdef __GNUC__
+__attribute__((noreturn))
+#endif
+void celt_fatal(const char *str, const char *file, int line);
+
+#if defined(CELT_C) && !defined(OVERRIDE_celt_fatal)
+#include <stdio.h>
+#include <stdlib.h>
+#ifdef __GNUC__
+__attribute__((noreturn))
+#endif
+void celt_fatal(const char *str, const char *file, int line)
+{
+   fprintf (stderr, "Fatal (internal) error in %s, line %d: %s\n", file, line, str);
+#if defined(_MSC_VER)
+   _set_abort_behavior( 0, _WRITE_ABORT_MSG);
+#endif
+   abort();
+}
+#endif
+
+#define celt_assert(cond) {if (!(cond)) {CELT_FATAL("assertion failed: " #cond);}}
+#define celt_assert2(cond, message) {if (!(cond)) {CELT_FATAL("assertion failed: " #cond "\n" message);}}
+#define MUST_SUCCEED(call) celt_assert((call) == OPUS_OK)
+#else
+#define celt_assert(cond) ((void)(cond))
+#define celt_assert2(cond, message) ((void)(cond))
+#define MUST_SUCCEED(call) do {if((call) != OPUS_OK) {RESTORE_STACK; return OPUS_INTERNAL_ERROR;} } while (0)
+#endif
+
+#if defined(ENABLE_ASSERTIONS)
+#define celt_sig_assert(cond) {if (!(cond)) {CELT_FATAL("signal assertion failed: " #cond);}}
+#else
+#define celt_sig_assert(cond) ((void)(cond))
+#endif
+
+#define IMUL32(a,b) ((a)*(b))
+
+#define MIN16(a,b) ((a) < (b) ? (a) : (b))   /**< Minimum 16-bit value.   */
+#define MAX16(a,b) ((a) > (b) ? (a) : (b))   /**< Maximum 16-bit value.   */
+#define MIN32(a,b) ((a) < (b) ? (a) : (b))   /**< Minimum 32-bit value.   */
+#define MAX32(a,b) ((a) > (b) ? (a) : (b))   /**< Maximum 32-bit value.   */
+#define IMIN(a,b) ((a) < (b) ? (a) : (b))   /**< Minimum int value.   */
+#define IMAX(a,b) ((a) > (b) ? (a) : (b))   /**< Maximum int value.   */
+#define FMIN(a,b) ((a) < (b) ? (a) : (b))   /**< Minimum float value.   */
+#define FMAX(a,b) ((a) > (b) ? (a) : (b))   /**< Maximum float value.   */
+#define UADD32(a,b) ((a)+(b))
+#define USUB32(a,b) ((a)-(b))
+#define MAXG(a,b) MAX32(a, b)
+#define MING(a,b) MIN32(a, b)
+
+/* Throughout the code, we use the following scaling for signals:
+   FLOAT: used for float API, normalized to +/-1.
+   INT16: used for 16-bit API, normalized to +/- 32768
+   RES: internal Opus resolution, defined as +/-1. in float builds, or either 16-bit or 24-bit int for fixed-point builds
+   SIG: internal CELT resolution: defined as +/- 32768. in float builds, or Q27 in fixed-point builds (int16 shifted by 12)
+*/
+
+
+/* Set this if opus_int64 is a native type of the CPU. */
+/* Assume that all LP64 architectures have fast 64-bit types; also x86_64
+   (which can be ILP32 for x32) and Win64 (which is LLP64). */
+#if defined(__x86_64__) || defined(__LP64__) || defined(_WIN64) || defined (__mips)
+#define OPUS_FAST_INT64 1
+#else
+#define OPUS_FAST_INT64 0
+#endif
+
+#ifdef FIXED_POINT
+#define ARG_FIXED(arg) , arg
+#else
+#define ARG_FIXED(arg)
+#endif
+
+#define PRINT_MIPS(file)
+
+#ifdef FIXED_POINT
+
+typedef opus_int16 opus_val16;
+typedef opus_int32 opus_val32;
+typedef opus_int64 opus_val64;
+
+typedef opus_val32 celt_sig;
+typedef opus_val32 celt_norm;
+typedef opus_val32 celt_ener;
+typedef opus_val32 celt_glog;
+
+#ifdef ENABLE_RES24
+typedef opus_val32 opus_res;
+#define RES_SHIFT 8
+#define SIG2RES(a)      PSHR32(a, SIG_SHIFT-RES_SHIFT)
+#define RES2INT16(a)    SAT16(PSHR32(a, RES_SHIFT))
+#define RES2INT24(a)    (a)
+#define RES2FLOAT(a)    ((1.f/32768.f/256.f)*(a))
+#define INT16TORES(a)   SHL32(EXTEND32(a), RES_SHIFT)
+#define INT24TORES(a)   (a)
+#define ADD_RES(a, b)   ADD32(a, b)
+#define FLOAT2RES(a)    FLOAT2INT24(a)
+#define RES2SIG(a)      SHL32((a), SIG_SHIFT-RES_SHIFT)
+#define MULT16_RES_Q15(a,b) MULT16_32_Q15(a,b)
+#define MAX_ENCODING_DEPTH 24
+#else
+typedef opus_val16 opus_res;
+#define RES_SHIFT 0
+#define SIG2RES(a)      SIG2WORD16(a)
+#define RES2INT16(a)    (a)
+#define RES2INT24(a)    SHL32(EXTEND32(a), 8)
+#define RES2FLOAT(a)    ((1.f/32768.f)*(a))
+#define INT16TORES(a)   (a)
+#define INT24TORES(a)   SAT16(PSHR32(a, 8))
+#define ADD_RES(a, b)   SAT16(ADD32((a), (b)));
+#define FLOAT2RES(a)    FLOAT2INT16(a)
+#define RES2SIG(a)      SHL32(EXTEND32(a), SIG_SHIFT)
+#define MULT16_RES_Q15(a,b) MULT16_16_Q15(a,b)
+#define MAX_ENCODING_DEPTH 16
+#endif
+
+#define RES2VAL16(a)    RES2INT16(a)
+#define INT16TOSIG(a)   SHL32(EXTEND32(a), SIG_SHIFT)
+#define INT24TOSIG(a)   SHL32(a, SIG_SHIFT-8)
+
+#define NORM_SHIFT 24
+#ifdef ENABLE_QEXT
+typedef opus_val32 celt_coef;
+#define COEF_ONE Q31ONE
+#define MULT_COEF_32(a, b) MULT32_32_P31(a,b)
+#define MAC_COEF_32_ARM(c, a, b) ADD32((c), MULT32_32_Q32(a,b))
+#define MULT_COEF(a, b) MULT32_32_Q31(a,b)
+#define MULT_COEF_TAPS(a, b) SHL32(MULT16_16(a,b), 1)
+#define COEF2VAL16(x) EXTRACT16(SHR32(x, 16))
+#else
+typedef opus_val16 celt_coef;
+#define COEF_ONE Q15ONE
+#define MULT_COEF_32(a, b) MULT16_32_Q15(a,b)
+#define MAC_COEF_32_ARM(a, b, c) MAC16_32_Q16(a,b,c)
+#define MULT_COEF(a, b) MULT16_16_Q15(a,b)
+#define MULT_COEF_TAPS(a, b) MULT16_16_P15(a,b)
+#define COEF2VAL16(x) (x)
+#endif
+
+#define celt_isnan(x) 0
+
+#define Q15ONE 32767
+#define Q31ONE 2147483647
+
+#define SIG_SHIFT 12
+/* Safe saturation value for 32-bit signals. We need to make sure that we can
+   add two sig values and that the first stages of the MDCT don't cause an overflow.
+   The most constraining is the ARM_ASM comb filter where we shift left by one
+   and then add two values. Because of that, we use 2^29-1. SIG_SAT must be large
+   enough to fit a full-scale high-freq tone through the prefilter and comb filter,
+   meaning 1.85*1.75*2^(15+SIG_SHIFT) =  434529895.
+   so the limit should be about 2^31*sqrt(.5). */
+#define SIG_SAT (536870911)
+
+#define NORM_SCALING (1<<NORM_SHIFT)
+
+#define DB_SHIFT 24
+
+#define EPSILON 1
+#define VERY_SMALL 0
+#define VERY_LARGE16 ((opus_val16)32767)
+#define Q15_ONE ((opus_val16)32767)
+
+
+#define ABS16(x) ((x) < 0 ? (-(x)) : (x))
+#define ABS32(x) ((x) < 0 ? (-(x)) : (x))
+
+static OPUS_INLINE opus_int16 SAT16(opus_int32 x) {
+   return x > 32767 ? 32767 : x < -32768 ? -32768 : (opus_int16)x;
+}
+
+#ifdef FIXED_DEBUG
+#include "fixed_debug.h"
+#else
+
+#include "fixed_generic.h"
+
+#ifdef OPUS_ARM_PRESUME_AARCH64_NEON_INTR
+#include "arm/fixed_arm64.h"
+#elif defined (OPUS_ARM_INLINE_EDSP)
+#include "arm/fixed_armv5e.h"
+#elif defined (OPUS_ARM_INLINE_ASM)
+#include "arm/fixed_armv4.h"
+#elif defined (BFIN_ASM)
+#include "fixed_bfin.h"
+#elif defined (TI_C5X_ASM)
+#include "fixed_c5x.h"
+#elif defined (TI_C6X_ASM)
+#include "fixed_c6x.h"
+#endif
+
+#endif
+
+#else /* FIXED_POINT */
+
+typedef float opus_val16;
+typedef float opus_val32;
+typedef float opus_val64;
+
+typedef float celt_sig;
+typedef float celt_norm;
+typedef float celt_ener;
+typedef float celt_glog;
+
+typedef float opus_res;
+typedef float celt_coef;
+
+#ifdef FLOAT_APPROX
+/* This code should reliably detect NaN/inf even when -ffast-math is used.
+   Assumes IEEE 754 format. */
+static OPUS_INLINE int celt_isnan(float x)
+{
+   union {float f; opus_uint32 i;} in;
+   in.f = x;
+   return ((in.i>>23)&0xFF)==0xFF && (in.i&0x007FFFFF)!=0;
+}
+#else
+#ifdef __FAST_MATH__
+#error Cannot build libopus with -ffast-math unless FLOAT_APPROX is defined. This could result in crashes on extreme (e.g. NaN) input
+#endif
+#define celt_isnan(x) ((x)!=(x))
+#endif
+
+#define Q15ONE 1.0f
+#define Q31ONE 1.0f
+#define COEF_ONE 1.0f
+#define COEF2VAL16(x) (x)
+
+#define NORM_SCALING 1.f
+
+#define EPSILON 1e-15f
+#define VERY_SMALL 1e-30f
+#define VERY_LARGE16 1e15f
+#define Q15_ONE ((opus_val16)1.f)
+
+/* This appears to be the same speed as C99's fabsf() but it's more portable. */
+#define ABS16(x) ((float)fabs(x))
+#define ABS32(x) ((float)fabs(x))
+
+#define QCONST16(x,bits) (x)
+#define QCONST32(x,bits) (x)
+#define GCONST(x) (x)
+
+#define NEG16(x) (-(x))
+#define NEG32(x) (-(x))
+#define NEG32_ovflw(x) (-(x))
+#define EXTRACT16(x) (x)
+#define EXTEND32(x) (x)
+#define SHR16(a,shift) (a)
+#define SHL16(a,shift) (a)
+#define SHR32(a,shift) (a)
+#define SHL32(a,shift) (a)
+#define PSHR32(a,shift) (a)
+#define VSHR32(a,shift) (a)
+
+#define SHR64(a,shift) (a)
+
+#define PSHR(a,shift)   (a)
+#define SHR(a,shift)    (a)
+#define SHL(a,shift)    (a)
+#define SATURATE(x,a)   (x)
+#define SATURATE16(x)   (x)
+
+#define ROUND16(a,shift)  (a)
+#define SROUND16(a,shift) (a)
+#define HALF16(x)       (.5f*(x))
+#define HALF32(x)       (.5f*(x))
+
+#define ADD16(a,b) ((a)+(b))
+#define SUB16(a,b) ((a)-(b))
+#define ADD32(a,b) ((a)+(b))
+#define SUB32(a,b) ((a)-(b))
+#define ADD32_ovflw(a,b) ((a)+(b))
+#define SUB32_ovflw(a,b) ((a)-(b))
+#define SHL32_ovflw(a,shift) (a)
+#define PSHR32_ovflw(a,shift) (a)
+
+#define MULT16_16_16(a,b)     ((a)*(b))
+#define MULT16_16(a,b)     ((opus_val32)(a)*(opus_val32)(b))
+#define MAC16_16(c,a,b)     ((c)+(opus_val32)(a)*(opus_val32)(b))
+
+#define MULT16_32_Q15(a,b)     ((a)*(b))
+#define MULT16_32_Q16(a,b)     ((a)*(b))
+
+#define MULT32_32_Q16(a,b)     ((a)*(b))
+#define MULT32_32_Q31(a,b)     ((a)*(b))
+#define MULT32_32_P31(a,b)     ((a)*(b))
+#define MULT32_32_P31_ovflw(a,b) ((a)*(b))
+
+#define MAC16_32_Q15(c,a,b)     ((c)+(a)*(b))
+#define MAC16_32_Q16(c,a,b)     ((c)+(a)*(b))
+#define MAC_COEF_32_ARM(c,a,b)     ((c)+(a)*(b))
+
+#define MULT16_16_Q11_32(a,b)     ((a)*(b))
+#define MULT16_16_Q11(a,b)     ((a)*(b))
+#define MULT16_16_Q13(a,b)     ((a)*(b))
+#define MULT16_16_Q14(a,b)     ((a)*(b))
+#define MULT16_16_Q15(a,b)     ((a)*(b))
+#define MULT16_16_P15(a,b)     ((a)*(b))
+#define MULT16_16_P13(a,b)     ((a)*(b))
+#define MULT16_16_P14(a,b)     ((a)*(b))
+#define MULT16_32_P16(a,b)     ((a)*(b))
+
+#define MULT_COEF_32(a, b)      ((a)*(b))
+#define MULT_COEF(a, b)   ((a)*(b))
+#define MULT_COEF_TAPS(a, b)   ((a)*(b))
+
+#define DIV32_16(a,b)     (((opus_val32)(a))/(opus_val16)(b))
+#define DIV32(a,b)     (((opus_val32)(a))/(opus_val32)(b))
+
+#define SIG2RES(a)      ((1/CELT_SIG_SCALE)*(a))
+#define RES2INT16(a)    FLOAT2INT16(a)
+#define RES2INT24(a)    float2int(32768.f*256.f*(a))
+#define RES2FLOAT(a)    (a)
+#define INT16TORES(a)   ((a)*(1/CELT_SIG_SCALE))
+#define INT24TORES(a)   ((1.f/32768.f/256.f)*(a))
+#define ADD_RES(a, b)   ADD32(a, b)
+#define FLOAT2RES(a)    (a)
+#define RES2SIG(a)      (CELT_SIG_SCALE*(a))
+#define MULT16_RES_Q15(a,b) MULT16_16_Q15(a,b)
+
+#define RES2VAL16(a)    (a)
+#define FLOAT2SIG(a)    ((a)*CELT_SIG_SCALE)
+#define INT16TOSIG(a)   ((float)(a))
+#define INT24TOSIG(a)   ((float)(a)*(1.f/256.f))
+#define MAX_ENCODING_DEPTH 24
+
+#endif /* !FIXED_POINT */
+
+#ifndef GLOBAL_STACK_SIZE
+#ifdef FIXED_POINT
+#define GLOBAL_STACK_SIZE 120000
+#else
+#define GLOBAL_STACK_SIZE 120000
+#endif
+#endif
+
+#endif /* ARCH_H */
--- a/lib/libopus/celt/bands.c
+++ b/lib/libopus/celt/bands.c
--- a/lib/libopus/celt/bands.h
+++ b/lib/libopus/celt/bands.h
@ -0,0 +1,125 @@
+/* Copyright (c) 2007-2008 CSIRO
+   Copyright (c) 2007-2009 Xiph.Org Foundation
+   Copyright (c) 2008-2009 Gregory Maxwell
+   Written by Jean-Marc Valin and Gregory Maxwell */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef BANDS_H
+#define BANDS_H
+
+#include "arch.h"
+#include "modes.h"
+#include "entenc.h"
+#include "entdec.h"
+#include "rate.h"
+
+opus_int16 bitexact_cos(opus_int16 x);
+int bitexact_log2tan(int isin,int icos);
+
+/** Compute the amplitude (sqrt energy) in each of the bands
+ * @param m Mode data
+ * @param X Spectrum
+ * @param bandE Square root of the energy for each band (returned)
+ */
+void compute_band_energies(const CELTMode *m, const celt_sig *X, celt_ener *bandE, int end, int C, int LM, int arch);
+
+/*void compute_noise_energies(const CELTMode *m, const celt_sig *X, const opus_val16 *tonality, celt_ener *bandE);*/
+
+/** Normalise each band of X such that the energy in each band is
+    equal to 1
+ * @param m Mode data
+ * @param X Spectrum (returned normalised)
+ * @param bandE Square root of the energy for each band
+ */
+void normalise_bands(const CELTMode *m, const celt_sig * OPUS_RESTRICT freq, celt_norm * OPUS_RESTRICT X, const celt_ener *bandE, int end, int C, int M);
+
+/** Denormalise each band of X to restore full amplitude
+ * @param m Mode data
+ * @param X Spectrum (returned de-normalised)
+ * @param bandE Square root of the energy for each band
+ */
+void denormalise_bands(const CELTMode *m, const celt_norm * OPUS_RESTRICT X,
+      celt_sig * OPUS_RESTRICT freq, const celt_glog *bandE, int start,
+      int end, int M, int downsample, int silence);
+
+#define SPREAD_NONE       (0)
+#define SPREAD_LIGHT      (1)
+#define SPREAD_NORMAL     (2)
+#define SPREAD_AGGRESSIVE (3)
+
+int spreading_decision(const CELTMode *m, const celt_norm *X, int *average,
+      int last_decision, int *hf_average, int *tapset_decision, int update_hf,
+      int end, int C, int M, const int *spread_weight);
+
+#ifdef MEASURE_NORM_MSE
+void measure_norm_mse(const CELTMode *m, float *X, float *X0, float *bandE, float *bandE0, int M, int N, int C);
+#endif
+
+void haar1(celt_norm *X, int N0, int stride);
+
+/** Quantisation/encoding of the residual spectrum
+ * @param encode flag that indicates whether we're encoding (1) or decoding (0)
+ * @param m Mode data
+ * @param start First band to process
+ * @param end Last band to process + 1
+ * @param X Residual (normalised)
+ * @param Y Residual (normalised) for second channel (or NULL for mono)
+ * @param collapse_masks Anti-collapse tracking mask
+ * @param bandE Square root of the energy for each band
+ * @param pulses Bit allocation (per band) for PVQ
+ * @param shortBlocks Zero for long blocks, non-zero for short blocks
+ * @param spread Amount of spreading to use
+ * @param dual_stereo Zero for MS stereo, non-zero for dual stereo
+ * @param intensity First band to use intensity stereo
+ * @param tf_res Time-frequency resolution change
+ * @param total_bits Total number of bits that can be used for the frame (including the ones already spent)
+ * @param balance Number of unallocated bits
+ * @param en Entropy coder state
+ * @param LM log2() of the number of 2.5 subframes in the frame
+ * @param codedBands Last band to receive bits + 1
+ * @param seed Random generator seed
+ * @param arch Run-time architecture (see opus_select_arch())
+ */
+void quant_all_bands(int encode, const CELTMode *m, int start, int end,
+      celt_norm * X, celt_norm * Y, unsigned char *collapse_masks,
+      const celt_ener *bandE, int *pulses, int shortBlocks, int spread,
+      int dual_stereo, int intensity, int *tf_res, opus_int32 total_bits,
+      opus_int32 balance, ec_ctx *ec, int M, int codedBands, opus_uint32 *seed,
+      int complexity, int arch, int disable_inv
+      ARG_QEXT(ec_ctx *ext_ec) ARG_QEXT(int *extra_pulses)
+      ARG_QEXT(opus_int32 total_ext_bits) ARG_QEXT(const int *cap));
+
+void anti_collapse(const CELTMode *m, celt_norm *X_,
+      unsigned char *collapse_masks, int LM, int C, int size, int start,
+      int end, const celt_glog *logE, const celt_glog *prev1logE,
+      const celt_glog *prev2logE, const int *pulses, opus_uint32 seed,
+      int encode, int arch);
+
+opus_uint32 celt_lcg_rand(opus_uint32 seed);
+
+int hysteresis_decision(opus_val16 val, const opus_val16 *thresholds, const opus_val16 *hysteresis, int N, int prev);
+
+#endif /* BANDS_H */
--- a/lib/libopus/celt/celt.c
+++ b/lib/libopus/celt/celt.c
@ -0,0 +1,373 @@
+/* Copyright (c) 2007-2008 CSIRO
+   Copyright (c) 2007-2010 Xiph.Org Foundation
+   Copyright (c) 2008 Gregory Maxwell
+   Written by Jean-Marc Valin and Gregory Maxwell */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#define CELT_C
+
+#include "os_support.h"
+#include "mdct.h"
+#include <math.h>
+#include "celt.h"
+#include "pitch.h"
+#include "bands.h"
+#include "modes.h"
+#include "entcode.h"
+#include "quant_bands.h"
+#include "rate.h"
+#include "stack_alloc.h"
+#include "mathops.h"
+#include "float_cast.h"
+#include <stdarg.h>
+#include "celt_lpc.h"
+#include "vq.h"
+
+#ifndef PACKAGE_VERSION
+#define PACKAGE_VERSION "unknown"
+#endif
+
+#if defined(FIXED_POINT) && defined(__mips)
+#include "mips/celt_mipsr1.h"
+#endif
+
+
+int resampling_factor(opus_int32 rate)
+{
+   int ret;
+   switch (rate)
+   {
+#ifdef ENABLE_QEXT
+   case 96000:
+#endif
+   case 48000:
+      ret = 1;
+      break;
+   case 24000:
+      ret = 2;
+      break;
+   case 16000:
+      ret = 3;
+      break;
+   case 12000:
+      ret = 4;
+      break;
+   case 8000:
+      ret = 6;
+      break;
+   default:
+#ifndef CUSTOM_MODES
+      celt_assert(0);
+#endif
+      ret = 0;
+      break;
+   }
+   return ret;
+}
+
+
+#if !defined(OVERRIDE_COMB_FILTER_CONST) || defined(NON_STATIC_COMB_FILTER_CONST_C)
+/* This version should be faster on ARM */
+#ifdef OPUS_ARM_ASM
+#ifndef NON_STATIC_COMB_FILTER_CONST_C
+static
+#endif
+void comb_filter_const_c(opus_val32 *y, opus_val32 *x, int T, int N,
+      celt_coef g10, celt_coef g11, celt_coef g12)
+{
+   opus_val32 x0, x1, x2, x3, x4;
+   int i;
+   x4 = SHL32(x[-T-2], 1);
+   x3 = SHL32(x[-T-1], 1);
+   x2 = SHL32(x[-T], 1);
+   x1 = SHL32(x[-T+1], 1);
+   for (i=0;i<N-4;i+=5)
+   {
+      opus_val32 t;
+      x0=SHL32(x[i-T+2],1);
+      t = MAC_COEF_32_ARM(x[i], g10, x2);
+      t = MAC_COEF_32_ARM(t, g11, ADD32(x1,x3));
+      t = MAC_COEF_32_ARM(t, g12, ADD32(x0,x4));
+      t = SATURATE(t, SIG_SAT);
+      y[i] = t;
+      x4=SHL32(x[i-T+3],1);
+      t = MAC_COEF_32_ARM(x[i+1], g10, x1);
+      t = MAC_COEF_32_ARM(t, g11, ADD32(x0,x2));
+      t = MAC_COEF_32_ARM(t, g12, ADD32(x4,x3));
+      t = SATURATE(t, SIG_SAT);
+      y[i+1] = t;
+      x3=SHL32(x[i-T+4],1);
+      t = MAC_COEF_32_ARM(x[i+2], g10, x0);
+      t = MAC_COEF_32_ARM(t, g11, ADD32(x4,x1));
+      t = MAC_COEF_32_ARM(t, g12, ADD32(x3,x2));
+      t = SATURATE(t, SIG_SAT);
+      y[i+2] = t;
+      x2=SHL32(x[i-T+5],1);
+      t = MAC_COEF_32_ARM(x[i+3], g10, x4);
+      t = MAC_COEF_32_ARM(t, g11, ADD32(x3,x0));
+      t = MAC_COEF_32_ARM(t, g12, ADD32(x2,x1));
+      t = SATURATE(t, SIG_SAT);
+      y[i+3] = t;
+      x1=SHL32(x[i-T+6],1);
+      t = MAC_COEF_32_ARM(x[i+4], g10, x3);
+      t = MAC_COEF_32_ARM(t, g11, ADD32(x2,x4));
+      t = MAC_COEF_32_ARM(t, g12, ADD32(x1,x0));
+      t = SATURATE(t, SIG_SAT);
+      y[i+4] = t;
+   }
+#ifdef CUSTOM_MODES
+   for (;i<N;i++)
+   {
+      opus_val32 t;
+      x0=SHL32(x[i-T+2],1);
+      t = MAC_COEF_32_ARM(x[i], g10, x2);
+      t = MAC_COEF_32_ARM(t, g11, ADD32(x1,x3));
+      t = MAC_COEF_32_ARM(t, g12, ADD32(x0,x4));
+      t = SATURATE(t, SIG_SAT);
+      y[i] = t;
+      x4=x3;
+      x3=x2;
+      x2=x1;
+      x1=x0;
+   }
+#endif
+}
+#else
+#ifndef NON_STATIC_COMB_FILTER_CONST_C
+static
+#endif
+void comb_filter_const_c(opus_val32 *y, opus_val32 *x, int T, int N,
+      celt_coef g10, celt_coef g11, celt_coef g12)
+{
+   opus_val32 x0, x1, x2, x3, x4;
+   int i;
+   x4 = x[-T-2];
+   x3 = x[-T-1];
+   x2 = x[-T];
+   x1 = x[-T+1];
+   for (i=0;i<N;i++)
+   {
+      x0=x[i-T+2];
+      y[i] = x[i]
+               + MULT_COEF_32(g10,x2)
+               + MULT_COEF_32(g11,ADD32(x1,x3))
+               + MULT_COEF_32(g12,ADD32(x0,x4));
+#ifdef FIXED_POINT
+      /* A bit of bias seems to help here. */
+      y[i] = SUB32(y[i], 1);
+#endif
+      y[i] = SATURATE(y[i], SIG_SAT);
+      x4=x3;
+      x3=x2;
+      x2=x1;
+      x1=x0;
+   }
+
+}
+#endif
+#endif
+
+#ifdef ENABLE_QEXT
+void comb_filter_qext(opus_val32 *y, opus_val32 *x, int T0, int T1, int N,
+      opus_val16 g0, opus_val16 g1, int tapset0, int tapset1,
+      const celt_coef *window, int overlap, int arch)
+{
+   VARDECL(opus_val32, mem_buf);
+   VARDECL(opus_val32, buf);
+   celt_coef new_window[120];
+   int s;
+   int i;
+   int N2;
+   int overlap2;
+   SAVE_STACK;
+   /* Using ALLOC() instead of a regular stack allocation to minimize real stack use when using the pseudostack.
+      This is useful on some embedded systems. */
+   ALLOC(mem_buf, COMBFILTER_MAXPERIOD+960, opus_val32);
+   ALLOC(buf, COMBFILTER_MAXPERIOD+960, opus_val32);
+   N2 = N/2;
+   overlap2=overlap/2;
+   /* At 96 kHz, we double the period and the spacing between taps, which is equivalent
+      to creating a mirror image of the filter around 24 kHz. It also means we can process
+      the even and odd samples completely independently. */
+   for (s=0;s<2;s++) {
+      opus_val32 *yptr;
+      for (i=0;i<overlap2;i++) new_window[i] = window[2*i+s];
+      for (i=0;i<COMBFILTER_MAXPERIOD+N2;i++) mem_buf[i] = x[2*i+s-2*COMBFILTER_MAXPERIOD];
+      if (x==y) {
+         yptr = mem_buf+COMBFILTER_MAXPERIOD;
+      } else {
+         for (i=0;i<N2;i++) buf[i] = y[2*i+s];
+         yptr = buf;
+      }
+      comb_filter(yptr, mem_buf+COMBFILTER_MAXPERIOD, T0, T1, N2, g0, g1, tapset0, tapset1, new_window, overlap2, arch);
+      for (i=0;i<N2;i++) y[2*i+s] = yptr[i];
+   }
+   RESTORE_STACK;
+   return;
+}
+#endif
+
+#ifndef OVERRIDE_comb_filter
+void comb_filter(opus_val32 *y, opus_val32 *x, int T0, int T1, int N,
+      opus_val16 g0, opus_val16 g1, int tapset0, int tapset1,
+      const celt_coef *window, int overlap, int arch)
+{
+   int i;
+   /* printf ("%d %d %f %f\n", T0, T1, g0, g1); */
+   celt_coef g00, g01, g02, g10, g11, g12;
+   opus_val32 x0, x1, x2, x3, x4;
+   static const opus_val16 gains[3][3] = {
+         {QCONST16(0.3066406250f, 15), QCONST16(0.2170410156f, 15), QCONST16(0.1296386719f, 15)},
+         {QCONST16(0.4638671875f, 15), QCONST16(0.2680664062f, 15), QCONST16(0.f, 15)},
+         {QCONST16(0.7998046875f, 15), QCONST16(0.1000976562f, 15), QCONST16(0.f, 15)}};
+#ifdef ENABLE_QEXT
+   if (overlap==240) {
+      comb_filter_qext(y, x, T0, T1, N, g0, g1, tapset0, tapset1, window, overlap, arch);
+      return;
+   }
+#endif
+   if (g0==0 && g1==0)
+   {
+      /* OPT: Happens to work without the OPUS_MOVE(), but only because the current encoder already copies x to y */
+      if (x!=y)
+         OPUS_MOVE(y, x, N);
+      return;
+   }
+   /* When the gain is zero, T0 and/or T1 is set to zero. We need
+      to have then be at least 2 to avoid processing garbage data. */
+   T0 = IMAX(T0, COMBFILTER_MINPERIOD);
+   T1 = IMAX(T1, COMBFILTER_MINPERIOD);
+   g00 = MULT_COEF_TAPS(g0, gains[tapset0][0]);
+   g01 = MULT_COEF_TAPS(g0, gains[tapset0][1]);
+   g02 = MULT_COEF_TAPS(g0, gains[tapset0][2]);
+   g10 = MULT_COEF_TAPS(g1, gains[tapset1][0]);
+   g11 = MULT_COEF_TAPS(g1, gains[tapset1][1]);
+   g12 = MULT_COEF_TAPS(g1, gains[tapset1][2]);
+   x1 = x[-T1+1];
+   x2 = x[-T1  ];
+   x3 = x[-T1-1];
+   x4 = x[-T1-2];
+   /* If the filter didn't change, we don't need the overlap */
+   if (g0==g1 && T0==T1 && tapset0==tapset1)
+      overlap=0;
+   for (i=0;i<overlap;i++)
+   {
+      celt_coef f;
+      x0=x[i-T1+2];
+      f = MULT_COEF(window[i],window[i]);
+      y[i] = x[i]
+               + MULT_COEF_32(MULT_COEF((COEF_ONE-f),g00),x[i-T0])
+               + MULT_COEF_32(MULT_COEF((COEF_ONE-f),g01),ADD32(x[i-T0+1],x[i-T0-1]))
+               + MULT_COEF_32(MULT_COEF((COEF_ONE-f),g02),ADD32(x[i-T0+2],x[i-T0-2]))
+               + MULT_COEF_32(MULT_COEF(f,g10),x2)
+               + MULT_COEF_32(MULT_COEF(f,g11),ADD32(x1,x3))
+               + MULT_COEF_32(MULT_COEF(f,g12),ADD32(x0,x4));
+#ifdef FIXED_POINT
+      /* A bit of bias seems to help here. */
+      y[i] = SUB32(y[i], 3);
+#endif
+      y[i] = SATURATE(y[i], SIG_SAT);
+      x4=x3;
+      x3=x2;
+      x2=x1;
+      x1=x0;
+
+   }
+   if (g1==0)
+   {
+      /* OPT: Happens to work without the OPUS_MOVE(), but only because the current encoder already copies x to y */
+      if (x!=y)
+         OPUS_MOVE(y+overlap, x+overlap, N-overlap);
+      return;
+   }
+
+   /* Compute the part with the constant filter. */
+   comb_filter_const(y+i, x+i, T1, N-i, g10, g11, g12, arch);
+}
+#endif /* OVERRIDE_comb_filter */
+
+/* TF change table. Positive values mean better frequency resolution (longer
+   effective window), whereas negative values mean better time resolution
+   (shorter effective window). The second index is computed as:
+   4*isTransient + 2*tf_select + per_band_flag */
+const signed char tf_select_table[4][8] = {
+    /*isTransient=0     isTransient=1 */
+      {0, -1, 0, -1,    0,-1, 0,-1}, /* 2.5 ms */
+      {0, -1, 0, -2,    1, 0, 1,-1}, /* 5 ms */
+      {0, -2, 0, -3,    2, 0, 1,-1}, /* 10 ms */
+      {0, -2, 0, -3,    3, 0, 1,-1}, /* 20 ms */
+};
+
+
+void init_caps(const CELTMode *m,int *cap,int LM,int C)
+{
+   int i;
+   for (i=0;i<m->nbEBands;i++)
+   {
+      int N;
+      N=(m->eBands[i+1]-m->eBands[i])<<LM;
+      cap[i] = (m->cache.caps[m->nbEBands*(2*LM+C-1)+i]+64)*C*N>>2;
+   }
+}
+
+
+
+const char *opus_strerror(int error)
+{
+   static const char * const error_strings[8] = {
+      "success",
+      "invalid argument",
+      "buffer too small",
+      "internal error",
+      "corrupted stream",
+      "request not implemented",
+      "invalid state",
+      "memory allocation failed"
+   };
+   if (error > 0 || error < -7)
+      return "unknown error";
+   else
+      return error_strings[-error];
+}
+
+const char *opus_get_version_string(void)
+{
+    return "libopus " PACKAGE_VERSION
+    /* Applications may rely on the presence of this substring in the version
+       string to determine if they have a fixed-point or floating-point build
+       at runtime. */
+#ifdef FIXED_POINT
+          "-fixed"
+#endif
+#ifdef FUZZING
+          "-fuzzing"
+#endif
+          ;
+}
--- a/lib/libopus/celt/celt.h
+++ b/lib/libopus/celt/celt.h
@ -0,0 +1,274 @@
+/* Copyright (c) 2007-2008 CSIRO
+   Copyright (c) 2007-2009 Xiph.Org Foundation
+   Copyright (c) 2008 Gregory Maxwell
+   Written by Jean-Marc Valin and Gregory Maxwell */
+/**
+  @file celt.h
+  @brief Contains all the functions for encoding and decoding audio
+ */
+
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef CELT_H
+#define CELT_H
+
+#include "opus_types.h"
+#include "opus_defines.h"
+#include "opus_custom.h"
+#include "entenc.h"
+#include "entdec.h"
+#include "arch.h"
+#include "kiss_fft.h"
+
+#ifdef ENABLE_QEXT
+#define ARG_QEXT(arg) , arg
+#else
+#define ARG_QEXT(arg)
+#endif
+
+
+#ifdef ENABLE_DEEP_PLC
+#include "lpcnet.h"
+#endif
+
+#define CELTEncoder OpusCustomEncoder
+#define CELTDecoder OpusCustomDecoder
+#define CELTMode OpusCustomMode
+
+#define QEXT_EXTENSION_ID 124
+
+#define LEAK_BANDS 19
+
+typedef struct {
+   int valid;
+   float tonality;
+   float tonality_slope;
+   float noisiness;
+   float activity;
+   float music_prob;
+   float music_prob_min;
+   float music_prob_max;
+   int   bandwidth;
+   float activity_probability;
+   float max_pitch_ratio;
+   /* Store as Q6 char to save space. */
+   unsigned char leak_boost[LEAK_BANDS];
+} AnalysisInfo;
+
+typedef struct {
+   int signalType;
+   int offset;
+} SILKInfo;
+
+#define celt_check_mode_ptr_ptr(ptr) ((ptr) + ((ptr) - (const CELTMode**)(ptr)))
+
+#define celt_check_analysis_ptr(ptr) ((ptr) + ((ptr) - (const AnalysisInfo*)(ptr)))
+
+#define celt_check_silkinfo_ptr(ptr) ((ptr) + ((ptr) - (const SILKInfo*)(ptr)))
+
+#define celt_check_glog_ptr(ptr) ((ptr) + ((ptr) - (celt_glog*)(ptr)))
+
+/* Encoder/decoder Requests */
+
+
+#define CELT_SET_PREDICTION_REQUEST    10002
+/** Controls the use of interframe prediction.
+    0=Independent frames
+    1=Short term interframe prediction allowed
+    2=Long term prediction allowed
+ */
+#define CELT_SET_PREDICTION(x) CELT_SET_PREDICTION_REQUEST, opus_check_int(x)
+
+#define CELT_SET_INPUT_CLIPPING_REQUEST    10004
+#define CELT_SET_INPUT_CLIPPING(x) CELT_SET_INPUT_CLIPPING_REQUEST, opus_check_int(x)
+
+#define CELT_GET_AND_CLEAR_ERROR_REQUEST   10007
+#define CELT_GET_AND_CLEAR_ERROR(x) CELT_GET_AND_CLEAR_ERROR_REQUEST, opus_check_int_ptr(x)
+
+#define CELT_SET_CHANNELS_REQUEST    10008
+#define CELT_SET_CHANNELS(x) CELT_SET_CHANNELS_REQUEST, opus_check_int(x)
+
+
+/* Internal */
+#define CELT_SET_START_BAND_REQUEST    10010
+#define CELT_SET_START_BAND(x) CELT_SET_START_BAND_REQUEST, opus_check_int(x)
+
+#define CELT_SET_END_BAND_REQUEST    10012
+#define CELT_SET_END_BAND(x) CELT_SET_END_BAND_REQUEST, opus_check_int(x)
+
+#define CELT_GET_MODE_REQUEST    10015
+/** Get the CELTMode used by an encoder or decoder */
+#define CELT_GET_MODE(x) CELT_GET_MODE_REQUEST, celt_check_mode_ptr_ptr(x)
+
+#define CELT_SET_SIGNALLING_REQUEST    10016
+#define CELT_SET_SIGNALLING(x) CELT_SET_SIGNALLING_REQUEST, opus_check_int(x)
+
+#define CELT_SET_TONALITY_REQUEST    10018
+#define CELT_SET_TONALITY(x) CELT_SET_TONALITY_REQUEST, opus_check_int(x)
+#define CELT_SET_TONALITY_SLOPE_REQUEST    10020
+#define CELT_SET_TONALITY_SLOPE(x) CELT_SET_TONALITY_SLOPE_REQUEST, opus_check_int(x)
+
+#define CELT_SET_ANALYSIS_REQUEST    10022
+#define CELT_SET_ANALYSIS(x) CELT_SET_ANALYSIS_REQUEST, celt_check_analysis_ptr(x)
+
+#define OPUS_SET_LFE_REQUEST    10024
+#define OPUS_SET_LFE(x) OPUS_SET_LFE_REQUEST, opus_check_int(x)
+
+#define OPUS_SET_ENERGY_MASK_REQUEST    10026
+#define OPUS_SET_ENERGY_MASK(x) OPUS_SET_ENERGY_MASK_REQUEST, celt_check_glog_ptr(x)
+
+#define CELT_SET_SILK_INFO_REQUEST    10028
+#define CELT_SET_SILK_INFO(x) CELT_SET_SILK_INFO_REQUEST, celt_check_silkinfo_ptr(x)
+
+
+static OPUS_INLINE opus_int32 bits_to_bitrate(opus_int32 bits, opus_int32 Fs, opus_int32 frame_size) {
+   return bits*(6*Fs/frame_size)/6;
+}
+
+static OPUS_INLINE opus_int32 bitrate_to_bits(opus_int32 bitrate, opus_int32 Fs, opus_int32 frame_size) {
+   return bitrate*6/(6*Fs/frame_size);
+}
+
+/* Encoder stuff */
+
+int celt_encoder_get_size(int channels);
+
+int celt_encode_with_ec(OpusCustomEncoder * OPUS_RESTRICT st, const opus_res * pcm, int frame_size, unsigned char *compressed, int nbCompressedBytes, ec_enc *enc);
+
+int celt_encoder_init(CELTEncoder *st, opus_int32 sampling_rate, int channels,
+                      int arch);
+
+
+
+/* Decoder stuff */
+
+int celt_decoder_get_size(int channels);
+
+
+int celt_decoder_init(CELTDecoder *st, opus_int32 sampling_rate, int channels);
+
+int celt_decode_with_ec_dred(CELTDecoder * OPUS_RESTRICT st, const unsigned char *data,
+      int len, opus_res * OPUS_RESTRICT pcm, int frame_size, ec_dec *dec, int accum
+#ifdef ENABLE_DEEP_PLC
+      ,LPCNetPLCState *lpcnet
+#endif
+      ARG_QEXT(const unsigned char *qext_payload) ARG_QEXT(int qext_payload_len)
+      );
+
+int celt_decode_with_ec(OpusCustomDecoder * OPUS_RESTRICT st, const unsigned char *data,
+      int len, opus_res * OPUS_RESTRICT pcm, int frame_size, ec_dec *dec, int accum);
+
+#define celt_encoder_ctl opus_custom_encoder_ctl
+#define celt_decoder_ctl opus_custom_decoder_ctl
+
+
+#if defined(CUSTOM_MODES) || defined(ENABLE_OPUS_CUSTOM_API)
+#define OPUS_CUSTOM_NOSTATIC
+#else
+#define OPUS_CUSTOM_NOSTATIC static OPUS_INLINE
+#endif
+
+static const unsigned char trim_icdf[11] = {126, 124, 119, 109, 87, 41, 19, 9, 4, 2, 0};
+/* Probs: NONE: 21.875%, LIGHT: 6.25%, NORMAL: 65.625%, AGGRESSIVE: 6.25% */
+static const unsigned char spread_icdf[4] = {25, 23, 2, 0};
+
+static const unsigned char tapset_icdf[3]={2,1,0};
+
+#if defined(CUSTOM_MODES) || defined(ENABLE_OPUS_CUSTOM_API)
+static const unsigned char toOpusTable[20] = {
+      0xE0, 0xE8, 0xF0, 0xF8,
+      0xC0, 0xC8, 0xD0, 0xD8,
+      0xA0, 0xA8, 0xB0, 0xB8,
+      0x00, 0x00, 0x00, 0x00,
+      0x80, 0x88, 0x90, 0x98,
+};
+
+static const unsigned char fromOpusTable[16] = {
+      0x80, 0x88, 0x90, 0x98,
+      0x40, 0x48, 0x50, 0x58,
+      0x20, 0x28, 0x30, 0x38,
+      0x00, 0x08, 0x10, 0x18
+};
+
+static OPUS_INLINE int toOpus(unsigned char c)
+{
+   int ret=0;
+   if (c<0xA0)
+      ret = toOpusTable[c>>3];
+   if (ret == 0)
+      return -1;
+   else
+      return ret|(c&0x7);
+}
+
+static OPUS_INLINE int fromOpus(unsigned char c)
+{
+   if (c<0x80)
+      return -1;
+   else
+      return fromOpusTable[(c>>3)-16] | (c&0x7);
+}
+#endif /* CUSTOM_MODES */
+
+#define COMBFILTER_MAXPERIOD 1024
+#define COMBFILTER_MINPERIOD 15
+
+extern const signed char tf_select_table[4][8];
+
+#if defined(ENABLE_HARDENING) || defined(ENABLE_ASSERTIONS)
+void validate_celt_decoder(CELTDecoder *st);
+#define VALIDATE_CELT_DECODER(st) validate_celt_decoder(st)
+#else
+#define VALIDATE_CELT_DECODER(st)
+#endif
+
+int resampling_factor(opus_int32 rate);
+
+void celt_preemphasis(const opus_res * OPUS_RESTRICT pcmp, celt_sig * OPUS_RESTRICT inp,
+                        int N, int CC, int upsample, const opus_val16 *coef, celt_sig *mem, int clip);
+
+void comb_filter(opus_val32 *y, opus_val32 *x, int T0, int T1, int N,
+      opus_val16 g0, opus_val16 g1, int tapset0, int tapset1,
+      const celt_coef *window, int overlap, int arch);
+
+void init_caps(const CELTMode *m,int *cap,int LM,int C);
+
+#ifdef RESYNTH
+void deemphasis(celt_sig *in[], opus_res *pcm, int N, int C, int downsample, const opus_val16 *coef, celt_sig *mem, int accum);
+void celt_synthesis(const CELTMode *mode, celt_norm *X, celt_sig * out_syn[],
+      celt_glog *oldBandE, int start, int effEnd, int C, int CC, int isTransient,
+      int LM, int downsample, int silence, int arch ARG_QEXT(const CELTMode *qext_mode) ARG_QEXT(const celt_glog *qext_bandLogE) ARG_QEXT(int qext_end));
+#endif
+
+#ifdef ENABLE_QEXT
+#define QEXT_SCALE(x) ((qext_scale)*(x))
+#define QEXT_SCALE2(x, qext_scale) ((qext_scale)*(x))
+#else
+#define QEXT_SCALE(x) (x)
+#define QEXT_SCALE2(x, qext_scale) (x)
+#endif
+
+#endif /* CELT_H */
--- a/lib/libopus/celt/celt_decoder.c
+++ b/lib/libopus/celt/celt_decoder.c
--- a/lib/libopus/celt/celt_encoder.c
+++ b/lib/libopus/celt/celt_encoder.c
--- a/lib/libopus/celt/celt_lpc.c
+++ b/lib/libopus/celt/celt_lpc.c
@ -0,0 +1,374 @@
+/* Copyright (c) 2009-2010 Xiph.Org Foundation
+   Written by Jean-Marc Valin */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "celt_lpc.h"
+#include "stack_alloc.h"
+#include "mathops.h"
+#include "pitch.h"
+
+void _celt_lpc(
+      opus_val16       *_lpc, /* out: [0...p-1] LPC coefficients      */
+const opus_val32 *ac,  /* in:  [0...p] autocorrelation values  */
+int          p
+)
+{
+   int i, j;
+   opus_val32 r;
+   opus_val32 error = ac[0];
+#ifdef FIXED_POINT
+   opus_val32 lpc[CELT_LPC_ORDER];
+#else
+   float *lpc = _lpc;
+#endif
+
+   OPUS_CLEAR(lpc, p);
+#ifdef FIXED_POINT
+   if (ac[0] != 0)
+#else
+   if (ac[0] > 1e-10f)
+#endif
+   {
+      for (i = 0; i < p; i++) {
+         /* Sum up this iteration's reflection coefficient */
+         opus_val32 rr = 0;
+#if defined (FIXED_POINT) && OPUS_FAST_INT64
+         opus_int64 acc = 0;
+         for (j = 0; j < i; j++)
+            acc += (opus_int64)(lpc[j]) * (opus_int64)(ac[i - j]);
+         rr = (opus_val32)SHR64(acc, 31);
+#else
+         for (j = 0; j < i; j++)
+            rr += MULT32_32_Q31(lpc[j],ac[i - j]);
+#endif
+         rr += SHR32(ac[i + 1],6);
+         r = -frac_div32(SHL32(rr,6), error);
+         /*  Update LPC coefficients and total error */
+         lpc[i] = SHR32(r,6);
+         for (j = 0; j < (i+1)>>1; j++)
+         {
+            opus_val32 tmp1, tmp2;
+            tmp1 = lpc[j];
+            tmp2 = lpc[i-1-j];
+            lpc[j]     = tmp1 + MULT32_32_Q31(r,tmp2);
+            lpc[i-1-j] = tmp2 + MULT32_32_Q31(r,tmp1);
+         }
+
+         error = error - MULT32_32_Q31(MULT32_32_Q31(r,r),error);
+         /* Bail out once we get 30 dB gain */
+#ifdef FIXED_POINT
+         if (error<=SHR32(ac[0],10))
+            break;
+#else
+         if (error<=.001f*ac[0])
+            break;
+#endif
+      }
+   }
+#ifdef FIXED_POINT
+   {
+      /* Convert the int32 lpcs to int16 and ensure there are no wrap-arounds.
+         This reuses the logic in silk_LPC_fit() and silk_bwexpander_32(). Any bug
+         fixes should also be applied there. */
+      int iter, idx = 0;
+      opus_val32 maxabs, absval, chirp_Q16, chirp_minus_one_Q16;
+
+      for (iter = 0; iter < 10; iter++) {
+         maxabs = 0;
+         for (i = 0; i < p; i++) {
+            absval = ABS32(lpc[i]);
+            if (absval > maxabs) {
+               maxabs = absval;
+               idx = i;
+            }
+         }
+         maxabs = PSHR32(maxabs, 13);  /* Q25->Q12 */
+
+         if (maxabs > 32767) {
+            maxabs = MIN32(maxabs, 163838);
+            chirp_Q16 = QCONST32(0.999, 16) - DIV32(SHL32(maxabs - 32767, 14),
+                                                    SHR32(MULT32_32_32(maxabs, idx + 1), 2));
+            chirp_minus_one_Q16 = chirp_Q16 - 65536;
+
+            /* Apply bandwidth expansion. */
+            for (i = 0; i < p - 1; i++) {
+               lpc[i] = MULT32_32_Q16(chirp_Q16, lpc[i]);
+               chirp_Q16 += PSHR32(MULT32_32_32(chirp_Q16, chirp_minus_one_Q16), 16);
+            }
+            lpc[p - 1] = MULT32_32_Q16(chirp_Q16, lpc[p - 1]);
+         } else {
+            break;
+         }
+      }
+
+      if (iter == 10) {
+         /* If the coeffs still do not fit into the 16 bit range after 10 iterations,
+            fall back to the A(z)=1 filter. */
+         OPUS_CLEAR(lpc, p);
+         _lpc[0] = 4096;  /* Q12 */
+      } else {
+         for (i = 0; i < p; i++) {
+            _lpc[i] = EXTRACT16(PSHR32(lpc[i], 13));  /* Q25->Q12 */
+         }
+      }
+   }
+#endif
+}
+
+
+void celt_fir_c(
+         const opus_val16 *x,
+         const opus_val16 *num,
+         opus_val16 *y,
+         int N,
+         int ord,
+         int arch)
+{
+   int i,j;
+   VARDECL(opus_val16, rnum);
+   SAVE_STACK;
+   celt_assert(x != y);
+   ALLOC(rnum, ord, opus_val16);
+   for(i=0;i<ord;i++)
+      rnum[i] = num[ord-i-1];
+   for (i=0;i<N-3;i+=4)
+   {
+      opus_val32 sum[4];
+      sum[0] = SHL32(EXTEND32(x[i  ]), SIG_SHIFT);
+      sum[1] = SHL32(EXTEND32(x[i+1]), SIG_SHIFT);
+      sum[2] = SHL32(EXTEND32(x[i+2]), SIG_SHIFT);
+      sum[3] = SHL32(EXTEND32(x[i+3]), SIG_SHIFT);
+#if defined(OPUS_CHECK_ASM) && defined(FIXED_POINT)
+      {
+         opus_val32 sum_c[4];
+         memcpy(sum_c, sum, sizeof(sum_c));
+         xcorr_kernel_c(rnum, x+i-ord, sum_c, ord);
+#endif
+         xcorr_kernel(rnum, x+i-ord, sum, ord, arch);
+#if defined(OPUS_CHECK_ASM) && defined(FIXED_POINT)
+         celt_assert(memcmp(sum, sum_c, sizeof(sum)) == 0);
+      }
+#endif
+      y[i  ] = SROUND16(sum[0], SIG_SHIFT);
+      y[i+1] = SROUND16(sum[1], SIG_SHIFT);
+      y[i+2] = SROUND16(sum[2], SIG_SHIFT);
+      y[i+3] = SROUND16(sum[3], SIG_SHIFT);
+   }
+   for (;i<N;i++)
+   {
+      opus_val32 sum = SHL32(EXTEND32(x[i]), SIG_SHIFT);
+      for (j=0;j<ord;j++)
+         sum = MAC16_16(sum,rnum[j],x[i+j-ord]);
+      y[i] = SROUND16(sum, SIG_SHIFT);
+   }
+   RESTORE_STACK;
+}
+
+void celt_iir(const opus_val32 *_x,
+         const opus_val16 *den,
+         opus_val32 *_y,
+         int N,
+         int ord,
+         opus_val16 *mem,
+         int arch)
+{
+#ifdef SMALL_FOOTPRINT
+   int i,j;
+   (void)arch;
+   for (i=0;i<N;i++)
+   {
+      opus_val32 sum = _x[i];
+      for (j=0;j<ord;j++)
+      {
+         sum -= MULT16_16(den[j],mem[j]);
+      }
+      for (j=ord-1;j>=1;j--)
+      {
+         mem[j]=mem[j-1];
+      }
+      mem[0] = SROUND16(sum, SIG_SHIFT);
+      _y[i] = sum;
+   }
+#else
+   int i,j;
+   VARDECL(opus_val16, rden);
+   VARDECL(opus_val16, y);
+   SAVE_STACK;
+
+   celt_assert((ord&3)==0);
+   ALLOC(rden, ord, opus_val16);
+   ALLOC(y, N+ord, opus_val16);
+   for(i=0;i<ord;i++)
+      rden[i] = den[ord-i-1];
+   for(i=0;i<ord;i++)
+      y[i] = -mem[ord-i-1];
+   for(;i<N+ord;i++)
+      y[i]=0;
+   for (i=0;i<N-3;i+=4)
+   {
+      /* Unroll by 4 as if it were an FIR filter */
+      opus_val32 sum[4];
+      sum[0]=_x[i];
+      sum[1]=_x[i+1];
+      sum[2]=_x[i+2];
+      sum[3]=_x[i+3];
+#if defined(OPUS_CHECK_ASM) && defined(FIXED_POINT)
+      {
+         opus_val32 sum_c[4];
+         memcpy(sum_c, sum, sizeof(sum_c));
+         xcorr_kernel_c(rden, y+i, sum_c, ord);
+#endif
+         xcorr_kernel(rden, y+i, sum, ord, arch);
+#if defined(OPUS_CHECK_ASM) && defined(FIXED_POINT)
+         celt_assert(memcmp(sum, sum_c, sizeof(sum)) == 0);
+      }
+#endif
+      /* Patch up the result to compensate for the fact that this is an IIR */
+      y[i+ord  ] = -SROUND16(sum[0],SIG_SHIFT);
+      _y[i  ] = sum[0];
+      sum[1] = MAC16_16(sum[1], y[i+ord  ], den[0]);
+      y[i+ord+1] = -SROUND16(sum[1],SIG_SHIFT);
+      _y[i+1] = sum[1];
+      sum[2] = MAC16_16(sum[2], y[i+ord+1], den[0]);
+      sum[2] = MAC16_16(sum[2], y[i+ord  ], den[1]);
+      y[i+ord+2] = -SROUND16(sum[2],SIG_SHIFT);
+      _y[i+2] = sum[2];
+
+      sum[3] = MAC16_16(sum[3], y[i+ord+2], den[0]);
+      sum[3] = MAC16_16(sum[3], y[i+ord+1], den[1]);
+      sum[3] = MAC16_16(sum[3], y[i+ord  ], den[2]);
+      y[i+ord+3] = -SROUND16(sum[3],SIG_SHIFT);
+      _y[i+3] = sum[3];
+   }
+   for (;i<N;i++)
+   {
+      opus_val32 sum = _x[i];
+      for (j=0;j<ord;j++)
+         sum -= MULT16_16(rden[j],y[i+j]);
+      y[i+ord] = SROUND16(sum,SIG_SHIFT);
+      _y[i] = sum;
+   }
+   for(i=0;i<ord;i++)
+      mem[i] = _y[N-i-1];
+   RESTORE_STACK;
+#endif
+}
+
+int _celt_autocorr(
+                   const opus_val16 *x,   /*  in: [0...n-1] samples x   */
+                   opus_val32       *ac,  /* out: [0...lag-1] ac values */
+                   const celt_coef  *window,
+                   int          overlap,
+                   int          lag,
+                   int          n,
+                   int          arch
+                  )
+{
+   opus_val32 d;
+   int i, k;
+   int fastN=n-lag;
+   int shift;
+   const opus_val16 *xptr;
+   VARDECL(opus_val16, xx);
+   SAVE_STACK;
+   ALLOC(xx, n, opus_val16);
+   celt_assert(n>0);
+   celt_assert(overlap>=0);
+   if (overlap == 0)
+   {
+      xptr = x;
+   } else {
+      for (i=0;i<n;i++)
+         xx[i] = x[i];
+      for (i=0;i<overlap;i++)
+      {
+         opus_val16 w = COEF2VAL16(window[i]);
+         xx[i] = MULT16_16_Q15(x[i],w);
+         xx[n-i-1] = MULT16_16_Q15(x[n-i-1],w);
+      }
+      xptr = xx;
+   }
+   shift=0;
+#ifdef FIXED_POINT
+   {
+      opus_val32 ac0;
+      int ac0_shift = celt_ilog2(n + (n>>4));
+      ac0 = 1+(n<<7);
+      if (n&1) ac0 += SHR32(MULT16_16(xptr[0],xptr[0]),ac0_shift);
+      for(i=(n&1);i<n;i+=2)
+      {
+         ac0 += SHR32(MULT16_16(xptr[i],xptr[i]),ac0_shift);
+         ac0 += SHR32(MULT16_16(xptr[i+1],xptr[i+1]),ac0_shift);
+      }
+      /* Consider the effect of rounding-to-nearest when scaling the signal. */
+      ac0 += SHR32(ac0,7);
+
+      shift = celt_ilog2(ac0)-30+ac0_shift+1;
+      shift = (shift)/2;
+      if (shift>0)
+      {
+         for(i=0;i<n;i++)
+            xx[i] = PSHR32(xptr[i], shift);
+         xptr = xx;
+      } else
+         shift = 0;
+   }
+#endif
+   celt_pitch_xcorr(xptr, xptr, ac, fastN, lag+1, arch);
+   for (k=0;k<=lag;k++)
+   {
+      for (i = k+fastN, d = 0; i < n; i++)
+         d = MAC16_16(d, xptr[i], xptr[i-k]);
+      ac[k] += d;
+   }
+#ifdef FIXED_POINT
+   shift = 2*shift;
+   if (shift<=0)
+      ac[0] += SHL32((opus_int32)1, -shift);
+   if (ac[0] < 268435456)
+   {
+      int shift2 = 29 - EC_ILOG(ac[0]);
+      for (i=0;i<=lag;i++)
+         ac[i] = SHL32(ac[i], shift2);
+      shift -= shift2;
+   } else if (ac[0] >= 536870912)
+   {
+      int shift2=1;
+      if (ac[0] >= 1073741824)
+         shift2++;
+      for (i=0;i<=lag;i++)
+         ac[i] = SHR32(ac[i], shift2);
+      shift += shift2;
+   }
+#endif
+
+   RESTORE_STACK;
+   return shift;
+}
--- a/lib/libopus/celt/celt_lpc.h
+++ b/lib/libopus/celt/celt_lpc.h
@ -0,0 +1,66 @@
+/* Copyright (c) 2009-2010 Xiph.Org Foundation
+   Written by Jean-Marc Valin */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef CELT_LPC_H
+#define CELT_LPC_H
+
+#include "arch.h"
+#include "cpu_support.h"
+
+#if defined(OPUS_X86_MAY_HAVE_SSE4_1)
+#include "x86/celt_lpc_sse.h"
+#endif
+
+#define CELT_LPC_ORDER 24
+
+void _celt_lpc(opus_val16 *_lpc, const opus_val32 *ac, int p);
+
+void celt_fir_c(
+         const opus_val16 *x,
+         const opus_val16 *num,
+         opus_val16 *y,
+         int N,
+         int ord,
+         int arch);
+
+#if !defined(OVERRIDE_CELT_FIR)
+#define celt_fir(x, num, y, N, ord, arch) \
+    (celt_fir_c(x, num, y, N, ord, arch))
+#endif
+
+void celt_iir(const opus_val32 *x,
+         const opus_val16 *den,
+         opus_val32 *y,
+         int N,
+         int ord,
+         opus_val16 *mem,
+         int arch);
+
+int _celt_autocorr(const opus_val16 *x, opus_val32 *ac,
+         const celt_coef *window, int overlap, int lag, int n, int arch);
+
+#endif /* CELT_LPC_H */
--- a/lib/libopus/celt/cpu_support.h
+++ b/lib/libopus/celt/cpu_support.h
@ -0,0 +1,72 @@
+/* Copyright (c) 2010 Xiph.Org Foundation
+ * Copyright (c) 2013 Parrot */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef CPU_SUPPORT_H
+#define CPU_SUPPORT_H
+
+#include "opus_types.h"
+#include "opus_defines.h"
+
+#if defined(OPUS_HAVE_RTCD) && \
+  (defined(OPUS_ARM_ASM) || defined(OPUS_ARM_MAY_HAVE_NEON_INTR))
+#include "arm/armcpu.h"
+
+/* We currently support 5 ARM variants:
+ * arch[0] -> ARMv4
+ * arch[1] -> ARMv5E
+ * arch[2] -> ARMv6
+ * arch[3] -> NEON
+ * arch[4] -> NEON+DOTPROD
+ */
+#define OPUS_ARCHMASK 7
+
+#elif defined(OPUS_HAVE_RTCD) && \
+  ((defined(OPUS_X86_MAY_HAVE_SSE) && !defined(OPUS_X86_PRESUME_SSE)) || \
+  (defined(OPUS_X86_MAY_HAVE_SSE2) && !defined(OPUS_X86_PRESUME_SSE2)) || \
+  (defined(OPUS_X86_MAY_HAVE_SSE4_1) && !defined(OPUS_X86_PRESUME_SSE4_1)) || \
+  (defined(OPUS_X86_MAY_HAVE_AVX2) && !defined(OPUS_X86_PRESUME_AVX2)))
+
+#include "x86/x86cpu.h"
+/* We currently support 5 x86 variants:
+ * arch[0] -> non-sse
+ * arch[1] -> sse
+ * arch[2] -> sse2
+ * arch[3] -> sse4.1
+ * arch[4] -> avx
+ */
+#define OPUS_ARCHMASK 7
+int opus_select_arch(void);
+
+#else
+#define OPUS_ARCHMASK 0
+
+static OPUS_INLINE int opus_select_arch(void)
+{
+  return 0;
+}
+#endif
+#endif
--- a/lib/libopus/celt/cwrs.c
+++ b/lib/libopus/celt/cwrs.c
@ -0,0 +1,719 @@
+/* Copyright (c) 2007-2008 CSIRO
+   Copyright (c) 2007-2009 Xiph.Org Foundation
+   Copyright (c) 2007-2009 Timothy B. Terriberry
+   Written by Timothy B. Terriberry and Jean-Marc Valin */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "os_support.h"
+#include "cwrs.h"
+#include "mathops.h"
+#include "arch.h"
+
+#if defined(CUSTOM_MODES) || defined(ENABLE_QEXT)
+#define CWRS_EXTRA_ROWS
+#endif
+
+#if defined(CUSTOM_MODES)
+
+/*Guaranteed to return a conservatively large estimate of the binary logarithm
+   with frac bits of fractional precision.
+  Tested for all possible 32-bit inputs with frac=4, where the maximum
+   overestimation is 0.06254243 bits.*/
+int log2_frac(opus_uint32 val, int frac)
+{
+  int l;
+  l=EC_ILOG(val);
+  if(val&(val-1)){
+    /*This is (val>>l-16), but guaranteed to round up, even if adding a bias
+       before the shift would cause overflow (e.g., for 0xFFFFxxxx).
+       Doesn't work for val=0, but that case fails the test above.*/
+    if(l>16)val=((val-1)>>(l-16))+1;
+    else val<<=16-l;
+    l=(l-1)<<frac;
+    /*Note that we always need one iteration, since the rounding up above means
+       that we might need to adjust the integer part of the logarithm.*/
+    do{
+      int b;
+      b=(int)(val>>16);
+      l+=b<<frac;
+      val=(val+b)>>b;
+      val=(val*val+0x7FFF)>>15;
+    }
+    while(frac-->0);
+    /*If val is not exactly 0x8000, then we have to round up the remainder.*/
+    return l+(val>0x8000);
+  }
+  /*Exact powers of two require no rounding.*/
+  else return (l-1)<<frac;
+}
+#endif
+
+/*Although derived separately, the pulse vector coding scheme is equivalent to
+   a Pyramid Vector Quantizer \cite{Fis86}.
+  Some additional notes about an early version appear at
+   https://people.xiph.org/~tterribe/notes/cwrs.html, but the codebook ordering
+   and the definitions of some terms have evolved since that was written.
+
+  The conversion from a pulse vector to an integer index (encoding) and back
+   (decoding) is governed by two related functions, V(N,K) and U(N,K).
+
+  V(N,K) = the number of combinations, with replacement, of N items, taken K
+   at a time, when a sign bit is added to each item taken at least once (i.e.,
+   the number of N-dimensional unit pulse vectors with K pulses).
+  One way to compute this is via
+    V(N,K) = K>0 ? sum(k=1...K,2**k*choose(N,k)*choose(K-1,k-1)) : 1,
+   where choose() is the binomial function.
+  A table of values for N<10 and K<10 looks like:
+  V[10][10] = {
+    {1,  0,   0,    0,    0,     0,     0,      0,      0,       0},
+    {1,  2,   2,    2,    2,     2,     2,      2,      2,       2},
+    {1,  4,   8,   12,   16,    20,    24,     28,     32,      36},
+    {1,  6,  18,   38,   66,   102,   146,    198,    258,     326},
+    {1,  8,  32,   88,  192,   360,   608,    952,   1408,    1992},
+    {1, 10,  50,  170,  450,  1002,  1970,   3530,   5890,    9290},
+    {1, 12,  72,  292,  912,  2364,  5336,  10836,  20256,   35436},
+    {1, 14,  98,  462, 1666,  4942, 12642,  28814,  59906,  115598},
+    {1, 16, 128,  688, 2816,  9424, 27008,  68464, 157184,  332688},
+    {1, 18, 162,  978, 4482, 16722, 53154, 148626, 374274,  864146}
+  };
+
+  U(N,K) = the number of such combinations wherein N-1 objects are taken at
+   most K-1 at a time.
+  This is given by
+    U(N,K) = sum(k=0...K-1,V(N-1,k))
+           = K>0 ? (V(N-1,K-1) + V(N,K-1))/2 : 0.
+  The latter expression also makes clear that U(N,K) is half the number of such
+   combinations wherein the first object is taken at least once.
+  Although it may not be clear from either of these definitions, U(N,K) is the
+   natural function to work with when enumerating the pulse vector codebooks,
+   not V(N,K).
+  U(N,K) is not well-defined for N=0, but with the extension
+    U(0,K) = K>0 ? 0 : 1,
+   the function becomes symmetric: U(N,K) = U(K,N), with a similar table:
+  U[10][10] = {
+    {1, 0,  0,   0,    0,    0,     0,     0,      0,      0},
+    {0, 1,  1,   1,    1,    1,     1,     1,      1,      1},
+    {0, 1,  3,   5,    7,    9,    11,    13,     15,     17},
+    {0, 1,  5,  13,   25,   41,    61,    85,    113,    145},
+    {0, 1,  7,  25,   63,  129,   231,   377,    575,    833},
+    {0, 1,  9,  41,  129,  321,   681,  1289,   2241,   3649},
+    {0, 1, 11,  61,  231,  681,  1683,  3653,   7183,  13073},
+    {0, 1, 13,  85,  377, 1289,  3653,  8989,  19825,  40081},
+    {0, 1, 15, 113,  575, 2241,  7183, 19825,  48639, 108545},
+    {0, 1, 17, 145,  833, 3649, 13073, 40081, 108545, 265729}
+  };
+
+  With this extension, V(N,K) may be written in terms of U(N,K):
+    V(N,K) = U(N,K) + U(N,K+1)
+   for all N>=0, K>=0.
+  Thus U(N,K+1) represents the number of combinations where the first element
+   is positive or zero, and U(N,K) represents the number of combinations where
+   it is negative.
+  With a large enough table of U(N,K) values, we could write O(N) encoding
+   and O(min(N*log(K),N+K)) decoding routines, but such a table would be
+   prohibitively large for small embedded devices (K may be as large as 32767
+   for small N, and N may be as large as 200).
+
+  Both functions obey the same recurrence relation:
+    V(N,K) = V(N-1,K) + V(N,K-1) + V(N-1,K-1),
+    U(N,K) = U(N-1,K) + U(N,K-1) + U(N-1,K-1),
+   for all N>0, K>0, with different initial conditions at N=0 or K=0.
+  This allows us to construct a row of one of the tables above given the
+   previous row or the next row.
+  Thus we can derive O(NK) encoding and decoding routines with O(K) memory
+   using only addition and subtraction.
+
+  When encoding, we build up from the U(2,K) row and work our way forwards.
+  When decoding, we need to start at the U(N,K) row and work our way backwards,
+   which requires a means of computing U(N,K).
+  U(N,K) may be computed from two previous values with the same N:
+    U(N,K) = ((2*N-1)*U(N,K-1) - U(N,K-2))/(K-1) + U(N,K-2)
+   for all N>1, and since U(N,K) is symmetric, a similar relation holds for two
+   previous values with the same K:
+    U(N,K>1) = ((2*K-1)*U(N-1,K) - U(N-2,K))/(N-1) + U(N-2,K)
+   for all K>1.
+  This allows us to construct an arbitrary row of the U(N,K) table by starting
+   with the first two values, which are constants.
+  This saves roughly 2/3 the work in our O(NK) decoding routine, but costs O(K)
+   multiplications.
+  Similar relations can be derived for V(N,K), but are not used here.
+
+  For N>0 and K>0, U(N,K) and V(N,K) take on the form of an (N-1)-degree
+   polynomial for fixed N.
+  The first few are
+    U(1,K) = 1,
+    U(2,K) = 2*K-1,
+    U(3,K) = (2*K-2)*K+1,
+    U(4,K) = (((4*K-6)*K+8)*K-3)/3,
+    U(5,K) = ((((2*K-4)*K+10)*K-8)*K+3)/3,
+   and
+    V(1,K) = 2,
+    V(2,K) = 4*K,
+    V(3,K) = 4*K*K+2,
+    V(4,K) = 8*(K*K+2)*K/3,
+    V(5,K) = ((4*K*K+20)*K*K+6)/3,
+   for all K>0.
+  This allows us to derive O(N) encoding and O(N*log(K)) decoding routines for
+   small N (and indeed decoding is also O(N) for N<3).
+
+  @ARTICLE{Fis86,
+    author="Thomas R. Fischer",
+    title="A Pyramid Vector Quantizer",
+    journal="IEEE Transactions on Information Theory",
+    volume="IT-32",
+    number=4,
+    pages="568--583",
+    month=Jul,
+    year=1986
+  }*/
+
+#if !defined(SMALL_FOOTPRINT)
+
+/*U(N,K) = U(K,N) := N>0?K>0?U(N-1,K)+U(N,K-1)+U(N-1,K-1):0:K>0?1:0*/
+# define CELT_PVQ_U(_n,_k) (CELT_PVQ_U_ROW[IMIN(_n,_k)][IMAX(_n,_k)])
+/*V(N,K) := U(N,K)+U(N,K+1) = the number of PVQ codewords for a band of size N
+   with K pulses allocated to it.*/
+# define CELT_PVQ_V(_n,_k) (CELT_PVQ_U(_n,_k)+CELT_PVQ_U(_n,(_k)+1))
+
+/*For each V(N,K) supported, we will access element U(min(N,K+1),max(N,K+1)).
+  Thus, the number of entries in row I is the larger of the maximum number of
+   pulses we will ever allocate for a given N=I (K=128, or however many fit in
+   32 bits, whichever is smaller), plus one, and the maximum N for which
+   K=I-1 pulses fit in 32 bits.
+  The largest band size in an Opus Custom mode is 208.
+  Otherwise, we can limit things to the set of N which can be achieved by
+   splitting a band from a standard Opus mode: 176, 144, 96, 88, 72, 64, 48,
+   44, 36, 32, 24, 22, 18, 16, 8, 4, 2).*/
+#if defined(CWRS_EXTRA_ROWS)
+static const opus_uint32 CELT_PVQ_U_DATA[1488]={
+#else
+static const opus_uint32 CELT_PVQ_U_DATA[1272]={
+#endif
+  /*N=0, K=0...176:*/
+  1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+#if defined(CWRS_EXTRA_ROWS)
+  /*...208:*/
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0,
+#endif
+  /*N=1, K=1...176:*/
+  1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+  1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+  1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+  1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+  1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+  1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+  1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+#if defined(CWRS_EXTRA_ROWS)
+  /*...208:*/
+  1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+  1, 1, 1, 1, 1, 1,
+#endif
+  /*N=2, K=2...176:*/
+  3, 5, 7, 9, 11, 13, 15, 17, 19, 21, 23, 25, 27, 29, 31, 33, 35, 37, 39, 41,
+  43, 45, 47, 49, 51, 53, 55, 57, 59, 61, 63, 65, 67, 69, 71, 73, 75, 77, 79,
+  81, 83, 85, 87, 89, 91, 93, 95, 97, 99, 101, 103, 105, 107, 109, 111, 113,
+  115, 117, 119, 121, 123, 125, 127, 129, 131, 133, 135, 137, 139, 141, 143,
+  145, 147, 149, 151, 153, 155, 157, 159, 161, 163, 165, 167, 169, 171, 173,
+  175, 177, 179, 181, 183, 185, 187, 189, 191, 193, 195, 197, 199, 201, 203,
+  205, 207, 209, 211, 213, 215, 217, 219, 221, 223, 225, 227, 229, 231, 233,
+  235, 237, 239, 241, 243, 245, 247, 249, 251, 253, 255, 257, 259, 261, 263,
+  265, 267, 269, 271, 273, 275, 277, 279, 281, 283, 285, 287, 289, 291, 293,
+  295, 297, 299, 301, 303, 305, 307, 309, 311, 313, 315, 317, 319, 321, 323,
+  325, 327, 329, 331, 333, 335, 337, 339, 341, 343, 345, 347, 349, 351,
+#if defined(CWRS_EXTRA_ROWS)
+  /*...208:*/
+  353, 355, 357, 359, 361, 363, 365, 367, 369, 371, 373, 375, 377, 379, 381,
+  383, 385, 387, 389, 391, 393, 395, 397, 399, 401, 403, 405, 407, 409, 411,
+  413, 415,
+#endif
+  /*N=3, K=3...176:*/
+  13, 25, 41, 61, 85, 113, 145, 181, 221, 265, 313, 365, 421, 481, 545, 613,
+  685, 761, 841, 925, 1013, 1105, 1201, 1301, 1405, 1513, 1625, 1741, 1861,
+  1985, 2113, 2245, 2381, 2521, 2665, 2813, 2965, 3121, 3281, 3445, 3613, 3785,
+  3961, 4141, 4325, 4513, 4705, 4901, 5101, 5305, 5513, 5725, 5941, 6161, 6385,
+  6613, 6845, 7081, 7321, 7565, 7813, 8065, 8321, 8581, 8845, 9113, 9385, 9661,
+  9941, 10225, 10513, 10805, 11101, 11401, 11705, 12013, 12325, 12641, 12961,
+  13285, 13613, 13945, 14281, 14621, 14965, 15313, 15665, 16021, 16381, 16745,
+  17113, 17485, 17861, 18241, 18625, 19013, 19405, 19801, 20201, 20605, 21013,
+  21425, 21841, 22261, 22685, 23113, 23545, 23981, 24421, 24865, 25313, 25765,
+  26221, 26681, 27145, 27613, 28085, 28561, 29041, 29525, 30013, 30505, 31001,
+  31501, 32005, 32513, 33025, 33541, 34061, 34585, 35113, 35645, 36181, 36721,
+  37265, 37813, 38365, 38921, 39481, 40045, 40613, 41185, 41761, 42341, 42925,
+  43513, 44105, 44701, 45301, 45905, 46513, 47125, 47741, 48361, 48985, 49613,
+  50245, 50881, 51521, 52165, 52813, 53465, 54121, 54781, 55445, 56113, 56785,
+  57461, 58141, 58825, 59513, 60205, 60901, 61601,
+#if defined(CWRS_EXTRA_ROWS)
+  /*...208:*/
+  62305, 63013, 63725, 64441, 65161, 65885, 66613, 67345, 68081, 68821, 69565,
+  70313, 71065, 71821, 72581, 73345, 74113, 74885, 75661, 76441, 77225, 78013,
+  78805, 79601, 80401, 81205, 82013, 82825, 83641, 84461, 85285, 86113,
+#endif
+  /*N=4, K=4...176:*/
+  63, 129, 231, 377, 575, 833, 1159, 1561, 2047, 2625, 3303, 4089, 4991, 6017,
+  7175, 8473, 9919, 11521, 13287, 15225, 17343, 19649, 22151, 24857, 27775,
+  30913, 34279, 37881, 41727, 45825, 50183, 54809, 59711, 64897, 70375, 76153,
+  82239, 88641, 95367, 102425, 109823, 117569, 125671, 134137, 142975, 152193,
+  161799, 171801, 182207, 193025, 204263, 215929, 228031, 240577, 253575,
+  267033, 280959, 295361, 310247, 325625, 341503, 357889, 374791, 392217,
+  410175, 428673, 447719, 467321, 487487, 508225, 529543, 551449, 573951,
+  597057, 620775, 645113, 670079, 695681, 721927, 748825, 776383, 804609,
+  833511, 863097, 893375, 924353, 956039, 988441, 1021567, 1055425, 1090023,
+  1125369, 1161471, 1198337, 1235975, 1274393, 1313599, 1353601, 1394407,
+  1436025, 1478463, 1521729, 1565831, 1610777, 1656575, 1703233, 1750759,
+  1799161, 1848447, 1898625, 1949703, 2001689, 2054591, 2108417, 2163175,
+  2218873, 2275519, 2333121, 2391687, 2451225, 2511743, 2573249, 2635751,
+  2699257, 2763775, 2829313, 2895879, 2963481, 3032127, 3101825, 3172583,
+  3244409, 3317311, 3391297, 3466375, 3542553, 3619839, 3698241, 3777767,
+  3858425, 3940223, 4023169, 4107271, 4192537, 4278975, 4366593, 4455399,
+  4545401, 4636607, 4729025, 4822663, 4917529, 5013631, 5110977, 5209575,
+  5309433, 5410559, 5512961, 5616647, 5721625, 5827903, 5935489, 6044391,
+  6154617, 6266175, 6379073, 6493319, 6608921, 6725887, 6844225, 6963943,
+  7085049, 7207551,
+#if defined(CWRS_EXTRA_ROWS)
+  /*...208:*/
+  7331457, 7456775, 7583513, 7711679, 7841281, 7972327, 8104825, 8238783,
+  8374209, 8511111, 8649497, 8789375, 8930753, 9073639, 9218041, 9363967,
+  9511425, 9660423, 9810969, 9963071, 10116737, 10271975, 10428793, 10587199,
+  10747201, 10908807, 11072025, 11236863, 11403329, 11571431, 11741177,
+  11912575,
+#endif
+  /*N=5, K=5...176:*/
+  321, 681, 1289, 2241, 3649, 5641, 8361, 11969, 16641, 22569, 29961, 39041,
+  50049, 63241, 78889, 97281, 118721, 143529, 172041, 204609, 241601, 283401,
+  330409, 383041, 441729, 506921, 579081, 658689, 746241, 842249, 947241,
+  1061761, 1186369, 1321641, 1468169, 1626561, 1797441, 1981449, 2179241,
+  2391489, 2618881, 2862121, 3121929, 3399041, 3694209, 4008201, 4341801,
+  4695809, 5071041, 5468329, 5888521, 6332481, 6801089, 7295241, 7815849,
+  8363841, 8940161, 9545769, 10181641, 10848769, 11548161, 12280841, 13047849,
+  13850241, 14689089, 15565481, 16480521, 17435329, 18431041, 19468809,
+  20549801, 21675201, 22846209, 24064041, 25329929, 26645121, 28010881,
+  29428489, 30899241, 32424449, 34005441, 35643561, 37340169, 39096641,
+  40914369, 42794761, 44739241, 46749249, 48826241, 50971689, 53187081,
+  55473921, 57833729, 60268041, 62778409, 65366401, 68033601, 70781609,
+  73612041, 76526529, 79526721, 82614281, 85790889, 89058241, 92418049,
+  95872041, 99421961, 103069569, 106816641, 110664969, 114616361, 118672641,
+  122835649, 127107241, 131489289, 135983681, 140592321, 145317129, 150160041,
+  155123009, 160208001, 165417001, 170752009, 176215041, 181808129, 187533321,
+  193392681, 199388289, 205522241, 211796649, 218213641, 224775361, 231483969,
+  238341641, 245350569, 252512961, 259831041, 267307049, 274943241, 282741889,
+  290705281, 298835721, 307135529, 315607041, 324252609, 333074601, 342075401,
+  351257409, 360623041, 370174729, 379914921, 389846081, 399970689, 410291241,
+  420810249, 431530241, 442453761, 453583369, 464921641, 476471169, 488234561,
+  500214441, 512413449, 524834241, 537479489, 550351881, 563454121, 576788929,
+  590359041, 604167209, 618216201, 632508801,
+#if defined(CWRS_EXTRA_ROWS)
+  /*...208:*/
+  647047809, 661836041, 676876329, 692171521, 707724481, 723538089, 739615241,
+  755958849, 772571841, 789457161, 806617769, 824056641, 841776769, 859781161,
+  878072841, 896654849, 915530241, 934702089, 954173481, 973947521, 994027329,
+  1014416041, 1035116809, 1056132801, 1077467201, 1099123209, 1121104041,
+  1143412929, 1166053121, 1189027881, 1212340489, 1235994241,
+#endif
+  /*N=6, K=6...96:*/
+  1683, 3653, 7183, 13073, 22363, 36365, 56695, 85305, 124515, 177045, 246047,
+  335137, 448427, 590557, 766727, 982729, 1244979, 1560549, 1937199, 2383409,
+  2908411, 3522221, 4235671, 5060441, 6009091, 7095093, 8332863, 9737793,
+  11326283, 13115773, 15124775, 17372905, 19880915, 22670725, 25765455,
+  29189457, 32968347, 37129037, 41699767, 46710137, 52191139, 58175189,
+  64696159, 71789409, 79491819, 87841821, 96879431, 106646281, 117185651,
+  128542501, 140763503, 153897073, 167993403, 183104493, 199284183, 216588185,
+  235074115, 254801525, 275831935, 298228865, 322057867, 347386557, 374284647,
+  402823977, 433078547, 465124549, 499040399, 534906769, 572806619, 612825229,
+  655050231, 699571641, 746481891, 795875861, 847850911, 902506913, 959946283,
+  1020274013, 1083597703, 1150027593, 1219676595, 1292660325, 1369097135,
+  1449108145, 1532817275, 1620351277, 1711839767, 1807415257, 1907213187,
+  2011371957, 2120032959,
+#if defined(CWRS_EXTRA_ROWS)
+  /*...109:*/
+  2233340609U, 2351442379U, 2474488829U, 2602633639U, 2736033641U, 2874848851U,
+  3019242501U, 3169381071U, 3325434321U, 3487575323U, 3655980493U, 3830829623U,
+  4012305913U,
+#endif
+  /*N=7, K=7...54*/
+  8989, 19825, 40081, 75517, 134245, 227305, 369305, 579125, 880685, 1303777,
+  1884961, 2668525, 3707509, 5064793, 6814249, 9041957, 11847485, 15345233,
+  19665841, 24957661, 31388293, 39146185, 48442297, 59511829, 72616013,
+  88043969, 106114625, 127178701, 151620757, 179861305, 212358985, 249612805,
+  292164445, 340600625, 395555537, 457713341, 527810725, 606639529, 695049433,
+  793950709, 904317037, 1027188385, 1163673953, 1314955181, 1482288821,
+  1667010073, 1870535785, 2094367717,
+#if defined(CWRS_EXTRA_ROWS)
+  /*...60:*/
+  2340095869U, 2609401873U, 2904062449U, 3225952925U, 3577050821U, 3959439497U,
+#endif
+  /*N=8, K=8...37*/
+  48639, 108545, 224143, 433905, 795455, 1392065, 2340495, 3800305, 5984767,
+  9173505, 13726991, 20103025, 28875327, 40754369, 56610575, 77500017,
+  104692735, 139703809, 184327311, 240673265, 311207743, 398796225, 506750351,
+  638878193, 799538175, 993696769, 1226990095, 1505789553, 1837271615,
+  2229491905U,
+#if defined(CWRS_EXTRA_ROWS)
+  /*...40:*/
+  2691463695U, 3233240945U, 3866006015U,
+#endif
+  /*N=9, K=9...28:*/
+  265729, 598417, 1256465, 2485825, 4673345, 8405905, 14546705, 24331777,
+  39490049, 62390545, 96220561, 145198913, 214828609, 312193553, 446304145,
+  628496897, 872893441, 1196924561, 1621925137, 2173806145U,
+#if defined(CWRS_EXTRA_ROWS)
+  /*...29:*/
+  2883810113U,
+#endif
+  /*N=10, K=10...24:*/
+  1462563, 3317445, 7059735, 14218905, 27298155, 50250765, 89129247, 152951073,
+  254831667, 413442773, 654862247, 1014889769, 1541911931, 2300409629U,
+  3375210671U,
+  /*N=11, K=11...19:*/
+  8097453, 18474633, 39753273, 81270333, 158819253, 298199265, 540279585,
+  948062325, 1616336765,
+#if defined(CWRS_EXTRA_ROWS)
+  /*...20:*/
+  2684641785U,
+#endif
+  /*N=12, K=12...18:*/
+  45046719, 103274625, 224298231, 464387817, 921406335, 1759885185,
+  3248227095U,
+  /*N=13, K=13...16:*/
+  251595969, 579168825, 1267854873, 2653649025U,
+  /*N=14, K=14:*/
+  1409933619
+};
+
+#if defined(CWRS_EXTRA_ROWS)
+static const opus_uint32 *const CELT_PVQ_U_ROW[15]={
+  CELT_PVQ_U_DATA+   0,CELT_PVQ_U_DATA+ 208,CELT_PVQ_U_DATA+ 415,
+  CELT_PVQ_U_DATA+ 621,CELT_PVQ_U_DATA+ 826,CELT_PVQ_U_DATA+1030,
+  CELT_PVQ_U_DATA+1233,CELT_PVQ_U_DATA+1336,CELT_PVQ_U_DATA+1389,
+  CELT_PVQ_U_DATA+1421,CELT_PVQ_U_DATA+1441,CELT_PVQ_U_DATA+1455,
+  CELT_PVQ_U_DATA+1464,CELT_PVQ_U_DATA+1470,CELT_PVQ_U_DATA+1473
+};
+#else
+static const opus_uint32 *const CELT_PVQ_U_ROW[15]={
+  CELT_PVQ_U_DATA+   0,CELT_PVQ_U_DATA+ 176,CELT_PVQ_U_DATA+ 351,
+  CELT_PVQ_U_DATA+ 525,CELT_PVQ_U_DATA+ 698,CELT_PVQ_U_DATA+ 870,
+  CELT_PVQ_U_DATA+1041,CELT_PVQ_U_DATA+1131,CELT_PVQ_U_DATA+1178,
+  CELT_PVQ_U_DATA+1207,CELT_PVQ_U_DATA+1226,CELT_PVQ_U_DATA+1240,
+  CELT_PVQ_U_DATA+1248,CELT_PVQ_U_DATA+1254,CELT_PVQ_U_DATA+1257
+};
+#endif
+
+#if defined(CUSTOM_MODES)
+void get_required_bits(opus_int16 *_bits,int _n,int _maxk,int _frac){
+  int k;
+  /*_maxk==0 => there's nothing to do.*/
+  celt_assert(_maxk>0);
+  _bits[0]=0;
+  for(k=1;k<=_maxk;k++)_bits[k]=log2_frac(CELT_PVQ_V(_n,k),_frac);
+}
+#endif
+
+static opus_uint32 icwrs(int _n,const int *_y){
+  opus_uint32 i;
+  int         j;
+  int         k;
+  celt_assert(_n>=2);
+  j=_n-1;
+  i=_y[j]<0;
+  k=abs(_y[j]);
+  do{
+    j--;
+    i+=CELT_PVQ_U(_n-j,k);
+    k+=abs(_y[j]);
+    if(_y[j]<0)i+=CELT_PVQ_U(_n-j,k+1);
+  }
+  while(j>0);
+  return i;
+}
+
+void encode_pulses(const int *_y,int _n,int _k,ec_enc *_enc){
+  celt_assert(_k>0);
+  ec_enc_uint(_enc,icwrs(_n,_y),CELT_PVQ_V(_n,_k));
+}
+
+static opus_val32 cwrsi(int _n,int _k,opus_uint32 _i,int *_y){
+  opus_uint32 p;
+  int         s;
+  int         k0;
+  opus_int16  val;
+  opus_val32  yy=0;
+  celt_assert(_k>0);
+  celt_assert(_n>1);
+  while(_n>2){
+    opus_uint32 q;
+    /*Lots of pulses case:*/
+    if(_k>=_n){
+      const opus_uint32 *row;
+      row=CELT_PVQ_U_ROW[_n];
+      /*Are the pulses in this dimension negative?*/
+      p=row[_k+1];
+      s=-(_i>=p);
+      _i-=p&s;
+      /*Count how many pulses were placed in this dimension.*/
+      k0=_k;
+      q=row[_n];
+      if(q>_i){
+        celt_sig_assert(p>q);
+        _k=_n;
+        do p=CELT_PVQ_U_ROW[--_k][_n];
+        while(p>_i);
+      }
+      else for(p=row[_k];p>_i;p=row[_k])_k--;
+      _i-=p;
+      val=(k0-_k+s)^s;
+      *_y++=val;
+      yy=MAC16_16(yy,val,val);
+    }
+    /*Lots of dimensions case:*/
+    else{
+      /*Are there any pulses in this dimension at all?*/
+      p=CELT_PVQ_U_ROW[_k][_n];
+      q=CELT_PVQ_U_ROW[_k+1][_n];
+      if(p<=_i&&_i<q){
+        _i-=p;
+        *_y++=0;
+      }
+      else{
+        /*Are the pulses in this dimension negative?*/
+        s=-(_i>=q);
+        _i-=q&s;
+        /*Count how many pulses were placed in this dimension.*/
+        k0=_k;
+        do p=CELT_PVQ_U_ROW[--_k][_n];
+        while(p>_i);
+        _i-=p;
+        val=(k0-_k+s)^s;
+        *_y++=val;
+        yy=MAC16_16(yy,val,val);
+      }
+    }
+    _n--;
+  }
+  /*_n==2*/
+  p=2*_k+1;
+  s=-(_i>=p);
+  _i-=p&s;
+  k0=_k;
+  _k=(_i+1)>>1;
+  if(_k)_i-=2*_k-1;
+  val=(k0-_k+s)^s;
+  *_y++=val;
+  yy=MAC16_16(yy,val,val);
+  /*_n==1*/
+  s=-(int)_i;
+  val=(_k+s)^s;
+  *_y=val;
+  yy=MAC16_16(yy,val,val);
+  return yy;
+}
+
+opus_val32 decode_pulses(int *_y,int _n,int _k,ec_dec *_dec){
+  return cwrsi(_n,_k,ec_dec_uint(_dec,CELT_PVQ_V(_n,_k)),_y);
+}
+
+#else /* SMALL_FOOTPRINT */
+
+/*Computes the next row/column of any recurrence that obeys the relation
+   u[i][j]=u[i-1][j]+u[i][j-1]+u[i-1][j-1].
+  _ui0 is the base case for the new row/column.*/
+static OPUS_INLINE void unext(opus_uint32 *_ui,unsigned _len,opus_uint32 _ui0){
+  opus_uint32 ui1;
+  unsigned      j;
+  /*This do-while will overrun the array if we don't have storage for at least
+     2 values.*/
+  j=1; do {
+    ui1=UADD32(UADD32(_ui[j],_ui[j-1]),_ui0);
+    _ui[j-1]=_ui0;
+    _ui0=ui1;
+  } while (++j<_len);
+  _ui[j-1]=_ui0;
+}
+
+/*Computes the previous row/column of any recurrence that obeys the relation
+   u[i-1][j]=u[i][j]-u[i][j-1]-u[i-1][j-1].
+  _ui0 is the base case for the new row/column.*/
+static OPUS_INLINE void uprev(opus_uint32 *_ui,unsigned _n,opus_uint32 _ui0){
+  opus_uint32 ui1;
+  unsigned      j;
+  /*This do-while will overrun the array if we don't have storage for at least
+     2 values.*/
+  j=1; do {
+    ui1=USUB32(USUB32(_ui[j],_ui[j-1]),_ui0);
+    _ui[j-1]=_ui0;
+    _ui0=ui1;
+  } while (++j<_n);
+  _ui[j-1]=_ui0;
+}
+
+/*Compute V(_n,_k), as well as U(_n,0..._k+1).
+  _u: On exit, _u[i] contains U(_n,i) for i in [0..._k+1].*/
+static opus_uint32 ncwrs_urow(unsigned _n,unsigned _k,opus_uint32 *_u){
+  opus_uint32 um2;
+  unsigned      len;
+  unsigned      k;
+  len=_k+2;
+  /*We require storage at least 3 values (e.g., _k>0).*/
+  celt_assert(len>=3);
+  _u[0]=0;
+  _u[1]=um2=1;
+  /*If _n==0, _u[0] should be 1 and the rest should be 0.*/
+  /*If _n==1, _u[i] should be 1 for i>1.*/
+  celt_assert(_n>=2);
+  /*If _k==0, the following do-while loop will overflow the buffer.*/
+  celt_assert(_k>0);
+  k=2;
+  do _u[k]=(k<<1)-1;
+  while(++k<len);
+  for(k=2;k<_n;k++)unext(_u+1,_k+1,1);
+  return _u[_k]+_u[_k+1];
+}
+
+/*Returns the _i'th combination of _k elements chosen from a set of size _n
+   with associated sign bits.
+  _y: Returns the vector of pulses.
+  _u: Must contain entries [0..._k+1] of row _n of U() on input.
+      Its contents will be destructively modified.*/
+static opus_val32 cwrsi(int _n,int _k,opus_uint32 _i,int *_y,opus_uint32 *_u){
+  int j;
+  opus_int16 val;
+  opus_val32 yy=0;
+  celt_assert(_n>0);
+  j=0;
+  do{
+    opus_uint32 p;
+    int           s;
+    int           yj;
+    p=_u[_k+1];
+    s=-(_i>=p);
+    _i-=p&s;
+    yj=_k;
+    p=_u[_k];
+    while(p>_i)p=_u[--_k];
+    _i-=p;
+    yj-=_k;
+    val=(yj+s)^s;
+    _y[j]=val;
+    yy=MAC16_16(yy,val,val);
+    uprev(_u,_k+2,0);
+  }
+  while(++j<_n);
+  return yy;
+}
+
+/*Returns the index of the given combination of K elements chosen from a set
+   of size 1 with associated sign bits.
+  _y: The vector of pulses, whose sum of absolute values is K.
+  _k: Returns K.*/
+static OPUS_INLINE opus_uint32 icwrs1(const int *_y,int *_k){
+  *_k=abs(_y[0]);
+  return _y[0]<0;
+}
+
+/*Returns the index of the given combination of K elements chosen from a set
+   of size _n with associated sign bits.
+  _y:  The vector of pulses, whose sum of absolute values must be _k.
+  _nc: Returns V(_n,_k).*/
+static OPUS_INLINE opus_uint32 icwrs(int _n,int _k,opus_uint32 *_nc,const int *_y,
+ opus_uint32 *_u){
+  opus_uint32 i;
+  int         j;
+  int         k;
+  /*We can't unroll the first two iterations of the loop unless _n>=2.*/
+  celt_assert(_n>=2);
+  _u[0]=0;
+  for(k=1;k<=_k+1;k++)_u[k]=(k<<1)-1;
+  i=icwrs1(_y+_n-1,&k);
+  j=_n-2;
+  i+=_u[k];
+  k+=abs(_y[j]);
+  if(_y[j]<0)i+=_u[k+1];
+  while(j-->0){
+    unext(_u,_k+2,0);
+    i+=_u[k];
+    k+=abs(_y[j]);
+    if(_y[j]<0)i+=_u[k+1];
+  }
+  *_nc=_u[k]+_u[k+1];
+  return i;
+}
+
+#if defined(CUSTOM_MODES)
+void get_required_bits(opus_int16 *_bits,int _n,int _maxk,int _frac){
+  int k;
+  /*_maxk==0 => there's nothing to do.*/
+  celt_assert(_maxk>0);
+  _bits[0]=0;
+  if (_n==1)
+  {
+    for (k=1;k<=_maxk;k++)
+      _bits[k] = 1<<_frac;
+  }
+  else {
+    VARDECL(opus_uint32,u);
+    SAVE_STACK;
+    ALLOC(u,_maxk+2U,opus_uint32);
+    ncwrs_urow(_n,_maxk,u);
+    for(k=1;k<=_maxk;k++)
+      _bits[k]=log2_frac(u[k]+u[k+1],_frac);
+    RESTORE_STACK;
+  }
+}
+#endif /* CUSTOM_MODES */
+
+void encode_pulses(const int *_y,int _n,int _k,ec_enc *_enc){
+  opus_uint32 i;
+  VARDECL(opus_uint32,u);
+  opus_uint32 nc;
+  SAVE_STACK;
+  celt_assert(_k>0);
+  ALLOC(u,_k+2U,opus_uint32);
+  i=icwrs(_n,_k,&nc,_y,u);
+  ec_enc_uint(_enc,i,nc);
+  RESTORE_STACK;
+}
+
+opus_val32 decode_pulses(int *_y,int _n,int _k,ec_dec *_dec){
+  VARDECL(opus_uint32,u);
+  int ret;
+  SAVE_STACK;
+  celt_assert(_k>0);
+  ALLOC(u,_k+2U,opus_uint32);
+  ret = cwrsi(_n,_k,ec_dec_uint(_dec,ncwrs_urow(_n,_k,u)),_y,u);
+  RESTORE_STACK;
+  return ret;
+}
+
+#endif /* SMALL_FOOTPRINT */
--- a/lib/libopus/celt/cwrs.h
+++ b/lib/libopus/celt/cwrs.h
@ -0,0 +1,48 @@
+/* Copyright (c) 2007-2008 CSIRO
+   Copyright (c) 2007-2009 Xiph.Org Foundation
+   Copyright (c) 2007-2009 Timothy B. Terriberry
+   Written by Timothy B. Terriberry and Jean-Marc Valin */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef CWRS_H
+#define CWRS_H
+
+#include "arch.h"
+#include "stack_alloc.h"
+#include "entenc.h"
+#include "entdec.h"
+
+#ifdef CUSTOM_MODES
+int log2_frac(opus_uint32 val, int frac);
+#endif
+
+void get_required_bits(opus_int16 *bits, int N, int K, int frac);
+
+void encode_pulses(const int *_y, int N, int K, ec_enc *enc);
+
+opus_val32 decode_pulses(int *_y, int N, int K, ec_dec *dec);
+
+#endif /* CWRS_H */
--- a/lib/libopus/celt/ecintrin.h
+++ b/lib/libopus/celt/ecintrin.h
@ -0,0 +1,91 @@
+/* Copyright (c) 2003-2008 Timothy B. Terriberry
+   Copyright (c) 2008 Xiph.Org Foundation */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+/*Some common macros for potential platform-specific optimization.*/
+#include "opus_types.h"
+#include <math.h>
+#include <limits.h>
+#include "arch.h"
+#if !defined(_ecintrin_H)
+# define _ecintrin_H (1)
+
+/*Some specific platforms may have optimized intrinsic or OPUS_INLINE assembly
+   versions of these functions which can substantially improve performance.
+  We define macros for them to allow easy incorporation of these non-ANSI
+   features.*/
+
+/*Modern gcc (4.x) can compile the naive versions of min and max with cmov if
+   given an appropriate architecture, but the branchless bit-twiddling versions
+   are just as fast, and do not require any special target architecture.
+  Earlier gcc versions (3.x) compiled both code to the same assembly
+   instructions, because of the way they represented ((_b)>(_a)) internally.*/
+# define EC_MINI(_a,_b)      ((_a)+(((_b)-(_a))&-((_b)<(_a))))
+
+/*Count leading zeros.
+  This macro should only be used for implementing ec_ilog(), if it is defined.
+  All other code should use EC_ILOG() instead.*/
+#if defined(_MSC_VER) && (_MSC_VER >= 1400)
+#if defined(_MSC_VER) && (_MSC_VER >= 1910)
+# include <intrin0.h> /* Improve compiler throughput. */
+#else
+# include <intrin.h>
+#endif
+/*In _DEBUG mode this is not an intrinsic by default.*/
+# pragma intrinsic(_BitScanReverse)
+
+static __inline int ec_bsr(unsigned long _x){
+  unsigned long ret;
+  _BitScanReverse(&ret,_x);
+  return (int)ret;
+}
+# define EC_CLZ0    (1)
+# define EC_CLZ(_x) (-ec_bsr(_x))
+#elif defined(ENABLE_TI_DSPLIB)
+# include "dsplib.h"
+# define EC_CLZ0    (31)
+# define EC_CLZ(_x) (_lnorm(_x))
+#elif __GNUC_PREREQ(3,4)
+# if INT_MAX>=2147483647
+#  define EC_CLZ0    ((int)sizeof(unsigned)*CHAR_BIT)
+#  define EC_CLZ(_x) (__builtin_clz(_x))
+# elif LONG_MAX>=2147483647L
+#  define EC_CLZ0    ((int)sizeof(unsigned long)*CHAR_BIT)
+#  define EC_CLZ(_x) (__builtin_clzl(_x))
+# endif
+#endif
+
+#if defined(EC_CLZ)
+/*Note that __builtin_clz is not defined when _x==0, according to the gcc
+   documentation (and that of the BSR instruction that implements it on x86).
+  The majority of the time we can never pass it zero.
+  When we need to, it can be special cased.*/
+# define EC_ILOG(_x) (EC_CLZ0-EC_CLZ(_x))
+#else
+int ec_ilog(opus_uint32 _v);
+# define EC_ILOG(_x) (ec_ilog(_x))
+#endif
+#endif
--- a/lib/libopus/celt/entcode.c
+++ b/lib/libopus/celt/entcode.c
@ -0,0 +1,153 @@
+/* Copyright (c) 2001-2011 Timothy B. Terriberry
+*/
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "entcode.h"
+#include "arch.h"
+
+#if !defined(EC_CLZ)
+/*This is a fallback for systems where we don't know how to access
+   a BSR or CLZ instruction (see ecintrin.h).
+  If you are optimizing Opus on a new platform and it has a native CLZ or
+   BZR (e.g. cell, MIPS, x86, etc) then making it available to Opus will be
+   an easy performance win.*/
+int ec_ilog(opus_uint32 _v){
+  /*On a Pentium M, this branchless version tested as the fastest on
+     1,000,000,000 random 32-bit integers, edging out a similar version with
+     branches, and a 256-entry LUT version.*/
+  int ret;
+  int m;
+  ret=!!_v;
+  m=!!(_v&0xFFFF0000)<<4;
+  _v>>=m;
+  ret|=m;
+  m=!!(_v&0xFF00)<<3;
+  _v>>=m;
+  ret|=m;
+  m=!!(_v&0xF0)<<2;
+  _v>>=m;
+  ret|=m;
+  m=!!(_v&0xC)<<1;
+  _v>>=m;
+  ret|=m;
+  ret+=!!(_v&0x2);
+  return ret;
+}
+#endif
+
+#if 1
+/* This is a faster version of ec_tell_frac() that takes advantage
+   of the low (1/8 bit) resolution to use just a linear function
+   followed by a lookup to determine the exact transition thresholds. */
+opus_uint32 ec_tell_frac(ec_ctx *_this){
+  static const unsigned correction[8] =
+    {35733, 38967, 42495, 46340,
+     50535, 55109, 60097, 65535};
+  opus_uint32 nbits;
+  opus_uint32 r;
+  int         l;
+  unsigned    b;
+  nbits=_this->nbits_total<<BITRES;
+  l=EC_ILOG(_this->rng);
+  r=_this->rng>>(l-16);
+  b = (r>>12)-8;
+  b += r>correction[b];
+  l = (l<<3)+b;
+  return nbits-l;
+}
+#else
+opus_uint32 ec_tell_frac(ec_ctx *_this){
+  opus_uint32 nbits;
+  opus_uint32 r;
+  int         l;
+  int         i;
+  /*To handle the non-integral number of bits still left in the encoder/decoder
+     state, we compute the worst-case number of bits of val that must be
+     encoded to ensure that the value is inside the range for any possible
+     subsequent bits.
+    The computation here is independent of val itself (the decoder does not
+     even track that value), even though the real number of bits used after
+     ec_enc_done() may be 1 smaller if rng is a power of two and the
+     corresponding trailing bits of val are all zeros.
+    If we did try to track that special case, then coding a value with a
+     probability of 1/(1<<n) might sometimes appear to use more than n bits.
+    This may help explain the surprising result that a newly initialized
+     encoder or decoder claims to have used 1 bit.*/
+  nbits=_this->nbits_total<<BITRES;
+  l=EC_ILOG(_this->rng);
+  r=_this->rng>>(l-16);
+  for(i=BITRES;i-->0;){
+    int b;
+    r=r*r>>15;
+    b=(int)(r>>16);
+    l=l<<1|b;
+    r>>=b;
+  }
+  return nbits-l;
+}
+#endif
+
+#ifdef USE_SMALL_DIV_TABLE
+/* Result of 2^32/(2*i+1), except for i=0. */
+const opus_uint32 SMALL_DIV_TABLE[129] = {
+   0xFFFFFFFF, 0x55555555, 0x33333333, 0x24924924,
+   0x1C71C71C, 0x1745D174, 0x13B13B13, 0x11111111,
+   0x0F0F0F0F, 0x0D79435E, 0x0C30C30C, 0x0B21642C,
+   0x0A3D70A3, 0x097B425E, 0x08D3DCB0, 0x08421084,
+   0x07C1F07C, 0x07507507, 0x06EB3E45, 0x06906906,
+   0x063E7063, 0x05F417D0, 0x05B05B05, 0x0572620A,
+   0x05397829, 0x05050505, 0x04D4873E, 0x04A7904A,
+   0x047DC11F, 0x0456C797, 0x04325C53, 0x04104104,
+   0x03F03F03, 0x03D22635, 0x03B5CC0E, 0x039B0AD1,
+   0x0381C0E0, 0x0369D036, 0x03531DEC, 0x033D91D2,
+   0x0329161F, 0x03159721, 0x03030303, 0x02F14990,
+   0x02E05C0B, 0x02D02D02, 0x02C0B02C, 0x02B1DA46,
+   0x02A3A0FD, 0x0295FAD4, 0x0288DF0C, 0x027C4597,
+   0x02702702, 0x02647C69, 0x02593F69, 0x024E6A17,
+   0x0243F6F0, 0x0239E0D5, 0x02302302, 0x0226B902,
+   0x021D9EAD, 0x0214D021, 0x020C49BA, 0x02040810,
+   0x01FC07F0, 0x01F44659, 0x01ECC07B, 0x01E573AC,
+   0x01DE5D6E, 0x01D77B65, 0x01D0CB58, 0x01CA4B30,
+   0x01C3F8F0, 0x01BDD2B8, 0x01B7D6C3, 0x01B20364,
+   0x01AC5701, 0x01A6D01A, 0x01A16D3F, 0x019C2D14,
+   0x01970E4F, 0x01920FB4, 0x018D3018, 0x01886E5F,
+   0x0183C977, 0x017F405F, 0x017AD220, 0x01767DCE,
+   0x01724287, 0x016E1F76, 0x016A13CD, 0x01661EC6,
+   0x01623FA7, 0x015E75BB, 0x015AC056, 0x01571ED3,
+   0x01539094, 0x01501501, 0x014CAB88, 0x0149539E,
+   0x01460CBC, 0x0142D662, 0x013FB013, 0x013C995A,
+   0x013991C2, 0x013698DF, 0x0133AE45, 0x0130D190,
+   0x012E025C, 0x012B404A, 0x01288B01, 0x0125E227,
+   0x01234567, 0x0120B470, 0x011E2EF3, 0x011BB4A4,
+   0x01194538, 0x0116E068, 0x011485F0, 0x0112358E,
+   0x010FEF01, 0x010DB20A, 0x010B7E6E, 0x010953F3,
+   0x01073260, 0x0105197F, 0x0103091B, 0x01010101
+};
+#endif
--- a/lib/libopus/celt/entcode.h
+++ b/lib/libopus/celt/entcode.h
@ -0,0 +1,152 @@
+/* Copyright (c) 2001-2011 Timothy B. Terriberry
+   Copyright (c) 2008-2009 Xiph.Org Foundation */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#include "opus_types.h"
+#include "opus_defines.h"
+
+#if !defined(_entcode_H)
+# define _entcode_H (1)
+# include <limits.h>
+# include <stddef.h>
+# include "ecintrin.h"
+
+extern const opus_uint32 SMALL_DIV_TABLE[129];
+
+#ifdef OPUS_ARM_ASM
+#define USE_SMALL_DIV_TABLE
+#endif
+
+/*OPT: ec_window must be at least 32 bits, but if you have fast arithmetic on a
+   larger type, you can speed up the decoder by using it here.*/
+typedef opus_uint32           ec_window;
+typedef struct ec_ctx         ec_ctx;
+typedef struct ec_ctx         ec_enc;
+typedef struct ec_ctx         ec_dec;
+
+# define EC_WINDOW_SIZE ((int)sizeof(ec_window)*CHAR_BIT)
+
+/*The number of bits to use for the range-coded part of unsigned integers.*/
+# define EC_UINT_BITS   (8)
+
+/*The resolution of fractional-precision bit usage measurements, i.e.,
+   3 => 1/8th bits.*/
+# define BITRES 3
+
+/*The entropy encoder/decoder context.
+  We use the same structure for both, so that common functions like ec_tell()
+   can be used on either one.*/
+struct ec_ctx{
+   /*Buffered input/output.*/
+   unsigned char *buf;
+   /*The size of the buffer.*/
+   opus_uint32    storage;
+   /*The offset at which the last byte containing raw bits was read/written.*/
+   opus_uint32    end_offs;
+   /*Bits that will be read from/written at the end.*/
+   ec_window      end_window;
+   /*Number of valid bits in end_window.*/
+   int            nend_bits;
+   /*The total number of whole bits read/written.
+     This does not include partial bits currently in the range coder.*/
+   int            nbits_total;
+   /*The offset at which the next range coder byte will be read/written.*/
+   opus_uint32    offs;
+   /*The number of values in the current range.*/
+   opus_uint32    rng;
+   /*In the decoder: the difference between the top of the current range and
+      the input value, minus one.
+     In the encoder: the low end of the current range.*/
+   opus_uint32    val;
+   /*In the decoder: the saved normalization factor from ec_decode().
+     In the encoder: the number of outstanding carry propagating symbols.*/
+   opus_uint32    ext;
+   /*A buffered input/output symbol, awaiting carry propagation.*/
+   int            rem;
+   /*Nonzero if an error occurred.*/
+   int            error;
+};
+
+static OPUS_INLINE opus_uint32 ec_range_bytes(ec_ctx *_this){
+  return _this->offs;
+}
+
+static OPUS_INLINE unsigned char *ec_get_buffer(ec_ctx *_this){
+  return _this->buf;
+}
+
+static OPUS_INLINE int ec_get_error(ec_ctx *_this){
+  return _this->error;
+}
+
+/*Returns the number of bits "used" by the encoded or decoded symbols so far.
+  This same number can be computed in either the encoder or the decoder, and is
+   suitable for making coding decisions.
+  Return: The number of bits.
+          This will always be slightly larger than the exact value (e.g., all
+           rounding error is in the positive direction).*/
+static OPUS_INLINE int ec_tell(ec_ctx *_this){
+  return _this->nbits_total-EC_ILOG(_this->rng);
+}
+
+/*Returns the number of bits "used" by the encoded or decoded symbols so far.
+  This same number can be computed in either the encoder or the decoder, and is
+   suitable for making coding decisions.
+  Return: The number of bits scaled by 2**BITRES.
+          This will always be slightly larger than the exact value (e.g., all
+           rounding error is in the positive direction).*/
+opus_uint32 ec_tell_frac(ec_ctx *_this);
+
+/* Tested exhaustively for all n and for 1<=d<=256 */
+static OPUS_INLINE opus_uint32 celt_udiv(opus_uint32 n, opus_uint32 d) {
+   celt_sig_assert(d>0);
+#ifdef USE_SMALL_DIV_TABLE
+   if (d>256)
+      return n/d;
+   else {
+      opus_uint32 t, q;
+      t = EC_ILOG(d&-d);
+      q = (opus_uint64)SMALL_DIV_TABLE[d>>t]*(n>>(t-1))>>32;
+      return q+(n-q*d >= d);
+   }
+#else
+   return n/d;
+#endif
+}
+
+static OPUS_INLINE opus_int32 celt_sudiv(opus_int32 n, opus_int32 d) {
+   celt_sig_assert(d>0);
+#ifdef USE_SMALL_DIV_TABLE
+   if (n<0)
+      return -(opus_int32)celt_udiv(-n, d);
+   else
+      return celt_udiv(n, d);
+#else
+   return n/d;
+#endif
+}
+
+#endif
--- a/lib/libopus/celt/entdec.c
+++ b/lib/libopus/celt/entdec.c
@ -0,0 +1,266 @@
+/* Copyright (c) 2001-2011 Timothy B. Terriberry
+   Copyright (c) 2008-2009 Xiph.Org Foundation */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include <stddef.h>
+#include "os_support.h"
+#include "arch.h"
+#include "entdec.h"
+#include "mfrngcod.h"
+
+/*A range decoder.
+  This is an entropy decoder based upon \cite{Mar79}, which is itself a
+   rediscovery of the FIFO arithmetic code introduced by \cite{Pas76}.
+  It is very similar to arithmetic encoding, except that encoding is done with
+   digits in any base, instead of with bits, and so it is faster when using
+   larger bases (i.e.: a byte).
+  The author claims an average waste of $\frac{1}{2}\log_b(2b)$ bits, where $b$
+   is the base, longer than the theoretical optimum, but to my knowledge there
+   is no published justification for this claim.
+  This only seems true when using near-infinite precision arithmetic so that
+   the process is carried out with no rounding errors.
+
+  An excellent description of implementation details is available at
+   http://www.arturocampos.com/ac_range.html
+  A recent work \cite{MNW98} which proposes several changes to arithmetic
+   encoding for efficiency actually re-discovers many of the principles
+   behind range encoding, and presents a good theoretical analysis of them.
+
+  End of stream is handled by writing out the smallest number of bits that
+   ensures that the stream will be correctly decoded regardless of the value of
+   any subsequent bits.
+  ec_tell() can be used to determine how many bits were needed to decode
+   all the symbols thus far; other data can be packed in the remaining bits of
+   the input buffer.
+  @PHDTHESIS{Pas76,
+    author="Richard Clark Pasco",
+    title="Source coding algorithms for fast data compression",
+    school="Dept. of Electrical Engineering, Stanford University",
+    address="Stanford, CA",
+    month=May,
+    year=1976
+  }
+  @INPROCEEDINGS{Mar79,
+   author="Martin, G.N.N.",
+   title="Range encoding: an algorithm for removing redundancy from a digitised
+    message",
+   booktitle="Video & Data Recording Conference",
+   year=1979,
+   address="Southampton",
+   month=Jul
+  }
+  @ARTICLE{MNW98,
+   author="Alistair Moffat and Radford Neal and Ian H. Witten",
+   title="Arithmetic Coding Revisited",
+   journal="{ACM} Transactions on Information Systems",
+   year=1998,
+   volume=16,
+   number=3,
+   pages="256--294",
+   month=Jul,
+   URL="http://www.stanford.edu/class/ee398a/handouts/papers/Moffat98ArithmCoding.pdf"
+  }*/
+
+static int ec_read_byte(ec_dec *_this){
+  return _this->offs<_this->storage?_this->buf[_this->offs++]:0;
+}
+
+static int ec_read_byte_from_end(ec_dec *_this){
+  return _this->end_offs<_this->storage?
+   _this->buf[_this->storage-++(_this->end_offs)]:0;
+}
+
+/*Normalizes the contents of val and rng so that rng lies entirely in the
+   high-order symbol.*/
+static void ec_dec_normalize(ec_dec *_this){
+  /*If the range is too small, rescale it and input some bits.*/
+  while(_this->rng<=EC_CODE_BOT){
+    int sym;
+    _this->nbits_total+=EC_SYM_BITS;
+    _this->rng<<=EC_SYM_BITS;
+    /*Use up the remaining bits from our last symbol.*/
+    sym=_this->rem;
+    /*Read the next value from the input.*/
+    _this->rem=ec_read_byte(_this);
+    /*Take the rest of the bits we need from this new symbol.*/
+    sym=(sym<<EC_SYM_BITS|_this->rem)>>(EC_SYM_BITS-EC_CODE_EXTRA);
+    /*And subtract them from val, capped to be less than EC_CODE_TOP.*/
+    _this->val=((_this->val<<EC_SYM_BITS)+(EC_SYM_MAX&~sym))&(EC_CODE_TOP-1);
+  }
+}
+
+void ec_dec_init(ec_dec *_this,unsigned char *_buf,opus_uint32 _storage){
+  _this->buf=_buf;
+  _this->storage=_storage;
+  _this->end_offs=0;
+  _this->end_window=0;
+  _this->nend_bits=0;
+  /*This is the offset from which ec_tell() will subtract partial bits.
+    The final value after the ec_dec_normalize() call will be the same as in
+     the encoder, but we have to compensate for the bits that are added there.*/
+  _this->nbits_total=EC_CODE_BITS+1
+   -((EC_CODE_BITS-EC_CODE_EXTRA)/EC_SYM_BITS)*EC_SYM_BITS;
+  _this->offs=0;
+  _this->rng=1U<<EC_CODE_EXTRA;
+  _this->rem=ec_read_byte(_this);
+  _this->val=_this->rng-1-(_this->rem>>(EC_SYM_BITS-EC_CODE_EXTRA));
+  _this->error=0;
+  /*Normalize the interval.*/
+  ec_dec_normalize(_this);
+}
+
+unsigned ec_decode(ec_dec *_this,unsigned _ft){
+  unsigned s;
+  _this->ext=celt_udiv(_this->rng,_ft);
+  s=(unsigned)(_this->val/_this->ext);
+  return _ft-EC_MINI(s+1,_ft);
+}
+
+unsigned ec_decode_bin(ec_dec *_this,unsigned _bits){
+   unsigned s;
+   _this->ext=_this->rng>>_bits;
+   s=(unsigned)(_this->val/_this->ext);
+   return (1U<<_bits)-EC_MINI(s+1U,1U<<_bits);
+}
+
+void ec_dec_update(ec_dec *_this,unsigned _fl,unsigned _fh,unsigned _ft){
+  opus_uint32 s;
+  s=IMUL32(_this->ext,_ft-_fh);
+  _this->val-=s;
+  _this->rng=_fl>0?IMUL32(_this->ext,_fh-_fl):_this->rng-s;
+  ec_dec_normalize(_this);
+}
+
+/*The probability of having a "one" is 1/(1<<_logp).*/
+int ec_dec_bit_logp(ec_dec *_this,unsigned _logp){
+  opus_uint32 r;
+  opus_uint32 d;
+  opus_uint32 s;
+  int         ret;
+  r=_this->rng;
+  d=_this->val;
+  s=r>>_logp;
+  ret=d<s;
+  if(!ret)_this->val=d-s;
+  _this->rng=ret?s:r-s;
+  ec_dec_normalize(_this);
+  return ret;
+}
+
+int ec_dec_icdf(ec_dec *_this,const unsigned char *_icdf,unsigned _ftb){
+  opus_uint32 r;
+  opus_uint32 d;
+  opus_uint32 s;
+  opus_uint32 t;
+  int         ret;
+  s=_this->rng;
+  d=_this->val;
+  r=s>>_ftb;
+  ret=-1;
+  do{
+    t=s;
+    s=IMUL32(r,_icdf[++ret]);
+  }
+  while(d<s);
+  _this->val=d-s;
+  _this->rng=t-s;
+  ec_dec_normalize(_this);
+  return ret;
+}
+
+int ec_dec_icdf16(ec_dec *_this,const opus_uint16 *_icdf,unsigned _ftb){
+  opus_uint32 r;
+  opus_uint32 d;
+  opus_uint32 s;
+  opus_uint32 t;
+  int         ret;
+  s=_this->rng;
+  d=_this->val;
+  r=s>>_ftb;
+  ret=-1;
+  do{
+    t=s;
+    s=IMUL32(r,_icdf[++ret]);
+  }
+  while(d<s);
+  _this->val=d-s;
+  _this->rng=t-s;
+  ec_dec_normalize(_this);
+  return ret;
+}
+
+opus_uint32 ec_dec_uint(ec_dec *_this,opus_uint32 _ft){
+  unsigned ft;
+  unsigned s;
+  int      ftb;
+  /*In order to optimize EC_ILOG(), it is undefined for the value 0.*/
+  celt_assert(_ft>1);
+  _ft--;
+  ftb=EC_ILOG(_ft);
+  if(ftb>EC_UINT_BITS){
+    opus_uint32 t;
+    ftb-=EC_UINT_BITS;
+    ft=(unsigned)(_ft>>ftb)+1;
+    s=ec_decode(_this,ft);
+    ec_dec_update(_this,s,s+1,ft);
+    t=(opus_uint32)s<<ftb|ec_dec_bits(_this,ftb);
+    if(t<=_ft)return t;
+    _this->error=1;
+    return _ft;
+  }
+  else{
+    _ft++;
+    s=ec_decode(_this,(unsigned)_ft);
+    ec_dec_update(_this,s,s+1,(unsigned)_ft);
+    return s;
+  }
+}
+
+opus_uint32 ec_dec_bits(ec_dec *_this,unsigned _bits){
+  ec_window   window;
+  int         available;
+  opus_uint32 ret;
+  window=_this->end_window;
+  available=_this->nend_bits;
+  if((unsigned)available<_bits){
+    do{
+      window|=(ec_window)ec_read_byte_from_end(_this)<<available;
+      available+=EC_SYM_BITS;
+    }
+    while(available<=EC_WINDOW_SIZE-EC_SYM_BITS);
+  }
+  ret=(opus_uint32)window&(((opus_uint32)1<<_bits)-1U);
+  window>>=_bits;
+  available-=_bits;
+  _this->end_window=window;
+  _this->nend_bits=available;
+  _this->nbits_total+=_bits;
+  return ret;
+}
--- a/lib/libopus/celt/entdec.h
+++ b/lib/libopus/celt/entdec.h
@ -0,0 +1,110 @@
+/* Copyright (c) 2001-2011 Timothy B. Terriberry
+   Copyright (c) 2008-2009 Xiph.Org Foundation */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#if !defined(_entdec_H)
+# define _entdec_H (1)
+# include <limits.h>
+# include "entcode.h"
+
+/*Initializes the decoder.
+  _buf: The input buffer to use.
+  Return: 0 on success, or a negative value on error.*/
+void ec_dec_init(ec_dec *_this,unsigned char *_buf,opus_uint32 _storage);
+
+/*Calculates the cumulative frequency for the next symbol.
+  This can then be fed into the probability model to determine what that
+   symbol is, and the additional frequency information required to advance to
+   the next symbol.
+  This function cannot be called more than once without a corresponding call to
+   ec_dec_update(), or decoding will not proceed correctly.
+  _ft: The total frequency of the symbols in the alphabet the next symbol was
+        encoded with.
+  Return: A cumulative frequency representing the encoded symbol.
+          If the cumulative frequency of all the symbols before the one that
+           was encoded was fl, and the cumulative frequency of all the symbols
+           up to and including the one encoded is fh, then the returned value
+           will fall in the range [fl,fh).*/
+unsigned ec_decode(ec_dec *_this,unsigned _ft);
+
+/*Equivalent to ec_decode() with _ft==1<<_bits.*/
+unsigned ec_decode_bin(ec_dec *_this,unsigned _bits);
+
+/*Advance the decoder past the next symbol using the frequency information the
+   symbol was encoded with.
+  Exactly one call to ec_decode() must have been made so that all necessary
+   intermediate calculations are performed.
+  _fl:  The cumulative frequency of all symbols that come before the symbol
+         decoded.
+  _fh:  The cumulative frequency of all symbols up to and including the symbol
+         decoded.
+        Together with _fl, this defines the range [_fl,_fh) in which the value
+         returned above must fall.
+  _ft:  The total frequency of the symbols in the alphabet the symbol decoded
+         was encoded in.
+        This must be the same as passed to the preceding call to ec_decode().*/
+void ec_dec_update(ec_dec *_this,unsigned _fl,unsigned _fh,unsigned _ft);
+
+/* Decode a bit that has a 1/(1<<_logp) probability of being a one */
+int ec_dec_bit_logp(ec_dec *_this,unsigned _logp);
+
+/*Decodes a symbol given an "inverse" CDF table.
+  No call to ec_dec_update() is necessary after this call.
+  _icdf: The "inverse" CDF, such that symbol s falls in the range
+          [s>0?ft-_icdf[s-1]:0,ft-_icdf[s]), where ft=1<<_ftb.
+         The values must be monotonically non-increasing, and the last value
+          must be 0.
+  _ftb: The number of bits of precision in the cumulative distribution.
+  Return: The decoded symbol s.*/
+int ec_dec_icdf(ec_dec *_this,const unsigned char *_icdf,unsigned _ftb);
+
+/*Decodes a symbol given an "inverse" CDF table.
+  No call to ec_dec_update() is necessary after this call.
+  _icdf: The "inverse" CDF, such that symbol s falls in the range
+          [s>0?ft-_icdf[s-1]:0,ft-_icdf[s]), where ft=1<<_ftb.
+         The values must be monotonically non-increasing, and the last value
+          must be 0.
+  _ftb: The number of bits of precision in the cumulative distribution.
+  Return: The decoded symbol s.*/
+int ec_dec_icdf16(ec_dec *_this,const opus_uint16 *_icdf,unsigned _ftb);
+
+/*Extracts a raw unsigned integer with a non-power-of-2 range from the stream.
+  The bits must have been encoded with ec_enc_uint().
+  No call to ec_dec_update() is necessary after this call.
+  _ft: The number of integers that can be decoded (one more than the max).
+       This must be at least 2, and no more than 2**32-1.
+  Return: The decoded bits.*/
+opus_uint32 ec_dec_uint(ec_dec *_this,opus_uint32 _ft);
+
+/*Extracts a sequence of raw bits from the stream.
+  The bits must have been encoded with ec_enc_bits().
+  No call to ec_dec_update() is necessary after this call.
+  _ftb: The number of bits to extract.
+        This must be between 0 and 25, inclusive.
+  Return: The decoded bits.*/
+opus_uint32 ec_dec_bits(ec_dec *_this,unsigned _ftb);
+
+#endif
--- a/lib/libopus/celt/entenc.c
+++ b/lib/libopus/celt/entenc.c
@ -0,0 +1,305 @@
+/* Copyright (c) 2001-2011 Timothy B. Terriberry
+   Copyright (c) 2008-2009 Xiph.Org Foundation */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#if defined(HAVE_CONFIG_H)
+# include "config.h"
+#endif
+#include "os_support.h"
+#include "arch.h"
+#include "entenc.h"
+#include "mfrngcod.h"
+
+/*A range encoder.
+  See entdec.c and the references for implementation details \cite{Mar79,MNW98}.
+
+  @INPROCEEDINGS{Mar79,
+   author="Martin, G.N.N.",
+   title="Range encoding: an algorithm for removing redundancy from a digitised
+    message",
+   booktitle="Video \& Data Recording Conference",
+   year=1979,
+   address="Southampton",
+   month=Jul
+  }
+  @ARTICLE{MNW98,
+   author="Alistair Moffat and Radford Neal and Ian H. Witten",
+   title="Arithmetic Coding Revisited",
+   journal="{ACM} Transactions on Information Systems",
+   year=1998,
+   volume=16,
+   number=3,
+   pages="256--294",
+   month=Jul,
+   URL="http://www.stanford.edu/class/ee398/handouts/papers/Moffat98ArithmCoding.pdf"
+  }*/
+
+static int ec_write_byte(ec_enc *_this,unsigned _value){
+  if(_this->offs+_this->end_offs>=_this->storage)return -1;
+  _this->buf[_this->offs++]=(unsigned char)_value;
+  return 0;
+}
+
+static int ec_write_byte_at_end(ec_enc *_this,unsigned _value){
+  if(_this->offs+_this->end_offs>=_this->storage)return -1;
+  _this->buf[_this->storage-++(_this->end_offs)]=(unsigned char)_value;
+  return 0;
+}
+
+/*Outputs a symbol, with a carry bit.
+  If there is a potential to propagate a carry over several symbols, they are
+   buffered until it can be determined whether or not an actual carry will
+   occur.
+  If the counter for the buffered symbols overflows, then the stream becomes
+   undecodable.
+  This gives a theoretical limit of a few billion symbols in a single packet on
+   32-bit systems.
+  The alternative is to truncate the range in order to force a carry, but
+   requires similar carry tracking in the decoder, needlessly slowing it down.*/
+static void ec_enc_carry_out(ec_enc *_this,int _c){
+  if(_c!=EC_SYM_MAX){
+    /*No further carry propagation possible, flush buffer.*/
+    int carry;
+    carry=_c>>EC_SYM_BITS;
+    /*Don't output a byte on the first write.
+      This compare should be taken care of by branch-prediction thereafter.*/
+    if(_this->rem>=0)_this->error|=ec_write_byte(_this,_this->rem+carry);
+    if(_this->ext>0){
+      unsigned sym;
+      sym=(EC_SYM_MAX+carry)&EC_SYM_MAX;
+      do _this->error|=ec_write_byte(_this,sym);
+      while(--(_this->ext)>0);
+    }
+    _this->rem=_c&EC_SYM_MAX;
+  }
+  else _this->ext++;
+}
+
+static OPUS_INLINE void ec_enc_normalize(ec_enc *_this){
+  /*If the range is too small, output some bits and rescale it.*/
+  while(_this->rng<=EC_CODE_BOT){
+    ec_enc_carry_out(_this,(int)(_this->val>>EC_CODE_SHIFT));
+    /*Move the next-to-high-order symbol into the high-order position.*/
+    _this->val=(_this->val<<EC_SYM_BITS)&(EC_CODE_TOP-1);
+    _this->rng<<=EC_SYM_BITS;
+    _this->nbits_total+=EC_SYM_BITS;
+  }
+}
+
+void ec_enc_init(ec_enc *_this,unsigned char *_buf,opus_uint32 _size){
+  _this->buf=_buf;
+  _this->end_offs=0;
+  _this->end_window=0;
+  _this->nend_bits=0;
+  /*This is the offset from which ec_tell() will subtract partial bits.*/
+  _this->nbits_total=EC_CODE_BITS+1;
+  _this->offs=0;
+  _this->rng=EC_CODE_TOP;
+  _this->rem=-1;
+  _this->val=0;
+  _this->ext=0;
+  _this->storage=_size;
+  _this->error=0;
+}
+
+void ec_encode(ec_enc *_this,unsigned _fl,unsigned _fh,unsigned _ft){
+  opus_uint32 r;
+  r=celt_udiv(_this->rng,_ft);
+  if(_fl>0){
+    _this->val+=_this->rng-IMUL32(r,(_ft-_fl));
+    _this->rng=IMUL32(r,(_fh-_fl));
+  }
+  else _this->rng-=IMUL32(r,(_ft-_fh));
+  ec_enc_normalize(_this);
+}
+
+void ec_encode_bin(ec_enc *_this,unsigned _fl,unsigned _fh,unsigned _bits){
+  opus_uint32 r;
+  r=_this->rng>>_bits;
+  if(_fl>0){
+    _this->val+=_this->rng-IMUL32(r,((1U<<_bits)-_fl));
+    _this->rng=IMUL32(r,(_fh-_fl));
+  }
+  else _this->rng-=IMUL32(r,((1U<<_bits)-_fh));
+  ec_enc_normalize(_this);
+}
+
+/*The probability of having a "one" is 1/(1<<_logp).*/
+void ec_enc_bit_logp(ec_enc *_this,int _val,unsigned _logp){
+  opus_uint32 r;
+  opus_uint32 s;
+  opus_uint32 l;
+  r=_this->rng;
+  l=_this->val;
+  s=r>>_logp;
+  r-=s;
+  if(_val)_this->val=l+r;
+  _this->rng=_val?s:r;
+  ec_enc_normalize(_this);
+}
+
+void ec_enc_icdf(ec_enc *_this,int _s,const unsigned char *_icdf,unsigned _ftb){
+  opus_uint32 r;
+  r=_this->rng>>_ftb;
+  if(_s>0){
+    _this->val+=_this->rng-IMUL32(r,_icdf[_s-1]);
+    _this->rng=IMUL32(r,_icdf[_s-1]-_icdf[_s]);
+  }
+  else _this->rng-=IMUL32(r,_icdf[_s]);
+  ec_enc_normalize(_this);
+}
+
+void ec_enc_icdf16(ec_enc *_this,int _s,const opus_uint16 *_icdf,unsigned _ftb){
+  opus_uint32 r;
+  r=_this->rng>>_ftb;
+  if(_s>0){
+    _this->val+=_this->rng-IMUL32(r,_icdf[_s-1]);
+    _this->rng=IMUL32(r,_icdf[_s-1]-_icdf[_s]);
+  }
+  else _this->rng-=IMUL32(r,_icdf[_s]);
+  ec_enc_normalize(_this);
+}
+
+void ec_enc_uint(ec_enc *_this,opus_uint32 _fl,opus_uint32 _ft){
+  unsigned  ft;
+  unsigned  fl;
+  int       ftb;
+  /*In order to optimize EC_ILOG(), it is undefined for the value 0.*/
+  celt_assert(_ft>1);
+  _ft--;
+  ftb=EC_ILOG(_ft);
+  if(ftb>EC_UINT_BITS){
+    ftb-=EC_UINT_BITS;
+    ft=(_ft>>ftb)+1;
+    fl=(unsigned)(_fl>>ftb);
+    ec_encode(_this,fl,fl+1,ft);
+    ec_enc_bits(_this,_fl&(((opus_uint32)1<<ftb)-1U),ftb);
+  }
+  else ec_encode(_this,_fl,_fl+1,_ft+1);
+}
+
+void ec_enc_bits(ec_enc *_this,opus_uint32 _fl,unsigned _bits){
+  ec_window window;
+  int       used;
+  window=_this->end_window;
+  used=_this->nend_bits;
+  celt_assert(_bits>0);
+  if(used+_bits>EC_WINDOW_SIZE){
+    do{
+      _this->error|=ec_write_byte_at_end(_this,(unsigned)window&EC_SYM_MAX);
+      window>>=EC_SYM_BITS;
+      used-=EC_SYM_BITS;
+    }
+    while(used>=EC_SYM_BITS);
+  }
+  window|=(ec_window)_fl<<used;
+  used+=_bits;
+  _this->end_window=window;
+  _this->nend_bits=used;
+  _this->nbits_total+=_bits;
+}
+
+void ec_enc_patch_initial_bits(ec_enc *_this,unsigned _val,unsigned _nbits){
+  int      shift;
+  unsigned mask;
+  celt_assert(_nbits<=EC_SYM_BITS);
+  shift=EC_SYM_BITS-_nbits;
+  mask=((1<<_nbits)-1)<<shift;
+  if(_this->offs>0){
+    /*The first byte has been finalized.*/
+    _this->buf[0]=(unsigned char)((_this->buf[0]&~mask)|_val<<shift);
+  }
+  else if(_this->rem>=0){
+    /*The first byte is still awaiting carry propagation.*/
+    _this->rem=(_this->rem&~mask)|_val<<shift;
+  }
+  else if(_this->rng<=(EC_CODE_TOP>>_nbits)){
+    /*The renormalization loop has never been run.*/
+    _this->val=(_this->val&~((opus_uint32)mask<<EC_CODE_SHIFT))|
+     (opus_uint32)_val<<(EC_CODE_SHIFT+shift);
+  }
+  /*The encoder hasn't even encoded _nbits of data yet.*/
+  else _this->error=-1;
+}
+
+void ec_enc_shrink(ec_enc *_this,opus_uint32 _size){
+  celt_assert(_this->offs+_this->end_offs<=_size);
+  OPUS_MOVE(_this->buf+_size-_this->end_offs,
+   _this->buf+_this->storage-_this->end_offs,_this->end_offs);
+  _this->storage=_size;
+}
+
+void ec_enc_done(ec_enc *_this){
+  ec_window   window;
+  int         used;
+  opus_uint32 msk;
+  opus_uint32 end;
+  int         l;
+  /*We output the minimum number of bits that ensures that the symbols encoded
+     thus far will be decoded correctly regardless of the bits that follow.*/
+  l=EC_CODE_BITS-EC_ILOG(_this->rng);
+  msk=(EC_CODE_TOP-1)>>l;
+  end=(_this->val+msk)&~msk;
+  if((end|msk)>=_this->val+_this->rng){
+    l++;
+    msk>>=1;
+    end=(_this->val+msk)&~msk;
+  }
+  while(l>0){
+    ec_enc_carry_out(_this,(int)(end>>EC_CODE_SHIFT));
+    end=(end<<EC_SYM_BITS)&(EC_CODE_TOP-1);
+    l-=EC_SYM_BITS;
+  }
+  /*If we have a buffered byte flush it into the output buffer.*/
+  if(_this->rem>=0||_this->ext>0)ec_enc_carry_out(_this,0);
+  /*If we have buffered extra bits, flush them as well.*/
+  window=_this->end_window;
+  used=_this->nend_bits;
+  while(used>=EC_SYM_BITS){
+    _this->error|=ec_write_byte_at_end(_this,(unsigned)window&EC_SYM_MAX);
+    window>>=EC_SYM_BITS;
+    used-=EC_SYM_BITS;
+  }
+  /*Clear any excess space and add any remaining extra bits to the last byte.*/
+  if(!_this->error){
+    if (_this->buf) OPUS_CLEAR(_this->buf+_this->offs,
+     _this->storage-_this->offs-_this->end_offs);
+    if(used>0){
+      /*If there's no range coder data at all, give up.*/
+      if(_this->end_offs>=_this->storage)_this->error=-1;
+      else{
+        l=-l;
+        /*If we've busted, don't add too many extra bits to the last byte; it
+           would corrupt the range coder data, and that's more important.*/
+        if(_this->offs+_this->end_offs>=_this->storage&&l<used){
+          window&=(1<<l)-1;
+          _this->error=-1;
+        }
+        _this->buf[_this->storage-_this->end_offs-1]|=(unsigned char)window;
+      }
+    }
+  }
+}
--- a/lib/libopus/celt/entenc.h
+++ b/lib/libopus/celt/entenc.h
@ -0,0 +1,119 @@
+/* Copyright (c) 2001-2011 Timothy B. Terriberry
+   Copyright (c) 2008-2009 Xiph.Org Foundation */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#if !defined(_entenc_H)
+# define _entenc_H (1)
+# include <stddef.h>
+# include "entcode.h"
+
+/*Initializes the encoder.
+  _buf:  The buffer to store output bytes in.
+  _size: The size of the buffer, in chars.*/
+void ec_enc_init(ec_enc *_this,unsigned char *_buf,opus_uint32 _size);
+/*Encodes a symbol given its frequency information.
+  The frequency information must be discernible by the decoder, assuming it
+   has read only the previous symbols from the stream.
+  It is allowable to change the frequency information, or even the entire
+   source alphabet, so long as the decoder can tell from the context of the
+   previously encoded information that it is supposed to do so as well.
+  _fl: The cumulative frequency of all symbols that come before the one to be
+        encoded.
+  _fh: The cumulative frequency of all symbols up to and including the one to
+        be encoded.
+       Together with _fl, this defines the range [_fl,_fh) in which the
+        decoded value will fall.
+  _ft: The sum of the frequencies of all the symbols*/
+void ec_encode(ec_enc *_this,unsigned _fl,unsigned _fh,unsigned _ft);
+
+/*Equivalent to ec_encode() with _ft==1<<_bits.*/
+void ec_encode_bin(ec_enc *_this,unsigned _fl,unsigned _fh,unsigned _bits);
+
+/* Encode a bit that has a 1/(1<<_logp) probability of being a one */
+void ec_enc_bit_logp(ec_enc *_this,int _val,unsigned _logp);
+
+/*Encodes a symbol given an "inverse" CDF table.
+  _s:    The index of the symbol to encode.
+  _icdf: The "inverse" CDF, such that symbol _s falls in the range
+          [_s>0?ft-_icdf[_s-1]:0,ft-_icdf[_s]), where ft=1<<_ftb.
+         The values must be monotonically non-increasing, and the last value
+          must be 0.
+  _ftb: The number of bits of precision in the cumulative distribution.*/
+void ec_enc_icdf(ec_enc *_this,int _s,const unsigned char *_icdf,unsigned _ftb);
+
+/*Encodes a symbol given an "inverse" CDF table.
+  _s:    The index of the symbol to encode.
+  _icdf: The "inverse" CDF, such that symbol _s falls in the range
+          [_s>0?ft-_icdf[_s-1]:0,ft-_icdf[_s]), where ft=1<<_ftb.
+         The values must be monotonically non-increasing, and the last value
+          must be 0.
+  _ftb: The number of bits of precision in the cumulative distribution.*/
+void ec_enc_icdf16(ec_enc *_this,int _s,const opus_uint16 *_icdf,unsigned _ftb);
+
+/*Encodes a raw unsigned integer in the stream.
+  _fl: The integer to encode.
+  _ft: The number of integers that can be encoded (one more than the max).
+       This must be at least 2, and no more than 2**32-1.*/
+void ec_enc_uint(ec_enc *_this,opus_uint32 _fl,opus_uint32 _ft);
+
+/*Encodes a sequence of raw bits in the stream.
+  _fl:  The bits to encode.
+  _ftb: The number of bits to encode.
+        This must be between 1 and 25, inclusive.*/
+void ec_enc_bits(ec_enc *_this,opus_uint32 _fl,unsigned _ftb);
+
+/*Overwrites a few bits at the very start of an existing stream, after they
+   have already been encoded.
+  This makes it possible to have a few flags up front, where it is easy for
+   decoders to access them without parsing the whole stream, even if their
+   values are not determined until late in the encoding process, without having
+   to buffer all the intermediate symbols in the encoder.
+  In order for this to work, at least _nbits bits must have already been
+   encoded using probabilities that are an exact power of two.
+  The encoder can verify the number of encoded bits is sufficient, but cannot
+   check this latter condition.
+  _val:   The bits to encode (in the least _nbits significant bits).
+          They will be decoded in order from most-significant to least.
+  _nbits: The number of bits to overwrite.
+          This must be no more than 8.*/
+void ec_enc_patch_initial_bits(ec_enc *_this,unsigned _val,unsigned _nbits);
+
+/*Compacts the data to fit in the target size.
+  This moves up the raw bits at the end of the current buffer so they are at
+   the end of the new buffer size.
+  The caller must ensure that the amount of data that's already been written
+   will fit in the new size.
+  _size: The number of bytes in the new buffer.
+         This must be large enough to contain the bits already written, and
+          must be no larger than the existing size.*/
+void ec_enc_shrink(ec_enc *_this,opus_uint32 _size);
+
+/*Indicates that there are no more symbols to encode.
+  All remaining output bytes are flushed to the output buffer.
+  ec_enc_init() must be called before the encoder can be used again.*/
+void ec_enc_done(ec_enc *_this);
+
+#endif
--- a/lib/libopus/celt/fixed_c5x.h
+++ b/lib/libopus/celt/fixed_c5x.h
@ -0,0 +1,79 @@
+/* Copyright (C) 2003 Jean-Marc Valin */
+/**
+   @file fixed_c5x.h
+   @brief Fixed-point operations for the TI C5x DSP family
+*/
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef FIXED_C5X_H
+#define FIXED_C5X_H
+
+#include "dsplib.h"
+
+#undef IMUL32
+static OPUS_INLINE long IMUL32(long i, long j)
+{
+   long ac0, ac1;
+   ac0 = _lmpy(i>>16,j);
+   ac1 = ac0 + _lmpy(i,j>>16);
+   return _lmpyu(i,j) + (ac1<<16);
+}
+
+#undef MAX16
+#define MAX16(a,b) _max(a,b)
+
+#undef MIN16
+#define MIN16(a,b) _min(a,b)
+
+#undef MAX32
+#define MAX32(a,b) _lmax(a,b)
+
+#undef MIN32
+#define MIN32(a,b) _lmin(a,b)
+
+#undef VSHR32
+#define VSHR32(a, shift) _lshl(a,-(shift))
+
+#undef MULT16_16_Q15
+#define MULT16_16_Q15(a,b) (_smpy(a,b))
+
+#undef MULT16_16SU
+#define MULT16_16SU(a,b) _lmpysu(a,b)
+
+#undef MULT_16_16
+#define MULT_16_16(a,b) _lmpy(a,b)
+
+/* FIXME: This is technically incorrect and is bound to cause problems. Is there any cleaner solution? */
+#undef MULT16_32_Q15
+#define MULT16_32_Q15(a,b) ADD32(SHL(MULT16_16((a),SHR((b),16)),1), SHR(MULT16_16SU((a),(b)),15))
+
+#define celt_ilog2(x) (30 - _lnorm(x))
+#define OVERRIDE_CELT_ILOG2
+
+#define celt_maxabs16(x, len) MAX32(EXTEND32(maxval((DATA *)x, len)),-EXTEND32(minval((DATA *)x, len)))
+#define OVERRIDE_CELT_MAXABS16
+
+#endif /* FIXED_C5X_H */
--- a/lib/libopus/celt/fixed_c6x.h
+++ b/lib/libopus/celt/fixed_c6x.h
@ -0,0 +1,70 @@
+/* Copyright (C) 2008 CSIRO */
+/**
+   @file fixed_c6x.h
+   @brief Fixed-point operations for the TI C6x DSP family
+*/
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef FIXED_C6X_H
+#define FIXED_C6X_H
+
+#undef MULT16_16SU
+#define MULT16_16SU(a,b) _mpysu(a,b)
+
+#undef MULT_16_16
+#define MULT_16_16(a,b) _mpy(a,b)
+
+#define celt_ilog2(x) (30 - _norm(x))
+#define OVERRIDE_CELT_ILOG2
+
+#undef MULT16_32_Q15
+#define MULT16_32_Q15(a,b) (_mpylill(a, b) >> 15)
+
+#if 0
+#include "dsplib.h"
+
+#undef MAX16
+#define MAX16(a,b) _max(a,b)
+
+#undef MIN16
+#define MIN16(a,b) _min(a,b)
+
+#undef MAX32
+#define MAX32(a,b) _lmax(a,b)
+
+#undef MIN32
+#define MIN32(a,b) _lmin(a,b)
+
+#undef VSHR32
+#define VSHR32(a, shift) _lshl(a,-(shift))
+
+#undef MULT16_16_Q15
+#define MULT16_16_Q15(a,b) (_smpy(a,b))
+
+#define celt_maxabs16(x, len) MAX32(EXTEND32(maxval((DATA *)x, len)),-EXTEND32(minval((DATA *)x, len)))
+#define OVERRIDE_CELT_MAXABS16
+
+#endif /* FIXED_C6X_H */
--- a/lib/libopus/celt/fixed_debug.h
+++ b/lib/libopus/celt/fixed_debug.h
@ -0,0 +1,848 @@
+/* Copyright (C) 2003-2008 Jean-Marc Valin
+   Copyright (C) 2007-2012 Xiph.Org Foundation */
+/**
+   @file fixed_debug.h
+   @brief Fixed-point operations with debugging
+*/
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef FIXED_DEBUG_H
+#define FIXED_DEBUG_H
+
+#include <stdio.h>
+#include "opus_defines.h"
+
+#ifdef CELT_C
+OPUS_EXPORT opus_int64 celt_mips=0;
+#else
+extern opus_int64 celt_mips;
+#endif
+
+#define MULT16_16U(a,b) ((opus_uint32)(a)*(opus_uint32)(b))
+#define MULT16_16SU(a,b) ((opus_val32)(opus_val16)(a)*(opus_val32)(opus_uint16)(b))
+#define MULT32_32_Q31(a,b) ADD32(ADD32(SHL32(MULT16_16(SHR32((a),16),SHR((b),16)),1), SHR32(MULT16_16SU(SHR32((a),16),((b)&0x0000ffff)),15)), SHR32(MULT16_16SU(SHR32((b),16),((a)&0x0000ffff)),15))
+#define MULT32_32_P31(a,b) ADD32(SHL32(MULT16_16(SHR((a),16),SHR((b),16)),1), SHR32(128+(opus_int32)(MULT16_16U(((a)&0x0000ffff),((b)&0x0000ffff))>>(16+7)) + SHR32(MULT16_16SU(SHR((a),16),((b)&0x0000ffff)),7) + SHR32(MULT16_16SU(SHR((b),16),((a)&0x0000ffff)),7), 8) )
+#define MULT32_32_P31_ovflw(a,b) ADD32_ovflw(SHL32(MULT16_16(SHR((a),16),SHR((b),16)),1), SHR32(128+(opus_int32)(MULT16_16U(((a)&0x0000ffff),((b)&0x0000ffff))>>(16+7)) + SHR32(MULT16_16SU(SHR((a),16),((b)&0x0000ffff)),7) + SHR32(MULT16_16SU(SHR((b),16),((a)&0x0000ffff)),7), 8) )
+#define MULT32_32_Q32(a,b) ADD32(ADD32(MULT16_16(SHR((a),16),SHR((b),16)), SHR(MULT16_16SU(SHR((a),16),((b)&0x0000ffff)),16)), SHR(MULT16_16SU(SHR((b),16),((a)&0x0000ffff)),16))
+
+/** 16x32 multiplication, followed by a 16-bit shift right. Results fits in 32 bits */
+#define MULT16_32_Q16(a,b) ADD32(MULT16_16((a),SHR32((b),16)), SHR32(MULT16_16SU((a),((b)&0x0000ffff)),16))
+
+#define MULT16_32_P16(a,b) MULT16_32_PX(a,b,16)
+
+#define QCONST16(x,bits) ((opus_val16)(.5+(x)*(((opus_val32)1)<<(bits))))
+#define QCONST32(x,bits) ((opus_val32)(.5+(x)*(((opus_val64)1)<<(bits))))
+#define GCONST2(x,bits) ((celt_glog)(.5+(x)*(((celt_glog)1)<<(bits))))
+#define GCONST(x) GCONST2((x),DB_SHIFT)
+
+#define VERIFY_SHORT(x) ((x)<=32767&&(x)>=-32768)
+#define VERIFY_INT(x) ((x)<=2147483647LL&&(x)>=-2147483648LL)
+#define VERIFY_UINT(x) ((x)<=(2147483647LLU<<1))
+
+#define SHR(a,b) SHR32(a,b)
+#define PSHR(a,b) PSHR32(a,b)
+
+/** Add two 32-bit values, ignore any overflows */
+#define ADD32_ovflw(a,b) (celt_mips+=2,(opus_val32)((opus_uint32)(a)+(opus_uint32)(b)))
+/** Subtract two 32-bit values, ignore any overflows */
+#define SUB32_ovflw(a,b) (celt_mips+=2,(opus_val32)((opus_uint32)(a)-(opus_uint32)(b)))
+/* Avoid MSVC warning C4146: unary minus operator applied to unsigned type */
+/** Negate 32-bit value, ignore any overflows */
+#define NEG32_ovflw(a) (celt_mips+=2,(opus_val32)(0-(opus_uint32)(a)))
+/** 32-bit shift left, ignoring overflows */
+#define SHL32_ovflw(a,shift) ((opus_int32)((opus_uint32)(a)<<(shift)))
+/** 32-bit arithmetic shift right with rounding-to-nearest, ignoring overflows */
+#define PSHR32_ovflw(a,shift) (SHR32(ADD32_ovflw(a, (EXTEND32(1)<<(shift)>>1)),shift))
+
+static OPUS_INLINE short NEG16(int x)
+{
+   int res;
+   if (!VERIFY_SHORT(x))
+   {
+      fprintf (stderr, "NEG16: input is not short: %d\n", (int)x);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res = -x;
+   if (!VERIFY_SHORT(res))
+   {
+      fprintf (stderr, "NEG16: output is not short: %d\n", (int)res);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   celt_mips++;
+   return res;
+}
+static OPUS_INLINE int NEG32(opus_int64 x)
+{
+   opus_int64 res;
+   if (!VERIFY_INT(x))
+   {
+      fprintf (stderr, "NEG16: input is not int: %d\n", (int)x);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res = -x;
+   if (!VERIFY_INT(res))
+   {
+      fprintf (stderr, "NEG16: output is not int: %d\n", (int)res);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   celt_mips+=2;
+   return res;
+}
+
+#define EXTRACT16(x) EXTRACT16_(x, __FILE__, __LINE__)
+static OPUS_INLINE short EXTRACT16_(int x, char *file, int line)
+{
+   int res;
+   if (!VERIFY_SHORT(x))
+   {
+      fprintf (stderr, "EXTRACT16: input is not short: %d in %s: line %d\n", x, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res = x;
+   celt_mips++;
+   return res;
+}
+
+#define EXTEND32(x) EXTEND32_(x, __FILE__, __LINE__)
+static OPUS_INLINE int EXTEND32_(int x, char *file, int line)
+{
+   int res;
+   if (!VERIFY_SHORT(x))
+   {
+      fprintf (stderr, "EXTEND32: input is not short: %d in %s: line %d\n", x, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res = x;
+   celt_mips++;
+   return res;
+}
+
+#define SHR16(a, shift) SHR16_(a, shift, __FILE__, __LINE__)
+static OPUS_INLINE short SHR16_(int a, int shift, char *file, int line)
+{
+   int res;
+   if (!VERIFY_SHORT(a) || !VERIFY_SHORT(shift))
+   {
+      fprintf (stderr, "SHR16: inputs are not short: %d >> %d in %s: line %d\n", a, shift, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res = a>>shift;
+   if (!VERIFY_SHORT(res))
+   {
+      fprintf (stderr, "SHR16: output is not short: %d in %s: line %d\n", res, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   celt_mips++;
+   return res;
+}
+#define SHL16(a, shift) SHL16_(a, shift, __FILE__, __LINE__)
+static OPUS_INLINE short SHL16_(int a, int shift, char *file, int line)
+{
+   opus_int32 res;
+   if (!VERIFY_SHORT(a) || !VERIFY_SHORT(shift))
+   {
+      fprintf (stderr, "SHL16: inputs are not short: %d %d in %s: line %d\n", a, shift, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res = (opus_int32)((opus_uint32)a<<shift);
+   if (!VERIFY_SHORT(res))
+   {
+      fprintf (stderr, "SHL16: output is not short: %d in %s: line %d\n", res, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   celt_mips++;
+   return res;
+}
+
+static OPUS_INLINE int SHR32(opus_int64 a, int shift)
+{
+   opus_int64  res;
+   if (!VERIFY_INT(a) || !VERIFY_SHORT(shift))
+   {
+      fprintf (stderr, "SHR32: inputs are not int: %d %d\n", (int)a, shift);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res = a>>shift;
+   if (!VERIFY_INT(res))
+   {
+      fprintf (stderr, "SHR32: output is not int: %d\n", (int)res);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   celt_mips+=2;
+   return res;
+}
+#define SHL32(a, shift) SHL32_(a, shift, __FILE__, __LINE__)
+static OPUS_INLINE int SHL32_(opus_int64 a, int shift, char *file, int line)
+{
+   opus_int64  res;
+   if (!VERIFY_INT(a) || !VERIFY_SHORT(shift))
+   {
+      fprintf (stderr, "SHL32: inputs are not int: %lld %d in %s: line %d\n", (long long)a, shift, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res = (opus_int64)((opus_uint64)a<<shift);
+   if (!VERIFY_INT(res))
+   {
+      fprintf (stderr, "SHL32: output is not int: %lld<<%d = %lld in %s: line %d\n", (long long)a, shift, (long long)res, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   celt_mips+=2;
+   return res;
+}
+
+#define PSHR32(a,shift) (celt_mips--,SHR32(ADD32((a),(((opus_val32)(1)<<((shift))>>1))),shift))
+#define VSHR32(a, shift) (((shift)>0) ? SHR32(a, shift) : SHL32(a, -(shift)))
+
+#define SHR64(a,shift) (celt_mips++,(a) >> (shift))
+
+#define ROUND16(x,a) (celt_mips--,EXTRACT16(PSHR32((x),(a))))
+#define SROUND16(x,a) (celt_mips--,EXTRACT16(SATURATE(PSHR32(x,a), 32767)));
+
+#define HALF16(x)  (SHR16(x,1))
+#define HALF32(x)  (SHR32(x,1))
+
+#define ADD16(a, b) ADD16_(a, b, __FILE__, __LINE__)
+static OPUS_INLINE short ADD16_(int a, int b, char *file, int line)
+{
+   int res;
+   if (!VERIFY_SHORT(a) || !VERIFY_SHORT(b))
+   {
+      fprintf (stderr, "ADD16: inputs are not short: %d %d in %s: line %d\n", a, b, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res = a+b;
+   if (!VERIFY_SHORT(res))
+   {
+      fprintf (stderr, "ADD16: output is not short: %d+%d=%d in %s: line %d\n", a,b,res, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   celt_mips++;
+   return res;
+}
+
+#define SUB16(a, b) SUB16_(a, b, __FILE__, __LINE__)
+static OPUS_INLINE short SUB16_(int a, int b, char *file, int line)
+{
+   int res;
+   if (!VERIFY_SHORT(a) || !VERIFY_SHORT(b))
+   {
+      fprintf (stderr, "SUB16: inputs are not short: %d %d in %s: line %d\n", a, b, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res = a-b;
+   if (!VERIFY_SHORT(res))
+   {
+      fprintf (stderr, "SUB16: output is not short: %d in %s: line %d\n", res, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   celt_mips++;
+   return res;
+}
+
+#define ADD32(a, b) ADD32_(a, b, __FILE__, __LINE__)
+static OPUS_INLINE int ADD32_(opus_int64 a, opus_int64 b, char *file, int line)
+{
+   opus_int64 res;
+   if (!VERIFY_INT(a) || !VERIFY_INT(b))
+   {
+      fprintf (stderr, "ADD32: inputs are not int: %d %d in %s: line %d\n", (int)a, (int)b, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res = a+b;
+   if (!VERIFY_INT(res))
+   {
+      fprintf (stderr, "ADD32: output is not int: %d in %s: line %d\n", (int)res, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   celt_mips+=2;
+   return res;
+}
+
+#define SUB32(a, b) SUB32_(a, b, __FILE__, __LINE__)
+static OPUS_INLINE int SUB32_(opus_int64 a, opus_int64 b, char *file, int line)
+{
+   opus_int64 res;
+   if (!VERIFY_INT(a) || !VERIFY_INT(b))
+   {
+      fprintf (stderr, "SUB32: inputs are not int: %d %d in %s: line %d\n", (int)a, (int)b, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res = a-b;
+   if (!VERIFY_INT(res))
+   {
+      fprintf (stderr, "SUB32: output is not int: %d in %s: line %d\n", (int)res, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   celt_mips+=2;
+   return res;
+}
+
+#undef UADD32
+#define UADD32(a, b) UADD32_(a, b, __FILE__, __LINE__)
+static OPUS_INLINE unsigned int UADD32_(opus_uint64 a, opus_uint64 b, char *file, int line)
+{
+   opus_uint64 res;
+   if (!VERIFY_UINT(a) || !VERIFY_UINT(b))
+   {
+      fprintf (stderr, "UADD32: inputs are not uint32: %llu %llu in %s: line %d\n", (unsigned long long)a, (unsigned long long)b, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res = a+b;
+   if (!VERIFY_UINT(res))
+   {
+      fprintf (stderr, "UADD32: output is not uint32: %llu in %s: line %d\n", (unsigned long long)res, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   celt_mips+=2;
+   return res;
+}
+
+#undef USUB32
+#define USUB32(a, b) USUB32_(a, b, __FILE__, __LINE__)
+static OPUS_INLINE unsigned int USUB32_(opus_uint64 a, opus_uint64 b, char *file, int line)
+{
+   opus_uint64 res;
+   if (!VERIFY_UINT(a) || !VERIFY_UINT(b))
+   {
+      fprintf (stderr, "USUB32: inputs are not uint32: %llu %llu in %s: line %d\n", (unsigned long long)a, (unsigned long long)b, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   if (a<b)
+   {
+      fprintf (stderr, "USUB32: inputs underflow: %llu < %llu in %s: line %d\n", (unsigned long long)a, (unsigned long long)b, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res = a-b;
+   if (!VERIFY_UINT(res))
+   {
+      fprintf (stderr, "USUB32: output is not uint32: %llu - %llu = %llu in %s: line %d\n", (unsigned long long)a, (unsigned long long)b, (unsigned long long)res, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   celt_mips+=2;
+   return res;
+}
+
+/* result fits in 16 bits */
+static OPUS_INLINE short MULT16_16_16(int a, int b)
+{
+   int res;
+   if (!VERIFY_SHORT(a) || !VERIFY_SHORT(b))
+   {
+      fprintf (stderr, "MULT16_16_16: inputs are not short: %d %d\n", a, b);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res = a*b;
+   if (!VERIFY_SHORT(res))
+   {
+      fprintf (stderr, "MULT16_16_16: output is not short: %d\n", res);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   celt_mips++;
+   return res;
+}
+
+/* result fits in 32 bits */
+static OPUS_INLINE int MULT32_32_32(opus_int64 a, opus_int64 b)
+{
+   opus_int64 res;
+   if (!VERIFY_INT(a) || !VERIFY_INT(b))
+   {
+      fprintf (stderr, "MULT32_32_32: inputs are not int: %lld %lld\n", (long long)a, (long long)b);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res = a*b;
+   if (!VERIFY_INT(res))
+   {
+      fprintf (stderr, "MULT32_32_32: output is not int: %lld\n", (long long)res);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   celt_mips+=5;
+   return res;
+}
+
+static OPUS_INLINE int MULT32_32_Q16(opus_int64 a, opus_int64 b)
+{
+   opus_int64 res;
+   if (!VERIFY_INT(a) || !VERIFY_INT(b))
+   {
+      fprintf (stderr, "MULT32_32_Q16: inputs are not int: %lld %lld\n", (long long)a, (long long)b);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res = ((opus_int64)(a)*(opus_int64)(b)) >> 16;
+   if (!VERIFY_INT(res))
+   {
+      fprintf (stderr, "MULT32_32_Q16: output is not int: %lld*%lld=%lld\n", (long long)a, (long long)b, (long long)res);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   celt_mips+=5;
+   return res;
+}
+
+#define MULT16_16(a, b) MULT16_16_(a, b, __FILE__, __LINE__)
+static OPUS_INLINE int MULT16_16_(int a, int b, char *file, int line)
+{
+   opus_int64 res;
+   if (!VERIFY_SHORT(a) || !VERIFY_SHORT(b))
+   {
+      fprintf (stderr, "MULT16_16: inputs are not short: %d %d in %s: line %d\n", a, b, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res = ((opus_int64)a)*b;
+   if (!VERIFY_INT(res))
+   {
+      fprintf (stderr, "MULT16_16: output is not int: %d in %s: line %d\n", (int)res, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   celt_mips++;
+   return res;
+}
+
+#define MAC16_16(c,a,b)     (celt_mips-=2,ADD32((c),MULT16_16((a),(b))))
+
+#define MULT16_32_QX(a, b, Q) MULT16_32_QX_(a, b, Q, __FILE__, __LINE__)
+static OPUS_INLINE int MULT16_32_QX_(int a, opus_int64 b, int Q, char *file, int line)
+{
+   opus_int64 res;
+   if (!VERIFY_SHORT(a) || !VERIFY_INT(b))
+   {
+      fprintf (stderr, "MULT16_32_Q%d: inputs are not short+int: %d %d in %s: line %d\n", Q, (int)a, (int)b, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   if (ABS32(b)>=((opus_int64)(1)<<(16+Q)))
+   {
+      fprintf (stderr, "MULT16_32_Q%d: second operand too large: %d %d in %s: line %d\n", Q, (int)a, (int)b, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res = (((opus_int64)a)*(opus_int64)b) >> Q;
+   if (!VERIFY_INT(res))
+   {
+      fprintf (stderr, "MULT16_32_Q%d: output is not int: %d*%d=%d in %s: line %d\n", Q, (int)a, (int)b,(int)res, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   if (Q==15)
+      celt_mips+=3;
+   else
+      celt_mips+=4;
+   return res;
+}
+
+#define MULT16_32_PX(a, b, Q) MULT16_32_PX_(a, b, Q, __FILE__, __LINE__)
+static OPUS_INLINE int MULT16_32_PX_(int a, opus_int64 b, int Q, char *file, int line)
+{
+   opus_int64 res;
+   if (!VERIFY_SHORT(a) || !VERIFY_INT(b))
+   {
+      fprintf (stderr, "MULT16_32_P%d: inputs are not short+int: %d %d in %s: line %d\n\n", Q, (int)a, (int)b, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   if (ABS32(b)>=((opus_int64)(1)<<(16+Q)))
+   {
+      fprintf (stderr, "MULT16_32_Q%d: second operand too large: %d %d in %s: line %d\n\n", Q, (int)a, (int)b,file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res = ((((opus_int64)a)*(opus_int64)b) + (((opus_val32)(1)<<Q)>>1))>> Q;
+   if (!VERIFY_INT(res))
+   {
+      fprintf (stderr, "MULT16_32_P%d: output is not int: %d*%d=%d in %s: line %d\n\n", Q, (int)a, (int)b,(int)res, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   if (Q==15)
+      celt_mips+=4;
+   else
+      celt_mips+=5;
+   return res;
+}
+
+#define MULT16_32_Q15(a,b) MULT16_32_QX(a,b,15)
+#define MAC16_32_Q15(c,a,b) (celt_mips-=2,ADD32((c),MULT16_32_Q15((a),(b))))
+#define MAC16_32_Q16(c,a,b) (celt_mips-=2,ADD32((c),MULT16_32_Q16((a),(b))))
+
+static OPUS_INLINE int SATURATE(int a, int b)
+{
+   if (a>b)
+      a=b;
+   if (a<-b)
+      a = -b;
+   celt_mips+=3;
+   return a;
+}
+
+static OPUS_INLINE opus_int16 SATURATE16(opus_int32 a)
+{
+   celt_mips+=3;
+   if (a>32767)
+      return 32767;
+   else if (a<-32768)
+      return -32768;
+   else return a;
+}
+
+static OPUS_INLINE int MULT16_16_Q11_32(int a, int b)
+{
+   opus_int64 res;
+   if (!VERIFY_SHORT(a) || !VERIFY_SHORT(b))
+   {
+      fprintf (stderr, "MULT16_16_Q11: inputs are not short: %d %d\n", a, b);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res = ((opus_int64)a)*b;
+   res >>= 11;
+   if (!VERIFY_INT(res))
+   {
+      fprintf (stderr, "MULT16_16_Q11: output is not short: %d*%d=%d\n", (int)a, (int)b, (int)res);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   celt_mips+=3;
+   return res;
+}
+static OPUS_INLINE short MULT16_16_Q13(int a, int b)
+{
+   opus_int64 res;
+   if (!VERIFY_SHORT(a) || !VERIFY_SHORT(b))
+   {
+      fprintf (stderr, "MULT16_16_Q13: inputs are not short: %d %d\n", a, b);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res = ((opus_int64)a)*b;
+   res >>= 13;
+   if (!VERIFY_SHORT(res))
+   {
+      fprintf (stderr, "MULT16_16_Q13: output is not short: %d*%d=%d\n", a, b, (int)res);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   celt_mips+=3;
+   return res;
+}
+static OPUS_INLINE short MULT16_16_Q14(int a, int b)
+{
+   opus_int64 res;
+   if (!VERIFY_SHORT(a) || !VERIFY_SHORT(b))
+   {
+      fprintf (stderr, "MULT16_16_Q14: inputs are not short: %d %d\n", a, b);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res = ((opus_int64)a)*b;
+   res >>= 14;
+   if (!VERIFY_SHORT(res))
+   {
+      fprintf (stderr, "MULT16_16_Q14: output is not short: %d\n", (int)res);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   celt_mips+=3;
+   return res;
+}
+
+#define MULT16_16_Q15(a, b) MULT16_16_Q15_(a, b, __FILE__, __LINE__)
+static OPUS_INLINE short MULT16_16_Q15_(int a, int b, char *file, int line)
+{
+   opus_int64 res;
+   if (!VERIFY_SHORT(a) || !VERIFY_SHORT(b))
+   {
+      fprintf (stderr, "MULT16_16_Q15: inputs are not short: %d %d in %s: line %d\n", a, b, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res = ((opus_int64)a)*b;
+   res >>= 15;
+   if (!VERIFY_SHORT(res))
+   {
+      fprintf (stderr, "MULT16_16_Q15: output is not short: %d in %s: line %d\n", (int)res, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   celt_mips+=1;
+   return res;
+}
+
+static OPUS_INLINE short MULT16_16_P13(int a, int b)
+{
+   opus_int64 res;
+   if (!VERIFY_SHORT(a) || !VERIFY_SHORT(b))
+   {
+      fprintf (stderr, "MULT16_16_P13: inputs are not short: %d %d\n", a, b);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res = ((opus_int64)a)*b;
+   res += 4096;
+   if (!VERIFY_INT(res))
+   {
+      fprintf (stderr, "MULT16_16_P13: overflow: %d*%d=%d\n", a, b, (int)res);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res >>= 13;
+   if (!VERIFY_SHORT(res))
+   {
+      fprintf (stderr, "MULT16_16_P13: output is not short: %d*%d=%d\n", a, b, (int)res);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   celt_mips+=4;
+   return res;
+}
+static OPUS_INLINE short MULT16_16_P14(int a, int b)
+{
+   opus_int64 res;
+   if (!VERIFY_SHORT(a) || !VERIFY_SHORT(b))
+   {
+      fprintf (stderr, "MULT16_16_P14: inputs are not short: %d %d\n", a, b);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res = ((opus_int64)a)*b;
+   res += 8192;
+   if (!VERIFY_INT(res))
+   {
+      fprintf (stderr, "MULT16_16_P14: overflow: %d*%d=%d\n", a, b, (int)res);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res >>= 14;
+   if (!VERIFY_SHORT(res))
+   {
+      fprintf (stderr, "MULT16_16_P14: output is not short: %d*%d=%d\n", a, b, (int)res);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   celt_mips+=4;
+   return res;
+}
+static OPUS_INLINE short MULT16_16_P15(int a, int b)
+{
+   opus_int64 res;
+   if (!VERIFY_SHORT(a) || !VERIFY_SHORT(b))
+   {
+      fprintf (stderr, "MULT16_16_P15: inputs are not short: %d %d\n", a, b);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res = ((opus_int64)a)*b;
+   res += 16384;
+   if (!VERIFY_INT(res))
+   {
+      fprintf (stderr, "MULT16_16_P15: overflow: %d*%d=%d\n", a, b, (int)res);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res >>= 15;
+   if (!VERIFY_SHORT(res))
+   {
+      fprintf (stderr, "MULT16_16_P15: output is not short: %d*%d=%d\n", a, b, (int)res);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   celt_mips+=2;
+   return res;
+}
+
+#define DIV32_16(a, b) DIV32_16_(a, b, __FILE__, __LINE__)
+
+static OPUS_INLINE int DIV32_16_(opus_int64 a, opus_int64 b, char *file, int line)
+{
+   opus_int64 res;
+   if (b==0)
+   {
+      fprintf(stderr, "DIV32_16: divide by zero: %d/%d in %s: line %d\n", (int)a, (int)b, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+      return 0;
+   }
+   if (!VERIFY_INT(a) || !VERIFY_SHORT(b))
+   {
+      fprintf (stderr, "DIV32_16: inputs are not int/short: %d %d in %s: line %d\n", (int)a, (int)b, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res = a/b;
+   if (!VERIFY_SHORT(res))
+   {
+      fprintf (stderr, "DIV32_16: output is not short: %d / %d = %d in %s: line %d\n", (int)a,(int)b,(int)res, file, line);
+      if (res>32767)
+         res = 32767;
+      if (res<-32768)
+         res = -32768;
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   celt_mips+=35;
+   return res;
+}
+
+#define DIV32(a, b) DIV32_(a, b, __FILE__, __LINE__)
+static OPUS_INLINE int DIV32_(opus_int64 a, opus_int64 b, char *file, int line)
+{
+   opus_int64 res;
+   if (b==0)
+   {
+      fprintf(stderr, "DIV32: divide by zero: %d/%d in %s: line %d\n", (int)a, (int)b, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+      return 0;
+   }
+
+   if (!VERIFY_INT(a) || !VERIFY_INT(b))
+   {
+      fprintf (stderr, "DIV32: inputs are not int/short: %d %d in %s: line %d\n", (int)a, (int)b, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   res = a/b;
+   if (!VERIFY_INT(res))
+   {
+      fprintf (stderr, "DIV32: output is not int: %d in %s: line %d\n", (int)res, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+      celt_assert(0);
+#endif
+   }
+   celt_mips+=70;
+   return res;
+}
+
+static OPUS_INLINE opus_val16 SIG2WORD16_generic(celt_sig x)
+{
+   x = PSHR32(x, SIG_SHIFT);
+   x = MAX32(x, -32768);
+   x = MIN32(x, 32767);
+   return EXTRACT16(x);
+}
+#define SIG2WORD16(x) (SIG2WORD16_generic(x))
+
+
+#undef PRINT_MIPS
+#define PRINT_MIPS(file) do {fprintf (file, "total complexity = %llu MIPS\n", (unsigned long long)celt_mips);} while (0);
+
+#endif
--- a/lib/libopus/celt/fixed_generic.h
+++ b/lib/libopus/celt/fixed_generic.h
@ -0,0 +1,218 @@
+/* Copyright (C) 2007-2009 Xiph.Org Foundation
+   Copyright (C) 2003-2008 Jean-Marc Valin
+   Copyright (C) 2007-2008 CSIRO */
+/**
+   @file fixed_generic.h
+   @brief Generic fixed-point operations
+*/
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef FIXED_GENERIC_H
+#define FIXED_GENERIC_H
+
+/** Multiply a 16-bit signed value by a 16-bit unsigned value. The result is a 32-bit signed value */
+#define MULT16_16SU(a,b) ((opus_val32)(opus_val16)(a)*(opus_val32)(opus_uint16)(b))
+
+/** 16x32 multiplication, followed by a 16-bit shift right. Results fits in 32 bits */
+#if OPUS_FAST_INT64
+#define MULT16_32_Q16(a,b) ((opus_val32)SHR((opus_int64)((opus_val16)(a))*(b),16))
+#else
+#define MULT16_32_Q16(a,b) ADD32(MULT16_16((a),SHR((b),16)), SHR(MULT16_16SU((a),((b)&0x0000ffff)),16))
+#endif
+
+/** 16x32 multiplication, followed by a 16-bit shift right (round-to-nearest). Results fits in 32 bits */
+#if OPUS_FAST_INT64
+#define MULT16_32_P16(a,b) ((opus_val32)PSHR((opus_int64)((opus_val16)(a))*(b),16))
+#else
+#define MULT16_32_P16(a,b) ADD32(MULT16_16((a),SHR((b),16)), PSHR(MULT16_16SU((a),((b)&0x0000ffff)),16))
+#endif
+
+/** 16x32 multiplication, followed by a 15-bit shift right. Results fits in 32 bits */
+#if OPUS_FAST_INT64
+#define MULT16_32_Q15(a,b) ((opus_val32)SHR((opus_int64)((opus_val16)(a))*(b),15))
+#else
+#define MULT16_32_Q15(a,b) ADD32(SHL(MULT16_16((a),SHR((b),16)),1), SHR(MULT16_16SU((a),((b)&0x0000ffff)),15))
+#endif
+
+/** 32x32 multiplication, followed by a 16-bit shift right. Results fits in 32 bits */
+#if OPUS_FAST_INT64
+#define MULT32_32_Q16(a,b) ((opus_val32)SHR((opus_int64)(a)*(opus_int64)(b),16))
+#else
+#define MULT32_32_Q16(a,b) (ADD32(ADD32(ADD32((opus_val32)(SHR32(((opus_uint32)((a)&0x0000ffff)*(opus_uint32)((b)&0x0000ffff)),16)), MULT16_16SU(SHR32(a,16),((b)&0x0000ffff))), MULT16_16SU(SHR32(b,16),((a)&0x0000ffff))), SHL32(MULT16_16(SHR32(a,16),SHR32(b,16)),16)))
+#endif
+
+/** 32x32 multiplication, followed by a 31-bit shift right. Results fits in 32 bits */
+#if OPUS_FAST_INT64
+#define MULT32_32_Q31(a,b) ((opus_val32)SHR((opus_int64)(a)*(opus_int64)(b),31))
+#else
+#define MULT32_32_Q31(a,b) ADD32(ADD32(SHL(MULT16_16(SHR((a),16),SHR((b),16)),1), SHR(MULT16_16SU(SHR((a),16),((b)&0x0000ffff)),15)), SHR(MULT16_16SU(SHR((b),16),((a)&0x0000ffff)),15))
+#endif
+
+/** 32x32 multiplication, followed by a 31-bit shift right (with rounding). Results fits in 32 bits */
+#if OPUS_FAST_INT64
+#define MULT32_32_P31(a,b) ((opus_val32)SHR(1073741824+(opus_int64)(a)*(opus_int64)(b),31))
+#define MULT32_32_P31_ovflw(a,b) MULT32_32_P31(a,b)
+#else
+#define MULT16_16U(a,b) ((opus_uint32)(a)*(opus_uint32)(b))
+#define MULT32_32_P31(a,b) ADD32(SHL(MULT16_16(SHR((a),16),SHR((b),16)),1), SHR32(128+(opus_int32)SHR(MULT16_16U(((a)&0x0000ffff),((b)&0x0000ffff)),16+7) + SHR32(MULT16_16SU(SHR((a),16),((b)&0x0000ffff)),7) + SHR32(MULT16_16SU(SHR((b),16),((a)&0x0000ffff)),7), 8) )
+#define MULT32_32_P31_ovflw(a,b) ADD32_ovflw(SHL(MULT16_16(SHR((a),16),SHR((b),16)),1), SHR32(128+(opus_int32)SHR(MULT16_16U(((a)&0x0000ffff),((b)&0x0000ffff)),16+7) + SHR32(MULT16_16SU(SHR((a),16),((b)&0x0000ffff)),7) + SHR32(MULT16_16SU(SHR((b),16),((a)&0x0000ffff)),7), 8) )
+#endif
+
+/** 32x32 multiplication, followed by a 32-bit shift right. Results fits in 32 bits */
+#if OPUS_FAST_INT64
+#define MULT32_32_Q32(a,b) ((opus_val32)SHR((opus_int64)(a)*(opus_int64)(b),32))
+#else
+#define MULT32_32_Q32(a,b) ADD32(ADD32(MULT16_16(SHR((a),16),SHR((b),16)), SHR(MULT16_16SU(SHR((a),16),((b)&0x0000ffff)),16)), SHR(MULT16_16SU(SHR((b),16),((a)&0x0000ffff)),16))
+#endif
+
+/** Compile-time conversion of float constant to 16-bit value */
+#define QCONST16(x,bits) ((opus_val16)(.5+(x)*(((opus_val32)1)<<(bits))))
+
+/** Compile-time conversion of float constant to 32-bit value */
+#define QCONST32(x,bits) ((opus_val32)(.5+(x)*(((opus_int64)1)<<(bits))))
+
+/** Compile-time conversion of float constant to log gain value */
+#define GCONST2(x,bits) ((celt_glog)(.5+(x)*(((celt_glog)1)<<(bits))))
+
+/** Compile-time conversion of float constant to DB_SHIFT log gain value */
+#define GCONST(x) GCONST2((x),DB_SHIFT)
+
+/** Negate a 16-bit value */
+#define NEG16(x) (-(x))
+/** Negate a 32-bit value */
+#define NEG32(x) (-(x))
+
+/** Change a 32-bit value into a 16-bit value. The value is assumed to fit in 16-bit, otherwise the result is undefined */
+#define EXTRACT16(x) ((opus_val16)(x))
+/** Change a 16-bit value into a 32-bit value */
+#define EXTEND32(x) ((opus_val32)(x))
+
+/** Arithmetic shift-right of a 16-bit value */
+#define SHR16(a,shift) ((a) >> (shift))
+/** Arithmetic shift-left of a 16-bit value */
+#define SHL16(a,shift) ((opus_int16)((opus_uint16)(a)<<(shift)))
+/** Arithmetic shift-right of a 32-bit value */
+#define SHR32(a,shift) ((a) >> (shift))
+/** Arithmetic shift-left of a 32-bit value */
+#define SHL32(a,shift) ((opus_int32)((opus_uint32)(a)<<(shift)))
+
+/** 32-bit arithmetic shift right with rounding-to-nearest instead of rounding down */
+#define PSHR32(a,shift) (SHR32((a)+((EXTEND32(1)<<((shift))>>1)),shift))
+/** 32-bit arithmetic shift right where the argument can be negative */
+#define VSHR32(a, shift) (((shift)>0) ? SHR32(a, shift) : SHL32(a, -(shift)))
+
+/** Arithmetic shift-right of a 64-bit value */
+#define SHR64(a,shift) ((a) >> (shift))
+
+/** "RAW" macros, should not be used outside of this header file */
+#define SHR(a,shift) ((a) >> (shift))
+#define SHL(a,shift) SHL32(a,shift)
+#define PSHR(a,shift) (SHR((a)+((EXTEND32(1)<<((shift))>>1)),shift))
+#define SATURATE(x,a) (((x)>(a) ? (a) : (x)<-(a) ? -(a) : (x)))
+
+#define SATURATE16(x) (EXTRACT16((x)>32767 ? 32767 : (x)<-32768 ? -32768 : (x)))
+
+/** Shift by a and round-to-nearest 32-bit value. Result is a 16-bit value */
+#define ROUND16(x,a) (EXTRACT16(PSHR32((x),(a))))
+/** Shift by a and round-to-nearest 32-bit value. Result is a saturated 16-bit value */
+#define SROUND16(x,a) EXTRACT16(SATURATE(PSHR32(x,a), 32767));
+
+/** Divide by two */
+#define HALF16(x)  (SHR16(x,1))
+#define HALF32(x)  (SHR32(x,1))
+
+/** Add two 16-bit values */
+#define ADD16(a,b) ((opus_val16)((opus_val16)(a)+(opus_val16)(b)))
+/** Subtract two 16-bit values */
+#define SUB16(a,b) ((opus_val16)(a)-(opus_val16)(b))
+/** Add two 32-bit values */
+#define ADD32(a,b) ((opus_val32)(a)+(opus_val32)(b))
+/** Subtract two 32-bit values */
+#define SUB32(a,b) ((opus_val32)(a)-(opus_val32)(b))
+
+/** Add two 32-bit values, ignore any overflows */
+#define ADD32_ovflw(a,b) ((opus_val32)((opus_uint32)(a)+(opus_uint32)(b)))
+/** Subtract two 32-bit values, ignore any overflows */
+#define SUB32_ovflw(a,b) ((opus_val32)((opus_uint32)(a)-(opus_uint32)(b)))
+/* Avoid MSVC warning C4146: unary minus operator applied to unsigned type */
+/** Negate 32-bit value, ignore any overflows */
+#define NEG32_ovflw(a) ((opus_val32)(0-(opus_uint32)(a)))
+/** 32-bit shift left, ignoring overflows */
+#define SHL32_ovflw(a,shift) SHL32(a,shift)
+/** 32-bit arithmetic shift right with rounding-to-nearest, ignoring overflows */
+#define PSHR32_ovflw(a,shift) (SHR32(ADD32_ovflw(a, (EXTEND32(1)<<(shift)>>1)),shift))
+
+/** 16x16 multiplication where the result fits in 16 bits */
+#define MULT16_16_16(a,b)     ((((opus_val16)(a))*((opus_val16)(b))))
+
+/** 32x32 multiplication where the result fits in 32 bits */
+#define MULT32_32_32(a,b)     ((((opus_val32)(a))*((opus_val32)(b))))
+
+/* (opus_val32)(opus_val16) gives TI compiler a hint that it's 16x16->32 multiply */
+/** 16x16 multiplication where the result fits in 32 bits */
+#define MULT16_16(a,b)     (((opus_val32)(opus_val16)(a))*((opus_val32)(opus_val16)(b)))
+
+/** 16x16 multiply-add where the result fits in 32 bits */
+#define MAC16_16(c,a,b) (ADD32((c),MULT16_16((a),(b))))
+/** 16x32 multiply, followed by a 15-bit shift right and 32-bit add.
+    b must fit in 31 bits.
+    Result fits in 32 bits. */
+#define MAC16_32_Q15(c,a,b) ADD32((c),ADD32(MULT16_16((a),SHR((b),15)), SHR(MULT16_16((a),((b)&0x00007fff)),15)))
+
+/** 16x32 multiplication, followed by a 16-bit shift right and 32-bit add.
+    Results fits in 32 bits */
+#define MAC16_32_Q16(c,a,b) ADD32((c),ADD32(MULT16_16((a),SHR((b),16)), SHR(MULT16_16SU((a),((b)&0x0000ffff)),16)))
+
+#define MULT16_16_Q11_32(a,b) (SHR(MULT16_16((a),(b)),11))
+#define MULT16_16_Q11(a,b) (SHR(MULT16_16((a),(b)),11))
+#define MULT16_16_Q13(a,b) (SHR(MULT16_16((a),(b)),13))
+#define MULT16_16_Q14(a,b) (SHR(MULT16_16((a),(b)),14))
+#define MULT16_16_Q15(a,b) (SHR(MULT16_16((a),(b)),15))
+
+#define MULT16_16_P13(a,b) (SHR(ADD32(4096,MULT16_16((a),(b))),13))
+#define MULT16_16_P14(a,b) (SHR(ADD32(8192,MULT16_16((a),(b))),14))
+#define MULT16_16_P15(a,b) (SHR(ADD32(16384,MULT16_16((a),(b))),15))
+
+/** Divide a 32-bit value by a 16-bit value. Result fits in 16 bits */
+#define DIV32_16(a,b) ((opus_val16)(((opus_val32)(a))/((opus_val16)(b))))
+
+/** Divide a 32-bit value by a 32-bit value. Result fits in 32 bits */
+#define DIV32(a,b) (((opus_val32)(a))/((opus_val32)(b)))
+
+#if defined(__mips)
+#include "mips/fixed_generic_mipsr1.h"
+#endif
+
+static OPUS_INLINE opus_val16 SIG2WORD16_generic(celt_sig x)
+{
+   x = PSHR32(x, SIG_SHIFT);
+   x = MAX32(x, -32768);
+   x = MIN32(x, 32767);
+   return EXTRACT16(x);
+}
+#define SIG2WORD16(x) (SIG2WORD16_generic(x))
+
+#endif
--- a/lib/libopus/celt/float_cast.h
+++ b/lib/libopus/celt/float_cast.h
@ -0,0 +1,176 @@
+/* Copyright (C) 2001 Erik de Castro Lopo <erikd AT mega-nerd DOT com> */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+/* Version 1.1 */
+
+#ifndef FLOAT_CAST_H
+#define FLOAT_CAST_H
+
+
+#include "arch.h"
+
+/*============================================================================
+**      On Intel Pentium processors (especially PIII and probably P4), converting
+**      from float to int is very slow. To meet the C specs, the code produced by
+**      most C compilers targeting Pentium needs to change the FPU rounding mode
+**      before the float to int conversion is performed.
+**
+**      Changing the FPU rounding mode causes the FPU pipeline to be flushed. It
+**      is this flushing of the pipeline which is so slow.
+**
+**      Fortunately the ISO C99 specifications define the functions lrint, lrintf,
+**      llrint and llrintf which fix this problem as a side effect.
+**
+**      On Unix-like systems, the configure process should have detected the
+**      presence of these functions. If they weren't found we have to replace them
+**      here with a standard C cast.
+*/
+
+/*
+**      The C99 prototypes for lrint and lrintf are as follows:
+**
+**              long int lrintf (float x) ;
+**              long int lrint  (double x) ;
+*/
+
+/*      The presence of the required functions are detected during the configure
+**      process and the values HAVE_LRINT and HAVE_LRINTF are set accordingly in
+**      the config.h file.
+*/
+
+/* With GCC, when SSE is available, the fastest conversion is cvtss2si. */
+#if defined(__GNUC__) && defined(__SSE__)
+
+#include <xmmintrin.h>
+static OPUS_INLINE opus_int32 float2int(float x) {return _mm_cvt_ss2si(_mm_set_ss(x));}
+
+#elif (defined(_MSC_VER) && _MSC_VER >= 1400) && (defined(_M_X64) || (defined(_M_IX86_FP) && _M_IX86_FP >= 1))
+
+        #include <xmmintrin.h>
+        static OPUS_INLINE opus_int32 float2int(float value)
+        {
+                /* _mm_load_ss will generate same code as _mm_set_ss
+                ** in _MSC_VER >= 1914 /02 so keep __mm_load__ss
+                ** for backward compatibility.
+                */
+                return _mm_cvtss_si32(_mm_load_ss(&value));
+        }
+
+#elif (defined(_MSC_VER) && _MSC_VER >= 1400) && defined (_M_IX86)
+
+        #include <math.h>
+
+        /*      Win32 doesn't seem to have these functions.
+        **      Therefore implement OPUS_INLINE versions of these functions here.
+        */
+
+        static OPUS_INLINE opus_int32
+        float2int (float flt)
+        {       int intgr;
+
+                _asm
+                {       fld flt
+                        fistp intgr
+                } ;
+
+                return intgr ;
+        }
+#elif defined(__aarch64__)
+
+        #include <arm_neon.h>
+        static OPUS_INLINE opus_int32 float2int(float flt)
+        {
+                return vcvtns_s32_f32(flt);
+        }
+
+#elif defined(HAVE_LRINTF) && defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L
+
+/*      These defines enable functionality introduced with the 1999 ISO C
+**      standard. They must be defined before the inclusion of math.h to
+**      engage them. If optimisation is enabled, these functions will be
+**      inlined. With optimisation switched off, you have to link in the
+**      maths library using -lm.
+*/
+
+#define _ISOC9X_SOURCE 1
+#define _ISOC99_SOURCE 1
+
+#define __USE_ISOC9X 1
+#define __USE_ISOC99 1
+
+#include <math.h>
+#define float2int(x) lrintf(x)
+
+#elif defined(HAVE_LRINT) && defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L
+
+#define _ISOC9X_SOURCE 1
+#define _ISOC99_SOURCE 1
+
+#define __USE_ISOC9X 1
+#define __USE_ISOC99 1
+
+#include <math.h>
+#define float2int(x) lrint(x)
+
+#else
+
+#if (defined(__GNUC__) && defined(__STDC__) && __STDC__ && __STDC_VERSION__ >= 199901L)
+        /* supported by gcc in C99 mode, but not by all other compilers */
+        #warning "Don't have the functions lrint() and lrintf ()."
+        #warning "Replacing these functions with a standard C cast."
+#endif /* __STDC_VERSION__ >= 199901L */
+        #include <math.h>
+        #define float2int(flt) ((int)(floor(.5+flt)))
+#endif
+
+#ifndef DISABLE_FLOAT_API
+static OPUS_INLINE opus_int16 FLOAT2INT16(float x)
+{
+   x = x*CELT_SIG_SCALE;
+   x = MAX32(x, -32768);
+   x = MIN32(x, 32767);
+   return (opus_int16)float2int(x);
+}
+
+static OPUS_INLINE opus_int32 FLOAT2INT24(float x)
+{
+   x = x*(CELT_SIG_SCALE*256.f);
+   x = MAX32(x, -16777216);
+   x = MIN32(x, 16777216);
+   return float2int(x);
+}
+#ifdef FIXED_POINT
+static OPUS_INLINE opus_int32 FLOAT2SIG(float x)
+{
+   x = x*((opus_int32)32768<<SIG_SHIFT);
+   x = MAX32(x, -(65536<<SIG_SHIFT));
+   x = MIN32(x, 65536<<SIG_SHIFT);
+   return float2int(x);
+}
+#endif
+#endif /* DISABLE_FLOAT_API */
+
+#endif /* FLOAT_CAST_H */
--- a/lib/libopus/celt/kiss_fft.c
+++ b/lib/libopus/celt/kiss_fft.c
@ -0,0 +1,650 @@
+/*Copyright (c) 2003-2004, Mark Borgerding
+  Lots of modifications by Jean-Marc Valin
+  Copyright (c) 2005-2007, Xiph.Org Foundation
+  Copyright (c) 2008,      Xiph.Org Foundation, CSIRO
+
+  All rights reserved.
+
+  Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are met:
+
+    * Redistributions of source code must retain the above copyright notice,
+       this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above copyright notice,
+       this list of conditions and the following disclaimer in the
+       documentation and/or other materials provided with the distribution.
+
+  THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+  AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+  IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+  ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+  LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+  CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+  SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+  INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+  CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+  ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+  POSSIBILITY OF SUCH DAMAGE.*/
+
+/* This code is originally from Mark Borgerding's KISS-FFT but has been
+   heavily modified to better suit Opus */
+
+#ifndef SKIP_CONFIG_H
+#  ifdef HAVE_CONFIG_H
+#    include "config.h"
+#  endif
+#endif
+
+#include "_kiss_fft_guts.h"
+#include "arch.h"
+#include "os_support.h"
+#include "mathops.h"
+#include "stack_alloc.h"
+
+#ifndef M_PI
+#define M_PI 3.141592653
+#endif
+
+/* The guts header contains all the multiplication and addition macros that are defined for
+   complex numbers.  It also declares the kf_ internal functions.
+*/
+
+static void kf_bfly2(
+                     kiss_fft_cpx * Fout,
+                     int m,
+                     int N
+                    )
+{
+   kiss_fft_cpx * Fout2;
+   int i;
+   (void)m;
+#ifdef CUSTOM_MODES
+   if (m==1)
+   {
+      celt_assert(m==1);
+      for (i=0;i<N;i++)
+      {
+         kiss_fft_cpx t;
+         Fout2 = Fout + 1;
+         t = *Fout2;
+         C_SUB( *Fout2 ,  *Fout , t );
+         C_ADDTO( *Fout ,  t );
+         Fout += 2;
+      }
+   } else
+#endif
+   {
+      celt_coef tw;
+      tw = QCONST32(0.7071067812f, COEF_SHIFT-1);
+      /* We know that m==4 here because the radix-2 is just after a radix-4 */
+      celt_assert(m==4);
+      for (i=0;i<N;i++)
+      {
+         kiss_fft_cpx t;
+         Fout2 = Fout + 4;
+         t = Fout2[0];
+         C_SUB( Fout2[0] ,  Fout[0] , t );
+         C_ADDTO( Fout[0] ,  t );
+
+         t.r = S_MUL(ADD32_ovflw(Fout2[1].r, Fout2[1].i), tw);
+         t.i = S_MUL(SUB32_ovflw(Fout2[1].i, Fout2[1].r), tw);
+         C_SUB( Fout2[1] ,  Fout[1] , t );
+         C_ADDTO( Fout[1] ,  t );
+
+         t.r = Fout2[2].i;
+         t.i = NEG32_ovflw(Fout2[2].r);
+         C_SUB( Fout2[2] ,  Fout[2] , t );
+         C_ADDTO( Fout[2] ,  t );
+
+         t.r = S_MUL(SUB32_ovflw(Fout2[3].i, Fout2[3].r), tw);
+         t.i = S_MUL(NEG32_ovflw(ADD32_ovflw(Fout2[3].i, Fout2[3].r)), tw);
+         C_SUB( Fout2[3] ,  Fout[3] , t );
+         C_ADDTO( Fout[3] ,  t );
+         Fout += 8;
+      }
+   }
+}
+
+static void kf_bfly4(
+                     kiss_fft_cpx * Fout,
+                     const size_t fstride,
+                     const kiss_fft_state *st,
+                     int m,
+                     int N,
+                     int mm
+                    )
+{
+   int i;
+
+   if (m==1)
+   {
+      /* Degenerate case where all the twiddles are 1. */
+      for (i=0;i<N;i++)
+      {
+         kiss_fft_cpx scratch0, scratch1;
+
+         C_SUB( scratch0 , *Fout, Fout[2] );
+         C_ADDTO(*Fout, Fout[2]);
+         C_ADD( scratch1 , Fout[1] , Fout[3] );
+         C_SUB( Fout[2], *Fout, scratch1 );
+         C_ADDTO( *Fout , scratch1 );
+         C_SUB( scratch1 , Fout[1] , Fout[3] );
+
+         Fout[1].r = ADD32_ovflw(scratch0.r, scratch1.i);
+         Fout[1].i = SUB32_ovflw(scratch0.i, scratch1.r);
+         Fout[3].r = SUB32_ovflw(scratch0.r, scratch1.i);
+         Fout[3].i = ADD32_ovflw(scratch0.i, scratch1.r);
+         Fout+=4;
+      }
+   } else {
+      int j;
+      kiss_fft_cpx scratch[6];
+      const kiss_twiddle_cpx *tw1,*tw2,*tw3;
+      const int m2=2*m;
+      const int m3=3*m;
+      kiss_fft_cpx * Fout_beg = Fout;
+      for (i=0;i<N;i++)
+      {
+         Fout = Fout_beg + i*mm;
+         tw3 = tw2 = tw1 = st->twiddles;
+         /* m is guaranteed to be a multiple of 4. */
+         for (j=0;j<m;j++)
+         {
+            C_MUL(scratch[0],Fout[m] , *tw1 );
+            C_MUL(scratch[1],Fout[m2] , *tw2 );
+            C_MUL(scratch[2],Fout[m3] , *tw3 );
+
+            C_SUB( scratch[5] , *Fout, scratch[1] );
+            C_ADDTO(*Fout, scratch[1]);
+            C_ADD( scratch[3] , scratch[0] , scratch[2] );
+            C_SUB( scratch[4] , scratch[0] , scratch[2] );
+            C_SUB( Fout[m2], *Fout, scratch[3] );
+            tw1 += fstride;
+            tw2 += fstride*2;
+            tw3 += fstride*3;
+            C_ADDTO( *Fout , scratch[3] );
+
+            Fout[m].r = ADD32_ovflw(scratch[5].r, scratch[4].i);
+            Fout[m].i = SUB32_ovflw(scratch[5].i, scratch[4].r);
+            Fout[m3].r = SUB32_ovflw(scratch[5].r, scratch[4].i);
+            Fout[m3].i = ADD32_ovflw(scratch[5].i, scratch[4].r);
+            ++Fout;
+         }
+      }
+   }
+}
+
+
+#ifndef RADIX_TWO_ONLY
+
+static void kf_bfly3(
+                     kiss_fft_cpx * Fout,
+                     const size_t fstride,
+                     const kiss_fft_state *st,
+                     int m,
+                     int N,
+                     int mm
+                    )
+{
+   int i;
+   size_t k;
+   const size_t m2 = 2*m;
+   const kiss_twiddle_cpx *tw1,*tw2;
+   kiss_fft_cpx scratch[5];
+   kiss_twiddle_cpx epi3;
+
+   kiss_fft_cpx * Fout_beg = Fout;
+#ifdef FIXED_POINT
+   /*epi3.r = -16384;*/ /* Unused */
+   epi3.i = -QCONST32(0.86602540f, COEF_SHIFT-1);
+#else
+   epi3 = st->twiddles[fstride*m];
+#endif
+   for (i=0;i<N;i++)
+   {
+      Fout = Fout_beg + i*mm;
+      tw1=tw2=st->twiddles;
+      /* For non-custom modes, m is guaranteed to be a multiple of 4. */
+      k=m;
+      do {
+
+         C_MUL(scratch[1],Fout[m] , *tw1);
+         C_MUL(scratch[2],Fout[m2] , *tw2);
+
+         C_ADD(scratch[3],scratch[1],scratch[2]);
+         C_SUB(scratch[0],scratch[1],scratch[2]);
+         tw1 += fstride;
+         tw2 += fstride*2;
+
+         Fout[m].r = SUB32_ovflw(Fout->r, HALF_OF(scratch[3].r));
+         Fout[m].i = SUB32_ovflw(Fout->i, HALF_OF(scratch[3].i));
+
+         C_MULBYSCALAR( scratch[0] , epi3.i );
+
+         C_ADDTO(*Fout,scratch[3]);
+
+         Fout[m2].r = ADD32_ovflw(Fout[m].r, scratch[0].i);
+         Fout[m2].i = SUB32_ovflw(Fout[m].i, scratch[0].r);
+
+         Fout[m].r = SUB32_ovflw(Fout[m].r, scratch[0].i);
+         Fout[m].i = ADD32_ovflw(Fout[m].i, scratch[0].r);
+
+         ++Fout;
+      } while(--k);
+   }
+}
+
+
+#ifndef OVERRIDE_kf_bfly5
+static void kf_bfly5(
+                     kiss_fft_cpx * Fout,
+                     const size_t fstride,
+                     const kiss_fft_state *st,
+                     int m,
+                     int N,
+                     int mm
+                    )
+{
+   kiss_fft_cpx *Fout0,*Fout1,*Fout2,*Fout3,*Fout4;
+   int i, u;
+   kiss_fft_cpx scratch[13];
+   const kiss_twiddle_cpx *tw;
+   kiss_twiddle_cpx ya,yb;
+   kiss_fft_cpx * Fout_beg = Fout;
+
+#ifdef FIXED_POINT
+   ya.r = QCONST32(0.30901699f, COEF_SHIFT-1);
+   ya.i = -QCONST32(0.95105652f, COEF_SHIFT-1);
+   yb.r = -QCONST32(0.80901699f, COEF_SHIFT-1);
+   yb.i = -QCONST32(0.58778525f, COEF_SHIFT-1);
+#else
+   ya = st->twiddles[fstride*m];
+   yb = st->twiddles[fstride*2*m];
+#endif
+   tw=st->twiddles;
+
+   for (i=0;i<N;i++)
+   {
+      Fout = Fout_beg + i*mm;
+      Fout0=Fout;
+      Fout1=Fout0+m;
+      Fout2=Fout0+2*m;
+      Fout3=Fout0+3*m;
+      Fout4=Fout0+4*m;
+
+      /* For non-custom modes, m is guaranteed to be a multiple of 4. */
+      for ( u=0; u<m; ++u ) {
+         scratch[0] = *Fout0;
+
+         C_MUL(scratch[1] ,*Fout1, tw[u*fstride]);
+         C_MUL(scratch[2] ,*Fout2, tw[2*u*fstride]);
+         C_MUL(scratch[3] ,*Fout3, tw[3*u*fstride]);
+         C_MUL(scratch[4] ,*Fout4, tw[4*u*fstride]);
+
+         C_ADD( scratch[7],scratch[1],scratch[4]);
+         C_SUB( scratch[10],scratch[1],scratch[4]);
+         C_ADD( scratch[8],scratch[2],scratch[3]);
+         C_SUB( scratch[9],scratch[2],scratch[3]);
+
+         Fout0->r = ADD32_ovflw(Fout0->r, ADD32_ovflw(scratch[7].r, scratch[8].r));
+         Fout0->i = ADD32_ovflw(Fout0->i, ADD32_ovflw(scratch[7].i, scratch[8].i));
+
+         scratch[5].r = ADD32_ovflw(scratch[0].r, ADD32_ovflw(S_MUL(scratch[7].r,ya.r), S_MUL(scratch[8].r,yb.r)));
+         scratch[5].i = ADD32_ovflw(scratch[0].i, ADD32_ovflw(S_MUL(scratch[7].i,ya.r), S_MUL(scratch[8].i,yb.r)));
+
+         scratch[6].r =  ADD32_ovflw(S_MUL(scratch[10].i,ya.i), S_MUL(scratch[9].i,yb.i));
+         scratch[6].i = NEG32_ovflw(ADD32_ovflw(S_MUL(scratch[10].r,ya.i), S_MUL(scratch[9].r,yb.i)));
+
+         C_SUB(*Fout1,scratch[5],scratch[6]);
+         C_ADD(*Fout4,scratch[5],scratch[6]);
+
+         scratch[11].r = ADD32_ovflw(scratch[0].r, ADD32_ovflw(S_MUL(scratch[7].r,yb.r), S_MUL(scratch[8].r,ya.r)));
+         scratch[11].i = ADD32_ovflw(scratch[0].i, ADD32_ovflw(S_MUL(scratch[7].i,yb.r), S_MUL(scratch[8].i,ya.r)));
+         scratch[12].r = SUB32_ovflw(S_MUL(scratch[9].i,ya.i), S_MUL(scratch[10].i,yb.i));
+         scratch[12].i = SUB32_ovflw(S_MUL(scratch[10].r,yb.i), S_MUL(scratch[9].r,ya.i));
+
+         C_ADD(*Fout2,scratch[11],scratch[12]);
+         C_SUB(*Fout3,scratch[11],scratch[12]);
+
+         ++Fout0;++Fout1;++Fout2;++Fout3;++Fout4;
+      }
+   }
+}
+#endif /* OVERRIDE_kf_bfly5 */
+
+
+#endif
+
+
+#ifdef CUSTOM_MODES
+
+static
+void compute_bitrev_table(
+         int Fout,
+         opus_int16 *f,
+         const size_t fstride,
+         int in_stride,
+         opus_int16 * factors,
+         const kiss_fft_state *st
+            )
+{
+   const int p=*factors++; /* the radix  */
+   const int m=*factors++; /* stage's fft length/p */
+
+    /*printf ("fft %d %d %d %d %d %d\n", p*m, m, p, s2, fstride*in_stride, N);*/
+   if (m==1)
+   {
+      int j;
+      for (j=0;j<p;j++)
+      {
+         *f = Fout+j;
+         f += fstride*in_stride;
+      }
+   } else {
+      int j;
+      for (j=0;j<p;j++)
+      {
+         compute_bitrev_table( Fout , f, fstride*p, in_stride, factors,st);
+         f += fstride*in_stride;
+         Fout += m;
+      }
+   }
+}
+
+/*  facbuf is populated by p1,m1,p2,m2, ...
+    where
+    p[i] * m[i] = m[i-1]
+    m0 = n                  */
+static
+int kf_factor(int n,opus_int16 * facbuf)
+{
+    int p=4;
+    int i;
+    int stages=0;
+    int nbak = n;
+
+    /*factor out powers of 4, powers of 2, then any remaining primes */
+    do {
+        while (n % p) {
+            switch (p) {
+                case 4: p = 2; break;
+                case 2: p = 3; break;
+                default: p += 2; break;
+            }
+            if (p>32000 || (opus_int32)p*(opus_int32)p > n)
+                p = n;          /* no more factors, skip to end */
+        }
+        n /= p;
+#ifdef RADIX_TWO_ONLY
+        if (p!=2 && p != 4)
+#else
+        if (p>5)
+#endif
+        {
+           return 0;
+        }
+        facbuf[2*stages] = p;
+        if (p==2 && stages > 1)
+        {
+           facbuf[2*stages] = 4;
+           facbuf[2] = 2;
+        }
+        stages++;
+    } while (n > 1);
+    n = nbak;
+    /* Reverse the order to get the radix 4 at the end, so we can use the
+       fast degenerate case. It turns out that reversing the order also
+       improves the noise behaviour. */
+    for (i=0;i<stages/2;i++)
+    {
+       int tmp;
+       tmp = facbuf[2*i];
+       facbuf[2*i] = facbuf[2*(stages-i-1)];
+       facbuf[2*(stages-i-1)] = tmp;
+    }
+    for (i=0;i<stages;i++)
+    {
+        n /= facbuf[2*i];
+        facbuf[2*i+1] = n;
+    }
+    return 1;
+}
+
+static void compute_twiddles(kiss_twiddle_cpx *twiddles, int nfft)
+{
+   int i;
+#ifdef FIXED_POINT
+   for (i=0;i<nfft;++i) {
+      opus_val32 phase = -i;
+#ifdef ENABLE_QEXT
+      twiddles[i].r = (int)MIN32(2147483647, floor(.5+2147483648*cos((2*M_PI/nfft)*phase)));
+      twiddles[i].i = (int)MIN32(2147483647, floor(.5+2147483648*sin((2*M_PI/nfft)*phase)));
+#else
+      kf_cexp2(twiddles+i, DIV32(SHL32(phase,17),nfft));
+#endif
+   }
+#else
+   for (i=0;i<nfft;++i) {
+      const double pi=3.14159265358979323846264338327;
+      double phase = ( -2*pi /nfft ) * i;
+      kf_cexp(twiddles+i, phase );
+   }
+#endif
+}
+
+int opus_fft_alloc_arch_c(kiss_fft_state *st) {
+   (void)st;
+   return 0;
+}
+
+/*
+ *
+ * Allocates all necessary storage space for the fft and ifft.
+ * The return value is a contiguous block of memory.  As such,
+ * It can be freed with free().
+ * */
+kiss_fft_state *opus_fft_alloc_twiddles(int nfft,void * mem,size_t * lenmem,
+                                        const kiss_fft_state *base, int arch)
+{
+    kiss_fft_state *st=NULL;
+    size_t memneeded = sizeof(struct kiss_fft_state); /* twiddle factors*/
+
+    if ( lenmem==NULL ) {
+        st = ( kiss_fft_state*)KISS_FFT_MALLOC( memneeded );
+    }else{
+        if (mem != NULL && *lenmem >= memneeded)
+            st = (kiss_fft_state*)mem;
+        *lenmem = memneeded;
+    }
+    if (st) {
+        opus_int16 *bitrev;
+        kiss_twiddle_cpx *twiddles;
+
+        st->nfft=nfft;
+#ifdef FIXED_POINT
+        st->scale_shift = celt_ilog2(st->nfft);
+# ifdef ENABLE_QEXT
+        if (st->nfft == 1<<st->scale_shift)
+           st->scale = QCONST32(1.0f, 30);
+        else
+           st->scale = (((opus_int64)1073741824<<st->scale_shift)+st->nfft/2)/st->nfft;
+# else
+        if (st->nfft == 1<<st->scale_shift)
+           st->scale = Q15ONE;
+        else
+           st->scale = (1073741824+st->nfft/2)/st->nfft>>(15-st->scale_shift);
+# endif
+#else
+        st->scale = 1.f/nfft;
+#endif
+        if (base != NULL)
+        {
+           st->twiddles = base->twiddles;
+           st->shift = 0;
+           while (st->shift < 32 && nfft<<st->shift != base->nfft)
+              st->shift++;
+           if (st->shift>=32)
+              goto fail;
+        } else {
+           st->twiddles = twiddles = (kiss_twiddle_cpx*)KISS_FFT_MALLOC(sizeof(kiss_twiddle_cpx)*nfft);
+           compute_twiddles(twiddles, nfft);
+           st->shift = -1;
+        }
+        if (!kf_factor(nfft,st->factors))
+        {
+           goto fail;
+        }
+
+        /* bitrev */
+        st->bitrev = bitrev = (opus_int16*)KISS_FFT_MALLOC(sizeof(opus_int16)*nfft);
+        if (st->bitrev==NULL)
+            goto fail;
+        compute_bitrev_table(0, bitrev, 1,1, st->factors,st);
+
+        /* Initialize architecture specific fft parameters */
+        if (opus_fft_alloc_arch(st, arch))
+            goto fail;
+    }
+    return st;
+fail:
+    opus_fft_free(st, arch);
+    return NULL;
+}
+
+kiss_fft_state *opus_fft_alloc(int nfft,void * mem,size_t * lenmem, int arch)
+{
+   return opus_fft_alloc_twiddles(nfft, mem, lenmem, NULL, arch);
+}
+
+void opus_fft_free_arch_c(kiss_fft_state *st) {
+   (void)st;
+}
+
+void opus_fft_free(const kiss_fft_state *cfg, int arch)
+{
+   if (cfg)
+   {
+      opus_fft_free_arch((kiss_fft_state *)cfg, arch);
+      opus_free((opus_int16*)cfg->bitrev);
+      if (cfg->shift < 0)
+         opus_free((kiss_twiddle_cpx*)cfg->twiddles);
+      opus_free((kiss_fft_state*)cfg);
+   }
+}
+
+#endif /* CUSTOM_MODES */
+
+#ifdef FIXED_POINT
+#ifndef OVERRIDE_fft_downshift
+static void fft_downshift(kiss_fft_cpx *x, int N, int *total, int step) {
+   int shift;
+   shift = IMIN(step, *total);
+   *total -= shift;
+   if (shift == 1) {
+      int i;
+      for (i=0;i<N;i++) {
+         x[i].r = SHR32(x[i].r, 1);
+         x[i].i = SHR32(x[i].i, 1);
+      }
+   } else if (shift>0) {
+      int i;
+      for (i=0;i<N;i++) {
+         x[i].r = PSHR32(x[i].r, shift);
+         x[i].i = PSHR32(x[i].i, shift);
+      }
+   }
+}
+#endif /* OVERRIDE_fft_downshift */
+#else
+#define fft_downshift(x, N, total, step)
+#endif
+
+void opus_fft_impl(const kiss_fft_state *st,kiss_fft_cpx *fout ARG_FIXED(int downshift))
+{
+    int m2, m;
+    int p;
+    int L;
+    int fstride[MAXFACTORS];
+    int i;
+    int shift;
+
+    /* st->shift can be -1 */
+    shift = st->shift>0 ? st->shift : 0;
+
+    fstride[0] = 1;
+    L=0;
+    do {
+       p = st->factors[2*L];
+       m = st->factors[2*L+1];
+       fstride[L+1] = fstride[L]*p;
+       L++;
+    } while(m!=1);
+    m = st->factors[2*L-1];
+    for (i=L-1;i>=0;i--)
+    {
+       if (i!=0)
+          m2 = st->factors[2*i-1];
+       else
+          m2 = 1;
+       switch (st->factors[2*i])
+       {
+       case 2:
+          fft_downshift(fout, st->nfft, &downshift, 1);
+          kf_bfly2(fout, m, fstride[i]);
+          break;
+       case 4:
+          fft_downshift(fout, st->nfft, &downshift, 2);
+          kf_bfly4(fout,fstride[i]<<shift,st,m, fstride[i], m2);
+          break;
+ #ifndef RADIX_TWO_ONLY
+       case 3:
+          fft_downshift(fout, st->nfft, &downshift, 2);
+          kf_bfly3(fout,fstride[i]<<shift,st,m, fstride[i], m2);
+          break;
+       case 5:
+          fft_downshift(fout, st->nfft, &downshift, 3);
+          kf_bfly5(fout,fstride[i]<<shift,st,m, fstride[i], m2);
+          break;
+ #endif
+       }
+       m = m2;
+    }
+    fft_downshift(fout, st->nfft, &downshift, downshift);
+}
+
+void opus_fft_c(const kiss_fft_state *st,const kiss_fft_cpx *fin,kiss_fft_cpx *fout)
+{
+   int i;
+   celt_coef scale;
+#ifdef FIXED_POINT
+   /* Allows us to scale with MULT16_32_Q16(), which is faster than
+      MULT16_32_Q15() on ARM. */
+   int scale_shift = st->scale_shift-1;
+#endif
+   scale = st->scale;
+
+   celt_assert2 (fin != fout, "In-place FFT not supported");
+   /* Bit-reverse the input */
+   for (i=0;i<st->nfft;i++)
+   {
+      kiss_fft_cpx x = fin[i];
+      fout[st->bitrev[i]].r = S_MUL2(x.r, scale);
+      fout[st->bitrev[i]].i = S_MUL2(x.i, scale);
+   }
+   opus_fft_impl(st, fout ARG_FIXED(scale_shift));
+}
+
+
+void opus_ifft_c(const kiss_fft_state *st,const kiss_fft_cpx *fin,kiss_fft_cpx *fout)
+{
+   int i;
+   celt_assert2 (fin != fout, "In-place FFT not supported");
+   /* Bit-reverse the input */
+   for (i=0;i<st->nfft;i++)
+      fout[st->bitrev[i]] = fin[i];
+   for (i=0;i<st->nfft;i++)
+      fout[i].i = -fout[i].i;
+   opus_fft_impl(st, fout ARG_FIXED(0));
+   for (i=0;i<st->nfft;i++)
+      fout[i].i = -fout[i].i;
+}
--- a/lib/libopus/celt/kiss_fft.h
+++ b/lib/libopus/celt/kiss_fft.h
@ -0,0 +1,209 @@
+/*Copyright (c) 2003-2004, Mark Borgerding
+  Lots of modifications by Jean-Marc Valin
+  Copyright (c) 2005-2007, Xiph.Org Foundation
+  Copyright (c) 2008,      Xiph.Org Foundation, CSIRO
+
+  All rights reserved.
+
+  Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are met:
+
+    * Redistributions of source code must retain the above copyright notice,
+       this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above copyright notice,
+       this list of conditions and the following disclaimer in the
+       documentation and/or other materials provided with the distribution.
+
+  THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+  AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+  IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+  ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+  LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+  CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+  SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+  INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+  CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+  ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+  POSSIBILITY OF SUCH DAMAGE.*/
+
+#ifndef KISS_FFT_H
+#define KISS_FFT_H
+
+#include <stdlib.h>
+#include <math.h>
+#include "arch.h"
+#include "cpu_support.h"
+
+#ifdef USE_SIMD
+# include <xmmintrin.h>
+# define kiss_fft_scalar __m128
+#define KISS_FFT_MALLOC(nbytes) memalign(16,nbytes)
+#else
+#define KISS_FFT_MALLOC opus_alloc
+#endif
+
+#ifdef FIXED_POINT
+#include "arch.h"
+
+#  define kiss_fft_scalar opus_int32
+#  ifdef ENABLE_QEXT
+#   define COEF_SHIFT 32
+#  else
+#   define COEF_SHIFT 16
+#  endif
+
+#  define kiss_twiddle_scalar celt_coef
+
+/* Some 32-bit CPUs would load/store a kiss_twiddle_cpx with a single memory
+ * access, and could benefit from additional alignment.
+ */
+#  define KISS_TWIDDLE_CPX_ALIGNMENT (sizeof(opus_int32))
+
+#else
+
+# ifndef kiss_fft_scalar
+/*  default is float */
+#   define kiss_fft_scalar float
+#   define kiss_twiddle_scalar float
+#   define KF_SUFFIX _celt_single
+# endif
+#endif
+
+#if defined(__GNUC__) && defined(KISS_TWIDDLE_CPX_ALIGNMENT)
+#define KISS_TWIDDLE_CPX_ALIGNED __attribute__((aligned(KISS_TWIDDLE_CPX_ALIGNMENT)))
+#else
+#define KISS_TWIDDLE_CPX_ALIGNED
+#endif
+
+typedef struct {
+    kiss_fft_scalar r;
+    kiss_fft_scalar i;
+}kiss_fft_cpx;
+
+typedef struct {
+   kiss_twiddle_scalar r;
+   kiss_twiddle_scalar i;
+} KISS_TWIDDLE_CPX_ALIGNED kiss_twiddle_cpx;
+
+#define MAXFACTORS 8
+/* e.g. an fft of length 128 has 4 factors
+ as far as kissfft is concerned
+ 4*4*4*2
+ */
+
+typedef struct arch_fft_state{
+   int is_supported;
+   void *priv;
+} arch_fft_state;
+
+typedef struct kiss_fft_state{
+    int nfft;
+    celt_coef scale;
+#ifdef FIXED_POINT
+    int scale_shift;
+#endif
+    int shift;
+    opus_int16 factors[2*MAXFACTORS];
+    const opus_int16 *bitrev;
+    const kiss_twiddle_cpx *twiddles;
+    arch_fft_state *arch_fft;
+} kiss_fft_state;
+
+#if defined(HAVE_ARM_NE10)
+#include "arm/fft_arm.h"
+#endif
+
+/*typedef struct kiss_fft_state* kiss_fft_cfg;*/
+
+/**
+ *  opus_fft_alloc
+ *
+ *  Initialize a FFT (or IFFT) algorithm's cfg/state buffer.
+ *
+ *  typical usage:      kiss_fft_cfg mycfg=opus_fft_alloc(1024,0,NULL,NULL);
+ *
+ *  The return value from fft_alloc is a cfg buffer used internally
+ *  by the fft routine or NULL.
+ *
+ *  If lenmem is NULL, then opus_fft_alloc will allocate a cfg buffer using malloc.
+ *  The returned value should be free()d when done to avoid memory leaks.
+ *
+ *  The state can be placed in a user supplied buffer 'mem':
+ *  If lenmem is not NULL and mem is not NULL and *lenmem is large enough,
+ *      then the function places the cfg in mem and the size used in *lenmem
+ *      and returns mem.
+ *
+ *  If lenmem is not NULL and ( mem is NULL or *lenmem is not large enough),
+ *      then the function returns NULL and places the minimum cfg
+ *      buffer size in *lenmem.
+ * */
+
+kiss_fft_state *opus_fft_alloc_twiddles(int nfft,void * mem,size_t * lenmem, const kiss_fft_state *base, int arch);
+
+kiss_fft_state *opus_fft_alloc(int nfft,void * mem,size_t * lenmem, int arch);
+
+/**
+ * opus_fft(cfg,in_out_buf)
+ *
+ * Perform an FFT on a complex input buffer.
+ * for a forward FFT,
+ * fin should be  f[0] , f[1] , ... ,f[nfft-1]
+ * fout will be   F[0] , F[1] , ... ,F[nfft-1]
+ * Note that each element is complex and can be accessed like
+    f[k].r and f[k].i
+ * */
+void opus_fft_c(const kiss_fft_state *cfg,const kiss_fft_cpx *fin,kiss_fft_cpx *fout);
+void opus_ifft_c(const kiss_fft_state *cfg,const kiss_fft_cpx *fin,kiss_fft_cpx *fout);
+
+void opus_fft_impl(const kiss_fft_state *st,kiss_fft_cpx *fout ARG_FIXED(int downshift));
+void opus_ifft_impl(const kiss_fft_state *st,kiss_fft_cpx *fout);
+
+void opus_fft_free(const kiss_fft_state *cfg, int arch);
+
+
+void opus_fft_free_arch_c(kiss_fft_state *st);
+int opus_fft_alloc_arch_c(kiss_fft_state *st);
+
+#if !defined(OVERRIDE_OPUS_FFT)
+/* Is run-time CPU detection enabled on this platform? */
+#if defined(OPUS_HAVE_RTCD) && (defined(HAVE_ARM_NE10))
+
+extern int (*const OPUS_FFT_ALLOC_ARCH_IMPL[OPUS_ARCHMASK+1])(
+ kiss_fft_state *st);
+
+#define opus_fft_alloc_arch(_st, arch) \
+         ((*OPUS_FFT_ALLOC_ARCH_IMPL[(arch)&OPUS_ARCHMASK])(_st))
+
+extern void (*const OPUS_FFT_FREE_ARCH_IMPL[OPUS_ARCHMASK+1])(
+ kiss_fft_state *st);
+#define opus_fft_free_arch(_st, arch) \
+         ((*OPUS_FFT_FREE_ARCH_IMPL[(arch)&OPUS_ARCHMASK])(_st))
+
+extern void (*const OPUS_FFT[OPUS_ARCHMASK+1])(const kiss_fft_state *cfg,
+ const kiss_fft_cpx *fin, kiss_fft_cpx *fout);
+#define opus_fft(_cfg, _fin, _fout, arch) \
+   ((*OPUS_FFT[(arch)&OPUS_ARCHMASK])(_cfg, _fin, _fout))
+
+extern void (*const OPUS_IFFT[OPUS_ARCHMASK+1])(const kiss_fft_state *cfg,
+ const kiss_fft_cpx *fin, kiss_fft_cpx *fout);
+#define opus_ifft(_cfg, _fin, _fout, arch) \
+   ((*OPUS_IFFT[(arch)&OPUS_ARCHMASK])(_cfg, _fin, _fout))
+
+#else /* else for if defined(OPUS_HAVE_RTCD) && (defined(HAVE_ARM_NE10)) */
+
+#define opus_fft_alloc_arch(_st, arch) \
+         ((void)(arch), opus_fft_alloc_arch_c(_st))
+
+#define opus_fft_free_arch(_st, arch) \
+         ((void)(arch), opus_fft_free_arch_c(_st))
+
+#define opus_fft(_cfg, _fin, _fout, arch) \
+         ((void)(arch), opus_fft_c(_cfg, _fin, _fout))
+
+#define opus_ifft(_cfg, _fin, _fout, arch) \
+         ((void)(arch), opus_ifft_c(_cfg, _fin, _fout))
+
+#endif /* end if defined(OPUS_HAVE_RTCD) && (defined(HAVE_ARM_NE10)) */
+#endif /* end if !defined(OVERRIDE_OPUS_FFT) */
+
+#endif
--- a/lib/libopus/celt/laplace.c
+++ b/lib/libopus/celt/laplace.c
@ -0,0 +1,235 @@
+/* Copyright (c) 2007 CSIRO
+   Copyright (c) 2007-2009 Xiph.Org Foundation
+   Written by Jean-Marc Valin */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "laplace.h"
+#include "mathops.h"
+
+/* The minimum probability of an energy delta (out of 32768). */
+#define LAPLACE_LOG_MINP (0)
+#define LAPLACE_MINP (1<<LAPLACE_LOG_MINP)
+/* The minimum number of guaranteed representable energy deltas (in one
+    direction). */
+#define LAPLACE_NMIN (16)
+
+/* When called, decay is positive and at most 11456. */
+static unsigned ec_laplace_get_freq1(unsigned fs0, int decay)
+{
+   unsigned ft;
+   ft = 32768 - LAPLACE_MINP*(2*LAPLACE_NMIN) - fs0;
+   return ft*(opus_int32)(16384-decay)>>15;
+}
+
+void ec_laplace_encode(ec_enc *enc, int *value, unsigned fs, int decay)
+{
+   unsigned fl;
+   int val = *value;
+   fl = 0;
+   if (val)
+   {
+      int s;
+      int i;
+      s = -(val<0);
+      val = (val+s)^s;
+      fl = fs;
+      fs = ec_laplace_get_freq1(fs, decay);
+      /* Search the decaying part of the PDF.*/
+      for (i=1; fs > 0 && i < val; i++)
+      {
+         fs *= 2;
+         fl += fs+2*LAPLACE_MINP;
+         fs = (fs*(opus_int32)decay)>>15;
+      }
+      /* Everything beyond that has probability LAPLACE_MINP. */
+      if (!fs)
+      {
+         int di;
+         int ndi_max;
+         ndi_max = (32768-fl+LAPLACE_MINP-1)>>LAPLACE_LOG_MINP;
+         ndi_max = (ndi_max-s)>>1;
+         di = IMIN(val - i, ndi_max - 1);
+         fl += (2*di+1+s)*LAPLACE_MINP;
+         fs = IMIN(LAPLACE_MINP, 32768-fl);
+         *value = (i+di+s)^s;
+      }
+      else
+      {
+         fs += LAPLACE_MINP;
+         fl += fs&~s;
+      }
+      celt_assert(fl+fs<=32768);
+      celt_assert(fs>0);
+   }
+   ec_encode_bin(enc, fl, fl+fs, 15);
+}
+
+int ec_laplace_decode(ec_dec *dec, unsigned fs, int decay)
+{
+   int val=0;
+   unsigned fl;
+   unsigned fm;
+   fm = ec_decode_bin(dec, 15);
+   fl = 0;
+   if (fm >= fs)
+   {
+      val++;
+      fl = fs;
+      fs = ec_laplace_get_freq1(fs, decay)+LAPLACE_MINP;
+      /* Search the decaying part of the PDF.*/
+      while(fs > LAPLACE_MINP && fm >= fl+2*fs)
+      {
+         fs *= 2;
+         fl += fs;
+         fs = ((fs-2*LAPLACE_MINP)*(opus_int32)decay)>>15;
+         fs += LAPLACE_MINP;
+         val++;
+      }
+      /* Everything beyond that has probability LAPLACE_MINP. */
+      if (fs <= LAPLACE_MINP)
+      {
+         int di;
+         di = (fm-fl)>>(LAPLACE_LOG_MINP+1);
+         val += di;
+         fl += 2*di*LAPLACE_MINP;
+      }
+      if (fm < fl+fs)
+         val = -val;
+      else
+         fl += fs;
+   }
+   celt_assert(fl<32768);
+   celt_assert(fs>0);
+   celt_assert(fl<=fm);
+   celt_assert(fm<IMIN(fl+fs,32768));
+   ec_dec_update(dec, fl, IMIN(fl+fs,32768), 32768);
+   return val;
+}
+
+void ec_laplace_encode_p0(ec_enc *enc, int value, opus_uint16 p0, opus_uint16 decay)
+{
+   int s;
+   opus_uint16 sign_icdf[3];
+   sign_icdf[0] = 32768-p0;
+   sign_icdf[1] = sign_icdf[0]/2;
+   sign_icdf[2] = 0;
+   s = value == 0 ? 0 : (value > 0 ? 1 : 2);
+   ec_enc_icdf16(enc, s, sign_icdf, 15);
+   value = abs(value);
+   if (value)
+   {
+      int i;
+      opus_uint16 icdf[8];
+      icdf[0] = IMAX(7, decay);
+      for (i=1;i<7;i++)
+      {
+         icdf[i] = IMAX(7-i, (icdf[i-1] * (opus_int32)decay) >> 15);
+      }
+      icdf[7] = 0;
+      value--;
+      do {
+         ec_enc_icdf16(enc, IMIN(value, 7), icdf, 15);
+         value -= 7;
+      } while (value >= 0);
+   }
+}
+
+int ec_laplace_decode_p0(ec_dec *dec, opus_uint16 p0, opus_uint16 decay)
+{
+   int s;
+   int value;
+   opus_uint16 sign_icdf[3];
+   sign_icdf[0] = 32768-p0;
+   sign_icdf[1] = sign_icdf[0]/2;
+   sign_icdf[2] = 0;
+   s = ec_dec_icdf16(dec, sign_icdf, 15);
+   if (s==2) s = -1;
+   if (s != 0)
+   {
+      int i;
+      int v;
+      opus_uint16 icdf[8];
+      icdf[0] = IMAX(7, decay);
+      for (i=1;i<7;i++)
+      {
+         icdf[i] = IMAX(7-i, (icdf[i-1] * (opus_int32)decay) >> 15);
+      }
+      icdf[7] = 0;
+      value = 1;
+      do {
+         v = ec_dec_icdf16(dec, icdf, 15);
+         value += v;
+      } while (v == 7);
+      return s*value;
+   } else return 0;
+}
+
+#if 0
+
+#include <stdio.h>
+#define NB_VALS 10
+#define DATA_SIZE 10000
+int main() {
+   ec_enc enc;
+   ec_dec dec;
+   unsigned char *ptr;
+   int i;
+   int decay, p0;
+   int val[NB_VALS] = {6, 7, 8, 9, 10, 11, 12, 13, 14, 15};
+   /*for (i=0;i<NB_VALS;i++) {
+      val[i] = -log(rand()/(float)RAND_MAX);
+      if (rand()%2) val[i] = -val[i];
+   }*/
+   p0 = 16000;
+   decay = 16000;
+   ptr = (unsigned char *)malloc(DATA_SIZE);
+   ec_enc_init(&enc,ptr,DATA_SIZE);
+   for (i=0;i<NB_VALS;i++) {
+      printf("%d ", val[i]);
+   }
+   printf("\n");
+   for (i=0;i<NB_VALS;i++) {
+      ec_laplace_encode_p0(&enc, val[i], p0, decay);
+   }
+
+   ec_enc_done(&enc);
+
+   ec_dec_init(&dec,ec_get_buffer(&enc),ec_range_bytes(&enc));
+
+   for (i=0;i<NB_VALS;i++) {
+      val[i] = ec_laplace_decode_p0(&dec, p0, decay);
+   }
+   for (i=0;i<NB_VALS;i++) {
+      printf("%d ", val[i]);
+   }
+   printf("\n");
+}
+
+#endif
--- a/lib/libopus/celt/laplace.h
+++ b/lib/libopus/celt/laplace.h
@ -0,0 +1,57 @@
+/* Copyright (c) 2007 CSIRO
+   Copyright (c) 2007-2009 Xiph.Org Foundation
+   Written by Jean-Marc Valin */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef LAPLACE_H
+#define LAPLACE_H
+
+#include "entenc.h"
+#include "entdec.h"
+
+/** Encode a value that is assumed to be the realisation of a
+    Laplace-distributed random process
+ @param enc Entropy encoder state
+ @param value Value to encode
+ @param fs Probability of 0, multiplied by 32768
+ @param decay Probability of the value +/- 1, multiplied by 16384
+*/
+void ec_laplace_encode(ec_enc *enc, int *value, unsigned fs, int decay);
+
+/** Decode a value that is assumed to be the realisation of a
+    Laplace-distributed random process
+ @param dec Entropy decoder state
+ @param fs Probability of 0, multiplied by 32768
+ @param decay Probability of the value +/- 1, multiplied by 16384
+ @return Value decoded
+ */
+int ec_laplace_decode(ec_dec *dec, unsigned fs, int decay);
+
+
+int ec_laplace_decode_p0(ec_dec *dec, opus_uint16 p0, opus_uint16 decay);
+void ec_laplace_encode_p0(ec_enc *enc, int value, opus_uint16 p0, opus_uint16 decay);
+
+#endif
--- a/lib/libopus/celt/mathops.c
+++ b/lib/libopus/celt/mathops.c
@ -0,0 +1,336 @@
+/* Copyright (c) 2002-2008 Jean-Marc Valin
+   Copyright (c) 2007-2008 CSIRO
+   Copyright (c) 2007-2009 Xiph.Org Foundation
+   Copyright (c) 2024 Arm Limited
+   Written by Jean-Marc Valin */
+/**
+   @file mathops.h
+   @brief Various math functions
+*/
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "float_cast.h"
+#include "mathops.h"
+
+/*Compute floor(sqrt(_val)) with exact arithmetic.
+  _val must be greater than 0.
+  This has been tested on all possible 32-bit inputs greater than 0.*/
+unsigned isqrt32(opus_uint32 _val){
+  unsigned b;
+  unsigned g;
+  int      bshift;
+  /*Uses the second method from
+     http://www.azillionmonkeys.com/qed/sqroot.html
+    The main idea is to search for the largest binary digit b such that
+     (g+b)*(g+b) <= _val, and add it to the solution g.*/
+  g=0;
+  bshift=(EC_ILOG(_val)-1)>>1;
+  b=1U<<bshift;
+  do{
+    opus_uint32 t;
+    t=(((opus_uint32)g<<1)+b)<<bshift;
+    if(t<=_val){
+      g+=b;
+      _val-=t;
+    }
+    b>>=1;
+    bshift--;
+  }
+  while(bshift>=0);
+  return g;
+}
+
+#ifdef FIXED_POINT
+
+opus_val32 frac_div32_q29(opus_val32 a, opus_val32 b)
+{
+   opus_val16 rcp;
+   opus_val32 result, rem;
+   int shift = celt_ilog2(b)-29;
+   a = VSHR32(a,shift);
+   b = VSHR32(b,shift);
+   /* 16-bit reciprocal */
+   rcp = ROUND16(celt_rcp(ROUND16(b,16)),3);
+   result = MULT16_32_Q15(rcp, a);
+   rem = PSHR32(a,2)-MULT32_32_Q31(result, b);
+   result = ADD32(result, SHL32(MULT16_32_Q15(rcp, rem),2));
+   return result;
+}
+
+opus_val32 frac_div32(opus_val32 a, opus_val32 b) {
+   opus_val32 result = frac_div32_q29(a,b);
+   if (result >= 536870912)       /*  2^29 */
+      return 2147483647;          /*  2^31 - 1 */
+   else if (result <= -536870912) /* -2^29 */
+      return -2147483647;         /* -2^31 */
+   else
+      return SHL32(result, 2);
+}
+
+/** Reciprocal sqrt approximation in the range [0.25,1) (Q16 in, Q14 out) */
+opus_val16 celt_rsqrt_norm(opus_val32 x)
+{
+   opus_val16 n;
+   opus_val16 r;
+   opus_val16 r2;
+   opus_val16 y;
+   /* Range of n is [-16384,32767] ([-0.5,1) in Q15). */
+   n = x-32768;
+   /* Get a rough initial guess for the root.
+      The optimal minimax quadratic approximation (using relative error) is
+       r = 1.437799046117536+n*(-0.823394375837328+n*0.4096419668459485).
+      Coefficients here, and the final result r, are Q14.*/
+   r = ADD16(23557, MULT16_16_Q15(n, ADD16(-13490, MULT16_16_Q15(n, 6713))));
+   /* We want y = x*r*r-1 in Q15, but x is 32-bit Q16 and r is Q14.
+      We can compute the result from n and r using Q15 multiplies with some
+       adjustment, carefully done to avoid overflow.
+      Range of y is [-1564,1594]. */
+   r2 = MULT16_16_Q15(r, r);
+   y = SHL16(SUB16(ADD16(MULT16_16_Q15(r2, n), r2), 16384), 1);
+   /* Apply a 2nd-order Householder iteration: r += r*y*(y*0.375-0.5).
+      This yields the Q14 reciprocal square root of the Q16 x, with a maximum
+       relative error of 1.04956E-4, a (relative) RMSE of 2.80979E-5, and a
+       peak absolute error of 2.26591/16384. */
+   return ADD16(r, MULT16_16_Q15(r, MULT16_16_Q15(y,
+              SUB16(MULT16_16_Q15(y, 12288), 16384))));
+}
+
+/** Reciprocal sqrt approximation in the range [0.25,1) (Q31 in, Q29 out) */
+opus_val32 celt_rsqrt_norm32(opus_val32 x)
+{
+   opus_int32 tmp;
+   /* Use the first-order Newton-Raphson method to refine the root estimate.
+    * r = r * (1.5 - 0.5*x*r*r) */
+   opus_int32 r_q29 = SHL32(celt_rsqrt_norm(SHR32(x, 31-16)), 15);
+   /* Split evaluation in steps to avoid exploding macro expansion. */
+   tmp = MULT32_32_Q31(r_q29, r_q29);
+   tmp = MULT32_32_Q31(1073741824 /* Q31 */, tmp);
+   tmp = MULT32_32_Q31(x, tmp);
+   return SHL32(MULT32_32_Q31(r_q29, SUB32(201326592 /* Q27 */, tmp)), 4);
+}
+
+/** Sqrt approximation (QX input, QX/2 output) */
+opus_val32 celt_sqrt(opus_val32 x)
+{
+   int k;
+   opus_val16 n;
+   opus_val32 rt;
+   /* These coeffs are optimized in fixed-point to minimize both RMS and max error
+      of sqrt(x) over .25<x<1 without exceeding 32767.
+      The RMS error is 3.4e-5 and the max is 8.2e-5. */
+   static const opus_val16 C[6] = {23171, 11574, -2901, 1592, -1002, 336};
+   if (x==0)
+      return 0;
+   else if (x>=1073741824)
+      return 32767;
+   k = (celt_ilog2(x)>>1)-7;
+   x = VSHR32(x, 2*k);
+   n = x-32768;
+   rt = ADD32(C[0], MULT16_16_Q15(n, ADD16(C[1], MULT16_16_Q15(n, ADD16(C[2],
+              MULT16_16_Q15(n, ADD16(C[3], MULT16_16_Q15(n, ADD16(C[4], MULT16_16_Q15(n, (C[5])))))))))));
+   rt = VSHR32(rt,7-k);
+   return rt;
+}
+
+/* Perform fixed-point arithmetic to approximate the square root. When the input
+ * is in Qx format, the output will be in Q(x/2 + 16) format. */
+opus_val32 celt_sqrt32(opus_val32 x)
+{
+   int k;
+   opus_int32 x_frac;
+   if (x==0)
+      return 0;
+   else if (x>=1073741824)
+      return 2147483647; /* 2^31 -1 */
+   k = (celt_ilog2(x)>>1);
+   x_frac = VSHR32(x, 2*(k-14)-1);
+   x_frac = MULT32_32_Q31(celt_rsqrt_norm32(x_frac), x_frac);
+   if (k < 12) return PSHR32(x_frac, 12-k);
+   else return SHL32(x_frac, k-12);
+}
+
+#define L1 32767
+#define L2 -7651
+#define L3 8277
+#define L4 -626
+
+static OPUS_INLINE opus_val16 _celt_cos_pi_2(opus_val16 x)
+{
+   opus_val16 x2;
+
+   x2 = MULT16_16_P15(x,x);
+   return ADD16(1,MIN16(32766,ADD32(SUB16(L1,x2), MULT16_16_P15(x2, ADD32(L2, MULT16_16_P15(x2, ADD32(L3, MULT16_16_P15(L4, x2
+                                                                                ))))))));
+}
+
+#undef L1
+#undef L2
+#undef L3
+#undef L4
+
+opus_val16 celt_cos_norm(opus_val32 x)
+{
+   x = x&0x0001ffff;
+   if (x>SHL32(EXTEND32(1), 16))
+      x = SUB32(SHL32(EXTEND32(1), 17),x);
+   if (x&0x00007fff)
+   {
+      if (x<SHL32(EXTEND32(1), 15))
+      {
+         return _celt_cos_pi_2(EXTRACT16(x));
+      } else {
+         return NEG16(_celt_cos_pi_2(EXTRACT16(65536-x)));
+      }
+   } else {
+      if (x&0x0000ffff)
+         return 0;
+      else if (x&0x0001ffff)
+         return -32767;
+      else
+         return 32767;
+   }
+}
+
+/* Calculates the cosine of (PI*0.5*x) where the input x ranges from -1 to 1 and
+ * is in Q30 format. The output will also be in Q31 format. */
+opus_val32 celt_cos_norm32(opus_val32 x)
+{
+   static const opus_val32 COS_NORM_COEFF_A0 = 134217720;   /* Q27 */
+   static const opus_val32 COS_NORM_COEFF_A1 = -662336704;  /* Q29 */
+   static const opus_val32 COS_NORM_COEFF_A2 = 544710848;   /* Q31 */
+   static const opus_val32 COS_NORM_COEFF_A3 = -178761936;  /* Q33 */
+   static const opus_val32 COS_NORM_COEFF_A4 = 29487206;    /* Q35 */
+   opus_int32 x_sq_q29, tmp;
+   /* The expected x is in the range of [-1.0f, 1.0f] */
+   celt_sig_assert((x >= -1073741824) && (x <= 1073741824));
+   /* Make cos(+/- pi/2) exactly zero. */
+   if (ABS32(x) == 1<<30) return 0;
+   x_sq_q29 = MULT32_32_Q31(x, x);
+   /* Split evaluation in steps to avoid exploding macro expansion. */
+   tmp = ADD32(COS_NORM_COEFF_A3, MULT32_32_Q31(x_sq_q29, COS_NORM_COEFF_A4));
+   tmp = ADD32(COS_NORM_COEFF_A2, MULT32_32_Q31(x_sq_q29, tmp));
+   tmp = ADD32(COS_NORM_COEFF_A1, MULT32_32_Q31(x_sq_q29, tmp));
+   return SHL32(ADD32(COS_NORM_COEFF_A0, MULT32_32_Q31(x_sq_q29, tmp)), 4);
+}
+
+/* Computes a 16 bit approximate reciprocal (1/x) for a normalized Q15 input,
+ * resulting in a Q15 output. */
+opus_val16 celt_rcp_norm16(opus_val16 x)
+{
+   opus_val16 r;
+   /* Start with a linear approximation:
+      r = 1.8823529411764706-0.9411764705882353*n.
+      The coefficients and the result are Q14 in the range [15420,30840].*/
+   r = ADD16(30840, MULT16_16_Q15(-15420, x));
+   /* Perform two Newton iterations:
+      r -= r*((r*n)+(r-1.Q15))
+         = r*((r*n)+(r-1.Q15)). */
+   r = SUB16(r, MULT16_16_Q15(r,
+             ADD16(MULT16_16_Q15(r, x), ADD16(r, -32768))));
+   /* We subtract an extra 1 in the second iteration to avoid overflow; it also
+       neatly compensates for truncation error in the rest of the process. */
+   return SUB16(r, ADD16(1, MULT16_16_Q15(r,
+                ADD16(MULT16_16_Q15(r, x), ADD16(r, -32768)))));
+}
+
+/* Computes a 32 bit approximated reciprocal (1/x) for a normalized Q31 input,
+ * resulting in a Q30 output. The expected input range is [0.5f, 1.0f) in Q31
+ * and the expected output range is [1.0f, 2.0f) in Q30. */
+opus_val32 celt_rcp_norm32(opus_val32 x)
+{
+   opus_val32 r_q30;
+   celt_sig_assert(x >= 1073741824);
+   r_q30 = SHL32(EXTEND32(celt_rcp_norm16(SHR32(x, 15)-32768)), 16);
+   /* Solving f(y) = a - 1/y using the Newton Method
+    * Note: f(y)' = 1/y^2
+    * r = r - f(r)/f(r)' = r - (x * r*r - r)
+    *   = r - r*(r*x - 1)
+    * where
+    *   - r means 1/y's approximation.
+    *   - x means a, the input of function.
+    * Please note that:
+    *   - It adds 1 to avoid overflow
+    *   - -1.0f in Q30 is -1073741824. */
+   return SUB32(r_q30, ADD32(SHL32(
+                MULT32_32_Q31(ADD32(MULT32_32_Q31(r_q30, x), -1073741824),
+                              r_q30), 1), 1));
+}
+
+/** Reciprocal approximation (Q15 input, Q16 output) */
+opus_val32 celt_rcp(opus_val32 x)
+{
+   int i;
+   opus_val16 r;
+   celt_sig_assert(x>0);
+   i = celt_ilog2(x);
+
+   /* Compute the reciprocal of a Q15 number in the range [0, 1). */
+   r = celt_rcp_norm16(VSHR32(x,i-15)-32768);
+
+   /* r is now the Q15 solution to 2/(n+1), with a maximum relative error
+       of 7.05346E-5, a (relative) RMSE of 2.14418E-5, and a peak absolute
+       error of 1.24665/32768. */
+   return VSHR32(EXTEND32(r),i-16);
+}
+
+#endif
+
+#ifndef DISABLE_FLOAT_API
+
+void celt_float2int16_c(const float * OPUS_RESTRICT in, short * OPUS_RESTRICT out, int cnt)
+{
+   int i;
+   for (i = 0; i < cnt; i++)
+   {
+      out[i] = FLOAT2INT16(in[i]);
+   }
+}
+
+int opus_limit2_checkwithin1_c(float * samples, int cnt)
+{
+   int i;
+   if (cnt <= 0)
+   {
+      return 1;
+   }
+
+   for (i = 0; i < cnt; i++)
+   {
+      float clippedVal = samples[i];
+      clippedVal = FMAX(-2.0f, clippedVal);
+      clippedVal = FMIN(2.0f, clippedVal);
+      samples[i] = clippedVal;
+   }
+
+   /* C implementation can't provide quick hint. Assume it might exceed -1/+1. */
+   return 0;
+}
+
+#endif /* DISABLE_FLOAT_API */
--- a/lib/libopus/celt/mathops.h
+++ b/lib/libopus/celt/mathops.h
@ -0,0 +1,662 @@
+/* Copyright (c) 2002-2008 Jean-Marc Valin
+   Copyright (c) 2007-2008 CSIRO
+   Copyright (c) 2007-2009 Xiph.Org Foundation
+   Copyright (c) 2024 Arm Limited
+   Written by Jean-Marc Valin, and Yunho Huh */
+/**
+   @file mathops.h
+   @brief Various math functions
+*/
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef MATHOPS_H
+#define MATHOPS_H
+
+#include "arch.h"
+#include "entcode.h"
+#include "os_support.h"
+
+
+#if defined(OPUS_ARM_MAY_HAVE_NEON_INTR)
+#include "arm/mathops_arm.h"
+#endif
+
+#define PI 3.1415926535897931
+
+/* Multiplies two 16-bit fractional values. Bit-exactness of this macro is important */
+#define FRAC_MUL16(a,b) ((16384+((opus_int32)(opus_int16)(a)*(opus_int16)(b)))>>15)
+
+unsigned isqrt32(opus_uint32 _val);
+
+/* CELT doesn't need it for fixed-point, by analysis.c does. */
+#if !defined(FIXED_POINT) || defined(ANALYSIS_C)
+#define cA 0.43157974f
+#define cB 0.67848403f
+#define cC 0.08595542f
+#define cE ((float)PI/2)
+static OPUS_INLINE float fast_atan2f(float y, float x) {
+   float x2, y2;
+   x2 = x*x;
+   y2 = y*y;
+   /* For very small values, we don't care about the answer, so
+      we can just return 0. */
+   if (x2 + y2 < 1e-18f)
+   {
+      return 0;
+   }
+   if(x2<y2){
+      float den = (y2 + cB*x2) * (y2 + cC*x2);
+      return -x*y*(y2 + cA*x2) / den + (y<0 ? -cE : cE);
+   }else{
+      float den = (x2 + cB*y2) * (x2 + cC*y2);
+      return  x*y*(x2 + cA*y2) / den + (y<0 ? -cE : cE) - (x*y<0 ? -cE : cE);
+   }
+}
+#undef cA
+#undef cB
+#undef cC
+#undef cE
+#endif
+
+
+#ifndef OVERRIDE_CELT_MAXABS16
+static OPUS_INLINE opus_val32 celt_maxabs16(const opus_val16 *x, int len)
+{
+   int i;
+   opus_val16 maxval = 0;
+   opus_val16 minval = 0;
+   for (i=0;i<len;i++)
+   {
+      maxval = MAX16(maxval, x[i]);
+      minval = MIN16(minval, x[i]);
+   }
+   return MAX32(EXTEND32(maxval),-EXTEND32(minval));
+}
+#endif
+
+#if defined(ENABLE_RES24) && defined(FIXED_POINT)
+static OPUS_INLINE opus_res celt_maxabs_res(const opus_res *x, int len)
+{
+   int i;
+   opus_res maxval = 0;
+   opus_res minval = 0;
+   for (i=0;i<len;i++)
+   {
+      maxval = MAX32(maxval, x[i]);
+      minval = MIN32(minval, x[i]);
+   }
+   /* opus_res should never reach such amplitude, so we should be safe. */
+   celt_sig_assert(minval != -2147483648);
+   return MAX32(maxval,-minval);
+}
+#else
+#define celt_maxabs_res celt_maxabs16
+#endif
+
+
+#ifndef OVERRIDE_CELT_MAXABS32
+#ifdef FIXED_POINT
+static OPUS_INLINE opus_val32 celt_maxabs32(const opus_val32 *x, int len)
+{
+   int i;
+   opus_val32 maxval = 0;
+   opus_val32 minval = 0;
+   for (i=0;i<len;i++)
+   {
+      maxval = MAX32(maxval, x[i]);
+      minval = MIN32(minval, x[i]);
+   }
+   return MAX32(maxval, -minval);
+}
+#else
+#define celt_maxabs32(x,len) celt_maxabs16(x,len)
+#endif
+#endif
+
+#ifndef FIXED_POINT
+/* Calculates the arctangent of x using a Remez approximation of order 15,
+ * incorporating only odd-powered terms. */
+static OPUS_INLINE float celt_atan_norm(float x)
+{
+   #define ATAN2_2_OVER_PI 0.636619772367581f
+   float x_sq = x * x;
+
+   /* Polynomial coefficients approximated in the [0, 1] range.
+    * Lolremez command: lolremez --degree 6 --range "0:1"
+    *                   "(atan(sqrt(x))-sqrt(x))/(x*sqrt(x))" "1/(sqrt(x)*x)"
+    * Please note that ATAN2_COEFF_A01 is fixed to 1.0f. */
+   #define ATAN2_COEFF_A03 -3.3331659436225891113281250000e-01f
+   #define ATAN2_COEFF_A05 1.99627041816711425781250000000e-01f
+   #define ATAN2_COEFF_A07 -1.3976582884788513183593750000e-01f
+   #define ATAN2_COEFF_A09 9.79423448443412780761718750000e-02f
+   #define ATAN2_COEFF_A11 -5.7773590087890625000000000000e-02f
+   #define ATAN2_COEFF_A13 2.30401363223791122436523437500e-02f
+   #define ATAN2_COEFF_A15 -4.3554059229791164398193359375e-03f
+   return ATAN2_2_OVER_PI * (x + x * x_sq * (ATAN2_COEFF_A03
+                + x_sq * (ATAN2_COEFF_A05
+                + x_sq * (ATAN2_COEFF_A07
+                + x_sq * (ATAN2_COEFF_A09
+                + x_sq * (ATAN2_COEFF_A11
+                + x_sq * (ATAN2_COEFF_A13
+                + x_sq * (ATAN2_COEFF_A15))))))));
+}
+
+/* Calculates the arctangent of y/x, returning an approximate value in radians.
+ * Please refer to the linked wiki page (https://en.wikipedia.org/wiki/Atan2)
+ * to learn how atan2 results are computed. */
+static OPUS_INLINE float celt_atan2p_norm(float y, float x)
+{
+   celt_sig_assert(x>=0 && y>=0);
+
+   /* For very small values, we don't care about the answer. */
+   if ((x*x + y*y) < 1e-18f)
+   {
+      return 0;
+   }
+
+   if (y < x)
+   {
+      return celt_atan_norm(y / x);
+   } else {
+      return 1.f - celt_atan_norm(x / y);
+   }
+}
+#endif
+
+#if !defined(FIXED_POINT) || defined(ENABLE_QEXT)
+/* Computes estimated cosine values for (PI/2 * x) using only terms with even
+ * exponents. */
+static OPUS_INLINE float celt_cos_norm2(float x)
+{
+   float x_norm_sq;
+   int output_sign;
+   /* Restrict x to [-1, 3]. */
+   x -= 4*floor(.25*(x+1));
+   /* Negative sign for [1, 3]. */
+   output_sign = 1 - 2*(x>1);
+   /* Restrict to [-1, 1]. */
+   x -= 2*(x>1);
+
+   /* The cosine function, cos(x), has a Taylor series representation consisting
+    * exclusively of even-powered polynomial terms. */
+   x_norm_sq = x * x;
+
+   /* Polynomial coefficients approximated in the [0, 1] range using only terms
+    * with even exponents.
+    * Lolremez command: lolremez --degree 4 --range 0:1 "cos(sqrt(x)*pi*0.5)" */
+   #define COS_COEFF_A0 9.999999403953552246093750000000e-01f
+   #define COS_COEFF_A2 -1.233698248863220214843750000000000f
+   #define COS_COEFF_A4 2.536507546901702880859375000000e-01f
+   #define COS_COEFF_A6 -2.08106283098459243774414062500e-02f
+   #define COS_COEFF_A8 8.581906440667808055877685546875e-04f
+   return output_sign * (COS_COEFF_A0 + x_norm_sq * (COS_COEFF_A2 +
+                               x_norm_sq * (COS_COEFF_A4 +
+                               x_norm_sq * (COS_COEFF_A6 +
+                               x_norm_sq * (COS_COEFF_A8)))));
+}
+
+#endif
+
+#ifndef FIXED_POINT
+
+#define celt_sqrt(x) ((float)sqrt(x))
+#define celt_sqrt32(x) ((float)sqrt(x))
+#define celt_rsqrt(x) (1.f/celt_sqrt(x))
+#define celt_rsqrt_norm(x) (celt_rsqrt(x))
+#define celt_rsqrt_norm32(x) (celt_rsqrt(x))
+#define celt_cos_norm(x) ((float)cos((.5f*PI)*(x)))
+#define celt_rcp(x) (1.f/(x))
+#define celt_div(a,b) ((a)/(b))
+#define frac_div32(a,b) ((float)(a)/(b))
+#define frac_div32_q29(a,b) frac_div32(a,b)
+
+#ifdef FLOAT_APPROX
+/* Calculates the base-2 logarithm (log2(x)) of a number. It is designed for
+ * systems using radix-2 floating-point representation, with the exponent
+ * located at bits 23 to 30 and an offset of 127. Note that special cases like
+ * denormalized numbers, positive/negative infinity, and NaN are not handled.
+ * log2(x) = log2(x^exponent * mantissa)
+ *         = exponent + log2(mantissa) */
+
+/* Log2 x normalization single precision coefficients calculated by
+ * 1 / (1 + 0.125 * index).
+ * Coefficients in Double Precision
+ * double log2_x_norm_coeff[8] = {
+ *    1.0000000000000000000, 8.888888888888888e-01,
+ *    8.000000000000000e-01, 7.272727272727273e-01,
+ *    6.666666666666666e-01, 6.153846153846154e-01,
+ *    5.714285714285714e-01, 5.333333333333333e-01} */
+static const float log2_x_norm_coeff[8] = {
+   1.000000000000000000000000000f, 8.88888895511627197265625e-01f,
+   8.00000000000000000000000e-01f, 7.27272748947143554687500e-01f,
+   6.66666686534881591796875e-01f, 6.15384638309478759765625e-01f,
+   5.71428596973419189453125e-01f, 5.33333361148834228515625e-01f};
+
+/* Log2 y normalization single precision coefficients calculated by
+ * log2(1 + 0.125 * index).
+ * Coefficients in Double Precision
+ * double log2_y_norm_coeff[8] = {
+ *    0.0000000000000000000, 1.699250014423124e-01,
+ *    3.219280948873623e-01, 4.594316186372973e-01,
+ *    5.849625007211562e-01, 7.004397181410922e-01,
+ *    8.073549220576041e-01, 9.068905956085185e-01}; */
+static const float log2_y_norm_coeff[8] = {
+   0.0000000000000000000000000000f, 1.699250042438507080078125e-01f,
+   3.219280838966369628906250e-01f, 4.594316184520721435546875e-01f,
+   5.849624872207641601562500e-01f, 7.004396915435791015625000e-01f,
+   8.073549270629882812500000e-01f, 9.068905711174011230468750e-01f};
+
+static OPUS_INLINE float celt_log2(float x)
+{
+   opus_int32 integer;
+   opus_int32 range_idx;
+   union {
+      float f;
+      opus_uint32 i;
+   } in;
+   in.f = x;
+   integer = (opus_int32)(in.i>>23)-127;
+   in.i = (opus_int32)in.i - (opus_int32)((opus_uint32)integer<<23);
+
+   /* Normalize the mantissa range from [1, 2] to [1,1.125], and then shift x
+    * by 1.0625 to [-0.0625, 0.0625]. */
+   range_idx = (in.i >> 20) & 0x7;
+   in.f = in.f * log2_x_norm_coeff[range_idx] - 1.0625f;
+
+   /* Polynomial coefficients approximated in the [1, 1.125] range.
+    * Lolremez command: lolremez --degree 4 --range -0.0625:0.0625
+    *                   "log(x+1.0625)/log(2)"
+    * Coefficients in Double Precision
+    * A0: 8.7462840624502679e-2    A1: 1.3578296070972002
+    * A2: -6.3897703690210047e-1   A3: 4.0197125617419959e-1
+    * A4: -2.8415445877832832e-1 */
+   #define LOG2_COEFF_A0 8.74628424644470214843750000e-02f
+   #define LOG2_COEFF_A1 1.357829570770263671875000000000f
+   #define LOG2_COEFF_A2 -6.3897705078125000000000000e-01f
+   #define LOG2_COEFF_A3 4.01971250772476196289062500e-01f
+   #define LOG2_COEFF_A4 -2.8415444493293762207031250e-01f
+   in.f = LOG2_COEFF_A0 + in.f * (LOG2_COEFF_A1
+               + in.f * (LOG2_COEFF_A2
+               + in.f * (LOG2_COEFF_A3
+               + in.f * (LOG2_COEFF_A4))));
+   return integer + in.f + log2_y_norm_coeff[range_idx];
+}
+
+/* Calculates an approximation of 2^x. The approximation was achieved by
+ * employing a base-2 exponential function and utilizing a Remez approximation
+ * of order 5, ensuring a controlled relative error.
+ * exp2(x) = exp2(integer + fraction)
+ *         = exp2(integer) * exp2(fraction) */
+static OPUS_INLINE float celt_exp2(float x)
+{
+   opus_int32 integer;
+   float frac;
+   union {
+      float f;
+      opus_uint32 i;
+   } res;
+   integer = (int)floor(x);
+   if (integer < -50)
+      return 0;
+   frac = x-integer;
+
+   /* Polynomial coefficients approximated in the [0, 1] range.
+    * Lolremez command: lolremez --degree 5 --range 0:1
+    *                   "exp(x*0.693147180559945)" "exp(x*0.693147180559945)"
+    * NOTE: log(2) ~ 0.693147180559945 */
+   #define EXP2_COEFF_A0 9.999999403953552246093750000000e-01f
+   #define EXP2_COEFF_A1 6.931530833244323730468750000000e-01f
+   #define EXP2_COEFF_A2 2.401536107063293457031250000000e-01f
+   #define EXP2_COEFF_A3 5.582631751894950866699218750000e-02f
+   #define EXP2_COEFF_A4 8.989339694380760192871093750000e-03f
+   #define EXP2_COEFF_A5 1.877576694823801517486572265625e-03f
+   res.f = EXP2_COEFF_A0 + frac * (EXP2_COEFF_A1
+               + frac * (EXP2_COEFF_A2
+               + frac * (EXP2_COEFF_A3
+               + frac * (EXP2_COEFF_A4
+               + frac * (EXP2_COEFF_A5)))));
+   res.i = (opus_uint32)((opus_int32)res.i + (opus_int32)((opus_uint32)integer<<23)) & 0x7fffffff;
+   return res.f;
+}
+
+#else
+#define celt_log2(x) ((float)(1.442695040888963387*log(x)))
+#define celt_exp2(x) ((float)exp(0.6931471805599453094*(x)))
+#endif
+
+#define celt_exp2_db celt_exp2
+#define celt_log2_db celt_log2
+
+#define celt_sin(x) celt_cos_norm2((0.5f*PI) * (x) - 1.0f)
+#define celt_log(x) (celt_log2(x) * 0.6931471805599453f)
+#define celt_exp(x) (celt_exp2((x) * 1.4426950408889634f))
+
+#endif
+
+#ifdef FIXED_POINT
+
+#include "os_support.h"
+
+#ifndef OVERRIDE_CELT_ILOG2
+/** Integer log in base2. Undefined for zero and negative numbers */
+static OPUS_INLINE opus_int16 celt_ilog2(opus_int32 x)
+{
+   celt_sig_assert(x>0);
+   return EC_ILOG(x)-1;
+}
+#endif
+
+
+/** Integer log in base2. Defined for zero, but not for negative numbers */
+static OPUS_INLINE opus_int16 celt_zlog2(opus_val32 x)
+{
+   return x <= 0 ? 0 : celt_ilog2(x);
+}
+
+opus_val16 celt_rsqrt_norm(opus_val32 x);
+
+opus_val32 celt_rsqrt_norm32(opus_val32 x);
+
+opus_val32 celt_sqrt(opus_val32 x);
+
+opus_val32 celt_sqrt32(opus_val32 x);
+
+opus_val16 celt_cos_norm(opus_val32 x);
+
+opus_val32 celt_cos_norm32(opus_val32 x);
+
+/** Base-2 logarithm approximation (log2(x)). (Q14 input, Q10 output) */
+static OPUS_INLINE opus_val16 celt_log2(opus_val32 x)
+{
+   int i;
+   opus_val16 n, frac;
+   /* -0.41509302963303146, 0.9609890551383969, -0.31836011537636605,
+       0.15530808010959576, -0.08556153059057618 */
+   static const opus_val16 C[5] = {-6801+(1<<(13-10)), 15746, -5217, 2545, -1401};
+   if (x==0)
+      return -32767;
+   i = celt_ilog2(x);
+   n = VSHR32(x,i-15)-32768-16384;
+   frac = ADD16(C[0], MULT16_16_Q15(n, ADD16(C[1], MULT16_16_Q15(n, ADD16(C[2], MULT16_16_Q15(n, ADD16(C[3], MULT16_16_Q15(n, C[4]))))))));
+   return SHL32(i-13,10)+SHR32(frac,14-10);
+}
+
+/*
+ K0 = 1
+ K1 = log(2)
+ K2 = 3-4*log(2)
+ K3 = 3*log(2) - 2
+*/
+#define D0 16383
+#define D1 22804
+#define D2 14819
+#define D3 10204
+
+static OPUS_INLINE opus_val32 celt_exp2_frac(opus_val16 x)
+{
+   opus_val16 frac;
+   frac = SHL16(x, 4);
+   return ADD16(D0, MULT16_16_Q15(frac, ADD16(D1, MULT16_16_Q15(frac, ADD16(D2 , MULT16_16_Q15(D3,frac))))));
+}
+
+#undef D0
+#undef D1
+#undef D2
+#undef D3
+
+/** Base-2 exponential approximation (2^x). (Q10 input, Q16 output) */
+static OPUS_INLINE opus_val32 celt_exp2(opus_val16 x)
+{
+   int integer;
+   opus_val16 frac;
+   integer = SHR16(x,10);
+   if (integer>14)
+      return 0x7f000000;
+   else if (integer < -15)
+      return 0;
+   frac = celt_exp2_frac(x-SHL16(integer,10));
+   return VSHR32(EXTEND32(frac), -integer-2);
+}
+
+#ifdef ENABLE_QEXT
+
+/* Calculates the base-2 logarithm of a Q14 input value. The result is returned
+ * in Q(DB_SHIFT). If the input value is 0, the function will output -32.0f. */
+static OPUS_INLINE opus_val32 celt_log2_db(opus_val32 x) {
+   /* Q30 */
+   static const opus_val32 log2_x_norm_coeff[8] = {
+      1073741824, 954437184, 858993472, 780903168,
+      715827904,  660764224, 613566784, 572662336};
+   /* Q24 */
+   static const opus_val32 log2_y_norm_coeff[8] = {
+      0,       2850868,  5401057,  7707983,
+      9814042, 11751428, 13545168, 15215099};
+   static const opus_val32 LOG2_COEFF_A0 = 1467383;     /* Q24 */
+   static const opus_val32 LOG2_COEFF_A1 = 182244800;   /* Q27 */
+   static const opus_val32 LOG2_COEFF_A2 = -21440512;   /* Q25 */
+   static const opus_val32 LOG2_COEFF_A3 = 107903336;   /* Q28 */
+   static const opus_val32 LOG2_COEFF_A4 = -610217024;  /* Q31 */
+
+   opus_int32 integer, norm_coeff_idx, tmp;
+   opus_val32 mantissa;
+   if (x==0) {
+      return -536870912; /* -32.0f */
+   }
+   integer =  SUB32(celt_ilog2(x), 14);  /* Q0 */
+   mantissa = VSHR32(x, integer + 14 - 29);  /* Q29 */
+   norm_coeff_idx = SHR32(mantissa, 29 - 3) & 0x7;
+   /* mantissa is in Q28 (29 + Q_NORM_CONST - 31 where Q_NORM_CONST is Q30)
+    * 285212672 (Q28) is 1.0625f. */
+   mantissa = SUB32(MULT32_32_Q31(mantissa, log2_x_norm_coeff[norm_coeff_idx]),
+                    285212672);
+
+   /* q_a3(Q28): q_mantissa + q_a4 - 31
+    * q_a2(Q25): q_mantissa + q_a3 - 31
+    * q_a1(Q27): q_mantissa + q_a2 - 31 + 5
+    * q_a0(Q24): q_mantissa + q_a1 - 31
+    * where  q_mantissa is Q28 */
+   /* Split evaluation in steps to avoid exploding macro expansion. */
+   tmp = MULT32_32_Q31(mantissa, LOG2_COEFF_A4);
+   tmp = MULT32_32_Q31(mantissa, ADD32(LOG2_COEFF_A3, tmp));
+   tmp = SHL32(MULT32_32_Q31(mantissa, ADD32(LOG2_COEFF_A2, tmp)), 5 /* SHL32 for LOG2_COEFF_A1 */);
+   tmp = MULT32_32_Q31(mantissa, ADD32(LOG2_COEFF_A1, tmp));
+   return ADD32(log2_y_norm_coeff[norm_coeff_idx],
+          ADD32(SHL32(integer, DB_SHIFT),
+          ADD32(LOG2_COEFF_A0, tmp)));
+}
+
+/* Calculates exp2 for Q28 within a specific range (0 to 1.0) using fixed-point
+ * arithmetic. The input number must be adjusted for Q DB_SHIFT. */
+static OPUS_INLINE opus_val32 celt_exp2_db_frac(opus_val32 x)
+{
+   /* Approximation constants. */
+   static const opus_int32 EXP2_COEFF_A0 = 268435440;   /* Q28 */
+   static const opus_int32 EXP2_COEFF_A1 = 744267456;   /* Q30 */
+   static const opus_int32 EXP2_COEFF_A2 = 1031451904;  /* Q32 */
+   static const opus_int32 EXP2_COEFF_A3 = 959088832;   /* Q34 */
+   static const opus_int32 EXP2_COEFF_A4 = 617742720;   /* Q36 */
+   static const opus_int32 EXP2_COEFF_A5 = 516104352;   /* Q38 */
+   opus_int32 tmp;
+   /* Converts input value from Q24 to Q29. */
+   opus_val32 x_q29 = SHL32(x, 29 - 24);
+   /* Split evaluation in steps to avoid exploding macro expansion. */
+   tmp = ADD32(EXP2_COEFF_A4, MULT32_32_Q31(x_q29, EXP2_COEFF_A5));
+   tmp = ADD32(EXP2_COEFF_A3, MULT32_32_Q31(x_q29, tmp));
+   tmp = ADD32(EXP2_COEFF_A2, MULT32_32_Q31(x_q29, tmp));
+   tmp = ADD32(EXP2_COEFF_A1, MULT32_32_Q31(x_q29, tmp));
+   return ADD32(EXP2_COEFF_A0, MULT32_32_Q31(x_q29, tmp));
+}
+
+/* Calculates exp2 for Q16 using fixed-point arithmetic. The input number must
+ * be adjusted for Q DB_SHIFT. */
+static OPUS_INLINE opus_val32 celt_exp2_db(opus_val32 x)
+{
+   int integer;
+   opus_val32 frac;
+   integer = SHR32(x,DB_SHIFT);
+   if (integer>14)
+      return 0x7f000000;
+   else if (integer <= -17)
+      return 0;
+   frac = celt_exp2_db_frac(x-SHL32(integer, DB_SHIFT));  /* Q28 */
+   return VSHR32(frac, -integer + 28 - 16);  /* Q16 */
+}
+#else
+
+#define celt_log2_db(x) SHL32(EXTEND32(celt_log2(x)), DB_SHIFT-10)
+#define celt_exp2_db_frac(x) SHL32(celt_exp2_frac(PSHR32(x, DB_SHIFT-10)), 14)
+#define celt_exp2_db(x) celt_exp2(PSHR32(x, DB_SHIFT-10))
+
+#endif
+
+
+opus_val32 celt_rcp(opus_val32 x);
+opus_val32 celt_rcp_norm32(opus_val32 x);
+
+#define celt_div(a,b) MULT32_32_Q31((opus_val32)(a),celt_rcp(b))
+
+opus_val32 frac_div32_q29(opus_val32 a, opus_val32 b);
+opus_val32 frac_div32(opus_val32 a, opus_val32 b);
+
+/* Computes atan(x) multiplied by 2/PI. The input value (x) should be within the
+ * range of -1 to 1 and represented in Q30 format. The function will return the
+ * result in Q30 format. */
+static OPUS_INLINE opus_val32 celt_atan_norm(opus_val32 x)
+{
+   /* Approximation constants. */
+   static const opus_int32 ATAN_2_OVER_PI = 1367130551;   /* Q31 */
+   static const opus_int32 ATAN_COEFF_A03 = -715791936;   /* Q31 */
+   static const opus_int32 ATAN_COEFF_A05 = 857391616;    /* Q32 */
+   static const opus_int32 ATAN_COEFF_A07 = -1200579328;  /* Q33 */
+   static const opus_int32 ATAN_COEFF_A09 = 1682636672;   /* Q34 */
+   static const opus_int32 ATAN_COEFF_A11 = -1985085440;  /* Q35 */
+   static const opus_int32 ATAN_COEFF_A13 = 1583306112;   /* Q36 */
+   static const opus_int32 ATAN_COEFF_A15 = -598602432;   /* Q37 */
+   opus_int32 x_sq_q30;
+   opus_int32 x_q31;
+   opus_int32 tmp;
+   /* The expected x is in the range of [-1.0f, 1.0f] */
+   celt_sig_assert((x <= 1073741824) && (x >= -1073741824));
+
+   /* If x = 1.0f, returns 0.5f */
+   if (x == 1073741824)
+   {
+      return 536870912; /* 0.5f (Q30) */
+   }
+   /* If x = 1.0f, returns 0.5f */
+   if (x == -1073741824)
+   {
+      return -536870912; /* -0.5f (Q30) */
+   }
+   x_q31 = SHL32(x, 1);
+   x_sq_q30 = MULT32_32_Q31(x_q31, x);
+   /* Split evaluation in steps to avoid exploding macro expansion. */
+   tmp = MULT32_32_Q31(x_sq_q30, ATAN_COEFF_A15);
+   tmp = MULT32_32_Q31(x_sq_q30, ADD32(ATAN_COEFF_A13, tmp));
+   tmp = MULT32_32_Q31(x_sq_q30, ADD32(ATAN_COEFF_A11, tmp));
+   tmp = MULT32_32_Q31(x_sq_q30, ADD32(ATAN_COEFF_A09, tmp));
+   tmp = MULT32_32_Q31(x_sq_q30, ADD32(ATAN_COEFF_A07, tmp));
+   tmp = MULT32_32_Q31(x_sq_q30, ADD32(ATAN_COEFF_A05, tmp));
+   tmp = MULT32_32_Q31(x_sq_q30, ADD32(ATAN_COEFF_A03, tmp));
+   tmp = ADD32(x, MULT32_32_Q31(x_q31, tmp));
+   return MULT32_32_Q31(ATAN_2_OVER_PI, tmp);
+}
+
+/* Calculates the arctangent of y/x, multiplies the result by 2/pi, and returns
+ * the value in Q30 format. Both input values (x and y) must be within the range
+ * of 0 to 1 and represented in Q30 format. Inputs must be zero or greater, and
+ * at least one input must be non-zero. */
+static OPUS_INLINE opus_val32 celt_atan2p_norm(opus_val32 y, opus_val32 x)
+{
+   celt_sig_assert(x>=0 && y>=0);
+   if (y==0 && x==0) {
+      return 0;
+   } else if (y < x) {
+      return celt_atan_norm(SHR32(frac_div32(y, x), 1));
+   } else {
+      celt_sig_assert(y > 0);
+      return 1073741824 /* 1.0f Q30 */ -
+             celt_atan_norm(SHR32(frac_div32(x, y), 1));
+   }
+}
+
+#define M1 32767
+#define M2 -21
+#define M3 -11943
+#define M4 4936
+
+/* Atan approximation using a 4th order polynomial. Input is in Q15 format
+   and normalized by pi/4. Output is in Q15 format */
+static OPUS_INLINE opus_val16 celt_atan01(opus_val16 x)
+{
+   return MULT16_16_P15(x, ADD32(M1, MULT16_16_P15(x, ADD32(M2, MULT16_16_P15(x, ADD32(M3, MULT16_16_P15(M4, x)))))));
+}
+
+#undef M1
+#undef M2
+#undef M3
+#undef M4
+
+/* atan2() approximation valid for positive input values */
+static OPUS_INLINE opus_val16 celt_atan2p(opus_val16 y, opus_val16 x)
+{
+   if (x==0 && y==0) {
+      return 0;
+   } else if (y < x)
+   {
+      opus_val32 arg;
+      arg = celt_div(SHL32(EXTEND32(y),15),x);
+      if (arg >= 32767)
+         arg = 32767;
+      return SHR16(celt_atan01(EXTRACT16(arg)),1);
+   } else {
+      opus_val32 arg;
+      arg = celt_div(SHL32(EXTEND32(x),15),y);
+      if (arg >= 32767)
+         arg = 32767;
+      return 25736-SHR16(celt_atan01(EXTRACT16(arg)),1);
+   }
+}
+
+#endif /* FIXED_POINT */
+
+#ifndef DISABLE_FLOAT_API
+
+void celt_float2int16_c(const float * OPUS_RESTRICT in, short * OPUS_RESTRICT out, int cnt);
+
+#ifndef OVERRIDE_FLOAT2INT16
+#define celt_float2int16(in, out, cnt, arch) ((void)(arch), celt_float2int16_c(in, out, cnt))
+#endif
+
+int opus_limit2_checkwithin1_c(float *samples, int cnt);
+
+#ifndef OVERRIDE_LIMIT2_CHECKWITHIN1
+#define opus_limit2_checkwithin1(samples, cnt, arch) ((void)(arch), opus_limit2_checkwithin1_c(samples, cnt))
+#endif
+
+#endif /* DISABLE_FLOAT_API */
+
+#endif /* MATHOPS_H */
--- a/lib/libopus/celt/mdct.c
+++ b/lib/libopus/celt/mdct.c
@ -0,0 +1,390 @@
+/* Copyright (c) 2007-2008 CSIRO
+   Copyright (c) 2007-2008 Xiph.Org Foundation
+   Written by Jean-Marc Valin */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+/* This is a simple MDCT implementation that uses a N/4 complex FFT
+   to do most of the work. It should be relatively straightforward to
+   plug in pretty much and FFT here.
+
+   This replaces the Vorbis FFT (and uses the exact same API), which
+   was a bit too messy and that was ending up duplicating code
+   (might as well use the same FFT everywhere).
+
+   The algorithm is similar to (and inspired from) Fabrice Bellard's
+   MDCT implementation in FFMPEG, but has differences in signs, ordering
+   and scaling in many places.
+*/
+
+#ifndef SKIP_CONFIG_H
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+#endif
+
+#include "mdct.h"
+#include "kiss_fft.h"
+#include "_kiss_fft_guts.h"
+#include <math.h>
+#include "os_support.h"
+#include "mathops.h"
+#include "stack_alloc.h"
+
+#if defined(FIXED_POINT) && defined(__mips) && __mips == 32
+#include "mips/mdct_mipsr1.h"
+#endif
+
+#ifndef M_PI
+#define M_PI 3.141592653
+#endif
+
+#ifdef CUSTOM_MODES
+
+int clt_mdct_init(mdct_lookup *l,int N, int maxshift, int arch)
+{
+   int i;
+   kiss_twiddle_scalar *trig;
+   int shift;
+   int N2=N>>1;
+   l->n = N;
+   l->maxshift = maxshift;
+   for (i=0;i<=maxshift;i++)
+   {
+      if (i==0)
+         l->kfft[i] = opus_fft_alloc(N>>2>>i, 0, 0, arch);
+      else
+         l->kfft[i] = opus_fft_alloc_twiddles(N>>2>>i, 0, 0, l->kfft[0], arch);
+#ifndef ENABLE_TI_DSPLIB55
+      if (l->kfft[i]==NULL)
+         return 0;
+#endif
+   }
+   l->trig = trig = (kiss_twiddle_scalar*)opus_alloc((N-(N2>>maxshift))*sizeof(kiss_twiddle_scalar));
+   if (l->trig==NULL)
+     return 0;
+   for (shift=0;shift<=maxshift;shift++)
+   {
+      /* We have enough points that sine isn't necessary */
+#if defined(FIXED_POINT)
+#ifndef ENABLE_QEXT
+      for (i=0;i<N2;i++)
+         trig[i] = TRIG_UPSCALE*celt_cos_norm(DIV32(ADD32(SHL32(EXTEND32(i),17),N2+16384),N));
+#else
+      for (i=0;i<N2;i++)
+         trig[i] = (kiss_twiddle_scalar)MAX32(-2147483647,MIN32(2147483647,floor(.5+2147483648*cos(2*M_PI*(i+.125)/N))));
+#endif
+#else
+      for (i=0;i<N2;i++)
+         trig[i] = (kiss_twiddle_scalar)cos(2*PI*(i+.125)/N);
+#endif
+      trig += N2;
+      N2 >>= 1;
+      N >>= 1;
+   }
+   return 1;
+}
+
+void clt_mdct_clear(mdct_lookup *l, int arch)
+{
+   int i;
+   for (i=0;i<=l->maxshift;i++)
+      opus_fft_free(l->kfft[i], arch);
+   opus_free((kiss_twiddle_scalar*)l->trig);
+}
+
+#endif /* CUSTOM_MODES */
+
+/* Forward MDCT trashes the input array */
+#ifndef OVERRIDE_clt_mdct_forward
+void clt_mdct_forward_c(const mdct_lookup *l, kiss_fft_scalar *in, kiss_fft_scalar * OPUS_RESTRICT out,
+      const celt_coef *window, int overlap, int shift, int stride, int arch)
+{
+   int i;
+   int N, N2, N4;
+   VARDECL(kiss_fft_scalar, f);
+   VARDECL(kiss_fft_cpx, f2);
+   const kiss_fft_state *st = l->kfft[shift];
+   const kiss_twiddle_scalar *trig;
+   celt_coef scale;
+#ifdef FIXED_POINT
+   /* Allows us to scale with MULT16_32_Q16(), which is faster than
+      MULT16_32_Q15() on ARM. */
+   int scale_shift = st->scale_shift-1;
+   int headroom;
+#endif
+   SAVE_STACK;
+   (void)arch;
+   scale = st->scale;
+
+   N = l->n;
+   trig = l->trig;
+   for (i=0;i<shift;i++)
+   {
+      N >>= 1;
+      trig += N;
+   }
+   N2 = N>>1;
+   N4 = N>>2;
+
+   ALLOC(f, N2, kiss_fft_scalar);
+   ALLOC(f2, N4, kiss_fft_cpx);
+
+   /* Consider the input to be composed of four blocks: [a, b, c, d] */
+   /* Window, shuffle, fold */
+   {
+      /* Temp pointers to make it really clear to the compiler what we're doing */
+      const kiss_fft_scalar * OPUS_RESTRICT xp1 = in+(overlap>>1);
+      const kiss_fft_scalar * OPUS_RESTRICT xp2 = in+N2-1+(overlap>>1);
+      kiss_fft_scalar * OPUS_RESTRICT yp = f;
+      const celt_coef * OPUS_RESTRICT wp1 = window+(overlap>>1);
+      const celt_coef * OPUS_RESTRICT wp2 = window+(overlap>>1)-1;
+      for(i=0;i<((overlap+3)>>2);i++)
+      {
+         /* Real part arranged as -d-cR, Imag part arranged as -b+aR*/
+         *yp++ = S_MUL(xp1[N2], *wp2) + S_MUL(*xp2, *wp1);
+         *yp++ = S_MUL(*xp1, *wp1)    - S_MUL(xp2[-N2], *wp2);
+         xp1+=2;
+         xp2-=2;
+         wp1+=2;
+         wp2-=2;
+      }
+      wp1 = window;
+      wp2 = window+overlap-1;
+      for(;i<N4-((overlap+3)>>2);i++)
+      {
+         /* Real part arranged as a-bR, Imag part arranged as -c-dR */
+         *yp++ = *xp2;
+         *yp++ = *xp1;
+         xp1+=2;
+         xp2-=2;
+      }
+      for(;i<N4;i++)
+      {
+         /* Real part arranged as a-bR, Imag part arranged as -c-dR */
+         *yp++ =  -S_MUL(xp1[-N2], *wp1) + S_MUL(*xp2, *wp2);
+         *yp++ = S_MUL(*xp1, *wp2)     + S_MUL(xp2[N2], *wp1);
+         xp1+=2;
+         xp2-=2;
+         wp1+=2;
+         wp2-=2;
+      }
+   }
+   /* Pre-rotation */
+   {
+      kiss_fft_scalar * OPUS_RESTRICT yp = f;
+      const kiss_twiddle_scalar *t = &trig[0];
+#ifdef FIXED_POINT
+      opus_val32 maxval=1;
+#endif
+      for(i=0;i<N4;i++)
+      {
+         kiss_fft_cpx yc;
+         kiss_twiddle_scalar t0, t1;
+         kiss_fft_scalar re, im, yr, yi;
+         t0 = t[i];
+         t1 = t[N4+i];
+         re = *yp++;
+         im = *yp++;
+         yr = S_MUL(re,t0)  -  S_MUL(im,t1);
+         yi = S_MUL(im,t0)  +  S_MUL(re,t1);
+         /* For QEXT, it's best to scale before the FFT, but otherwise it's best to scale after.
+            For floating-point it doesn't matter. */
+#ifdef ENABLE_QEXT
+         yc.r = yr;
+         yc.i = yi;
+#else
+         yc.r = S_MUL2(yr, scale);
+         yc.i = S_MUL2(yi, scale);
+#endif
+#ifdef FIXED_POINT
+         maxval = MAX32(maxval, MAX32(ABS32(yc.r), ABS32(yc.i)));
+#endif
+         f2[st->bitrev[i]] = yc;
+      }
+#ifdef FIXED_POINT
+      headroom = IMAX(0, IMIN(scale_shift, 28-celt_ilog2(maxval)));
+#endif
+   }
+
+   /* N/4 complex FFT, does not downscale anymore */
+   opus_fft_impl(st, f2 ARG_FIXED(scale_shift-headroom));
+
+   /* Post-rotate */
+   {
+      /* Temp pointers to make it really clear to the compiler what we're doing */
+      const kiss_fft_cpx * OPUS_RESTRICT fp = f2;
+      kiss_fft_scalar * OPUS_RESTRICT yp1 = out;
+      kiss_fft_scalar * OPUS_RESTRICT yp2 = out+stride*(N2-1);
+      const kiss_twiddle_scalar *t = &trig[0];
+      /* Temp pointers to make it really clear to the compiler what we're doing */
+      for(i=0;i<N4;i++)
+      {
+         kiss_fft_scalar yr, yi;
+         kiss_fft_scalar t0, t1;
+#ifdef ENABLE_QEXT
+         t0 = S_MUL2(t[i], scale);
+         t1 = S_MUL2(t[N4+i], scale);
+#else
+         t0 = t[i];
+         t1 = t[N4+i];
+#endif
+         yr = PSHR32(S_MUL(fp->i,t1) - S_MUL(fp->r,t0), headroom);
+         yi = PSHR32(S_MUL(fp->r,t1) + S_MUL(fp->i,t0), headroom);
+         *yp1 = yr;
+         *yp2 = yi;
+         fp++;
+         yp1 += 2*stride;
+         yp2 -= 2*stride;
+      }
+   }
+   RESTORE_STACK;
+}
+#endif /* OVERRIDE_clt_mdct_forward */
+
+#ifndef OVERRIDE_clt_mdct_backward
+void clt_mdct_backward_c(const mdct_lookup *l, kiss_fft_scalar *in, kiss_fft_scalar * OPUS_RESTRICT out,
+      const celt_coef * OPUS_RESTRICT window, int overlap, int shift, int stride, int arch)
+{
+   int i;
+   int N, N2, N4;
+   const kiss_twiddle_scalar *trig;
+#ifdef FIXED_POINT
+   int pre_shift, post_shift, fft_shift;
+#endif
+   (void) arch;
+
+   N = l->n;
+   trig = l->trig;
+   for (i=0;i<shift;i++)
+   {
+      N >>= 1;
+      trig += N;
+   }
+   N2 = N>>1;
+   N4 = N>>2;
+
+#ifdef FIXED_POINT
+   {
+      opus_val32 sumval=N2;
+      opus_val32 maxval=0;
+      for (i=0;i<N2;i++) {
+         maxval = MAX32(maxval, ABS32(in[i*stride]));
+         sumval = ADD32_ovflw(sumval, ABS32(SHR32(in[i*stride],11)));
+      }
+      pre_shift = IMAX(0, 29-celt_zlog2(1+maxval));
+      /* Worst-case where all the energy goes to a single sample. */
+      post_shift = IMAX(0, 19-celt_ilog2(ABS32(sumval)));
+      post_shift = IMIN(post_shift, pre_shift);
+      fft_shift = pre_shift - post_shift;
+   }
+#endif
+   /* Pre-rotate */
+   {
+      /* Temp pointers to make it really clear to the compiler what we're doing */
+      const kiss_fft_scalar * OPUS_RESTRICT xp1 = in;
+      const kiss_fft_scalar * OPUS_RESTRICT xp2 = in+stride*(N2-1);
+      kiss_fft_scalar * OPUS_RESTRICT yp = out+(overlap>>1);
+      const kiss_twiddle_scalar * OPUS_RESTRICT t = &trig[0];
+      const opus_int16 * OPUS_RESTRICT bitrev = l->kfft[shift]->bitrev;
+      for(i=0;i<N4;i++)
+      {
+         int rev;
+         kiss_fft_scalar yr, yi;
+         opus_val32 x1, x2;
+         rev = *bitrev++;
+         x1 = SHL32_ovflw(*xp1, pre_shift);
+         x2 = SHL32_ovflw(*xp2, pre_shift);
+         yr = ADD32_ovflw(S_MUL(x2, t[i]), S_MUL(x1, t[N4+i]));
+         yi = SUB32_ovflw(S_MUL(x1, t[i]), S_MUL(x2, t[N4+i]));
+         /* We swap real and imag because we use an FFT instead of an IFFT. */
+         yp[2*rev+1] = yr;
+         yp[2*rev] = yi;
+         /* Storing the pre-rotation directly in the bitrev order. */
+         xp1+=2*stride;
+         xp2-=2*stride;
+      }
+   }
+
+   opus_fft_impl(l->kfft[shift], (kiss_fft_cpx*)(out+(overlap>>1)) ARG_FIXED(fft_shift));
+
+   /* Post-rotate and de-shuffle from both ends of the buffer at once to make
+      it in-place. */
+   {
+      kiss_fft_scalar * yp0 = out+(overlap>>1);
+      kiss_fft_scalar * yp1 = out+(overlap>>1)+N2-2;
+      const kiss_twiddle_scalar *t = &trig[0];
+      /* Loop to (N4+1)>>1 to handle odd N4. When N4 is odd, the
+         middle pair will be computed twice. */
+      for(i=0;i<(N4+1)>>1;i++)
+      {
+         kiss_fft_scalar re, im, yr, yi;
+         kiss_twiddle_scalar t0, t1;
+         /* We swap real and imag because we're using an FFT instead of an IFFT. */
+         re = yp0[1];
+         im = yp0[0];
+         t0 = t[i];
+         t1 = t[N4+i];
+         /* We'd scale up by 2 here, but instead it's done when mixing the windows */
+         yr = PSHR32_ovflw(ADD32_ovflw(S_MUL(re,t0), S_MUL(im,t1)), post_shift);
+         yi = PSHR32_ovflw(SUB32_ovflw(S_MUL(re,t1), S_MUL(im,t0)), post_shift);
+         /* We swap real and imag because we're using an FFT instead of an IFFT. */
+         re = yp1[1];
+         im = yp1[0];
+         yp0[0] = yr;
+         yp1[1] = yi;
+
+         t0 = t[(N4-i-1)];
+         t1 = t[(N2-i-1)];
+         /* We'd scale up by 2 here, but instead it's done when mixing the windows */
+         yr = PSHR32_ovflw(ADD32_ovflw(S_MUL(re,t0), S_MUL(im,t1)), post_shift);
+         yi = PSHR32_ovflw(SUB32_ovflw(S_MUL(re,t1), S_MUL(im,t0)), post_shift);
+         yp1[0] = yr;
+         yp0[1] = yi;
+         yp0 += 2;
+         yp1 -= 2;
+      }
+   }
+
+   /* Mirror on both sides for TDAC */
+   {
+      kiss_fft_scalar * OPUS_RESTRICT xp1 = out+overlap-1;
+      kiss_fft_scalar * OPUS_RESTRICT yp1 = out;
+      const celt_coef * OPUS_RESTRICT wp1 = window;
+      const celt_coef * OPUS_RESTRICT wp2 = window+overlap-1;
+
+      for(i = 0; i < overlap/2; i++)
+      {
+         kiss_fft_scalar x1, x2;
+         x1 = *xp1;
+         x2 = *yp1;
+         *yp1++ = SUB32_ovflw(S_MUL(x2, *wp2), S_MUL(x1, *wp1));
+         *xp1-- = ADD32_ovflw(S_MUL(x2, *wp1), S_MUL(x1, *wp2));
+         wp1++;
+         wp2--;
+      }
+   }
+}
+#endif /* OVERRIDE_clt_mdct_backward */
--- a/lib/libopus/celt/mdct.h
+++ b/lib/libopus/celt/mdct.h
@ -0,0 +1,111 @@
+/* Copyright (c) 2007-2008 CSIRO
+   Copyright (c) 2007-2008 Xiph.Org Foundation
+   Written by Jean-Marc Valin */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+/* This is a simple MDCT implementation that uses a N/4 complex FFT
+   to do most of the work. It should be relatively straightforward to
+   plug in pretty much and FFT here.
+
+   This replaces the Vorbis FFT (and uses the exact same API), which
+   was a bit too messy and that was ending up duplicating code
+   (might as well use the same FFT everywhere).
+
+   The algorithm is similar to (and inspired from) Fabrice Bellard's
+   MDCT implementation in FFMPEG, but has differences in signs, ordering
+   and scaling in many places.
+*/
+
+#ifndef MDCT_H
+#define MDCT_H
+
+#include "opus_defines.h"
+#include "kiss_fft.h"
+#include "arch.h"
+
+typedef struct {
+   int n;
+   int maxshift;
+   const kiss_fft_state *kfft[4];
+   const kiss_twiddle_scalar * OPUS_RESTRICT trig;
+} mdct_lookup;
+
+#if defined(HAVE_ARM_NE10)
+#include "arm/mdct_arm.h"
+#endif
+
+int clt_mdct_init(mdct_lookup *l,int N, int maxshift, int arch);
+void clt_mdct_clear(mdct_lookup *l, int arch);
+
+/** Compute a forward MDCT and scale by 4/N, trashes the input array */
+void clt_mdct_forward_c(const mdct_lookup *l, kiss_fft_scalar *in,
+                        kiss_fft_scalar * OPUS_RESTRICT out,
+                        const celt_coef *window, int overlap,
+                        int shift, int stride, int arch);
+
+/** Compute a backward MDCT (no scaling) and performs weighted overlap-add
+    (scales implicitly by 1/2) */
+void clt_mdct_backward_c(const mdct_lookup *l, kiss_fft_scalar *in,
+      kiss_fft_scalar * OPUS_RESTRICT out,
+      const celt_coef * OPUS_RESTRICT window,
+      int overlap, int shift, int stride, int arch);
+
+#if !defined(OVERRIDE_OPUS_MDCT)
+/* Is run-time CPU detection enabled on this platform? */
+#if defined(OPUS_HAVE_RTCD) && defined(HAVE_ARM_NE10)
+
+extern void (*const CLT_MDCT_FORWARD_IMPL[OPUS_ARCHMASK+1])(
+      const mdct_lookup *l, kiss_fft_scalar *in,
+      kiss_fft_scalar * OPUS_RESTRICT out, const celt_coef *window,
+      int overlap, int shift, int stride, int arch);
+
+#define clt_mdct_forward(_l, _in, _out, _window, _overlap, _shift, _stride, _arch) \
+   ((*CLT_MDCT_FORWARD_IMPL[(arch)&OPUS_ARCHMASK])(_l, _in, _out, \
+                                                   _window, _overlap, _shift, \
+                                                   _stride, _arch))
+
+extern void (*const CLT_MDCT_BACKWARD_IMPL[OPUS_ARCHMASK+1])(
+      const mdct_lookup *l, kiss_fft_scalar *in,
+      kiss_fft_scalar * OPUS_RESTRICT out, const celt_coef *window,
+      int overlap, int shift, int stride, int arch);
+
+#define clt_mdct_backward(_l, _in, _out, _window, _overlap, _shift, _stride, _arch) \
+   (*CLT_MDCT_BACKWARD_IMPL[(arch)&OPUS_ARCHMASK])(_l, _in, _out, \
+                                                   _window, _overlap, _shift, \
+                                                   _stride, _arch)
+
+#else /* if defined(OPUS_HAVE_RTCD) && defined(HAVE_ARM_NE10) */
+
+#define clt_mdct_forward(_l, _in, _out, _window, _overlap, _shift, _stride, _arch) \
+   clt_mdct_forward_c(_l, _in, _out, _window, _overlap, _shift, _stride, _arch)
+
+#define clt_mdct_backward(_l, _in, _out, _window, _overlap, _shift, _stride, _arch) \
+   clt_mdct_backward_c(_l, _in, _out, _window, _overlap, _shift, _stride, _arch)
+
+#endif /* end if defined(OPUS_HAVE_RTCD) && defined(HAVE_ARM_NE10) && !defined(FIXED_POINT) */
+#endif /* end if !defined(OVERRIDE_OPUS_MDCT) */
+
+#endif
--- a/lib/libopus/celt/mfrngcod.h
+++ b/lib/libopus/celt/mfrngcod.h
@ -0,0 +1,48 @@
+/* Copyright (c) 2001-2008 Timothy B. Terriberry
+   Copyright (c) 2008-2009 Xiph.Org Foundation */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#if !defined(_mfrngcode_H)
+# define _mfrngcode_H (1)
+# include "entcode.h"
+
+/*Constants used by the entropy encoder/decoder.*/
+
+/*The number of bits to output at a time.*/
+# define EC_SYM_BITS   (8)
+/*The total number of bits in each of the state registers.*/
+# define EC_CODE_BITS  (32)
+/*The maximum symbol value.*/
+# define EC_SYM_MAX    ((1U<<EC_SYM_BITS)-1)
+/*Bits to shift by to move a symbol into the high-order position.*/
+# define EC_CODE_SHIFT (EC_CODE_BITS-EC_SYM_BITS-1)
+/*Carry bit of the high-order range symbol.*/
+# define EC_CODE_TOP   (((opus_uint32)1U)<<(EC_CODE_BITS-1))
+/*Low-order bit of the high-order range symbol.*/
+# define EC_CODE_BOT   (EC_CODE_TOP>>EC_SYM_BITS)
+/*The number of bits available for the last, partial symbol in the code field.*/
+# define EC_CODE_EXTRA ((EC_CODE_BITS-2)%EC_SYM_BITS+1)
+#endif
--- a/lib/libopus/celt/modes.c
+++ b/lib/libopus/celt/modes.c
@ -0,0 +1,520 @@
+/* Copyright (c) 2007-2008 CSIRO
+   Copyright (c) 2007-2009 Xiph.Org Foundation
+   Copyright (c) 2008 Gregory Maxwell
+   Written by Jean-Marc Valin and Gregory Maxwell */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "celt.h"
+#include "modes.h"
+#include "rate.h"
+#include "os_support.h"
+#include "stack_alloc.h"
+#include "quant_bands.h"
+#include "cpu_support.h"
+
+static const opus_int16 eband5ms[] = {
+/*0  200 400 600 800  1k 1.2 1.4 1.6  2k 2.4 2.8 3.2  4k 4.8 5.6 6.8  8k 9.6 12k 15.6 */
+  0,  1,  2,  3,  4,  5,  6,  7,  8, 10, 12, 14, 16, 20, 24, 28, 34, 40, 48, 60, 78, 100
+};
+
+/* Alternate tuning (partially derived from Vorbis) */
+#define BITALLOC_SIZE 11
+/* Bit allocation table in units of 1/32 bit/sample (0.1875 dB SNR) */
+static const unsigned char band_allocation[] = {
+/*0  200 400 600 800  1k 1.2 1.4 1.6  2k 2.4 2.8 3.2  4k 4.8 5.6 6.8  8k 9.6 12k 15.6 */
+  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,
+ 90, 80, 75, 69, 63, 56, 49, 40, 34, 29, 20, 18, 10,  0,  0,  0,  0,  0,  0,  0,  0,
+110,100, 90, 84, 78, 71, 65, 58, 51, 45, 39, 32, 26, 20, 12,  0,  0,  0,  0,  0,  0,
+118,110,103, 93, 86, 80, 75, 70, 65, 59, 53, 47, 40, 31, 23, 15,  4,  0,  0,  0,  0,
+126,119,112,104, 95, 89, 83, 78, 72, 66, 60, 54, 47, 39, 32, 25, 17, 12,  1,  0,  0,
+134,127,120,114,103, 97, 91, 85, 78, 72, 66, 60, 54, 47, 41, 35, 29, 23, 16, 10,  1,
+144,137,130,124,113,107,101, 95, 88, 82, 76, 70, 64, 57, 51, 45, 39, 33, 26, 15,  1,
+152,145,138,132,123,117,111,105, 98, 92, 86, 80, 74, 67, 61, 55, 49, 43, 36, 20,  1,
+162,155,148,142,133,127,121,115,108,102, 96, 90, 84, 77, 71, 65, 59, 53, 46, 30,  1,
+172,165,158,152,143,137,131,125,118,112,106,100, 94, 87, 81, 75, 69, 63, 56, 45, 20,
+200,200,200,200,200,200,200,200,198,193,188,183,178,173,168,163,158,153,148,129,104,
+};
+
+#ifndef CUSTOM_MODES_ONLY
+ #ifdef FIXED_POINT
+  #include "static_modes_fixed.h"
+ #else
+  #include "static_modes_float.h"
+ #endif
+#endif /* CUSTOM_MODES_ONLY */
+
+#ifndef M_PI
+#define M_PI 3.1415926535897931
+#endif
+
+#ifdef CUSTOM_MODES
+
+/* Defining 25 critical bands for the full 0-20 kHz audio bandwidth
+   Taken from http://ccrma.stanford.edu/~jos/bbt/Bark_Frequency_Scale.html */
+#define BARK_BANDS 25
+static const opus_int16 bark_freq[BARK_BANDS+1] = {
+      0,   100,   200,   300,   400,
+    510,   630,   770,   920,  1080,
+   1270,  1480,  1720,  2000,  2320,
+   2700,  3150,  3700,  4400,  5300,
+   6400,  7700,  9500, 12000, 15500,
+  20000};
+
+static opus_int16 *compute_ebands(opus_int32 Fs, int frame_size, int res, int *nbEBands)
+{
+   opus_int16 *eBands;
+   int i, j, lin, low, high, nBark, offset=0;
+
+   /* All modes that have 2.5 ms short blocks use the same definition */
+   if (Fs == 400*(opus_int32)frame_size)
+   {
+      *nbEBands = sizeof(eband5ms)/sizeof(eband5ms[0])-1;
+      eBands = opus_alloc(sizeof(opus_int16)*(*nbEBands+1));
+      for (i=0;i<*nbEBands+1;i++)
+         eBands[i] = eband5ms[i];
+      return eBands;
+   }
+   /* Find the number of critical bands supported by our sampling rate */
+   for (nBark=1;nBark<BARK_BANDS;nBark++)
+    if (bark_freq[nBark+1]*2 >= Fs)
+       break;
+
+   /* Find where the linear part ends (i.e. where the spacing is more than min_width */
+   for (lin=0;lin<nBark;lin++)
+      if (bark_freq[lin+1]-bark_freq[lin] >= res)
+         break;
+
+   low = (bark_freq[lin]+res/2)/res;
+   high = nBark-lin;
+   *nbEBands = low+high;
+   eBands = opus_alloc(sizeof(opus_int16)*(*nbEBands+2));
+
+   if (eBands==NULL)
+      return NULL;
+
+   /* Linear spacing (min_width) */
+   for (i=0;i<low;i++)
+      eBands[i] = i;
+   if (low>0)
+      offset = eBands[low-1]*res - bark_freq[lin-1];
+   /* Spacing follows critical bands */
+   for (i=0;i<high;i++)
+   {
+      int target = bark_freq[lin+i];
+      /* Round to an even value */
+      eBands[i+low] = (target+offset/2+res)/(2*res)*2;
+      offset = eBands[i+low]*res - target;
+   }
+   /* Enforce the minimum spacing at the boundary */
+   for (i=0;i<*nbEBands;i++)
+      if (eBands[i] < i)
+         eBands[i] = i;
+   /* Round to an even value */
+   eBands[*nbEBands] = (bark_freq[nBark]+res)/(2*res)*2;
+   if (eBands[*nbEBands] > frame_size)
+      eBands[*nbEBands] = frame_size;
+   for (i=1;i<*nbEBands-1;i++)
+   {
+      if (eBands[i+1]-eBands[i] < eBands[i]-eBands[i-1])
+      {
+         eBands[i] -= (2*eBands[i]-eBands[i-1]-eBands[i+1])/2;
+      }
+   }
+   /* Remove any empty bands. */
+   for (i=j=0;i<*nbEBands;i++)
+      if(eBands[i+1]>eBands[j])
+         eBands[++j]=eBands[i+1];
+   *nbEBands=j;
+
+   for (i=1;i<*nbEBands;i++)
+   {
+      /* Every band must be smaller than the last band. */
+      celt_assert(eBands[i]-eBands[i-1]<=eBands[*nbEBands]-eBands[*nbEBands-1]);
+      /* Each band must be no larger than twice the size of the previous one. */
+      celt_assert(eBands[i+1]-eBands[i]<=2*(eBands[i]-eBands[i-1]));
+   }
+
+   return eBands;
+}
+
+static void compute_allocation_table(CELTMode *mode)
+{
+   int i, j;
+   unsigned char *allocVectors;
+   int maxBands = sizeof(eband5ms)/sizeof(eband5ms[0])-1;
+
+   mode->nbAllocVectors = BITALLOC_SIZE;
+   allocVectors = opus_alloc(sizeof(unsigned char)*(BITALLOC_SIZE*mode->nbEBands));
+   if (allocVectors==NULL)
+   {
+      mode->allocVectors = NULL;
+      return;
+   }
+
+   /* Check for standard mode */
+   if (mode->Fs == 400*(opus_int32)mode->shortMdctSize)
+   {
+      for (i=0;i<BITALLOC_SIZE*mode->nbEBands;i++)
+         allocVectors[i] = band_allocation[i];
+      mode->allocVectors = allocVectors;
+      return;
+   }
+   /* If not the standard mode, interpolate */
+   /* Compute per-codec-band allocation from per-critical-band matrix */
+   for (i=0;i<BITALLOC_SIZE;i++)
+   {
+      for (j=0;j<mode->nbEBands;j++)
+      {
+         int k;
+         for (k=0;k<maxBands;k++)
+         {
+            if (400*(opus_int32)eband5ms[k] > mode->eBands[j]*(opus_int32)mode->Fs/mode->shortMdctSize)
+               break;
+         }
+         if (k>maxBands-1)
+            allocVectors[i*mode->nbEBands+j] = band_allocation[i*maxBands + maxBands-1];
+         else {
+            opus_int32 a0, a1;
+            a1 = mode->eBands[j]*(opus_int32)mode->Fs/mode->shortMdctSize - 400*(opus_int32)eband5ms[k-1];
+            a0 = 400*(opus_int32)eband5ms[k] - mode->eBands[j]*(opus_int32)mode->Fs/mode->shortMdctSize;
+            allocVectors[i*mode->nbEBands+j] = (a0*band_allocation[i*maxBands+k-1]
+                                             + a1*band_allocation[i*maxBands+k])/(a0+a1);
+         }
+      }
+   }
+
+   /*printf ("\n");
+   for (i=0;i<BITALLOC_SIZE;i++)
+   {
+      for (j=0;j<mode->nbEBands;j++)
+         printf ("%d ", allocVectors[i*mode->nbEBands+j]);
+      printf ("\n");
+   }
+   exit(0);*/
+
+   mode->allocVectors = allocVectors;
+}
+
+#endif /* CUSTOM_MODES */
+
+CELTMode *opus_custom_mode_create(opus_int32 Fs, int frame_size, int *error)
+{
+   int i;
+#ifdef CUSTOM_MODES
+   CELTMode *mode=NULL;
+   int res;
+   celt_coef *window;
+   opus_int16 *logN;
+   int LM;
+   int arch = opus_select_arch();
+   ALLOC_STACK;
+#if !defined(VAR_ARRAYS) && !defined(USE_ALLOCA)
+   if (global_stack==NULL)
+      goto failure;
+#endif
+#endif
+
+#ifndef CUSTOM_MODES_ONLY
+   for (i=0;i<TOTAL_MODES;i++)
+   {
+      int j;
+      for (j=0;j<4;j++)
+      {
+         if (Fs == static_mode_list[i]->Fs &&
+               (frame_size<<j) == static_mode_list[i]->shortMdctSize*static_mode_list[i]->nbShortMdcts)
+         {
+            if (error)
+               *error = OPUS_OK;
+            return (CELTMode*)static_mode_list[i];
+         }
+      }
+   }
+#endif /* CUSTOM_MODES_ONLY */
+
+#ifndef CUSTOM_MODES
+   if (error)
+      *error = OPUS_BAD_ARG;
+   return NULL;
+#else
+
+   /* The good thing here is that permutation of the arguments will automatically be invalid */
+
+   if (Fs < 8000 || Fs > 96000)
+   {
+      if (error)
+         *error = OPUS_BAD_ARG;
+      return NULL;
+   }
+#ifdef ENABLE_QEXT
+   if (frame_size < 40 || frame_size > 2048 || frame_size%2!=0)
+#else
+   if (frame_size < 40 || frame_size > 1024 || frame_size%2!=0)
+#endif
+   {
+      if (error)
+         *error = OPUS_BAD_ARG;
+      return NULL;
+   }
+   /* Frames of less than 1ms are not supported. */
+   if ((opus_int32)frame_size*1000 < Fs)
+   {
+      if (error)
+         *error = OPUS_BAD_ARG;
+      return NULL;
+   }
+
+   if ((opus_int32)frame_size*75 >= Fs && (frame_size%16)==0)
+   {
+     LM = 3;
+   } else if ((opus_int32)frame_size*150 >= Fs && (frame_size%8)==0)
+   {
+     LM = 2;
+   } else if ((opus_int32)frame_size*300 >= Fs && (frame_size%4)==0)
+   {
+     LM = 1;
+   } else
+   {
+     LM = 0;
+   }
+
+   /* Shorts longer than 3.3ms are not supported. */
+   if ((opus_int32)(frame_size>>LM)*300 > Fs)
+   {
+      if (error)
+         *error = OPUS_BAD_ARG;
+      return NULL;
+   }
+
+   mode = opus_alloc(sizeof(CELTMode));
+   if (mode==NULL)
+      goto failure;
+   mode->Fs = Fs;
+
+   /* Pre/de-emphasis depends on sampling rate. The "standard" pre-emphasis
+      is defined as A(z) = 1 - 0.85*z^-1 at 48 kHz. Other rates should
+      approximate that. */
+#ifdef ENABLE_QEXT
+   if(Fs == 96000) /* 96 kHz */
+   {
+      mode->preemph[0] =  QCONST16(0.9230041504f, 15);
+      mode->preemph[1] =  QCONST16(0.2200012207f, 15);
+      mode->preemph[2] =  QCONST16(1.5128347184f, SIG_SHIFT); /* exact 1/preemph[3] */
+      mode->preemph[3] =  QCONST16(0.6610107422f, 13);
+   } else
+#endif
+   if(Fs < 12000) /* 8 kHz */
+   {
+      mode->preemph[0] =  QCONST16(0.3500061035f, 15);
+      mode->preemph[1] = -QCONST16(0.1799926758f, 15);
+      mode->preemph[2] =  QCONST16(0.2719968125f, SIG_SHIFT); /* exact 1/preemph[3] */
+      mode->preemph[3] =  QCONST16(3.6765136719f, 13);
+   } else if(Fs < 24000) /* 16 kHz */
+   {
+      mode->preemph[0] =  QCONST16(0.6000061035f, 15);
+      mode->preemph[1] = -QCONST16(0.1799926758f, 15);
+      mode->preemph[2] =  QCONST16(0.4424998650f, SIG_SHIFT); /* exact 1/preemph[3] */
+      mode->preemph[3] =  QCONST16(2.2598876953f, 13);
+   } else if(Fs < 40000) /* 32 kHz */
+   {
+      mode->preemph[0] =  QCONST16(0.7799987793f, 15);
+      mode->preemph[1] = -QCONST16(0.1000061035f, 15);
+      mode->preemph[2] =  QCONST16(0.7499771125f, SIG_SHIFT); /* exact 1/preemph[3] */
+      mode->preemph[3] =  QCONST16(1.3333740234f, 13);
+   } else /* 48 kHz */
+   {
+      mode->preemph[0] =  QCONST16(0.8500061035f, 15);
+      mode->preemph[1] =  QCONST16(0.0f, 15);
+      mode->preemph[2] =  QCONST16(1.f, SIG_SHIFT);
+      mode->preemph[3] =  QCONST16(1.f, 13);
+   }
+
+   mode->maxLM = LM;
+   mode->nbShortMdcts = 1<<LM;
+   mode->shortMdctSize = frame_size/mode->nbShortMdcts;
+   res = (mode->Fs+mode->shortMdctSize)/(2*mode->shortMdctSize);
+
+   mode->eBands = compute_ebands(Fs, mode->shortMdctSize, res, &mode->nbEBands);
+   if (mode->eBands==NULL)
+      goto failure;
+#if !defined(SMALL_FOOTPRINT)
+   /* Make sure we don't allocate a band larger than our PVQ table.
+      208 should be enough, but let's be paranoid. */
+   if ((mode->eBands[mode->nbEBands] - mode->eBands[mode->nbEBands-1])<<LM >
+    208) {
+       goto failure;
+   }
+#endif
+
+   mode->effEBands = mode->nbEBands;
+   while (mode->eBands[mode->effEBands] > mode->shortMdctSize)
+      mode->effEBands--;
+
+   /* Overlap must be divisible by 4 */
+   mode->overlap = ((mode->shortMdctSize>>2)<<2);
+
+   compute_allocation_table(mode);
+   if (mode->allocVectors==NULL)
+      goto failure;
+
+   window = (celt_coef*)opus_alloc(mode->overlap*sizeof(*window));
+   if (window==NULL)
+      goto failure;
+
+#ifndef FIXED_POINT
+   for (i=0;i<mode->overlap;i++)
+      window[i] = Q15ONE*sin(.5*M_PI* sin(.5*M_PI*(i+.5)/mode->overlap) * sin(.5*M_PI*(i+.5)/mode->overlap));
+#else
+# ifdef ENABLE_QEXT
+   for (i=0;i<mode->overlap;i++)
+      window[i] = MIN32(2147483647, 2147483648*sin(.5*M_PI* sin(.5*M_PI*(i+.5)/mode->overlap) * sin(.5*M_PI*(i+.5)/mode->overlap)));
+# else
+   for (i=0;i<mode->overlap;i++)
+      window[i] = MIN32(32767,floor(.5+32768.*sin(.5*M_PI* sin(.5*M_PI*(i+.5)/mode->overlap) * sin(.5*M_PI*(i+.5)/mode->overlap))));
+# endif
+#endif
+   mode->window = window;
+
+   logN = (opus_int16*)opus_alloc(mode->nbEBands*sizeof(opus_int16));
+   if (logN==NULL)
+      goto failure;
+
+   for (i=0;i<mode->nbEBands;i++)
+      logN[i] = log2_frac(mode->eBands[i+1]-mode->eBands[i], BITRES);
+   mode->logN = logN;
+
+   compute_pulse_cache(mode, mode->maxLM);
+#ifdef ENABLE_QEXT
+   OPUS_CLEAR(&mode->qext_cache, 1);
+   if ( (mode->Fs == 48000 && (mode->shortMdctSize==120 || mode->shortMdctSize==90)) || (mode->Fs == 96000 && (mode->shortMdctSize==240 || mode->shortMdctSize==180)) ) {
+      CELTMode dummy;
+      compute_qext_mode(&dummy, mode);
+      compute_pulse_cache(&dummy, dummy.maxLM);
+      OPUS_COPY(&mode->qext_cache, &dummy.cache, 1);
+   }
+#endif
+
+   if (clt_mdct_init(&mode->mdct, 2*mode->shortMdctSize*mode->nbShortMdcts,
+           mode->maxLM, arch) == 0)
+      goto failure;
+
+   if (error)
+      *error = OPUS_OK;
+
+   return mode;
+failure:
+   if (error)
+      *error = OPUS_ALLOC_FAIL;
+   if (mode!=NULL)
+      opus_custom_mode_destroy(mode);
+   return NULL;
+#endif /* !CUSTOM_MODES */
+}
+
+#if defined(CUSTOM_MODES) || defined(ENABLE_OPUS_CUSTOM_API)
+void opus_custom_mode_destroy(CELTMode *mode)
+{
+   int arch = opus_select_arch();
+
+   if (mode == NULL)
+      return;
+#ifndef CUSTOM_MODES_ONLY
+   {
+     int i;
+     for (i=0;i<TOTAL_MODES;i++)
+     {
+        if (mode == static_mode_list[i])
+        {
+           return;
+        }
+     }
+   }
+#endif /* CUSTOM_MODES_ONLY */
+#ifdef CUSTOM_MODES
+#ifdef ENABLE_QEXT
+   if (mode->qext_cache.index) opus_free((opus_int16*)mode->qext_cache.index);
+   if (mode->qext_cache.bits) opus_free((unsigned char*)mode->qext_cache.bits);
+   if (mode->qext_cache.caps) opus_free((unsigned char*)mode->qext_cache.caps);
+#endif
+   opus_free((opus_int16*)mode->eBands);
+   opus_free((unsigned char*)mode->allocVectors);
+
+   opus_free((opus_val16*)mode->window);
+   opus_free((opus_int16*)mode->logN);
+
+   opus_free((opus_int16*)mode->cache.index);
+   opus_free((unsigned char*)mode->cache.bits);
+   opus_free((unsigned char*)mode->cache.caps);
+   clt_mdct_clear(&mode->mdct, arch);
+
+   opus_free((CELTMode *)mode);
+#else
+   (void)arch;
+   celt_assert(0);
+#endif
+}
+#endif
+
+#ifdef ENABLE_QEXT
+
+static const opus_int16 qext_eBands_180[] = {
+/* 20k  22k  24k  26k  28k  30k  32k  34k  36k  38k  40k  42k  44k  47k  48k */
+    74,  82,  90,  98, 106, 114, 122, 130, 138, 146, 154, 162, 168, 174, 180
+};
+
+static const opus_int16 qext_logN_180[] = {24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 24, 21, 21, 21};
+
+/* Extra bands. */
+static const opus_int16 qext_eBands_240[] = {
+/* 20k  22k  24k  26k  28k  30k  32k  34k  36k  38k  40k  42k  44k  47k  48k */
+   100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240
+};
+
+static const opus_int16 qext_logN_240[] = {27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27, 27};
+
+void compute_qext_mode(CELTMode *qext, const CELTMode *m)
+{
+   OPUS_COPY(qext, m, 1);
+   if (m->shortMdctSize*48000 == 120*m->Fs) {
+      qext->eBands = qext_eBands_240;
+      qext->logN = qext_logN_240;
+   } else if (m->shortMdctSize*48000 == 90*m->Fs) {
+      qext->eBands = qext_eBands_180;
+      qext->logN = qext_logN_180;
+   } else {
+      celt_assert(0);
+   }
+   qext->nbEBands = qext->effEBands = NB_QEXT_BANDS;
+   while (qext->eBands[qext->effEBands] > qext->shortMdctSize)
+      qext->effEBands--;
+   qext->nbAllocVectors = 0;
+   qext->allocVectors = NULL;
+   OPUS_COPY(&qext->cache, &m->qext_cache, 1);
+}
+#endif
--- a/lib/libopus/celt/modes.h
+++ b/lib/libopus/celt/modes.h
@ -0,0 +1,85 @@
+/* Copyright (c) 2007-2008 CSIRO
+   Copyright (c) 2007-2009 Xiph.Org Foundation
+   Copyright (c) 2008 Gregory Maxwell
+   Written by Jean-Marc Valin and Gregory Maxwell */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef MODES_H
+#define MODES_H
+
+#include "opus_types.h"
+#include "celt.h"
+#include "arch.h"
+#include "mdct.h"
+#include "entenc.h"
+#include "entdec.h"
+
+#define MAX_PERIOD 1024
+
+#define DEC_PITCH_BUF_SIZE 2048
+
+typedef struct {
+   int size;
+   const opus_int16 *index;
+   const unsigned char *bits;
+   const unsigned char *caps;
+} PulseCache;
+
+/** Mode definition (opaque)
+ @brief Mode definition
+ */
+struct OpusCustomMode {
+   opus_int32 Fs;
+   int          overlap;
+
+   int          nbEBands;
+   int          effEBands;
+   opus_val16    preemph[4];
+   const opus_int16   *eBands;   /**< Definition for each "pseudo-critical band" */
+
+   int         maxLM;
+   int         nbShortMdcts;
+   int         shortMdctSize;
+
+   int          nbAllocVectors; /**< Number of lines in the matrix below */
+   const unsigned char   *allocVectors;   /**< Number of bits in each band for several rates */
+   const opus_int16 *logN;
+
+   const celt_coef *window;
+   mdct_lookup mdct;
+   PulseCache cache;
+#ifdef ENABLE_QEXT
+   PulseCache qext_cache;
+#endif
+};
+
+#ifdef ENABLE_QEXT
+#define QEXT_PACKET_SIZE_CAP 3825
+#define NB_QEXT_BANDS 14
+void compute_qext_mode(CELTMode *qext, const CELTMode *m);
+#endif
+
+#endif
--- a/lib/libopus/celt/os_support.h
+++ b/lib/libopus/celt/os_support.h
@ -0,0 +1,98 @@
+/* Copyright (C) 2007 Jean-Marc Valin
+
+   File: os_support.h
+   This is the (tiny) OS abstraction layer. Aside from math.h, this is the
+   only place where system headers are allowed.
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions are
+   met:
+
+   1. Redistributions of source code must retain the above copyright notice,
+   this list of conditions and the following disclaimer.
+
+   2. Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
+   IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
+   OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+   DISCLAIMED. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT,
+   INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+   (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+   SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
+   HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
+   STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
+   ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+   POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef OS_SUPPORT_H
+#define OS_SUPPORT_H
+
+#ifdef CUSTOM_SUPPORT
+#  include "custom_support.h"
+#endif
+
+#include "opus_types.h"
+#include "opus_defines.h"
+
+#include <string.h>
+#include <stdlib.h>
+
+/** Opus wrapper for malloc(). To do your own dynamic allocation replace this function, opus_realloc, and opus_free */
+#ifndef OVERRIDE_OPUS_ALLOC
+static OPUS_INLINE void *opus_alloc (size_t size)
+{
+   return malloc(size);
+}
+#endif
+
+#ifndef OVERRIDE_OPUS_REALLOC
+static OPUS_INLINE void *opus_realloc (void *ptr, size_t size)
+{
+   return realloc(ptr, size);
+}
+#endif
+
+/** Used only for non-threadsafe pseudostack.
+    If desired, this can always return the same area of memory rather than allocating a new one every time. */
+#ifndef OVERRIDE_OPUS_ALLOC_SCRATCH
+static OPUS_INLINE void *opus_alloc_scratch (size_t size)
+{
+   /* Scratch space doesn't need to be cleared */
+   return opus_alloc(size);
+}
+#endif
+
+/** Opus wrapper for free(). To do your own dynamic allocation replace this function, opus_realloc, and opus_free */
+#ifndef OVERRIDE_OPUS_FREE
+static OPUS_INLINE void opus_free (void *ptr)
+{
+   free(ptr);
+}
+#endif
+
+/** Copy n elements from src to dst. The 0* term provides compile-time type checking  */
+#ifndef OVERRIDE_OPUS_COPY
+#define OPUS_COPY(dst, src, n) (memcpy((dst), (src), (n)*sizeof(*(dst)) + 0*((dst)-(src)) ))
+#endif
+
+/** Copy n elements from src to dst, allowing overlapping regions. The 0* term
+    provides compile-time type checking */
+#ifndef OVERRIDE_OPUS_MOVE
+#define OPUS_MOVE(dst, src, n) (memmove((dst), (src), (n)*sizeof(*(dst)) + 0*((dst)-(src)) ))
+#endif
+
+/** Set n elements of dst to zero */
+#ifndef OVERRIDE_OPUS_CLEAR
+#define OPUS_CLEAR(dst, n) (memset((dst), 0, (n)*sizeof(*(dst))))
+#endif
+
+/*#ifdef __GNUC__
+#pragma GCC poison printf sprintf
+#pragma GCC poison malloc free realloc calloc
+#endif*/
+
+#endif /* OS_SUPPORT_H */
--- a/lib/libopus/celt/pitch.c
+++ b/lib/libopus/celt/pitch.c
@ -0,0 +1,560 @@
+/* Copyright (c) 2007-2008 CSIRO
+   Copyright (c) 2007-2009 Xiph.Org Foundation
+   Written by Jean-Marc Valin */
+/**
+   @file pitch.c
+   @brief Pitch analysis
+ */
+
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "pitch.h"
+#include "os_support.h"
+#include "modes.h"
+#include "stack_alloc.h"
+#include "mathops.h"
+#include "celt_lpc.h"
+
+static void find_best_pitch(opus_val32 *xcorr, opus_val16 *y, int len,
+                            int max_pitch, int *best_pitch
+#ifdef FIXED_POINT
+                            , int yshift, opus_val32 maxcorr
+#endif
+                            )
+{
+   int i, j;
+   opus_val32 Syy=1;
+   opus_val16 best_num[2];
+   opus_val32 best_den[2];
+#ifdef FIXED_POINT
+   int xshift;
+
+   xshift = celt_ilog2(maxcorr)-14;
+#endif
+
+   best_num[0] = -1;
+   best_num[1] = -1;
+   best_den[0] = 0;
+   best_den[1] = 0;
+   best_pitch[0] = 0;
+   best_pitch[1] = 1;
+   for (j=0;j<len;j++)
+      Syy = ADD32(Syy, SHR32(MULT16_16(y[j],y[j]), yshift));
+   for (i=0;i<max_pitch;i++)
+   {
+      if (xcorr[i]>0)
+      {
+         opus_val16 num;
+         opus_val32 xcorr16;
+         xcorr16 = EXTRACT16(VSHR32(xcorr[i], xshift));
+#ifndef FIXED_POINT
+         /* Considering the range of xcorr16, this should avoid both underflows
+            and overflows (inf) when squaring xcorr16 */
+         xcorr16 *= 1e-12f;
+#endif
+         num = MULT16_16_Q15(xcorr16,xcorr16);
+         if (MULT16_32_Q15(num,best_den[1]) > MULT16_32_Q15(best_num[1],Syy))
+         {
+            if (MULT16_32_Q15(num,best_den[0]) > MULT16_32_Q15(best_num[0],Syy))
+            {
+               best_num[1] = best_num[0];
+               best_den[1] = best_den[0];
+               best_pitch[1] = best_pitch[0];
+               best_num[0] = num;
+               best_den[0] = Syy;
+               best_pitch[0] = i;
+            } else {
+               best_num[1] = num;
+               best_den[1] = Syy;
+               best_pitch[1] = i;
+            }
+         }
+      }
+      Syy += SHR32(MULT16_16(y[i+len],y[i+len]),yshift) - SHR32(MULT16_16(y[i],y[i]),yshift);
+      Syy = MAX32(1, Syy);
+   }
+}
+
+static void celt_fir5(opus_val16 *x,
+         const opus_val16 *num,
+         int N)
+{
+   int i;
+   opus_val16 num0, num1, num2, num3, num4;
+   opus_val32 mem0, mem1, mem2, mem3, mem4;
+   num0=num[0];
+   num1=num[1];
+   num2=num[2];
+   num3=num[3];
+   num4=num[4];
+   mem0=0;
+   mem1=0;
+   mem2=0;
+   mem3=0;
+   mem4=0;
+   for (i=0;i<N;i++)
+   {
+      opus_val32 sum = SHL32(EXTEND32(x[i]), SIG_SHIFT);
+      sum = MAC16_16(sum,num0,mem0);
+      sum = MAC16_16(sum,num1,mem1);
+      sum = MAC16_16(sum,num2,mem2);
+      sum = MAC16_16(sum,num3,mem3);
+      sum = MAC16_16(sum,num4,mem4);
+      mem4 = mem3;
+      mem3 = mem2;
+      mem2 = mem1;
+      mem1 = mem0;
+      mem0 = x[i];
+      x[i] = ROUND16(sum, SIG_SHIFT);
+   }
+}
+
+
+void pitch_downsample(celt_sig * OPUS_RESTRICT x[], opus_val16 * OPUS_RESTRICT x_lp,
+      int len, int C, int factor, int arch)
+{
+   int i;
+   opus_val32 ac[5];
+   opus_val16 tmp=Q15ONE;
+   opus_val16 lpc[4];
+   opus_val16 lpc2[5];
+   opus_val16 c1 = QCONST16(.8f,15);
+   int offset;
+#ifdef FIXED_POINT
+   int shift;
+   opus_val32 maxabs;
+#endif
+   offset = factor/2;
+#ifdef FIXED_POINT
+   maxabs = celt_maxabs32(x[0], len*factor);
+   if (C==2)
+   {
+      opus_val32 maxabs_1 = celt_maxabs32(x[1], len*factor);
+      maxabs = MAX32(maxabs, maxabs_1);
+   }
+   if (maxabs<1)
+      maxabs=1;
+   shift = celt_ilog2(maxabs)-10;
+   if (shift<0)
+      shift=0;
+   if (C==2)
+      shift++;
+   for (i=1;i<len;i++)
+      x_lp[i] = SHR32(x[0][(factor*i-offset)], shift+2) + SHR32(x[0][(factor*i+offset)], shift+2) + SHR32(x[0][factor*i], shift+1);
+   x_lp[0] = SHR32(x[0][offset], shift+2) + SHR32(x[0][0], shift+1);
+   if (C==2)
+   {
+      for (i=1;i<len;i++)
+         x_lp[i] += SHR32(x[1][(factor*i-offset)], shift+2) + SHR32(x[1][(factor*i+offset)], shift+2) + SHR32(x[1][factor*i], shift+1);
+      x_lp[0] += SHR32(x[1][offset], shift+2) + SHR32(x[1][0], shift+1);
+   }
+#else
+   for (i=1;i<len;i++)
+      x_lp[i] = .25f*x[0][(factor*i-offset)] + .25f*x[0][(factor*i+offset)] + .5f*x[0][factor*i];
+   x_lp[0] = .25f*x[0][offset] + .5f*x[0][0];
+   if (C==2)
+   {
+      for (i=1;i<len;i++)
+         x_lp[i] += .25f*x[1][(factor*i-offset)] + .25f*x[1][(factor*i+offset)] + .5f*x[1][factor*i];
+      x_lp[0] += .25f*x[1][offset] + .5f*x[1][0];
+   }
+#endif
+   _celt_autocorr(x_lp, ac, NULL, 0,
+                  4, len, arch);
+
+   /* Noise floor -40 dB */
+#ifdef FIXED_POINT
+   ac[0] += SHR32(ac[0],13);
+#else
+   ac[0] *= 1.0001f;
+#endif
+   /* Lag windowing */
+   for (i=1;i<=4;i++)
+   {
+      /*ac[i] *= exp(-.5*(2*M_PI*.002*i)*(2*M_PI*.002*i));*/
+#ifdef FIXED_POINT
+      ac[i] -= MULT16_32_Q15(2*i*i, ac[i]);
+#else
+      ac[i] -= ac[i]*(.008f*i)*(.008f*i);
+#endif
+   }
+
+   _celt_lpc(lpc, ac, 4);
+   for (i=0;i<4;i++)
+   {
+      tmp = MULT16_16_Q15(QCONST16(.9f,15), tmp);
+      lpc[i] = MULT16_16_Q15(lpc[i], tmp);
+   }
+   /* Add a zero */
+   lpc2[0] = lpc[0] + QCONST16(.8f,SIG_SHIFT);
+   lpc2[1] = lpc[1] + MULT16_16_Q15(c1,lpc[0]);
+   lpc2[2] = lpc[2] + MULT16_16_Q15(c1,lpc[1]);
+   lpc2[3] = lpc[3] + MULT16_16_Q15(c1,lpc[2]);
+   lpc2[4] = MULT16_16_Q15(c1,lpc[3]);
+   celt_fir5(x_lp, lpc2, len);
+}
+
+/* Pure C implementation. */
+#ifdef FIXED_POINT
+opus_val32
+#else
+void
+#endif
+celt_pitch_xcorr_c(const opus_val16 *_x, const opus_val16 *_y,
+      opus_val32 *xcorr, int len, int max_pitch, int arch)
+{
+
+#if 0 /* This is a simple version of the pitch correlation that should work
+         well on DSPs like Blackfin and TI C5x/C6x */
+   int i, j;
+#ifdef FIXED_POINT
+   opus_val32 maxcorr=1;
+#endif
+#if !defined(OVERRIDE_PITCH_XCORR)
+   (void)arch;
+#endif
+   for (i=0;i<max_pitch;i++)
+   {
+      opus_val32 sum = 0;
+      for (j=0;j<len;j++)
+         sum = MAC16_16(sum, _x[j], _y[i+j]);
+      xcorr[i] = sum;
+#ifdef FIXED_POINT
+      maxcorr = MAX32(maxcorr, sum);
+#endif
+   }
+#ifdef FIXED_POINT
+   return maxcorr;
+#endif
+
+#else /* Unrolled version of the pitch correlation -- runs faster on x86 and ARM */
+   int i;
+   /*The EDSP version requires that max_pitch is at least 1, and that _x is
+      32-bit aligned.
+     Since it's hard to put asserts in assembly, put them here.*/
+#ifdef FIXED_POINT
+   opus_val32 maxcorr=1;
+#endif
+   celt_assert(max_pitch>0);
+   celt_sig_assert(((size_t)_x&3)==0);
+   for (i=0;i<max_pitch-3;i+=4)
+   {
+      opus_val32 sum[4]={0,0,0,0};
+#if defined(OPUS_CHECK_ASM) && defined(FIXED_POINT)
+      {
+         opus_val32 sum_c[4]={0,0,0,0};
+         xcorr_kernel_c(_x, _y+i, sum_c, len);
+#endif
+         xcorr_kernel(_x, _y+i, sum, len, arch);
+#if defined(OPUS_CHECK_ASM) && defined(FIXED_POINT)
+         celt_assert(memcmp(sum, sum_c, sizeof(sum)) == 0);
+      }
+#endif
+      xcorr[i]=sum[0];
+      xcorr[i+1]=sum[1];
+      xcorr[i+2]=sum[2];
+      xcorr[i+3]=sum[3];
+#ifdef FIXED_POINT
+      sum[0] = MAX32(sum[0], sum[1]);
+      sum[2] = MAX32(sum[2], sum[3]);
+      sum[0] = MAX32(sum[0], sum[2]);
+      maxcorr = MAX32(maxcorr, sum[0]);
+#endif
+   }
+   /* In case max_pitch isn't a multiple of 4, do non-unrolled version. */
+   for (;i<max_pitch;i++)
+   {
+      opus_val32 sum;
+      sum = celt_inner_prod(_x, _y+i, len, arch);
+      xcorr[i] = sum;
+#ifdef FIXED_POINT
+      maxcorr = MAX32(maxcorr, sum);
+#endif
+   }
+#ifdef FIXED_POINT
+   return maxcorr;
+#endif
+#endif
+}
+
+void pitch_search(const opus_val16 * OPUS_RESTRICT x_lp, opus_val16 * OPUS_RESTRICT y,
+                  int len, int max_pitch, int *pitch, int arch)
+{
+   int i, j;
+   int lag;
+   int best_pitch[2]={0,0};
+   VARDECL(opus_val16, x_lp4);
+   VARDECL(opus_val16, y_lp4);
+   VARDECL(opus_val32, xcorr);
+#ifdef FIXED_POINT
+   opus_val32 maxcorr;
+   opus_val32 xmax, ymax;
+   int shift=0;
+#endif
+   int offset;
+
+   SAVE_STACK;
+
+   celt_assert(len>0);
+   celt_assert(max_pitch>0);
+   lag = len+max_pitch;
+
+   ALLOC(x_lp4, len>>2, opus_val16);
+   ALLOC(y_lp4, lag>>2, opus_val16);
+   ALLOC(xcorr, max_pitch>>1, opus_val32);
+
+   /* Downsample by 2 again */
+   for (j=0;j<len>>2;j++)
+      x_lp4[j] = x_lp[2*j];
+   for (j=0;j<lag>>2;j++)
+      y_lp4[j] = y[2*j];
+
+#ifdef FIXED_POINT
+   xmax = celt_maxabs16(x_lp4, len>>2);
+   ymax = celt_maxabs16(y_lp4, lag>>2);
+   shift = celt_ilog2(MAX32(1, MAX32(xmax, ymax))) - 14 + celt_ilog2(len)/2;
+   if (shift>0)
+   {
+      for (j=0;j<len>>2;j++)
+         x_lp4[j] = SHR16(x_lp4[j], shift);
+      for (j=0;j<lag>>2;j++)
+         y_lp4[j] = SHR16(y_lp4[j], shift);
+      /* Use double the shift for a MAC */
+      shift *= 2;
+   } else {
+      shift = 0;
+   }
+#endif
+
+   /* Coarse search with 4x decimation */
+
+#ifdef FIXED_POINT
+   maxcorr =
+#endif
+   celt_pitch_xcorr(x_lp4, y_lp4, xcorr, len>>2, max_pitch>>2, arch);
+
+   find_best_pitch(xcorr, y_lp4, len>>2, max_pitch>>2, best_pitch
+#ifdef FIXED_POINT
+                   , 0, maxcorr
+#endif
+                   );
+
+   /* Finer search with 2x decimation */
+#ifdef FIXED_POINT
+   maxcorr=1;
+#endif
+   for (i=0;i<max_pitch>>1;i++)
+   {
+      opus_val32 sum;
+      xcorr[i] = 0;
+      if (abs(i-2*best_pitch[0])>2 && abs(i-2*best_pitch[1])>2)
+         continue;
+#ifdef FIXED_POINT
+      sum = 0;
+      for (j=0;j<len>>1;j++)
+         sum += SHR32(MULT16_16(x_lp[j],y[i+j]), shift);
+#else
+      sum = celt_inner_prod(x_lp, y+i, len>>1, arch);
+#endif
+      xcorr[i] = MAX32(-1, sum);
+#ifdef FIXED_POINT
+      maxcorr = MAX32(maxcorr, sum);
+#endif
+   }
+   find_best_pitch(xcorr, y, len>>1, max_pitch>>1, best_pitch
+#ifdef FIXED_POINT
+                   , shift+1, maxcorr
+#endif
+                   );
+
+   /* Refine by pseudo-interpolation */
+   if (best_pitch[0]>0 && best_pitch[0]<(max_pitch>>1)-1)
+   {
+      opus_val32 a, b, c;
+      a = xcorr[best_pitch[0]-1];
+      b = xcorr[best_pitch[0]];
+      c = xcorr[best_pitch[0]+1];
+      if ((c-a) > MULT16_32_Q15(QCONST16(.7f,15),b-a))
+         offset = 1;
+      else if ((a-c) > MULT16_32_Q15(QCONST16(.7f,15),b-c))
+         offset = -1;
+      else
+         offset = 0;
+   } else {
+      offset = 0;
+   }
+   *pitch = 2*best_pitch[0]-offset;
+
+   RESTORE_STACK;
+}
+
+#ifdef FIXED_POINT
+static opus_val16 compute_pitch_gain(opus_val32 xy, opus_val32 xx, opus_val32 yy)
+{
+   opus_val32 x2y2;
+   int sx, sy, shift;
+   opus_val32 g;
+   opus_val16 den;
+   if (xy == 0 || xx == 0 || yy == 0)
+      return 0;
+   sx = celt_ilog2(xx)-14;
+   sy = celt_ilog2(yy)-14;
+   shift = sx + sy;
+   x2y2 = SHR32(MULT16_16(VSHR32(xx, sx), VSHR32(yy, sy)), 14);
+   if (shift & 1) {
+      if (x2y2 < 32768)
+      {
+         x2y2 <<= 1;
+         shift--;
+      } else {
+         x2y2 >>= 1;
+         shift++;
+      }
+   }
+   den = celt_rsqrt_norm(x2y2);
+   g = MULT16_32_Q15(den, xy);
+   g = VSHR32(g, (shift>>1)-1);
+   return EXTRACT16(MAX32(-Q15ONE, MIN32(g, Q15ONE)));
+}
+#else
+static opus_val16 compute_pitch_gain(opus_val32 xy, opus_val32 xx, opus_val32 yy)
+{
+   return xy/celt_sqrt(1+xx*yy);
+}
+#endif
+
+static const int second_check[16] = {0, 0, 3, 2, 3, 2, 5, 2, 3, 2, 3, 2, 5, 2, 3, 2};
+opus_val16 remove_doubling(opus_val16 *x, int maxperiod, int minperiod,
+      int N, int *T0_, int prev_period, opus_val16 prev_gain, int arch)
+{
+   int k, i, T, T0;
+   opus_val16 g, g0;
+   opus_val16 pg;
+   opus_val32 xy,xx,yy,xy2;
+   opus_val32 xcorr[3];
+   opus_val32 best_xy, best_yy;
+   int offset;
+   int minperiod0;
+   VARDECL(opus_val32, yy_lookup);
+   SAVE_STACK;
+
+   minperiod0 = minperiod;
+   maxperiod /= 2;
+   minperiod /= 2;
+   *T0_ /= 2;
+   prev_period /= 2;
+   N /= 2;
+   x += maxperiod;
+   if (*T0_>=maxperiod)
+      *T0_=maxperiod-1;
+
+   T = T0 = *T0_;
+   ALLOC(yy_lookup, maxperiod+1, opus_val32);
+   dual_inner_prod(x, x, x-T0, N, &xx, &xy, arch);
+   yy_lookup[0] = xx;
+   yy=xx;
+   for (i=1;i<=maxperiod;i++)
+   {
+      yy = yy+MULT16_16(x[-i],x[-i])-MULT16_16(x[N-i],x[N-i]);
+      yy_lookup[i] = MAX32(0, yy);
+   }
+   yy = yy_lookup[T0];
+   best_xy = xy;
+   best_yy = yy;
+   g = g0 = compute_pitch_gain(xy, xx, yy);
+   /* Look for any pitch at T/k */
+   for (k=2;k<=15;k++)
+   {
+      int T1, T1b;
+      opus_val16 g1;
+      opus_val16 cont=0;
+      opus_val16 thresh;
+      T1 = celt_udiv(2*T0+k, 2*k);
+      if (T1 < minperiod)
+         break;
+      /* Look for another strong correlation at T1b */
+      if (k==2)
+      {
+         if (T1+T0>maxperiod)
+            T1b = T0;
+         else
+            T1b = T0+T1;
+      } else
+      {
+         T1b = celt_udiv(2*second_check[k]*T0+k, 2*k);
+      }
+      dual_inner_prod(x, &x[-T1], &x[-T1b], N, &xy, &xy2, arch);
+      xy = HALF32(xy + xy2);
+      yy = HALF32(yy_lookup[T1] + yy_lookup[T1b]);
+      g1 = compute_pitch_gain(xy, xx, yy);
+      if (abs(T1-prev_period)<=1)
+         cont = prev_gain;
+      else if (abs(T1-prev_period)<=2 && 5*k*k < T0)
+         cont = HALF16(prev_gain);
+      else
+         cont = 0;
+      thresh = MAX16(QCONST16(.3f,15), MULT16_16_Q15(QCONST16(.7f,15),g0)-cont);
+      /* Bias against very high pitch (very short period) to avoid false-positives
+         due to short-term correlation */
+      if (T1<3*minperiod)
+         thresh = MAX16(QCONST16(.4f,15), MULT16_16_Q15(QCONST16(.85f,15),g0)-cont);
+      else if (T1<2*minperiod)
+         thresh = MAX16(QCONST16(.5f,15), MULT16_16_Q15(QCONST16(.9f,15),g0)-cont);
+      if (g1 > thresh)
+      {
+         best_xy = xy;
+         best_yy = yy;
+         T = T1;
+         g = g1;
+      }
+   }
+   best_xy = MAX32(0, best_xy);
+   if (best_yy <= best_xy)
+      pg = Q15ONE;
+   else
+      pg = SHR32(frac_div32(best_xy,best_yy+1),16);
+
+   for (k=0;k<3;k++)
+      xcorr[k] = celt_inner_prod(x, x-(T+k-1), N, arch);
+   if ((xcorr[2]-xcorr[0]) > MULT16_32_Q15(QCONST16(.7f,15),xcorr[1]-xcorr[0]))
+      offset = 1;
+   else if ((xcorr[0]-xcorr[2]) > MULT16_32_Q15(QCONST16(.7f,15),xcorr[1]-xcorr[2]))
+      offset = -1;
+   else
+      offset = 0;
+   if (pg > g)
+      pg = g;
+   *T0_ = 2*T+offset;
+
+   if (*T0_<minperiod0)
+      *T0_=minperiod0;
+   RESTORE_STACK;
+   return pg;
+}
--- a/lib/libopus/celt/pitch.h
+++ b/lib/libopus/celt/pitch.h
@ -0,0 +1,203 @@
+/* Copyright (c) 2007-2008 CSIRO
+   Copyright (c) 2007-2009 Xiph.Org Foundation
+   Written by Jean-Marc Valin */
+/**
+   @file pitch.h
+   @brief Pitch analysis
+ */
+
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef PITCH_H
+#define PITCH_H
+
+#include "modes.h"
+#include "cpu_support.h"
+
+#if (defined(OPUS_X86_MAY_HAVE_SSE) && !defined(FIXED_POINT)) \
+  || ((defined(OPUS_X86_MAY_HAVE_SSE4_1) || defined(OPUS_X86_MAY_HAVE_SSE2)) && defined(FIXED_POINT))
+#include "x86/pitch_sse.h"
+#endif
+
+#if defined(FIXED_POINT) && defined(__mips)
+#include "mips/pitch_mipsr1.h"
+#endif
+
+#if (defined(OPUS_ARM_ASM) || defined(OPUS_ARM_MAY_HAVE_NEON_INTR))
+# include "arm/pitch_arm.h"
+#endif
+
+void pitch_downsample(celt_sig * OPUS_RESTRICT x[], opus_val16 * OPUS_RESTRICT x_lp,
+      int len, int C, int factor, int arch);
+
+void pitch_search(const opus_val16 * OPUS_RESTRICT x_lp, opus_val16 * OPUS_RESTRICT y,
+                  int len, int max_pitch, int *pitch, int arch);
+
+opus_val16 remove_doubling(opus_val16 *x, int maxperiod, int minperiod,
+      int N, int *T0, int prev_period, opus_val16 prev_gain, int arch);
+
+
+/* OPT: This is the kernel you really want to optimize. It gets used a lot
+   by the prefilter and by the PLC. */
+static OPUS_INLINE void xcorr_kernel_c(const opus_val16 * x, const opus_val16 * y, opus_val32 sum[4], int len)
+{
+   int j;
+   opus_val16 y_0, y_1, y_2, y_3;
+   celt_assert(len>=3);
+   y_3=0; /* gcc doesn't realize that y_3 can't be used uninitialized */
+   y_0=*y++;
+   y_1=*y++;
+   y_2=*y++;
+   for (j=0;j<len-3;j+=4)
+   {
+      opus_val16 tmp;
+      tmp = *x++;
+      y_3=*y++;
+      sum[0] = MAC16_16(sum[0],tmp,y_0);
+      sum[1] = MAC16_16(sum[1],tmp,y_1);
+      sum[2] = MAC16_16(sum[2],tmp,y_2);
+      sum[3] = MAC16_16(sum[3],tmp,y_3);
+      tmp=*x++;
+      y_0=*y++;
+      sum[0] = MAC16_16(sum[0],tmp,y_1);
+      sum[1] = MAC16_16(sum[1],tmp,y_2);
+      sum[2] = MAC16_16(sum[2],tmp,y_3);
+      sum[3] = MAC16_16(sum[3],tmp,y_0);
+      tmp=*x++;
+      y_1=*y++;
+      sum[0] = MAC16_16(sum[0],tmp,y_2);
+      sum[1] = MAC16_16(sum[1],tmp,y_3);
+      sum[2] = MAC16_16(sum[2],tmp,y_0);
+      sum[3] = MAC16_16(sum[3],tmp,y_1);
+      tmp=*x++;
+      y_2=*y++;
+      sum[0] = MAC16_16(sum[0],tmp,y_3);
+      sum[1] = MAC16_16(sum[1],tmp,y_0);
+      sum[2] = MAC16_16(sum[2],tmp,y_1);
+      sum[3] = MAC16_16(sum[3],tmp,y_2);
+   }
+   if (j++<len)
+   {
+      opus_val16 tmp = *x++;
+      y_3=*y++;
+      sum[0] = MAC16_16(sum[0],tmp,y_0);
+      sum[1] = MAC16_16(sum[1],tmp,y_1);
+      sum[2] = MAC16_16(sum[2],tmp,y_2);
+      sum[3] = MAC16_16(sum[3],tmp,y_3);
+   }
+   if (j++<len)
+   {
+      opus_val16 tmp=*x++;
+      y_0=*y++;
+      sum[0] = MAC16_16(sum[0],tmp,y_1);
+      sum[1] = MAC16_16(sum[1],tmp,y_2);
+      sum[2] = MAC16_16(sum[2],tmp,y_3);
+      sum[3] = MAC16_16(sum[3],tmp,y_0);
+   }
+   if (j<len)
+   {
+      opus_val16 tmp=*x++;
+      y_1=*y++;
+      sum[0] = MAC16_16(sum[0],tmp,y_2);
+      sum[1] = MAC16_16(sum[1],tmp,y_3);
+      sum[2] = MAC16_16(sum[2],tmp,y_0);
+      sum[3] = MAC16_16(sum[3],tmp,y_1);
+   }
+}
+
+#ifndef OVERRIDE_XCORR_KERNEL
+#define xcorr_kernel(x, y, sum, len, arch) \
+    ((void)(arch),xcorr_kernel_c(x, y, sum, len))
+#endif /* OVERRIDE_XCORR_KERNEL */
+
+
+static OPUS_INLINE void dual_inner_prod_c(const opus_val16 *x, const opus_val16 *y01, const opus_val16 *y02,
+      int N, opus_val32 *xy1, opus_val32 *xy2)
+{
+   int i;
+   opus_val32 xy01=0;
+   opus_val32 xy02=0;
+   for (i=0;i<N;i++)
+   {
+      xy01 = MAC16_16(xy01, x[i], y01[i]);
+      xy02 = MAC16_16(xy02, x[i], y02[i]);
+   }
+   *xy1 = xy01;
+   *xy2 = xy02;
+}
+
+#ifndef OVERRIDE_DUAL_INNER_PROD
+# define dual_inner_prod(x, y01, y02, N, xy1, xy2, arch) \
+    ((void)(arch),dual_inner_prod_c(x, y01, y02, N, xy1, xy2))
+#endif
+
+/*We make sure a C version is always available for cases where the overhead of
+  vectorization and passing around an arch flag aren't worth it.*/
+static OPUS_INLINE opus_val32 celt_inner_prod_c(const opus_val16 *x,
+      const opus_val16 *y, int N)
+{
+   int i;
+   opus_val32 xy=0;
+   for (i=0;i<N;i++)
+      xy = MAC16_16(xy, x[i], y[i]);
+   return xy;
+}
+
+#if !defined(OVERRIDE_CELT_INNER_PROD)
+# define celt_inner_prod(x, y, N, arch) \
+    ((void)(arch),celt_inner_prod_c(x, y, N))
+#endif
+
+#ifdef NON_STATIC_COMB_FILTER_CONST_C
+void comb_filter_const_c(opus_val32 *y, opus_val32 *x, int T, int N,
+     opus_val16 g10, opus_val16 g11, opus_val16 g12);
+#endif
+
+
+#ifdef FIXED_POINT
+opus_val32
+#else
+void
+#endif
+celt_pitch_xcorr_c(const opus_val16 *_x, const opus_val16 *_y,
+      opus_val32 *xcorr, int len, int max_pitch, int arch);
+
+#ifndef OVERRIDE_PITCH_XCORR
+# define celt_pitch_xcorr celt_pitch_xcorr_c
+#endif
+
+#ifdef NON_STATIC_COMB_FILTER_CONST_C
+void comb_filter_const_c(opus_val32 *y, opus_val32 *x, int T, int N,
+                         opus_val16 g10, opus_val16 g11, opus_val16 g12);
+#endif
+
+#ifndef OVERRIDE_COMB_FILTER_CONST
+# define comb_filter_const(y, x, T, N, g10, g11, g12, arch) \
+    ((void)(arch),comb_filter_const_c(y, x, T, N, g10, g11, g12))
+#endif
+
+
+#endif
--- a/lib/libopus/celt/quant_bands.c
+++ b/lib/libopus/celt/quant_bands.c
@ -0,0 +1,572 @@
+/* Copyright (c) 2007-2008 CSIRO
+   Copyright (c) 2007-2009 Xiph.Org Foundation
+   Written by Jean-Marc Valin */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "quant_bands.h"
+#include "laplace.h"
+#include <math.h>
+#include "os_support.h"
+#include "arch.h"
+#include "mathops.h"
+#include "stack_alloc.h"
+#include "rate.h"
+
+#ifdef FIXED_POINT
+/* Mean energy in each band quantized in Q4 */
+const signed char eMeans[25] = {
+      103,100, 92, 85, 81,
+       77, 72, 70, 78, 75,
+       73, 71, 78, 74, 69,
+       72, 70, 74, 76, 71,
+       60, 60, 60, 60, 60
+};
+#else
+/* Mean energy in each band quantized in Q4 and converted back to float */
+const opus_val16 eMeans[25] = {
+      6.437500f, 6.250000f, 5.750000f, 5.312500f, 5.062500f,
+      4.812500f, 4.500000f, 4.375000f, 4.875000f, 4.687500f,
+      4.562500f, 4.437500f, 4.875000f, 4.625000f, 4.312500f,
+      4.500000f, 4.375000f, 4.625000f, 4.750000f, 4.437500f,
+      3.750000f, 3.750000f, 3.750000f, 3.750000f, 3.750000f
+};
+#endif
+/* prediction coefficients: 0.9, 0.8, 0.65, 0.5 */
+#ifdef FIXED_POINT
+static const opus_val16 pred_coef[4] = {29440, 26112, 21248, 16384};
+static const opus_val16 beta_coef[4] = {30147, 22282, 12124, 6554};
+static const opus_val16 beta_intra = 4915;
+#else
+static const opus_val16 pred_coef[4] = {29440/32768., 26112/32768., 21248/32768., 16384/32768.};
+static const opus_val16 beta_coef[4] = {30147/32768., 22282/32768., 12124/32768., 6554/32768.};
+static const opus_val16 beta_intra = 4915/32768.;
+#endif
+
+/*Parameters of the Laplace-like probability models used for the coarse energy.
+  There is one pair of parameters for each frame size, prediction type
+   (inter/intra), and band number.
+  The first number of each pair is the probability of 0, and the second is the
+   decay rate, both in Q8 precision.*/
+static const unsigned char e_prob_model[4][2][42] = {
+   /*120 sample frames.*/
+   {
+      /*Inter*/
+      {
+          72, 127,  65, 129,  66, 128,  65, 128,  64, 128,  62, 128,  64, 128,
+          64, 128,  92,  78,  92,  79,  92,  78,  90,  79, 116,  41, 115,  40,
+         114,  40, 132,  26, 132,  26, 145,  17, 161,  12, 176,  10, 177,  11
+      },
+      /*Intra*/
+      {
+          24, 179,  48, 138,  54, 135,  54, 132,  53, 134,  56, 133,  55, 132,
+          55, 132,  61, 114,  70,  96,  74,  88,  75,  88,  87,  74,  89,  66,
+          91,  67, 100,  59, 108,  50, 120,  40, 122,  37,  97,  43,  78,  50
+      }
+   },
+   /*240 sample frames.*/
+   {
+      /*Inter*/
+      {
+          83,  78,  84,  81,  88,  75,  86,  74,  87,  71,  90,  73,  93,  74,
+          93,  74, 109,  40, 114,  36, 117,  34, 117,  34, 143,  17, 145,  18,
+         146,  19, 162,  12, 165,  10, 178,   7, 189,   6, 190,   8, 177,   9
+      },
+      /*Intra*/
+      {
+          23, 178,  54, 115,  63, 102,  66,  98,  69,  99,  74,  89,  71,  91,
+          73,  91,  78,  89,  86,  80,  92,  66,  93,  64, 102,  59, 103,  60,
+         104,  60, 117,  52, 123,  44, 138,  35, 133,  31,  97,  38,  77,  45
+      }
+   },
+   /*480 sample frames.*/
+   {
+      /*Inter*/
+      {
+          61,  90,  93,  60, 105,  42, 107,  41, 110,  45, 116,  38, 113,  38,
+         112,  38, 124,  26, 132,  27, 136,  19, 140,  20, 155,  14, 159,  16,
+         158,  18, 170,  13, 177,  10, 187,   8, 192,   6, 175,   9, 159,  10
+      },
+      /*Intra*/
+      {
+          21, 178,  59, 110,  71,  86,  75,  85,  84,  83,  91,  66,  88,  73,
+          87,  72,  92,  75,  98,  72, 105,  58, 107,  54, 115,  52, 114,  55,
+         112,  56, 129,  51, 132,  40, 150,  33, 140,  29,  98,  35,  77,  42
+      }
+   },
+   /*960 sample frames.*/
+   {
+      /*Inter*/
+      {
+          42, 121,  96,  66, 108,  43, 111,  40, 117,  44, 123,  32, 120,  36,
+         119,  33, 127,  33, 134,  34, 139,  21, 147,  23, 152,  20, 158,  25,
+         154,  26, 166,  21, 173,  16, 184,  13, 184,  10, 150,  13, 139,  15
+      },
+      /*Intra*/
+      {
+          22, 178,  63, 114,  74,  82,  84,  83,  92,  82, 103,  62,  96,  72,
+          96,  67, 101,  73, 107,  72, 113,  55, 118,  52, 125,  52, 118,  52,
+         117,  55, 135,  49, 137,  39, 157,  32, 145,  29,  97,  33,  77,  40
+      }
+   }
+};
+
+static const unsigned char small_energy_icdf[3]={2,1,0};
+
+static opus_val32 loss_distortion(const celt_glog *eBands, celt_glog *oldEBands, int start, int end, int len, int C)
+{
+   int c, i;
+   opus_val32 dist = 0;
+   c=0; do {
+      for (i=start;i<end;i++)
+      {
+         celt_glog d = PSHR32(SUB32(eBands[i+c*len], oldEBands[i+c*len]), DB_SHIFT-7);
+         dist = MAC16_16(dist, d,d);
+      }
+   } while (++c<C);
+   return MIN32(200,SHR32(dist,14));
+}
+
+static int quant_coarse_energy_impl(const CELTMode *m, int start, int end,
+      const celt_glog *eBands, celt_glog *oldEBands,
+      opus_int32 budget, opus_int32 tell,
+      const unsigned char *prob_model, celt_glog *error, ec_enc *enc,
+      int C, int LM, int intra, celt_glog max_decay, int lfe)
+{
+   int i, c;
+   int badness = 0;
+   opus_val32 prev[2] = {0,0};
+   opus_val16 coef;
+   opus_val16 beta;
+
+   if (tell+3 <= budget)
+      ec_enc_bit_logp(enc, intra, 3);
+   if (intra)
+   {
+      coef = 0;
+      beta = beta_intra;
+   } else {
+      beta = beta_coef[LM];
+      coef = pred_coef[LM];
+   }
+
+   /* Encode at a fixed coarse resolution */
+   for (i=start;i<end;i++)
+   {
+      c=0;
+      do {
+         int bits_left;
+         int qi, qi0;
+         opus_val32 q;
+         celt_glog x;
+         opus_val32 f, tmp;
+         celt_glog oldE;
+         celt_glog decay_bound;
+         x = eBands[i+c*m->nbEBands];
+         oldE = MAXG(-GCONST(9.f), oldEBands[i+c*m->nbEBands]);
+#ifdef FIXED_POINT
+         f = x - MULT16_32_Q15(coef,oldE) - prev[c];
+         /* Rounding to nearest integer here is really important! */
+         qi = (f+QCONST32(.5f,DB_SHIFT))>>DB_SHIFT;
+         decay_bound = MAXG(-GCONST(28.f), SUB32((opus_val32)oldEBands[i+c*m->nbEBands],max_decay));
+#else
+         f = x-coef*oldE-prev[c];
+         /* Rounding to nearest integer here is really important! */
+         qi = (int)floor(.5f+f);
+         decay_bound = MAXG(-GCONST(28.f), oldEBands[i+c*m->nbEBands]) - max_decay;
+#endif
+         /* Prevent the energy from going down too quickly (e.g. for bands
+            that have just one bin) */
+         if (qi < 0 && x < decay_bound)
+         {
+            qi += (int)SHR32(SUB32(decay_bound,x), DB_SHIFT);
+            if (qi > 0)
+               qi = 0;
+         }
+         qi0 = qi;
+         /* If we don't have enough bits to encode all the energy, just assume
+             something safe. */
+         tell = ec_tell(enc);
+         bits_left = budget-tell-3*C*(end-i);
+         if (i!=start && bits_left < 30)
+         {
+            if (bits_left < 24)
+               qi = IMIN(1, qi);
+            if (bits_left < 16)
+               qi = IMAX(-1, qi);
+         }
+         if (lfe && i>=2)
+            qi = IMIN(qi, 0);
+         if (budget-tell >= 15)
+         {
+            int pi;
+            pi = 2*IMIN(i,20);
+            ec_laplace_encode(enc, &qi,
+                  prob_model[pi]<<7, prob_model[pi+1]<<6);
+         }
+         else if(budget-tell >= 2)
+         {
+            qi = IMAX(-1, IMIN(qi, 1));
+            ec_enc_icdf(enc, 2*qi^-(qi<0), small_energy_icdf, 2);
+         }
+         else if(budget-tell >= 1)
+         {
+            qi = IMIN(0, qi);
+            ec_enc_bit_logp(enc, -qi, 1);
+         }
+         else
+            qi = -1;
+         error[i+c*m->nbEBands] = f - SHL32(qi,DB_SHIFT);
+         badness += abs(qi0-qi);
+         q = (opus_val32)SHL32(EXTEND32(qi),DB_SHIFT);
+
+         tmp = MULT16_32_Q15(coef,oldE) + prev[c] + q;
+#ifdef FIXED_POINT
+         tmp = MAX32(-GCONST(28.f), tmp);
+#endif
+         oldEBands[i+c*m->nbEBands] = tmp;
+         prev[c] = prev[c] + q - MULT16_32_Q15(beta,q);
+      } while (++c < C);
+   }
+   return lfe ? 0 : badness;
+}
+
+void quant_coarse_energy(const CELTMode *m, int start, int end, int effEnd,
+      const celt_glog *eBands, celt_glog *oldEBands, opus_uint32 budget,
+      celt_glog *error, ec_enc *enc, int C, int LM, int nbAvailableBytes,
+      int force_intra, opus_val32 *delayedIntra, int two_pass, int loss_rate, int lfe)
+{
+   int intra;
+   celt_glog max_decay;
+   VARDECL(celt_glog, oldEBands_intra);
+   VARDECL(celt_glog, error_intra);
+   ec_enc enc_start_state;
+   opus_uint32 tell;
+   int badness1=0;
+   opus_int32 intra_bias;
+   opus_val32 new_distortion;
+   SAVE_STACK;
+
+   intra = force_intra || (!two_pass && *delayedIntra>2*C*(end-start) && nbAvailableBytes > (end-start)*C);
+   intra_bias = (opus_int32)((budget**delayedIntra*loss_rate)/(C*512));
+   new_distortion = loss_distortion(eBands, oldEBands, start, effEnd, m->nbEBands, C);
+
+   tell = ec_tell(enc);
+   if (tell+3 > budget)
+      two_pass = intra = 0;
+
+   max_decay = GCONST(16.f);
+   if (end-start>10)
+   {
+#ifdef FIXED_POINT
+      max_decay = SHL32(MIN32(SHR32(max_decay,DB_SHIFT-3), EXTEND32(nbAvailableBytes)),DB_SHIFT-3);
+#else
+      max_decay = MIN32(max_decay, .125f*nbAvailableBytes);
+#endif
+   }
+   if (lfe)
+      max_decay = GCONST(3.f);
+   enc_start_state = *enc;
+
+   ALLOC(oldEBands_intra, C*m->nbEBands, celt_glog);
+   ALLOC(error_intra, C*m->nbEBands, celt_glog);
+   OPUS_COPY(oldEBands_intra, oldEBands, C*m->nbEBands);
+
+   if (two_pass || intra)
+   {
+      badness1 = quant_coarse_energy_impl(m, start, end, eBands, oldEBands_intra, budget,
+            tell, e_prob_model[LM][1], error_intra, enc, C, LM, 1, max_decay, lfe);
+   }
+
+   if (!intra)
+   {
+      unsigned char *intra_buf;
+      ec_enc enc_intra_state;
+      opus_int32 tell_intra;
+      opus_uint32 nstart_bytes;
+      opus_uint32 nintra_bytes;
+      opus_uint32 save_bytes;
+      int badness2;
+      VARDECL(unsigned char, intra_bits);
+
+      tell_intra = ec_tell_frac(enc);
+
+      enc_intra_state = *enc;
+
+      nstart_bytes = ec_range_bytes(&enc_start_state);
+      nintra_bytes = ec_range_bytes(&enc_intra_state);
+      intra_buf = ec_get_buffer(&enc_intra_state) + nstart_bytes;
+      save_bytes = nintra_bytes-nstart_bytes;
+      if (save_bytes == 0)
+         save_bytes = ALLOC_NONE;
+      ALLOC(intra_bits, save_bytes, unsigned char);
+      /* Copy bits from intra bit-stream */
+      OPUS_COPY(intra_bits, intra_buf, nintra_bytes - nstart_bytes);
+
+      *enc = enc_start_state;
+
+      badness2 = quant_coarse_energy_impl(m, start, end, eBands, oldEBands, budget,
+            tell, e_prob_model[LM][intra], error, enc, C, LM, 0, max_decay, lfe);
+
+      if (two_pass && (badness1 < badness2 || (badness1 == badness2 && ((opus_int32)ec_tell_frac(enc))+intra_bias > tell_intra)))
+      {
+         *enc = enc_intra_state;
+         /* Copy intra bits to bit-stream */
+         OPUS_COPY(intra_buf, intra_bits, nintra_bytes - nstart_bytes);
+         OPUS_COPY(oldEBands, oldEBands_intra, C*m->nbEBands);
+         OPUS_COPY(error, error_intra, C*m->nbEBands);
+         intra = 1;
+      }
+   } else {
+      OPUS_COPY(oldEBands, oldEBands_intra, C*m->nbEBands);
+      OPUS_COPY(error, error_intra, C*m->nbEBands);
+   }
+
+   if (intra)
+      *delayedIntra = new_distortion;
+   else
+      *delayedIntra = ADD32(MULT16_32_Q15(MULT16_16_Q15(pred_coef[LM], pred_coef[LM]),*delayedIntra),
+            new_distortion);
+
+   RESTORE_STACK;
+}
+
+void quant_fine_energy(const CELTMode *m, int start, int end, celt_glog *oldEBands, celt_glog *error, int *prev_quant, int *extra_quant, ec_enc *enc, int C)
+{
+   int i, c;
+   /* Encode finer resolution */
+   for (i=start;i<end;i++)
+   {
+      opus_int16 extra, prev;
+      extra = 1<<extra_quant[i];
+      if (extra_quant[i] <= 0)
+         continue;
+      if (ec_tell(enc)+C*extra_quant[i] > (opus_int32)enc->storage*8) continue;
+      prev = (prev_quant!=NULL) ? prev_quant[i] : 0;
+      c=0;
+      do {
+         int q2;
+         celt_glog offset;
+#ifdef FIXED_POINT
+         /* Has to be without rounding */
+         q2 = VSHR32(ADD32(error[i+c*m->nbEBands], SHR32(GCONST(.5f), prev)), DB_SHIFT-extra_quant[i]-prev);
+#else
+         q2 = (int)floor((error[i+c*m->nbEBands]*(1<<prev)+.5f)*extra);
+#endif
+         if (q2 > extra-1)
+            q2 = extra-1;
+         if (q2<0)
+            q2 = 0;
+         ec_enc_bits(enc, q2, extra_quant[i]);
+#ifdef FIXED_POINT
+         offset = SUB32(VSHR32(2*q2+1, extra_quant[i]-DB_SHIFT+1), GCONST(.5f));
+         offset = SHR32(offset, prev);
+#else
+         offset = (q2+.5f)*(1<<(14-extra_quant[i]))*(1.f/16384) - .5f;
+         offset *= (1<<(14-prev))*(1.f/16384);
+#endif
+         oldEBands[i+c*m->nbEBands] += offset;
+         error[i+c*m->nbEBands] -= offset;
+         /*printf ("%f ", error[i] - offset);*/
+      } while (++c < C);
+   }
+}
+
+void quant_energy_finalise(const CELTMode *m, int start, int end, celt_glog *oldEBands, celt_glog *error, int *fine_quant, int *fine_priority, int bits_left, ec_enc *enc, int C)
+{
+   int i, prio, c;
+
+   /* Use up the remaining bits */
+   for (prio=0;prio<2;prio++)
+   {
+      for (i=start;i<end && bits_left>=C ;i++)
+      {
+         if (fine_quant[i] >= MAX_FINE_BITS || fine_priority[i]!=prio)
+            continue;
+         c=0;
+         do {
+            int q2;
+            celt_glog offset;
+            q2 = error[i+c*m->nbEBands]<0 ? 0 : 1;
+            ec_enc_bits(enc, q2, 1);
+#ifdef FIXED_POINT
+            offset = SHR32(SHL32(q2,DB_SHIFT)-GCONST(.5f),fine_quant[i]+1);
+#else
+            offset = (q2-.5f)*(1<<(14-fine_quant[i]-1))*(1.f/16384);
+#endif
+            if (oldEBands != NULL) oldEBands[i+c*m->nbEBands] += offset;
+            error[i+c*m->nbEBands] -= offset;
+            bits_left--;
+         } while (++c < C);
+      }
+   }
+}
+
+void unquant_coarse_energy(const CELTMode *m, int start, int end, celt_glog *oldEBands, int intra, ec_dec *dec, int C, int LM)
+{
+   const unsigned char *prob_model = e_prob_model[LM][intra];
+   int i, c;
+   opus_val64 prev[2] = {0, 0};
+   opus_val16 coef;
+   opus_val16 beta;
+   opus_int32 budget;
+   opus_int32 tell;
+
+   if (intra)
+   {
+      coef = 0;
+      beta = beta_intra;
+   } else {
+      beta = beta_coef[LM];
+      coef = pred_coef[LM];
+   }
+
+   budget = dec->storage*8;
+
+   /* Decode at a fixed coarse resolution */
+   for (i=start;i<end;i++)
+   {
+      c=0;
+      do {
+         int qi;
+         opus_val32 q;
+         opus_val32 tmp;
+         /* It would be better to express this invariant as a
+            test on C at function entry, but that isn't enough
+            to make the static analyzer happy. */
+         celt_sig_assert(c<2);
+         tell = ec_tell(dec);
+         if(budget-tell>=15)
+         {
+            int pi;
+            pi = 2*IMIN(i,20);
+            qi = ec_laplace_decode(dec,
+                  prob_model[pi]<<7, prob_model[pi+1]<<6);
+         }
+         else if(budget-tell>=2)
+         {
+            qi = ec_dec_icdf(dec, small_energy_icdf, 2);
+            qi = (qi>>1)^-(qi&1);
+         }
+         else if(budget-tell>=1)
+         {
+            qi = -ec_dec_bit_logp(dec, 1);
+         }
+         else
+            qi = -1;
+         q = (opus_val32)SHL32(EXTEND32(qi),DB_SHIFT);
+
+         oldEBands[i+c*m->nbEBands] = MAXG(-GCONST(9.f), oldEBands[i+c*m->nbEBands]);
+         tmp = MULT16_32_Q15(coef,oldEBands[i+c*m->nbEBands]) + prev[c] + q;
+#ifdef FIXED_POINT
+         tmp = MIN32(GCONST(28.f), MAX32(-GCONST(28.f), tmp));
+#endif
+         oldEBands[i+c*m->nbEBands] = tmp;
+         prev[c] = prev[c] + q - MULT16_32_Q15(beta,q);
+      } while (++c < C);
+   }
+}
+
+void unquant_fine_energy(const CELTMode *m, int start, int end, celt_glog *oldEBands, int *prev_quant, int *extra_quant, ec_dec *dec, int C)
+{
+   int i, c;
+   /* Decode finer resolution */
+   for (i=start;i<end;i++)
+   {
+      opus_int16 extra, prev;
+      extra = extra_quant[i];
+      if (extra_quant[i] <= 0)
+         continue;
+      if (ec_tell(dec)+C*extra_quant[i] > (opus_int32)dec->storage*8) continue;
+      prev = (prev_quant!=NULL) ? prev_quant[i] : 0;
+      c=0;
+      do {
+         int q2;
+         celt_glog offset;
+         q2 = ec_dec_bits(dec, extra);
+#ifdef FIXED_POINT
+         offset = SUB32(VSHR32(2*q2+1, extra-DB_SHIFT+1), GCONST(.5f));
+         offset = SHR32(offset, prev);
+#else
+         offset = (q2+.5f)*(1<<(14-extra))*(1.f/16384) - .5f;
+         offset *= (1<<(14-prev))*(1.f/16384);
+#endif
+         oldEBands[i+c*m->nbEBands] += offset;
+      } while (++c < C);
+   }
+}
+
+void unquant_energy_finalise(const CELTMode *m, int start, int end, celt_glog *oldEBands, int *fine_quant,  int *fine_priority, int bits_left, ec_dec *dec, int C)
+{
+   int i, prio, c;
+
+   /* Use up the remaining bits */
+   for (prio=0;prio<2;prio++)
+   {
+      for (i=start;i<end && bits_left>=C ;i++)
+      {
+         if (fine_quant[i] >= MAX_FINE_BITS || fine_priority[i]!=prio)
+            continue;
+         c=0;
+         do {
+            int q2;
+            celt_glog offset;
+            q2 = ec_dec_bits(dec, 1);
+#ifdef FIXED_POINT
+            offset = SHR32(SHL32(q2,DB_SHIFT)-GCONST(.5f),fine_quant[i]+1);
+#else
+            offset = (q2-.5f)*(1<<(14-fine_quant[i]-1))*(1.f/16384);
+#endif
+            if (oldEBands != NULL) oldEBands[i+c*m->nbEBands] += offset;
+            bits_left--;
+         } while (++c < C);
+      }
+   }
+}
+
+void amp2Log2(const CELTMode *m, int effEnd, int end,
+      celt_ener *bandE, celt_glog *bandLogE, int C)
+{
+   int c, i;
+   c=0;
+   do {
+      for (i=0;i<effEnd;i++)
+      {
+         bandLogE[i+c*m->nbEBands] =
+               celt_log2_db(bandE[i+c*m->nbEBands])
+               - SHL32((celt_glog)eMeans[i],DB_SHIFT-4);
+#ifdef FIXED_POINT
+         /* Compensate for bandE[] being Q12 but celt_log2() taking a Q14 input. */
+         bandLogE[i+c*m->nbEBands] += GCONST(2.f);
+#endif
+      }
+      for (i=effEnd;i<end;i++)
+         bandLogE[c*m->nbEBands+i] = -GCONST(14.f);
+   } while (++c < C);
+}
--- a/lib/libopus/celt/quant_bands.h
+++ b/lib/libopus/celt/quant_bands.h
@ -0,0 +1,66 @@
+/* Copyright (c) 2007-2008 CSIRO
+   Copyright (c) 2007-2009 Xiph.Org Foundation
+   Written by Jean-Marc Valin */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef QUANT_BANDS
+#define QUANT_BANDS
+
+#include "arch.h"
+#include "modes.h"
+#include "entenc.h"
+#include "entdec.h"
+#include "mathops.h"
+
+#ifdef FIXED_POINT
+extern const signed char eMeans[25];
+#else
+extern const opus_val16 eMeans[25];
+#endif
+
+void amp2Log2(const CELTMode *m, int effEnd, int end,
+      celt_ener *bandE, celt_glog *bandLogE, int C);
+
+void log2Amp(const CELTMode *m, int start, int end,
+      celt_ener *eBands, const celt_glog *oldEBands, int C);
+
+void quant_coarse_energy(const CELTMode *m, int start, int end, int effEnd,
+      const celt_glog *eBands, celt_glog *oldEBands, opus_uint32 budget,
+      celt_glog *error, ec_enc *enc, int C, int LM,
+      int nbAvailableBytes, int force_intra, opus_val32 *delayedIntra,
+      int two_pass, int loss_rate, int lfe);
+
+void quant_fine_energy(const CELTMode *m, int start, int end, celt_glog *oldEBands, celt_glog *error, int *fine_quant, int *extra_quant, ec_enc *enc, int C);
+
+void quant_energy_finalise(const CELTMode *m, int start, int end, celt_glog *oldEBands, celt_glog *error, int *fine_quant, int *fine_priority, int bits_left, ec_enc *enc, int C);
+
+void unquant_coarse_energy(const CELTMode *m, int start, int end, celt_glog *oldEBands, int intra, ec_dec *dec, int C, int LM);
+
+void unquant_fine_energy(const CELTMode *m, int start, int end, celt_glog *oldEBands, int *fine_quant, int *extra_quant, ec_dec *dec, int C);
+
+void unquant_energy_finalise(const CELTMode *m, int start, int end, celt_glog *oldEBands, int *fine_quant, int *fine_priority, int bits_left, ec_dec *dec, int C);
+
+#endif /* QUANT_BANDS */
--- a/lib/libopus/celt/rate.c
+++ b/lib/libopus/celt/rate.c
@ -0,0 +1,876 @@
+/* Copyright (c) 2007-2008 CSIRO
+   Copyright (c) 2007-2009 Xiph.Org Foundation
+   Written by Jean-Marc Valin */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include <math.h>
+#include "modes.h"
+#include "cwrs.h"
+#include "arch.h"
+#include "os_support.h"
+
+#include "entcode.h"
+#include "rate.h"
+#include "quant_bands.h"
+
+static const unsigned char LOG2_FRAC_TABLE[24]={
+   0,
+   8,13,
+  16,19,21,23,
+  24,26,27,28,29,30,31,32,
+  32,33,34,34,35,36,36,37,37
+};
+
+#if defined(CUSTOM_MODES)
+
+/*Determines if V(N,K) fits in a 32-bit unsigned integer.
+  N and K are themselves limited to 15 bits.*/
+static int fits_in32(int _n, int _k)
+{
+   static const opus_int16 maxN[15] = {
+      32767, 32767, 32767, 1476, 283, 109,  60,  40,
+       29,  24,  20,  18,  16,  14,  13};
+   static const opus_int16 maxK[15] = {
+      32767, 32767, 32767, 32767, 1172, 238,  95,  53,
+       36,  27,  22,  18,  16,  15,  13};
+   if (_n>=14)
+   {
+      if (_k>=14)
+         return 0;
+      else
+         return _n <= maxN[_k];
+   } else {
+      return _k <= maxK[_n];
+   }
+}
+
+void compute_pulse_cache(CELTMode *m, int LM)
+{
+   int C;
+   int i;
+   int j;
+   int curr=0;
+   int nbEntries=0;
+   int entryN[100], entryK[100], entryI[100];
+   const opus_int16 *eBands = m->eBands;
+   PulseCache *cache = &m->cache;
+   opus_int16 *cindex;
+   unsigned char *bits;
+   unsigned char *cap;
+
+   cindex = (opus_int16 *)opus_alloc(sizeof(cache->index[0])*m->nbEBands*(LM+2));
+   cache->index = cindex;
+
+   /* Scan for all unique band sizes */
+   for (i=0;i<=LM+1;i++)
+   {
+      for (j=0;j<m->nbEBands;j++)
+      {
+         int k;
+         int N = (eBands[j+1]-eBands[j])<<i>>1;
+         cindex[i*m->nbEBands+j] = -1;
+         /* Find other bands that have the same size */
+         for (k=0;k<=i;k++)
+         {
+            int n;
+            for (n=0;n<m->nbEBands && (k!=i || n<j);n++)
+            {
+               if (N == (eBands[n+1]-eBands[n])<<k>>1)
+               {
+                  cindex[i*m->nbEBands+j] = cindex[k*m->nbEBands+n];
+                  break;
+               }
+            }
+         }
+         if (cache->index[i*m->nbEBands+j] == -1 && N!=0)
+         {
+            int K;
+            entryN[nbEntries] = N;
+            K = 0;
+            while (fits_in32(N,get_pulses(K+1)) && K<MAX_PSEUDO)
+               K++;
+            entryK[nbEntries] = K;
+            cindex[i*m->nbEBands+j] = curr;
+            entryI[nbEntries] = curr;
+
+            curr += K+1;
+            nbEntries++;
+         }
+      }
+   }
+   bits = (unsigned char *)opus_alloc(sizeof(unsigned char)*curr);
+   cache->bits = bits;
+   cache->size = curr;
+   /* Compute the cache for all unique sizes */
+   for (i=0;i<nbEntries;i++)
+   {
+      unsigned char *ptr = bits+entryI[i];
+      opus_int16 tmp[CELT_MAX_PULSES+1];
+      get_required_bits(tmp, entryN[i], get_pulses(entryK[i]), BITRES);
+      for (j=1;j<=entryK[i];j++)
+         ptr[j] = tmp[get_pulses(j)]-1;
+      ptr[0] = entryK[i];
+   }
+
+   /* Compute the maximum rate for each band at which we'll reliably use as
+       many bits as we ask for. */
+   cache->caps = cap = (unsigned char *)opus_alloc(sizeof(cache->caps[0])*(LM+1)*2*m->nbEBands);
+   for (i=0;i<=LM;i++)
+   {
+      for (C=1;C<=2;C++)
+      {
+         for (j=0;j<m->nbEBands;j++)
+         {
+            int N0;
+            int max_bits;
+            N0 = m->eBands[j+1]-m->eBands[j];
+            /* N=1 bands only have a sign bit and fine bits. */
+            if (N0<<i == 1)
+               max_bits = C*(1+MAX_FINE_BITS)<<BITRES;
+            else
+            {
+               const unsigned char *pcache;
+               opus_int32           num;
+               opus_int32           den;
+               int                  LM0;
+               int                  N;
+               int                  offset;
+               int                  ndof;
+               int                  qb;
+               int                  k;
+               LM0 = 0;
+               /* Even-sized bands bigger than N=2 can be split one more time.
+                  As of commit 44203907 all bands >1 are even, including custom modes.*/
+               if (N0 > 2)
+               {
+                  N0>>=1;
+                  LM0--;
+               }
+               /* N0=1 bands can't be split down to N<2. */
+               else if (N0 <= 1)
+               {
+                  LM0=IMIN(i,1);
+                  N0<<=LM0;
+               }
+               /* Compute the cost for the lowest-level PVQ of a fully split
+                   band. */
+               pcache = bits + cindex[(LM0+1)*m->nbEBands+j];
+               max_bits = pcache[pcache[0]]+1;
+               /* Add in the cost of coding regular splits. */
+               N = N0;
+               for(k=0;k<i-LM0;k++){
+                  max_bits <<= 1;
+                  /* Offset the number of qtheta bits by log2(N)/2
+                      + QTHETA_OFFSET compared to their "fair share" of
+                      total/N */
+                  offset = ((m->logN[j]+(opus_int32)((opus_uint32)(LM0+k)<<BITRES))>>1)-QTHETA_OFFSET;
+                  /* The number of qtheta bits we'll allocate if the remainder
+                      is to be max_bits.
+                     The average measured cost for theta is 0.89701 times qb,
+                      approximated here as 459/512. */
+                  num=459*(opus_int32)((2*N-1)*offset+max_bits);
+                  den=((opus_int32)(2*N-1)<<9)-459;
+                  qb = IMIN((num+(den>>1))/den, 57);
+                  celt_assert(qb >= 0);
+                  max_bits += qb;
+                  N <<= 1;
+               }
+               /* Add in the cost of a stereo split, if necessary. */
+               if (C==2)
+               {
+                  max_bits <<= 1;
+                  offset = ((m->logN[j]+(i<<BITRES))>>1)-(N==2?QTHETA_OFFSET_TWOPHASE:QTHETA_OFFSET);
+                  ndof = 2*N-1-(N==2);
+                  /* The average measured cost for theta with the step PDF is
+                      0.95164 times qb, approximated here as 487/512. */
+                  num = (N==2?512:487)*(opus_int32)(max_bits+ndof*offset);
+                  den = ((opus_int32)ndof<<9)-(N==2?512:487);
+                  qb = IMIN((num+(den>>1))/den, (N==2?64:61));
+                  celt_assert(qb >= 0);
+                  max_bits += qb;
+               }
+               /* Add the fine bits we'll use. */
+               /* Compensate for the extra DoF in stereo */
+               ndof = C*N + ((C==2 && N>2) ? 1 : 0);
+               /* Offset the number of fine bits by log2(N)/2 + FINE_OFFSET
+                   compared to their "fair share" of total/N */
+               offset = ((m->logN[j] + (i<<BITRES))>>1)-FINE_OFFSET;
+               /* N=2 is the only point that doesn't match the curve */
+               if (N==2)
+                  offset += 1<<BITRES>>2;
+               /* The number of fine bits we'll allocate if the remainder is
+                   to be max_bits. */
+               num = max_bits+ndof*offset;
+               den = (ndof-1)<<BITRES;
+               qb = IMIN((num+(den>>1))/den, MAX_FINE_BITS);
+               celt_assert(qb >= 0);
+               max_bits += C*qb<<BITRES;
+            }
+            max_bits = (4*max_bits/(C*((m->eBands[j+1]-m->eBands[j])<<i)))-64;
+            celt_assert(max_bits >= 0);
+            celt_assert(max_bits < 256);
+            *cap++ = (unsigned char)max_bits;
+         }
+      }
+   }
+}
+
+#endif /* CUSTOM_MODES */
+
+#define ALLOC_STEPS 6
+
+static OPUS_INLINE int interp_bits2pulses(const CELTMode *m, int start, int end, int skip_start,
+      const int *bits1, const int *bits2, const int *thresh, const int *cap, opus_int32 total, opus_int32 *_balance,
+      int skip_rsv, int *intensity, int intensity_rsv, int *dual_stereo, int dual_stereo_rsv, int *bits,
+      int *ebits, int *fine_priority, int C, int LM, ec_ctx *ec, int encode, int prev, int signalBandwidth)
+{
+   opus_int32 psum;
+   int lo, hi;
+   int i, j;
+   int logM;
+   int stereo;
+   int codedBands=-1;
+   int alloc_floor;
+   opus_int32 left, percoeff;
+   int done;
+   opus_int32 balance;
+   SAVE_STACK;
+
+   alloc_floor = C<<BITRES;
+   stereo = C>1;
+
+   logM = LM<<BITRES;
+   lo = 0;
+   hi = 1<<ALLOC_STEPS;
+   for (i=0;i<ALLOC_STEPS;i++)
+   {
+      int mid = (lo+hi)>>1;
+      psum = 0;
+      done = 0;
+      for (j=end;j-->start;)
+      {
+         int tmp = bits1[j] + (mid*(opus_int32)bits2[j]>>ALLOC_STEPS);
+         if (tmp >= thresh[j] || done)
+         {
+            done = 1;
+            /* Don't allocate more than we can actually use */
+            psum += IMIN(tmp, cap[j]);
+         } else {
+            if (tmp >= alloc_floor)
+               psum += alloc_floor;
+         }
+      }
+      if (psum > total)
+         hi = mid;
+      else
+         lo = mid;
+   }
+   psum = 0;
+   /*printf ("interp bisection gave %d\n", lo);*/
+   done = 0;
+   for (j=end;j-->start;)
+   {
+      int tmp = bits1[j] + ((opus_int32)lo*bits2[j]>>ALLOC_STEPS);
+      if (tmp < thresh[j] && !done)
+      {
+         if (tmp >= alloc_floor)
+            tmp = alloc_floor;
+         else
+            tmp = 0;
+      } else
+         done = 1;
+      /* Don't allocate more than we can actually use */
+      tmp = IMIN(tmp, cap[j]);
+      bits[j] = tmp;
+      psum += tmp;
+   }
+
+   /* Decide which bands to skip, working backwards from the end. */
+   for (codedBands=end;;codedBands--)
+   {
+      int band_width;
+      int band_bits;
+      int rem;
+      j = codedBands-1;
+      /* Never skip the first band, nor a band that has been boosted by
+          dynalloc.
+         In the first case, we'd be coding a bit to signal we're going to waste
+          all the other bits.
+         In the second case, we'd be coding a bit to redistribute all the bits
+          we just signaled should be concentrated in this band. */
+      if (j<=skip_start)
+      {
+         /* Give the bit we reserved to end skipping back. */
+         total += skip_rsv;
+         break;
+      }
+      /*Figure out how many left-over bits we would be adding to this band.
+        This can include bits we've stolen back from higher, skipped bands.*/
+      left = total-psum;
+      percoeff = celt_udiv(left, m->eBands[codedBands]-m->eBands[start]);
+      left -= (m->eBands[codedBands]-m->eBands[start])*percoeff;
+      rem = IMAX(left-(m->eBands[j]-m->eBands[start]),0);
+      band_width = m->eBands[codedBands]-m->eBands[j];
+      band_bits = (int)(bits[j] + percoeff*band_width + rem);
+      /*Only code a skip decision if we're above the threshold for this band.
+        Otherwise it is force-skipped.
+        This ensures that we have enough bits to code the skip flag.*/
+      if (band_bits >= IMAX(thresh[j], alloc_floor+(1<<BITRES)))
+      {
+         if (encode)
+         {
+            /*This if() block is the only part of the allocation function that
+               is not a mandatory part of the bitstream: any bands we choose to
+               skip here must be explicitly signaled.*/
+            int depth_threshold;
+            /*We choose a threshold with some hysteresis to keep bands from
+               fluctuating in and out, but we try not to fold below a certain point. */
+            if (codedBands > 17)
+               depth_threshold = j<prev ? 7 : 9;
+            else
+               depth_threshold = 0;
+#ifdef FUZZING
+            (void)signalBandwidth;
+            (void)depth_threshold;
+            if ((rand()&0x1) == 0)
+#else
+            if (codedBands<=start+2 || (band_bits > (depth_threshold*band_width<<LM<<BITRES)>>4 && j<=signalBandwidth))
+#endif
+            {
+               ec_enc_bit_logp(ec, 1, 1);
+               break;
+            }
+            ec_enc_bit_logp(ec, 0, 1);
+         } else if (ec_dec_bit_logp(ec, 1)) {
+            break;
+         }
+         /*We used a bit to skip this band.*/
+         psum += 1<<BITRES;
+         band_bits -= 1<<BITRES;
+      }
+      /*Reclaim the bits originally allocated to this band.*/
+      psum -= bits[j]+intensity_rsv;
+      if (intensity_rsv > 0)
+         intensity_rsv = LOG2_FRAC_TABLE[j-start];
+      psum += intensity_rsv;
+      if (band_bits >= alloc_floor)
+      {
+         /*If we have enough for a fine energy bit per channel, use it.*/
+         psum += alloc_floor;
+         bits[j] = alloc_floor;
+      } else {
+         /*Otherwise this band gets nothing at all.*/
+         bits[j] = 0;
+      }
+   }
+
+   celt_assert(codedBands > start);
+   /* Code the intensity and dual stereo parameters. */
+   if (intensity_rsv > 0)
+   {
+      if (encode)
+      {
+         *intensity = IMIN(*intensity, codedBands);
+         ec_enc_uint(ec, *intensity-start, codedBands+1-start);
+      }
+      else
+         *intensity = start+ec_dec_uint(ec, codedBands+1-start);
+   }
+   else
+      *intensity = 0;
+   if (*intensity <= start)
+   {
+      total += dual_stereo_rsv;
+      dual_stereo_rsv = 0;
+   }
+   if (dual_stereo_rsv > 0)
+   {
+      if (encode)
+         ec_enc_bit_logp(ec, *dual_stereo, 1);
+      else
+         *dual_stereo = ec_dec_bit_logp(ec, 1);
+   }
+   else
+      *dual_stereo = 0;
+
+   /* Allocate the remaining bits */
+   left = total-psum;
+   percoeff = celt_udiv(left, m->eBands[codedBands]-m->eBands[start]);
+   left -= (m->eBands[codedBands]-m->eBands[start])*percoeff;
+   for (j=start;j<codedBands;j++)
+      bits[j] += ((int)percoeff*(m->eBands[j+1]-m->eBands[j]));
+   for (j=start;j<codedBands;j++)
+   {
+      int tmp = (int)IMIN(left, m->eBands[j+1]-m->eBands[j]);
+      bits[j] += tmp;
+      left -= tmp;
+   }
+   /*for (j=0;j<end;j++)printf("%d ", bits[j]);printf("\n");*/
+
+   balance = 0;
+   for (j=start;j<codedBands;j++)
+   {
+      int N0, N, den;
+      int offset;
+      int NClogN;
+      opus_int32 excess, bit;
+
+      celt_assert(bits[j] >= 0);
+      N0 = m->eBands[j+1]-m->eBands[j];
+      N=N0<<LM;
+      bit = (opus_int32)bits[j]+balance;
+
+      if (N>1)
+      {
+         excess = MAX32(bit-cap[j],0);
+         bits[j] = bit-excess;
+
+         /* Compensate for the extra DoF in stereo */
+         den=(C*N+ ((C==2 && N>2 && !*dual_stereo && j<*intensity) ? 1 : 0));
+
+         NClogN = den*(m->logN[j] + logM);
+
+         /* Offset for the number of fine bits by log2(N)/2 + FINE_OFFSET
+            compared to their "fair share" of total/N */
+         offset = (NClogN>>1)-den*FINE_OFFSET;
+
+         /* N=2 is the only point that doesn't match the curve */
+         if (N==2)
+            offset += den<<BITRES>>2;
+
+         /* Changing the offset for allocating the second and third
+             fine energy bit */
+         if (bits[j] + offset < den*2<<BITRES)
+            offset += NClogN>>2;
+         else if (bits[j] + offset < den*3<<BITRES)
+            offset += NClogN>>3;
+
+         /* Divide with rounding */
+         ebits[j] = IMAX(0, (bits[j] + offset + (den<<(BITRES-1))));
+         ebits[j] = celt_udiv(ebits[j], den)>>BITRES;
+
+         /* Make sure not to bust */
+         if (C*ebits[j] > (bits[j]>>BITRES))
+            ebits[j] = bits[j] >> stereo >> BITRES;
+
+         /* More than that is useless because that's about as far as PVQ can go */
+         ebits[j] = IMIN(ebits[j], MAX_FINE_BITS);
+
+         /* If we rounded down or capped this band, make it a candidate for the
+             final fine energy pass */
+         fine_priority[j] = ebits[j]*(den<<BITRES) >= bits[j]+offset;
+
+         /* Remove the allocated fine bits; the rest are assigned to PVQ */
+         bits[j] -= C*ebits[j]<<BITRES;
+
+      } else {
+         /* For N=1, all bits go to fine energy except for a single sign bit */
+         excess = MAX32(0,bit-(C<<BITRES));
+         bits[j] = bit-excess;
+         ebits[j] = 0;
+         fine_priority[j] = 1;
+      }
+
+      /* Fine energy can't take advantage of the re-balancing in
+          quant_all_bands().
+         Instead, do the re-balancing here.*/
+      if(excess > 0)
+      {
+         int extra_fine;
+         int extra_bits;
+         extra_fine = IMIN(excess>>(stereo+BITRES),MAX_FINE_BITS-ebits[j]);
+         ebits[j] += extra_fine;
+         extra_bits = extra_fine*C<<BITRES;
+         fine_priority[j] = extra_bits >= excess-balance;
+         excess -= extra_bits;
+      }
+      balance = excess;
+
+      celt_assert(bits[j] >= 0);
+      celt_assert(ebits[j] >= 0);
+   }
+   /* Save any remaining bits over the cap for the rebalancing in
+       quant_all_bands(). */
+   *_balance = balance;
+
+   /* The skipped bands use all their bits for fine energy. */
+   for (;j<end;j++)
+   {
+      ebits[j] = bits[j] >> stereo >> BITRES;
+      celt_assert(C*ebits[j]<<BITRES == bits[j]);
+      bits[j] = 0;
+      fine_priority[j] = ebits[j]<1;
+   }
+   RESTORE_STACK;
+   return codedBands;
+}
+
+int clt_compute_allocation(const CELTMode *m, int start, int end, const int *offsets, const int *cap, int alloc_trim, int *intensity, int *dual_stereo,
+      opus_int32 total, opus_int32 *balance, int *pulses, int *ebits, int *fine_priority, int C, int LM, ec_ctx *ec, int encode, int prev, int signalBandwidth)
+{
+   int lo, hi, len, j;
+   int codedBands;
+   int skip_start;
+   int skip_rsv;
+   int intensity_rsv;
+   int dual_stereo_rsv;
+   VARDECL(int, bits1);
+   VARDECL(int, bits2);
+   VARDECL(int, thresh);
+   VARDECL(int, trim_offset);
+   SAVE_STACK;
+
+   total = IMAX(total, 0);
+   len = m->nbEBands;
+   skip_start = start;
+   /* Reserve a bit to signal the end of manually skipped bands. */
+   skip_rsv = total >= 1<<BITRES ? 1<<BITRES : 0;
+   total -= skip_rsv;
+   /* Reserve bits for the intensity and dual stereo parameters. */
+   intensity_rsv = dual_stereo_rsv = 0;
+   if (C==2)
+   {
+      intensity_rsv = LOG2_FRAC_TABLE[end-start];
+      if (intensity_rsv>total)
+         intensity_rsv = 0;
+      else
+      {
+         total -= intensity_rsv;
+         dual_stereo_rsv = total>=1<<BITRES ? 1<<BITRES : 0;
+         total -= dual_stereo_rsv;
+      }
+   }
+   ALLOC(bits1, len, int);
+   ALLOC(bits2, len, int);
+   ALLOC(thresh, len, int);
+   ALLOC(trim_offset, len, int);
+
+   for (j=start;j<end;j++)
+   {
+      /* Below this threshold, we're sure not to allocate any PVQ bits */
+      thresh[j] = IMAX((C)<<BITRES, (3*(m->eBands[j+1]-m->eBands[j])<<LM<<BITRES)>>4);
+      /* Tilt of the allocation curve */
+      trim_offset[j] = C*(m->eBands[j+1]-m->eBands[j])*(alloc_trim-5-LM)*(end-j-1)
+            *(1<<(LM+BITRES))>>6;
+      /* Giving less resolution to single-coefficient bands because they get
+         more benefit from having one coarse value per coefficient*/
+      if ((m->eBands[j+1]-m->eBands[j])<<LM==1)
+         trim_offset[j] -= C<<BITRES;
+   }
+   lo = 1;
+   hi = m->nbAllocVectors - 1;
+   do
+   {
+      int done = 0;
+      int psum = 0;
+      int mid = (lo+hi) >> 1;
+      for (j=end;j-->start;)
+      {
+         int bitsj;
+         int N = m->eBands[j+1]-m->eBands[j];
+         bitsj = C*N*m->allocVectors[mid*len+j]<<LM>>2;
+         if (bitsj > 0)
+            bitsj = IMAX(0, bitsj + trim_offset[j]);
+         bitsj += offsets[j];
+         if (bitsj >= thresh[j] || done)
+         {
+            done = 1;
+            /* Don't allocate more than we can actually use */
+            psum += IMIN(bitsj, cap[j]);
+         } else {
+            if (bitsj >= C<<BITRES)
+               psum += C<<BITRES;
+         }
+      }
+      if (psum > total)
+         hi = mid - 1;
+      else
+         lo = mid + 1;
+      /*printf ("lo = %d, hi = %d\n", lo, hi);*/
+   }
+   while (lo <= hi);
+   hi = lo--;
+   /*printf ("interp between %d and %d\n", lo, hi);*/
+   for (j=start;j<end;j++)
+   {
+      int bits1j, bits2j;
+      int N = m->eBands[j+1]-m->eBands[j];
+      bits1j = C*N*m->allocVectors[lo*len+j]<<LM>>2;
+      bits2j = hi>=m->nbAllocVectors ?
+            cap[j] : C*N*m->allocVectors[hi*len+j]<<LM>>2;
+      if (bits1j > 0)
+         bits1j = IMAX(0, bits1j + trim_offset[j]);
+      if (bits2j > 0)
+         bits2j = IMAX(0, bits2j + trim_offset[j]);
+      if (lo > 0)
+         bits1j += offsets[j];
+      bits2j += offsets[j];
+      if (offsets[j]>0)
+         skip_start = j;
+      bits2j = IMAX(0,bits2j-bits1j);
+      bits1[j] = bits1j;
+      bits2[j] = bits2j;
+   }
+   codedBands = interp_bits2pulses(m, start, end, skip_start, bits1, bits2, thresh, cap,
+         total, balance, skip_rsv, intensity, intensity_rsv, dual_stereo, dual_stereo_rsv,
+         pulses, ebits, fine_priority, C, LM, ec, encode, prev, signalBandwidth);
+   RESTORE_STACK;
+   return codedBands;
+}
+#ifdef ENABLE_QEXT
+
+static const unsigned char last_zero[3] = {64, 50, 0};
+static const unsigned char last_cap[3] = {110, 60, 0};
+static const unsigned char last_other[4] = {120, 112, 70, 0};
+
+static void ec_enc_depth(ec_enc *enc, opus_int32 depth, opus_int32 cap, opus_int32 *last) {
+   int sym = 3;
+   if (depth==*last) sym = 2;
+   if (depth==cap) sym = 1;
+   if (depth==0) sym = 0;
+   if (*last == 0) {
+      ec_enc_icdf(enc, IMIN(sym, 2), last_zero, 7);
+   } else if (*last == cap) {
+      ec_enc_icdf(enc, IMIN(sym, 2), last_cap, 7);
+   } else {
+      ec_enc_icdf(enc, sym, last_other, 7);
+   }
+   /* We accept some redundancy if depth==last (for last different from 0 and cap). */
+   if (sym == 3) ec_enc_uint(enc, depth-1, cap);
+   *last = depth;
+}
+
+static int ec_dec_depth(ec_dec *dec, opus_int32 cap, opus_int32 *last) {
+   int depth, sym;
+   if (*last == 0) {
+      sym = ec_dec_icdf(dec, last_zero, 7);
+      if (sym==2) sym=3;
+   } else if (*last == cap) {
+      sym = ec_dec_icdf(dec, last_cap, 7);
+      if (sym==2) sym=3;
+   } else {
+      sym = ec_dec_icdf(dec, last_other, 7);
+   }
+   if (sym==0) depth=0;
+   else if (sym==1) depth=cap;
+   else if (sym==2) depth=*last;
+   else depth = 1 + ec_dec_uint(dec, cap);
+   *last = depth;
+   return depth;
+}
+
+#define MSWAP16(a,b) do {opus_val16 tmp = a;a=b;b=tmp;} while(0)
+static opus_val16 median_of_5_val16(const opus_val16 *x)
+{
+   opus_val16 t0, t1, t2, t3, t4;
+   t2 = x[2];
+   if (x[0] > x[1])
+   {
+      t0 = x[1];
+      t1 = x[0];
+   } else {
+      t0 = x[0];
+      t1 = x[1];
+   }
+   if (x[3] > x[4])
+   {
+      t3 = x[4];
+      t4 = x[3];
+   } else {
+      t3 = x[3];
+      t4 = x[4];
+   }
+   if (t0 > t3)
+   {
+      MSWAP16(t0, t3);
+      MSWAP16(t1, t4);
+   }
+   if (t2 > t1)
+   {
+      if (t1 < t3)
+         return MIN16(t2, t3);
+      else
+         return MIN16(t4, t1);
+   } else {
+      if (t2 < t3)
+         return MIN16(t1, t3);
+      else
+         return MIN16(t2, t4);
+   }
+}
+
+void clt_compute_extra_allocation(const CELTMode *m, const CELTMode *qext_mode, int start, int end, int qext_end, const celt_glog *bandLogE, const celt_glog *qext_bandLogE,
+      opus_int32 total, int *extra_pulses, int *extra_equant, int C, int LM, ec_ctx *ec, int encode, opus_val16 tone_freq, opus_val32 toneishness)
+{
+   int i;
+   opus_int32 last=0;
+   opus_val32 sum;
+   opus_val32 fill;
+   int iter;
+   int tot_bands;
+   int tot_samples;
+   VARDECL(int, depth);
+   VARDECL(opus_int32, cap);
+#ifdef FUZZING
+   float depth_std;
+#endif
+   SAVE_STACK;
+#ifdef FUZZING
+   depth_std = -10.f*log(1e-8+(float)rand()/(float)RAND_MAX);
+   depth_std = FMAX(0, FMIN(48, depth_std));
+#endif
+   if (qext_mode != NULL) {
+      celt_assert(end==m->nbEBands);
+      tot_bands = end + qext_end;
+      tot_samples = qext_mode->eBands[qext_end]*C<<LM;
+   } else {
+      tot_bands = end;
+      tot_samples = (m->eBands[end]-m->eBands[start])*C<<LM;
+   }
+   ALLOC(cap, tot_bands, opus_int32);
+   for (i=start;i<end;i++) cap[i] = 12;
+   if (qext_mode != NULL) {
+      for (i=0;i<qext_end;i++) cap[end+i] = 14;
+   }
+   if (total <= 0) {
+      for (i=start;i<m->nbEBands+qext_end;i++) {
+         extra_pulses[i] = extra_equant[i] = 0;
+      }
+      RESTORE_STACK;
+      return;
+   }
+   ALLOC(depth, tot_bands, int);
+   if (encode) {
+      VARDECL(opus_val16, flatE);
+      VARDECL(int, Ncoef);
+      VARDECL(opus_val16, min);
+      VARDECL(opus_val16, follower);
+
+      ALLOC(flatE, tot_bands, opus_val16);
+      ALLOC(min, tot_bands, opus_val16);
+      ALLOC(Ncoef, tot_bands, int);
+      for (i=start;i<end;i++) {
+         Ncoef[i] = (m->eBands[i+1]-m->eBands[i])*C<<LM;
+      }
+      /* Remove the effect of band width, eMeans and pre-emphasis to compute the real (flat) spectrum. */
+      for (i=start;i<end;i++) {
+         flatE[i] = PSHR32(bandLogE[i] - GCONST(0.0625f)*m->logN[i] + SHL32(eMeans[i],DB_SHIFT-4) - GCONST(.0062f)*(i+5)*(i+5), DB_SHIFT-10);
+         min[i] = 0;
+      }
+      if (C==2) {
+         for (i=start;i<end;i++) {
+            flatE[i] = MAXG(flatE[i], PSHR32(bandLogE[m->nbEBands+i] - GCONST(0.0625f)*m->logN[i] + SHL32(eMeans[i],DB_SHIFT-4) - GCONST(.0062f)*(i+5)*(i+5), DB_SHIFT-10));
+         }
+      }
+      flatE[end-1] += QCONST16(2.f, 10);
+      if (qext_mode != NULL) {
+         opus_val16 min_depth = 0;
+         /* If we have enough bits, give at least 1 bit of depth to all higher bands. */
+         if (total >= 3*C*(qext_mode->eBands[qext_end]-qext_mode->eBands[start])<<LM<<BITRES && (toneishness < QCONST32(.98f, 29) || tone_freq > 1.33f))
+            min_depth = QCONST16(1.f, 10);
+         for (i=0;i<qext_end;i++) {
+            Ncoef[end+i] = (qext_mode->eBands[i+1]-qext_mode->eBands[i])*C<<LM;
+            min[end+i] = min_depth;
+         }
+         for (i=0;i<qext_end;i++) {
+            flatE[end+i] = PSHR32(qext_bandLogE[i] - GCONST(0.0625f)*qext_mode->logN[i] + SHL32(eMeans[i],DB_SHIFT-4) - GCONST(.0062f)*(end+i+5)*(end+i+5), DB_SHIFT-10);
+         }
+         if (C==2) {
+            for (i=0;i<qext_end;i++) {
+               flatE[end+i] = MAXG(flatE[end+i], PSHR32(qext_bandLogE[NB_QEXT_BANDS+i] - GCONST(0.0625f)*qext_mode->logN[i] + SHL32(eMeans[i],DB_SHIFT-4) - GCONST(.0062f)*(end+i+5)*(end+i+5), DB_SHIFT-10));
+            }
+         }
+      }
+      ALLOC(follower, tot_bands, opus_val16);
+      for (i=start+2;i<tot_bands-2;i++) {
+         follower[i] = median_of_5_val16(&flatE[i-2]);
+      }
+      follower[start] = follower[start+1] = follower[start+2];
+      follower[tot_bands-1] = follower[tot_bands-2] = follower[tot_bands-3];
+      for (i=start+1;i<tot_bands;i++) {
+         follower[i] = MAX16(follower[i], follower[i-1]-QCONST16(1.f, 10));
+      }
+      for (i=tot_bands-2;i>=start;i--) {
+         follower[i] = MAX16(follower[i], follower[i+1]-QCONST16(1.f, 10));
+      }
+      for (i=start;i<tot_bands;i++) flatE[i] -= MULT16_16_Q15(Q15ONE-PSHR32(toneishness, 14), follower[i]);
+      if (qext_mode != NULL) {
+         for (i=0;i<qext_end;i++) flatE[end+i] = flatE[end+i] + QCONST16(3.f, 10) + QCONST16(.2f, 10)*i;
+      }
+      /* Approximate fill level assuming all bands contribute fully. */
+      sum = 0;
+      for (i=start;i<tot_bands;i++) {
+         sum += MULT16_16(Ncoef[i], flatE[i]);
+      }
+      total >>= BITRES;
+      fill = (SHL32(total, 10) + sum)/tot_samples;
+      /* Iteratively refine the fill level considering the depth min and cap. */
+      for (iter=0;iter<10;iter++) {
+         sum = 0;
+         for (i=start;i<tot_bands;i++)
+            sum += Ncoef[i] * MIN32(SHL32(cap[i], 10), MAX32(min[i], flatE[i]-fill));
+         fill -= (SHL32(total, 10) - sum)/tot_samples;
+      }
+      for (i=start;i<tot_bands;i++) {
+#ifdef FIXED_POINT
+         depth[i] = PSHR32(MIN32(SHL32(cap[i], 10), MAX32(min[i], flatE[i]-fill)), 10-2);
+#else
+         depth[i] = (int)floor(.5+4*MIN32(SHL32(cap[i], 10), MAX32(min[i], flatE[i]-fill)));
+#endif
+#ifdef FUZZING
+         depth[i] = (int)-depth_std*log(1e-8+(float)rand()/(float)RAND_MAX);
+         depth[i] = IMAX(0, IMIN(cap[i]<<2, depth[i]));
+#endif
+         if (ec_tell_frac(ec) + 80 < ec->storage*8<<BITRES)
+            ec_enc_depth(ec, depth[i], 4*cap[i], &last);
+         else
+            depth[i] = 0;
+      }
+   } else {
+      for (i=start;i<tot_bands;i++) {
+         if (ec_tell_frac(ec) + 80 < ec->storage*8<<BITRES)
+            depth[i] = ec_dec_depth(ec, 4*cap[i], &last);
+         else
+            depth[i] = 0;
+      }
+   }
+   for (i=start;i<end;i++) {
+      extra_equant[i] = (depth[i]+3)>>2;
+      extra_pulses[i] = ((((m->eBands[i+1]-m->eBands[i])<<LM)-1)*C * depth[i] * (1<<BITRES) + 2)>>2;
+   }
+   if (qext_mode) {
+      for (i=0;i<qext_end;i++) {
+         extra_equant[end+i] = (depth[end+i]+3)>>2;
+         extra_pulses[end+i] = ((((qext_mode->eBands[i+1]-qext_mode->eBands[i])<<LM)-1)*C * depth[end+i] * (1<<BITRES) + 2)>>2;
+      }
+   }
+   RESTORE_STACK;
+}
+#endif
--- a/lib/libopus/celt/rate.h
+++ b/lib/libopus/celt/rate.h
@ -0,0 +1,104 @@
+/* Copyright (c) 2007-2008 CSIRO
+   Copyright (c) 2007-2009 Xiph.Org Foundation
+   Written by Jean-Marc Valin */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef RATE_H
+#define RATE_H
+
+#define MAX_PSEUDO 40
+#define LOG_MAX_PSEUDO 6
+
+#define CELT_MAX_PULSES 128
+
+#define MAX_FINE_BITS 8
+
+#define FINE_OFFSET 21
+#define QTHETA_OFFSET 4
+#define QTHETA_OFFSET_TWOPHASE 16
+
+#include "cwrs.h"
+#include "modes.h"
+
+void compute_pulse_cache(CELTMode *m, int LM);
+
+static OPUS_INLINE int get_pulses(int i)
+{
+   return i<8 ? i : (8 + (i&7)) << ((i>>3)-1);
+}
+
+static OPUS_INLINE int bits2pulses(const CELTMode *m, int band, int LM, int bits)
+{
+   int i;
+   int lo, hi;
+   const unsigned char *cache;
+
+   LM++;
+   cache = m->cache.bits + m->cache.index[LM*m->nbEBands+band];
+
+   lo = 0;
+   hi = cache[0];
+   bits--;
+   for (i=0;i<LOG_MAX_PSEUDO;i++)
+   {
+      int mid = (lo+hi+1)>>1;
+      /* OPT: Make sure this is implemented with a conditional move */
+      if ((int)cache[mid] >= bits)
+         hi = mid;
+      else
+         lo = mid;
+   }
+   if (bits- (lo == 0 ? -1 : (int)cache[lo]) <= (int)cache[hi]-bits)
+      return lo;
+   else
+      return hi;
+}
+
+static OPUS_INLINE int pulses2bits(const CELTMode *m, int band, int LM, int pulses)
+{
+   const unsigned char *cache;
+
+   LM++;
+   cache = m->cache.bits + m->cache.index[LM*m->nbEBands+band];
+   return pulses == 0 ? 0 : cache[pulses]+1;
+}
+
+/** Compute the pulse allocation, i.e. how many pulses will go in each
+  * band.
+ @param m mode
+ @param offsets Requested increase or decrease in the number of bits for
+                each band
+ @param total Number of bands
+ @param pulses Number of pulses per band (returned)
+ @return Total number of bits allocated
+*/
+int clt_compute_allocation(const CELTMode *m, int start, int end, const int *offsets, const int *cap, int alloc_trim, int *intensity, int *dual_stereo,
+      opus_int32 total, opus_int32 *balance, int *pulses, int *ebits, int *fine_priority, int C, int LM, ec_ctx *ec, int encode, int prev, int signalBandwidth);
+
+void clt_compute_extra_allocation(const CELTMode *m, const CELTMode *qext_mode, int start, int end, int qext_end, const celt_glog *bandLogE, const celt_glog *qext_bandLogE,
+      opus_int32 total, int *extra_pulses, int *extra_equant, int C, int LM, ec_ctx *ec, int encode, opus_val16 tone_freq, opus_val32 toneishness);
+
+#endif
--- a/lib/libopus/celt/stack_alloc.h
+++ b/lib/libopus/celt/stack_alloc.h
@ -0,0 +1,209 @@
+/* Copyright (C) 2002-2003 Jean-Marc Valin
+   Copyright (C) 2007-2009 Xiph.Org Foundation */
+/**
+   @file stack_alloc.h
+   @brief Temporary memory allocation on stack
+*/
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef STACK_ALLOC_H
+#define STACK_ALLOC_H
+
+#include "opus_types.h"
+#include "opus_defines.h"
+
+#if (!defined (VAR_ARRAYS) && !defined (USE_ALLOCA) && !defined (NONTHREADSAFE_PSEUDOSTACK))
+#error "Opus requires one of VAR_ARRAYS, USE_ALLOCA, or NONTHREADSAFE_PSEUDOSTACK be defined to select the temporary allocation mode."
+#endif
+
+#ifdef USE_ALLOCA
+# ifdef _WIN32
+#  include <malloc.h>
+# else
+#  ifdef HAVE_ALLOCA_H
+#   include <alloca.h>
+#  else
+#   include <stdlib.h>
+#  endif
+# endif
+#endif
+
+/**
+ * @def ALIGN(stack, size)
+ *
+ * Aligns the stack to a 'size' boundary
+ *
+ * @param stack Stack
+ * @param size  New size boundary
+ */
+
+/**
+ * @def PUSH(stack, size, type)
+ *
+ * Allocates 'size' elements of type 'type' on the stack
+ *
+ * @param stack Stack
+ * @param size  Number of elements
+ * @param type  Type of element
+ */
+
+/**
+ * @def VARDECL(var)
+ *
+ * Declare variable on stack
+ *
+ * @param var Variable to declare
+ */
+
+/**
+ * @def ALLOC(var, size, type)
+ *
+ * Allocate 'size' elements of 'type' on stack
+ *
+ * @param var  Name of variable to allocate
+ * @param size Number of elements
+ * @param type Type of element
+ */
+
+#if defined(VAR_ARRAYS)
+
+#define VARDECL(type, var)
+#define ALLOC(var, size, type) type var[size]
+#define SAVE_STACK
+#define RESTORE_STACK
+#define ALLOC_STACK
+/* C99 does not allow VLAs of size zero */
+#define ALLOC_NONE 1
+
+#elif defined(USE_ALLOCA)
+
+#define VARDECL(type, var) type *var
+
+# ifdef _WIN32
+#  define ALLOC(var, size, type) var = ((type*)_alloca(sizeof(type)*(size)))
+# else
+#  define ALLOC(var, size, type) var = ((type*)alloca(sizeof(type)*(size)))
+# endif
+
+#define SAVE_STACK
+#define RESTORE_STACK
+#define ALLOC_STACK
+#define ALLOC_NONE 0
+
+#else
+
+#ifdef CELT_C
+char *scratch_ptr=0;
+char *global_stack=0;
+#else
+extern char *global_stack;
+extern char *scratch_ptr;
+#endif /* CELT_C */
+
+#if __STDC_VERSION__ >= 201112L
+#  include <stdalign.h>
+#  define ALIGNOF(T) alignof(T)
+#elif defined(__GNUC__) || defined(__clang__)
+#  define ALIGNOF(T) __alignof__(T)
+#else
+# include <stddef.h>
+# ifdef __cplusplus
+template <typename T>
+struct alignment_helper {
+    char c;
+    T member;
+};
+# define ALIGNOF(T) (offsetof(alignment_helper<T>, member))
+# else
+# define ALIGNOF(T) (offsetof(struct { char c; T member; }, member))
+# endif
+#endif
+
+#ifdef ENABLE_VALGRIND
+
+#include <valgrind/memcheck.h>
+
+#ifdef CELT_C
+char *global_stack_top=0;
+#else
+extern char *global_stack_top;
+#endif /* CELT_C */
+
+#define ALIGN(stack, size) ((stack) += ((size) - (long)(stack)) & ((size) - 1))
+#define PUSH(stack, size, type) (VALGRIND_MAKE_MEM_NOACCESS(stack, global_stack_top-stack),ALIGN((stack),ALIGNOF(type)),VALGRIND_MAKE_MEM_UNDEFINED(stack, ((size)*sizeof(type)/sizeof(char))),(stack)+=(2*(size)*sizeof(type)/sizeof(char)),(type*)((stack)-(2*(size)*sizeof(type)/sizeof(char))))
+#define RESTORE_STACK ((global_stack = _saved_stack),VALGRIND_MAKE_MEM_NOACCESS(global_stack, global_stack_top-global_stack))
+#define ALLOC_STACK char *_saved_stack; ((global_stack = (global_stack==0) ? ((global_stack_top=(char*)opus_alloc_scratch(GLOBAL_STACK_SIZE*2)+(GLOBAL_STACK_SIZE*2))-(GLOBAL_STACK_SIZE*2)) : global_stack),VALGRIND_MAKE_MEM_NOACCESS(global_stack, global_stack_top-global_stack)); _saved_stack = global_stack;
+
+#else
+
+#define ALIGN(stack, size) ((stack) += ((size) - (long)(stack)) & ((size) - 1))
+#ifdef ENABLE_HARDENING
+#include "arch.h"
+#define PUSH(stack, size, type) (ALIGN((stack),ALIGNOF(type)),(void)(((int)((size)*(sizeof(type)/(sizeof(char)))) <= (scratch_ptr)+GLOBAL_STACK_SIZE-(stack))?0:CELT_FATAL("pseudostack overflow")),(stack)+=(size)*(sizeof(type)/(sizeof(char))),(type*)(void*)((stack)-(size)*(sizeof(type)/(sizeof(char)))))
+#else
+#define PUSH(stack, size, type) (ALIGN((stack),ALIGNOF(type)),(stack)+=(size)*(sizeof(type)/(sizeof(char))),(type*)(void*)((stack)-(size)*(sizeof(type)/(sizeof(char)))))
+#endif
+
+#if 0 /* Set this to 1 to instrument pseudostack usage */
+#define RESTORE_STACK (printf("%ld %s:%d\n", global_stack-scratch_ptr, __FILE__, __LINE__),global_stack = _saved_stack)
+#else
+#define RESTORE_STACK (global_stack = _saved_stack)
+#endif
+#define ALLOC_STACK char *_saved_stack; (global_stack = (global_stack==0) ? (scratch_ptr=(char*)opus_alloc_scratch(GLOBAL_STACK_SIZE)) : global_stack); _saved_stack = global_stack;
+
+#endif /* ENABLE_VALGRIND */
+
+#include "os_support.h"
+#define VARDECL(type, var) type *var
+#define ALLOC(var, size, type) var = PUSH(global_stack, size, type)
+#define SAVE_STACK char *_saved_stack = global_stack;
+#define ALLOC_NONE 0
+
+#endif /* VAR_ARRAYS */
+
+
+#ifdef ENABLE_VALGRIND
+
+#include <valgrind/memcheck.h>
+#define OPUS_CHECK_ARRAY(ptr, len) VALGRIND_CHECK_MEM_IS_DEFINED(ptr, len*sizeof(*ptr))
+#define OPUS_CHECK_VALUE(value) VALGRIND_CHECK_VALUE_IS_DEFINED(value)
+#define OPUS_CHECK_ARRAY_COND(ptr, len) VALGRIND_CHECK_MEM_IS_DEFINED(ptr, len*sizeof(*ptr))
+#define OPUS_CHECK_VALUE_COND(value) VALGRIND_CHECK_VALUE_IS_DEFINED(value)
+#define OPUS_PRINT_INT(value) do {fprintf(stderr, #value " = %d at %s:%d\n", value, __FILE__, __LINE__);}while(0)
+#define OPUS_FPRINTF fprintf
+
+#else
+
+static OPUS_INLINE int _opus_false(void) {return 0;}
+#define OPUS_CHECK_ARRAY(ptr, len) _opus_false()
+#define OPUS_CHECK_VALUE(value) _opus_false()
+#define OPUS_PRINT_INT(value) do{}while(0)
+#define OPUS_FPRINTF (void)
+
+#endif
+
+
+#endif /* STACK_ALLOC_H */
--- a/lib/libopus/celt/static_modes_fixed.h
+++ b/lib/libopus/celt/static_modes_fixed.h
--- a/lib/libopus/celt/static_modes_fixed_arm_ne10.h
+++ b/lib/libopus/celt/static_modes_fixed_arm_ne10.h
@ -0,0 +1,388 @@
+/* The contents of this file was automatically generated by
+ * dump_mode_arm_ne10.c with arguments: 48000 960
+ * It contains static definitions for some pre-defined modes. */
+#include <NE10_types.h>
+
+#ifndef NE10_FFT_PARAMS48000_960
+#define NE10_FFT_PARAMS48000_960
+static const ne10_int32_t ne10_factors_480[64] = {
+4, 40, 4, 30, 2, 15, 5, 3, 3, 1, 1, 0, 0, 0, 0,
+0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+0, 0, 0, 0, };
+static const ne10_int32_t ne10_factors_240[64] = {
+3, 20, 4, 15, 5, 3, 3, 1, 1, 0, 0, 0, 0, 0, 0,
+0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+0, 0, 0, 0, };
+static const ne10_int32_t ne10_factors_120[64] = {
+3, 10, 2, 15, 5, 3, 3, 1, 1, 0, 0, 0, 0, 0, 0,
+0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+0, 0, 0, 0, };
+static const ne10_int32_t ne10_factors_60[64] = {
+2, 5, 5, 3, 3, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0,
+0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+0, 0, 0, 0, };
+static const ne10_fft_cpx_int32_t ne10_twiddles_480[480] = {
+{0,0}, {2147483647,0}, {2147483647,0},
+{2147483647,0}, {1961823921,-873460313}, {1436946998,-1595891394},
+{2147483647,0}, {1436946998,-1595891394}, {-224473265,-2135719496},
+{2147483647,0}, {663608871,-2042378339}, {-1737350854,-1262259096},
+{2147483647,0}, {-224473265,-2135719496}, {-2100555935,446487152},
+{2147483647,0}, {2100555974,-446486968}, {1961823921,-873460313},
+{1737350743,-1262259248}, {1436946998,-1595891394}, {1073741769,-1859775424},
+{663608871,-2042378339}, {224473078,-2135719516}, {-224473265,-2135719496},
+{-663609049,-2042378281}, {-1073741932,-1859775330}, {-1436947137,-1595891268},
+{-1737350854,-1262259096}, {-1961823997,-873460141}, {-2100556013,-446486785},
+{2147483647,0}, {2144540595,-112390613}, {2135719506,-224473172},
+{2121044558,-335940465}, {2100555974,-446486968}, {2074309912,-555809682},
+{2042378310,-663608960}, {2004848691,-769589332}, {1961823921,-873460313},
+{1913421927,-974937199}, {1859775377,-1073741851}, {1801031311,-1169603450},
+{1737350743,-1262259248}, {1668908218,-1351455280}, {1595891331,-1436947067},
+{1518500216,-1518500282}, {1436946998,-1595891394}, {1351455207,-1668908277},
+{1262259172,-1737350799}, {1169603371,-1801031362}, {1073741769,-1859775424},
+{974937230,-1913421912}, {873460227,-1961823959}, {769589125,-2004848771},
+{663608871,-2042378339}, {555809715,-2074309903}, {446486876,-2100555994},
+{335940246,-2121044593}, {224473078,-2135719516}, {112390647,-2144540593},
+{2147483647,0}, {2135719506,-224473172}, {2100555974,-446486968},
+{2042378310,-663608960}, {1961823921,-873460313}, {1859775377,-1073741851},
+{1737350743,-1262259248}, {1595891331,-1436947067}, {1436946998,-1595891394},
+{1262259172,-1737350799}, {1073741769,-1859775424}, {873460227,-1961823959},
+{663608871,-2042378339}, {446486876,-2100555994}, {224473078,-2135719516},
+{-94,-2147483647}, {-224473265,-2135719496}, {-446487060,-2100555955},
+{-663609049,-2042378281}, {-873460398,-1961823883}, {-1073741932,-1859775330},
+{-1262259116,-1737350839}, {-1436947137,-1595891268}, {-1595891628,-1436946738},
+{-1737350854,-1262259096}, {-1859775343,-1073741910}, {-1961823997,-873460141},
+{-2042378447,-663608538}, {-2100556013,-446486785}, {-2135719499,-224473240},
+{2147483647,0}, {2121044558,-335940465}, {2042378310,-663608960},
+{1913421927,-974937199}, {1737350743,-1262259248}, {1518500216,-1518500282},
+{1262259172,-1737350799}, {974937230,-1913421912}, {663608871,-2042378339},
+{335940246,-2121044593}, {-94,-2147483647}, {-335940431,-2121044564},
+{-663609049,-2042378281}, {-974937397,-1913421827}, {-1262259116,-1737350839},
+{-1518500258,-1518500240}, {-1737350854,-1262259096}, {-1913422071,-974936918},
+{-2042378447,-663608538}, {-2121044568,-335940406}, {-2147483647,188},
+{-2121044509,335940777}, {-2042378331,663608895}, {-1913421900,974937252},
+{-1737350633,1262259400}, {-1518499993,1518500506}, {-1262258813,1737351059},
+{-974936606,1913422229}, {-663609179,2042378239}, {-335940566,2121044542},
+{2147483647,0}, {2147299667,-28109693}, {2146747758,-56214570},
+{2145828015,-84309815}, {2144540595,-112390613}, {2142885719,-140452154},
+{2140863671,-168489630}, {2138474797,-196498235}, {2135719506,-224473172},
+{2132598271,-252409646}, {2129111626,-280302871}, {2125260168,-308148068},
+{2121044558,-335940465}, {2116465518,-363675300}, {2111523833,-391347822},
+{2106220349,-418953288}, {2100555974,-446486968}, {2094531681,-473944146},
+{2088148500,-501320115}, {2081407525,-528610186}, {2074309912,-555809682},
+{2066856885,-582913912}, {2059049696,-609918325}, {2050889698,-636818231},
+{2042378310,-663608960}, {2033516972,-690285983}, {2024307180,-716844791},
+{2014750533,-743280770}, {2004848691,-769589332}, {1994603329,-795766029},
+{1984016179,-821806435}, {1973089077,-847706028}, {1961823921,-873460313},
+{1950222618,-899064934}, {1938287127,-924515564}, {1926019520,-949807783},
+{1913421927,-974937199}, {1900496481,-999899565}, {1887245364,-1024690661},
+{1873670877,-1049306180}, {1859775377,-1073741851}, {1845561215,-1097993541},
+{1831030826,-1122057097}, {1816186632,-1145928502}, {1801031311,-1169603450},
+{1785567394,-1193077993}, {1769797456,-1216348214}, {1753724345,-1239409914},
+{1737350743,-1262259248}, {1720679456,-1284892300}, {1703713340,-1307305194},
+{1686455222,-1329494189}, {1668908218,-1351455280}, {1651075255,-1373184807},
+{1632959307,-1394679144}, {1614563642,-1415934412}, {1595891331,-1436947067},
+{1576945572,-1457713510}, {1557729613,-1478230181}, {1538246655,-1498493658},
+{1518500216,-1518500282}, {1498493590,-1538246721}, {1478230113,-1557729677},
+{1457713441,-1576945636}, {1436946998,-1595891394}, {1415934341,-1614563704},
+{1394679073,-1632959368}, {1373184735,-1651075315}, {1351455207,-1668908277},
+{1329494115,-1686455280}, {1307305120,-1703713397}, {1284892225,-1720679512},
+{1262259172,-1737350799}, {1239409837,-1753724400}, {1216348136,-1769797510},
+{1193077915,-1785567446}, {1169603371,-1801031362}, {1145928423,-1816186682},
+{1122057017,-1831030875}, {1097993571,-1845561197}, {1073741769,-1859775424},
+{1049305987,-1873670985}, {1024690635,-1887245378}, {999899482,-1900496524},
+{974937230,-1913421912}, {949807699,-1926019561}, {924515422,-1938287195},
+{899064965,-1950222603}, {873460227,-1961823959}, {847705824,-1973089164},
+{821806407,-1984016190}, {795765941,-1994603364}, {769589125,-2004848771},
+{743280682,-2014750566}, {716844642,-2024307233}, {690286016,-2033516961},
+{663608871,-2042378339}, {636818019,-2050889764}, {609918296,-2059049705},
+{582913822,-2066856911}, {555809715,-2074309903}, {528610126,-2081407540},
+{501319962,-2088148536}, {473944148,-2094531680}, {446486876,-2100555994},
+{418953102,-2106220386}, {391347792,-2111523838}, {363675176,-2116465540},
+{335940246,-2121044593}, {308148006,-2125260177}, {280302715,-2129111646},
+{252409648,-2132598271}, {224473078,-2135719516}, {196498046,-2138474814},
+{168489600,-2140863674}, {140452029,-2142885728}, {112390647,-2144540593},
+{84309753,-2145828017}, {56214412,-2146747762}, {28109695,-2147299667},
+{2147483647,0}, {2146747758,-56214570}, {2144540595,-112390613},
+{2140863671,-168489630}, {2135719506,-224473172}, {2129111626,-280302871},
+{2121044558,-335940465}, {2111523833,-391347822}, {2100555974,-446486968},
+{2088148500,-501320115}, {2074309912,-555809682}, {2059049696,-609918325},
+{2042378310,-663608960}, {2024307180,-716844791}, {2004848691,-769589332},
+{1984016179,-821806435}, {1961823921,-873460313}, {1938287127,-924515564},
+{1913421927,-974937199}, {1887245364,-1024690661}, {1859775377,-1073741851},
+{1831030826,-1122057097}, {1801031311,-1169603450}, {1769797456,-1216348214},
+{1737350743,-1262259248}, {1703713340,-1307305194}, {1668908218,-1351455280},
+{1632959307,-1394679144}, {1595891331,-1436947067}, {1557729613,-1478230181},
+{1518500216,-1518500282}, {1478230113,-1557729677}, {1436946998,-1595891394},
+{1394679073,-1632959368}, {1351455207,-1668908277}, {1307305120,-1703713397},
+{1262259172,-1737350799}, {1216348136,-1769797510}, {1169603371,-1801031362},
+{1122057017,-1831030875}, {1073741769,-1859775424}, {1024690635,-1887245378},
+{974937230,-1913421912}, {924515422,-1938287195}, {873460227,-1961823959},
+{821806407,-1984016190}, {769589125,-2004848771}, {716844642,-2024307233},
+{663608871,-2042378339}, {609918296,-2059049705}, {555809715,-2074309903},
+{501319962,-2088148536}, {446486876,-2100555994}, {391347792,-2111523838},
+{335940246,-2121044593}, {280302715,-2129111646}, {224473078,-2135719516},
+{168489600,-2140863674}, {112390647,-2144540593}, {56214412,-2146747762},
+{-94,-2147483647}, {-56214600,-2146747757}, {-112390835,-2144540584},
+{-168489787,-2140863659}, {-224473265,-2135719496}, {-280302901,-2129111622},
+{-335940431,-2121044564}, {-391347977,-2111523804}, {-446487060,-2100555955},
+{-501320144,-2088148493}, {-555809896,-2074309855}, {-609918476,-2059049651},
+{-663609049,-2042378281}, {-716844819,-2024307170}, {-769589300,-2004848703},
+{-821806581,-1984016118}, {-873460398,-1961823883}, {-924515591,-1938287114},
+{-974937397,-1913421827}, {-1024690575,-1887245411}, {-1073741932,-1859775330},
+{-1122057395,-1831030643}, {-1169603421,-1801031330}, {-1216348291,-1769797403},
+{-1262259116,-1737350839}, {-1307305268,-1703713283}, {-1351455453,-1668908078},
+{-1394679021,-1632959413}, {-1436947137,-1595891268}, {-1478230435,-1557729372},
+{-1518500258,-1518500240}, {-1557729742,-1478230045}, {-1595891628,-1436946738},
+{-1632959429,-1394679001}, {-1668908417,-1351455035}, {-1703713298,-1307305248},
+{-1737350854,-1262259096}, {-1769797708,-1216347848}, {-1801031344,-1169603400},
+{-1831030924,-1122056937}, {-1859775343,-1073741910}, {-1887245423,-1024690552},
+{-1913422071,-974936918}, {-1938287125,-924515568}, {-1961823997,-873460141},
+{-1984016324,-821806084}, {-2004848713,-769589276}, {-2024307264,-716844553},
+{-2042378447,-663608538}, {-2059049731,-609918206}, {-2074309994,-555809377},
+{-2088148499,-501320119}, {-2100556013,-446486785}, {-2111523902,-391347448},
+{-2121044568,-335940406}, {-2129111659,-280302621}, {-2135719499,-224473240},
+{-2140863681,-168489506}, {-2144540612,-112390298}, {-2146747758,-56214574},
+{2147483647,0}, {2145828015,-84309815}, {2140863671,-168489630},
+{2132598271,-252409646}, {2121044558,-335940465}, {2106220349,-418953288},
+{2088148500,-501320115}, {2066856885,-582913912}, {2042378310,-663608960},
+{2014750533,-743280770}, {1984016179,-821806435}, {1950222618,-899064934},
+{1913421927,-974937199}, {1873670877,-1049306180}, {1831030826,-1122057097},
+{1785567394,-1193077993}, {1737350743,-1262259248}, {1686455222,-1329494189},
+{1632959307,-1394679144}, {1576945572,-1457713510}, {1518500216,-1518500282},
+{1457713441,-1576945636}, {1394679073,-1632959368}, {1329494115,-1686455280},
+{1262259172,-1737350799}, {1193077915,-1785567446}, {1122057017,-1831030875},
+{1049305987,-1873670985}, {974937230,-1913421912}, {899064965,-1950222603},
+{821806407,-1984016190}, {743280682,-2014750566}, {663608871,-2042378339},
+{582913822,-2066856911}, {501319962,-2088148536}, {418953102,-2106220386},
+{335940246,-2121044593}, {252409648,-2132598271}, {168489600,-2140863674},
+{84309753,-2145828017}, {-94,-2147483647}, {-84309940,-2145828010},
+{-168489787,-2140863659}, {-252409834,-2132598249}, {-335940431,-2121044564},
+{-418953286,-2106220349}, {-501320144,-2088148493}, {-582914003,-2066856860},
+{-663609049,-2042378281}, {-743280858,-2014750501}, {-821806581,-1984016118},
+{-899065136,-1950222525}, {-974937397,-1913421827}, {-1049306374,-1873670768},
+{-1122057395,-1831030643}, {-1193078284,-1785567199}, {-1262259116,-1737350839},
+{-1329494061,-1686455323}, {-1394679021,-1632959413}, {-1457713485,-1576945595},
+{-1518500258,-1518500240}, {-1576945613,-1457713466}, {-1632959429,-1394679001},
+{-1686455338,-1329494041}, {-1737350854,-1262259096}, {-1785567498,-1193077837},
+{-1831030924,-1122056937}, {-1873671031,-1049305905}, {-1913422071,-974936918},
+{-1950222750,-899064648}, {-1984016324,-821806084}, {-2014750687,-743280354},
+{-2042378447,-663608538}, {-2066856867,-582913978}, {-2088148499,-501320119},
+{-2106220354,-418953261}, {-2121044568,-335940406}, {-2132598282,-252409555},
+{-2140863681,-168489506}, {-2145828021,-84309659}, {-2147483647,188},
+{-2145828006,84310034}, {-2140863651,168489881}, {-2132598237,252409928},
+{-2121044509,335940777}, {-2106220281,418953629}, {-2088148411,501320484},
+{-2066856765,582914339}, {-2042378331,663608895}, {-2014750557,743280706},
+{-1984016181,821806431}, {-1950222593,899064989}, {-1913421900,974937252},
+{-1873670848,1049306232}, {-1831030728,1122057257}, {-1785567289,1193078149},
+{-1737350633,1262259400}, {-1686455106,1329494336}, {-1632959185,1394679287},
+{-1576945358,1457713742}, {-1518499993,1518500506}, {-1457713209,1576945850},
+{-1394678735,1632959656}, {-1329493766,1686455555}, {-1262258813,1737351059},
+{-1193077546,1785567692}, {-1122056638,1831031107}, {-1049305599,1873671202},
+{-974936606,1913422229}, {-899064330,1950222896}, {-821805761,1984016458},
+{-743280025,2014750808}, {-663609179,2042378239}, {-582914134,2066856823},
+{-501320277,2088148461}, {-418953420,2106220322}, {-335940566,2121044542},
+{-252409716,2132598263}, {-168489668,2140863668}, {-84309821,2145828015},
+};
+static const ne10_fft_cpx_int32_t ne10_twiddles_240[240] = {
+{0,0}, {2147483647,0}, {2147483647,0},
+{2147483647,0}, {1961823921,-873460313}, {1436946998,-1595891394},
+{2147483647,0}, {1436946998,-1595891394}, {-224473265,-2135719496},
+{2147483647,0}, {663608871,-2042378339}, {-1737350854,-1262259096},
+{2147483647,0}, {-224473265,-2135719496}, {-2100555935,446487152},
+{2147483647,0}, {2135719506,-224473172}, {2100555974,-446486968},
+{2042378310,-663608960}, {1961823921,-873460313}, {1859775377,-1073741851},
+{1737350743,-1262259248}, {1595891331,-1436947067}, {1436946998,-1595891394},
+{1262259172,-1737350799}, {1073741769,-1859775424}, {873460227,-1961823959},
+{663608871,-2042378339}, {446486876,-2100555994}, {224473078,-2135719516},
+{2147483647,0}, {2100555974,-446486968}, {1961823921,-873460313},
+{1737350743,-1262259248}, {1436946998,-1595891394}, {1073741769,-1859775424},
+{663608871,-2042378339}, {224473078,-2135719516}, {-224473265,-2135719496},
+{-663609049,-2042378281}, {-1073741932,-1859775330}, {-1436947137,-1595891268},
+{-1737350854,-1262259096}, {-1961823997,-873460141}, {-2100556013,-446486785},
+{2147483647,0}, {2042378310,-663608960}, {1737350743,-1262259248},
+{1262259172,-1737350799}, {663608871,-2042378339}, {-94,-2147483647},
+{-663609049,-2042378281}, {-1262259116,-1737350839}, {-1737350854,-1262259096},
+{-2042378447,-663608538}, {-2147483647,188}, {-2042378331,663608895},
+{-1737350633,1262259400}, {-1262258813,1737351059}, {-663609179,2042378239},
+{2147483647,0}, {2146747758,-56214570}, {2144540595,-112390613},
+{2140863671,-168489630}, {2135719506,-224473172}, {2129111626,-280302871},
+{2121044558,-335940465}, {2111523833,-391347822}, {2100555974,-446486968},
+{2088148500,-501320115}, {2074309912,-555809682}, {2059049696,-609918325},
+{2042378310,-663608960}, {2024307180,-716844791}, {2004848691,-769589332},
+{1984016179,-821806435}, {1961823921,-873460313}, {1938287127,-924515564},
+{1913421927,-974937199}, {1887245364,-1024690661}, {1859775377,-1073741851},
+{1831030826,-1122057097}, {1801031311,-1169603450}, {1769797456,-1216348214},
+{1737350743,-1262259248}, {1703713340,-1307305194}, {1668908218,-1351455280},
+{1632959307,-1394679144}, {1595891331,-1436947067}, {1557729613,-1478230181},
+{1518500216,-1518500282}, {1478230113,-1557729677}, {1436946998,-1595891394},
+{1394679073,-1632959368}, {1351455207,-1668908277}, {1307305120,-1703713397},
+{1262259172,-1737350799}, {1216348136,-1769797510}, {1169603371,-1801031362},
+{1122057017,-1831030875}, {1073741769,-1859775424}, {1024690635,-1887245378},
+{974937230,-1913421912}, {924515422,-1938287195}, {873460227,-1961823959},
+{821806407,-1984016190}, {769589125,-2004848771}, {716844642,-2024307233},
+{663608871,-2042378339}, {609918296,-2059049705}, {555809715,-2074309903},
+{501319962,-2088148536}, {446486876,-2100555994}, {391347792,-2111523838},
+{335940246,-2121044593}, {280302715,-2129111646}, {224473078,-2135719516},
+{168489600,-2140863674}, {112390647,-2144540593}, {56214412,-2146747762},
+{2147483647,0}, {2144540595,-112390613}, {2135719506,-224473172},
+{2121044558,-335940465}, {2100555974,-446486968}, {2074309912,-555809682},
+{2042378310,-663608960}, {2004848691,-769589332}, {1961823921,-873460313},
+{1913421927,-974937199}, {1859775377,-1073741851}, {1801031311,-1169603450},
+{1737350743,-1262259248}, {1668908218,-1351455280}, {1595891331,-1436947067},
+{1518500216,-1518500282}, {1436946998,-1595891394}, {1351455207,-1668908277},
+{1262259172,-1737350799}, {1169603371,-1801031362}, {1073741769,-1859775424},
+{974937230,-1913421912}, {873460227,-1961823959}, {769589125,-2004848771},
+{663608871,-2042378339}, {555809715,-2074309903}, {446486876,-2100555994},
+{335940246,-2121044593}, {224473078,-2135719516}, {112390647,-2144540593},
+{-94,-2147483647}, {-112390835,-2144540584}, {-224473265,-2135719496},
+{-335940431,-2121044564}, {-446487060,-2100555955}, {-555809896,-2074309855},
+{-663609049,-2042378281}, {-769589300,-2004848703}, {-873460398,-1961823883},
+{-974937397,-1913421827}, {-1073741932,-1859775330}, {-1169603421,-1801031330},
+{-1262259116,-1737350839}, {-1351455453,-1668908078}, {-1436947137,-1595891268},
+{-1518500258,-1518500240}, {-1595891628,-1436946738}, {-1668908417,-1351455035},
+{-1737350854,-1262259096}, {-1801031344,-1169603400}, {-1859775343,-1073741910},
+{-1913422071,-974936918}, {-1961823997,-873460141}, {-2004848713,-769589276},
+{-2042378447,-663608538}, {-2074309994,-555809377}, {-2100556013,-446486785},
+{-2121044568,-335940406}, {-2135719499,-224473240}, {-2144540612,-112390298},
+{2147483647,0}, {2140863671,-168489630}, {2121044558,-335940465},
+{2088148500,-501320115}, {2042378310,-663608960}, {1984016179,-821806435},
+{1913421927,-974937199}, {1831030826,-1122057097}, {1737350743,-1262259248},
+{1632959307,-1394679144}, {1518500216,-1518500282}, {1394679073,-1632959368},
+{1262259172,-1737350799}, {1122057017,-1831030875}, {974937230,-1913421912},
+{821806407,-1984016190}, {663608871,-2042378339}, {501319962,-2088148536},
+{335940246,-2121044593}, {168489600,-2140863674}, {-94,-2147483647},
+{-168489787,-2140863659}, {-335940431,-2121044564}, {-501320144,-2088148493},
+{-663609049,-2042378281}, {-821806581,-1984016118}, {-974937397,-1913421827},
+{-1122057395,-1831030643}, {-1262259116,-1737350839}, {-1394679021,-1632959413},
+{-1518500258,-1518500240}, {-1632959429,-1394679001}, {-1737350854,-1262259096},
+{-1831030924,-1122056937}, {-1913422071,-974936918}, {-1984016324,-821806084},
+{-2042378447,-663608538}, {-2088148499,-501320119}, {-2121044568,-335940406},
+{-2140863681,-168489506}, {-2147483647,188}, {-2140863651,168489881},
+{-2121044509,335940777}, {-2088148411,501320484}, {-2042378331,663608895},
+{-1984016181,821806431}, {-1913421900,974937252}, {-1831030728,1122057257},
+{-1737350633,1262259400}, {-1632959185,1394679287}, {-1518499993,1518500506},
+{-1394678735,1632959656}, {-1262258813,1737351059}, {-1122056638,1831031107},
+{-974936606,1913422229}, {-821805761,1984016458}, {-663609179,2042378239},
+{-501320277,2088148461}, {-335940566,2121044542}, {-168489668,2140863668},
+};
+static const ne10_fft_cpx_int32_t ne10_twiddles_120[120] = {
+{0,0}, {2147483647,0}, {2147483647,0},
+{2147483647,0}, {1961823921,-873460313}, {1436946998,-1595891394},
+{2147483647,0}, {1436946998,-1595891394}, {-224473265,-2135719496},
+{2147483647,0}, {663608871,-2042378339}, {-1737350854,-1262259096},
+{2147483647,0}, {-224473265,-2135719496}, {-2100555935,446487152},
+{2147483647,0}, {2100555974,-446486968}, {1961823921,-873460313},
+{1737350743,-1262259248}, {1436946998,-1595891394}, {1073741769,-1859775424},
+{663608871,-2042378339}, {224473078,-2135719516}, {-224473265,-2135719496},
+{-663609049,-2042378281}, {-1073741932,-1859775330}, {-1436947137,-1595891268},
+{-1737350854,-1262259096}, {-1961823997,-873460141}, {-2100556013,-446486785},
+{2147483647,0}, {2144540595,-112390613}, {2135719506,-224473172},
+{2121044558,-335940465}, {2100555974,-446486968}, {2074309912,-555809682},
+{2042378310,-663608960}, {2004848691,-769589332}, {1961823921,-873460313},
+{1913421927,-974937199}, {1859775377,-1073741851}, {1801031311,-1169603450},
+{1737350743,-1262259248}, {1668908218,-1351455280}, {1595891331,-1436947067},
+{1518500216,-1518500282}, {1436946998,-1595891394}, {1351455207,-1668908277},
+{1262259172,-1737350799}, {1169603371,-1801031362}, {1073741769,-1859775424},
+{974937230,-1913421912}, {873460227,-1961823959}, {769589125,-2004848771},
+{663608871,-2042378339}, {555809715,-2074309903}, {446486876,-2100555994},
+{335940246,-2121044593}, {224473078,-2135719516}, {112390647,-2144540593},
+{2147483647,0}, {2135719506,-224473172}, {2100555974,-446486968},
+{2042378310,-663608960}, {1961823921,-873460313}, {1859775377,-1073741851},
+{1737350743,-1262259248}, {1595891331,-1436947067}, {1436946998,-1595891394},
+{1262259172,-1737350799}, {1073741769,-1859775424}, {873460227,-1961823959},
+{663608871,-2042378339}, {446486876,-2100555994}, {224473078,-2135719516},
+{-94,-2147483647}, {-224473265,-2135719496}, {-446487060,-2100555955},
+{-663609049,-2042378281}, {-873460398,-1961823883}, {-1073741932,-1859775330},
+{-1262259116,-1737350839}, {-1436947137,-1595891268}, {-1595891628,-1436946738},
+{-1737350854,-1262259096}, {-1859775343,-1073741910}, {-1961823997,-873460141},
+{-2042378447,-663608538}, {-2100556013,-446486785}, {-2135719499,-224473240},
+{2147483647,0}, {2121044558,-335940465}, {2042378310,-663608960},
+{1913421927,-974937199}, {1737350743,-1262259248}, {1518500216,-1518500282},
+{1262259172,-1737350799}, {974937230,-1913421912}, {663608871,-2042378339},
+{335940246,-2121044593}, {-94,-2147483647}, {-335940431,-2121044564},
+{-663609049,-2042378281}, {-974937397,-1913421827}, {-1262259116,-1737350839},
+{-1518500258,-1518500240}, {-1737350854,-1262259096}, {-1913422071,-974936918},
+{-2042378447,-663608538}, {-2121044568,-335940406}, {-2147483647,188},
+{-2121044509,335940777}, {-2042378331,663608895}, {-1913421900,974937252},
+{-1737350633,1262259400}, {-1518499993,1518500506}, {-1262258813,1737351059},
+{-974936606,1913422229}, {-663609179,2042378239}, {-335940566,2121044542},
+};
+static const ne10_fft_cpx_int32_t ne10_twiddles_60[60] = {
+{0,0}, {2147483647,0}, {2147483647,0},
+{2147483647,0}, {1961823921,-873460313}, {1436946998,-1595891394},
+{2147483647,0}, {1436946998,-1595891394}, {-224473265,-2135719496},
+{2147483647,0}, {663608871,-2042378339}, {-1737350854,-1262259096},
+{2147483647,0}, {-224473265,-2135719496}, {-2100555935,446487152},
+{2147483647,0}, {2135719506,-224473172}, {2100555974,-446486968},
+{2042378310,-663608960}, {1961823921,-873460313}, {1859775377,-1073741851},
+{1737350743,-1262259248}, {1595891331,-1436947067}, {1436946998,-1595891394},
+{1262259172,-1737350799}, {1073741769,-1859775424}, {873460227,-1961823959},
+{663608871,-2042378339}, {446486876,-2100555994}, {224473078,-2135719516},
+{2147483647,0}, {2100555974,-446486968}, {1961823921,-873460313},
+{1737350743,-1262259248}, {1436946998,-1595891394}, {1073741769,-1859775424},
+{663608871,-2042378339}, {224473078,-2135719516}, {-224473265,-2135719496},
+{-663609049,-2042378281}, {-1073741932,-1859775330}, {-1436947137,-1595891268},
+{-1737350854,-1262259096}, {-1961823997,-873460141}, {-2100556013,-446486785},
+{2147483647,0}, {2042378310,-663608960}, {1737350743,-1262259248},
+{1262259172,-1737350799}, {663608871,-2042378339}, {-94,-2147483647},
+{-663609049,-2042378281}, {-1262259116,-1737350839}, {-1737350854,-1262259096},
+{-2042378447,-663608538}, {-2147483647,188}, {-2042378331,663608895},
+{-1737350633,1262259400}, {-1262258813,1737351059}, {-663609179,2042378239},
+};
+static const ne10_fft_state_int32_t ne10_fft_state_int32_t_480 = {
+120,
+(ne10_int32_t *)ne10_factors_480,
+(ne10_fft_cpx_int32_t *)ne10_twiddles_480,
+NULL,
+(ne10_fft_cpx_int32_t *)&ne10_twiddles_480[120],
+};
+static const arch_fft_state cfg_arch_480 = {
+1,
+(void *)&ne10_fft_state_int32_t_480,
+};
+
+static const ne10_fft_state_int32_t ne10_fft_state_int32_t_240 = {
+60,
+(ne10_int32_t *)ne10_factors_240,
+(ne10_fft_cpx_int32_t *)ne10_twiddles_240,
+NULL,
+(ne10_fft_cpx_int32_t *)&ne10_twiddles_240[60],
+};
+static const arch_fft_state cfg_arch_240 = {
+1,
+(void *)&ne10_fft_state_int32_t_240,
+};
+
+static const ne10_fft_state_int32_t ne10_fft_state_int32_t_120 = {
+30,
+(ne10_int32_t *)ne10_factors_120,
+(ne10_fft_cpx_int32_t *)ne10_twiddles_120,
+NULL,
+(ne10_fft_cpx_int32_t *)&ne10_twiddles_120[30],
+};
+static const arch_fft_state cfg_arch_120 = {
+1,
+(void *)&ne10_fft_state_int32_t_120,
+};
+
+static const ne10_fft_state_int32_t ne10_fft_state_int32_t_60 = {
+15,
+(ne10_int32_t *)ne10_factors_60,
+(ne10_fft_cpx_int32_t *)ne10_twiddles_60,
+NULL,
+(ne10_fft_cpx_int32_t *)&ne10_twiddles_60[15],
+};
+static const arch_fft_state cfg_arch_60 = {
+1,
+(void *)&ne10_fft_state_int32_t_60,
+};
+
+#endif  /* end NE10_FFT_PARAMS48000_960 */
--- a/lib/libopus/celt/static_modes_float.h
+++ b/lib/libopus/celt/static_modes_float.h
--- a/lib/libopus/celt/static_modes_float_arm_ne10.h
+++ b/lib/libopus/celt/static_modes_float_arm_ne10.h
@ -0,0 +1,404 @@
+/* The contents of this file was automatically generated by
+ * dump_mode_arm_ne10.c with arguments: 48000 960
+ * It contains static definitions for some pre-defined modes. */
+#include <NE10_types.h>
+
+#ifndef NE10_FFT_PARAMS48000_960
+#define NE10_FFT_PARAMS48000_960
+static const ne10_int32_t ne10_factors_480[64] = {
+4, 40, 4, 30, 2, 15, 5, 3, 3, 1, 1, 0, 0, 0, 0,
+0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+0, 0, 0, 0, };
+static const ne10_int32_t ne10_factors_240[64] = {
+3, 20, 4, 15, 5, 3, 3, 1, 1, 0, 0, 0, 0, 0, 0,
+0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+0, 0, 0, 0, };
+static const ne10_int32_t ne10_factors_120[64] = {
+3, 10, 2, 15, 5, 3, 3, 1, 1, 0, 0, 0, 0, 0, 0,
+0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+0, 0, 0, 0, };
+static const ne10_int32_t ne10_factors_60[64] = {
+2, 5, 5, 3, 3, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0,
+0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+0, 0, 0, 0, };
+static const ne10_fft_cpx_float32_t ne10_twiddles_480[480] = {
+{1.0000000f,0.0000000f}, {1.0000000f,-0.0000000f}, {1.0000000f,-0.0000000f},
+{1.0000000f,-0.0000000f}, {0.91354543f,-0.40673664f}, {0.66913056f,-0.74314487f},
+{1.0000000f,-0.0000000f}, {0.66913056f,-0.74314487f}, {-0.10452851f,-0.99452192f},
+{1.0000000f,-0.0000000f}, {0.30901697f,-0.95105654f}, {-0.80901700f,-0.58778518f},
+{1.0000000f,-0.0000000f}, {-0.10452851f,-0.99452192f}, {-0.97814757f,0.20791179f},
+{1.0000000f,-0.0000000f}, {0.97814763f,-0.20791170f}, {0.91354543f,-0.40673664f},
+{0.80901700f,-0.58778524f}, {0.66913056f,-0.74314487f}, {0.49999997f,-0.86602545f},
+{0.30901697f,-0.95105654f}, {0.10452842f,-0.99452192f}, {-0.10452851f,-0.99452192f},
+{-0.30901703f,-0.95105648f}, {-0.50000006f,-0.86602533f}, {-0.66913068f,-0.74314475f},
+{-0.80901700f,-0.58778518f}, {-0.91354549f,-0.40673658f}, {-0.97814763f,-0.20791161f},
+{1.0000000f,-0.0000000f}, {0.99862951f,-0.052335959f}, {0.99452192f,-0.10452846f},
+{0.98768836f,-0.15643448f}, {0.97814763f,-0.20791170f}, {0.96592581f,-0.25881904f},
+{0.95105648f,-0.30901700f}, {0.93358040f,-0.35836795f}, {0.91354543f,-0.40673664f},
+{0.89100653f,-0.45399052f}, {0.86602545f,-0.50000000f}, {0.83867055f,-0.54463905f},
+{0.80901700f,-0.58778524f}, {0.77714598f,-0.62932038f}, {0.74314475f,-0.66913062f},
+{0.70710677f,-0.70710683f}, {0.66913056f,-0.74314487f}, {0.62932038f,-0.77714598f},
+{0.58778524f,-0.80901700f}, {0.54463899f,-0.83867055f}, {0.49999997f,-0.86602545f},
+{0.45399052f,-0.89100653f}, {0.40673661f,-0.91354549f}, {0.35836786f,-0.93358046f},
+{0.30901697f,-0.95105654f}, {0.25881907f,-0.96592581f}, {0.20791166f,-0.97814763f},
+{0.15643437f,-0.98768836f}, {0.10452842f,-0.99452192f}, {0.052335974f,-0.99862951f},
+{1.0000000f,-0.0000000f}, {0.99452192f,-0.10452846f}, {0.97814763f,-0.20791170f},
+{0.95105648f,-0.30901700f}, {0.91354543f,-0.40673664f}, {0.86602545f,-0.50000000f},
+{0.80901700f,-0.58778524f}, {0.74314475f,-0.66913062f}, {0.66913056f,-0.74314487f},
+{0.58778524f,-0.80901700f}, {0.49999997f,-0.86602545f}, {0.40673661f,-0.91354549f},
+{0.30901697f,-0.95105654f}, {0.20791166f,-0.97814763f}, {0.10452842f,-0.99452192f},
+{-4.3711388e-08f,-1.0000000f}, {-0.10452851f,-0.99452192f}, {-0.20791174f,-0.97814757f},
+{-0.30901703f,-0.95105648f}, {-0.40673670f,-0.91354543f}, {-0.50000006f,-0.86602533f},
+{-0.58778518f,-0.80901700f}, {-0.66913068f,-0.74314475f}, {-0.74314493f,-0.66913044f},
+{-0.80901700f,-0.58778518f}, {-0.86602539f,-0.50000006f}, {-0.91354549f,-0.40673658f},
+{-0.95105654f,-0.30901679f}, {-0.97814763f,-0.20791161f}, {-0.99452192f,-0.10452849f},
+{1.0000000f,-0.0000000f}, {0.98768836f,-0.15643448f}, {0.95105648f,-0.30901700f},
+{0.89100653f,-0.45399052f}, {0.80901700f,-0.58778524f}, {0.70710677f,-0.70710683f},
+{0.58778524f,-0.80901700f}, {0.45399052f,-0.89100653f}, {0.30901697f,-0.95105654f},
+{0.15643437f,-0.98768836f}, {-4.3711388e-08f,-1.0000000f}, {-0.15643445f,-0.98768836f},
+{-0.30901703f,-0.95105648f}, {-0.45399061f,-0.89100647f}, {-0.58778518f,-0.80901700f},
+{-0.70710677f,-0.70710677f}, {-0.80901700f,-0.58778518f}, {-0.89100659f,-0.45399037f},
+{-0.95105654f,-0.30901679f}, {-0.98768836f,-0.15643445f}, {-1.0000000f,8.7422777e-08f},
+{-0.98768830f,0.15643461f}, {-0.95105654f,0.30901697f}, {-0.89100653f,0.45399055f},
+{-0.80901694f,0.58778536f}, {-0.70710665f,0.70710689f}, {-0.58778507f,0.80901712f},
+{-0.45399022f,0.89100665f}, {-0.30901709f,0.95105648f}, {-0.15643452f,0.98768830f},
+{1.0000000f,-0.0000000f}, {0.99991435f,-0.013089596f}, {0.99965733f,-0.026176950f},
+{0.99922901f,-0.039259817f}, {0.99862951f,-0.052335959f}, {0.99785894f,-0.065403134f},
+{0.99691731f,-0.078459099f}, {0.99580491f,-0.091501623f}, {0.99452192f,-0.10452846f},
+{0.99306846f,-0.11753740f}, {0.99144489f,-0.13052620f}, {0.98965138f,-0.14349262f},
+{0.98768836f,-0.15643448f}, {0.98555607f,-0.16934951f}, {0.98325491f,-0.18223552f},
+{0.98078525f,-0.19509032f}, {0.97814763f,-0.20791170f}, {0.97534233f,-0.22069745f},
+{0.97236991f,-0.23344538f}, {0.96923089f,-0.24615330f}, {0.96592581f,-0.25881904f},
+{0.96245521f,-0.27144045f}, {0.95881975f,-0.28401536f}, {0.95501995f,-0.29654160f},
+{0.95105648f,-0.30901700f}, {0.94693011f,-0.32143945f}, {0.94264150f,-0.33380687f},
+{0.93819129f,-0.34611708f}, {0.93358040f,-0.35836795f}, {0.92880952f,-0.37055743f},
+{0.92387956f,-0.38268346f}, {0.91879117f,-0.39474389f}, {0.91354543f,-0.40673664f},
+{0.90814316f,-0.41865975f}, {0.90258527f,-0.43051112f}, {0.89687270f,-0.44228873f},
+{0.89100653f,-0.45399052f}, {0.88498765f,-0.46561453f}, {0.87881708f,-0.47715878f},
+{0.87249601f,-0.48862126f}, {0.86602545f,-0.50000000f}, {0.85940641f,-0.51129311f},
+{0.85264015f,-0.52249855f}, {0.84572786f,-0.53361452f}, {0.83867055f,-0.54463905f},
+{0.83146960f,-0.55557024f}, {0.82412618f,-0.56640625f}, {0.81664151f,-0.57714522f},
+{0.80901700f,-0.58778524f}, {0.80125380f,-0.59832460f}, {0.79335332f,-0.60876143f},
+{0.78531694f,-0.61909395f}, {0.77714598f,-0.62932038f}, {0.76884180f,-0.63943899f},
+{0.76040596f,-0.64944810f}, {0.75183982f,-0.65934587f}, {0.74314475f,-0.66913062f},
+{0.73432249f,-0.67880076f}, {0.72537434f,-0.68835455f}, {0.71630192f,-0.69779050f},
+{0.70710677f,-0.70710683f}, {0.69779044f,-0.71630198f}, {0.68835455f,-0.72537440f},
+{0.67880070f,-0.73432255f}, {0.66913056f,-0.74314487f}, {0.65934581f,-0.75183982f},
+{0.64944804f,-0.76040596f}, {0.63943899f,-0.76884186f}, {0.62932038f,-0.77714598f},
+{0.61909395f,-0.78531694f}, {0.60876137f,-0.79335338f}, {0.59832460f,-0.80125386f},
+{0.58778524f,-0.80901700f}, {0.57714516f,-0.81664151f}, {0.56640625f,-0.82412618f},
+{0.55557019f,-0.83146960f}, {0.54463899f,-0.83867055f}, {0.53361452f,-0.84572786f},
+{0.52249849f,-0.85264015f}, {0.51129311f,-0.85940641f}, {0.49999997f,-0.86602545f},
+{0.48862118f,-0.87249601f}, {0.47715876f,-0.87881708f}, {0.46561447f,-0.88498765f},
+{0.45399052f,-0.89100653f}, {0.44228867f,-0.89687276f}, {0.43051103f,-0.90258533f},
+{0.41865975f,-0.90814316f}, {0.40673661f,-0.91354549f}, {0.39474380f,-0.91879129f},
+{0.38268343f,-0.92387956f}, {0.37055740f,-0.92880958f}, {0.35836786f,-0.93358046f},
+{0.34611705f,-0.93819135f}, {0.33380681f,-0.94264150f}, {0.32143947f,-0.94693011f},
+{0.30901697f,-0.95105654f}, {0.29654151f,-0.95501995f}, {0.28401533f,-0.95881975f},
+{0.27144039f,-0.96245527f}, {0.25881907f,-0.96592581f}, {0.24615327f,-0.96923089f},
+{0.23344530f,-0.97236991f}, {0.22069745f,-0.97534233f}, {0.20791166f,-0.97814763f},
+{0.19509023f,-0.98078531f}, {0.18223552f,-0.98325491f}, {0.16934945f,-0.98555607f},
+{0.15643437f,-0.98768836f}, {0.14349259f,-0.98965138f}, {0.13052613f,-0.99144489f},
+{0.11753740f,-0.99306846f}, {0.10452842f,-0.99452192f}, {0.091501534f,-0.99580491f},
+{0.078459084f,-0.99691731f}, {0.065403074f,-0.99785894f}, {0.052335974f,-0.99862951f},
+{0.039259788f,-0.99922901f}, {0.026176875f,-0.99965733f}, {0.013089597f,-0.99991435f},
+{1.0000000f,-0.0000000f}, {0.99965733f,-0.026176950f}, {0.99862951f,-0.052335959f},
+{0.99691731f,-0.078459099f}, {0.99452192f,-0.10452846f}, {0.99144489f,-0.13052620f},
+{0.98768836f,-0.15643448f}, {0.98325491f,-0.18223552f}, {0.97814763f,-0.20791170f},
+{0.97236991f,-0.23344538f}, {0.96592581f,-0.25881904f}, {0.95881975f,-0.28401536f},
+{0.95105648f,-0.30901700f}, {0.94264150f,-0.33380687f}, {0.93358040f,-0.35836795f},
+{0.92387956f,-0.38268346f}, {0.91354543f,-0.40673664f}, {0.90258527f,-0.43051112f},
+{0.89100653f,-0.45399052f}, {0.87881708f,-0.47715878f}, {0.86602545f,-0.50000000f},
+{0.85264015f,-0.52249855f}, {0.83867055f,-0.54463905f}, {0.82412618f,-0.56640625f},
+{0.80901700f,-0.58778524f}, {0.79335332f,-0.60876143f}, {0.77714598f,-0.62932038f},
+{0.76040596f,-0.64944810f}, {0.74314475f,-0.66913062f}, {0.72537434f,-0.68835455f},
+{0.70710677f,-0.70710683f}, {0.68835455f,-0.72537440f}, {0.66913056f,-0.74314487f},
+{0.64944804f,-0.76040596f}, {0.62932038f,-0.77714598f}, {0.60876137f,-0.79335338f},
+{0.58778524f,-0.80901700f}, {0.56640625f,-0.82412618f}, {0.54463899f,-0.83867055f},
+{0.52249849f,-0.85264015f}, {0.49999997f,-0.86602545f}, {0.47715876f,-0.87881708f},
+{0.45399052f,-0.89100653f}, {0.43051103f,-0.90258533f}, {0.40673661f,-0.91354549f},
+{0.38268343f,-0.92387956f}, {0.35836786f,-0.93358046f}, {0.33380681f,-0.94264150f},
+{0.30901697f,-0.95105654f}, {0.28401533f,-0.95881975f}, {0.25881907f,-0.96592581f},
+{0.23344530f,-0.97236991f}, {0.20791166f,-0.97814763f}, {0.18223552f,-0.98325491f},
+{0.15643437f,-0.98768836f}, {0.13052613f,-0.99144489f}, {0.10452842f,-0.99452192f},
+{0.078459084f,-0.99691731f}, {0.052335974f,-0.99862951f}, {0.026176875f,-0.99965733f},
+{-4.3711388e-08f,-1.0000000f}, {-0.026176963f,-0.99965733f}, {-0.052336060f,-0.99862951f},
+{-0.078459173f,-0.99691731f}, {-0.10452851f,-0.99452192f}, {-0.13052621f,-0.99144489f},
+{-0.15643445f,-0.98768836f}, {-0.18223560f,-0.98325491f}, {-0.20791174f,-0.97814757f},
+{-0.23344538f,-0.97236991f}, {-0.25881916f,-0.96592581f}, {-0.28401542f,-0.95881969f},
+{-0.30901703f,-0.95105648f}, {-0.33380687f,-0.94264150f}, {-0.35836795f,-0.93358040f},
+{-0.38268352f,-0.92387950f}, {-0.40673670f,-0.91354543f}, {-0.43051112f,-0.90258527f},
+{-0.45399061f,-0.89100647f}, {-0.47715873f,-0.87881708f}, {-0.50000006f,-0.86602533f},
+{-0.52249867f,-0.85264009f}, {-0.54463905f,-0.83867055f}, {-0.56640631f,-0.82412612f},
+{-0.58778518f,-0.80901700f}, {-0.60876143f,-0.79335332f}, {-0.62932050f,-0.77714586f},
+{-0.64944804f,-0.76040596f}, {-0.66913068f,-0.74314475f}, {-0.68835467f,-0.72537428f},
+{-0.70710677f,-0.70710677f}, {-0.72537446f,-0.68835449f}, {-0.74314493f,-0.66913044f},
+{-0.76040596f,-0.64944804f}, {-0.77714604f,-0.62932026f}, {-0.79335332f,-0.60876143f},
+{-0.80901700f,-0.58778518f}, {-0.82412624f,-0.56640613f}, {-0.83867055f,-0.54463899f},
+{-0.85264021f,-0.52249849f}, {-0.86602539f,-0.50000006f}, {-0.87881714f,-0.47715873f},
+{-0.89100659f,-0.45399037f}, {-0.90258527f,-0.43051112f}, {-0.91354549f,-0.40673658f},
+{-0.92387956f,-0.38268328f}, {-0.93358040f,-0.35836792f}, {-0.94264150f,-0.33380675f},
+{-0.95105654f,-0.30901679f}, {-0.95881975f,-0.28401530f}, {-0.96592587f,-0.25881892f},
+{-0.97236991f,-0.23344538f}, {-0.97814763f,-0.20791161f}, {-0.98325491f,-0.18223536f},
+{-0.98768836f,-0.15643445f}, {-0.99144489f,-0.13052608f}, {-0.99452192f,-0.10452849f},
+{-0.99691737f,-0.078459039f}, {-0.99862957f,-0.052335810f}, {-0.99965733f,-0.026176952f},
+{1.0000000f,-0.0000000f}, {0.99922901f,-0.039259817f}, {0.99691731f,-0.078459099f},
+{0.99306846f,-0.11753740f}, {0.98768836f,-0.15643448f}, {0.98078525f,-0.19509032f},
+{0.97236991f,-0.23344538f}, {0.96245521f,-0.27144045f}, {0.95105648f,-0.30901700f},
+{0.93819129f,-0.34611708f}, {0.92387956f,-0.38268346f}, {0.90814316f,-0.41865975f},
+{0.89100653f,-0.45399052f}, {0.87249601f,-0.48862126f}, {0.85264015f,-0.52249855f},
+{0.83146960f,-0.55557024f}, {0.80901700f,-0.58778524f}, {0.78531694f,-0.61909395f},
+{0.76040596f,-0.64944810f}, {0.73432249f,-0.67880076f}, {0.70710677f,-0.70710683f},
+{0.67880070f,-0.73432255f}, {0.64944804f,-0.76040596f}, {0.61909395f,-0.78531694f},
+{0.58778524f,-0.80901700f}, {0.55557019f,-0.83146960f}, {0.52249849f,-0.85264015f},
+{0.48862118f,-0.87249601f}, {0.45399052f,-0.89100653f}, {0.41865975f,-0.90814316f},
+{0.38268343f,-0.92387956f}, {0.34611705f,-0.93819135f}, {0.30901697f,-0.95105654f},
+{0.27144039f,-0.96245527f}, {0.23344530f,-0.97236991f}, {0.19509023f,-0.98078531f},
+{0.15643437f,-0.98768836f}, {0.11753740f,-0.99306846f}, {0.078459084f,-0.99691731f},
+{0.039259788f,-0.99922901f}, {-4.3711388e-08f,-1.0000000f}, {-0.039259877f,-0.99922901f},
+{-0.078459173f,-0.99691731f}, {-0.11753749f,-0.99306846f}, {-0.15643445f,-0.98768836f},
+{-0.19509032f,-0.98078525f}, {-0.23344538f,-0.97236991f}, {-0.27144048f,-0.96245521f},
+{-0.30901703f,-0.95105648f}, {-0.34611711f,-0.93819129f}, {-0.38268352f,-0.92387950f},
+{-0.41865984f,-0.90814310f}, {-0.45399061f,-0.89100647f}, {-0.48862135f,-0.87249595f},
+{-0.52249867f,-0.85264009f}, {-0.55557036f,-0.83146954f}, {-0.58778518f,-0.80901700f},
+{-0.61909389f,-0.78531694f}, {-0.64944804f,-0.76040596f}, {-0.67880076f,-0.73432249f},
+{-0.70710677f,-0.70710677f}, {-0.73432249f,-0.67880070f}, {-0.76040596f,-0.64944804f},
+{-0.78531694f,-0.61909389f}, {-0.80901700f,-0.58778518f}, {-0.83146966f,-0.55557019f},
+{-0.85264021f,-0.52249849f}, {-0.87249607f,-0.48862115f}, {-0.89100659f,-0.45399037f},
+{-0.90814322f,-0.41865960f}, {-0.92387956f,-0.38268328f}, {-0.93819135f,-0.34611690f},
+{-0.95105654f,-0.30901679f}, {-0.96245521f,-0.27144048f}, {-0.97236991f,-0.23344538f},
+{-0.98078531f,-0.19509031f}, {-0.98768836f,-0.15643445f}, {-0.99306846f,-0.11753736f},
+{-0.99691737f,-0.078459039f}, {-0.99922901f,-0.039259743f}, {-1.0000000f,8.7422777e-08f},
+{-0.99922901f,0.039259918f}, {-0.99691731f,0.078459218f}, {-0.99306846f,0.11753753f},
+{-0.98768830f,0.15643461f}, {-0.98078525f,0.19509049f}, {-0.97236985f,0.23344554f},
+{-0.96245515f,0.27144065f}, {-0.95105654f,0.30901697f}, {-0.93819135f,0.34611705f},
+{-0.92387956f,0.38268346f}, {-0.90814316f,0.41865975f}, {-0.89100653f,0.45399055f},
+{-0.87249601f,0.48862129f}, {-0.85264015f,0.52249861f}, {-0.83146960f,0.55557030f},
+{-0.80901694f,0.58778536f}, {-0.78531688f,0.61909401f}, {-0.76040590f,0.64944816f},
+{-0.73432243f,0.67880082f}, {-0.70710665f,0.70710689f}, {-0.67880058f,0.73432261f},
+{-0.64944792f,0.76040608f}, {-0.61909378f,0.78531706f}, {-0.58778507f,0.80901712f},
+{-0.55557001f,0.83146977f}, {-0.52249837f,0.85264033f}, {-0.48862100f,0.87249613f},
+{-0.45399022f,0.89100665f}, {-0.41865945f,0.90814328f}, {-0.38268313f,0.92387968f},
+{-0.34611672f,0.93819147f}, {-0.30901709f,0.95105648f}, {-0.27144054f,0.96245521f},
+{-0.23344545f,0.97236991f}, {-0.19509038f,0.98078525f}, {-0.15643452f,0.98768830f},
+{-0.11753743f,0.99306846f}, {-0.078459114f,0.99691731f}, {-0.039259821f,0.99922901f},
+};
+static const ne10_fft_cpx_float32_t ne10_twiddles_240[240] = {
+{1.0000000f,0.0000000f}, {1.0000000f,-0.0000000f}, {1.0000000f,-0.0000000f},
+{1.0000000f,-0.0000000f}, {0.91354543f,-0.40673664f}, {0.66913056f,-0.74314487f},
+{1.0000000f,-0.0000000f}, {0.66913056f,-0.74314487f}, {-0.10452851f,-0.99452192f},
+{1.0000000f,-0.0000000f}, {0.30901697f,-0.95105654f}, {-0.80901700f,-0.58778518f},
+{1.0000000f,-0.0000000f}, {-0.10452851f,-0.99452192f}, {-0.97814757f,0.20791179f},
+{1.0000000f,-0.0000000f}, {0.99452192f,-0.10452846f}, {0.97814763f,-0.20791170f},
+{0.95105648f,-0.30901700f}, {0.91354543f,-0.40673664f}, {0.86602545f,-0.50000000f},
+{0.80901700f,-0.58778524f}, {0.74314475f,-0.66913062f}, {0.66913056f,-0.74314487f},
+{0.58778524f,-0.80901700f}, {0.49999997f,-0.86602545f}, {0.40673661f,-0.91354549f},
+{0.30901697f,-0.95105654f}, {0.20791166f,-0.97814763f}, {0.10452842f,-0.99452192f},
+{1.0000000f,-0.0000000f}, {0.97814763f,-0.20791170f}, {0.91354543f,-0.40673664f},
+{0.80901700f,-0.58778524f}, {0.66913056f,-0.74314487f}, {0.49999997f,-0.86602545f},
+{0.30901697f,-0.95105654f}, {0.10452842f,-0.99452192f}, {-0.10452851f,-0.99452192f},
+{-0.30901703f,-0.95105648f}, {-0.50000006f,-0.86602533f}, {-0.66913068f,-0.74314475f},
+{-0.80901700f,-0.58778518f}, {-0.91354549f,-0.40673658f}, {-0.97814763f,-0.20791161f},
+{1.0000000f,-0.0000000f}, {0.95105648f,-0.30901700f}, {0.80901700f,-0.58778524f},
+{0.58778524f,-0.80901700f}, {0.30901697f,-0.95105654f}, {-4.3711388e-08f,-1.0000000f},
+{-0.30901703f,-0.95105648f}, {-0.58778518f,-0.80901700f}, {-0.80901700f,-0.58778518f},
+{-0.95105654f,-0.30901679f}, {-1.0000000f,8.7422777e-08f}, {-0.95105654f,0.30901697f},
+{-0.80901694f,0.58778536f}, {-0.58778507f,0.80901712f}, {-0.30901709f,0.95105648f},
+{1.0000000f,-0.0000000f}, {0.99965733f,-0.026176950f}, {0.99862951f,-0.052335959f},
+{0.99691731f,-0.078459099f}, {0.99452192f,-0.10452846f}, {0.99144489f,-0.13052620f},
+{0.98768836f,-0.15643448f}, {0.98325491f,-0.18223552f}, {0.97814763f,-0.20791170f},
+{0.97236991f,-0.23344538f}, {0.96592581f,-0.25881904f}, {0.95881975f,-0.28401536f},
+{0.95105648f,-0.30901700f}, {0.94264150f,-0.33380687f}, {0.93358040f,-0.35836795f},
+{0.92387956f,-0.38268346f}, {0.91354543f,-0.40673664f}, {0.90258527f,-0.43051112f},
+{0.89100653f,-0.45399052f}, {0.87881708f,-0.47715878f}, {0.86602545f,-0.50000000f},
+{0.85264015f,-0.52249855f}, {0.83867055f,-0.54463905f}, {0.82412618f,-0.56640625f},
+{0.80901700f,-0.58778524f}, {0.79335332f,-0.60876143f}, {0.77714598f,-0.62932038f},
+{0.76040596f,-0.64944810f}, {0.74314475f,-0.66913062f}, {0.72537434f,-0.68835455f},
+{0.70710677f,-0.70710683f}, {0.68835455f,-0.72537440f}, {0.66913056f,-0.74314487f},
+{0.64944804f,-0.76040596f}, {0.62932038f,-0.77714598f}, {0.60876137f,-0.79335338f},
+{0.58778524f,-0.80901700f}, {0.56640625f,-0.82412618f}, {0.54463899f,-0.83867055f},
+{0.52249849f,-0.85264015f}, {0.49999997f,-0.86602545f}, {0.47715876f,-0.87881708f},
+{0.45399052f,-0.89100653f}, {0.43051103f,-0.90258533f}, {0.40673661f,-0.91354549f},
+{0.38268343f,-0.92387956f}, {0.35836786f,-0.93358046f}, {0.33380681f,-0.94264150f},
+{0.30901697f,-0.95105654f}, {0.28401533f,-0.95881975f}, {0.25881907f,-0.96592581f},
+{0.23344530f,-0.97236991f}, {0.20791166f,-0.97814763f}, {0.18223552f,-0.98325491f},
+{0.15643437f,-0.98768836f}, {0.13052613f,-0.99144489f}, {0.10452842f,-0.99452192f},
+{0.078459084f,-0.99691731f}, {0.052335974f,-0.99862951f}, {0.026176875f,-0.99965733f},
+{1.0000000f,-0.0000000f}, {0.99862951f,-0.052335959f}, {0.99452192f,-0.10452846f},
+{0.98768836f,-0.15643448f}, {0.97814763f,-0.20791170f}, {0.96592581f,-0.25881904f},
+{0.95105648f,-0.30901700f}, {0.93358040f,-0.35836795f}, {0.91354543f,-0.40673664f},
+{0.89100653f,-0.45399052f}, {0.86602545f,-0.50000000f}, {0.83867055f,-0.54463905f},
+{0.80901700f,-0.58778524f}, {0.77714598f,-0.62932038f}, {0.74314475f,-0.66913062f},
+{0.70710677f,-0.70710683f}, {0.66913056f,-0.74314487f}, {0.62932038f,-0.77714598f},
+{0.58778524f,-0.80901700f}, {0.54463899f,-0.83867055f}, {0.49999997f,-0.86602545f},
+{0.45399052f,-0.89100653f}, {0.40673661f,-0.91354549f}, {0.35836786f,-0.93358046f},
+{0.30901697f,-0.95105654f}, {0.25881907f,-0.96592581f}, {0.20791166f,-0.97814763f},
+{0.15643437f,-0.98768836f}, {0.10452842f,-0.99452192f}, {0.052335974f,-0.99862951f},
+{-4.3711388e-08f,-1.0000000f}, {-0.052336060f,-0.99862951f}, {-0.10452851f,-0.99452192f},
+{-0.15643445f,-0.98768836f}, {-0.20791174f,-0.97814757f}, {-0.25881916f,-0.96592581f},
+{-0.30901703f,-0.95105648f}, {-0.35836795f,-0.93358040f}, {-0.40673670f,-0.91354543f},
+{-0.45399061f,-0.89100647f}, {-0.50000006f,-0.86602533f}, {-0.54463905f,-0.83867055f},
+{-0.58778518f,-0.80901700f}, {-0.62932050f,-0.77714586f}, {-0.66913068f,-0.74314475f},
+{-0.70710677f,-0.70710677f}, {-0.74314493f,-0.66913044f}, {-0.77714604f,-0.62932026f},
+{-0.80901700f,-0.58778518f}, {-0.83867055f,-0.54463899f}, {-0.86602539f,-0.50000006f},
+{-0.89100659f,-0.45399037f}, {-0.91354549f,-0.40673658f}, {-0.93358040f,-0.35836792f},
+{-0.95105654f,-0.30901679f}, {-0.96592587f,-0.25881892f}, {-0.97814763f,-0.20791161f},
+{-0.98768836f,-0.15643445f}, {-0.99452192f,-0.10452849f}, {-0.99862957f,-0.052335810f},
+{1.0000000f,-0.0000000f}, {0.99691731f,-0.078459099f}, {0.98768836f,-0.15643448f},
+{0.97236991f,-0.23344538f}, {0.95105648f,-0.30901700f}, {0.92387956f,-0.38268346f},
+{0.89100653f,-0.45399052f}, {0.85264015f,-0.52249855f}, {0.80901700f,-0.58778524f},
+{0.76040596f,-0.64944810f}, {0.70710677f,-0.70710683f}, {0.64944804f,-0.76040596f},
+{0.58778524f,-0.80901700f}, {0.52249849f,-0.85264015f}, {0.45399052f,-0.89100653f},
+{0.38268343f,-0.92387956f}, {0.30901697f,-0.95105654f}, {0.23344530f,-0.97236991f},
+{0.15643437f,-0.98768836f}, {0.078459084f,-0.99691731f}, {-4.3711388e-08f,-1.0000000f},
+{-0.078459173f,-0.99691731f}, {-0.15643445f,-0.98768836f}, {-0.23344538f,-0.97236991f},
+{-0.30901703f,-0.95105648f}, {-0.38268352f,-0.92387950f}, {-0.45399061f,-0.89100647f},
+{-0.52249867f,-0.85264009f}, {-0.58778518f,-0.80901700f}, {-0.64944804f,-0.76040596f},
+{-0.70710677f,-0.70710677f}, {-0.76040596f,-0.64944804f}, {-0.80901700f,-0.58778518f},
+{-0.85264021f,-0.52249849f}, {-0.89100659f,-0.45399037f}, {-0.92387956f,-0.38268328f},
+{-0.95105654f,-0.30901679f}, {-0.97236991f,-0.23344538f}, {-0.98768836f,-0.15643445f},
+{-0.99691737f,-0.078459039f}, {-1.0000000f,8.7422777e-08f}, {-0.99691731f,0.078459218f},
+{-0.98768830f,0.15643461f}, {-0.97236985f,0.23344554f}, {-0.95105654f,0.30901697f},
+{-0.92387956f,0.38268346f}, {-0.89100653f,0.45399055f}, {-0.85264015f,0.52249861f},
+{-0.80901694f,0.58778536f}, {-0.76040590f,0.64944816f}, {-0.70710665f,0.70710689f},
+{-0.64944792f,0.76040608f}, {-0.58778507f,0.80901712f}, {-0.52249837f,0.85264033f},
+{-0.45399022f,0.89100665f}, {-0.38268313f,0.92387968f}, {-0.30901709f,0.95105648f},
+{-0.23344545f,0.97236991f}, {-0.15643452f,0.98768830f}, {-0.078459114f,0.99691731f},
+};
+static const ne10_fft_cpx_float32_t ne10_twiddles_120[120] = {
+{1.0000000f,0.0000000f}, {1.0000000f,-0.0000000f}, {1.0000000f,-0.0000000f},
+{1.0000000f,-0.0000000f}, {0.91354543f,-0.40673664f}, {0.66913056f,-0.74314487f},
+{1.0000000f,-0.0000000f}, {0.66913056f,-0.74314487f}, {-0.10452851f,-0.99452192f},
+{1.0000000f,-0.0000000f}, {0.30901697f,-0.95105654f}, {-0.80901700f,-0.58778518f},
+{1.0000000f,-0.0000000f}, {-0.10452851f,-0.99452192f}, {-0.97814757f,0.20791179f},
+{1.0000000f,-0.0000000f}, {0.97814763f,-0.20791170f}, {0.91354543f,-0.40673664f},
+{0.80901700f,-0.58778524f}, {0.66913056f,-0.74314487f}, {0.49999997f,-0.86602545f},
+{0.30901697f,-0.95105654f}, {0.10452842f,-0.99452192f}, {-0.10452851f,-0.99452192f},
+{-0.30901703f,-0.95105648f}, {-0.50000006f,-0.86602533f}, {-0.66913068f,-0.74314475f},
+{-0.80901700f,-0.58778518f}, {-0.91354549f,-0.40673658f}, {-0.97814763f,-0.20791161f},
+{1.0000000f,-0.0000000f}, {0.99862951f,-0.052335959f}, {0.99452192f,-0.10452846f},
+{0.98768836f,-0.15643448f}, {0.97814763f,-0.20791170f}, {0.96592581f,-0.25881904f},
+{0.95105648f,-0.30901700f}, {0.93358040f,-0.35836795f}, {0.91354543f,-0.40673664f},
+{0.89100653f,-0.45399052f}, {0.86602545f,-0.50000000f}, {0.83867055f,-0.54463905f},
+{0.80901700f,-0.58778524f}, {0.77714598f,-0.62932038f}, {0.74314475f,-0.66913062f},
+{0.70710677f,-0.70710683f}, {0.66913056f,-0.74314487f}, {0.62932038f,-0.77714598f},
+{0.58778524f,-0.80901700f}, {0.54463899f,-0.83867055f}, {0.49999997f,-0.86602545f},
+{0.45399052f,-0.89100653f}, {0.40673661f,-0.91354549f}, {0.35836786f,-0.93358046f},
+{0.30901697f,-0.95105654f}, {0.25881907f,-0.96592581f}, {0.20791166f,-0.97814763f},
+{0.15643437f,-0.98768836f}, {0.10452842f,-0.99452192f}, {0.052335974f,-0.99862951f},
+{1.0000000f,-0.0000000f}, {0.99452192f,-0.10452846f}, {0.97814763f,-0.20791170f},
+{0.95105648f,-0.30901700f}, {0.91354543f,-0.40673664f}, {0.86602545f,-0.50000000f},
+{0.80901700f,-0.58778524f}, {0.74314475f,-0.66913062f}, {0.66913056f,-0.74314487f},
+{0.58778524f,-0.80901700f}, {0.49999997f,-0.86602545f}, {0.40673661f,-0.91354549f},
+{0.30901697f,-0.95105654f}, {0.20791166f,-0.97814763f}, {0.10452842f,-0.99452192f},
+{-4.3711388e-08f,-1.0000000f}, {-0.10452851f,-0.99452192f}, {-0.20791174f,-0.97814757f},
+{-0.30901703f,-0.95105648f}, {-0.40673670f,-0.91354543f}, {-0.50000006f,-0.86602533f},
+{-0.58778518f,-0.80901700f}, {-0.66913068f,-0.74314475f}, {-0.74314493f,-0.66913044f},
+{-0.80901700f,-0.58778518f}, {-0.86602539f,-0.50000006f}, {-0.91354549f,-0.40673658f},
+{-0.95105654f,-0.30901679f}, {-0.97814763f,-0.20791161f}, {-0.99452192f,-0.10452849f},
+{1.0000000f,-0.0000000f}, {0.98768836f,-0.15643448f}, {0.95105648f,-0.30901700f},
+{0.89100653f,-0.45399052f}, {0.80901700f,-0.58778524f}, {0.70710677f,-0.70710683f},
+{0.58778524f,-0.80901700f}, {0.45399052f,-0.89100653f}, {0.30901697f,-0.95105654f},
+{0.15643437f,-0.98768836f}, {-4.3711388e-08f,-1.0000000f}, {-0.15643445f,-0.98768836f},
+{-0.30901703f,-0.95105648f}, {-0.45399061f,-0.89100647f}, {-0.58778518f,-0.80901700f},
+{-0.70710677f,-0.70710677f}, {-0.80901700f,-0.58778518f}, {-0.89100659f,-0.45399037f},
+{-0.95105654f,-0.30901679f}, {-0.98768836f,-0.15643445f}, {-1.0000000f,8.7422777e-08f},
+{-0.98768830f,0.15643461f}, {-0.95105654f,0.30901697f}, {-0.89100653f,0.45399055f},
+{-0.80901694f,0.58778536f}, {-0.70710665f,0.70710689f}, {-0.58778507f,0.80901712f},
+{-0.45399022f,0.89100665f}, {-0.30901709f,0.95105648f}, {-0.15643452f,0.98768830f},
+};
+static const ne10_fft_cpx_float32_t ne10_twiddles_60[60] = {
+{1.0000000f,0.0000000f}, {1.0000000f,-0.0000000f}, {1.0000000f,-0.0000000f},
+{1.0000000f,-0.0000000f}, {0.91354543f,-0.40673664f}, {0.66913056f,-0.74314487f},
+{1.0000000f,-0.0000000f}, {0.66913056f,-0.74314487f}, {-0.10452851f,-0.99452192f},
+{1.0000000f,-0.0000000f}, {0.30901697f,-0.95105654f}, {-0.80901700f,-0.58778518f},
+{1.0000000f,-0.0000000f}, {-0.10452851f,-0.99452192f}, {-0.97814757f,0.20791179f},
+{1.0000000f,-0.0000000f}, {0.99452192f,-0.10452846f}, {0.97814763f,-0.20791170f},
+{0.95105648f,-0.30901700f}, {0.91354543f,-0.40673664f}, {0.86602545f,-0.50000000f},
+{0.80901700f,-0.58778524f}, {0.74314475f,-0.66913062f}, {0.66913056f,-0.74314487f},
+{0.58778524f,-0.80901700f}, {0.49999997f,-0.86602545f}, {0.40673661f,-0.91354549f},
+{0.30901697f,-0.95105654f}, {0.20791166f,-0.97814763f}, {0.10452842f,-0.99452192f},
+{1.0000000f,-0.0000000f}, {0.97814763f,-0.20791170f}, {0.91354543f,-0.40673664f},
+{0.80901700f,-0.58778524f}, {0.66913056f,-0.74314487f}, {0.49999997f,-0.86602545f},
+{0.30901697f,-0.95105654f}, {0.10452842f,-0.99452192f}, {-0.10452851f,-0.99452192f},
+{-0.30901703f,-0.95105648f}, {-0.50000006f,-0.86602533f}, {-0.66913068f,-0.74314475f},
+{-0.80901700f,-0.58778518f}, {-0.91354549f,-0.40673658f}, {-0.97814763f,-0.20791161f},
+{1.0000000f,-0.0000000f}, {0.95105648f,-0.30901700f}, {0.80901700f,-0.58778524f},
+{0.58778524f,-0.80901700f}, {0.30901697f,-0.95105654f}, {-4.3711388e-08f,-1.0000000f},
+{-0.30901703f,-0.95105648f}, {-0.58778518f,-0.80901700f}, {-0.80901700f,-0.58778518f},
+{-0.95105654f,-0.30901679f}, {-1.0000000f,8.7422777e-08f}, {-0.95105654f,0.30901697f},
+{-0.80901694f,0.58778536f}, {-0.58778507f,0.80901712f}, {-0.30901709f,0.95105648f},
+};
+static const ne10_fft_state_float32_t ne10_fft_state_float32_t_480 = {
+120,
+(ne10_int32_t *)ne10_factors_480,
+(ne10_fft_cpx_float32_t *)ne10_twiddles_480,
+NULL,
+(ne10_fft_cpx_float32_t *)&ne10_twiddles_480[120],
+/* is_forward_scaled = true */
+(ne10_int32_t) 1,
+/* is_backward_scaled = false */
+(ne10_int32_t) 0,
+};
+static const arch_fft_state cfg_arch_480 = {
+1,
+(void *)&ne10_fft_state_float32_t_480,
+};
+
+static const ne10_fft_state_float32_t ne10_fft_state_float32_t_240 = {
+60,
+(ne10_int32_t *)ne10_factors_240,
+(ne10_fft_cpx_float32_t *)ne10_twiddles_240,
+NULL,
+(ne10_fft_cpx_float32_t *)&ne10_twiddles_240[60],
+/* is_forward_scaled = true */
+(ne10_int32_t) 1,
+/* is_backward_scaled = false */
+(ne10_int32_t) 0,
+};
+static const arch_fft_state cfg_arch_240 = {
+1,
+(void *)&ne10_fft_state_float32_t_240,
+};
+
+static const ne10_fft_state_float32_t ne10_fft_state_float32_t_120 = {
+30,
+(ne10_int32_t *)ne10_factors_120,
+(ne10_fft_cpx_float32_t *)ne10_twiddles_120,
+NULL,
+(ne10_fft_cpx_float32_t *)&ne10_twiddles_120[30],
+/* is_forward_scaled = true */
+(ne10_int32_t) 1,
+/* is_backward_scaled = false */
+(ne10_int32_t) 0,
+};
+static const arch_fft_state cfg_arch_120 = {
+1,
+(void *)&ne10_fft_state_float32_t_120,
+};
+
+static const ne10_fft_state_float32_t ne10_fft_state_float32_t_60 = {
+15,
+(ne10_int32_t *)ne10_factors_60,
+(ne10_fft_cpx_float32_t *)ne10_twiddles_60,
+NULL,
+(ne10_fft_cpx_float32_t *)&ne10_twiddles_60[15],
+/* is_forward_scaled = true */
+(ne10_int32_t) 1,
+/* is_backward_scaled = false */
+(ne10_int32_t) 0,
+};
+static const arch_fft_state cfg_arch_60 = {
+1,
+(void *)&ne10_fft_state_float32_t_60,
+};
+
+#endif  /* end NE10_FFT_PARAMS48000_960 */
--- a/lib/libopus/celt/vq.c
+++ b/lib/libopus/celt/vq.c
@ -0,0 +1,867 @@
+/* Copyright (c) 2007-2008 CSIRO
+   Copyright (c) 2007-2009 Xiph.Org Foundation
+   Written by Jean-Marc Valin */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "mathops.h"
+#include "cwrs.h"
+#include "vq.h"
+#include "arch.h"
+#include "os_support.h"
+#include "bands.h"
+#include "rate.h"
+#include "pitch.h"
+#include "SigProc_FIX.h"
+
+#if defined(FIXED_POINT)
+void norm_scaleup(celt_norm *X, int N, int shift) {
+   int i;
+   celt_assert(shift >= 0);
+   if (shift <= 0) return;
+   for (i=0;i<N;i++) X[i] = SHL32(X[i], shift);
+}
+
+void norm_scaledown(celt_norm *X, int N, int shift) {
+   int i;
+   celt_assert(shift >= 0);
+   if (shift <= 0) return;
+   for (i=0;i<N;i++) X[i] = PSHR32(X[i], shift);
+}
+
+opus_val32 celt_inner_prod_norm(const celt_norm *x, const celt_norm *y, int len, int arch) {
+   int i;
+   opus_val32 sum = 0;
+   (void)arch;
+   for (i=0;i<len;i++) sum += x[i]*y[i];
+   return sum;
+}
+opus_val32 celt_inner_prod_norm_shift(const celt_norm *x, const celt_norm *y, int len, int arch) {
+   int i;
+   opus_val64 sum = 0;
+   (void)arch;
+   for (i=0;i<len;i++) sum += x[i]*(opus_val64)y[i];
+   return sum>>2*(NORM_SHIFT-14);
+}
+#endif
+
+#ifndef OVERRIDE_vq_exp_rotation1
+static void exp_rotation1(celt_norm *X, int len, int stride, opus_val16 c, opus_val16 s)
+{
+   int i;
+   opus_val16 ms;
+   celt_norm *Xptr;
+   Xptr = X;
+   ms = NEG16(s);
+   norm_scaledown(X, len, NORM_SHIFT-14);
+   for (i=0;i<len-stride;i++)
+   {
+      celt_norm x1, x2;
+      x1 = Xptr[0];
+      x2 = Xptr[stride];
+      Xptr[stride] = EXTRACT16(PSHR32(MAC16_16(MULT16_16(c, x2),  s, x1), 15));
+      *Xptr++      = EXTRACT16(PSHR32(MAC16_16(MULT16_16(c, x1), ms, x2), 15));
+   }
+   Xptr = &X[len-2*stride-1];
+   for (i=len-2*stride-1;i>=0;i--)
+   {
+      celt_norm x1, x2;
+      x1 = Xptr[0];
+      x2 = Xptr[stride];
+      Xptr[stride] = EXTRACT16(PSHR32(MAC16_16(MULT16_16(c, x2),  s, x1), 15));
+      *Xptr--      = EXTRACT16(PSHR32(MAC16_16(MULT16_16(c, x1), ms, x2), 15));
+   }
+   norm_scaleup(X, len, NORM_SHIFT-14);
+}
+#endif /* OVERRIDE_vq_exp_rotation1 */
+
+void exp_rotation(celt_norm *X, int len, int dir, int stride, int K, int spread)
+{
+   static const int SPREAD_FACTOR[3]={15,10,5};
+   int i;
+   opus_val16 c, s;
+   opus_val16 gain, theta;
+   int stride2=0;
+   int factor;
+
+   if (2*K>=len || spread==SPREAD_NONE)
+      return;
+   factor = SPREAD_FACTOR[spread-1];
+
+   gain = celt_div((opus_val32)MULT16_16(Q15_ONE,len),(opus_val32)(len+factor*K));
+   theta = HALF16(MULT16_16_Q15(gain,gain));
+
+   c = celt_cos_norm(EXTEND32(theta));
+   s = celt_cos_norm(EXTEND32(SUB16(Q15ONE,theta))); /*  sin(theta) */
+
+   if (len>=8*stride)
+   {
+      stride2 = 1;
+      /* This is just a simple (equivalent) way of computing sqrt(len/stride) with rounding.
+         It's basically incrementing long as (stride2+0.5)^2 < len/stride. */
+      while ((stride2*stride2+stride2)*stride + (stride>>2) < len)
+         stride2++;
+   }
+   /*NOTE: As a minor optimization, we could be passing around log2(B), not B, for both this and for
+      extract_collapse_mask().*/
+   len = celt_udiv(len, stride);
+   for (i=0;i<stride;i++)
+   {
+      if (dir < 0)
+      {
+         if (stride2)
+            exp_rotation1(X+i*len, len, stride2, s, c);
+         exp_rotation1(X+i*len, len, 1, c, s);
+      } else {
+         exp_rotation1(X+i*len, len, 1, c, -s);
+         if (stride2)
+            exp_rotation1(X+i*len, len, stride2, s, -c);
+      }
+   }
+}
+
+/** Normalizes the decoded integer pvq codeword to unit norm. */
+static void normalise_residual(int * OPUS_RESTRICT iy, celt_norm * OPUS_RESTRICT X,
+      int N, opus_val32 Ryy, opus_val32 gain, int shift)
+{
+   int i;
+#ifdef FIXED_POINT
+   int k;
+#endif
+   opus_val32 t;
+   opus_val32 g;
+
+#ifdef FIXED_POINT
+   k = celt_ilog2(Ryy)>>1;
+#endif
+   t = VSHR32(Ryy, 2*(k-7)-15);
+   g = MULT32_32_Q31(celt_rsqrt_norm32(t),gain);
+   i=0;
+   (void)shift;
+#if defined(FIXED_POINT) && defined(ENABLE_QEXT)
+   if (shift>0) {
+      int tot_shift = NORM_SHIFT+1-k-shift;
+      if (tot_shift >= 0) {
+         do X[i] = MULT32_32_Q31(g, SHL32(iy[i], tot_shift));
+         while (++i < N);
+      } else {
+         do X[i] = MULT32_32_Q31(g, PSHR32(iy[i], -tot_shift));
+         while (++i < N);
+      }
+   } else
+#endif
+   do X[i] = VSHR32(MULT16_32_Q15(iy[i], g), k+15-NORM_SHIFT);
+   while (++i < N);
+}
+
+static unsigned extract_collapse_mask(int *iy, int N, int B)
+{
+   unsigned collapse_mask;
+   int N0;
+   int i;
+   if (B<=1)
+      return 1;
+   /*NOTE: As a minor optimization, we could be passing around log2(B), not B, for both this and for
+      exp_rotation().*/
+   N0 = celt_udiv(N, B);
+   collapse_mask = 0;
+   i=0; do {
+      int j;
+      unsigned tmp=0;
+      j=0; do {
+         tmp |= iy[i*N0+j];
+      } while (++j<N0);
+      collapse_mask |= (tmp!=0)<<i;
+   } while (++i<B);
+   return collapse_mask;
+}
+
+opus_val16 op_pvq_search_c(celt_norm *X, int *iy, int K, int N, int arch)
+{
+   VARDECL(celt_norm, y);
+   VARDECL(int, signx);
+   int i, j;
+   int pulsesLeft;
+   opus_val32 sum;
+   opus_val32 xy;
+   opus_val16 yy;
+   SAVE_STACK;
+
+   (void)arch;
+   ALLOC(y, N, celt_norm);
+   ALLOC(signx, N, int);
+#ifdef FIXED_POINT
+   {
+      int shift = (celt_ilog2(1+celt_inner_prod_norm_shift(X, X, N, arch))+1)/2;
+      shift = IMAX(0, shift+(NORM_SHIFT-14)-14);
+      norm_scaledown(X, N, shift);
+   }
+#endif
+   /* Get rid of the sign */
+   sum = 0;
+   j=0; do {
+      signx[j] = X[j]<0;
+      /* OPT: Make sure the compiler doesn't use a branch on ABS16(). */
+      X[j] = ABS16(X[j]);
+      iy[j] = 0;
+      y[j] = 0;
+   } while (++j<N);
+
+   xy = yy = 0;
+
+   pulsesLeft = K;
+
+   /* Do a pre-search by projecting on the pyramid */
+   if (K > (N>>1))
+   {
+      opus_val16 rcp;
+      j=0; do {
+         sum += X[j];
+      }  while (++j<N);
+
+      /* If X is too small, just replace it with a pulse at 0 */
+#ifdef FIXED_POINT
+      if (sum <= K)
+#else
+      /* Prevents infinities and NaNs from causing too many pulses
+         to be allocated. 64 is an approximation of infinity here. */
+      if (!(sum > EPSILON && sum < 64))
+#endif
+      {
+         X[0] = QCONST16(1.f,14);
+         j=1; do
+            X[j]=0;
+         while (++j<N);
+         sum = QCONST16(1.f,14);
+      }
+#ifdef FIXED_POINT
+      rcp = EXTRACT16(MULT16_32_Q16(K, celt_rcp(sum)));
+#else
+      /* Using K+e with e < 1 guarantees we cannot get more than K pulses. */
+      rcp = EXTRACT16(MULT16_32_Q16(K+0.8f, celt_rcp(sum)));
+#endif
+      j=0; do {
+#ifdef FIXED_POINT
+         /* It's really important to round *towards zero* here */
+         iy[j] = MULT16_16_Q15(X[j],rcp);
+#else
+         iy[j] = (int)floor(rcp*X[j]);
+#endif
+         y[j] = (celt_norm)iy[j];
+         yy = MAC16_16(yy, y[j],y[j]);
+         xy = MAC16_16(xy, X[j],y[j]);
+         y[j] *= 2;
+         pulsesLeft -= iy[j];
+      }  while (++j<N);
+   }
+   celt_sig_assert(pulsesLeft>=0);
+
+   /* This should never happen, but just in case it does (e.g. on silence)
+      we fill the first bin with pulses. */
+#ifdef FIXED_POINT_DEBUG
+   celt_sig_assert(pulsesLeft<=N+3);
+#endif
+   if (pulsesLeft > N+3)
+   {
+      opus_val16 tmp = (opus_val16)pulsesLeft;
+      yy = MAC16_16(yy, tmp, tmp);
+      yy = MAC16_16(yy, tmp, y[0]);
+      iy[0] += pulsesLeft;
+      pulsesLeft=0;
+   }
+
+   for (i=0;i<pulsesLeft;i++)
+   {
+      opus_val16 Rxy, Ryy;
+      int best_id;
+      opus_val32 best_num;
+      opus_val16 best_den;
+#ifdef FIXED_POINT
+      int rshift;
+#endif
+#ifdef FIXED_POINT
+      rshift = 1+celt_ilog2(K-pulsesLeft+i+1);
+#endif
+      best_id = 0;
+      /* The squared magnitude term gets added anyway, so we might as well
+         add it outside the loop */
+      yy = ADD16(yy, 1);
+
+      /* Calculations for position 0 are out of the loop, in part to reduce
+         mispredicted branches (since the if condition is usually false)
+         in the loop. */
+      /* Temporary sums of the new pulse(s) */
+      Rxy = EXTRACT16(SHR32(ADD32(xy, EXTEND32(X[0])),rshift));
+      /* We're multiplying y[j] by two so we don't have to do it here */
+      Ryy = ADD16(yy, y[0]);
+
+      /* Approximate score: we maximise Rxy/sqrt(Ryy) (we're guaranteed that
+         Rxy is positive because the sign is pre-computed) */
+      Rxy = MULT16_16_Q15(Rxy,Rxy);
+      best_den = Ryy;
+      best_num = Rxy;
+      j=1;
+      do {
+         /* Temporary sums of the new pulse(s) */
+         Rxy = EXTRACT16(SHR32(ADD32(xy, EXTEND32(X[j])),rshift));
+         /* We're multiplying y[j] by two so we don't have to do it here */
+         Ryy = ADD16(yy, y[j]);
+
+         /* Approximate score: we maximise Rxy/sqrt(Ryy) (we're guaranteed that
+            Rxy is positive because the sign is pre-computed) */
+         Rxy = MULT16_16_Q15(Rxy,Rxy);
+         /* The idea is to check for num/den >= best_num/best_den, but that way
+            we can do it without any division */
+         /* OPT: It's not clear whether a cmov is faster than a branch here
+            since the condition is more often false than true and using
+            a cmov introduces data dependencies across iterations. The optimal
+            choice may be architecture-dependent. */
+         if (opus_unlikely(MULT16_16(best_den, Rxy) > MULT16_16(Ryy, best_num)))
+         {
+            best_den = Ryy;
+            best_num = Rxy;
+            best_id = j;
+         }
+      } while (++j<N);
+
+      /* Updating the sums of the new pulse(s) */
+      xy = ADD32(xy, EXTEND32(X[best_id]));
+      /* We're multiplying y[j] by two so we don't have to do it here */
+      yy = ADD16(yy, y[best_id]);
+
+      /* Only now that we've made the final choice, update y/iy */
+      /* Multiplying y[j] by 2 so we don't have to do it everywhere else */
+      y[best_id] += 2;
+      iy[best_id]++;
+   }
+
+   /* Put the original sign back */
+   j=0;
+   do {
+      /*iy[j] = signx[j] ? -iy[j] : iy[j];*/
+      /* OPT: The is more likely to be compiled without a branch than the code above
+         but has the same performance otherwise. */
+      iy[j] = (iy[j]^-signx[j]) + signx[j];
+   } while (++j<N);
+   RESTORE_STACK;
+   return yy;
+}
+
+#ifdef ENABLE_QEXT
+#include "macros.h"
+
+static opus_val32 op_pvq_search_N2(const celt_norm *X, int *iy, int *up_iy, int K, int up, int *refine, int shift) {
+   opus_val32 sum;
+   opus_val32 rcp_sum;
+   int offset;
+   sum = ABS32(X[0]) + ABS32(X[1]);
+   if (sum < EPSILON) {
+      iy[0] = K;
+      up_iy[0] = up*K;
+      iy[1]=up_iy[1]=0;
+      *refine=0;
+#ifdef FIXED_POINT
+      return (opus_val64)K*K*up*up>>2*shift;
+#else
+      (void)shift;
+      return K*(float)K*up*up;
+#endif
+   }
+#ifdef FIXED_POINT
+   int sum_shift;
+   opus_val32 X0;
+   sum_shift = 30-celt_ilog2(sum);
+   rcp_sum = celt_rcp_norm32(SHL32(sum, sum_shift));
+   X0 = MULT32_32_Q31(SHL32(X[0], sum_shift), rcp_sum);
+   iy[0] = PSHR32(MULT32_32_Q31(SHL32(K, 8), X0), 7);
+   up_iy[0] = PSHR32(MULT32_32_Q31(SHL32(up*K, 8), X0), 7);
+#else
+   rcp_sum = 1.f/sum;
+   iy[0] = (int)floor(.5f+K*X[0]*rcp_sum);
+   up_iy[0] = (int)floor(.5f+up*K*X[0]*rcp_sum);
+#endif
+   up_iy[0] = IMAX(up*iy[0] - (up-1)/2, IMIN(up*iy[0] + (up-1)/2, up_iy[0]));
+   offset = up_iy[0] - up*iy[0];
+   iy[1] = K-abs(iy[0]);
+   up_iy[1] = up*K-abs(up_iy[0]);
+   if (X[1] < 0) {
+      iy[1] = -iy[1];
+      up_iy[1] = -up_iy[1];
+      offset = -offset;
+   }
+   *refine = offset;
+#ifdef FIXED_POINT
+   return (up_iy[0]*(opus_val64)up_iy[0] + up_iy[1]*(opus_val64)up_iy[1] + (1<<2*shift>>1))>>2*shift;
+#else
+   return up_iy[0]*(opus_val64)up_iy[0] + up_iy[1]*(opus_val64)up_iy[1];
+#endif
+}
+
+static int op_pvq_refine(const opus_val32 *Xn, int *iy, int *iy0, int K, int up, int margin, int N) {
+   int i;
+   int dir;
+   VARDECL(opus_val32, rounding);
+   int iysum = 0;
+   SAVE_STACK;
+   ALLOC(rounding, N, opus_val32);
+   for (i=0;i<N;i++) {
+      opus_val32 tmp;
+      tmp = MULT32_32_Q31(SHL32(K, 8), Xn[i]);
+#ifdef FIXED_POINT
+      iy[i] = (tmp+64) >> 7;
+#else
+      iy[i] = (int)floor(.5+tmp);
+#endif
+      rounding[i] = tmp - SHL32(iy[i], 7);
+   }
+   if (iy != iy0) {
+      for (i=0;i<N;i++) iy[i] = IMIN(up*iy0[i]+up-1, IMAX(up*iy0[i]-up+1, iy[i]));
+   }
+   for (i=0;i<N;i++) iysum += iy[i];
+   if (abs(iysum - K) > 32) {
+      RESTORE_STACK;
+      return 1;
+   }
+   dir = iysum < K ? 1 : -1;
+   while (iysum != K) {
+      opus_val32 roundval=-1000000*dir;
+      int roundpos=0;
+      for (i=0;i<N;i++) {
+         if ((rounding[i]-roundval)*dir > 0 && abs(iy[i]-up*iy0[i]) < (margin-1) && !(dir==-1 && iy[i] == 0)) {
+            roundval = rounding[i];
+            roundpos = i;
+         }
+      }
+      iy[roundpos] += dir;
+      rounding[roundpos] -= SHL32(dir, 15);
+      iysum+=dir;
+   }
+   RESTORE_STACK;
+   return 0;
+}
+
+static opus_val32 op_pvq_search_extra(const celt_norm *X, int *iy, int *up_iy, int K, int up, int *refine, int N, int shift) {
+   opus_val32 rcp_sum;
+   opus_val32 sum=0;
+   int i;
+   int failed=0;
+   opus_val64 yy=0;
+   VARDECL(opus_val32, Xn);
+   SAVE_STACK;
+   for (i=0;i<N;i++) sum += ABS32(X[i]);
+   ALLOC(Xn, N, opus_val32);
+   if (sum < EPSILON)
+      failed = 1;
+   else {
+#ifdef FIXED_POINT
+      int sum_shift = 30-celt_ilog2(sum);
+      rcp_sum = celt_rcp_norm32(SHL32(sum, sum_shift));
+      for (i=0;i<N;i++) {
+         Xn[i] = MULT32_32_Q31(SHL32(ABS32(X[i]), sum_shift), rcp_sum);
+      }
+#else
+      rcp_sum = celt_rcp(sum);
+      for (i=0;i<N;i++) {
+         Xn[i] = ABS32(X[i])*rcp_sum;
+      }
+#endif
+   }
+   failed = failed || op_pvq_refine(Xn, iy, iy, K, 1, K+1, N);
+   failed = failed || op_pvq_refine(Xn, up_iy, iy, up*K, up, up, N);
+   if (failed) {
+      iy[0] = K;
+      for (i=1;i<N;i++) iy[i] = 0;
+      up_iy[0] = up*K;
+      for (i=1;i<N;i++) up_iy[i] = 0;
+   }
+   for (i=0;i<N;i++) {
+      yy += up_iy[i]*(opus_val64)up_iy[i];
+      if (X[i] < 0) {
+         iy[i] = -iy[i];
+         up_iy[i] = -up_iy[i];
+      }
+      refine[i] = up_iy[i]-up*iy[i];
+   }
+   RESTORE_STACK;
+#ifdef FIXED_POINT
+   return (yy + (1<<2*shift>>1))>>2*shift;
+#else
+   (void)shift;
+   return yy;
+#endif
+}
+#endif
+
+#ifdef ENABLE_QEXT
+/* Take advantage of the fact that "large" refine values are much less likely
+   than smaller ones. */
+static void ec_enc_refine(ec_enc *enc, opus_int32 refine, opus_int32 up, int extra_bits, int use_entropy) {
+   int large;
+   large = abs(refine)>up/2;
+   ec_enc_bit_logp(enc, large, use_entropy ? 3 : 1);
+   if (large) {
+      ec_enc_bits(enc, refine < 0, 1);
+      ec_enc_bits(enc, abs(refine)-up/2-1, extra_bits-1);
+   } else {
+      ec_enc_bits(enc, refine+up/2, extra_bits);
+   }
+}
+
+static int ec_dec_refine(ec_enc *dec, opus_int32 up, int extra_bits, int use_entropy) {
+   int large, refine;
+   large = ec_dec_bit_logp(dec, use_entropy ? 3 : 1);
+   if (large) {
+      int sign = ec_dec_bits(dec, 1);
+      refine = ec_dec_bits(dec, extra_bits-1) + up/2+1;
+      if (sign) refine = -refine;
+   } else {
+      refine = (opus_int32)ec_dec_bits(dec, extra_bits)-up/2;
+   }
+   return refine;
+}
+#endif
+
+unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc,
+      opus_val32 gain, int resynth
+      ARG_QEXT(ec_enc *ext_enc) ARG_QEXT(int extra_bits), int arch)
+{
+   VARDECL(int, iy);
+   opus_val32 yy;
+   unsigned collapse_mask;
+#ifdef ENABLE_QEXT
+   int yy_shift = 0;
+#endif
+   SAVE_STACK;
+
+   celt_assert2(K>0, "alg_quant() needs at least one pulse");
+   celt_assert2(N>1, "alg_quant() needs at least two dimensions");
+
+   /* Covers vectorization by up to 4. */
+   ALLOC(iy, N+3, int);
+
+   exp_rotation(X, N, 1, B, K, spread);
+
+#ifdef ENABLE_QEXT
+   if (N==2 && extra_bits >= 2) {
+      int refine;
+      int up_iy[2];
+      int up;
+      yy_shift = IMAX(0, extra_bits-7);
+      up = (1<<extra_bits)-1;
+      yy = op_pvq_search_N2(X, iy, up_iy, K, up, &refine, yy_shift);
+      collapse_mask = extract_collapse_mask(up_iy, N, B);
+      encode_pulses(iy, N, K, enc);
+      ec_enc_uint(ext_enc, refine+(up-1)/2, up);
+      if (resynth)
+         normalise_residual(up_iy, X, N, yy, gain, yy_shift);
+   } else if (extra_bits >= 2) {
+      int i;
+      VARDECL(int, up_iy);
+      VARDECL(int, refine);
+      int up, use_entropy;
+      ALLOC(up_iy, N, int);
+      ALLOC(refine, N, int);
+      yy_shift = IMAX(0, extra_bits-7);
+      up = (1<<extra_bits)-1;
+      yy = op_pvq_search_extra(X, iy, up_iy, K, up, refine, N, yy_shift);
+      collapse_mask = extract_collapse_mask(up_iy, N, B);
+      encode_pulses(iy, N, K, enc);
+      use_entropy = (ext_enc->storage*8 - ec_tell(ext_enc)) > (unsigned)(N-1)*(extra_bits+3)+1;
+      for (i=0;i<N-1;i++) ec_enc_refine(ext_enc, refine[i], up, extra_bits, use_entropy);
+      if (iy[N-1]==0) ec_enc_bits(ext_enc, up_iy[N-1]<0, 1);
+      if (resynth)
+         normalise_residual(up_iy, X, N, yy, gain, yy_shift);
+   } else
+#endif
+   {
+      yy = op_pvq_search(X, iy, K, N, arch);
+      collapse_mask = extract_collapse_mask(iy, N, B);
+      encode_pulses(iy, N, K, enc);
+      if (resynth)
+         normalise_residual(iy, X, N, yy, gain, 0);
+   }
+
+   if (resynth)
+      exp_rotation(X, N, -1, B, K, spread);
+
+   RESTORE_STACK;
+   return collapse_mask;
+}
+
+/** Decode pulse vector and combine the result with the pitch vector to produce
+    the final normalised signal in the current band. */
+unsigned alg_unquant(celt_norm *X, int N, int K, int spread, int B,
+      ec_dec *dec, opus_val32 gain
+      ARG_QEXT(ec_enc *ext_dec) ARG_QEXT(int extra_bits))
+{
+   opus_val32 Ryy;
+   unsigned collapse_mask;
+   VARDECL(int, iy);
+   int yy_shift=0;
+   SAVE_STACK;
+
+   celt_assert2(K>0, "alg_unquant() needs at least one pulse");
+   celt_assert2(N>1, "alg_unquant() needs at least two dimensions");
+   ALLOC(iy, N, int);
+   Ryy = decode_pulses(iy, N, K, dec);
+#ifdef ENABLE_QEXT
+   if (N==2 && extra_bits >= 2) {
+      int up;
+      int refine;
+      yy_shift = IMAX(0, extra_bits-7);
+      up = (1<<extra_bits)-1;
+      refine = (opus_int32)ec_dec_uint(ext_dec, up) - (up-1)/2;
+      iy[0] *= up;
+      iy[1] *= up;
+      if (iy[1] == 0) {
+         iy[1] = (iy[0] > 0) ? -refine : refine;
+         iy[0] += (refine*(opus_int64)iy[0] > 0) ? -refine : refine;
+      } else if (iy[1] > 0) {
+         iy[0] += refine;
+         iy[1] -= refine*(iy[0]>0?1:-1);
+      } else {
+         iy[0] -= refine;
+         iy[1] -= refine*(iy[0]>0?1:-1);
+      }
+#ifdef FIXED_POINT
+      Ryy = (iy[0]*(opus_val64)iy[0] + iy[1]*(opus_val64)iy[1] + (1<<2*yy_shift>>1)) >> 2*yy_shift;
+#else
+      Ryy = iy[0]*(opus_val64)iy[0] + iy[1]*(opus_val64)iy[1];
+#endif
+   } else if (extra_bits >= 2) {
+      int i;
+      opus_val64 yy64;
+      VARDECL(int, refine);
+      int up, use_entropy;
+      int sign=0;
+      ALLOC(refine, N, int);
+      yy_shift = IMAX(0, extra_bits-7);
+      up = (1<<extra_bits)-1;
+      use_entropy = (ext_dec->storage*8 - ec_tell(ext_dec)) > (unsigned)(N-1)*(extra_bits+3)+1;
+      for (i=0;i<N-1;i++) refine[i] = ec_dec_refine(ext_dec, up, extra_bits, use_entropy);
+      if (iy[N-1]==0) sign = ec_dec_bits(ext_dec, 1);
+      else sign = iy[N-1] < 0;
+      for (i=0;i<N-1;i++) {
+         iy[i] = iy[i]*up + refine[i];
+      }
+      iy[N-1] = up*K;
+      for (i=0;i<N-1;i++) iy[N-1] -= abs(iy[i]);
+      if (sign) iy[N-1] = -iy[N-1];
+      yy64 = 0;
+      for (i=0;i<N;i++) yy64 += iy[i]*(opus_val64)iy[i];
+#ifdef FIXED_POINT
+      Ryy = (yy64 + (1<<2*yy_shift>>1)) >> 2*yy_shift;
+#else
+      Ryy = yy64;
+#endif
+   }
+#endif
+   normalise_residual(iy, X, N, Ryy, gain, yy_shift);
+   exp_rotation(X, N, -1, B, K, spread);
+   collapse_mask = extract_collapse_mask(iy, N, B);
+   RESTORE_STACK;
+   return collapse_mask;
+}
+
+#ifndef OVERRIDE_renormalise_vector
+void renormalise_vector(celt_norm *X, int N, opus_val32 gain, int arch)
+{
+   int i;
+#ifdef FIXED_POINT
+   int k;
+#endif
+   opus_val32 E;
+   opus_val16 g;
+   opus_val32 t;
+   celt_norm *xptr;
+   norm_scaledown(X, N, NORM_SHIFT-14);
+   E = EPSILON + celt_inner_prod_norm(X, X, N, arch);
+#ifdef FIXED_POINT
+   k = celt_ilog2(E)>>1;
+#endif
+   t = VSHR32(E, 2*(k-7));
+   g = MULT32_32_Q31(celt_rsqrt_norm(t),gain);
+
+   xptr = X;
+   for (i=0;i<N;i++)
+   {
+      *xptr = EXTRACT16(PSHR32(MULT16_16(g, *xptr), k+15-14));
+      xptr++;
+   }
+   norm_scaleup(X, N, NORM_SHIFT-14);
+   /*return celt_sqrt(E);*/
+}
+#endif /* OVERRIDE_renormalise_vector */
+
+opus_int32 stereo_itheta(const celt_norm *X, const celt_norm *Y, int stereo, int N, int arch)
+{
+   int i;
+   int itheta;
+   opus_val32 mid, side;
+   opus_val32 Emid, Eside;
+
+   Emid = Eside = 0;
+   if (stereo)
+   {
+      for (i=0;i<N;i++)
+      {
+         celt_norm m, s;
+         m = PSHR32(ADD32(X[i], Y[i]), NORM_SHIFT-13);
+         s = PSHR32(SUB32(X[i], Y[i]), NORM_SHIFT-13);
+         Emid = MAC16_16(Emid, m, m);
+         Eside = MAC16_16(Eside, s, s);
+      }
+   } else {
+      Emid += celt_inner_prod_norm_shift(X, X, N, arch);
+      Eside += celt_inner_prod_norm_shift(Y, Y, N, arch);
+   }
+   mid = celt_sqrt32(Emid);
+   side = celt_sqrt32(Eside);
+#if defined(FIXED_POINT)
+   itheta = celt_atan2p_norm(side, mid);
+#else
+   itheta = (int)floor(.5f+65536.f*16384*celt_atan2p_norm(side,mid));
+#endif
+
+   return itheta;
+}
+
+#ifdef ENABLE_QEXT
+
+static void cubic_synthesis(celt_norm *X, int *iy, int N, int K, int face, int sign, opus_val32 gain) {
+   int i;
+   opus_val32 sum=0;
+   opus_val32 mag;
+#ifdef FIXED_POINT
+   int sum_shift;
+   int shift = IMAX(celt_ilog2(K) + celt_ilog2(N)/2 - 13, 0);
+#endif
+   for (i=0;i<N;i++) {
+      X[i] = (1+2*iy[i])-K;
+   }
+   X[face] = sign ? -K : K;
+   for (i=0;i<N;i++) {
+      sum += PSHR32(MULT16_16(X[i],X[i]), 2*shift);
+   }
+#ifdef FIXED_POINT
+   sum_shift = (29-celt_ilog2(sum))>>1;
+   mag = celt_rsqrt_norm32(SHL32(sum, 2*sum_shift+1));
+   for (i=0;i<N;i++) {
+      X[i] = VSHR32(MULT16_32_Q15(X[i],MULT32_32_Q31(mag,gain)), shift-sum_shift+29-NORM_SHIFT);
+   }
+#else
+   mag = 1.f/sqrt(sum);
+   for (i=0;i<N;i++) {
+      X[i] *= mag*gain;
+   }
+#endif
+}
+
+unsigned cubic_quant(celt_norm *X, int N, int res, int B, ec_enc *enc, opus_val32 gain, int resynth) {
+   int i;
+   int face=0;
+   int K;
+   VARDECL(int, iy);
+   celt_norm faceval=-1;
+   opus_val32 norm;
+   int sign;
+   SAVE_STACK;
+   ALLOC(iy, N, int);
+   K = 1<<res;
+   /* Using odd K on transients to avoid adding pre-echo. */
+   if (B!=1) K=IMAX(1, K-1);
+   if (K==1) {
+      if (resynth) OPUS_CLEAR(X, N);
+      RESTORE_STACK;
+      return 0;
+   }
+   for (i=0;i<N;i++) {
+      if (ABS32(X[i]) > faceval) {
+         faceval = ABS32(X[i]);
+         face = i;
+      }
+   }
+   sign = X[face]<0;
+   ec_enc_uint(enc, face, N);
+   ec_enc_bits(enc, sign, 1);
+#ifdef FIXED_POINT
+   if (faceval != 0) {
+      int face_shift = 30-celt_ilog2(faceval);
+      norm = celt_rcp_norm32(SHL32(faceval, face_shift));
+      norm = MULT16_32_Q15(K, norm);
+      for (i=0;i<N;i++) {
+         /* By computing X[i]+faceval inside the shift, the result is guaranteed non-negative. */
+         iy[i] = IMIN(K-1, (MULT32_32_Q31(SHL32(X[i]+faceval, face_shift-1), norm)) >> 15);
+      }
+   } else {
+      OPUS_CLEAR(iy, N);
+   }
+#else
+   norm = .5f*K/(faceval+EPSILON);
+   for (i=0;i<N;i++) {
+      iy[i] = IMIN(K-1, (int)floor((X[i]+faceval)*norm));
+   }
+#endif
+   for (i=0;i<N;i++) {
+      if (i != face) ec_enc_bits(enc, iy[i], res);
+   }
+   if (resynth) {
+      cubic_synthesis(X, iy, N, K, face, sign, gain);
+   }
+   RESTORE_STACK;
+   return (1<<B)-1;
+}
+
+unsigned cubic_unquant(celt_norm *X, int N, int res, int B, ec_dec *dec, opus_val32 gain) {
+   int i;
+   int face;
+   int sign;
+   int K;
+   VARDECL(int, iy);
+   SAVE_STACK;
+   ALLOC(iy, N, int);
+   K = 1<<res;
+   /* Using odd K on transients to avoid adding pre-echo. */
+   if (B!=1) K=IMAX(1, K-1);
+   if (K==1) {
+      OPUS_CLEAR(X, N);
+      RESTORE_STACK;
+      return 0;
+   }
+   face = ec_dec_uint(dec, N);
+   sign = ec_dec_bits(dec, 1);
+   for (i=0;i<N;i++) {
+      if (i != face) iy[i] = ec_dec_bits(dec, res);
+   }
+   iy[face]=0;
+   cubic_synthesis(X, iy, N, K, face, sign, gain);
+   RESTORE_STACK;
+   return (1<<B)-1;
+}
+#endif
--- a/lib/libopus/celt/vq.h
+++ b/lib/libopus/celt/vq.h
@ -0,0 +1,98 @@
+/* Copyright (c) 2007-2008 CSIRO
+   Copyright (c) 2007-2009 Xiph.Org Foundation
+   Written by Jean-Marc Valin */
+/**
+   @file vq.h
+   @brief Vector quantisation of the residual
+ */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef VQ_H
+#define VQ_H
+
+#include "entenc.h"
+#include "entdec.h"
+#include "modes.h"
+
+#if (defined(OPUS_X86_MAY_HAVE_SSE2) && !defined(FIXED_POINT))
+#include "x86/vq_sse.h"
+#endif
+
+#if defined(FIXED_POINT)
+opus_val32 celt_inner_prod_norm(const celt_norm *x, const celt_norm *y, int len, int arch);
+opus_val32 celt_inner_prod_norm_shift(const celt_norm *x, const celt_norm *y, int len, int arch);
+
+void norm_scaleup(celt_norm *X, int N, int shift);
+void norm_scaledown(celt_norm *X, int N, int shift);
+
+#else
+#define celt_inner_prod_norm celt_inner_prod
+#define celt_inner_prod_norm_shift celt_inner_prod
+#define norm_scaleup(X, N, shift)
+#define norm_scaledown(X, N, shift)
+#endif
+
+void exp_rotation(celt_norm *X, int len, int dir, int stride, int K, int spread);
+
+opus_val16 op_pvq_search_c(celt_norm *X, int *iy, int K, int N, int arch);
+
+#if !defined(OVERRIDE_OP_PVQ_SEARCH)
+#define op_pvq_search(x, iy, K, N, arch) \
+    (op_pvq_search_c(x, iy, K, N, arch))
+#endif
+
+/** Algebraic pulse-vector quantiser. The signal x is replaced by the sum of
+  * the pitch and a combination of pulses such that its norm is still equal
+  * to 1. This is the function that will typically require the most CPU.
+ * @param X Residual signal to quantise/encode (returns quantised version)
+ * @param N Number of samples to encode
+ * @param K Number of pulses to use
+ * @param enc Entropy encoder state
+ * @ret A mask indicating which blocks in the band received pulses
+*/
+unsigned alg_quant(celt_norm *X, int N, int K, int spread, int B, ec_enc *enc,
+      opus_val32 gain, int resynth
+      ARG_QEXT(ec_enc *ext_enc) ARG_QEXT(int extra_bits), int arch);
+
+/** Algebraic pulse decoder
+ * @param X Decoded normalised spectrum (returned)
+ * @param N Number of samples to decode
+ * @param K Number of pulses to use
+ * @param dec Entropy decoder state
+ * @ret A mask indicating which blocks in the band received pulses
+ */
+unsigned alg_unquant(celt_norm *X, int N, int K, int spread, int B,
+      ec_dec *dec, opus_val32 gain
+      ARG_QEXT(ec_enc *ext_dec) ARG_QEXT(int extra_bits));
+
+void renormalise_vector(celt_norm *X, int N, opus_val32 gain, int arch);
+
+opus_int32 stereo_itheta(const celt_norm *X, const celt_norm *Y, int stereo, int N, int arch);
+
+unsigned cubic_quant(celt_norm *X, int N, int K, int B, ec_enc *enc, opus_val32 gain, int resynth);
+unsigned cubic_unquant(celt_norm *X, int N, int K, int B, ec_dec *dec, opus_val32 gain);
+
+#endif /* VQ_H */
--- a/lib/libopus/celt/x86/celt_lpc_sse.h
+++ b/lib/libopus/celt/x86/celt_lpc_sse.h
@ -0,0 +1,67 @@
+/* Copyright (c) 2014, Cisco Systems, INC
+   Written by XiangMingZhu WeiZhou MinPeng YanWang
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef CELT_LPC_SSE_H
+#define CELT_LPC_SSE_H
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#if defined(OPUS_X86_MAY_HAVE_SSE4_1) && defined(FIXED_POINT)
+
+void celt_fir_sse4_1(
+         const opus_val16 *x,
+         const opus_val16 *num,
+         opus_val16 *y,
+         int N,
+         int ord,
+         int arch);
+
+#if defined(OPUS_X86_PRESUME_SSE4_1)
+#define OVERRIDE_CELT_FIR
+#define celt_fir(x, num, y, N, ord, arch) \
+    ((void)arch, celt_fir_sse4_1(x, num, y, N, ord, arch))
+
+#elif defined(OPUS_HAVE_RTCD)
+
+extern void (*const CELT_FIR_IMPL[OPUS_ARCHMASK + 1])(
+         const opus_val16 *x,
+         const opus_val16 *num,
+         opus_val16 *y,
+         int N,
+         int ord,
+         int arch);
+
+#define OVERRIDE_CELT_FIR
+#  define celt_fir(x, num, y, N, ord, arch) \
+    ((*CELT_FIR_IMPL[(arch) & OPUS_ARCHMASK])(x, num, y, N, ord, arch))
+
+#endif
+#endif
+
+#endif
--- a/lib/libopus/celt/x86/celt_lpc_sse4_1.c
+++ b/lib/libopus/celt/x86/celt_lpc_sse4_1.c
@ -0,0 +1,96 @@
+/* Copyright (c) 2014, Cisco Systems, INC
+   Written by XiangMingZhu WeiZhou MinPeng YanWang
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include <xmmintrin.h>
+#include <emmintrin.h>
+#include <smmintrin.h>
+#include "celt_lpc.h"
+#include "stack_alloc.h"
+#include "mathops.h"
+#include "pitch.h"
+#include "x86cpu.h"
+
+#if defined(FIXED_POINT)
+
+void celt_fir_sse4_1(const opus_val16 *x,
+         const opus_val16 *num,
+         opus_val16 *y,
+         int N,
+         int ord,
+         int arch)
+{
+    int i,j;
+    VARDECL(opus_val16, rnum);
+
+    __m128i vecNoA;
+    opus_int32 noA ;
+    SAVE_STACK;
+
+   ALLOC(rnum, ord, opus_val16);
+   for(i=0;i<ord;i++)
+      rnum[i] = num[ord-i-1];
+   noA = EXTEND32(1) << SIG_SHIFT >> 1;
+   vecNoA = _mm_set_epi32(noA, noA, noA, noA);
+
+   for (i=0;i<N-3;i+=4)
+   {
+      opus_val32 sums[4] = {0};
+      __m128i vecSum, vecX;
+#if defined(OPUS_CHECK_ASM)
+      {
+         opus_val32 sums_c[4] = {0};
+         xcorr_kernel_c(rnum, x+i-ord, sums_c, ord);
+#endif
+         xcorr_kernel(rnum, x+i-ord, sums, ord, arch);
+#if defined(OPUS_CHECK_ASM)
+         celt_assert(memcmp(sums, sums_c, sizeof(sums)) == 0);
+      }
+#endif
+      vecSum = _mm_loadu_si128((__m128i *)(void*)sums);
+      vecSum = _mm_add_epi32(vecSum, vecNoA);
+      vecSum = _mm_srai_epi32(vecSum, SIG_SHIFT);
+      vecX = OP_CVTEPI16_EPI32_M64(x + i);
+      vecSum = _mm_add_epi32(vecSum, vecX);
+      vecSum = _mm_packs_epi32(vecSum, vecSum);
+      _mm_storel_epi64((__m128i *)(void *)(y + i), vecSum);
+   }
+   for (;i<N;i++)
+   {
+      opus_val32 sum = 0;
+      for (j=0;j<ord;j++)
+         sum = MAC16_16(sum, rnum[j], x[i+j-ord]);
+      y[i] = SATURATE16(ADD32(EXTEND32(x[i]), PSHR32(sum, SIG_SHIFT)));
+   }
+
+   RESTORE_STACK;
+}
+
+#endif
--- a/lib/libopus/celt/x86/pitch_avx.c
+++ b/lib/libopus/celt/x86/pitch_avx.c
@ -0,0 +1,101 @@
+/* Copyright (c) 2023 Amazon */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+
+#include <immintrin.h>
+#include "x86cpu.h"
+#include "pitch.h"
+
+#if defined(OPUS_X86_MAY_HAVE_AVX2) && !defined(FIXED_POINT)
+
+/* Like the "regular" xcorr_kernel(), but computes 8 results at a time. */
+static void xcorr_kernel_avx(const float *x, const float *y, float sum[8], int len)
+{
+    __m256 xsum0, xsum1, xsum2, xsum3, xsum4, xsum5, xsum6, xsum7;
+    xsum7 = xsum6 = xsum5 = xsum4 = xsum3 = xsum2 = xsum1 = xsum0 = _mm256_setzero_ps();
+    int i;
+    __m256 x0;
+    /* Compute 8 inner products using partial sums. */
+    for (i=0;i<len-7;i+=8)
+    {
+        x0 = _mm256_loadu_ps(x+i);
+        xsum0 = _mm256_fmadd_ps(x0, _mm256_loadu_ps(y+i  ), xsum0);
+        xsum1 = _mm256_fmadd_ps(x0, _mm256_loadu_ps(y+i+1), xsum1);
+        xsum2 = _mm256_fmadd_ps(x0, _mm256_loadu_ps(y+i+2), xsum2);
+        xsum3 = _mm256_fmadd_ps(x0, _mm256_loadu_ps(y+i+3), xsum3);
+        xsum4 = _mm256_fmadd_ps(x0, _mm256_loadu_ps(y+i+4), xsum4);
+        xsum5 = _mm256_fmadd_ps(x0, _mm256_loadu_ps(y+i+5), xsum5);
+        xsum6 = _mm256_fmadd_ps(x0, _mm256_loadu_ps(y+i+6), xsum6);
+        xsum7 = _mm256_fmadd_ps(x0, _mm256_loadu_ps(y+i+7), xsum7);
+    }
+    if (i != len) {
+        static const int mask[15] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0, 0};
+        __m256i m;
+        m = _mm256_loadu_si256((__m256i*)(void*)(mask + 7+i-len));
+        x0 = _mm256_maskload_ps(x+i, m);
+        xsum0 = _mm256_fmadd_ps(x0, _mm256_maskload_ps(y+i  , m), xsum0);
+        xsum1 = _mm256_fmadd_ps(x0, _mm256_maskload_ps(y+i+1, m), xsum1);
+        xsum2 = _mm256_fmadd_ps(x0, _mm256_maskload_ps(y+i+2, m), xsum2);
+        xsum3 = _mm256_fmadd_ps(x0, _mm256_maskload_ps(y+i+3, m), xsum3);
+        xsum4 = _mm256_fmadd_ps(x0, _mm256_maskload_ps(y+i+4, m), xsum4);
+        xsum5 = _mm256_fmadd_ps(x0, _mm256_maskload_ps(y+i+5, m), xsum5);
+        xsum6 = _mm256_fmadd_ps(x0, _mm256_maskload_ps(y+i+6, m), xsum6);
+        xsum7 = _mm256_fmadd_ps(x0, _mm256_maskload_ps(y+i+7, m), xsum7);
+    }
+    /* 8 horizontal adds. */
+    /* Compute [0 4] [1 5] [2 6] [3 7] */
+    xsum0 = _mm256_add_ps(_mm256_permute2f128_ps(xsum0, xsum4, 2<<4), _mm256_permute2f128_ps(xsum0, xsum4, 1 | (3<<4)));
+    xsum1 = _mm256_add_ps(_mm256_permute2f128_ps(xsum1, xsum5, 2<<4), _mm256_permute2f128_ps(xsum1, xsum5, 1 | (3<<4)));
+    xsum2 = _mm256_add_ps(_mm256_permute2f128_ps(xsum2, xsum6, 2<<4), _mm256_permute2f128_ps(xsum2, xsum6, 1 | (3<<4)));
+    xsum3 = _mm256_add_ps(_mm256_permute2f128_ps(xsum3, xsum7, 2<<4), _mm256_permute2f128_ps(xsum3, xsum7, 1 | (3<<4)));
+    /* Compute [0 1 4 5] [2 3 6 7] */
+    xsum0 = _mm256_hadd_ps(xsum0, xsum1);
+    xsum1 = _mm256_hadd_ps(xsum2, xsum3);
+    /* Compute [0 1 2 3 4 5 6 7] */
+    xsum0 = _mm256_hadd_ps(xsum0, xsum1);
+    _mm256_storeu_ps(sum, xsum0);
+}
+
+void celt_pitch_xcorr_avx2(const float *_x, const float *_y, float *xcorr, int len, int max_pitch, int arch)
+{
+   int i;
+   celt_assert(max_pitch>0);
+   (void)arch;
+   for (i=0;i<max_pitch-7;i+=8)
+   {
+      xcorr_kernel_avx(_x, _y+i, &xcorr[i], len);
+   }
+   for (;i<max_pitch;i++)
+   {
+      xcorr[i] = celt_inner_prod(_x, _y+i, len, arch);
+   }
+}
+
+#endif
--- a/lib/libopus/celt/x86/pitch_sse.c
+++ b/lib/libopus/celt/x86/pitch_sse.c
@ -0,0 +1,185 @@
+/* Copyright (c) 2014, Cisco Systems, INC
+   Written by XiangMingZhu WeiZhou MinPeng YanWang
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "macros.h"
+#include "celt_lpc.h"
+#include "stack_alloc.h"
+#include "mathops.h"
+#include "pitch.h"
+
+#if defined(OPUS_X86_MAY_HAVE_SSE) && !defined(FIXED_POINT)
+
+#include <xmmintrin.h>
+#include "arch.h"
+
+void xcorr_kernel_sse(const opus_val16 *x, const opus_val16 *y, opus_val32 sum[4], int len)
+{
+   int j;
+   __m128 xsum1, xsum2;
+   xsum1 = _mm_loadu_ps(sum);
+   xsum2 = _mm_setzero_ps();
+
+   for (j = 0; j < len-3; j += 4)
+   {
+      __m128 x0 = _mm_loadu_ps(x+j);
+      __m128 yj = _mm_loadu_ps(y+j);
+      __m128 y3 = _mm_loadu_ps(y+j+3);
+
+      xsum1 = _mm_add_ps(xsum1,_mm_mul_ps(_mm_shuffle_ps(x0,x0,0x00),yj));
+      xsum2 = _mm_add_ps(xsum2,_mm_mul_ps(_mm_shuffle_ps(x0,x0,0x55),
+                                          _mm_shuffle_ps(yj,y3,0x49)));
+      xsum1 = _mm_add_ps(xsum1,_mm_mul_ps(_mm_shuffle_ps(x0,x0,0xaa),
+                                          _mm_shuffle_ps(yj,y3,0x9e)));
+      xsum2 = _mm_add_ps(xsum2,_mm_mul_ps(_mm_shuffle_ps(x0,x0,0xff),y3));
+   }
+   if (j < len)
+   {
+      xsum1 = _mm_add_ps(xsum1,_mm_mul_ps(_mm_load1_ps(x+j),_mm_loadu_ps(y+j)));
+      if (++j < len)
+      {
+         xsum2 = _mm_add_ps(xsum2,_mm_mul_ps(_mm_load1_ps(x+j),_mm_loadu_ps(y+j)));
+         if (++j < len)
+         {
+            xsum1 = _mm_add_ps(xsum1,_mm_mul_ps(_mm_load1_ps(x+j),_mm_loadu_ps(y+j)));
+         }
+      }
+   }
+   _mm_storeu_ps(sum,_mm_add_ps(xsum1,xsum2));
+}
+
+
+void dual_inner_prod_sse(const opus_val16 *x, const opus_val16 *y01, const opus_val16 *y02,
+      int N, opus_val32 *xy1, opus_val32 *xy2)
+{
+   int i;
+   __m128 xsum1, xsum2;
+   xsum1 = _mm_setzero_ps();
+   xsum2 = _mm_setzero_ps();
+   for (i=0;i<N-3;i+=4)
+   {
+      __m128 xi = _mm_loadu_ps(x+i);
+      __m128 y1i = _mm_loadu_ps(y01+i);
+      __m128 y2i = _mm_loadu_ps(y02+i);
+      xsum1 = _mm_add_ps(xsum1,_mm_mul_ps(xi, y1i));
+      xsum2 = _mm_add_ps(xsum2,_mm_mul_ps(xi, y2i));
+   }
+   /* Horizontal sum */
+   xsum1 = _mm_add_ps(xsum1, _mm_movehl_ps(xsum1, xsum1));
+   xsum1 = _mm_add_ss(xsum1, _mm_shuffle_ps(xsum1, xsum1, 0x55));
+   _mm_store_ss(xy1, xsum1);
+   xsum2 = _mm_add_ps(xsum2, _mm_movehl_ps(xsum2, xsum2));
+   xsum2 = _mm_add_ss(xsum2, _mm_shuffle_ps(xsum2, xsum2, 0x55));
+   _mm_store_ss(xy2, xsum2);
+   for (;i<N;i++)
+   {
+      *xy1 = MAC16_16(*xy1, x[i], y01[i]);
+      *xy2 = MAC16_16(*xy2, x[i], y02[i]);
+   }
+}
+
+opus_val32 celt_inner_prod_sse(const opus_val16 *x, const opus_val16 *y,
+      int N)
+{
+   int i;
+   float xy;
+   __m128 sum;
+   sum = _mm_setzero_ps();
+   /* FIXME: We should probably go 8-way and use 2 sums. */
+   for (i=0;i<N-3;i+=4)
+   {
+      __m128 xi = _mm_loadu_ps(x+i);
+      __m128 yi = _mm_loadu_ps(y+i);
+      sum = _mm_add_ps(sum,_mm_mul_ps(xi, yi));
+   }
+   /* Horizontal sum */
+   sum = _mm_add_ps(sum, _mm_movehl_ps(sum, sum));
+   sum = _mm_add_ss(sum, _mm_shuffle_ps(sum, sum, 0x55));
+   _mm_store_ss(&xy, sum);
+   for (;i<N;i++)
+   {
+      xy = MAC16_16(xy, x[i], y[i]);
+   }
+   return xy;
+}
+
+void comb_filter_const_sse(opus_val32 *y, opus_val32 *x, int T, int N,
+      opus_val16 g10, opus_val16 g11, opus_val16 g12)
+{
+   int i;
+   __m128 x0v;
+   __m128 g10v, g11v, g12v;
+   g10v = _mm_load1_ps(&g10);
+   g11v = _mm_load1_ps(&g11);
+   g12v = _mm_load1_ps(&g12);
+   x0v = _mm_loadu_ps(&x[-T-2]);
+   for (i=0;i<N-3;i+=4)
+   {
+      __m128 yi, yi2, x1v, x2v, x3v, x4v;
+      const opus_val32 *xp = &x[i-T-2];
+      yi = _mm_loadu_ps(x+i);
+      x4v = _mm_loadu_ps(xp+4);
+#if 0
+      /* Slower version with all loads */
+      x1v = _mm_loadu_ps(xp+1);
+      x2v = _mm_loadu_ps(xp+2);
+      x3v = _mm_loadu_ps(xp+3);
+#else
+      x2v = _mm_shuffle_ps(x0v, x4v, 0x4e);
+      x1v = _mm_shuffle_ps(x0v, x2v, 0x99);
+      x3v = _mm_shuffle_ps(x2v, x4v, 0x99);
+#endif
+
+      yi = _mm_add_ps(yi, _mm_mul_ps(g10v,x2v));
+#if 0 /* Set to 1 to make it bit-exact with the non-SSE version */
+      yi = _mm_add_ps(yi, _mm_mul_ps(g11v,_mm_add_ps(x3v,x1v)));
+      yi = _mm_add_ps(yi, _mm_mul_ps(g12v,_mm_add_ps(x4v,x0v)));
+#else
+      /* Use partial sums */
+      yi2 = _mm_add_ps(_mm_mul_ps(g11v,_mm_add_ps(x3v,x1v)),
+                       _mm_mul_ps(g12v,_mm_add_ps(x4v,x0v)));
+      yi = _mm_add_ps(yi, yi2);
+#endif
+      x0v=x4v;
+      _mm_storeu_ps(y+i, yi);
+   }
+#ifdef CUSTOM_MODES
+   for (;i<N;i++)
+   {
+      y[i] = x[i]
+               + MULT16_32_Q15(g10,x[i-T])
+               + MULT16_32_Q15(g11,ADD32(x[i-T+1],x[i-T-1]))
+               + MULT16_32_Q15(g12,ADD32(x[i-T+2],x[i-T-2]));
+   }
+#endif
+}
+
+
+#endif
--- a/lib/libopus/celt/x86/pitch_sse.h
+++ b/lib/libopus/celt/x86/pitch_sse.h
@ -0,0 +1,216 @@
+/* Copyright (c) 2013 Jean-Marc Valin and John Ridges
+   Copyright (c) 2014, Cisco Systems, INC MingXiang WeiZhou MinPeng YanWang*/
+/**
+   @file pitch_sse.h
+   @brief Pitch analysis
+ */
+
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef PITCH_SSE_H
+#define PITCH_SSE_H
+
+#if defined(HAVE_CONFIG_H)
+#include "config.h"
+#endif
+
+#if defined(OPUS_X86_MAY_HAVE_SSE4_1) && defined(FIXED_POINT)
+void xcorr_kernel_sse4_1(
+                    const opus_int16 *x,
+                    const opus_int16 *y,
+                    opus_val32       sum[4],
+                    int              len);
+#endif
+
+#if defined(OPUS_X86_MAY_HAVE_SSE) && !defined(FIXED_POINT)
+void xcorr_kernel_sse(
+                    const opus_val16 *x,
+                    const opus_val16 *y,
+                    opus_val32       sum[4],
+                    int              len);
+#endif
+
+#if defined(OPUS_X86_PRESUME_SSE4_1) && defined(FIXED_POINT)
+#define OVERRIDE_XCORR_KERNEL
+#define xcorr_kernel(x, y, sum, len, arch) \
+    ((void)arch, xcorr_kernel_sse4_1(x, y, sum, len))
+
+#elif defined(OPUS_X86_PRESUME_SSE) && !defined(FIXED_POINT)
+#define OVERRIDE_XCORR_KERNEL
+#define xcorr_kernel(x, y, sum, len, arch) \
+    ((void)arch, xcorr_kernel_sse(x, y, sum, len))
+
+#elif defined(OPUS_HAVE_RTCD) &&  ((defined(OPUS_X86_MAY_HAVE_SSE4_1) && defined(FIXED_POINT)) || (defined(OPUS_X86_MAY_HAVE_SSE) && !defined(FIXED_POINT)))
+
+extern void (*const XCORR_KERNEL_IMPL[OPUS_ARCHMASK + 1])(
+                    const opus_val16 *x,
+                    const opus_val16 *y,
+                    opus_val32       sum[4],
+                    int              len);
+
+#define OVERRIDE_XCORR_KERNEL
+#define xcorr_kernel(x, y, sum, len, arch) \
+    ((*XCORR_KERNEL_IMPL[(arch) & OPUS_ARCHMASK])(x, y, sum, len))
+
+#endif
+
+#if defined(OPUS_X86_MAY_HAVE_SSE4_1) && defined(FIXED_POINT)
+opus_val32 celt_inner_prod_sse4_1(
+    const opus_int16 *x,
+    const opus_int16 *y,
+    int               N);
+#endif
+
+#if defined(OPUS_X86_MAY_HAVE_SSE2) && defined(FIXED_POINT)
+opus_val32 celt_inner_prod_sse2(
+    const opus_int16 *x,
+    const opus_int16 *y,
+    int               N);
+#endif
+
+#if defined(OPUS_X86_MAY_HAVE_SSE) && !defined(FIXED_POINT)
+opus_val32 celt_inner_prod_sse(
+    const opus_val16 *x,
+    const opus_val16 *y,
+    int               N);
+#endif
+
+
+#if defined(OPUS_X86_PRESUME_SSE4_1) && defined(FIXED_POINT)
+#define OVERRIDE_CELT_INNER_PROD
+#define celt_inner_prod(x, y, N, arch) \
+    ((void)arch, celt_inner_prod_sse4_1(x, y, N))
+
+#elif defined(OPUS_X86_PRESUME_SSE2) && defined(FIXED_POINT) && !defined(OPUS_X86_MAY_HAVE_SSE4_1)
+#define OVERRIDE_CELT_INNER_PROD
+#define celt_inner_prod(x, y, N, arch) \
+    ((void)arch, celt_inner_prod_sse2(x, y, N))
+
+#elif defined(OPUS_X86_PRESUME_SSE) && !defined(FIXED_POINT)
+#define OVERRIDE_CELT_INNER_PROD
+#define celt_inner_prod(x, y, N, arch) \
+    ((void)arch, celt_inner_prod_sse(x, y, N))
+
+
+#elif defined(OPUS_HAVE_RTCD) && (((defined(OPUS_X86_MAY_HAVE_SSE4_1) || defined(OPUS_X86_MAY_HAVE_SSE2)) && defined(FIXED_POINT)) || \
+    (defined(OPUS_X86_MAY_HAVE_SSE) && !defined(FIXED_POINT)))
+
+extern opus_val32 (*const CELT_INNER_PROD_IMPL[OPUS_ARCHMASK + 1])(
+                    const opus_val16 *x,
+                    const opus_val16 *y,
+                    int               N);
+
+#define OVERRIDE_CELT_INNER_PROD
+#define celt_inner_prod(x, y, N, arch) \
+    ((*CELT_INNER_PROD_IMPL[(arch) & OPUS_ARCHMASK])(x, y, N))
+
+#endif
+
+#if defined(OPUS_X86_MAY_HAVE_SSE) && !defined(FIXED_POINT)
+
+void dual_inner_prod_sse(const opus_val16 *x,
+    const opus_val16 *y01,
+    const opus_val16 *y02,
+    int               N,
+    opus_val32       *xy1,
+    opus_val32       *xy2);
+
+void comb_filter_const_sse(opus_val32 *y,
+    opus_val32 *x,
+    int         T,
+    int         N,
+    opus_val16  g10,
+    opus_val16  g11,
+    opus_val16  g12);
+
+
+#if defined(OPUS_X86_PRESUME_SSE)
+#define OVERRIDE_DUAL_INNER_PROD
+#define OVERRIDE_COMB_FILTER_CONST
+# define dual_inner_prod(x, y01, y02, N, xy1, xy2, arch) \
+    ((void)(arch),dual_inner_prod_sse(x, y01, y02, N, xy1, xy2))
+
+# define comb_filter_const(y, x, T, N, g10, g11, g12, arch) \
+    ((void)(arch),comb_filter_const_sse(y, x, T, N, g10, g11, g12))
+#elif defined(OPUS_HAVE_RTCD)
+
+#define OVERRIDE_DUAL_INNER_PROD
+#define OVERRIDE_COMB_FILTER_CONST
+extern void (*const DUAL_INNER_PROD_IMPL[OPUS_ARCHMASK + 1])(
+              const opus_val16 *x,
+              const opus_val16 *y01,
+              const opus_val16 *y02,
+              int               N,
+              opus_val32       *xy1,
+              opus_val32       *xy2);
+
+#define dual_inner_prod(x, y01, y02, N, xy1, xy2, arch) \
+    ((*DUAL_INNER_PROD_IMPL[(arch) & OPUS_ARCHMASK])(x, y01, y02, N, xy1, xy2))
+
+extern void (*const COMB_FILTER_CONST_IMPL[OPUS_ARCHMASK + 1])(
+              opus_val32 *y,
+              opus_val32 *x,
+              int         T,
+              int         N,
+              opus_val16  g10,
+              opus_val16  g11,
+              opus_val16  g12);
+
+#define comb_filter_const(y, x, T, N, g10, g11, g12, arch) \
+    ((*COMB_FILTER_CONST_IMPL[(arch) & OPUS_ARCHMASK])(y, x, T, N, g10, g11, g12))
+
+#define NON_STATIC_COMB_FILTER_CONST_C
+
+#endif
+
+void celt_pitch_xcorr_avx2(const float *_x, const float *_y, float *xcorr, int len, int max_pitch, int arch);
+
+#if defined(OPUS_X86_PRESUME_AVX2)
+
+#define OVERRIDE_PITCH_XCORR
+# define celt_pitch_xcorr celt_pitch_xcorr_avx2
+
+#elif defined(OPUS_HAVE_RTCD) && defined(OPUS_X86_MAY_HAVE_AVX2)
+
+#define OVERRIDE_PITCH_XCORR
+extern void (*const PITCH_XCORR_IMPL[OPUS_ARCHMASK + 1])(
+              const float *_x,
+              const float *_y,
+              float *xcorr,
+              int len,
+              int max_pitch,
+              int arch
+              );
+
+#define celt_pitch_xcorr(_x, _y, xcorr, len, max_pitch, arch) \
+    ((*PITCH_XCORR_IMPL[(arch) & OPUS_ARCHMASK])(_x, _y, xcorr, len, max_pitch, arch))
+
+
+#endif /* OPUS_X86_PRESUME_AVX2 && !OPUS_HAVE_RTCD */
+
+#endif /* OPUS_X86_MAY_HAVE_SSE && !FIXED_POINT */
+
+#endif
--- a/lib/libopus/celt/x86/pitch_sse2.c
+++ b/lib/libopus/celt/x86/pitch_sse2.c
@ -0,0 +1,95 @@
+/* Copyright (c) 2014, Cisco Systems, INC
+   Written by XiangMingZhu WeiZhou MinPeng YanWang
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include <xmmintrin.h>
+#include <emmintrin.h>
+
+#include "macros.h"
+#include "celt_lpc.h"
+#include "stack_alloc.h"
+#include "mathops.h"
+#include "pitch.h"
+
+#if defined(OPUS_X86_MAY_HAVE_SSE2) && defined(FIXED_POINT)
+opus_val32 celt_inner_prod_sse2(const opus_val16 *x, const opus_val16 *y,
+      int N)
+{
+    opus_int  i, dataSize16;
+    opus_int32 sum;
+
+    __m128i inVec1_76543210, inVec1_FEDCBA98, acc1;
+    __m128i inVec2_76543210, inVec2_FEDCBA98, acc2;
+
+    sum = 0;
+    dataSize16 = N & ~15;
+
+    acc1 = _mm_setzero_si128();
+    acc2 = _mm_setzero_si128();
+
+    for (i=0;i<dataSize16;i+=16)
+    {
+        inVec1_76543210 = _mm_loadu_si128((__m128i *)(void*)(&x[i + 0]));
+        inVec2_76543210 = _mm_loadu_si128((__m128i *)(void*)(&y[i + 0]));
+
+        inVec1_FEDCBA98 = _mm_loadu_si128((__m128i *)(void*)(&x[i + 8]));
+        inVec2_FEDCBA98 = _mm_loadu_si128((__m128i *)(void*)(&y[i + 8]));
+
+        inVec1_76543210 = _mm_madd_epi16(inVec1_76543210, inVec2_76543210);
+        inVec1_FEDCBA98 = _mm_madd_epi16(inVec1_FEDCBA98, inVec2_FEDCBA98);
+
+        acc1 = _mm_add_epi32(acc1, inVec1_76543210);
+        acc2 = _mm_add_epi32(acc2, inVec1_FEDCBA98);
+    }
+
+    acc1 = _mm_add_epi32( acc1, acc2 );
+
+    if (N - i >= 8)
+    {
+        inVec1_76543210 = _mm_loadu_si128((__m128i *)(void*)(&x[i + 0]));
+        inVec2_76543210 = _mm_loadu_si128((__m128i *)(void*)(&y[i + 0]));
+
+        inVec1_76543210 = _mm_madd_epi16(inVec1_76543210, inVec2_76543210);
+
+        acc1 = _mm_add_epi32(acc1, inVec1_76543210);
+        i += 8;
+    }
+
+    acc1 = _mm_add_epi32(acc1, _mm_unpackhi_epi64( acc1, acc1));
+    acc1 = _mm_add_epi32(acc1, _mm_shufflelo_epi16( acc1, 0x0E));
+    sum += _mm_cvtsi128_si32(acc1);
+
+    for (;i<N;i++) {
+        sum = silk_SMLABB(sum, x[i], y[i]);
+    }
+
+    return sum;
+}
+#endif
--- a/lib/libopus/celt/x86/pitch_sse4_1.c
+++ b/lib/libopus/celt/x86/pitch_sse4_1.c
@ -0,0 +1,240 @@
+/* Copyright (c) 2014, Cisco Systems, INC
+   Written by XiangMingZhu WeiZhou MinPeng YanWang
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include <xmmintrin.h>
+#include <emmintrin.h>
+
+#include "macros.h"
+#include "celt_lpc.h"
+#include "stack_alloc.h"
+#include "mathops.h"
+#include "pitch.h"
+
+#if defined(OPUS_X86_MAY_HAVE_SSE4_1) && defined(FIXED_POINT)
+#include <smmintrin.h>
+#include "x86cpu.h"
+
+opus_val32 celt_inner_prod_sse4_1(const opus_val16 *x, const opus_val16 *y,
+      int N)
+{
+    opus_int  i, dataSize16;
+    opus_int32 sum;
+    __m128i inVec1_76543210, inVec1_FEDCBA98, acc1;
+    __m128i inVec2_76543210, inVec2_FEDCBA98, acc2;
+    __m128i inVec1_3210, inVec2_3210;
+
+    sum = 0;
+    dataSize16 = N & ~15;
+
+    acc1 = _mm_setzero_si128();
+    acc2 = _mm_setzero_si128();
+
+    for (i=0;i<dataSize16;i+=16) {
+        inVec1_76543210 = _mm_loadu_si128((__m128i *)(void*)(&x[i + 0]));
+        inVec2_76543210 = _mm_loadu_si128((__m128i *)(void*)(&y[i + 0]));
+
+        inVec1_FEDCBA98 = _mm_loadu_si128((__m128i *)(void*)(&x[i + 8]));
+        inVec2_FEDCBA98 = _mm_loadu_si128((__m128i *)(void*)(&y[i + 8]));
+
+        inVec1_76543210 = _mm_madd_epi16(inVec1_76543210, inVec2_76543210);
+        inVec1_FEDCBA98 = _mm_madd_epi16(inVec1_FEDCBA98, inVec2_FEDCBA98);
+
+        acc1 = _mm_add_epi32(acc1, inVec1_76543210);
+        acc2 = _mm_add_epi32(acc2, inVec1_FEDCBA98);
+    }
+
+    acc1 = _mm_add_epi32(acc1, acc2);
+
+    if (N - i >= 8)
+    {
+        inVec1_76543210 = _mm_loadu_si128((__m128i *)(void*)(&x[i + 0]));
+        inVec2_76543210 = _mm_loadu_si128((__m128i *)(void*)(&y[i + 0]));
+
+        inVec1_76543210 = _mm_madd_epi16(inVec1_76543210, inVec2_76543210);
+
+        acc1 = _mm_add_epi32(acc1, inVec1_76543210);
+        i += 8;
+    }
+
+    if (N - i >= 4)
+    {
+        inVec1_3210 = OP_CVTEPI16_EPI32_M64(&x[i + 0]);
+        inVec2_3210 = OP_CVTEPI16_EPI32_M64(&y[i + 0]);
+
+        inVec1_3210 = _mm_mullo_epi32(inVec1_3210, inVec2_3210);
+
+        acc1 = _mm_add_epi32(acc1, inVec1_3210);
+        i += 4;
+    }
+
+    acc1 = _mm_add_epi32(acc1, _mm_unpackhi_epi64(acc1, acc1));
+    acc1 = _mm_add_epi32(acc1, _mm_shufflelo_epi16(acc1, 0x0E));
+
+    sum += _mm_cvtsi128_si32(acc1);
+
+    for (;i<N;i++)
+    {
+        sum = silk_SMLABB(sum, x[i], y[i]);
+    }
+
+    return sum;
+}
+
+void xcorr_kernel_sse4_1(const opus_val16 * x, const opus_val16 * y, opus_val32 sum[ 4 ], int len)
+{
+    int j;
+
+    __m128i vecX, vecX0, vecX1, vecX2, vecX3;
+    __m128i vecY0, vecY1, vecY2, vecY3;
+    __m128i sum0, sum1, sum2, sum3, vecSum;
+    __m128i initSum;
+
+#ifdef OPUS_CHECK_ASM
+    opus_val32 sum_c[4];
+    for (j=0;j<4;j++) {
+      sum_c[j] = sum[j];
+    }
+    xcorr_kernel_c(x, y, sum_c, len);
+#endif
+
+    celt_assert(len >= 3);
+
+    sum0 = _mm_setzero_si128();
+    sum1 = _mm_setzero_si128();
+    sum2 = _mm_setzero_si128();
+    sum3 = _mm_setzero_si128();
+
+    for (j=0;j<(len-7);j+=8)
+    {
+        vecX = _mm_loadu_si128((__m128i *)(void*)(&x[j + 0]));
+        vecY0 = _mm_loadu_si128((__m128i *)(void*)(&y[j + 0]));
+        vecY1 = _mm_loadu_si128((__m128i *)(void*)(&y[j + 1]));
+        vecY2 = _mm_loadu_si128((__m128i *)(void*)(&y[j + 2]));
+        vecY3 = _mm_loadu_si128((__m128i *)(void*)(&y[j + 3]));
+
+        sum0 = _mm_add_epi32(sum0, _mm_madd_epi16(vecX, vecY0));
+        sum1 = _mm_add_epi32(sum1, _mm_madd_epi16(vecX, vecY1));
+        sum2 = _mm_add_epi32(sum2, _mm_madd_epi16(vecX, vecY2));
+        sum3 = _mm_add_epi32(sum3, _mm_madd_epi16(vecX, vecY3));
+    }
+
+    sum0 = _mm_add_epi32(sum0, _mm_unpackhi_epi64( sum0, sum0));
+    sum0 = _mm_add_epi32(sum0, _mm_shufflelo_epi16( sum0, 0x0E));
+
+    sum1 = _mm_add_epi32(sum1, _mm_unpackhi_epi64( sum1, sum1));
+    sum1 = _mm_add_epi32(sum1, _mm_shufflelo_epi16( sum1, 0x0E));
+
+    sum2 = _mm_add_epi32(sum2, _mm_unpackhi_epi64( sum2, sum2));
+    sum2 = _mm_add_epi32(sum2, _mm_shufflelo_epi16( sum2, 0x0E));
+
+    sum3 = _mm_add_epi32(sum3, _mm_unpackhi_epi64( sum3, sum3));
+    sum3 = _mm_add_epi32(sum3, _mm_shufflelo_epi16( sum3, 0x0E));
+
+    vecSum = _mm_unpacklo_epi64(_mm_unpacklo_epi32(sum0, sum1),
+          _mm_unpacklo_epi32(sum2, sum3));
+
+    for (;j<(len-3);j+=4)
+    {
+        vecX = OP_CVTEPI16_EPI32_M64(&x[j + 0]);
+        vecX0 = _mm_shuffle_epi32(vecX, 0x00);
+        vecX1 = _mm_shuffle_epi32(vecX, 0x55);
+        vecX2 = _mm_shuffle_epi32(vecX, 0xaa);
+        vecX3 = _mm_shuffle_epi32(vecX, 0xff);
+
+        vecY0 = OP_CVTEPI16_EPI32_M64(&y[j + 0]);
+        vecY1 = OP_CVTEPI16_EPI32_M64(&y[j + 1]);
+        vecY2 = OP_CVTEPI16_EPI32_M64(&y[j + 2]);
+        vecY3 = OP_CVTEPI16_EPI32_M64(&y[j + 3]);
+
+        sum0 = _mm_mullo_epi32(vecX0, vecY0);
+        sum1 = _mm_mullo_epi32(vecX1, vecY1);
+        sum2 = _mm_mullo_epi32(vecX2, vecY2);
+        sum3 = _mm_mullo_epi32(vecX3, vecY3);
+
+        sum0 = _mm_add_epi32(sum0, sum1);
+        sum2 = _mm_add_epi32(sum2, sum3);
+        vecSum = _mm_add_epi32(vecSum, sum0);
+        vecSum = _mm_add_epi32(vecSum, sum2);
+    }
+
+    vecX = OP_CVTEPI16_EPI32_M64(&x[len - 4]);
+    if (len - j == 3)
+    {
+        vecX0 = _mm_shuffle_epi32(vecX, 0x55);
+        vecX1 = _mm_shuffle_epi32(vecX, 0xaa);
+        vecX2 = _mm_shuffle_epi32(vecX, 0xff);
+
+        vecY0 = OP_CVTEPI16_EPI32_M64(&y[j + 0]);
+        vecY1 = OP_CVTEPI16_EPI32_M64(&y[j + 1]);
+        vecY2 = OP_CVTEPI16_EPI32_M64(&y[j + 2]);
+
+        sum0 = _mm_mullo_epi32(vecX0, vecY0);
+        sum1 = _mm_mullo_epi32(vecX1, vecY1);
+        sum2 = _mm_mullo_epi32(vecX2, vecY2);
+
+        vecSum = _mm_add_epi32(vecSum, sum0);
+        vecSum = _mm_add_epi32(vecSum, sum1);
+        vecSum = _mm_add_epi32(vecSum, sum2);
+    }
+    else if (len - j == 2)
+    {
+        vecX0 = _mm_shuffle_epi32(vecX, 0xaa);
+        vecX1 = _mm_shuffle_epi32(vecX, 0xff);
+
+        vecY0 = OP_CVTEPI16_EPI32_M64(&y[j + 0]);
+        vecY1 = OP_CVTEPI16_EPI32_M64(&y[j + 1]);
+
+        sum0 = _mm_mullo_epi32(vecX0, vecY0);
+        sum1 = _mm_mullo_epi32(vecX1, vecY1);
+
+        vecSum = _mm_add_epi32(vecSum, sum0);
+        vecSum = _mm_add_epi32(vecSum, sum1);
+    }
+    else if (len - j == 1)
+    {
+        vecX0 = _mm_shuffle_epi32(vecX, 0xff);
+
+        vecY0 = OP_CVTEPI16_EPI32_M64(&y[j + 0]);
+
+        sum0 = _mm_mullo_epi32(vecX0, vecY0);
+
+        vecSum = _mm_add_epi32(vecSum, sum0);
+    }
+
+    initSum = _mm_loadu_si128((__m128i *)(void*)(&sum[0]));
+    initSum = _mm_add_epi32(initSum, vecSum);
+    _mm_storeu_si128((__m128i *)(void*)sum, initSum);
+
+#ifdef OPUS_CHECK_ASM
+    celt_assert(!memcmp(sum_c, sum, sizeof(sum_c)));
+#endif
+}
+#endif
--- a/lib/libopus/celt/x86/vq_sse.h
+++ b/lib/libopus/celt/x86/vq_sse.h
@ -0,0 +1,52 @@
+/* Copyright (c) 2016  Jean-Marc Valin */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef VQ_SSE_H
+#define VQ_SSE_H
+
+#if defined(OPUS_X86_MAY_HAVE_SSE2) && !defined(FIXED_POINT)
+
+opus_val16 op_pvq_search_sse2(celt_norm *_X, int *iy, int K, int N, int arch);
+
+#if defined(OPUS_X86_PRESUME_SSE2)
+
+#define OVERRIDE_OP_PVQ_SEARCH
+#define op_pvq_search(x, iy, K, N, arch) \
+    (op_pvq_search_sse2(x, iy, K, N, arch))
+
+#elif defined(OPUS_HAVE_RTCD)
+
+#define OVERRIDE_OP_PVQ_SEARCH
+extern opus_val16 (*const OP_PVQ_SEARCH_IMPL[OPUS_ARCHMASK + 1])(
+      celt_norm *_X, int *iy, int K, int N, int arch);
+
+#  define op_pvq_search(X, iy, K, N, arch) \
+    ((*OP_PVQ_SEARCH_IMPL[(arch) & OPUS_ARCHMASK])(X, iy, K, N, arch))
+
+#endif
+#endif
+
+#endif
--- a/lib/libopus/celt/x86/vq_sse2.c
+++ b/lib/libopus/celt/x86/vq_sse2.c
@ -0,0 +1,217 @@
+/* Copyright (c) 2007-2008 CSIRO
+   Copyright (c) 2007-2009 Xiph.Org Foundation
+   Copyright (c) 2007-2016 Jean-Marc Valin */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include <xmmintrin.h>
+#include <emmintrin.h>
+#include "celt_lpc.h"
+#include "stack_alloc.h"
+#include "mathops.h"
+#include "vq.h"
+#include "x86cpu.h"
+
+
+#ifndef FIXED_POINT
+
+opus_val16 op_pvq_search_sse2(celt_norm *_X, int *iy, int K, int N, int arch)
+{
+   int i, j;
+   int pulsesLeft;
+   float xy, yy;
+   VARDECL(celt_norm, y);
+   VARDECL(celt_norm, X);
+   VARDECL(float, signy);
+   __m128 signmask;
+   __m128 sums;
+   __m128i fours;
+   SAVE_STACK;
+
+   (void)arch;
+   /* All bits set to zero, except for the sign bit. */
+   signmask = _mm_set_ps1(-0.f);
+   fours = _mm_set_epi32(4, 4, 4, 4);
+   ALLOC(y, N+3, celt_norm);
+   ALLOC(X, N+3, celt_norm);
+   ALLOC(signy, N+3, float);
+
+   OPUS_COPY(X, _X, N);
+   X[N] = X[N+1] = X[N+2] = 0;
+   sums = _mm_setzero_ps();
+   for (j=0;j<N;j+=4)
+   {
+      __m128 x4, s4;
+      x4 = _mm_loadu_ps(&X[j]);
+      s4 = _mm_cmplt_ps(x4, _mm_setzero_ps());
+      /* Get rid of the sign */
+      x4 = _mm_andnot_ps(signmask, x4);
+      sums = _mm_add_ps(sums, x4);
+      /* Clear y and iy in case we don't do the projection. */
+      _mm_storeu_ps(&y[j], _mm_setzero_ps());
+      _mm_storeu_si128((__m128i*)(void*)&iy[j], _mm_setzero_si128());
+      _mm_storeu_ps(&X[j], x4);
+      _mm_storeu_ps(&signy[j], s4);
+   }
+   sums = _mm_add_ps(sums, _mm_shuffle_ps(sums, sums, _MM_SHUFFLE(1, 0, 3, 2)));
+   sums = _mm_add_ps(sums, _mm_shuffle_ps(sums, sums, _MM_SHUFFLE(2, 3, 0, 1)));
+
+   xy = yy = 0;
+
+   pulsesLeft = K;
+
+   /* Do a pre-search by projecting on the pyramid */
+   if (K > (N>>1))
+   {
+      __m128i pulses_sum;
+      __m128 yy4, xy4;
+      __m128 rcp4;
+      opus_val32 sum = _mm_cvtss_f32(sums);
+      /* If X is too small, just replace it with a pulse at 0 */
+      /* Prevents infinities and NaNs from causing too many pulses
+         to be allocated. 64 is an approximation of infinity here. */
+      if (!(sum > EPSILON && sum < 64))
+      {
+         X[0] = QCONST16(1.f,14);
+         j=1; do
+            X[j]=0;
+         while (++j<N);
+         sums = _mm_set_ps1(1.f);
+      }
+      /* Using K+e with e < 1 guarantees we cannot get more than K pulses. */
+      rcp4 = _mm_mul_ps(_mm_set_ps1((float)(K+.8)), _mm_rcp_ps(sums));
+      xy4 = yy4 = _mm_setzero_ps();
+      pulses_sum = _mm_setzero_si128();
+      for (j=0;j<N;j+=4)
+      {
+         __m128 rx4, x4, y4;
+         __m128i iy4;
+         x4 = _mm_loadu_ps(&X[j]);
+         rx4 = _mm_mul_ps(x4, rcp4);
+         iy4 = _mm_cvttps_epi32(rx4);
+         pulses_sum = _mm_add_epi32(pulses_sum, iy4);
+         _mm_storeu_si128((__m128i*)(void*)&iy[j], iy4);
+         y4 = _mm_cvtepi32_ps(iy4);
+         xy4 = _mm_add_ps(xy4, _mm_mul_ps(x4, y4));
+         yy4 = _mm_add_ps(yy4, _mm_mul_ps(y4, y4));
+         /* double the y[] vector so we don't have to do it in the search loop. */
+         _mm_storeu_ps(&y[j], _mm_add_ps(y4, y4));
+      }
+      pulses_sum = _mm_add_epi32(pulses_sum, _mm_shuffle_epi32(pulses_sum, _MM_SHUFFLE(1, 0, 3, 2)));
+      pulses_sum = _mm_add_epi32(pulses_sum, _mm_shuffle_epi32(pulses_sum, _MM_SHUFFLE(2, 3, 0, 1)));
+      pulsesLeft -= _mm_cvtsi128_si32(pulses_sum);
+      xy4 = _mm_add_ps(xy4, _mm_shuffle_ps(xy4, xy4, _MM_SHUFFLE(1, 0, 3, 2)));
+      xy4 = _mm_add_ps(xy4, _mm_shuffle_ps(xy4, xy4, _MM_SHUFFLE(2, 3, 0, 1)));
+      xy = _mm_cvtss_f32(xy4);
+      yy4 = _mm_add_ps(yy4, _mm_shuffle_ps(yy4, yy4, _MM_SHUFFLE(1, 0, 3, 2)));
+      yy4 = _mm_add_ps(yy4, _mm_shuffle_ps(yy4, yy4, _MM_SHUFFLE(2, 3, 0, 1)));
+      yy = _mm_cvtss_f32(yy4);
+   }
+   X[N] = X[N+1] = X[N+2] = -100;
+   y[N] = y[N+1] = y[N+2] = 100;
+   celt_sig_assert(pulsesLeft>=0);
+
+   /* This should never happen, but just in case it does (e.g. on silence)
+      we fill the first bin with pulses. */
+   if (pulsesLeft > N+3)
+   {
+      opus_val16 tmp = (opus_val16)pulsesLeft;
+      yy = MAC16_16(yy, tmp, tmp);
+      yy = MAC16_16(yy, tmp, y[0]);
+      iy[0] += pulsesLeft;
+      pulsesLeft=0;
+   }
+
+   for (i=0;i<pulsesLeft;i++)
+   {
+      int best_id;
+      __m128 xy4, yy4;
+      __m128 max, max2;
+      __m128i count;
+      __m128i pos;
+      /* The squared magnitude term gets added anyway, so we might as well
+         add it outside the loop */
+      yy = ADD16(yy, 1);
+      xy4 = _mm_load1_ps(&xy);
+      yy4 = _mm_load1_ps(&yy);
+      max = _mm_setzero_ps();
+      pos = _mm_setzero_si128();
+      count = _mm_set_epi32(3, 2, 1, 0);
+      for (j=0;j<N;j+=4)
+      {
+         __m128 x4, y4, r4;
+         x4 = _mm_loadu_ps(&X[j]);
+         y4 = _mm_loadu_ps(&y[j]);
+         x4 = _mm_add_ps(x4, xy4);
+         y4 = _mm_add_ps(y4, yy4);
+         y4 = _mm_rsqrt_ps(y4);
+         r4 = _mm_mul_ps(x4, y4);
+         /* Update the index of the max. */
+         pos = _mm_max_epi16(pos, _mm_and_si128(count, _mm_castps_si128(_mm_cmpgt_ps(r4, max))));
+         /* Update the max. */
+         max = _mm_max_ps(max, r4);
+         /* Update the indices (+4) */
+         count = _mm_add_epi32(count, fours);
+      }
+      /* Horizontal max */
+      max2 = _mm_max_ps(max, _mm_shuffle_ps(max, max, _MM_SHUFFLE(1, 0, 3, 2)));
+      max2 = _mm_max_ps(max2, _mm_shuffle_ps(max2, max2, _MM_SHUFFLE(2, 3, 0, 1)));
+      /* Now that max2 contains the max at all positions, look at which value(s) of the
+         partial max is equal to the global max. */
+      pos = _mm_and_si128(pos, _mm_castps_si128(_mm_cmpeq_ps(max, max2)));
+      pos = _mm_max_epi16(pos, _mm_unpackhi_epi64(pos, pos));
+      pos = _mm_max_epi16(pos, _mm_shufflelo_epi16(pos, _MM_SHUFFLE(1, 0, 3, 2)));
+      best_id = _mm_cvtsi128_si32(pos);
+
+      /* Updating the sums of the new pulse(s) */
+      xy = ADD32(xy, EXTEND32(X[best_id]));
+      /* We're multiplying y[j] by two so we don't have to do it here */
+      yy = ADD16(yy, y[best_id]);
+
+      /* Only now that we've made the final choice, update y/iy */
+      /* Multiplying y[j] by 2 so we don't have to do it everywhere else */
+      y[best_id] += 2;
+      iy[best_id]++;
+   }
+
+   /* Put the original sign back */
+   for (j=0;j<N;j+=4)
+   {
+      __m128i y4;
+      __m128i s4;
+      y4 = _mm_loadu_si128((__m128i*)(void*)&iy[j]);
+      s4 = _mm_castps_si128(_mm_loadu_ps(&signy[j]));
+      y4 = _mm_xor_si128(_mm_add_epi32(y4, s4), s4);
+      _mm_storeu_si128((__m128i*)(void*)&iy[j], y4);
+   }
+   RESTORE_STACK;
+   return yy;
+}
+
+#endif
--- a/lib/libopus/celt/x86/x86_arch_macros.h
+++ b/lib/libopus/celt/x86/x86_arch_macros.h
@ -0,0 +1,47 @@
+/* Copyright (c) 2023 Amazon */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR
+   CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifdef _MSC_VER
+
+# ifdef OPUS_X86_MAY_HAVE_SSE
+#  ifndef __SSE__
+#   define __SSE__
+#  endif
+# endif
+
+# ifdef OPUS_X86_MAY_HAVE_SSE2
+#  ifndef __SSE2__
+#   define __SSE2__
+#  endif
+# endif
+
+# ifdef OPUS_X86_MAY_HAVE_SSE4_1
+#  ifndef __SSE4_1__
+#   define __SSE4_1__
+#  endif
+# endif
+
+#endif
--- a/lib/libopus/celt/x86/x86_celt_map.c
+++ b/lib/libopus/celt/x86/x86_celt_map.c
@ -0,0 +1,187 @@
+/* Copyright (c) 2014, Cisco Systems, INC
+   Written by XiangMingZhu WeiZhou MinPeng YanWang
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#if defined(HAVE_CONFIG_H)
+#include "config.h"
+#endif
+
+#include "x86/x86cpu.h"
+#include "celt_lpc.h"
+#include "pitch.h"
+#include "pitch_sse.h"
+#include "vq.h"
+
+#if defined(OPUS_HAVE_RTCD)
+
+# if defined(FIXED_POINT)
+
+#if defined(OPUS_X86_MAY_HAVE_SSE4_1) && !defined(OPUS_X86_PRESUME_SSE4_1)
+
+void (*const CELT_FIR_IMPL[OPUS_ARCHMASK + 1])(
+         const opus_val16 *x,
+         const opus_val16 *num,
+         opus_val16       *y,
+         int              N,
+         int              ord,
+         int              arch
+) = {
+  celt_fir_c,                /* non-sse */
+  celt_fir_c,
+  celt_fir_c,
+  MAY_HAVE_SSE4_1(celt_fir), /* sse4.1  */
+  MAY_HAVE_SSE4_1(celt_fir)  /* avx  */
+};
+
+void (*const XCORR_KERNEL_IMPL[OPUS_ARCHMASK + 1])(
+         const opus_val16 *x,
+         const opus_val16 *y,
+         opus_val32       sum[4],
+         int              len
+) = {
+  xcorr_kernel_c,                /* non-sse */
+  xcorr_kernel_c,
+  xcorr_kernel_c,
+  MAY_HAVE_SSE4_1(xcorr_kernel), /* sse4.1  */
+  MAY_HAVE_SSE4_1(xcorr_kernel)  /* avx  */
+};
+
+#endif
+
+#if (defined(OPUS_X86_MAY_HAVE_SSE4_1) && !defined(OPUS_X86_PRESUME_SSE4_1)) ||  \
+ (!defined(OPUS_X86_MAY_HAVE_SSE_4_1) && defined(OPUS_X86_MAY_HAVE_SSE2) && !defined(OPUS_X86_PRESUME_SSE2))
+
+opus_val32 (*const CELT_INNER_PROD_IMPL[OPUS_ARCHMASK + 1])(
+         const opus_val16 *x,
+         const opus_val16 *y,
+         int              N
+) = {
+  celt_inner_prod_c,                /* non-sse */
+  celt_inner_prod_c,
+  MAY_HAVE_SSE2(celt_inner_prod),
+  MAY_HAVE_SSE4_1(celt_inner_prod), /* sse4.1  */
+  MAY_HAVE_SSE4_1(celt_inner_prod)  /* avx  */
+};
+
+#endif
+
+# else
+
+#if defined(OPUS_X86_MAY_HAVE_AVX2) && !defined(OPUS_X86_PRESUME_AVX2)
+
+void (*const PITCH_XCORR_IMPL[OPUS_ARCHMASK + 1])(
+         const float *_x,
+         const float *_y,
+         float *xcorr,
+         int len,
+         int max_pitch,
+         int arch
+) = {
+  celt_pitch_xcorr_c,                /* non-sse */
+  celt_pitch_xcorr_c,
+  celt_pitch_xcorr_c,
+  celt_pitch_xcorr_c,
+  MAY_HAVE_AVX2(celt_pitch_xcorr)
+};
+
+#endif
+
+
+#if defined(OPUS_X86_MAY_HAVE_SSE) && !defined(OPUS_X86_PRESUME_SSE)
+
+void (*const XCORR_KERNEL_IMPL[OPUS_ARCHMASK + 1])(
+         const opus_val16 *x,
+         const opus_val16 *y,
+         opus_val32       sum[4],
+         int              len
+) = {
+  xcorr_kernel_c,                /* non-sse */
+  MAY_HAVE_SSE(xcorr_kernel),
+  MAY_HAVE_SSE(xcorr_kernel),
+  MAY_HAVE_SSE(xcorr_kernel),
+  MAY_HAVE_SSE(xcorr_kernel)
+};
+
+opus_val32 (*const CELT_INNER_PROD_IMPL[OPUS_ARCHMASK + 1])(
+         const opus_val16 *x,
+         const opus_val16 *y,
+         int              N
+) = {
+  celt_inner_prod_c,                /* non-sse */
+  MAY_HAVE_SSE(celt_inner_prod),
+  MAY_HAVE_SSE(celt_inner_prod),
+  MAY_HAVE_SSE(celt_inner_prod),
+  MAY_HAVE_SSE(celt_inner_prod)
+};
+
+void (*const DUAL_INNER_PROD_IMPL[OPUS_ARCHMASK + 1])(
+                    const opus_val16 *x,
+                    const opus_val16 *y01,
+                    const opus_val16 *y02,
+                    int               N,
+                    opus_val32       *xy1,
+                    opus_val32       *xy2
+) = {
+  dual_inner_prod_c,                /* non-sse */
+  MAY_HAVE_SSE(dual_inner_prod),
+  MAY_HAVE_SSE(dual_inner_prod),
+  MAY_HAVE_SSE(dual_inner_prod),
+  MAY_HAVE_SSE(dual_inner_prod)
+};
+
+void (*const COMB_FILTER_CONST_IMPL[OPUS_ARCHMASK + 1])(
+              opus_val32 *y,
+              opus_val32 *x,
+              int         T,
+              int         N,
+              opus_val16  g10,
+              opus_val16  g11,
+              opus_val16  g12
+) = {
+  comb_filter_const_c,                /* non-sse */
+  MAY_HAVE_SSE(comb_filter_const),
+  MAY_HAVE_SSE(comb_filter_const),
+  MAY_HAVE_SSE(comb_filter_const),
+  MAY_HAVE_SSE(comb_filter_const)
+};
+
+
+#endif
+
+#if defined(OPUS_X86_MAY_HAVE_SSE2) && !defined(OPUS_X86_PRESUME_SSE2)
+opus_val16 (*const OP_PVQ_SEARCH_IMPL[OPUS_ARCHMASK + 1])(
+      celt_norm *_X, int *iy, int K, int N, int arch
+) = {
+  op_pvq_search_c,                /* non-sse */
+  op_pvq_search_c,
+  MAY_HAVE_SSE2(op_pvq_search),
+  MAY_HAVE_SSE2(op_pvq_search),
+  MAY_HAVE_SSE2(op_pvq_search)
+};
+#endif
+
+#endif
+#endif
--- a/lib/libopus/celt/x86/x86cpu.c
+++ b/lib/libopus/celt/x86/x86cpu.c
@ -0,0 +1,184 @@
+/* Copyright (c) 2014, Cisco Systems, INC
+   Written by XiangMingZhu WeiZhou MinPeng YanWang
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "cpu_support.h"
+#include "macros.h"
+#include "main.h"
+#include "pitch.h"
+#include "x86cpu.h"
+
+#if defined(OPUS_HAVE_RTCD) && \
+  ((defined(OPUS_X86_MAY_HAVE_SSE) && !defined(OPUS_X86_PRESUME_SSE)) || \
+  (defined(OPUS_X86_MAY_HAVE_SSE2) && !defined(OPUS_X86_PRESUME_SSE2)) || \
+  (defined(OPUS_X86_MAY_HAVE_SSE4_1) && !defined(OPUS_X86_PRESUME_SSE4_1)) || \
+  (defined(OPUS_X86_MAY_HAVE_AVX2) && !defined(OPUS_X86_PRESUME_AVX2)))
+
+#if defined(_MSC_VER)
+
+#include <intrin.h>
+static _inline void cpuid(unsigned int CPUInfo[4], unsigned int InfoType)
+{
+    __cpuid((int*)CPUInfo, InfoType);
+}
+
+#else
+
+#if defined(CPU_INFO_BY_C)
+#include <cpuid.h>
+#endif
+
+static void cpuid(unsigned int CPUInfo[4], unsigned int InfoType)
+{
+#if defined(CPU_INFO_BY_ASM)
+#if defined(__i386__) && defined(__PIC__)
+/* %ebx is PIC register in 32-bit, so mustn't clobber it. */
+    __asm__ __volatile__ (
+        "xchg %%ebx, %1\n"
+        "cpuid\n"
+        "xchg %%ebx, %1\n":
+        "=a" (CPUInfo[0]),
+        "=r" (CPUInfo[1]),
+        "=c" (CPUInfo[2]),
+        "=d" (CPUInfo[3]) :
+        /* We clear ECX to avoid a valgrind false-positive prior to v3.17.0. */
+        "0" (InfoType), "2" (0)
+    );
+#else
+    __asm__ __volatile__ (
+        "cpuid":
+        "=a" (CPUInfo[0]),
+        "=b" (CPUInfo[1]),
+        "=c" (CPUInfo[2]),
+        "=d" (CPUInfo[3]) :
+        /* We clear ECX to avoid a valgrind false-positive prior to v3.17.0. */
+        "0" (InfoType), "2" (0)
+    );
+#endif
+#elif defined(CPU_INFO_BY_C)
+    /* We use __get_cpuid_count to clear ECX to avoid a valgrind false-positive
+        prior to v3.17.0.*/
+    if (!__get_cpuid_count(InfoType, 0, &(CPUInfo[0]), &(CPUInfo[1]), &(CPUInfo[2]), &(CPUInfo[3]))) {
+        /* Our function cannot fail, but __get_cpuid{_count} can.
+           Returning all zeroes will effectively disable all SIMD, which is
+            what we want on CPUs that don't support CPUID. */
+        CPUInfo[3] = CPUInfo[2] = CPUInfo[1] = CPUInfo[0] = 0;
+    }
+#else
+# error "Configured to use x86 RTCD, but no CPU detection method available. " \
+ "Reconfigure with --disable-rtcd (or send patches)."
+#endif
+}
+
+#endif
+
+typedef struct CPU_Feature{
+    /*  SIMD: 128-bit */
+    int HW_SSE;
+    int HW_SSE2;
+    int HW_SSE41;
+    /*  SIMD: 256-bit */
+    int HW_AVX2;
+} CPU_Feature;
+
+static void opus_cpu_feature_check(CPU_Feature *cpu_feature)
+{
+    unsigned int info[4];
+    unsigned int nIds = 0;
+
+    cpuid(info, 0);
+    nIds = info[0];
+
+    if (nIds >= 1){
+        cpuid(info, 1);
+        cpu_feature->HW_SSE = (info[3] & (1 << 25)) != 0;
+        cpu_feature->HW_SSE2 = (info[3] & (1 << 26)) != 0;
+        cpu_feature->HW_SSE41 = (info[2] & (1 << 19)) != 0;
+        cpu_feature->HW_AVX2 = (info[2] & (1 << 28)) != 0 && (info[2] & (1 << 12)) != 0;
+        if (cpu_feature->HW_AVX2 && nIds >= 7) {
+            cpuid(info, 7);
+            cpu_feature->HW_AVX2 = cpu_feature->HW_AVX2 && (info[1] & (1 << 5)) != 0;
+        } else {
+            cpu_feature->HW_AVX2 = 0;
+        }
+    }
+    else {
+        cpu_feature->HW_SSE = 0;
+        cpu_feature->HW_SSE2 = 0;
+        cpu_feature->HW_SSE41 = 0;
+        cpu_feature->HW_AVX2 = 0;
+    }
+}
+
+static int opus_select_arch_impl(void)
+{
+    CPU_Feature cpu_feature;
+    int arch;
+
+    opus_cpu_feature_check(&cpu_feature);
+
+    arch = 0;
+    if (!cpu_feature.HW_SSE)
+    {
+       return arch;
+    }
+    arch++;
+
+    if (!cpu_feature.HW_SSE2)
+    {
+       return arch;
+    }
+    arch++;
+
+    if (!cpu_feature.HW_SSE41)
+    {
+        return arch;
+    }
+    arch++;
+
+    if (!cpu_feature.HW_AVX2)
+    {
+        return arch;
+    }
+    arch++;
+
+    return arch;
+}
+
+int opus_select_arch(void) {
+    int arch = opus_select_arch_impl();
+#ifdef FUZZING
+    /* Randomly downgrade the architecture. */
+    arch = rand()%(arch+1);
+#endif
+    return arch;
+}
+
+#endif
--- a/lib/libopus/celt/x86/x86cpu.h
+++ b/lib/libopus/celt/x86/x86cpu.h
@ -0,0 +1,112 @@
+/* Copyright (c) 2014, Cisco Systems, INC
+   Written by XiangMingZhu WeiZhou MinPeng YanWang
+
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#if !defined(X86CPU_H)
+# define X86CPU_H
+
+# if defined(OPUS_X86_MAY_HAVE_SSE)
+#  define MAY_HAVE_SSE(name) name ## _sse
+# else
+#  define MAY_HAVE_SSE(name) name ## _c
+# endif
+
+# if defined(OPUS_X86_MAY_HAVE_SSE2)
+#  define MAY_HAVE_SSE2(name) name ## _sse2
+# else
+#  define MAY_HAVE_SSE2(name) name ## _c
+# endif
+
+# if defined(OPUS_X86_MAY_HAVE_SSE4_1)
+#  define MAY_HAVE_SSE4_1(name) name ## _sse4_1
+# else
+#  define MAY_HAVE_SSE4_1(name) name ## _c
+# endif
+
+# if defined(OPUS_X86_MAY_HAVE_AVX2)
+#  define MAY_HAVE_AVX2(name) name ## _avx2
+# else
+#  define MAY_HAVE_AVX2(name) name ## _c
+# endif
+
+# if defined(OPUS_HAVE_RTCD) && \
+  ((defined(OPUS_X86_MAY_HAVE_SSE) && !defined(OPUS_X86_PRESUME_SSE)) || \
+  (defined(OPUS_X86_MAY_HAVE_SSE2) && !defined(OPUS_X86_PRESUME_SSE2)) || \
+  (defined(OPUS_X86_MAY_HAVE_SSE4_1) && !defined(OPUS_X86_PRESUME_SSE4_1)) || \
+  (defined(OPUS_X86_MAY_HAVE_AVX2) && !defined(OPUS_X86_PRESUME_AVX2)))
+int opus_select_arch(void);
+# endif
+
+# if defined(OPUS_X86_MAY_HAVE_SSE2)
+#  include "opus_defines.h"
+
+/*MOVD should not impose any alignment restrictions, but the C standard does,
+   and UBSan will report errors if we actually make unaligned accesses.
+  Use this to work around those restrictions (which should hopefully all get
+   optimized to a single MOVD instruction).
+  GCC implemented _mm_loadu_si32() since GCC 11; HOWEVER, there is a bug!
+  https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99754
+  LLVM implemented _mm_loadu_si32() since Clang 8.0, however the
+   __clang_major__ version number macro is unreliable, as vendors
+   (specifically, Apple) will use different numbering schemes than upstream.
+  Clang's advice is "use feature detection", but they do not provide feature
+   detection support for specific SIMD functions.
+  We follow the approach from the SIMDe project and instead detect unrelated
+   features that should be available in the version we want (see
+   <https://github.com/simd-everywhere/simde/blob/master/simde/simde-detect-clang.h>).*/
+#  if defined(__clang__)
+#   if __has_warning("-Wextra-semi-stmt") || \
+ __has_builtin(__builtin_rotateleft32)
+#    define OPUS_CLANG_8 (1)
+#   endif
+#  endif
+#  if !defined(_MSC_VER) && !OPUS_GNUC_PREREQ(11,3) && !defined(OPUS_CLANG_8)
+#   include <string.h>
+#   include <emmintrin.h>
+
+#   ifdef _mm_loadu_si32
+#    undef _mm_loadu_si32
+#   endif
+#   define _mm_loadu_si32 WORKAROUND_mm_loadu_si32
+static inline __m128i WORKAROUND_mm_loadu_si32(void const* mem_addr) {
+  int val;
+  memcpy(&val, mem_addr, sizeof(val));
+  return _mm_cvtsi32_si128(val);
+}
+#  elif defined(_MSC_VER)
+    /* MSVC needs this for _mm_loadu_si32 */
+#   include <immintrin.h>
+#  endif
+
+#  define OP_CVTEPI8_EPI32_M32(x) \
+ (_mm_cvtepi8_epi32(_mm_loadu_si32(x)))
+
+#  define OP_CVTEPI16_EPI32_M64(x) \
+ (_mm_cvtepi16_epi32(_mm_loadl_epi64((__m128i *)(void*)(x))))
+
+# endif
+
+#endif
--- a/lib/libopus/include/opus.h
+++ b/lib/libopus/include/opus.h
--- a/lib/libopus/include/opus_custom.h
+++ b/lib/libopus/include/opus_custom.h
@ -0,0 +1,381 @@
+/* Copyright (c) 2007-2008 CSIRO
+   Copyright (c) 2007-2009 Xiph.Org Foundation
+   Copyright (c) 2008-2012 Gregory Maxwell
+   Written by Jean-Marc Valin and Gregory Maxwell */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+/**
+  @file opus_custom.h
+  @brief Opus-Custom reference implementation API
+ */
+
+#ifndef OPUS_CUSTOM_H
+#define OPUS_CUSTOM_H
+
+#include "opus_defines.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if defined(CUSTOM_MODES) || defined(ENABLE_OPUS_CUSTOM_API)
+# define OPUS_CUSTOM_EXPORT OPUS_EXPORT
+# define OPUS_CUSTOM_EXPORT_STATIC OPUS_EXPORT
+#else
+# define OPUS_CUSTOM_EXPORT
+# ifdef OPUS_BUILD
+#  define OPUS_CUSTOM_EXPORT_STATIC static OPUS_INLINE
+# else
+#  define OPUS_CUSTOM_EXPORT_STATIC
+# endif
+#endif
+
+/** @defgroup opus_custom Opus Custom
+  * @{
+  *  Opus Custom is an optional part of the Opus specification and
+  * reference implementation which uses a distinct API from the regular
+  * API and supports frame sizes that are not normally supported.\ Use
+  * of Opus Custom is discouraged for all but very special applications
+  * for which a frame size different from 2.5, 5, 10, or 20 ms is needed
+  * (for either complexity or latency reasons) and where interoperability
+  * is less important.
+  *
+  * In addition to the interoperability limitations the use of Opus custom
+  * disables a substantial chunk of the codec and generally lowers the
+  * quality available at a given bitrate. Normally when an application needs
+  * a different frame size from the codec it should buffer to match the
+  * sizes but this adds a small amount of delay which may be important
+  * in some very low latency applications. Some transports (especially
+  * constant rate RF transports) may also work best with frames of
+  * particular durations.
+  *
+  * Libopus only supports custom modes if they are enabled at compile time.
+  *
+  * The Opus Custom API is similar to the regular API but the
+  * @ref opus_encoder_create and @ref opus_decoder_create calls take
+  * an additional mode parameter which is a structure produced by
+  * a call to @ref opus_custom_mode_create. Both the encoder and decoder
+  * must create a mode using the same sample rate (fs) and frame size
+  * (frame size) so these parameters must either be signaled out of band
+  * or fixed in a particular implementation.
+  *
+  * Similar to regular Opus the custom modes support on the fly frame size
+  * switching, but the sizes available depend on the particular frame size in
+  * use. For some initial frame sizes on a single on the fly size is available.
+  */
+
+/** Contains the state of an encoder. One encoder state is needed
+    for each stream. It is initialized once at the beginning of the
+    stream. Do *not* re-initialize the state for every frame.
+   @brief Encoder state
+ */
+typedef struct OpusCustomEncoder OpusCustomEncoder;
+
+/** State of the decoder. One decoder state is needed for each stream.
+    It is initialized once at the beginning of the stream. Do *not*
+    re-initialize the state for every frame.
+   @brief Decoder state
+ */
+typedef struct OpusCustomDecoder OpusCustomDecoder;
+
+/** The mode contains all the information necessary to create an
+    encoder. Both the encoder and decoder need to be initialized
+    with exactly the same mode, otherwise the output will be
+    corrupted. The mode MUST NOT BE DESTROYED until the encoders and
+    decoders that use it are destroyed as well.
+   @brief Mode configuration
+ */
+typedef struct OpusCustomMode OpusCustomMode;
+
+/** Creates a new mode struct. This will be passed to an encoder or
+  * decoder. The mode MUST NOT BE DESTROYED until the encoders and
+  * decoders that use it are destroyed as well.
+  * @param [in] Fs <tt>int</tt>: Sampling rate (8000 to 96000 Hz)
+  * @param [in] frame_size <tt>int</tt>: Number of samples (per channel) to encode in each
+  *        packet (64 - 1024, prime factorization must contain zero or more 2s, 3s, or 5s and no other primes)
+  * @param [out] error <tt>int*</tt>: Returned error code (if NULL, no error will be returned)
+  * @return A newly created mode
+  */
+OPUS_CUSTOM_EXPORT OPUS_WARN_UNUSED_RESULT OpusCustomMode *opus_custom_mode_create(opus_int32 Fs, int frame_size, int *error);
+
+/** Destroys a mode struct. Only call this after all encoders and
+  * decoders using this mode are destroyed as well.
+  * @param [in] mode <tt>OpusCustomMode*</tt>: Mode to be freed.
+  */
+OPUS_CUSTOM_EXPORT void opus_custom_mode_destroy(OpusCustomMode *mode);
+
+
+#if !defined(OPUS_BUILD) || defined(CELT_ENCODER_C)
+
+/* Encoder */
+/** Gets the size of an OpusCustomEncoder structure.
+  * @param [in] mode <tt>OpusCustomMode *</tt>: Mode configuration
+  * @param [in] channels <tt>int</tt>: Number of channels
+  * @returns size
+  */
+OPUS_CUSTOM_EXPORT_STATIC OPUS_WARN_UNUSED_RESULT int opus_custom_encoder_get_size(
+    const OpusCustomMode *mode,
+    int channels
+) OPUS_ARG_NONNULL(1);
+
+#if defined(CUSTOM_MODES) || defined(ENABLE_OPUS_CUSTOM_API)
+/** Initializes a previously allocated encoder state
+  * The memory pointed to by st must be the size returned by opus_custom_encoder_get_size.
+  * This is intended for applications which use their own allocator instead of malloc.
+  * @see opus_custom_encoder_create(),opus_custom_encoder_get_size()
+  * To reset a previously initialized state use the OPUS_RESET_STATE CTL.
+  * @param [in] st <tt>OpusCustomEncoder*</tt>: Encoder state
+  * @param [in] mode <tt>OpusCustomMode *</tt>: Contains all the information about the characteristics of
+  *  the stream (must be the same characteristics as used for the
+  *  decoder)
+  * @param [in] channels <tt>int</tt>: Number of channels
+  * @return OPUS_OK Success or @ref opus_errorcodes
+  */
+OPUS_CUSTOM_EXPORT int opus_custom_encoder_init(
+    OpusCustomEncoder *st,
+    const OpusCustomMode *mode,
+    int channels
+) OPUS_ARG_NONNULL(1) OPUS_ARG_NONNULL(2);
+# endif
+#endif
+
+
+/** Creates a new encoder state. Each stream needs its own encoder
+  * state (can't be shared across simultaneous streams).
+  * @param [in] mode <tt>OpusCustomMode*</tt>: Contains all the information about the characteristics of
+  *  the stream (must be the same characteristics as used for the
+  *  decoder)
+  * @param [in] channels <tt>int</tt>: Number of channels
+  * @param [out] error <tt>int*</tt>: Returns an error code
+  * @return Newly created encoder state.
+*/
+OPUS_CUSTOM_EXPORT OPUS_WARN_UNUSED_RESULT OpusCustomEncoder *opus_custom_encoder_create(
+    const OpusCustomMode *mode,
+    int channels,
+    int *error
+) OPUS_ARG_NONNULL(1);
+
+
+/** Destroys an encoder state.
+  * @param[in] st <tt>OpusCustomEncoder*</tt>: State to be freed.
+  */
+OPUS_CUSTOM_EXPORT void opus_custom_encoder_destroy(OpusCustomEncoder *st);
+
+/** Encodes a frame of audio.
+  * @param [in] st <tt>OpusCustomEncoder*</tt>: Encoder state
+  * @param [in] pcm <tt>float*</tt>: PCM audio in float format, with a normal range of +/-1.0.
+  *          Samples with a range beyond +/-1.0 are supported but will
+  *          be clipped by decoders using the integer API and should
+  *          only be used if it is known that the far end supports
+  *          extended dynamic range. There must be exactly
+  *          frame_size samples per channel.
+  * @param [in] frame_size <tt>int</tt>: Number of samples per frame of input signal
+  * @param [out] compressed <tt>char *</tt>: The compressed data is written here. This may not alias pcm and must be at least maxCompressedBytes long.
+  * @param [in] maxCompressedBytes <tt>int</tt>: Maximum number of bytes to use for compressing the frame
+  *          (can change from one frame to another)
+  * @return Number of bytes written to "compressed".
+  *       If negative, an error has occurred (see error codes). It is IMPORTANT that
+  *       the length returned be somehow transmitted to the decoder. Otherwise, no
+  *       decoding is possible.
+  */
+OPUS_CUSTOM_EXPORT OPUS_WARN_UNUSED_RESULT int opus_custom_encode_float(
+    OpusCustomEncoder *st,
+    const float *pcm,
+    int frame_size,
+    unsigned char *compressed,
+    int maxCompressedBytes
+) OPUS_ARG_NONNULL(1) OPUS_ARG_NONNULL(2) OPUS_ARG_NONNULL(4);
+
+/** Encodes a frame of audio.
+  * @param [in] st <tt>OpusCustomEncoder*</tt>: Encoder state
+  * @param [in] pcm <tt>opus_int16*</tt>: PCM audio in signed 16-bit format (native endian).
+  *          There must be exactly frame_size samples per channel.
+  * @param [in] frame_size <tt>int</tt>: Number of samples per frame of input signal
+  * @param [out] compressed <tt>char *</tt>: The compressed data is written here. This may not alias pcm and must be at least maxCompressedBytes long.
+  * @param [in] maxCompressedBytes <tt>int</tt>: Maximum number of bytes to use for compressing the frame
+  *          (can change from one frame to another)
+  * @return Number of bytes written to "compressed".
+  *       If negative, an error has occurred (see error codes). It is IMPORTANT that
+  *       the length returned be somehow transmitted to the decoder. Otherwise, no
+  *       decoding is possible.
+ */
+OPUS_CUSTOM_EXPORT OPUS_WARN_UNUSED_RESULT int opus_custom_encode(
+    OpusCustomEncoder *st,
+    const opus_int16 *pcm,
+    int frame_size,
+    unsigned char *compressed,
+    int maxCompressedBytes
+) OPUS_ARG_NONNULL(1) OPUS_ARG_NONNULL(2) OPUS_ARG_NONNULL(4);
+
+/** Encodes a frame of audio.
+  * @param [in] st <tt>OpusCustomEncoder*</tt>: Encoder state
+  * @param [in] pcm <tt>opus_int32*</tt>: PCM audio in signed 32-bit format (native endian) representing (or slightly exceeding) 24-bit values.
+  *          There must be exactly frame_size samples per channel.
+  * @param [in] frame_size <tt>int</tt>: Number of samples per frame of input signal
+  * @param [out] compressed <tt>char *</tt>: The compressed data is written here. This may not alias pcm and must be at least maxCompressedBytes long.
+  * @param [in] maxCompressedBytes <tt>int</tt>: Maximum number of bytes to use for compressing the frame
+  *          (can change from one frame to another)
+  * @return Number of bytes written to "compressed".
+  *       If negative, an error has occurred (see error codes). It is IMPORTANT that
+  *       the length returned be somehow transmitted to the decoder. Otherwise, no
+  *       decoding is possible.
+ */
+OPUS_CUSTOM_EXPORT OPUS_WARN_UNUSED_RESULT int opus_custom_encode24(
+    OpusCustomEncoder *st,
+    const opus_int32 *pcm,
+    int frame_size,
+    unsigned char *compressed,
+    int maxCompressedBytes
+) OPUS_ARG_NONNULL(1) OPUS_ARG_NONNULL(2) OPUS_ARG_NONNULL(4);
+
+/** Perform a CTL function on an Opus custom encoder.
+  *
+  * Generally the request and subsequent arguments are generated
+  * by a convenience macro.
+  * @see opus_encoderctls
+  */
+OPUS_CUSTOM_EXPORT int opus_custom_encoder_ctl(OpusCustomEncoder * OPUS_RESTRICT st, int request, ...) OPUS_ARG_NONNULL(1);
+
+
+#if !defined(OPUS_BUILD) || defined(CELT_DECODER_C)
+/* Decoder */
+
+/** Gets the size of an OpusCustomDecoder structure.
+  * @param [in] mode <tt>OpusCustomMode *</tt>: Mode configuration
+  * @param [in] channels <tt>int</tt>: Number of channels
+  * @returns size
+  */
+OPUS_CUSTOM_EXPORT_STATIC OPUS_WARN_UNUSED_RESULT int opus_custom_decoder_get_size(
+    const OpusCustomMode *mode,
+    int channels
+) OPUS_ARG_NONNULL(1);
+
+/** Initializes a previously allocated decoder state
+  * The memory pointed to by st must be the size returned by opus_custom_decoder_get_size.
+  * This is intended for applications which use their own allocator instead of malloc.
+  * @see opus_custom_decoder_create(),opus_custom_decoder_get_size()
+  * To reset a previously initialized state use the OPUS_RESET_STATE CTL.
+  * @param [in] st <tt>OpusCustomDecoder*</tt>: Decoder state
+  * @param [in] mode <tt>OpusCustomMode *</tt>: Contains all the information about the characteristics of
+  *  the stream (must be the same characteristics as used for the
+  *  encoder)
+  * @param [in] channels <tt>int</tt>: Number of channels
+  * @return OPUS_OK Success or @ref opus_errorcodes
+  */
+OPUS_CUSTOM_EXPORT_STATIC int opus_custom_decoder_init(
+    OpusCustomDecoder *st,
+    const OpusCustomMode *mode,
+    int channels
+) OPUS_ARG_NONNULL(1) OPUS_ARG_NONNULL(2);
+
+#endif
+
+
+/** Creates a new decoder state. Each stream needs its own decoder state (can't
+  * be shared across simultaneous streams).
+  * @param [in] mode <tt>OpusCustomMode</tt>: Contains all the information about the characteristics of the
+  *          stream (must be the same characteristics as used for the encoder)
+  * @param [in] channels <tt>int</tt>: Number of channels
+  * @param [out] error <tt>int*</tt>: Returns an error code
+  * @return Newly created decoder state.
+  */
+OPUS_CUSTOM_EXPORT OPUS_WARN_UNUSED_RESULT OpusCustomDecoder *opus_custom_decoder_create(
+    const OpusCustomMode *mode,
+    int channels,
+    int *error
+) OPUS_ARG_NONNULL(1);
+
+/** Destroys a decoder state.
+  * @param[in] st <tt>OpusCustomDecoder*</tt>: State to be freed.
+  */
+OPUS_CUSTOM_EXPORT void opus_custom_decoder_destroy(OpusCustomDecoder *st);
+
+/** Decode an opus custom frame with floating point output
+  * @param [in] st <tt>OpusCustomDecoder*</tt>: Decoder state
+  * @param [in] data <tt>char*</tt>: Input payload. Use a NULL pointer to indicate packet loss
+  * @param [in] len <tt>int</tt>: Number of bytes in payload
+  * @param [out] pcm <tt>float*</tt>: Output signal (interleaved if 2 channels). length
+  *  is frame_size*channels*sizeof(float)
+  * @param [in] frame_size Number of samples per channel of available space in *pcm.
+  * @returns Number of decoded samples or @ref opus_errorcodes
+  */
+OPUS_CUSTOM_EXPORT OPUS_WARN_UNUSED_RESULT int opus_custom_decode_float(
+    OpusCustomDecoder *st,
+    const unsigned char *data,
+    int len,
+    float *pcm,
+    int frame_size
+) OPUS_ARG_NONNULL(1) OPUS_ARG_NONNULL(4);
+
+/** Decode an opus custom frame
+  * @param [in] st <tt>OpusCustomDecoder*</tt>: Decoder state
+  * @param [in] data <tt>char*</tt>: Input payload. Use a NULL pointer to indicate packet loss
+  * @param [in] len <tt>int</tt>: Number of bytes in payload
+  * @param [out] pcm <tt>opus_int16*</tt>: Output signal (interleaved if 2 channels). length
+  *  is frame_size*channels*sizeof(opus_int16)
+  * @param [in] frame_size Number of samples per channel of available space in *pcm.
+  * @returns Number of decoded samples or @ref opus_errorcodes
+  */
+OPUS_CUSTOM_EXPORT OPUS_WARN_UNUSED_RESULT int opus_custom_decode(
+    OpusCustomDecoder *st,
+    const unsigned char *data,
+    int len,
+    opus_int16 *pcm,
+    int frame_size
+) OPUS_ARG_NONNULL(1) OPUS_ARG_NONNULL(4);
+
+/** Decode an opus custom frame
+  * @param [in] st <tt>OpusCustomDecoder*</tt>: Decoder state
+  * @param [in] data <tt>char*</tt>: Input payload. Use a NULL pointer to indicate packet loss
+  * @param [in] len <tt>int</tt>: Number of bytes in payload
+  * @param [out] pcm <tt>opus_int32*</tt>: Output signal (interleaved if 2 channels) representing (or slightly exceeding) 24-bit values. length
+  *  is frame_size*channels*sizeof(opus_int32)
+  * @param [in] frame_size Number of samples per channel of available space in *pcm.
+  * @returns Number of decoded samples or @ref opus_errorcodes
+  */
+OPUS_CUSTOM_EXPORT OPUS_WARN_UNUSED_RESULT int opus_custom_decode24(
+    OpusCustomDecoder *st,
+    const unsigned char *data,
+    int len,
+    opus_int32 *pcm,
+    int frame_size
+) OPUS_ARG_NONNULL(1) OPUS_ARG_NONNULL(4);
+
+/** Perform a CTL function on an Opus custom decoder.
+  *
+  * Generally the request and subsequent arguments are generated
+  * by a convenience macro.
+  * @see opus_genericctls
+  */
+OPUS_CUSTOM_EXPORT int opus_custom_decoder_ctl(OpusCustomDecoder * OPUS_RESTRICT st, int request, ...) OPUS_ARG_NONNULL(1);
+
+/**@}*/
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* OPUS_CUSTOM_H */
--- a/lib/libopus/include/opus_defines.h
+++ b/lib/libopus/include/opus_defines.h
@ -0,0 +1,867 @@
+/* Copyright (c) 2010-2011 Xiph.Org Foundation, Skype Limited
+   Written by Jean-Marc Valin and Koen Vos */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+/**
+ * @file opus_defines.h
+ * @brief Opus reference implementation constants
+ */
+
+#ifndef OPUS_DEFINES_H
+#define OPUS_DEFINES_H
+
+#include "opus_types.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/** @defgroup opus_errorcodes Error codes
+ * @{
+ */
+/** No error @hideinitializer*/
+#define OPUS_OK                0
+/** One or more invalid/out of range arguments @hideinitializer*/
+#define OPUS_BAD_ARG          -1
+/** Not enough bytes allocated in the buffer @hideinitializer*/
+#define OPUS_BUFFER_TOO_SMALL -2
+/** An internal error was detected @hideinitializer*/
+#define OPUS_INTERNAL_ERROR   -3
+/** The compressed data passed is corrupted @hideinitializer*/
+#define OPUS_INVALID_PACKET   -4
+/** Invalid/unsupported request number @hideinitializer*/
+#define OPUS_UNIMPLEMENTED    -5
+/** An encoder or decoder structure is invalid or already freed @hideinitializer*/
+#define OPUS_INVALID_STATE    -6
+/** Memory allocation has failed @hideinitializer*/
+#define OPUS_ALLOC_FAIL       -7
+/**@}*/
+
+/** @cond OPUS_INTERNAL_DOC */
+/**Export control for opus functions */
+
+#ifndef OPUS_EXPORT
+# if defined(_WIN32)
+#  if defined(OPUS_BUILD) && defined(DLL_EXPORT)
+#   define OPUS_EXPORT __declspec(dllexport)
+#  else
+#   define OPUS_EXPORT
+#  endif
+# elif defined(__GNUC__) && defined(OPUS_BUILD)
+#  define OPUS_EXPORT __attribute__ ((visibility ("default")))
+# else
+#  define OPUS_EXPORT
+# endif
+#endif
+
+# if !defined(OPUS_GNUC_PREREQ)
+#  if defined(__GNUC__)&&defined(__GNUC_MINOR__)
+#   define OPUS_GNUC_PREREQ(_maj,_min) \
+ ((__GNUC__<<16)+__GNUC_MINOR__>=((_maj)<<16)+(_min))
+#  else
+#   define OPUS_GNUC_PREREQ(_maj,_min) 0
+#  endif
+# endif
+
+#if (!defined(__STDC_VERSION__) || (__STDC_VERSION__ < 199901L) )
+# if OPUS_GNUC_PREREQ(3,0)
+#  define OPUS_RESTRICT __restrict__
+# elif (defined(_MSC_VER) && _MSC_VER >= 1400)
+#  define OPUS_RESTRICT __restrict
+# else
+#  define OPUS_RESTRICT
+# endif
+#else
+# define OPUS_RESTRICT restrict
+#endif
+
+#if (!defined(__STDC_VERSION__) || (__STDC_VERSION__ < 199901L) )
+# if OPUS_GNUC_PREREQ(2,7)
+#  define OPUS_INLINE __inline__
+# elif (defined(_MSC_VER))
+#  define OPUS_INLINE __inline
+# else
+#  define OPUS_INLINE
+# endif
+#else
+# define OPUS_INLINE inline
+#endif
+
+/**Warning attributes for opus functions
+  * NONNULL is not used in OPUS_BUILD to avoid the compiler optimizing out
+  * some paranoid null checks. */
+#if defined(__GNUC__) && OPUS_GNUC_PREREQ(3, 4)
+# define OPUS_WARN_UNUSED_RESULT __attribute__ ((__warn_unused_result__))
+#else
+# define OPUS_WARN_UNUSED_RESULT
+#endif
+#if !defined(OPUS_BUILD) && defined(__GNUC__) && OPUS_GNUC_PREREQ(3, 4)
+# define OPUS_ARG_NONNULL(_x)  __attribute__ ((__nonnull__(_x)))
+#else
+# define OPUS_ARG_NONNULL(_x)
+#endif
+
+/** These are the actual Encoder CTL ID numbers.
+  * They should not be used directly by applications.
+  * In general, SETs should be even and GETs should be odd.*/
+#define OPUS_SET_APPLICATION_REQUEST         4000
+#define OPUS_GET_APPLICATION_REQUEST         4001
+#define OPUS_SET_BITRATE_REQUEST             4002
+#define OPUS_GET_BITRATE_REQUEST             4003
+#define OPUS_SET_MAX_BANDWIDTH_REQUEST       4004
+#define OPUS_GET_MAX_BANDWIDTH_REQUEST       4005
+#define OPUS_SET_VBR_REQUEST                 4006
+#define OPUS_GET_VBR_REQUEST                 4007
+#define OPUS_SET_BANDWIDTH_REQUEST           4008
+#define OPUS_GET_BANDWIDTH_REQUEST           4009
+#define OPUS_SET_COMPLEXITY_REQUEST          4010
+#define OPUS_GET_COMPLEXITY_REQUEST          4011
+#define OPUS_SET_INBAND_FEC_REQUEST          4012
+#define OPUS_GET_INBAND_FEC_REQUEST          4013
+#define OPUS_SET_PACKET_LOSS_PERC_REQUEST    4014
+#define OPUS_GET_PACKET_LOSS_PERC_REQUEST    4015
+#define OPUS_SET_DTX_REQUEST                 4016
+#define OPUS_GET_DTX_REQUEST                 4017
+#define OPUS_SET_VBR_CONSTRAINT_REQUEST      4020
+#define OPUS_GET_VBR_CONSTRAINT_REQUEST      4021
+#define OPUS_SET_FORCE_CHANNELS_REQUEST      4022
+#define OPUS_GET_FORCE_CHANNELS_REQUEST      4023
+#define OPUS_SET_SIGNAL_REQUEST              4024
+#define OPUS_GET_SIGNAL_REQUEST              4025
+#define OPUS_GET_LOOKAHEAD_REQUEST           4027
+/* #define OPUS_RESET_STATE 4028 */
+#define OPUS_GET_SAMPLE_RATE_REQUEST         4029
+#define OPUS_GET_FINAL_RANGE_REQUEST         4031
+#define OPUS_GET_PITCH_REQUEST               4033
+#define OPUS_SET_GAIN_REQUEST                4034
+#define OPUS_GET_GAIN_REQUEST                4045 /* Should have been 4035 */
+#define OPUS_SET_LSB_DEPTH_REQUEST           4036
+#define OPUS_GET_LSB_DEPTH_REQUEST           4037
+#define OPUS_GET_LAST_PACKET_DURATION_REQUEST 4039
+#define OPUS_SET_EXPERT_FRAME_DURATION_REQUEST 4040
+#define OPUS_GET_EXPERT_FRAME_DURATION_REQUEST 4041
+#define OPUS_SET_PREDICTION_DISABLED_REQUEST 4042
+#define OPUS_GET_PREDICTION_DISABLED_REQUEST 4043
+/* Don't use 4045, it's already taken by OPUS_GET_GAIN_REQUEST */
+#define OPUS_SET_PHASE_INVERSION_DISABLED_REQUEST 4046
+#define OPUS_GET_PHASE_INVERSION_DISABLED_REQUEST 4047
+#define OPUS_GET_IN_DTX_REQUEST              4049
+#define OPUS_SET_DRED_DURATION_REQUEST 4050
+#define OPUS_GET_DRED_DURATION_REQUEST 4051
+#define OPUS_SET_DNN_BLOB_REQUEST 4052
+/*#define OPUS_GET_DNN_BLOB_REQUEST 4053 */
+#define OPUS_SET_OSCE_BWE_REQUEST 4054
+#define OPUS_GET_OSCE_BWE_REQUEST 4055
+#define OPUS_SET_QEXT_REQUEST 4056
+#define OPUS_GET_QEXT_REQUEST 4057
+#define OPUS_SET_IGNORE_EXTENSIONS_REQUEST 4058
+#define OPUS_GET_IGNORE_EXTENSIONS_REQUEST 4059
+
+/** Defines for the presence of extended APIs. */
+#define OPUS_HAVE_OPUS_PROJECTION_H
+
+/* Macros to trigger compilation errors when the wrong types are provided to a CTL */
+#define opus_check_int(x) (((void)((x) == (opus_int32)0)), (opus_int32)(x))
+
+#ifdef DISABLE_PTR_CHECK
+/* Disable checks to prevent ubsan from complaining about NULL checks
+   in test_opus_api. */
+#define opus_check_int_ptr(ptr) (ptr)
+#define opus_check_uint_ptr(ptr) (ptr)
+#define opus_check_uint8_ptr(ptr) (ptr)
+#define opus_check_val16_ptr(ptr) (ptr)
+#define opus_check_void_ptr(ptr) (ptr)
+#else
+#define opus_check_int_ptr(ptr) ((ptr) + ((ptr) - (opus_int32*)(ptr)))
+#define opus_check_uint_ptr(ptr) ((ptr) + ((ptr) - (opus_uint32*)(ptr)))
+#define opus_check_uint8_ptr(ptr) ((ptr) + ((ptr) - (opus_uint8*)(ptr)))
+#define opus_check_val16_ptr(ptr) ((ptr) + ((ptr) - (opus_val16*)(ptr)))
+#define opus_check_void_ptr(x) ((void)((void *)0 == (x)), (x))
+#endif
+/** @endcond */
+
+/** @defgroup opus_ctlvalues Pre-defined values for CTL interface
+  * @see opus_genericctls, opus_encoderctls
+  * @{
+  */
+/* Values for the various encoder CTLs */
+#define OPUS_AUTO                           -1000 /**<Auto/default setting @hideinitializer*/
+#define OPUS_BITRATE_MAX                       -1 /**<Maximum bitrate @hideinitializer*/
+
+/** Best for most VoIP/videoconference applications where listening quality and intelligibility matter most
+ * @hideinitializer */
+#define OPUS_APPLICATION_VOIP                2048
+/** Best for broadcast/high-fidelity application where the decoded audio should be as close as possible to the input
+ * @hideinitializer */
+#define OPUS_APPLICATION_AUDIO               2049
+/** Only use when lowest-achievable latency is what matters most. Voice-optimized modes cannot be used.
+ * @hideinitializer */
+#define OPUS_APPLICATION_RESTRICTED_LOWDELAY 2051
+/** Experts only: forces SILK encoding; don't allocate CELT state at all. Disables OPUS_SET_APPLICATION. */
+#define OPUS_APPLICATION_RESTRICTED_SILK     2052
+/** Experts only: forces CELT encoding; don't allocate SILK state at all. Disables OPUS_SET_APPLICATION. */
+#define OPUS_APPLICATION_RESTRICTED_CELT     2053
+
+#define OPUS_SIGNAL_VOICE                    3001 /**< Signal being encoded is voice */
+#define OPUS_SIGNAL_MUSIC                    3002 /**< Signal being encoded is music */
+#define OPUS_BANDWIDTH_NARROWBAND            1101 /**< 4 kHz bandpass @hideinitializer*/
+#define OPUS_BANDWIDTH_MEDIUMBAND            1102 /**< 6 kHz bandpass @hideinitializer*/
+#define OPUS_BANDWIDTH_WIDEBAND              1103 /**< 8 kHz bandpass @hideinitializer*/
+#define OPUS_BANDWIDTH_SUPERWIDEBAND         1104 /**<12 kHz bandpass @hideinitializer*/
+#define OPUS_BANDWIDTH_FULLBAND              1105 /**<20 kHz bandpass @hideinitializer*/
+
+#define OPUS_FRAMESIZE_ARG                   5000 /**< Select frame size from the argument (default) */
+#define OPUS_FRAMESIZE_2_5_MS                5001 /**< Use 2.5 ms frames */
+#define OPUS_FRAMESIZE_5_MS                  5002 /**< Use 5 ms frames */
+#define OPUS_FRAMESIZE_10_MS                 5003 /**< Use 10 ms frames */
+#define OPUS_FRAMESIZE_20_MS                 5004 /**< Use 20 ms frames */
+#define OPUS_FRAMESIZE_40_MS                 5005 /**< Use 40 ms frames */
+#define OPUS_FRAMESIZE_60_MS                 5006 /**< Use 60 ms frames */
+#define OPUS_FRAMESIZE_80_MS                 5007 /**< Use 80 ms frames */
+#define OPUS_FRAMESIZE_100_MS                5008 /**< Use 100 ms frames */
+#define OPUS_FRAMESIZE_120_MS                5009 /**< Use 120 ms frames */
+
+/**@}*/
+
+
+/** @defgroup opus_encoderctls Encoder related CTLs
+  *
+  * These are convenience macros for use with the \c opus_encode_ctl
+  * interface. They are used to generate the appropriate series of
+  * arguments for that call, passing the correct type, size and so
+  * on as expected for each particular request.
+  *
+  * Some usage examples:
+  *
+  * @code
+  * int ret;
+  * ret = opus_encoder_ctl(enc_ctx, OPUS_SET_BANDWIDTH(OPUS_AUTO));
+  * if (ret != OPUS_OK) return ret;
+  *
+  * opus_int32 rate;
+  * opus_encoder_ctl(enc_ctx, OPUS_GET_BANDWIDTH(&rate));
+  *
+  * opus_encoder_ctl(enc_ctx, OPUS_RESET_STATE);
+  * @endcode
+  *
+  * @see opus_genericctls, opus_encoder
+  * @{
+  */
+
+/** Configures the encoder's computational complexity.
+  * The supported range is 0-10 inclusive with 10 representing the highest complexity.
+  * @see OPUS_GET_COMPLEXITY
+  * @param[in] x <tt>opus_int32</tt>: Allowed values: 0-10, inclusive.
+  *
+  * @hideinitializer */
+#define OPUS_SET_COMPLEXITY(x) OPUS_SET_COMPLEXITY_REQUEST, opus_check_int(x)
+/** Gets the encoder's complexity configuration.
+  * @see OPUS_SET_COMPLEXITY
+  * @param[out] x <tt>opus_int32 *</tt>: Returns a value in the range 0-10,
+  *                                      inclusive.
+  * @hideinitializer */
+#define OPUS_GET_COMPLEXITY(x) OPUS_GET_COMPLEXITY_REQUEST, opus_check_int_ptr(x)
+
+/** Configures the bitrate in the encoder.
+  * Rates from 500 to 512000 bits per second are meaningful, as well as the
+  * special values #OPUS_AUTO and #OPUS_BITRATE_MAX.
+  * The value #OPUS_BITRATE_MAX can be used to cause the codec to use as much
+  * rate as it can, which is useful for controlling the rate by adjusting the
+  * output buffer size.
+  * @see OPUS_GET_BITRATE
+  * @param[in] x <tt>opus_int32</tt>: Bitrate in bits per second. The default
+  *                                   is determined based on the number of
+  *                                   channels and the input sampling rate.
+  * @hideinitializer */
+#define OPUS_SET_BITRATE(x) OPUS_SET_BITRATE_REQUEST, opus_check_int(x)
+/** Gets the encoder's bitrate configuration.
+  * @see OPUS_SET_BITRATE
+  * @param[out] x <tt>opus_int32 *</tt>: Returns the bitrate in bits per second.
+  *                                      The default is determined based on the
+  *                                      number of channels and the input
+  *                                      sampling rate.
+  * @hideinitializer */
+#define OPUS_GET_BITRATE(x) OPUS_GET_BITRATE_REQUEST, opus_check_int_ptr(x)
+
+/** Enables or disables variable bitrate (VBR) in the encoder.
+  * The configured bitrate may not be met exactly because frames must
+  * be an integer number of bytes in length.
+  * @see OPUS_GET_VBR
+  * @see OPUS_SET_VBR_CONSTRAINT
+  * @param[in] x <tt>opus_int32</tt>: Allowed values:
+  * <dl>
+  * <dt>0</dt><dd>Hard CBR. For LPC/hybrid modes at very low bit-rate, this can
+  *               cause noticeable quality degradation.</dd>
+  * <dt>1</dt><dd>VBR (default). The exact type of VBR is controlled by
+  *               #OPUS_SET_VBR_CONSTRAINT.</dd>
+  * </dl>
+  * @hideinitializer */
+#define OPUS_SET_VBR(x) OPUS_SET_VBR_REQUEST, opus_check_int(x)
+/** Determine if variable bitrate (VBR) is enabled in the encoder.
+  * @see OPUS_SET_VBR
+  * @see OPUS_GET_VBR_CONSTRAINT
+  * @param[out] x <tt>opus_int32 *</tt>: Returns one of the following values:
+  * <dl>
+  * <dt>0</dt><dd>Hard CBR.</dd>
+  * <dt>1</dt><dd>VBR (default). The exact type of VBR may be retrieved via
+  *               #OPUS_GET_VBR_CONSTRAINT.</dd>
+  * </dl>
+  * @hideinitializer */
+#define OPUS_GET_VBR(x) OPUS_GET_VBR_REQUEST, opus_check_int_ptr(x)
+
+/** Enables or disables constrained VBR in the encoder.
+  * This setting is ignored when the encoder is in CBR mode.
+  * @warning Only the MDCT mode of Opus currently heeds the constraint.
+  *  Speech mode ignores it completely, hybrid mode may fail to obey it
+  *  if the LPC layer uses more bitrate than the constraint would have
+  *  permitted.
+  * @see OPUS_GET_VBR_CONSTRAINT
+  * @see OPUS_SET_VBR
+  * @param[in] x <tt>opus_int32</tt>: Allowed values:
+  * <dl>
+  * <dt>0</dt><dd>Unconstrained VBR.</dd>
+  * <dt>1</dt><dd>Constrained VBR (default). This creates a maximum of one
+  *               frame of buffering delay assuming a transport with a
+  *               serialization speed of the nominal bitrate.</dd>
+  * </dl>
+  * @hideinitializer */
+#define OPUS_SET_VBR_CONSTRAINT(x) OPUS_SET_VBR_CONSTRAINT_REQUEST, opus_check_int(x)
+/** Determine if constrained VBR is enabled in the encoder.
+  * @see OPUS_SET_VBR_CONSTRAINT
+  * @see OPUS_GET_VBR
+  * @param[out] x <tt>opus_int32 *</tt>: Returns one of the following values:
+  * <dl>
+  * <dt>0</dt><dd>Unconstrained VBR.</dd>
+  * <dt>1</dt><dd>Constrained VBR (default).</dd>
+  * </dl>
+  * @hideinitializer */
+#define OPUS_GET_VBR_CONSTRAINT(x) OPUS_GET_VBR_CONSTRAINT_REQUEST, opus_check_int_ptr(x)
+
+/** Configures mono/stereo forcing in the encoder.
+  * This can force the encoder to produce packets encoded as either mono or
+  * stereo, regardless of the format of the input audio. This is useful when
+  * the caller knows that the input signal is currently a mono source embedded
+  * in a stereo stream.
+  * @see OPUS_GET_FORCE_CHANNELS
+  * @param[in] x <tt>opus_int32</tt>: Allowed values:
+  * <dl>
+  * <dt>#OPUS_AUTO</dt><dd>Not forced (default)</dd>
+  * <dt>1</dt>         <dd>Forced mono</dd>
+  * <dt>2</dt>         <dd>Forced stereo</dd>
+  * </dl>
+  * @hideinitializer */
+#define OPUS_SET_FORCE_CHANNELS(x) OPUS_SET_FORCE_CHANNELS_REQUEST, opus_check_int(x)
+/** Gets the encoder's forced channel configuration.
+  * @see OPUS_SET_FORCE_CHANNELS
+  * @param[out] x <tt>opus_int32 *</tt>:
+  * <dl>
+  * <dt>#OPUS_AUTO</dt><dd>Not forced (default)</dd>
+  * <dt>1</dt>         <dd>Forced mono</dd>
+  * <dt>2</dt>         <dd>Forced stereo</dd>
+  * </dl>
+  * @hideinitializer */
+#define OPUS_GET_FORCE_CHANNELS(x) OPUS_GET_FORCE_CHANNELS_REQUEST, opus_check_int_ptr(x)
+
+/** Configures the maximum bandpass that the encoder will select automatically.
+  * Applications should normally use this instead of #OPUS_SET_BANDWIDTH
+  * (leaving that set to the default, #OPUS_AUTO). This allows the
+  * application to set an upper bound based on the type of input it is
+  * providing, but still gives the encoder the freedom to reduce the bandpass
+  * when the bitrate becomes too low, for better overall quality.
+  * @see OPUS_GET_MAX_BANDWIDTH
+  * @param[in] x <tt>opus_int32</tt>: Allowed values:
+  * <dl>
+  * <dt>OPUS_BANDWIDTH_NARROWBAND</dt>    <dd>4 kHz passband</dd>
+  * <dt>OPUS_BANDWIDTH_MEDIUMBAND</dt>    <dd>6 kHz passband</dd>
+  * <dt>OPUS_BANDWIDTH_WIDEBAND</dt>      <dd>8 kHz passband</dd>
+  * <dt>OPUS_BANDWIDTH_SUPERWIDEBAND</dt><dd>12 kHz passband</dd>
+  * <dt>OPUS_BANDWIDTH_FULLBAND</dt>     <dd>20 kHz passband (default)</dd>
+  * </dl>
+  * @hideinitializer */
+#define OPUS_SET_MAX_BANDWIDTH(x) OPUS_SET_MAX_BANDWIDTH_REQUEST, opus_check_int(x)
+
+/** Gets the encoder's configured maximum allowed bandpass.
+  * @see OPUS_SET_MAX_BANDWIDTH
+  * @param[out] x <tt>opus_int32 *</tt>: Allowed values:
+  * <dl>
+  * <dt>#OPUS_BANDWIDTH_NARROWBAND</dt>    <dd>4 kHz passband</dd>
+  * <dt>#OPUS_BANDWIDTH_MEDIUMBAND</dt>    <dd>6 kHz passband</dd>
+  * <dt>#OPUS_BANDWIDTH_WIDEBAND</dt>      <dd>8 kHz passband</dd>
+  * <dt>#OPUS_BANDWIDTH_SUPERWIDEBAND</dt><dd>12 kHz passband</dd>
+  * <dt>#OPUS_BANDWIDTH_FULLBAND</dt>     <dd>20 kHz passband (default)</dd>
+  * </dl>
+  * @hideinitializer */
+#define OPUS_GET_MAX_BANDWIDTH(x) OPUS_GET_MAX_BANDWIDTH_REQUEST, opus_check_int_ptr(x)
+
+/** Sets the encoder's bandpass to a specific value.
+  * This prevents the encoder from automatically selecting the bandpass based
+  * on the available bitrate. If an application knows the bandpass of the input
+  * audio it is providing, it should normally use #OPUS_SET_MAX_BANDWIDTH
+  * instead, which still gives the encoder the freedom to reduce the bandpass
+  * when the bitrate becomes too low, for better overall quality.
+  * @see OPUS_GET_BANDWIDTH
+  * @param[in] x <tt>opus_int32</tt>: Allowed values:
+  * <dl>
+  * <dt>#OPUS_AUTO</dt>                    <dd>(default)</dd>
+  * <dt>#OPUS_BANDWIDTH_NARROWBAND</dt>    <dd>4 kHz passband</dd>
+  * <dt>#OPUS_BANDWIDTH_MEDIUMBAND</dt>    <dd>6 kHz passband</dd>
+  * <dt>#OPUS_BANDWIDTH_WIDEBAND</dt>      <dd>8 kHz passband</dd>
+  * <dt>#OPUS_BANDWIDTH_SUPERWIDEBAND</dt><dd>12 kHz passband</dd>
+  * <dt>#OPUS_BANDWIDTH_FULLBAND</dt>     <dd>20 kHz passband</dd>
+  * </dl>
+  * @hideinitializer */
+#define OPUS_SET_BANDWIDTH(x) OPUS_SET_BANDWIDTH_REQUEST, opus_check_int(x)
+
+/** Configures the type of signal being encoded.
+  * This is a hint which helps the encoder's mode selection.
+  * @see OPUS_GET_SIGNAL
+  * @param[in] x <tt>opus_int32</tt>: Allowed values:
+  * <dl>
+  * <dt>#OPUS_AUTO</dt>        <dd>(default)</dd>
+  * <dt>#OPUS_SIGNAL_VOICE</dt><dd>Bias thresholds towards choosing LPC or Hybrid modes.</dd>
+  * <dt>#OPUS_SIGNAL_MUSIC</dt><dd>Bias thresholds towards choosing MDCT modes.</dd>
+  * </dl>
+  * @hideinitializer */
+#define OPUS_SET_SIGNAL(x) OPUS_SET_SIGNAL_REQUEST, opus_check_int(x)
+/** Gets the encoder's configured signal type.
+  * @see OPUS_SET_SIGNAL
+  * @param[out] x <tt>opus_int32 *</tt>: Returns one of the following values:
+  * <dl>
+  * <dt>#OPUS_AUTO</dt>        <dd>(default)</dd>
+  * <dt>#OPUS_SIGNAL_VOICE</dt><dd>Bias thresholds towards choosing LPC or Hybrid modes.</dd>
+  * <dt>#OPUS_SIGNAL_MUSIC</dt><dd>Bias thresholds towards choosing MDCT modes.</dd>
+  * </dl>
+  * @hideinitializer */
+#define OPUS_GET_SIGNAL(x) OPUS_GET_SIGNAL_REQUEST, opus_check_int_ptr(x)
+
+
+/** Configures the encoder's intended application.
+  * The initial value is a mandatory argument to the encoder_create function.
+  * @see OPUS_GET_APPLICATION
+  * @param[in] x <tt>opus_int32</tt>: Returns one of the following values:
+  * <dl>
+  * <dt>#OPUS_APPLICATION_VOIP</dt>
+  * <dd>Process signal for improved speech intelligibility.</dd>
+  * <dt>#OPUS_APPLICATION_AUDIO</dt>
+  * <dd>Favor faithfulness to the original input.</dd>
+  * <dt>#OPUS_APPLICATION_RESTRICTED_LOWDELAY</dt>
+  * <dd>Configure the minimum possible coding delay by disabling certain modes
+  * of operation.</dd>
+  * </dl>
+  * @hideinitializer */
+#define OPUS_SET_APPLICATION(x) OPUS_SET_APPLICATION_REQUEST, opus_check_int(x)
+/** Gets the encoder's configured application.
+  * @see OPUS_SET_APPLICATION
+  * @param[out] x <tt>opus_int32 *</tt>: Returns one of the following values:
+  * <dl>
+  * <dt>#OPUS_APPLICATION_VOIP</dt>
+  * <dd>Process signal for improved speech intelligibility.</dd>
+  * <dt>#OPUS_APPLICATION_AUDIO</dt>
+  * <dd>Favor faithfulness to the original input.</dd>
+  * <dt>#OPUS_APPLICATION_RESTRICTED_LOWDELAY</dt>
+  * <dd>Configure the minimum possible coding delay by disabling certain modes
+  * of operation.</dd>
+  * </dl>
+  * @hideinitializer */
+#define OPUS_GET_APPLICATION(x) OPUS_GET_APPLICATION_REQUEST, opus_check_int_ptr(x)
+
+/** Gets the total samples of delay added by the entire codec.
+  * This can be queried by the encoder and then the provided number of samples can be
+  * skipped on from the start of the decoder's output to provide time aligned input
+  * and output. From the perspective of a decoding application the real data begins this many
+  * samples late.
+  *
+  * The decoder contribution to this delay is identical for all decoders, but the
+  * encoder portion of the delay may vary from implementation to implementation,
+  * version to version, or even depend on the encoder's initial configuration.
+  * Applications needing delay compensation should call this CTL rather than
+  * hard-coding a value.
+  * @param[out] x <tt>opus_int32 *</tt>:   Number of lookahead samples
+  * @hideinitializer */
+#define OPUS_GET_LOOKAHEAD(x) OPUS_GET_LOOKAHEAD_REQUEST, opus_check_int_ptr(x)
+
+/** Configures the encoder's use of inband forward error correction (FEC).
+  * @note This is only applicable to the LPC layer
+  * @see OPUS_GET_INBAND_FEC
+  * @param[in] x <tt>opus_int32</tt>: Allowed values:
+  * <dl>
+  * <dt>0</dt><dd>Disable inband FEC (default).</dd>
+  * <dt>1</dt><dd>Inband FEC enabled. If the packet loss rate is sufficiently high, Opus will automatically switch to SILK even at high rates to enable use of that FEC.</dd>
+  * <dt>2</dt><dd>Inband FEC enabled, but does not necessarily switch to SILK if we have music.</dd>
+  * </dl>
+  * @hideinitializer */
+#define OPUS_SET_INBAND_FEC(x) OPUS_SET_INBAND_FEC_REQUEST, opus_check_int(x)
+/** Gets encoder's configured use of inband forward error correction.
+  * @see OPUS_SET_INBAND_FEC
+  * @param[out] x <tt>opus_int32 *</tt>: Returns one of the following values:
+  * <dl>
+  * <dt>0</dt><dd>Inband FEC disabled (default).</dd>
+  * <dt>1</dt><dd>Inband FEC enabled. If the packet loss rate is sufficiently high, Opus will automatically switch to SILK even at high rates to enable use of that FEC.</dd>
+  * <dt>2</dt><dd>Inband FEC enabled, but does not necessarily switch to SILK if we have music.</dd>
+  * </dl>
+  * @hideinitializer */
+#define OPUS_GET_INBAND_FEC(x) OPUS_GET_INBAND_FEC_REQUEST, opus_check_int_ptr(x)
+
+/** Configures the encoder's expected packet loss percentage.
+  * Higher values trigger progressively more loss resistant behavior in the encoder
+  * at the expense of quality at a given bitrate in the absence of packet loss, but
+  * greater quality under loss.
+  * @see OPUS_GET_PACKET_LOSS_PERC
+  * @param[in] x <tt>opus_int32</tt>:   Loss percentage in the range 0-100, inclusive (default: 0).
+  * @hideinitializer */
+#define OPUS_SET_PACKET_LOSS_PERC(x) OPUS_SET_PACKET_LOSS_PERC_REQUEST, opus_check_int(x)
+/** Gets the encoder's configured packet loss percentage.
+  * @see OPUS_SET_PACKET_LOSS_PERC
+  * @param[out] x <tt>opus_int32 *</tt>: Returns the configured loss percentage
+  *                                      in the range 0-100, inclusive (default: 0).
+  * @hideinitializer */
+#define OPUS_GET_PACKET_LOSS_PERC(x) OPUS_GET_PACKET_LOSS_PERC_REQUEST, opus_check_int_ptr(x)
+
+/** Configures the encoder's use of discontinuous transmission (DTX).
+  * @note This is only applicable to the LPC layer
+  * @see OPUS_GET_DTX
+  * @param[in] x <tt>opus_int32</tt>: Allowed values:
+  * <dl>
+  * <dt>0</dt><dd>Disable DTX (default).</dd>
+  * <dt>1</dt><dd>Enabled DTX.</dd>
+  * </dl>
+  * @hideinitializer */
+#define OPUS_SET_DTX(x) OPUS_SET_DTX_REQUEST, opus_check_int(x)
+/** Gets encoder's configured use of discontinuous transmission.
+  * @see OPUS_SET_DTX
+  * @param[out] x <tt>opus_int32 *</tt>: Returns one of the following values:
+  * <dl>
+  * <dt>0</dt><dd>DTX disabled (default).</dd>
+  * <dt>1</dt><dd>DTX enabled.</dd>
+  * </dl>
+  * @hideinitializer */
+#define OPUS_GET_DTX(x) OPUS_GET_DTX_REQUEST, opus_check_int_ptr(x)
+/** Configures the depth of signal being encoded.
+  *
+  * This is a hint which helps the encoder identify silence and near-silence.
+  * It represents the number of significant bits of linear intensity below
+  * which the signal contains ignorable quantization or other noise.
+  *
+  * For example, OPUS_SET_LSB_DEPTH(14) would be an appropriate setting
+  * for G.711 u-law input. OPUS_SET_LSB_DEPTH(16) would be appropriate
+  * for 16-bit linear pcm input with opus_encode_float().
+  *
+  * When using opus_encode() instead of opus_encode_float(), or when libopus
+  * is compiled for fixed-point, the encoder uses the minimum of the value
+  * set here and the value 16.
+  *
+  * @see OPUS_GET_LSB_DEPTH
+  * @param[in] x <tt>opus_int32</tt>: Input precision in bits, between 8 and 24
+  *                                   (default: 24).
+  * @hideinitializer */
+#define OPUS_SET_LSB_DEPTH(x) OPUS_SET_LSB_DEPTH_REQUEST, opus_check_int(x)
+/** Gets the encoder's configured signal depth.
+  * @see OPUS_SET_LSB_DEPTH
+  * @param[out] x <tt>opus_int32 *</tt>: Input precision in bits, between 8 and
+  *                                      24 (default: 24).
+  * @hideinitializer */
+#define OPUS_GET_LSB_DEPTH(x) OPUS_GET_LSB_DEPTH_REQUEST, opus_check_int_ptr(x)
+
+/** Configures the encoder's use of variable duration frames.
+  * When variable duration is enabled, the encoder is free to use a shorter frame
+  * size than the one requested in the opus_encode*() call.
+  * It is then the user's responsibility
+  * to verify how much audio was encoded by checking the ToC byte of the encoded
+  * packet. The part of the audio that was not encoded needs to be resent to the
+  * encoder for the next call. Do not use this option unless you <b>really</b>
+  * know what you are doing.
+  * @see OPUS_GET_EXPERT_FRAME_DURATION
+  * @param[in] x <tt>opus_int32</tt>: Allowed values:
+  * <dl>
+  * <dt>OPUS_FRAMESIZE_ARG</dt><dd>Select frame size from the argument (default).</dd>
+  * <dt>OPUS_FRAMESIZE_2_5_MS</dt><dd>Use 2.5 ms frames.</dd>
+  * <dt>OPUS_FRAMESIZE_5_MS</dt><dd>Use 5 ms frames.</dd>
+  * <dt>OPUS_FRAMESIZE_10_MS</dt><dd>Use 10 ms frames.</dd>
+  * <dt>OPUS_FRAMESIZE_20_MS</dt><dd>Use 20 ms frames.</dd>
+  * <dt>OPUS_FRAMESIZE_40_MS</dt><dd>Use 40 ms frames.</dd>
+  * <dt>OPUS_FRAMESIZE_60_MS</dt><dd>Use 60 ms frames.</dd>
+  * <dt>OPUS_FRAMESIZE_80_MS</dt><dd>Use 80 ms frames.</dd>
+  * <dt>OPUS_FRAMESIZE_100_MS</dt><dd>Use 100 ms frames.</dd>
+  * <dt>OPUS_FRAMESIZE_120_MS</dt><dd>Use 120 ms frames.</dd>
+  * </dl>
+  * @hideinitializer */
+#define OPUS_SET_EXPERT_FRAME_DURATION(x) OPUS_SET_EXPERT_FRAME_DURATION_REQUEST, opus_check_int(x)
+/** Gets the encoder's configured use of variable duration frames.
+  * @see OPUS_SET_EXPERT_FRAME_DURATION
+  * @param[out] x <tt>opus_int32 *</tt>: Returns one of the following values:
+  * <dl>
+  * <dt>OPUS_FRAMESIZE_ARG</dt><dd>Select frame size from the argument (default).</dd>
+  * <dt>OPUS_FRAMESIZE_2_5_MS</dt><dd>Use 2.5 ms frames.</dd>
+  * <dt>OPUS_FRAMESIZE_5_MS</dt><dd>Use 5 ms frames.</dd>
+  * <dt>OPUS_FRAMESIZE_10_MS</dt><dd>Use 10 ms frames.</dd>
+  * <dt>OPUS_FRAMESIZE_20_MS</dt><dd>Use 20 ms frames.</dd>
+  * <dt>OPUS_FRAMESIZE_40_MS</dt><dd>Use 40 ms frames.</dd>
+  * <dt>OPUS_FRAMESIZE_60_MS</dt><dd>Use 60 ms frames.</dd>
+  * <dt>OPUS_FRAMESIZE_80_MS</dt><dd>Use 80 ms frames.</dd>
+  * <dt>OPUS_FRAMESIZE_100_MS</dt><dd>Use 100 ms frames.</dd>
+  * <dt>OPUS_FRAMESIZE_120_MS</dt><dd>Use 120 ms frames.</dd>
+  * </dl>
+  * @hideinitializer */
+#define OPUS_GET_EXPERT_FRAME_DURATION(x) OPUS_GET_EXPERT_FRAME_DURATION_REQUEST, opus_check_int_ptr(x)
+
+/** If set to 1, disables almost all use of prediction, making frames almost
+  * completely independent. This reduces quality.
+  * @see OPUS_GET_PREDICTION_DISABLED
+  * @param[in] x <tt>opus_int32</tt>: Allowed values:
+  * <dl>
+  * <dt>0</dt><dd>Enable prediction (default).</dd>
+  * <dt>1</dt><dd>Disable prediction.</dd>
+  * </dl>
+  * @hideinitializer */
+#define OPUS_SET_PREDICTION_DISABLED(x) OPUS_SET_PREDICTION_DISABLED_REQUEST, opus_check_int(x)
+/** Gets the encoder's configured prediction status.
+  * @see OPUS_SET_PREDICTION_DISABLED
+  * @param[out] x <tt>opus_int32 *</tt>: Returns one of the following values:
+  * <dl>
+  * <dt>0</dt><dd>Prediction enabled (default).</dd>
+  * <dt>1</dt><dd>Prediction disabled.</dd>
+  * </dl>
+  * @hideinitializer */
+#define OPUS_GET_PREDICTION_DISABLED(x) OPUS_GET_PREDICTION_DISABLED_REQUEST, opus_check_int_ptr(x)
+
+/** If non-zero, enables Deep Redundancy (DRED) and use the specified maximum number of 10-ms redundant frames
+  * @hideinitializer */
+#define OPUS_SET_DRED_DURATION(x) OPUS_SET_DRED_DURATION_REQUEST, opus_check_int(x)
+/** Gets the encoder's configured Deep Redundancy (DRED) maximum number of frames.
+  * @hideinitializer */
+#define OPUS_GET_DRED_DURATION(x) OPUS_GET_DRED_DURATION_REQUEST, opus_check_int_ptr(x)
+
+/** Provide external DNN weights from binary object (only when explicitly built without the weights)
+  * @hideinitializer */
+#define OPUS_SET_DNN_BLOB(data, len) OPUS_SET_DNN_BLOB_REQUEST, opus_check_void_ptr(data), opus_check_int(len)
+
+/** If set to 1, enables quality extension (QEXT), otherwise disables it (default). Warning: This will *hurt* audio quality unless operating at
+    a very high bitrate.
+  * @hideinitializer */
+#define OPUS_SET_QEXT(x) OPUS_SET_QEXT_REQUEST, opus_check_int(x)
+/** Gets the encoder's configured quality extension (QEXT).
+  * @hideinitializer */
+#define OPUS_GET_QEXT(x) OPUS_GET_QEXT_REQUEST, opus_check_int_ptr(x)
+
+/**@}*/
+
+/** @defgroup opus_genericctls Generic CTLs
+  *
+  * These macros are used with the \c opus_decoder_ctl and
+  * \c opus_encoder_ctl calls to generate a particular
+  * request.
+  *
+  * When called on an \c OpusDecoder they apply to that
+  * particular decoder instance. When called on an
+  * \c OpusEncoder they apply to the corresponding setting
+  * on that encoder instance, if present.
+  *
+  * Some usage examples:
+  *
+  * @code
+  * int ret;
+  * opus_int32 pitch;
+  * ret = opus_decoder_ctl(dec_ctx, OPUS_GET_PITCH(&pitch));
+  * if (ret == OPUS_OK) return ret;
+  *
+  * opus_encoder_ctl(enc_ctx, OPUS_RESET_STATE);
+  * opus_decoder_ctl(dec_ctx, OPUS_RESET_STATE);
+  *
+  * opus_int32 enc_bw, dec_bw;
+  * opus_encoder_ctl(enc_ctx, OPUS_GET_BANDWIDTH(&enc_bw));
+  * opus_decoder_ctl(dec_ctx, OPUS_GET_BANDWIDTH(&dec_bw));
+  * if (enc_bw != dec_bw) {
+  *   printf("packet bandwidth mismatch!\n");
+  * }
+  * @endcode
+  *
+  * @see opus_encoder, opus_decoder_ctl, opus_encoder_ctl, opus_decoderctls, opus_encoderctls
+  * @{
+  */
+
+/** Resets the codec state to be equivalent to a freshly initialized state.
+  * This should be called when switching streams in order to prevent
+  * the back to back decoding from giving different results from
+  * one at a time decoding.
+  * @hideinitializer */
+#define OPUS_RESET_STATE 4028
+
+/** Gets the final state of the codec's entropy coder.
+  * This is used for testing purposes,
+  * The encoder and decoder state should be identical after coding a payload
+  * (assuming no data corruption or software bugs)
+  *
+  * @param[out] x <tt>opus_uint32 *</tt>: Entropy coder state
+  *
+  * @hideinitializer */
+#define OPUS_GET_FINAL_RANGE(x) OPUS_GET_FINAL_RANGE_REQUEST, opus_check_uint_ptr(x)
+
+/** Gets the encoder's configured bandpass or the decoder's last bandpass.
+  * @see OPUS_SET_BANDWIDTH
+  * @param[out] x <tt>opus_int32 *</tt>: Returns one of the following values:
+  * <dl>
+  * <dt>#OPUS_AUTO</dt>                    <dd>(default)</dd>
+  * <dt>#OPUS_BANDWIDTH_NARROWBAND</dt>    <dd>4 kHz passband</dd>
+  * <dt>#OPUS_BANDWIDTH_MEDIUMBAND</dt>    <dd>6 kHz passband</dd>
+  * <dt>#OPUS_BANDWIDTH_WIDEBAND</dt>      <dd>8 kHz passband</dd>
+  * <dt>#OPUS_BANDWIDTH_SUPERWIDEBAND</dt><dd>12 kHz passband</dd>
+  * <dt>#OPUS_BANDWIDTH_FULLBAND</dt>     <dd>20 kHz passband</dd>
+  * </dl>
+  * @hideinitializer */
+#define OPUS_GET_BANDWIDTH(x) OPUS_GET_BANDWIDTH_REQUEST, opus_check_int_ptr(x)
+
+/** Gets the sampling rate the encoder or decoder was initialized with.
+  * This simply returns the <code>Fs</code> value passed to opus_encoder_init()
+  * or opus_decoder_init().
+  * @param[out] x <tt>opus_int32 *</tt>: Sampling rate of encoder or decoder.
+  * @hideinitializer
+  */
+#define OPUS_GET_SAMPLE_RATE(x) OPUS_GET_SAMPLE_RATE_REQUEST, opus_check_int_ptr(x)
+
+/** If set to 1, disables the use of phase inversion for intensity stereo,
+  * improving the quality of mono downmixes, but slightly reducing normal
+  * stereo quality. Disabling phase inversion in the decoder does not comply
+  * with RFC 6716, although it does not cause any interoperability issue and
+  * is expected to become part of the Opus standard once RFC 6716 is updated
+  * by draft-ietf-codec-opus-update.
+  * @see OPUS_GET_PHASE_INVERSION_DISABLED
+  * @param[in] x <tt>opus_int32</tt>: Allowed values:
+  * <dl>
+  * <dt>0</dt><dd>Enable phase inversion (default).</dd>
+  * <dt>1</dt><dd>Disable phase inversion.</dd>
+  * </dl>
+  * @hideinitializer */
+#define OPUS_SET_PHASE_INVERSION_DISABLED(x) OPUS_SET_PHASE_INVERSION_DISABLED_REQUEST, opus_check_int(x)
+/** Gets the encoder's configured phase inversion status.
+  * @see OPUS_SET_PHASE_INVERSION_DISABLED
+  * @param[out] x <tt>opus_int32 *</tt>: Returns one of the following values:
+  * <dl>
+  * <dt>0</dt><dd>Stereo phase inversion enabled (default).</dd>
+  * <dt>1</dt><dd>Stereo phase inversion disabled.</dd>
+  * </dl>
+  * @hideinitializer */
+#define OPUS_GET_PHASE_INVERSION_DISABLED(x) OPUS_GET_PHASE_INVERSION_DISABLED_REQUEST, opus_check_int_ptr(x)
+/** Gets the DTX state of the encoder.
+  * Returns whether the last encoded frame was either a comfort noise update
+  * during DTX or not encoded because of DTX.
+  * @param[out] x <tt>opus_int32 *</tt>: Returns one of the following values:
+  * <dl>
+  * <dt>0</dt><dd>The encoder is not in DTX.</dd>
+  * <dt>1</dt><dd>The encoder is in DTX.</dd>
+  * </dl>
+  * @hideinitializer */
+#define OPUS_GET_IN_DTX(x) OPUS_GET_IN_DTX_REQUEST, opus_check_int_ptr(x)
+
+/**@}*/
+
+/** @defgroup opus_decoderctls Decoder related CTLs
+  * @see opus_genericctls, opus_encoderctls, opus_decoder
+  * @{
+  */
+
+/** Configures decoder gain adjustment.
+  * Scales the decoded output by a factor specified in Q8 dB units.
+  * This has a maximum range of -32768 to 32767 inclusive, and returns
+  * OPUS_BAD_ARG otherwise. The default is zero indicating no adjustment.
+  * This setting survives decoder reset.
+  *
+  * gain = pow(10, x/(20.0*256))
+  *
+  * @param[in] x <tt>opus_int32</tt>:   Amount to scale PCM signal by in Q8 dB units.
+  * @hideinitializer */
+#define OPUS_SET_GAIN(x) OPUS_SET_GAIN_REQUEST, opus_check_int(x)
+/** Gets the decoder's configured gain adjustment. @see OPUS_SET_GAIN
+  *
+  * @param[out] x <tt>opus_int32 *</tt>: Amount to scale PCM signal by in Q8 dB units.
+  * @hideinitializer */
+#define OPUS_GET_GAIN(x) OPUS_GET_GAIN_REQUEST, opus_check_int_ptr(x)
+
+/** Gets the duration (in samples) of the last packet successfully decoded or concealed.
+  * @param[out] x <tt>opus_int32 *</tt>: Number of samples (at current sampling rate).
+  * @hideinitializer */
+#define OPUS_GET_LAST_PACKET_DURATION(x) OPUS_GET_LAST_PACKET_DURATION_REQUEST, opus_check_int_ptr(x)
+
+/** Gets the pitch of the last decoded frame, if available.
+  * This can be used for any post-processing algorithm requiring the use of pitch,
+  * e.g. time stretching/shortening. If the last frame was not voiced, or if the
+  * pitch was not coded in the frame, then zero is returned.
+  *
+  * This CTL is only implemented for decoder instances.
+  *
+  * @param[out] x <tt>opus_int32 *</tt>: pitch period at 48 kHz (or 0 if not available)
+  *
+  * @hideinitializer */
+#define OPUS_GET_PITCH(x) OPUS_GET_PITCH_REQUEST, opus_check_int_ptr(x)
+
+/** Enables blind bandwidth extension for wideband signals if decoding sampling rate is 48 kHz.
+  * @param[in] x <tt>opus_int32 </tt>: 1 enables bandwidth extension, 0 disables it.
+  * The default is 0.
+  *
+  * @hideinitializer */
+ #define OPUS_SET_OSCE_BWE(x) OPUS_SET_OSCE_BWE_REQUEST, opus_check_int(x)
+/** Gets blind bandwidth extension flag for wideband signals if decoding sampling rate is 48 kHz.
+  * @param[out] x <tt>opus_int32 *</tt>: 1 if bwe enabled, 0 if disabled.
+  *
+  * @hideinitializer */
+ #define OPUS_GET_OSCE_BWE(x) OPUS_GET_OSCE_BWE_REQUEST, opus_check_int_ptr(x)
+
+/** If set to 1, the decoder will ignore all extensions found in the padding area
+  * (does not affect DRED, which is decoded separately).
+  * @hideinitializer */
+#define OPUS_SET_IGNORE_EXTENSIONS(x) OPUS_SET_IGNORE_EXTENSIONS_REQUEST, opus_check_int(x)
+/** Gets whether the decoder is ignoring extensions.
+  * @hideinitializer */
+#define OPUS_GET_IGNORE_EXTENSIONS(x) OPUS_GET_IGNORE_EXTENSIONS_REQUEST, opus_check_int_ptr(x)
+
+/**@}*/
+
+/** @defgroup opus_libinfo Opus library information functions
+  * @{
+  */
+
+/** Converts an opus error code into a human readable string.
+  *
+  * @param[in] error <tt>int</tt>: Error number
+  * @returns Error string
+  */
+OPUS_EXPORT const char *opus_strerror(int error);
+
+/** Gets the libopus version string.
+  *
+  * Applications may look for the substring "-fixed" in the version string to
+  * determine whether they have a fixed-point or floating-point build at
+  * runtime.
+  *
+  * @returns Version string
+  */
+OPUS_EXPORT const char *opus_get_version_string(void);
+/**@}*/
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* OPUS_DEFINES_H */
--- a/lib/libopus/include/opus_multistream.h
+++ b/lib/libopus/include/opus_multistream.h
@ -0,0 +1,736 @@
+/* Copyright (c) 2011 Xiph.Org Foundation
+   Written by Jean-Marc Valin */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+/**
+ * @file opus_multistream.h
+ * @brief Opus reference implementation multistream API
+ */
+
+#ifndef OPUS_MULTISTREAM_H
+#define OPUS_MULTISTREAM_H
+
+#include "opus.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/** @cond OPUS_INTERNAL_DOC */
+
+/** Macros to trigger compilation errors when the wrong types are provided to a
+  * CTL. */
+/**@{*/
+#define opus_check_encstate_ptr(ptr) ((ptr) + ((ptr) - (OpusEncoder**)(ptr)))
+#define opus_check_decstate_ptr(ptr) ((ptr) + ((ptr) - (OpusDecoder**)(ptr)))
+/**@}*/
+
+/** These are the actual encoder and decoder CTL ID numbers.
+  * They should not be used directly by applications.
+  * In general, SETs should be even and GETs should be odd.*/
+/**@{*/
+#define OPUS_MULTISTREAM_GET_ENCODER_STATE_REQUEST 5120
+#define OPUS_MULTISTREAM_GET_DECODER_STATE_REQUEST 5122
+/**@}*/
+
+/** @endcond */
+
+/** @defgroup opus_multistream_ctls Multistream specific encoder and decoder CTLs
+  *
+  * These are convenience macros that are specific to the
+  * opus_multistream_encoder_ctl() and opus_multistream_decoder_ctl()
+  * interface.
+  * The CTLs from @ref opus_genericctls, @ref opus_encoderctls, and
+  * @ref opus_decoderctls may be applied to a multistream encoder or decoder as
+  * well.
+  * In addition, you may retrieve the encoder or decoder state for an specific
+  * stream via #OPUS_MULTISTREAM_GET_ENCODER_STATE or
+  * #OPUS_MULTISTREAM_GET_DECODER_STATE and apply CTLs to it individually.
+  */
+/**@{*/
+
+/** Gets the encoder state for an individual stream of a multistream encoder.
+  * @param[in] x <tt>opus_int32</tt>: The index of the stream whose encoder you
+  *                                   wish to retrieve.
+  *                                   This must be non-negative and less than
+  *                                   the <code>streams</code> parameter used
+  *                                   to initialize the encoder.
+  * @param[out] y <tt>OpusEncoder**</tt>: Returns a pointer to the given
+  *                                       encoder state.
+  * @retval OPUS_BAD_ARG The index of the requested stream was out of range.
+  * @hideinitializer
+  */
+#define OPUS_MULTISTREAM_GET_ENCODER_STATE(x,y) OPUS_MULTISTREAM_GET_ENCODER_STATE_REQUEST, opus_check_int(x), opus_check_encstate_ptr(y)
+
+/** Gets the decoder state for an individual stream of a multistream decoder.
+  * @param[in] x <tt>opus_int32</tt>: The index of the stream whose decoder you
+  *                                   wish to retrieve.
+  *                                   This must be non-negative and less than
+  *                                   the <code>streams</code> parameter used
+  *                                   to initialize the decoder.
+  * @param[out] y <tt>OpusDecoder**</tt>: Returns a pointer to the given
+  *                                       decoder state.
+  * @retval OPUS_BAD_ARG The index of the requested stream was out of range.
+  * @hideinitializer
+  */
+#define OPUS_MULTISTREAM_GET_DECODER_STATE(x,y) OPUS_MULTISTREAM_GET_DECODER_STATE_REQUEST, opus_check_int(x), opus_check_decstate_ptr(y)
+
+/**@}*/
+
+/** @defgroup opus_multistream Opus Multistream API
+  * @{
+  *
+  * The multistream API allows individual Opus streams to be combined into a
+  * single packet, enabling support for up to 255 channels. Unlike an
+  * elementary Opus stream, the encoder and decoder must negotiate the channel
+  * configuration before the decoder can successfully interpret the data in the
+  * packets produced by the encoder. Some basic information, such as packet
+  * duration, can be computed without any special negotiation.
+  *
+  * The format for multistream Opus packets is defined in
+  * <a href="https://tools.ietf.org/html/rfc7845">RFC 7845</a>
+  * and is based on the self-delimited Opus framing described in Appendix B of
+  * <a href="https://tools.ietf.org/html/rfc6716">RFC 6716</a>.
+  * Normal Opus packets are just a degenerate case of multistream Opus packets,
+  * and can be encoded or decoded with the multistream API by setting
+  * <code>streams</code> to <code>1</code> when initializing the encoder or
+  * decoder.
+  *
+  * Multistream Opus streams can contain up to 255 elementary Opus streams.
+  * These may be either "uncoupled" or "coupled", indicating that the decoder
+  * is configured to decode them to either 1 or 2 channels, respectively.
+  * The streams are ordered so that all coupled streams appear at the
+  * beginning.
+  *
+  * A <code>mapping</code> table defines which decoded channel <code>i</code>
+  * should be used for each input/output (I/O) channel <code>j</code>. This table is
+  * typically provided as an unsigned char array.
+  * Let <code>i = mapping[j]</code> be the index for I/O channel <code>j</code>.
+  * If <code>i < 2*coupled_streams</code>, then I/O channel <code>j</code> is
+  * encoded as the left channel of stream <code>(i/2)</code> if <code>i</code>
+  * is even, or  as the right channel of stream <code>(i/2)</code> if
+  * <code>i</code> is odd. Otherwise, I/O channel <code>j</code> is encoded as
+  * mono in stream <code>(i - coupled_streams)</code>, unless it has the special
+  * value 255, in which case it is omitted from the encoding entirely (the
+  * decoder will reproduce it as silence). Each value <code>i</code> must either
+  * be the special value 255 or be less than <code>streams + coupled_streams</code>.
+  *
+  * The output channels specified by the encoder
+  * should use the
+  * <a href="https://www.xiph.org/vorbis/doc/Vorbis_I_spec.html#x1-810004.3.9">Vorbis
+  * channel ordering</a>. A decoder may wish to apply an additional permutation
+  * to the mapping the encoder used to achieve a different output channel
+  * order (e.g. for outputting in WAV order).
+  *
+  * Each multistream packet contains an Opus packet for each stream, and all of
+  * the Opus packets in a single multistream packet must have the same
+  * duration. Therefore the duration of a multistream packet can be extracted
+  * from the TOC sequence of the first stream, which is located at the
+  * beginning of the packet, just like an elementary Opus stream:
+  *
+  * @code
+  * int nb_samples;
+  * int nb_frames;
+  * nb_frames = opus_packet_get_nb_frames(data, len);
+  * if (nb_frames < 1)
+  *   return nb_frames;
+  * nb_samples = opus_packet_get_samples_per_frame(data, 48000) * nb_frames;
+  * @endcode
+  *
+  * The general encoding and decoding process proceeds exactly the same as in
+  * the normal @ref opus_encoder and @ref opus_decoder APIs.
+  * See their documentation for an overview of how to use the corresponding
+  * multistream functions.
+  */
+
+/** Opus multistream encoder state.
+  * This contains the complete state of a multistream Opus encoder.
+  * It is position independent and can be freely copied.
+  * @see opus_multistream_encoder_create
+  * @see opus_multistream_encoder_init
+  */
+typedef struct OpusMSEncoder OpusMSEncoder;
+
+/** Opus multistream decoder state.
+  * This contains the complete state of a multistream Opus decoder.
+  * It is position independent and can be freely copied.
+  * @see opus_multistream_decoder_create
+  * @see opus_multistream_decoder_init
+  */
+typedef struct OpusMSDecoder OpusMSDecoder;
+
+/**\name Multistream encoder functions */
+/**@{*/
+
+/** Gets the size of an OpusMSEncoder structure.
+  * @param streams <tt>int</tt>: The total number of streams to encode from the
+  *                              input.
+  *                              This must be no more than 255.
+  * @param coupled_streams <tt>int</tt>: Number of coupled (2 channel) streams
+  *                                      to encode.
+  *                                      This must be no larger than the total
+  *                                      number of streams.
+  *                                      Additionally, The total number of
+  *                                      encoded channels (<code>streams +
+  *                                      coupled_streams</code>) must be no
+  *                                      more than 255.
+  * @returns The size in bytes on success, or a negative error code
+  *          (see @ref opus_errorcodes) on error.
+  */
+OPUS_EXPORT OPUS_WARN_UNUSED_RESULT opus_int32 opus_multistream_encoder_get_size(
+      int streams,
+      int coupled_streams
+);
+
+OPUS_EXPORT OPUS_WARN_UNUSED_RESULT opus_int32 opus_multistream_surround_encoder_get_size(
+      int channels,
+      int mapping_family
+);
+
+
+/** Allocates and initializes a multistream encoder state.
+  * Call opus_multistream_encoder_destroy() to release
+  * this object when finished.
+  * @param Fs <tt>opus_int32</tt>: Sampling rate of the input signal (in Hz).
+  *                                This must be one of 8000, 12000, 16000,
+  *                                24000, or 48000.
+  * @param channels <tt>int</tt>: Number of channels in the input signal.
+  *                               This must be at most 255.
+  *                               It may be greater than the number of
+  *                               coded channels (<code>streams +
+  *                               coupled_streams</code>).
+  * @param streams <tt>int</tt>: The total number of streams to encode from the
+  *                              input.
+  *                              This must be no more than the number of channels.
+  * @param coupled_streams <tt>int</tt>: Number of coupled (2 channel) streams
+  *                                      to encode.
+  *                                      This must be no larger than the total
+  *                                      number of streams.
+  *                                      Additionally, The total number of
+  *                                      encoded channels (<code>streams +
+  *                                      coupled_streams</code>) must be no
+  *                                      more than the number of input channels.
+  * @param[in] mapping <code>const unsigned char[channels]</code>: Mapping from
+  *                    encoded channels to input channels, as described in
+  *                    @ref opus_multistream. As an extra constraint, the
+  *                    multistream encoder does not allow encoding coupled
+  *                    streams for which one channel is unused since this
+  *                    is never a good idea.
+  * @param application <tt>int</tt>: The target encoder application.
+  *                                  This must be one of the following:
+  * <dl>
+  * <dt>#OPUS_APPLICATION_VOIP</dt>
+  * <dd>Process signal for improved speech intelligibility.</dd>
+  * <dt>#OPUS_APPLICATION_AUDIO</dt>
+  * <dd>Favor faithfulness to the original input.</dd>
+  * <dt>#OPUS_APPLICATION_RESTRICTED_LOWDELAY</dt>
+  * <dd>Configure the minimum possible coding delay by disabling certain modes
+  * of operation.</dd>
+  * </dl>
+  * @param[out] error <tt>int *</tt>: Returns #OPUS_OK on success, or an error
+  *                                   code (see @ref opus_errorcodes) on
+  *                                   failure.
+  */
+OPUS_EXPORT OPUS_WARN_UNUSED_RESULT OpusMSEncoder *opus_multistream_encoder_create(
+      opus_int32 Fs,
+      int channels,
+      int streams,
+      int coupled_streams,
+      const unsigned char *mapping,
+      int application,
+      int *error
+) OPUS_ARG_NONNULL(5);
+
+OPUS_EXPORT OPUS_WARN_UNUSED_RESULT OpusMSEncoder *opus_multistream_surround_encoder_create(
+      opus_int32 Fs,
+      int channels,
+      int mapping_family,
+      int *streams,
+      int *coupled_streams,
+      unsigned char *mapping,
+      int application,
+      int *error
+) OPUS_ARG_NONNULL(4) OPUS_ARG_NONNULL(5) OPUS_ARG_NONNULL(6);
+
+/** Initialize a previously allocated multistream encoder state.
+  * The memory pointed to by \a st must be at least the size returned by
+  * opus_multistream_encoder_get_size().
+  * This is intended for applications which use their own allocator instead of
+  * malloc.
+  * To reset a previously initialized state, use the #OPUS_RESET_STATE CTL.
+  * @see opus_multistream_encoder_create
+  * @see opus_multistream_encoder_get_size
+  * @param st <tt>OpusMSEncoder*</tt>: Multistream encoder state to initialize.
+  * @param Fs <tt>opus_int32</tt>: Sampling rate of the input signal (in Hz).
+  *                                This must be one of 8000, 12000, 16000,
+  *                                24000, or 48000.
+  * @param channels <tt>int</tt>: Number of channels in the input signal.
+  *                               This must be at most 255.
+  *                               It may be greater than the number of
+  *                               coded channels (<code>streams +
+  *                               coupled_streams</code>).
+  * @param streams <tt>int</tt>: The total number of streams to encode from the
+  *                              input.
+  *                              This must be no more than the number of channels.
+  * @param coupled_streams <tt>int</tt>: Number of coupled (2 channel) streams
+  *                                      to encode.
+  *                                      This must be no larger than the total
+  *                                      number of streams.
+  *                                      Additionally, The total number of
+  *                                      encoded channels (<code>streams +
+  *                                      coupled_streams</code>) must be no
+  *                                      more than the number of input channels.
+  * @param[in] mapping <code>const unsigned char[channels]</code>: Mapping from
+  *                    encoded channels to input channels, as described in
+  *                    @ref opus_multistream. As an extra constraint, the
+  *                    multistream encoder does not allow encoding coupled
+  *                    streams for which one channel is unused since this
+  *                    is never a good idea.
+  * @param application <tt>int</tt>: The target encoder application.
+  *                                  This must be one of the following:
+  * <dl>
+  * <dt>#OPUS_APPLICATION_VOIP</dt>
+  * <dd>Process signal for improved speech intelligibility.</dd>
+  * <dt>#OPUS_APPLICATION_AUDIO</dt>
+  * <dd>Favor faithfulness to the original input.</dd>
+  * <dt>#OPUS_APPLICATION_RESTRICTED_LOWDELAY</dt>
+  * <dd>Configure the minimum possible coding delay by disabling certain modes
+  * of operation.</dd>
+  * </dl>
+  * @returns #OPUS_OK on success, or an error code (see @ref opus_errorcodes)
+  *          on failure.
+  */
+OPUS_EXPORT int opus_multistream_encoder_init(
+      OpusMSEncoder *st,
+      opus_int32 Fs,
+      int channels,
+      int streams,
+      int coupled_streams,
+      const unsigned char *mapping,
+      int application
+) OPUS_ARG_NONNULL(1) OPUS_ARG_NONNULL(6);
+
+OPUS_EXPORT int opus_multistream_surround_encoder_init(
+      OpusMSEncoder *st,
+      opus_int32 Fs,
+      int channels,
+      int mapping_family,
+      int *streams,
+      int *coupled_streams,
+      unsigned char *mapping,
+      int application
+) OPUS_ARG_NONNULL(1) OPUS_ARG_NONNULL(5) OPUS_ARG_NONNULL(6) OPUS_ARG_NONNULL(7);
+
+/** Encodes a multistream Opus frame.
+  * @param st <tt>OpusMSEncoder*</tt>: Multistream encoder state.
+  * @param[in] pcm <tt>const opus_int16*</tt>: The input signal as interleaved
+  *                                            samples.
+  *                                            This must contain
+  *                                            <code>frame_size*channels</code>
+  *                                            samples.
+  * @param frame_size <tt>int</tt>: Number of samples per channel in the input
+  *                                 signal.
+  *                                 This must be an Opus frame size for the
+  *                                 encoder's sampling rate.
+  *                                 For example, at 48 kHz the permitted values
+  *                                 are 120, 240, 480, 960, 1920, and 2880.
+  *                                 Passing in a duration of less than 10 ms
+  *                                 (480 samples at 48 kHz) will prevent the
+  *                                 encoder from using the LPC or hybrid modes.
+  * @param[out] data <tt>unsigned char*</tt>: Output payload.
+  *                                           This must contain storage for at
+  *                                           least \a max_data_bytes.
+  * @param [in] max_data_bytes <tt>opus_int32</tt>: Size of the allocated
+  *                                                 memory for the output
+  *                                                 payload. This may be
+  *                                                 used to impose an upper limit on
+  *                                                 the instant bitrate, but should
+  *                                                 not be used as the only bitrate
+  *                                                 control. Use #OPUS_SET_BITRATE to
+  *                                                 control the bitrate.
+  * @returns The length of the encoded packet (in bytes) on success or a
+  *          negative error code (see @ref opus_errorcodes) on failure.
+  */
+OPUS_EXPORT OPUS_WARN_UNUSED_RESULT int opus_multistream_encode(
+    OpusMSEncoder *st,
+    const opus_int16 *pcm,
+    int frame_size,
+    unsigned char *data,
+    opus_int32 max_data_bytes
+) OPUS_ARG_NONNULL(1) OPUS_ARG_NONNULL(2) OPUS_ARG_NONNULL(4);
+
+/** Encodes a multistream Opus frame.
+  * @param st <tt>OpusMSEncoder*</tt>: Multistream encoder state.
+  * @param[in] pcm <tt>const opus_int32*</tt>: The input signal as interleaved
+  *                                            samples representing (or slightly exceeding) 24-bit values.
+  *                                            This must contain
+  *                                            <code>frame_size*channels</code>
+  *                                            samples.
+  * @param frame_size <tt>int</tt>: Number of samples per channel in the input
+  *                                 signal.
+  *                                 This must be an Opus frame size for the
+  *                                 encoder's sampling rate.
+  *                                 For example, at 48 kHz the permitted values
+  *                                 are 120, 240, 480, 960, 1920, and 2880.
+  *                                 Passing in a duration of less than 10 ms
+  *                                 (480 samples at 48 kHz) will prevent the
+  *                                 encoder from using the LPC or hybrid modes.
+  * @param[out] data <tt>unsigned char*</tt>: Output payload.
+  *                                           This must contain storage for at
+  *                                           least \a max_data_bytes.
+  * @param [in] max_data_bytes <tt>opus_int32</tt>: Size of the allocated
+  *                                                 memory for the output
+  *                                                 payload. This may be
+  *                                                 used to impose an upper limit on
+  *                                                 the instant bitrate, but should
+  *                                                 not be used as the only bitrate
+  *                                                 control. Use #OPUS_SET_BITRATE to
+  *                                                 control the bitrate.
+  * @returns The length of the encoded packet (in bytes) on success or a
+  *          negative error code (see @ref opus_errorcodes) on failure.
+  */
+OPUS_EXPORT OPUS_WARN_UNUSED_RESULT int opus_multistream_encode24(
+    OpusMSEncoder *st,
+    const opus_int32 *pcm,
+    int frame_size,
+    unsigned char *data,
+    opus_int32 max_data_bytes
+) OPUS_ARG_NONNULL(1) OPUS_ARG_NONNULL(2) OPUS_ARG_NONNULL(4);
+
+/** Encodes a multistream Opus frame from floating point input.
+  * @param st <tt>OpusMSEncoder*</tt>: Multistream encoder state.
+  * @param[in] pcm <tt>const float*</tt>: The input signal as interleaved
+  *                                       samples with a normal range of
+  *                                       +/-1.0.
+  *                                       Samples with a range beyond +/-1.0
+  *                                       are supported but will be clipped by
+  *                                       decoders using the integer API and
+  *                                       should only be used if it is known
+  *                                       that the far end supports extended
+  *                                       dynamic range.
+  *                                       This must contain
+  *                                       <code>frame_size*channels</code>
+  *                                       samples.
+  * @param frame_size <tt>int</tt>: Number of samples per channel in the input
+  *                                 signal.
+  *                                 This must be an Opus frame size for the
+  *                                 encoder's sampling rate.
+  *                                 For example, at 48 kHz the permitted values
+  *                                 are 120, 240, 480, 960, 1920, and 2880.
+  *                                 Passing in a duration of less than 10 ms
+  *                                 (480 samples at 48 kHz) will prevent the
+  *                                 encoder from using the LPC or hybrid modes.
+  * @param[out] data <tt>unsigned char*</tt>: Output payload.
+  *                                           This must contain storage for at
+  *                                           least \a max_data_bytes.
+  * @param [in] max_data_bytes <tt>opus_int32</tt>: Size of the allocated
+  *                                                 memory for the output
+  *                                                 payload. This may be
+  *                                                 used to impose an upper limit on
+  *                                                 the instant bitrate, but should
+  *                                                 not be used as the only bitrate
+  *                                                 control. Use #OPUS_SET_BITRATE to
+  *                                                 control the bitrate.
+  * @returns The length of the encoded packet (in bytes) on success or a
+  *          negative error code (see @ref opus_errorcodes) on failure.
+  */
+OPUS_EXPORT OPUS_WARN_UNUSED_RESULT int opus_multistream_encode_float(
+      OpusMSEncoder *st,
+      const float *pcm,
+      int frame_size,
+      unsigned char *data,
+      opus_int32 max_data_bytes
+) OPUS_ARG_NONNULL(1) OPUS_ARG_NONNULL(2) OPUS_ARG_NONNULL(4);
+
+/** Frees an <code>OpusMSEncoder</code> allocated by
+  * opus_multistream_encoder_create().
+  * @param st <tt>OpusMSEncoder*</tt>: Multistream encoder state to be freed.
+  */
+OPUS_EXPORT void opus_multistream_encoder_destroy(OpusMSEncoder *st);
+
+/** Perform a CTL function on a multistream Opus encoder.
+  *
+  * Generally the request and subsequent arguments are generated by a
+  * convenience macro.
+  * @param st <tt>OpusMSEncoder*</tt>: Multistream encoder state.
+  * @param request This and all remaining parameters should be replaced by one
+  *                of the convenience macros in @ref opus_genericctls,
+  *                @ref opus_encoderctls, or @ref opus_multistream_ctls.
+  * @see opus_genericctls
+  * @see opus_encoderctls
+  * @see opus_multistream_ctls
+  */
+OPUS_EXPORT int opus_multistream_encoder_ctl(OpusMSEncoder *st, int request, ...) OPUS_ARG_NONNULL(1);
+
+/**@}*/
+
+/**\name Multistream decoder functions */
+/**@{*/
+
+/** Gets the size of an <code>OpusMSDecoder</code> structure.
+  * @param streams <tt>int</tt>: The total number of streams coded in the
+  *                              input.
+  *                              This must be no more than 255.
+  * @param coupled_streams <tt>int</tt>: Number streams to decode as coupled
+  *                                      (2 channel) streams.
+  *                                      This must be no larger than the total
+  *                                      number of streams.
+  *                                      Additionally, The total number of
+  *                                      coded channels (<code>streams +
+  *                                      coupled_streams</code>) must be no
+  *                                      more than 255.
+  * @returns The size in bytes on success, or a negative error code
+  *          (see @ref opus_errorcodes) on error.
+  */
+OPUS_EXPORT OPUS_WARN_UNUSED_RESULT opus_int32 opus_multistream_decoder_get_size(
+      int streams,
+      int coupled_streams
+);
+
+/** Allocates and initializes a multistream decoder state.
+  * Call opus_multistream_decoder_destroy() to release
+  * this object when finished.
+  * @param Fs <tt>opus_int32</tt>: Sampling rate to decode at (in Hz).
+  *                                This must be one of 8000, 12000, 16000,
+  *                                24000, or 48000.
+  * @param channels <tt>int</tt>: Number of channels to output.
+  *                               This must be at most 255.
+  *                               It may be different from the number of coded
+  *                               channels (<code>streams +
+  *                               coupled_streams</code>).
+  * @param streams <tt>int</tt>: The total number of streams coded in the
+  *                              input.
+  *                              This must be no more than 255.
+  * @param coupled_streams <tt>int</tt>: Number of streams to decode as coupled
+  *                                      (2 channel) streams.
+  *                                      This must be no larger than the total
+  *                                      number of streams.
+  *                                      Additionally, The total number of
+  *                                      coded channels (<code>streams +
+  *                                      coupled_streams</code>) must be no
+  *                                      more than 255.
+  * @param[in] mapping <code>const unsigned char[channels]</code>: Mapping from
+  *                    coded channels to output channels, as described in
+  *                    @ref opus_multistream.
+  * @param[out] error <tt>int *</tt>: Returns #OPUS_OK on success, or an error
+  *                                   code (see @ref opus_errorcodes) on
+  *                                   failure.
+  */
+OPUS_EXPORT OPUS_WARN_UNUSED_RESULT OpusMSDecoder *opus_multistream_decoder_create(
+      opus_int32 Fs,
+      int channels,
+      int streams,
+      int coupled_streams,
+      const unsigned char *mapping,
+      int *error
+) OPUS_ARG_NONNULL(5);
+
+/** Initialize a previously allocated decoder state object.
+  * The memory pointed to by \a st must be at least the size returned by
+  * opus_multistream_encoder_get_size().
+  * This is intended for applications which use their own allocator instead of
+  * malloc.
+  * To reset a previously initialized state, use the #OPUS_RESET_STATE CTL.
+  * @see opus_multistream_decoder_create
+  * @see opus_multistream_deocder_get_size
+  * @param st <tt>OpusMSEncoder*</tt>: Multistream encoder state to initialize.
+  * @param Fs <tt>opus_int32</tt>: Sampling rate to decode at (in Hz).
+  *                                This must be one of 8000, 12000, 16000,
+  *                                24000, or 48000.
+  * @param channels <tt>int</tt>: Number of channels to output.
+  *                               This must be at most 255.
+  *                               It may be different from the number of coded
+  *                               channels (<code>streams +
+  *                               coupled_streams</code>).
+  * @param streams <tt>int</tt>: The total number of streams coded in the
+  *                              input.
+  *                              This must be no more than 255.
+  * @param coupled_streams <tt>int</tt>: Number of streams to decode as coupled
+  *                                      (2 channel) streams.
+  *                                      This must be no larger than the total
+  *                                      number of streams.
+  *                                      Additionally, The total number of
+  *                                      coded channels (<code>streams +
+  *                                      coupled_streams</code>) must be no
+  *                                      more than 255.
+  * @param[in] mapping <code>const unsigned char[channels]</code>: Mapping from
+  *                    coded channels to output channels, as described in
+  *                    @ref opus_multistream.
+  * @returns #OPUS_OK on success, or an error code (see @ref opus_errorcodes)
+  *          on failure.
+  */
+OPUS_EXPORT int opus_multistream_decoder_init(
+      OpusMSDecoder *st,
+      opus_int32 Fs,
+      int channels,
+      int streams,
+      int coupled_streams,
+      const unsigned char *mapping
+) OPUS_ARG_NONNULL(1) OPUS_ARG_NONNULL(6);
+
+/** Decode a multistream Opus packet.
+  * @param st <tt>OpusMSDecoder*</tt>: Multistream decoder state.
+  * @param[in] data <tt>const unsigned char*</tt>: Input payload.
+  *                                                Use a <code>NULL</code>
+  *                                                pointer to indicate packet
+  *                                                loss.
+  * @param len <tt>opus_int32</tt>: Number of bytes in payload.
+  * @param[out] pcm <tt>opus_int16*</tt>: Output signal, with interleaved
+  *                                       samples.
+  *                                       This must contain room for
+  *                                       <code>frame_size*channels</code>
+  *                                       samples.
+  * @param frame_size <tt>int</tt>: The number of samples per channel of
+  *                                 available space in \a pcm.
+  *                                 If this is less than the maximum packet duration
+  *                                 (120 ms; 5760 for 48kHz), this function will not be capable
+  *                                 of decoding some packets. In the case of PLC (data==NULL)
+  *                                 or FEC (decode_fec=1), then frame_size needs to be exactly
+  *                                 the duration of audio that is missing, otherwise the
+  *                                 decoder will not be in the optimal state to decode the
+  *                                 next incoming packet. For the PLC and FEC cases, frame_size
+  *                                 <b>must</b> be a multiple of 2.5 ms.
+  * @param decode_fec <tt>int</tt>: Flag (0 or 1) to request that any in-band
+  *                                 forward error correction data be decoded.
+  *                                 If no such data is available, the frame is
+  *                                 decoded as if it were lost.
+  * @returns Number of samples decoded on success or a negative error code
+  *          (see @ref opus_errorcodes) on failure.
+  */
+OPUS_EXPORT OPUS_WARN_UNUSED_RESULT int opus_multistream_decode(
+    OpusMSDecoder *st,
+    const unsigned char *data,
+    opus_int32 len,
+    opus_int16 *pcm,
+    int frame_size,
+    int decode_fec
+) OPUS_ARG_NONNULL(1) OPUS_ARG_NONNULL(4);
+
+/** Decode a multistream Opus packet.
+  * @param st <tt>OpusMSDecoder*</tt>: Multistream decoder state.
+  * @param[in] data <tt>const unsigned char*</tt>: Input payload.
+  *                                                Use a <code>NULL</code>
+  *                                                pointer to indicate packet
+  *                                                loss.
+  * @param len <tt>opus_int32</tt>: Number of bytes in payload.
+  * @param[out] pcm <tt>opus_int32*</tt>: Output signal, with interleaved
+  *                                       samples representing (or slightly exceeding) 24-bit values.
+  *                                       This must contain room for
+  *                                       <code>frame_size*channels</code>
+  *                                       samples.
+  * @param frame_size <tt>int</tt>: The number of samples per channel of
+  *                                 available space in \a pcm.
+  *                                 If this is less than the maximum packet duration
+  *                                 (120 ms; 5760 for 48kHz), this function will not be capable
+  *                                 of decoding some packets. In the case of PLC (data==NULL)
+  *                                 or FEC (decode_fec=1), then frame_size needs to be exactly
+  *                                 the duration of audio that is missing, otherwise the
+  *                                 decoder will not be in the optimal state to decode the
+  *                                 next incoming packet. For the PLC and FEC cases, frame_size
+  *                                 <b>must</b> be a multiple of 2.5 ms.
+  * @param decode_fec <tt>int</tt>: Flag (0 or 1) to request that any in-band
+  *                                 forward error correction data be decoded.
+  *                                 If no such data is available, the frame is
+  *                                 decoded as if it were lost.
+  * @returns Number of samples decoded on success or a negative error code
+  *          (see @ref opus_errorcodes) on failure.
+  */
+OPUS_EXPORT OPUS_WARN_UNUSED_RESULT int opus_multistream_decode24(
+    OpusMSDecoder *st,
+    const unsigned char *data,
+    opus_int32 len,
+    opus_int32 *pcm,
+    int frame_size,
+    int decode_fec
+) OPUS_ARG_NONNULL(1) OPUS_ARG_NONNULL(4);
+
+/** Decode a multistream Opus packet with floating point output.
+  * @param st <tt>OpusMSDecoder*</tt>: Multistream decoder state.
+  * @param[in] data <tt>const unsigned char*</tt>: Input payload.
+  *                                                Use a <code>NULL</code>
+  *                                                pointer to indicate packet
+  *                                                loss.
+  * @param len <tt>opus_int32</tt>: Number of bytes in payload.
+  * @param[out] pcm <tt>opus_int16*</tt>: Output signal, with interleaved
+  *                                       samples.
+  *                                       This must contain room for
+  *                                       <code>frame_size*channels</code>
+  *                                       samples.
+  * @param frame_size <tt>int</tt>: The number of samples per channel of
+  *                                 available space in \a pcm.
+  *                                 If this is less than the maximum packet duration
+  *                                 (120 ms; 5760 for 48kHz), this function will not be capable
+  *                                 of decoding some packets. In the case of PLC (data==NULL)
+  *                                 or FEC (decode_fec=1), then frame_size needs to be exactly
+  *                                 the duration of audio that is missing, otherwise the
+  *                                 decoder will not be in the optimal state to decode the
+  *                                 next incoming packet. For the PLC and FEC cases, frame_size
+  *                                 <b>must</b> be a multiple of 2.5 ms.
+  * @param decode_fec <tt>int</tt>: Flag (0 or 1) to request that any in-band
+  *                                 forward error correction data be decoded.
+  *                                 If no such data is available, the frame is
+  *                                 decoded as if it were lost.
+  * @returns Number of samples decoded on success or a negative error code
+  *          (see @ref opus_errorcodes) on failure.
+  */
+OPUS_EXPORT OPUS_WARN_UNUSED_RESULT int opus_multistream_decode_float(
+    OpusMSDecoder *st,
+    const unsigned char *data,
+    opus_int32 len,
+    float *pcm,
+    int frame_size,
+    int decode_fec
+) OPUS_ARG_NONNULL(1) OPUS_ARG_NONNULL(4);
+
+/** Perform a CTL function on a multistream Opus decoder.
+  *
+  * Generally the request and subsequent arguments are generated by a
+  * convenience macro.
+  * @param st <tt>OpusMSDecoder*</tt>: Multistream decoder state.
+  * @param request This and all remaining parameters should be replaced by one
+  *                of the convenience macros in @ref opus_genericctls,
+  *                @ref opus_decoderctls, or @ref opus_multistream_ctls.
+  * @see opus_genericctls
+  * @see opus_decoderctls
+  * @see opus_multistream_ctls
+  */
+OPUS_EXPORT int opus_multistream_decoder_ctl(OpusMSDecoder *st, int request, ...) OPUS_ARG_NONNULL(1);
+
+/** Frees an <code>OpusMSDecoder</code> allocated by
+  * opus_multistream_decoder_create().
+  * @param st <tt>OpusMSDecoder</tt>: Multistream decoder state to be freed.
+  */
+OPUS_EXPORT void opus_multistream_decoder_destroy(OpusMSDecoder *st);
+
+/**@}*/
+
+/**@}*/
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* OPUS_MULTISTREAM_H */
--- a/lib/libopus/include/opus_projection.h
+++ b/lib/libopus/include/opus_projection.h
@ -0,0 +1,643 @@
+/* Copyright (c) 2017 Google Inc.
+   Written by Andrew Allen */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+
+/**
+ * @file opus_projection.h
+ * @brief Opus projection reference API
+ */
+
+#ifndef OPUS_PROJECTION_H
+#define OPUS_PROJECTION_H
+
+#include "opus_multistream.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/** @cond OPUS_INTERNAL_DOC */
+
+/** These are the actual encoder and decoder CTL ID numbers.
+  * They should not be used directly by applications.c
+  * In general, SETs should be even and GETs should be odd.*/
+/**@{*/
+#define OPUS_PROJECTION_GET_DEMIXING_MATRIX_GAIN_REQUEST    6001
+#define OPUS_PROJECTION_GET_DEMIXING_MATRIX_SIZE_REQUEST    6003
+#define OPUS_PROJECTION_GET_DEMIXING_MATRIX_REQUEST         6005
+/**@}*/
+
+
+/** @endcond */
+
+/** @defgroup opus_projection_ctls Projection specific encoder and decoder CTLs
+  *
+  * These are convenience macros that are specific to the
+  * opus_projection_encoder_ctl() and opus_projection_decoder_ctl()
+  * interface.
+  * The CTLs from @ref opus_genericctls, @ref opus_encoderctls,
+  * @ref opus_decoderctls, and @ref opus_multistream_ctls may be applied to a
+  * projection encoder or decoder as well.
+  */
+/**@{*/
+
+/** Gets the gain (in dB. S7.8-format) of the demixing matrix from the encoder.
+  * @param[out] x <tt>opus_int32 *</tt>: Returns the gain (in dB. S7.8-format)
+  *                                      of the demixing matrix.
+  * @hideinitializer
+  */
+#define OPUS_PROJECTION_GET_DEMIXING_MATRIX_GAIN(x) OPUS_PROJECTION_GET_DEMIXING_MATRIX_GAIN_REQUEST, opus_check_int_ptr(x)
+
+
+/** Gets the size in bytes of the demixing matrix from the encoder.
+  * @param[out] x <tt>opus_int32 *</tt>: Returns the size in bytes of the
+  *                                      demixing matrix.
+  * @hideinitializer
+  */
+#define OPUS_PROJECTION_GET_DEMIXING_MATRIX_SIZE(x) OPUS_PROJECTION_GET_DEMIXING_MATRIX_SIZE_REQUEST, opus_check_int_ptr(x)
+
+
+/** Copies the demixing matrix to the supplied pointer location.
+  * @param[out] x <tt>unsigned char *</tt>: Returns the demixing matrix to the
+  *                                         supplied pointer location.
+  * @param y <tt>opus_int32</tt>: The size in bytes of the reserved memory at the
+  *                              pointer location.
+  * @hideinitializer
+  */
+#define OPUS_PROJECTION_GET_DEMIXING_MATRIX(x,y) OPUS_PROJECTION_GET_DEMIXING_MATRIX_REQUEST, x, opus_check_int(y)
+
+
+/**@}*/
+
+/** Opus projection encoder state.
+ * This contains the complete state of a projection Opus encoder.
+ * It is position independent and can be freely copied.
+ * @see opus_projection_ambisonics_encoder_create
+ */
+typedef struct OpusProjectionEncoder OpusProjectionEncoder;
+
+
+/** Opus projection decoder state.
+  * This contains the complete state of a projection Opus decoder.
+  * It is position independent and can be freely copied.
+  * @see opus_projection_decoder_create
+  * @see opus_projection_decoder_init
+  */
+typedef struct OpusProjectionDecoder OpusProjectionDecoder;
+
+
+/**\name Projection encoder functions */
+/**@{*/
+
+/** Gets the size of an OpusProjectionEncoder structure.
+  * @param channels <tt>int</tt>: The total number of input channels to encode.
+  *                               This must be no more than 255.
+  * @param mapping_family <tt>int</tt>: The mapping family to use for selecting
+  *                                     the appropriate projection.
+  * @returns The size in bytes on success, or a negative error code
+  *          (see @ref opus_errorcodes) on error.
+  */
+OPUS_EXPORT OPUS_WARN_UNUSED_RESULT opus_int32 opus_projection_ambisonics_encoder_get_size(
+    int channels,
+    int mapping_family
+);
+
+
+/** Allocates and initializes a projection encoder state.
+  * Call opus_projection_encoder_destroy() to release
+  * this object when finished.
+  * @param Fs <tt>opus_int32</tt>: Sampling rate of the input signal (in Hz).
+  *                                This must be one of 8000, 12000, 16000,
+  *                                24000, or 48000.
+  * @param channels <tt>int</tt>: Number of channels in the input signal.
+  *                               This must be at most 255.
+  *                               It may be greater than the number of
+  *                               coded channels (<code>streams +
+  *                               coupled_streams</code>).
+  * @param mapping_family <tt>int</tt>: The mapping family to use for selecting
+  *                                     the appropriate projection.
+  * @param[out] streams <tt>int *</tt>: The total number of streams that will
+  *                                     be encoded from the input.
+  * @param[out] coupled_streams <tt>int *</tt>: Number of coupled (2 channel)
+  *                                 streams that will be encoded from the input.
+  * @param application <tt>int</tt>: The target encoder application.
+  *                                  This must be one of the following:
+  * <dl>
+  * <dt>#OPUS_APPLICATION_VOIP</dt>
+  * <dd>Process signal for improved speech intelligibility.</dd>
+  * <dt>#OPUS_APPLICATION_AUDIO</dt>
+  * <dd>Favor faithfulness to the original input.</dd>
+  * <dt>#OPUS_APPLICATION_RESTRICTED_LOWDELAY</dt>
+  * <dd>Configure the minimum possible coding delay by disabling certain modes
+  * of operation.</dd>
+  * </dl>
+  * @param[out] error <tt>int *</tt>: Returns #OPUS_OK on success, or an error
+  *                                   code (see @ref opus_errorcodes) on
+  *                                   failure.
+  */
+OPUS_EXPORT OPUS_WARN_UNUSED_RESULT OpusProjectionEncoder *opus_projection_ambisonics_encoder_create(
+    opus_int32 Fs,
+    int channels,
+    int mapping_family,
+    int *streams,
+    int *coupled_streams,
+    int application,
+    int *error
+) OPUS_ARG_NONNULL(4) OPUS_ARG_NONNULL(5);
+
+
+/** Initialize a previously allocated projection encoder state.
+  * The memory pointed to by \a st must be at least the size returned by
+  * opus_projection_ambisonics_encoder_get_size().
+  * This is intended for applications which use their own allocator instead of
+  * malloc.
+  * To reset a previously initialized state, use the #OPUS_RESET_STATE CTL.
+  * @see opus_projection_ambisonics_encoder_create
+  * @see opus_projection_ambisonics_encoder_get_size
+  * @param st <tt>OpusProjectionEncoder*</tt>: Projection encoder state to initialize.
+  * @param Fs <tt>opus_int32</tt>: Sampling rate of the input signal (in Hz).
+  *                                This must be one of 8000, 12000, 16000,
+  *                                24000, or 48000.
+  * @param channels <tt>int</tt>: Number of channels in the input signal.
+  *                               This must be at most 255.
+  *                               It may be greater than the number of
+  *                               coded channels (<code>streams +
+  *                               coupled_streams</code>).
+  * @param streams <tt>int</tt>: The total number of streams to encode from the
+  *                              input.
+  *                              This must be no more than the number of channels.
+  * @param coupled_streams <tt>int</tt>: Number of coupled (2 channel) streams
+  *                                      to encode.
+  *                                      This must be no larger than the total
+  *                                      number of streams.
+  *                                      Additionally, The total number of
+  *                                      encoded channels (<code>streams +
+  *                                      coupled_streams</code>) must be no
+  *                                      more than the number of input channels.
+  * @param application <tt>int</tt>: The target encoder application.
+  *                                  This must be one of the following:
+  * <dl>
+  * <dt>#OPUS_APPLICATION_VOIP</dt>
+  * <dd>Process signal for improved speech intelligibility.</dd>
+  * <dt>#OPUS_APPLICATION_AUDIO</dt>
+  * <dd>Favor faithfulness to the original input.</dd>
+  * <dt>#OPUS_APPLICATION_RESTRICTED_LOWDELAY</dt>
+  * <dd>Configure the minimum possible coding delay by disabling certain modes
+  * of operation.</dd>
+  * </dl>
+  * @returns #OPUS_OK on success, or an error code (see @ref opus_errorcodes)
+  *          on failure.
+  */
+OPUS_EXPORT int opus_projection_ambisonics_encoder_init(
+    OpusProjectionEncoder *st,
+    opus_int32 Fs,
+    int channels,
+    int mapping_family,
+    int *streams,
+    int *coupled_streams,
+    int application
+) OPUS_ARG_NONNULL(1) OPUS_ARG_NONNULL(5) OPUS_ARG_NONNULL(6);
+
+
+/** Encodes a projection Opus frame.
+  * @param st <tt>OpusProjectionEncoder*</tt>: Projection encoder state.
+  * @param[in] pcm <tt>const opus_int16*</tt>: The input signal as interleaved
+  *                                            samples.
+  *                                            This must contain
+  *                                            <code>frame_size*channels</code>
+  *                                            samples.
+  * @param frame_size <tt>int</tt>: Number of samples per channel in the input
+  *                                 signal.
+  *                                 This must be an Opus frame size for the
+  *                                 encoder's sampling rate.
+  *                                 For example, at 48 kHz the permitted values
+  *                                 are 120, 240, 480, 960, 1920, and 2880.
+  *                                 Passing in a duration of less than 10 ms
+  *                                 (480 samples at 48 kHz) will prevent the
+  *                                 encoder from using the LPC or hybrid modes.
+  * @param[out] data <tt>unsigned char*</tt>: Output payload.
+  *                                           This must contain storage for at
+  *                                           least \a max_data_bytes.
+  * @param [in] max_data_bytes <tt>opus_int32</tt>: Size of the allocated
+  *                                                 memory for the output
+  *                                                 payload. This may be
+  *                                                 used to impose an upper limit on
+  *                                                 the instant bitrate, but should
+  *                                                 not be used as the only bitrate
+  *                                                 control. Use #OPUS_SET_BITRATE to
+  *                                                 control the bitrate.
+  * @returns The length of the encoded packet (in bytes) on success or a
+  *          negative error code (see @ref opus_errorcodes) on failure.
+  */
+OPUS_EXPORT OPUS_WARN_UNUSED_RESULT int opus_projection_encode(
+    OpusProjectionEncoder *st,
+    const opus_int16 *pcm,
+    int frame_size,
+    unsigned char *data,
+    opus_int32 max_data_bytes
+) OPUS_ARG_NONNULL(1) OPUS_ARG_NONNULL(2) OPUS_ARG_NONNULL(4);
+
+/** Encodes a projection Opus frame.
+  * @param st <tt>OpusProjectionEncoder*</tt>: Projection encoder state.
+  * @param[in] pcm <tt>const opus_int32*</tt>: The input signal as interleaved
+  *                                            samples representing (or slightly exceeding) 24-bit values.
+  *                                            This must contain
+  *                                            <code>frame_size*channels</code>
+  *                                            samples.
+  * @param frame_size <tt>int</tt>: Number of samples per channel in the input
+  *                                 signal.
+  *                                 This must be an Opus frame size for the
+  *                                 encoder's sampling rate.
+  *                                 For example, at 48 kHz the permitted values
+  *                                 are 120, 240, 480, 960, 1920, and 2880.
+  *                                 Passing in a duration of less than 10 ms
+  *                                 (480 samples at 48 kHz) will prevent the
+  *                                 encoder from using the LPC or hybrid modes.
+  * @param[out] data <tt>unsigned char*</tt>: Output payload.
+  *                                           This must contain storage for at
+  *                                           least \a max_data_bytes.
+  * @param [in] max_data_bytes <tt>opus_int32</tt>: Size of the allocated
+  *                                                 memory for the output
+  *                                                 payload. This may be
+  *                                                 used to impose an upper limit on
+  *                                                 the instant bitrate, but should
+  *                                                 not be used as the only bitrate
+  *                                                 control. Use #OPUS_SET_BITRATE to
+  *                                                 control the bitrate.
+  * @returns The length of the encoded packet (in bytes) on success or a
+  *          negative error code (see @ref opus_errorcodes) on failure.
+  */
+OPUS_EXPORT OPUS_WARN_UNUSED_RESULT int opus_projection_encode24(
+    OpusProjectionEncoder *st,
+    const opus_int32 *pcm,
+    int frame_size,
+    unsigned char *data,
+    opus_int32 max_data_bytes
+) OPUS_ARG_NONNULL(1) OPUS_ARG_NONNULL(2) OPUS_ARG_NONNULL(4);
+
+
+/** Encodes a projection Opus frame from floating point input.
+  * @param st <tt>OpusProjectionEncoder*</tt>: Projection encoder state.
+  * @param[in] pcm <tt>const float*</tt>: The input signal as interleaved
+  *                                       samples with a normal range of
+  *                                       +/-1.0.
+  *                                       Samples with a range beyond +/-1.0
+  *                                       are supported but will be clipped by
+  *                                       decoders using the integer API and
+  *                                       should only be used if it is known
+  *                                       that the far end supports extended
+  *                                       dynamic range.
+  *                                       This must contain
+  *                                       <code>frame_size*channels</code>
+  *                                       samples.
+  * @param frame_size <tt>int</tt>: Number of samples per channel in the input
+  *                                 signal.
+  *                                 This must be an Opus frame size for the
+  *                                 encoder's sampling rate.
+  *                                 For example, at 48 kHz the permitted values
+  *                                 are 120, 240, 480, 960, 1920, and 2880.
+  *                                 Passing in a duration of less than 10 ms
+  *                                 (480 samples at 48 kHz) will prevent the
+  *                                 encoder from using the LPC or hybrid modes.
+  * @param[out] data <tt>unsigned char*</tt>: Output payload.
+  *                                           This must contain storage for at
+  *                                           least \a max_data_bytes.
+  * @param [in] max_data_bytes <tt>opus_int32</tt>: Size of the allocated
+  *                                                 memory for the output
+  *                                                 payload. This may be
+  *                                                 used to impose an upper limit on
+  *                                                 the instant bitrate, but should
+  *                                                 not be used as the only bitrate
+  *                                                 control. Use #OPUS_SET_BITRATE to
+  *                                                 control the bitrate.
+  * @returns The length of the encoded packet (in bytes) on success or a
+  *          negative error code (see @ref opus_errorcodes) on failure.
+  */
+OPUS_EXPORT OPUS_WARN_UNUSED_RESULT int opus_projection_encode_float(
+    OpusProjectionEncoder *st,
+    const float *pcm,
+    int frame_size,
+    unsigned char *data,
+    opus_int32 max_data_bytes
+) OPUS_ARG_NONNULL(1) OPUS_ARG_NONNULL(2) OPUS_ARG_NONNULL(4);
+
+
+/** Frees an <code>OpusProjectionEncoder</code> allocated by
+  * opus_projection_ambisonics_encoder_create().
+  * @param st <tt>OpusProjectionEncoder*</tt>: Projection encoder state to be freed.
+  */
+OPUS_EXPORT void opus_projection_encoder_destroy(OpusProjectionEncoder *st);
+
+
+/** Perform a CTL function on a projection Opus encoder.
+  *
+  * Generally the request and subsequent arguments are generated by a
+  * convenience macro.
+  * @param st <tt>OpusProjectionEncoder*</tt>: Projection encoder state.
+  * @param request This and all remaining parameters should be replaced by one
+  *                of the convenience macros in @ref opus_genericctls,
+  *                @ref opus_encoderctls, @ref opus_multistream_ctls, or
+  *                @ref opus_projection_ctls
+  * @see opus_genericctls
+  * @see opus_encoderctls
+  * @see opus_multistream_ctls
+  * @see opus_projection_ctls
+  */
+OPUS_EXPORT int opus_projection_encoder_ctl(OpusProjectionEncoder *st, int request, ...) OPUS_ARG_NONNULL(1);
+
+
+/**@}*/
+
+/**\name Projection decoder functions */
+/**@{*/
+
+/** Gets the size of an <code>OpusProjectionDecoder</code> structure.
+  * @param channels <tt>int</tt>: The total number of output channels.
+  *                               This must be no more than 255.
+  * @param streams <tt>int</tt>: The total number of streams coded in the
+  *                              input.
+  *                              This must be no more than 255.
+  * @param coupled_streams <tt>int</tt>: Number streams to decode as coupled
+  *                                      (2 channel) streams.
+  *                                      This must be no larger than the total
+  *                                      number of streams.
+  *                                      Additionally, The total number of
+  *                                      coded channels (<code>streams +
+  *                                      coupled_streams</code>) must be no
+  *                                      more than 255.
+  * @returns The size in bytes on success, or a negative error code
+  *          (see @ref opus_errorcodes) on error.
+  */
+OPUS_EXPORT OPUS_WARN_UNUSED_RESULT opus_int32 opus_projection_decoder_get_size(
+    int channels,
+    int streams,
+    int coupled_streams
+);
+
+
+/** Allocates and initializes a projection decoder state.
+  * Call opus_projection_decoder_destroy() to release
+  * this object when finished.
+  * @param Fs <tt>opus_int32</tt>: Sampling rate to decode at (in Hz).
+  *                                This must be one of 8000, 12000, 16000,
+  *                                24000, or 48000.
+  * @param channels <tt>int</tt>: Number of channels to output.
+  *                               This must be at most 255.
+  *                               It may be different from the number of coded
+  *                               channels (<code>streams +
+  *                               coupled_streams</code>).
+  * @param streams <tt>int</tt>: The total number of streams coded in the
+  *                              input.
+  *                              This must be no more than 255.
+  * @param coupled_streams <tt>int</tt>: Number of streams to decode as coupled
+  *                                      (2 channel) streams.
+  *                                      This must be no larger than the total
+  *                                      number of streams.
+  *                                      Additionally, The total number of
+  *                                      coded channels (<code>streams +
+  *                                      coupled_streams</code>) must be no
+  *                                      more than 255.
+  * @param[in] demixing_matrix <tt>const unsigned char[demixing_matrix_size]</tt>: Demixing matrix
+  *                         that mapping from coded channels to output channels,
+  *                         as described in @ref opus_projection and
+  *                         @ref opus_projection_ctls.
+  * @param demixing_matrix_size <tt>opus_int32</tt>: The size in bytes of the
+  *                                                  demixing matrix, as
+  *                                                  described in @ref
+  *                                                  opus_projection_ctls.
+  * @param[out] error <tt>int *</tt>: Returns #OPUS_OK on success, or an error
+  *                                   code (see @ref opus_errorcodes) on
+  *                                   failure.
+  */
+OPUS_EXPORT OPUS_WARN_UNUSED_RESULT OpusProjectionDecoder *opus_projection_decoder_create(
+    opus_int32 Fs,
+    int channels,
+    int streams,
+    int coupled_streams,
+    unsigned char *demixing_matrix,
+    opus_int32 demixing_matrix_size,
+    int *error
+) OPUS_ARG_NONNULL(5);
+
+
+/** Initialize a previously allocated projection decoder state object.
+  * The memory pointed to by \a st must be at least the size returned by
+  * opus_projection_decoder_get_size().
+  * This is intended for applications which use their own allocator instead of
+  * malloc.
+  * To reset a previously initialized state, use the #OPUS_RESET_STATE CTL.
+  * @see opus_projection_decoder_create
+  * @see opus_projection_deocder_get_size
+  * @param st <tt>OpusProjectionDecoder*</tt>: Projection encoder state to initialize.
+  * @param Fs <tt>opus_int32</tt>: Sampling rate to decode at (in Hz).
+  *                                This must be one of 8000, 12000, 16000,
+  *                                24000, or 48000.
+  * @param channels <tt>int</tt>: Number of channels to output.
+  *                               This must be at most 255.
+  *                               It may be different from the number of coded
+  *                               channels (<code>streams +
+  *                               coupled_streams</code>).
+  * @param streams <tt>int</tt>: The total number of streams coded in the
+  *                              input.
+  *                              This must be no more than 255.
+  * @param coupled_streams <tt>int</tt>: Number of streams to decode as coupled
+  *                                      (2 channel) streams.
+  *                                      This must be no larger than the total
+  *                                      number of streams.
+  *                                      Additionally, The total number of
+  *                                      coded channels (<code>streams +
+  *                                      coupled_streams</code>) must be no
+  *                                      more than 255.
+  * @param[in] demixing_matrix <tt>const unsigned char[demixing_matrix_size]</tt>: Demixing matrix
+  *                         that mapping from coded channels to output channels,
+  *                         as described in @ref opus_projection and
+  *                         @ref opus_projection_ctls.
+  * @param demixing_matrix_size <tt>opus_int32</tt>: The size in bytes of the
+  *                                                  demixing matrix, as
+  *                                                  described in @ref
+  *                                                  opus_projection_ctls.
+  * @returns #OPUS_OK on success, or an error code (see @ref opus_errorcodes)
+  *          on failure.
+  */
+OPUS_EXPORT int opus_projection_decoder_init(
+    OpusProjectionDecoder *st,
+    opus_int32 Fs,
+    int channels,
+    int streams,
+    int coupled_streams,
+    unsigned char *demixing_matrix,
+    opus_int32 demixing_matrix_size
+) OPUS_ARG_NONNULL(1) OPUS_ARG_NONNULL(6);
+
+
+/** Decode a projection Opus packet.
+  * @param st <tt>OpusProjectionDecoder*</tt>: Projection decoder state.
+  * @param[in] data <tt>const unsigned char*</tt>: Input payload.
+  *                                                Use a <code>NULL</code>
+  *                                                pointer to indicate packet
+  *                                                loss.
+  * @param len <tt>opus_int32</tt>: Number of bytes in payload.
+  * @param[out] pcm <tt>opus_int16*</tt>: Output signal, with interleaved
+  *                                       samples.
+  *                                       This must contain room for
+  *                                       <code>frame_size*channels</code>
+  *                                       samples.
+  * @param frame_size <tt>int</tt>: The number of samples per channel of
+  *                                 available space in \a pcm.
+  *                                 If this is less than the maximum packet duration
+  *                                 (120 ms; 5760 for 48kHz), this function will not be capable
+  *                                 of decoding some packets. In the case of PLC (data==NULL)
+  *                                 or FEC (decode_fec=1), then frame_size needs to be exactly
+  *                                 the duration of audio that is missing, otherwise the
+  *                                 decoder will not be in the optimal state to decode the
+  *                                 next incoming packet. For the PLC and FEC cases, frame_size
+  *                                 <b>must</b> be a multiple of 2.5 ms.
+  * @param decode_fec <tt>int</tt>: Flag (0 or 1) to request that any in-band
+  *                                 forward error correction data be decoded.
+  *                                 If no such data is available, the frame is
+  *                                 decoded as if it were lost.
+  * @returns Number of samples decoded on success or a negative error code
+  *          (see @ref opus_errorcodes) on failure.
+  */
+OPUS_EXPORT OPUS_WARN_UNUSED_RESULT int opus_projection_decode(
+    OpusProjectionDecoder *st,
+    const unsigned char *data,
+    opus_int32 len,
+    opus_int16 *pcm,
+    int frame_size,
+    int decode_fec
+) OPUS_ARG_NONNULL(1) OPUS_ARG_NONNULL(4);
+
+/** Decode a projection Opus packet.
+  * @param st <tt>OpusProjectionDecoder*</tt>: Projection decoder state.
+  * @param[in] data <tt>const unsigned char*</tt>: Input payload.
+  *                                                Use a <code>NULL</code>
+  *                                                pointer to indicate packet
+  *                                                loss.
+  * @param len <tt>opus_int32</tt>: Number of bytes in payload.
+  * @param[out] pcm <tt>opus_int32*</tt>: Output signal, with interleaved
+  *                                       samples representing (or slightly exceeding) 24-bit values.
+  *                                       This must contain room for
+  *                                       <code>frame_size*channels</code>
+  *                                       samples.
+  * @param frame_size <tt>int</tt>: The number of samples per channel of
+  *                                 available space in \a pcm.
+  *                                 If this is less than the maximum packet duration
+  *                                 (120 ms; 5760 for 48kHz), this function will not be capable
+  *                                 of decoding some packets. In the case of PLC (data==NULL)
+  *                                 or FEC (decode_fec=1), then frame_size needs to be exactly
+  *                                 the duration of audio that is missing, otherwise the
+  *                                 decoder will not be in the optimal state to decode the
+  *                                 next incoming packet. For the PLC and FEC cases, frame_size
+  *                                 <b>must</b> be a multiple of 2.5 ms.
+  * @param decode_fec <tt>int</tt>: Flag (0 or 1) to request that any in-band
+  *                                 forward error correction data be decoded.
+  *                                 If no such data is available, the frame is
+  *                                 decoded as if it were lost.
+  * @returns Number of samples decoded on success or a negative error code
+  *          (see @ref opus_errorcodes) on failure.
+  */
+OPUS_EXPORT OPUS_WARN_UNUSED_RESULT int opus_projection_decode24(
+    OpusProjectionDecoder *st,
+    const unsigned char *data,
+    opus_int32 len,
+    opus_int32 *pcm,
+    int frame_size,
+    int decode_fec
+) OPUS_ARG_NONNULL(1) OPUS_ARG_NONNULL(4);
+
+/** Decode a projection Opus packet with floating point output.
+  * @param st <tt>OpusProjectionDecoder*</tt>: Projection decoder state.
+  * @param[in] data <tt>const unsigned char*</tt>: Input payload.
+  *                                                Use a <code>NULL</code>
+  *                                                pointer to indicate packet
+  *                                                loss.
+  * @param len <tt>opus_int32</tt>: Number of bytes in payload.
+  * @param[out] pcm <tt>opus_int16*</tt>: Output signal, with interleaved
+  *                                       samples.
+  *                                       This must contain room for
+  *                                       <code>frame_size*channels</code>
+  *                                       samples.
+  * @param frame_size <tt>int</tt>: The number of samples per channel of
+  *                                 available space in \a pcm.
+  *                                 If this is less than the maximum packet duration
+  *                                 (120 ms; 5760 for 48kHz), this function will not be capable
+  *                                 of decoding some packets. In the case of PLC (data==NULL)
+  *                                 or FEC (decode_fec=1), then frame_size needs to be exactly
+  *                                 the duration of audio that is missing, otherwise the
+  *                                 decoder will not be in the optimal state to decode the
+  *                                 next incoming packet. For the PLC and FEC cases, frame_size
+  *                                 <b>must</b> be a multiple of 2.5 ms.
+  * @param decode_fec <tt>int</tt>: Flag (0 or 1) to request that any in-band
+  *                                 forward error correction data be decoded.
+  *                                 If no such data is available, the frame is
+  *                                 decoded as if it were lost.
+  * @returns Number of samples decoded on success or a negative error code
+  *          (see @ref opus_errorcodes) on failure.
+  */
+OPUS_EXPORT OPUS_WARN_UNUSED_RESULT int opus_projection_decode_float(
+    OpusProjectionDecoder *st,
+    const unsigned char *data,
+    opus_int32 len,
+    float *pcm,
+    int frame_size,
+    int decode_fec
+) OPUS_ARG_NONNULL(1) OPUS_ARG_NONNULL(4);
+
+
+/** Perform a CTL function on a projection Opus decoder.
+  *
+  * Generally the request and subsequent arguments are generated by a
+  * convenience macro.
+  * @param st <tt>OpusProjectionDecoder*</tt>: Projection decoder state.
+  * @param request This and all remaining parameters should be replaced by one
+  *                of the convenience macros in @ref opus_genericctls,
+  *                @ref opus_decoderctls, @ref opus_multistream_ctls, or
+  *                @ref opus_projection_ctls.
+  * @see opus_genericctls
+  * @see opus_decoderctls
+  * @see opus_multistream_ctls
+  * @see opus_projection_ctls
+  */
+OPUS_EXPORT int opus_projection_decoder_ctl(OpusProjectionDecoder *st, int request, ...) OPUS_ARG_NONNULL(1);
+
+
+/** Frees an <code>OpusProjectionDecoder</code> allocated by
+  * opus_projection_decoder_create().
+  * @param st <tt>OpusProjectionDecoder</tt>: Projection decoder state to be freed.
+  */
+OPUS_EXPORT void opus_projection_decoder_destroy(OpusProjectionDecoder *st);
+
+
+/**@}*/
+
+/**@}*/
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* OPUS_PROJECTION_H */
--- a/lib/libopus/include/opus_types.h
+++ b/lib/libopus/include/opus_types.h
@ -0,0 +1,166 @@
+/* (C) COPYRIGHT 1994-2002 Xiph.Org Foundation */
+/* Modified by Jean-Marc Valin */
+/*
+   Redistribution and use in source and binary forms, with or without
+   modification, are permitted provided that the following conditions
+   are met:
+
+   - Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+   - Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
+   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
+   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*/
+/* opus_types.h based on ogg_types.h from libogg */
+
+/**
+   @file opus_types.h
+   @brief Opus reference implementation types
+*/
+#ifndef OPUS_TYPES_H
+#define OPUS_TYPES_H
+
+#define opus_int         int                     /* used for counters etc; at least 16 bits */
+#define opus_int64       long long
+#define opus_int8        signed char
+
+#define opus_uint        unsigned int            /* used for counters etc; at least 16 bits */
+#define opus_uint64      unsigned long long
+#define opus_uint8       unsigned char
+
+/* Use the real stdint.h if it's there (taken from Paul Hsieh's pstdint.h) */
+#if (defined(__STDC__) && __STDC__ && defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L) || (defined(__GNUC__) && (defined(_STDINT_H) || defined(_STDINT_H_)) || defined (HAVE_STDINT_H))
+#include <stdint.h>
+#  undef opus_int64
+#  undef opus_int8
+#  undef opus_uint64
+#  undef opus_uint8
+   typedef int8_t opus_int8;
+   typedef uint8_t opus_uint8;
+   typedef int16_t opus_int16;
+   typedef uint16_t opus_uint16;
+   typedef int32_t opus_int32;
+   typedef uint32_t opus_uint32;
+   typedef int64_t opus_int64;
+   typedef uint64_t opus_uint64;
+#elif defined(_WIN32)
+
+#  if defined(__CYGWIN__)
+#    include <_G_config.h>
+     typedef _G_int32_t opus_int32;
+     typedef _G_uint32_t opus_uint32;
+     typedef _G_int16 opus_int16;
+     typedef _G_uint16 opus_uint16;
+#  elif defined(__MINGW32__)
+     typedef short opus_int16;
+     typedef unsigned short opus_uint16;
+     typedef int opus_int32;
+     typedef unsigned int opus_uint32;
+#  elif defined(__MWERKS__)
+     typedef int opus_int32;
+     typedef unsigned int opus_uint32;
+     typedef short opus_int16;
+     typedef unsigned short opus_uint16;
+#  else
+     /* MSVC/Borland */
+     typedef __int32 opus_int32;
+     typedef unsigned __int32 opus_uint32;
+     typedef __int16 opus_int16;
+     typedef unsigned __int16 opus_uint16;
+#  endif
+
+#elif defined(__MACOS__)
+
+#  include <sys/types.h>
+   typedef SInt16 opus_int16;
+   typedef UInt16 opus_uint16;
+   typedef SInt32 opus_int32;
+   typedef UInt32 opus_uint32;
+
+#elif (defined(__APPLE__) && defined(__MACH__)) /* MacOS X Framework build */
+
+#  include <sys/types.h>
+   typedef int16_t opus_int16;
+   typedef u_int16_t opus_uint16;
+   typedef int32_t opus_int32;
+   typedef u_int32_t opus_uint32;
+
+#elif defined(__BEOS__)
+
+   /* Be */
+#  include <inttypes.h>
+   typedef int16 opus_int16;
+   typedef u_int16 opus_uint16;
+   typedef int32_t opus_int32;
+   typedef u_int32_t opus_uint32;
+
+#elif defined (__EMX__)
+
+   /* OS/2 GCC */
+   typedef short opus_int16;
+   typedef unsigned short opus_uint16;
+   typedef int opus_int32;
+   typedef unsigned int opus_uint32;
+
+#elif defined (DJGPP)
+
+   /* DJGPP */
+   typedef short opus_int16;
+   typedef unsigned short opus_uint16;
+   typedef int opus_int32;
+   typedef unsigned int opus_uint32;
+
+#elif defined(R5900)
+
+   /* PS2 EE */
+   typedef int opus_int32;
+   typedef unsigned opus_uint32;
+   typedef short opus_int16;
+   typedef unsigned short opus_uint16;
+
+#elif defined(__SYMBIAN32__)
+
+   /* Symbian GCC */
+   typedef signed short opus_int16;
+   typedef unsigned short opus_uint16;
+   typedef signed int opus_int32;
+   typedef unsigned int opus_uint32;
+
+#elif defined(CONFIG_TI_C54X) || defined (CONFIG_TI_C55X)
+
+   typedef short opus_int16;
+   typedef unsigned short opus_uint16;
+   typedef long opus_int32;
+   typedef unsigned long opus_uint32;
+
+#elif defined(CONFIG_TI_C6X)
+
+   typedef short opus_int16;
+   typedef unsigned short opus_uint16;
+   typedef int opus_int32;
+   typedef unsigned int opus_uint32;
+
+#else
+
+   /* Give up, take a reasonable guess */
+   typedef short opus_int16;
+   typedef unsigned short opus_uint16;
+   typedef int opus_int32;
+   typedef unsigned int opus_uint32;
+
+#endif
+
+#endif  /* OPUS_TYPES_H */
--- a/lib/libopus/opus_config.h
+++ b/lib/libopus/opus_config.h
@ -0,0 +1,8 @@
+#ifndef OPUS_CONFIG_H
+#define OPUS_CONFIG_H
+
+#define OPUS_BUILD 1
+#define VAR_ARRAYS 1
+#define USE_ALLOCA 1
+
+#endif
--- a/lib/libopus/silk/A2NLSF.c
+++ b/lib/libopus/silk/A2NLSF.c
@ -0,0 +1,267 @@
+/***********************************************************************
+Copyright (c) 2006-2011, Skype Limited. All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+- Redistributions of source code must retain the above copyright notice,
+this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+***********************************************************************/
+
+/* Conversion between prediction filter coefficients and NLSFs  */
+/* Requires the order to be an even number                      */
+/* A piecewise linear approximation maps LSF <-> cos(LSF)       */
+/* Therefore the result is not accurate NLSFs, but the two      */
+/* functions are accurate inverses of each other                */
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "SigProc_FIX.h"
+#include "tables.h"
+
+/* Number of binary divisions, when not in low complexity mode */
+#define BIN_DIV_STEPS_A2NLSF_FIX      3 /* must be no higher than 16 - log2( LSF_COS_TAB_SZ_FIX ) */
+#define MAX_ITERATIONS_A2NLSF_FIX    16
+
+/* Helper function for A2NLSF(..)                    */
+/* Transforms polynomials from cos(n*f) to cos(f)^n  */
+static OPUS_INLINE void silk_A2NLSF_trans_poly(
+    opus_int32          *p,                     /* I/O    Polynomial                                */
+    const opus_int      dd                      /* I      Polynomial order (= filter order / 2 )    */
+)
+{
+    opus_int k, n;
+
+    for( k = 2; k <= dd; k++ ) {
+        for( n = dd; n > k; n-- ) {
+            p[ n - 2 ] -= p[ n ];
+        }
+        p[ k - 2 ] -= silk_LSHIFT( p[ k ], 1 );
+    }
+}
+/* Helper function for A2NLSF(..) */
+/* Polynomial evaluation          */
+static OPUS_INLINE opus_int32 silk_A2NLSF_eval_poly( /* return the polynomial evaluation, in Q16     */
+    opus_int32          *p,                     /* I    Polynomial, Q16                         */
+    const opus_int32    x,                      /* I    Evaluation point, Q12                   */
+    const opus_int      dd                      /* I    Order                                   */
+)
+{
+    opus_int   n;
+    opus_int32 x_Q16, y32;
+
+    y32 = p[ dd ];                                  /* Q16 */
+    x_Q16 = silk_LSHIFT( x, 4 );
+
+    if ( opus_likely( 8 == dd ) )
+    {
+        y32 = silk_SMLAWW( p[ 7 ], y32, x_Q16 );
+        y32 = silk_SMLAWW( p[ 6 ], y32, x_Q16 );
+        y32 = silk_SMLAWW( p[ 5 ], y32, x_Q16 );
+        y32 = silk_SMLAWW( p[ 4 ], y32, x_Q16 );
+        y32 = silk_SMLAWW( p[ 3 ], y32, x_Q16 );
+        y32 = silk_SMLAWW( p[ 2 ], y32, x_Q16 );
+        y32 = silk_SMLAWW( p[ 1 ], y32, x_Q16 );
+        y32 = silk_SMLAWW( p[ 0 ], y32, x_Q16 );
+    }
+    else
+    {
+        for( n = dd - 1; n >= 0; n-- ) {
+            y32 = silk_SMLAWW( p[ n ], y32, x_Q16 );    /* Q16 */
+        }
+    }
+    return y32;
+}
+
+static OPUS_INLINE void silk_A2NLSF_init(
+     const opus_int32    *a_Q16,
+     opus_int32          *P,
+     opus_int32          *Q,
+     const opus_int      dd
+)
+{
+    opus_int k;
+
+    /* Convert filter coefs to even and odd polynomials */
+    P[dd] = silk_LSHIFT( 1, 16 );
+    Q[dd] = silk_LSHIFT( 1, 16 );
+    for( k = 0; k < dd; k++ ) {
+        P[ k ] = -a_Q16[ dd - k - 1 ] - a_Q16[ dd + k ];    /* Q16 */
+        Q[ k ] = -a_Q16[ dd - k - 1 ] + a_Q16[ dd + k ];    /* Q16 */
+    }
+
+    /* Divide out zeros as we have that for even filter orders, */
+    /* z =  1 is always a root in Q, and                        */
+    /* z = -1 is always a root in P                             */
+    for( k = dd; k > 0; k-- ) {
+        P[ k - 1 ] -= P[ k ];
+        Q[ k - 1 ] += Q[ k ];
+    }
+
+    /* Transform polynomials from cos(n*f) to cos(f)^n */
+    silk_A2NLSF_trans_poly( P, dd );
+    silk_A2NLSF_trans_poly( Q, dd );
+}
+
+/* Compute Normalized Line Spectral Frequencies (NLSFs) from whitening filter coefficients      */
+/* If not all roots are found, the a_Q16 coefficients are bandwidth expanded until convergence. */
+void silk_A2NLSF(
+    opus_int16                  *NLSF,              /* O    Normalized Line Spectral Frequencies in Q15 (0..2^15-1) [d] */
+    opus_int32                  *a_Q16,             /* I/O  Monic whitening filter coefficients in Q16 [d]              */
+    const opus_int              d                   /* I    Filter order (must be even)                                 */
+)
+{
+    opus_int   i, k, m, dd, root_ix, ffrac;
+    opus_int32 xlo, xhi, xmid;
+    opus_int32 ylo, yhi, ymid, thr;
+    opus_int32 nom, den;
+    opus_int32 P[ SILK_MAX_ORDER_LPC / 2 + 1 ];
+    opus_int32 Q[ SILK_MAX_ORDER_LPC / 2 + 1 ];
+    opus_int32 *PQ[ 2 ];
+    opus_int32 *p;
+
+    /* Store pointers to array */
+    PQ[ 0 ] = P;
+    PQ[ 1 ] = Q;
+
+    dd = silk_RSHIFT( d, 1 );
+
+    silk_A2NLSF_init( a_Q16, P, Q, dd );
+
+    /* Find roots, alternating between P and Q */
+    p = P;                          /* Pointer to polynomial */
+
+    xlo = silk_LSFCosTab_FIX_Q12[ 0 ]; /* Q12*/
+    ylo = silk_A2NLSF_eval_poly( p, xlo, dd );
+
+    if( ylo < 0 ) {
+        /* Set the first NLSF to zero and move on to the next */
+        NLSF[ 0 ] = 0;
+        p = Q;                      /* Pointer to polynomial */
+        ylo = silk_A2NLSF_eval_poly( p, xlo, dd );
+        root_ix = 1;                /* Index of current root */
+    } else {
+        root_ix = 0;                /* Index of current root */
+    }
+    k = 1;                          /* Loop counter */
+    i = 0;                          /* Counter for bandwidth expansions applied */
+    thr = 0;
+    while( 1 ) {
+        /* Evaluate polynomial */
+        xhi = silk_LSFCosTab_FIX_Q12[ k ]; /* Q12 */
+        yhi = silk_A2NLSF_eval_poly( p, xhi, dd );
+
+        /* Detect zero crossing */
+        if( ( ylo <= 0 && yhi >= thr ) || ( ylo >= 0 && yhi <= -thr ) ) {
+            if( yhi == 0 ) {
+                /* If the root lies exactly at the end of the current       */
+                /* interval, look for the next root in the next interval    */
+                thr = 1;
+            } else {
+                thr = 0;
+            }
+            /* Binary division */
+            ffrac = -256;
+            for( m = 0; m < BIN_DIV_STEPS_A2NLSF_FIX; m++ ) {
+                /* Evaluate polynomial */
+                xmid = silk_RSHIFT_ROUND( xlo + xhi, 1 );
+                ymid = silk_A2NLSF_eval_poly( p, xmid, dd );
+
+                /* Detect zero crossing */
+                if( ( ylo <= 0 && ymid >= 0 ) || ( ylo >= 0 && ymid <= 0 ) ) {
+                    /* Reduce frequency */
+                    xhi = xmid;
+                    yhi = ymid;
+                } else {
+                    /* Increase frequency */
+                    xlo = xmid;
+                    ylo = ymid;
+                    ffrac = silk_ADD_RSHIFT( ffrac, 128, m );
+                }
+            }
+
+            /* Interpolate */
+            if( silk_abs( ylo ) < 65536 ) {
+                /* Avoid dividing by zero */
+                den = ylo - yhi;
+                nom = silk_LSHIFT( ylo, 8 - BIN_DIV_STEPS_A2NLSF_FIX ) + silk_RSHIFT( den, 1 );
+                if( den != 0 ) {
+                    ffrac += silk_DIV32( nom, den );
+                }
+            } else {
+                /* No risk of dividing by zero because abs(ylo - yhi) >= abs(ylo) >= 65536 */
+                ffrac += silk_DIV32( ylo, silk_RSHIFT( ylo - yhi, 8 - BIN_DIV_STEPS_A2NLSF_FIX ) );
+            }
+            NLSF[ root_ix ] = (opus_int16)silk_min_32( silk_LSHIFT( (opus_int32)k, 8 ) + ffrac, silk_int16_MAX );
+
+            silk_assert( NLSF[ root_ix ] >= 0 );
+
+            root_ix++;        /* Next root */
+            if( root_ix >= d ) {
+                /* Found all roots */
+                break;
+            }
+            /* Alternate pointer to polynomial */
+            p = PQ[ root_ix & 1 ];
+
+            /* Evaluate polynomial */
+            xlo = silk_LSFCosTab_FIX_Q12[ k - 1 ]; /* Q12*/
+            ylo = silk_LSHIFT( 1 - ( root_ix & 2 ), 12 );
+        } else {
+            /* Increment loop counter */
+            k++;
+            xlo = xhi;
+            ylo = yhi;
+            thr = 0;
+
+            if( k > LSF_COS_TAB_SZ_FIX ) {
+                i++;
+                if( i > MAX_ITERATIONS_A2NLSF_FIX ) {
+                    /* Set NLSFs to white spectrum and exit */
+                    NLSF[ 0 ] = (opus_int16)silk_DIV32_16( 1 << 15, d + 1 );
+                    for( k = 1; k < d; k++ ) {
+                        NLSF[ k ] = (opus_int16)silk_ADD16( NLSF[ k-1 ], NLSF[ 0 ] );
+                    }
+                    return;
+                }
+
+                /* Error: Apply progressively more bandwidth expansion and run again */
+                silk_bwexpander_32( a_Q16, d, 65536 - silk_LSHIFT( 1, i ) );
+
+                silk_A2NLSF_init( a_Q16, P, Q, dd );
+                p = P;                            /* Pointer to polynomial */
+                xlo = silk_LSFCosTab_FIX_Q12[ 0 ]; /* Q12*/
+                ylo = silk_A2NLSF_eval_poly( p, xlo, dd );
+                if( ylo < 0 ) {
+                    /* Set the first NLSF to zero and move on to the next */
+                    NLSF[ 0 ] = 0;
+                    p = Q;                        /* Pointer to polynomial */
+                    ylo = silk_A2NLSF_eval_poly( p, xlo, dd );
+                    root_ix = 1;                  /* Index of current root */
+                } else {
+                    root_ix = 0;                  /* Index of current root */
+                }
+                k = 1;                            /* Reset loop counter */
+            }
+        }
+    }
+}
--- a/lib/libopus/silk/API.h
+++ b/lib/libopus/silk/API.h
@ -0,0 +1,150 @@
+/***********************************************************************
+Copyright (c) 2006-2011, Skype Limited. All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+- Redistributions of source code must retain the above copyright notice,
+this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+***********************************************************************/
+
+#ifndef SILK_API_H
+#define SILK_API_H
+
+#include "control.h"
+#include "typedef.h"
+#include "errors.h"
+#include "entenc.h"
+#include "entdec.h"
+
+#ifdef ENABLE_DEEP_PLC
+#include "lpcnet_private.h"
+#endif
+
+#define SILK_MAX_FRAMES_PER_PACKET  3
+
+/* Struct for TOC (Table of Contents) */
+typedef struct {
+    opus_int    VADFlag;                                /* Voice activity for packet                            */
+    opus_int    VADFlags[ SILK_MAX_FRAMES_PER_PACKET ]; /* Voice activity for each frame in packet              */
+    opus_int    inbandFECFlag;                          /* Flag indicating if packet contains in-band FEC       */
+} silk_TOC_struct;
+
+/****************************************/
+/* Encoder functions                    */
+/****************************************/
+
+/***********************************************/
+/* Get size in bytes of the Silk encoder state */
+/***********************************************/
+opus_int silk_Get_Encoder_Size(                         /* O    Returns error code                              */
+    opus_int                        *encSizeBytes,      /* O    Number of bytes in SILK encoder state           */
+    opus_int                         channels           /* I    Number of channels                              */
+);
+
+/*************************/
+/* Init or reset encoder */
+/*************************/
+opus_int silk_InitEncoder(                              /* O    Returns error code                              */
+    void                            *encState,          /* I/O  State                                           */
+    int                              channels,          /* I    Number of channels                              */
+    int                              arch,              /* I    Run-time architecture                           */
+    silk_EncControlStruct           *encStatus          /* O    Encoder Status                                  */
+);
+
+/**************************/
+/* Encode frame with Silk */
+/**************************/
+/* Note: if prefillFlag is set, the input must contain 10 ms of audio, irrespective of what                     */
+/* encControl->payloadSize_ms is set to                                                                         */
+opus_int silk_Encode(                                   /* O    Returns error code                              */
+    void                            *encState,          /* I/O  State                                           */
+    silk_EncControlStruct           *encControl,        /* I    Control status                                  */
+    const opus_res                  *samplesIn,         /* I    Speech sample input vector                      */
+    opus_int                        nSamplesIn,         /* I    Number of samples in input vector               */
+    ec_enc                          *psRangeEnc,        /* I/O  Compressor data structure                       */
+    opus_int32                      *nBytesOut,         /* I/O  Number of bytes in payload (input: Max bytes)   */
+    const opus_int                  prefillFlag,        /* I    Flag to indicate prefilling buffers no coding   */
+    int                             activity            /* I    Decision of Opus voice activity detector        */
+);
+
+/****************************************/
+/* Decoder functions                    */
+/****************************************/
+
+
+/***********************************************/
+/* Load OSCE models from external data pointer */
+/***********************************************/
+opus_int silk_LoadOSCEModels(
+    void *decState,                                     /* O    I/O State                                       */
+    const unsigned char *data,                          /* I    pointer to binary blob                          */
+    int len                                             /* I    length of binary blob data                      */
+);
+
+/***********************************************/
+/* Get size in bytes of the Silk decoder state */
+/***********************************************/
+opus_int silk_Get_Decoder_Size(                         /* O    Returns error code                              */
+    opus_int                        *decSizeBytes       /* O    Number of bytes in SILK decoder state           */
+);
+
+/*************************/
+/* Init and Reset decoder */
+/*************************/
+opus_int silk_ResetDecoder(                              /* O    Returns error code                              */
+    void                            *decState            /* I/O  State                                           */
+);
+
+opus_int silk_InitDecoder(                              /* O    Returns error code                              */
+    void                            *decState           /* I/O  State                                           */
+);
+
+/******************/
+/* Decode a frame */
+/******************/
+opus_int silk_Decode(                                   /* O    Returns error code                              */
+    void*                           decState,           /* I/O  State                                           */
+    silk_DecControlStruct*          decControl,         /* I/O  Control Structure                               */
+    opus_int                        lostFlag,           /* I    0: no loss, 1 loss, 2 decode fec                */
+    opus_int                        newPacketFlag,      /* I    Indicates first decoder call for this packet    */
+    ec_dec                          *psRangeDec,        /* I/O  Compressor data structure                       */
+    opus_res                        *samplesOut,        /* O    Decoded output speech vector                    */
+    opus_int32                      *nSamplesOut,       /* O    Number of samples decoded                       */
+#ifdef ENABLE_DEEP_PLC
+    LPCNetPLCState                  *lpcnet,
+#endif
+    int                             arch                /* I    Run-time architecture                           */
+);
+
+#if 0
+/**************************************/
+/* Get table of contents for a packet */
+/**************************************/
+opus_int silk_get_TOC(
+    const opus_uint8                *payload,           /* I    Payload data                                */
+    const opus_int                  nBytesIn,           /* I    Number of input bytes                       */
+    const opus_int                  nFramesPerPayload,  /* I    Number of SILK frames per payload           */
+    silk_TOC_struct                 *Silk_TOC           /* O    Type of content                             */
+);
+#endif
+
+
+#endif
--- a/lib/libopus/silk/CNG.c
+++ b/lib/libopus/silk/CNG.c
@ -0,0 +1,188 @@
+/***********************************************************************
+Copyright (c) 2006-2011, Skype Limited. All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+- Redistributions of source code must retain the above copyright notice,
+this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+***********************************************************************/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "main.h"
+#include "stack_alloc.h"
+
+/* Generates excitation for CNG LPC synthesis */
+static OPUS_INLINE void silk_CNG_exc(
+    opus_int32                       exc_Q14[],          /* O    CNG excitation signal Q10                   */
+    opus_int32                       exc_buf_Q14[],      /* I    Random samples buffer Q10                   */
+    opus_int                         length,             /* I    Length                                      */
+    opus_int32                       *rand_seed          /* I/O  Seed to random index generator              */
+)
+{
+    opus_int32 seed;
+    opus_int   i, idx, exc_mask;
+
+    exc_mask = CNG_BUF_MASK_MAX;
+    while( exc_mask > length ) {
+        exc_mask = silk_RSHIFT( exc_mask, 1 );
+    }
+
+    seed = *rand_seed;
+    for( i = 0; i < length; i++ ) {
+        seed = silk_RAND( seed );
+        idx = (opus_int)( silk_RSHIFT( seed, 24 ) & exc_mask );
+        silk_assert( idx >= 0 );
+        silk_assert( idx <= CNG_BUF_MASK_MAX );
+        exc_Q14[ i ] = exc_buf_Q14[ idx ];
+    }
+    *rand_seed = seed;
+}
+
+void silk_CNG_Reset(
+    silk_decoder_state          *psDec                          /* I/O  Decoder state                               */
+)
+{
+    opus_int i, NLSF_step_Q15, NLSF_acc_Q15;
+
+    NLSF_step_Q15 = silk_DIV32_16( silk_int16_MAX, psDec->LPC_order + 1 );
+    NLSF_acc_Q15 = 0;
+    for( i = 0; i < psDec->LPC_order; i++ ) {
+        NLSF_acc_Q15 += NLSF_step_Q15;
+        psDec->sCNG.CNG_smth_NLSF_Q15[ i ] = NLSF_acc_Q15;
+    }
+    psDec->sCNG.CNG_smth_Gain_Q16 = 0;
+    psDec->sCNG.rand_seed = 3176576;
+}
+
+/* Updates CNG estimate, and applies the CNG when packet was lost   */
+void silk_CNG(
+    silk_decoder_state          *psDec,                         /* I/O  Decoder state                               */
+    silk_decoder_control        *psDecCtrl,                     /* I/O  Decoder control                             */
+    opus_int16                  frame[],                        /* I/O  Signal                                      */
+    opus_int                    length                          /* I    Length of residual                          */
+)
+{
+    opus_int   i, subfr;
+    opus_int32 LPC_pred_Q10, max_Gain_Q16, gain_Q16, gain_Q10;
+    opus_int16 A_Q12[ MAX_LPC_ORDER ];
+    silk_CNG_struct *psCNG = &psDec->sCNG;
+    SAVE_STACK;
+
+    if( psDec->fs_kHz != psCNG->fs_kHz ) {
+        /* Reset state */
+        silk_CNG_Reset( psDec );
+
+        psCNG->fs_kHz = psDec->fs_kHz;
+    }
+    if( psDec->lossCnt == 0 && psDec->prevSignalType == TYPE_NO_VOICE_ACTIVITY ) {
+        /* Update CNG parameters */
+
+        /* Smoothing of LSF's  */
+        for( i = 0; i < psDec->LPC_order; i++ ) {
+            psCNG->CNG_smth_NLSF_Q15[ i ] += silk_SMULWB( (opus_int32)psDec->prevNLSF_Q15[ i ] - (opus_int32)psCNG->CNG_smth_NLSF_Q15[ i ], CNG_NLSF_SMTH_Q16 );
+        }
+        /* Find the subframe with the highest gain */
+        max_Gain_Q16 = 0;
+        subfr        = 0;
+        for( i = 0; i < psDec->nb_subfr; i++ ) {
+            if( psDecCtrl->Gains_Q16[ i ] > max_Gain_Q16 ) {
+                max_Gain_Q16 = psDecCtrl->Gains_Q16[ i ];
+                subfr        = i;
+            }
+        }
+        /* Update CNG excitation buffer with excitation from this subframe */
+        silk_memmove( &psCNG->CNG_exc_buf_Q14[ psDec->subfr_length ], psCNG->CNG_exc_buf_Q14, ( psDec->nb_subfr - 1 ) * psDec->subfr_length * sizeof( opus_int32 ) );
+        silk_memcpy(   psCNG->CNG_exc_buf_Q14, &psDec->exc_Q14[ subfr * psDec->subfr_length ], psDec->subfr_length * sizeof( opus_int32 ) );
+
+        /* Smooth gains */
+        for( i = 0; i < psDec->nb_subfr; i++ ) {
+            psCNG->CNG_smth_Gain_Q16 += silk_SMULWB( psDecCtrl->Gains_Q16[ i ] - psCNG->CNG_smth_Gain_Q16, CNG_GAIN_SMTH_Q16 );
+            /* If the smoothed gain is 3 dB greater than this subframe's gain, use this subframe's gain to adapt faster. */
+            if( silk_SMULWW( psCNG->CNG_smth_Gain_Q16, CNG_GAIN_SMTH_THRESHOLD_Q16 ) > psDecCtrl->Gains_Q16[ i ] ) {
+                psCNG->CNG_smth_Gain_Q16 = psDecCtrl->Gains_Q16[ i ];
+            }
+        }
+    }
+
+    /* Add CNG when packet is lost or during DTX */
+    if( psDec->lossCnt ) {
+        VARDECL( opus_int32, CNG_sig_Q14 );
+        ALLOC( CNG_sig_Q14, length + MAX_LPC_ORDER, opus_int32 );
+
+        /* Generate CNG excitation */
+        gain_Q16 = silk_SMULWW( psDec->sPLC.randScale_Q14, psDec->sPLC.prevGain_Q16[1] );
+        if( gain_Q16 >= (1 << 21) || psCNG->CNG_smth_Gain_Q16 > (1 << 23) ) {
+            gain_Q16 = silk_SMULTT( gain_Q16, gain_Q16 );
+            gain_Q16 = silk_SUB_LSHIFT32(silk_SMULTT( psCNG->CNG_smth_Gain_Q16, psCNG->CNG_smth_Gain_Q16 ), gain_Q16, 5 );
+            gain_Q16 = silk_LSHIFT32( silk_SQRT_APPROX( gain_Q16 ), 16 );
+        } else {
+            gain_Q16 = silk_SMULWW( gain_Q16, gain_Q16 );
+            gain_Q16 = silk_SUB_LSHIFT32(silk_SMULWW( psCNG->CNG_smth_Gain_Q16, psCNG->CNG_smth_Gain_Q16 ), gain_Q16, 5 );
+            gain_Q16 = silk_LSHIFT32( silk_SQRT_APPROX( gain_Q16 ), 8 );
+        }
+        gain_Q10 = silk_RSHIFT( gain_Q16, 6 );
+
+        silk_CNG_exc( CNG_sig_Q14 + MAX_LPC_ORDER, psCNG->CNG_exc_buf_Q14, length, &psCNG->rand_seed );
+
+        /* Convert CNG NLSF to filter representation */
+        silk_NLSF2A( A_Q12, psCNG->CNG_smth_NLSF_Q15, psDec->LPC_order, psDec->arch );
+
+        /* Generate CNG signal, by synthesis filtering */
+        silk_memcpy( CNG_sig_Q14, psCNG->CNG_synth_state, MAX_LPC_ORDER * sizeof( opus_int32 ) );
+        celt_assert( psDec->LPC_order == 10 || psDec->LPC_order == 16 );
+        for( i = 0; i < length; i++ ) {
+            /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */
+            LPC_pred_Q10 = silk_RSHIFT( psDec->LPC_order, 1 );
+            LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i -  1 ], A_Q12[ 0 ] );
+            LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i -  2 ], A_Q12[ 1 ] );
+            LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i -  3 ], A_Q12[ 2 ] );
+            LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i -  4 ], A_Q12[ 3 ] );
+            LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i -  5 ], A_Q12[ 4 ] );
+            LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i -  6 ], A_Q12[ 5 ] );
+            LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i -  7 ], A_Q12[ 6 ] );
+            LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i -  8 ], A_Q12[ 7 ] );
+            LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i -  9 ], A_Q12[ 8 ] );
+            LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i - 10 ], A_Q12[ 9 ] );
+            if( psDec->LPC_order == 16 ) {
+                LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i - 11 ], A_Q12[ 10 ] );
+                LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i - 12 ], A_Q12[ 11 ] );
+                LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i - 13 ], A_Q12[ 12 ] );
+                LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i - 14 ], A_Q12[ 13 ] );
+                LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i - 15 ], A_Q12[ 14 ] );
+                LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, CNG_sig_Q14[ MAX_LPC_ORDER + i - 16 ], A_Q12[ 15 ] );
+            }
+
+            /* Update states */
+            CNG_sig_Q14[ MAX_LPC_ORDER + i ] = silk_ADD_SAT32( CNG_sig_Q14[ MAX_LPC_ORDER + i ], silk_LSHIFT_SAT32( LPC_pred_Q10, 4 ) );
+
+            /* Scale with Gain and add to input signal */
+            frame[ i ] = (opus_int16)silk_ADD_SAT16( frame[ i ], silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( CNG_sig_Q14[ MAX_LPC_ORDER + i ], gain_Q10 ), 8 ) ) );
+
+        }
+        silk_memcpy( psCNG->CNG_synth_state, &CNG_sig_Q14[ length ], MAX_LPC_ORDER * sizeof( opus_int32 ) );
+    } else {
+        silk_memset( psCNG->CNG_synth_state, 0, psDec->LPC_order *  sizeof( opus_int32 ) );
+    }
+    RESTORE_STACK;
+}
--- a/lib/libopus/silk/HP_variable_cutoff.c
+++ b/lib/libopus/silk/HP_variable_cutoff.c
@ -0,0 +1,77 @@
+/***********************************************************************
+Copyright (c) 2006-2011, Skype Limited. All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+- Redistributions of source code must retain the above copyright notice,
+this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+***********************************************************************/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+#ifdef FIXED_POINT
+#include "main_FIX.h"
+#else
+#include "main_FLP.h"
+#endif
+#include "tuning_parameters.h"
+
+/* High-pass filter with cutoff frequency adaptation based on pitch lag statistics */
+void silk_HP_variable_cutoff(
+    silk_encoder_state_Fxx          state_Fxx[]                         /* I/O  Encoder states                              */
+)
+{
+   opus_int   quality_Q15;
+   opus_int32 pitch_freq_Hz_Q16, pitch_freq_log_Q7, delta_freq_Q7;
+   silk_encoder_state *psEncC1 = &state_Fxx[ 0 ].sCmn;
+
+   /* Adaptive cutoff frequency: estimate low end of pitch frequency range */
+   if( psEncC1->prevSignalType == TYPE_VOICED ) {
+      /* difference, in log domain */
+      pitch_freq_Hz_Q16 = silk_DIV32_16( silk_LSHIFT( silk_MUL( psEncC1->fs_kHz, 1000 ), 16 ), psEncC1->prevLag );
+      pitch_freq_log_Q7 = silk_lin2log( pitch_freq_Hz_Q16 ) - ( 16 << 7 );
+
+      /* adjustment based on quality */
+      quality_Q15 = psEncC1->input_quality_bands_Q15[ 0 ];
+      pitch_freq_log_Q7 = silk_SMLAWB( pitch_freq_log_Q7, silk_SMULWB( silk_LSHIFT( -quality_Q15, 2 ), quality_Q15 ),
+            pitch_freq_log_Q7 - ( silk_lin2log( SILK_FIX_CONST( VARIABLE_HP_MIN_CUTOFF_HZ, 16 ) ) - ( 16 << 7 ) ) );
+
+      /* delta_freq = pitch_freq_log - psEnc->variable_HP_smth1; */
+      delta_freq_Q7 = pitch_freq_log_Q7 - silk_RSHIFT( psEncC1->variable_HP_smth1_Q15, 8 );
+      if( delta_freq_Q7 < 0 ) {
+         /* less smoothing for decreasing pitch frequency, to track something close to the minimum */
+         delta_freq_Q7 = silk_MUL( delta_freq_Q7, 3 );
+      }
+
+      /* limit delta, to reduce impact of outliers in pitch estimation */
+      delta_freq_Q7 = silk_LIMIT_32( delta_freq_Q7, -SILK_FIX_CONST( VARIABLE_HP_MAX_DELTA_FREQ, 7 ), SILK_FIX_CONST( VARIABLE_HP_MAX_DELTA_FREQ, 7 ) );
+
+      /* update smoother */
+      psEncC1->variable_HP_smth1_Q15 = silk_SMLAWB( psEncC1->variable_HP_smth1_Q15,
+            silk_SMULBB( psEncC1->speech_activity_Q8, delta_freq_Q7 ), SILK_FIX_CONST( VARIABLE_HP_SMTH_COEF1, 16 ) );
+
+      /* limit frequency range */
+      psEncC1->variable_HP_smth1_Q15 = silk_LIMIT_32( psEncC1->variable_HP_smth1_Q15,
+            silk_LSHIFT( silk_lin2log( VARIABLE_HP_MIN_CUTOFF_HZ ), 8 ),
+            silk_LSHIFT( silk_lin2log( VARIABLE_HP_MAX_CUTOFF_HZ ), 8 ) );
+   }
+}
--- a/lib/libopus/silk/Inlines.h
+++ b/lib/libopus/silk/Inlines.h
@ -0,0 +1,180 @@
+/***********************************************************************
+Copyright (c) 2006-2011, Skype Limited. All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+- Redistributions of source code must retain the above copyright notice,
+this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+***********************************************************************/
+
+/*! \file silk_Inlines.h
+ *  \brief silk_Inlines.h defines OPUS_INLINE signal processing functions.
+ */
+
+#ifndef SILK_FIX_INLINES_H
+#define SILK_FIX_INLINES_H
+
+
+/* count leading zeros of opus_int64 */
+static OPUS_INLINE opus_int32 silk_CLZ64( opus_int64 in )
+{
+    opus_int32 in_upper;
+
+    in_upper = (opus_int32)silk_RSHIFT64(in, 32);
+    if (in_upper == 0) {
+        /* Search in the lower 32 bits */
+        return 32 + silk_CLZ32( (opus_int32) in );
+    } else {
+        /* Search in the upper 32 bits */
+        return silk_CLZ32( in_upper );
+    }
+}
+
+/* get number of leading zeros and fractional part (the bits right after the leading one */
+static OPUS_INLINE void silk_CLZ_FRAC(
+    opus_int32 in,            /* I  input                               */
+    opus_int32 *lz,           /* O  number of leading zeros             */
+    opus_int32 *frac_Q7       /* O  the 7 bits right after the leading one */
+)
+{
+    opus_int32 lzeros = silk_CLZ32(in);
+
+    * lz = lzeros;
+    * frac_Q7 = silk_ROR32(in, 24 - lzeros) & 0x7f;
+}
+
+/* Approximation of square root                                          */
+/* Accuracy: < +/- 10%  for output values > 15                           */
+/*           < +/- 2.5% for output values > 120                          */
+static OPUS_INLINE opus_int32 silk_SQRT_APPROX( opus_int32 x )
+{
+    opus_int32 y, lz, frac_Q7;
+
+    if( x <= 0 ) {
+        return 0;
+    }
+
+    silk_CLZ_FRAC(x, &lz, &frac_Q7);
+
+    if( lz & 1 ) {
+        y = 32768;
+    } else {
+        y = 46214;        /* 46214 = sqrt(2) * 32768 */
+    }
+
+    /* get scaling right */
+    y >>= silk_RSHIFT(lz, 1);
+
+    /* increment using fractional part of input */
+    y = silk_SMLAWB(y, y, silk_SMULBB(213, frac_Q7));
+
+    return y;
+}
+
+/* Divide two int32 values and return result as int32 in a given Q-domain */
+static OPUS_INLINE opus_int32 silk_DIV32_varQ(   /* O    returns a good approximation of "(a32 << Qres) / b32" */
+    const opus_int32     a32,               /* I    numerator (Q0)                  */
+    const opus_int32     b32,               /* I    denominator (Q0)                */
+    const opus_int       Qres               /* I    Q-domain of result (>= 0)       */
+)
+{
+    opus_int   a_headrm, b_headrm, lshift;
+    opus_int32 b32_inv, a32_nrm, b32_nrm, result;
+
+    silk_assert( b32 != 0 );
+    silk_assert( Qres >= 0 );
+
+    /* Compute number of bits head room and normalize inputs */
+    a_headrm = silk_CLZ32( silk_abs(a32) ) - 1;
+    a32_nrm = silk_LSHIFT(a32, a_headrm);                                       /* Q: a_headrm                  */
+    b_headrm = silk_CLZ32( silk_abs(b32) ) - 1;
+    b32_nrm = silk_LSHIFT(b32, b_headrm);                                       /* Q: b_headrm                  */
+
+    /* Inverse of b32, with 14 bits of precision */
+    b32_inv = silk_DIV32_16( silk_int32_MAX >> 2, silk_RSHIFT(b32_nrm, 16) );   /* Q: 29 + 16 - b_headrm        */
+
+    /* First approximation */
+    result = silk_SMULWB(a32_nrm, b32_inv);                                     /* Q: 29 + a_headrm - b_headrm  */
+
+    /* Compute residual by subtracting product of denominator and first approximation */
+    /* It's OK to overflow because the final value of a32_nrm should always be small */
+    a32_nrm = silk_SUB32_ovflw(a32_nrm, silk_LSHIFT_ovflw( silk_SMMUL(b32_nrm, result), 3 ));  /* Q: a_headrm   */
+
+    /* Refinement */
+    result = silk_SMLAWB(result, a32_nrm, b32_inv);                             /* Q: 29 + a_headrm - b_headrm  */
+
+    /* Convert to Qres domain */
+    lshift = 29 + a_headrm - b_headrm - Qres;
+    if( lshift < 0 ) {
+        return silk_LSHIFT_SAT32(result, -lshift);
+    } else {
+        if( lshift < 32){
+            return silk_RSHIFT(result, lshift);
+        } else {
+            /* Avoid undefined result */
+            return 0;
+        }
+    }
+}
+
+/* Invert int32 value and return result as int32 in a given Q-domain */
+static OPUS_INLINE opus_int32 silk_INVERSE32_varQ(   /* O    returns a good approximation of "(1 << Qres) / b32" */
+    const opus_int32     b32,                   /* I    denominator (Q0)                */
+    const opus_int       Qres                   /* I    Q-domain of result (> 0)        */
+)
+{
+    opus_int   b_headrm, lshift;
+    opus_int32 b32_inv, b32_nrm, err_Q32, result;
+
+    silk_assert( b32 != 0 );
+    silk_assert( Qres > 0 );
+
+    /* Compute number of bits head room and normalize input */
+    b_headrm = silk_CLZ32( silk_abs(b32) ) - 1;
+    b32_nrm = silk_LSHIFT(b32, b_headrm);                                       /* Q: b_headrm                */
+
+    /* Inverse of b32, with 14 bits of precision */
+    b32_inv = silk_DIV32_16( silk_int32_MAX >> 2, silk_RSHIFT(b32_nrm, 16) );   /* Q: 29 + 16 - b_headrm    */
+
+    /* First approximation */
+    result = silk_LSHIFT(b32_inv, 16);                                          /* Q: 61 - b_headrm            */
+
+    /* Compute residual by subtracting product of denominator and first approximation from one */
+    err_Q32 = silk_LSHIFT( ((opus_int32)1<<29) - silk_SMULWB(b32_nrm, b32_inv), 3 );        /* Q32                        */
+
+    /* Refinement */
+    result = silk_SMLAWW(result, err_Q32, b32_inv);                             /* Q: 61 - b_headrm            */
+
+    /* Convert to Qres domain */
+    lshift = 61 - b_headrm - Qres;
+    if( lshift <= 0 ) {
+        return silk_LSHIFT_SAT32(result, -lshift);
+    } else {
+        if( lshift < 32){
+            return silk_RSHIFT(result, lshift);
+        }else{
+            /* Avoid undefined result */
+            return 0;
+        }
+    }
+}
+
+#endif /* SILK_FIX_INLINES_H */
--- a/lib/libopus/silk/LPC_analysis_filter.c
+++ b/lib/libopus/silk/LPC_analysis_filter.c
@ -0,0 +1,111 @@
+/***********************************************************************
+Copyright (c) 2006-2011, Skype Limited. All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+- Redistributions of source code must retain the above copyright notice,
+this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+***********************************************************************/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "SigProc_FIX.h"
+#include "celt_lpc.h"
+
+/*******************************************/
+/* LPC analysis filter                     */
+/* NB! State is kept internally and the    */
+/* filter always starts with zero state    */
+/* first d output samples are set to zero  */
+/*******************************************/
+
+/* OPT: Using celt_fir() for this function should be faster, but it may cause
+   integer overflows in intermediate values (not final results), which the
+   current implementation silences by casting to unsigned. Enabling
+   this should be safe in pretty much all cases, even though it is not technically
+   C89-compliant. */
+#define USE_CELT_FIR 0
+
+void silk_LPC_analysis_filter(
+    opus_int16                  *out,               /* O    Output signal                                               */
+    const opus_int16            *in,                /* I    Input signal                                                */
+    const opus_int16            *B,                 /* I    MA prediction coefficients, Q12 [order]                     */
+    const opus_int32            len,                /* I    Signal length                                               */
+    const opus_int32            d,                  /* I    Filter order                                                */
+    int                         arch                /* I    Run-time architecture                                       */
+)
+{
+    opus_int   j;
+#if defined(FIXED_POINT) && USE_CELT_FIR
+    opus_int16 num[SILK_MAX_ORDER_LPC];
+#else
+    int ix;
+    opus_int32       out32_Q12, out32;
+    const opus_int16 *in_ptr;
+#endif
+
+    celt_assert( d >= 6 );
+    celt_assert( (d & 1) == 0 );
+    celt_assert( d <= len );
+
+#if defined(FIXED_POINT) && USE_CELT_FIR
+    celt_assert( d <= SILK_MAX_ORDER_LPC );
+    for ( j = 0; j < d; j++ ) {
+        num[ j ] = -B[ j ];
+    }
+    celt_fir( in + d, num, out + d, len - d, d, arch );
+    for ( j = 0; j < d; j++ ) {
+        out[ j ] = 0;
+    }
+#else
+    (void)arch;
+    for( ix = d; ix < len; ix++ ) {
+        in_ptr = &in[ ix - 1 ];
+
+        out32_Q12 = silk_SMULBB( in_ptr[  0 ], B[ 0 ] );
+        /* Allowing wrap around so that two wraps can cancel each other. The rare
+           cases where the result wraps around can only be triggered by invalid streams*/
+        out32_Q12 = silk_SMLABB_ovflw( out32_Q12, in_ptr[ -1 ], B[ 1 ] );
+        out32_Q12 = silk_SMLABB_ovflw( out32_Q12, in_ptr[ -2 ], B[ 2 ] );
+        out32_Q12 = silk_SMLABB_ovflw( out32_Q12, in_ptr[ -3 ], B[ 3 ] );
+        out32_Q12 = silk_SMLABB_ovflw( out32_Q12, in_ptr[ -4 ], B[ 4 ] );
+        out32_Q12 = silk_SMLABB_ovflw( out32_Q12, in_ptr[ -5 ], B[ 5 ] );
+        for( j = 6; j < d; j += 2 ) {
+            out32_Q12 = silk_SMLABB_ovflw( out32_Q12, in_ptr[ -j     ], B[ j     ] );
+            out32_Q12 = silk_SMLABB_ovflw( out32_Q12, in_ptr[ -j - 1 ], B[ j + 1 ] );
+        }
+
+        /* Subtract prediction */
+        out32_Q12 = silk_SUB32_ovflw( silk_LSHIFT( (opus_int32)in_ptr[ 1 ], 12 ), out32_Q12 );
+
+        /* Scale to Q0 */
+        out32 = silk_RSHIFT_ROUND( out32_Q12, 12 );
+
+        /* Saturate output */
+        out[ ix ] = (opus_int16)silk_SAT16( out32 );
+    }
+
+    /* Set first d output samples to zero */
+    silk_memset( out, 0, d * sizeof( opus_int16 ) );
+#endif
+}
--- a/lib/libopus/silk/LPC_fit.c
+++ b/lib/libopus/silk/LPC_fit.c
@ -0,0 +1,82 @@
+/***********************************************************************
+Copyright (c) 2013, Koen Vos. All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+- Redistributions of source code must retain the above copyright notice,
+this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+***********************************************************************/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "SigProc_FIX.h"
+
+/* Convert int32 coefficients to int16 coefs and make sure there's no wrap-around.
+   This logic is reused in _celt_lpc(). Any bug fixes should also be applied there. */
+void silk_LPC_fit(
+    opus_int16                  *a_QOUT,            /* O    Output signal                                               */
+    opus_int32                    *a_QIN,             /* I/O  Input signal                                                */
+    const opus_int              QOUT,               /* I    Input Q domain                                              */
+    const opus_int              QIN,                /* I    Input Q domain                                              */
+    const opus_int              d                   /* I    Filter order                                                */
+)
+{
+    opus_int    i, k, idx = 0;
+    opus_int32    maxabs, absval, chirp_Q16;
+
+    /* Limit the maximum absolute value of the prediction coefficients, so that they'll fit in int16 */
+    for( i = 0; i < 10; i++ ) {
+        /* Find maximum absolute value and its index */
+        maxabs = 0;
+        for( k = 0; k < d; k++ ) {
+            absval = silk_abs( a_QIN[k] );
+            if( absval > maxabs ) {
+                maxabs = absval;
+                idx    = k;
+            }
+        }
+        maxabs = silk_RSHIFT_ROUND( maxabs, QIN - QOUT );
+
+        if( maxabs > silk_int16_MAX ) {
+            /* Reduce magnitude of prediction coefficients */
+            maxabs = silk_min( maxabs, 163838 );  /* ( silk_int32_MAX >> 14 ) + silk_int16_MAX = 163838 */
+            chirp_Q16 = SILK_FIX_CONST( 0.999, 16 ) - silk_DIV32( silk_LSHIFT( maxabs - silk_int16_MAX, 14 ),
+                                        silk_RSHIFT32( silk_MUL( maxabs, idx + 1), 2 ) );
+            silk_bwexpander_32( a_QIN, d, chirp_Q16 );
+        } else {
+            break;
+        }
+    }
+
+    if( i == 10 ) {
+        /* Reached the last iteration, clip the coefficients */
+        for( k = 0; k < d; k++ ) {
+            a_QOUT[ k ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( a_QIN[ k ], QIN - QOUT ) );
+            a_QIN[ k ] = silk_LSHIFT( (opus_int32)a_QOUT[ k ], QIN - QOUT );
+        }
+    } else {
+        for( k = 0; k < d; k++ ) {
+            a_QOUT[ k ] = (opus_int16)silk_RSHIFT_ROUND( a_QIN[ k ], QIN - QOUT );
+        }
+    }
+}
--- a/lib/libopus/silk/LPC_inv_pred_gain.c
+++ b/lib/libopus/silk/LPC_inv_pred_gain.c
@ -0,0 +1,141 @@
+/***********************************************************************
+Copyright (c) 2006-2011, Skype Limited. All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+- Redistributions of source code must retain the above copyright notice,
+this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+***********************************************************************/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "SigProc_FIX.h"
+#include "define.h"
+
+#define QA                          24
+#define A_LIMIT                     SILK_FIX_CONST( 0.99975, QA )
+
+#define MUL32_FRAC_Q(a32, b32, Q)   ((opus_int32)(silk_RSHIFT_ROUND64(silk_SMULL(a32, b32), Q)))
+
+/* Compute inverse of LPC prediction gain, and                          */
+/* test if LPC coefficients are stable (all poles within unit circle)   */
+static opus_int32 LPC_inverse_pred_gain_QA_c(               /* O   Returns inverse prediction gain in energy domain, Q30    */
+    opus_int32           A_QA[ SILK_MAX_ORDER_LPC ],        /* I   Prediction coefficients                                  */
+    const opus_int       order                              /* I   Prediction order                                         */
+)
+{
+    opus_int   k, n, mult2Q;
+    opus_int32 invGain_Q30, rc_Q31, rc_mult1_Q30, rc_mult2, tmp1, tmp2;
+
+    invGain_Q30 = SILK_FIX_CONST( 1, 30 );
+    for( k = order - 1; k > 0; k-- ) {
+        /* Check for stability */
+        if( ( A_QA[ k ] > A_LIMIT ) || ( A_QA[ k ] < -A_LIMIT ) ) {
+            return 0;
+        }
+
+        /* Set RC equal to negated AR coef */
+        rc_Q31 = -silk_LSHIFT( A_QA[ k ], 31 - QA );
+
+        /* rc_mult1_Q30 range: [ 1 : 2^30 ] */
+        rc_mult1_Q30 = silk_SUB32( SILK_FIX_CONST( 1, 30 ), silk_SMMUL( rc_Q31, rc_Q31 ) );
+        silk_assert( rc_mult1_Q30 > ( 1 << 15 ) );                   /* reduce A_LIMIT if fails */
+        silk_assert( rc_mult1_Q30 <= ( 1 << 30 ) );
+
+        /* Update inverse gain */
+        /* invGain_Q30 range: [ 0 : 2^30 ] */
+        invGain_Q30 = silk_LSHIFT( silk_SMMUL( invGain_Q30, rc_mult1_Q30 ), 2 );
+        silk_assert( invGain_Q30 >= 0           );
+        silk_assert( invGain_Q30 <= ( 1 << 30 ) );
+        if( invGain_Q30 < SILK_FIX_CONST( 1.0f / MAX_PREDICTION_POWER_GAIN, 30 ) ) {
+            return 0;
+        }
+
+        /* rc_mult2 range: [ 2^30 : silk_int32_MAX ] */
+        mult2Q = 32 - silk_CLZ32( silk_abs( rc_mult1_Q30 ) );
+        rc_mult2 = silk_INVERSE32_varQ( rc_mult1_Q30, mult2Q + 30 );
+
+        /* Update AR coefficient */
+        for( n = 0; n < (k + 1) >> 1; n++ ) {
+            opus_int64 tmp64;
+            tmp1 = A_QA[ n ];
+            tmp2 = A_QA[ k - n - 1 ];
+            tmp64 = silk_RSHIFT_ROUND64( silk_SMULL( silk_SUB_SAT32(tmp1,
+                  MUL32_FRAC_Q( tmp2, rc_Q31, 31 ) ), rc_mult2 ), mult2Q);
+            if( tmp64 > silk_int32_MAX || tmp64 < silk_int32_MIN ) {
+               return 0;
+            }
+            A_QA[ n ] = ( opus_int32 )tmp64;
+            tmp64 = silk_RSHIFT_ROUND64( silk_SMULL( silk_SUB_SAT32(tmp2,
+                  MUL32_FRAC_Q( tmp1, rc_Q31, 31 ) ), rc_mult2), mult2Q);
+            if( tmp64 > silk_int32_MAX || tmp64 < silk_int32_MIN ) {
+               return 0;
+            }
+            A_QA[ k - n - 1 ] = ( opus_int32 )tmp64;
+        }
+    }
+
+    /* Check for stability */
+    if( ( A_QA[ k ] > A_LIMIT ) || ( A_QA[ k ] < -A_LIMIT ) ) {
+        return 0;
+    }
+
+    /* Set RC equal to negated AR coef */
+    rc_Q31 = -silk_LSHIFT( A_QA[ 0 ], 31 - QA );
+
+    /* Range: [ 1 : 2^30 ] */
+    rc_mult1_Q30 = silk_SUB32( SILK_FIX_CONST( 1, 30 ), silk_SMMUL( rc_Q31, rc_Q31 ) );
+
+    /* Update inverse gain */
+    /* Range: [ 0 : 2^30 ] */
+    invGain_Q30 = silk_LSHIFT( silk_SMMUL( invGain_Q30, rc_mult1_Q30 ), 2 );
+    silk_assert( invGain_Q30 >= 0           );
+    silk_assert( invGain_Q30 <= ( 1 << 30 ) );
+    if( invGain_Q30 < SILK_FIX_CONST( 1.0f / MAX_PREDICTION_POWER_GAIN, 30 ) ) {
+        return 0;
+    }
+
+    return invGain_Q30;
+}
+
+/* For input in Q12 domain */
+opus_int32 silk_LPC_inverse_pred_gain_c(            /* O   Returns inverse prediction gain in energy domain, Q30        */
+    const opus_int16            *A_Q12,             /* I   Prediction coefficients, Q12 [order]                         */
+    const opus_int              order               /* I   Prediction order                                             */
+)
+{
+    opus_int   k;
+    opus_int32 Atmp_QA[ SILK_MAX_ORDER_LPC ];
+    opus_int32 DC_resp = 0;
+
+    /* Increase Q domain of the AR coefficients */
+    for( k = 0; k < order; k++ ) {
+        DC_resp += (opus_int32)A_Q12[ k ];
+        Atmp_QA[ k ] = silk_LSHIFT32( (opus_int32)A_Q12[ k ], QA - 12 );
+    }
+    /* If the DC is unstable, we don't even need to do the full calculations */
+    if( DC_resp >= 4096 ) {
+        return 0;
+    }
+    return LPC_inverse_pred_gain_QA_c( Atmp_QA, order );
+}
--- a/lib/libopus/silk/LP_variable_cutoff.c
+++ b/lib/libopus/silk/LP_variable_cutoff.c
@ -0,0 +1,135 @@
+/***********************************************************************
+Copyright (c) 2006-2011, Skype Limited. All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+- Redistributions of source code must retain the above copyright notice,
+this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+***********************************************************************/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+/*
+    Elliptic/Cauer filters designed with 0.1 dB passband ripple,
+    80 dB minimum stopband attenuation, and
+    [0.95 : 0.15 : 0.35] normalized cut off frequencies.
+*/
+
+#include "main.h"
+
+/* Helper function, interpolates the filter taps */
+static OPUS_INLINE void silk_LP_interpolate_filter_taps(
+    opus_int32           B_Q28[ TRANSITION_NB ],
+    opus_int32           A_Q28[ TRANSITION_NA ],
+    const opus_int       ind,
+    const opus_int32     fac_Q16
+)
+{
+    opus_int nb, na;
+
+    if( ind < TRANSITION_INT_NUM - 1 ) {
+        if( fac_Q16 > 0 ) {
+            if( fac_Q16 < 32768 ) { /* fac_Q16 is in range of a 16-bit int */
+                /* Piece-wise linear interpolation of B and A */
+                for( nb = 0; nb < TRANSITION_NB; nb++ ) {
+                    B_Q28[ nb ] = silk_SMLAWB(
+                        silk_Transition_LP_B_Q28[ ind     ][ nb ],
+                        silk_Transition_LP_B_Q28[ ind + 1 ][ nb ] -
+                        silk_Transition_LP_B_Q28[ ind     ][ nb ],
+                        fac_Q16 );
+                }
+                for( na = 0; na < TRANSITION_NA; na++ ) {
+                    A_Q28[ na ] = silk_SMLAWB(
+                        silk_Transition_LP_A_Q28[ ind     ][ na ],
+                        silk_Transition_LP_A_Q28[ ind + 1 ][ na ] -
+                        silk_Transition_LP_A_Q28[ ind     ][ na ],
+                        fac_Q16 );
+                }
+            } else { /* ( fac_Q16 - ( 1 << 16 ) ) is in range of a 16-bit int */
+                silk_assert( fac_Q16 - ( 1 << 16 ) == silk_SAT16( fac_Q16 - ( 1 << 16 ) ) );
+                /* Piece-wise linear interpolation of B and A */
+                for( nb = 0; nb < TRANSITION_NB; nb++ ) {
+                    B_Q28[ nb ] = silk_SMLAWB(
+                        silk_Transition_LP_B_Q28[ ind + 1 ][ nb ],
+                        silk_Transition_LP_B_Q28[ ind + 1 ][ nb ] -
+                        silk_Transition_LP_B_Q28[ ind     ][ nb ],
+                        fac_Q16 - ( (opus_int32)1 << 16 ) );
+                }
+                for( na = 0; na < TRANSITION_NA; na++ ) {
+                    A_Q28[ na ] = silk_SMLAWB(
+                        silk_Transition_LP_A_Q28[ ind + 1 ][ na ],
+                        silk_Transition_LP_A_Q28[ ind + 1 ][ na ] -
+                        silk_Transition_LP_A_Q28[ ind     ][ na ],
+                        fac_Q16 - ( (opus_int32)1 << 16 ) );
+                }
+            }
+        } else {
+            silk_memcpy( B_Q28, silk_Transition_LP_B_Q28[ ind ], TRANSITION_NB * sizeof( opus_int32 ) );
+            silk_memcpy( A_Q28, silk_Transition_LP_A_Q28[ ind ], TRANSITION_NA * sizeof( opus_int32 ) );
+        }
+    } else {
+        silk_memcpy( B_Q28, silk_Transition_LP_B_Q28[ TRANSITION_INT_NUM - 1 ], TRANSITION_NB * sizeof( opus_int32 ) );
+        silk_memcpy( A_Q28, silk_Transition_LP_A_Q28[ TRANSITION_INT_NUM - 1 ], TRANSITION_NA * sizeof( opus_int32 ) );
+    }
+}
+
+/* Low-pass filter with variable cutoff frequency based on  */
+/* piece-wise linear interpolation between elliptic filters */
+/* Start by setting psEncC->mode <> 0;                      */
+/* Deactivate by setting psEncC->mode = 0;                  */
+void silk_LP_variable_cutoff(
+    silk_LP_state               *psLP,                          /* I/O  LP filter state                             */
+    opus_int16                  *frame,                         /* I/O  Low-pass filtered output signal             */
+    const opus_int              frame_length                    /* I    Frame length                                */
+)
+{
+    opus_int32   B_Q28[ TRANSITION_NB ], A_Q28[ TRANSITION_NA ], fac_Q16 = 0;
+    opus_int     ind = 0;
+
+    silk_assert( psLP->transition_frame_no >= 0 && psLP->transition_frame_no <= TRANSITION_FRAMES );
+
+    /* Run filter if needed */
+    if( psLP->mode != 0 ) {
+        /* Calculate index and interpolation factor for interpolation */
+#if( TRANSITION_INT_STEPS == 64 )
+        fac_Q16 = silk_LSHIFT( TRANSITION_FRAMES - psLP->transition_frame_no, 16 - 6 );
+#else
+        fac_Q16 = silk_DIV32_16( silk_LSHIFT( TRANSITION_FRAMES - psLP->transition_frame_no, 16 ), TRANSITION_FRAMES );
+#endif
+        ind      = silk_RSHIFT( fac_Q16, 16 );
+        fac_Q16 -= silk_LSHIFT( ind, 16 );
+
+        silk_assert( ind >= 0 );
+        silk_assert( ind < TRANSITION_INT_NUM );
+
+        /* Interpolate filter coefficients */
+        silk_LP_interpolate_filter_taps( B_Q28, A_Q28, ind, fac_Q16 );
+
+        /* Update transition frame number for next frame */
+        psLP->transition_frame_no = silk_LIMIT( psLP->transition_frame_no + psLP->mode, 0, TRANSITION_FRAMES );
+
+        /* ARMA low-pass filtering */
+        silk_assert( TRANSITION_NB == 3 && TRANSITION_NA == 2 );
+        silk_biquad_alt_stride1( frame, B_Q28, A_Q28, psLP->In_LP_State, frame, frame_length);
+    }
+}
--- a/lib/libopus/silk/MacroCount.h
+++ b/lib/libopus/silk/MacroCount.h
@ -0,0 +1,709 @@
+/***********************************************************************
+Copyright (c) 2006-2011, Skype Limited. All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+- Redistributions of source code must retain the above copyright notice,
+this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+***********************************************************************/
+
+#ifndef SIGPROCFIX_API_MACROCOUNT_H
+#define SIGPROCFIX_API_MACROCOUNT_H
+
+#ifdef    silk_MACRO_COUNT
+#include <stdio.h>
+#define varDefine opus_int64 ops_count = 0;
+
+extern opus_int64 ops_count;
+
+static OPUS_INLINE opus_int64 silk_SaveCount(){
+    return(ops_count);
+}
+
+static OPUS_INLINE opus_int64 silk_SaveResetCount(){
+    opus_int64 ret;
+
+    ret = ops_count;
+    ops_count = 0;
+    return(ret);
+}
+
+static OPUS_INLINE silk_PrintCount(){
+    printf("ops_count = %d \n ", (opus_int32)ops_count);
+}
+
+#undef silk_MUL
+static OPUS_INLINE opus_int32 silk_MUL(opus_int32 a32, opus_int32 b32){
+    opus_int32 ret;
+    ops_count += 4;
+    ret = a32 * b32;
+    return ret;
+}
+
+#undef silk_MUL_uint
+static OPUS_INLINE opus_uint32 silk_MUL_uint(opus_uint32 a32, opus_uint32 b32){
+    opus_uint32 ret;
+    ops_count += 4;
+    ret = a32 * b32;
+    return ret;
+}
+#undef silk_MLA
+static OPUS_INLINE opus_int32 silk_MLA(opus_int32 a32, opus_int32 b32, opus_int32 c32){
+    opus_int32 ret;
+    ops_count += 4;
+    ret = a32 + b32 * c32;
+    return ret;
+}
+
+#undef silk_MLA_uint
+static OPUS_INLINE opus_int32 silk_MLA_uint(opus_uint32 a32, opus_uint32 b32, opus_uint32 c32){
+    opus_uint32 ret;
+    ops_count += 4;
+    ret = a32 + b32 * c32;
+    return ret;
+}
+
+#undef silk_SMULWB
+static OPUS_INLINE opus_int32 silk_SMULWB(opus_int32 a32, opus_int32 b32){
+    opus_int32 ret;
+    ops_count += 5;
+    ret = (a32 >> 16) * (opus_int32)((opus_int16)b32) + (((a32 & 0x0000FFFF) * (opus_int32)((opus_int16)b32)) >> 16);
+    return ret;
+}
+#undef    silk_SMLAWB
+static OPUS_INLINE opus_int32 silk_SMLAWB(opus_int32 a32, opus_int32 b32, opus_int32 c32){
+    opus_int32 ret;
+    ops_count += 5;
+    ret = ((a32) + ((((b32) >> 16) * (opus_int32)((opus_int16)(c32))) + ((((b32) & 0x0000FFFF) * (opus_int32)((opus_int16)(c32))) >> 16)));
+    return ret;
+}
+
+#undef silk_SMULWT
+static OPUS_INLINE opus_int32 silk_SMULWT(opus_int32 a32, opus_int32 b32){
+    opus_int32 ret;
+    ops_count += 4;
+    ret = (a32 >> 16) * (b32 >> 16) + (((a32 & 0x0000FFFF) * (b32 >> 16)) >> 16);
+    return ret;
+}
+#undef silk_SMLAWT
+static OPUS_INLINE opus_int32 silk_SMLAWT(opus_int32 a32, opus_int32 b32, opus_int32 c32){
+    opus_int32 ret;
+    ops_count += 4;
+    ret = a32 + ((b32 >> 16) * (c32 >> 16)) + (((b32 & 0x0000FFFF) * ((c32 >> 16)) >> 16));
+    return ret;
+}
+
+#undef silk_SMULBB
+static OPUS_INLINE opus_int32 silk_SMULBB(opus_int32 a32, opus_int32 b32){
+    opus_int32 ret;
+    ops_count += 1;
+    ret = (opus_int32)((opus_int16)a32) * (opus_int32)((opus_int16)b32);
+    return ret;
+}
+#undef silk_SMLABB
+static OPUS_INLINE opus_int32 silk_SMLABB(opus_int32 a32, opus_int32 b32, opus_int32 c32){
+    opus_int32 ret;
+    ops_count += 1;
+    ret = a32 + (opus_int32)((opus_int16)b32) * (opus_int32)((opus_int16)c32);
+    return ret;
+}
+
+#undef silk_SMULBT
+static OPUS_INLINE opus_int32 silk_SMULBT(opus_int32 a32, opus_int32 b32 ){
+    opus_int32 ret;
+    ops_count += 4;
+    ret = ((opus_int32)((opus_int16)a32)) * (b32 >> 16);
+    return ret;
+}
+
+#undef silk_SMLABT
+static OPUS_INLINE opus_int32 silk_SMLABT(opus_int32 a32, opus_int32 b32, opus_int32 c32){
+    opus_int32 ret;
+    ops_count += 1;
+    ret = a32 + ((opus_int32)((opus_int16)b32)) * (c32 >> 16);
+    return ret;
+}
+
+#undef silk_SMULTT
+static OPUS_INLINE opus_int32 silk_SMULTT(opus_int32 a32, opus_int32 b32){
+    opus_int32 ret;
+    ops_count += 1;
+    ret = (a32 >> 16) * (b32 >> 16);
+    return ret;
+}
+
+#undef    silk_SMLATT
+static OPUS_INLINE opus_int32 silk_SMLATT(opus_int32 a32, opus_int32 b32, opus_int32 c32){
+    opus_int32 ret;
+    ops_count += 1;
+    ret = a32 + (b32 >> 16) * (c32 >> 16);
+    return ret;
+}
+
+
+/* multiply-accumulate macros that allow overflow in the addition (ie, no asserts in debug mode)*/
+#undef    silk_MLA_ovflw
+#define silk_MLA_ovflw silk_MLA
+
+#undef silk_SMLABB_ovflw
+#define silk_SMLABB_ovflw silk_SMLABB
+
+#undef silk_SMLABT_ovflw
+#define silk_SMLABT_ovflw silk_SMLABT
+
+#undef silk_SMLATT_ovflw
+#define silk_SMLATT_ovflw silk_SMLATT
+
+#undef silk_SMLAWB_ovflw
+#define silk_SMLAWB_ovflw silk_SMLAWB
+
+#undef silk_SMLAWT_ovflw
+#define silk_SMLAWT_ovflw silk_SMLAWT
+
+#undef silk_SMULL
+static OPUS_INLINE opus_int64 silk_SMULL(opus_int32 a32, opus_int32 b32){
+    opus_int64 ret;
+    ops_count += 8;
+    ret = ((opus_int64)(a32) * /*(opus_int64)*/(b32));
+    return ret;
+}
+
+#undef    silk_SMLAL
+static OPUS_INLINE opus_int64 silk_SMLAL(opus_int64 a64, opus_int32 b32, opus_int32 c32){
+    opus_int64 ret;
+    ops_count += 8;
+    ret = a64 + ((opus_int64)(b32) * /*(opus_int64)*/(c32));
+    return ret;
+}
+#undef    silk_SMLALBB
+static OPUS_INLINE opus_int64 silk_SMLALBB(opus_int64 a64, opus_int16 b16, opus_int16 c16){
+    opus_int64 ret;
+    ops_count += 4;
+    ret = a64 + ((opus_int64)(b16) * /*(opus_int64)*/(c16));
+    return ret;
+}
+
+#undef    SigProcFIX_CLZ16
+static OPUS_INLINE opus_int32 SigProcFIX_CLZ16(opus_int16 in16)
+{
+    opus_int32 out32 = 0;
+    ops_count += 10;
+    if( in16 == 0 ) {
+        return 16;
+    }
+    /* test nibbles */
+    if( in16 & 0xFF00 ) {
+        if( in16 & 0xF000 ) {
+            in16 >>= 12;
+        } else {
+            out32 += 4;
+            in16 >>= 8;
+        }
+    } else {
+        if( in16 & 0xFFF0 ) {
+            out32 += 8;
+            in16 >>= 4;
+        } else {
+            out32 += 12;
+        }
+    }
+    /* test bits and return */
+    if( in16 & 0xC ) {
+        if( in16 & 0x8 )
+            return out32 + 0;
+        else
+            return out32 + 1;
+    } else {
+        if( in16 & 0xE )
+            return out32 + 2;
+        else
+            return out32 + 3;
+    }
+}
+
+#undef SigProcFIX_CLZ32
+static OPUS_INLINE opus_int32 SigProcFIX_CLZ32(opus_int32 in32)
+{
+    /* test highest 16 bits and convert to opus_int16 */
+    ops_count += 2;
+    if( in32 & 0xFFFF0000 ) {
+        return SigProcFIX_CLZ16((opus_int16)(in32 >> 16));
+    } else {
+        return SigProcFIX_CLZ16((opus_int16)in32) + 16;
+    }
+}
+
+#undef silk_DIV32
+static OPUS_INLINE opus_int32 silk_DIV32(opus_int32 a32, opus_int32 b32){
+    ops_count += 64;
+    return a32 / b32;
+}
+
+#undef silk_DIV32_16
+static OPUS_INLINE opus_int32 silk_DIV32_16(opus_int32 a32, opus_int32 b32){
+    ops_count += 32;
+    return a32 / b32;
+}
+
+#undef silk_SAT8
+static OPUS_INLINE opus_int8 silk_SAT8(opus_int64 a){
+    opus_int8 tmp;
+    ops_count += 1;
+    tmp = (opus_int8)((a) > silk_int8_MAX ? silk_int8_MAX  : \
+                    ((a) < silk_int8_MIN ? silk_int8_MIN  : (a)));
+    return(tmp);
+}
+
+#undef silk_SAT16
+static OPUS_INLINE opus_int16 silk_SAT16(opus_int64 a){
+    opus_int16 tmp;
+    ops_count += 1;
+    tmp = (opus_int16)((a) > silk_int16_MAX ? silk_int16_MAX  : \
+                     ((a) < silk_int16_MIN ? silk_int16_MIN  : (a)));
+    return(tmp);
+}
+#undef silk_SAT32
+static OPUS_INLINE opus_int32 silk_SAT32(opus_int64 a){
+    opus_int32 tmp;
+    ops_count += 1;
+    tmp = (opus_int32)((a) > silk_int32_MAX ? silk_int32_MAX  : \
+                     ((a) < silk_int32_MIN ? silk_int32_MIN  : (a)));
+    return(tmp);
+}
+#undef silk_POS_SAT32
+static OPUS_INLINE opus_int32 silk_POS_SAT32(opus_int64 a){
+    opus_int32 tmp;
+    ops_count += 1;
+    tmp = (opus_int32)((a) > silk_int32_MAX ? silk_int32_MAX : (a));
+    return(tmp);
+}
+
+#undef silk_ADD_POS_SAT8
+static OPUS_INLINE opus_int8 silk_ADD_POS_SAT8(opus_int64 a, opus_int64 b){
+    opus_int8 tmp;
+    ops_count += 1;
+    tmp = (opus_int8)((((a)+(b)) & 0x80) ? silk_int8_MAX  : ((a)+(b)));
+    return(tmp);
+}
+#undef silk_ADD_POS_SAT16
+static OPUS_INLINE opus_int16 silk_ADD_POS_SAT16(opus_int64 a, opus_int64 b){
+    opus_int16 tmp;
+    ops_count += 1;
+    tmp = (opus_int16)((((a)+(b)) & 0x8000) ? silk_int16_MAX : ((a)+(b)));
+    return(tmp);
+}
+
+#undef silk_ADD_POS_SAT32
+static OPUS_INLINE opus_int32 silk_ADD_POS_SAT32(opus_int64 a, opus_int64 b){
+    opus_int32 tmp;
+    ops_count += 1;
+    tmp = (opus_int32)((((a)+(b)) & 0x80000000) ? silk_int32_MAX : ((a)+(b)));
+    return(tmp);
+}
+
+#undef    silk_LSHIFT8
+static OPUS_INLINE opus_int8 silk_LSHIFT8(opus_int8 a, opus_int32 shift){
+    opus_int8 ret;
+    ops_count += 1;
+    ret = a << shift;
+    return ret;
+}
+#undef    silk_LSHIFT16
+static OPUS_INLINE opus_int16 silk_LSHIFT16(opus_int16 a, opus_int32 shift){
+    opus_int16 ret;
+    ops_count += 1;
+    ret = a << shift;
+    return ret;
+}
+#undef    silk_LSHIFT32
+static OPUS_INLINE opus_int32 silk_LSHIFT32(opus_int32 a, opus_int32 shift){
+    opus_int32 ret;
+    ops_count += 1;
+    ret = a << shift;
+    return ret;
+}
+#undef    silk_LSHIFT64
+static OPUS_INLINE opus_int64 silk_LSHIFT64(opus_int64 a, opus_int shift){
+    ops_count += 1;
+    return a << shift;
+}
+
+#undef    silk_LSHIFT_ovflw
+static OPUS_INLINE opus_int32 silk_LSHIFT_ovflw(opus_int32 a, opus_int32 shift){
+    ops_count += 1;
+    return a << shift;
+}
+
+#undef    silk_LSHIFT_uint
+static OPUS_INLINE opus_uint32 silk_LSHIFT_uint(opus_uint32 a, opus_int32 shift){
+    opus_uint32 ret;
+    ops_count += 1;
+    ret = a << shift;
+    return ret;
+}
+
+#undef    silk_RSHIFT8
+static OPUS_INLINE opus_int8 silk_RSHIFT8(opus_int8 a, opus_int32 shift){
+    ops_count += 1;
+    return a >> shift;
+}
+#undef    silk_RSHIFT16
+static OPUS_INLINE opus_int16 silk_RSHIFT16(opus_int16 a, opus_int32 shift){
+    ops_count += 1;
+    return a >> shift;
+}
+#undef    silk_RSHIFT32
+static OPUS_INLINE opus_int32 silk_RSHIFT32(opus_int32 a, opus_int32 shift){
+    ops_count += 1;
+    return a >> shift;
+}
+#undef    silk_RSHIFT64
+static OPUS_INLINE opus_int64 silk_RSHIFT64(opus_int64 a, opus_int64 shift){
+    ops_count += 1;
+    return a >> shift;
+}
+
+#undef    silk_RSHIFT_uint
+static OPUS_INLINE opus_uint32 silk_RSHIFT_uint(opus_uint32 a, opus_int32 shift){
+    ops_count += 1;
+    return a >> shift;
+}
+
+#undef    silk_ADD_LSHIFT
+static OPUS_INLINE opus_int32 silk_ADD_LSHIFT(opus_int32 a, opus_int32 b, opus_int32 shift){
+    opus_int32 ret;
+    ops_count += 1;
+    ret = a + (b << shift);
+    return ret;                /* shift >= 0*/
+}
+#undef    silk_ADD_LSHIFT32
+static OPUS_INLINE opus_int32 silk_ADD_LSHIFT32(opus_int32 a, opus_int32 b, opus_int32 shift){
+    opus_int32 ret;
+    ops_count += 1;
+    ret = a + (b << shift);
+    return ret;                /* shift >= 0*/
+}
+#undef    silk_ADD_LSHIFT_uint
+static OPUS_INLINE opus_uint32 silk_ADD_LSHIFT_uint(opus_uint32 a, opus_uint32 b, opus_int32 shift){
+    opus_uint32 ret;
+    ops_count += 1;
+    ret = a + (b << shift);
+    return ret;                /* shift >= 0*/
+}
+#undef    silk_ADD_RSHIFT
+static OPUS_INLINE opus_int32 silk_ADD_RSHIFT(opus_int32 a, opus_int32 b, opus_int32 shift){
+    opus_int32 ret;
+    ops_count += 1;
+    ret = a + (b >> shift);
+    return ret;                /* shift  > 0*/
+}
+#undef    silk_ADD_RSHIFT32
+static OPUS_INLINE opus_int32 silk_ADD_RSHIFT32(opus_int32 a, opus_int32 b, opus_int32 shift){
+    opus_int32 ret;
+    ops_count += 1;
+    ret = a + (b >> shift);
+    return ret;                /* shift  > 0*/
+}
+#undef    silk_ADD_RSHIFT_uint
+static OPUS_INLINE opus_uint32 silk_ADD_RSHIFT_uint(opus_uint32 a, opus_uint32 b, opus_int32 shift){
+    opus_uint32 ret;
+    ops_count += 1;
+    ret = a + (b >> shift);
+    return ret;                /* shift  > 0*/
+}
+#undef    silk_SUB_LSHIFT32
+static OPUS_INLINE opus_int32 silk_SUB_LSHIFT32(opus_int32 a, opus_int32 b, opus_int32 shift){
+    opus_int32 ret;
+    ops_count += 1;
+    ret = a - (b << shift);
+    return ret;                /* shift >= 0*/
+}
+#undef    silk_SUB_RSHIFT32
+static OPUS_INLINE opus_int32 silk_SUB_RSHIFT32(opus_int32 a, opus_int32 b, opus_int32 shift){
+    opus_int32 ret;
+    ops_count += 1;
+    ret = a - (b >> shift);
+    return ret;                /* shift  > 0*/
+}
+
+#undef    silk_RSHIFT_ROUND
+static OPUS_INLINE opus_int32 silk_RSHIFT_ROUND(opus_int32 a, opus_int32 shift){
+    opus_int32 ret;
+    ops_count += 3;
+    ret = shift == 1 ? (a >> 1) + (a & 1) : ((a >> (shift - 1)) + 1) >> 1;
+    return ret;
+}
+
+#undef    silk_RSHIFT_ROUND64
+static OPUS_INLINE opus_int64 silk_RSHIFT_ROUND64(opus_int64 a, opus_int32 shift){
+    opus_int64 ret;
+    ops_count += 6;
+    ret = shift == 1 ? (a >> 1) + (a & 1) : ((a >> (shift - 1)) + 1) >> 1;
+    return ret;
+}
+
+#undef    silk_abs_int64
+static OPUS_INLINE opus_int64 silk_abs_int64(opus_int64 a){
+    ops_count += 1;
+    return (((a) >  0)  ? (a) : -(a));            /* Be careful, silk_abs returns wrong when input equals to silk_intXX_MIN*/
+}
+
+#undef    silk_abs_int32
+static OPUS_INLINE opus_int32 silk_abs_int32(opus_int32 a){
+    ops_count += 1;
+    return silk_abs(a);
+}
+
+
+#undef silk_min
+static silk_min(a, b){
+    ops_count += 1;
+    return (((a) < (b)) ? (a) :  (b));
+}
+#undef silk_max
+static silk_max(a, b){
+    ops_count += 1;
+    return (((a) > (b)) ? (a) :  (b));
+}
+#undef silk_sign
+static silk_sign(a){
+    ops_count += 1;
+    return ((a) > 0 ? 1 : ( (a) < 0 ? -1 : 0 ));
+}
+
+#undef    silk_ADD16
+static OPUS_INLINE opus_int16 silk_ADD16(opus_int16 a, opus_int16 b){
+    opus_int16 ret;
+    ops_count += 1;
+    ret = a + b;
+    return ret;
+}
+
+#undef    silk_ADD32
+static OPUS_INLINE opus_int32 silk_ADD32(opus_int32 a, opus_int32 b){
+    opus_int32 ret;
+    ops_count += 1;
+    ret = a + b;
+    return ret;
+}
+
+#undef    silk_ADD64
+static OPUS_INLINE opus_int64 silk_ADD64(opus_int64 a, opus_int64 b){
+    opus_int64 ret;
+    ops_count += 2;
+    ret = a + b;
+    return ret;
+}
+
+#undef    silk_SUB16
+static OPUS_INLINE opus_int16 silk_SUB16(opus_int16 a, opus_int16 b){
+    opus_int16 ret;
+    ops_count += 1;
+    ret = a - b;
+    return ret;
+}
+
+#undef    silk_SUB32
+static OPUS_INLINE opus_int32 silk_SUB32(opus_int32 a, opus_int32 b){
+    opus_int32 ret;
+    ops_count += 1;
+    ret = a - b;
+    return ret;
+}
+
+#undef    silk_SUB64
+static OPUS_INLINE opus_int64 silk_SUB64(opus_int64 a, opus_int64 b){
+    opus_int64 ret;
+    ops_count += 2;
+    ret = a - b;
+    return ret;
+}
+
+#undef silk_ADD_SAT16
+static OPUS_INLINE opus_int16 silk_ADD_SAT16( opus_int16 a16, opus_int16 b16 ) {
+    opus_int16 res;
+    /* Nb will be counted in AKP_add32 and silk_SAT16*/
+    res = (opus_int16)silk_SAT16( silk_ADD32( (opus_int32)(a16), (b16) ) );
+    return res;
+}
+
+#undef silk_ADD_SAT32
+static OPUS_INLINE opus_int32 silk_ADD_SAT32(opus_int32 a32, opus_int32 b32){
+    opus_int32 res;
+    ops_count += 1;
+    res =    ((((a32) + (b32)) & 0x80000000) == 0 ?                                    \
+            ((((a32) & (b32)) & 0x80000000) != 0 ? silk_int32_MIN : (a32)+(b32)) :    \
+            ((((a32) | (b32)) & 0x80000000) == 0 ? silk_int32_MAX : (a32)+(b32)) );
+    return res;
+}
+
+#undef silk_ADD_SAT64
+static OPUS_INLINE opus_int64 silk_ADD_SAT64( opus_int64 a64, opus_int64 b64 ) {
+    opus_int64 res;
+    ops_count += 1;
+    res =    ((((a64) + (b64)) & 0x8000000000000000LL) == 0 ?                                \
+            ((((a64) & (b64)) & 0x8000000000000000LL) != 0 ? silk_int64_MIN : (a64)+(b64)) :    \
+            ((((a64) | (b64)) & 0x8000000000000000LL) == 0 ? silk_int64_MAX : (a64)+(b64)) );
+    return res;
+}
+
+#undef silk_SUB_SAT16
+static OPUS_INLINE opus_int16 silk_SUB_SAT16( opus_int16 a16, opus_int16 b16 ) {
+    opus_int16 res;
+    silk_assert(0);
+    /* Nb will be counted in sub-macros*/
+    res = (opus_int16)silk_SAT16( silk_SUB32( (opus_int32)(a16), (b16) ) );
+    return res;
+}
+
+#undef silk_SUB_SAT32
+static OPUS_INLINE opus_int32 silk_SUB_SAT32( opus_int32 a32, opus_int32 b32 ) {
+    opus_int32 res;
+    ops_count += 1;
+    res =     ((((a32)-(b32)) & 0x80000000) == 0 ?                                            \
+            (( (a32) & ((b32)^0x80000000) & 0x80000000) ? silk_int32_MIN : (a32)-(b32)) :    \
+            ((((a32)^0x80000000) & (b32)  & 0x80000000) ? silk_int32_MAX : (a32)-(b32)) );
+    return res;
+}
+
+#undef silk_SUB_SAT64
+static OPUS_INLINE opus_int64 silk_SUB_SAT64( opus_int64 a64, opus_int64 b64 ) {
+    opus_int64 res;
+    ops_count += 1;
+    res =    ((((a64)-(b64)) & 0x8000000000000000LL) == 0 ?                                                        \
+            (( (a64) & ((b64)^0x8000000000000000LL) & 0x8000000000000000LL) ? silk_int64_MIN : (a64)-(b64)) :    \
+            ((((a64)^0x8000000000000000LL) & (b64)  & 0x8000000000000000LL) ? silk_int64_MAX : (a64)-(b64)) );
+
+    return res;
+}
+
+#undef    silk_SMULWW
+static OPUS_INLINE opus_int32 silk_SMULWW(opus_int32 a32, opus_int32 b32){
+    opus_int32 ret;
+    /* Nb will be counted in sub-macros*/
+    ret = silk_MLA(silk_SMULWB((a32), (b32)), (a32), silk_RSHIFT_ROUND((b32), 16));
+    return ret;
+}
+
+#undef    silk_SMLAWW
+static OPUS_INLINE opus_int32 silk_SMLAWW(opus_int32 a32, opus_int32 b32, opus_int32 c32){
+    opus_int32 ret;
+    /* Nb will be counted in sub-macros*/
+    ret = silk_MLA(silk_SMLAWB((a32), (b32), (c32)), (b32), silk_RSHIFT_ROUND((c32), 16));
+    return ret;
+}
+
+#undef    silk_min_int
+static OPUS_INLINE opus_int silk_min_int(opus_int a, opus_int b)
+{
+    ops_count += 1;
+    return (((a) < (b)) ? (a) : (b));
+}
+
+#undef    silk_min_16
+static OPUS_INLINE opus_int16 silk_min_16(opus_int16 a, opus_int16 b)
+{
+    ops_count += 1;
+    return (((a) < (b)) ? (a) : (b));
+}
+#undef    silk_min_32
+static OPUS_INLINE opus_int32 silk_min_32(opus_int32 a, opus_int32 b)
+{
+    ops_count += 1;
+    return (((a) < (b)) ? (a) : (b));
+}
+#undef    silk_min_64
+static OPUS_INLINE opus_int64 silk_min_64(opus_int64 a, opus_int64 b)
+{
+    ops_count += 1;
+    return (((a) < (b)) ? (a) : (b));
+}
+
+/* silk_min() versions with typecast in the function call */
+#undef    silk_max_int
+static OPUS_INLINE opus_int silk_max_int(opus_int a, opus_int b)
+{
+    ops_count += 1;
+    return (((a) > (b)) ? (a) : (b));
+}
+#undef    silk_max_16
+static OPUS_INLINE opus_int16 silk_max_16(opus_int16 a, opus_int16 b)
+{
+    ops_count += 1;
+    return (((a) > (b)) ? (a) : (b));
+}
+#undef    silk_max_32
+static OPUS_INLINE opus_int32 silk_max_32(opus_int32 a, opus_int32 b)
+{
+    ops_count += 1;
+    return (((a) > (b)) ? (a) : (b));
+}
+
+#undef    silk_max_64
+static OPUS_INLINE opus_int64 silk_max_64(opus_int64 a, opus_int64 b)
+{
+    ops_count += 1;
+    return (((a) > (b)) ? (a) : (b));
+}
+
+
+#undef silk_LIMIT_int
+static OPUS_INLINE opus_int silk_LIMIT_int(opus_int a, opus_int limit1, opus_int limit2)
+{
+    opus_int ret;
+    ops_count += 6;
+
+    ret = ((limit1) > (limit2) ? ((a) > (limit1) ? (limit1) : ((a) < (limit2) ? (limit2) : (a))) \
+        : ((a) > (limit2) ? (limit2) : ((a) < (limit1) ? (limit1) : (a))));
+
+    return(ret);
+}
+
+#undef silk_LIMIT_16
+static OPUS_INLINE opus_int16 silk_LIMIT_16(opus_int16 a, opus_int16 limit1, opus_int16 limit2)
+{
+    opus_int16 ret;
+    ops_count += 6;
+
+    ret = ((limit1) > (limit2) ? ((a) > (limit1) ? (limit1) : ((a) < (limit2) ? (limit2) : (a))) \
+        : ((a) > (limit2) ? (limit2) : ((a) < (limit1) ? (limit1) : (a))));
+
+return(ret);
+}
+
+
+#undef silk_LIMIT_32
+static OPUS_INLINE opus_int32 silk_LIMIT_32(opus_int32 a, opus_int32 limit1, opus_int32 limit2)
+{
+    opus_int32 ret;
+    ops_count += 6;
+
+    ret = ((limit1) > (limit2) ? ((a) > (limit1) ? (limit1) : ((a) < (limit2) ? (limit2) : (a))) \
+        : ((a) > (limit2) ? (limit2) : ((a) < (limit1) ? (limit1) : (a))));
+    return(ret);
+}
+
+#else
+#define varDefine
+#define silk_SaveCount()
+
+#endif
+#endif
--- a/lib/libopus/silk/MacroDebug.h
+++ b/lib/libopus/silk/MacroDebug.h
@ -0,0 +1,945 @@
+/***********************************************************************
+Copyright (c) 2006-2011, Skype Limited. All rights reserved.
+Copyright (C) 2012 Xiph.Org Foundation
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+- Redistributions of source code must retain the above copyright notice,
+this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+***********************************************************************/
+
+#ifndef MACRO_DEBUG_H
+#define MACRO_DEBUG_H
+
+/* Redefine macro functions with extensive assertion in DEBUG mode.
+   As functions can't be undefined, this file can't work with SigProcFIX_MacroCount.h */
+
+#if ( defined (FIXED_DEBUG) || ( 0 && defined (_DEBUG) ) ) && !defined (silk_MACRO_COUNT)
+
+#undef silk_ADD16
+#define silk_ADD16(a,b) silk_ADD16_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_int16 silk_ADD16_(opus_int16 a, opus_int16 b, char *file, int line){
+    opus_int16 ret;
+
+    ret = a + b;
+    if ( ret != silk_ADD_SAT16( a, b ) )
+    {
+        fprintf (stderr, "silk_ADD16(%d, %d) in %s: line %d\n", a, b, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;
+}
+
+#undef silk_ADD32
+#define silk_ADD32(a,b) silk_ADD32_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_int32 silk_ADD32_(opus_int32 a, opus_int32 b, char *file, int line){
+    opus_int32 ret;
+
+    ret = (opus_int32)((opus_uint32)a + (opus_uint32)b);
+    if ( ret != silk_ADD_SAT32( a, b ) )
+    {
+        fprintf (stderr, "silk_ADD32(%d, %d) in %s: line %d\n", a, b, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;
+}
+
+#undef silk_ADD64
+#define silk_ADD64(a,b) silk_ADD64_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_int64 silk_ADD64_(opus_int64 a, opus_int64 b, char *file, int line){
+    opus_int64 ret;
+
+    ret = a + b;
+    if ( ret != silk_ADD_SAT64( a, b ) )
+    {
+        fprintf (stderr, "silk_ADD64(%lld, %lld) in %s: line %d\n", (long long)a, (long long)b, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;
+}
+
+#undef silk_SUB16
+#define silk_SUB16(a,b) silk_SUB16_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_int16 silk_SUB16_(opus_int16 a, opus_int16 b, char *file, int line){
+    opus_int16 ret;
+
+    ret = a - b;
+    if ( ret != silk_SUB_SAT16( a, b ) )
+    {
+        fprintf (stderr, "silk_SUB16(%d, %d) in %s: line %d\n", a, b, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;
+}
+
+#undef silk_SUB32
+#define silk_SUB32(a,b) silk_SUB32_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_int32 silk_SUB32_(opus_int32 a, opus_int32 b, char *file, int line){
+    opus_int64 ret;
+
+    ret = a - (opus_int64)b;
+    if ( ret != silk_SUB_SAT32( a, b ) )
+    {
+        fprintf (stderr, "silk_SUB32(%d, %d) in %s: line %d\n", a, b, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;
+}
+
+#undef silk_SUB64
+#define silk_SUB64(a,b) silk_SUB64_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_int64 silk_SUB64_(opus_int64 a, opus_int64 b, char *file, int line){
+    opus_int64 ret;
+
+    ret = a - b;
+    if ( ret != silk_SUB_SAT64( a, b ) )
+    {
+        fprintf (stderr, "silk_SUB64(%lld, %lld) in %s: line %d\n", (long long)a, (long long)b, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;
+}
+
+#undef silk_ADD_SAT16
+#define silk_ADD_SAT16(a,b) silk_ADD_SAT16_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_int16 silk_ADD_SAT16_( opus_int16 a16, opus_int16 b16, char *file, int line) {
+    opus_int16 res;
+    res = (opus_int16)silk_SAT16( silk_ADD32( (opus_int32)(a16), (b16) ) );
+    if ( res != silk_SAT16( (opus_int32)a16 + (opus_int32)b16 ) )
+    {
+        fprintf (stderr, "silk_ADD_SAT16(%d, %d) in %s: line %d\n", a16, b16, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return res;
+}
+
+#undef silk_ADD_SAT32
+#define silk_ADD_SAT32(a,b) silk_ADD_SAT32_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_int32 silk_ADD_SAT32_(opus_int32 a32, opus_int32 b32, char *file, int line){
+    opus_int32 res;
+    res =   ((((opus_uint32)(a32) + (opus_uint32)(b32)) & 0x80000000) == 0 ?       \
+            ((((a32) & (b32)) & 0x80000000) != 0 ? silk_int32_MIN : (a32)+(b32)) : \
+            ((((a32) | (b32)) & 0x80000000) == 0 ? silk_int32_MAX : (a32)+(b32)) );
+    if ( res != silk_SAT32( (opus_int64)a32 + (opus_int64)b32 ) )
+    {
+        fprintf (stderr, "silk_ADD_SAT32(%d, %d) in %s: line %d\n", a32, b32, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return res;
+}
+
+#undef silk_ADD_SAT64
+#define silk_ADD_SAT64(a,b) silk_ADD_SAT64_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_int64 silk_ADD_SAT64_( opus_int64 a64, opus_int64 b64, char *file, int line) {
+    opus_int64 res;
+    int        fail = 0;
+    res =   ((((a64) + (b64)) & 0x8000000000000000LL) == 0 ?                                 \
+            ((((a64) & (b64)) & 0x8000000000000000LL) != 0 ? silk_int64_MIN : (a64)+(b64)) : \
+            ((((a64) | (b64)) & 0x8000000000000000LL) == 0 ? silk_int64_MAX : (a64)+(b64)) );
+    if( res != a64 + b64 ) {
+        /* Check that we saturated to the correct extreme value */
+        if ( !(( res == silk_int64_MAX && ( ( a64 >> 1 ) + ( b64 >> 1 ) > ( silk_int64_MAX >> 3 ) ) ) ||
+               ( res == silk_int64_MIN && ( ( a64 >> 1 ) + ( b64 >> 1 ) < ( silk_int64_MIN >> 3 ) ) ) ) )
+        {
+            fail = 1;
+        }
+    } else {
+        /* Saturation not necessary */
+        fail = res != a64 + b64;
+    }
+    if ( fail )
+    {
+        fprintf (stderr, "silk_ADD_SAT64(%lld, %lld) in %s: line %d\n", (long long)a64, (long long)b64, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return res;
+}
+
+#undef silk_SUB_SAT16
+#define silk_SUB_SAT16(a,b) silk_SUB_SAT16_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_int16 silk_SUB_SAT16_( opus_int16 a16, opus_int16 b16, char *file, int line ) {
+    opus_int16 res;
+    res = (opus_int16)silk_SAT16( silk_SUB32( (opus_int32)(a16), (b16) ) );
+    if ( res != silk_SAT16( (opus_int32)a16 - (opus_int32)b16 ) )
+    {
+        fprintf (stderr, "silk_SUB_SAT16(%d, %d) in %s: line %d\n", a16, b16, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return res;
+}
+
+#undef silk_SUB_SAT32
+#define silk_SUB_SAT32(a,b) silk_SUB_SAT32_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_int32 silk_SUB_SAT32_( opus_int32 a32, opus_int32 b32, char *file, int line ) {
+    opus_int32 res;
+    res =   ((((opus_uint32)(a32)-(opus_uint32)(b32)) & 0x80000000) == 0 ?                \
+            (( (a32) & ((b32)^0x80000000) & 0x80000000) ? silk_int32_MIN : (a32)-(b32)) : \
+            ((((a32)^0x80000000) & (b32)  & 0x80000000) ? silk_int32_MAX : (a32)-(b32)) );
+    if ( res != silk_SAT32( (opus_int64)a32 - (opus_int64)b32 ) )
+    {
+        fprintf (stderr, "silk_SUB_SAT32(%d, %d) in %s: line %d\n", a32, b32, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return res;
+}
+
+#undef silk_SUB_SAT64
+#define silk_SUB_SAT64(a,b) silk_SUB_SAT64_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_int64 silk_SUB_SAT64_( opus_int64 a64, opus_int64 b64, char *file, int line ) {
+    opus_int64 res;
+    int        fail = 0;
+    res =   ((((a64)-(b64)) & 0x8000000000000000LL) == 0 ?                                                    \
+            (( (a64) & ((b64)^0x8000000000000000LL) & 0x8000000000000000LL) ? silk_int64_MIN : (a64)-(b64)) : \
+            ((((a64)^0x8000000000000000LL) & (b64)  & 0x8000000000000000LL) ? silk_int64_MAX : (a64)-(b64)) );
+    if( res != a64 - b64 ) {
+        /* Check that we saturated to the correct extreme value */
+        if( !(( res == silk_int64_MAX && ( ( a64 >> 1 ) + ( b64 >> 1 ) > ( silk_int64_MAX >> 3 ) ) ) ||
+              ( res == silk_int64_MIN && ( ( a64 >> 1 ) + ( b64 >> 1 ) < ( silk_int64_MIN >> 3 ) ) ) ))
+        {
+            fail = 1;
+        }
+    } else {
+        /* Saturation not necessary */
+        fail = res != a64 - b64;
+    }
+    if ( fail )
+    {
+        fprintf (stderr, "silk_SUB_SAT64(%lld, %lld) in %s: line %d\n", (long long)a64, (long long)b64, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return res;
+}
+
+#undef silk_MUL
+#define silk_MUL(a,b) silk_MUL_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_int32 silk_MUL_(opus_int32 a32, opus_int32 b32, char *file, int line){
+    opus_int32 ret;
+    opus_int64 ret64;
+    ret = (opus_int32)((opus_uint32)a32 * (opus_uint32)b32);
+    ret64 = (opus_int64)a32 * (opus_int64)b32;
+    if ( (opus_int64)ret != ret64 )
+    {
+        fprintf (stderr, "silk_MUL(%d, %d) in %s: line %d\n", a32, b32, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;
+}
+
+#undef silk_MUL_uint
+#define silk_MUL_uint(a,b) silk_MUL_uint_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_uint32 silk_MUL_uint_(opus_uint32 a32, opus_uint32 b32, char *file, int line){
+    opus_uint32 ret;
+    ret = a32 * b32;
+    if ( (opus_uint64)ret != (opus_uint64)a32 * (opus_uint64)b32 )
+    {
+        fprintf (stderr, "silk_MUL_uint(%u, %u) in %s: line %d\n", a32, b32, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;
+}
+
+#undef silk_MLA
+#define silk_MLA(a,b,c) silk_MLA_((a), (b), (c), __FILE__, __LINE__)
+static OPUS_INLINE opus_int32 silk_MLA_(opus_int32 a32, opus_int32 b32, opus_int32 c32, char *file, int line){
+    opus_int32 ret;
+    ret = a32 + b32 * c32;
+    if ( (opus_int64)ret != (opus_int64)a32 + (opus_int64)b32 * (opus_int64)c32 )
+    {
+        fprintf (stderr, "silk_MLA(%d, %d, %d) in %s: line %d\n", a32, b32, c32, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;
+}
+
+#undef silk_MLA_uint
+#define silk_MLA_uint(a,b,c) silk_MLA_uint_((a), (b), (c), __FILE__, __LINE__)
+static OPUS_INLINE opus_int32 silk_MLA_uint_(opus_uint32 a32, opus_uint32 b32, opus_uint32 c32, char *file, int line){
+    opus_uint32 ret;
+    ret = a32 + b32 * c32;
+    if ( (opus_int64)ret != (opus_int64)a32 + (opus_int64)b32 * (opus_int64)c32 )
+    {
+        fprintf (stderr, "silk_MLA_uint(%d, %d, %d) in %s: line %d\n", a32, b32, c32, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;
+}
+
+#undef silk_SMULWB
+#define silk_SMULWB(a,b) silk_SMULWB_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_int32 silk_SMULWB_(opus_int32 a32, opus_int32 b32, char *file, int line){
+    opus_int32 ret;
+    ret = (a32 >> 16) * (opus_int32)((opus_int16)b32) + (((a32 & 0x0000FFFF) * (opus_int32)((opus_int16)b32)) >> 16);
+    if ( (opus_int64)ret != ((opus_int64)a32 * (opus_int16)b32) >> 16 )
+    {
+        fprintf (stderr, "silk_SMULWB(%d, %d) in %s: line %d\n", a32, b32, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;
+}
+
+#undef silk_SMLAWB
+#define silk_SMLAWB(a,b,c) silk_SMLAWB_((a), (b), (c), __FILE__, __LINE__)
+static OPUS_INLINE opus_int32 silk_SMLAWB_(opus_int32 a32, opus_int32 b32, opus_int32 c32, char *file, int line){
+    opus_int32 ret;
+    ret = silk_ADD32_ovflw( a32, silk_SMULWB( b32, c32 ) );
+    if ( ret != silk_ADD_SAT32( a32, silk_SMULWB( b32, c32 ) ) )
+    {
+        fprintf (stderr, "silk_SMLAWB(%d, %d, %d) in %s: line %d\n", a32, b32, c32, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;
+}
+
+#undef silk_SMULWT
+#define silk_SMULWT(a,b) silk_SMULWT_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_int32 silk_SMULWT_(opus_int32 a32, opus_int32 b32, char *file, int line){
+    opus_int32 ret;
+    ret = (a32 >> 16) * (b32 >> 16) + (((a32 & 0x0000FFFF) * (b32 >> 16)) >> 16);
+    if ( (opus_int64)ret != ((opus_int64)a32 * (b32 >> 16)) >> 16 )
+    {
+        fprintf (stderr, "silk_SMULWT(%d, %d) in %s: line %d\n", a32, b32, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;
+}
+
+#undef silk_SMLAWT
+#define silk_SMLAWT(a,b,c) silk_SMLAWT_((a), (b), (c), __FILE__, __LINE__)
+static OPUS_INLINE opus_int32 silk_SMLAWT_(opus_int32 a32, opus_int32 b32, opus_int32 c32, char *file, int line){
+    opus_int32 ret;
+    ret = a32 + ((b32 >> 16) * (c32 >> 16)) + (((b32 & 0x0000FFFF) * ((c32 >> 16)) >> 16));
+    if ( (opus_int64)ret != (opus_int64)a32 + (((opus_int64)b32 * (c32 >> 16)) >> 16) )
+    {
+        fprintf (stderr, "silk_SMLAWT(%d, %d, %d) in %s: line %d\n", a32, b32, c32, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;
+}
+
+#undef silk_SMULL
+#define silk_SMULL(a,b) silk_SMULL_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_int64 silk_SMULL_(opus_int64 a64, opus_int64 b64, char *file, int line){
+    opus_int64 ret64;
+    int        fail = 0;
+    ret64 = a64 * b64;
+    if( b64 != 0 ) {
+        fail = a64 != (ret64 / b64);
+    } else if( a64 != 0 ) {
+        fail = b64 != (ret64 / a64);
+    }
+    if ( fail )
+    {
+        fprintf (stderr, "silk_SMULL(%lld, %lld) in %s: line %d\n", (long long)a64, (long long)b64, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret64;
+}
+
+/* no checking needed for silk_SMULBB */
+#undef silk_SMLABB
+#define silk_SMLABB(a,b,c) silk_SMLABB_((a), (b), (c), __FILE__, __LINE__)
+static OPUS_INLINE opus_int32 silk_SMLABB_(opus_int32 a32, opus_int32 b32, opus_int32 c32, char *file, int line){
+    opus_int32 ret;
+    ret = a32 + (opus_int32)((opus_int16)b32) * (opus_int32)((opus_int16)c32);
+    if ( (opus_int64)ret != (opus_int64)a32 + (opus_int64)b32 * (opus_int16)c32 )
+    {
+        fprintf (stderr, "silk_SMLABB(%d, %d, %d) in %s: line %d\n", a32, b32, c32, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;
+}
+
+/* no checking needed for silk_SMULBT */
+#undef silk_SMLABT
+#define silk_SMLABT(a,b,c) silk_SMLABT_((a), (b), (c), __FILE__, __LINE__)
+static OPUS_INLINE opus_int32 silk_SMLABT_(opus_int32 a32, opus_int32 b32, opus_int32 c32, char *file, int line){
+    opus_int32 ret;
+    ret = a32 + ((opus_int32)((opus_int16)b32)) * (c32 >> 16);
+    if ( (opus_int64)ret != (opus_int64)a32 + (opus_int64)b32 * (c32 >> 16) )
+    {
+        fprintf (stderr, "silk_SMLABT(%d, %d, %d) in %s: line %d\n", a32, b32, c32, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;
+}
+
+/* no checking needed for silk_SMULTT */
+#undef silk_SMLATT
+#define silk_SMLATT(a,b,c) silk_SMLATT_((a), (b), (c), __FILE__, __LINE__)
+static OPUS_INLINE opus_int32 silk_SMLATT_(opus_int32 a32, opus_int32 b32, opus_int32 c32, char *file, int line){
+    opus_int32 ret;
+    ret = a32 + (b32 >> 16) * (c32 >> 16);
+    if ( (opus_int64)ret != (opus_int64)a32 + (b32 >> 16) * (c32 >> 16) )
+    {
+        fprintf (stderr, "silk_SMLATT(%d, %d, %d) in %s: line %d\n", a32, b32, c32, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;
+}
+
+#undef silk_SMULWW
+#define silk_SMULWW(a,b) silk_SMULWW_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_int32 silk_SMULWW_(opus_int32 a32, opus_int32 b32, char *file, int line){
+    opus_int32 ret, tmp1, tmp2;
+    opus_int64 ret64;
+    int        fail = 0;
+
+    ret  = silk_SMULWB( a32, b32 );
+    tmp1 = silk_RSHIFT_ROUND( b32, 16 );
+    tmp2 = silk_MUL( a32, tmp1 );
+
+    fail |= (opus_int64)tmp2 != (opus_int64) a32 * (opus_int64) tmp1;
+
+    tmp1 = ret;
+    ret  = silk_ADD32( tmp1, tmp2 );
+    fail |= silk_ADD32( tmp1, tmp2 ) != silk_ADD_SAT32( tmp1, tmp2 );
+
+    ret64 = silk_RSHIFT64( silk_SMULL( a32, b32 ), 16 );
+    fail |= (opus_int64)ret != ret64;
+
+    if ( fail )
+    {
+        fprintf (stderr, "silk_SMULWW(%d, %d) in %s: line %d\n", a32, b32, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+
+    return ret;
+}
+
+#undef silk_SMLAWW
+#define silk_SMLAWW(a,b,c) silk_SMLAWW_((a), (b), (c), __FILE__, __LINE__)
+static OPUS_INLINE opus_int32 silk_SMLAWW_(opus_int32 a32, opus_int32 b32, opus_int32 c32, char *file, int line){
+    opus_int32 ret, tmp;
+
+    tmp = silk_SMULWW( b32, c32 );
+    ret = silk_ADD32( a32, tmp );
+    if ( ret != silk_ADD_SAT32( a32, tmp ) )
+    {
+        fprintf (stderr, "silk_SMLAWW(%d, %d, %d) in %s: line %d\n", a32, b32, c32, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;
+}
+
+/* no checking needed for silk_SMULL
+   no checking needed for silk_SMLAL
+   no checking needed for silk_SMLALBB
+   no checking needed for SigProcFIX_CLZ16
+   no checking needed for SigProcFIX_CLZ32*/
+
+#undef silk_DIV32
+#define silk_DIV32(a,b) silk_DIV32_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_int32 silk_DIV32_(opus_int32 a32, opus_int32 b32, char *file, int line){
+    if ( b32 == 0 )
+    {
+        fprintf (stderr, "silk_DIV32(%d, %d) in %s: line %d\n", a32, b32, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return a32 / b32;
+}
+
+#undef silk_DIV32_16
+#define silk_DIV32_16(a,b) silk_DIV32_16_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_int32 silk_DIV32_16_(opus_int32 a32, opus_int32 b32, char *file, int line){
+    int fail = 0;
+    fail |= b32 == 0;
+    fail |= b32 > silk_int16_MAX;
+    fail |= b32 < silk_int16_MIN;
+    if ( fail )
+    {
+        fprintf (stderr, "silk_DIV32_16(%d, %d) in %s: line %d\n", a32, b32, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return a32 / b32;
+}
+
+/* no checking needed for silk_SAT8
+   no checking needed for silk_SAT16
+   no checking needed for silk_SAT32
+   no checking needed for silk_POS_SAT32
+   no checking needed for silk_ADD_POS_SAT8
+   no checking needed for silk_ADD_POS_SAT16
+   no checking needed for silk_ADD_POS_SAT32 */
+
+#undef silk_LSHIFT8
+#define silk_LSHIFT8(a,b) silk_LSHIFT8_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_int8 silk_LSHIFT8_(opus_int8 a, opus_int32 shift, char *file, int line){
+    opus_int8 ret;
+    int       fail = 0;
+    ret = (opus_int8)((opus_uint8)a << shift);
+    fail |= shift < 0;
+    fail |= shift >= 8;
+    fail |= (opus_int64)ret != (opus_int64)(((opus_uint64)a) << shift);
+    if ( fail )
+    {
+        fprintf (stderr, "silk_LSHIFT8(%d, %d) in %s: line %d\n", a, shift, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;
+}
+
+#undef silk_LSHIFT16
+#define silk_LSHIFT16(a,b) silk_LSHIFT16_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_int16 silk_LSHIFT16_(opus_int16 a, opus_int32 shift, char *file, int line){
+    opus_int16 ret;
+    int        fail = 0;
+    ret = (opus_int16)((opus_uint16)a << shift);
+    fail |= shift < 0;
+    fail |= shift >= 16;
+    fail |= (opus_int64)ret != (opus_int64)(((opus_uint64)a) << shift);
+    if ( fail )
+    {
+        fprintf (stderr, "silk_LSHIFT16(%d, %d) in %s: line %d\n", a, shift, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;
+}
+
+#undef silk_LSHIFT32
+#define silk_LSHIFT32(a,b) silk_LSHIFT32_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_int32 silk_LSHIFT32_(opus_int32 a, opus_int32 shift, char *file, int line){
+    opus_int32 ret;
+    int        fail = 0;
+    ret = (opus_int32)((opus_uint32)a << shift);
+    fail |= shift < 0;
+    fail |= shift >= 32;
+    fail |= (opus_int64)ret != (opus_int64)(((opus_uint64)a) << shift);
+    if ( fail )
+    {
+        fprintf (stderr, "silk_LSHIFT32(%d, %d) in %s: line %d\n", a, shift, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;
+}
+
+#undef silk_LSHIFT64
+#define silk_LSHIFT64(a,b) silk_LSHIFT64_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_int64 silk_LSHIFT64_(opus_int64 a, opus_int shift, char *file, int line){
+    opus_int64 ret;
+    int        fail = 0;
+    ret = (opus_int64)((opus_uint64)a << shift);
+    fail |= shift < 0;
+    fail |= shift >= 64;
+    fail |= (ret>>shift) != ((opus_int64)a);
+    if ( fail )
+    {
+        fprintf (stderr, "silk_LSHIFT64(%lld, %d) in %s: line %d\n", (long long)a, shift, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;
+}
+
+#undef silk_LSHIFT_ovflw
+#define silk_LSHIFT_ovflw(a,b) silk_LSHIFT_ovflw_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_int32 silk_LSHIFT_ovflw_(opus_int32 a, opus_int32 shift, char *file, int line){
+    if ( (shift < 0) || (shift >= 32) ) /* no check for overflow */
+    {
+        fprintf (stderr, "silk_LSHIFT_ovflw(%d, %d) in %s: line %d\n", a, shift, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return a << shift;
+}
+
+#undef silk_LSHIFT_uint
+#define silk_LSHIFT_uint(a,b) silk_LSHIFT_uint_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_uint32 silk_LSHIFT_uint_(opus_uint32 a, opus_int32 shift, char *file, int line){
+    opus_uint32 ret;
+    ret = a << shift;
+    if ( (shift < 0) || ((opus_int64)ret != ((opus_int64)a) << shift))
+    {
+        fprintf (stderr, "silk_LSHIFT_uint(%u, %d) in %s: line %d\n", a, shift, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;
+}
+
+#undef silk_RSHIFT8
+#define silk_RSHITF8(a,b) silk_RSHIFT8_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_int8 silk_RSHIFT8_(opus_int8 a, opus_int32 shift, char *file, int line){
+    if ( (shift < 0) || (shift>=8) )
+    {
+        fprintf (stderr, "silk_RSHITF8(%d, %d) in %s: line %d\n", a, shift, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return a >> shift;
+}
+
+#undef silk_RSHIFT16
+#define silk_RSHITF16(a,b) silk_RSHIFT16_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_int16 silk_RSHIFT16_(opus_int16 a, opus_int32 shift, char *file, int line){
+    if ( (shift < 0) || (shift>=16) )
+    {
+        fprintf (stderr, "silk_RSHITF16(%d, %d) in %s: line %d\n", a, shift, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return a >> shift;
+}
+
+#undef silk_RSHIFT32
+#define silk_RSHIFT32(a,b) silk_RSHIFT32_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_int32 silk_RSHIFT32_(opus_int32 a, opus_int32 shift, char *file, int line){
+    if ( (shift < 0) || (shift>=32) )
+    {
+        fprintf (stderr, "silk_RSHITF32(%d, %d) in %s: line %d\n", a, shift, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return a >> shift;
+}
+
+#undef silk_RSHIFT64
+#define silk_RSHIFT64(a,b) silk_RSHIFT64_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_int64 silk_RSHIFT64_(opus_int64 a, opus_int64 shift, char *file, int line){
+    if ( (shift < 0) || (shift>=64) )
+    {
+        fprintf (stderr, "silk_RSHITF64(%lld, %lld) in %s: line %d\n", (long long)a, (long long)shift, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return a >> shift;
+}
+
+#undef silk_RSHIFT_uint
+#define silk_RSHIFT_uint(a,b) silk_RSHIFT_uint_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_uint32 silk_RSHIFT_uint_(opus_uint32 a, opus_int32 shift, char *file, int line){
+    if ( (shift < 0) || (shift>32) )
+    {
+        fprintf (stderr, "silk_RSHIFT_uint(%u, %d) in %s: line %d\n", a, shift, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return a >> shift;
+}
+
+#undef silk_ADD_LSHIFT
+#define silk_ADD_LSHIFT(a,b,c) silk_ADD_LSHIFT_((a), (b), (c), __FILE__, __LINE__)
+static OPUS_INLINE int silk_ADD_LSHIFT_(int a, int b, int shift, char *file, int line){
+    opus_int16 ret;
+    ret = a + (opus_int16)((opus_uint16)b << shift);
+    if ( (shift < 0) || (shift>15) || ((opus_int64)ret != (opus_int64)a + (opus_int64)(((opus_uint64)b) << shift)) )
+    {
+        fprintf (stderr, "silk_ADD_LSHIFT(%d, %d, %d) in %s: line %d\n", a, b, shift, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;                /* shift >= 0 */
+}
+
+#undef silk_ADD_LSHIFT32
+#define silk_ADD_LSHIFT32(a,b,c) silk_ADD_LSHIFT32_((a), (b), (c), __FILE__, __LINE__)
+static OPUS_INLINE opus_int32 silk_ADD_LSHIFT32_(opus_int32 a, opus_int32 b, opus_int32 shift, char *file, int line){
+    opus_int32 ret;
+    ret = silk_ADD32_ovflw(a, (opus_int32)((opus_uint32)b << shift));
+    if ( (shift < 0) || (shift>31) || ((opus_int64)ret != (opus_int64)a + (opus_int64)(((opus_uint64)b) << shift)) )
+    {
+        fprintf (stderr, "silk_ADD_LSHIFT32(%d, %d, %d) in %s: line %d\n", a, b, shift, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;                /* shift >= 0 */
+}
+
+#undef silk_ADD_LSHIFT_uint
+#define silk_ADD_LSHIFT_uint(a,b,c) silk_ADD_LSHIFT_uint_((a), (b), (c), __FILE__, __LINE__)
+static OPUS_INLINE opus_uint32 silk_ADD_LSHIFT_uint_(opus_uint32 a, opus_uint32 b, opus_int32 shift, char *file, int line){
+    opus_uint32 ret;
+    ret = a + (b << shift);
+    if ( (shift < 0) || (shift>32) || ((opus_int64)ret != (opus_int64)a + (((opus_int64)b) << shift)) )
+    {
+        fprintf (stderr, "silk_ADD_LSHIFT_uint(%u, %u, %d) in %s: line %d\n", a, b, shift, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;                /* shift >= 0 */
+}
+
+#undef silk_ADD_RSHIFT
+#define silk_ADD_RSHIFT(a,b,c) silk_ADD_RSHIFT_((a), (b), (c), __FILE__, __LINE__)
+static OPUS_INLINE int silk_ADD_RSHIFT_(int a, int b, int shift, char *file, int line){
+    opus_int16 ret;
+    ret = a + (b >> shift);
+    if ( (shift < 0) || (shift>15) || ((opus_int64)ret != (opus_int64)a + (((opus_int64)b) >> shift)) )
+    {
+        fprintf (stderr, "silk_ADD_RSHIFT(%d, %d, %d) in %s: line %d\n", a, b, shift, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;                /* shift  > 0 */
+}
+
+#undef silk_ADD_RSHIFT32
+#define silk_ADD_RSHIFT32(a,b,c) silk_ADD_RSHIFT32_((a), (b), (c), __FILE__, __LINE__)
+static OPUS_INLINE opus_int32 silk_ADD_RSHIFT32_(opus_int32 a, opus_int32 b, opus_int32 shift, char *file, int line){
+    opus_int32 ret;
+    ret = silk_ADD32_ovflw(a, (b >> shift));
+    if ( (shift < 0) || (shift>31) || ((opus_int64)ret != (opus_int64)a + (((opus_int64)b) >> shift)) )
+    {
+        fprintf (stderr, "silk_ADD_RSHIFT32(%d, %d, %d) in %s: line %d\n", a, b, shift, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;                /* shift  > 0 */
+}
+
+#undef silk_ADD_RSHIFT_uint
+#define silk_ADD_RSHIFT_uint(a,b,c) silk_ADD_RSHIFT_uint_((a), (b), (c), __FILE__, __LINE__)
+static OPUS_INLINE opus_uint32 silk_ADD_RSHIFT_uint_(opus_uint32 a, opus_uint32 b, opus_int32 shift, char *file, int line){
+    opus_uint32 ret;
+    ret = a + (b >> shift);
+    if ( (shift < 0) || (shift>32) || ((opus_int64)ret != (opus_int64)a + (((opus_int64)b) >> shift)) )
+    {
+        fprintf (stderr, "silk_ADD_RSHIFT_uint(%u, %u, %d) in %s: line %d\n", a, b, shift, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;                /* shift  > 0 */
+}
+
+#undef silk_SUB_LSHIFT32
+#define silk_SUB_LSHIFT32(a,b,c) silk_SUB_LSHIFT32_((a), (b), (c), __FILE__, __LINE__)
+static OPUS_INLINE opus_int32 silk_SUB_LSHIFT32_(opus_int32 a, opus_int32 b, opus_int32 shift, char *file, int line){
+    opus_int32 ret;
+    ret = silk_SUB32_ovflw(a, (opus_int32)((opus_uint32)b << shift));
+    if ( (shift < 0) || (shift>31) || ((opus_int64)ret != (opus_int64)a - (opus_int64)(((opus_uint64)b) << shift)) )
+    {
+        fprintf (stderr, "silk_SUB_LSHIFT32(%d, %d, %d) in %s: line %d\n", a, b, shift, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;                /* shift >= 0 */
+}
+
+#undef silk_SUB_RSHIFT32
+#define silk_SUB_RSHIFT32(a,b,c) silk_SUB_RSHIFT32_((a), (b), (c), __FILE__, __LINE__)
+static OPUS_INLINE opus_int32 silk_SUB_RSHIFT32_(opus_int32 a, opus_int32 b, opus_int32 shift, char *file, int line){
+    opus_int32 ret;
+    ret = silk_SUB32_ovflw(a, (b >> shift));
+    if ( (shift < 0) || (shift>31) || ((opus_int64)ret != (opus_int64)a - (((opus_int64)b) >> shift)) )
+    {
+        fprintf (stderr, "silk_SUB_RSHIFT32(%d, %d, %d) in %s: line %d\n", a, b, shift, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;                /* shift  > 0 */
+}
+
+#undef silk_RSHIFT_ROUND
+#define silk_RSHIFT_ROUND(a,b) silk_RSHIFT_ROUND_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_int32 silk_RSHIFT_ROUND_(opus_int32 a, opus_int32 shift, char *file, int line){
+    opus_int32 ret;
+    ret = shift == 1 ? (a >> 1) + (a & 1) : ((a >> (shift - 1)) + 1) >> 1;
+    /* the macro definition can't handle a shift of zero */
+    if ( (shift <= 0) || (shift>31) || ((opus_int64)ret != ((opus_int64)a + ((opus_int64)1 << (shift - 1))) >> shift) )
+    {
+        fprintf (stderr, "silk_RSHIFT_ROUND(%d, %d) in %s: line %d\n", a, shift, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return ret;
+}
+
+#undef silk_RSHIFT_ROUND64
+#define silk_RSHIFT_ROUND64(a,b) silk_RSHIFT_ROUND64_((a), (b), __FILE__, __LINE__)
+static OPUS_INLINE opus_int64 silk_RSHIFT_ROUND64_(opus_int64 a, opus_int32 shift, char *file, int line){
+    opus_int64 ret;
+    /* the macro definition can't handle a shift of zero */
+    if ( (shift <= 0) || (shift>=64) )
+    {
+        fprintf (stderr, "silk_RSHIFT_ROUND64(%lld, %d) in %s: line %d\n", (long long)a, shift, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    ret = shift == 1 ? (a >> 1) + (a & 1) : ((a >> (shift - 1)) + 1) >> 1;
+    return ret;
+}
+
+/* silk_abs is used on floats also, so doesn't work... */
+/*#undef silk_abs
+static OPUS_INLINE opus_int32 silk_abs(opus_int32 a){
+    silk_assert(a != 0x80000000);
+    return (((a) >  0)  ? (a) : -(a));            // Be careful, silk_abs returns wrong when input equals to silk_intXX_MIN
+}*/
+
+#undef silk_abs_int64
+#define silk_abs_int64(a) silk_abs_int64_((a), __FILE__, __LINE__)
+static OPUS_INLINE opus_int64 silk_abs_int64_(opus_int64 a, char *file, int line){
+    if ( a == silk_int64_MIN )
+    {
+        fprintf (stderr, "silk_abs_int64(%lld) in %s: line %d\n", (long long)a, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return (((a) >  0)  ? (a) : -(a));            /* Be careful, silk_abs returns wrong when input equals to silk_intXX_MIN */
+}
+
+#undef silk_abs_int32
+#define silk_abs_int32(a) silk_abs_int32_((a), __FILE__, __LINE__)
+static OPUS_INLINE opus_int32 silk_abs_int32_(opus_int32 a, char *file, int line){
+    if ( a == silk_int32_MIN )
+    {
+        fprintf (stderr, "silk_abs_int32(%d) in %s: line %d\n", a, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return silk_abs(a);
+}
+
+#undef silk_CHECK_FIT8
+#define silk_CHECK_FIT8(a) silk_CHECK_FIT8_((a), __FILE__, __LINE__)
+static OPUS_INLINE opus_int8 silk_CHECK_FIT8_( opus_int64 a, char *file, int line ){
+    opus_int8 ret;
+    ret = (opus_int8)a;
+    if ( (opus_int64)ret != a )
+    {
+        fprintf (stderr, "silk_CHECK_FIT8(%lld) in %s: line %d\n", (long long)a, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return( ret );
+}
+
+#undef silk_CHECK_FIT16
+#define silk_CHECK_FIT16(a) silk_CHECK_FIT16_((a), __FILE__, __LINE__)
+static OPUS_INLINE opus_int16 silk_CHECK_FIT16_( opus_int64 a, char *file, int line ){
+    opus_int16 ret;
+    ret = (opus_int16)a;
+    if ( (opus_int64)ret != a )
+    {
+        fprintf (stderr, "silk_CHECK_FIT16(%lld) in %s: line %d\n", (long long)a, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return( ret );
+}
+
+#undef silk_CHECK_FIT32
+#define silk_CHECK_FIT32(a) silk_CHECK_FIT32_((a), __FILE__, __LINE__)
+static OPUS_INLINE opus_int32 silk_CHECK_FIT32_( opus_int64 a, char *file, int line ){
+    opus_int32 ret;
+    ret = (opus_int32)a;
+    if ( (opus_int64)ret != a )
+    {
+        fprintf (stderr, "silk_CHECK_FIT32(%lld) in %s: line %d\n", (long long)a, file, line);
+#ifdef FIXED_DEBUG_ASSERT
+        silk_assert( 0 );
+#endif
+    }
+    return( ret );
+}
+
+/* no checking for silk_NSHIFT_MUL_32_32
+   no checking for silk_NSHIFT_MUL_16_16
+   no checking needed for silk_min
+   no checking needed for silk_max
+   no checking needed for silk_sign
+*/
+
+#endif
+#endif /* MACRO_DEBUG_H */
--- a/lib/libopus/silk/NLSF2A.c
+++ b/lib/libopus/silk/NLSF2A.c
@ -0,0 +1,140 @@
+/***********************************************************************
+Copyright (c) 2006-2011, Skype Limited. All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+- Redistributions of source code must retain the above copyright notice,
+this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+***********************************************************************/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+/* conversion between prediction filter coefficients and LSFs   */
+/* order should be even                                         */
+/* a piecewise linear approximation maps LSF <-> cos(LSF)       */
+/* therefore the result is not accurate LSFs, but the two       */
+/* functions are accurate inverses of each other                */
+
+#include "SigProc_FIX.h"
+#include "tables.h"
+
+#define QA      16
+
+/* helper function for NLSF2A(..) */
+static OPUS_INLINE void silk_NLSF2A_find_poly(
+    opus_int32          *out,      /* O    intermediate polynomial, QA [dd+1]        */
+    const opus_int32    *cLSF,     /* I    vector of interleaved 2*cos(LSFs), QA [d] */
+    opus_int            dd         /* I    polynomial order (= 1/2 * filter order)   */
+)
+{
+    opus_int   k, n;
+    opus_int32 ftmp;
+
+    out[0] = silk_LSHIFT( 1, QA );
+    out[1] = -cLSF[0];
+    for( k = 1; k < dd; k++ ) {
+        ftmp = cLSF[2*k];            /* QA*/
+        out[k+1] = silk_LSHIFT( out[k-1], 1 ) - (opus_int32)silk_RSHIFT_ROUND64( silk_SMULL( ftmp, out[k] ), QA );
+        for( n = k; n > 1; n-- ) {
+            out[n] += out[n-2] - (opus_int32)silk_RSHIFT_ROUND64( silk_SMULL( ftmp, out[n-1] ), QA );
+        }
+        out[1] -= ftmp;
+    }
+}
+
+/* compute whitening filter coefficients from normalized line spectral frequencies */
+void silk_NLSF2A(
+    opus_int16                  *a_Q12,             /* O    monic whitening filter coefficients in Q12,  [ d ]          */
+    const opus_int16            *NLSF,              /* I    normalized line spectral frequencies in Q15, [ d ]          */
+    const opus_int              d,                  /* I    filter order (should be even)                               */
+    int                         arch                /* I    Run-time architecture                                       */
+)
+{
+    /* This ordering was found to maximize quality. It improves numerical accuracy of
+       silk_NLSF2A_find_poly() compared to "standard" ordering. */
+    static const unsigned char ordering16[16] = {
+      0, 15, 8, 7, 4, 11, 12, 3, 2, 13, 10, 5, 6, 9, 14, 1
+    };
+    static const unsigned char ordering10[10] = {
+      0, 9, 6, 3, 4, 5, 8, 1, 2, 7
+    };
+    const unsigned char *ordering;
+    opus_int   k, i, dd;
+    opus_int32 cos_LSF_QA[ SILK_MAX_ORDER_LPC ];
+    opus_int32 P[ SILK_MAX_ORDER_LPC / 2 + 1 ], Q[ SILK_MAX_ORDER_LPC / 2 + 1 ];
+    opus_int32 Ptmp, Qtmp, f_int, f_frac, cos_val, delta;
+    opus_int32 a32_QA1[ SILK_MAX_ORDER_LPC ];
+
+    silk_assert( LSF_COS_TAB_SZ_FIX == 128 );
+    celt_assert( d==10 || d==16 );
+
+    /* convert LSFs to 2*cos(LSF), using piecewise linear curve from table */
+    ordering = d == 16 ? ordering16 : ordering10;
+    for( k = 0; k < d; k++ ) {
+        silk_assert( NLSF[k] >= 0 );
+
+        /* f_int on a scale 0-127 (rounded down) */
+        f_int = silk_RSHIFT( NLSF[k], 15 - 7 );
+
+        /* f_frac, range: 0..255 */
+        f_frac = NLSF[k] - silk_LSHIFT( f_int, 15 - 7 );
+
+        silk_assert(f_int >= 0);
+        silk_assert(f_int < LSF_COS_TAB_SZ_FIX );
+
+        /* Read start and end value from table */
+        cos_val = silk_LSFCosTab_FIX_Q12[ f_int ];                /* Q12 */
+        delta   = silk_LSFCosTab_FIX_Q12[ f_int + 1 ] - cos_val;  /* Q12, with a range of 0..200 */
+
+        /* Linear interpolation */
+        cos_LSF_QA[ordering[k]] = silk_RSHIFT_ROUND( silk_LSHIFT( cos_val, 8 ) + silk_MUL( delta, f_frac ), 20 - QA ); /* QA */
+    }
+
+    dd = silk_RSHIFT( d, 1 );
+
+    /* generate even and odd polynomials using convolution */
+    silk_NLSF2A_find_poly( P, &cos_LSF_QA[ 0 ], dd );
+    silk_NLSF2A_find_poly( Q, &cos_LSF_QA[ 1 ], dd );
+
+    /* convert even and odd polynomials to opus_int32 Q12 filter coefs */
+    for( k = 0; k < dd; k++ ) {
+        Ptmp = P[ k+1 ] + P[ k ];
+        Qtmp = Q[ k+1 ] - Q[ k ];
+
+        /* the Ptmp and Qtmp values at this stage need to fit in int32 */
+        a32_QA1[ k ]     = -Qtmp - Ptmp;        /* QA+1 */
+        a32_QA1[ d-k-1 ] =  Qtmp - Ptmp;        /* QA+1 */
+    }
+
+    /* Convert int32 coefficients to Q12 int16 coefs */
+    silk_LPC_fit( a_Q12, a32_QA1, 12, QA + 1, d );
+
+    for( i = 0; silk_LPC_inverse_pred_gain( a_Q12, d, arch ) == 0 && i < MAX_LPC_STABILIZE_ITERATIONS; i++ ) {
+        /* Prediction coefficients are (too close to) unstable; apply bandwidth expansion   */
+        /* on the unscaled coefficients, convert to Q12 and measure again                   */
+        silk_bwexpander_32( a32_QA1, d, 65536 - silk_LSHIFT( 2, i ) );
+        for( k = 0; k < d; k++ ) {
+            a_Q12[ k ] = (opus_int16)silk_RSHIFT_ROUND( a32_QA1[ k ], QA + 1 - 12 );            /* QA+1 -> Q12 */
+        }
+    }
+}
--- a/lib/libopus/silk/NLSF_VQ.c
+++ b/lib/libopus/silk/NLSF_VQ.c
@ -0,0 +1,76 @@
+/***********************************************************************
+Copyright (c) 2006-2011, Skype Limited. All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+- Redistributions of source code must retain the above copyright notice,
+this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+***********************************************************************/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "main.h"
+
+/* Compute quantization errors for an LPC_order element input vector for a VQ codebook */
+void silk_NLSF_VQ(
+    opus_int32                  err_Q24[],                      /* O    Quantization errors [K]                     */
+    const opus_int16            in_Q15[],                       /* I    Input vectors to be quantized [LPC_order]   */
+    const opus_uint8            pCB_Q8[],                       /* I    Codebook vectors [K*LPC_order]              */
+    const opus_int16            pWght_Q9[],                     /* I    Codebook weights [K*LPC_order]              */
+    const opus_int              K,                              /* I    Number of codebook vectors                  */
+    const opus_int              LPC_order                       /* I    Number of LPCs                              */
+)
+{
+    opus_int         i, m;
+    opus_int32       diff_Q15, diffw_Q24, sum_error_Q24, pred_Q24;
+    const opus_int16 *w_Q9_ptr;
+    const opus_uint8 *cb_Q8_ptr;
+
+    celt_assert( ( LPC_order & 1 ) == 0 );
+
+    /* Loop over codebook */
+    cb_Q8_ptr = pCB_Q8;
+    w_Q9_ptr = pWght_Q9;
+    for( i = 0; i < K; i++ ) {
+        sum_error_Q24 = 0;
+        pred_Q24 = 0;
+        for( m = LPC_order-2; m >= 0; m -= 2 ) {
+            /* Compute weighted absolute predictive quantization error for index m + 1 */
+            diff_Q15 = silk_SUB_LSHIFT32( in_Q15[ m + 1 ], (opus_int32)cb_Q8_ptr[ m + 1 ], 7 ); /* range: [ -32767 : 32767 ]*/
+            diffw_Q24 = silk_SMULBB( diff_Q15, w_Q9_ptr[ m + 1 ] );
+            sum_error_Q24 = silk_ADD32( sum_error_Q24, silk_abs( silk_SUB_RSHIFT32( diffw_Q24, pred_Q24, 1 ) ) );
+            pred_Q24 = diffw_Q24;
+
+            /* Compute weighted absolute predictive quantization error for index m */
+            diff_Q15 = silk_SUB_LSHIFT32( in_Q15[ m ], (opus_int32)cb_Q8_ptr[ m ], 7 ); /* range: [ -32767 : 32767 ]*/
+            diffw_Q24 = silk_SMULBB( diff_Q15, w_Q9_ptr[ m ] );
+            sum_error_Q24 = silk_ADD32( sum_error_Q24, silk_abs( silk_SUB_RSHIFT32( diffw_Q24, pred_Q24, 1 ) ) );
+            pred_Q24 = diffw_Q24;
+
+            silk_assert( sum_error_Q24 >= 0 );
+        }
+        err_Q24[ i ] = sum_error_Q24;
+        cb_Q8_ptr += LPC_order;
+        w_Q9_ptr += LPC_order;
+    }
+}
--- a/lib/libopus/silk/NLSF_VQ_weights_laroia.c
+++ b/lib/libopus/silk/NLSF_VQ_weights_laroia.c
@ -0,0 +1,80 @@
+/***********************************************************************
+Copyright (c) 2006-2011, Skype Limited. All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+- Redistributions of source code must retain the above copyright notice,
+this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+***********************************************************************/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "define.h"
+#include "SigProc_FIX.h"
+
+/*
+R. Laroia, N. Phamdo and N. Farvardin, "Robust and Efficient Quantization of Speech LSP
+Parameters Using Structured Vector Quantization", Proc. IEEE Int. Conf. Acoust., Speech,
+Signal Processing, pp. 641-644, 1991.
+*/
+
+/* Laroia low complexity NLSF weights */
+void silk_NLSF_VQ_weights_laroia(
+    opus_int16                  *pNLSFW_Q_OUT,      /* O     Pointer to input vector weights [D]                        */
+    const opus_int16            *pNLSF_Q15,         /* I     Pointer to input vector         [D]                        */
+    const opus_int              D                   /* I     Input vector dimension (even)                              */
+)
+{
+    opus_int   k;
+    opus_int32 tmp1_int, tmp2_int;
+
+    celt_assert( D > 0 );
+    celt_assert( ( D & 1 ) == 0 );
+
+    /* First value */
+    tmp1_int = silk_max_int( pNLSF_Q15[ 0 ], 1 );
+    tmp1_int = silk_DIV32_16( (opus_int32)1 << ( 15 + NLSF_W_Q ), tmp1_int );
+    tmp2_int = silk_max_int( pNLSF_Q15[ 1 ] - pNLSF_Q15[ 0 ], 1 );
+    tmp2_int = silk_DIV32_16( (opus_int32)1 << ( 15 + NLSF_W_Q ), tmp2_int );
+    pNLSFW_Q_OUT[ 0 ] = (opus_int16)silk_min_int( tmp1_int + tmp2_int, silk_int16_MAX );
+    silk_assert( pNLSFW_Q_OUT[ 0 ] > 0 );
+
+    /* Main loop */
+    for( k = 1; k < D - 1; k += 2 ) {
+        tmp1_int = silk_max_int( pNLSF_Q15[ k + 1 ] - pNLSF_Q15[ k ], 1 );
+        tmp1_int = silk_DIV32_16( (opus_int32)1 << ( 15 + NLSF_W_Q ), tmp1_int );
+        pNLSFW_Q_OUT[ k ] = (opus_int16)silk_min_int( tmp1_int + tmp2_int, silk_int16_MAX );
+        silk_assert( pNLSFW_Q_OUT[ k ] > 0 );
+
+        tmp2_int = silk_max_int( pNLSF_Q15[ k + 2 ] - pNLSF_Q15[ k + 1 ], 1 );
+        tmp2_int = silk_DIV32_16( (opus_int32)1 << ( 15 + NLSF_W_Q ), tmp2_int );
+        pNLSFW_Q_OUT[ k + 1 ] = (opus_int16)silk_min_int( tmp1_int + tmp2_int, silk_int16_MAX );
+        silk_assert( pNLSFW_Q_OUT[ k + 1 ] > 0 );
+    }
+
+    /* Last value */
+    tmp1_int = silk_max_int( ( 1 << 15 ) - pNLSF_Q15[ D - 1 ], 1 );
+    tmp1_int = silk_DIV32_16( (opus_int32)1 << ( 15 + NLSF_W_Q ), tmp1_int );
+    pNLSFW_Q_OUT[ D - 1 ] = (opus_int16)silk_min_int( tmp1_int + tmp2_int, silk_int16_MAX );
+    silk_assert( pNLSFW_Q_OUT[ D - 1 ] > 0 );
+}
--- a/lib/libopus/silk/NLSF_decode.c
+++ b/lib/libopus/silk/NLSF_decode.c
@ -0,0 +1,93 @@
+/***********************************************************************
+Copyright (c) 2006-2011, Skype Limited. All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+- Redistributions of source code must retain the above copyright notice,
+this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+***********************************************************************/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "main.h"
+
+/* Predictive dequantizer for NLSF residuals */
+static OPUS_INLINE void silk_NLSF_residual_dequant(          /* O    Returns RD value in Q30                     */
+          opus_int16         x_Q10[],                        /* O    Output [ order ]                            */
+    const opus_int8          indices[],                      /* I    Quantization indices [ order ]              */
+    const opus_uint8         pred_coef_Q8[],                 /* I    Backward predictor coefs [ order ]          */
+    const opus_int           quant_step_size_Q16,            /* I    Quantization step size                      */
+    const opus_int16         order                           /* I    Number of input values                      */
+)
+{
+    opus_int     i, out_Q10, pred_Q10;
+
+    out_Q10 = 0;
+    for( i = order-1; i >= 0; i-- ) {
+        pred_Q10 = silk_RSHIFT( silk_SMULBB( out_Q10, (opus_int16)pred_coef_Q8[ i ] ), 8 );
+        out_Q10  = silk_LSHIFT( indices[ i ], 10 );
+        if( out_Q10 > 0 ) {
+            out_Q10 = silk_SUB16( out_Q10, SILK_FIX_CONST( NLSF_QUANT_LEVEL_ADJ, 10 ) );
+        } else if( out_Q10 < 0 ) {
+            out_Q10 = silk_ADD16( out_Q10, SILK_FIX_CONST( NLSF_QUANT_LEVEL_ADJ, 10 ) );
+        }
+        out_Q10  = silk_SMLAWB( pred_Q10, (opus_int32)out_Q10, quant_step_size_Q16 );
+        x_Q10[ i ] = out_Q10;
+    }
+}
+
+
+/***********************/
+/* NLSF vector decoder */
+/***********************/
+void silk_NLSF_decode(
+          opus_int16            *pNLSF_Q15,                     /* O    Quantized NLSF vector [ LPC_ORDER ]         */
+          opus_int8             *NLSFIndices,                   /* I    Codebook path vector [ LPC_ORDER + 1 ]      */
+    const silk_NLSF_CB_struct   *psNLSF_CB                      /* I    Codebook object                             */
+)
+{
+    opus_int         i;
+    opus_uint8       pred_Q8[  MAX_LPC_ORDER ];
+    opus_int16       ec_ix[    MAX_LPC_ORDER ];
+    opus_int16       res_Q10[  MAX_LPC_ORDER ];
+    opus_int32       NLSF_Q15_tmp;
+    const opus_uint8 *pCB_element;
+    const opus_int16 *pCB_Wght_Q9;
+
+    /* Unpack entropy table indices and predictor for current CB1 index */
+    silk_NLSF_unpack( ec_ix, pred_Q8, psNLSF_CB, NLSFIndices[ 0 ] );
+
+    /* Predictive residual dequantizer */
+    silk_NLSF_residual_dequant( res_Q10, &NLSFIndices[ 1 ], pred_Q8, psNLSF_CB->quantStepSize_Q16, psNLSF_CB->order );
+
+    /* Apply inverse square-rooted weights to first stage and add to output */
+    pCB_element = &psNLSF_CB->CB1_NLSF_Q8[ NLSFIndices[ 0 ] * psNLSF_CB->order ];
+    pCB_Wght_Q9 = &psNLSF_CB->CB1_Wght_Q9[ NLSFIndices[ 0 ] * psNLSF_CB->order ];
+    for( i = 0; i < psNLSF_CB->order; i++ ) {
+        NLSF_Q15_tmp = silk_ADD_LSHIFT32( silk_DIV32_16( silk_LSHIFT( (opus_int32)res_Q10[ i ], 14 ), pCB_Wght_Q9[ i ] ), (opus_int16)pCB_element[ i ], 7 );
+        pNLSF_Q15[ i ] = (opus_int16)silk_LIMIT( NLSF_Q15_tmp, 0, 32767 );
+    }
+
+    /* NLSF stabilization */
+    silk_NLSF_stabilize( pNLSF_Q15, psNLSF_CB->deltaMin_Q15, psNLSF_CB->order );
+}
--- a/lib/libopus/silk/NLSF_del_dec_quant.c
+++ b/lib/libopus/silk/NLSF_del_dec_quant.c
@ -0,0 +1,215 @@
+/***********************************************************************
+Copyright (c) 2006-2011, Skype Limited. All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+- Redistributions of source code must retain the above copyright notice,
+this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+***********************************************************************/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "main.h"
+
+/* Delayed-decision quantizer for NLSF residuals */
+opus_int32 silk_NLSF_del_dec_quant(                             /* O    Returns RD value in Q25                     */
+    opus_int8                   indices[],                      /* O    Quantization indices [ order ]              */
+    const opus_int16            x_Q10[],                        /* I    Input [ order ]                             */
+    const opus_int16            w_Q5[],                         /* I    Weights [ order ]                           */
+    const opus_uint8            pred_coef_Q8[],                 /* I    Backward predictor coefs [ order ]          */
+    const opus_int16            ec_ix[],                        /* I    Indices to entropy coding tables [ order ]  */
+    const opus_uint8            ec_rates_Q5[],                  /* I    Rates []                                    */
+    const opus_int              quant_step_size_Q16,            /* I    Quantization step size                      */
+    const opus_int16            inv_quant_step_size_Q6,         /* I    Inverse quantization step size              */
+    const opus_int32            mu_Q20,                         /* I    R/D tradeoff                                */
+    const opus_int16            order                           /* I    Number of input values                      */
+)
+{
+    opus_int         i, j, nStates, ind_tmp, ind_min_max, ind_max_min, in_Q10, res_Q10;
+    opus_int         pred_Q10, diff_Q10, rate0_Q5, rate1_Q5;
+    opus_int16       out0_Q10, out1_Q10;
+    opus_int32       RD_tmp_Q25, min_Q25, min_max_Q25, max_min_Q25;
+    opus_int         ind_sort[         NLSF_QUANT_DEL_DEC_STATES ];
+    opus_int8        ind[              NLSF_QUANT_DEL_DEC_STATES ][ MAX_LPC_ORDER ];
+    opus_int16       prev_out_Q10[ 2 * NLSF_QUANT_DEL_DEC_STATES ];
+    opus_int32       RD_Q25[       2 * NLSF_QUANT_DEL_DEC_STATES ];
+    opus_int32       RD_min_Q25[       NLSF_QUANT_DEL_DEC_STATES ];
+    opus_int32       RD_max_Q25[       NLSF_QUANT_DEL_DEC_STATES ];
+    const opus_uint8 *rates_Q5;
+
+    opus_int out0_Q10_table[2 * NLSF_QUANT_MAX_AMPLITUDE_EXT];
+    opus_int out1_Q10_table[2 * NLSF_QUANT_MAX_AMPLITUDE_EXT];
+
+    for (i = -NLSF_QUANT_MAX_AMPLITUDE_EXT; i <= NLSF_QUANT_MAX_AMPLITUDE_EXT-1; i++)
+    {
+        out0_Q10 = silk_LSHIFT( i, 10 );
+        out1_Q10 = silk_ADD16( out0_Q10, 1024 );
+        if( i > 0 ) {
+            out0_Q10 = silk_SUB16( out0_Q10, SILK_FIX_CONST( NLSF_QUANT_LEVEL_ADJ, 10 ) );
+            out1_Q10 = silk_SUB16( out1_Q10, SILK_FIX_CONST( NLSF_QUANT_LEVEL_ADJ, 10 ) );
+        } else if( i == 0 ) {
+            out1_Q10 = silk_SUB16( out1_Q10, SILK_FIX_CONST( NLSF_QUANT_LEVEL_ADJ, 10 ) );
+        } else if( i == -1 ) {
+            out0_Q10 = silk_ADD16( out0_Q10, SILK_FIX_CONST( NLSF_QUANT_LEVEL_ADJ, 10 ) );
+        } else {
+            out0_Q10 = silk_ADD16( out0_Q10, SILK_FIX_CONST( NLSF_QUANT_LEVEL_ADJ, 10 ) );
+            out1_Q10 = silk_ADD16( out1_Q10, SILK_FIX_CONST( NLSF_QUANT_LEVEL_ADJ, 10 ) );
+        }
+        out0_Q10_table[ i + NLSF_QUANT_MAX_AMPLITUDE_EXT ] = silk_RSHIFT( silk_SMULBB( out0_Q10, quant_step_size_Q16 ), 16 );
+        out1_Q10_table[ i + NLSF_QUANT_MAX_AMPLITUDE_EXT ] = silk_RSHIFT( silk_SMULBB( out1_Q10, quant_step_size_Q16 ), 16 );
+    }
+
+    silk_assert( (NLSF_QUANT_DEL_DEC_STATES & (NLSF_QUANT_DEL_DEC_STATES-1)) == 0 );     /* must be power of two */
+
+    nStates = 1;
+    RD_Q25[ 0 ] = 0;
+    prev_out_Q10[ 0 ] = 0;
+    for( i = order - 1; i >= 0; i-- ) {
+        rates_Q5 = &ec_rates_Q5[ ec_ix[ i ] ];
+        in_Q10 = x_Q10[ i ];
+        for( j = 0; j < nStates; j++ ) {
+            pred_Q10 = silk_RSHIFT( silk_SMULBB( (opus_int16)pred_coef_Q8[ i ], prev_out_Q10[ j ] ), 8 );
+            res_Q10  = silk_SUB16( in_Q10, pred_Q10 );
+            ind_tmp  = silk_RSHIFT( silk_SMULBB( inv_quant_step_size_Q6, res_Q10 ), 16 );
+            ind_tmp  = silk_LIMIT( ind_tmp, -NLSF_QUANT_MAX_AMPLITUDE_EXT, NLSF_QUANT_MAX_AMPLITUDE_EXT-1 );
+            ind[ j ][ i ] = (opus_int8)ind_tmp;
+
+            /* compute outputs for ind_tmp and ind_tmp + 1 */
+            out0_Q10 = out0_Q10_table[ ind_tmp + NLSF_QUANT_MAX_AMPLITUDE_EXT ];
+            out1_Q10 = out1_Q10_table[ ind_tmp + NLSF_QUANT_MAX_AMPLITUDE_EXT ];
+
+            out0_Q10  = silk_ADD16( out0_Q10, pred_Q10 );
+            out1_Q10  = silk_ADD16( out1_Q10, pred_Q10 );
+            prev_out_Q10[ j           ] = out0_Q10;
+            prev_out_Q10[ j + nStates ] = out1_Q10;
+
+            /* compute RD for ind_tmp and ind_tmp + 1 */
+            if( ind_tmp + 1 >= NLSF_QUANT_MAX_AMPLITUDE ) {
+                if( ind_tmp + 1 == NLSF_QUANT_MAX_AMPLITUDE ) {
+                    rate0_Q5 = rates_Q5[ ind_tmp + NLSF_QUANT_MAX_AMPLITUDE ];
+                    rate1_Q5 = 280;
+                } else {
+                    rate0_Q5 = silk_SMLABB( 280 - 43 * NLSF_QUANT_MAX_AMPLITUDE, 43, ind_tmp );
+                    rate1_Q5 = silk_ADD16( rate0_Q5, 43 );
+                }
+            } else if( ind_tmp <= -NLSF_QUANT_MAX_AMPLITUDE ) {
+                if( ind_tmp == -NLSF_QUANT_MAX_AMPLITUDE ) {
+                    rate0_Q5 = 280;
+                    rate1_Q5 = rates_Q5[ ind_tmp + 1 + NLSF_QUANT_MAX_AMPLITUDE ];
+                } else {
+                    rate0_Q5 = silk_SMLABB( 280 - 43 * NLSF_QUANT_MAX_AMPLITUDE, -43, ind_tmp );
+                    rate1_Q5 = silk_SUB16( rate0_Q5, 43 );
+                }
+            } else {
+                rate0_Q5 = rates_Q5[ ind_tmp +     NLSF_QUANT_MAX_AMPLITUDE ];
+                rate1_Q5 = rates_Q5[ ind_tmp + 1 + NLSF_QUANT_MAX_AMPLITUDE ];
+            }
+            RD_tmp_Q25            = RD_Q25[ j ];
+            diff_Q10              = silk_SUB16( in_Q10, out0_Q10 );
+            RD_Q25[ j ]           = silk_SMLABB( silk_MLA( RD_tmp_Q25, silk_SMULBB( diff_Q10, diff_Q10 ), w_Q5[ i ] ), mu_Q20, rate0_Q5 );
+            diff_Q10              = silk_SUB16( in_Q10, out1_Q10 );
+            RD_Q25[ j + nStates ] = silk_SMLABB( silk_MLA( RD_tmp_Q25, silk_SMULBB( diff_Q10, diff_Q10 ), w_Q5[ i ] ), mu_Q20, rate1_Q5 );
+        }
+
+        if( nStates <= NLSF_QUANT_DEL_DEC_STATES/2 ) {
+            /* double number of states and copy */
+            for( j = 0; j < nStates; j++ ) {
+                ind[ j + nStates ][ i ] = ind[ j ][ i ] + 1;
+            }
+            nStates = silk_LSHIFT( nStates, 1 );
+            for( j = nStates; j < NLSF_QUANT_DEL_DEC_STATES; j++ ) {
+                ind[ j ][ i ] = ind[ j - nStates ][ i ];
+            }
+        } else {
+            /* sort lower and upper half of RD_Q25, pairwise */
+            for( j = 0; j < NLSF_QUANT_DEL_DEC_STATES; j++ ) {
+                if( RD_Q25[ j ] > RD_Q25[ j + NLSF_QUANT_DEL_DEC_STATES ] ) {
+                    RD_max_Q25[ j ]                         = RD_Q25[ j ];
+                    RD_min_Q25[ j ]                         = RD_Q25[ j + NLSF_QUANT_DEL_DEC_STATES ];
+                    RD_Q25[ j ]                             = RD_min_Q25[ j ];
+                    RD_Q25[ j + NLSF_QUANT_DEL_DEC_STATES ] = RD_max_Q25[ j ];
+                    /* swap prev_out values */
+                    out0_Q10 = prev_out_Q10[ j ];
+                    prev_out_Q10[ j ] = prev_out_Q10[ j + NLSF_QUANT_DEL_DEC_STATES ];
+                    prev_out_Q10[ j + NLSF_QUANT_DEL_DEC_STATES ] = out0_Q10;
+                    ind_sort[ j ] = j + NLSF_QUANT_DEL_DEC_STATES;
+                } else {
+                    RD_min_Q25[ j ] = RD_Q25[ j ];
+                    RD_max_Q25[ j ] = RD_Q25[ j + NLSF_QUANT_DEL_DEC_STATES ];
+                    ind_sort[ j ] = j;
+                }
+            }
+            /* compare the highest RD values of the winning half with the lowest one in the losing half, and copy if necessary */
+            /* afterwards ind_sort[] will contain the indices of the NLSF_QUANT_DEL_DEC_STATES winning RD values */
+            while( 1 ) {
+                min_max_Q25 = silk_int32_MAX;
+                max_min_Q25 = 0;
+                ind_min_max = 0;
+                ind_max_min = 0;
+                for( j = 0; j < NLSF_QUANT_DEL_DEC_STATES; j++ ) {
+                    if( min_max_Q25 > RD_max_Q25[ j ] ) {
+                        min_max_Q25 = RD_max_Q25[ j ];
+                        ind_min_max = j;
+                    }
+                    if( max_min_Q25 < RD_min_Q25[ j ] ) {
+                        max_min_Q25 = RD_min_Q25[ j ];
+                        ind_max_min = j;
+                    }
+                }
+                if( min_max_Q25 >= max_min_Q25 ) {
+                    break;
+                }
+                /* copy ind_min_max to ind_max_min */
+                ind_sort[     ind_max_min ] = ind_sort[     ind_min_max ] ^ NLSF_QUANT_DEL_DEC_STATES;
+                RD_Q25[       ind_max_min ] = RD_Q25[       ind_min_max + NLSF_QUANT_DEL_DEC_STATES ];
+                prev_out_Q10[ ind_max_min ] = prev_out_Q10[ ind_min_max + NLSF_QUANT_DEL_DEC_STATES ];
+                RD_min_Q25[   ind_max_min ] = 0;
+                RD_max_Q25[   ind_min_max ] = silk_int32_MAX;
+                silk_memcpy( ind[ ind_max_min ], ind[ ind_min_max ], MAX_LPC_ORDER * sizeof( opus_int8 ) );
+            }
+            /* increment index if it comes from the upper half */
+            for( j = 0; j < NLSF_QUANT_DEL_DEC_STATES; j++ ) {
+                ind[ j ][ i ] += silk_RSHIFT( ind_sort[ j ], NLSF_QUANT_DEL_DEC_STATES_LOG2 );
+            }
+        }
+    }
+
+    /* last sample: find winner, copy indices and return RD value */
+    ind_tmp = 0;
+    min_Q25 = silk_int32_MAX;
+    for( j = 0; j < 2 * NLSF_QUANT_DEL_DEC_STATES; j++ ) {
+        if( min_Q25 > RD_Q25[ j ] ) {
+            min_Q25 = RD_Q25[ j ];
+            ind_tmp = j;
+        }
+    }
+    for( j = 0; j < order; j++ ) {
+        indices[ j ] = ind[ ind_tmp & ( NLSF_QUANT_DEL_DEC_STATES - 1 ) ][ j ];
+        silk_assert( indices[ j ] >= -NLSF_QUANT_MAX_AMPLITUDE_EXT );
+        silk_assert( indices[ j ] <=  NLSF_QUANT_MAX_AMPLITUDE_EXT );
+    }
+    indices[ 0 ] += silk_RSHIFT( ind_tmp, NLSF_QUANT_DEL_DEC_STATES_LOG2 );
+    silk_assert( indices[ 0 ] <= NLSF_QUANT_MAX_AMPLITUDE_EXT );
+    silk_assert( min_Q25 >= 0 );
+    return min_Q25;
+}
--- a/lib/libopus/silk/NLSF_encode.c
+++ b/lib/libopus/silk/NLSF_encode.c
@ -0,0 +1,124 @@
+/***********************************************************************
+Copyright (c) 2006-2011, Skype Limited. All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+- Redistributions of source code must retain the above copyright notice,
+this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+***********************************************************************/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "main.h"
+#include "stack_alloc.h"
+
+/***********************/
+/* NLSF vector encoder */
+/***********************/
+opus_int32 silk_NLSF_encode(                                    /* O    Returns RD value in Q25                     */
+          opus_int8             *NLSFIndices,                   /* I    Codebook path vector [ LPC_ORDER + 1 ]      */
+          opus_int16            *pNLSF_Q15,                     /* I/O  (Un)quantized NLSF vector [ LPC_ORDER ]     */
+    const silk_NLSF_CB_struct   *psNLSF_CB,                     /* I    Codebook object                             */
+    const opus_int16            *pW_Q2,                         /* I    NLSF weight vector [ LPC_ORDER ]            */
+    const opus_int              NLSF_mu_Q20,                    /* I    Rate weight for the RD optimization         */
+    const opus_int              nSurvivors,                     /* I    Max survivors after first stage             */
+    const opus_int              signalType                      /* I    Signal type: 0/1/2                          */
+)
+{
+    opus_int         i, s, ind1, bestIndex, prob_Q8, bits_q7;
+    opus_int32       W_tmp_Q9, ret;
+    VARDECL( opus_int32, err_Q24 );
+    VARDECL( opus_int32, RD_Q25 );
+    VARDECL( opus_int, tempIndices1 );
+    VARDECL( opus_int8, tempIndices2 );
+    opus_int16       res_Q10[      MAX_LPC_ORDER ];
+    opus_int16       NLSF_tmp_Q15[ MAX_LPC_ORDER ];
+    opus_int16       W_adj_Q5[     MAX_LPC_ORDER ];
+    opus_uint8       pred_Q8[      MAX_LPC_ORDER ];
+    opus_int16       ec_ix[        MAX_LPC_ORDER ];
+    const opus_uint8 *pCB_element, *iCDF_ptr;
+    const opus_int16 *pCB_Wght_Q9;
+    SAVE_STACK;
+
+    celt_assert( signalType >= 0 && signalType <= 2 );
+    silk_assert( NLSF_mu_Q20 <= 32767 && NLSF_mu_Q20 >= 0 );
+
+    /* NLSF stabilization */
+    silk_NLSF_stabilize( pNLSF_Q15, psNLSF_CB->deltaMin_Q15, psNLSF_CB->order );
+
+    /* First stage: VQ */
+    ALLOC( err_Q24, psNLSF_CB->nVectors, opus_int32 );
+    silk_NLSF_VQ( err_Q24, pNLSF_Q15, psNLSF_CB->CB1_NLSF_Q8, psNLSF_CB->CB1_Wght_Q9, psNLSF_CB->nVectors, psNLSF_CB->order );
+
+    /* Sort the quantization errors */
+    ALLOC( tempIndices1, nSurvivors, opus_int );
+    silk_insertion_sort_increasing( err_Q24, tempIndices1, psNLSF_CB->nVectors, nSurvivors );
+
+    ALLOC( RD_Q25, nSurvivors, opus_int32 );
+    ALLOC( tempIndices2, nSurvivors * MAX_LPC_ORDER, opus_int8 );
+
+    /* Loop over survivors */
+    for( s = 0; s < nSurvivors; s++ ) {
+        ind1 = tempIndices1[ s ];
+
+        /* Residual after first stage */
+        pCB_element = &psNLSF_CB->CB1_NLSF_Q8[ ind1 * psNLSF_CB->order ];
+        pCB_Wght_Q9 = &psNLSF_CB->CB1_Wght_Q9[ ind1 * psNLSF_CB->order ];
+        for( i = 0; i < psNLSF_CB->order; i++ ) {
+            NLSF_tmp_Q15[ i ] = silk_LSHIFT16( (opus_int16)pCB_element[ i ], 7 );
+            W_tmp_Q9 = pCB_Wght_Q9[ i ];
+            res_Q10[ i ] = (opus_int16)silk_RSHIFT( silk_SMULBB( pNLSF_Q15[ i ] - NLSF_tmp_Q15[ i ], W_tmp_Q9 ), 14 );
+            W_adj_Q5[ i ] = silk_DIV32_varQ( (opus_int32)pW_Q2[ i ], silk_SMULBB( W_tmp_Q9, W_tmp_Q9 ), 21 );
+        }
+
+        /* Unpack entropy table indices and predictor for current CB1 index */
+        silk_NLSF_unpack( ec_ix, pred_Q8, psNLSF_CB, ind1 );
+
+        /* Trellis quantizer */
+        RD_Q25[ s ] = silk_NLSF_del_dec_quant( &tempIndices2[ s * MAX_LPC_ORDER ], res_Q10, W_adj_Q5, pred_Q8, ec_ix,
+            psNLSF_CB->ec_Rates_Q5, psNLSF_CB->quantStepSize_Q16, psNLSF_CB->invQuantStepSize_Q6, NLSF_mu_Q20, psNLSF_CB->order );
+
+        /* Add rate for first stage */
+        iCDF_ptr = &psNLSF_CB->CB1_iCDF[ ( signalType >> 1 ) * psNLSF_CB->nVectors ];
+        if( ind1 == 0 ) {
+            prob_Q8 = 256 - iCDF_ptr[ ind1 ];
+        } else {
+            prob_Q8 = iCDF_ptr[ ind1 - 1 ] - iCDF_ptr[ ind1 ];
+        }
+        bits_q7 = ( 8 << 7 ) - silk_lin2log( prob_Q8 );
+        RD_Q25[ s ] = silk_SMLABB( RD_Q25[ s ], bits_q7, silk_RSHIFT( NLSF_mu_Q20, 2 ) );
+    }
+
+    /* Find the lowest rate-distortion error */
+    silk_insertion_sort_increasing( RD_Q25, &bestIndex, nSurvivors, 1 );
+
+    NLSFIndices[ 0 ] = (opus_int8)tempIndices1[ bestIndex ];
+    silk_memcpy( &NLSFIndices[ 1 ], &tempIndices2[ bestIndex * MAX_LPC_ORDER ], psNLSF_CB->order * sizeof( opus_int8 ) );
+
+    /* Decode */
+    silk_NLSF_decode( pNLSF_Q15, NLSFIndices, psNLSF_CB );
+
+    ret = RD_Q25[ 0 ];
+    RESTORE_STACK;
+    return ret;
+}
--- a/lib/libopus/silk/NLSF_stabilize.c
+++ b/lib/libopus/silk/NLSF_stabilize.c
@ -0,0 +1,142 @@
+/***********************************************************************
+Copyright (c) 2006-2011, Skype Limited. All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+- Redistributions of source code must retain the above copyright notice,
+this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+***********************************************************************/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+/* NLSF stabilizer:                                         */
+/*                                                          */
+/* - Moves NLSFs further apart if they are too close        */
+/* - Moves NLSFs away from borders if they are too close    */
+/* - High effort to achieve a modification with minimum     */
+/*     Euclidean distance to input vector                   */
+/* - Output are sorted NLSF coefficients                    */
+/*                                                          */
+
+#include "SigProc_FIX.h"
+
+/* Constant Definitions */
+#define MAX_LOOPS        20
+
+/* NLSF stabilizer, for a single input data vector */
+void silk_NLSF_stabilize(
+          opus_int16            *NLSF_Q15,          /* I/O   Unstable/stabilized normalized LSF vector in Q15 [L]       */
+    const opus_int16            *NDeltaMin_Q15,     /* I     Min distance vector, NDeltaMin_Q15[L] must be >= 1 [L+1]   */
+    const opus_int              L                   /* I     Number of NLSF parameters in the input vector              */
+)
+{
+    opus_int   i, I=0, k, loops;
+    opus_int16 center_freq_Q15;
+    opus_int32 diff_Q15, min_diff_Q15, min_center_Q15, max_center_Q15;
+
+    /* This is necessary to ensure an output within range of a opus_int16 */
+    silk_assert( NDeltaMin_Q15[L] >= 1 );
+
+    for( loops = 0; loops < MAX_LOOPS; loops++ ) {
+        /**************************/
+        /* Find smallest distance */
+        /**************************/
+        /* First element */
+        min_diff_Q15 = NLSF_Q15[0] - NDeltaMin_Q15[0];
+        I = 0;
+        /* Middle elements */
+        for( i = 1; i <= L-1; i++ ) {
+            diff_Q15 = NLSF_Q15[i] - ( NLSF_Q15[i-1] + NDeltaMin_Q15[i] );
+            if( diff_Q15 < min_diff_Q15 ) {
+                min_diff_Q15 = diff_Q15;
+                I = i;
+            }
+        }
+        /* Last element */
+        diff_Q15 = ( 1 << 15 ) - ( NLSF_Q15[L-1] + NDeltaMin_Q15[L] );
+        if( diff_Q15 < min_diff_Q15 ) {
+            min_diff_Q15 = diff_Q15;
+            I = L;
+        }
+
+        /***************************************************/
+        /* Now check if the smallest distance non-negative */
+        /***************************************************/
+        if( min_diff_Q15 >= 0 ) {
+            return;
+        }
+
+        if( I == 0 ) {
+            /* Move away from lower limit */
+            NLSF_Q15[0] = NDeltaMin_Q15[0];
+
+        } else if( I == L) {
+            /* Move away from higher limit */
+            NLSF_Q15[L-1] = ( 1 << 15 ) - NDeltaMin_Q15[L];
+
+        } else {
+            /* Find the lower extreme for the location of the current center frequency */
+            min_center_Q15 = 0;
+            for( k = 0; k < I; k++ ) {
+                min_center_Q15 += NDeltaMin_Q15[k];
+            }
+            min_center_Q15 += silk_RSHIFT( NDeltaMin_Q15[I], 1 );
+
+            /* Find the upper extreme for the location of the current center frequency */
+            max_center_Q15 = 1 << 15;
+            for( k = L; k > I; k-- ) {
+                max_center_Q15 -= NDeltaMin_Q15[k];
+            }
+            max_center_Q15 -= silk_RSHIFT( NDeltaMin_Q15[I], 1 );
+
+            /* Move apart, sorted by value, keeping the same center frequency */
+            center_freq_Q15 = (opus_int16)silk_LIMIT_32( silk_RSHIFT_ROUND( (opus_int32)NLSF_Q15[I-1] + (opus_int32)NLSF_Q15[I], 1 ),
+                min_center_Q15, max_center_Q15 );
+            NLSF_Q15[I-1] = center_freq_Q15 - silk_RSHIFT( NDeltaMin_Q15[I], 1 );
+            NLSF_Q15[I] = NLSF_Q15[I-1] + NDeltaMin_Q15[I];
+        }
+    }
+
+    /* Safe and simple fall back method, which is less ideal than the above */
+    if( loops == MAX_LOOPS )
+    {
+        /* Insertion sort (fast for already almost sorted arrays):   */
+        /* Best case:  O(n)   for an already sorted array            */
+        /* Worst case: O(n^2) for an inversely sorted array          */
+        silk_insertion_sort_increasing_all_values_int16( &NLSF_Q15[0], L );
+
+        /* First NLSF should be no less than NDeltaMin[0] */
+        NLSF_Q15[0] = silk_max_int( NLSF_Q15[0], NDeltaMin_Q15[0] );
+
+        /* Keep delta_min distance between the NLSFs */
+        for( i = 1; i < L; i++ )
+            NLSF_Q15[i] = silk_max_int( NLSF_Q15[i], silk_ADD_SAT16( NLSF_Q15[i-1], NDeltaMin_Q15[i] ) );
+
+        /* Last NLSF should be no higher than 1 - NDeltaMin[L] */
+        NLSF_Q15[L-1] = silk_min_int( NLSF_Q15[L-1], (1<<15) - NDeltaMin_Q15[L] );
+
+        /* Keep NDeltaMin distance between the NLSFs */
+        for( i = L-2; i >= 0; i-- )
+            NLSF_Q15[i] = silk_min_int( NLSF_Q15[i], NLSF_Q15[i+1] - NDeltaMin_Q15[i+1] );
+    }
+}
--- a/lib/libopus/silk/NLSF_unpack.c
+++ b/lib/libopus/silk/NLSF_unpack.c
@ -0,0 +1,54 @@
+/***********************************************************************
+Copyright (c) 2006-2011, Skype Limited. All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+- Redistributions of source code must retain the above copyright notice,
+this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+***********************************************************************/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "main.h"
+
+/* Unpack predictor values and indices for entropy coding tables */
+void silk_NLSF_unpack(
+          opus_int16            ec_ix[],                        /* O    Indices to entropy tables [ LPC_ORDER ]     */
+          opus_uint8            pred_Q8[],                      /* O    LSF predictor [ LPC_ORDER ]                 */
+    const silk_NLSF_CB_struct   *psNLSF_CB,                     /* I    Codebook object                             */
+    const opus_int              CB1_index                       /* I    Index of vector in first LSF codebook       */
+)
+{
+    opus_int   i;
+    opus_uint8 entry;
+    const opus_uint8 *ec_sel_ptr;
+
+    ec_sel_ptr = &psNLSF_CB->ec_sel[ CB1_index * psNLSF_CB->order / 2 ];
+    for( i = 0; i < psNLSF_CB->order; i += 2 ) {
+        entry = *ec_sel_ptr++;
+        ec_ix  [ i     ] = silk_SMULBB( silk_RSHIFT( entry, 1 ) & 7, 2 * NLSF_QUANT_MAX_AMPLITUDE + 1 );
+        pred_Q8[ i     ] = psNLSF_CB->pred_Q8[ i + ( entry & 1 ) * ( psNLSF_CB->order - 1 ) ];
+        ec_ix  [ i + 1 ] = silk_SMULBB( silk_RSHIFT( entry, 5 ) & 7, 2 * NLSF_QUANT_MAX_AMPLITUDE + 1 );
+        pred_Q8[ i + 1 ] = psNLSF_CB->pred_Q8[ i + ( silk_RSHIFT( entry, 4 ) & 1 ) * ( psNLSF_CB->order - 1 ) + 1 ];
+    }
+}
--- a/lib/libopus/silk/NSQ.c
+++ b/lib/libopus/silk/NSQ.c
@ -0,0 +1,437 @@
+/***********************************************************************
+Copyright (c) 2006-2011, Skype Limited. All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+- Redistributions of source code must retain the above copyright notice,
+this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+***********************************************************************/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "main.h"
+#include "stack_alloc.h"
+#include "NSQ.h"
+
+
+static OPUS_INLINE void silk_nsq_scale_states(
+    const silk_encoder_state *psEncC,           /* I    Encoder State                   */
+    silk_nsq_state      *NSQ,                   /* I/O  NSQ state                       */
+    const opus_int16    x16[],                  /* I    input                           */
+    opus_int32          x_sc_Q10[],             /* O    input scaled with 1/Gain        */
+    const opus_int16    sLTP[],                 /* I    re-whitened LTP state in Q0     */
+    opus_int32          sLTP_Q15[],             /* O    LTP state matching scaled input */
+    opus_int            subfr,                  /* I    subframe number                 */
+    const opus_int      LTP_scale_Q14,          /* I                                    */
+    const opus_int32    Gains_Q16[ MAX_NB_SUBFR ], /* I                                 */
+    const opus_int      pitchL[ MAX_NB_SUBFR ], /* I    Pitch lag                       */
+    const opus_int      signal_type             /* I    Signal type                     */
+);
+
+#if !defined(OPUS_X86_MAY_HAVE_SSE4_1)
+static OPUS_INLINE void silk_noise_shape_quantizer(
+    silk_nsq_state      *NSQ,                   /* I/O  NSQ state                       */
+    opus_int            signalType,             /* I    Signal type                     */
+    const opus_int32    x_sc_Q10[],             /* I                                    */
+    opus_int8           pulses[],               /* O                                    */
+    opus_int16          xq[],                   /* O                                    */
+    opus_int32          sLTP_Q15[],             /* I/O  LTP state                       */
+    const opus_int16    a_Q12[],                /* I    Short term prediction coefs     */
+    const opus_int16    b_Q14[],                /* I    Long term prediction coefs      */
+    const opus_int16    AR_shp_Q13[],           /* I    Noise shaping AR coefs          */
+    opus_int            lag,                    /* I    Pitch lag                       */
+    opus_int32          HarmShapeFIRPacked_Q14, /* I                                    */
+    opus_int            Tilt_Q14,               /* I    Spectral tilt                   */
+    opus_int32          LF_shp_Q14,             /* I                                    */
+    opus_int32          Gain_Q16,               /* I                                    */
+    opus_int            Lambda_Q10,             /* I                                    */
+    opus_int            offset_Q10,             /* I                                    */
+    opus_int            length,                 /* I    Input length                    */
+    opus_int            shapingLPCOrder,        /* I    Noise shaping AR filter order   */
+    opus_int            predictLPCOrder,        /* I    Prediction filter order         */
+    int                 arch                    /* I    Architecture                    */
+);
+#endif
+
+void silk_NSQ_c
+(
+    const silk_encoder_state    *psEncC,                                      /* I    Encoder State                   */
+    silk_nsq_state              *NSQ,                                         /* I/O  NSQ state                       */
+    SideInfoIndices             *psIndices,                                   /* I/O  Quantization Indices            */
+    const opus_int16            x16[],                                        /* I    Input                           */
+    opus_int8                   pulses[],                                     /* O    Quantized pulse signal          */
+    const opus_int16            *PredCoef_Q12,                                /* I    Short term prediction coefs     */
+    const opus_int16            LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ],      /* I    Long term prediction coefs      */
+    const opus_int16            AR_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I    Noise shaping coefs             */
+    const opus_int              HarmShapeGain_Q14[ MAX_NB_SUBFR ],            /* I    Long term shaping coefs         */
+    const opus_int              Tilt_Q14[ MAX_NB_SUBFR ],                     /* I    Spectral tilt                   */
+    const opus_int32            LF_shp_Q14[ MAX_NB_SUBFR ],                   /* I    Low frequency shaping coefs     */
+    const opus_int32            Gains_Q16[ MAX_NB_SUBFR ],                    /* I    Quantization step sizes         */
+    const opus_int              pitchL[ MAX_NB_SUBFR ],                       /* I    Pitch lags                      */
+    const opus_int              Lambda_Q10,                                   /* I    Rate/distortion tradeoff        */
+    const opus_int              LTP_scale_Q14                                 /* I    LTP state scaling               */
+)
+{
+    opus_int            k, lag, start_idx, LSF_interpolation_flag;
+    const opus_int16    *A_Q12, *B_Q14, *AR_shp_Q13;
+    opus_int16          *pxq;
+    VARDECL( opus_int32, sLTP_Q15 );
+    VARDECL( opus_int16, sLTP );
+    opus_int32          HarmShapeFIRPacked_Q14;
+    opus_int            offset_Q10;
+    VARDECL( opus_int32, x_sc_Q10 );
+    SAVE_STACK;
+
+    NSQ->rand_seed = psIndices->Seed;
+
+    /* Set unvoiced lag to the previous one, overwrite later for voiced */
+    lag = NSQ->lagPrev;
+
+    silk_assert( NSQ->prev_gain_Q16 != 0 );
+
+    offset_Q10 = silk_Quantization_Offsets_Q10[ psIndices->signalType >> 1 ][ psIndices->quantOffsetType ];
+
+    if( psIndices->NLSFInterpCoef_Q2 == 4 ) {
+        LSF_interpolation_flag = 0;
+    } else {
+        LSF_interpolation_flag = 1;
+    }
+
+    ALLOC( sLTP_Q15, psEncC->ltp_mem_length + psEncC->frame_length, opus_int32 );
+    ALLOC( sLTP, psEncC->ltp_mem_length + psEncC->frame_length, opus_int16 );
+    ALLOC( x_sc_Q10, psEncC->subfr_length, opus_int32 );
+    /* Set up pointers to start of sub frame */
+    NSQ->sLTP_shp_buf_idx = psEncC->ltp_mem_length;
+    NSQ->sLTP_buf_idx     = psEncC->ltp_mem_length;
+    pxq                   = &NSQ->xq[ psEncC->ltp_mem_length ];
+    for( k = 0; k < psEncC->nb_subfr; k++ ) {
+        A_Q12      = &PredCoef_Q12[ (( k >> 1 ) | ( 1 - LSF_interpolation_flag )) * MAX_LPC_ORDER ];
+        B_Q14      = &LTPCoef_Q14[ k * LTP_ORDER ];
+        AR_shp_Q13 = &AR_Q13[ k * MAX_SHAPE_LPC_ORDER ];
+
+        /* Noise shape parameters */
+        silk_assert( HarmShapeGain_Q14[ k ] >= 0 );
+        HarmShapeFIRPacked_Q14  =                          silk_RSHIFT( HarmShapeGain_Q14[ k ], 2 );
+        HarmShapeFIRPacked_Q14 |= silk_LSHIFT( (opus_int32)silk_RSHIFT( HarmShapeGain_Q14[ k ], 1 ), 16 );
+
+        NSQ->rewhite_flag = 0;
+        if( psIndices->signalType == TYPE_VOICED ) {
+            /* Voiced */
+            lag = pitchL[ k ];
+
+            /* Re-whitening */
+            if( ( k & ( 3 - silk_LSHIFT( LSF_interpolation_flag, 1 ) ) ) == 0 ) {
+                /* Rewhiten with new A coefs */
+                start_idx = psEncC->ltp_mem_length - lag - psEncC->predictLPCOrder - LTP_ORDER / 2;
+                celt_assert( start_idx > 0 );
+
+                silk_LPC_analysis_filter( &sLTP[ start_idx ], &NSQ->xq[ start_idx + k * psEncC->subfr_length ],
+                    A_Q12, psEncC->ltp_mem_length - start_idx, psEncC->predictLPCOrder, psEncC->arch );
+
+                NSQ->rewhite_flag = 1;
+                NSQ->sLTP_buf_idx = psEncC->ltp_mem_length;
+            }
+        }
+
+        silk_nsq_scale_states( psEncC, NSQ, x16, x_sc_Q10, sLTP, sLTP_Q15, k, LTP_scale_Q14, Gains_Q16, pitchL, psIndices->signalType );
+
+        silk_noise_shape_quantizer( NSQ, psIndices->signalType, x_sc_Q10, pulses, pxq, sLTP_Q15, A_Q12, B_Q14,
+            AR_shp_Q13, lag, HarmShapeFIRPacked_Q14, Tilt_Q14[ k ], LF_shp_Q14[ k ], Gains_Q16[ k ], Lambda_Q10,
+            offset_Q10, psEncC->subfr_length, psEncC->shapingLPCOrder, psEncC->predictLPCOrder, psEncC->arch );
+
+        x16    += psEncC->subfr_length;
+        pulses += psEncC->subfr_length;
+        pxq    += psEncC->subfr_length;
+    }
+
+    /* Update lagPrev for next frame */
+    NSQ->lagPrev = pitchL[ psEncC->nb_subfr - 1 ];
+
+    /* Save quantized speech and noise shaping signals */
+    silk_memmove( NSQ->xq,           &NSQ->xq[           psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int16 ) );
+    silk_memmove( NSQ->sLTP_shp_Q14, &NSQ->sLTP_shp_Q14[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int32 ) );
+    RESTORE_STACK;
+}
+
+/******************************/
+/* silk_noise_shape_quantizer */
+/******************************/
+
+#if !defined(OPUS_X86_MAY_HAVE_SSE4_1)
+static OPUS_INLINE
+#endif
+void silk_noise_shape_quantizer(
+    silk_nsq_state      *NSQ,                   /* I/O  NSQ state                       */
+    opus_int            signalType,             /* I    Signal type                     */
+    const opus_int32    x_sc_Q10[],             /* I                                    */
+    opus_int8           pulses[],               /* O                                    */
+    opus_int16          xq[],                   /* O                                    */
+    opus_int32          sLTP_Q15[],             /* I/O  LTP state                       */
+    const opus_int16    a_Q12[],                /* I    Short term prediction coefs     */
+    const opus_int16    b_Q14[],                /* I    Long term prediction coefs      */
+    const opus_int16    AR_shp_Q13[],           /* I    Noise shaping AR coefs          */
+    opus_int            lag,                    /* I    Pitch lag                       */
+    opus_int32          HarmShapeFIRPacked_Q14, /* I                                    */
+    opus_int            Tilt_Q14,               /* I    Spectral tilt                   */
+    opus_int32          LF_shp_Q14,             /* I                                    */
+    opus_int32          Gain_Q16,               /* I                                    */
+    opus_int            Lambda_Q10,             /* I                                    */
+    opus_int            offset_Q10,             /* I                                    */
+    opus_int            length,                 /* I    Input length                    */
+    opus_int            shapingLPCOrder,        /* I    Noise shaping AR filter order   */
+    opus_int            predictLPCOrder,        /* I    Prediction filter order         */
+    int                 arch                    /* I    Architecture                    */
+)
+{
+    opus_int     i;
+    opus_int32   LTP_pred_Q13, LPC_pred_Q10, n_AR_Q12, n_LTP_Q13;
+    opus_int32   n_LF_Q12, r_Q10, rr_Q10, q1_Q0, q1_Q10, q2_Q10, rd1_Q20, rd2_Q20;
+    opus_int32   exc_Q14, LPC_exc_Q14, xq_Q14, Gain_Q10;
+    opus_int32   tmp1, tmp2, sLF_AR_shp_Q14;
+    opus_int32   *psLPC_Q14, *shp_lag_ptr, *pred_lag_ptr;
+#ifdef silk_short_prediction_create_arch_coef
+    opus_int32   a_Q12_arch[MAX_LPC_ORDER];
+#endif
+
+    shp_lag_ptr  = &NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - lag + HARM_SHAPE_FIR_TAPS / 2 ];
+    pred_lag_ptr = &sLTP_Q15[ NSQ->sLTP_buf_idx - lag + LTP_ORDER / 2 ];
+    Gain_Q10     = silk_RSHIFT( Gain_Q16, 6 );
+
+    /* Set up short term AR state */
+    psLPC_Q14 = &NSQ->sLPC_Q14[ NSQ_LPC_BUF_LENGTH - 1 ];
+
+#ifdef silk_short_prediction_create_arch_coef
+    silk_short_prediction_create_arch_coef(a_Q12_arch, a_Q12, predictLPCOrder);
+#endif
+
+    for( i = 0; i < length; i++ ) {
+        /* Generate dither */
+        NSQ->rand_seed = silk_RAND( NSQ->rand_seed );
+
+        /* Short-term prediction */
+        LPC_pred_Q10 = silk_noise_shape_quantizer_short_prediction(psLPC_Q14, a_Q12, a_Q12_arch, predictLPCOrder, arch);
+
+        /* Long-term prediction */
+        if( signalType == TYPE_VOICED ) {
+            /* Unrolled loop */
+            /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */
+            LTP_pred_Q13 = 2;
+            LTP_pred_Q13 = silk_SMLAWB( LTP_pred_Q13, pred_lag_ptr[  0 ], b_Q14[ 0 ] );
+            LTP_pred_Q13 = silk_SMLAWB( LTP_pred_Q13, pred_lag_ptr[ -1 ], b_Q14[ 1 ] );
+            LTP_pred_Q13 = silk_SMLAWB( LTP_pred_Q13, pred_lag_ptr[ -2 ], b_Q14[ 2 ] );
+            LTP_pred_Q13 = silk_SMLAWB( LTP_pred_Q13, pred_lag_ptr[ -3 ], b_Q14[ 3 ] );
+            LTP_pred_Q13 = silk_SMLAWB( LTP_pred_Q13, pred_lag_ptr[ -4 ], b_Q14[ 4 ] );
+            pred_lag_ptr++;
+        } else {
+            LTP_pred_Q13 = 0;
+        }
+
+        /* Noise shape feedback */
+        celt_assert( ( shapingLPCOrder & 1 ) == 0 );   /* check that order is even */
+        n_AR_Q12 = silk_NSQ_noise_shape_feedback_loop(&NSQ->sDiff_shp_Q14, NSQ->sAR2_Q14, AR_shp_Q13, shapingLPCOrder, arch);
+
+        n_AR_Q12 = silk_SMLAWB( n_AR_Q12, NSQ->sLF_AR_shp_Q14, Tilt_Q14 );
+
+        n_LF_Q12 = silk_SMULWB( NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - 1 ], LF_shp_Q14 );
+        n_LF_Q12 = silk_SMLAWT( n_LF_Q12, NSQ->sLF_AR_shp_Q14, LF_shp_Q14 );
+
+        celt_assert( lag > 0 || signalType != TYPE_VOICED );
+
+        /* Combine prediction and noise shaping signals */
+        tmp1 = silk_SUB32_ovflw( silk_LSHIFT32( LPC_pred_Q10, 2 ), n_AR_Q12 );  /* Q12 */
+        tmp1 = silk_SUB32_ovflw( tmp1, n_LF_Q12 );                              /* Q12 */
+        if( lag > 0 ) {
+            /* Symmetric, packed FIR coefficients */
+            n_LTP_Q13 = silk_SMULWB( silk_ADD_SAT32( shp_lag_ptr[ 0 ], shp_lag_ptr[ -2 ] ), HarmShapeFIRPacked_Q14 );
+            n_LTP_Q13 = silk_SMLAWT( n_LTP_Q13, shp_lag_ptr[ -1 ],                      HarmShapeFIRPacked_Q14 );
+            n_LTP_Q13 = silk_LSHIFT( n_LTP_Q13, 1 );
+            shp_lag_ptr++;
+
+            tmp2 = silk_SUB32( LTP_pred_Q13, n_LTP_Q13 );                       /* Q13 */
+            tmp1 = silk_ADD32_ovflw( tmp2, silk_LSHIFT32( tmp1, 1 ) );          /* Q13 */
+            tmp1 = silk_RSHIFT_ROUND( tmp1, 3 );                                /* Q10 */
+        } else {
+            tmp1 = silk_RSHIFT_ROUND( tmp1, 2 );                                /* Q10 */
+        }
+
+        r_Q10 = silk_SUB32( x_sc_Q10[ i ], tmp1 );                              /* residual error Q10 */
+
+        /* Flip sign depending on dither */
+        if( NSQ->rand_seed < 0 ) {
+            r_Q10 = -r_Q10;
+        }
+        r_Q10 = silk_LIMIT_32( r_Q10, -(31 << 10), 30 << 10 );
+
+        /* Find two quantization level candidates and measure their rate-distortion */
+        q1_Q10 = silk_SUB32( r_Q10, offset_Q10 );
+        q1_Q0 = silk_RSHIFT( q1_Q10, 10 );
+        if (Lambda_Q10 > 2048) {
+            /* For aggressive RDO, the bias becomes more than one pulse. */
+            int rdo_offset = Lambda_Q10/2 - 512;
+            if (q1_Q10 > rdo_offset) {
+                q1_Q0 = silk_RSHIFT( q1_Q10 - rdo_offset, 10 );
+            } else if (q1_Q10 < -rdo_offset) {
+                q1_Q0 = silk_RSHIFT( q1_Q10 + rdo_offset, 10 );
+            } else if (q1_Q10 < 0) {
+                q1_Q0 = -1;
+            } else {
+                q1_Q0 = 0;
+            }
+        }
+        if( q1_Q0 > 0 ) {
+            q1_Q10  = silk_SUB32( silk_LSHIFT( q1_Q0, 10 ), QUANT_LEVEL_ADJUST_Q10 );
+            q1_Q10  = silk_ADD32( q1_Q10, offset_Q10 );
+            q2_Q10  = silk_ADD32( q1_Q10, 1024 );
+            rd1_Q20 = silk_SMULBB( q1_Q10, Lambda_Q10 );
+            rd2_Q20 = silk_SMULBB( q2_Q10, Lambda_Q10 );
+        } else if( q1_Q0 == 0 ) {
+            q1_Q10  = offset_Q10;
+            q2_Q10  = silk_ADD32( q1_Q10, 1024 - QUANT_LEVEL_ADJUST_Q10 );
+            rd1_Q20 = silk_SMULBB( q1_Q10, Lambda_Q10 );
+            rd2_Q20 = silk_SMULBB( q2_Q10, Lambda_Q10 );
+        } else if( q1_Q0 == -1 ) {
+            q2_Q10  = offset_Q10;
+            q1_Q10  = silk_SUB32( q2_Q10, 1024 - QUANT_LEVEL_ADJUST_Q10 );
+            rd1_Q20 = silk_SMULBB( -q1_Q10, Lambda_Q10 );
+            rd2_Q20 = silk_SMULBB(  q2_Q10, Lambda_Q10 );
+        } else {            /* Q1_Q0 < -1 */
+            q1_Q10  = silk_ADD32( silk_LSHIFT( q1_Q0, 10 ), QUANT_LEVEL_ADJUST_Q10 );
+            q1_Q10  = silk_ADD32( q1_Q10, offset_Q10 );
+            q2_Q10  = silk_ADD32( q1_Q10, 1024 );
+            rd1_Q20 = silk_SMULBB( -q1_Q10, Lambda_Q10 );
+            rd2_Q20 = silk_SMULBB( -q2_Q10, Lambda_Q10 );
+        }
+        rr_Q10  = silk_SUB32( r_Q10, q1_Q10 );
+        rd1_Q20 = silk_SMLABB( rd1_Q20, rr_Q10, rr_Q10 );
+        rr_Q10  = silk_SUB32( r_Q10, q2_Q10 );
+        rd2_Q20 = silk_SMLABB( rd2_Q20, rr_Q10, rr_Q10 );
+
+        if( rd2_Q20 < rd1_Q20 ) {
+            q1_Q10 = q2_Q10;
+        }
+
+        pulses[ i ] = (opus_int8)silk_RSHIFT_ROUND( q1_Q10, 10 );
+
+        /* Excitation */
+        exc_Q14 = silk_LSHIFT( q1_Q10, 4 );
+        if ( NSQ->rand_seed < 0 ) {
+           exc_Q14 = -exc_Q14;
+        }
+
+        /* Add predictions */
+        LPC_exc_Q14 = silk_ADD_LSHIFT32( exc_Q14, LTP_pred_Q13, 1 );
+        xq_Q14      = silk_ADD32_ovflw( LPC_exc_Q14, silk_LSHIFT32( LPC_pred_Q10, 4 ) );
+
+        /* Scale XQ back to normal level before saving */
+        xq[ i ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( xq_Q14, Gain_Q10 ), 8 ) );
+
+        /* Update states */
+        psLPC_Q14++;
+        *psLPC_Q14 = xq_Q14;
+        NSQ->sDiff_shp_Q14 = silk_SUB32_ovflw( xq_Q14, silk_LSHIFT32( x_sc_Q10[ i ], 4 ) );
+        sLF_AR_shp_Q14 = silk_SUB32_ovflw( NSQ->sDiff_shp_Q14, silk_LSHIFT32( n_AR_Q12, 2 ) );
+        NSQ->sLF_AR_shp_Q14 = sLF_AR_shp_Q14;
+
+        NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx ] = silk_SUB32_ovflw(sLF_AR_shp_Q14, silk_LSHIFT32(n_LF_Q12, 2));
+        sLTP_Q15[ NSQ->sLTP_buf_idx ] = silk_LSHIFT( LPC_exc_Q14, 1 );
+        NSQ->sLTP_shp_buf_idx++;
+        NSQ->sLTP_buf_idx++;
+
+        /* Make dither dependent on quantized signal */
+        NSQ->rand_seed = silk_ADD32_ovflw( NSQ->rand_seed, pulses[ i ] );
+    }
+
+    /* Update LPC synth buffer */
+    silk_memcpy( NSQ->sLPC_Q14, &NSQ->sLPC_Q14[ length ], NSQ_LPC_BUF_LENGTH * sizeof( opus_int32 ) );
+}
+
+static OPUS_INLINE void silk_nsq_scale_states(
+    const silk_encoder_state *psEncC,           /* I    Encoder State                   */
+    silk_nsq_state      *NSQ,                   /* I/O  NSQ state                       */
+    const opus_int16    x16[],                  /* I    input                           */
+    opus_int32          x_sc_Q10[],             /* O    input scaled with 1/Gain        */
+    const opus_int16    sLTP[],                 /* I    re-whitened LTP state in Q0     */
+    opus_int32          sLTP_Q15[],             /* O    LTP state matching scaled input */
+    opus_int            subfr,                  /* I    subframe number                 */
+    const opus_int      LTP_scale_Q14,          /* I                                    */
+    const opus_int32    Gains_Q16[ MAX_NB_SUBFR ], /* I                                 */
+    const opus_int      pitchL[ MAX_NB_SUBFR ], /* I    Pitch lag                       */
+    const opus_int      signal_type             /* I    Signal type                     */
+)
+{
+    opus_int   i, lag;
+    opus_int32 gain_adj_Q16, inv_gain_Q31, inv_gain_Q26;
+
+    lag          = pitchL[ subfr ];
+    inv_gain_Q31 = silk_INVERSE32_varQ( silk_max( Gains_Q16[ subfr ], 1 ), 47 );
+    silk_assert( inv_gain_Q31 != 0 );
+
+    /* Scale input */
+    inv_gain_Q26 = silk_RSHIFT_ROUND( inv_gain_Q31, 5 );
+    for( i = 0; i < psEncC->subfr_length; i++ ) {
+        x_sc_Q10[ i ] = silk_SMULWW( x16[ i ], inv_gain_Q26 );
+    }
+
+    /* After rewhitening the LTP state is un-scaled, so scale with inv_gain_Q16 */
+    if( NSQ->rewhite_flag ) {
+        if( subfr == 0 ) {
+            /* Do LTP downscaling */
+            inv_gain_Q31 = silk_LSHIFT( silk_SMULWB( inv_gain_Q31, LTP_scale_Q14 ), 2 );
+        }
+        for( i = NSQ->sLTP_buf_idx - lag - LTP_ORDER / 2; i < NSQ->sLTP_buf_idx; i++ ) {
+            silk_assert( i < MAX_FRAME_LENGTH );
+            sLTP_Q15[ i ] = silk_SMULWB( inv_gain_Q31, sLTP[ i ] );
+        }
+    }
+
+    /* Adjust for changing gain */
+    if( Gains_Q16[ subfr ] != NSQ->prev_gain_Q16 ) {
+        gain_adj_Q16 =  silk_DIV32_varQ( NSQ->prev_gain_Q16, Gains_Q16[ subfr ], 16 );
+
+        /* Scale long-term shaping state */
+        for( i = NSQ->sLTP_shp_buf_idx - psEncC->ltp_mem_length; i < NSQ->sLTP_shp_buf_idx; i++ ) {
+            NSQ->sLTP_shp_Q14[ i ] = silk_SMULWW( gain_adj_Q16, NSQ->sLTP_shp_Q14[ i ] );
+        }
+
+        /* Scale long-term prediction state */
+        if( signal_type == TYPE_VOICED && NSQ->rewhite_flag == 0 ) {
+            for( i = NSQ->sLTP_buf_idx - lag - LTP_ORDER / 2; i < NSQ->sLTP_buf_idx; i++ ) {
+                sLTP_Q15[ i ] = silk_SMULWW( gain_adj_Q16, sLTP_Q15[ i ] );
+            }
+        }
+
+        NSQ->sLF_AR_shp_Q14 = silk_SMULWW( gain_adj_Q16, NSQ->sLF_AR_shp_Q14 );
+        NSQ->sDiff_shp_Q14 = silk_SMULWW( gain_adj_Q16, NSQ->sDiff_shp_Q14 );
+
+        /* Scale short-term prediction and shaping states */
+        for( i = 0; i < NSQ_LPC_BUF_LENGTH; i++ ) {
+            NSQ->sLPC_Q14[ i ] = silk_SMULWW( gain_adj_Q16, NSQ->sLPC_Q14[ i ] );
+        }
+        for( i = 0; i < MAX_SHAPE_LPC_ORDER; i++ ) {
+            NSQ->sAR2_Q14[ i ] = silk_SMULWW( gain_adj_Q16, NSQ->sAR2_Q14[ i ] );
+        }
+
+        /* Save inverse gain */
+        NSQ->prev_gain_Q16 = Gains_Q16[ subfr ];
+    }
+}
--- a/lib/libopus/silk/NSQ.h
+++ b/lib/libopus/silk/NSQ.h
@ -0,0 +1,105 @@
+/***********************************************************************
+Copyright (c) 2014 Vidyo.
+Copyright (c) 2006-2011, Skype Limited. All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+- Redistributions of source code must retain the above copyright notice,
+this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+***********************************************************************/
+#ifndef SILK_NSQ_H
+#define SILK_NSQ_H
+
+#include "SigProc_FIX.h"
+
+#undef silk_short_prediction_create_arch_coef
+
+static OPUS_INLINE opus_int32 silk_noise_shape_quantizer_short_prediction_c(const opus_int32 *buf32, const opus_int16 *coef16, opus_int order)
+{
+    opus_int32 out;
+    silk_assert( order == 10 || order == 16 );
+
+    /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */
+    out = silk_RSHIFT( order, 1 );
+    out = silk_SMLAWB( out, buf32[  0 ], coef16[ 0 ] );
+    out = silk_SMLAWB( out, buf32[ -1 ], coef16[ 1 ] );
+    out = silk_SMLAWB( out, buf32[ -2 ], coef16[ 2 ] );
+    out = silk_SMLAWB( out, buf32[ -3 ], coef16[ 3 ] );
+    out = silk_SMLAWB( out, buf32[ -4 ], coef16[ 4 ] );
+    out = silk_SMLAWB( out, buf32[ -5 ], coef16[ 5 ] );
+    out = silk_SMLAWB( out, buf32[ -6 ], coef16[ 6 ] );
+    out = silk_SMLAWB( out, buf32[ -7 ], coef16[ 7 ] );
+    out = silk_SMLAWB( out, buf32[ -8 ], coef16[ 8 ] );
+    out = silk_SMLAWB( out, buf32[ -9 ], coef16[ 9 ] );
+
+    if( order == 16 )
+    {
+        out = silk_SMLAWB( out, buf32[ -10 ], coef16[ 10 ] );
+        out = silk_SMLAWB( out, buf32[ -11 ], coef16[ 11 ] );
+        out = silk_SMLAWB( out, buf32[ -12 ], coef16[ 12 ] );
+        out = silk_SMLAWB( out, buf32[ -13 ], coef16[ 13 ] );
+        out = silk_SMLAWB( out, buf32[ -14 ], coef16[ 14 ] );
+        out = silk_SMLAWB( out, buf32[ -15 ], coef16[ 15 ] );
+    }
+    return out;
+}
+
+#define silk_noise_shape_quantizer_short_prediction(in, coef, coefRev, order, arch)  ((void)arch,silk_noise_shape_quantizer_short_prediction_c(in, coef, order))
+
+static OPUS_INLINE opus_int32 silk_NSQ_noise_shape_feedback_loop_c(const opus_int32 *data0, opus_int32 *data1, const opus_int16 *coef, opus_int order)
+{
+    opus_int32 out;
+    opus_int32 tmp1, tmp2;
+    opus_int j;
+
+    tmp2 = data0[0];
+    tmp1 = data1[0];
+    data1[0] = tmp2;
+
+    out = silk_RSHIFT(order, 1);
+    out = silk_SMLAWB(out, tmp2, coef[0]);
+
+    for (j = 2; j < order; j += 2) {
+        tmp2 = data1[j - 1];
+        data1[j - 1] = tmp1;
+        out = silk_SMLAWB(out, tmp1, coef[j - 1]);
+        tmp1 = data1[j + 0];
+        data1[j + 0] = tmp2;
+        out = silk_SMLAWB(out, tmp2, coef[j]);
+    }
+    data1[order - 1] = tmp1;
+    out = silk_SMLAWB(out, tmp1, coef[order - 1]);
+    /* Q11 -> Q12 */
+    out = silk_LSHIFT32( out, 1 );
+    return out;
+}
+
+#define silk_NSQ_noise_shape_feedback_loop(data0, data1, coef, order, arch)  ((void)arch,silk_NSQ_noise_shape_feedback_loop_c(data0, data1, coef, order))
+
+#if defined(OPUS_ARM_MAY_HAVE_NEON_INTR)
+#include "arm/NSQ_neon.h"
+#endif
+
+#if defined(__mips)
+#include "mips/NSQ_mips.h"
+#endif
+
+#endif /* SILK_NSQ_H */
--- a/lib/libopus/silk/NSQ_del_dec.c
+++ b/lib/libopus/silk/NSQ_del_dec.c
@ -0,0 +1,730 @@
+/***********************************************************************
+Copyright (c) 2006-2011, Skype Limited. All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+- Redistributions of source code must retain the above copyright notice,
+this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+***********************************************************************/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "main.h"
+#include "stack_alloc.h"
+#include "NSQ.h"
+
+
+typedef struct {
+    opus_int32 sLPC_Q14[ MAX_SUB_FRAME_LENGTH + NSQ_LPC_BUF_LENGTH ];
+    opus_int32 RandState[ DECISION_DELAY ];
+    opus_int32 Q_Q10[     DECISION_DELAY ];
+    opus_int32 Xq_Q14[    DECISION_DELAY ];
+    opus_int32 Pred_Q15[  DECISION_DELAY ];
+    opus_int32 Shape_Q14[ DECISION_DELAY ];
+    opus_int32 sAR2_Q14[ MAX_SHAPE_LPC_ORDER ];
+    opus_int32 LF_AR_Q14;
+    opus_int32 Diff_Q14;
+    opus_int32 Seed;
+    opus_int32 SeedInit;
+    opus_int32 RD_Q10;
+} NSQ_del_dec_struct;
+
+typedef struct {
+    opus_int32 Q_Q10;
+    opus_int32 RD_Q10;
+    opus_int32 xq_Q14;
+    opus_int32 LF_AR_Q14;
+    opus_int32 Diff_Q14;
+    opus_int32 sLTP_shp_Q14;
+    opus_int32 LPC_exc_Q14;
+} NSQ_sample_struct;
+
+typedef NSQ_sample_struct  NSQ_sample_pair[ 2 ];
+
+static OPUS_INLINE void silk_nsq_del_dec_scale_states(
+    const silk_encoder_state *psEncC,               /* I    Encoder State                       */
+    silk_nsq_state      *NSQ,                       /* I/O  NSQ state                           */
+    NSQ_del_dec_struct  psDelDec[],                 /* I/O  Delayed decision states             */
+    const opus_int16    x16[],                      /* I    Input                               */
+    opus_int32          x_sc_Q10[],                 /* O    Input scaled with 1/Gain in Q10     */
+    const opus_int16    sLTP[],                     /* I    Re-whitened LTP state in Q0         */
+    opus_int32          sLTP_Q15[],                 /* O    LTP state matching scaled input     */
+    opus_int            subfr,                      /* I    Subframe number                     */
+    opus_int            nStatesDelayedDecision,     /* I    Number of del dec states            */
+    const opus_int      LTP_scale_Q14,              /* I    LTP state scaling                   */
+    const opus_int32    Gains_Q16[ MAX_NB_SUBFR ],  /* I                                        */
+    const opus_int      pitchL[ MAX_NB_SUBFR ],     /* I    Pitch lag                           */
+    const opus_int      signal_type,                /* I    Signal type                         */
+    const opus_int      decisionDelay               /* I    Decision delay                      */
+);
+
+/******************************************/
+/* Noise shape quantizer for one subframe */
+/******************************************/
+static OPUS_INLINE void silk_noise_shape_quantizer_del_dec(
+    silk_nsq_state      *NSQ,                   /* I/O  NSQ state                           */
+    NSQ_del_dec_struct  psDelDec[],             /* I/O  Delayed decision states             */
+    opus_int            signalType,             /* I    Signal type                         */
+    const opus_int32    x_Q10[],                /* I                                        */
+    opus_int8           pulses[],               /* O                                        */
+    opus_int16          xq[],                   /* O                                        */
+    opus_int32          sLTP_Q15[],             /* I/O  LTP filter state                    */
+    opus_int32          delayedGain_Q10[],      /* I/O  Gain delay buffer                   */
+    const opus_int16    a_Q12[],                /* I    Short term prediction coefs         */
+    const opus_int16    b_Q14[],                /* I    Long term prediction coefs          */
+    const opus_int16    AR_shp_Q13[],           /* I    Noise shaping coefs                 */
+    opus_int            lag,                    /* I    Pitch lag                           */
+    opus_int32          HarmShapeFIRPacked_Q14, /* I                                        */
+    opus_int            Tilt_Q14,               /* I    Spectral tilt                       */
+    opus_int32          LF_shp_Q14,             /* I                                        */
+    opus_int32          Gain_Q16,               /* I                                        */
+    opus_int            Lambda_Q10,             /* I                                        */
+    opus_int            offset_Q10,             /* I                                        */
+    opus_int            length,                 /* I    Input length                        */
+    opus_int            subfr,                  /* I    Subframe number                     */
+    opus_int            shapingLPCOrder,        /* I    Shaping LPC filter order            */
+    opus_int            predictLPCOrder,        /* I    Prediction filter order             */
+    opus_int            warping_Q16,            /* I                                        */
+    opus_int            nStatesDelayedDecision, /* I    Number of states in decision tree   */
+    opus_int            *smpl_buf_idx,          /* I/O  Index to newest samples in buffers  */
+    opus_int            decisionDelay,          /* I                                        */
+    int                 arch                    /* I                                        */
+);
+
+void silk_NSQ_del_dec_c(
+    const silk_encoder_state    *psEncC,                                      /* I    Encoder State                   */
+    silk_nsq_state              *NSQ,                                         /* I/O  NSQ state                       */
+    SideInfoIndices             *psIndices,                                   /* I/O  Quantization Indices            */
+    const opus_int16            x16[],                                        /* I    Input                           */
+    opus_int8                   pulses[],                                     /* O    Quantized pulse signal          */
+    const opus_int16            *PredCoef_Q12,                                /* I    Short term prediction coefs     */
+    const opus_int16            LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ],      /* I    Long term prediction coefs      */
+    const opus_int16            AR_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I    Noise shaping coefs             */
+    const opus_int              HarmShapeGain_Q14[ MAX_NB_SUBFR ],            /* I    Long term shaping coefs         */
+    const opus_int              Tilt_Q14[ MAX_NB_SUBFR ],                     /* I    Spectral tilt                   */
+    const opus_int32            LF_shp_Q14[ MAX_NB_SUBFR ],                   /* I    Low frequency shaping coefs     */
+    const opus_int32            Gains_Q16[ MAX_NB_SUBFR ],                    /* I    Quantization step sizes         */
+    const opus_int              pitchL[ MAX_NB_SUBFR ],                       /* I    Pitch lags                      */
+    const opus_int              Lambda_Q10,                                   /* I    Rate/distortion tradeoff        */
+    const opus_int              LTP_scale_Q14                                 /* I    LTP state scaling               */
+)
+{
+    opus_int            i, k, lag, start_idx, LSF_interpolation_flag, Winner_ind, subfr;
+    opus_int            last_smple_idx, smpl_buf_idx, decisionDelay;
+    const opus_int16    *A_Q12, *B_Q14, *AR_shp_Q13;
+    opus_int16          *pxq;
+    VARDECL( opus_int32, sLTP_Q15 );
+    VARDECL( opus_int16, sLTP );
+    opus_int32          HarmShapeFIRPacked_Q14;
+    opus_int            offset_Q10;
+    opus_int32          RDmin_Q10, Gain_Q10;
+    VARDECL( opus_int32, x_sc_Q10 );
+    VARDECL( opus_int32, delayedGain_Q10 );
+    VARDECL( NSQ_del_dec_struct, psDelDec );
+    NSQ_del_dec_struct  *psDD;
+    SAVE_STACK;
+
+    /* Set unvoiced lag to the previous one, overwrite later for voiced */
+    lag = NSQ->lagPrev;
+
+    silk_assert( NSQ->prev_gain_Q16 != 0 );
+
+    /* Initialize delayed decision states */
+    ALLOC( psDelDec, psEncC->nStatesDelayedDecision, NSQ_del_dec_struct );
+    silk_memset( psDelDec, 0, psEncC->nStatesDelayedDecision * sizeof( NSQ_del_dec_struct ) );
+    for( k = 0; k < psEncC->nStatesDelayedDecision; k++ ) {
+        psDD                 = &psDelDec[ k ];
+        psDD->Seed           = ( k + psIndices->Seed ) & 3;
+        psDD->SeedInit       = psDD->Seed;
+        psDD->RD_Q10         = 0;
+        psDD->LF_AR_Q14      = NSQ->sLF_AR_shp_Q14;
+        psDD->Diff_Q14       = NSQ->sDiff_shp_Q14;
+        psDD->Shape_Q14[ 0 ] = NSQ->sLTP_shp_Q14[ psEncC->ltp_mem_length - 1 ];
+        silk_memcpy( psDD->sLPC_Q14, NSQ->sLPC_Q14, NSQ_LPC_BUF_LENGTH * sizeof( opus_int32 ) );
+        silk_memcpy( psDD->sAR2_Q14, NSQ->sAR2_Q14, sizeof( NSQ->sAR2_Q14 ) );
+    }
+
+    offset_Q10   = silk_Quantization_Offsets_Q10[ psIndices->signalType >> 1 ][ psIndices->quantOffsetType ];
+    smpl_buf_idx = 0; /* index of oldest samples */
+
+    decisionDelay = silk_min_int( DECISION_DELAY, psEncC->subfr_length );
+
+    /* For voiced frames limit the decision delay to lower than the pitch lag */
+    if( psIndices->signalType == TYPE_VOICED ) {
+        for( k = 0; k < psEncC->nb_subfr; k++ ) {
+            decisionDelay = silk_min_int( decisionDelay, pitchL[ k ] - LTP_ORDER / 2 - 1 );
+        }
+    } else {
+        if( lag > 0 ) {
+            decisionDelay = silk_min_int( decisionDelay, lag - LTP_ORDER / 2 - 1 );
+        }
+    }
+
+    if( psIndices->NLSFInterpCoef_Q2 == 4 ) {
+        LSF_interpolation_flag = 0;
+    } else {
+        LSF_interpolation_flag = 1;
+    }
+
+    ALLOC( sLTP_Q15, psEncC->ltp_mem_length + psEncC->frame_length, opus_int32 );
+    ALLOC( sLTP, psEncC->ltp_mem_length + psEncC->frame_length, opus_int16 );
+    ALLOC( x_sc_Q10, psEncC->subfr_length, opus_int32 );
+    ALLOC( delayedGain_Q10, DECISION_DELAY, opus_int32 );
+    /* Set up pointers to start of sub frame */
+    pxq                   = &NSQ->xq[ psEncC->ltp_mem_length ];
+    NSQ->sLTP_shp_buf_idx = psEncC->ltp_mem_length;
+    NSQ->sLTP_buf_idx     = psEncC->ltp_mem_length;
+    subfr = 0;
+    for( k = 0; k < psEncC->nb_subfr; k++ ) {
+        A_Q12      = &PredCoef_Q12[ ( ( k >> 1 ) | ( 1 - LSF_interpolation_flag ) ) * MAX_LPC_ORDER ];
+        B_Q14      = &LTPCoef_Q14[ k * LTP_ORDER           ];
+        AR_shp_Q13 = &AR_Q13[     k * MAX_SHAPE_LPC_ORDER ];
+
+        /* Noise shape parameters */
+        silk_assert( HarmShapeGain_Q14[ k ] >= 0 );
+        HarmShapeFIRPacked_Q14  =                          silk_RSHIFT( HarmShapeGain_Q14[ k ], 2 );
+        HarmShapeFIRPacked_Q14 |= silk_LSHIFT( (opus_int32)silk_RSHIFT( HarmShapeGain_Q14[ k ], 1 ), 16 );
+
+        NSQ->rewhite_flag = 0;
+        if( psIndices->signalType == TYPE_VOICED ) {
+            /* Voiced */
+            lag = pitchL[ k ];
+
+            /* Re-whitening */
+            if( ( k & ( 3 - silk_LSHIFT( LSF_interpolation_flag, 1 ) ) ) == 0 ) {
+                if( k == 2 ) {
+                    /* RESET DELAYED DECISIONS */
+                    /* Find winner */
+                    RDmin_Q10 = psDelDec[ 0 ].RD_Q10;
+                    Winner_ind = 0;
+                    for( i = 1; i < psEncC->nStatesDelayedDecision; i++ ) {
+                        if( psDelDec[ i ].RD_Q10 < RDmin_Q10 ) {
+                            RDmin_Q10 = psDelDec[ i ].RD_Q10;
+                            Winner_ind = i;
+                        }
+                    }
+                    for( i = 0; i < psEncC->nStatesDelayedDecision; i++ ) {
+                        if( i != Winner_ind ) {
+                            psDelDec[ i ].RD_Q10 += ( silk_int32_MAX >> 4 );
+                            silk_assert( psDelDec[ i ].RD_Q10 >= 0 );
+                        }
+                    }
+
+                    /* Copy final part of signals from winner state to output and long-term filter states */
+                    psDD = &psDelDec[ Winner_ind ];
+                    last_smple_idx = smpl_buf_idx + decisionDelay;
+                    for( i = 0; i < decisionDelay; i++ ) {
+                        last_smple_idx = ( last_smple_idx - 1 ) % DECISION_DELAY;
+                        if( last_smple_idx < 0 ) last_smple_idx += DECISION_DELAY;
+                        pulses[   i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDD->Q_Q10[ last_smple_idx ], 10 );
+                        pxq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND(
+                            silk_SMULWW( psDD->Xq_Q14[ last_smple_idx ], Gains_Q16[ 1 ] ), 14 ) );
+                        NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - decisionDelay + i ] = psDD->Shape_Q14[ last_smple_idx ];
+                    }
+
+                    subfr = 0;
+                }
+
+                /* Rewhiten with new A coefs */
+                start_idx = psEncC->ltp_mem_length - lag - psEncC->predictLPCOrder - LTP_ORDER / 2;
+                celt_assert( start_idx > 0 );
+
+                silk_LPC_analysis_filter( &sLTP[ start_idx ], &NSQ->xq[ start_idx + k * psEncC->subfr_length ],
+                    A_Q12, psEncC->ltp_mem_length - start_idx, psEncC->predictLPCOrder, psEncC->arch );
+
+                NSQ->sLTP_buf_idx = psEncC->ltp_mem_length;
+                NSQ->rewhite_flag = 1;
+            }
+        }
+
+        silk_nsq_del_dec_scale_states( psEncC, NSQ, psDelDec, x16, x_sc_Q10, sLTP, sLTP_Q15, k,
+            psEncC->nStatesDelayedDecision, LTP_scale_Q14, Gains_Q16, pitchL, psIndices->signalType, decisionDelay );
+
+        silk_noise_shape_quantizer_del_dec( NSQ, psDelDec, psIndices->signalType, x_sc_Q10, pulses, pxq, sLTP_Q15,
+            delayedGain_Q10, A_Q12, B_Q14, AR_shp_Q13, lag, HarmShapeFIRPacked_Q14, Tilt_Q14[ k ], LF_shp_Q14[ k ],
+            Gains_Q16[ k ], Lambda_Q10, offset_Q10, psEncC->subfr_length, subfr++, psEncC->shapingLPCOrder,
+            psEncC->predictLPCOrder, psEncC->warping_Q16, psEncC->nStatesDelayedDecision, &smpl_buf_idx, decisionDelay, psEncC->arch );
+
+        x16    += psEncC->subfr_length;
+        pulses += psEncC->subfr_length;
+        pxq    += psEncC->subfr_length;
+    }
+
+    /* Find winner */
+    RDmin_Q10 = psDelDec[ 0 ].RD_Q10;
+    Winner_ind = 0;
+    for( k = 1; k < psEncC->nStatesDelayedDecision; k++ ) {
+        if( psDelDec[ k ].RD_Q10 < RDmin_Q10 ) {
+            RDmin_Q10 = psDelDec[ k ].RD_Q10;
+            Winner_ind = k;
+        }
+    }
+
+    /* Copy final part of signals from winner state to output and long-term filter states */
+    psDD = &psDelDec[ Winner_ind ];
+    psIndices->Seed = psDD->SeedInit;
+    last_smple_idx = smpl_buf_idx + decisionDelay;
+    Gain_Q10 = silk_RSHIFT32( Gains_Q16[ psEncC->nb_subfr - 1 ], 6 );
+    for( i = 0; i < decisionDelay; i++ ) {
+        last_smple_idx = ( last_smple_idx - 1 ) % DECISION_DELAY;
+        if( last_smple_idx < 0 ) last_smple_idx += DECISION_DELAY;
+
+        pulses[   i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDD->Q_Q10[ last_smple_idx ], 10 );
+        pxq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND(
+            silk_SMULWW( psDD->Xq_Q14[ last_smple_idx ], Gain_Q10 ), 8 ) );
+        NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - decisionDelay + i ] = psDD->Shape_Q14[ last_smple_idx ];
+    }
+    silk_memcpy( NSQ->sLPC_Q14, &psDD->sLPC_Q14[ psEncC->subfr_length ], NSQ_LPC_BUF_LENGTH * sizeof( opus_int32 ) );
+    silk_memcpy( NSQ->sAR2_Q14, psDD->sAR2_Q14, sizeof( psDD->sAR2_Q14 ) );
+
+    /* Update states */
+    NSQ->sLF_AR_shp_Q14 = psDD->LF_AR_Q14;
+    NSQ->sDiff_shp_Q14  = psDD->Diff_Q14;
+    NSQ->lagPrev        = pitchL[ psEncC->nb_subfr - 1 ];
+
+    /* Save quantized speech signal */
+    silk_memmove( NSQ->xq,           &NSQ->xq[           psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int16 ) );
+    silk_memmove( NSQ->sLTP_shp_Q14, &NSQ->sLTP_shp_Q14[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int32 ) );
+    RESTORE_STACK;
+}
+
+/******************************************/
+/* Noise shape quantizer for one subframe */
+/******************************************/
+#ifndef OVERRIDE_silk_noise_shape_quantizer_del_dec
+static OPUS_INLINE void silk_noise_shape_quantizer_del_dec(
+    silk_nsq_state      *NSQ,                   /* I/O  NSQ state                           */
+    NSQ_del_dec_struct  psDelDec[],             /* I/O  Delayed decision states             */
+    opus_int            signalType,             /* I    Signal type                         */
+    const opus_int32    x_Q10[],                /* I                                        */
+    opus_int8           pulses[],               /* O                                        */
+    opus_int16          xq[],                   /* O                                        */
+    opus_int32          sLTP_Q15[],             /* I/O  LTP filter state                    */
+    opus_int32          delayedGain_Q10[],      /* I/O  Gain delay buffer                   */
+    const opus_int16    a_Q12[],                /* I    Short term prediction coefs         */
+    const opus_int16    b_Q14[],                /* I    Long term prediction coefs          */
+    const opus_int16    AR_shp_Q13[],           /* I    Noise shaping coefs                 */
+    opus_int            lag,                    /* I    Pitch lag                           */
+    opus_int32          HarmShapeFIRPacked_Q14, /* I                                        */
+    opus_int            Tilt_Q14,               /* I    Spectral tilt                       */
+    opus_int32          LF_shp_Q14,             /* I                                        */
+    opus_int32          Gain_Q16,               /* I                                        */
+    opus_int            Lambda_Q10,             /* I                                        */
+    opus_int            offset_Q10,             /* I                                        */
+    opus_int            length,                 /* I    Input length                        */
+    opus_int            subfr,                  /* I    Subframe number                     */
+    opus_int            shapingLPCOrder,        /* I    Shaping LPC filter order            */
+    opus_int            predictLPCOrder,        /* I    Prediction filter order             */
+    opus_int            warping_Q16,            /* I                                        */
+    opus_int            nStatesDelayedDecision, /* I    Number of states in decision tree   */
+    opus_int            *smpl_buf_idx,          /* I/O  Index to newest samples in buffers  */
+    opus_int            decisionDelay,          /* I                                        */
+    int                 arch                    /* I                                        */
+)
+{
+    opus_int     i, j, k, Winner_ind, RDmin_ind, RDmax_ind, last_smple_idx;
+    opus_int32   Winner_rand_state;
+    opus_int32   LTP_pred_Q14, LPC_pred_Q14, n_AR_Q14, n_LTP_Q14;
+    opus_int32   n_LF_Q14, r_Q10, rr_Q10, rd1_Q10, rd2_Q10, RDmin_Q10, RDmax_Q10;
+    opus_int32   q1_Q0, q1_Q10, q2_Q10, exc_Q14, LPC_exc_Q14, xq_Q14, Gain_Q10;
+    opus_int32   tmp1, tmp2, sLF_AR_shp_Q14;
+    opus_int32   *pred_lag_ptr, *shp_lag_ptr, *psLPC_Q14;
+#ifdef silk_short_prediction_create_arch_coef
+    opus_int32   a_Q12_arch[MAX_LPC_ORDER];
+#endif
+
+    VARDECL( NSQ_sample_pair, psSampleState );
+    NSQ_del_dec_struct *psDD;
+    NSQ_sample_struct  *psSS;
+    SAVE_STACK;
+
+    celt_assert( nStatesDelayedDecision > 0 );
+    ALLOC( psSampleState, nStatesDelayedDecision, NSQ_sample_pair );
+
+    shp_lag_ptr  = &NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - lag + HARM_SHAPE_FIR_TAPS / 2 ];
+    pred_lag_ptr = &sLTP_Q15[ NSQ->sLTP_buf_idx - lag + LTP_ORDER / 2 ];
+    Gain_Q10     = silk_RSHIFT( Gain_Q16, 6 );
+
+#ifdef silk_short_prediction_create_arch_coef
+    silk_short_prediction_create_arch_coef(a_Q12_arch, a_Q12, predictLPCOrder);
+#endif
+
+    for( i = 0; i < length; i++ ) {
+        /* Perform common calculations used in all states */
+
+        /* Long-term prediction */
+        if( signalType == TYPE_VOICED ) {
+            /* Unrolled loop */
+            /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */
+            LTP_pred_Q14 = 2;
+            LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[  0 ], b_Q14[ 0 ] );
+            LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -1 ], b_Q14[ 1 ] );
+            LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -2 ], b_Q14[ 2 ] );
+            LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -3 ], b_Q14[ 3 ] );
+            LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -4 ], b_Q14[ 4 ] );
+            LTP_pred_Q14 = silk_LSHIFT( LTP_pred_Q14, 1 );                          /* Q13 -> Q14 */
+            pred_lag_ptr++;
+        } else {
+            LTP_pred_Q14 = 0;
+        }
+
+        /* Long-term shaping */
+        if( lag > 0 ) {
+            /* Symmetric, packed FIR coefficients */
+            n_LTP_Q14 = silk_SMULWB( silk_ADD_SAT32( shp_lag_ptr[ 0 ], shp_lag_ptr[ -2 ] ), HarmShapeFIRPacked_Q14 );
+            n_LTP_Q14 = silk_SMLAWT( n_LTP_Q14, shp_lag_ptr[ -1 ], HarmShapeFIRPacked_Q14 );
+            n_LTP_Q14 = silk_SUB_LSHIFT32( LTP_pred_Q14, n_LTP_Q14, 2 );            /* Q12 -> Q14 */
+            shp_lag_ptr++;
+        } else {
+            n_LTP_Q14 = 0;
+        }
+
+        for( k = 0; k < nStatesDelayedDecision; k++ ) {
+            /* Delayed decision state */
+            psDD = &psDelDec[ k ];
+
+            /* Sample state */
+            psSS = psSampleState[ k ];
+
+            /* Generate dither */
+            psDD->Seed = silk_RAND( psDD->Seed );
+
+            /* Pointer used in short term prediction and shaping */
+            psLPC_Q14 = &psDD->sLPC_Q14[ NSQ_LPC_BUF_LENGTH - 1 + i ];
+            /* Short-term prediction */
+            LPC_pred_Q14 = silk_noise_shape_quantizer_short_prediction(psLPC_Q14, a_Q12, a_Q12_arch, predictLPCOrder, arch);
+            LPC_pred_Q14 = silk_LSHIFT( LPC_pred_Q14, 4 );                              /* Q10 -> Q14 */
+
+            /* Noise shape feedback */
+            celt_assert( ( shapingLPCOrder & 1 ) == 0 );   /* check that order is even */
+            /* Output of lowpass section */
+            tmp2 = silk_SMLAWB( psDD->Diff_Q14, psDD->sAR2_Q14[ 0 ], warping_Q16 );
+            /* Output of allpass section */
+            tmp1 = silk_SMLAWB( psDD->sAR2_Q14[ 0 ], silk_SUB32_ovflw(psDD->sAR2_Q14[ 1 ], tmp2), warping_Q16 );
+            psDD->sAR2_Q14[ 0 ] = tmp2;
+            n_AR_Q14 = silk_RSHIFT( shapingLPCOrder, 1 );
+            n_AR_Q14 = silk_SMLAWB( n_AR_Q14, tmp2, AR_shp_Q13[ 0 ] );
+            /* Loop over allpass sections */
+            for( j = 2; j < shapingLPCOrder; j += 2 ) {
+                /* Output of allpass section */
+                tmp2 = silk_SMLAWB( psDD->sAR2_Q14[ j - 1 ], silk_SUB32_ovflw(psDD->sAR2_Q14[ j + 0 ], tmp1), warping_Q16 );
+                psDD->sAR2_Q14[ j - 1 ] = tmp1;
+                n_AR_Q14 = silk_SMLAWB( n_AR_Q14, tmp1, AR_shp_Q13[ j - 1 ] );
+                /* Output of allpass section */
+                tmp1 = silk_SMLAWB( psDD->sAR2_Q14[ j + 0 ], silk_SUB32_ovflw(psDD->sAR2_Q14[ j + 1 ], tmp2), warping_Q16 );
+                psDD->sAR2_Q14[ j + 0 ] = tmp2;
+                n_AR_Q14 = silk_SMLAWB( n_AR_Q14, tmp2, AR_shp_Q13[ j ] );
+            }
+            psDD->sAR2_Q14[ shapingLPCOrder - 1 ] = tmp1;
+            n_AR_Q14 = silk_SMLAWB( n_AR_Q14, tmp1, AR_shp_Q13[ shapingLPCOrder - 1 ] );
+
+            n_AR_Q14 = silk_LSHIFT( n_AR_Q14, 1 );                                      /* Q11 -> Q12 */
+            n_AR_Q14 = silk_SMLAWB( n_AR_Q14, psDD->LF_AR_Q14, Tilt_Q14 );              /* Q12 */
+            n_AR_Q14 = silk_LSHIFT( n_AR_Q14, 2 );                                      /* Q12 -> Q14 */
+
+            n_LF_Q14 = silk_SMULWB( psDD->Shape_Q14[ *smpl_buf_idx ], LF_shp_Q14 );     /* Q12 */
+            n_LF_Q14 = silk_SMLAWT( n_LF_Q14, psDD->LF_AR_Q14, LF_shp_Q14 );            /* Q12 */
+            n_LF_Q14 = silk_LSHIFT( n_LF_Q14, 2 );                                      /* Q12 -> Q14 */
+
+            /* Input minus prediction plus noise feedback                       */
+            /* r = x[ i ] - LTP_pred - LPC_pred + n_AR + n_Tilt + n_LF + n_LTP  */
+            tmp1 = silk_ADD_SAT32( n_AR_Q14, n_LF_Q14 );                                /* Q14 */
+            tmp2 = silk_ADD32_ovflw( n_LTP_Q14, LPC_pred_Q14 );                         /* Q13 */
+            tmp1 = silk_SUB_SAT32( tmp2, tmp1 );                                        /* Q13 */
+            tmp1 = silk_RSHIFT_ROUND( tmp1, 4 );                                        /* Q10 */
+
+            r_Q10 = silk_SUB32( x_Q10[ i ], tmp1 );                                     /* residual error Q10 */
+
+            /* Flip sign depending on dither */
+            if ( psDD->Seed < 0 ) {
+                r_Q10 = -r_Q10;
+            }
+            r_Q10 = silk_LIMIT_32( r_Q10, -(31 << 10), 30 << 10 );
+
+            /* Find two quantization level candidates and measure their rate-distortion */
+            q1_Q10 = silk_SUB32( r_Q10, offset_Q10 );
+            q1_Q0 = silk_RSHIFT( q1_Q10, 10 );
+            if (Lambda_Q10 > 2048) {
+                /* For aggressive RDO, the bias becomes more than one pulse. */
+                int rdo_offset = Lambda_Q10/2 - 512;
+                if (q1_Q10 > rdo_offset) {
+                    q1_Q0 = silk_RSHIFT( q1_Q10 - rdo_offset, 10 );
+                } else if (q1_Q10 < -rdo_offset) {
+                    q1_Q0 = silk_RSHIFT( q1_Q10 + rdo_offset, 10 );
+                } else if (q1_Q10 < 0) {
+                    q1_Q0 = -1;
+                } else {
+                    q1_Q0 = 0;
+                }
+            }
+            if( q1_Q0 > 0 ) {
+                q1_Q10  = silk_SUB32( silk_LSHIFT( q1_Q0, 10 ), QUANT_LEVEL_ADJUST_Q10 );
+                q1_Q10  = silk_ADD32( q1_Q10, offset_Q10 );
+                q2_Q10  = silk_ADD32( q1_Q10, 1024 );
+                rd1_Q10 = silk_SMULBB( q1_Q10, Lambda_Q10 );
+                rd2_Q10 = silk_SMULBB( q2_Q10, Lambda_Q10 );
+            } else if( q1_Q0 == 0 ) {
+                q1_Q10  = offset_Q10;
+                q2_Q10  = silk_ADD32( q1_Q10, 1024 - QUANT_LEVEL_ADJUST_Q10 );
+                rd1_Q10 = silk_SMULBB( q1_Q10, Lambda_Q10 );
+                rd2_Q10 = silk_SMULBB( q2_Q10, Lambda_Q10 );
+            } else if( q1_Q0 == -1 ) {
+                q2_Q10  = offset_Q10;
+                q1_Q10  = silk_SUB32( q2_Q10, 1024 - QUANT_LEVEL_ADJUST_Q10 );
+                rd1_Q10 = silk_SMULBB( -q1_Q10, Lambda_Q10 );
+                rd2_Q10 = silk_SMULBB(  q2_Q10, Lambda_Q10 );
+            } else {            /* q1_Q0 < -1 */
+                q1_Q10  = silk_ADD32( silk_LSHIFT( q1_Q0, 10 ), QUANT_LEVEL_ADJUST_Q10 );
+                q1_Q10  = silk_ADD32( q1_Q10, offset_Q10 );
+                q2_Q10  = silk_ADD32( q1_Q10, 1024 );
+                rd1_Q10 = silk_SMULBB( -q1_Q10, Lambda_Q10 );
+                rd2_Q10 = silk_SMULBB( -q2_Q10, Lambda_Q10 );
+            }
+            rr_Q10  = silk_SUB32( r_Q10, q1_Q10 );
+            rd1_Q10 = silk_RSHIFT( silk_SMLABB( rd1_Q10, rr_Q10, rr_Q10 ), 10 );
+            rr_Q10  = silk_SUB32( r_Q10, q2_Q10 );
+            rd2_Q10 = silk_RSHIFT( silk_SMLABB( rd2_Q10, rr_Q10, rr_Q10 ), 10 );
+
+            if( rd1_Q10 < rd2_Q10 ) {
+                psSS[ 0 ].RD_Q10 = silk_ADD32( psDD->RD_Q10, rd1_Q10 );
+                psSS[ 1 ].RD_Q10 = silk_ADD32( psDD->RD_Q10, rd2_Q10 );
+                psSS[ 0 ].Q_Q10  = q1_Q10;
+                psSS[ 1 ].Q_Q10  = q2_Q10;
+            } else {
+                psSS[ 0 ].RD_Q10 = silk_ADD32( psDD->RD_Q10, rd2_Q10 );
+                psSS[ 1 ].RD_Q10 = silk_ADD32( psDD->RD_Q10, rd1_Q10 );
+                psSS[ 0 ].Q_Q10  = q2_Q10;
+                psSS[ 1 ].Q_Q10  = q1_Q10;
+            }
+
+            /* Update states for best quantization */
+
+            /* Quantized excitation */
+            exc_Q14 = silk_LSHIFT32( psSS[ 0 ].Q_Q10, 4 );
+            if ( psDD->Seed < 0 ) {
+                exc_Q14 = -exc_Q14;
+            }
+
+            /* Add predictions */
+            LPC_exc_Q14 = silk_ADD32( exc_Q14, LTP_pred_Q14 );
+            xq_Q14      = silk_ADD32_ovflw( LPC_exc_Q14, LPC_pred_Q14 );
+
+            /* Update states */
+            psSS[ 0 ].Diff_Q14     = silk_SUB32_ovflw( xq_Q14, silk_LSHIFT32( x_Q10[ i ], 4 ) );
+            sLF_AR_shp_Q14         = silk_SUB32_ovflw( psSS[ 0 ].Diff_Q14, n_AR_Q14 );
+            psSS[ 0 ].sLTP_shp_Q14 = silk_SUB_SAT32( sLF_AR_shp_Q14, n_LF_Q14 );
+            psSS[ 0 ].LF_AR_Q14    = sLF_AR_shp_Q14;
+            psSS[ 0 ].LPC_exc_Q14  = LPC_exc_Q14;
+            psSS[ 0 ].xq_Q14       = xq_Q14;
+
+            /* Update states for second best quantization */
+
+            /* Quantized excitation */
+            exc_Q14 = silk_LSHIFT32( psSS[ 1 ].Q_Q10, 4 );
+            if ( psDD->Seed < 0 ) {
+                exc_Q14 = -exc_Q14;
+            }
+
+            /* Add predictions */
+            LPC_exc_Q14 = silk_ADD32( exc_Q14, LTP_pred_Q14 );
+            xq_Q14      = silk_ADD32_ovflw( LPC_exc_Q14, LPC_pred_Q14 );
+
+            /* Update states */
+            psSS[ 1 ].Diff_Q14     = silk_SUB32_ovflw( xq_Q14, silk_LSHIFT32( x_Q10[ i ], 4 ) );
+            sLF_AR_shp_Q14         = silk_SUB32_ovflw( psSS[ 1 ].Diff_Q14, n_AR_Q14 );
+            psSS[ 1 ].sLTP_shp_Q14 = silk_SUB_SAT32( sLF_AR_shp_Q14, n_LF_Q14 );
+            psSS[ 1 ].LF_AR_Q14    = sLF_AR_shp_Q14;
+            psSS[ 1 ].LPC_exc_Q14  = LPC_exc_Q14;
+            psSS[ 1 ].xq_Q14       = xq_Q14;
+        }
+
+        *smpl_buf_idx  = ( *smpl_buf_idx - 1 ) % DECISION_DELAY;
+        if( *smpl_buf_idx < 0 ) *smpl_buf_idx += DECISION_DELAY;
+        last_smple_idx = ( *smpl_buf_idx + decisionDelay ) % DECISION_DELAY;
+
+        /* Find winner */
+        RDmin_Q10 = psSampleState[ 0 ][ 0 ].RD_Q10;
+        Winner_ind = 0;
+        for( k = 1; k < nStatesDelayedDecision; k++ ) {
+            if( psSampleState[ k ][ 0 ].RD_Q10 < RDmin_Q10 ) {
+                RDmin_Q10  = psSampleState[ k ][ 0 ].RD_Q10;
+                Winner_ind = k;
+            }
+        }
+
+        /* Increase RD values of expired states */
+        Winner_rand_state = psDelDec[ Winner_ind ].RandState[ last_smple_idx ];
+        for( k = 0; k < nStatesDelayedDecision; k++ ) {
+            if( psDelDec[ k ].RandState[ last_smple_idx ] != Winner_rand_state ) {
+                psSampleState[ k ][ 0 ].RD_Q10 = silk_ADD32( psSampleState[ k ][ 0 ].RD_Q10, silk_int32_MAX >> 4 );
+                psSampleState[ k ][ 1 ].RD_Q10 = silk_ADD32( psSampleState[ k ][ 1 ].RD_Q10, silk_int32_MAX >> 4 );
+                silk_assert( psSampleState[ k ][ 0 ].RD_Q10 >= 0 );
+            }
+        }
+
+        /* Find worst in first set and best in second set */
+        RDmax_Q10  = psSampleState[ 0 ][ 0 ].RD_Q10;
+        RDmin_Q10  = psSampleState[ 0 ][ 1 ].RD_Q10;
+        RDmax_ind = 0;
+        RDmin_ind = 0;
+        for( k = 1; k < nStatesDelayedDecision; k++ ) {
+            /* find worst in first set */
+            if( psSampleState[ k ][ 0 ].RD_Q10 > RDmax_Q10 ) {
+                RDmax_Q10  = psSampleState[ k ][ 0 ].RD_Q10;
+                RDmax_ind = k;
+            }
+            /* find best in second set */
+            if( psSampleState[ k ][ 1 ].RD_Q10 < RDmin_Q10 ) {
+                RDmin_Q10  = psSampleState[ k ][ 1 ].RD_Q10;
+                RDmin_ind = k;
+            }
+        }
+
+        /* Replace a state if best from second set outperforms worst in first set */
+        if( RDmin_Q10 < RDmax_Q10 ) {
+            silk_memcpy( ( (opus_int32 *)&psDelDec[ RDmax_ind ] ) + i,
+                         ( (opus_int32 *)&psDelDec[ RDmin_ind ] ) + i, sizeof( NSQ_del_dec_struct ) - i * sizeof( opus_int32) );
+            silk_memcpy( &psSampleState[ RDmax_ind ][ 0 ], &psSampleState[ RDmin_ind ][ 1 ], sizeof( NSQ_sample_struct ) );
+        }
+
+        /* Write samples from winner to output and long-term filter states */
+        psDD = &psDelDec[ Winner_ind ];
+        if( subfr > 0 || i >= decisionDelay ) {
+            pulses[  i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDD->Q_Q10[ last_smple_idx ], 10 );
+            xq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND(
+                silk_SMULWW( psDD->Xq_Q14[ last_smple_idx ], delayedGain_Q10[ last_smple_idx ] ), 8 ) );
+            NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - decisionDelay ] = psDD->Shape_Q14[ last_smple_idx ];
+            sLTP_Q15[          NSQ->sLTP_buf_idx     - decisionDelay ] = psDD->Pred_Q15[  last_smple_idx ];
+        }
+        NSQ->sLTP_shp_buf_idx++;
+        NSQ->sLTP_buf_idx++;
+
+        /* Update states */
+        for( k = 0; k < nStatesDelayedDecision; k++ ) {
+            psDD                                     = &psDelDec[ k ];
+            psSS                                     = &psSampleState[ k ][ 0 ];
+            psDD->LF_AR_Q14                          = psSS->LF_AR_Q14;
+            psDD->Diff_Q14                           = psSS->Diff_Q14;
+            psDD->sLPC_Q14[ NSQ_LPC_BUF_LENGTH + i ] = psSS->xq_Q14;
+            psDD->Xq_Q14[    *smpl_buf_idx ]         = psSS->xq_Q14;
+            psDD->Q_Q10[     *smpl_buf_idx ]         = psSS->Q_Q10;
+            psDD->Pred_Q15[  *smpl_buf_idx ]         = silk_LSHIFT32( psSS->LPC_exc_Q14, 1 );
+            psDD->Shape_Q14[ *smpl_buf_idx ]         = psSS->sLTP_shp_Q14;
+            psDD->Seed                               = silk_ADD32_ovflw( psDD->Seed, silk_RSHIFT_ROUND( psSS->Q_Q10, 10 ) );
+            psDD->RandState[ *smpl_buf_idx ]         = psDD->Seed;
+            psDD->RD_Q10                             = psSS->RD_Q10;
+        }
+        delayedGain_Q10[     *smpl_buf_idx ]         = Gain_Q10;
+    }
+    /* Update LPC states */
+    for( k = 0; k < nStatesDelayedDecision; k++ ) {
+        psDD = &psDelDec[ k ];
+        silk_memcpy( psDD->sLPC_Q14, &psDD->sLPC_Q14[ length ], NSQ_LPC_BUF_LENGTH * sizeof( opus_int32 ) );
+    }
+    RESTORE_STACK;
+}
+#endif /* OVERRIDE_silk_noise_shape_quantizer_del_dec */
+
+static OPUS_INLINE void silk_nsq_del_dec_scale_states(
+    const silk_encoder_state *psEncC,               /* I    Encoder State                       */
+    silk_nsq_state      *NSQ,                       /* I/O  NSQ state                           */
+    NSQ_del_dec_struct  psDelDec[],                 /* I/O  Delayed decision states             */
+    const opus_int16    x16[],                      /* I    Input                               */
+    opus_int32          x_sc_Q10[],                 /* O    Input scaled with 1/Gain in Q10     */
+    const opus_int16    sLTP[],                     /* I    Re-whitened LTP state in Q0         */
+    opus_int32          sLTP_Q15[],                 /* O    LTP state matching scaled input     */
+    opus_int            subfr,                      /* I    Subframe number                     */
+    opus_int            nStatesDelayedDecision,     /* I    Number of del dec states            */
+    const opus_int      LTP_scale_Q14,              /* I    LTP state scaling                   */
+    const opus_int32    Gains_Q16[ MAX_NB_SUBFR ],  /* I                                        */
+    const opus_int      pitchL[ MAX_NB_SUBFR ],     /* I    Pitch lag                           */
+    const opus_int      signal_type,                /* I    Signal type                         */
+    const opus_int      decisionDelay               /* I    Decision delay                      */
+)
+{
+    opus_int            i, k, lag;
+    opus_int32          gain_adj_Q16, inv_gain_Q31, inv_gain_Q26;
+    NSQ_del_dec_struct  *psDD;
+
+    lag          = pitchL[ subfr ];
+    inv_gain_Q31 = silk_INVERSE32_varQ( silk_max( Gains_Q16[ subfr ], 1 ), 47 );
+    silk_assert( inv_gain_Q31 != 0 );
+
+    /* Scale input */
+    inv_gain_Q26 = silk_RSHIFT_ROUND( inv_gain_Q31, 5 );
+    for( i = 0; i < psEncC->subfr_length; i++ ) {
+        x_sc_Q10[ i ] = silk_SMULWW( x16[ i ], inv_gain_Q26 );
+    }
+
+    /* After rewhitening the LTP state is un-scaled, so scale with inv_gain_Q16 */
+    if( NSQ->rewhite_flag ) {
+        if( subfr == 0 ) {
+            /* Do LTP downscaling */
+            inv_gain_Q31 = silk_LSHIFT( silk_SMULWB( inv_gain_Q31, LTP_scale_Q14 ), 2 );
+        }
+        for( i = NSQ->sLTP_buf_idx - lag - LTP_ORDER / 2; i < NSQ->sLTP_buf_idx; i++ ) {
+            silk_assert( i < MAX_FRAME_LENGTH );
+            sLTP_Q15[ i ] = silk_SMULWB( inv_gain_Q31, sLTP[ i ] );
+        }
+    }
+
+    /* Adjust for changing gain */
+    if( Gains_Q16[ subfr ] != NSQ->prev_gain_Q16 ) {
+        gain_adj_Q16 =  silk_DIV32_varQ( NSQ->prev_gain_Q16, Gains_Q16[ subfr ], 16 );
+
+        /* Scale long-term shaping state */
+        for( i = NSQ->sLTP_shp_buf_idx - psEncC->ltp_mem_length; i < NSQ->sLTP_shp_buf_idx; i++ ) {
+            NSQ->sLTP_shp_Q14[ i ] = silk_SMULWW( gain_adj_Q16, NSQ->sLTP_shp_Q14[ i ] );
+        }
+
+        /* Scale long-term prediction state */
+        if( signal_type == TYPE_VOICED && NSQ->rewhite_flag == 0 ) {
+            for( i = NSQ->sLTP_buf_idx - lag - LTP_ORDER / 2; i < NSQ->sLTP_buf_idx - decisionDelay; i++ ) {
+                sLTP_Q15[ i ] = silk_SMULWW( gain_adj_Q16, sLTP_Q15[ i ] );
+            }
+        }
+
+        for( k = 0; k < nStatesDelayedDecision; k++ ) {
+            psDD = &psDelDec[ k ];
+
+            /* Scale scalar states */
+            psDD->LF_AR_Q14 = silk_SMULWW( gain_adj_Q16, psDD->LF_AR_Q14 );
+            psDD->Diff_Q14 = silk_SMULWW( gain_adj_Q16, psDD->Diff_Q14 );
+
+            /* Scale short-term prediction and shaping states */
+            for( i = 0; i < NSQ_LPC_BUF_LENGTH; i++ ) {
+                psDD->sLPC_Q14[ i ] = silk_SMULWW( gain_adj_Q16, psDD->sLPC_Q14[ i ] );
+            }
+            for( i = 0; i < MAX_SHAPE_LPC_ORDER; i++ ) {
+                psDD->sAR2_Q14[ i ] = silk_SMULWW( gain_adj_Q16, psDD->sAR2_Q14[ i ] );
+            }
+            for( i = 0; i < DECISION_DELAY; i++ ) {
+                psDD->Pred_Q15[  i ] = silk_SMULWW( gain_adj_Q16, psDD->Pred_Q15[  i ] );
+                psDD->Shape_Q14[ i ] = silk_SMULWW( gain_adj_Q16, psDD->Shape_Q14[ i ] );
+            }
+        }
+
+        /* Save inverse gain */
+        NSQ->prev_gain_Q16 = Gains_Q16[ subfr ];
+    }
+}
--- a/lib/libopus/silk/PLC.c
+++ b/lib/libopus/silk/PLC.c
@ -0,0 +1,493 @@
+/***********************************************************************
+Copyright (c) 2006-2011, Skype Limited. All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+- Redistributions of source code must retain the above copyright notice,
+this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+***********************************************************************/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "main.h"
+#include "stack_alloc.h"
+#include "PLC.h"
+
+#ifdef ENABLE_DEEP_PLC
+#include "lpcnet.h"
+#endif
+
+#define NB_ATT 2
+static const opus_int16 HARM_ATT_Q15[NB_ATT]              = { 32440, 31130 }; /* 0.99, 0.95 */
+static const opus_int16 PLC_RAND_ATTENUATE_V_Q15[NB_ATT]  = { 31130, 26214 }; /* 0.95, 0.8 */
+static const opus_int16 PLC_RAND_ATTENUATE_UV_Q15[NB_ATT] = { 32440, 29491 }; /* 0.99, 0.9 */
+
+static OPUS_INLINE void silk_PLC_update(
+    silk_decoder_state                  *psDec,             /* I/O Decoder state        */
+    silk_decoder_control                *psDecCtrl          /* I/O Decoder control      */
+);
+
+static OPUS_INLINE void silk_PLC_conceal(
+    silk_decoder_state                  *psDec,             /* I/O Decoder state        */
+    silk_decoder_control                *psDecCtrl,         /* I/O Decoder control      */
+    opus_int16                          frame[],            /* O LPC residual signal    */
+#ifdef ENABLE_DEEP_PLC
+    LPCNetPLCState                      *lpcnet,
+#endif
+    int                                 arch                /* I  Run-time architecture */
+);
+
+
+void silk_PLC_Reset(
+    silk_decoder_state                  *psDec              /* I/O Decoder state        */
+)
+{
+    psDec->sPLC.pitchL_Q8 = silk_LSHIFT( psDec->frame_length, 8 - 1 );
+    psDec->sPLC.prevGain_Q16[ 0 ] = SILK_FIX_CONST( 1, 16 );
+    psDec->sPLC.prevGain_Q16[ 1 ] = SILK_FIX_CONST( 1, 16 );
+    psDec->sPLC.subfr_length = 20;
+    psDec->sPLC.nb_subfr = 2;
+}
+
+void silk_PLC(
+    silk_decoder_state                  *psDec,             /* I/O Decoder state        */
+    silk_decoder_control                *psDecCtrl,         /* I/O Decoder control      */
+    opus_int16                          frame[],            /* I/O  signal              */
+    opus_int                            lost,               /* I Loss flag              */
+#ifdef ENABLE_DEEP_PLC
+    LPCNetPLCState                      *lpcnet,
+#endif
+    int                                 arch                /* I Run-time architecture  */
+)
+{
+    /* PLC control function */
+    if( psDec->fs_kHz != psDec->sPLC.fs_kHz ) {
+        silk_PLC_Reset( psDec );
+        psDec->sPLC.fs_kHz = psDec->fs_kHz;
+    }
+
+    if( lost ) {
+        /****************************/
+        /* Generate Signal          */
+        /****************************/
+        silk_PLC_conceal( psDec, psDecCtrl, frame,
+#ifdef ENABLE_DEEP_PLC
+            lpcnet,
+#endif
+            arch );
+
+        psDec->lossCnt++;
+    } else {
+        /****************************/
+        /* Update state             */
+        /****************************/
+        silk_PLC_update( psDec, psDecCtrl );
+#ifdef ENABLE_DEEP_PLC
+        if ( lpcnet != NULL && psDec->sPLC.fs_kHz == 16 ) {
+            int k;
+            for( k = 0; k < psDec->nb_subfr; k += 2 ) {
+                lpcnet_plc_update( lpcnet, frame + k * psDec->subfr_length );
+            }
+        }
+#endif
+    }
+}
+
+/**************************************************/
+/* Update state of PLC                            */
+/**************************************************/
+static OPUS_INLINE void silk_PLC_update(
+    silk_decoder_state                  *psDec,             /* I/O Decoder state        */
+    silk_decoder_control                *psDecCtrl          /* I/O Decoder control      */
+)
+{
+    opus_int32 LTP_Gain_Q14, temp_LTP_Gain_Q14;
+    opus_int   i, j;
+    silk_PLC_struct *psPLC;
+
+    psPLC = &psDec->sPLC;
+
+    /* Update parameters used in case of packet loss */
+    psDec->prevSignalType = psDec->indices.signalType;
+    LTP_Gain_Q14 = 0;
+    if( psDec->indices.signalType == TYPE_VOICED ) {
+        /* Find the parameters for the last subframe which contains a pitch pulse */
+        for( j = 0; j * psDec->subfr_length < psDecCtrl->pitchL[ psDec->nb_subfr - 1 ]; j++ ) {
+            if( j == psDec->nb_subfr ) {
+                break;
+            }
+            temp_LTP_Gain_Q14 = 0;
+            for( i = 0; i < LTP_ORDER; i++ ) {
+                temp_LTP_Gain_Q14 += psDecCtrl->LTPCoef_Q14[ ( psDec->nb_subfr - 1 - j ) * LTP_ORDER  + i ];
+            }
+            if( temp_LTP_Gain_Q14 > LTP_Gain_Q14 ) {
+                LTP_Gain_Q14 = temp_LTP_Gain_Q14;
+                silk_memcpy( psPLC->LTPCoef_Q14,
+                    &psDecCtrl->LTPCoef_Q14[ silk_SMULBB( psDec->nb_subfr - 1 - j, LTP_ORDER ) ],
+                    LTP_ORDER * sizeof( opus_int16 ) );
+
+                psPLC->pitchL_Q8 = silk_LSHIFT( psDecCtrl->pitchL[ psDec->nb_subfr - 1 - j ], 8 );
+            }
+        }
+
+        silk_memset( psPLC->LTPCoef_Q14, 0, LTP_ORDER * sizeof( opus_int16 ) );
+        psPLC->LTPCoef_Q14[ LTP_ORDER / 2 ] = LTP_Gain_Q14;
+
+        /* Limit LT coefs */
+        if( LTP_Gain_Q14 < V_PITCH_GAIN_START_MIN_Q14 ) {
+            opus_int   scale_Q10;
+            opus_int32 tmp;
+
+            tmp = silk_LSHIFT( V_PITCH_GAIN_START_MIN_Q14, 10 );
+            scale_Q10 = silk_DIV32( tmp, silk_max( LTP_Gain_Q14, 1 ) );
+            for( i = 0; i < LTP_ORDER; i++ ) {
+                psPLC->LTPCoef_Q14[ i ] = silk_RSHIFT( silk_SMULBB( psPLC->LTPCoef_Q14[ i ], scale_Q10 ), 10 );
+            }
+        } else if( LTP_Gain_Q14 > V_PITCH_GAIN_START_MAX_Q14 ) {
+            opus_int   scale_Q14;
+            opus_int32 tmp;
+
+            tmp = silk_LSHIFT( V_PITCH_GAIN_START_MAX_Q14, 14 );
+            scale_Q14 = silk_DIV32( tmp, silk_max( LTP_Gain_Q14, 1 ) );
+            for( i = 0; i < LTP_ORDER; i++ ) {
+                psPLC->LTPCoef_Q14[ i ] = silk_RSHIFT( silk_SMULBB( psPLC->LTPCoef_Q14[ i ], scale_Q14 ), 14 );
+            }
+        }
+    } else {
+        psPLC->pitchL_Q8 = silk_LSHIFT( silk_SMULBB( psDec->fs_kHz, 18 ), 8 );
+        silk_memset( psPLC->LTPCoef_Q14, 0, LTP_ORDER * sizeof( opus_int16 ));
+    }
+
+    /* Save LPC coefficients */
+    silk_memcpy( psPLC->prevLPC_Q12, psDecCtrl->PredCoef_Q12[ 1 ], psDec->LPC_order * sizeof( opus_int16 ) );
+    psPLC->prevLTP_scale_Q14 = psDecCtrl->LTP_scale_Q14;
+
+    /* Save last two gains */
+    silk_memcpy( psPLC->prevGain_Q16, &psDecCtrl->Gains_Q16[ psDec->nb_subfr - 2 ], 2 * sizeof( opus_int32 ) );
+
+    psPLC->subfr_length = psDec->subfr_length;
+    psPLC->nb_subfr = psDec->nb_subfr;
+}
+
+static OPUS_INLINE void silk_PLC_energy(opus_int32 *energy1, opus_int *shift1, opus_int32 *energy2, opus_int *shift2,
+      const opus_int32 *exc_Q14, const opus_int32 *prevGain_Q10, int subfr_length, int nb_subfr)
+{
+    int i, k;
+    VARDECL( opus_int16, exc_buf );
+    opus_int16 *exc_buf_ptr;
+    SAVE_STACK;
+    ALLOC( exc_buf, 2*subfr_length, opus_int16 );
+    /* Find random noise component */
+    /* Scale previous excitation signal */
+    exc_buf_ptr = exc_buf;
+    for( k = 0; k < 2; k++ ) {
+        for( i = 0; i < subfr_length; i++ ) {
+            exc_buf_ptr[ i ] = (opus_int16)silk_SAT16( silk_RSHIFT(
+                silk_SMULWW( exc_Q14[ i + ( k + nb_subfr - 2 ) * subfr_length ], prevGain_Q10[ k ] ), 8 ) );
+        }
+        exc_buf_ptr += subfr_length;
+    }
+    /* Find the subframe with lowest energy of the last two and use that as random noise generator */
+    silk_sum_sqr_shift( energy1, shift1, exc_buf,                  subfr_length );
+    silk_sum_sqr_shift( energy2, shift2, &exc_buf[ subfr_length ], subfr_length );
+    RESTORE_STACK;
+}
+
+static OPUS_INLINE void silk_PLC_conceal(
+    silk_decoder_state                  *psDec,             /* I/O Decoder state        */
+    silk_decoder_control                *psDecCtrl,         /* I/O Decoder control      */
+    opus_int16                          frame[],            /* O LPC residual signal    */
+#ifdef ENABLE_DEEP_PLC
+    LPCNetPLCState                      *lpcnet,
+#endif
+    int                                 arch                /* I Run-time architecture  */
+)
+{
+    opus_int   i, j, k;
+    opus_int   lag, idx, sLTP_buf_idx, shift1, shift2;
+    opus_int32 rand_seed, harm_Gain_Q15, rand_Gain_Q15, inv_gain_Q30;
+    opus_int32 energy1, energy2, *rand_ptr, *pred_lag_ptr;
+    opus_int32 LPC_pred_Q10, LTP_pred_Q12;
+    opus_int16 rand_scale_Q14;
+    opus_int16 *B_Q14;
+    opus_int32 *sLPC_Q14_ptr;
+    opus_int16 A_Q12[ MAX_LPC_ORDER ];
+#ifdef SMALL_FOOTPRINT
+    opus_int16 *sLTP;
+#else
+    VARDECL( opus_int16, sLTP );
+#endif
+    VARDECL( opus_int32, sLTP_Q14 );
+    silk_PLC_struct *psPLC = &psDec->sPLC;
+    opus_int32 prevGain_Q10[2];
+    SAVE_STACK;
+
+    ALLOC( sLTP_Q14, psDec->ltp_mem_length + psDec->frame_length, opus_int32 );
+#ifdef SMALL_FOOTPRINT
+    /* Ugly hack that breaks aliasing rules to save stack: put sLTP at the very end of sLTP_Q14. */
+    sLTP = ((opus_int16*)&sLTP_Q14[psDec->ltp_mem_length + psDec->frame_length])-psDec->ltp_mem_length;
+#else
+    ALLOC( sLTP, psDec->ltp_mem_length, opus_int16 );
+#endif
+
+    prevGain_Q10[0] = silk_RSHIFT( psPLC->prevGain_Q16[ 0 ], 6);
+    prevGain_Q10[1] = silk_RSHIFT( psPLC->prevGain_Q16[ 1 ], 6);
+
+    if( psDec->first_frame_after_reset ) {
+       silk_memset( psPLC->prevLPC_Q12, 0, sizeof( psPLC->prevLPC_Q12 ) );
+    }
+
+    silk_PLC_energy(&energy1, &shift1, &energy2, &shift2, psDec->exc_Q14, prevGain_Q10, psDec->subfr_length, psDec->nb_subfr);
+
+    if( silk_RSHIFT( energy1, shift2 ) < silk_RSHIFT( energy2, shift1 ) ) {
+        /* First sub-frame has lowest energy */
+        rand_ptr = &psDec->exc_Q14[ silk_max_int( 0, ( psPLC->nb_subfr - 1 ) * psPLC->subfr_length - RAND_BUF_SIZE ) ];
+    } else {
+        /* Second sub-frame has lowest energy */
+        rand_ptr = &psDec->exc_Q14[ silk_max_int( 0, psPLC->nb_subfr * psPLC->subfr_length - RAND_BUF_SIZE ) ];
+    }
+
+    /* Set up Gain to random noise component */
+    B_Q14          = psPLC->LTPCoef_Q14;
+    rand_scale_Q14 = psPLC->randScale_Q14;
+
+    /* Set up attenuation gains */
+    harm_Gain_Q15 = HARM_ATT_Q15[ silk_min_int( NB_ATT - 1, psDec->lossCnt ) ];
+    if( psDec->prevSignalType == TYPE_VOICED ) {
+        rand_Gain_Q15 = PLC_RAND_ATTENUATE_V_Q15[  silk_min_int( NB_ATT - 1, psDec->lossCnt ) ];
+    } else {
+        rand_Gain_Q15 = PLC_RAND_ATTENUATE_UV_Q15[ silk_min_int( NB_ATT - 1, psDec->lossCnt ) ];
+    }
+
+    /* LPC concealment. Apply BWE to previous LPC */
+    silk_bwexpander( psPLC->prevLPC_Q12, psDec->LPC_order, SILK_FIX_CONST( BWE_COEF, 16 ) );
+
+    /* Preload LPC coefficients to array on stack. Gives small performance gain */
+    silk_memcpy( A_Q12, psPLC->prevLPC_Q12, psDec->LPC_order * sizeof( opus_int16 ) );
+
+    /* First Lost frame */
+    if( psDec->lossCnt == 0 ) {
+        rand_scale_Q14 = 1 << 14;
+
+        /* Reduce random noise Gain for voiced frames */
+        if( psDec->prevSignalType == TYPE_VOICED ) {
+            for( i = 0; i < LTP_ORDER; i++ ) {
+                rand_scale_Q14 -= B_Q14[ i ];
+            }
+            rand_scale_Q14 = silk_max_16( 3277, rand_scale_Q14 ); /* 0.2 */
+            rand_scale_Q14 = (opus_int16)silk_RSHIFT( silk_SMULBB( rand_scale_Q14, psPLC->prevLTP_scale_Q14 ), 14 );
+        } else {
+            /* Reduce random noise for unvoiced frames with high LPC gain */
+            opus_int32 invGain_Q30, down_scale_Q30;
+
+            invGain_Q30 = silk_LPC_inverse_pred_gain( psPLC->prevLPC_Q12, psDec->LPC_order, arch );
+
+            down_scale_Q30 = silk_min_32( silk_RSHIFT( (opus_int32)1 << 30, LOG2_INV_LPC_GAIN_HIGH_THRES ), invGain_Q30 );
+            down_scale_Q30 = silk_max_32( silk_RSHIFT( (opus_int32)1 << 30, LOG2_INV_LPC_GAIN_LOW_THRES ), down_scale_Q30 );
+            down_scale_Q30 = silk_LSHIFT( down_scale_Q30, LOG2_INV_LPC_GAIN_HIGH_THRES );
+
+            rand_Gain_Q15 = silk_RSHIFT( silk_SMULWB( down_scale_Q30, rand_Gain_Q15 ), 14 );
+        }
+    }
+
+    rand_seed    = psPLC->rand_seed;
+    lag          = silk_RSHIFT_ROUND( psPLC->pitchL_Q8, 8 );
+    sLTP_buf_idx = psDec->ltp_mem_length;
+
+    /* Rewhiten LTP state */
+    idx = psDec->ltp_mem_length - lag - psDec->LPC_order - LTP_ORDER / 2;
+    celt_assert( idx > 0 );
+    silk_LPC_analysis_filter( &sLTP[ idx ], &psDec->outBuf[ idx ], A_Q12, psDec->ltp_mem_length - idx, psDec->LPC_order, arch );
+    /* Scale LTP state */
+    inv_gain_Q30 = silk_INVERSE32_varQ( psPLC->prevGain_Q16[ 1 ], 46 );
+    inv_gain_Q30 = silk_min( inv_gain_Q30, silk_int32_MAX >> 1 );
+    for( i = idx + psDec->LPC_order; i < psDec->ltp_mem_length; i++ ) {
+        sLTP_Q14[ i ] = silk_SMULWB( inv_gain_Q30, sLTP[ i ] );
+    }
+
+    /***************************/
+    /* LTP synthesis filtering */
+    /***************************/
+    for( k = 0; k < psDec->nb_subfr; k++ ) {
+        /* Set up pointer */
+        pred_lag_ptr = &sLTP_Q14[ sLTP_buf_idx - lag + LTP_ORDER / 2 ];
+        for( i = 0; i < psDec->subfr_length; i++ ) {
+            /* Unrolled loop */
+            /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */
+            LTP_pred_Q12 = 2;
+            LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[  0 ], B_Q14[ 0 ] );
+            LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -1 ], B_Q14[ 1 ] );
+            LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -2 ], B_Q14[ 2 ] );
+            LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -3 ], B_Q14[ 3 ] );
+            LTP_pred_Q12 = silk_SMLAWB( LTP_pred_Q12, pred_lag_ptr[ -4 ], B_Q14[ 4 ] );
+            pred_lag_ptr++;
+
+            /* Generate LPC excitation */
+            rand_seed = silk_RAND( rand_seed );
+            idx = silk_RSHIFT( rand_seed, 25 ) & RAND_BUF_MASK;
+            sLTP_Q14[ sLTP_buf_idx ] = silk_LSHIFT32( silk_SMLAWB( LTP_pred_Q12, rand_ptr[ idx ], rand_scale_Q14 ), 2 );
+            sLTP_buf_idx++;
+        }
+
+        /* Gradually reduce LTP gain */
+        for( j = 0; j < LTP_ORDER; j++ ) {
+            B_Q14[ j ] = silk_RSHIFT( silk_SMULBB( harm_Gain_Q15, B_Q14[ j ] ), 15 );
+        }
+        /* Gradually reduce excitation gain */
+        rand_scale_Q14 = silk_RSHIFT( silk_SMULBB( rand_scale_Q14, rand_Gain_Q15 ), 15 );
+
+        /* Slowly increase pitch lag */
+        psPLC->pitchL_Q8 = silk_SMLAWB( psPLC->pitchL_Q8, psPLC->pitchL_Q8, PITCH_DRIFT_FAC_Q16 );
+        psPLC->pitchL_Q8 = silk_min_32( psPLC->pitchL_Q8, silk_LSHIFT( silk_SMULBB( MAX_PITCH_LAG_MS, psDec->fs_kHz ), 8 ) );
+        lag = silk_RSHIFT_ROUND( psPLC->pitchL_Q8, 8 );
+    }
+
+    /***************************/
+    /* LPC synthesis filtering */
+    /***************************/
+    sLPC_Q14_ptr = &sLTP_Q14[ psDec->ltp_mem_length - MAX_LPC_ORDER ];
+
+    /* Copy LPC state */
+    silk_memcpy( sLPC_Q14_ptr, psDec->sLPC_Q14_buf, MAX_LPC_ORDER * sizeof( opus_int32 ) );
+
+    celt_assert( psDec->LPC_order >= 10 ); /* check that unrolling works */
+    for( i = 0; i < psDec->frame_length; i++ ) {
+        /* partly unrolled */
+        /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */
+        LPC_pred_Q10 = silk_RSHIFT( psDec->LPC_order, 1 );
+        LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i -  1 ], A_Q12[ 0 ] );
+        LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i -  2 ], A_Q12[ 1 ] );
+        LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i -  3 ], A_Q12[ 2 ] );
+        LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i -  4 ], A_Q12[ 3 ] );
+        LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i -  5 ], A_Q12[ 4 ] );
+        LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i -  6 ], A_Q12[ 5 ] );
+        LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i -  7 ], A_Q12[ 6 ] );
+        LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i -  8 ], A_Q12[ 7 ] );
+        LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i -  9 ], A_Q12[ 8 ] );
+        LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - 10 ], A_Q12[ 9 ] );
+        for( j = 10; j < psDec->LPC_order; j++ ) {
+            LPC_pred_Q10 = silk_SMLAWB( LPC_pred_Q10, sLPC_Q14_ptr[ MAX_LPC_ORDER + i - j - 1 ], A_Q12[ j ] );
+        }
+
+        /* Add prediction to LPC excitation */
+        sLPC_Q14_ptr[ MAX_LPC_ORDER + i ] = silk_ADD_SAT32( sLPC_Q14_ptr[ MAX_LPC_ORDER + i ],
+                                            silk_LSHIFT_SAT32( LPC_pred_Q10, 4 ));
+
+        /* Scale with Gain */
+        frame[ i ] = (opus_int16)silk_SAT16( silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( sLPC_Q14_ptr[ MAX_LPC_ORDER + i ], prevGain_Q10[ 1 ] ), 8 ) ) );
+    }
+#ifdef ENABLE_DEEP_PLC
+    if ( lpcnet != NULL && lpcnet->loaded && psDec->sPLC.fs_kHz == 16 ) {
+        int run_deep_plc = psDec->sPLC.enable_deep_plc || lpcnet->fec_fill_pos != 0;
+        if( run_deep_plc ) {
+            for( k = 0; k < psDec->nb_subfr; k += 2 ) {
+                lpcnet_plc_conceal( lpcnet, frame + k * psDec->subfr_length );
+            }
+            /* We *should* be able to copy only from psDec->frame_length-MAX_LPC_ORDER, i.e. the last MAX_LPC_ORDER samples. */
+            for( i = 0; i < psDec->frame_length; i++ ) {
+                sLPC_Q14_ptr[ MAX_LPC_ORDER + i ] = (int)floor(.5 + frame[ i ] * (float)(1 << 24) / prevGain_Q10[ 1 ] );
+            }
+        } else {
+          for( k = 0; k < psDec->nb_subfr; k += 2 ) {
+              lpcnet_plc_update( lpcnet, frame + k * psDec->subfr_length );
+          }
+        }
+    }
+#endif
+
+    /* Save LPC state */
+    silk_memcpy( psDec->sLPC_Q14_buf, &sLPC_Q14_ptr[ psDec->frame_length ], MAX_LPC_ORDER * sizeof( opus_int32 ) );
+
+    /**************************************/
+    /* Update states                      */
+    /**************************************/
+    psPLC->rand_seed     = rand_seed;
+    psPLC->randScale_Q14 = rand_scale_Q14;
+    for( i = 0; i < MAX_NB_SUBFR; i++ ) {
+        psDecCtrl->pitchL[ i ] = lag;
+    }
+    RESTORE_STACK;
+}
+
+/* Glues concealed frames with new good received frames */
+void silk_PLC_glue_frames(
+    silk_decoder_state                  *psDec,             /* I/O decoder state        */
+    opus_int16                          frame[],            /* I/O signal               */
+    opus_int                            length              /* I length of signal       */
+)
+{
+    opus_int   i, energy_shift;
+    opus_int32 energy;
+    silk_PLC_struct *psPLC;
+    psPLC = &psDec->sPLC;
+
+    if( psDec->lossCnt ) {
+        /* Calculate energy in concealed residual */
+        silk_sum_sqr_shift( &psPLC->conc_energy, &psPLC->conc_energy_shift, frame, length );
+
+        psPLC->last_frame_lost = 1;
+    } else {
+        if( psDec->sPLC.last_frame_lost ) {
+            /* Calculate residual in decoded signal if last frame was lost */
+            silk_sum_sqr_shift( &energy, &energy_shift, frame, length );
+
+            /* Normalize energies */
+            if( energy_shift > psPLC->conc_energy_shift ) {
+                psPLC->conc_energy = silk_RSHIFT( psPLC->conc_energy, energy_shift - psPLC->conc_energy_shift );
+            } else if( energy_shift < psPLC->conc_energy_shift ) {
+                energy = silk_RSHIFT( energy, psPLC->conc_energy_shift - energy_shift );
+            }
+
+            /* Fade in the energy difference */
+            if( energy > psPLC->conc_energy ) {
+                opus_int32 frac_Q24, LZ;
+                opus_int32 gain_Q16, slope_Q16;
+
+                LZ = silk_CLZ32( psPLC->conc_energy );
+                LZ = LZ - 1;
+                psPLC->conc_energy = silk_LSHIFT( psPLC->conc_energy, LZ );
+                energy = silk_RSHIFT( energy, silk_max_32( 24 - LZ, 0 ) );
+
+                frac_Q24 = silk_DIV32( psPLC->conc_energy, silk_max( energy, 1 ) );
+
+                gain_Q16 = silk_LSHIFT( silk_SQRT_APPROX( frac_Q24 ), 4 );
+                slope_Q16 = silk_DIV32_16( ( (opus_int32)1 << 16 ) - gain_Q16, length );
+                /* Make slope 4x steeper to avoid missing onsets after DTX */
+                slope_Q16 = silk_LSHIFT( slope_Q16, 2 );
+#ifdef ENABLE_DEEP_PLC
+                if ( psDec->sPLC.fs_kHz != 16 )
+#endif
+                {
+                    for( i = 0; i < length; i++ ) {
+                        frame[ i ] = silk_SMULWB( gain_Q16, frame[ i ] );
+                        gain_Q16 += slope_Q16;
+                        if( gain_Q16 > (opus_int32)1 << 16 ) {
+                            break;
+                        }
+                    }
+                }
+            }
+        }
+        psPLC->last_frame_lost = 0;
+    }
+}
--- a/lib/libopus/silk/PLC.h
+++ b/lib/libopus/silk/PLC.h
@ -0,0 +1,64 @@
+/***********************************************************************
+Copyright (c) 2006-2011, Skype Limited. All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+- Redistributions of source code must retain the above copyright notice,
+this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+***********************************************************************/
+
+#ifndef SILK_PLC_H
+#define SILK_PLC_H
+
+#include "main.h"
+
+#define BWE_COEF                        0.99
+#define V_PITCH_GAIN_START_MIN_Q14      11469               /* 0.7 in Q14               */
+#define V_PITCH_GAIN_START_MAX_Q14      15565               /* 0.95 in Q14              */
+#define MAX_PITCH_LAG_MS                18
+#define RAND_BUF_SIZE                   128
+#define RAND_BUF_MASK                   ( RAND_BUF_SIZE - 1 )
+#define LOG2_INV_LPC_GAIN_HIGH_THRES    3                   /* 2^3 = 8 dB LPC gain      */
+#define LOG2_INV_LPC_GAIN_LOW_THRES     8                   /* 2^8 = 24 dB LPC gain     */
+#define PITCH_DRIFT_FAC_Q16             655                 /* 0.01 in Q16              */
+
+void silk_PLC_Reset(
+    silk_decoder_state                  *psDec              /* I/O Decoder state        */
+);
+
+void silk_PLC(
+    silk_decoder_state                  *psDec,             /* I/O Decoder state        */
+    silk_decoder_control                *psDecCtrl,         /* I/O Decoder control      */
+    opus_int16                          frame[],            /* I/O  signal              */
+    opus_int                            lost,               /* I Loss flag              */
+#ifdef ENABLE_DEEP_PLC
+    LPCNetPLCState                      *lpcnet,
+#endif
+    int                                 arch                /* I Run-time architecture  */
+);
+
+void silk_PLC_glue_frames(
+    silk_decoder_state                  *psDec,             /* I/O decoder state        */
+    opus_int16                          frame[],            /* I/O signal               */
+    opus_int                            length              /* I length of signal       */
+);
+
+#endif
--- a/lib/libopus/silk/SigProc_FIX.h
+++ b/lib/libopus/silk/SigProc_FIX.h
@ -0,0 +1,641 @@
+/***********************************************************************
+Copyright (c) 2006-2011, Skype Limited. All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+- Redistributions of source code must retain the above copyright notice,
+this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+***********************************************************************/
+
+#ifndef SILK_SIGPROC_FIX_H
+#define SILK_SIGPROC_FIX_H
+
+
+/*#define silk_MACRO_COUNT */          /* Used to enable WMOPS counting */
+
+#define SILK_MAX_ORDER_LPC            24            /* max order of the LPC analysis in schur() and k2a() */
+
+#include <string.h>                                 /* for memset(), memcpy(), memmove() */
+#include "typedef.h"
+#include "resampler_structs.h"
+#include "macros.h"
+#include "cpu_support.h"
+
+#if defined(OPUS_X86_MAY_HAVE_SSE4_1)
+#include "x86/SigProc_FIX_sse.h"
+#endif
+
+#if (defined(OPUS_ARM_ASM) || defined(OPUS_ARM_MAY_HAVE_NEON_INTR))
+#include "arm/biquad_alt_arm.h"
+#include "arm/LPC_inv_pred_gain_arm.h"
+#endif
+
+/********************************************************************/
+/*                    SIGNAL PROCESSING FUNCTIONS                   */
+/********************************************************************/
+
+/*!
+ * Initialize/reset the resampler state for a given pair of input/output sampling rates
+*/
+opus_int silk_resampler_init(
+    silk_resampler_state_struct *S,                 /* I/O  Resampler state                                             */
+    opus_int32                  Fs_Hz_in,           /* I    Input sampling rate (Hz)                                    */
+    opus_int32                  Fs_Hz_out,          /* I    Output sampling rate (Hz)                                   */
+    opus_int                    forEnc              /* I    If 1: encoder; if 0: decoder                                */
+);
+
+/*!
+ * Resampler: convert from one sampling rate to another
+ */
+opus_int silk_resampler(
+    silk_resampler_state_struct *S,                 /* I/O  Resampler state                                             */
+    opus_int16                  out[],              /* O    Output signal                                               */
+    const opus_int16            in[],               /* I    Input signal                                                */
+    opus_int32                  inLen               /* I    Number of input samples                                     */
+);
+
+/*!
+* Downsample 2x, mediocre quality
+*/
+void silk_resampler_down2(
+    opus_int32                  *S,                 /* I/O  State vector [ 2 ]                                          */
+    opus_int16                  *out,               /* O    Output signal [ len ]                                       */
+    const opus_int16            *in,                /* I    Input signal [ floor(len/2) ]                               */
+    opus_int32                  inLen               /* I    Number of input samples                                     */
+);
+
+/*!
+ * Downsample by a factor 2/3, low quality
+*/
+void silk_resampler_down2_3(
+    opus_int32                  *S,                 /* I/O  State vector [ 6 ]                                          */
+    opus_int16                  *out,               /* O    Output signal [ floor(2*inLen/3) ]                          */
+    const opus_int16            *in,                /* I    Input signal [ inLen ]                                      */
+    opus_int32                  inLen               /* I    Number of input samples                                     */
+);
+
+/*!
+ * second order ARMA filter;
+ * slower than biquad() but uses more precise coefficients
+ * can handle (slowly) varying coefficients
+ */
+void silk_biquad_alt_stride1(
+    const opus_int16            *in,                /* I     input signal                                               */
+    const opus_int32            *B_Q28,             /* I     MA coefficients [3]                                        */
+    const opus_int32            *A_Q28,             /* I     AR coefficients [2]                                        */
+    opus_int32                  *S,                 /* I/O   State vector [2]                                           */
+    opus_int16                  *out,               /* O     output signal                                              */
+    const opus_int32            len                 /* I     signal length (must be even)                               */
+);
+
+void silk_biquad_alt_stride2_c(
+    const opus_int16            *in,                /* I     input signal                                               */
+    const opus_int32            *B_Q28,             /* I     MA coefficients [3]                                        */
+    const opus_int32            *A_Q28,             /* I     AR coefficients [2]                                        */
+    opus_int32                  *S,                 /* I/O   State vector [4]                                           */
+    opus_int16                  *out,               /* O     output signal                                              */
+    const opus_int32            len                 /* I     signal length (must be even)                               */
+);
+
+/* Variable order MA prediction error filter. */
+void silk_LPC_analysis_filter(
+    opus_int16                  *out,               /* O    Output signal                                               */
+    const opus_int16            *in,                /* I    Input signal                                                */
+    const opus_int16            *B,                 /* I    MA prediction coefficients, Q12 [order]                     */
+    const opus_int32            len,                /* I    Signal length                                               */
+    const opus_int32            d,                  /* I    Filter order                                                */
+    int                         arch                /* I    Run-time architecture                                       */
+);
+
+/* Chirp (bandwidth expand) LP AR filter */
+void silk_bwexpander(
+    opus_int16                  *ar,                /* I/O  AR filter to be expanded (without leading 1)                */
+    const opus_int              d,                  /* I    Length of ar                                                */
+    opus_int32                  chirp_Q16           /* I    Chirp factor (typically in the range 0 to 1)                */
+);
+
+/* Chirp (bandwidth expand) LP AR filter */
+void silk_bwexpander_32(
+    opus_int32                  *ar,                /* I/O  AR filter to be expanded (without leading 1)                */
+    const opus_int              d,                  /* I    Length of ar                                                */
+    opus_int32                  chirp_Q16           /* I    Chirp factor in Q16                                         */
+);
+
+/* Compute inverse of LPC prediction gain, and                           */
+/* test if LPC coefficients are stable (all poles within unit circle)    */
+opus_int32 silk_LPC_inverse_pred_gain_c(            /* O   Returns inverse prediction gain in energy domain, Q30        */
+    const opus_int16            *A_Q12,             /* I   Prediction coefficients, Q12 [order]                         */
+    const opus_int              order               /* I   Prediction order                                             */
+);
+
+/* Split signal in two decimated bands using first-order allpass filters */
+void silk_ana_filt_bank_1(
+    const opus_int16            *in,                /* I    Input signal [N]                                            */
+    opus_int32                  *S,                 /* I/O  State vector [2]                                            */
+    opus_int16                  *outL,              /* O    Low band [N/2]                                              */
+    opus_int16                  *outH,              /* O    High band [N/2]                                             */
+    const opus_int32            N                   /* I    Number of input samples                                     */
+);
+
+#if !defined(OVERRIDE_silk_biquad_alt_stride2)
+#define silk_biquad_alt_stride2(in, B_Q28, A_Q28, S, out, len, arch) ((void)(arch), silk_biquad_alt_stride2_c(in, B_Q28, A_Q28, S, out, len))
+#endif
+
+#if !defined(OVERRIDE_silk_LPC_inverse_pred_gain)
+#define silk_LPC_inverse_pred_gain(A_Q12, order, arch)     ((void)(arch), silk_LPC_inverse_pred_gain_c(A_Q12, order))
+#endif
+
+/********************************************************************/
+/*                        SCALAR FUNCTIONS                          */
+/********************************************************************/
+
+/* Approximation of 128 * log2() (exact inverse of approx 2^() below) */
+/* Convert input to a log scale    */
+opus_int32 silk_lin2log(
+    const opus_int32            inLin               /* I  input in linear scale                                         */
+);
+
+/* Approximation of a sigmoid function */
+opus_int silk_sigm_Q15(
+    opus_int                    in_Q5               /* I                                                                */
+);
+
+/* Approximation of 2^() (exact inverse of approx log2() above) */
+/* Convert input to a linear scale */
+opus_int32 silk_log2lin(
+    const opus_int32            inLog_Q7            /* I  input on log scale                                            */
+);
+
+/* Compute number of bits to right shift the sum of squares of a vector    */
+/* of int16s to make it fit in an int32                                    */
+void silk_sum_sqr_shift(
+    opus_int32                  *energy,            /* O   Energy of x, after shifting to the right                     */
+    opus_int                    *shift,             /* O   Number of bits right shift applied to energy                 */
+    const opus_int16            *x,                 /* I   Input vector                                                 */
+    opus_int                    len                 /* I   Length of input vector                                       */
+);
+
+/* Calculates the reflection coefficients from the correlation sequence    */
+/* Faster than schur64(), but much less accurate.                          */
+/* uses SMLAWB(), requiring armv5E and higher.                             */
+opus_int32 silk_schur(                              /* O    Returns residual energy                                     */
+    opus_int16                  *rc_Q15,            /* O    reflection coefficients [order] Q15                         */
+    const opus_int32            *c,                 /* I    correlations [order+1]                                      */
+    const opus_int32            order               /* I    prediction order                                            */
+);
+
+/* Calculates the reflection coefficients from the correlation sequence    */
+/* Slower than schur(), but more accurate.                                 */
+/* Uses SMULL(), available on armv4                                        */
+opus_int32 silk_schur64(                            /* O    returns residual energy                                     */
+    opus_int32                  rc_Q16[],           /* O    Reflection coefficients [order] Q16                         */
+    const opus_int32            c[],                /* I    Correlations [order+1]                                      */
+    opus_int32                  order               /* I    Prediction order                                            */
+);
+
+/* Step up function, converts reflection coefficients to prediction coefficients */
+void silk_k2a(
+    opus_int32                  *A_Q24,             /* O    Prediction coefficients [order] Q24                         */
+    const opus_int16            *rc_Q15,            /* I    Reflection coefficients [order] Q15                         */
+    const opus_int32            order               /* I    Prediction order                                            */
+);
+
+/* Step up function, converts reflection coefficients to prediction coefficients */
+void silk_k2a_Q16(
+    opus_int32                  *A_Q24,             /* O    Prediction coefficients [order] Q24                         */
+    const opus_int32            *rc_Q16,            /* I    Reflection coefficients [order] Q16                         */
+    const opus_int32            order               /* I    Prediction order                                            */
+);
+
+/* Apply sine window to signal vector.                              */
+/* Window types:                                                    */
+/*    1 -> sine window from 0 to pi/2                               */
+/*    2 -> sine window from pi/2 to pi                              */
+/* every other sample of window is linearly interpolated, for speed */
+void silk_apply_sine_window(
+    opus_int16                  px_win[],           /* O    Pointer to windowed signal                                  */
+    const opus_int16            px[],               /* I    Pointer to input signal                                     */
+    const opus_int              win_type,           /* I    Selects a window type                                       */
+    const opus_int              length              /* I    Window length, multiple of 4                                */
+);
+
+/* Compute autocorrelation */
+void silk_autocorr(
+    opus_int32                  *results,           /* O    Result (length correlationCount)                            */
+    opus_int                    *scale,             /* O    Scaling of the correlation vector                           */
+    const opus_int16            *inputData,         /* I    Input data to correlate                                     */
+    const opus_int              inputDataSize,      /* I    Length of input                                             */
+    const opus_int              correlationCount,   /* I    Number of correlation taps to compute                       */
+    int                         arch                /* I    Run-time architecture                                       */
+);
+
+void silk_decode_pitch(
+    opus_int16                  lagIndex,           /* I                                                                */
+    opus_int8                   contourIndex,       /* O                                                                */
+    opus_int                    pitch_lags[],       /* O    4 pitch values                                              */
+    const opus_int              Fs_kHz,             /* I    sampling frequency (kHz)                                    */
+    const opus_int              nb_subfr            /* I    number of sub frames                                        */
+);
+
+opus_int silk_pitch_analysis_core(                  /* O    Voicing estimate: 0 voiced, 1 unvoiced                      */
+    const opus_int16            *frame,             /* I    Signal of length PE_FRAME_LENGTH_MS*Fs_kHz                  */
+    opus_int                    *pitch_out,         /* O    4 pitch lag values                                          */
+    opus_int16                  *lagIndex,          /* O    Lag Index                                                   */
+    opus_int8                   *contourIndex,      /* O    Pitch contour Index                                         */
+    opus_int                    *LTPCorr_Q15,       /* I/O  Normalized correlation; input: value from previous frame    */
+    opus_int                    prevLag,            /* I    Last lag of previous frame; set to zero is unvoiced         */
+    const opus_int32            search_thres1_Q16,  /* I    First stage threshold for lag candidates 0 - 1              */
+    const opus_int              search_thres2_Q13,  /* I    Final threshold for lag candidates 0 - 1                    */
+    const opus_int              Fs_kHz,             /* I    Sample frequency (kHz)                                      */
+    const opus_int              complexity,         /* I    Complexity setting, 0-2, where 2 is highest                 */
+    const opus_int              nb_subfr,           /* I    number of 5 ms subframes                                    */
+    int                         arch                /* I    Run-time architecture                                       */
+);
+
+/* Compute Normalized Line Spectral Frequencies (NLSFs) from whitening filter coefficients      */
+/* If not all roots are found, the a_Q16 coefficients are bandwidth expanded until convergence. */
+void silk_A2NLSF(
+    opus_int16                  *NLSF,              /* O    Normalized Line Spectral Frequencies in Q15 (0..2^15-1) [d] */
+    opus_int32                  *a_Q16,             /* I/O  Monic whitening filter coefficients in Q16 [d]              */
+    const opus_int              d                   /* I    Filter order (must be even)                                 */
+);
+
+/* compute whitening filter coefficients from normalized line spectral frequencies */
+void silk_NLSF2A(
+    opus_int16                  *a_Q12,             /* O    monic whitening filter coefficients in Q12,  [ d ]          */
+    const opus_int16            *NLSF,              /* I    normalized line spectral frequencies in Q15, [ d ]          */
+    const opus_int              d,                  /* I    filter order (should be even)                               */
+    int                         arch                /* I    Run-time architecture                                       */
+);
+
+/* Convert int32 coefficients to int16 coefs and make sure there's no wrap-around */
+void silk_LPC_fit(
+    opus_int16                  *a_QOUT,            /* O    Output signal                                               */
+    opus_int32                  *a_QIN,             /* I/O  Input signal                                                */
+    const opus_int              QOUT,               /* I    Input Q domain                                              */
+    const opus_int              QIN,                /* I    Input Q domain                                              */
+    const opus_int              d                   /* I    Filter order                                                */
+);
+
+void silk_insertion_sort_increasing(
+    opus_int32                  *a,                 /* I/O   Unsorted / Sorted vector                                   */
+    opus_int                    *idx,               /* O     Index vector for the sorted elements                       */
+    const opus_int              L,                  /* I     Vector length                                              */
+    const opus_int              K                   /* I     Number of correctly sorted positions                       */
+);
+
+void silk_insertion_sort_decreasing_int16(
+    opus_int16                  *a,                 /* I/O   Unsorted / Sorted vector                                   */
+    opus_int                    *idx,               /* O     Index vector for the sorted elements                       */
+    const opus_int              L,                  /* I     Vector length                                              */
+    const opus_int              K                   /* I     Number of correctly sorted positions                       */
+);
+
+void silk_insertion_sort_increasing_all_values_int16(
+     opus_int16                 *a,                 /* I/O   Unsorted / Sorted vector                                   */
+     const opus_int             L                   /* I     Vector length                                              */
+);
+
+/* NLSF stabilizer, for a single input data vector */
+void silk_NLSF_stabilize(
+          opus_int16            *NLSF_Q15,          /* I/O   Unstable/stabilized normalized LSF vector in Q15 [L]       */
+    const opus_int16            *NDeltaMin_Q15,     /* I     Min distance vector, NDeltaMin_Q15[L] must be >= 1 [L+1]   */
+    const opus_int              L                   /* I     Number of NLSF parameters in the input vector              */
+);
+
+/* Laroia low complexity NLSF weights */
+void silk_NLSF_VQ_weights_laroia(
+    opus_int16                  *pNLSFW_Q_OUT,      /* O     Pointer to input vector weights [D]                        */
+    const opus_int16            *pNLSF_Q15,         /* I     Pointer to input vector         [D]                        */
+    const opus_int              D                   /* I     Input vector dimension (even)                              */
+);
+
+/* Compute reflection coefficients from input signal */
+void silk_burg_modified_c(
+    opus_int32                  *res_nrg,           /* O    Residual energy                                             */
+    opus_int                    *res_nrg_Q,         /* O    Residual energy Q value                                     */
+    opus_int32                  A_Q16[],            /* O    Prediction coefficients (length order)                      */
+    const opus_int16            x[],                /* I    Input signal, length: nb_subfr * ( D + subfr_length )       */
+    const opus_int32            minInvGain_Q30,     /* I    Inverse of max prediction gain                              */
+    const opus_int              subfr_length,       /* I    Input signal subframe length (incl. D preceding samples)    */
+    const opus_int              nb_subfr,           /* I    Number of subframes stacked in x                            */
+    const opus_int              D,                  /* I    Order                                                       */
+    int                         arch                /* I    Run-time architecture                                       */
+);
+
+/* Copy and multiply a vector by a constant */
+void silk_scale_copy_vector16(
+    opus_int16                  *data_out,
+    const opus_int16            *data_in,
+    opus_int32                  gain_Q16,           /* I    Gain in Q16                                                 */
+    const opus_int              dataSize            /* I    Length                                                      */
+);
+
+/* Some for the LTP related function requires Q26 to work.*/
+void silk_scale_vector32_Q26_lshift_18(
+    opus_int32                  *data1,             /* I/O  Q0/Q18                                                      */
+    opus_int32                  gain_Q26,           /* I    Q26                                                         */
+    opus_int                    dataSize            /* I    length                                                      */
+);
+
+/********************************************************************/
+/*                        INLINE ARM MATH                           */
+/********************************************************************/
+
+/*    return sum( inVec1[i] * inVec2[i] ) */
+
+opus_int32 silk_inner_prod_aligned(
+    const opus_int16 *const     inVec1,             /*    I input vector 1                                              */
+    const opus_int16 *const     inVec2,             /*    I input vector 2                                              */
+    const opus_int              len,                /*    I vector lengths                                              */
+    int                         arch                /*    I Run-time architecture                                       */
+);
+
+
+opus_int32 silk_inner_prod_aligned_scale(
+    const opus_int16 *const     inVec1,             /*    I input vector 1                                              */
+    const opus_int16 *const     inVec2,             /*    I input vector 2                                              */
+    const opus_int              scale,              /*    I number of bits to shift                                     */
+    const opus_int              len                 /*    I vector lengths                                              */
+);
+
+opus_int64 silk_inner_prod16_c(
+    const opus_int16            *inVec1,            /*    I input vector 1                                              */
+    const opus_int16            *inVec2,            /*    I input vector 2                                              */
+    const opus_int              len                 /*    I vector lengths                                              */
+);
+
+/********************************************************************/
+/*                                MACROS                            */
+/********************************************************************/
+
+/* Rotate a32 right by 'rot' bits. Negative rot values result in rotating
+   left. Output is 32bit int.
+   Note: contemporary compilers recognize the C expression below and
+   compile it into a 'ror' instruction if available. No need for OPUS_INLINE ASM! */
+static OPUS_INLINE opus_int32 silk_ROR32( opus_int32 a32, opus_int rot )
+{
+    opus_uint32 x = (opus_uint32) a32;
+    opus_uint32 r = (opus_uint32) rot;
+    opus_uint32 m = (opus_uint32) -rot;
+    if( rot == 0 ) {
+        return a32;
+    } else if( rot < 0 ) {
+        return (opus_int32) ((x << m) | (x >> (32 - m)));
+    } else {
+        return (opus_int32) ((x << (32 - r)) | (x >> r));
+    }
+}
+
+/* Allocate opus_int16 aligned to 4-byte memory address */
+#if EMBEDDED_ARM
+#define silk_DWORD_ALIGN __attribute__((aligned(4)))
+#else
+#define silk_DWORD_ALIGN
+#endif
+
+/* Useful Macros that can be adjusted to other platforms */
+#define silk_memcpy(dest, src, size)        memcpy((dest), (src), (size))
+#define silk_memset(dest, src, size)        memset((dest), (src), (size))
+#define silk_memmove(dest, src, size)       memmove((dest), (src), (size))
+
+/* Fixed point macros */
+
+/* (a32 * b32) output have to be 32bit int */
+#define silk_MUL(a32, b32)                  ((a32) * (b32))
+
+/* (a32 * b32) output have to be 32bit uint */
+#define silk_MUL_uint(a32, b32)             silk_MUL(a32, b32)
+
+/* a32 + (b32 * c32) output have to be 32bit int */
+#define silk_MLA(a32, b32, c32)             silk_ADD32((a32),((b32) * (c32)))
+
+/* a32 + (b32 * c32) output have to be 32bit uint */
+#define silk_MLA_uint(a32, b32, c32)        silk_MLA(a32, b32, c32)
+
+/* ((a32 >> 16)  * (b32 >> 16)) output have to be 32bit int */
+#define silk_SMULTT(a32, b32)               (((a32) >> 16) * ((b32) >> 16))
+
+/* a32 + ((a32 >> 16)  * (b32 >> 16)) output have to be 32bit int */
+#define silk_SMLATT(a32, b32, c32)          silk_ADD32((a32),((b32) >> 16) * ((c32) >> 16))
+
+#define silk_SMLALBB(a64, b16, c16)         silk_ADD64((a64),(opus_int64)((opus_int32)(b16) * (opus_int32)(c16)))
+
+/* (a32 * b32) */
+#define silk_SMULL(a32, b32)                ((opus_int64)(a32) * /*(opus_int64)*/(b32))
+
+/* Adds two signed 32-bit values in a way that can overflow, while not relying on undefined behaviour
+   (just standard two's complement implementation-specific behaviour) */
+#define silk_ADD32_ovflw(a, b)              ((opus_int32)((opus_uint32)(a) + (opus_uint32)(b)))
+/* Subtractss two signed 32-bit values in a way that can overflow, while not relying on undefined behaviour
+   (just standard two's complement implementation-specific behaviour) */
+#define silk_SUB32_ovflw(a, b)              ((opus_int32)((opus_uint32)(a) - (opus_uint32)(b)))
+
+/* Multiply-accumulate macros that allow overflow in the addition (ie, no asserts in debug mode) */
+#define silk_MLA_ovflw(a32, b32, c32)       silk_ADD32_ovflw((a32), (opus_uint32)(b32) * (opus_uint32)(c32))
+#define silk_SMLABB_ovflw(a32, b32, c32)    (silk_ADD32_ovflw((a32) , ((opus_int32)((opus_int16)(b32))) * (opus_int32)((opus_int16)(c32))))
+
+#define silk_DIV32_16(a32, b16)             ((opus_int32)((a32) / (b16)))
+#define silk_DIV32(a32, b32)                ((opus_int32)((a32) / (b32)))
+
+/* These macros enables checking for overflow in silk_API_Debug.h*/
+#define silk_ADD16(a, b)                    ((a) + (b))
+#define silk_ADD32(a, b)                    ((a) + (b))
+#define silk_ADD64(a, b)                    ((a) + (b))
+
+#define silk_SUB16(a, b)                    ((a) - (b))
+#define silk_SUB32(a, b)                    ((a) - (b))
+#define silk_SUB64(a, b)                    ((a) - (b))
+
+#define silk_SAT8(a)                        ((a) > silk_int8_MAX ? silk_int8_MAX  :       \
+                                            ((a) < silk_int8_MIN ? silk_int8_MIN  : (a)))
+#define silk_SAT16(a)                       ((a) > silk_int16_MAX ? silk_int16_MAX :      \
+                                            ((a) < silk_int16_MIN ? silk_int16_MIN : (a)))
+#define silk_SAT32(a)                       ((a) > silk_int32_MAX ? silk_int32_MAX :      \
+                                            ((a) < silk_int32_MIN ? silk_int32_MIN : (a)))
+
+#define silk_CHECK_FIT8(a)                  (a)
+#define silk_CHECK_FIT16(a)                 (a)
+#define silk_CHECK_FIT32(a)                 (a)
+
+#define silk_ADD_SAT16(a, b)                (opus_int16)silk_SAT16( silk_ADD32( (opus_int32)(a), (b) ) )
+#define silk_ADD_SAT64(a, b)                ((((a) + (b)) & 0x8000000000000000LL) == 0 ?                            \
+                                            ((((a) & (b)) & 0x8000000000000000LL) != 0 ? silk_int64_MIN : (a)+(b)) : \
+                                            ((((a) | (b)) & 0x8000000000000000LL) == 0 ? silk_int64_MAX : (a)+(b)) )
+
+#define silk_SUB_SAT16(a, b)                (opus_int16)silk_SAT16( silk_SUB32( (opus_int32)(a), (b) ) )
+#define silk_SUB_SAT64(a, b)                ((((a)-(b)) & 0x8000000000000000LL) == 0 ?                                               \
+                                            (( (a) & ((b)^0x8000000000000000LL) & 0x8000000000000000LL) ? silk_int64_MIN : (a)-(b)) : \
+                                            ((((a)^0x8000000000000000LL) & (b)  & 0x8000000000000000LL) ? silk_int64_MAX : (a)-(b)) )
+
+/* Saturation for positive input values */
+#define silk_POS_SAT32(a)                   ((a) > silk_int32_MAX ? silk_int32_MAX : (a))
+
+/* Add with saturation for positive input values */
+#define silk_ADD_POS_SAT8(a, b)             ((((a)+(b)) & 0x80)                 ? silk_int8_MAX  : ((a)+(b)))
+#define silk_ADD_POS_SAT16(a, b)            ((((a)+(b)) & 0x8000)               ? silk_int16_MAX : ((a)+(b)))
+#define silk_ADD_POS_SAT32(a, b)            ((((opus_uint32)(a)+(opus_uint32)(b)) & 0x80000000) ? silk_int32_MAX : ((a)+(b)))
+
+#define silk_LSHIFT8(a, shift)              ((opus_int8)((opus_uint8)(a)<<(shift)))         /* shift >= 0, shift < 8  */
+#define silk_LSHIFT16(a, shift)             ((opus_int16)((opus_uint16)(a)<<(shift)))       /* shift >= 0, shift < 16 */
+#define silk_LSHIFT32(a, shift)             ((opus_int32)((opus_uint32)(a)<<(shift)))       /* shift >= 0, shift < 32 */
+#define silk_LSHIFT64(a, shift)             ((opus_int64)((opus_uint64)(a)<<(shift)))       /* shift >= 0, shift < 64 */
+#define silk_LSHIFT(a, shift)               silk_LSHIFT32(a, shift)                         /* shift >= 0, shift < 32 */
+
+#define silk_RSHIFT8(a, shift)              ((a)>>(shift))                                  /* shift >= 0, shift < 8  */
+#define silk_RSHIFT16(a, shift)             ((a)>>(shift))                                  /* shift >= 0, shift < 16 */
+#define silk_RSHIFT32(a, shift)             ((a)>>(shift))                                  /* shift >= 0, shift < 32 */
+#define silk_RSHIFT64(a, shift)             ((a)>>(shift))                                  /* shift >= 0, shift < 64 */
+#define silk_RSHIFT(a, shift)               silk_RSHIFT32(a, shift)                         /* shift >= 0, shift < 32 */
+
+/* saturates before shifting */
+#define silk_LSHIFT_SAT32(a, shift)         (silk_LSHIFT32( silk_LIMIT( (a), silk_RSHIFT32( silk_int32_MIN, (shift) ), \
+                                                    silk_RSHIFT32( silk_int32_MAX, (shift) ) ), (shift) ))
+
+#define silk_LSHIFT_ovflw(a, shift)         ((opus_int32)((opus_uint32)(a) << (shift)))     /* shift >= 0, allowed to overflow */
+#define silk_LSHIFT_uint(a, shift)          ((a) << (shift))                                /* shift >= 0 */
+#define silk_RSHIFT_uint(a, shift)          ((a) >> (shift))                                /* shift >= 0 */
+
+#define silk_ADD_LSHIFT(a, b, shift)        ((a) + silk_LSHIFT((b), (shift)))               /* shift >= 0 */
+#define silk_ADD_LSHIFT32(a, b, shift)      silk_ADD32((a), silk_LSHIFT32((b), (shift)))    /* shift >= 0 */
+#define silk_ADD_LSHIFT_uint(a, b, shift)   ((a) + silk_LSHIFT_uint((b), (shift)))          /* shift >= 0 */
+#define silk_ADD_RSHIFT(a, b, shift)        ((a) + silk_RSHIFT((b), (shift)))               /* shift >= 0 */
+#define silk_ADD_RSHIFT32(a, b, shift)      silk_ADD32((a), silk_RSHIFT32((b), (shift)))    /* shift >= 0 */
+#define silk_ADD_RSHIFT_uint(a, b, shift)   ((a) + silk_RSHIFT_uint((b), (shift)))          /* shift >= 0 */
+#define silk_SUB_LSHIFT32(a, b, shift)      silk_SUB32((a), silk_LSHIFT32((b), (shift)))    /* shift >= 0 */
+#define silk_SUB_RSHIFT32(a, b, shift)      silk_SUB32((a), silk_RSHIFT32((b), (shift)))    /* shift >= 0 */
+
+/* Requires that shift > 0 */
+#define silk_RSHIFT_ROUND(a, shift)         ((shift) == 1 ? ((a) >> 1) + ((a) & 1) : (((a) >> ((shift) - 1)) + 1) >> 1)
+#define silk_RSHIFT_ROUND64(a, shift)       ((shift) == 1 ? ((a) >> 1) + ((a) & 1) : (((a) >> ((shift) - 1)) + 1) >> 1)
+
+/* Number of rightshift required to fit the multiplication */
+#define silk_NSHIFT_MUL_32_32(a, b)         ( -(31- (32-silk_CLZ32(silk_abs(a)) + (32-silk_CLZ32(silk_abs(b))))) )
+#define silk_NSHIFT_MUL_16_16(a, b)         ( -(15- (16-silk_CLZ16(silk_abs(a)) + (16-silk_CLZ16(silk_abs(b))))) )
+
+
+#define silk_min(a, b)                      (((a) < (b)) ? (a) : (b))
+#define silk_max(a, b)                      (((a) > (b)) ? (a) : (b))
+
+/* Macro to convert floating-point constants to fixed-point */
+#define SILK_FIX_CONST( C, Q )              ((opus_int32)((C) * ((opus_int64)1 << (Q)) + 0.5))
+
+/* silk_min() versions with typecast in the function call */
+static OPUS_INLINE opus_int silk_min_int(opus_int a, opus_int b)
+{
+    return (((a) < (b)) ? (a) : (b));
+}
+static OPUS_INLINE opus_int16 silk_min_16(opus_int16 a, opus_int16 b)
+{
+    return (((a) < (b)) ? (a) : (b));
+}
+static OPUS_INLINE opus_int32 silk_min_32(opus_int32 a, opus_int32 b)
+{
+    return (((a) < (b)) ? (a) : (b));
+}
+static OPUS_INLINE opus_int64 silk_min_64(opus_int64 a, opus_int64 b)
+{
+    return (((a) < (b)) ? (a) : (b));
+}
+
+/* silk_min() versions with typecast in the function call */
+static OPUS_INLINE opus_int silk_max_int(opus_int a, opus_int b)
+{
+    return (((a) > (b)) ? (a) : (b));
+}
+static OPUS_INLINE opus_int16 silk_max_16(opus_int16 a, opus_int16 b)
+{
+    return (((a) > (b)) ? (a) : (b));
+}
+static OPUS_INLINE opus_int32 silk_max_32(opus_int32 a, opus_int32 b)
+{
+    return (((a) > (b)) ? (a) : (b));
+}
+static OPUS_INLINE opus_int64 silk_max_64(opus_int64 a, opus_int64 b)
+{
+    return (((a) > (b)) ? (a) : (b));
+}
+
+#define silk_LIMIT( a, limit1, limit2)      ((limit1) > (limit2) ? ((a) > (limit1) ? (limit1) : ((a) < (limit2) ? (limit2) : (a))) \
+                                                                 : ((a) > (limit2) ? (limit2) : ((a) < (limit1) ? (limit1) : (a))))
+
+#define silk_LIMIT_int                      silk_LIMIT
+#define silk_LIMIT_16                       silk_LIMIT
+#define silk_LIMIT_32                       silk_LIMIT
+
+#define silk_abs(a)                         (((a) >  0)  ? (a) : -(a))            /* Be careful, silk_abs returns wrong when input equals to silk_intXX_MIN */
+#define silk_abs_int(a)                     (((a) ^ ((a) >> (8 * sizeof(a) - 1))) - ((a) >> (8 * sizeof(a) - 1)))
+#define silk_abs_int32(a)                   (((a) ^ ((a) >> 31)) - ((a) >> 31))
+#define silk_abs_int64(a)                   (((a) >  0)  ? (a) : -(a))
+
+#define silk_sign(a)                        ((a) > 0 ? 1 : ( (a) < 0 ? -1 : 0 ))
+
+/* PSEUDO-RANDOM GENERATOR                                                          */
+/* Make sure to store the result as the seed for the next call (also in between     */
+/* frames), otherwise result won't be random at all. When only using some of the    */
+/* bits, take the most significant bits by right-shifting.                          */
+#define RAND_MULTIPLIER                     196314165
+#define RAND_INCREMENT                      907633515
+#define silk_RAND(seed)                     (silk_MLA_ovflw((RAND_INCREMENT), (seed), (RAND_MULTIPLIER)))
+
+/*  Add some multiplication functions that can be easily mapped to ARM/MIPS32. */
+
+/*    silk_SMMUL: Signed top word multiply.
+          ARMv6        2 instruction cycles.
+          ARMv3M+      3 instruction cycles. use SMULL and ignore LSB registers.(except xM)
+          MIPS32       2 instructions mul+mfhi
+          MIPS32r6     1 instruction muh */
+/*#define silk_SMMUL(a32, b32)                (opus_int32)silk_RSHIFT(silk_SMLAL(silk_SMULWB((a32), (b32)), (a32), silk_RSHIFT_ROUND((b32), 16)), 16)*/
+/* the following seems faster on x86 */
+#define silk_SMMUL(a32, b32)                (opus_int32)silk_RSHIFT64(silk_SMULL((a32), (b32)), 32)
+
+#if !defined(OVERRIDE_silk_burg_modified)
+#define silk_burg_modified(res_nrg, res_nrg_Q, A_Q16, x, minInvGain_Q30, subfr_length, nb_subfr, D, arch) \
+    ((void)(arch), silk_burg_modified_c(res_nrg, res_nrg_Q, A_Q16, x, minInvGain_Q30, subfr_length, nb_subfr, D, arch))
+#endif
+
+#if !defined(OVERRIDE_silk_inner_prod16)
+#define silk_inner_prod16(inVec1, inVec2, len, arch) \
+    ((void)(arch),silk_inner_prod16_c(inVec1, inVec2, len))
+#endif
+
+#include "Inlines.h"
+#include "MacroCount.h"
+#include "MacroDebug.h"
+
+#ifdef OPUS_ARM_INLINE_ASM
+#include "arm/SigProc_FIX_armv4.h"
+#endif
+
+#ifdef OPUS_ARM_INLINE_EDSP
+#include "arm/SigProc_FIX_armv5e.h"
+#endif
+
+#if defined(FIXED_POINT) && defined(__mips_dsp) && __mips == 32
+#include "mips/sigproc_fix_mipsr1.h"
+#endif
+
+#ifdef OPUS_XTENSA_LX7
+#include "xtensa/SigProc_FIX_lx7.h"
+#endif
+
+
+#endif /* SILK_SIGPROC_FIX_H */
--- a/lib/libopus/silk/VAD.c
+++ b/lib/libopus/silk/VAD.c
@ -0,0 +1,360 @@
+/***********************************************************************
+Copyright (c) 2006-2011, Skype Limited. All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions
+are met:
+- Redistributions of source code must retain the above copyright notice,
+this list of conditions and the following disclaimer.
+- Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in the
+documentation and/or other materials provided with the distribution.
+- Neither the name of Internet Society, IETF or IETF Trust, nor the
+names of specific contributors, may be used to endorse or promote
+products derived from this software without specific prior written
+permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+***********************************************************************/
+
+#ifdef HAVE_CONFIG_H
+#include "config.h"
+#endif
+
+#include "main.h"
+#include "stack_alloc.h"
+
+/* Silk VAD noise level estimation */
+# if !defined(OPUS_X86_MAY_HAVE_SSE4_1)
+static OPUS_INLINE void silk_VAD_GetNoiseLevels(
+    const opus_int32             pX[ VAD_N_BANDS ], /* I    subband energies                            */
+    silk_VAD_state              *psSilk_VAD         /* I/O  Pointer to Silk VAD state                   */
+);
+#endif
+
+/**********************************/
+/* Initialization of the Silk VAD */
+/**********************************/
+opus_int silk_VAD_Init(                                         /* O    Return value, 0 if success                  */
+    silk_VAD_state              *psSilk_VAD                     /* I/O  Pointer to Silk VAD state                   */
+)
+{
+    opus_int b, ret = 0;
+
+    /* reset state memory */
+    silk_memset( psSilk_VAD, 0, sizeof( silk_VAD_state ) );
+
+    /* init noise levels */
+    /* Initialize array with approx pink noise levels (psd proportional to inverse of frequency) */
+    for( b = 0; b < VAD_N_BANDS; b++ ) {
+        psSilk_VAD->NoiseLevelBias[ b ] = silk_max_32( silk_DIV32_16( VAD_NOISE_LEVELS_BIAS, b + 1 ), 1 );
+    }
+
+    /* Initialize state */
+    for( b = 0; b < VAD_N_BANDS; b++ ) {
+        psSilk_VAD->NL[ b ]     = silk_MUL( 100, psSilk_VAD->NoiseLevelBias[ b ] );
+        psSilk_VAD->inv_NL[ b ] = silk_DIV32( silk_int32_MAX, psSilk_VAD->NL[ b ] );
+    }
+    psSilk_VAD->counter = 15;
+
+    /* init smoothed energy-to-noise ratio*/
+    for( b = 0; b < VAD_N_BANDS; b++ ) {
+        psSilk_VAD->NrgRatioSmth_Q8[ b ] = 100 * 256;       /* 100 * 256 --> 20 dB SNR */
+    }
+
+    return( ret );
+}
+
+/* Weighting factors for tilt measure */
+static const opus_int32 tiltWeights[ VAD_N_BANDS ] = { 30000, 6000, -12000, -12000 };
+
+/***************************************/
+/* Get the speech activity level in Q8 */
+/***************************************/
+opus_int silk_VAD_GetSA_Q8_c(                                   /* O    Return value, 0 if success                  */
+    silk_encoder_state          *psEncC,                        /* I/O  Encoder state                               */
+    const opus_int16            pIn[]                           /* I    PCM input                                   */
+)
+{
+    opus_int   SA_Q15, pSNR_dB_Q7, input_tilt;
+    opus_int   decimated_framelength1, decimated_framelength2;
+    opus_int   decimated_framelength;
+    opus_int   dec_subframe_length, dec_subframe_offset, SNR_Q7, i, b, s;
+    opus_int32 sumSquared, smooth_coef_Q16;
+    opus_int16 HPstateTmp;
+    VARDECL( opus_int16, X );
+    opus_int32 Xnrg[ VAD_N_BANDS ];
+    opus_int32 NrgToNoiseRatio_Q8[ VAD_N_BANDS ];
+    opus_int32 speech_nrg, x_tmp;
+    opus_int   X_offset[ VAD_N_BANDS ];
+    opus_int   ret = 0;
+    silk_VAD_state *psSilk_VAD = &psEncC->sVAD;
+    SAVE_STACK;
+
+    /* Safety checks */
+    silk_assert( VAD_N_BANDS == 4 );
+    celt_assert( MAX_FRAME_LENGTH >= psEncC->frame_length );
+    celt_assert( psEncC->frame_length <= 512 );
+    celt_assert( psEncC->frame_length == 8 * silk_RSHIFT( psEncC->frame_length, 3 ) );
+
+    /***********************/
+    /* Filter and Decimate */
+    /***********************/
+    decimated_framelength1 = silk_RSHIFT( psEncC->frame_length, 1 );
+    decimated_framelength2 = silk_RSHIFT( psEncC->frame_length, 2 );
+    decimated_framelength = silk_RSHIFT( psEncC->frame_length, 3 );
+    /* Decimate into 4 bands:
+       0       L      3L       L              3L                             5L
+               -      --       -              --                             --
+               8       8       2               4                              4
+
+       [0-1 kHz| temp. |1-2 kHz|    2-4 kHz    |            4-8 kHz           |
+
+       They're arranged to allow the minimal ( frame_length / 4 ) extra
+       scratch space during the downsampling process */
+    X_offset[ 0 ] = 0;
+    X_offset[ 1 ] = decimated_framelength + decimated_framelength2;
+    X_offset[ 2 ] = X_offset[ 1 ] + decimated_framelength;
+    X_offset[ 3 ] = X_offset[ 2 ] + decimated_framelength2;
+    ALLOC( X, X_offset[ 3 ] + decimated_framelength1, opus_int16 );
+
+    /* 0-8 kHz to 0-4 kHz and 4-8 kHz */
+    silk_ana_filt_bank_1( pIn, &psSilk_VAD->AnaState[  0 ],
+        X, &X[ X_offset[ 3 ] ], psEncC->frame_length );
+
+    /* 0-4 kHz to 0-2 kHz and 2-4 kHz */
+    silk_ana_filt_bank_1( X, &psSilk_VAD->AnaState1[ 0 ],
+        X, &X[ X_offset[ 2 ] ], decimated_framelength1 );
+
+    /* 0-2 kHz to 0-1 kHz and 1-2 kHz */
+    silk_ana_filt_bank_1( X, &psSilk_VAD->AnaState2[ 0 ],
+        X, &X[ X_offset[ 1 ] ], decimated_framelength2 );
+
+    /*********************************************/
+    /* HP filter on lowest band (differentiator) */
+    /*********************************************/
+    X[ decimated_framelength - 1 ] = silk_RSHIFT( X[ decimated_framelength - 1 ], 1 );
+    HPstateTmp = X[ decimated_framelength - 1 ];
+    for( i = decimated_framelength - 1; i > 0; i-- ) {
+        X[ i - 1 ]  = silk_RSHIFT( X[ i - 1 ], 1 );
+        X[ i ]     -= X[ i - 1 ];
+    }
+    X[ 0 ] -= psSilk_VAD->HPstate;
+    psSilk_VAD->HPstate = HPstateTmp;
+
+    /*************************************/
+    /* Calculate the energy in each band */
+    /*************************************/
+    for( b = 0; b < VAD_N_BANDS; b++ ) {
+        /* Find the decimated framelength in the non-uniformly divided bands */
+        decimated_framelength = silk_RSHIFT( psEncC->frame_length, silk_min_int( VAD_N_BANDS - b, VAD_N_BANDS - 1 ) );
+
+        /* Split length into subframe lengths */
+        dec_subframe_length = silk_RSHIFT( decimated_framelength, VAD_INTERNAL_SUBFRAMES_LOG2 );
+        dec_subframe_offset = 0;
+
+        /* Compute energy per sub-frame */
+        /* initialize with summed energy of last subframe */
+        Xnrg[ b ] = psSilk_VAD->XnrgSubfr[ b ];
+        for( s = 0; s < VAD_INTERNAL_SUBFRAMES; s++ ) {
+            sumSquared = 0;
+            for( i = 0; i < dec_subframe_length; i++ ) {
+                /* The energy will be less than dec_subframe_length * ( silk_int16_MIN / 8 ) ^ 2.            */
+                /* Therefore we can accumulate with no risk of overflow (unless dec_subframe_length > 128)  */
+                x_tmp = silk_RSHIFT(
+                    X[ X_offset[ b ] + i + dec_subframe_offset ], 3 );
+                sumSquared = silk_SMLABB( sumSquared, x_tmp, x_tmp );
+
+                /* Safety check */
+                silk_assert( sumSquared >= 0 );
+            }
+
+            /* Add/saturate summed energy of current subframe */
+            if( s < VAD_INTERNAL_SUBFRAMES - 1 ) {
+                Xnrg[ b ] = silk_ADD_POS_SAT32( Xnrg[ b ], sumSquared );
+            } else {
+                /* Look-ahead subframe */
+                Xnrg[ b ] = silk_ADD_POS_SAT32( Xnrg[ b ], silk_RSHIFT( sumSquared, 1 ) );
+            }
+
+            dec_subframe_offset += dec_subframe_length;
+        }
+        psSilk_VAD->XnrgSubfr[ b ] = sumSquared;
+    }
+
+    /********************/
+    /* Noise estimation */
+    /********************/
+    silk_VAD_GetNoiseLevels( &Xnrg[ 0 ], psSilk_VAD );
+
+    /***********************************************/
+    /* Signal-plus-noise to noise ratio estimation */
+    /***********************************************/
+    sumSquared = 0;
+    input_tilt = 0;
+    for( b = 0; b < VAD_N_BANDS; b++ ) {
+        speech_nrg = Xnrg[ b ] - psSilk_VAD->NL[ b ];
+        if( speech_nrg > 0 ) {
+            /* Divide, with sufficient resolution */
+            if( ( Xnrg[ b ] & 0xFF800000 ) == 0 ) {
+                NrgToNoiseRatio_Q8[ b ] = silk_DIV32( silk_LSHIFT( Xnrg[ b ], 8 ), psSilk_VAD->NL[ b ] + 1 );
+            } else {
+                NrgToNoiseRatio_Q8[ b ] = silk_DIV32( Xnrg[ b ], silk_RSHIFT( psSilk_VAD->NL[ b ], 8 ) + 1 );
+            }
+
+            /* Convert to log domain */
+            SNR_Q7 = silk_lin2log( NrgToNoiseRatio_Q8[ b ] ) - 8 * 128;
+
+            /* Sum-of-squares */
+            sumSquared = silk_SMLABB( sumSquared, SNR_Q7, SNR_Q7 );          /* Q14 */
+
+            /* Tilt measure */
+            if( speech_nrg < ( (opus_int32)1 << 20 ) ) {
+                /* Scale down SNR value for small subband speech energies */
+                SNR_Q7 = silk_SMULWB( silk_LSHIFT( silk_SQRT_APPROX( speech_nrg ), 6 ), SNR_Q7 );
+            }
+            input_tilt = silk_SMLAWB( input_tilt, tiltWeights[ b ], SNR_Q7 );
+        } else {
+            NrgToNoiseRatio_Q8[ b ] = 256;
+        }
+    }
+
+    /* Mean-of-squares */
+    sumSquared = silk_DIV32_16( sumSquared, VAD_N_BANDS ); /* Q14 */
+
+    /* Root-mean-square approximation, scale to dBs, and write to output pointer */
+    pSNR_dB_Q7 = (opus_int16)( 3 * silk_SQRT_APPROX( sumSquared ) ); /* Q7 */
+
+    /*********************************/
+    /* Speech Probability Estimation */
+    /*********************************/
+    SA_Q15 = silk_sigm_Q15( silk_SMULWB( VAD_SNR_FACTOR_Q16, pSNR_dB_Q7 ) - VAD_NEGATIVE_OFFSET_Q5 );
+
+    /**************************/
+    /* Frequency Tilt Measure */
+    /**************************/
+    psEncC->input_tilt_Q15 = silk_LSHIFT( silk_sigm_Q15( input_tilt ) - 16384, 1 );
+
+    /**************************************************/
+    /* Scale the sigmoid output based on power levels */
+    /**************************************************/
+    speech_nrg = 0;
+    for( b = 0; b < VAD_N_BANDS; b++ ) {
+        /* Accumulate signal-without-noise energies, higher frequency bands have more weight */
+        speech_nrg += ( b + 1 ) * silk_RSHIFT( Xnrg[ b ] - psSilk_VAD->NL[ b ], 4 );
+    }
+
+    if( psEncC->frame_length == 20 * psEncC->fs_kHz ) {
+        speech_nrg = silk_RSHIFT32( speech_nrg, 1 );
+    }
+    /* Power scaling */
+    if( speech_nrg <= 0 ) {
+        SA_Q15 = silk_RSHIFT( SA_Q15, 1 );
+    } else if( speech_nrg < 16384 ) {
+        speech_nrg = silk_LSHIFT32( speech_nrg, 16 );
+
+        /* square-root */
+        speech_nrg = silk_SQRT_APPROX( speech_nrg );
+        SA_Q15 = silk_SMULWB( 32768 + speech_nrg, SA_Q15 );
+    }
+
+    /* Copy the resulting speech activity in Q8 */
+    psEncC->speech_activity_Q8 = silk_min_int( silk_RSHIFT( SA_Q15, 7 ), silk_uint8_MAX );
+
+    /***********************************/
+    /* Energy Level and SNR estimation */
+    /***********************************/
+    /* Smoothing coefficient */
+    smooth_coef_Q16 = silk_SMULWB( VAD_SNR_SMOOTH_COEF_Q18, silk_SMULWB( (opus_int32)SA_Q15, SA_Q15 ) );
+
+    if( psEncC->frame_length == 10 * psEncC->fs_kHz ) {
+        smooth_coef_Q16 >>= 1;
+    }
+
+    for( b = 0; b < VAD_N_BANDS; b++ ) {
+        /* compute smoothed energy-to-noise ratio per band */
+        psSilk_VAD->NrgRatioSmth_Q8[ b ] = silk_SMLAWB( psSilk_VAD->NrgRatioSmth_Q8[ b ],
+            NrgToNoiseRatio_Q8[ b ] - psSilk_VAD->NrgRatioSmth_Q8[ b ], smooth_coef_Q16 );
+
+        /* signal to noise ratio in dB per band */
+        SNR_Q7 = 3 * ( silk_lin2log( psSilk_VAD->NrgRatioSmth_Q8[b] ) - 8 * 128 );
+        /* quality = sigmoid( 0.25 * ( SNR_dB - 16 ) ); */
+        psEncC->input_quality_bands_Q15[ b ] = silk_sigm_Q15( silk_RSHIFT( SNR_Q7 - 16 * 128, 4 ) );
+    }
+
+    RESTORE_STACK;
+    return( ret );
+}
+
+/**************************/
+/* Noise level estimation */
+/**************************/
+# if  !defined(OPUS_X86_MAY_HAVE_SSE4_1)
+static OPUS_INLINE
+#endif
+void silk_VAD_GetNoiseLevels(
+    const opus_int32            pX[ VAD_N_BANDS ],  /* I    subband energies                            */
+    silk_VAD_state              *psSilk_VAD         /* I/O  Pointer to Silk VAD state                   */
+)
+{
+    opus_int   k;
+    opus_int32 nl, nrg, inv_nrg;
+    opus_int   coef, min_coef;
+
+    /* Initially faster smoothing */
+    if( psSilk_VAD->counter < 1000 ) { /* 1000 = 20 sec */
+        min_coef = silk_DIV32_16( silk_int16_MAX, silk_RSHIFT( psSilk_VAD->counter, 4 ) + 1 );
+        /* Increment frame counter */
+        psSilk_VAD->counter++;
+    } else {
+        min_coef = 0;
+    }
+
+    for( k = 0; k < VAD_N_BANDS; k++ ) {
+        /* Get old noise level estimate for current band */
+        nl = psSilk_VAD->NL[ k ];
+        silk_assert( nl >= 0 );
+
+        /* Add bias */
+        nrg = silk_ADD_POS_SAT32( pX[ k ], psSilk_VAD->NoiseLevelBias[ k ] );
+        silk_assert( nrg > 0 );
+
+        /* Invert energies */
+        inv_nrg = silk_DIV32( silk_int32_MAX, nrg );
+        silk_assert( inv_nrg >= 0 );
+
+        /* Less update when subband energy is high */
+        if( nrg > silk_LSHIFT( nl, 3 ) ) {
+            coef = VAD_NOISE_LEVEL_SMOOTH_COEF_Q16 >> 3;
+        } else if( nrg < nl ) {
+            coef = VAD_NOISE_LEVEL_SMOOTH_COEF_Q16;
+        } else {
+            coef = silk_SMULWB( silk_SMULWW( inv_nrg, nl ), VAD_NOISE_LEVEL_SMOOTH_COEF_Q16 << 1 );
+        }
+
+        /* Initially faster smoothing */
+        coef = silk_max_int( coef, min_coef );
+
+        /* Smooth inverse energies */
+        psSilk_VAD->inv_NL[ k ] = silk_SMLAWB( psSilk_VAD->inv_NL[ k ], inv_nrg - psSilk_VAD->inv_NL[ k ], coef );
+        silk_assert( psSilk_VAD->inv_NL[ k ] >= 0 );
+
+        /* Compute noise level by inverting again */
+        nl = silk_DIV32( silk_int32_MAX, psSilk_VAD->inv_NL[ k ] );
+        silk_assert( nl >= 0 );
+
+        /* Limit noise levels (guarantee 7 bits of head room) */
+        nl = silk_min( nl, 0x00FFFFFF );
+
+        /* Store as part of state */
+        psSilk_VAD->NL[ k ] = nl;
+    }
+}
--- a/Show More
+++ b/Show More