eden-emu/opus
18
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Add rtcd for silk_inner_product_FLP()

Browse source
This commit is contained in:
Jean-Marc Valin 2023-11-21 02:56:04 -05:00
parent b93e4a149c
commit 239d223d84
No known key found for this signature in database
GPG key ID: 531A52533318F00A
13 changed files with 65 additions and 28 deletions
Download patch file Download diff file Expand all files Collapse all files

8
silk/float/SigProc_FLP.h
View file

@ -74,7 +74,8 @@ void silk_autocorrelation_FLP(
silk_float *results, /* O result (length correlationCount) */ silk_float *results, /* O result (length correlationCount) */
const silk_float *inputData, /* I input data to correlate */ const silk_float *inputData, /* I input data to correlate */
opus_int inputDataSize, /* I length of input */ opus_int inputDataSize, /* I length of input */
opus_int correlationCount /* I number of correlation taps to compute */ opus_int correlationCount, /* I number of correlation taps to compute */
int arch
); );
opus_int silk_pitch_analysis_core_FLP( /* O Voicing estimate: 0 voiced, 1 unvoiced */ opus_int silk_pitch_analysis_core_FLP( /* O Voicing estimate: 0 voiced, 1 unvoiced */
@ -106,7 +107,8 @@ silk_float silk_burg_modified_FLP( /* O returns residual energy
const silk_float minInvGain, /* I minimum inverse prediction gain */ const silk_float minInvGain, /* I minimum inverse prediction gain */
const opus_int subfr_length, /* I input signal subframe length (incl. D preceding samples) */ 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 nb_subfr, /* I number of subframes stacked in x */
const opus_int D /* I order */ const opus_int D, /* I order */
int arch
); );
/* multiply a vector by a constant */ /* multiply a vector by a constant */
@ -132,7 +134,7 @@ double silk_inner_product_FLP_c(
); );
#ifndef OVERRIDE_inner_product_FLP #ifndef OVERRIDE_inner_product_FLP
#define silk_inner_product_FLP(data1, data2, dataSize) silk_inner_product_FLP_c(data1, data2, dataSize) #define silk_inner_product_FLP(data1, data2, dataSize, arch) ((void)arch,silk_inner_product_FLP_c(data1, data2, dataSize))
#endif #endif

5
silk/float/autocorrelation_FLP.c
View file

@ -37,7 +37,8 @@ void silk_autocorrelation_FLP(
silk_float *results, /* O result (length correlationCount) */ silk_float *results, /* O result (length correlationCount) */
const silk_float *inputData, /* I input data to correlate */ const silk_float *inputData, /* I input data to correlate */
opus_int inputDataSize, /* I length of input */ opus_int inputDataSize, /* I length of input */
opus_int correlationCount /* I number of correlation taps to compute */ opus_int correlationCount, /* I number of correlation taps to compute */
int arch
) )
{ {
opus_int i; opus_int i;
@ -47,6 +48,6 @@ void silk_autocorrelation_FLP(
} }
for( i = 0; i < correlationCount; i++ ) { for( i = 0; i < correlationCount; i++ ) {
results[ i ] = (silk_float)silk_inner_product_FLP( inputData, inputData + i, inputDataSize - i ); results[ i ] = (silk_float)silk_inner_product_FLP( inputData, inputData + i, inputDataSize - i, arch );
} }
} }

5
silk/float/burg_modified_FLP.c
View file

@ -42,7 +42,8 @@ silk_float silk_burg_modified_FLP( /* O returns residual energy
const silk_float minInvGain, /* I minimum inverse prediction gain */ const silk_float minInvGain, /* I minimum inverse prediction gain */
const opus_int subfr_length, /* I input signal subframe length (incl. D preceding samples) */ 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 nb_subfr, /* I number of subframes stacked in x */
const opus_int D /* I order */ const opus_int D, /* I order */
int arch
) )
{ {
opus_int k, n, s, reached_max_gain; opus_int k, n, s, reached_max_gain;
@ -60,7 +61,7 @@ silk_float silk_burg_modified_FLP( /* O returns residual energy
for( s = 0; s < nb_subfr; s++ ) { for( s = 0; s < nb_subfr; s++ ) {
x_ptr = x + s * subfr_length; x_ptr = x + s * subfr_length;
for( n = 1; n < D + 1; n++ ) { for( n = 1; n < D + 1; n++ ) {
C_first_row[ n - 1 ] += silk_inner_product_FLP( x_ptr, x_ptr + n, subfr_length - n ); C_first_row[ n - 1 ] += silk_inner_product_FLP( x_ptr, x_ptr + n, subfr_length - n, arch );
} }
} }
silk_memcpy( C_last_row, C_first_row, SILK_MAX_ORDER_LPC * sizeof( double ) ); silk_memcpy( C_last_row, C_first_row, SILK_MAX_ORDER_LPC * sizeof( double ) );

10
silk/float/corrMatrix_FLP.c
View file

@ -41,7 +41,8 @@ void silk_corrVector_FLP(
const silk_float *t, /* I Target vector [L] */ const silk_float *t, /* I Target vector [L] */
const opus_int L, /* I Length of vecors */ const opus_int L, /* I Length of vecors */
const opus_int Order, /* I Max lag for correlation */ const opus_int Order, /* I Max lag for correlation */
silk_float *Xt /* O X'*t correlation vector [order] */ silk_float *Xt, /* O X'*t correlation vector [order] */
int arch
) )
{ {
opus_int lag; opus_int lag;
@ -50,7 +51,7 @@ void silk_corrVector_FLP(
ptr1 = &x[ Order - 1 ]; /* Points to first sample of column 0 of X: X[:,0] */ ptr1 = &x[ Order - 1 ]; /* Points to first sample of column 0 of X: X[:,0] */
for( lag = 0; lag < Order; lag++ ) { for( lag = 0; lag < Order; lag++ ) {
/* Calculate X[:,lag]'*t */ /* Calculate X[:,lag]'*t */
Xt[ lag ] = (silk_float)silk_inner_product_FLP( ptr1, t, L ); Xt[ lag ] = (silk_float)silk_inner_product_FLP( ptr1, t, L, arch );
ptr1--; /* Next column of X */ ptr1--; /* Next column of X */
} }
} }
@ -60,7 +61,8 @@ void silk_corrMatrix_FLP(
const silk_float *x, /* I x vector [ L+order-1 ] used to create X */ const silk_float *x, /* I x vector [ L+order-1 ] used to create X */
const opus_int L, /* I Length of vectors */ const opus_int L, /* I Length of vectors */
const opus_int Order, /* I Max lag for correlation */ const opus_int Order, /* I Max lag for correlation */
silk_float *XX /* O X'*X correlation matrix [order x order] */ silk_float *XX, /* O X'*X correlation matrix [order x order] */
int arch
) )
{ {
opus_int j, lag; opus_int j, lag;
@ -79,7 +81,7 @@ void silk_corrMatrix_FLP(
ptr2 = &x[ Order - 2 ]; /* First sample of column 1 of X */ ptr2 = &x[ Order - 2 ]; /* First sample of column 1 of X */
for( lag = 1; lag < Order; lag++ ) { for( lag = 1; lag < Order; lag++ ) {
/* Calculate X[:,0]'*X[:,lag] */ /* Calculate X[:,0]'*X[:,lag] */
energy = silk_inner_product_FLP( ptr1, ptr2, L ); energy = silk_inner_product_FLP( ptr1, ptr2, L, arch );
matrix_ptr( XX, lag, 0, Order ) = ( silk_float )energy; matrix_ptr( XX, lag, 0, Order ) = ( silk_float )energy;
matrix_ptr( XX, 0, lag, Order ) = ( silk_float )energy; matrix_ptr( XX, 0, lag, Order ) = ( silk_float )energy;
/* Calculate X[:,j]'*X[:,j + lag] */ /* Calculate X[:,j]'*X[:,j + lag] */

7
silk/float/find_LPC_FLP.c
View file

@ -38,7 +38,8 @@ void silk_find_LPC_FLP(
silk_encoder_state *psEncC, /* I/O Encoder state */ silk_encoder_state *psEncC, /* I/O Encoder state */
opus_int16 NLSF_Q15[], /* O NLSFs */ opus_int16 NLSF_Q15[], /* O NLSFs */
const silk_float x[], /* I Input signal */ const silk_float x[], /* I Input signal */
const silk_float minInvGain /* I Inverse of max prediction gain */ const silk_float minInvGain, /* I Inverse of max prediction gain */
int arch
) )
{ {
opus_int k, subfr_length; opus_int k, subfr_length;
@ -56,12 +57,12 @@ void silk_find_LPC_FLP(
psEncC->indices.NLSFInterpCoef_Q2 = 4; psEncC->indices.NLSFInterpCoef_Q2 = 4;
/* Burg AR analysis for the full frame */ /* Burg AR analysis for the full frame */
res_nrg = silk_burg_modified_FLP( a, x, minInvGain, subfr_length, psEncC->nb_subfr, psEncC->predictLPCOrder ); res_nrg = silk_burg_modified_FLP( a, x, minInvGain, subfr_length, psEncC->nb_subfr, psEncC->predictLPCOrder, arch );
if( psEncC->useInterpolatedNLSFs && !psEncC->first_frame_after_reset && psEncC->nb_subfr == MAX_NB_SUBFR ) { if( psEncC->useInterpolatedNLSFs && !psEncC->first_frame_after_reset && psEncC->nb_subfr == MAX_NB_SUBFR ) {
/* Optimal solution for last 10 ms; subtract residual energy here, as that's easier than */ /* Optimal solution for last 10 ms; subtract residual energy here, as that's easier than */
/* adding it to the residual energy of the first 10 ms in each iteration of the search below */ /* adding it to the residual energy of the first 10 ms in each iteration of the search below */
res_nrg -= silk_burg_modified_FLP( a_tmp, x + ( MAX_NB_SUBFR / 2 ) * subfr_length, minInvGain, subfr_length, MAX_NB_SUBFR / 2, psEncC->predictLPCOrder ); res_nrg -= silk_burg_modified_FLP( a_tmp, x + ( MAX_NB_SUBFR / 2 ) * subfr_length, minInvGain, subfr_length, MAX_NB_SUBFR / 2, psEncC->predictLPCOrder, arch );
/* Convert to NLSFs */ /* Convert to NLSFs */
silk_A2NLSF_FLP( NLSF_Q15, a_tmp, psEncC->predictLPCOrder ); silk_A2NLSF_FLP( NLSF_Q15, a_tmp, psEncC->predictLPCOrder );

7
silk/float/find_LTP_FLP.c
View file

@ -38,7 +38,8 @@ void silk_find_LTP_FLP(
const silk_float r_ptr[], /* I LPC residual */ const silk_float r_ptr[], /* I LPC residual */
const opus_int lag[ MAX_NB_SUBFR ], /* I LTP lags */ const opus_int lag[ MAX_NB_SUBFR ], /* I LTP lags */
const opus_int subfr_length, /* I Subframe length */ const opus_int subfr_length, /* I Subframe length */
const opus_int nb_subfr /* I number of subframes */ const opus_int nb_subfr, /* I number of subframes */
int arch
) )
{ {
opus_int k; opus_int k;
@ -50,8 +51,8 @@ void silk_find_LTP_FLP(
XX_ptr = XX; XX_ptr = XX;
for( k = 0; k < nb_subfr; k++ ) { for( k = 0; k < nb_subfr; k++ ) {
lag_ptr = r_ptr - ( lag[ k ] + LTP_ORDER / 2 ); lag_ptr = r_ptr - ( lag[ k ] + LTP_ORDER / 2 );
silk_corrMatrix_FLP( lag_ptr, subfr_length, LTP_ORDER, XX_ptr ); silk_corrMatrix_FLP( lag_ptr, subfr_length, LTP_ORDER, XX_ptr, arch );
silk_corrVector_FLP( lag_ptr, r_ptr, subfr_length, LTP_ORDER, xX_ptr ); silk_corrVector_FLP( lag_ptr, r_ptr, subfr_length, LTP_ORDER, xX_ptr, arch );
xx = ( silk_float )silk_energy_FLP( r_ptr, subfr_length + LTP_ORDER ); xx = ( silk_float )silk_energy_FLP( r_ptr, subfr_length + LTP_ORDER );
temp = 1.0f / silk_max( xx, LTP_CORR_INV_MAX * 0.5f * ( XX_ptr[ 0 ] + XX_ptr[ 24 ] ) + 1.0f ); temp = 1.0f / silk_max( xx, LTP_CORR_INV_MAX * 0.5f * ( XX_ptr[ 0 ] + XX_ptr[ 24 ] ) + 1.0f );
silk_scale_vector_FLP( XX_ptr, temp, LTP_ORDER * LTP_ORDER ); silk_scale_vector_FLP( XX_ptr, temp, LTP_ORDER * LTP_ORDER );

2
silk/float/find_pitch_lags_FLP.c
View file

@ -82,7 +82,7 @@ void silk_find_pitch_lags_FLP(
silk_apply_sine_window_FLP( Wsig_ptr, x_buf_ptr, 2, psEnc->sCmn.la_pitch ); silk_apply_sine_window_FLP( Wsig_ptr, x_buf_ptr, 2, psEnc->sCmn.la_pitch );
/* Calculate autocorrelation sequence */ /* Calculate autocorrelation sequence */
silk_autocorrelation_FLP( auto_corr, Wsig, psEnc->sCmn.pitch_LPC_win_length, psEnc->sCmn.pitchEstimationLPCOrder + 1 ); silk_autocorrelation_FLP( auto_corr, Wsig, psEnc->sCmn.pitch_LPC_win_length, psEnc->sCmn.pitchEstimationLPCOrder + 1, arch );
/* Add white noise, as a fraction of the energy */ /* Add white noise, as a fraction of the energy */
auto_corr[ 0 ] += auto_corr[ 0 ] * FIND_PITCH_WHITE_NOISE_FRACTION + 1; auto_corr[ 0 ] += auto_corr[ 0 ] * FIND_PITCH_WHITE_NOISE_FRACTION + 1;

4
silk/float/find_pred_coefs_FLP.c
View file

@ -63,7 +63,7 @@ void silk_find_pred_coefs_FLP(
celt_assert( psEnc->sCmn.ltp_mem_length - psEnc->sCmn.predictLPCOrder >= psEncCtrl->pitchL[ 0 ] + LTP_ORDER / 2 ); celt_assert( psEnc->sCmn.ltp_mem_length - psEnc->sCmn.predictLPCOrder >= psEncCtrl->pitchL[ 0 ] + LTP_ORDER / 2 );
/* LTP analysis */ /* LTP analysis */
silk_find_LTP_FLP( XXLTP, xXLTP, res_pitch, psEncCtrl->pitchL, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr ); silk_find_LTP_FLP( XXLTP, xXLTP, res_pitch, psEncCtrl->pitchL, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.arch );
/* Quantize LTP gain parameters */ /* Quantize LTP gain parameters */
silk_quant_LTP_gains_FLP( psEncCtrl->LTPCoef, psEnc->sCmn.indices.LTPIndex, &psEnc->sCmn.indices.PERIndex, silk_quant_LTP_gains_FLP( psEncCtrl->LTPCoef, psEnc->sCmn.indices.LTPIndex, &psEnc->sCmn.indices.PERIndex,
@ -102,7 +102,7 @@ void silk_find_pred_coefs_FLP(
} }
/* LPC_in_pre contains the LTP-filtered input for voiced, and the unfiltered input for unvoiced */ /* LPC_in_pre contains the LTP-filtered input for voiced, and the unfiltered input for unvoiced */
silk_find_LPC_FLP( &psEnc->sCmn, NLSF_Q15, LPC_in_pre, minInvGain ); silk_find_LPC_FLP( &psEnc->sCmn, NLSF_Q15, LPC_in_pre, minInvGain, psEnc->sCmn.arch );
/* Quantize LSFs */ /* Quantize LSFs */
silk_process_NLSFs_FLP( &psEnc->sCmn, psEncCtrl->PredCoef, NLSF_Q15, psEnc->sCmn.prev_NLSFq_Q15 ); silk_process_NLSFs_FLP( &psEnc->sCmn, psEncCtrl->PredCoef, NLSF_Q15, psEnc->sCmn.prev_NLSFq_Q15 );

12
silk/float/main_FLP.h
View file

@ -138,7 +138,8 @@ void silk_find_LPC_FLP(
silk_encoder_state *psEncC, /* I/O Encoder state */ silk_encoder_state *psEncC, /* I/O Encoder state */
opus_int16 NLSF_Q15[], /* O NLSFs */ opus_int16 NLSF_Q15[], /* O NLSFs */
const silk_float x[], /* I Input signal */ const silk_float x[], /* I Input signal */
const silk_float minInvGain /* I Prediction gain from LTP (dB) */ const silk_float minInvGain, /* I Prediction gain from LTP (dB) */
int arch
); );
/* LTP analysis */ /* LTP analysis */
@ -148,7 +149,8 @@ void silk_find_LTP_FLP(
const silk_float r_ptr[], /* I LPC residual */ const silk_float r_ptr[], /* I LPC residual */
const opus_int lag[ MAX_NB_SUBFR ], /* I LTP lags */ const opus_int lag[ MAX_NB_SUBFR ], /* I LTP lags */
const opus_int subfr_length, /* I Subframe length */ const opus_int subfr_length, /* I Subframe length */
const opus_int nb_subfr /* I number of subframes */ const opus_int nb_subfr, /* I number of subframes */
int arch
); );
void silk_LTP_analysis_filter_FLP( void silk_LTP_analysis_filter_FLP(
@ -221,7 +223,8 @@ void silk_corrMatrix_FLP(
const silk_float *x, /* I x vector [ L+order-1 ] used to create X */ const silk_float *x, /* I x vector [ L+order-1 ] used to create X */
const opus_int L, /* I Length of vectors */ const opus_int L, /* I Length of vectors */
const opus_int Order, /* I Max lag for correlation */ const opus_int Order, /* I Max lag for correlation */
silk_float *XX /* O X'*X correlation matrix [order x order] */ silk_float *XX, /* O X'*X correlation matrix [order x order] */
int arch
); );
/* Calculates correlation vector X'*t */ /* Calculates correlation vector X'*t */
@ -230,7 +233,8 @@ void silk_corrVector_FLP(
const silk_float *t, /* I Target vector [L] */ const silk_float *t, /* I Target vector [L] */
const opus_int L, /* I Length of vecors */ const opus_int L, /* I Length of vecors */
const opus_int Order, /* I Max lag for correlation */ const opus_int Order, /* I Max lag for correlation */
silk_float *Xt /* O X'*t correlation vector [order] */ silk_float *Xt, /* O X'*t correlation vector [order] */
int arch
); );
/* Apply sine window to signal vector. */ /* Apply sine window to signal vector. */

2
silk/float/noise_shape_analysis_FLP.c
View file

@ -255,7 +255,7 @@ void silk_noise_shape_analysis_FLP(
psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder ); psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder );
} else { } else {
/* Calculate regular auto correlation */ /* Calculate regular auto correlation */
silk_autocorrelation_FLP( auto_corr, x_windowed, psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder + 1 ); silk_autocorrelation_FLP( auto_corr, x_windowed, psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder + 1, psEnc->sCmn.arch );
} }
/* Add white noise, as a fraction of energy */ /* Add white noise, as a fraction of energy */

2
silk/float/pitch_analysis_core_FLP.c
View file

@ -291,7 +291,7 @@ opus_int silk_pitch_analysis_core_FLP( /* O Voicing estimate: 0 voiced,
for( j = 0; j < length_d_comp; j++ ) { for( j = 0; j < length_d_comp; j++ ) {
d = d_comp[ j ]; d = d_comp[ j ];
basis_ptr = target_ptr - d; basis_ptr = target_ptr - d;
cross_corr = silk_inner_product_FLP( basis_ptr, target_ptr, sf_length_8kHz ); cross_corr = silk_inner_product_FLP( basis_ptr, target_ptr, sf_length_8kHz, arch );
if( cross_corr > 0.0f ) { if( cross_corr > 0.0f ) {
energy = silk_energy_FLP( basis_ptr, sf_length_8kHz ); energy = silk_energy_FLP( basis_ptr, sf_length_8kHz );
C[ k ][ d ] = (silk_float)( 2 * cross_corr / ( energy + energy_tmp ) ); C[ k ][ d ] = (silk_float)( 2 * cross_corr / ( energy + energy_tmp ) );

11
silk/x86/main_sse.h
View file

@ -278,11 +278,18 @@ double silk_inner_product_FLP_avx2(
#if defined (OPUS_X86_PRESUME_AVX2) #if defined (OPUS_X86_PRESUME_AVX2)
#define OVERRIDE_inner_product_FLP #define OVERRIDE_inner_product_FLP
#define silk_inner_product_FLP(data1, data2, dataSize) silk_inner_product_FLP_avx2(data1, data2, dataSize) #define silk_inner_product_FLP(data1, data2, dataSize, arch) ((void)arch,silk_inner_product_FLP_avx2(data1, data2, dataSize))
#elif defined(OPUS_HAVE_RTCD) && defined(OPUS_X86_MAY_HAVE_AVX2) #elif defined(OPUS_HAVE_RTCD) && defined(OPUS_X86_MAY_HAVE_AVX2)
/*#define OVERRIDE_inner_product_FLP*/ #define OVERRIDE_inner_product_FLP
extern double (*const SILK_INNER_PRODUCT_FLP_IMPL[OPUS_ARCHMASK + 1])(
const silk_float *data1,
const silk_float *data2,
opus_int dataSize
);
#define silk_inner_product_FLP(data1, data2, dataSize, arch) ((void)arch,(*SILK_INNER_PRODUCT_FLP_IMPL[(arch) & OPUS_ARCHMASK])(data1, data2, dataSize))
#endif #endif

18
silk/x86/x86_silk_map.c
View file

@ -32,6 +32,7 @@
#include "celt/x86/x86cpu.h" #include "celt/x86/x86cpu.h"
#include "structs.h" #include "structs.h"
#include "SigProc_FIX.h" #include "SigProc_FIX.h"
#include "SigProc_FLP.h"
#include "pitch.h" #include "pitch.h"
#include "main.h" #include "main.h"
@ -156,4 +157,21 @@ void (*const SILK_BURG_MODIFIED_IMPL[ OPUS_ARCHMASK + 1 ] )(
}; };
#endif #endif
#ifndef FIXED_POINT
double (*const SILK_INNER_PRODUCT_FLP_IMPL[ OPUS_ARCHMASK + 1 ] )(
const silk_float *data1,
const silk_float *data2,
opus_int dataSize
) = {
silk_inner_product_FLP_c, /* non-sse */
silk_inner_product_FLP_c,
silk_inner_product_FLP_c,
silk_inner_product_FLP_c, /* sse4.1 */
MAY_HAVE_AVX2( silk_inner_product_FLP ) /* avx */
};
#endif
#endif #endif