mozilla
/
gecko-dev
зеркало из https://github.com/mozilla/gecko-dev.git
1
0
Форкнуть 0
Код Задачи Пакеты Проекты Релизы Вики Активность

Bug 1239078 - Update libopus to 1.1.2 release. r=kinetik 

This release fixes two bugs:

- Resetting the encoder or decoder state with OPUS_RESET_STATE would
  disable some run-time selected architecture-specific optimizations.

- In hybrid mode discontinuous transmission (DTX) operation, the
  comfort noise above 8 kHz was incorrectly estimated and could
  oscillate in time.
This commit is contained in:
Ralph Giles 2016-01-12 12:12:39 -08:00
Родитель 7095f1cc1d
Коммит cb77a8b5fe
25 изменённых файлов: 1133 добавлений и 53 удалений
Показать статистику Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы

2
media/libopus/README_MOZILLA
Просмотреть файл

@ -8,4 +8,4 @@ files after the copy step.
The upstream repository is https://git.xiph.org/opus.git
The git tag/revision used was v1.1.1.
The git tag/revision used was v1.1.2.

35
media/libopus/celt/celt_decoder.c
Просмотреть файл

@ -457,10 +457,9 @@ static void celt_decode_lost(CELTDecoder * OPUS_RESTRICT st, int N, int LM)
VARDECL(celt_norm, X);
#endif
opus_uint32 seed;
opus_val16 *plcLogE;
int end;
int effEnd;
opus_val16 decay;
end = st->end;
effEnd = IMAX(start, IMIN(end, mode->effEBands));
@ -472,19 +471,13 @@ static void celt_decode_lost(CELTDecoder * OPUS_RESTRICT st, int N, int LM)
ALLOC(X, C*N, celt_norm); /**< Interleaved normalised MDCTs */
#endif
if (loss_count >= 5)
plcLogE = backgroundLogE;
else {
/* Energy decay */
opus_val16 decay = loss_count==0 ?
QCONST16(1.5f, DB_SHIFT) : QCONST16(.5f, DB_SHIFT);
c=0; do
{
for (i=start;i<end;i++)
oldBandE[c*nbEBands+i] -= decay;
} while (++c<C);
plcLogE = oldBandE;
}
/* Energy decay */
decay = loss_count==0 ? QCONST16(1.5f, DB_SHIFT) : QCONST16(.5f, DB_SHIFT);
c=0; do
{
for (i=start;i<end;i++)
oldBandE[c*nbEBands+i] = MAX16(backgroundLogE[c*nbEBands+i], oldBandE[c*nbEBands+i] - decay);
} while (++c<C);
seed = st->rng;
for (c=0;c<C;c++)
{
@ -510,7 +503,7 @@ static void celt_decode_lost(CELTDecoder * OPUS_RESTRICT st, int N, int LM)
DECODE_BUFFER_SIZE-N+(overlap>>1));
} while (++c<C);
celt_synthesis(mode, X, out_syn, plcLogE, start, effEnd, C, C, 0, LM, st->downsample, 0, st->arch);
celt_synthesis(mode, X, out_syn, oldBandE, start, effEnd, C, C, 0, LM, st->downsample, 0, st->arch);
} else {
/* Pitch-based PLC */
const opus_val16 *window;
@ -1037,10 +1030,18 @@ int celt_decode_with_ec(CELTDecoder * OPUS_RESTRICT st, const unsigned char *dat
/* In case start or end were to change */
if (!isTransient)
{
opus_val16 max_background_increase;
OPUS_COPY(oldLogE2, oldLogE, 2*nbEBands);
OPUS_COPY(oldLogE, oldBandE, 2*nbEBands);
/* In normal circumstances, we only allow the noise floor to increase by
up to 2.4 dB/second, but when we're in DTX, we allow up to 6 dB
increase for each update.*/
if (st->loss_count < 10)
max_background_increase = M*QCONST16(0.001f,DB_SHIFT);
else
max_background_increase = QCONST16(1.f,DB_SHIFT);
for (i=0;i<2*nbEBands;i++)
backgroundLogE[i] = MIN16(backgroundLogE[i] + M*QCONST16(0.001f,DB_SHIFT), oldBandE[i]);
backgroundLogE[i] = MIN16(backgroundLogE[i] + max_background_increase, oldBandE[i]);
} else {
for (i=0;i<2*nbEBands;i++)
oldLogE[i] = MIN16(oldLogE[i], oldBandE[i]);

26
media/libopus/celt/celt_encoder.c
Просмотреть файл

@ -343,9 +343,9 @@ static int transient_analysis(const opus_val32 * OPUS_RESTRICT in, int len, int
{
int id;
#ifdef FIXED_POINT
id = IMAX(0,IMIN(127,MULT16_32_Q15(tmp[i],norm))); /* Do not round to nearest */
id = MAX32(0,MIN32(127,MULT16_32_Q15(tmp[i]+EPSILON,norm))); /* Do not round to nearest */
#else
id = IMAX(0,IMIN(127,(int)floor(64*norm*tmp[i]))); /* Do not round to nearest */
id = (int)MAX32(0,MIN32(127,floor(64*norm*(tmp[i]+EPSILON)))); /* Do not round to nearest */
#endif
unmask += inv_table[id];
}
@ -375,8 +375,8 @@ static int transient_analysis(const opus_val32 * OPUS_RESTRICT in, int len, int
/* Looks for sudden increases of energy to decide whether we need to patch
the transient decision */
int patch_transient_decision(opus_val16 *newE, opus_val16 *oldE, int nbEBands,
int end, int C)
static int patch_transient_decision(opus_val16 *newE, opus_val16 *oldE, int nbEBands,
int start, int end, int C)
{
int i, c;
opus_val32 mean_diff=0;
@ -385,28 +385,28 @@ int patch_transient_decision(opus_val16 *newE, opus_val16 *oldE, int nbEBands,
avoid false detection caused by irrelevant bands */
if (C==1)
{
spread_old[0] = oldE[0];
for (i=1;i<end;i++)
spread_old[start] = oldE[start];
for (i=start+1;i<end;i++)
spread_old[i] = MAX16(spread_old[i-1]-QCONST16(1.0f, DB_SHIFT), oldE[i]);
} else {
spread_old[0] = MAX16(oldE[0],oldE[nbEBands]);
for (i=1;i<end;i++)
spread_old[start] = MAX16(oldE[start],oldE[start+nbEBands]);
for (i=start+1;i<end;i++)
spread_old[i] = MAX16(spread_old[i-1]-QCONST16(1.0f, DB_SHIFT),
MAX16(oldE[i],oldE[i+nbEBands]));
}
for (i=end-2;i>=0;i--)
for (i=end-2;i>=start;i--)
spread_old[i] = MAX16(spread_old[i], spread_old[i+1]-QCONST16(1.0f, DB_SHIFT));
/* Compute mean increase */
c=0; do {
for (i=2;i<end-1;i++)
for (i=IMAX(2,start);i<end-1;i++)
{
opus_val16 x1, x2;
x1 = MAX16(0, newE[i]);
x1 = MAX16(0, newE[i + c*nbEBands]);
x2 = MAX16(0, spread_old[i]);
mean_diff = ADD32(mean_diff, EXTEND32(MAX16(0, SUB16(x1, x2))));
}
} while (++c<C);
mean_diff = DIV32(mean_diff, C*(end-3));
mean_diff = DIV32(mean_diff, C*(end-1-IMAX(2,start)));
/*printf("%f %f %d\n", mean_diff, max_diff, count);*/
return mean_diff > QCONST16(1.f, DB_SHIFT);
}
@ -1735,7 +1735,7 @@ int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm,
time-domain analysis */
if (LM>0 && ec_tell(enc)+3<=total_bits && !isTransient && st->complexity>=5 && !st->lfe)
{
if (patch_transient_decision(bandLogE, oldBandE, nbEBands, end, C))
if (patch_transient_decision(bandLogE, oldBandE, nbEBands, start, end, C))
{
isTransient = 1;
shortBlocks = M;

8
media/libopus/celt/cpu_support.h
Просмотреть файл

@ -45,16 +45,18 @@
#elif (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_SSE4_1) && !defined(OPUS_X86_PRESUME_SSE4_1)) || \
(defined(OPUS_X86_MAY_HAVE_AVX) && !defined(OPUS_X86_PRESUME_AVX))
#include "x86/x86cpu.h"
/* We currently support 4 x86 variants:
/* 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 3
#define OPUS_ARCHMASK 7
int opus_select_arch(void);
#else

2
media/libopus/celt/quant_bands.c
Просмотреть файл

@ -292,7 +292,7 @@ void quant_coarse_energy(const CELTMode *m, int start, int end, int effEnd,
#endif
}
if (lfe)
max_decay=3;
max_decay = QCONST16(3.f,DB_SHIFT);
enc_start_state = *enc;
ALLOC(oldEBands_intra, C*m->nbEBands, opus_val16);

2
media/libopus/celt/rate.c
Просмотреть файл

@ -131,7 +131,7 @@ void compute_pulse_cache(CELTMode *m, int LM)
for (i=0;i<nbEntries;i++)
{
unsigned char *ptr = bits+entryI[i];
opus_int16 tmp[MAX_PULSES+1];
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;

2
media/libopus/celt/rate.c.orig
Просмотреть файл

@ -131,7 +131,7 @@ void compute_pulse_cache(CELTMode *m, int LM)
for (i=0;i<nbEntries;i++)
{
unsigned char *ptr = bits+entryI[i];
opus_int16 tmp[MAX_PULSES+1];
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;

2
media/libopus/celt/rate.h
Просмотреть файл

@ -32,7 +32,7 @@
#define MAX_PSEUDO 40
#define LOG_MAX_PSEUDO 6
#define MAX_PULSES 128
#define CELT_MAX_PULSES 128
#define MAX_FINE_BITS 8

132
media/libopus/celt/x86/celt_lpc_sse.c Normal file
Просмотреть файл

@ -0,0 +1,132 @@
/* 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,
opus_val16 *mem,
int arch)
{
int i,j;
VARDECL(opus_val16, rnum);
VARDECL(opus_val16, x);
__m128i vecNoA;
opus_int32 noA ;
SAVE_STACK;
ALLOC(rnum, ord, opus_val16);
ALLOC(x, N+ord, opus_val16);
for(i=0;i<ord;i++)
rnum[i] = num[ord-i-1];
for(i=0;i<ord;i++)
x[i] = mem[ord-i-1];
for (i=0;i<N-7;i+=8)
{
x[i+ord ]=_x[i ];
x[i+ord+1]=_x[i+1];
x[i+ord+2]=_x[i+2];
x[i+ord+3]=_x[i+3];
x[i+ord+4]=_x[i+4];
x[i+ord+5]=_x[i+5];
x[i+ord+6]=_x[i+6];
x[i+ord+7]=_x[i+7];
}
for (;i<N-3;i+=4)
{
x[i+ord ]=_x[i ];
x[i+ord+1]=_x[i+1];
x[i+ord+2]=_x[i+2];
x[i+ord+3]=_x[i+3];
}
for (;i<N;i++)
x[i+ord]=_x[i];
for(i=0;i<ord;i++)
mem[i] = _x[N-i-1];
#ifdef SMALL_FOOTPRINT
for (i=0;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]);
}
_y[i] = SATURATE16(PSHR32(sum, SIG_SHIFT));
}
#else
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;
xcorr_kernel(rnum, x+i, sums, ord, arch);
vecSum = _mm_loadu_si128((__m128i *)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 *)(_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]);
_y[i] = SATURATE16(ADD32(EXTEND32(_x[i]), PSHR32(sum, SIG_SHIFT)));
}
#endif
RESTORE_STACK;
}
#endif

95
media/libopus/celt/x86/pitch_sse2.c Normal file
Просмотреть файл

@ -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 *)(&x[i + 0]));
inVec2_76543210 = _mm_loadu_si128((__m128i *)(&y[i + 0]));
inVec1_FEDCBA98 = _mm_loadu_si128((__m128i *)(&x[i + 8]));
inVec2_FEDCBA98 = _mm_loadu_si128((__m128i *)(&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 *)(&x[i + 0]));
inVec2_76543210 = _mm_loadu_si128((__m128i *)(&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

7
media/libopus/celt/x86/x86_celt_map.c
Просмотреть файл

@ -53,6 +53,7 @@ void (*const CELT_FIR_IMPL[OPUS_ARCHMASK + 1])(
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])(
@ -65,6 +66,7 @@ void (*const XCORR_KERNEL_IMPL[OPUS_ARCHMASK + 1])(
xcorr_kernel_c,
xcorr_kernel_c,
MAY_HAVE_SSE4_1(xcorr_kernel), /* sse4.1 */
MAY_HAVE_SSE4_1(xcorr_kernel) /* avx */
};
#endif
@ -81,6 +83,7 @@ opus_val32 (*const CELT_INNER_PROD_IMPL[OPUS_ARCHMASK + 1])(
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
@ -99,6 +102,7 @@ void (*const XCORR_KERNEL_IMPL[OPUS_ARCHMASK + 1])(
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])(
@ -110,6 +114,7 @@ opus_val32 (*const CELT_INNER_PROD_IMPL[OPUS_ARCHMASK + 1])(
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])(
@ -124,6 +129,7 @@ void (*const DUAL_INNER_PROD_IMPL[OPUS_ARCHMASK + 1])(
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])(
@ -139,6 +145,7 @@ void (*const COMB_FILTER_CONST_IMPL[OPUS_ARCHMASK + 1])(
MAY_HAVE_SSE(comb_filter_const),
MAY_HAVE_SSE(comb_filter_const),
MAY_HAVE_SSE(comb_filter_const),
MAY_HAVE_SSE(comb_filter_const)
};

13
media/libopus/celt/x86/x86cpu.c
Просмотреть файл

@ -37,7 +37,8 @@
#if (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_SSE4_1) && !defined(OPUS_X86_PRESUME_SSE4_1)) || \
(defined(OPUS_X86_MAY_HAVE_AVX) && !defined(OPUS_X86_PRESUME_AVX))
#if defined(_MSC_VER)
@ -91,6 +92,8 @@ typedef struct CPU_Feature{
int HW_SSE;
int HW_SSE2;
int HW_SSE41;
/* SIMD: 256-bit */
int HW_AVX;
} CPU_Feature;
static void opus_cpu_feature_check(CPU_Feature *cpu_feature)
@ -106,11 +109,13 @@ static void opus_cpu_feature_check(CPU_Feature *cpu_feature)
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_AVX = (info[2] & (1 << 28)) != 0;
}
else {
cpu_feature->HW_SSE = 0;
cpu_feature->HW_SSE2 = 0;
cpu_feature->HW_SSE41 = 0;
cpu_feature->HW_AVX = 0;
}
}
@ -140,6 +145,12 @@ int opus_select_arch(void)
}
arch++;
if (!cpu_feature.HW_AVX)
{
return arch;
}
arch++;
return arch;
}

6
media/libopus/celt/x86/x86cpu.h
Просмотреть файл

@ -46,6 +46,12 @@
# define MAY_HAVE_SSE4_1(name) name ## _c
# endif
# if defined(OPUS_X86_MAY_HAVE_AVX)
# define MAY_HAVE_AVX(name) name ## _avx
# else
# define MAY_HAVE_AVX(name) name ## _c
# endif
# if defined(OPUS_HAVE_RTCD)
int opus_select_arch(void);
# endif

9
media/libopus/include/opus_defines.h
Просмотреть файл

@ -523,10 +523,19 @@ extern "C" {
* @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).

2
media/libopus/moz.build
Просмотреть файл

@ -20,7 +20,7 @@ ALLOW_COMPILER_WARNINGS = True
FINAL_LIBRARY = 'gkmedias'
DEFINES['OPUS_BUILD'] = True
DEFINES['OPUS_VERSION'] = '"v1.1.1-mozilla"'
DEFINES['OPUS_VERSION'] = '"v1.1.2-mozilla"'
DEFINES['USE_ALLOCA'] = True
# Don't export symbols

4
media/libopus/silk/decode_pulses.c
Просмотреть файл

@ -69,9 +69,9 @@ void silk_decode_pulses(
sum_pulses[ i ] = ec_dec_icdf( psRangeDec, cdf_ptr, 8 );
/* LSB indication */
while( sum_pulses[ i ] == MAX_PULSES + 1 ) {
while( sum_pulses[ i ] == SILK_MAX_PULSES + 1 ) {
nLshifts[ i ]++;
/* When we've already got 10 LSBs, we shift the table to not allow (MAX_PULSES + 1) */
/* When we've already got 10 LSBs, we shift the table to not allow (SILK_MAX_PULSES + 1) */
sum_pulses[ i ] = ec_dec_icdf( psRangeDec,
silk_pulses_per_block_iCDF[ N_RATE_LEVELS - 1] + ( nLshifts[ i ] == 10 ), 8 );
}

2
media/libopus/silk/define.h
Просмотреть файл

@ -169,7 +169,7 @@ extern "C"
#define N_RATE_LEVELS 10
/* Maximum sum of pulses per shell coding frame */
#define MAX_PULSES 16
#define SILK_MAX_PULSES 16
#define MAX_MATRIX_SIZE MAX_LPC_ORDER /* Max of LPC Order and LTP order */

6
media/libopus/silk/encode_pulses.c
Просмотреть файл

@ -142,7 +142,7 @@ void silk_encode_pulses(
sumBits_Q5 = silk_rate_levels_BITS_Q5[ signalType >> 1 ][ k ];
for( i = 0; i < iter; i++ ) {
if( nRshifts[ i ] > 0 ) {
sumBits_Q5 += nBits_ptr[ MAX_PULSES + 1 ];
sumBits_Q5 += nBits_ptr[ SILK_MAX_PULSES + 1 ];
} else {
sumBits_Q5 += nBits_ptr[ sum_pulses[ i ] ];
}
@ -162,9 +162,9 @@ void silk_encode_pulses(
if( nRshifts[ i ] == 0 ) {
ec_enc_icdf( psRangeEnc, sum_pulses[ i ], cdf_ptr, 8 );
} else {
ec_enc_icdf( psRangeEnc, MAX_PULSES + 1, cdf_ptr, 8 );
ec_enc_icdf( psRangeEnc, SILK_MAX_PULSES + 1, cdf_ptr, 8 );
for( k = 0; k < nRshifts[ i ] - 1; k++ ) {
ec_enc_icdf( psRangeEnc, MAX_PULSES + 1, silk_pulses_per_block_iCDF[ N_RATE_LEVELS - 1 ], 8 );
ec_enc_icdf( psRangeEnc, SILK_MAX_PULSES + 1, silk_pulses_per_block_iCDF[ N_RATE_LEVELS - 1 ], 8 );
}
ec_enc_icdf( psRangeEnc, sum_pulses[ i ], silk_pulses_per_block_iCDF[ N_RATE_LEVELS - 1 ], 8 );
}

88
media/libopus/silk/fixed/x86/vector_ops_FIX_sse.c Normal file
Просмотреть файл

@ -0,0 +1,88 @@
/* 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 "main.h"
#include "SigProc_FIX.h"
#include "pitch.h"
opus_int64 silk_inner_prod16_aligned_64_sse4_1(
const opus_int16 *inVec1, /* I input vector 1 */
const opus_int16 *inVec2, /* I input vector 2 */
const opus_int len /* I vector lengths */
)
{
opus_int i, dataSize8;
opus_int64 sum;
__m128i xmm_tempa;
__m128i inVec1_76543210, acc1;
__m128i inVec2_76543210, acc2;
sum = 0;
dataSize8 = len & ~7;
acc1 = _mm_setzero_si128();
acc2 = _mm_setzero_si128();
for( i = 0; i < dataSize8; i += 8 ) {
inVec1_76543210 = _mm_loadu_si128( (__m128i *)(&inVec1[i + 0] ) );
inVec2_76543210 = _mm_loadu_si128( (__m128i *)(&inVec2[i + 0] ) );
/* only when all 4 operands are -32768 (0x8000), this results in wrap around */
inVec1_76543210 = _mm_madd_epi16( inVec1_76543210, inVec2_76543210 );
xmm_tempa = _mm_cvtepi32_epi64( inVec1_76543210 );
/* equal shift right 8 bytes */
inVec1_76543210 = _mm_shuffle_epi32( inVec1_76543210, _MM_SHUFFLE( 0, 0, 3, 2 ) );
inVec1_76543210 = _mm_cvtepi32_epi64( inVec1_76543210 );
acc1 = _mm_add_epi64( acc1, xmm_tempa );
acc2 = _mm_add_epi64( acc2, inVec1_76543210 );
}
acc1 = _mm_add_epi64( acc1, acc2 );
/* equal shift right 8 bytes */
acc2 = _mm_shuffle_epi32( acc1, _MM_SHUFFLE( 0, 0, 3, 2 ) );
acc1 = _mm_add_epi64( acc1, acc2 );
_mm_storel_epi64( (__m128i *)&sum, acc1 );
for( ; i < len; i++ ) {
sum = silk_SMLABB( sum, inVec1[ i ], inVec2[ i ] );
}
return sum;
}

6
media/libopus/silk/tables.h
Просмотреть файл

@ -47,8 +47,8 @@ extern const opus_uint8 silk_pitch_contour_NB_iCDF[ 11 ];
extern const opus_uint8 silk_pitch_contour_10_ms_iCDF[ 12 ]; /* 12 */
extern const opus_uint8 silk_pitch_contour_10_ms_NB_iCDF[ 3 ]; /* 3 */
extern const opus_uint8 silk_pulses_per_block_iCDF[ N_RATE_LEVELS ][ MAX_PULSES + 2 ]; /* 180 */
extern const opus_uint8 silk_pulses_per_block_BITS_Q5[ N_RATE_LEVELS - 1 ][ MAX_PULSES + 2 ]; /* 162 */
extern const opus_uint8 silk_pulses_per_block_iCDF[ N_RATE_LEVELS ][ SILK_MAX_PULSES + 2 ]; /* 180 */
extern const opus_uint8 silk_pulses_per_block_BITS_Q5[ N_RATE_LEVELS - 1 ][ SILK_MAX_PULSES + 2 ]; /* 162 */
extern const opus_uint8 silk_rate_levels_iCDF[ 2 ][ N_RATE_LEVELS - 1 ]; /* 18 */
extern const opus_uint8 silk_rate_levels_BITS_Q5[ 2 ][ N_RATE_LEVELS - 1 ]; /* 18 */
@ -59,7 +59,7 @@ extern const opus_uint8 silk_shell_code_table0[ 152 ];
extern const opus_uint8 silk_shell_code_table1[ 152 ]; /* 152 */
extern const opus_uint8 silk_shell_code_table2[ 152 ]; /* 152 */
extern const opus_uint8 silk_shell_code_table3[ 152 ]; /* 152 */
extern const opus_uint8 silk_shell_code_table_offsets[ MAX_PULSES + 1 ]; /* 17 */
extern const opus_uint8 silk_shell_code_table_offsets[ SILK_MAX_PULSES + 1 ]; /* 17 */
extern const opus_uint8 silk_lsb_iCDF[ 2 ]; /* 2 */

720
media/libopus/silk/x86/NSQ_sse.c Normal file
Просмотреть файл

@ -0,0 +1,720 @@
/* 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 "main.h"
#include "celt/x86/x86cpu.h"
#include "stack_alloc.h"
static OPUS_INLINE void silk_nsq_scale_states_sse4_1(
const silk_encoder_state *psEncC, /* I Encoder State */
silk_nsq_state *NSQ, /* I/O NSQ state */
const opus_int32 x_Q3[], /* I input in Q3 */
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 */
);
static OPUS_INLINE void silk_noise_shape_quantizer_10_16_sse4_1(
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 offset_Q10, /* I */
opus_int length, /* I Input length */
opus_int32 table[][4] /* I */
);
void silk_NSQ_sse4_1(
const silk_encoder_state *psEncC, /* I/O Encoder State */
silk_nsq_state *NSQ, /* I/O NSQ state */
SideInfoIndices *psIndices, /* I/O Quantization Indices */
const opus_int32 x_Q3[], /* I Prefiltered input signal */
opus_int8 pulses[], /* O Quantized pulse signal */
const opus_int16 PredCoef_Q12[ 2 * MAX_LPC_ORDER ], /* I Short term prediction coefs */
const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ], /* I Long term prediction coefs */
const opus_int16 AR2_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 );
opus_int32 table[ 64 ][ 4 ];
opus_int32 tmp1;
opus_int32 q1_Q10, q2_Q10, rd1_Q20, rd2_Q20;
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 ];
/* 0 */
q1_Q10 = offset_Q10;
q2_Q10 = offset_Q10 + ( 1024 - QUANT_LEVEL_ADJUST_Q10 );
rd1_Q20 = q1_Q10 * Lambda_Q10;
rd2_Q20 = q2_Q10 * Lambda_Q10;
table[ 32 ][ 0 ] = q1_Q10;
table[ 32 ][ 1 ] = q2_Q10;
table[ 32 ][ 2 ] = 2 * (q1_Q10 - q2_Q10);
table[ 32 ][ 3 ] = (rd1_Q20 - rd2_Q20) + (q1_Q10 * q1_Q10 - q2_Q10 * q2_Q10);
/* -1 */
q1_Q10 = offset_Q10 - ( 1024 - QUANT_LEVEL_ADJUST_Q10 );
q2_Q10 = offset_Q10;
rd1_Q20 = - q1_Q10 * Lambda_Q10;
rd2_Q20 = q2_Q10 * Lambda_Q10;
table[ 31 ][ 0 ] = q1_Q10;
table[ 31 ][ 1 ] = q2_Q10;
table[ 31 ][ 2 ] = 2 * (q1_Q10 - q2_Q10);
table[ 31 ][ 3 ] = (rd1_Q20 - rd2_Q20) + (q1_Q10 * q1_Q10 - q2_Q10 * q2_Q10);
/* > 0 */
for (k = 1; k <= 31; k++)
{
tmp1 = offset_Q10 + silk_LSHIFT( k, 10 );
q1_Q10 = tmp1 - QUANT_LEVEL_ADJUST_Q10;
q2_Q10 = tmp1 - QUANT_LEVEL_ADJUST_Q10 + 1024;
rd1_Q20 = q1_Q10 * Lambda_Q10;
rd2_Q20 = q2_Q10 * Lambda_Q10;
table[ 32 + k ][ 0 ] = q1_Q10;
table[ 32 + k ][ 1 ] = q2_Q10;
table[ 32 + k ][ 2 ] = 2 * (q1_Q10 - q2_Q10);
table[ 32 + k ][ 3 ] = (rd1_Q20 - rd2_Q20) + (q1_Q10 * q1_Q10 - q2_Q10 * q2_Q10);
}
/* < -1 */
for (k = -32; k <= -2; k++)
{
tmp1 = offset_Q10 + silk_LSHIFT( k, 10 );
q1_Q10 = tmp1 + QUANT_LEVEL_ADJUST_Q10;
q2_Q10 = tmp1 + QUANT_LEVEL_ADJUST_Q10 + 1024;
rd1_Q20 = - q1_Q10 * Lambda_Q10;
rd2_Q20 = - q2_Q10 * Lambda_Q10;
table[ 32 + k ][ 0 ] = q1_Q10;
table[ 32 + k ][ 1 ] = q2_Q10;
table[ 32 + k ][ 2 ] = 2 * (q1_Q10 - q2_Q10);
table[ 32 + k ][ 3 ] = (rd1_Q20 - rd2_Q20) + (q1_Q10 * q1_Q10 - q2_Q10 * q2_Q10);
}
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 = &AR2_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;
silk_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_sse4_1( psEncC, NSQ, x_Q3, x_sc_Q10, sLTP, sLTP_Q15, k, LTP_scale_Q14, Gains_Q16, pitchL, psIndices->signalType );
if ( opus_likely( ( 10 == psEncC->shapingLPCOrder ) && ( 16 == psEncC->predictLPCOrder) ) )
{
silk_noise_shape_quantizer_10_16_sse4_1( 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 ],
offset_Q10, psEncC->subfr_length, &(table[32]) );
}
else
{
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 );
}
x_Q3 += 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 */
/* DEBUG_STORE_DATA( enc.pcm, &NSQ->xq[ psEncC->ltp_mem_length ], psEncC->frame_length * sizeof( opus_int16 ) ) */
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_10_16 */
/***********************************/
static OPUS_INLINE void silk_noise_shape_quantizer_10_16_sse4_1(
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 offset_Q10, /* I */
opus_int length, /* I Input length */
opus_int32 table[][4] /* I */
)
{
opus_int i;
opus_int32 LTP_pred_Q13, LPC_pred_Q10, n_AR_Q12, n_LTP_Q13;
opus_int32 n_LF_Q12, r_Q10, q1_Q0, q1_Q10, q2_Q10;
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;
__m128i xmm_tempa, xmm_tempb;
__m128i xmm_one;
__m128i psLPC_Q14_hi_01234567, psLPC_Q14_hi_89ABCDEF;
__m128i psLPC_Q14_lo_01234567, psLPC_Q14_lo_89ABCDEF;
__m128i a_Q12_01234567, a_Q12_89ABCDEF;
__m128i sAR2_Q14_hi_76543210, sAR2_Q14_lo_76543210;
__m128i AR_shp_Q13_76543210;
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 ];
sLF_AR_shp_Q14 = NSQ->sLF_AR_shp_Q14;
xq_Q14 = psLPC_Q14[ 0 ];
LTP_pred_Q13 = 0;
/* load a_Q12 */
xmm_one = _mm_set_epi8( 1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14 );
/* load a_Q12[0] - a_Q12[7] */
a_Q12_01234567 = _mm_loadu_si128( (__m128i *)(&a_Q12[ 0 ] ) );
/* load a_Q12[ 8 ] - a_Q12[ 15 ] */
a_Q12_89ABCDEF = _mm_loadu_si128( (__m128i *)(&a_Q12[ 8 ] ) );
a_Q12_01234567 = _mm_shuffle_epi8( a_Q12_01234567, xmm_one );
a_Q12_89ABCDEF = _mm_shuffle_epi8( a_Q12_89ABCDEF, xmm_one );
/* load AR_shp_Q13 */
AR_shp_Q13_76543210 = _mm_loadu_si128( (__m128i *)(&AR_shp_Q13[0] ) );
/* load psLPC_Q14 */
xmm_one = _mm_set_epi8(15, 14, 11, 10, 7, 6, 3, 2, 13, 12, 9, 8, 5, 4, 1, 0 );
xmm_tempa = _mm_loadu_si128( (__m128i *)(&psLPC_Q14[-16]) );
xmm_tempb = _mm_loadu_si128( (__m128i *)(&psLPC_Q14[-12]) );
xmm_tempa = _mm_shuffle_epi8( xmm_tempa, xmm_one );
xmm_tempb = _mm_shuffle_epi8( xmm_tempb, xmm_one );
psLPC_Q14_hi_89ABCDEF = _mm_unpackhi_epi64( xmm_tempa, xmm_tempb );
psLPC_Q14_lo_89ABCDEF = _mm_unpacklo_epi64( xmm_tempa, xmm_tempb );
xmm_tempa = _mm_loadu_si128( (__m128i *)(&psLPC_Q14[ -8 ]) );
xmm_tempb = _mm_loadu_si128( (__m128i *)(&psLPC_Q14[ -4 ]) );
xmm_tempa = _mm_shuffle_epi8( xmm_tempa, xmm_one );
xmm_tempb = _mm_shuffle_epi8( xmm_tempb, xmm_one );
psLPC_Q14_hi_01234567 = _mm_unpackhi_epi64( xmm_tempa, xmm_tempb );
psLPC_Q14_lo_01234567 = _mm_unpacklo_epi64( xmm_tempa, xmm_tempb );
/* load sAR2_Q14 */
xmm_tempa = _mm_loadu_si128( (__m128i *)(&(NSQ->sAR2_Q14[ 0 ]) ) );
xmm_tempb = _mm_loadu_si128( (__m128i *)(&(NSQ->sAR2_Q14[ 4 ]) ) );
xmm_tempa = _mm_shuffle_epi8( xmm_tempa, xmm_one );
xmm_tempb = _mm_shuffle_epi8( xmm_tempb, xmm_one );
sAR2_Q14_hi_76543210 = _mm_unpackhi_epi64( xmm_tempa, xmm_tempb );
sAR2_Q14_lo_76543210 = _mm_unpacklo_epi64( xmm_tempa, xmm_tempb );
/* prepare 1 in 8 * 16bit */
xmm_one = _mm_set1_epi16(1);
for( i = 0; i < length; i++ )
{
/* Short-term prediction */
__m128i xmm_hi_07, xmm_hi_8F, xmm_lo_07, xmm_lo_8F;
/* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */
LPC_pred_Q10 = 8; /* silk_RSHIFT( predictLPCOrder, 1 ); */
/* shift psLPC_Q14 */
psLPC_Q14_hi_89ABCDEF = _mm_alignr_epi8( psLPC_Q14_hi_01234567, psLPC_Q14_hi_89ABCDEF, 2 );
psLPC_Q14_lo_89ABCDEF = _mm_alignr_epi8( psLPC_Q14_lo_01234567, psLPC_Q14_lo_89ABCDEF, 2 );
psLPC_Q14_hi_01234567 = _mm_srli_si128( psLPC_Q14_hi_01234567, 2 );
psLPC_Q14_lo_01234567 = _mm_srli_si128( psLPC_Q14_lo_01234567, 2 );
psLPC_Q14_hi_01234567 = _mm_insert_epi16( psLPC_Q14_hi_01234567, (xq_Q14 >> 16), 7 );
psLPC_Q14_lo_01234567 = _mm_insert_epi16( psLPC_Q14_lo_01234567, (xq_Q14), 7 );
/* high part, use pmaddwd, results in 4 32-bit */
xmm_hi_07 = _mm_madd_epi16( psLPC_Q14_hi_01234567, a_Q12_01234567 );
xmm_hi_8F = _mm_madd_epi16( psLPC_Q14_hi_89ABCDEF, a_Q12_89ABCDEF );
/* low part, use pmulhw, results in 8 16-bit, note we need simulate unsigned * signed, _mm_srai_epi16(psLPC_Q14_lo_01234567, 15) */
xmm_tempa = _mm_cmpgt_epi16( _mm_setzero_si128(), psLPC_Q14_lo_01234567 );
xmm_tempb = _mm_cmpgt_epi16( _mm_setzero_si128(), psLPC_Q14_lo_89ABCDEF );
xmm_tempa = _mm_and_si128( xmm_tempa, a_Q12_01234567 );
xmm_tempb = _mm_and_si128( xmm_tempb, a_Q12_89ABCDEF );
xmm_lo_07 = _mm_mulhi_epi16( psLPC_Q14_lo_01234567, a_Q12_01234567 );
xmm_lo_8F = _mm_mulhi_epi16( psLPC_Q14_lo_89ABCDEF, a_Q12_89ABCDEF );
xmm_lo_07 = _mm_add_epi16( xmm_lo_07, xmm_tempa );
xmm_lo_8F = _mm_add_epi16( xmm_lo_8F, xmm_tempb );
xmm_lo_07 = _mm_madd_epi16( xmm_lo_07, xmm_one );
xmm_lo_8F = _mm_madd_epi16( xmm_lo_8F, xmm_one );
/* accumulate */
xmm_hi_07 = _mm_add_epi32( xmm_hi_07, xmm_hi_8F );
xmm_lo_07 = _mm_add_epi32( xmm_lo_07, xmm_lo_8F );
xmm_hi_07 = _mm_add_epi32( xmm_hi_07, xmm_lo_07 );
xmm_hi_07 = _mm_add_epi32( xmm_hi_07, _mm_unpackhi_epi64(xmm_hi_07, xmm_hi_07 ) );
xmm_hi_07 = _mm_add_epi32( xmm_hi_07, _mm_shufflelo_epi16(xmm_hi_07, 0x0E ) );
LPC_pred_Q10 += _mm_cvtsi128_si32( xmm_hi_07 );
/* Long-term prediction */
if ( opus_likely( signalType == TYPE_VOICED ) ) {
/* Unrolled loop */
/* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */
LTP_pred_Q13 = 2;
{
__m128i b_Q14_3210, b_Q14_0123, pred_lag_ptr_0123;
b_Q14_3210 = OP_CVTEPI16_EPI32_M64( b_Q14 );
b_Q14_0123 = _mm_shuffle_epi32( b_Q14_3210, 0x1B );
/* loaded: [0] [-1] [-2] [-3] */
pred_lag_ptr_0123 = _mm_loadu_si128( (__m128i *)(&pred_lag_ptr[ -3 ] ) );
/* shuffle to [-3] [-2] [-1] [0] and to new xmm */
xmm_tempa = _mm_shuffle_epi32( pred_lag_ptr_0123, 0x1B );
/*64-bit multiply, a[2] * b[-2], a[0] * b[0] */
xmm_tempa = _mm_mul_epi32( xmm_tempa, b_Q14_3210 );
/* right shift 2 bytes (16 bits), zero extended */
xmm_tempa = _mm_srli_si128( xmm_tempa, 2 );
/* a[1] * b[-1], a[3] * b[-3] */
pred_lag_ptr_0123 = _mm_mul_epi32( pred_lag_ptr_0123, b_Q14_0123 );
pred_lag_ptr_0123 = _mm_srli_si128( pred_lag_ptr_0123, 2 );
pred_lag_ptr_0123 = _mm_add_epi32( pred_lag_ptr_0123, xmm_tempa );
/* equal shift right 8 bytes*/
xmm_tempa = _mm_shuffle_epi32( pred_lag_ptr_0123, _MM_SHUFFLE( 0, 0, 3, 2 ) );
xmm_tempa = _mm_add_epi32( xmm_tempa, pred_lag_ptr_0123 );
LTP_pred_Q13 += _mm_cvtsi128_si32( xmm_tempa );
LTP_pred_Q13 = silk_SMLAWB( LTP_pred_Q13, pred_lag_ptr[ -4 ], b_Q14[ 4 ] );
pred_lag_ptr++;
}
}
/* Noise shape feedback */
NSQ->sAR2_Q14[ 9 ] = NSQ->sAR2_Q14[ 8 ];
NSQ->sAR2_Q14[ 8 ] = _mm_cvtsi128_si32( _mm_srli_si128(_mm_unpackhi_epi16( sAR2_Q14_lo_76543210, sAR2_Q14_hi_76543210 ), 12 ) );
sAR2_Q14_hi_76543210 = _mm_slli_si128( sAR2_Q14_hi_76543210, 2 );
sAR2_Q14_lo_76543210 = _mm_slli_si128( sAR2_Q14_lo_76543210, 2 );
sAR2_Q14_hi_76543210 = _mm_insert_epi16( sAR2_Q14_hi_76543210, (xq_Q14 >> 16), 0 );
sAR2_Q14_lo_76543210 = _mm_insert_epi16( sAR2_Q14_lo_76543210, (xq_Q14), 0 );
/* high part, use pmaddwd, results in 4 32-bit */
xmm_hi_07 = _mm_madd_epi16( sAR2_Q14_hi_76543210, AR_shp_Q13_76543210 );
/* low part, use pmulhw, results in 8 16-bit, note we need simulate unsigned * signed,_mm_srai_epi16(sAR2_Q14_lo_76543210, 15) */
xmm_tempa = _mm_cmpgt_epi16( _mm_setzero_si128(), sAR2_Q14_lo_76543210 );
xmm_tempa = _mm_and_si128( xmm_tempa, AR_shp_Q13_76543210 );
xmm_lo_07 = _mm_mulhi_epi16( sAR2_Q14_lo_76543210, AR_shp_Q13_76543210 );
xmm_lo_07 = _mm_add_epi16( xmm_lo_07, xmm_tempa );
xmm_lo_07 = _mm_madd_epi16( xmm_lo_07, xmm_one );
/* accumulate */
xmm_hi_07 = _mm_add_epi32( xmm_hi_07, xmm_lo_07 );
xmm_hi_07 = _mm_add_epi32( xmm_hi_07, _mm_unpackhi_epi64(xmm_hi_07, xmm_hi_07 ) );
xmm_hi_07 = _mm_add_epi32( xmm_hi_07, _mm_shufflelo_epi16(xmm_hi_07, 0x0E ) );
n_AR_Q12 = 5 + _mm_cvtsi128_si32( xmm_hi_07 );
n_AR_Q12 = silk_SMLAWB( n_AR_Q12, NSQ->sAR2_Q14[ 8 ], AR_shp_Q13[ 8 ] );
n_AR_Q12 = silk_SMLAWB( n_AR_Q12, NSQ->sAR2_Q14[ 9 ], AR_shp_Q13[ 9 ] );
n_AR_Q12 = silk_LSHIFT32( n_AR_Q12, 1 ); /* Q11 -> Q12 */
n_AR_Q12 = silk_SMLAWB( n_AR_Q12, 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, sLF_AR_shp_Q14, LF_shp_Q14 );
silk_assert( lag > 0 || signalType != TYPE_VOICED );
/* Combine prediction and noise shaping signals */
tmp1 = silk_SUB32( silk_LSHIFT32( LPC_pred_Q10, 2 ), n_AR_Q12 ); /* Q12 */
tmp1 = silk_SUB32( tmp1, n_LF_Q12 ); /* Q12 */
if( lag > 0 ) {
/* Symmetric, packed FIR coefficients */
n_LTP_Q13 = silk_SMULWB( silk_ADD32( 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_ADD_LSHIFT32( tmp2, 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 */
/* Generate dither */
NSQ->rand_seed = silk_RAND( NSQ->rand_seed );
/* Flip sign depending on dither */
tmp2 = -r_Q10;
if ( NSQ->rand_seed < 0 ) r_Q10 = tmp2;
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 );
q1_Q10 = table[q1_Q0][0];
q2_Q10 = table[q1_Q0][1];
if (r_Q10 * table[q1_Q0][2] - table[q1_Q0][3] < 0)
{
q1_Q10 = q2_Q10;
}
pulses[ i ] = (opus_int8)silk_RSHIFT_ROUND( q1_Q10, 10 );
/* Excitation */
exc_Q14 = silk_LSHIFT( q1_Q10, 4 );
tmp2 = -exc_Q14;
if ( NSQ->rand_seed < 0 ) exc_Q14 = tmp2;
/* Add predictions */
LPC_exc_Q14 = silk_ADD_LSHIFT32( exc_Q14, LTP_pred_Q13, 1 );
xq_Q14 = silk_ADD_LSHIFT32( LPC_exc_Q14, LPC_pred_Q10, 4 );
/* Update states */
psLPC_Q14++;
*psLPC_Q14 = xq_Q14;
sLF_AR_shp_Q14 = silk_SUB_LSHIFT32( xq_Q14, n_AR_Q12, 2 );
NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx ] = silk_SUB_LSHIFT32( sLF_AR_shp_Q14, 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 ] );
}
NSQ->sLF_AR_shp_Q14 = sLF_AR_shp_Q14;
/* Scale XQ back to normal level before saving */
psLPC_Q14 = &NSQ->sLPC_Q14[ NSQ_LPC_BUF_LENGTH ];
/* write back sAR2_Q14 */
xmm_tempa = _mm_unpackhi_epi16( sAR2_Q14_lo_76543210, sAR2_Q14_hi_76543210 );
xmm_tempb = _mm_unpacklo_epi16( sAR2_Q14_lo_76543210, sAR2_Q14_hi_76543210 );
_mm_storeu_si128( (__m128i *)(&NSQ->sAR2_Q14[ 4 ]), xmm_tempa );
_mm_storeu_si128( (__m128i *)(&NSQ->sAR2_Q14[ 0 ]), xmm_tempb );
/* xq[ i ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( psLPC_Q14[ i ], Gain_Q10 ), 8 ) ); */
{
__m128i xmm_Gain_Q10;
__m128i xmm_xq_Q14_3210, xmm_xq_Q14_x3x1, xmm_xq_Q14_7654, xmm_xq_Q14_x7x5;
/* prepare (1 << 7) in packed 4 32-bits */
xmm_tempa = _mm_set1_epi32( (1 << 7) );
/* prepare Gain_Q10 in packed 4 32-bits */
xmm_Gain_Q10 = _mm_set1_epi32( Gain_Q10 );
/* process xq */
for (i = 0; i < length - 7; i += 8)
{
xmm_xq_Q14_3210 = _mm_loadu_si128( (__m128i *)(&(psLPC_Q14[ i + 0 ] ) ) );
xmm_xq_Q14_7654 = _mm_loadu_si128( (__m128i *)(&(psLPC_Q14[ i + 4 ] ) ) );
/* equal shift right 4 bytes*/
xmm_xq_Q14_x3x1 = _mm_shuffle_epi32( xmm_xq_Q14_3210, _MM_SHUFFLE( 0, 3, 2, 1 ) );
/* equal shift right 4 bytes*/
xmm_xq_Q14_x7x5 = _mm_shuffle_epi32( xmm_xq_Q14_7654, _MM_SHUFFLE( 0, 3, 2, 1 ) );
xmm_xq_Q14_3210 = _mm_mul_epi32( xmm_xq_Q14_3210, xmm_Gain_Q10 );
xmm_xq_Q14_x3x1 = _mm_mul_epi32( xmm_xq_Q14_x3x1, xmm_Gain_Q10 );
xmm_xq_Q14_7654 = _mm_mul_epi32( xmm_xq_Q14_7654, xmm_Gain_Q10 );
xmm_xq_Q14_x7x5 = _mm_mul_epi32( xmm_xq_Q14_x7x5, xmm_Gain_Q10 );
xmm_xq_Q14_3210 = _mm_srli_epi64( xmm_xq_Q14_3210, 16 );
xmm_xq_Q14_x3x1 = _mm_slli_epi64( xmm_xq_Q14_x3x1, 16 );
xmm_xq_Q14_7654 = _mm_srli_epi64( xmm_xq_Q14_7654, 16 );
xmm_xq_Q14_x7x5 = _mm_slli_epi64( xmm_xq_Q14_x7x5, 16 );
xmm_xq_Q14_3210 = _mm_blend_epi16( xmm_xq_Q14_3210, xmm_xq_Q14_x3x1, 0xCC );
xmm_xq_Q14_7654 = _mm_blend_epi16( xmm_xq_Q14_7654, xmm_xq_Q14_x7x5, 0xCC );
/* silk_RSHIFT_ROUND(xq, 8) */
xmm_xq_Q14_3210 = _mm_add_epi32( xmm_xq_Q14_3210, xmm_tempa );
xmm_xq_Q14_7654 = _mm_add_epi32( xmm_xq_Q14_7654, xmm_tempa );
xmm_xq_Q14_3210 = _mm_srai_epi32( xmm_xq_Q14_3210, 8 );
xmm_xq_Q14_7654 = _mm_srai_epi32( xmm_xq_Q14_7654, 8 );
/* silk_SAT16 */
xmm_xq_Q14_3210 = _mm_packs_epi32( xmm_xq_Q14_3210, xmm_xq_Q14_7654 );
/* save to xq */
_mm_storeu_si128( (__m128i *)(&xq[ i ] ), xmm_xq_Q14_3210 );
}
}
for ( ; i < length; i++)
{
xq[i] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( silk_SMULWW( psLPC_Q14[ i ], Gain_Q10 ), 8 ) );
}
/* 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_sse4_1(
const silk_encoder_state *psEncC, /* I Encoder State */
silk_nsq_state *NSQ, /* I/O NSQ state */
const opus_int32 x_Q3[], /* I input in Q3 */
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_Q23;
__m128i xmm_inv_gain_Q23, xmm_x_Q3_x2x0, xmm_x_Q3_x3x1;
lag = pitchL[ subfr ];
inv_gain_Q31 = silk_INVERSE32_varQ( silk_max( Gains_Q16[ subfr ], 1 ), 47 );
silk_assert( inv_gain_Q31 != 0 );
/* Calculate gain adjustment factor */
if( Gains_Q16[ subfr ] != NSQ->prev_gain_Q16 ) {
gain_adj_Q16 = silk_DIV32_varQ( NSQ->prev_gain_Q16, Gains_Q16[ subfr ], 16 );
} else {
gain_adj_Q16 = (opus_int32)1 << 16;
}
/* Scale input */
inv_gain_Q23 = silk_RSHIFT_ROUND( inv_gain_Q31, 8 );
/* prepare inv_gain_Q23 in packed 4 32-bits */
xmm_inv_gain_Q23 = _mm_set1_epi32(inv_gain_Q23);
for( i = 0; i < psEncC->subfr_length - 3; i += 4 ) {
xmm_x_Q3_x2x0 = _mm_loadu_si128( (__m128i *)(&(x_Q3[ i ] ) ) );
/* equal shift right 4 bytes*/
xmm_x_Q3_x3x1 = _mm_shuffle_epi32( xmm_x_Q3_x2x0, _MM_SHUFFLE( 0, 3, 2, 1 ) );
xmm_x_Q3_x2x0 = _mm_mul_epi32( xmm_x_Q3_x2x0, xmm_inv_gain_Q23 );
xmm_x_Q3_x3x1 = _mm_mul_epi32( xmm_x_Q3_x3x1, xmm_inv_gain_Q23 );
xmm_x_Q3_x2x0 = _mm_srli_epi64( xmm_x_Q3_x2x0, 16 );
xmm_x_Q3_x3x1 = _mm_slli_epi64( xmm_x_Q3_x3x1, 16 );
xmm_x_Q3_x2x0 = _mm_blend_epi16( xmm_x_Q3_x2x0, xmm_x_Q3_x3x1, 0xCC );
_mm_storeu_si128( (__m128i *)(&(x_sc_Q10[ i ] ) ), xmm_x_Q3_x2x0 );
}
for( ; i < psEncC->subfr_length; i++ ) {
x_sc_Q10[ i ] = silk_SMULWW( x_Q3[ i ], inv_gain_Q23 );
}
/* Save inverse gain */
NSQ->prev_gain_Q16 = Gains_Q16[ subfr ];
/* 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( gain_adj_Q16 != (opus_int32)1 << 16 ) {
/* Scale long-term shaping state */
__m128i xmm_gain_adj_Q16, xmm_sLTP_shp_Q14_x2x0, xmm_sLTP_shp_Q14_x3x1;
/* prepare gain_adj_Q16 in packed 4 32-bits */
xmm_gain_adj_Q16 = _mm_set1_epi32(gain_adj_Q16);
for( i = NSQ->sLTP_shp_buf_idx - psEncC->ltp_mem_length; i < NSQ->sLTP_shp_buf_idx - 3; i += 4 )
{
xmm_sLTP_shp_Q14_x2x0 = _mm_loadu_si128( (__m128i *)(&(NSQ->sLTP_shp_Q14[ i ] ) ) );
/* equal shift right 4 bytes*/
xmm_sLTP_shp_Q14_x3x1 = _mm_shuffle_epi32( xmm_sLTP_shp_Q14_x2x0, _MM_SHUFFLE( 0, 3, 2, 1 ) );
xmm_sLTP_shp_Q14_x2x0 = _mm_mul_epi32( xmm_sLTP_shp_Q14_x2x0, xmm_gain_adj_Q16 );
xmm_sLTP_shp_Q14_x3x1 = _mm_mul_epi32( xmm_sLTP_shp_Q14_x3x1, xmm_gain_adj_Q16 );
xmm_sLTP_shp_Q14_x2x0 = _mm_srli_epi64( xmm_sLTP_shp_Q14_x2x0, 16 );
xmm_sLTP_shp_Q14_x3x1 = _mm_slli_epi64( xmm_sLTP_shp_Q14_x3x1, 16 );
xmm_sLTP_shp_Q14_x2x0 = _mm_blend_epi16( xmm_sLTP_shp_Q14_x2x0, xmm_sLTP_shp_Q14_x3x1, 0xCC );
_mm_storeu_si128( (__m128i *)(&(NSQ->sLTP_shp_Q14[ i ] ) ), xmm_sLTP_shp_Q14_x2x0 );
}
for( ; 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 );
/* 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 ] );
}
}
}

7
media/libopus/silk/x86/x86_silk_map.c
Просмотреть файл

@ -50,6 +50,7 @@ opus_int64 (*const SILK_INNER_PROD16_ALIGNED_64_IMPL[ OPUS_ARCHMASK + 1 ] )(
silk_inner_prod16_aligned_64_c,
silk_inner_prod16_aligned_64_c,
MAY_HAVE_SSE4_1( silk_inner_prod16_aligned_64 ), /* sse4.1 */
MAY_HAVE_SSE4_1( silk_inner_prod16_aligned_64 ) /* avx */
};
#endif
@ -62,6 +63,7 @@ opus_int (*const SILK_VAD_GETSA_Q8_IMPL[ OPUS_ARCHMASK + 1 ] )(
silk_VAD_GetSA_Q8_c,
silk_VAD_GetSA_Q8_c,
MAY_HAVE_SSE4_1( silk_VAD_GetSA_Q8 ), /* sse4.1 */
MAY_HAVE_SSE4_1( silk_VAD_GetSA_Q8 ) /* avx */
};
void (*const SILK_NSQ_IMPL[ OPUS_ARCHMASK + 1 ] )(
@ -85,6 +87,7 @@ void (*const SILK_NSQ_IMPL[ OPUS_ARCHMASK + 1 ] )(
silk_NSQ_c,
silk_NSQ_c,
MAY_HAVE_SSE4_1( silk_NSQ ), /* sse4.1 */
MAY_HAVE_SSE4_1( silk_NSQ ) /* avx */
};
void (*const SILK_VQ_WMAT_EC_IMPL[ OPUS_ARCHMASK + 1 ] )(
@ -104,6 +107,7 @@ void (*const SILK_VQ_WMAT_EC_IMPL[ OPUS_ARCHMASK + 1 ] )(
silk_VQ_WMat_EC_c,
silk_VQ_WMat_EC_c,
MAY_HAVE_SSE4_1( silk_VQ_WMat_EC ), /* sse4.1 */
MAY_HAVE_SSE4_1( silk_VQ_WMat_EC ) /* avx */
};
void (*const SILK_NSQ_DEL_DEC_IMPL[ OPUS_ARCHMASK + 1 ] )(
@ -127,6 +131,7 @@ void (*const SILK_NSQ_DEL_DEC_IMPL[ OPUS_ARCHMASK + 1 ] )(
silk_NSQ_del_dec_c,
silk_NSQ_del_dec_c,
MAY_HAVE_SSE4_1( silk_NSQ_del_dec ), /* sse4.1 */
MAY_HAVE_SSE4_1( silk_NSQ_del_dec ) /* avx */
};
#if defined(FIXED_POINT)
@ -144,6 +149,7 @@ void (*const SILK_WARPED_LPC_ANALYSIS_FILTER_FIX_IMPL[ OPUS_ARCHMASK + 1 ] )(
silk_warped_LPC_analysis_filter_FIX_c,
silk_warped_LPC_analysis_filter_FIX_c,
MAY_HAVE_SSE4_1( silk_warped_LPC_analysis_filter_FIX ), /* sse4.1 */
MAY_HAVE_SSE4_1( silk_warped_LPC_analysis_filter_FIX ) /* avx */
};
void (*const SILK_BURG_MODIFIED_IMPL[ OPUS_ARCHMASK + 1 ] )(
@ -161,6 +167,7 @@ void (*const SILK_BURG_MODIFIED_IMPL[ OPUS_ARCHMASK + 1 ] )(
silk_burg_modified_c,
silk_burg_modified_c,
MAY_HAVE_SSE4_1( silk_burg_modified ), /* sse4.1 */
MAY_HAVE_SSE4_1( silk_burg_modified ) /* avx */
};
#endif

2
media/libopus/sources.mozbuild
Просмотреть файл

@ -19,8 +19,10 @@ celt_sources = [
]
opus_nonunified_sources = [
# Disabled because of name clash of opus_custom_encoder_get_size.
'celt/celt_decoder.c',
'celt/celt_encoder.c',
# Disabled for (safe) warning about QA redefinition.
'silk/LPC_inv_pred_gain.c',
'silk/NLSF2A.c',
]

2
media/libopus/src/opus_decoder.c
Просмотреть файл

@ -59,6 +59,7 @@ struct OpusDecoder {
opus_int32 Fs; /** Sampling rate (at the API level) */
silk_DecControlStruct DecControl;
int decode_gain;
int arch;
/* Everything beyond this point gets cleared on a reset */
#define OPUS_DECODER_RESET_START stream_channels
@ -75,7 +76,6 @@ struct OpusDecoder {
#endif
opus_uint32 rangeFinal;
int arch;
};

6
media/libopus/src/opus_encoder.c
Просмотреть файл

@ -81,6 +81,7 @@ struct OpusEncoder {
int lsb_depth;
int encoder_buffer;
int lfe;
int arch;
#ifndef DISABLE_FLOAT_API
TonalityAnalysisState analysis;
#endif
@ -104,10 +105,8 @@ struct OpusEncoder {
opus_val16 delay_buffer[MAX_ENCODER_BUFFER*2];
#ifndef DISABLE_FLOAT_API
int detected_bandwidth;
int analysis_offset;
#endif
opus_uint32 rangeFinal;
int arch;
};
/* Transition tables for the voice and music. First column is the
@ -2455,12 +2454,13 @@ int opus_encoder_ctl(OpusEncoder *st, int request, ...)
{
void *silk_enc;
silk_EncControlStruct dummy;
char *start;
silk_enc = (char*)st+st->silk_enc_offset;
#ifndef DISABLE_FLOAT_API
tonality_analysis_reset(&st->analysis);
#endif
char *start = (char*)&st->OPUS_ENCODER_RESET_START;
start = (char*)&st->OPUS_ENCODER_RESET_START;
OPUS_CLEAR(start, sizeof(OpusEncoder) - (start - (char*)st));
celt_encoder_ctl(celt_enc, OPUS_RESET_STATE);