1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libopus/silk/NSQ_del_dec.c Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,719 @@
1.4 +/***********************************************************************
1.5 +Copyright (c) 2006-2011, Skype Limited. All rights reserved.
1.6 +Redistribution and use in source and binary forms, with or without
1.7 +modification, are permitted provided that the following conditions
1.8 +are met:
1.9 +- Redistributions of source code must retain the above copyright notice,
1.10 +this list of conditions and the following disclaimer.
1.11 +- Redistributions in binary form must reproduce the above copyright
1.12 +notice, this list of conditions and the following disclaimer in the
1.13 +documentation and/or other materials provided with the distribution.
1.14 +- Neither the name of Internet Society, IETF or IETF Trust, nor the
1.15 +names of specific contributors, may be used to endorse or promote
1.16 +products derived from this software without specific prior written
1.17 +permission.
1.18 +THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
1.19 +AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
1.20 +IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
1.21 +ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
1.22 +LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
1.23 +CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
1.24 +SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
1.25 +INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
1.26 +CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
1.27 +ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
1.28 +POSSIBILITY OF SUCH DAMAGE.
1.29 +***********************************************************************/
1.30 +
1.31 +#ifdef HAVE_CONFIG_H
1.32 +#include "config.h"
1.33 +#endif
1.34 +
1.35 +#include "main.h"
1.36 +#include "stack_alloc.h"
1.37 +
1.38 +typedef struct {
1.39 + opus_int32 sLPC_Q14[ MAX_SUB_FRAME_LENGTH + NSQ_LPC_BUF_LENGTH ];
1.40 + opus_int32 RandState[ DECISION_DELAY ];
1.41 + opus_int32 Q_Q10[ DECISION_DELAY ];
1.42 + opus_int32 Xq_Q14[ DECISION_DELAY ];
1.43 + opus_int32 Pred_Q15[ DECISION_DELAY ];
1.44 + opus_int32 Shape_Q14[ DECISION_DELAY ];
1.45 + opus_int32 sAR2_Q14[ MAX_SHAPE_LPC_ORDER ];
1.46 + opus_int32 LF_AR_Q14;
1.47 + opus_int32 Seed;
1.48 + opus_int32 SeedInit;
1.49 + opus_int32 RD_Q10;
1.50 +} NSQ_del_dec_struct;
1.51 +
1.52 +typedef struct {
1.53 + opus_int32 Q_Q10;
1.54 + opus_int32 RD_Q10;
1.55 + opus_int32 xq_Q14;
1.56 + opus_int32 LF_AR_Q14;
1.57 + opus_int32 sLTP_shp_Q14;
1.58 + opus_int32 LPC_exc_Q14;
1.59 +} NSQ_sample_struct;
1.60 +
1.61 +typedef NSQ_sample_struct NSQ_sample_pair[ 2 ];
1.62 +
1.63 +static OPUS_INLINE void silk_nsq_del_dec_scale_states(
1.64 + const silk_encoder_state *psEncC, /* I Encoder State */
1.65 + silk_nsq_state *NSQ, /* I/O NSQ state */
1.66 + NSQ_del_dec_struct psDelDec[], /* I/O Delayed decision states */
1.67 + const opus_int32 x_Q3[], /* I Input in Q3 */
1.68 + opus_int32 x_sc_Q10[], /* O Input scaled with 1/Gain in Q10 */
1.69 + const opus_int16 sLTP[], /* I Re-whitened LTP state in Q0 */
1.70 + opus_int32 sLTP_Q15[], /* O LTP state matching scaled input */
1.71 + opus_int subfr, /* I Subframe number */
1.72 + opus_int nStatesDelayedDecision, /* I Number of del dec states */
1.73 + const opus_int LTP_scale_Q14, /* I LTP state scaling */
1.74 + const opus_int32 Gains_Q16[ MAX_NB_SUBFR ], /* I */
1.75 + const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lag */
1.76 + const opus_int signal_type, /* I Signal type */
1.77 + const opus_int decisionDelay /* I Decision delay */
1.78 +);
1.79 +
1.80 +/******************************************/
1.81 +/* Noise shape quantizer for one subframe */
1.82 +/******************************************/
1.83 +static OPUS_INLINE void silk_noise_shape_quantizer_del_dec(
1.84 + silk_nsq_state *NSQ, /* I/O NSQ state */
1.85 + NSQ_del_dec_struct psDelDec[], /* I/O Delayed decision states */
1.86 + opus_int signalType, /* I Signal type */
1.87 + const opus_int32 x_Q10[], /* I */
1.88 + opus_int8 pulses[], /* O */
1.89 + opus_int16 xq[], /* O */
1.90 + opus_int32 sLTP_Q15[], /* I/O LTP filter state */
1.91 + opus_int32 delayedGain_Q10[], /* I/O Gain delay buffer */
1.92 + const opus_int16 a_Q12[], /* I Short term prediction coefs */
1.93 + const opus_int16 b_Q14[], /* I Long term prediction coefs */
1.94 + const opus_int16 AR_shp_Q13[], /* I Noise shaping coefs */
1.95 + opus_int lag, /* I Pitch lag */
1.96 + opus_int32 HarmShapeFIRPacked_Q14, /* I */
1.97 + opus_int Tilt_Q14, /* I Spectral tilt */
1.98 + opus_int32 LF_shp_Q14, /* I */
1.99 + opus_int32 Gain_Q16, /* I */
1.100 + opus_int Lambda_Q10, /* I */
1.101 + opus_int offset_Q10, /* I */
1.102 + opus_int length, /* I Input length */
1.103 + opus_int subfr, /* I Subframe number */
1.104 + opus_int shapingLPCOrder, /* I Shaping LPC filter order */
1.105 + opus_int predictLPCOrder, /* I Prediction filter order */
1.106 + opus_int warping_Q16, /* I */
1.107 + opus_int nStatesDelayedDecision, /* I Number of states in decision tree */
1.108 + opus_int *smpl_buf_idx, /* I Index to newest samples in buffers */
1.109 + opus_int decisionDelay /* I */
1.110 +);
1.111 +
1.112 +void silk_NSQ_del_dec(
1.113 + const silk_encoder_state *psEncC, /* I/O Encoder State */
1.114 + silk_nsq_state *NSQ, /* I/O NSQ state */
1.115 + SideInfoIndices *psIndices, /* I/O Quantization Indices */
1.116 + const opus_int32 x_Q3[], /* I Prefiltered input signal */
1.117 + opus_int8 pulses[], /* O Quantized pulse signal */
1.118 + const opus_int16 PredCoef_Q12[ 2 * MAX_LPC_ORDER ], /* I Short term prediction coefs */
1.119 + const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ], /* I Long term prediction coefs */
1.120 + const opus_int16 AR2_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */
1.121 + const opus_int HarmShapeGain_Q14[ MAX_NB_SUBFR ], /* I Long term shaping coefs */
1.122 + const opus_int Tilt_Q14[ MAX_NB_SUBFR ], /* I Spectral tilt */
1.123 + const opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ], /* I Low frequency shaping coefs */
1.124 + const opus_int32 Gains_Q16[ MAX_NB_SUBFR ], /* I Quantization step sizes */
1.125 + const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lags */
1.126 + const opus_int Lambda_Q10, /* I Rate/distortion tradeoff */
1.127 + const opus_int LTP_scale_Q14 /* I LTP state scaling */
1.128 +)
1.129 +{
1.130 + opus_int i, k, lag, start_idx, LSF_interpolation_flag, Winner_ind, subfr;
1.131 + opus_int last_smple_idx, smpl_buf_idx, decisionDelay;
1.132 + const opus_int16 *A_Q12, *B_Q14, *AR_shp_Q13;
1.133 + opus_int16 *pxq;
1.134 + VARDECL( opus_int32, sLTP_Q15 );
1.135 + VARDECL( opus_int16, sLTP );
1.136 + opus_int32 HarmShapeFIRPacked_Q14;
1.137 + opus_int offset_Q10;
1.138 + opus_int32 RDmin_Q10, Gain_Q10;
1.139 + VARDECL( opus_int32, x_sc_Q10 );
1.140 + VARDECL( opus_int32, delayedGain_Q10 );
1.141 + VARDECL( NSQ_del_dec_struct, psDelDec );
1.142 + NSQ_del_dec_struct *psDD;
1.143 + SAVE_STACK;
1.144 +
1.145 + /* Set unvoiced lag to the previous one, overwrite later for voiced */
1.146 + lag = NSQ->lagPrev;
1.147 +
1.148 + silk_assert( NSQ->prev_gain_Q16 != 0 );
1.149 +
1.150 + /* Initialize delayed decision states */
1.151 + ALLOC( psDelDec, psEncC->nStatesDelayedDecision, NSQ_del_dec_struct );
1.152 + silk_memset( psDelDec, 0, psEncC->nStatesDelayedDecision * sizeof( NSQ_del_dec_struct ) );
1.153 + for( k = 0; k < psEncC->nStatesDelayedDecision; k++ ) {
1.154 + psDD = &psDelDec[ k ];
1.155 + psDD->Seed = ( k + psIndices->Seed ) & 3;
1.156 + psDD->SeedInit = psDD->Seed;
1.157 + psDD->RD_Q10 = 0;
1.158 + psDD->LF_AR_Q14 = NSQ->sLF_AR_shp_Q14;
1.159 + psDD->Shape_Q14[ 0 ] = NSQ->sLTP_shp_Q14[ psEncC->ltp_mem_length - 1 ];
1.160 + silk_memcpy( psDD->sLPC_Q14, NSQ->sLPC_Q14, NSQ_LPC_BUF_LENGTH * sizeof( opus_int32 ) );
1.161 + silk_memcpy( psDD->sAR2_Q14, NSQ->sAR2_Q14, sizeof( NSQ->sAR2_Q14 ) );
1.162 + }
1.163 +
1.164 + offset_Q10 = silk_Quantization_Offsets_Q10[ psIndices->signalType >> 1 ][ psIndices->quantOffsetType ];
1.165 + smpl_buf_idx = 0; /* index of oldest samples */
1.166 +
1.167 + decisionDelay = silk_min_int( DECISION_DELAY, psEncC->subfr_length );
1.168 +
1.169 + /* For voiced frames limit the decision delay to lower than the pitch lag */
1.170 + if( psIndices->signalType == TYPE_VOICED ) {
1.171 + for( k = 0; k < psEncC->nb_subfr; k++ ) {
1.172 + decisionDelay = silk_min_int( decisionDelay, pitchL[ k ] - LTP_ORDER / 2 - 1 );
1.173 + }
1.174 + } else {
1.175 + if( lag > 0 ) {
1.176 + decisionDelay = silk_min_int( decisionDelay, lag - LTP_ORDER / 2 - 1 );
1.177 + }
1.178 + }
1.179 +
1.180 + if( psIndices->NLSFInterpCoef_Q2 == 4 ) {
1.181 + LSF_interpolation_flag = 0;
1.182 + } else {
1.183 + LSF_interpolation_flag = 1;
1.184 + }
1.185 +
1.186 + ALLOC( sLTP_Q15,
1.187 + psEncC->ltp_mem_length + psEncC->frame_length, opus_int32 );
1.188 + ALLOC( sLTP, psEncC->ltp_mem_length + psEncC->frame_length, opus_int16 );
1.189 + ALLOC( x_sc_Q10, psEncC->subfr_length, opus_int32 );
1.190 + ALLOC( delayedGain_Q10, DECISION_DELAY, opus_int32 );
1.191 + /* Set up pointers to start of sub frame */
1.192 + pxq = &NSQ->xq[ psEncC->ltp_mem_length ];
1.193 + NSQ->sLTP_shp_buf_idx = psEncC->ltp_mem_length;
1.194 + NSQ->sLTP_buf_idx = psEncC->ltp_mem_length;
1.195 + subfr = 0;
1.196 + for( k = 0; k < psEncC->nb_subfr; k++ ) {
1.197 + A_Q12 = &PredCoef_Q12[ ( ( k >> 1 ) | ( 1 - LSF_interpolation_flag ) ) * MAX_LPC_ORDER ];
1.198 + B_Q14 = <PCoef_Q14[ k * LTP_ORDER ];
1.199 + AR_shp_Q13 = &AR2_Q13[ k * MAX_SHAPE_LPC_ORDER ];
1.200 +
1.201 + /* Noise shape parameters */
1.202 + silk_assert( HarmShapeGain_Q14[ k ] >= 0 );
1.203 + HarmShapeFIRPacked_Q14 = silk_RSHIFT( HarmShapeGain_Q14[ k ], 2 );
1.204 + HarmShapeFIRPacked_Q14 |= silk_LSHIFT( (opus_int32)silk_RSHIFT( HarmShapeGain_Q14[ k ], 1 ), 16 );
1.205 +
1.206 + NSQ->rewhite_flag = 0;
1.207 + if( psIndices->signalType == TYPE_VOICED ) {
1.208 + /* Voiced */
1.209 + lag = pitchL[ k ];
1.210 +
1.211 + /* Re-whitening */
1.212 + if( ( k & ( 3 - silk_LSHIFT( LSF_interpolation_flag, 1 ) ) ) == 0 ) {
1.213 + if( k == 2 ) {
1.214 + /* RESET DELAYED DECISIONS */
1.215 + /* Find winner */
1.216 + RDmin_Q10 = psDelDec[ 0 ].RD_Q10;
1.217 + Winner_ind = 0;
1.218 + for( i = 1; i < psEncC->nStatesDelayedDecision; i++ ) {
1.219 + if( psDelDec[ i ].RD_Q10 < RDmin_Q10 ) {
1.220 + RDmin_Q10 = psDelDec[ i ].RD_Q10;
1.221 + Winner_ind = i;
1.222 + }
1.223 + }
1.224 + for( i = 0; i < psEncC->nStatesDelayedDecision; i++ ) {
1.225 + if( i != Winner_ind ) {
1.226 + psDelDec[ i ].RD_Q10 += ( silk_int32_MAX >> 4 );
1.227 + silk_assert( psDelDec[ i ].RD_Q10 >= 0 );
1.228 + }
1.229 + }
1.230 +
1.231 + /* Copy final part of signals from winner state to output and long-term filter states */
1.232 + psDD = &psDelDec[ Winner_ind ];
1.233 + last_smple_idx = smpl_buf_idx + decisionDelay;
1.234 + for( i = 0; i < decisionDelay; i++ ) {
1.235 + last_smple_idx = ( last_smple_idx - 1 ) & DECISION_DELAY_MASK;
1.236 + pulses[ i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDD->Q_Q10[ last_smple_idx ], 10 );
1.237 + pxq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND(
1.238 + silk_SMULWW( psDD->Xq_Q14[ last_smple_idx ], Gains_Q16[ 1 ] ), 14 ) );
1.239 + NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - decisionDelay + i ] = psDD->Shape_Q14[ last_smple_idx ];
1.240 + }
1.241 +
1.242 + subfr = 0;
1.243 + }
1.244 +
1.245 + /* Rewhiten with new A coefs */
1.246 + start_idx = psEncC->ltp_mem_length - lag - psEncC->predictLPCOrder - LTP_ORDER / 2;
1.247 + silk_assert( start_idx > 0 );
1.248 +
1.249 + silk_LPC_analysis_filter( &sLTP[ start_idx ], &NSQ->xq[ start_idx + k * psEncC->subfr_length ],
1.250 + A_Q12, psEncC->ltp_mem_length - start_idx, psEncC->predictLPCOrder );
1.251 +
1.252 + NSQ->sLTP_buf_idx = psEncC->ltp_mem_length;
1.253 + NSQ->rewhite_flag = 1;
1.254 + }
1.255 + }
1.256 +
1.257 + silk_nsq_del_dec_scale_states( psEncC, NSQ, psDelDec, x_Q3, x_sc_Q10, sLTP, sLTP_Q15, k,
1.258 + psEncC->nStatesDelayedDecision, LTP_scale_Q14, Gains_Q16, pitchL, psIndices->signalType, decisionDelay );
1.259 +
1.260 + silk_noise_shape_quantizer_del_dec( NSQ, psDelDec, psIndices->signalType, x_sc_Q10, pulses, pxq, sLTP_Q15,
1.261 + delayedGain_Q10, A_Q12, B_Q14, AR_shp_Q13, lag, HarmShapeFIRPacked_Q14, Tilt_Q14[ k ], LF_shp_Q14[ k ],
1.262 + Gains_Q16[ k ], Lambda_Q10, offset_Q10, psEncC->subfr_length, subfr++, psEncC->shapingLPCOrder,
1.263 + psEncC->predictLPCOrder, psEncC->warping_Q16, psEncC->nStatesDelayedDecision, &smpl_buf_idx, decisionDelay );
1.264 +
1.265 + x_Q3 += psEncC->subfr_length;
1.266 + pulses += psEncC->subfr_length;
1.267 + pxq += psEncC->subfr_length;
1.268 + }
1.269 +
1.270 + /* Find winner */
1.271 + RDmin_Q10 = psDelDec[ 0 ].RD_Q10;
1.272 + Winner_ind = 0;
1.273 + for( k = 1; k < psEncC->nStatesDelayedDecision; k++ ) {
1.274 + if( psDelDec[ k ].RD_Q10 < RDmin_Q10 ) {
1.275 + RDmin_Q10 = psDelDec[ k ].RD_Q10;
1.276 + Winner_ind = k;
1.277 + }
1.278 + }
1.279 +
1.280 + /* Copy final part of signals from winner state to output and long-term filter states */
1.281 + psDD = &psDelDec[ Winner_ind ];
1.282 + psIndices->Seed = psDD->SeedInit;
1.283 + last_smple_idx = smpl_buf_idx + decisionDelay;
1.284 + Gain_Q10 = silk_RSHIFT32( Gains_Q16[ psEncC->nb_subfr - 1 ], 6 );
1.285 + for( i = 0; i < decisionDelay; i++ ) {
1.286 + last_smple_idx = ( last_smple_idx - 1 ) & DECISION_DELAY_MASK;
1.287 + pulses[ i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDD->Q_Q10[ last_smple_idx ], 10 );
1.288 + pxq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND(
1.289 + silk_SMULWW( psDD->Xq_Q14[ last_smple_idx ], Gain_Q10 ), 8 ) );
1.290 + NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - decisionDelay + i ] = psDD->Shape_Q14[ last_smple_idx ];
1.291 + }
1.292 + silk_memcpy( NSQ->sLPC_Q14, &psDD->sLPC_Q14[ psEncC->subfr_length ], NSQ_LPC_BUF_LENGTH * sizeof( opus_int32 ) );
1.293 + silk_memcpy( NSQ->sAR2_Q14, psDD->sAR2_Q14, sizeof( psDD->sAR2_Q14 ) );
1.294 +
1.295 + /* Update states */
1.296 + NSQ->sLF_AR_shp_Q14 = psDD->LF_AR_Q14;
1.297 + NSQ->lagPrev = pitchL[ psEncC->nb_subfr - 1 ];
1.298 +
1.299 + /* Save quantized speech signal */
1.300 + /* DEBUG_STORE_DATA( enc.pcm, &NSQ->xq[psEncC->ltp_mem_length], psEncC->frame_length * sizeof( opus_int16 ) ) */
1.301 + silk_memmove( NSQ->xq, &NSQ->xq[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int16 ) );
1.302 + silk_memmove( NSQ->sLTP_shp_Q14, &NSQ->sLTP_shp_Q14[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int32 ) );
1.303 + RESTORE_STACK;
1.304 +}
1.305 +
1.306 +/******************************************/
1.307 +/* Noise shape quantizer for one subframe */
1.308 +/******************************************/
1.309 +static OPUS_INLINE void silk_noise_shape_quantizer_del_dec(
1.310 + silk_nsq_state *NSQ, /* I/O NSQ state */
1.311 + NSQ_del_dec_struct psDelDec[], /* I/O Delayed decision states */
1.312 + opus_int signalType, /* I Signal type */
1.313 + const opus_int32 x_Q10[], /* I */
1.314 + opus_int8 pulses[], /* O */
1.315 + opus_int16 xq[], /* O */
1.316 + opus_int32 sLTP_Q15[], /* I/O LTP filter state */
1.317 + opus_int32 delayedGain_Q10[], /* I/O Gain delay buffer */
1.318 + const opus_int16 a_Q12[], /* I Short term prediction coefs */
1.319 + const opus_int16 b_Q14[], /* I Long term prediction coefs */
1.320 + const opus_int16 AR_shp_Q13[], /* I Noise shaping coefs */
1.321 + opus_int lag, /* I Pitch lag */
1.322 + opus_int32 HarmShapeFIRPacked_Q14, /* I */
1.323 + opus_int Tilt_Q14, /* I Spectral tilt */
1.324 + opus_int32 LF_shp_Q14, /* I */
1.325 + opus_int32 Gain_Q16, /* I */
1.326 + opus_int Lambda_Q10, /* I */
1.327 + opus_int offset_Q10, /* I */
1.328 + opus_int length, /* I Input length */
1.329 + opus_int subfr, /* I Subframe number */
1.330 + opus_int shapingLPCOrder, /* I Shaping LPC filter order */
1.331 + opus_int predictLPCOrder, /* I Prediction filter order */
1.332 + opus_int warping_Q16, /* I */
1.333 + opus_int nStatesDelayedDecision, /* I Number of states in decision tree */
1.334 + opus_int *smpl_buf_idx, /* I Index to newest samples in buffers */
1.335 + opus_int decisionDelay /* I */
1.336 +)
1.337 +{
1.338 + opus_int i, j, k, Winner_ind, RDmin_ind, RDmax_ind, last_smple_idx;
1.339 + opus_int32 Winner_rand_state;
1.340 + opus_int32 LTP_pred_Q14, LPC_pred_Q14, n_AR_Q14, n_LTP_Q14;
1.341 + opus_int32 n_LF_Q14, r_Q10, rr_Q10, rd1_Q10, rd2_Q10, RDmin_Q10, RDmax_Q10;
1.342 + opus_int32 q1_Q0, q1_Q10, q2_Q10, exc_Q14, LPC_exc_Q14, xq_Q14, Gain_Q10;
1.343 + opus_int32 tmp1, tmp2, sLF_AR_shp_Q14;
1.344 + opus_int32 *pred_lag_ptr, *shp_lag_ptr, *psLPC_Q14;
1.345 + VARDECL( NSQ_sample_pair, psSampleState );
1.346 + NSQ_del_dec_struct *psDD;
1.347 + NSQ_sample_struct *psSS;
1.348 + SAVE_STACK;
1.349 +
1.350 + silk_assert( nStatesDelayedDecision > 0 );
1.351 + ALLOC( psSampleState, nStatesDelayedDecision, NSQ_sample_pair );
1.352 +
1.353 + shp_lag_ptr = &NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - lag + HARM_SHAPE_FIR_TAPS / 2 ];
1.354 + pred_lag_ptr = &sLTP_Q15[ NSQ->sLTP_buf_idx - lag + LTP_ORDER / 2 ];
1.355 + Gain_Q10 = silk_RSHIFT( Gain_Q16, 6 );
1.356 +
1.357 + for( i = 0; i < length; i++ ) {
1.358 + /* Perform common calculations used in all states */
1.359 +
1.360 + /* Long-term prediction */
1.361 + if( signalType == TYPE_VOICED ) {
1.362 + /* Unrolled loop */
1.363 + /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */
1.364 + LTP_pred_Q14 = 2;
1.365 + LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ 0 ], b_Q14[ 0 ] );
1.366 + LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -1 ], b_Q14[ 1 ] );
1.367 + LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -2 ], b_Q14[ 2 ] );
1.368 + LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -3 ], b_Q14[ 3 ] );
1.369 + LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -4 ], b_Q14[ 4 ] );
1.370 + LTP_pred_Q14 = silk_LSHIFT( LTP_pred_Q14, 1 ); /* Q13 -> Q14 */
1.371 + pred_lag_ptr++;
1.372 + } else {
1.373 + LTP_pred_Q14 = 0;
1.374 + }
1.375 +
1.376 + /* Long-term shaping */
1.377 + if( lag > 0 ) {
1.378 + /* Symmetric, packed FIR coefficients */
1.379 + n_LTP_Q14 = silk_SMULWB( silk_ADD32( shp_lag_ptr[ 0 ], shp_lag_ptr[ -2 ] ), HarmShapeFIRPacked_Q14 );
1.380 + n_LTP_Q14 = silk_SMLAWT( n_LTP_Q14, shp_lag_ptr[ -1 ], HarmShapeFIRPacked_Q14 );
1.381 + n_LTP_Q14 = silk_SUB_LSHIFT32( LTP_pred_Q14, n_LTP_Q14, 2 ); /* Q12 -> Q14 */
1.382 + shp_lag_ptr++;
1.383 + } else {
1.384 + n_LTP_Q14 = 0;
1.385 + }
1.386 +
1.387 + for( k = 0; k < nStatesDelayedDecision; k++ ) {
1.388 + /* Delayed decision state */
1.389 + psDD = &psDelDec[ k ];
1.390 +
1.391 + /* Sample state */
1.392 + psSS = psSampleState[ k ];
1.393 +
1.394 + /* Generate dither */
1.395 + psDD->Seed = silk_RAND( psDD->Seed );
1.396 +
1.397 + /* Pointer used in short term prediction and shaping */
1.398 + psLPC_Q14 = &psDD->sLPC_Q14[ NSQ_LPC_BUF_LENGTH - 1 + i ];
1.399 + /* Short-term prediction */
1.400 + silk_assert( predictLPCOrder == 10 || predictLPCOrder == 16 );
1.401 + /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */
1.402 + LPC_pred_Q14 = silk_RSHIFT( predictLPCOrder, 1 );
1.403 + LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ 0 ], a_Q12[ 0 ] );
1.404 + LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -1 ], a_Q12[ 1 ] );
1.405 + LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -2 ], a_Q12[ 2 ] );
1.406 + LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -3 ], a_Q12[ 3 ] );
1.407 + LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -4 ], a_Q12[ 4 ] );
1.408 + LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -5 ], a_Q12[ 5 ] );
1.409 + LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -6 ], a_Q12[ 6 ] );
1.410 + LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -7 ], a_Q12[ 7 ] );
1.411 + LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -8 ], a_Q12[ 8 ] );
1.412 + LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -9 ], a_Q12[ 9 ] );
1.413 + if( predictLPCOrder == 16 ) {
1.414 + LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -10 ], a_Q12[ 10 ] );
1.415 + LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -11 ], a_Q12[ 11 ] );
1.416 + LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -12 ], a_Q12[ 12 ] );
1.417 + LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -13 ], a_Q12[ 13 ] );
1.418 + LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -14 ], a_Q12[ 14 ] );
1.419 + LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -15 ], a_Q12[ 15 ] );
1.420 + }
1.421 + LPC_pred_Q14 = silk_LSHIFT( LPC_pred_Q14, 4 ); /* Q10 -> Q14 */
1.422 +
1.423 + /* Noise shape feedback */
1.424 + silk_assert( ( shapingLPCOrder & 1 ) == 0 ); /* check that order is even */
1.425 + /* Output of lowpass section */
1.426 + tmp2 = silk_SMLAWB( psLPC_Q14[ 0 ], psDD->sAR2_Q14[ 0 ], warping_Q16 );
1.427 + /* Output of allpass section */
1.428 + tmp1 = silk_SMLAWB( psDD->sAR2_Q14[ 0 ], psDD->sAR2_Q14[ 1 ] - tmp2, warping_Q16 );
1.429 + psDD->sAR2_Q14[ 0 ] = tmp2;
1.430 + n_AR_Q14 = silk_RSHIFT( shapingLPCOrder, 1 );
1.431 + n_AR_Q14 = silk_SMLAWB( n_AR_Q14, tmp2, AR_shp_Q13[ 0 ] );
1.432 + /* Loop over allpass sections */
1.433 + for( j = 2; j < shapingLPCOrder; j += 2 ) {
1.434 + /* Output of allpass section */
1.435 + tmp2 = silk_SMLAWB( psDD->sAR2_Q14[ j - 1 ], psDD->sAR2_Q14[ j + 0 ] - tmp1, warping_Q16 );
1.436 + psDD->sAR2_Q14[ j - 1 ] = tmp1;
1.437 + n_AR_Q14 = silk_SMLAWB( n_AR_Q14, tmp1, AR_shp_Q13[ j - 1 ] );
1.438 + /* Output of allpass section */
1.439 + tmp1 = silk_SMLAWB( psDD->sAR2_Q14[ j + 0 ], psDD->sAR2_Q14[ j + 1 ] - tmp2, warping_Q16 );
1.440 + psDD->sAR2_Q14[ j + 0 ] = tmp2;
1.441 + n_AR_Q14 = silk_SMLAWB( n_AR_Q14, tmp2, AR_shp_Q13[ j ] );
1.442 + }
1.443 + psDD->sAR2_Q14[ shapingLPCOrder - 1 ] = tmp1;
1.444 + n_AR_Q14 = silk_SMLAWB( n_AR_Q14, tmp1, AR_shp_Q13[ shapingLPCOrder - 1 ] );
1.445 +
1.446 + n_AR_Q14 = silk_LSHIFT( n_AR_Q14, 1 ); /* Q11 -> Q12 */
1.447 + n_AR_Q14 = silk_SMLAWB( n_AR_Q14, psDD->LF_AR_Q14, Tilt_Q14 ); /* Q12 */
1.448 + n_AR_Q14 = silk_LSHIFT( n_AR_Q14, 2 ); /* Q12 -> Q14 */
1.449 +
1.450 + n_LF_Q14 = silk_SMULWB( psDD->Shape_Q14[ *smpl_buf_idx ], LF_shp_Q14 ); /* Q12 */
1.451 + n_LF_Q14 = silk_SMLAWT( n_LF_Q14, psDD->LF_AR_Q14, LF_shp_Q14 ); /* Q12 */
1.452 + n_LF_Q14 = silk_LSHIFT( n_LF_Q14, 2 ); /* Q12 -> Q14 */
1.453 +
1.454 + /* Input minus prediction plus noise feedback */
1.455 + /* r = x[ i ] - LTP_pred - LPC_pred + n_AR + n_Tilt + n_LF + n_LTP */
1.456 + tmp1 = silk_ADD32( n_AR_Q14, n_LF_Q14 ); /* Q14 */
1.457 + tmp2 = silk_ADD32( n_LTP_Q14, LPC_pred_Q14 ); /* Q13 */
1.458 + tmp1 = silk_SUB32( tmp2, tmp1 ); /* Q13 */
1.459 + tmp1 = silk_RSHIFT_ROUND( tmp1, 4 ); /* Q10 */
1.460 +
1.461 + r_Q10 = silk_SUB32( x_Q10[ i ], tmp1 ); /* residual error Q10 */
1.462 +
1.463 + /* Flip sign depending on dither */
1.464 + if ( psDD->Seed < 0 ) {
1.465 + r_Q10 = -r_Q10;
1.466 + }
1.467 + r_Q10 = silk_LIMIT_32( r_Q10, -(31 << 10), 30 << 10 );
1.468 +
1.469 + /* Find two quantization level candidates and measure their rate-distortion */
1.470 + q1_Q10 = silk_SUB32( r_Q10, offset_Q10 );
1.471 + q1_Q0 = silk_RSHIFT( q1_Q10, 10 );
1.472 + if( q1_Q0 > 0 ) {
1.473 + q1_Q10 = silk_SUB32( silk_LSHIFT( q1_Q0, 10 ), QUANT_LEVEL_ADJUST_Q10 );
1.474 + q1_Q10 = silk_ADD32( q1_Q10, offset_Q10 );
1.475 + q2_Q10 = silk_ADD32( q1_Q10, 1024 );
1.476 + rd1_Q10 = silk_SMULBB( q1_Q10, Lambda_Q10 );
1.477 + rd2_Q10 = silk_SMULBB( q2_Q10, Lambda_Q10 );
1.478 + } else if( q1_Q0 == 0 ) {
1.479 + q1_Q10 = offset_Q10;
1.480 + q2_Q10 = silk_ADD32( q1_Q10, 1024 - QUANT_LEVEL_ADJUST_Q10 );
1.481 + rd1_Q10 = silk_SMULBB( q1_Q10, Lambda_Q10 );
1.482 + rd2_Q10 = silk_SMULBB( q2_Q10, Lambda_Q10 );
1.483 + } else if( q1_Q0 == -1 ) {
1.484 + q2_Q10 = offset_Q10;
1.485 + q1_Q10 = silk_SUB32( q2_Q10, 1024 - QUANT_LEVEL_ADJUST_Q10 );
1.486 + rd1_Q10 = silk_SMULBB( -q1_Q10, Lambda_Q10 );
1.487 + rd2_Q10 = silk_SMULBB( q2_Q10, Lambda_Q10 );
1.488 + } else { /* q1_Q0 < -1 */
1.489 + q1_Q10 = silk_ADD32( silk_LSHIFT( q1_Q0, 10 ), QUANT_LEVEL_ADJUST_Q10 );
1.490 + q1_Q10 = silk_ADD32( q1_Q10, offset_Q10 );
1.491 + q2_Q10 = silk_ADD32( q1_Q10, 1024 );
1.492 + rd1_Q10 = silk_SMULBB( -q1_Q10, Lambda_Q10 );
1.493 + rd2_Q10 = silk_SMULBB( -q2_Q10, Lambda_Q10 );
1.494 + }
1.495 + rr_Q10 = silk_SUB32( r_Q10, q1_Q10 );
1.496 + rd1_Q10 = silk_RSHIFT( silk_SMLABB( rd1_Q10, rr_Q10, rr_Q10 ), 10 );
1.497 + rr_Q10 = silk_SUB32( r_Q10, q2_Q10 );
1.498 + rd2_Q10 = silk_RSHIFT( silk_SMLABB( rd2_Q10, rr_Q10, rr_Q10 ), 10 );
1.499 +
1.500 + if( rd1_Q10 < rd2_Q10 ) {
1.501 + psSS[ 0 ].RD_Q10 = silk_ADD32( psDD->RD_Q10, rd1_Q10 );
1.502 + psSS[ 1 ].RD_Q10 = silk_ADD32( psDD->RD_Q10, rd2_Q10 );
1.503 + psSS[ 0 ].Q_Q10 = q1_Q10;
1.504 + psSS[ 1 ].Q_Q10 = q2_Q10;
1.505 + } else {
1.506 + psSS[ 0 ].RD_Q10 = silk_ADD32( psDD->RD_Q10, rd2_Q10 );
1.507 + psSS[ 1 ].RD_Q10 = silk_ADD32( psDD->RD_Q10, rd1_Q10 );
1.508 + psSS[ 0 ].Q_Q10 = q2_Q10;
1.509 + psSS[ 1 ].Q_Q10 = q1_Q10;
1.510 + }
1.511 +
1.512 + /* Update states for best quantization */
1.513 +
1.514 + /* Quantized excitation */
1.515 + exc_Q14 = silk_LSHIFT32( psSS[ 0 ].Q_Q10, 4 );
1.516 + if ( psDD->Seed < 0 ) {
1.517 + exc_Q14 = -exc_Q14;
1.518 + }
1.519 +
1.520 + /* Add predictions */
1.521 + LPC_exc_Q14 = silk_ADD32( exc_Q14, LTP_pred_Q14 );
1.522 + xq_Q14 = silk_ADD32( LPC_exc_Q14, LPC_pred_Q14 );
1.523 +
1.524 + /* Update states */
1.525 + sLF_AR_shp_Q14 = silk_SUB32( xq_Q14, n_AR_Q14 );
1.526 + psSS[ 0 ].sLTP_shp_Q14 = silk_SUB32( sLF_AR_shp_Q14, n_LF_Q14 );
1.527 + psSS[ 0 ].LF_AR_Q14 = sLF_AR_shp_Q14;
1.528 + psSS[ 0 ].LPC_exc_Q14 = LPC_exc_Q14;
1.529 + psSS[ 0 ].xq_Q14 = xq_Q14;
1.530 +
1.531 + /* Update states for second best quantization */
1.532 +
1.533 + /* Quantized excitation */
1.534 + exc_Q14 = silk_LSHIFT32( psSS[ 1 ].Q_Q10, 4 );
1.535 + if ( psDD->Seed < 0 ) {
1.536 + exc_Q14 = -exc_Q14;
1.537 + }
1.538 +
1.539 +
1.540 + /* Add predictions */
1.541 + LPC_exc_Q14 = silk_ADD32( exc_Q14, LTP_pred_Q14 );
1.542 + xq_Q14 = silk_ADD32( LPC_exc_Q14, LPC_pred_Q14 );
1.543 +
1.544 + /* Update states */
1.545 + sLF_AR_shp_Q14 = silk_SUB32( xq_Q14, n_AR_Q14 );
1.546 + psSS[ 1 ].sLTP_shp_Q14 = silk_SUB32( sLF_AR_shp_Q14, n_LF_Q14 );
1.547 + psSS[ 1 ].LF_AR_Q14 = sLF_AR_shp_Q14;
1.548 + psSS[ 1 ].LPC_exc_Q14 = LPC_exc_Q14;
1.549 + psSS[ 1 ].xq_Q14 = xq_Q14;
1.550 + }
1.551 +
1.552 + *smpl_buf_idx = ( *smpl_buf_idx - 1 ) & DECISION_DELAY_MASK; /* Index to newest samples */
1.553 + last_smple_idx = ( *smpl_buf_idx + decisionDelay ) & DECISION_DELAY_MASK; /* Index to decisionDelay old samples */
1.554 +
1.555 + /* Find winner */
1.556 + RDmin_Q10 = psSampleState[ 0 ][ 0 ].RD_Q10;
1.557 + Winner_ind = 0;
1.558 + for( k = 1; k < nStatesDelayedDecision; k++ ) {
1.559 + if( psSampleState[ k ][ 0 ].RD_Q10 < RDmin_Q10 ) {
1.560 + RDmin_Q10 = psSampleState[ k ][ 0 ].RD_Q10;
1.561 + Winner_ind = k;
1.562 + }
1.563 + }
1.564 +
1.565 + /* Increase RD values of expired states */
1.566 + Winner_rand_state = psDelDec[ Winner_ind ].RandState[ last_smple_idx ];
1.567 + for( k = 0; k < nStatesDelayedDecision; k++ ) {
1.568 + if( psDelDec[ k ].RandState[ last_smple_idx ] != Winner_rand_state ) {
1.569 + psSampleState[ k ][ 0 ].RD_Q10 = silk_ADD32( psSampleState[ k ][ 0 ].RD_Q10, silk_int32_MAX >> 4 );
1.570 + psSampleState[ k ][ 1 ].RD_Q10 = silk_ADD32( psSampleState[ k ][ 1 ].RD_Q10, silk_int32_MAX >> 4 );
1.571 + silk_assert( psSampleState[ k ][ 0 ].RD_Q10 >= 0 );
1.572 + }
1.573 + }
1.574 +
1.575 + /* Find worst in first set and best in second set */
1.576 + RDmax_Q10 = psSampleState[ 0 ][ 0 ].RD_Q10;
1.577 + RDmin_Q10 = psSampleState[ 0 ][ 1 ].RD_Q10;
1.578 + RDmax_ind = 0;
1.579 + RDmin_ind = 0;
1.580 + for( k = 1; k < nStatesDelayedDecision; k++ ) {
1.581 + /* find worst in first set */
1.582 + if( psSampleState[ k ][ 0 ].RD_Q10 > RDmax_Q10 ) {
1.583 + RDmax_Q10 = psSampleState[ k ][ 0 ].RD_Q10;
1.584 + RDmax_ind = k;
1.585 + }
1.586 + /* find best in second set */
1.587 + if( psSampleState[ k ][ 1 ].RD_Q10 < RDmin_Q10 ) {
1.588 + RDmin_Q10 = psSampleState[ k ][ 1 ].RD_Q10;
1.589 + RDmin_ind = k;
1.590 + }
1.591 + }
1.592 +
1.593 + /* Replace a state if best from second set outperforms worst in first set */
1.594 + if( RDmin_Q10 < RDmax_Q10 ) {
1.595 + silk_memcpy( ( (opus_int32 *)&psDelDec[ RDmax_ind ] ) + i,
1.596 + ( (opus_int32 *)&psDelDec[ RDmin_ind ] ) + i, sizeof( NSQ_del_dec_struct ) - i * sizeof( opus_int32) );
1.597 + silk_memcpy( &psSampleState[ RDmax_ind ][ 0 ], &psSampleState[ RDmin_ind ][ 1 ], sizeof( NSQ_sample_struct ) );
1.598 + }
1.599 +
1.600 + /* Write samples from winner to output and long-term filter states */
1.601 + psDD = &psDelDec[ Winner_ind ];
1.602 + if( subfr > 0 || i >= decisionDelay ) {
1.603 + pulses[ i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDD->Q_Q10[ last_smple_idx ], 10 );
1.604 + xq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND(
1.605 + silk_SMULWW( psDD->Xq_Q14[ last_smple_idx ], delayedGain_Q10[ last_smple_idx ] ), 8 ) );
1.606 + NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - decisionDelay ] = psDD->Shape_Q14[ last_smple_idx ];
1.607 + sLTP_Q15[ NSQ->sLTP_buf_idx - decisionDelay ] = psDD->Pred_Q15[ last_smple_idx ];
1.608 + }
1.609 + NSQ->sLTP_shp_buf_idx++;
1.610 + NSQ->sLTP_buf_idx++;
1.611 +
1.612 + /* Update states */
1.613 + for( k = 0; k < nStatesDelayedDecision; k++ ) {
1.614 + psDD = &psDelDec[ k ];
1.615 + psSS = &psSampleState[ k ][ 0 ];
1.616 + psDD->LF_AR_Q14 = psSS->LF_AR_Q14;
1.617 + psDD->sLPC_Q14[ NSQ_LPC_BUF_LENGTH + i ] = psSS->xq_Q14;
1.618 + psDD->Xq_Q14[ *smpl_buf_idx ] = psSS->xq_Q14;
1.619 + psDD->Q_Q10[ *smpl_buf_idx ] = psSS->Q_Q10;
1.620 + psDD->Pred_Q15[ *smpl_buf_idx ] = silk_LSHIFT32( psSS->LPC_exc_Q14, 1 );
1.621 + psDD->Shape_Q14[ *smpl_buf_idx ] = psSS->sLTP_shp_Q14;
1.622 + psDD->Seed = silk_ADD32_ovflw( psDD->Seed, silk_RSHIFT_ROUND( psSS->Q_Q10, 10 ) );
1.623 + psDD->RandState[ *smpl_buf_idx ] = psDD->Seed;
1.624 + psDD->RD_Q10 = psSS->RD_Q10;
1.625 + }
1.626 + delayedGain_Q10[ *smpl_buf_idx ] = Gain_Q10;
1.627 + }
1.628 + /* Update LPC states */
1.629 + for( k = 0; k < nStatesDelayedDecision; k++ ) {
1.630 + psDD = &psDelDec[ k ];
1.631 + silk_memcpy( psDD->sLPC_Q14, &psDD->sLPC_Q14[ length ], NSQ_LPC_BUF_LENGTH * sizeof( opus_int32 ) );
1.632 + }
1.633 + RESTORE_STACK;
1.634 +}
1.635 +
1.636 +static OPUS_INLINE void silk_nsq_del_dec_scale_states(
1.637 + const silk_encoder_state *psEncC, /* I Encoder State */
1.638 + silk_nsq_state *NSQ, /* I/O NSQ state */
1.639 + NSQ_del_dec_struct psDelDec[], /* I/O Delayed decision states */
1.640 + const opus_int32 x_Q3[], /* I Input in Q3 */
1.641 + opus_int32 x_sc_Q10[], /* O Input scaled with 1/Gain in Q10 */
1.642 + const opus_int16 sLTP[], /* I Re-whitened LTP state in Q0 */
1.643 + opus_int32 sLTP_Q15[], /* O LTP state matching scaled input */
1.644 + opus_int subfr, /* I Subframe number */
1.645 + opus_int nStatesDelayedDecision, /* I Number of del dec states */
1.646 + const opus_int LTP_scale_Q14, /* I LTP state scaling */
1.647 + const opus_int32 Gains_Q16[ MAX_NB_SUBFR ], /* I */
1.648 + const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lag */
1.649 + const opus_int signal_type, /* I Signal type */
1.650 + const opus_int decisionDelay /* I Decision delay */
1.651 +)
1.652 +{
1.653 + opus_int i, k, lag;
1.654 + opus_int32 gain_adj_Q16, inv_gain_Q31, inv_gain_Q23;
1.655 + NSQ_del_dec_struct *psDD;
1.656 +
1.657 + lag = pitchL[ subfr ];
1.658 + inv_gain_Q31 = silk_INVERSE32_varQ( silk_max( Gains_Q16[ subfr ], 1 ), 47 );
1.659 + silk_assert( inv_gain_Q31 != 0 );
1.660 +
1.661 + /* Calculate gain adjustment factor */
1.662 + if( Gains_Q16[ subfr ] != NSQ->prev_gain_Q16 ) {
1.663 + gain_adj_Q16 = silk_DIV32_varQ( NSQ->prev_gain_Q16, Gains_Q16[ subfr ], 16 );
1.664 + } else {
1.665 + gain_adj_Q16 = (opus_int32)1 << 16;
1.666 + }
1.667 +
1.668 + /* Scale input */
1.669 + inv_gain_Q23 = silk_RSHIFT_ROUND( inv_gain_Q31, 8 );
1.670 + for( i = 0; i < psEncC->subfr_length; i++ ) {
1.671 + x_sc_Q10[ i ] = silk_SMULWW( x_Q3[ i ], inv_gain_Q23 );
1.672 + }
1.673 +
1.674 + /* Save inverse gain */
1.675 + NSQ->prev_gain_Q16 = Gains_Q16[ subfr ];
1.676 +
1.677 + /* After rewhitening the LTP state is un-scaled, so scale with inv_gain_Q16 */
1.678 + if( NSQ->rewhite_flag ) {
1.679 + if( subfr == 0 ) {
1.680 + /* Do LTP downscaling */
1.681 + inv_gain_Q31 = silk_LSHIFT( silk_SMULWB( inv_gain_Q31, LTP_scale_Q14 ), 2 );
1.682 + }
1.683 + for( i = NSQ->sLTP_buf_idx - lag - LTP_ORDER / 2; i < NSQ->sLTP_buf_idx; i++ ) {
1.684 + silk_assert( i < MAX_FRAME_LENGTH );
1.685 + sLTP_Q15[ i ] = silk_SMULWB( inv_gain_Q31, sLTP[ i ] );
1.686 + }
1.687 + }
1.688 +
1.689 + /* Adjust for changing gain */
1.690 + if( gain_adj_Q16 != (opus_int32)1 << 16 ) {
1.691 + /* Scale long-term shaping state */
1.692 + for( i = NSQ->sLTP_shp_buf_idx - psEncC->ltp_mem_length; i < NSQ->sLTP_shp_buf_idx; i++ ) {
1.693 + NSQ->sLTP_shp_Q14[ i ] = silk_SMULWW( gain_adj_Q16, NSQ->sLTP_shp_Q14[ i ] );
1.694 + }
1.695 +
1.696 + /* Scale long-term prediction state */
1.697 + if( signal_type == TYPE_VOICED && NSQ->rewhite_flag == 0 ) {
1.698 + for( i = NSQ->sLTP_buf_idx - lag - LTP_ORDER / 2; i < NSQ->sLTP_buf_idx - decisionDelay; i++ ) {
1.699 + sLTP_Q15[ i ] = silk_SMULWW( gain_adj_Q16, sLTP_Q15[ i ] );
1.700 + }
1.701 + }
1.702 +
1.703 + for( k = 0; k < nStatesDelayedDecision; k++ ) {
1.704 + psDD = &psDelDec[ k ];
1.705 +
1.706 + /* Scale scalar states */
1.707 + psDD->LF_AR_Q14 = silk_SMULWW( gain_adj_Q16, psDD->LF_AR_Q14 );
1.708 +
1.709 + /* Scale short-term prediction and shaping states */
1.710 + for( i = 0; i < NSQ_LPC_BUF_LENGTH; i++ ) {
1.711 + psDD->sLPC_Q14[ i ] = silk_SMULWW( gain_adj_Q16, psDD->sLPC_Q14[ i ] );
1.712 + }
1.713 + for( i = 0; i < MAX_SHAPE_LPC_ORDER; i++ ) {
1.714 + psDD->sAR2_Q14[ i ] = silk_SMULWW( gain_adj_Q16, psDD->sAR2_Q14[ i ] );
1.715 + }
1.716 + for( i = 0; i < DECISION_DELAY; i++ ) {
1.717 + psDD->Pred_Q15[ i ] = silk_SMULWW( gain_adj_Q16, psDD->Pred_Q15[ i ] );
1.718 + psDD->Shape_Q14[ i ] = silk_SMULWW( gain_adj_Q16, psDD->Shape_Q14[ i ] );
1.719 + }
1.720 + }
1.721 + }
1.722 +}
