ref: 0a698d65226470b37a457a6b3758a486cefe53db
parent: f78e5080ab79567305bfc8755e4dad5678473f7b
author: Koen Vos <[email protected]>
date: Fri Jun 3 10:18:03 EDT 2016
merge back in the LTP gain safety limiter
--- a/silk/VQ_WMat_EC.c
+++ b/silk/VQ_WMat_EC.c
@@ -36,15 +36,18 @@
opus_int8 *ind, /* O index of best codebook vector */
opus_int32 *res_nrg_Q15, /* O best residual energy */
opus_int32 *rate_dist_Q8, /* O best total bitrate */
+ opus_int *gain_Q7, /* O sum of absolute LTP coefficients */
const opus_int32 *XX_Q17, /* I correlation matrix */
const opus_int32 *xX_Q17, /* I correlation vector */
const opus_int8 *cb_Q7, /* I codebook */
+ const opus_uint8 *cb_gain_Q7, /* I codebook effective gain */
const opus_uint8 *cl_Q5, /* I code length for each codebook vector */
const opus_int subfr_len, /* I number of samples per subframe */
+ const opus_int32 max_gain_Q7, /* I maximum sum of absolute LTP coefficients */
const opus_int L /* I number of vectors in codebook */
)
{- opus_int k;
+ opus_int k, gain_tmp_Q7;
const opus_int8 *cb_row_Q7;
opus_int32 neg_xX_Q24[ 5 ];
opus_int32 sum1_Q15, sum2_Q24;
@@ -64,10 +67,15 @@
/* In things go really bad, at least *ind is set to something safe. */
*ind = 0;
for( k = 0; k < L; k++ ) {+ opus_int32 penalty;
+ gain_tmp_Q7 = cb_gain_Q7[k];
/* Weighted rate */
/* Quantization error: 1 - 2 * xX * cb + cb' * XX * cb */
sum1_Q15 = SILK_FIX_CONST( 1.001, 15 );
+ /* Penalty for too large gain */
+ penalty = silk_LSHIFT32( silk_max( silk_SUB32( gain_tmp_Q7, max_gain_Q7 ), 0 ), 11 );
+
/* first row of XX_Q17 */
sum2_Q24 = silk_MLA( neg_xX_Q24[ 0 ], XX_Q17[ 1 ], cb_row_Q7[ 1 ] );
sum2_Q24 = silk_MLA( sum2_Q24, XX_Q17[ 2 ], cb_row_Q7[ 2 ] );
@@ -106,13 +114,14 @@
/* find best */
if( sum1_Q15 >= 0 ) {/* Translate residual energy to bits using high-rate assumption (6 dB ==> 1 bit/sample) */
- bits_res_Q8 = silk_SMULBB( subfr_len, silk_lin2log( sum1_Q15 ) - (15 << 7) );
+ bits_res_Q8 = silk_SMULBB( subfr_len, silk_lin2log( sum1_Q15 + penalty) - (15 << 7) );
/* In the following line we reduce the codelength component by half ("-1"); seems to slghtly improve quality */bits_tot_Q8 = silk_ADD_LSHIFT32( bits_res_Q8, cl_Q5[ k ], 3-1 );
if( bits_tot_Q8 <= *rate_dist_Q8 ) {*rate_dist_Q8 = bits_tot_Q8;
- *res_nrg_Q15 = sum1_Q15;
+ *res_nrg_Q15 = sum1_Q15 + penalty;
*ind = (opus_int8)k;
+ *gain_Q7 = gain_tmp_Q7;
}
}
--- a/silk/fixed/find_pred_coefs_FIX.c
+++ b/silk/fixed/find_pred_coefs_FIX.c
@@ -91,7 +91,7 @@
/* Quantize LTP gain parameters */
silk_quant_LTP_gains( psEncCtrl->LTPCoef_Q14, psEnc->sCmn.indices.LTPIndex, &psEnc->sCmn.indices.PERIndex,
- &psEncCtrl->LTPredCodGain_Q7, XXLTP_Q17, xXLTP_Q17, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.arch );
+ &psEnc->sCmn.sum_log_gain_Q7, &psEncCtrl->LTPredCodGain_Q7, XXLTP_Q17, xXLTP_Q17, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.arch );
/* Control LTP scaling */
silk_LTP_scale_ctrl_FIX( psEnc, psEncCtrl, condCoding );
@@ -116,6 +116,7 @@
silk_memset( psEncCtrl->LTPCoef_Q14, 0, psEnc->sCmn.nb_subfr * LTP_ORDER * sizeof( opus_int16 ) );
psEncCtrl->LTPredCodGain_Q7 = 0;
+ psEnc->sCmn.sum_log_gain_Q7 = 0;
}
/* Limit on total predictive coding gain */
--- a/silk/float/find_pred_coefs_FLP.c
+++ b/silk/float/find_pred_coefs_FLP.c
@@ -66,7 +66,7 @@
/* Quantize LTP gain parameters */
silk_quant_LTP_gains_FLP( psEncCtrl->LTPCoef, psEnc->sCmn.indices.LTPIndex, &psEnc->sCmn.indices.PERIndex,
- &psEncCtrl->LTPredCodGain, XXLTP, xXLTP, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.arch );
+ &psEnc->sCmn.sum_log_gain_Q7, &psEncCtrl->LTPredCodGain, XXLTP, xXLTP, psEnc->sCmn.subfr_length, psEnc->sCmn.nb_subfr, psEnc->sCmn.arch );
/* Control LTP scaling */
silk_LTP_scale_ctrl_FLP( psEnc, psEncCtrl, condCoding );
@@ -89,6 +89,7 @@
}
silk_memset( psEncCtrl->LTPCoef, 0, psEnc->sCmn.nb_subfr * LTP_ORDER * sizeof( silk_float ) );
psEncCtrl->LTPredCodGain = 0.0f;
+ psEnc->sCmn.sum_log_gain_Q7 = 0;
}
/* Limit on total predictive coding gain */
--- a/silk/float/main_FLP.h
+++ b/silk/float/main_FLP.h
@@ -188,6 +188,7 @@
silk_float B[ MAX_NB_SUBFR * LTP_ORDER ], /* O Quantized LTP gains */
opus_int8 cbk_index[ MAX_NB_SUBFR ], /* O Codebook index */
opus_int8 *periodicity_index, /* O Periodicity index */
+ opus_int32 *sum_log_gain_Q7, /* I/O Cumulative max prediction gain */
silk_float *pred_gain_dB, /* O LTP prediction gain */
const silk_float XX[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ], /* I Correlation matrix */
const silk_float xX[ MAX_NB_SUBFR * LTP_ORDER ], /* I Correlation vector */
--- a/silk/float/wrappers_FLP.c
+++ b/silk/float/wrappers_FLP.c
@@ -175,6 +175,7 @@
silk_float B[ MAX_NB_SUBFR * LTP_ORDER ], /* O Quantized LTP gains */
opus_int8 cbk_index[ MAX_NB_SUBFR ], /* O Codebook index */
opus_int8 *periodicity_index, /* O Periodicity index */
+ opus_int32 *sum_log_gain_Q7, /* I/O Cumulative max prediction gain */
silk_float *pred_gain_dB, /* O LTP prediction gain */
const silk_float XX[ MAX_NB_SUBFR * LTP_ORDER * LTP_ORDER ], /* I Correlation matrix */
const silk_float xX[ MAX_NB_SUBFR * LTP_ORDER ], /* I Correlation vector */
@@ -195,7 +196,7 @@
xX_Q17[ i ] = (opus_int32)silk_float2int( xX[ i ] * 131072.0f );
}
- silk_quant_LTP_gains( B_Q14, cbk_index, periodicity_index, &pred_gain_dB_Q7, XX_Q17, xX_Q17, subfr_len, nb_subfr, arch );
+ silk_quant_LTP_gains( B_Q14, cbk_index, periodicity_index, sum_log_gain_Q7, &pred_gain_dB_Q7, XX_Q17, xX_Q17, subfr_len, nb_subfr, arch );
for( i = 0; i < nb_subfr * LTP_ORDER; i++ ) {B[ i ] = (silk_float)B_Q14[ i ] * ( 1.0f / 16384.0f );
--- a/silk/main.h
+++ b/silk/main.h
@@ -208,6 +208,7 @@
opus_int16 B_Q14[ MAX_NB_SUBFR * LTP_ORDER ], /* O Quantized LTP gains */
opus_int8 cbk_index[ MAX_NB_SUBFR ], /* O Codebook Index */
opus_int8 *periodicity_index, /* O Periodicity Index */
+ opus_int32 *sum_gain_dB_Q7, /* I/O Cumulative max prediction gain */
opus_int *pred_gain_dB_Q7, /* O LTP prediction gain */
const opus_int32 XX_Q17[ MAX_NB_SUBFR*LTP_ORDER*LTP_ORDER ], /* I Correlation matrix in Q18 */
const opus_int32 xX_Q17[ MAX_NB_SUBFR*LTP_ORDER ], /* I Correlation vector in Q18 */
@@ -221,17 +222,20 @@
opus_int8 *ind, /* O index of best codebook vector */
opus_int32 *res_nrg_Q15, /* O best residual energy */
opus_int32 *rate_dist_Q8, /* O best total bitrate */
+ opus_int *gain_Q7, /* O sum of absolute LTP coefficients */
const opus_int32 *XX_Q17, /* I correlation matrix */
const opus_int32 *xX_Q17, /* I correlation vector */
const opus_int8 *cb_Q7, /* I codebook */
+ const opus_uint8 *cb_gain_Q7, /* I codebook effective gain */
const opus_uint8 *cl_Q5, /* I code length for each codebook vector */
const opus_int subfr_len, /* I number of samples per subframe */
+ const opus_int32 max_gain_Q7, /* I maximum sum of absolute LTP coefficients */
const opus_int L /* I number of vectors in codebook */
);
#if !defined(OVERRIDE_silk_VQ_WMat_EC)
-#define silk_VQ_WMat_EC(ind, rate_dist_Q15, gain_Q8, XX_Q17, xX_Q17, cb_Q7, cl_Q5, subfr_len, L, arch) \
- ((void)(arch),silk_VQ_WMat_EC_c(ind, rate_dist_Q15, gain_Q8, XX_Q17, xX_Q17, cb_Q7, cl_Q5, subfr_len, L))
+#define silk_VQ_WMat_EC(ind, res_nrg_Q15, rate_dist_Q8, gain_Q7, XX_Q17, xX_Q17, cb_Q7, cb_gain_Q7, cl_Q5, subfr_len, max_gain_Q7, L, arch) \
+ ((void)(arch),silk_VQ_WMat_EC_c(ind, res_nrg_Q15, rate_dist_Q8, gain_Q7, XX_Q17, xX_Q17, cb_Q7, cb_gain_Q7, cl_Q5, subfr_len, max_gain_Q7, L))
#endif
/************************************/
--- a/silk/quant_LTP_gains.c
+++ b/silk/quant_LTP_gains.c
@@ -30,11 +30,13 @@
#endif
#include "main.h"
+#include "tuning_parameters.h"
void silk_quant_LTP_gains(
opus_int16 B_Q14[ MAX_NB_SUBFR * LTP_ORDER ], /* O Quantized LTP gains */
opus_int8 cbk_index[ MAX_NB_SUBFR ], /* O Codebook Index */
opus_int8 *periodicity_index, /* O Periodicity Index */
+ opus_int32 *sum_log_gain_Q7, /* I/O Cumulative max prediction gain */
opus_int *pred_gain_dB_Q7, /* O LTP prediction gain */
const opus_int32 XX_Q17[ MAX_NB_SUBFR*LTP_ORDER*LTP_ORDER ], /* I Correlation matrix in Q18 */
const opus_int32 xX_Q17[ MAX_NB_SUBFR*LTP_ORDER ], /* I Correlation vector in Q18 */
@@ -47,8 +49,10 @@
opus_int8 temp_idx[ MAX_NB_SUBFR ];
const opus_uint8 *cl_ptr_Q5;
const opus_int8 *cbk_ptr_Q7;
+ const opus_uint8 *cbk_gain_ptr_Q7;
const opus_int32 *XX_Q17_ptr, *xX_Q17_ptr;
opus_int32 res_nrg_Q15_subfr, res_nrg_Q15, rate_dist_Q7_subfr, rate_dist_Q7, min_rate_dist_Q7;
+ opus_int32 sum_log_gain_tmp_Q7, best_sum_log_gain_Q7, max_gain_Q7, gain_Q7;
/***************************************************/
/* iterate over different codebooks with different */
@@ -55,9 +59,15 @@
/* rates/distortions, and choose best */
/***************************************************/
min_rate_dist_Q7 = silk_int32_MAX;
+ best_sum_log_gain_Q7 = 0;
for( k = 0; k < 3; k++ ) {+ /* Safety margin for pitch gain control, to take into account factors
+ such as state rescaling/rewhitening. */
+ opus_int32 gain_safety = SILK_FIX_CONST( 0.4, 7 );
+
cl_ptr_Q5 = silk_LTP_gain_BITS_Q5_ptrs[ k ];
cbk_ptr_Q7 = silk_LTP_vq_ptrs_Q7[ k ];
+ cbk_gain_ptr_Q7 = silk_LTP_vq_gain_ptrs_Q7[ k ];
cbk_size = silk_LTP_vq_sizes[ k ];
/* Set up pointers to first subframe */
@@ -66,16 +76,22 @@
res_nrg_Q15 = 0;
rate_dist_Q7 = 0;
+ sum_log_gain_tmp_Q7 = *sum_log_gain_Q7;
for( j = 0; j < nb_subfr; j++ ) {+ max_gain_Q7 = silk_log2lin( ( SILK_FIX_CONST( MAX_SUM_LOG_GAIN_DB / 6.0, 7 ) - sum_log_gain_tmp_Q7 )
+ + SILK_FIX_CONST( 7, 7 ) ) - gain_safety;
silk_VQ_WMat_EC(
&temp_idx[ j ], /* O index of best codebook vector */
&res_nrg_Q15_subfr, /* O residual energy */
&rate_dist_Q7_subfr, /* O best weighted quantization error + mu * rate */
+ &gain_Q7, /* O sum of absolute LTP coefficients */
XX_Q17_ptr, /* I correlation matrix */
xX_Q17_ptr, /* I correlation vector */
cbk_ptr_Q7, /* I codebook */
+ cbk_gain_ptr_Q7, /* I codebook effective gains */
cl_ptr_Q5, /* I code length for each codebook vector */
subfr_len, /* I number of samples per subframe */
+ max_gain_Q7, /* I maximum sum of absolute LTP coefficients */
cbk_size, /* I number of vectors in codebook */
arch /* I Run-time architecture */
);
@@ -82,6 +98,8 @@
res_nrg_Q15 = silk_ADD_POS_SAT32( res_nrg_Q15, res_nrg_Q15_subfr );
rate_dist_Q7 = silk_ADD_POS_SAT32( rate_dist_Q7, rate_dist_Q7_subfr );
+ sum_log_gain_tmp_Q7 = silk_max(0, sum_log_gain_tmp_Q7
+ + silk_lin2log( gain_safety + gain_Q7 ) - SILK_FIX_CONST( 7, 7 ));
XX_Q17_ptr += LTP_ORDER * LTP_ORDER;
xX_Q17_ptr += LTP_ORDER;
@@ -91,6 +109,7 @@
min_rate_dist_Q7 = rate_dist_Q7;
*periodicity_index = (opus_int8)k;
silk_memcpy( cbk_index, temp_idx, nb_subfr * sizeof( opus_int8 ) );
+ best_sum_log_gain_Q7 = sum_log_gain_tmp_Q7;
}
}
@@ -107,5 +126,6 @@
res_nrg_Q15 = silk_RSHIFT32( res_nrg_Q15, 2 );
}
+ *sum_log_gain_Q7 = best_sum_log_gain_Q7;
*pred_gain_dB_Q7 = (opus_int)silk_SMULBB( -3, silk_lin2log( res_nrg_Q15 ) - ( 15 << 7 ) );
}
--- a/silk/tables.h
+++ b/silk/tables.h
@@ -77,6 +77,7 @@
extern const opus_uint8 * const silk_LTP_gain_iCDF_ptrs[ NB_LTP_CBKS ]; /* 3 */
extern const opus_uint8 * const silk_LTP_gain_BITS_Q5_ptrs[ NB_LTP_CBKS ]; /* 3 */
extern const opus_int8 * const silk_LTP_vq_ptrs_Q7[ NB_LTP_CBKS ]; /* 168 */
+extern const opus_uint8 * const silk_LTP_vq_gain_ptrs_Q7[NB_LTP_CBKS];
extern const opus_int8 silk_LTP_vq_sizes[ NB_LTP_CBKS ]; /* 3 */
extern const opus_uint8 silk_LTPscale_iCDF[ 3 ]; /* 4 */
--- a/silk/tables_LTP.c
+++ b/silk/tables_LTP.c
@@ -265,6 +265,30 @@
(opus_int8 *)&silk_LTP_gain_vq_2[0][0]
};
+/* Maximum frequency-dependent response of the pitch taps above,
+ computed as max(abs(freqz(taps))) */
+static const opus_uint8 silk_LTP_gain_vq_0_gain[8] = {+ 46, 2, 90, 87, 93, 91, 82, 98
+};
+
+static const opus_uint8 silk_LTP_gain_vq_1_gain[16] = {+ 109, 120, 118, 12, 113, 115, 117, 119,
+ 99, 59, 87, 111, 63, 111, 112, 80
+};
+
+static const opus_uint8 silk_LTP_gain_vq_2_gain[32] = {+ 126, 124, 125, 124, 129, 121, 126, 23,
+ 132, 127, 127, 127, 126, 127, 122, 133,
+ 130, 134, 101, 118, 119, 145, 126, 86,
+ 124, 120, 123, 119, 170, 173, 107, 109
+};
+
+const opus_uint8 * const silk_LTP_vq_gain_ptrs_Q7[NB_LTP_CBKS] = {+ &silk_LTP_gain_vq_0_gain[0],
+ &silk_LTP_gain_vq_1_gain[0],
+ &silk_LTP_gain_vq_2_gain[0]
+};
+
const opus_int8 silk_LTP_vq_sizes[NB_LTP_CBKS] = {8, 16, 32
};
--- a/silk/tuning_parameters.h
+++ b/silk/tuning_parameters.h
@@ -50,8 +50,11 @@
/* Linear prediction */
/*********************/
-/* LPC analysis defines: regularization and bandwidth expansion */
+/* LPC analysis regularization */
#define FIND_LPC_COND_FAC 1e-5f
+
+/* Max cumulative LTP gain */
+#define MAX_SUM_LOG_GAIN_DB 250.0f
/* LTP analysis defines */
#define LTP_CORR_INV_MAX 0.015f