ref: c41aeebea836cf2f3a67b3ac5d7bbfb0924ca1e7
dir: /silk/x86/VAD_sse.c/
/* 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 "stack_alloc.h"
/* 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_sse4_1( /* 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 );
silk_assert( MAX_FRAME_LENGTH >= psEncC->frame_length );
silk_assert( psEncC->frame_length <= 512 );
silk_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++ ) {
__m128i xmm_X, xmm_acc;
sumSquared = 0;
xmm_acc = _mm_setzero_si128();
for( i = 0; i < dec_subframe_length - 7; i += 8 )
{
xmm_X = _mm_loadu_si128( (__m128i *)&(X[ X_offset[ b ] + i + dec_subframe_offset ] ) );
xmm_X = _mm_srai_epi16( xmm_X, 3 );
xmm_X = _mm_madd_epi16( xmm_X, xmm_X );
xmm_acc = _mm_add_epi32( xmm_acc, xmm_X );
}
xmm_acc = _mm_add_epi32( xmm_acc, _mm_unpackhi_epi64( xmm_acc, xmm_acc ) );
xmm_acc = _mm_add_epi32( xmm_acc, _mm_shufflelo_epi16( xmm_acc, 0x0E ) );
sumSquared += _mm_cvtsi128_si32( xmm_acc );
for( ; 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 );
}
/* Power scaling */
if( speech_nrg <= 0 ) {
SA_Q15 = silk_RSHIFT( SA_Q15, 1 );
} else if( speech_nrg < 32768 ) {
if( psEncC->frame_length == 10 * psEncC->fs_kHz ) {
speech_nrg = silk_LSHIFT_SAT32( speech_nrg, 16 );
} else {
speech_nrg = silk_LSHIFT_SAT32( speech_nrg, 15 );
}
/* 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 );
}