NSQ_del_dec.c
1
2
3
4
5
6
7
8
9
10
/***********************************************************************
Copyright (c) 2006-2011, Skype Limited. All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions
are met:
- Redistributions of source code must retain the above copyright notice,
this list of conditions and the following disclaimer.
- Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in the
documentation and/or other materials provided with the distribution.
11
- Neither the name of Internet Society, IETF or IETF Trust, nor the
12
13
14
names of specific contributors, may be used to endorse or promote
products derived from this software without specific prior written
permission.
15
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
POSSIBILITY OF SUCH DAMAGE.
***********************************************************************/
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif
#include "main.h"
33
34
35
#include "stack_alloc.h"
#include "NSQ.h"
36
37
38
39
40
41
42
43
44
45
typedef struct {
opus_int32 sLPC_Q14[ MAX_SUB_FRAME_LENGTH + NSQ_LPC_BUF_LENGTH ];
opus_int32 RandState[ DECISION_DELAY ];
opus_int32 Q_Q10[ DECISION_DELAY ];
opus_int32 Xq_Q14[ DECISION_DELAY ];
opus_int32 Pred_Q15[ DECISION_DELAY ];
opus_int32 Shape_Q14[ DECISION_DELAY ];
opus_int32 sAR2_Q14[ MAX_SHAPE_LPC_ORDER ];
opus_int32 LF_AR_Q14;
46
opus_int32 Diff_Q14;
47
48
49
50
51
52
53
54
55
56
opus_int32 Seed;
opus_int32 SeedInit;
opus_int32 RD_Q10;
} NSQ_del_dec_struct;
typedef struct {
opus_int32 Q_Q10;
opus_int32 RD_Q10;
opus_int32 xq_Q14;
opus_int32 LF_AR_Q14;
57
opus_int32 Diff_Q14;
58
59
60
61
opus_int32 sLTP_shp_Q14;
opus_int32 LPC_exc_Q14;
} NSQ_sample_struct;
62
63
64
65
66
67
typedef NSQ_sample_struct NSQ_sample_pair[ 2 ];
#if defined(MIPSr1_ASM)
#include "mips/NSQ_del_dec_mipsr1.h"
#endif
static OPUS_INLINE void silk_nsq_del_dec_scale_states(
68
69
70
const silk_encoder_state *psEncC, /* I Encoder State */
silk_nsq_state *NSQ, /* I/O NSQ state */
NSQ_del_dec_struct psDelDec[], /* I/O Delayed decision states */
71
const opus_int16 x16[], /* I Input */
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
opus_int32 x_sc_Q10[], /* O Input scaled with 1/Gain in Q10 */
const opus_int16 sLTP[], /* I Re-whitened LTP state in Q0 */
opus_int32 sLTP_Q15[], /* O LTP state matching scaled input */
opus_int subfr, /* I Subframe number */
opus_int nStatesDelayedDecision, /* I Number of del dec states */
const opus_int LTP_scale_Q14, /* I LTP state scaling */
const opus_int32 Gains_Q16[ MAX_NB_SUBFR ], /* I */
const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lag */
const opus_int signal_type, /* I Signal type */
const opus_int decisionDelay /* I Decision delay */
);
/******************************************/
/* Noise shape quantizer for one subframe */
/******************************************/
87
static OPUS_INLINE void silk_noise_shape_quantizer_del_dec(
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
silk_nsq_state *NSQ, /* I/O NSQ state */
NSQ_del_dec_struct psDelDec[], /* I/O Delayed decision states */
opus_int signalType, /* I Signal type */
const opus_int32 x_Q10[], /* I */
opus_int8 pulses[], /* O */
opus_int16 xq[], /* O */
opus_int32 sLTP_Q15[], /* I/O LTP filter state */
opus_int32 delayedGain_Q10[], /* I/O Gain delay buffer */
const opus_int16 a_Q12[], /* I Short term prediction coefs */
const opus_int16 b_Q14[], /* I Long term prediction coefs */
const opus_int16 AR_shp_Q13[], /* I Noise shaping coefs */
opus_int lag, /* I Pitch lag */
opus_int32 HarmShapeFIRPacked_Q14, /* I */
opus_int Tilt_Q14, /* I Spectral tilt */
opus_int32 LF_shp_Q14, /* I */
opus_int32 Gain_Q16, /* I */
opus_int Lambda_Q10, /* I */
opus_int offset_Q10, /* I */
opus_int length, /* I Input length */
opus_int subfr, /* I Subframe number */
opus_int shapingLPCOrder, /* I Shaping LPC filter order */
opus_int predictLPCOrder, /* I Prediction filter order */
opus_int warping_Q16, /* I */
opus_int nStatesDelayedDecision, /* I Number of states in decision tree */
112
113
114
opus_int *smpl_buf_idx, /* I/O Index to newest samples in buffers */
opus_int decisionDelay, /* I */
int arch /* I */
117
118
void silk_NSQ_del_dec_c(
const silk_encoder_state *psEncC, /* I Encoder State */
119
120
silk_nsq_state *NSQ, /* I/O NSQ state */
SideInfoIndices *psIndices, /* I/O Quantization Indices */
121
const opus_int16 x16[], /* I Input */
122
123
124
opus_int8 pulses[], /* O Quantized pulse signal */
const opus_int16 PredCoef_Q12[ 2 * MAX_LPC_ORDER ], /* I Short term prediction coefs */
const opus_int16 LTPCoef_Q14[ LTP_ORDER * MAX_NB_SUBFR ], /* I Long term prediction coefs */
125
const opus_int16 AR_Q13[ MAX_NB_SUBFR * MAX_SHAPE_LPC_ORDER ], /* I Noise shaping coefs */
126
127
128
129
130
131
132
133
134
135
136
const opus_int HarmShapeGain_Q14[ MAX_NB_SUBFR ], /* I Long term shaping coefs */
const opus_int Tilt_Q14[ MAX_NB_SUBFR ], /* I Spectral tilt */
const opus_int32 LF_shp_Q14[ MAX_NB_SUBFR ], /* I Low frequency shaping coefs */
const opus_int32 Gains_Q16[ MAX_NB_SUBFR ], /* I Quantization step sizes */
const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lags */
const opus_int Lambda_Q10, /* I Rate/distortion tradeoff */
const opus_int LTP_scale_Q14 /* I LTP state scaling */
)
{
opus_int i, k, lag, start_idx, LSF_interpolation_flag, Winner_ind, subfr;
opus_int last_smple_idx, smpl_buf_idx, decisionDelay;
137
const opus_int16 *A_Q12, *B_Q14, *AR_shp_Q13;
138
opus_int16 *pxq;
139
140
VARDECL( opus_int32, sLTP_Q15 );
VARDECL( opus_int16, sLTP );
141
142
143
opus_int32 HarmShapeFIRPacked_Q14;
opus_int offset_Q10;
opus_int32 RDmin_Q10, Gain_Q10;
144
145
146
VARDECL( opus_int32, x_sc_Q10 );
VARDECL( opus_int32, delayedGain_Q10 );
VARDECL( NSQ_del_dec_struct, psDelDec );
147
NSQ_del_dec_struct *psDD;
148
SAVE_STACK;
149
150
151
152
153
154
155
/* Set unvoiced lag to the previous one, overwrite later for voiced */
lag = NSQ->lagPrev;
silk_assert( NSQ->prev_gain_Q16 != 0 );
/* Initialize delayed decision states */
156
ALLOC( psDelDec, psEncC->nStatesDelayedDecision, NSQ_del_dec_struct );
157
158
159
160
161
162
163
silk_memset( psDelDec, 0, psEncC->nStatesDelayedDecision * sizeof( NSQ_del_dec_struct ) );
for( k = 0; k < psEncC->nStatesDelayedDecision; k++ ) {
psDD = &psDelDec[ k ];
psDD->Seed = ( k + psIndices->Seed ) & 3;
psDD->SeedInit = psDD->Seed;
psDD->RD_Q10 = 0;
psDD->LF_AR_Q14 = NSQ->sLF_AR_shp_Q14;
164
psDD->Diff_Q14 = NSQ->sDiff_shp_Q14;
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
psDD->Shape_Q14[ 0 ] = NSQ->sLTP_shp_Q14[ psEncC->ltp_mem_length - 1 ];
silk_memcpy( psDD->sLPC_Q14, NSQ->sLPC_Q14, NSQ_LPC_BUF_LENGTH * sizeof( opus_int32 ) );
silk_memcpy( psDD->sAR2_Q14, NSQ->sAR2_Q14, sizeof( NSQ->sAR2_Q14 ) );
}
offset_Q10 = silk_Quantization_Offsets_Q10[ psIndices->signalType >> 1 ][ psIndices->quantOffsetType ];
smpl_buf_idx = 0; /* index of oldest samples */
decisionDelay = silk_min_int( DECISION_DELAY, psEncC->subfr_length );
/* For voiced frames limit the decision delay to lower than the pitch lag */
if( psIndices->signalType == TYPE_VOICED ) {
for( k = 0; k < psEncC->nb_subfr; k++ ) {
decisionDelay = silk_min_int( decisionDelay, pitchL[ k ] - LTP_ORDER / 2 - 1 );
}
} else {
if( lag > 0 ) {
decisionDelay = silk_min_int( decisionDelay, lag - LTP_ORDER / 2 - 1 );
}
}
if( psIndices->NLSFInterpCoef_Q2 == 4 ) {
LSF_interpolation_flag = 0;
} else {
LSF_interpolation_flag = 1;
}
192
193
194
195
ALLOC( sLTP_Q15, psEncC->ltp_mem_length + psEncC->frame_length, opus_int32 );
ALLOC( sLTP, psEncC->ltp_mem_length + psEncC->frame_length, opus_int16 );
ALLOC( x_sc_Q10, psEncC->subfr_length, opus_int32 );
ALLOC( delayedGain_Q10, DECISION_DELAY, opus_int32 );
196
197
198
199
200
201
202
203
/* Set up pointers to start of sub frame */
pxq = &NSQ->xq[ psEncC->ltp_mem_length ];
NSQ->sLTP_shp_buf_idx = psEncC->ltp_mem_length;
NSQ->sLTP_buf_idx = psEncC->ltp_mem_length;
subfr = 0;
for( k = 0; k < psEncC->nb_subfr; k++ ) {
A_Q12 = &PredCoef_Q12[ ( ( k >> 1 ) | ( 1 - LSF_interpolation_flag ) ) * MAX_LPC_ORDER ];
B_Q14 = <PCoef_Q14[ k * LTP_ORDER ];
204
AR_shp_Q13 = &AR_Q13[ k * MAX_SHAPE_LPC_ORDER ];
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
/* Noise shape parameters */
silk_assert( HarmShapeGain_Q14[ k ] >= 0 );
HarmShapeFIRPacked_Q14 = silk_RSHIFT( HarmShapeGain_Q14[ k ], 2 );
HarmShapeFIRPacked_Q14 |= silk_LSHIFT( (opus_int32)silk_RSHIFT( HarmShapeGain_Q14[ k ], 1 ), 16 );
NSQ->rewhite_flag = 0;
if( psIndices->signalType == TYPE_VOICED ) {
/* Voiced */
lag = pitchL[ k ];
/* Re-whitening */
if( ( k & ( 3 - silk_LSHIFT( LSF_interpolation_flag, 1 ) ) ) == 0 ) {
if( k == 2 ) {
/* RESET DELAYED DECISIONS */
/* Find winner */
RDmin_Q10 = psDelDec[ 0 ].RD_Q10;
Winner_ind = 0;
for( i = 1; i < psEncC->nStatesDelayedDecision; i++ ) {
if( psDelDec[ i ].RD_Q10 < RDmin_Q10 ) {
RDmin_Q10 = psDelDec[ i ].RD_Q10;
Winner_ind = i;
}
}
for( i = 0; i < psEncC->nStatesDelayedDecision; i++ ) {
if( i != Winner_ind ) {
psDelDec[ i ].RD_Q10 += ( silk_int32_MAX >> 4 );
silk_assert( psDelDec[ i ].RD_Q10 >= 0 );
}
}
/* Copy final part of signals from winner state to output and long-term filter states */
psDD = &psDelDec[ Winner_ind ];
last_smple_idx = smpl_buf_idx + decisionDelay;
for( i = 0; i < decisionDelay; i++ ) {
240
241
last_smple_idx = ( last_smple_idx - 1 ) % DECISION_DELAY;
if( last_smple_idx < 0 ) last_smple_idx += DECISION_DELAY;
242
243
244
245
246
247
248
249
250
251
252
pulses[ i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDD->Q_Q10[ last_smple_idx ], 10 );
pxq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND(
silk_SMULWW( psDD->Xq_Q14[ last_smple_idx ], Gains_Q16[ 1 ] ), 14 ) );
NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - decisionDelay + i ] = psDD->Shape_Q14[ last_smple_idx ];
}
subfr = 0;
}
/* Rewhiten with new A coefs */
start_idx = psEncC->ltp_mem_length - lag - psEncC->predictLPCOrder - LTP_ORDER / 2;
253
celt_assert( start_idx > 0 );
254
255
silk_LPC_analysis_filter( &sLTP[ start_idx ], &NSQ->xq[ start_idx + k * psEncC->subfr_length ],
256
A_Q12, psEncC->ltp_mem_length - start_idx, psEncC->predictLPCOrder, psEncC->arch );
257
258
259
260
261
262
NSQ->sLTP_buf_idx = psEncC->ltp_mem_length;
NSQ->rewhite_flag = 1;
}
}
263
silk_nsq_del_dec_scale_states( psEncC, NSQ, psDelDec, x16, x_sc_Q10, sLTP, sLTP_Q15, k,
264
265
266
267
268
psEncC->nStatesDelayedDecision, LTP_scale_Q14, Gains_Q16, pitchL, psIndices->signalType, decisionDelay );
silk_noise_shape_quantizer_del_dec( NSQ, psDelDec, psIndices->signalType, x_sc_Q10, pulses, pxq, sLTP_Q15,
delayedGain_Q10, A_Q12, B_Q14, AR_shp_Q13, lag, HarmShapeFIRPacked_Q14, Tilt_Q14[ k ], LF_shp_Q14[ k ],
Gains_Q16[ k ], Lambda_Q10, offset_Q10, psEncC->subfr_length, subfr++, psEncC->shapingLPCOrder,
269
psEncC->predictLPCOrder, psEncC->warping_Q16, psEncC->nStatesDelayedDecision, &smpl_buf_idx, decisionDelay, psEncC->arch );
271
x16 += psEncC->subfr_length;
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
pulses += psEncC->subfr_length;
pxq += psEncC->subfr_length;
}
/* Find winner */
RDmin_Q10 = psDelDec[ 0 ].RD_Q10;
Winner_ind = 0;
for( k = 1; k < psEncC->nStatesDelayedDecision; k++ ) {
if( psDelDec[ k ].RD_Q10 < RDmin_Q10 ) {
RDmin_Q10 = psDelDec[ k ].RD_Q10;
Winner_ind = k;
}
}
/* Copy final part of signals from winner state to output and long-term filter states */
psDD = &psDelDec[ Winner_ind ];
psIndices->Seed = psDD->SeedInit;
last_smple_idx = smpl_buf_idx + decisionDelay;
Gain_Q10 = silk_RSHIFT32( Gains_Q16[ psEncC->nb_subfr - 1 ], 6 );
for( i = 0; i < decisionDelay; i++ ) {
292
293
294
last_smple_idx = ( last_smple_idx - 1 ) % DECISION_DELAY;
if( last_smple_idx < 0 ) last_smple_idx += DECISION_DELAY;
295
296
297
298
299
300
301
302
303
304
pulses[ i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDD->Q_Q10[ last_smple_idx ], 10 );
pxq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND(
silk_SMULWW( psDD->Xq_Q14[ last_smple_idx ], Gain_Q10 ), 8 ) );
NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - decisionDelay + i ] = psDD->Shape_Q14[ last_smple_idx ];
}
silk_memcpy( NSQ->sLPC_Q14, &psDD->sLPC_Q14[ psEncC->subfr_length ], NSQ_LPC_BUF_LENGTH * sizeof( opus_int32 ) );
silk_memcpy( NSQ->sAR2_Q14, psDD->sAR2_Q14, sizeof( psDD->sAR2_Q14 ) );
/* Update states */
NSQ->sLF_AR_shp_Q14 = psDD->LF_AR_Q14;
305
NSQ->sDiff_shp_Q14 = psDD->Diff_Q14;
306
307
308
309
310
NSQ->lagPrev = pitchL[ psEncC->nb_subfr - 1 ];
/* Save quantized speech signal */
silk_memmove( NSQ->xq, &NSQ->xq[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int16 ) );
silk_memmove( NSQ->sLTP_shp_Q14, &NSQ->sLTP_shp_Q14[ psEncC->frame_length ], psEncC->ltp_mem_length * sizeof( opus_int32 ) );
311
RESTORE_STACK;
312
313
314
315
316
}
/******************************************/
/* Noise shape quantizer for one subframe */
/******************************************/
317
318
#ifndef OVERRIDE_silk_noise_shape_quantizer_del_dec
static OPUS_INLINE void silk_noise_shape_quantizer_del_dec(
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
silk_nsq_state *NSQ, /* I/O NSQ state */
NSQ_del_dec_struct psDelDec[], /* I/O Delayed decision states */
opus_int signalType, /* I Signal type */
const opus_int32 x_Q10[], /* I */
opus_int8 pulses[], /* O */
opus_int16 xq[], /* O */
opus_int32 sLTP_Q15[], /* I/O LTP filter state */
opus_int32 delayedGain_Q10[], /* I/O Gain delay buffer */
const opus_int16 a_Q12[], /* I Short term prediction coefs */
const opus_int16 b_Q14[], /* I Long term prediction coefs */
const opus_int16 AR_shp_Q13[], /* I Noise shaping coefs */
opus_int lag, /* I Pitch lag */
opus_int32 HarmShapeFIRPacked_Q14, /* I */
opus_int Tilt_Q14, /* I Spectral tilt */
opus_int32 LF_shp_Q14, /* I */
opus_int32 Gain_Q16, /* I */
opus_int Lambda_Q10, /* I */
opus_int offset_Q10, /* I */
opus_int length, /* I Input length */
opus_int subfr, /* I Subframe number */
opus_int shapingLPCOrder, /* I Shaping LPC filter order */
opus_int predictLPCOrder, /* I Prediction filter order */
opus_int warping_Q16, /* I */
opus_int nStatesDelayedDecision, /* I Number of states in decision tree */
343
344
345
opus_int *smpl_buf_idx, /* I/O Index to newest samples in buffers */
opus_int decisionDelay, /* I */
int arch /* I */
346
347
348
349
350
351
352
353
354
)
{
opus_int i, j, k, Winner_ind, RDmin_ind, RDmax_ind, last_smple_idx;
opus_int32 Winner_rand_state;
opus_int32 LTP_pred_Q14, LPC_pred_Q14, n_AR_Q14, n_LTP_Q14;
opus_int32 n_LF_Q14, r_Q10, rr_Q10, rd1_Q10, rd2_Q10, RDmin_Q10, RDmax_Q10;
opus_int32 q1_Q0, q1_Q10, q2_Q10, exc_Q14, LPC_exc_Q14, xq_Q14, Gain_Q10;
opus_int32 tmp1, tmp2, sLF_AR_shp_Q14;
opus_int32 *pred_lag_ptr, *shp_lag_ptr, *psLPC_Q14;
355
356
357
358
359
#ifdef silk_short_prediction_create_arch_coef
opus_int32 a_Q12_arch[MAX_LPC_ORDER];
#endif
VARDECL( NSQ_sample_pair, psSampleState );
360
361
NSQ_del_dec_struct *psDD;
NSQ_sample_struct *psSS;
362
SAVE_STACK;
364
365
celt_assert( nStatesDelayedDecision > 0 );
ALLOC( psSampleState, nStatesDelayedDecision, NSQ_sample_pair );
366
367
368
369
370
shp_lag_ptr = &NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - lag + HARM_SHAPE_FIR_TAPS / 2 ];
pred_lag_ptr = &sLTP_Q15[ NSQ->sLTP_buf_idx - lag + LTP_ORDER / 2 ];
Gain_Q10 = silk_RSHIFT( Gain_Q16, 6 );
371
372
373
374
#ifdef silk_short_prediction_create_arch_coef
silk_short_prediction_create_arch_coef(a_Q12_arch, a_Q12, predictLPCOrder);
#endif
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
for( i = 0; i < length; i++ ) {
/* Perform common calculations used in all states */
/* Long-term prediction */
if( signalType == TYPE_VOICED ) {
/* Unrolled loop */
/* Avoids introducing a bias because silk_SMLAWB() always rounds to -inf */
LTP_pred_Q14 = 2;
LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ 0 ], b_Q14[ 0 ] );
LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -1 ], b_Q14[ 1 ] );
LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -2 ], b_Q14[ 2 ] );
LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -3 ], b_Q14[ 3 ] );
LTP_pred_Q14 = silk_SMLAWB( LTP_pred_Q14, pred_lag_ptr[ -4 ], b_Q14[ 4 ] );
LTP_pred_Q14 = silk_LSHIFT( LTP_pred_Q14, 1 ); /* Q13 -> Q14 */
pred_lag_ptr++;
} else {
LTP_pred_Q14 = 0;
}
/* Long-term shaping */
if( lag > 0 ) {
/* Symmetric, packed FIR coefficients */
n_LTP_Q14 = silk_SMULWB( silk_ADD32( shp_lag_ptr[ 0 ], shp_lag_ptr[ -2 ] ), HarmShapeFIRPacked_Q14 );
n_LTP_Q14 = silk_SMLAWT( n_LTP_Q14, shp_lag_ptr[ -1 ], HarmShapeFIRPacked_Q14 );
n_LTP_Q14 = silk_SUB_LSHIFT32( LTP_pred_Q14, n_LTP_Q14, 2 ); /* Q12 -> Q14 */
shp_lag_ptr++;
} else {
n_LTP_Q14 = 0;
}
for( k = 0; k < nStatesDelayedDecision; k++ ) {
/* Delayed decision state */
psDD = &psDelDec[ k ];
/* Sample state */
psSS = psSampleState[ k ];
/* Generate dither */
psDD->Seed = silk_RAND( psDD->Seed );
/* Pointer used in short term prediction and shaping */
psLPC_Q14 = &psDD->sLPC_Q14[ NSQ_LPC_BUF_LENGTH - 1 + i ];
/* Short-term prediction */
418
LPC_pred_Q14 = silk_noise_shape_quantizer_short_prediction(psLPC_Q14, a_Q12, a_Q12_arch, predictLPCOrder, arch);
419
420
421
LPC_pred_Q14 = silk_LSHIFT( LPC_pred_Q14, 4 ); /* Q10 -> Q14 */
/* Noise shape feedback */
422
celt_assert( ( shapingLPCOrder & 1 ) == 0 ); /* check that order is even */
423
/* Output of lowpass section */
424
tmp2 = silk_SMLAWB( psDD->Diff_Q14, psDD->sAR2_Q14[ 0 ], warping_Q16 );
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
/* Output of allpass section */
tmp1 = silk_SMLAWB( psDD->sAR2_Q14[ 0 ], psDD->sAR2_Q14[ 1 ] - tmp2, warping_Q16 );
psDD->sAR2_Q14[ 0 ] = tmp2;
n_AR_Q14 = silk_RSHIFT( shapingLPCOrder, 1 );
n_AR_Q14 = silk_SMLAWB( n_AR_Q14, tmp2, AR_shp_Q13[ 0 ] );
/* Loop over allpass sections */
for( j = 2; j < shapingLPCOrder; j += 2 ) {
/* Output of allpass section */
tmp2 = silk_SMLAWB( psDD->sAR2_Q14[ j - 1 ], psDD->sAR2_Q14[ j + 0 ] - tmp1, warping_Q16 );
psDD->sAR2_Q14[ j - 1 ] = tmp1;
n_AR_Q14 = silk_SMLAWB( n_AR_Q14, tmp1, AR_shp_Q13[ j - 1 ] );
/* Output of allpass section */
tmp1 = silk_SMLAWB( psDD->sAR2_Q14[ j + 0 ], psDD->sAR2_Q14[ j + 1 ] - tmp2, warping_Q16 );
psDD->sAR2_Q14[ j + 0 ] = tmp2;
n_AR_Q14 = silk_SMLAWB( n_AR_Q14, tmp2, AR_shp_Q13[ j ] );
}
psDD->sAR2_Q14[ shapingLPCOrder - 1 ] = tmp1;
n_AR_Q14 = silk_SMLAWB( n_AR_Q14, tmp1, AR_shp_Q13[ shapingLPCOrder - 1 ] );
n_AR_Q14 = silk_LSHIFT( n_AR_Q14, 1 ); /* Q11 -> Q12 */
n_AR_Q14 = silk_SMLAWB( n_AR_Q14, psDD->LF_AR_Q14, Tilt_Q14 ); /* Q12 */
n_AR_Q14 = silk_LSHIFT( n_AR_Q14, 2 ); /* Q12 -> Q14 */
n_LF_Q14 = silk_SMULWB( psDD->Shape_Q14[ *smpl_buf_idx ], LF_shp_Q14 ); /* Q12 */
n_LF_Q14 = silk_SMLAWT( n_LF_Q14, psDD->LF_AR_Q14, LF_shp_Q14 ); /* Q12 */
n_LF_Q14 = silk_LSHIFT( n_LF_Q14, 2 ); /* Q12 -> Q14 */
/* Input minus prediction plus noise feedback */
/* r = x[ i ] - LTP_pred - LPC_pred + n_AR + n_Tilt + n_LF + n_LTP */
tmp1 = silk_ADD32( n_AR_Q14, n_LF_Q14 ); /* Q14 */
tmp2 = silk_ADD32( n_LTP_Q14, LPC_pred_Q14 ); /* Q13 */
tmp1 = silk_SUB32( tmp2, tmp1 ); /* Q13 */
tmp1 = silk_RSHIFT_ROUND( tmp1, 4 ); /* Q10 */
r_Q10 = silk_SUB32( x_Q10[ i ], tmp1 ); /* residual error Q10 */
/* Flip sign depending on dither */
if ( psDD->Seed < 0 ) {
r_Q10 = -r_Q10;
}
r_Q10 = silk_LIMIT_32( r_Q10, -(31 << 10), 30 << 10 );
/* Find two quantization level candidates and measure their rate-distortion */
q1_Q10 = silk_SUB32( r_Q10, offset_Q10 );
q1_Q0 = silk_RSHIFT( q1_Q10, 10 );
470
471
472
473
474
475
476
477
478
479
480
481
482
if (Lambda_Q10 > 2048) {
/* For aggressive RDO, the bias becomes more than one pulse. */
int rdo_offset = Lambda_Q10/2 - 512;
if (q1_Q10 > rdo_offset) {
q1_Q0 = silk_RSHIFT( q1_Q10 - rdo_offset, 10 );
} else if (q1_Q10 < -rdo_offset) {
q1_Q0 = silk_RSHIFT( q1_Q10 + rdo_offset, 10 );
} else if (q1_Q10 < 0) {
q1_Q0 = -1;
} else {
q1_Q0 = 0;
}
}
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
if( q1_Q0 > 0 ) {
q1_Q10 = silk_SUB32( silk_LSHIFT( q1_Q0, 10 ), QUANT_LEVEL_ADJUST_Q10 );
q1_Q10 = silk_ADD32( q1_Q10, offset_Q10 );
q2_Q10 = silk_ADD32( q1_Q10, 1024 );
rd1_Q10 = silk_SMULBB( q1_Q10, Lambda_Q10 );
rd2_Q10 = silk_SMULBB( q2_Q10, Lambda_Q10 );
} else if( q1_Q0 == 0 ) {
q1_Q10 = offset_Q10;
q2_Q10 = silk_ADD32( q1_Q10, 1024 - QUANT_LEVEL_ADJUST_Q10 );
rd1_Q10 = silk_SMULBB( q1_Q10, Lambda_Q10 );
rd2_Q10 = silk_SMULBB( q2_Q10, Lambda_Q10 );
} else if( q1_Q0 == -1 ) {
q2_Q10 = offset_Q10;
q1_Q10 = silk_SUB32( q2_Q10, 1024 - QUANT_LEVEL_ADJUST_Q10 );
rd1_Q10 = silk_SMULBB( -q1_Q10, Lambda_Q10 );
rd2_Q10 = silk_SMULBB( q2_Q10, Lambda_Q10 );
} else { /* q1_Q0 < -1 */
q1_Q10 = silk_ADD32( silk_LSHIFT( q1_Q0, 10 ), QUANT_LEVEL_ADJUST_Q10 );
q1_Q10 = silk_ADD32( q1_Q10, offset_Q10 );
q2_Q10 = silk_ADD32( q1_Q10, 1024 );
rd1_Q10 = silk_SMULBB( -q1_Q10, Lambda_Q10 );
rd2_Q10 = silk_SMULBB( -q2_Q10, Lambda_Q10 );
}
rr_Q10 = silk_SUB32( r_Q10, q1_Q10 );
rd1_Q10 = silk_RSHIFT( silk_SMLABB( rd1_Q10, rr_Q10, rr_Q10 ), 10 );
rr_Q10 = silk_SUB32( r_Q10, q2_Q10 );
rd2_Q10 = silk_RSHIFT( silk_SMLABB( rd2_Q10, rr_Q10, rr_Q10 ), 10 );
if( rd1_Q10 < rd2_Q10 ) {
psSS[ 0 ].RD_Q10 = silk_ADD32( psDD->RD_Q10, rd1_Q10 );
psSS[ 1 ].RD_Q10 = silk_ADD32( psDD->RD_Q10, rd2_Q10 );
psSS[ 0 ].Q_Q10 = q1_Q10;
psSS[ 1 ].Q_Q10 = q2_Q10;
} else {
psSS[ 0 ].RD_Q10 = silk_ADD32( psDD->RD_Q10, rd2_Q10 );
psSS[ 1 ].RD_Q10 = silk_ADD32( psDD->RD_Q10, rd1_Q10 );
psSS[ 0 ].Q_Q10 = q2_Q10;
psSS[ 1 ].Q_Q10 = q1_Q10;
}
/* Update states for best quantization */
/* Quantized excitation */
exc_Q14 = silk_LSHIFT32( psSS[ 0 ].Q_Q10, 4 );
if ( psDD->Seed < 0 ) {
exc_Q14 = -exc_Q14;
}
/* Add predictions */
LPC_exc_Q14 = silk_ADD32( exc_Q14, LTP_pred_Q14 );
xq_Q14 = silk_ADD32( LPC_exc_Q14, LPC_pred_Q14 );
/* Update states */
536
537
psSS[ 0 ].Diff_Q14 = silk_SUB_LSHIFT32( xq_Q14, x_Q10[ i ], 4 );
sLF_AR_shp_Q14 = silk_SUB32( psSS[ 0 ].Diff_Q14, n_AR_Q14 );
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
psSS[ 0 ].sLTP_shp_Q14 = silk_SUB32( sLF_AR_shp_Q14, n_LF_Q14 );
psSS[ 0 ].LF_AR_Q14 = sLF_AR_shp_Q14;
psSS[ 0 ].LPC_exc_Q14 = LPC_exc_Q14;
psSS[ 0 ].xq_Q14 = xq_Q14;
/* Update states for second best quantization */
/* Quantized excitation */
exc_Q14 = silk_LSHIFT32( psSS[ 1 ].Q_Q10, 4 );
if ( psDD->Seed < 0 ) {
exc_Q14 = -exc_Q14;
}
/* Add predictions */
LPC_exc_Q14 = silk_ADD32( exc_Q14, LTP_pred_Q14 );
xq_Q14 = silk_ADD32( LPC_exc_Q14, LPC_pred_Q14 );
/* Update states */
556
557
psSS[ 1 ].Diff_Q14 = silk_SUB_LSHIFT32( xq_Q14, x_Q10[ i ], 4 );
sLF_AR_shp_Q14 = silk_SUB32( psSS[ 1 ].Diff_Q14, n_AR_Q14 );
558
559
560
561
562
563
psSS[ 1 ].sLTP_shp_Q14 = silk_SUB32( sLF_AR_shp_Q14, n_LF_Q14 );
psSS[ 1 ].LF_AR_Q14 = sLF_AR_shp_Q14;
psSS[ 1 ].LPC_exc_Q14 = LPC_exc_Q14;
psSS[ 1 ].xq_Q14 = xq_Q14;
}
564
565
566
*smpl_buf_idx = ( *smpl_buf_idx - 1 ) % DECISION_DELAY;
if( *smpl_buf_idx < 0 ) *smpl_buf_idx += DECISION_DELAY;
last_smple_idx = ( *smpl_buf_idx + decisionDelay ) % DECISION_DELAY;
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
/* Find winner */
RDmin_Q10 = psSampleState[ 0 ][ 0 ].RD_Q10;
Winner_ind = 0;
for( k = 1; k < nStatesDelayedDecision; k++ ) {
if( psSampleState[ k ][ 0 ].RD_Q10 < RDmin_Q10 ) {
RDmin_Q10 = psSampleState[ k ][ 0 ].RD_Q10;
Winner_ind = k;
}
}
/* Increase RD values of expired states */
Winner_rand_state = psDelDec[ Winner_ind ].RandState[ last_smple_idx ];
for( k = 0; k < nStatesDelayedDecision; k++ ) {
if( psDelDec[ k ].RandState[ last_smple_idx ] != Winner_rand_state ) {
psSampleState[ k ][ 0 ].RD_Q10 = silk_ADD32( psSampleState[ k ][ 0 ].RD_Q10, silk_int32_MAX >> 4 );
psSampleState[ k ][ 1 ].RD_Q10 = silk_ADD32( psSampleState[ k ][ 1 ].RD_Q10, silk_int32_MAX >> 4 );
silk_assert( psSampleState[ k ][ 0 ].RD_Q10 >= 0 );
}
}
/* Find worst in first set and best in second set */
RDmax_Q10 = psSampleState[ 0 ][ 0 ].RD_Q10;
RDmin_Q10 = psSampleState[ 0 ][ 1 ].RD_Q10;
RDmax_ind = 0;
RDmin_ind = 0;
for( k = 1; k < nStatesDelayedDecision; k++ ) {
/* find worst in first set */
if( psSampleState[ k ][ 0 ].RD_Q10 > RDmax_Q10 ) {
RDmax_Q10 = psSampleState[ k ][ 0 ].RD_Q10;
RDmax_ind = k;
}
/* find best in second set */
if( psSampleState[ k ][ 1 ].RD_Q10 < RDmin_Q10 ) {
RDmin_Q10 = psSampleState[ k ][ 1 ].RD_Q10;
RDmin_ind = k;
}
}
/* Replace a state if best from second set outperforms worst in first set */
if( RDmin_Q10 < RDmax_Q10 ) {
silk_memcpy( ( (opus_int32 *)&psDelDec[ RDmax_ind ] ) + i,
( (opus_int32 *)&psDelDec[ RDmin_ind ] ) + i, sizeof( NSQ_del_dec_struct ) - i * sizeof( opus_int32) );
silk_memcpy( &psSampleState[ RDmax_ind ][ 0 ], &psSampleState[ RDmin_ind ][ 1 ], sizeof( NSQ_sample_struct ) );
}
/* Write samples from winner to output and long-term filter states */
psDD = &psDelDec[ Winner_ind ];
if( subfr > 0 || i >= decisionDelay ) {
pulses[ i - decisionDelay ] = (opus_int8)silk_RSHIFT_ROUND( psDD->Q_Q10[ last_smple_idx ], 10 );
xq[ i - decisionDelay ] = (opus_int16)silk_SAT16( silk_RSHIFT_ROUND(
silk_SMULWW( psDD->Xq_Q14[ last_smple_idx ], delayedGain_Q10[ last_smple_idx ] ), 8 ) );
NSQ->sLTP_shp_Q14[ NSQ->sLTP_shp_buf_idx - decisionDelay ] = psDD->Shape_Q14[ last_smple_idx ];
sLTP_Q15[ NSQ->sLTP_buf_idx - decisionDelay ] = psDD->Pred_Q15[ last_smple_idx ];
}
NSQ->sLTP_shp_buf_idx++;
NSQ->sLTP_buf_idx++;
/* Update states */
for( k = 0; k < nStatesDelayedDecision; k++ ) {
psDD = &psDelDec[ k ];
psSS = &psSampleState[ k ][ 0 ];
psDD->LF_AR_Q14 = psSS->LF_AR_Q14;
630
psDD->Diff_Q14 = psSS->Diff_Q14;
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
psDD->sLPC_Q14[ NSQ_LPC_BUF_LENGTH + i ] = psSS->xq_Q14;
psDD->Xq_Q14[ *smpl_buf_idx ] = psSS->xq_Q14;
psDD->Q_Q10[ *smpl_buf_idx ] = psSS->Q_Q10;
psDD->Pred_Q15[ *smpl_buf_idx ] = silk_LSHIFT32( psSS->LPC_exc_Q14, 1 );
psDD->Shape_Q14[ *smpl_buf_idx ] = psSS->sLTP_shp_Q14;
psDD->Seed = silk_ADD32_ovflw( psDD->Seed, silk_RSHIFT_ROUND( psSS->Q_Q10, 10 ) );
psDD->RandState[ *smpl_buf_idx ] = psDD->Seed;
psDD->RD_Q10 = psSS->RD_Q10;
}
delayedGain_Q10[ *smpl_buf_idx ] = Gain_Q10;
}
/* Update LPC states */
for( k = 0; k < nStatesDelayedDecision; k++ ) {
psDD = &psDelDec[ k ];
silk_memcpy( psDD->sLPC_Q14, &psDD->sLPC_Q14[ length ], NSQ_LPC_BUF_LENGTH * sizeof( opus_int32 ) );
}
647
RESTORE_STACK;
649
#endif /* OVERRIDE_silk_noise_shape_quantizer_del_dec */
651
static OPUS_INLINE void silk_nsq_del_dec_scale_states(
652
653
654
const silk_encoder_state *psEncC, /* I Encoder State */
silk_nsq_state *NSQ, /* I/O NSQ state */
NSQ_del_dec_struct psDelDec[], /* I/O Delayed decision states */
655
const opus_int16 x16[], /* I Input */
656
657
658
659
660
661
662
663
664
665
666
667
668
opus_int32 x_sc_Q10[], /* O Input scaled with 1/Gain in Q10 */
const opus_int16 sLTP[], /* I Re-whitened LTP state in Q0 */
opus_int32 sLTP_Q15[], /* O LTP state matching scaled input */
opus_int subfr, /* I Subframe number */
opus_int nStatesDelayedDecision, /* I Number of del dec states */
const opus_int LTP_scale_Q14, /* I LTP state scaling */
const opus_int32 Gains_Q16[ MAX_NB_SUBFR ], /* I */
const opus_int pitchL[ MAX_NB_SUBFR ], /* I Pitch lag */
const opus_int signal_type, /* I Signal type */
const opus_int decisionDelay /* I Decision delay */
)
{
opus_int i, k, lag;
669
opus_int32 gain_adj_Q16, inv_gain_Q31, inv_gain_Q26;
670
671
672
673
674
675
676
NSQ_del_dec_struct *psDD;
lag = pitchL[ subfr ];
inv_gain_Q31 = silk_INVERSE32_varQ( silk_max( Gains_Q16[ subfr ], 1 ), 47 );
silk_assert( inv_gain_Q31 != 0 );
/* Scale input */
677
inv_gain_Q26 = silk_RSHIFT_ROUND( inv_gain_Q31, 5 );
678
for( i = 0; i < psEncC->subfr_length; i++ ) {
679
x_sc_Q10[ i ] = silk_SMULWW( x16[ i ], inv_gain_Q26 );
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
}
/* After rewhitening the LTP state is un-scaled, so scale with inv_gain_Q16 */
if( NSQ->rewhite_flag ) {
if( subfr == 0 ) {
/* Do LTP downscaling */
inv_gain_Q31 = silk_LSHIFT( silk_SMULWB( inv_gain_Q31, LTP_scale_Q14 ), 2 );
}
for( i = NSQ->sLTP_buf_idx - lag - LTP_ORDER / 2; i < NSQ->sLTP_buf_idx; i++ ) {
silk_assert( i < MAX_FRAME_LENGTH );
sLTP_Q15[ i ] = silk_SMULWB( inv_gain_Q31, sLTP[ i ] );
}
}
/* Adjust for changing gain */
695
696
697
if( Gains_Q16[ subfr ] != NSQ->prev_gain_Q16 ) {
gain_adj_Q16 = silk_DIV32_varQ( NSQ->prev_gain_Q16, Gains_Q16[ subfr ], 16 );
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
/* Scale long-term shaping state */
for( i = NSQ->sLTP_shp_buf_idx - psEncC->ltp_mem_length; i < NSQ->sLTP_shp_buf_idx; i++ ) {
NSQ->sLTP_shp_Q14[ i ] = silk_SMULWW( gain_adj_Q16, NSQ->sLTP_shp_Q14[ i ] );
}
/* Scale long-term prediction state */
if( signal_type == TYPE_VOICED && NSQ->rewhite_flag == 0 ) {
for( i = NSQ->sLTP_buf_idx - lag - LTP_ORDER / 2; i < NSQ->sLTP_buf_idx - decisionDelay; i++ ) {
sLTP_Q15[ i ] = silk_SMULWW( gain_adj_Q16, sLTP_Q15[ i ] );
}
}
for( k = 0; k < nStatesDelayedDecision; k++ ) {
psDD = &psDelDec[ k ];
/* Scale scalar states */
psDD->LF_AR_Q14 = silk_SMULWW( gain_adj_Q16, psDD->LF_AR_Q14 );
715
psDD->Diff_Q14 = silk_SMULWW( gain_adj_Q16, psDD->Diff_Q14 );
716
717
718
719
720
721
722
723
724
725
726
727
728
/* Scale short-term prediction and shaping states */
for( i = 0; i < NSQ_LPC_BUF_LENGTH; i++ ) {
psDD->sLPC_Q14[ i ] = silk_SMULWW( gain_adj_Q16, psDD->sLPC_Q14[ i ] );
}
for( i = 0; i < MAX_SHAPE_LPC_ORDER; i++ ) {
psDD->sAR2_Q14[ i ] = silk_SMULWW( gain_adj_Q16, psDD->sAR2_Q14[ i ] );
}
for( i = 0; i < DECISION_DELAY; i++ ) {
psDD->Pred_Q15[ i ] = silk_SMULWW( gain_adj_Q16, psDD->Pred_Q15[ i ] );
psDD->Shape_Q14[ i ] = silk_SMULWW( gain_adj_Q16, psDD->Shape_Q14[ i ] );
}
}
729
730
731
/* Save inverse gain */
NSQ->prev_gain_Q16 = Gains_Q16[ subfr ];