1 /*********************************************************************** 2 Copyright (c) 2006-2011, Skype Limited. All rights reserved. 3 Redistribution and use in source and binary forms, with or without 4 modification, are permitted provided that the following conditions 5 are met: 6 - Redistributions of source code must retain the above copyright notice, 7 this list of conditions and the following disclaimer. 8 - Redistributions in binary form must reproduce the above copyright 9 notice, this list of conditions and the following disclaimer in the 10 documentation and/or other materials provided with the distribution. 11 - Neither the name of Internet Society, IETF or IETF Trust, nor the 12 names of specific contributors, may be used to endorse or promote 13 products derived from this software without specific prior written 14 permission. 15 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" 16 AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 17 IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 18 ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE 19 LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 20 CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 21 SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 22 INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 23 CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 24 ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 25 POSSIBILITY OF SUCH DAMAGE. 26 ***********************************************************************/ 27 28 #ifdef HAVE_CONFIG_H 29 #include "config.h" 30 #endif 31 32 #include "main.h" 33 #include "stack_alloc.h" 34 35 typedef struct { 36 opus_int32 sLPC_Q14[ MAX_SUB_FRAME_LENGTH + NSQ_LPC_BUF_LENGTH ]; 37 opus_int32 RandState[ DECISION_DELAY ]; 38 opus_int32 Q_Q10[ DECISION_DELAY ]; 39 opus_int32 Xq_Q14[ DECISION_DELAY ]; 40 opus_int32 Pred_Q15[ DECISION_DELAY ]; 41 opus_int32 Shape_Q14[ DECISION_DELAY ]; 42 opus_int32 sAR2_Q14[ MAX_SHAPE_LPC_ORDER ]; 43 opus_int32 LF_AR_Q14; 44 opus_int32 Seed; 45 opus_int32 SeedInit; 46 opus_int32 RD_Q10; 47 } NSQ_del_dec_struct; 48 49 typedef struct { 50 opus_int32 Q_Q10; 51 opus_int32 RD_Q10; 52 opus_int32 xq_Q14; 53 opus_int32 LF_AR_Q14; 54 opus_int32 sLTP_shp_Q14; 55 opus_int32 LPC_exc_Q14; 56 } NSQ_sample_struct; 57 58 typedef NSQ_sample_struct NSQ_sample_pair[ 2 ]; 59 60 static inline void silk_nsq_del_dec_scale_states( 61 const silk_encoder_state *psEncC, /* I Encoder State */ 62 silk_nsq_state *NSQ, /* I/O NSQ state */ 63 NSQ_del_dec_struct psDelDec[], /* I/O Delayed decision states */ 64 const opus_int32 x_Q3[], /* I Input in Q3 */ 65 opus_int32 x_sc_Q10[], /* O Input scaled with 1/Gain in Q10 */ 66 const opus_int16 sLTP[], /* I Re-whitened LTP state in Q0 */ 67 opus_int32 sLTP_Q15[], /* O LTP state matching scaled input */ 68 opus_int subfr, /* I Subframe number */ 69 opus_int nStatesDelayedDecision, /* I Number of del dec states */ 70 const opus_int LTP_scale_Q14, /* I LTP state scaling */ 71 const opus_int32 Gains_Q16[ MAX_NB_SUBFR ], /* I */ 72 const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lag */ 73 const opus_int signal_type, /* I Signal type */ 74 const opus_int decisionDelay /* I Decision delay */ 75 ); 76 77 /******************************************/ 78 /* Noise shape quantizer for one subframe */ 79 /******************************************/ 80 static inline void silk_noise_shape_quantizer_del_dec( 81 silk_nsq_state *NSQ, /* I/O NSQ state */ 82 NSQ_del_dec_struct psDelDec[], /* I/O Delayed decision states */ 83 opus_int signalType, /* I Signal type */ 84 const opus_int32 x_Q10[], /* I */ 85 opus_int8 pulses[], /* O */ 86 opus_int16 xq[], /* O */ 87 opus_int32 sLTP_Q15[], /* I/O LTP filter state */ 88 opus_int32 delayedGain_Q10[], /* I/O Gain delay buffer */ 89 const opus_int16 a_Q12[], /* I Short term prediction coefs */ 90 const opus_int16 b_Q14[], /* I Long term prediction coefs */ 91 const opus_int16 AR_shp_Q13[], /* I Noise shaping coefs */ 92 opus_int lag, /* I Pitch lag */ 93 opus_int32 HarmShapeFIRPacked_Q14, /* I */ 94 opus_int Tilt_Q14, /* I Spectral tilt */ 95 opus_int32 LF_shp_Q14, /* I */ 96 opus_int32 Gain_Q16, /* I */ 97 opus_int Lambda_Q10, /* I */ 98 opus_int offset_Q10, /* I */ 99 opus_int length, /* I Input length */ 100 opus_int subfr, /* I Subframe number */ 101 opus_int shapingLPCOrder, /* I Shaping LPC filter order */ 102 opus_int predictLPCOrder, /* I Prediction filter order */ 103 opus_int warping_Q16, /* I */ 104 opus_int nStatesDelayedDecision, /* I Number of states in decision tree */ 105 opus_int *smpl_buf_idx, /* I Index to newest samples in buffers */ 106 opus_int decisionDelay /* I */ 107 ); 108 109 void silk_NSQ_del_dec( 110 const silk_encoder_state *psEncC, /* I/O Encoder State */ 111 silk_nsq_state *NSQ, /* I/O NSQ state */ 112 SideInfoIndices *psIndices, /* I/O Quantization Indices */ 113 const opus_int32 x_Q3[], /* I Prefiltered input signal */ 114 opus_int8 pulses[], /* O Quantized pulse signal */ 115 const opus_int16 PredCoef_Q12[ 2 * MAX_LPC_ORDER ], /* I Short term prediction coefs */ 116 const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ], /* I Long term prediction coefs */ 117 const opus_int16 AR2_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */ 118 const opus_int HarmShapeGain_Q14[ MAX_NB_SUBFR ], /* I Long term shaping coefs */ 119 const opus_int Tilt_Q14[ MAX_NB_SUBFR ], /* I Spectral tilt */ 120 const opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ], /* I Low frequency shaping coefs */ 121 const opus_int32 Gains_Q16[ MAX_NB_SUBFR ], /* I Quantization step sizes */ 122 const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lags */ 123 const opus_int Lambda_Q10, /* I Rate/distortion tradeoff */ 124 const opus_int LTP_scale_Q14 /* I LTP state scaling */ 125 ) 126 { 127 opus_int i, k, lag, start_idx, LSF_interpolation_flag, Winner_ind, subfr; 128 opus_int last_smple_idx, smpl_buf_idx, decisionDelay; 129 const opus_int16 *A_Q12, *B_Q14, *AR_shp_Q13; 130 opus_int16 *pxq; 131 VARDECL( opus_int32, sLTP_Q15 ); 132 VARDECL( opus_int16, sLTP ); 133 opus_int32 HarmShapeFIRPacked_Q14; 134 opus_int offset_Q10; 135 opus_int32 RDmin_Q10, Gain_Q10; 136 VARDECL( opus_int32, x_sc_Q10 ); 137 VARDECL( opus_int32, delayedGain_Q10 ); 138 VARDECL( NSQ_del_dec_struct, psDelDec ); 139 NSQ_del_dec_struct *psDD; 140 SAVE_STACK; 141 142 /* Set unvoiced lag to the previous one, overwrite later for voiced */ 143 lag = NSQ->lagPrev; 144 145 silk_assert( NSQ->prev_gain_Q16 != 0 ); 146 147 /* Initialize delayed decision states */ 148 ALLOC( psDelDec, psEncC->nStatesDelayedDecision, NSQ_del_dec_struct ); 149 silk_memset( psDelDec, 0, psEncC->nStatesDelayedDecision * sizeof( NSQ_del_dec_struct ) ); 150 for( k = 0; k < psEncC->nStatesDelayedDecision; k++ ) { 151 psDD = &psDelDec[ k ]; 152 psDD->Seed = ( k + psIndices->Seed ) & 3; 153 psDD->SeedInit = psDD->Seed; 154 psDD->RD_Q10 = 0; 155 psDD->LF_AR_Q14 = NSQ->sLF_AR_shp_Q14; 156 psDD->Shape_Q14[ 0 ] = NSQ->sLTP_shp_Q14[ psEncC->ltp_mem_length - 1 ]; 157 silk_memcpy( psDD->sLPC_Q14, NSQ->sLPC_Q14, NSQ_LPC_BUF_LENGTH * sizeof( opus_int32 ) ); 158 silk_memcpy( psDD->sAR2_Q14, NSQ->sAR2_Q14, sizeof( NSQ->sAR2_Q14 ) ); 159 } 160 161 offset_Q10 = silk_Quantization_Offsets_Q10[ psIndices->signalType >> 1 ][ psIndices->quantOffsetType ]; 162 smpl_buf_idx = 0; /* index of oldest samples */ 163 164 decisionDelay = silk_min_int( DECISION_DELAY, psEncC->subfr_length ); 165 166 /* For voiced frames limit the decision delay to lower than the pitch lag */ 167 if( psIndices->signalType == TYPE_VOICED ) { 168 for( k = 0; k < psEncC->nb_subfr; k++ ) { 169 decisionDelay = silk_min_int( decisionDelay, pitchL[ k ] - LTP_ORDER / 2 - 1 ); 170 } 171 } else { 172 if( lag > 0 ) { 173 decisionDelay = silk_min_int( decisionDelay, lag - LTP_ORDER / 2 - 1 ); 174 } 175 } 176 177 if( psIndices->NLSFInterpCoef_Q2 == 4 ) { 178 LSF_interpolation_flag = 0; 179 } else { 180 LSF_interpolation_flag = 1; 181 } 182 183 ALLOC( sLTP_Q15, 184 psEncC->ltp_mem_length + psEncC->frame_length, opus_int32 ); 185 ALLOC( sLTP, psEncC->ltp_mem_length + psEncC->frame_length, opus_int16 ); 186 ALLOC( x_sc_Q10, psEncC->subfr_length, opus_int32 ); 187 ALLOC( delayedGain_Q10, DECISION_DELAY, opus_int32 ); 188 /* Set up pointers to start of sub frame */ 189 pxq = &NSQ->xq[ psEncC->ltp_mem_length ]; 190 NSQ->sLTP_shp_buf_idx = psEncC->ltp_mem_length; 191 NSQ->sLTP_buf_idx = psEncC->ltp_mem_length; 192 subfr = 0; 193 for( k = 0; k < psEncC->nb_subfr; k++ ) { 194 A_Q12 = &PredCoef_Q12[ ( ( k >> 1 ) | ( 1 - LSF_interpolation_flag ) ) * MAX_LPC_ORDER ]; 195 B_Q14 = <PCoef_Q14[ k * LTP_ORDER ]; 196 AR_shp_Q13 = &AR2_Q13[ k * MAX_SHAPE_LPC_ORDER ]; 197 198 /* Noise shape parameters */ 199 silk_assert( HarmShapeGain_Q14[ k ] >= 0 ); 200 HarmShapeFIRPacked_Q14 = silk_RSHIFT( HarmShapeGain_Q14[ k ], 2 ); 201 HarmShapeFIRPacked_Q14 |= silk_LSHIFT( (opus_int32)silk_RSHIFT( HarmShapeGain_Q14[ k ], 1 ), 16 ); 202 203 NSQ->rewhite_flag = 0; 204 if( psIndices->signalType == TYPE_VOICED ) { 205 /* Voiced */ 206 lag = pitchL[ k ]; 207 208 /* Re-whitening */ 209 if( ( k & ( 3 - silk_LSHIFT( LSF_interpolation_flag, 1 ) ) ) == 0 ) { 210 if( k == 2 ) { 211 /* RESET DELAYED DECISIONS */ 212 /* Find winner */ 213 RDmin_Q10 = psDelDec[ 0 ].RD_Q10; 214 Winner_ind = 0; 215 for( i = 1; i < psEncC->nStatesDelayedDecision; i++ ) { 216 if( psDelDec[ i ].RD_Q10 < RDmin_Q10 ) { 217 RDmin_Q10 = psDelDec[ i ].RD_Q10; 218 Winner_ind = i; 219 } 220 } 221 for( i = 0; i < psEncC->nStatesDelayedDecision; i++ ) { 222 if( i != Winner_ind ) { 223 psDelDec[ i ].RD_Q10 += ( silk_int32_MAX >> 4 ); 224 silk_assert( psDelDec[ i ].RD_Q10 >= 0 ); 225 } 226 } 227 228 /* Copy final part of signals from winner state to output and long-term filter states */ 229 psDD = &psDelDec[ Winner_ind ]; 230 last_smple_idx = smpl_buf_idx + decisionDelay; 231 for( i = 0; i < decisionDelay; i++ ) { 232 last_smple_idx = ( last_smple_idx - 1 ) & DECISION_DELAY_MASK; 233 pulses[ i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDD->Q_Q10[ last_smple_idx ], 10 ); 234 pxq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( 235 silk_SMULWW( psDD->Xq_Q14[ last_smple_idx ], Gains_Q16[ 1 ] ), 14 ) ); 236 NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - decisionDelay + i ] = psDD->Shape_Q14[ last_smple_idx ]; 237 } 238 239 subfr = 0; 240 } 241 242 /* Rewhiten with new A coefs */ 243 start_idx = psEncC->ltp_mem_length - lag - psEncC->predictLPCOrder - LTP_ORDER / 2; 244 silk_assert( start_idx > 0 ); 245 246 silk_LPC_analysis_filter( &sLTP[ start_idx ], &NSQ->xq[ start_idx + k * psEncC->subfr_length ], 247 A_Q12, psEncC->ltp_mem_length - start_idx, psEncC->predictLPCOrder ); 248 249 NSQ->sLTP_buf_idx = psEncC->ltp_mem_length; 250 NSQ->rewhite_flag = 1; 251 } 252 } 253 254 silk_nsq_del_dec_scale_states( psEncC, NSQ, psDelDec, x_Q3, x_sc_Q10, sLTP, sLTP_Q15, k, 255 psEncC->nStatesDelayedDecision, LTP_scale_Q14, Gains_Q16, pitchL, psIndices->signalType, decisionDelay ); 256 257 silk_noise_shape_quantizer_del_dec( NSQ, psDelDec, psIndices->signalType, x_sc_Q10, pulses, pxq, sLTP_Q15, 258 delayedGain_Q10, A_Q12, B_Q14, AR_shp_Q13, lag, HarmShapeFIRPacked_Q14, Tilt_Q14[ k ], LF_shp_Q14[ k ], 259 Gains_Q16[ k ], Lambda_Q10, offset_Q10, psEncC->subfr_length, subfr++, psEncC->shapingLPCOrder, 260 psEncC->predictLPCOrder, psEncC->warping_Q16, psEncC->nStatesDelayedDecision, &smpl_buf_idx, decisionDelay ); 261 262 x_Q3 += psEncC->subfr_length; 263 pulses += psEncC->subfr_length; 264 pxq += psEncC->subfr_length; 265 } 266 267 /* Find winner */ 268 RDmin_Q10 = psDelDec[ 0 ].RD_Q10; 269 Winner_ind = 0; 270 for( k = 1; k < psEncC->nStatesDelayedDecision; k++ ) { 271 if( psDelDec[ k ].RD_Q10 < RDmin_Q10 ) { 272 RDmin_Q10 = psDelDec[ k ].RD_Q10; 273 Winner_ind = k; 274 } 275 } 276 277 /* Copy final part of signals from winner state to output and long-term filter states */ 278 psDD = &psDelDec[ Winner_ind ]; 279 psIndices->Seed = psDD->SeedInit; 280 last_smple_idx = smpl_buf_idx + decisionDelay; 281 Gain_Q10 = silk_RSHIFT32( Gains_Q16[ psEncC->nb_subfr - 1 ], 6 ); 282 for( i = 0; i < decisionDelay; i++ ) { 283 last_smple_idx = ( last_smple_idx - 1 ) & DECISION_DELAY_MASK; 284 pulses[ i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDD->Q_Q10[ last_smple_idx ], 10 ); 285 pxq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( 286 silk_SMULWW( psDD->Xq_Q14[ last_smple_idx ], Gain_Q10 ), 8 ) ); 287 NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - decisionDelay + i ] = psDD->Shape_Q14[ last_smple_idx ]; 288 } 289 silk_memcpy( NSQ->sLPC_Q14, &psDD->sLPC_Q14[ psEncC->subfr_length ], NSQ_LPC_BUF_LENGTH * sizeof( opus_int32 ) ); 290 silk_memcpy( NSQ->sAR2_Q14, psDD->sAR2_Q14, sizeof( psDD->sAR2_Q14 ) ); 291 292 /* Update states */ 293 NSQ->sLF_AR_shp_Q14 = psDD->LF_AR_Q14; 294 NSQ->lagPrev = pitchL[ psEncC->nb_subfr - 1 ]; 295 296 /* Save quantized speech signal */ 297 /* DEBUG_STORE_DATA( enc.pcm, &NSQ->xq[psEncC->ltp_mem_length], psEncC->frame_length * sizeof( opus_int16 ) ) */ 298 silk_memmove( NSQ->xq, &NSQ->xq[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int16 ) ); 299 silk_memmove( NSQ->sLTP_shp_Q14, &NSQ->sLTP_shp_Q14[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int32 ) ); 300 RESTORE_STACK; 301 } 302 303 /******************************************/ 304 /* Noise shape quantizer for one subframe */ 305 /******************************************/ 306 static inline void silk_noise_shape_quantizer_del_dec( 307 silk_nsq_state *NSQ, /* I/O NSQ state */ 308 NSQ_del_dec_struct psDelDec[], /* I/O Delayed decision states */ 309 opus_int signalType, /* I Signal type */ 310 const opus_int32 x_Q10[], /* I */ 311 opus_int8 pulses[], /* O */ 312 opus_int16 xq[], /* O */ 313 opus_int32 sLTP_Q15[], /* I/O LTP filter state */ 314 opus_int32 delayedGain_Q10[], /* I/O Gain delay buffer */ 315 const opus_int16 a_Q12[], /* I Short term prediction coefs */ 316 const opus_int16 b_Q14[], /* I Long term prediction coefs */ 317 const opus_int16 AR_shp_Q13[], /* I Noise shaping coefs */ 318 opus_int lag, /* I Pitch lag */ 319 opus_int32 HarmShapeFIRPacked_Q14, /* I */ 320 opus_int Tilt_Q14, /* I Spectral tilt */ 321 opus_int32 LF_shp_Q14, /* I */ 322 opus_int32 Gain_Q16, /* I */ 323 opus_int Lambda_Q10, /* I */ 324 opus_int offset_Q10, /* I */ 325 opus_int length, /* I Input length */ 326 opus_int subfr, /* I Subframe number */ 327 opus_int shapingLPCOrder, /* I Shaping LPC filter order */ 328 opus_int predictLPCOrder, /* I Prediction filter order */ 329 opus_int warping_Q16, /* I */ 330 opus_int nStatesDelayedDecision, /* I Number of states in decision tree */ 331 opus_int *smpl_buf_idx, /* I Index to newest samples in buffers */ 332 opus_int decisionDelay /* I */ 333 ) 334 { 335 opus_int i, j, k, Winner_ind, RDmin_ind, RDmax_ind, last_smple_idx; 336 opus_int32 Winner_rand_state; 337 opus_int32 LTP_pred_Q14, LPC_pred_Q14, n_AR_Q14, n_LTP_Q14; 338 opus_int32 n_LF_Q14, r_Q10, rr_Q10, rd1_Q10, rd2_Q10, RDmin_Q10, RDmax_Q10; 339 opus_int32 q1_Q0, q1_Q10, q2_Q10, exc_Q14, LPC_exc_Q14, xq_Q14, Gain_Q10; 340 opus_int32 tmp1, tmp2, sLF_AR_shp_Q14; 341 opus_int32 *pred_lag_ptr, *shp_lag_ptr, *psLPC_Q14; 342 VARDECL( NSQ_sample_pair, psSampleState ); 343 NSQ_del_dec_struct *psDD; 344 NSQ_sample_struct *psSS; 345 SAVE_STACK; 346 347 silk_assert( nStatesDelayedDecision > 0 ); 348 ALLOC( psSampleState, nStatesDelayedDecision, NSQ_sample_pair ); 349 350 shp_lag_ptr = &NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - lag + HARM_SHAPE_FIR_TAPS / 2 ]; 351 pred_lag_ptr = &sLTP_Q15[ NSQ->sLTP_buf_idx - lag + LTP_ORDER / 2 ]; 352 Gain_Q10 = silk_RSHIFT( Gain_Q16, 6 ); 353 354 for( i = 0; i < length; i++ ) { 355 /* Perform common calculations used in all states */ 356 357 /* Long-term prediction */ 358 if( signalType == TYPE_VOICED ) { 359 /* Unrolled loop */ 360 /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */ 361 LTP_pred_Q14 = 2; 362 LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ 0 ], b_Q14[ 0 ] ); 363 LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -1 ], b_Q14[ 1 ] ); 364 LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -2 ], b_Q14[ 2 ] ); 365 LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -3 ], b_Q14[ 3 ] ); 366 LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -4 ], b_Q14[ 4 ] ); 367 LTP_pred_Q14 = silk_LSHIFT( LTP_pred_Q14, 1 ); /* Q13 -> Q14 */ 368 pred_lag_ptr++; 369 } else { 370 LTP_pred_Q14 = 0; 371 } 372 373 /* Long-term shaping */ 374 if( lag > 0 ) { 375 /* Symmetric, packed FIR coefficients */ 376 n_LTP_Q14 = silk_SMULWB( silk_ADD32( shp_lag_ptr[ 0 ], shp_lag_ptr[ -2 ] ), HarmShapeFIRPacked_Q14 ); 377 n_LTP_Q14 = silk_SMLAWT( n_LTP_Q14, shp_lag_ptr[ -1 ], HarmShapeFIRPacked_Q14 ); 378 n_LTP_Q14 = silk_SUB_LSHIFT32( LTP_pred_Q14, n_LTP_Q14, 2 ); /* Q12 -> Q14 */ 379 shp_lag_ptr++; 380 } else { 381 n_LTP_Q14 = 0; 382 } 383 384 for( k = 0; k < nStatesDelayedDecision; k++ ) { 385 /* Delayed decision state */ 386 psDD = &psDelDec[ k ]; 387 388 /* Sample state */ 389 psSS = psSampleState[ k ]; 390 391 /* Generate dither */ 392 psDD->Seed = silk_RAND( psDD->Seed ); 393 394 /* Pointer used in short term prediction and shaping */ 395 psLPC_Q14 = &psDD->sLPC_Q14[ NSQ_LPC_BUF_LENGTH - 1 + i ]; 396 /* Short-term prediction */ 397 silk_assert( predictLPCOrder == 10 || predictLPCOrder == 16 ); 398 /* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */ 399 LPC_pred_Q14 = silk_RSHIFT( predictLPCOrder, 1 ); 400 LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ 0 ], a_Q12[ 0 ] ); 401 LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -1 ], a_Q12[ 1 ] ); 402 LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -2 ], a_Q12[ 2 ] ); 403 LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -3 ], a_Q12[ 3 ] ); 404 LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -4 ], a_Q12[ 4 ] ); 405 LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -5 ], a_Q12[ 5 ] ); 406 LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -6 ], a_Q12[ 6 ] ); 407 LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -7 ], a_Q12[ 7 ] ); 408 LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -8 ], a_Q12[ 8 ] ); 409 LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -9 ], a_Q12[ 9 ] ); 410 if( predictLPCOrder == 16 ) { 411 LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -10 ], a_Q12[ 10 ] ); 412 LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -11 ], a_Q12[ 11 ] ); 413 LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -12 ], a_Q12[ 12 ] ); 414 LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -13 ], a_Q12[ 13 ] ); 415 LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -14 ], a_Q12[ 14 ] ); 416 LPC_pred_Q14 = silk_SMLAWB( LPC_pred_Q14, psLPC_Q14[ -15 ], a_Q12[ 15 ] ); 417 } 418 LPC_pred_Q14 = silk_LSHIFT( LPC_pred_Q14, 4 ); /* Q10 -> Q14 */ 419 420 /* Noise shape feedback */ 421 silk_assert( ( shapingLPCOrder & 1 ) == 0 ); /* check that order is even */ 422 /* Output of lowpass section */ 423 tmp2 = silk_SMLAWB( psLPC_Q14[ 0 ], psDD->sAR2_Q14[ 0 ], warping_Q16 ); 424 /* Output of allpass section */ 425 tmp1 = silk_SMLAWB( psDD->sAR2_Q14[ 0 ], psDD->sAR2_Q14[ 1 ] - tmp2, warping_Q16 ); 426 psDD->sAR2_Q14[ 0 ] = tmp2; 427 n_AR_Q14 = silk_RSHIFT( shapingLPCOrder, 1 ); 428 n_AR_Q14 = silk_SMLAWB( n_AR_Q14, tmp2, AR_shp_Q13[ 0 ] ); 429 /* Loop over allpass sections */ 430 for( j = 2; j < shapingLPCOrder; j += 2 ) { 431 /* Output of allpass section */ 432 tmp2 = silk_SMLAWB( psDD->sAR2_Q14[ j - 1 ], psDD->sAR2_Q14[ j + 0 ] - tmp1, warping_Q16 ); 433 psDD->sAR2_Q14[ j - 1 ] = tmp1; 434 n_AR_Q14 = silk_SMLAWB( n_AR_Q14, tmp1, AR_shp_Q13[ j - 1 ] ); 435 /* Output of allpass section */ 436 tmp1 = silk_SMLAWB( psDD->sAR2_Q14[ j + 0 ], psDD->sAR2_Q14[ j + 1 ] - tmp2, warping_Q16 ); 437 psDD->sAR2_Q14[ j + 0 ] = tmp2; 438 n_AR_Q14 = silk_SMLAWB( n_AR_Q14, tmp2, AR_shp_Q13[ j ] ); 439 } 440 psDD->sAR2_Q14[ shapingLPCOrder - 1 ] = tmp1; 441 n_AR_Q14 = silk_SMLAWB( n_AR_Q14, tmp1, AR_shp_Q13[ shapingLPCOrder - 1 ] ); 442 443 n_AR_Q14 = silk_LSHIFT( n_AR_Q14, 1 ); /* Q11 -> Q12 */ 444 n_AR_Q14 = silk_SMLAWB( n_AR_Q14, psDD->LF_AR_Q14, Tilt_Q14 ); /* Q12 */ 445 n_AR_Q14 = silk_LSHIFT( n_AR_Q14, 2 ); /* Q12 -> Q14 */ 446 447 n_LF_Q14 = silk_SMULWB( psDD->Shape_Q14[ *smpl_buf_idx ], LF_shp_Q14 ); /* Q12 */ 448 n_LF_Q14 = silk_SMLAWT( n_LF_Q14, psDD->LF_AR_Q14, LF_shp_Q14 ); /* Q12 */ 449 n_LF_Q14 = silk_LSHIFT( n_LF_Q14, 2 ); /* Q12 -> Q14 */ 450 451 /* Input minus prediction plus noise feedback */ 452 /* r = x[ i ] - LTP_pred - LPC_pred + n_AR + n_Tilt + n_LF + n_LTP */ 453 tmp1 = silk_ADD32( n_AR_Q14, n_LF_Q14 ); /* Q14 */ 454 tmp2 = silk_ADD32( n_LTP_Q14, LPC_pred_Q14 ); /* Q13 */ 455 tmp1 = silk_SUB32( tmp2, tmp1 ); /* Q13 */ 456 tmp1 = silk_RSHIFT_ROUND( tmp1, 4 ); /* Q10 */ 457 458 r_Q10 = silk_SUB32( x_Q10[ i ], tmp1 ); /* residual error Q10 */ 459 460 /* Flip sign depending on dither */ 461 if ( psDD->Seed < 0 ) { 462 r_Q10 = -r_Q10; 463 } 464 r_Q10 = silk_LIMIT_32( r_Q10, -(31 << 10), 30 << 10 ); 465 466 /* Find two quantization level candidates and measure their rate-distortion */ 467 q1_Q10 = silk_SUB32( r_Q10, offset_Q10 ); 468 q1_Q0 = silk_RSHIFT( q1_Q10, 10 ); 469 if( q1_Q0 > 0 ) { 470 q1_Q10 = silk_SUB32( silk_LSHIFT( q1_Q0, 10 ), QUANT_LEVEL_ADJUST_Q10 ); 471 q1_Q10 = silk_ADD32( q1_Q10, offset_Q10 ); 472 q2_Q10 = silk_ADD32( q1_Q10, 1024 ); 473 rd1_Q10 = silk_SMULBB( q1_Q10, Lambda_Q10 ); 474 rd2_Q10 = silk_SMULBB( q2_Q10, Lambda_Q10 ); 475 } else if( q1_Q0 == 0 ) { 476 q1_Q10 = offset_Q10; 477 q2_Q10 = silk_ADD32( q1_Q10, 1024 - QUANT_LEVEL_ADJUST_Q10 ); 478 rd1_Q10 = silk_SMULBB( q1_Q10, Lambda_Q10 ); 479 rd2_Q10 = silk_SMULBB( q2_Q10, Lambda_Q10 ); 480 } else if( q1_Q0 == -1 ) { 481 q2_Q10 = offset_Q10; 482 q1_Q10 = silk_SUB32( q2_Q10, 1024 - QUANT_LEVEL_ADJUST_Q10 ); 483 rd1_Q10 = silk_SMULBB( -q1_Q10, Lambda_Q10 ); 484 rd2_Q10 = silk_SMULBB( q2_Q10, Lambda_Q10 ); 485 } else { /* q1_Q0 < -1 */ 486 q1_Q10 = silk_ADD32( silk_LSHIFT( q1_Q0, 10 ), QUANT_LEVEL_ADJUST_Q10 ); 487 q1_Q10 = silk_ADD32( q1_Q10, offset_Q10 ); 488 q2_Q10 = silk_ADD32( q1_Q10, 1024 ); 489 rd1_Q10 = silk_SMULBB( -q1_Q10, Lambda_Q10 ); 490 rd2_Q10 = silk_SMULBB( -q2_Q10, Lambda_Q10 ); 491 } 492 rr_Q10 = silk_SUB32( r_Q10, q1_Q10 ); 493 rd1_Q10 = silk_RSHIFT( silk_SMLABB( rd1_Q10, rr_Q10, rr_Q10 ), 10 ); 494 rr_Q10 = silk_SUB32( r_Q10, q2_Q10 ); 495 rd2_Q10 = silk_RSHIFT( silk_SMLABB( rd2_Q10, rr_Q10, rr_Q10 ), 10 ); 496 497 if( rd1_Q10 < rd2_Q10 ) { 498 psSS[ 0 ].RD_Q10 = silk_ADD32( psDD->RD_Q10, rd1_Q10 ); 499 psSS[ 1 ].RD_Q10 = silk_ADD32( psDD->RD_Q10, rd2_Q10 ); 500 psSS[ 0 ].Q_Q10 = q1_Q10; 501 psSS[ 1 ].Q_Q10 = q2_Q10; 502 } else { 503 psSS[ 0 ].RD_Q10 = silk_ADD32( psDD->RD_Q10, rd2_Q10 ); 504 psSS[ 1 ].RD_Q10 = silk_ADD32( psDD->RD_Q10, rd1_Q10 ); 505 psSS[ 0 ].Q_Q10 = q2_Q10; 506 psSS[ 1 ].Q_Q10 = q1_Q10; 507 } 508 509 /* Update states for best quantization */ 510 511 /* Quantized excitation */ 512 exc_Q14 = silk_LSHIFT32( psSS[ 0 ].Q_Q10, 4 ); 513 if ( psDD->Seed < 0 ) { 514 exc_Q14 = -exc_Q14; 515 } 516 517 /* Add predictions */ 518 LPC_exc_Q14 = silk_ADD32( exc_Q14, LTP_pred_Q14 ); 519 xq_Q14 = silk_ADD32( LPC_exc_Q14, LPC_pred_Q14 ); 520 521 /* Update states */ 522 sLF_AR_shp_Q14 = silk_SUB32( xq_Q14, n_AR_Q14 ); 523 psSS[ 0 ].sLTP_shp_Q14 = silk_SUB32( sLF_AR_shp_Q14, n_LF_Q14 ); 524 psSS[ 0 ].LF_AR_Q14 = sLF_AR_shp_Q14; 525 psSS[ 0 ].LPC_exc_Q14 = LPC_exc_Q14; 526 psSS[ 0 ].xq_Q14 = xq_Q14; 527 528 /* Update states for second best quantization */ 529 530 /* Quantized excitation */ 531 exc_Q14 = silk_LSHIFT32( psSS[ 1 ].Q_Q10, 4 ); 532 if ( psDD->Seed < 0 ) { 533 exc_Q14 = -exc_Q14; 534 } 535 536 537 /* Add predictions */ 538 LPC_exc_Q14 = silk_ADD32( exc_Q14, LTP_pred_Q14 ); 539 xq_Q14 = silk_ADD32( LPC_exc_Q14, LPC_pred_Q14 ); 540 541 /* Update states */ 542 sLF_AR_shp_Q14 = silk_SUB32( xq_Q14, n_AR_Q14 ); 543 psSS[ 1 ].sLTP_shp_Q14 = silk_SUB32( sLF_AR_shp_Q14, n_LF_Q14 ); 544 psSS[ 1 ].LF_AR_Q14 = sLF_AR_shp_Q14; 545 psSS[ 1 ].LPC_exc_Q14 = LPC_exc_Q14; 546 psSS[ 1 ].xq_Q14 = xq_Q14; 547 } 548 549 *smpl_buf_idx = ( *smpl_buf_idx - 1 ) & DECISION_DELAY_MASK; /* Index to newest samples */ 550 last_smple_idx = ( *smpl_buf_idx + decisionDelay ) & DECISION_DELAY_MASK; /* Index to decisionDelay old samples */ 551 552 /* Find winner */ 553 RDmin_Q10 = psSampleState[ 0 ][ 0 ].RD_Q10; 554 Winner_ind = 0; 555 for( k = 1; k < nStatesDelayedDecision; k++ ) { 556 if( psSampleState[ k ][ 0 ].RD_Q10 < RDmin_Q10 ) { 557 RDmin_Q10 = psSampleState[ k ][ 0 ].RD_Q10; 558 Winner_ind = k; 559 } 560 } 561 562 /* Increase RD values of expired states */ 563 Winner_rand_state = psDelDec[ Winner_ind ].RandState[ last_smple_idx ]; 564 for( k = 0; k < nStatesDelayedDecision; k++ ) { 565 if( psDelDec[ k ].RandState[ last_smple_idx ] != Winner_rand_state ) { 566 psSampleState[ k ][ 0 ].RD_Q10 = silk_ADD32( psSampleState[ k ][ 0 ].RD_Q10, silk_int32_MAX >> 4 ); 567 psSampleState[ k ][ 1 ].RD_Q10 = silk_ADD32( psSampleState[ k ][ 1 ].RD_Q10, silk_int32_MAX >> 4 ); 568 silk_assert( psSampleState[ k ][ 0 ].RD_Q10 >= 0 ); 569 } 570 } 571 572 /* Find worst in first set and best in second set */ 573 RDmax_Q10 = psSampleState[ 0 ][ 0 ].RD_Q10; 574 RDmin_Q10 = psSampleState[ 0 ][ 1 ].RD_Q10; 575 RDmax_ind = 0; 576 RDmin_ind = 0; 577 for( k = 1; k < nStatesDelayedDecision; k++ ) { 578 /* find worst in first set */ 579 if( psSampleState[ k ][ 0 ].RD_Q10 > RDmax_Q10 ) { 580 RDmax_Q10 = psSampleState[ k ][ 0 ].RD_Q10; 581 RDmax_ind = k; 582 } 583 /* find best in second set */ 584 if( psSampleState[ k ][ 1 ].RD_Q10 < RDmin_Q10 ) { 585 RDmin_Q10 = psSampleState[ k ][ 1 ].RD_Q10; 586 RDmin_ind = k; 587 } 588 } 589 590 /* Replace a state if best from second set outperforms worst in first set */ 591 if( RDmin_Q10 < RDmax_Q10 ) { 592 silk_memcpy( ( (opus_int32 *)&psDelDec[ RDmax_ind ] ) + i, 593 ( (opus_int32 *)&psDelDec[ RDmin_ind ] ) + i, sizeof( NSQ_del_dec_struct ) - i * sizeof( opus_int32) ); 594 silk_memcpy( &psSampleState[ RDmax_ind ][ 0 ], &psSampleState[ RDmin_ind ][ 1 ], sizeof( NSQ_sample_struct ) ); 595 } 596 597 /* Write samples from winner to output and long-term filter states */ 598 psDD = &psDelDec[ Winner_ind ]; 599 if( subfr > 0 || i >= decisionDelay ) { 600 pulses[ i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDD->Q_Q10[ last_smple_idx ], 10 ); 601 xq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND( 602 silk_SMULWW( psDD->Xq_Q14[ last_smple_idx ], delayedGain_Q10[ last_smple_idx ] ), 8 ) ); 603 NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - decisionDelay ] = psDD->Shape_Q14[ last_smple_idx ]; 604 sLTP_Q15[ NSQ->sLTP_buf_idx - decisionDelay ] = psDD->Pred_Q15[ last_smple_idx ]; 605 } 606 NSQ->sLTP_shp_buf_idx++; 607 NSQ->sLTP_buf_idx++; 608 609 /* Update states */ 610 for( k = 0; k < nStatesDelayedDecision; k++ ) { 611 psDD = &psDelDec[ k ]; 612 psSS = &psSampleState[ k ][ 0 ]; 613 psDD->LF_AR_Q14 = psSS->LF_AR_Q14; 614 psDD->sLPC_Q14[ NSQ_LPC_BUF_LENGTH + i ] = psSS->xq_Q14; 615 psDD->Xq_Q14[ *smpl_buf_idx ] = psSS->xq_Q14; 616 psDD->Q_Q10[ *smpl_buf_idx ] = psSS->Q_Q10; 617 psDD->Pred_Q15[ *smpl_buf_idx ] = silk_LSHIFT32( psSS->LPC_exc_Q14, 1 ); 618 psDD->Shape_Q14[ *smpl_buf_idx ] = psSS->sLTP_shp_Q14; 619 psDD->Seed = silk_ADD32_ovflw( psDD->Seed, silk_RSHIFT_ROUND( psSS->Q_Q10, 10 ) ); 620 psDD->RandState[ *smpl_buf_idx ] = psDD->Seed; 621 psDD->RD_Q10 = psSS->RD_Q10; 622 } 623 delayedGain_Q10[ *smpl_buf_idx ] = Gain_Q10; 624 } 625 /* Update LPC states */ 626 for( k = 0; k < nStatesDelayedDecision; k++ ) { 627 psDD = &psDelDec[ k ]; 628 silk_memcpy( psDD->sLPC_Q14, &psDD->sLPC_Q14[ length ], NSQ_LPC_BUF_LENGTH * sizeof( opus_int32 ) ); 629 } 630 RESTORE_STACK; 631 } 632 633 static inline void silk_nsq_del_dec_scale_states( 634 const silk_encoder_state *psEncC, /* I Encoder State */ 635 silk_nsq_state *NSQ, /* I/O NSQ state */ 636 NSQ_del_dec_struct psDelDec[], /* I/O Delayed decision states */ 637 const opus_int32 x_Q3[], /* I Input in Q3 */ 638 opus_int32 x_sc_Q10[], /* O Input scaled with 1/Gain in Q10 */ 639 const opus_int16 sLTP[], /* I Re-whitened LTP state in Q0 */ 640 opus_int32 sLTP_Q15[], /* O LTP state matching scaled input */ 641 opus_int subfr, /* I Subframe number */ 642 opus_int nStatesDelayedDecision, /* I Number of del dec states */ 643 const opus_int LTP_scale_Q14, /* I LTP state scaling */ 644 const opus_int32 Gains_Q16[ MAX_NB_SUBFR ], /* I */ 645 const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lag */ 646 const opus_int signal_type, /* I Signal type */ 647 const opus_int decisionDelay /* I Decision delay */ 648 ) 649 { 650 opus_int i, k, lag; 651 opus_int32 gain_adj_Q16, inv_gain_Q31, inv_gain_Q23; 652 NSQ_del_dec_struct *psDD; 653 654 lag = pitchL[ subfr ]; 655 inv_gain_Q31 = silk_INVERSE32_varQ( silk_max( Gains_Q16[ subfr ], 1 ), 47 ); 656 silk_assert( inv_gain_Q31 != 0 ); 657 658 /* Calculate gain adjustment factor */ 659 if( Gains_Q16[ subfr ] != NSQ->prev_gain_Q16 ) { 660 gain_adj_Q16 = silk_DIV32_varQ( NSQ->prev_gain_Q16, Gains_Q16[ subfr ], 16 ); 661 } else { 662 gain_adj_Q16 = (opus_int32)1 << 16; 663 } 664 665 /* Scale input */ 666 inv_gain_Q23 = silk_RSHIFT_ROUND( inv_gain_Q31, 8 ); 667 for( i = 0; i < psEncC->subfr_length; i++ ) { 668 x_sc_Q10[ i ] = silk_SMULWW( x_Q3[ i ], inv_gain_Q23 ); 669 } 670 671 /* Save inverse gain */ 672 NSQ->prev_gain_Q16 = Gains_Q16[ subfr ]; 673 674 /* After rewhitening the LTP state is un-scaled, so scale with inv_gain_Q16 */ 675 if( NSQ->rewhite_flag ) { 676 if( subfr == 0 ) { 677 /* Do LTP downscaling */ 678 inv_gain_Q31 = silk_LSHIFT( silk_SMULWB( inv_gain_Q31, LTP_scale_Q14 ), 2 ); 679 } 680 for( i = NSQ->sLTP_buf_idx - lag - LTP_ORDER / 2; i < NSQ->sLTP_buf_idx; i++ ) { 681 silk_assert( i < MAX_FRAME_LENGTH ); 682 sLTP_Q15[ i ] = silk_SMULWB( inv_gain_Q31, sLTP[ i ] ); 683 } 684 } 685 686 /* Adjust for changing gain */ 687 if( gain_adj_Q16 != (opus_int32)1 << 16 ) { 688 /* Scale long-term shaping state */ 689 for( i = NSQ->sLTP_shp_buf_idx - psEncC->ltp_mem_length; i < NSQ->sLTP_shp_buf_idx; i++ ) { 690 NSQ->sLTP_shp_Q14[ i ] = silk_SMULWW( gain_adj_Q16, NSQ->sLTP_shp_Q14[ i ] ); 691 } 692 693 /* Scale long-term prediction state */ 694 if( signal_type == TYPE_VOICED && NSQ->rewhite_flag == 0 ) { 695 for( i = NSQ->sLTP_buf_idx - lag - LTP_ORDER / 2; i < NSQ->sLTP_buf_idx - decisionDelay; i++ ) { 696 sLTP_Q15[ i ] = silk_SMULWW( gain_adj_Q16, sLTP_Q15[ i ] ); 697 } 698 } 699 700 for( k = 0; k < nStatesDelayedDecision; k++ ) { 701 psDD = &psDelDec[ k ]; 702 703 /* Scale scalar states */ 704 psDD->LF_AR_Q14 = silk_SMULWW( gain_adj_Q16, psDD->LF_AR_Q14 ); 705 706 /* Scale short-term prediction and shaping states */ 707 for( i = 0; i < NSQ_LPC_BUF_LENGTH; i++ ) { 708 psDD->sLPC_Q14[ i ] = silk_SMULWW( gain_adj_Q16, psDD->sLPC_Q14[ i ] ); 709 } 710 for( i = 0; i < MAX_SHAPE_LPC_ORDER; i++ ) { 711 psDD->sAR2_Q14[ i ] = silk_SMULWW( gain_adj_Q16, psDD->sAR2_Q14[ i ] ); 712 } 713 for( i = 0; i < DECISION_DELAY; i++ ) { 714 psDD->Pred_Q15[ i ] = silk_SMULWW( gain_adj_Q16, psDD->Pred_Q15[ i ] ); 715 psDD->Shape_Q14[ i ] = silk_SMULWW( gain_adj_Q16, psDD->Shape_Q14[ i ] ); 716 } 717 } 718 } 719 } 720
