ref: 239d223d84c20d5c146ce6d5200b1b148ab85af6
parent: b93e4a149c1bb77e65f4e2c66249553d2a9c4428
author: Jean-Marc Valin <jmvalin@amazon.com>
date: Mon Nov 20 21:56:04 EST 2023
Add rtcd for silk_inner_product_FLP()
--- a/silk/float/SigProc_FLP.h
+++ b/silk/float/SigProc_FLP.h
@@ -74,7 +74,8 @@
silk_float *results, /* O result (length correlationCount) */
const silk_float *inputData, /* I input data to correlate */
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 */
@@ -106,7 +107,8 @@
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 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 */
@@ -132,7 +134,7 @@
);
#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
--- a/silk/float/autocorrelation_FLP.c
+++ b/silk/float/autocorrelation_FLP.c
@@ -37,7 +37,8 @@
silk_float *results, /* O result (length correlationCount) */
const silk_float *inputData, /* I input data to correlate */
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;
@@ -47,6 +48,6 @@
}
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 );
}
}
--- a/silk/float/burg_modified_FLP.c
+++ b/silk/float/burg_modified_FLP.c
@@ -42,7 +42,8 @@
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 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;
@@ -60,7 +61,7 @@
for( s = 0; s < nb_subfr; s++ ) {x_ptr = x + s * subfr_length;
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 ) );
--- a/silk/float/corrMatrix_FLP.c
+++ b/silk/float/corrMatrix_FLP.c
@@ -41,7 +41,8 @@
const silk_float *t, /* I Target vector [L] */
const opus_int L, /* I Length of vecors */
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;
@@ -50,7 +51,7 @@
ptr1 = &x[ Order - 1 ]; /* Points to first sample of column 0 of X: X[:,0] */
for( lag = 0; lag < Order; lag++ ) {/* 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 */
}
}
@@ -60,7 +61,8 @@
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 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;
@@ -79,7 +81,7 @@
ptr2 = &x[ Order - 2 ]; /* First sample of column 1 of X */
for( lag = 1; lag < Order; 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, 0, lag, Order ) = ( silk_float )energy;
/* Calculate X[:,j]'*X[:,j + lag] */
--- a/silk/float/find_LPC_FLP.c
+++ b/silk/float/find_LPC_FLP.c
@@ -38,7 +38,8 @@
silk_encoder_state *psEncC, /* I/O Encoder state */
opus_int16 NLSF_Q15[], /* O NLSFs */
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;
@@ -56,12 +57,12 @@
psEncC->indices.NLSFInterpCoef_Q2 = 4;
/* 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 ) {/* 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 */
- 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 */
silk_A2NLSF_FLP( NLSF_Q15, a_tmp, psEncC->predictLPCOrder );
--- a/silk/float/find_LTP_FLP.c
+++ b/silk/float/find_LTP_FLP.c
@@ -38,7 +38,8 @@
const silk_float r_ptr[], /* I LPC residual */
const opus_int lag[ MAX_NB_SUBFR ], /* I LTP lags */
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;
@@ -50,8 +51,8 @@
XX_ptr = XX;
for( k = 0; k < nb_subfr; k++ ) {lag_ptr = r_ptr - ( lag[ k ] + LTP_ORDER / 2 );
- silk_corrMatrix_FLP( lag_ptr, subfr_length, LTP_ORDER, XX_ptr );
- silk_corrVector_FLP( lag_ptr, r_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, arch );
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 );
silk_scale_vector_FLP( XX_ptr, temp, LTP_ORDER * LTP_ORDER );
--- a/silk/float/find_pitch_lags_FLP.c
+++ b/silk/float/find_pitch_lags_FLP.c
@@ -82,7 +82,7 @@
silk_apply_sine_window_FLP( Wsig_ptr, x_buf_ptr, 2, psEnc->sCmn.la_pitch );
/* 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 */
auto_corr[ 0 ] += auto_corr[ 0 ] * FIND_PITCH_WHITE_NOISE_FRACTION + 1;
--- a/silk/float/find_pred_coefs_FLP.c
+++ b/silk/float/find_pred_coefs_FLP.c
@@ -63,7 +63,7 @@
celt_assert( psEnc->sCmn.ltp_mem_length - psEnc->sCmn.predictLPCOrder >= psEncCtrl->pitchL[ 0 ] + LTP_ORDER / 2 );
/* 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 */
silk_quant_LTP_gains_FLP( psEncCtrl->LTPCoef, psEnc->sCmn.indices.LTPIndex, &psEnc->sCmn.indices.PERIndex,
@@ -102,7 +102,7 @@
}
/* 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 */
silk_process_NLSFs_FLP( &psEnc->sCmn, psEncCtrl->PredCoef, NLSF_Q15, psEnc->sCmn.prev_NLSFq_Q15 );
--- a/silk/float/main_FLP.h
+++ b/silk/float/main_FLP.h
@@ -138,7 +138,8 @@
silk_encoder_state *psEncC, /* I/O Encoder state */
opus_int16 NLSF_Q15[], /* O NLSFs */
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 */
@@ -148,7 +149,8 @@
const silk_float r_ptr[], /* I LPC residual */
const opus_int lag[ MAX_NB_SUBFR ], /* I LTP lags */
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(
@@ -221,7 +223,8 @@
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 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 */
@@ -230,7 +233,8 @@
const silk_float *t, /* I Target vector [L] */
const opus_int L, /* I Length of vecors */
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. */
--- a/silk/float/noise_shape_analysis_FLP.c
+++ b/silk/float/noise_shape_analysis_FLP.c
@@ -255,7 +255,7 @@
psEnc->sCmn.shapeWinLength, psEnc->sCmn.shapingLPCOrder );
} else {/* 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 */
--- a/silk/float/pitch_analysis_core_FLP.c
+++ b/silk/float/pitch_analysis_core_FLP.c
@@ -291,7 +291,7 @@
for( j = 0; j < length_d_comp; j++ ) {d = d_comp[ j ];
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 ) {energy = silk_energy_FLP( basis_ptr, sf_length_8kHz );
C[ k ][ d ] = (silk_float)( 2 * cross_corr / ( energy + energy_tmp ) );
--- a/silk/x86/main_sse.h
+++ b/silk/x86/main_sse.h
@@ -278,11 +278,18 @@
#if defined (OPUS_X86_PRESUME_AVX2)
#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)
-/*#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
--- a/silk/x86/x86_silk_map.c
+++ b/silk/x86/x86_silk_map.c
@@ -32,6 +32,7 @@
#include "celt/x86/x86cpu.h"
#include "structs.h"
#include "SigProc_FIX.h"
+#include "SigProc_FLP.h"
#include "pitch.h"
#include "main.h"
@@ -156,4 +157,21 @@
};
#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
--
⑨