|
1 /* Copyright (c) 2007-2008 CSIRO |
|
2 Copyright (c) 2007-2010 Xiph.Org Foundation |
|
3 Copyright (c) 2008 Gregory Maxwell |
|
4 Written by Jean-Marc Valin and Gregory Maxwell */ |
|
5 /* |
|
6 Redistribution and use in source and binary forms, with or without |
|
7 modification, are permitted provided that the following conditions |
|
8 are met: |
|
9 |
|
10 - Redistributions of source code must retain the above copyright |
|
11 notice, this list of conditions and the following disclaimer. |
|
12 |
|
13 - Redistributions in binary form must reproduce the above copyright |
|
14 notice, this list of conditions and the following disclaimer in the |
|
15 documentation and/or other materials provided with the distribution. |
|
16 |
|
17 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS |
|
18 ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT |
|
19 LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR |
|
20 A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER |
|
21 OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, |
|
22 EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, |
|
23 PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR |
|
24 PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF |
|
25 LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING |
|
26 NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS |
|
27 SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
|
28 */ |
|
29 |
|
30 #ifdef HAVE_CONFIG_H |
|
31 #include "config.h" |
|
32 #endif |
|
33 |
|
34 #define CELT_ENCODER_C |
|
35 |
|
36 #include "cpu_support.h" |
|
37 #include "os_support.h" |
|
38 #include "mdct.h" |
|
39 #include <math.h> |
|
40 #include "celt.h" |
|
41 #include "pitch.h" |
|
42 #include "bands.h" |
|
43 #include "modes.h" |
|
44 #include "entcode.h" |
|
45 #include "quant_bands.h" |
|
46 #include "rate.h" |
|
47 #include "stack_alloc.h" |
|
48 #include "mathops.h" |
|
49 #include "float_cast.h" |
|
50 #include <stdarg.h> |
|
51 #include "celt_lpc.h" |
|
52 #include "vq.h" |
|
53 |
|
54 |
|
55 /** Encoder state |
|
56 @brief Encoder state |
|
57 */ |
|
58 struct OpusCustomEncoder { |
|
59 const OpusCustomMode *mode; /**< Mode used by the encoder */ |
|
60 int overlap; |
|
61 int channels; |
|
62 int stream_channels; |
|
63 |
|
64 int force_intra; |
|
65 int clip; |
|
66 int disable_pf; |
|
67 int complexity; |
|
68 int upsample; |
|
69 int start, end; |
|
70 |
|
71 opus_int32 bitrate; |
|
72 int vbr; |
|
73 int signalling; |
|
74 int constrained_vbr; /* If zero, VBR can do whatever it likes with the rate */ |
|
75 int loss_rate; |
|
76 int lsb_depth; |
|
77 int variable_duration; |
|
78 int lfe; |
|
79 int arch; |
|
80 |
|
81 /* Everything beyond this point gets cleared on a reset */ |
|
82 #define ENCODER_RESET_START rng |
|
83 |
|
84 opus_uint32 rng; |
|
85 int spread_decision; |
|
86 opus_val32 delayedIntra; |
|
87 int tonal_average; |
|
88 int lastCodedBands; |
|
89 int hf_average; |
|
90 int tapset_decision; |
|
91 |
|
92 int prefilter_period; |
|
93 opus_val16 prefilter_gain; |
|
94 int prefilter_tapset; |
|
95 #ifdef RESYNTH |
|
96 int prefilter_period_old; |
|
97 opus_val16 prefilter_gain_old; |
|
98 int prefilter_tapset_old; |
|
99 #endif |
|
100 int consec_transient; |
|
101 AnalysisInfo analysis; |
|
102 |
|
103 opus_val32 preemph_memE[2]; |
|
104 opus_val32 preemph_memD[2]; |
|
105 |
|
106 /* VBR-related parameters */ |
|
107 opus_int32 vbr_reservoir; |
|
108 opus_int32 vbr_drift; |
|
109 opus_int32 vbr_offset; |
|
110 opus_int32 vbr_count; |
|
111 opus_val32 overlap_max; |
|
112 opus_val16 stereo_saving; |
|
113 int intensity; |
|
114 opus_val16 *energy_mask; |
|
115 opus_val16 spec_avg; |
|
116 |
|
117 #ifdef RESYNTH |
|
118 /* +MAX_PERIOD/2 to make space for overlap */ |
|
119 celt_sig syn_mem[2][2*MAX_PERIOD+MAX_PERIOD/2]; |
|
120 #endif |
|
121 |
|
122 celt_sig in_mem[1]; /* Size = channels*mode->overlap */ |
|
123 /* celt_sig prefilter_mem[], Size = channels*COMBFILTER_MAXPERIOD */ |
|
124 /* opus_val16 oldBandE[], Size = channels*mode->nbEBands */ |
|
125 /* opus_val16 oldLogE[], Size = channels*mode->nbEBands */ |
|
126 /* opus_val16 oldLogE2[], Size = channels*mode->nbEBands */ |
|
127 }; |
|
128 |
|
129 int celt_encoder_get_size(int channels) |
|
130 { |
|
131 CELTMode *mode = opus_custom_mode_create(48000, 960, NULL); |
|
132 return opus_custom_encoder_get_size(mode, channels); |
|
133 } |
|
134 |
|
135 OPUS_CUSTOM_NOSTATIC int opus_custom_encoder_get_size(const CELTMode *mode, int channels) |
|
136 { |
|
137 int size = sizeof(struct CELTEncoder) |
|
138 + (channels*mode->overlap-1)*sizeof(celt_sig) /* celt_sig in_mem[channels*mode->overlap]; */ |
|
139 + channels*COMBFILTER_MAXPERIOD*sizeof(celt_sig) /* celt_sig prefilter_mem[channels*COMBFILTER_MAXPERIOD]; */ |
|
140 + 3*channels*mode->nbEBands*sizeof(opus_val16); /* opus_val16 oldBandE[channels*mode->nbEBands]; */ |
|
141 /* opus_val16 oldLogE[channels*mode->nbEBands]; */ |
|
142 /* opus_val16 oldLogE2[channels*mode->nbEBands]; */ |
|
143 return size; |
|
144 } |
|
145 |
|
146 #ifdef CUSTOM_MODES |
|
147 CELTEncoder *opus_custom_encoder_create(const CELTMode *mode, int channels, int *error) |
|
148 { |
|
149 int ret; |
|
150 CELTEncoder *st = (CELTEncoder *)opus_alloc(opus_custom_encoder_get_size(mode, channels)); |
|
151 /* init will handle the NULL case */ |
|
152 ret = opus_custom_encoder_init(st, mode, channels); |
|
153 if (ret != OPUS_OK) |
|
154 { |
|
155 opus_custom_encoder_destroy(st); |
|
156 st = NULL; |
|
157 } |
|
158 if (error) |
|
159 *error = ret; |
|
160 return st; |
|
161 } |
|
162 #endif /* CUSTOM_MODES */ |
|
163 |
|
164 static int opus_custom_encoder_init_arch(CELTEncoder *st, const CELTMode *mode, |
|
165 int channels, int arch) |
|
166 { |
|
167 if (channels < 0 || channels > 2) |
|
168 return OPUS_BAD_ARG; |
|
169 |
|
170 if (st==NULL || mode==NULL) |
|
171 return OPUS_ALLOC_FAIL; |
|
172 |
|
173 OPUS_CLEAR((char*)st, opus_custom_encoder_get_size(mode, channels)); |
|
174 |
|
175 st->mode = mode; |
|
176 st->overlap = mode->overlap; |
|
177 st->stream_channels = st->channels = channels; |
|
178 |
|
179 st->upsample = 1; |
|
180 st->start = 0; |
|
181 st->end = st->mode->effEBands; |
|
182 st->signalling = 1; |
|
183 |
|
184 st->arch = arch; |
|
185 |
|
186 st->constrained_vbr = 1; |
|
187 st->clip = 1; |
|
188 |
|
189 st->bitrate = OPUS_BITRATE_MAX; |
|
190 st->vbr = 0; |
|
191 st->force_intra = 0; |
|
192 st->complexity = 5; |
|
193 st->lsb_depth=24; |
|
194 |
|
195 opus_custom_encoder_ctl(st, OPUS_RESET_STATE); |
|
196 |
|
197 return OPUS_OK; |
|
198 } |
|
199 |
|
200 #ifdef CUSTOM_MODES |
|
201 int opus_custom_encoder_init(CELTEncoder *st, const CELTMode *mode, int channels) |
|
202 { |
|
203 return opus_custom_encoder_init_arch(st, mode, channels, opus_select_arch()); |
|
204 } |
|
205 #endif |
|
206 |
|
207 int celt_encoder_init(CELTEncoder *st, opus_int32 sampling_rate, int channels, |
|
208 int arch) |
|
209 { |
|
210 int ret; |
|
211 ret = opus_custom_encoder_init_arch(st, |
|
212 opus_custom_mode_create(48000, 960, NULL), channels, arch); |
|
213 if (ret != OPUS_OK) |
|
214 return ret; |
|
215 st->upsample = resampling_factor(sampling_rate); |
|
216 return OPUS_OK; |
|
217 } |
|
218 |
|
219 #ifdef CUSTOM_MODES |
|
220 void opus_custom_encoder_destroy(CELTEncoder *st) |
|
221 { |
|
222 opus_free(st); |
|
223 } |
|
224 #endif /* CUSTOM_MODES */ |
|
225 |
|
226 |
|
227 static int transient_analysis(const opus_val32 * OPUS_RESTRICT in, int len, int C, |
|
228 opus_val16 *tf_estimate, int *tf_chan) |
|
229 { |
|
230 int i; |
|
231 VARDECL(opus_val16, tmp); |
|
232 opus_val32 mem0,mem1; |
|
233 int is_transient = 0; |
|
234 opus_int32 mask_metric = 0; |
|
235 int c; |
|
236 opus_val16 tf_max; |
|
237 int len2; |
|
238 /* Table of 6*64/x, trained on real data to minimize the average error */ |
|
239 static const unsigned char inv_table[128] = { |
|
240 255,255,156,110, 86, 70, 59, 51, 45, 40, 37, 33, 31, 28, 26, 25, |
|
241 23, 22, 21, 20, 19, 18, 17, 16, 16, 15, 15, 14, 13, 13, 12, 12, |
|
242 12, 12, 11, 11, 11, 10, 10, 10, 9, 9, 9, 9, 9, 9, 8, 8, |
|
243 8, 8, 8, 7, 7, 7, 7, 7, 7, 6, 6, 6, 6, 6, 6, 6, |
|
244 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5, 5, 5, 5, |
|
245 5, 5, 5, 5, 5, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, |
|
246 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 3, 3, |
|
247 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 2, |
|
248 }; |
|
249 SAVE_STACK; |
|
250 ALLOC(tmp, len, opus_val16); |
|
251 |
|
252 len2=len/2; |
|
253 for (c=0;c<C;c++) |
|
254 { |
|
255 opus_val32 mean; |
|
256 opus_int32 unmask=0; |
|
257 opus_val32 norm; |
|
258 opus_val16 maxE; |
|
259 mem0=0; |
|
260 mem1=0; |
|
261 /* High-pass filter: (1 - 2*z^-1 + z^-2) / (1 - z^-1 + .5*z^-2) */ |
|
262 for (i=0;i<len;i++) |
|
263 { |
|
264 opus_val32 x,y; |
|
265 x = SHR32(in[i+c*len],SIG_SHIFT); |
|
266 y = ADD32(mem0, x); |
|
267 #ifdef FIXED_POINT |
|
268 mem0 = mem1 + y - SHL32(x,1); |
|
269 mem1 = x - SHR32(y,1); |
|
270 #else |
|
271 mem0 = mem1 + y - 2*x; |
|
272 mem1 = x - .5f*y; |
|
273 #endif |
|
274 tmp[i] = EXTRACT16(SHR32(y,2)); |
|
275 /*printf("%f ", tmp[i]);*/ |
|
276 } |
|
277 /*printf("\n");*/ |
|
278 /* First few samples are bad because we don't propagate the memory */ |
|
279 for (i=0;i<12;i++) |
|
280 tmp[i] = 0; |
|
281 |
|
282 #ifdef FIXED_POINT |
|
283 /* Normalize tmp to max range */ |
|
284 { |
|
285 int shift=0; |
|
286 shift = 14-celt_ilog2(1+celt_maxabs16(tmp, len)); |
|
287 if (shift!=0) |
|
288 { |
|
289 for (i=0;i<len;i++) |
|
290 tmp[i] = SHL16(tmp[i], shift); |
|
291 } |
|
292 } |
|
293 #endif |
|
294 |
|
295 mean=0; |
|
296 mem0=0; |
|
297 /* Grouping by two to reduce complexity */ |
|
298 /* Forward pass to compute the post-echo threshold*/ |
|
299 for (i=0;i<len2;i++) |
|
300 { |
|
301 opus_val16 x2 = PSHR32(MULT16_16(tmp[2*i],tmp[2*i]) + MULT16_16(tmp[2*i+1],tmp[2*i+1]),16); |
|
302 mean += x2; |
|
303 #ifdef FIXED_POINT |
|
304 /* FIXME: Use PSHR16() instead */ |
|
305 tmp[i] = mem0 + PSHR32(x2-mem0,4); |
|
306 #else |
|
307 tmp[i] = mem0 + MULT16_16_P15(QCONST16(.0625f,15),x2-mem0); |
|
308 #endif |
|
309 mem0 = tmp[i]; |
|
310 } |
|
311 |
|
312 mem0=0; |
|
313 maxE=0; |
|
314 /* Backward pass to compute the pre-echo threshold */ |
|
315 for (i=len2-1;i>=0;i--) |
|
316 { |
|
317 #ifdef FIXED_POINT |
|
318 /* FIXME: Use PSHR16() instead */ |
|
319 tmp[i] = mem0 + PSHR32(tmp[i]-mem0,3); |
|
320 #else |
|
321 tmp[i] = mem0 + MULT16_16_P15(QCONST16(0.125f,15),tmp[i]-mem0); |
|
322 #endif |
|
323 mem0 = tmp[i]; |
|
324 maxE = MAX16(maxE, mem0); |
|
325 } |
|
326 /*for (i=0;i<len2;i++)printf("%f ", tmp[i]/mean);printf("\n");*/ |
|
327 |
|
328 /* Compute the ratio of the "frame energy" over the harmonic mean of the energy. |
|
329 This essentially corresponds to a bitrate-normalized temporal noise-to-mask |
|
330 ratio */ |
|
331 |
|
332 /* As a compromise with the old transient detector, frame energy is the |
|
333 geometric mean of the energy and half the max */ |
|
334 #ifdef FIXED_POINT |
|
335 /* Costs two sqrt() to avoid overflows */ |
|
336 mean = MULT16_16(celt_sqrt(mean), celt_sqrt(MULT16_16(maxE,len2>>1))); |
|
337 #else |
|
338 mean = celt_sqrt(mean * maxE*.5*len2); |
|
339 #endif |
|
340 /* Inverse of the mean energy in Q15+6 */ |
|
341 norm = SHL32(EXTEND32(len2),6+14)/ADD32(EPSILON,SHR32(mean,1)); |
|
342 /* Compute harmonic mean discarding the unreliable boundaries |
|
343 The data is smooth, so we only take 1/4th of the samples */ |
|
344 unmask=0; |
|
345 for (i=12;i<len2-5;i+=4) |
|
346 { |
|
347 int id; |
|
348 #ifdef FIXED_POINT |
|
349 id = IMAX(0,IMIN(127,MULT16_32_Q15(tmp[i],norm))); /* Do not round to nearest */ |
|
350 #else |
|
351 id = IMAX(0,IMIN(127,(int)floor(64*norm*tmp[i]))); /* Do not round to nearest */ |
|
352 #endif |
|
353 unmask += inv_table[id]; |
|
354 } |
|
355 /*printf("%d\n", unmask);*/ |
|
356 /* Normalize, compensate for the 1/4th of the sample and the factor of 6 in the inverse table */ |
|
357 unmask = 64*unmask*4/(6*(len2-17)); |
|
358 if (unmask>mask_metric) |
|
359 { |
|
360 *tf_chan = c; |
|
361 mask_metric = unmask; |
|
362 } |
|
363 } |
|
364 is_transient = mask_metric>200; |
|
365 |
|
366 /* Arbitrary metric for VBR boost */ |
|
367 tf_max = MAX16(0,celt_sqrt(27*mask_metric)-42); |
|
368 /* *tf_estimate = 1 + MIN16(1, sqrt(MAX16(0, tf_max-30))/20); */ |
|
369 *tf_estimate = celt_sqrt(MAX16(0, SHL32(MULT16_16(QCONST16(0.0069,14),MIN16(163,tf_max)),14)-QCONST32(0.139,28))); |
|
370 /*printf("%d %f\n", tf_max, mask_metric);*/ |
|
371 RESTORE_STACK; |
|
372 #ifdef FUZZING |
|
373 is_transient = rand()&0x1; |
|
374 #endif |
|
375 /*printf("%d %f %d\n", is_transient, (float)*tf_estimate, tf_max);*/ |
|
376 return is_transient; |
|
377 } |
|
378 |
|
379 /* Looks for sudden increases of energy to decide whether we need to patch |
|
380 the transient decision */ |
|
381 int patch_transient_decision(opus_val16 *newE, opus_val16 *oldE, int nbEBands, |
|
382 int end, int C) |
|
383 { |
|
384 int i, c; |
|
385 opus_val32 mean_diff=0; |
|
386 opus_val16 spread_old[26]; |
|
387 /* Apply an aggressive (-6 dB/Bark) spreading function to the old frame to |
|
388 avoid false detection caused by irrelevant bands */ |
|
389 if (C==1) |
|
390 { |
|
391 spread_old[0] = oldE[0]; |
|
392 for (i=1;i<end;i++) |
|
393 spread_old[i] = MAX16(spread_old[i-1]-QCONST16(1.0f, DB_SHIFT), oldE[i]); |
|
394 } else { |
|
395 spread_old[0] = MAX16(oldE[0],oldE[nbEBands]); |
|
396 for (i=1;i<end;i++) |
|
397 spread_old[i] = MAX16(spread_old[i-1]-QCONST16(1.0f, DB_SHIFT), |
|
398 MAX16(oldE[i],oldE[i+nbEBands])); |
|
399 } |
|
400 for (i=end-2;i>=0;i--) |
|
401 spread_old[i] = MAX16(spread_old[i], spread_old[i+1]-QCONST16(1.0f, DB_SHIFT)); |
|
402 /* Compute mean increase */ |
|
403 c=0; do { |
|
404 for (i=2;i<end-1;i++) |
|
405 { |
|
406 opus_val16 x1, x2; |
|
407 x1 = MAX16(0, newE[i]); |
|
408 x2 = MAX16(0, spread_old[i]); |
|
409 mean_diff = ADD32(mean_diff, EXTEND32(MAX16(0, SUB16(x1, x2)))); |
|
410 } |
|
411 } while (++c<C); |
|
412 mean_diff = DIV32(mean_diff, C*(end-3)); |
|
413 /*printf("%f %f %d\n", mean_diff, max_diff, count);*/ |
|
414 return mean_diff > QCONST16(1.f, DB_SHIFT); |
|
415 } |
|
416 |
|
417 /** Apply window and compute the MDCT for all sub-frames and |
|
418 all channels in a frame */ |
|
419 static void compute_mdcts(const CELTMode *mode, int shortBlocks, celt_sig * OPUS_RESTRICT in, |
|
420 celt_sig * OPUS_RESTRICT out, int C, int CC, int LM, int upsample) |
|
421 { |
|
422 const int overlap = OVERLAP(mode); |
|
423 int N; |
|
424 int B; |
|
425 int shift; |
|
426 int i, b, c; |
|
427 if (shortBlocks) |
|
428 { |
|
429 B = shortBlocks; |
|
430 N = mode->shortMdctSize; |
|
431 shift = mode->maxLM; |
|
432 } else { |
|
433 B = 1; |
|
434 N = mode->shortMdctSize<<LM; |
|
435 shift = mode->maxLM-LM; |
|
436 } |
|
437 c=0; do { |
|
438 for (b=0;b<B;b++) |
|
439 { |
|
440 /* Interleaving the sub-frames while doing the MDCTs */ |
|
441 clt_mdct_forward(&mode->mdct, in+c*(B*N+overlap)+b*N, &out[b+c*N*B], mode->window, overlap, shift, B); |
|
442 } |
|
443 } while (++c<CC); |
|
444 if (CC==2&&C==1) |
|
445 { |
|
446 for (i=0;i<B*N;i++) |
|
447 out[i] = ADD32(HALF32(out[i]), HALF32(out[B*N+i])); |
|
448 } |
|
449 if (upsample != 1) |
|
450 { |
|
451 c=0; do |
|
452 { |
|
453 int bound = B*N/upsample; |
|
454 for (i=0;i<bound;i++) |
|
455 out[c*B*N+i] *= upsample; |
|
456 for (;i<B*N;i++) |
|
457 out[c*B*N+i] = 0; |
|
458 } while (++c<C); |
|
459 } |
|
460 } |
|
461 |
|
462 |
|
463 void celt_preemphasis(const opus_val16 * OPUS_RESTRICT pcmp, celt_sig * OPUS_RESTRICT inp, |
|
464 int N, int CC, int upsample, const opus_val16 *coef, celt_sig *mem, int clip) |
|
465 { |
|
466 int i; |
|
467 opus_val16 coef0; |
|
468 celt_sig m; |
|
469 int Nu; |
|
470 |
|
471 coef0 = coef[0]; |
|
472 |
|
473 |
|
474 Nu = N/upsample; |
|
475 if (upsample!=1) |
|
476 { |
|
477 for (i=0;i<N;i++) |
|
478 inp[i] = 0; |
|
479 } |
|
480 for (i=0;i<Nu;i++) |
|
481 { |
|
482 celt_sig x; |
|
483 |
|
484 x = SCALEIN(pcmp[CC*i]); |
|
485 #ifndef FIXED_POINT |
|
486 /* Replace NaNs with zeros */ |
|
487 if (!(x==x)) |
|
488 x = 0; |
|
489 #endif |
|
490 inp[i*upsample] = x; |
|
491 } |
|
492 |
|
493 #ifndef FIXED_POINT |
|
494 if (clip) |
|
495 { |
|
496 /* Clip input to avoid encoding non-portable files */ |
|
497 for (i=0;i<Nu;i++) |
|
498 inp[i*upsample] = MAX32(-65536.f, MIN32(65536.f,inp[i*upsample])); |
|
499 } |
|
500 #else |
|
501 (void)clip; /* Avoids a warning about clip being unused. */ |
|
502 #endif |
|
503 m = *mem; |
|
504 #ifdef CUSTOM_MODES |
|
505 if (coef[1] != 0) |
|
506 { |
|
507 opus_val16 coef1 = coef[1]; |
|
508 opus_val16 coef2 = coef[2]; |
|
509 for (i=0;i<N;i++) |
|
510 { |
|
511 celt_sig x, tmp; |
|
512 x = inp[i]; |
|
513 /* Apply pre-emphasis */ |
|
514 tmp = MULT16_16(coef2, x); |
|
515 inp[i] = tmp + m; |
|
516 m = MULT16_32_Q15(coef1, inp[i]) - MULT16_32_Q15(coef0, tmp); |
|
517 } |
|
518 } else |
|
519 #endif |
|
520 { |
|
521 for (i=0;i<N;i++) |
|
522 { |
|
523 celt_sig x; |
|
524 x = SHL32(inp[i], SIG_SHIFT); |
|
525 /* Apply pre-emphasis */ |
|
526 inp[i] = x + m; |
|
527 m = - MULT16_32_Q15(coef0, x); |
|
528 } |
|
529 } |
|
530 *mem = m; |
|
531 } |
|
532 |
|
533 |
|
534 |
|
535 static opus_val32 l1_metric(const celt_norm *tmp, int N, int LM, opus_val16 bias) |
|
536 { |
|
537 int i; |
|
538 opus_val32 L1; |
|
539 L1 = 0; |
|
540 for (i=0;i<N;i++) |
|
541 L1 += EXTEND32(ABS16(tmp[i])); |
|
542 /* When in doubt, prefer good freq resolution */ |
|
543 L1 = MAC16_32_Q15(L1, LM*bias, L1); |
|
544 return L1; |
|
545 |
|
546 } |
|
547 |
|
548 static int tf_analysis(const CELTMode *m, int len, int isTransient, |
|
549 int *tf_res, int lambda, celt_norm *X, int N0, int LM, |
|
550 int *tf_sum, opus_val16 tf_estimate, int tf_chan) |
|
551 { |
|
552 int i; |
|
553 VARDECL(int, metric); |
|
554 int cost0; |
|
555 int cost1; |
|
556 VARDECL(int, path0); |
|
557 VARDECL(int, path1); |
|
558 VARDECL(celt_norm, tmp); |
|
559 VARDECL(celt_norm, tmp_1); |
|
560 int sel; |
|
561 int selcost[2]; |
|
562 int tf_select=0; |
|
563 opus_val16 bias; |
|
564 |
|
565 SAVE_STACK; |
|
566 bias = MULT16_16_Q14(QCONST16(.04f,15), MAX16(-QCONST16(.25f,14), QCONST16(.5f,14)-tf_estimate)); |
|
567 /*printf("%f ", bias);*/ |
|
568 |
|
569 ALLOC(metric, len, int); |
|
570 ALLOC(tmp, (m->eBands[len]-m->eBands[len-1])<<LM, celt_norm); |
|
571 ALLOC(tmp_1, (m->eBands[len]-m->eBands[len-1])<<LM, celt_norm); |
|
572 ALLOC(path0, len, int); |
|
573 ALLOC(path1, len, int); |
|
574 |
|
575 *tf_sum = 0; |
|
576 for (i=0;i<len;i++) |
|
577 { |
|
578 int j, k, N; |
|
579 int narrow; |
|
580 opus_val32 L1, best_L1; |
|
581 int best_level=0; |
|
582 N = (m->eBands[i+1]-m->eBands[i])<<LM; |
|
583 /* band is too narrow to be split down to LM=-1 */ |
|
584 narrow = (m->eBands[i+1]-m->eBands[i])==1; |
|
585 for (j=0;j<N;j++) |
|
586 tmp[j] = X[tf_chan*N0 + j+(m->eBands[i]<<LM)]; |
|
587 /* Just add the right channel if we're in stereo */ |
|
588 /*if (C==2) |
|
589 for (j=0;j<N;j++) |
|
590 tmp[j] = ADD16(SHR16(tmp[j], 1),SHR16(X[N0+j+(m->eBands[i]<<LM)], 1));*/ |
|
591 L1 = l1_metric(tmp, N, isTransient ? LM : 0, bias); |
|
592 best_L1 = L1; |
|
593 /* Check the -1 case for transients */ |
|
594 if (isTransient && !narrow) |
|
595 { |
|
596 for (j=0;j<N;j++) |
|
597 tmp_1[j] = tmp[j]; |
|
598 haar1(tmp_1, N>>LM, 1<<LM); |
|
599 L1 = l1_metric(tmp_1, N, LM+1, bias); |
|
600 if (L1<best_L1) |
|
601 { |
|
602 best_L1 = L1; |
|
603 best_level = -1; |
|
604 } |
|
605 } |
|
606 /*printf ("%f ", L1);*/ |
|
607 for (k=0;k<LM+!(isTransient||narrow);k++) |
|
608 { |
|
609 int B; |
|
610 |
|
611 if (isTransient) |
|
612 B = (LM-k-1); |
|
613 else |
|
614 B = k+1; |
|
615 |
|
616 haar1(tmp, N>>k, 1<<k); |
|
617 |
|
618 L1 = l1_metric(tmp, N, B, bias); |
|
619 |
|
620 if (L1 < best_L1) |
|
621 { |
|
622 best_L1 = L1; |
|
623 best_level = k+1; |
|
624 } |
|
625 } |
|
626 /*printf ("%d ", isTransient ? LM-best_level : best_level);*/ |
|
627 /* metric is in Q1 to be able to select the mid-point (-0.5) for narrower bands */ |
|
628 if (isTransient) |
|
629 metric[i] = 2*best_level; |
|
630 else |
|
631 metric[i] = -2*best_level; |
|
632 *tf_sum += (isTransient ? LM : 0) - metric[i]/2; |
|
633 /* For bands that can't be split to -1, set the metric to the half-way point to avoid |
|
634 biasing the decision */ |
|
635 if (narrow && (metric[i]==0 || metric[i]==-2*LM)) |
|
636 metric[i]-=1; |
|
637 /*printf("%d ", metric[i]);*/ |
|
638 } |
|
639 /*printf("\n");*/ |
|
640 /* Search for the optimal tf resolution, including tf_select */ |
|
641 tf_select = 0; |
|
642 for (sel=0;sel<2;sel++) |
|
643 { |
|
644 cost0 = 0; |
|
645 cost1 = isTransient ? 0 : lambda; |
|
646 for (i=1;i<len;i++) |
|
647 { |
|
648 int curr0, curr1; |
|
649 curr0 = IMIN(cost0, cost1 + lambda); |
|
650 curr1 = IMIN(cost0 + lambda, cost1); |
|
651 cost0 = curr0 + abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*sel+0]); |
|
652 cost1 = curr1 + abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*sel+1]); |
|
653 } |
|
654 cost0 = IMIN(cost0, cost1); |
|
655 selcost[sel]=cost0; |
|
656 } |
|
657 /* For now, we're conservative and only allow tf_select=1 for transients. |
|
658 * If tests confirm it's useful for non-transients, we could allow it. */ |
|
659 if (selcost[1]<selcost[0] && isTransient) |
|
660 tf_select=1; |
|
661 cost0 = 0; |
|
662 cost1 = isTransient ? 0 : lambda; |
|
663 /* Viterbi forward pass */ |
|
664 for (i=1;i<len;i++) |
|
665 { |
|
666 int curr0, curr1; |
|
667 int from0, from1; |
|
668 |
|
669 from0 = cost0; |
|
670 from1 = cost1 + lambda; |
|
671 if (from0 < from1) |
|
672 { |
|
673 curr0 = from0; |
|
674 path0[i]= 0; |
|
675 } else { |
|
676 curr0 = from1; |
|
677 path0[i]= 1; |
|
678 } |
|
679 |
|
680 from0 = cost0 + lambda; |
|
681 from1 = cost1; |
|
682 if (from0 < from1) |
|
683 { |
|
684 curr1 = from0; |
|
685 path1[i]= 0; |
|
686 } else { |
|
687 curr1 = from1; |
|
688 path1[i]= 1; |
|
689 } |
|
690 cost0 = curr0 + abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*tf_select+0]); |
|
691 cost1 = curr1 + abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*tf_select+1]); |
|
692 } |
|
693 tf_res[len-1] = cost0 < cost1 ? 0 : 1; |
|
694 /* Viterbi backward pass to check the decisions */ |
|
695 for (i=len-2;i>=0;i--) |
|
696 { |
|
697 if (tf_res[i+1] == 1) |
|
698 tf_res[i] = path1[i+1]; |
|
699 else |
|
700 tf_res[i] = path0[i+1]; |
|
701 } |
|
702 /*printf("%d %f\n", *tf_sum, tf_estimate);*/ |
|
703 RESTORE_STACK; |
|
704 #ifdef FUZZING |
|
705 tf_select = rand()&0x1; |
|
706 tf_res[0] = rand()&0x1; |
|
707 for (i=1;i<len;i++) |
|
708 tf_res[i] = tf_res[i-1] ^ ((rand()&0xF) == 0); |
|
709 #endif |
|
710 return tf_select; |
|
711 } |
|
712 |
|
713 static void tf_encode(int start, int end, int isTransient, int *tf_res, int LM, int tf_select, ec_enc *enc) |
|
714 { |
|
715 int curr, i; |
|
716 int tf_select_rsv; |
|
717 int tf_changed; |
|
718 int logp; |
|
719 opus_uint32 budget; |
|
720 opus_uint32 tell; |
|
721 budget = enc->storage*8; |
|
722 tell = ec_tell(enc); |
|
723 logp = isTransient ? 2 : 4; |
|
724 /* Reserve space to code the tf_select decision. */ |
|
725 tf_select_rsv = LM>0 && tell+logp+1 <= budget; |
|
726 budget -= tf_select_rsv; |
|
727 curr = tf_changed = 0; |
|
728 for (i=start;i<end;i++) |
|
729 { |
|
730 if (tell+logp<=budget) |
|
731 { |
|
732 ec_enc_bit_logp(enc, tf_res[i] ^ curr, logp); |
|
733 tell = ec_tell(enc); |
|
734 curr = tf_res[i]; |
|
735 tf_changed |= curr; |
|
736 } |
|
737 else |
|
738 tf_res[i] = curr; |
|
739 logp = isTransient ? 4 : 5; |
|
740 } |
|
741 /* Only code tf_select if it would actually make a difference. */ |
|
742 if (tf_select_rsv && |
|
743 tf_select_table[LM][4*isTransient+0+tf_changed]!= |
|
744 tf_select_table[LM][4*isTransient+2+tf_changed]) |
|
745 ec_enc_bit_logp(enc, tf_select, 1); |
|
746 else |
|
747 tf_select = 0; |
|
748 for (i=start;i<end;i++) |
|
749 tf_res[i] = tf_select_table[LM][4*isTransient+2*tf_select+tf_res[i]]; |
|
750 /*for(i=0;i<end;i++)printf("%d ", isTransient ? tf_res[i] : LM+tf_res[i]);printf("\n");*/ |
|
751 } |
|
752 |
|
753 |
|
754 static int alloc_trim_analysis(const CELTMode *m, const celt_norm *X, |
|
755 const opus_val16 *bandLogE, int end, int LM, int C, int N0, |
|
756 AnalysisInfo *analysis, opus_val16 *stereo_saving, opus_val16 tf_estimate, |
|
757 int intensity, opus_val16 surround_trim) |
|
758 { |
|
759 int i; |
|
760 opus_val32 diff=0; |
|
761 int c; |
|
762 int trim_index = 5; |
|
763 opus_val16 trim = QCONST16(5.f, 8); |
|
764 opus_val16 logXC, logXC2; |
|
765 if (C==2) |
|
766 { |
|
767 opus_val16 sum = 0; /* Q10 */ |
|
768 opus_val16 minXC; /* Q10 */ |
|
769 /* Compute inter-channel correlation for low frequencies */ |
|
770 for (i=0;i<8;i++) |
|
771 { |
|
772 int j; |
|
773 opus_val32 partial = 0; |
|
774 for (j=m->eBands[i]<<LM;j<m->eBands[i+1]<<LM;j++) |
|
775 partial = MAC16_16(partial, X[j], X[N0+j]); |
|
776 sum = ADD16(sum, EXTRACT16(SHR32(partial, 18))); |
|
777 } |
|
778 sum = MULT16_16_Q15(QCONST16(1.f/8, 15), sum); |
|
779 sum = MIN16(QCONST16(1.f, 10), ABS16(sum)); |
|
780 minXC = sum; |
|
781 for (i=8;i<intensity;i++) |
|
782 { |
|
783 int j; |
|
784 opus_val32 partial = 0; |
|
785 for (j=m->eBands[i]<<LM;j<m->eBands[i+1]<<LM;j++) |
|
786 partial = MAC16_16(partial, X[j], X[N0+j]); |
|
787 minXC = MIN16(minXC, ABS16(EXTRACT16(SHR32(partial, 18)))); |
|
788 } |
|
789 minXC = MIN16(QCONST16(1.f, 10), ABS16(minXC)); |
|
790 /*printf ("%f\n", sum);*/ |
|
791 if (sum > QCONST16(.995f,10)) |
|
792 trim_index-=4; |
|
793 else if (sum > QCONST16(.92f,10)) |
|
794 trim_index-=3; |
|
795 else if (sum > QCONST16(.85f,10)) |
|
796 trim_index-=2; |
|
797 else if (sum > QCONST16(.8f,10)) |
|
798 trim_index-=1; |
|
799 /* mid-side savings estimations based on the LF average*/ |
|
800 logXC = celt_log2(QCONST32(1.001f, 20)-MULT16_16(sum, sum)); |
|
801 /* mid-side savings estimations based on min correlation */ |
|
802 logXC2 = MAX16(HALF16(logXC), celt_log2(QCONST32(1.001f, 20)-MULT16_16(minXC, minXC))); |
|
803 #ifdef FIXED_POINT |
|
804 /* Compensate for Q20 vs Q14 input and convert output to Q8 */ |
|
805 logXC = PSHR32(logXC-QCONST16(6.f, DB_SHIFT),DB_SHIFT-8); |
|
806 logXC2 = PSHR32(logXC2-QCONST16(6.f, DB_SHIFT),DB_SHIFT-8); |
|
807 #endif |
|
808 |
|
809 trim += MAX16(-QCONST16(4.f, 8), MULT16_16_Q15(QCONST16(.75f,15),logXC)); |
|
810 *stereo_saving = MIN16(*stereo_saving + QCONST16(0.25f, 8), -HALF16(logXC2)); |
|
811 } |
|
812 |
|
813 /* Estimate spectral tilt */ |
|
814 c=0; do { |
|
815 for (i=0;i<end-1;i++) |
|
816 { |
|
817 diff += bandLogE[i+c*m->nbEBands]*(opus_int32)(2+2*i-end); |
|
818 } |
|
819 } while (++c<C); |
|
820 diff /= C*(end-1); |
|
821 /*printf("%f\n", diff);*/ |
|
822 if (diff > QCONST16(2.f, DB_SHIFT)) |
|
823 trim_index--; |
|
824 if (diff > QCONST16(8.f, DB_SHIFT)) |
|
825 trim_index--; |
|
826 if (diff < -QCONST16(4.f, DB_SHIFT)) |
|
827 trim_index++; |
|
828 if (diff < -QCONST16(10.f, DB_SHIFT)) |
|
829 trim_index++; |
|
830 trim -= MAX16(-QCONST16(2.f, 8), MIN16(QCONST16(2.f, 8), SHR16(diff+QCONST16(1.f, DB_SHIFT),DB_SHIFT-8)/6 )); |
|
831 trim -= SHR16(surround_trim, DB_SHIFT-8); |
|
832 trim -= 2*SHR16(tf_estimate, 14-8); |
|
833 #ifndef DISABLE_FLOAT_API |
|
834 if (analysis->valid) |
|
835 { |
|
836 trim -= MAX16(-QCONST16(2.f, 8), MIN16(QCONST16(2.f, 8), |
|
837 (opus_val16)(QCONST16(2.f, 8)*(analysis->tonality_slope+.05f)))); |
|
838 } |
|
839 #endif |
|
840 |
|
841 #ifdef FIXED_POINT |
|
842 trim_index = PSHR32(trim, 8); |
|
843 #else |
|
844 trim_index = (int)floor(.5f+trim); |
|
845 #endif |
|
846 if (trim_index<0) |
|
847 trim_index = 0; |
|
848 if (trim_index>10) |
|
849 trim_index = 10; |
|
850 /*printf("%d\n", trim_index);*/ |
|
851 #ifdef FUZZING |
|
852 trim_index = rand()%11; |
|
853 #endif |
|
854 return trim_index; |
|
855 } |
|
856 |
|
857 static int stereo_analysis(const CELTMode *m, const celt_norm *X, |
|
858 int LM, int N0) |
|
859 { |
|
860 int i; |
|
861 int thetas; |
|
862 opus_val32 sumLR = EPSILON, sumMS = EPSILON; |
|
863 |
|
864 /* Use the L1 norm to model the entropy of the L/R signal vs the M/S signal */ |
|
865 for (i=0;i<13;i++) |
|
866 { |
|
867 int j; |
|
868 for (j=m->eBands[i]<<LM;j<m->eBands[i+1]<<LM;j++) |
|
869 { |
|
870 opus_val32 L, R, M, S; |
|
871 /* We cast to 32-bit first because of the -32768 case */ |
|
872 L = EXTEND32(X[j]); |
|
873 R = EXTEND32(X[N0+j]); |
|
874 M = ADD32(L, R); |
|
875 S = SUB32(L, R); |
|
876 sumLR = ADD32(sumLR, ADD32(ABS32(L), ABS32(R))); |
|
877 sumMS = ADD32(sumMS, ADD32(ABS32(M), ABS32(S))); |
|
878 } |
|
879 } |
|
880 sumMS = MULT16_32_Q15(QCONST16(0.707107f, 15), sumMS); |
|
881 thetas = 13; |
|
882 /* We don't need thetas for lower bands with LM<=1 */ |
|
883 if (LM<=1) |
|
884 thetas -= 8; |
|
885 return MULT16_32_Q15((m->eBands[13]<<(LM+1))+thetas, sumMS) |
|
886 > MULT16_32_Q15(m->eBands[13]<<(LM+1), sumLR); |
|
887 } |
|
888 |
|
889 static opus_val16 dynalloc_analysis(const opus_val16 *bandLogE, const opus_val16 *bandLogE2, |
|
890 int nbEBands, int start, int end, int C, int *offsets, int lsb_depth, const opus_int16 *logN, |
|
891 int isTransient, int vbr, int constrained_vbr, const opus_int16 *eBands, int LM, |
|
892 int effectiveBytes, opus_int32 *tot_boost_, int lfe, opus_val16 *surround_dynalloc) |
|
893 { |
|
894 int i, c; |
|
895 opus_int32 tot_boost=0; |
|
896 opus_val16 maxDepth; |
|
897 VARDECL(opus_val16, follower); |
|
898 VARDECL(opus_val16, noise_floor); |
|
899 SAVE_STACK; |
|
900 ALLOC(follower, C*nbEBands, opus_val16); |
|
901 ALLOC(noise_floor, C*nbEBands, opus_val16); |
|
902 for (i=0;i<nbEBands;i++) |
|
903 offsets[i] = 0; |
|
904 /* Dynamic allocation code */ |
|
905 maxDepth=-QCONST16(31.9f, DB_SHIFT); |
|
906 for (i=0;i<end;i++) |
|
907 { |
|
908 /* Noise floor must take into account eMeans, the depth, the width of the bands |
|
909 and the preemphasis filter (approx. square of bark band ID) */ |
|
910 noise_floor[i] = MULT16_16(QCONST16(0.0625f, DB_SHIFT),logN[i]) |
|
911 +QCONST16(.5f,DB_SHIFT)+SHL16(9-lsb_depth,DB_SHIFT)-SHL16(eMeans[i],6) |
|
912 +MULT16_16(QCONST16(.0062,DB_SHIFT),(i+5)*(i+5)); |
|
913 } |
|
914 c=0;do |
|
915 { |
|
916 for (i=0;i<end;i++) |
|
917 maxDepth = MAX16(maxDepth, bandLogE[c*nbEBands+i]-noise_floor[i]); |
|
918 } while (++c<C); |
|
919 /* Make sure that dynamic allocation can't make us bust the budget */ |
|
920 if (effectiveBytes > 50 && LM>=1 && !lfe) |
|
921 { |
|
922 int last=0; |
|
923 c=0;do |
|
924 { |
|
925 follower[c*nbEBands] = bandLogE2[c*nbEBands]; |
|
926 for (i=1;i<end;i++) |
|
927 { |
|
928 /* The last band to be at least 3 dB higher than the previous one |
|
929 is the last we'll consider. Otherwise, we run into problems on |
|
930 bandlimited signals. */ |
|
931 if (bandLogE2[c*nbEBands+i] > bandLogE2[c*nbEBands+i-1]+QCONST16(.5f,DB_SHIFT)) |
|
932 last=i; |
|
933 follower[c*nbEBands+i] = MIN16(follower[c*nbEBands+i-1]+QCONST16(1.5f,DB_SHIFT), bandLogE2[c*nbEBands+i]); |
|
934 } |
|
935 for (i=last-1;i>=0;i--) |
|
936 follower[c*nbEBands+i] = MIN16(follower[c*nbEBands+i], MIN16(follower[c*nbEBands+i+1]+QCONST16(2.f,DB_SHIFT), bandLogE2[c*nbEBands+i])); |
|
937 for (i=0;i<end;i++) |
|
938 follower[c*nbEBands+i] = MAX16(follower[c*nbEBands+i], noise_floor[i]); |
|
939 } while (++c<C); |
|
940 if (C==2) |
|
941 { |
|
942 for (i=start;i<end;i++) |
|
943 { |
|
944 /* Consider 24 dB "cross-talk" */ |
|
945 follower[nbEBands+i] = MAX16(follower[nbEBands+i], follower[ i]-QCONST16(4.f,DB_SHIFT)); |
|
946 follower[ i] = MAX16(follower[ i], follower[nbEBands+i]-QCONST16(4.f,DB_SHIFT)); |
|
947 follower[i] = HALF16(MAX16(0, bandLogE[i]-follower[i]) + MAX16(0, bandLogE[nbEBands+i]-follower[nbEBands+i])); |
|
948 } |
|
949 } else { |
|
950 for (i=start;i<end;i++) |
|
951 { |
|
952 follower[i] = MAX16(0, bandLogE[i]-follower[i]); |
|
953 } |
|
954 } |
|
955 for (i=start;i<end;i++) |
|
956 follower[i] = MAX16(follower[i], surround_dynalloc[i]); |
|
957 /* For non-transient CBR/CVBR frames, halve the dynalloc contribution */ |
|
958 if ((!vbr || constrained_vbr)&&!isTransient) |
|
959 { |
|
960 for (i=start;i<end;i++) |
|
961 follower[i] = HALF16(follower[i]); |
|
962 } |
|
963 for (i=start;i<end;i++) |
|
964 { |
|
965 int width; |
|
966 int boost; |
|
967 int boost_bits; |
|
968 |
|
969 if (i<8) |
|
970 follower[i] *= 2; |
|
971 if (i>=12) |
|
972 follower[i] = HALF16(follower[i]); |
|
973 follower[i] = MIN16(follower[i], QCONST16(4, DB_SHIFT)); |
|
974 |
|
975 width = C*(eBands[i+1]-eBands[i])<<LM; |
|
976 if (width<6) |
|
977 { |
|
978 boost = (int)SHR32(EXTEND32(follower[i]),DB_SHIFT); |
|
979 boost_bits = boost*width<<BITRES; |
|
980 } else if (width > 48) { |
|
981 boost = (int)SHR32(EXTEND32(follower[i])*8,DB_SHIFT); |
|
982 boost_bits = (boost*width<<BITRES)/8; |
|
983 } else { |
|
984 boost = (int)SHR32(EXTEND32(follower[i])*width/6,DB_SHIFT); |
|
985 boost_bits = boost*6<<BITRES; |
|
986 } |
|
987 /* For CBR and non-transient CVBR frames, limit dynalloc to 1/4 of the bits */ |
|
988 if ((!vbr || (constrained_vbr&&!isTransient)) |
|
989 && (tot_boost+boost_bits)>>BITRES>>3 > effectiveBytes/4) |
|
990 { |
|
991 opus_int32 cap = ((effectiveBytes/4)<<BITRES<<3); |
|
992 offsets[i] = cap-tot_boost; |
|
993 tot_boost = cap; |
|
994 break; |
|
995 } else { |
|
996 offsets[i] = boost; |
|
997 tot_boost += boost_bits; |
|
998 } |
|
999 } |
|
1000 } |
|
1001 *tot_boost_ = tot_boost; |
|
1002 RESTORE_STACK; |
|
1003 return maxDepth; |
|
1004 } |
|
1005 |
|
1006 |
|
1007 static int run_prefilter(CELTEncoder *st, celt_sig *in, celt_sig *prefilter_mem, int CC, int N, |
|
1008 int prefilter_tapset, int *pitch, opus_val16 *gain, int *qgain, int enabled, int nbAvailableBytes) |
|
1009 { |
|
1010 int c; |
|
1011 VARDECL(celt_sig, _pre); |
|
1012 celt_sig *pre[2]; |
|
1013 const CELTMode *mode; |
|
1014 int pitch_index; |
|
1015 opus_val16 gain1; |
|
1016 opus_val16 pf_threshold; |
|
1017 int pf_on; |
|
1018 int qg; |
|
1019 SAVE_STACK; |
|
1020 |
|
1021 mode = st->mode; |
|
1022 ALLOC(_pre, CC*(N+COMBFILTER_MAXPERIOD), celt_sig); |
|
1023 |
|
1024 pre[0] = _pre; |
|
1025 pre[1] = _pre + (N+COMBFILTER_MAXPERIOD); |
|
1026 |
|
1027 |
|
1028 c=0; do { |
|
1029 OPUS_COPY(pre[c], prefilter_mem+c*COMBFILTER_MAXPERIOD, COMBFILTER_MAXPERIOD); |
|
1030 OPUS_COPY(pre[c]+COMBFILTER_MAXPERIOD, in+c*(N+st->overlap)+st->overlap, N); |
|
1031 } while (++c<CC); |
|
1032 |
|
1033 if (enabled) |
|
1034 { |
|
1035 VARDECL(opus_val16, pitch_buf); |
|
1036 ALLOC(pitch_buf, (COMBFILTER_MAXPERIOD+N)>>1, opus_val16); |
|
1037 |
|
1038 pitch_downsample(pre, pitch_buf, COMBFILTER_MAXPERIOD+N, CC, st->arch); |
|
1039 /* Don't search for the fir last 1.5 octave of the range because |
|
1040 there's too many false-positives due to short-term correlation */ |
|
1041 pitch_search(pitch_buf+(COMBFILTER_MAXPERIOD>>1), pitch_buf, N, |
|
1042 COMBFILTER_MAXPERIOD-3*COMBFILTER_MINPERIOD, &pitch_index, |
|
1043 st->arch); |
|
1044 pitch_index = COMBFILTER_MAXPERIOD-pitch_index; |
|
1045 |
|
1046 gain1 = remove_doubling(pitch_buf, COMBFILTER_MAXPERIOD, COMBFILTER_MINPERIOD, |
|
1047 N, &pitch_index, st->prefilter_period, st->prefilter_gain); |
|
1048 if (pitch_index > COMBFILTER_MAXPERIOD-2) |
|
1049 pitch_index = COMBFILTER_MAXPERIOD-2; |
|
1050 gain1 = MULT16_16_Q15(QCONST16(.7f,15),gain1); |
|
1051 /*printf("%d %d %f %f\n", pitch_change, pitch_index, gain1, st->analysis.tonality);*/ |
|
1052 if (st->loss_rate>2) |
|
1053 gain1 = HALF32(gain1); |
|
1054 if (st->loss_rate>4) |
|
1055 gain1 = HALF32(gain1); |
|
1056 if (st->loss_rate>8) |
|
1057 gain1 = 0; |
|
1058 } else { |
|
1059 gain1 = 0; |
|
1060 pitch_index = COMBFILTER_MINPERIOD; |
|
1061 } |
|
1062 |
|
1063 /* Gain threshold for enabling the prefilter/postfilter */ |
|
1064 pf_threshold = QCONST16(.2f,15); |
|
1065 |
|
1066 /* Adjusting the threshold based on rate and continuity */ |
|
1067 if (abs(pitch_index-st->prefilter_period)*10>pitch_index) |
|
1068 pf_threshold += QCONST16(.2f,15); |
|
1069 if (nbAvailableBytes<25) |
|
1070 pf_threshold += QCONST16(.1f,15); |
|
1071 if (nbAvailableBytes<35) |
|
1072 pf_threshold += QCONST16(.1f,15); |
|
1073 if (st->prefilter_gain > QCONST16(.4f,15)) |
|
1074 pf_threshold -= QCONST16(.1f,15); |
|
1075 if (st->prefilter_gain > QCONST16(.55f,15)) |
|
1076 pf_threshold -= QCONST16(.1f,15); |
|
1077 |
|
1078 /* Hard threshold at 0.2 */ |
|
1079 pf_threshold = MAX16(pf_threshold, QCONST16(.2f,15)); |
|
1080 if (gain1<pf_threshold) |
|
1081 { |
|
1082 gain1 = 0; |
|
1083 pf_on = 0; |
|
1084 qg = 0; |
|
1085 } else { |
|
1086 /*This block is not gated by a total bits check only because |
|
1087 of the nbAvailableBytes check above.*/ |
|
1088 if (ABS16(gain1-st->prefilter_gain)<QCONST16(.1f,15)) |
|
1089 gain1=st->prefilter_gain; |
|
1090 |
|
1091 #ifdef FIXED_POINT |
|
1092 qg = ((gain1+1536)>>10)/3-1; |
|
1093 #else |
|
1094 qg = (int)floor(.5f+gain1*32/3)-1; |
|
1095 #endif |
|
1096 qg = IMAX(0, IMIN(7, qg)); |
|
1097 gain1 = QCONST16(0.09375f,15)*(qg+1); |
|
1098 pf_on = 1; |
|
1099 } |
|
1100 /*printf("%d %f\n", pitch_index, gain1);*/ |
|
1101 |
|
1102 c=0; do { |
|
1103 int offset = mode->shortMdctSize-st->overlap; |
|
1104 st->prefilter_period=IMAX(st->prefilter_period, COMBFILTER_MINPERIOD); |
|
1105 OPUS_COPY(in+c*(N+st->overlap), st->in_mem+c*(st->overlap), st->overlap); |
|
1106 if (offset) |
|
1107 comb_filter(in+c*(N+st->overlap)+st->overlap, pre[c]+COMBFILTER_MAXPERIOD, |
|
1108 st->prefilter_period, st->prefilter_period, offset, -st->prefilter_gain, -st->prefilter_gain, |
|
1109 st->prefilter_tapset, st->prefilter_tapset, NULL, 0); |
|
1110 |
|
1111 comb_filter(in+c*(N+st->overlap)+st->overlap+offset, pre[c]+COMBFILTER_MAXPERIOD+offset, |
|
1112 st->prefilter_period, pitch_index, N-offset, -st->prefilter_gain, -gain1, |
|
1113 st->prefilter_tapset, prefilter_tapset, mode->window, st->overlap); |
|
1114 OPUS_COPY(st->in_mem+c*(st->overlap), in+c*(N+st->overlap)+N, st->overlap); |
|
1115 |
|
1116 if (N>COMBFILTER_MAXPERIOD) |
|
1117 { |
|
1118 OPUS_MOVE(prefilter_mem+c*COMBFILTER_MAXPERIOD, pre[c]+N, COMBFILTER_MAXPERIOD); |
|
1119 } else { |
|
1120 OPUS_MOVE(prefilter_mem+c*COMBFILTER_MAXPERIOD, prefilter_mem+c*COMBFILTER_MAXPERIOD+N, COMBFILTER_MAXPERIOD-N); |
|
1121 OPUS_MOVE(prefilter_mem+c*COMBFILTER_MAXPERIOD+COMBFILTER_MAXPERIOD-N, pre[c]+COMBFILTER_MAXPERIOD, N); |
|
1122 } |
|
1123 } while (++c<CC); |
|
1124 |
|
1125 RESTORE_STACK; |
|
1126 *gain = gain1; |
|
1127 *pitch = pitch_index; |
|
1128 *qgain = qg; |
|
1129 return pf_on; |
|
1130 } |
|
1131 |
|
1132 static int compute_vbr(const CELTMode *mode, AnalysisInfo *analysis, opus_int32 base_target, |
|
1133 int LM, opus_int32 bitrate, int lastCodedBands, int C, int intensity, |
|
1134 int constrained_vbr, opus_val16 stereo_saving, int tot_boost, |
|
1135 opus_val16 tf_estimate, int pitch_change, opus_val16 maxDepth, |
|
1136 int variable_duration, int lfe, int has_surround_mask, opus_val16 surround_masking, |
|
1137 opus_val16 temporal_vbr) |
|
1138 { |
|
1139 /* The target rate in 8th bits per frame */ |
|
1140 opus_int32 target; |
|
1141 int coded_bins; |
|
1142 int coded_bands; |
|
1143 opus_val16 tf_calibration; |
|
1144 int nbEBands; |
|
1145 const opus_int16 *eBands; |
|
1146 |
|
1147 nbEBands = mode->nbEBands; |
|
1148 eBands = mode->eBands; |
|
1149 |
|
1150 coded_bands = lastCodedBands ? lastCodedBands : nbEBands; |
|
1151 coded_bins = eBands[coded_bands]<<LM; |
|
1152 if (C==2) |
|
1153 coded_bins += eBands[IMIN(intensity, coded_bands)]<<LM; |
|
1154 |
|
1155 target = base_target; |
|
1156 |
|
1157 /*printf("%f %f %f %f %d %d ", st->analysis.activity, st->analysis.tonality, tf_estimate, st->stereo_saving, tot_boost, coded_bands);*/ |
|
1158 #ifndef DISABLE_FLOAT_API |
|
1159 if (analysis->valid && analysis->activity<.4) |
|
1160 target -= (opus_int32)((coded_bins<<BITRES)*(.4f-analysis->activity)); |
|
1161 #endif |
|
1162 /* Stereo savings */ |
|
1163 if (C==2) |
|
1164 { |
|
1165 int coded_stereo_bands; |
|
1166 int coded_stereo_dof; |
|
1167 opus_val16 max_frac; |
|
1168 coded_stereo_bands = IMIN(intensity, coded_bands); |
|
1169 coded_stereo_dof = (eBands[coded_stereo_bands]<<LM)-coded_stereo_bands; |
|
1170 /* Maximum fraction of the bits we can save if the signal is mono. */ |
|
1171 max_frac = DIV32_16(MULT16_16(QCONST16(0.8f, 15), coded_stereo_dof), coded_bins); |
|
1172 stereo_saving = MIN16(stereo_saving, QCONST16(1.f, 8)); |
|
1173 /*printf("%d %d %d ", coded_stereo_dof, coded_bins, tot_boost);*/ |
|
1174 target -= (opus_int32)MIN32(MULT16_32_Q15(max_frac,target), |
|
1175 SHR32(MULT16_16(stereo_saving-QCONST16(0.1f,8),(coded_stereo_dof<<BITRES)),8)); |
|
1176 } |
|
1177 /* Boost the rate according to dynalloc (minus the dynalloc average for calibration). */ |
|
1178 target += tot_boost-(16<<LM); |
|
1179 /* Apply transient boost, compensating for average boost. */ |
|
1180 tf_calibration = variable_duration==OPUS_FRAMESIZE_VARIABLE ? |
|
1181 QCONST16(0.02f,14) : QCONST16(0.04f,14); |
|
1182 target += (opus_int32)SHL32(MULT16_32_Q15(tf_estimate-tf_calibration, target),1); |
|
1183 |
|
1184 #ifndef DISABLE_FLOAT_API |
|
1185 /* Apply tonality boost */ |
|
1186 if (analysis->valid && !lfe) |
|
1187 { |
|
1188 opus_int32 tonal_target; |
|
1189 float tonal; |
|
1190 |
|
1191 /* Tonality boost (compensating for the average). */ |
|
1192 tonal = MAX16(0.f,analysis->tonality-.15f)-0.09f; |
|
1193 tonal_target = target + (opus_int32)((coded_bins<<BITRES)*1.2f*tonal); |
|
1194 if (pitch_change) |
|
1195 tonal_target += (opus_int32)((coded_bins<<BITRES)*.8f); |
|
1196 /*printf("%f %f ", analysis->tonality, tonal);*/ |
|
1197 target = tonal_target; |
|
1198 } |
|
1199 #endif |
|
1200 |
|
1201 if (has_surround_mask&&!lfe) |
|
1202 { |
|
1203 opus_int32 surround_target = target + (opus_int32)SHR32(MULT16_16(surround_masking,coded_bins<<BITRES), DB_SHIFT); |
|
1204 /*printf("%f %d %d %d %d %d %d ", surround_masking, coded_bins, st->end, st->intensity, surround_target, target, st->bitrate);*/ |
|
1205 target = IMAX(target/4, surround_target); |
|
1206 } |
|
1207 |
|
1208 { |
|
1209 opus_int32 floor_depth; |
|
1210 int bins; |
|
1211 bins = eBands[nbEBands-2]<<LM; |
|
1212 /*floor_depth = SHR32(MULT16_16((C*bins<<BITRES),celt_log2(SHL32(MAX16(1,sample_max),13))), DB_SHIFT);*/ |
|
1213 floor_depth = (opus_int32)SHR32(MULT16_16((C*bins<<BITRES),maxDepth), DB_SHIFT); |
|
1214 floor_depth = IMAX(floor_depth, target>>2); |
|
1215 target = IMIN(target, floor_depth); |
|
1216 /*printf("%f %d\n", maxDepth, floor_depth);*/ |
|
1217 } |
|
1218 |
|
1219 if ((!has_surround_mask||lfe) && (constrained_vbr || bitrate<64000)) |
|
1220 { |
|
1221 opus_val16 rate_factor; |
|
1222 #ifdef FIXED_POINT |
|
1223 rate_factor = MAX16(0,(bitrate-32000)); |
|
1224 #else |
|
1225 rate_factor = MAX16(0,(1.f/32768)*(bitrate-32000)); |
|
1226 #endif |
|
1227 if (constrained_vbr) |
|
1228 rate_factor = MIN16(rate_factor, QCONST16(0.67f, 15)); |
|
1229 target = base_target + (opus_int32)MULT16_32_Q15(rate_factor, target-base_target); |
|
1230 |
|
1231 } |
|
1232 |
|
1233 if (!has_surround_mask && tf_estimate < QCONST16(.2f, 14)) |
|
1234 { |
|
1235 opus_val16 amount; |
|
1236 opus_val16 tvbr_factor; |
|
1237 amount = MULT16_16_Q15(QCONST16(.0000031f, 30), IMAX(0, IMIN(32000, 96000-bitrate))); |
|
1238 tvbr_factor = SHR32(MULT16_16(temporal_vbr, amount), DB_SHIFT); |
|
1239 target += (opus_int32)MULT16_32_Q15(tvbr_factor, target); |
|
1240 } |
|
1241 |
|
1242 /* Don't allow more than doubling the rate */ |
|
1243 target = IMIN(2*base_target, target); |
|
1244 |
|
1245 return target; |
|
1246 } |
|
1247 |
|
1248 int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm, int frame_size, unsigned char *compressed, int nbCompressedBytes, ec_enc *enc) |
|
1249 { |
|
1250 int i, c, N; |
|
1251 opus_int32 bits; |
|
1252 ec_enc _enc; |
|
1253 VARDECL(celt_sig, in); |
|
1254 VARDECL(celt_sig, freq); |
|
1255 VARDECL(celt_norm, X); |
|
1256 VARDECL(celt_ener, bandE); |
|
1257 VARDECL(opus_val16, bandLogE); |
|
1258 VARDECL(opus_val16, bandLogE2); |
|
1259 VARDECL(int, fine_quant); |
|
1260 VARDECL(opus_val16, error); |
|
1261 VARDECL(int, pulses); |
|
1262 VARDECL(int, cap); |
|
1263 VARDECL(int, offsets); |
|
1264 VARDECL(int, fine_priority); |
|
1265 VARDECL(int, tf_res); |
|
1266 VARDECL(unsigned char, collapse_masks); |
|
1267 celt_sig *prefilter_mem; |
|
1268 opus_val16 *oldBandE, *oldLogE, *oldLogE2; |
|
1269 int shortBlocks=0; |
|
1270 int isTransient=0; |
|
1271 const int CC = st->channels; |
|
1272 const int C = st->stream_channels; |
|
1273 int LM, M; |
|
1274 int tf_select; |
|
1275 int nbFilledBytes, nbAvailableBytes; |
|
1276 int effEnd; |
|
1277 int codedBands; |
|
1278 int tf_sum; |
|
1279 int alloc_trim; |
|
1280 int pitch_index=COMBFILTER_MINPERIOD; |
|
1281 opus_val16 gain1 = 0; |
|
1282 int dual_stereo=0; |
|
1283 int effectiveBytes; |
|
1284 int dynalloc_logp; |
|
1285 opus_int32 vbr_rate; |
|
1286 opus_int32 total_bits; |
|
1287 opus_int32 total_boost; |
|
1288 opus_int32 balance; |
|
1289 opus_int32 tell; |
|
1290 int prefilter_tapset=0; |
|
1291 int pf_on; |
|
1292 int anti_collapse_rsv; |
|
1293 int anti_collapse_on=0; |
|
1294 int silence=0; |
|
1295 int tf_chan = 0; |
|
1296 opus_val16 tf_estimate; |
|
1297 int pitch_change=0; |
|
1298 opus_int32 tot_boost; |
|
1299 opus_val32 sample_max; |
|
1300 opus_val16 maxDepth; |
|
1301 const OpusCustomMode *mode; |
|
1302 int nbEBands; |
|
1303 int overlap; |
|
1304 const opus_int16 *eBands; |
|
1305 int secondMdct; |
|
1306 int signalBandwidth; |
|
1307 int transient_got_disabled=0; |
|
1308 opus_val16 surround_masking=0; |
|
1309 opus_val16 temporal_vbr=0; |
|
1310 opus_val16 surround_trim = 0; |
|
1311 opus_int32 equiv_rate = 510000; |
|
1312 VARDECL(opus_val16, surround_dynalloc); |
|
1313 ALLOC_STACK; |
|
1314 |
|
1315 mode = st->mode; |
|
1316 nbEBands = mode->nbEBands; |
|
1317 overlap = mode->overlap; |
|
1318 eBands = mode->eBands; |
|
1319 tf_estimate = 0; |
|
1320 if (nbCompressedBytes<2 || pcm==NULL) |
|
1321 { |
|
1322 RESTORE_STACK; |
|
1323 return OPUS_BAD_ARG; |
|
1324 } |
|
1325 |
|
1326 frame_size *= st->upsample; |
|
1327 for (LM=0;LM<=mode->maxLM;LM++) |
|
1328 if (mode->shortMdctSize<<LM==frame_size) |
|
1329 break; |
|
1330 if (LM>mode->maxLM) |
|
1331 { |
|
1332 RESTORE_STACK; |
|
1333 return OPUS_BAD_ARG; |
|
1334 } |
|
1335 M=1<<LM; |
|
1336 N = M*mode->shortMdctSize; |
|
1337 |
|
1338 prefilter_mem = st->in_mem+CC*(st->overlap); |
|
1339 oldBandE = (opus_val16*)(st->in_mem+CC*(st->overlap+COMBFILTER_MAXPERIOD)); |
|
1340 oldLogE = oldBandE + CC*nbEBands; |
|
1341 oldLogE2 = oldLogE + CC*nbEBands; |
|
1342 |
|
1343 if (enc==NULL) |
|
1344 { |
|
1345 tell=1; |
|
1346 nbFilledBytes=0; |
|
1347 } else { |
|
1348 tell=ec_tell(enc); |
|
1349 nbFilledBytes=(tell+4)>>3; |
|
1350 } |
|
1351 |
|
1352 #ifdef CUSTOM_MODES |
|
1353 if (st->signalling && enc==NULL) |
|
1354 { |
|
1355 int tmp = (mode->effEBands-st->end)>>1; |
|
1356 st->end = IMAX(1, mode->effEBands-tmp); |
|
1357 compressed[0] = tmp<<5; |
|
1358 compressed[0] |= LM<<3; |
|
1359 compressed[0] |= (C==2)<<2; |
|
1360 /* Convert "standard mode" to Opus header */ |
|
1361 if (mode->Fs==48000 && mode->shortMdctSize==120) |
|
1362 { |
|
1363 int c0 = toOpus(compressed[0]); |
|
1364 if (c0<0) |
|
1365 { |
|
1366 RESTORE_STACK; |
|
1367 return OPUS_BAD_ARG; |
|
1368 } |
|
1369 compressed[0] = c0; |
|
1370 } |
|
1371 compressed++; |
|
1372 nbCompressedBytes--; |
|
1373 } |
|
1374 #else |
|
1375 celt_assert(st->signalling==0); |
|
1376 #endif |
|
1377 |
|
1378 /* Can't produce more than 1275 output bytes */ |
|
1379 nbCompressedBytes = IMIN(nbCompressedBytes,1275); |
|
1380 nbAvailableBytes = nbCompressedBytes - nbFilledBytes; |
|
1381 |
|
1382 if (st->vbr && st->bitrate!=OPUS_BITRATE_MAX) |
|
1383 { |
|
1384 opus_int32 den=mode->Fs>>BITRES; |
|
1385 vbr_rate=(st->bitrate*frame_size+(den>>1))/den; |
|
1386 #ifdef CUSTOM_MODES |
|
1387 if (st->signalling) |
|
1388 vbr_rate -= 8<<BITRES; |
|
1389 #endif |
|
1390 effectiveBytes = vbr_rate>>(3+BITRES); |
|
1391 } else { |
|
1392 opus_int32 tmp; |
|
1393 vbr_rate = 0; |
|
1394 tmp = st->bitrate*frame_size; |
|
1395 if (tell>1) |
|
1396 tmp += tell; |
|
1397 if (st->bitrate!=OPUS_BITRATE_MAX) |
|
1398 nbCompressedBytes = IMAX(2, IMIN(nbCompressedBytes, |
|
1399 (tmp+4*mode->Fs)/(8*mode->Fs)-!!st->signalling)); |
|
1400 effectiveBytes = nbCompressedBytes; |
|
1401 } |
|
1402 if (st->bitrate != OPUS_BITRATE_MAX) |
|
1403 equiv_rate = st->bitrate - (40*C+20)*((400>>LM) - 50); |
|
1404 |
|
1405 if (enc==NULL) |
|
1406 { |
|
1407 ec_enc_init(&_enc, compressed, nbCompressedBytes); |
|
1408 enc = &_enc; |
|
1409 } |
|
1410 |
|
1411 if (vbr_rate>0) |
|
1412 { |
|
1413 /* Computes the max bit-rate allowed in VBR mode to avoid violating the |
|
1414 target rate and buffering. |
|
1415 We must do this up front so that bust-prevention logic triggers |
|
1416 correctly if we don't have enough bits. */ |
|
1417 if (st->constrained_vbr) |
|
1418 { |
|
1419 opus_int32 vbr_bound; |
|
1420 opus_int32 max_allowed; |
|
1421 /* We could use any multiple of vbr_rate as bound (depending on the |
|
1422 delay). |
|
1423 This is clamped to ensure we use at least two bytes if the encoder |
|
1424 was entirely empty, but to allow 0 in hybrid mode. */ |
|
1425 vbr_bound = vbr_rate; |
|
1426 max_allowed = IMIN(IMAX(tell==1?2:0, |
|
1427 (vbr_rate+vbr_bound-st->vbr_reservoir)>>(BITRES+3)), |
|
1428 nbAvailableBytes); |
|
1429 if(max_allowed < nbAvailableBytes) |
|
1430 { |
|
1431 nbCompressedBytes = nbFilledBytes+max_allowed; |
|
1432 nbAvailableBytes = max_allowed; |
|
1433 ec_enc_shrink(enc, nbCompressedBytes); |
|
1434 } |
|
1435 } |
|
1436 } |
|
1437 total_bits = nbCompressedBytes*8; |
|
1438 |
|
1439 effEnd = st->end; |
|
1440 if (effEnd > mode->effEBands) |
|
1441 effEnd = mode->effEBands; |
|
1442 |
|
1443 ALLOC(in, CC*(N+st->overlap), celt_sig); |
|
1444 |
|
1445 sample_max=MAX32(st->overlap_max, celt_maxabs16(pcm, C*(N-overlap)/st->upsample)); |
|
1446 st->overlap_max=celt_maxabs16(pcm+C*(N-overlap)/st->upsample, C*overlap/st->upsample); |
|
1447 sample_max=MAX32(sample_max, st->overlap_max); |
|
1448 #ifdef FIXED_POINT |
|
1449 silence = (sample_max==0); |
|
1450 #else |
|
1451 silence = (sample_max <= (opus_val16)1/(1<<st->lsb_depth)); |
|
1452 #endif |
|
1453 #ifdef FUZZING |
|
1454 if ((rand()&0x3F)==0) |
|
1455 silence = 1; |
|
1456 #endif |
|
1457 if (tell==1) |
|
1458 ec_enc_bit_logp(enc, silence, 15); |
|
1459 else |
|
1460 silence=0; |
|
1461 if (silence) |
|
1462 { |
|
1463 /*In VBR mode there is no need to send more than the minimum. */ |
|
1464 if (vbr_rate>0) |
|
1465 { |
|
1466 effectiveBytes=nbCompressedBytes=IMIN(nbCompressedBytes, nbFilledBytes+2); |
|
1467 total_bits=nbCompressedBytes*8; |
|
1468 nbAvailableBytes=2; |
|
1469 ec_enc_shrink(enc, nbCompressedBytes); |
|
1470 } |
|
1471 /* Pretend we've filled all the remaining bits with zeros |
|
1472 (that's what the initialiser did anyway) */ |
|
1473 tell = nbCompressedBytes*8; |
|
1474 enc->nbits_total+=tell-ec_tell(enc); |
|
1475 } |
|
1476 c=0; do { |
|
1477 celt_preemphasis(pcm+c, in+c*(N+st->overlap)+st->overlap, N, CC, st->upsample, |
|
1478 mode->preemph, st->preemph_memE+c, st->clip); |
|
1479 } while (++c<CC); |
|
1480 |
|
1481 |
|
1482 |
|
1483 /* Find pitch period and gain */ |
|
1484 { |
|
1485 int enabled; |
|
1486 int qg; |
|
1487 enabled = ((st->lfe&&nbAvailableBytes>3) || nbAvailableBytes>12*C) && st->start==0 && !silence && !st->disable_pf |
|
1488 && st->complexity >= 5 && !(st->consec_transient && LM!=3 && st->variable_duration==OPUS_FRAMESIZE_VARIABLE); |
|
1489 |
|
1490 prefilter_tapset = st->tapset_decision; |
|
1491 pf_on = run_prefilter(st, in, prefilter_mem, CC, N, prefilter_tapset, &pitch_index, &gain1, &qg, enabled, nbAvailableBytes); |
|
1492 if ((gain1 > QCONST16(.4f,15) || st->prefilter_gain > QCONST16(.4f,15)) && (!st->analysis.valid || st->analysis.tonality > .3) |
|
1493 && (pitch_index > 1.26*st->prefilter_period || pitch_index < .79*st->prefilter_period)) |
|
1494 pitch_change = 1; |
|
1495 if (pf_on==0) |
|
1496 { |
|
1497 if(st->start==0 && tell+16<=total_bits) |
|
1498 ec_enc_bit_logp(enc, 0, 1); |
|
1499 } else { |
|
1500 /*This block is not gated by a total bits check only because |
|
1501 of the nbAvailableBytes check above.*/ |
|
1502 int octave; |
|
1503 ec_enc_bit_logp(enc, 1, 1); |
|
1504 pitch_index += 1; |
|
1505 octave = EC_ILOG(pitch_index)-5; |
|
1506 ec_enc_uint(enc, octave, 6); |
|
1507 ec_enc_bits(enc, pitch_index-(16<<octave), 4+octave); |
|
1508 pitch_index -= 1; |
|
1509 ec_enc_bits(enc, qg, 3); |
|
1510 ec_enc_icdf(enc, prefilter_tapset, tapset_icdf, 2); |
|
1511 } |
|
1512 } |
|
1513 |
|
1514 isTransient = 0; |
|
1515 shortBlocks = 0; |
|
1516 if (st->complexity >= 1 && !st->lfe) |
|
1517 { |
|
1518 isTransient = transient_analysis(in, N+st->overlap, CC, |
|
1519 &tf_estimate, &tf_chan); |
|
1520 } |
|
1521 if (LM>0 && ec_tell(enc)+3<=total_bits) |
|
1522 { |
|
1523 if (isTransient) |
|
1524 shortBlocks = M; |
|
1525 } else { |
|
1526 isTransient = 0; |
|
1527 transient_got_disabled=1; |
|
1528 } |
|
1529 |
|
1530 ALLOC(freq, CC*N, celt_sig); /**< Interleaved signal MDCTs */ |
|
1531 ALLOC(bandE,nbEBands*CC, celt_ener); |
|
1532 ALLOC(bandLogE,nbEBands*CC, opus_val16); |
|
1533 |
|
1534 secondMdct = shortBlocks && st->complexity>=8; |
|
1535 ALLOC(bandLogE2, C*nbEBands, opus_val16); |
|
1536 if (secondMdct) |
|
1537 { |
|
1538 compute_mdcts(mode, 0, in, freq, C, CC, LM, st->upsample); |
|
1539 compute_band_energies(mode, freq, bandE, effEnd, C, M); |
|
1540 amp2Log2(mode, effEnd, st->end, bandE, bandLogE2, C); |
|
1541 for (i=0;i<C*nbEBands;i++) |
|
1542 bandLogE2[i] += HALF16(SHL16(LM, DB_SHIFT)); |
|
1543 } |
|
1544 |
|
1545 compute_mdcts(mode, shortBlocks, in, freq, C, CC, LM, st->upsample); |
|
1546 if (CC==2&&C==1) |
|
1547 tf_chan = 0; |
|
1548 compute_band_energies(mode, freq, bandE, effEnd, C, M); |
|
1549 |
|
1550 if (st->lfe) |
|
1551 { |
|
1552 for (i=2;i<st->end;i++) |
|
1553 { |
|
1554 bandE[i] = IMIN(bandE[i], MULT16_32_Q15(QCONST16(1e-4f,15),bandE[0])); |
|
1555 bandE[i] = MAX32(bandE[i], EPSILON); |
|
1556 } |
|
1557 } |
|
1558 amp2Log2(mode, effEnd, st->end, bandE, bandLogE, C); |
|
1559 |
|
1560 ALLOC(surround_dynalloc, C*nbEBands, opus_val16); |
|
1561 for(i=0;i<st->end;i++) |
|
1562 surround_dynalloc[i] = 0; |
|
1563 /* This computes how much masking takes place between surround channels */ |
|
1564 if (st->start==0&&st->energy_mask&&!st->lfe) |
|
1565 { |
|
1566 int mask_end; |
|
1567 int midband; |
|
1568 int count_dynalloc; |
|
1569 opus_val32 mask_avg=0; |
|
1570 opus_val32 diff=0; |
|
1571 int count=0; |
|
1572 mask_end = IMAX(2,st->lastCodedBands); |
|
1573 for (c=0;c<C;c++) |
|
1574 { |
|
1575 for(i=0;i<mask_end;i++) |
|
1576 { |
|
1577 opus_val16 mask; |
|
1578 mask = MAX16(MIN16(st->energy_mask[nbEBands*c+i], |
|
1579 QCONST16(.25f, DB_SHIFT)), -QCONST16(2.0f, DB_SHIFT)); |
|
1580 if (mask > 0) |
|
1581 mask = HALF16(mask); |
|
1582 mask_avg += MULT16_16(mask, eBands[i+1]-eBands[i]); |
|
1583 count += eBands[i+1]-eBands[i]; |
|
1584 diff += MULT16_16(mask, 1+2*i-mask_end); |
|
1585 } |
|
1586 } |
|
1587 mask_avg = DIV32_16(mask_avg,count); |
|
1588 mask_avg += QCONST16(.2f, DB_SHIFT); |
|
1589 diff = diff*6/(C*(mask_end-1)*(mask_end+1)*mask_end); |
|
1590 /* Again, being conservative */ |
|
1591 diff = HALF32(diff); |
|
1592 diff = MAX32(MIN32(diff, QCONST32(.031f, DB_SHIFT)), -QCONST32(.031f, DB_SHIFT)); |
|
1593 /* Find the band that's in the middle of the coded spectrum */ |
|
1594 for (midband=0;eBands[midband+1] < eBands[mask_end]/2;midband++); |
|
1595 count_dynalloc=0; |
|
1596 for(i=0;i<mask_end;i++) |
|
1597 { |
|
1598 opus_val32 lin; |
|
1599 opus_val16 unmask; |
|
1600 lin = mask_avg + diff*(i-midband); |
|
1601 if (C==2) |
|
1602 unmask = MAX16(st->energy_mask[i], st->energy_mask[nbEBands+i]); |
|
1603 else |
|
1604 unmask = st->energy_mask[i]; |
|
1605 unmask = MIN16(unmask, QCONST16(.0f, DB_SHIFT)); |
|
1606 unmask -= lin; |
|
1607 if (unmask > QCONST16(.25f, DB_SHIFT)) |
|
1608 { |
|
1609 surround_dynalloc[i] = unmask - QCONST16(.25f, DB_SHIFT); |
|
1610 count_dynalloc++; |
|
1611 } |
|
1612 } |
|
1613 if (count_dynalloc>=3) |
|
1614 { |
|
1615 /* If we need dynalloc in many bands, it's probably because our |
|
1616 initial masking rate was too low. */ |
|
1617 mask_avg += QCONST16(.25f, DB_SHIFT); |
|
1618 if (mask_avg>0) |
|
1619 { |
|
1620 /* Something went really wrong in the original calculations, |
|
1621 disabling masking. */ |
|
1622 mask_avg = 0; |
|
1623 diff = 0; |
|
1624 for(i=0;i<mask_end;i++) |
|
1625 surround_dynalloc[i] = 0; |
|
1626 } else { |
|
1627 for(i=0;i<mask_end;i++) |
|
1628 surround_dynalloc[i] = MAX16(0, surround_dynalloc[i]-QCONST16(.25f, DB_SHIFT)); |
|
1629 } |
|
1630 } |
|
1631 mask_avg += QCONST16(.2f, DB_SHIFT); |
|
1632 /* Convert to 1/64th units used for the trim */ |
|
1633 surround_trim = 64*diff; |
|
1634 /*printf("%d %d ", mask_avg, surround_trim);*/ |
|
1635 surround_masking = mask_avg; |
|
1636 } |
|
1637 /* Temporal VBR (but not for LFE) */ |
|
1638 if (!st->lfe) |
|
1639 { |
|
1640 opus_val16 follow=-QCONST16(10.0f,DB_SHIFT); |
|
1641 opus_val32 frame_avg=0; |
|
1642 opus_val16 offset = shortBlocks?HALF16(SHL16(LM, DB_SHIFT)):0; |
|
1643 for(i=st->start;i<st->end;i++) |
|
1644 { |
|
1645 follow = MAX16(follow-QCONST16(1.f, DB_SHIFT), bandLogE[i]-offset); |
|
1646 if (C==2) |
|
1647 follow = MAX16(follow, bandLogE[i+nbEBands]-offset); |
|
1648 frame_avg += follow; |
|
1649 } |
|
1650 frame_avg /= (st->end-st->start); |
|
1651 temporal_vbr = SUB16(frame_avg,st->spec_avg); |
|
1652 temporal_vbr = MIN16(QCONST16(3.f, DB_SHIFT), MAX16(-QCONST16(1.5f, DB_SHIFT), temporal_vbr)); |
|
1653 st->spec_avg += MULT16_16_Q15(QCONST16(.02f, 15), temporal_vbr); |
|
1654 } |
|
1655 /*for (i=0;i<21;i++) |
|
1656 printf("%f ", bandLogE[i]); |
|
1657 printf("\n");*/ |
|
1658 |
|
1659 if (!secondMdct) |
|
1660 { |
|
1661 for (i=0;i<C*nbEBands;i++) |
|
1662 bandLogE2[i] = bandLogE[i]; |
|
1663 } |
|
1664 |
|
1665 /* Last chance to catch any transient we might have missed in the |
|
1666 time-domain analysis */ |
|
1667 if (LM>0 && ec_tell(enc)+3<=total_bits && !isTransient && st->complexity>=5 && !st->lfe) |
|
1668 { |
|
1669 if (patch_transient_decision(bandLogE, oldBandE, nbEBands, st->end, C)) |
|
1670 { |
|
1671 isTransient = 1; |
|
1672 shortBlocks = M; |
|
1673 compute_mdcts(mode, shortBlocks, in, freq, C, CC, LM, st->upsample); |
|
1674 compute_band_energies(mode, freq, bandE, effEnd, C, M); |
|
1675 amp2Log2(mode, effEnd, st->end, bandE, bandLogE, C); |
|
1676 /* Compensate for the scaling of short vs long mdcts */ |
|
1677 for (i=0;i<C*nbEBands;i++) |
|
1678 bandLogE2[i] += HALF16(SHL16(LM, DB_SHIFT)); |
|
1679 tf_estimate = QCONST16(.2f,14); |
|
1680 } |
|
1681 } |
|
1682 |
|
1683 if (LM>0 && ec_tell(enc)+3<=total_bits) |
|
1684 ec_enc_bit_logp(enc, isTransient, 3); |
|
1685 |
|
1686 ALLOC(X, C*N, celt_norm); /**< Interleaved normalised MDCTs */ |
|
1687 |
|
1688 /* Band normalisation */ |
|
1689 normalise_bands(mode, freq, X, bandE, effEnd, C, M); |
|
1690 |
|
1691 ALLOC(tf_res, nbEBands, int); |
|
1692 /* Disable variable tf resolution for hybrid and at very low bitrate */ |
|
1693 if (effectiveBytes>=15*C && st->start==0 && st->complexity>=2 && !st->lfe) |
|
1694 { |
|
1695 int lambda; |
|
1696 if (effectiveBytes<40) |
|
1697 lambda = 12; |
|
1698 else if (effectiveBytes<60) |
|
1699 lambda = 6; |
|
1700 else if (effectiveBytes<100) |
|
1701 lambda = 4; |
|
1702 else |
|
1703 lambda = 3; |
|
1704 lambda*=2; |
|
1705 tf_select = tf_analysis(mode, effEnd, isTransient, tf_res, lambda, X, N, LM, &tf_sum, tf_estimate, tf_chan); |
|
1706 for (i=effEnd;i<st->end;i++) |
|
1707 tf_res[i] = tf_res[effEnd-1]; |
|
1708 } else { |
|
1709 tf_sum = 0; |
|
1710 for (i=0;i<st->end;i++) |
|
1711 tf_res[i] = isTransient; |
|
1712 tf_select=0; |
|
1713 } |
|
1714 |
|
1715 ALLOC(error, C*nbEBands, opus_val16); |
|
1716 quant_coarse_energy(mode, st->start, st->end, effEnd, bandLogE, |
|
1717 oldBandE, total_bits, error, enc, |
|
1718 C, LM, nbAvailableBytes, st->force_intra, |
|
1719 &st->delayedIntra, st->complexity >= 4, st->loss_rate, st->lfe); |
|
1720 |
|
1721 tf_encode(st->start, st->end, isTransient, tf_res, LM, tf_select, enc); |
|
1722 |
|
1723 if (ec_tell(enc)+4<=total_bits) |
|
1724 { |
|
1725 if (st->lfe) |
|
1726 { |
|
1727 st->tapset_decision = 0; |
|
1728 st->spread_decision = SPREAD_NORMAL; |
|
1729 } else if (shortBlocks || st->complexity < 3 || nbAvailableBytes < 10*C || st->start != 0) |
|
1730 { |
|
1731 if (st->complexity == 0) |
|
1732 st->spread_decision = SPREAD_NONE; |
|
1733 else |
|
1734 st->spread_decision = SPREAD_NORMAL; |
|
1735 } else { |
|
1736 /* Disable new spreading+tapset estimator until we can show it works |
|
1737 better than the old one. So far it seems like spreading_decision() |
|
1738 works best. */ |
|
1739 #if 0 |
|
1740 if (st->analysis.valid) |
|
1741 { |
|
1742 static const opus_val16 spread_thresholds[3] = {-QCONST16(.6f, 15), -QCONST16(.2f, 15), -QCONST16(.07f, 15)}; |
|
1743 static const opus_val16 spread_histeresis[3] = {QCONST16(.15f, 15), QCONST16(.07f, 15), QCONST16(.02f, 15)}; |
|
1744 static const opus_val16 tapset_thresholds[2] = {QCONST16(.0f, 15), QCONST16(.15f, 15)}; |
|
1745 static const opus_val16 tapset_histeresis[2] = {QCONST16(.1f, 15), QCONST16(.05f, 15)}; |
|
1746 st->spread_decision = hysteresis_decision(-st->analysis.tonality, spread_thresholds, spread_histeresis, 3, st->spread_decision); |
|
1747 st->tapset_decision = hysteresis_decision(st->analysis.tonality_slope, tapset_thresholds, tapset_histeresis, 2, st->tapset_decision); |
|
1748 } else |
|
1749 #endif |
|
1750 { |
|
1751 st->spread_decision = spreading_decision(mode, X, |
|
1752 &st->tonal_average, st->spread_decision, &st->hf_average, |
|
1753 &st->tapset_decision, pf_on&&!shortBlocks, effEnd, C, M); |
|
1754 } |
|
1755 /*printf("%d %d\n", st->tapset_decision, st->spread_decision);*/ |
|
1756 /*printf("%f %d %f %d\n\n", st->analysis.tonality, st->spread_decision, st->analysis.tonality_slope, st->tapset_decision);*/ |
|
1757 } |
|
1758 ec_enc_icdf(enc, st->spread_decision, spread_icdf, 5); |
|
1759 } |
|
1760 |
|
1761 ALLOC(offsets, nbEBands, int); |
|
1762 |
|
1763 maxDepth = dynalloc_analysis(bandLogE, bandLogE2, nbEBands, st->start, st->end, C, offsets, |
|
1764 st->lsb_depth, mode->logN, isTransient, st->vbr, st->constrained_vbr, |
|
1765 eBands, LM, effectiveBytes, &tot_boost, st->lfe, surround_dynalloc); |
|
1766 /* For LFE, everything interesting is in the first band */ |
|
1767 if (st->lfe) |
|
1768 offsets[0] = IMIN(8, effectiveBytes/3); |
|
1769 ALLOC(cap, nbEBands, int); |
|
1770 init_caps(mode,cap,LM,C); |
|
1771 |
|
1772 dynalloc_logp = 6; |
|
1773 total_bits<<=BITRES; |
|
1774 total_boost = 0; |
|
1775 tell = ec_tell_frac(enc); |
|
1776 for (i=st->start;i<st->end;i++) |
|
1777 { |
|
1778 int width, quanta; |
|
1779 int dynalloc_loop_logp; |
|
1780 int boost; |
|
1781 int j; |
|
1782 width = C*(eBands[i+1]-eBands[i])<<LM; |
|
1783 /* quanta is 6 bits, but no more than 1 bit/sample |
|
1784 and no less than 1/8 bit/sample */ |
|
1785 quanta = IMIN(width<<BITRES, IMAX(6<<BITRES, width)); |
|
1786 dynalloc_loop_logp = dynalloc_logp; |
|
1787 boost = 0; |
|
1788 for (j = 0; tell+(dynalloc_loop_logp<<BITRES) < total_bits-total_boost |
|
1789 && boost < cap[i]; j++) |
|
1790 { |
|
1791 int flag; |
|
1792 flag = j<offsets[i]; |
|
1793 ec_enc_bit_logp(enc, flag, dynalloc_loop_logp); |
|
1794 tell = ec_tell_frac(enc); |
|
1795 if (!flag) |
|
1796 break; |
|
1797 boost += quanta; |
|
1798 total_boost += quanta; |
|
1799 dynalloc_loop_logp = 1; |
|
1800 } |
|
1801 /* Making dynalloc more likely */ |
|
1802 if (j) |
|
1803 dynalloc_logp = IMAX(2, dynalloc_logp-1); |
|
1804 offsets[i] = boost; |
|
1805 } |
|
1806 |
|
1807 if (C==2) |
|
1808 { |
|
1809 static const opus_val16 intensity_thresholds[21]= |
|
1810 /* 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 off*/ |
|
1811 { 1, 2, 3, 4, 5, 6, 7, 8,16,24,36,44,50,56,62,67,72,79,88,106,134}; |
|
1812 static const opus_val16 intensity_histeresis[21]= |
|
1813 { 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 2, 3, 3, 4, 5, 6, 8, 8}; |
|
1814 |
|
1815 /* Always use MS for 2.5 ms frames until we can do a better analysis */ |
|
1816 if (LM!=0) |
|
1817 dual_stereo = stereo_analysis(mode, X, LM, N); |
|
1818 |
|
1819 st->intensity = hysteresis_decision((opus_val16)(equiv_rate/1000), |
|
1820 intensity_thresholds, intensity_histeresis, 21, st->intensity); |
|
1821 st->intensity = IMIN(st->end,IMAX(st->start, st->intensity)); |
|
1822 } |
|
1823 |
|
1824 alloc_trim = 5; |
|
1825 if (tell+(6<<BITRES) <= total_bits - total_boost) |
|
1826 { |
|
1827 if (st->lfe) |
|
1828 alloc_trim = 5; |
|
1829 else |
|
1830 alloc_trim = alloc_trim_analysis(mode, X, bandLogE, |
|
1831 st->end, LM, C, N, &st->analysis, &st->stereo_saving, tf_estimate, st->intensity, surround_trim); |
|
1832 ec_enc_icdf(enc, alloc_trim, trim_icdf, 7); |
|
1833 tell = ec_tell_frac(enc); |
|
1834 } |
|
1835 |
|
1836 /* Variable bitrate */ |
|
1837 if (vbr_rate>0) |
|
1838 { |
|
1839 opus_val16 alpha; |
|
1840 opus_int32 delta; |
|
1841 /* The target rate in 8th bits per frame */ |
|
1842 opus_int32 target, base_target; |
|
1843 opus_int32 min_allowed; |
|
1844 int lm_diff = mode->maxLM - LM; |
|
1845 |
|
1846 /* Don't attempt to use more than 510 kb/s, even for frames smaller than 20 ms. |
|
1847 The CELT allocator will just not be able to use more than that anyway. */ |
|
1848 nbCompressedBytes = IMIN(nbCompressedBytes,1275>>(3-LM)); |
|
1849 base_target = vbr_rate - ((40*C+20)<<BITRES); |
|
1850 |
|
1851 if (st->constrained_vbr) |
|
1852 base_target += (st->vbr_offset>>lm_diff); |
|
1853 |
|
1854 target = compute_vbr(mode, &st->analysis, base_target, LM, equiv_rate, |
|
1855 st->lastCodedBands, C, st->intensity, st->constrained_vbr, |
|
1856 st->stereo_saving, tot_boost, tf_estimate, pitch_change, maxDepth, |
|
1857 st->variable_duration, st->lfe, st->energy_mask!=NULL, surround_masking, |
|
1858 temporal_vbr); |
|
1859 |
|
1860 /* The current offset is removed from the target and the space used |
|
1861 so far is added*/ |
|
1862 target=target+tell; |
|
1863 /* In VBR mode the frame size must not be reduced so much that it would |
|
1864 result in the encoder running out of bits. |
|
1865 The margin of 2 bytes ensures that none of the bust-prevention logic |
|
1866 in the decoder will have triggered so far. */ |
|
1867 min_allowed = ((tell+total_boost+(1<<(BITRES+3))-1)>>(BITRES+3)) + 2 - nbFilledBytes; |
|
1868 |
|
1869 nbAvailableBytes = (target+(1<<(BITRES+2)))>>(BITRES+3); |
|
1870 nbAvailableBytes = IMAX(min_allowed,nbAvailableBytes); |
|
1871 nbAvailableBytes = IMIN(nbCompressedBytes,nbAvailableBytes+nbFilledBytes) - nbFilledBytes; |
|
1872 |
|
1873 /* By how much did we "miss" the target on that frame */ |
|
1874 delta = target - vbr_rate; |
|
1875 |
|
1876 target=nbAvailableBytes<<(BITRES+3); |
|
1877 |
|
1878 /*If the frame is silent we don't adjust our drift, otherwise |
|
1879 the encoder will shoot to very high rates after hitting a |
|
1880 span of silence, but we do allow the bitres to refill. |
|
1881 This means that we'll undershoot our target in CVBR/VBR modes |
|
1882 on files with lots of silence. */ |
|
1883 if(silence) |
|
1884 { |
|
1885 nbAvailableBytes = 2; |
|
1886 target = 2*8<<BITRES; |
|
1887 delta = 0; |
|
1888 } |
|
1889 |
|
1890 if (st->vbr_count < 970) |
|
1891 { |
|
1892 st->vbr_count++; |
|
1893 alpha = celt_rcp(SHL32(EXTEND32(st->vbr_count+20),16)); |
|
1894 } else |
|
1895 alpha = QCONST16(.001f,15); |
|
1896 /* How many bits have we used in excess of what we're allowed */ |
|
1897 if (st->constrained_vbr) |
|
1898 st->vbr_reservoir += target - vbr_rate; |
|
1899 /*printf ("%d\n", st->vbr_reservoir);*/ |
|
1900 |
|
1901 /* Compute the offset we need to apply in order to reach the target */ |
|
1902 if (st->constrained_vbr) |
|
1903 { |
|
1904 st->vbr_drift += (opus_int32)MULT16_32_Q15(alpha,(delta*(1<<lm_diff))-st->vbr_offset-st->vbr_drift); |
|
1905 st->vbr_offset = -st->vbr_drift; |
|
1906 } |
|
1907 /*printf ("%d\n", st->vbr_drift);*/ |
|
1908 |
|
1909 if (st->constrained_vbr && st->vbr_reservoir < 0) |
|
1910 { |
|
1911 /* We're under the min value -- increase rate */ |
|
1912 int adjust = (-st->vbr_reservoir)/(8<<BITRES); |
|
1913 /* Unless we're just coding silence */ |
|
1914 nbAvailableBytes += silence?0:adjust; |
|
1915 st->vbr_reservoir = 0; |
|
1916 /*printf ("+%d\n", adjust);*/ |
|
1917 } |
|
1918 nbCompressedBytes = IMIN(nbCompressedBytes,nbAvailableBytes+nbFilledBytes); |
|
1919 /*printf("%d\n", nbCompressedBytes*50*8);*/ |
|
1920 /* This moves the raw bits to take into account the new compressed size */ |
|
1921 ec_enc_shrink(enc, nbCompressedBytes); |
|
1922 } |
|
1923 |
|
1924 /* Bit allocation */ |
|
1925 ALLOC(fine_quant, nbEBands, int); |
|
1926 ALLOC(pulses, nbEBands, int); |
|
1927 ALLOC(fine_priority, nbEBands, int); |
|
1928 |
|
1929 /* bits = packet size - where we are - safety*/ |
|
1930 bits = (((opus_int32)nbCompressedBytes*8)<<BITRES) - ec_tell_frac(enc) - 1; |
|
1931 anti_collapse_rsv = isTransient&&LM>=2&&bits>=((LM+2)<<BITRES) ? (1<<BITRES) : 0; |
|
1932 bits -= anti_collapse_rsv; |
|
1933 signalBandwidth = st->end-1; |
|
1934 #ifndef DISABLE_FLOAT_API |
|
1935 if (st->analysis.valid) |
|
1936 { |
|
1937 int min_bandwidth; |
|
1938 if (equiv_rate < (opus_int32)32000*C) |
|
1939 min_bandwidth = 13; |
|
1940 else if (equiv_rate < (opus_int32)48000*C) |
|
1941 min_bandwidth = 16; |
|
1942 else if (equiv_rate < (opus_int32)60000*C) |
|
1943 min_bandwidth = 18; |
|
1944 else if (equiv_rate < (opus_int32)80000*C) |
|
1945 min_bandwidth = 19; |
|
1946 else |
|
1947 min_bandwidth = 20; |
|
1948 signalBandwidth = IMAX(st->analysis.bandwidth, min_bandwidth); |
|
1949 } |
|
1950 #endif |
|
1951 if (st->lfe) |
|
1952 signalBandwidth = 1; |
|
1953 codedBands = compute_allocation(mode, st->start, st->end, offsets, cap, |
|
1954 alloc_trim, &st->intensity, &dual_stereo, bits, &balance, pulses, |
|
1955 fine_quant, fine_priority, C, LM, enc, 1, st->lastCodedBands, signalBandwidth); |
|
1956 if (st->lastCodedBands) |
|
1957 st->lastCodedBands = IMIN(st->lastCodedBands+1,IMAX(st->lastCodedBands-1,codedBands)); |
|
1958 else |
|
1959 st->lastCodedBands = codedBands; |
|
1960 |
|
1961 quant_fine_energy(mode, st->start, st->end, oldBandE, error, fine_quant, enc, C); |
|
1962 |
|
1963 /* Residual quantisation */ |
|
1964 ALLOC(collapse_masks, C*nbEBands, unsigned char); |
|
1965 quant_all_bands(1, mode, st->start, st->end, X, C==2 ? X+N : NULL, collapse_masks, |
|
1966 bandE, pulses, shortBlocks, st->spread_decision, dual_stereo, st->intensity, tf_res, |
|
1967 nbCompressedBytes*(8<<BITRES)-anti_collapse_rsv, balance, enc, LM, codedBands, &st->rng); |
|
1968 |
|
1969 if (anti_collapse_rsv > 0) |
|
1970 { |
|
1971 anti_collapse_on = st->consec_transient<2; |
|
1972 #ifdef FUZZING |
|
1973 anti_collapse_on = rand()&0x1; |
|
1974 #endif |
|
1975 ec_enc_bits(enc, anti_collapse_on, 1); |
|
1976 } |
|
1977 quant_energy_finalise(mode, st->start, st->end, oldBandE, error, fine_quant, fine_priority, nbCompressedBytes*8-ec_tell(enc), enc, C); |
|
1978 |
|
1979 if (silence) |
|
1980 { |
|
1981 for (i=0;i<C*nbEBands;i++) |
|
1982 oldBandE[i] = -QCONST16(28.f,DB_SHIFT); |
|
1983 } |
|
1984 |
|
1985 #ifdef RESYNTH |
|
1986 /* Re-synthesis of the coded audio if required */ |
|
1987 { |
|
1988 celt_sig *out_mem[2]; |
|
1989 |
|
1990 if (anti_collapse_on) |
|
1991 { |
|
1992 anti_collapse(mode, X, collapse_masks, LM, C, N, |
|
1993 st->start, st->end, oldBandE, oldLogE, oldLogE2, pulses, st->rng); |
|
1994 } |
|
1995 |
|
1996 if (silence) |
|
1997 { |
|
1998 for (i=0;i<C*N;i++) |
|
1999 freq[i] = 0; |
|
2000 } else { |
|
2001 /* Synthesis */ |
|
2002 denormalise_bands(mode, X, freq, oldBandE, st->start, effEnd, C, M); |
|
2003 } |
|
2004 |
|
2005 c=0; do { |
|
2006 OPUS_MOVE(st->syn_mem[c], st->syn_mem[c]+N, 2*MAX_PERIOD-N+overlap/2); |
|
2007 } while (++c<CC); |
|
2008 |
|
2009 if (CC==2&&C==1) |
|
2010 { |
|
2011 for (i=0;i<N;i++) |
|
2012 freq[N+i] = freq[i]; |
|
2013 } |
|
2014 |
|
2015 c=0; do { |
|
2016 out_mem[c] = st->syn_mem[c]+2*MAX_PERIOD-N; |
|
2017 } while (++c<CC); |
|
2018 |
|
2019 compute_inv_mdcts(mode, shortBlocks, freq, out_mem, CC, LM); |
|
2020 |
|
2021 c=0; do { |
|
2022 st->prefilter_period=IMAX(st->prefilter_period, COMBFILTER_MINPERIOD); |
|
2023 st->prefilter_period_old=IMAX(st->prefilter_period_old, COMBFILTER_MINPERIOD); |
|
2024 comb_filter(out_mem[c], out_mem[c], st->prefilter_period_old, st->prefilter_period, mode->shortMdctSize, |
|
2025 st->prefilter_gain_old, st->prefilter_gain, st->prefilter_tapset_old, st->prefilter_tapset, |
|
2026 mode->window, st->overlap); |
|
2027 if (LM!=0) |
|
2028 comb_filter(out_mem[c]+mode->shortMdctSize, out_mem[c]+mode->shortMdctSize, st->prefilter_period, pitch_index, N-mode->shortMdctSize, |
|
2029 st->prefilter_gain, gain1, st->prefilter_tapset, prefilter_tapset, |
|
2030 mode->window, overlap); |
|
2031 } while (++c<CC); |
|
2032 |
|
2033 /* We reuse freq[] as scratch space for the de-emphasis */ |
|
2034 deemphasis(out_mem, (opus_val16*)pcm, N, CC, st->upsample, mode->preemph, st->preemph_memD, freq); |
|
2035 st->prefilter_period_old = st->prefilter_period; |
|
2036 st->prefilter_gain_old = st->prefilter_gain; |
|
2037 st->prefilter_tapset_old = st->prefilter_tapset; |
|
2038 } |
|
2039 #endif |
|
2040 |
|
2041 st->prefilter_period = pitch_index; |
|
2042 st->prefilter_gain = gain1; |
|
2043 st->prefilter_tapset = prefilter_tapset; |
|
2044 #ifdef RESYNTH |
|
2045 if (LM!=0) |
|
2046 { |
|
2047 st->prefilter_period_old = st->prefilter_period; |
|
2048 st->prefilter_gain_old = st->prefilter_gain; |
|
2049 st->prefilter_tapset_old = st->prefilter_tapset; |
|
2050 } |
|
2051 #endif |
|
2052 |
|
2053 if (CC==2&&C==1) { |
|
2054 for (i=0;i<nbEBands;i++) |
|
2055 oldBandE[nbEBands+i]=oldBandE[i]; |
|
2056 } |
|
2057 |
|
2058 if (!isTransient) |
|
2059 { |
|
2060 for (i=0;i<CC*nbEBands;i++) |
|
2061 oldLogE2[i] = oldLogE[i]; |
|
2062 for (i=0;i<CC*nbEBands;i++) |
|
2063 oldLogE[i] = oldBandE[i]; |
|
2064 } else { |
|
2065 for (i=0;i<CC*nbEBands;i++) |
|
2066 oldLogE[i] = MIN16(oldLogE[i], oldBandE[i]); |
|
2067 } |
|
2068 /* In case start or end were to change */ |
|
2069 c=0; do |
|
2070 { |
|
2071 for (i=0;i<st->start;i++) |
|
2072 { |
|
2073 oldBandE[c*nbEBands+i]=0; |
|
2074 oldLogE[c*nbEBands+i]=oldLogE2[c*nbEBands+i]=-QCONST16(28.f,DB_SHIFT); |
|
2075 } |
|
2076 for (i=st->end;i<nbEBands;i++) |
|
2077 { |
|
2078 oldBandE[c*nbEBands+i]=0; |
|
2079 oldLogE[c*nbEBands+i]=oldLogE2[c*nbEBands+i]=-QCONST16(28.f,DB_SHIFT); |
|
2080 } |
|
2081 } while (++c<CC); |
|
2082 |
|
2083 if (isTransient || transient_got_disabled) |
|
2084 st->consec_transient++; |
|
2085 else |
|
2086 st->consec_transient=0; |
|
2087 st->rng = enc->rng; |
|
2088 |
|
2089 /* If there's any room left (can only happen for very high rates), |
|
2090 it's already filled with zeros */ |
|
2091 ec_enc_done(enc); |
|
2092 |
|
2093 #ifdef CUSTOM_MODES |
|
2094 if (st->signalling) |
|
2095 nbCompressedBytes++; |
|
2096 #endif |
|
2097 |
|
2098 RESTORE_STACK; |
|
2099 if (ec_get_error(enc)) |
|
2100 return OPUS_INTERNAL_ERROR; |
|
2101 else |
|
2102 return nbCompressedBytes; |
|
2103 } |
|
2104 |
|
2105 |
|
2106 #ifdef CUSTOM_MODES |
|
2107 |
|
2108 #ifdef FIXED_POINT |
|
2109 int opus_custom_encode(CELTEncoder * OPUS_RESTRICT st, const opus_int16 * pcm, int frame_size, unsigned char *compressed, int nbCompressedBytes) |
|
2110 { |
|
2111 return celt_encode_with_ec(st, pcm, frame_size, compressed, nbCompressedBytes, NULL); |
|
2112 } |
|
2113 |
|
2114 #ifndef DISABLE_FLOAT_API |
|
2115 int opus_custom_encode_float(CELTEncoder * OPUS_RESTRICT st, const float * pcm, int frame_size, unsigned char *compressed, int nbCompressedBytes) |
|
2116 { |
|
2117 int j, ret, C, N; |
|
2118 VARDECL(opus_int16, in); |
|
2119 ALLOC_STACK; |
|
2120 |
|
2121 if (pcm==NULL) |
|
2122 return OPUS_BAD_ARG; |
|
2123 |
|
2124 C = st->channels; |
|
2125 N = frame_size; |
|
2126 ALLOC(in, C*N, opus_int16); |
|
2127 |
|
2128 for (j=0;j<C*N;j++) |
|
2129 in[j] = FLOAT2INT16(pcm[j]); |
|
2130 |
|
2131 ret=celt_encode_with_ec(st,in,frame_size,compressed,nbCompressedBytes, NULL); |
|
2132 #ifdef RESYNTH |
|
2133 for (j=0;j<C*N;j++) |
|
2134 ((float*)pcm)[j]=in[j]*(1.f/32768.f); |
|
2135 #endif |
|
2136 RESTORE_STACK; |
|
2137 return ret; |
|
2138 } |
|
2139 #endif /* DISABLE_FLOAT_API */ |
|
2140 #else |
|
2141 |
|
2142 int opus_custom_encode(CELTEncoder * OPUS_RESTRICT st, const opus_int16 * pcm, int frame_size, unsigned char *compressed, int nbCompressedBytes) |
|
2143 { |
|
2144 int j, ret, C, N; |
|
2145 VARDECL(celt_sig, in); |
|
2146 ALLOC_STACK; |
|
2147 |
|
2148 if (pcm==NULL) |
|
2149 return OPUS_BAD_ARG; |
|
2150 |
|
2151 C=st->channels; |
|
2152 N=frame_size; |
|
2153 ALLOC(in, C*N, celt_sig); |
|
2154 for (j=0;j<C*N;j++) { |
|
2155 in[j] = SCALEOUT(pcm[j]); |
|
2156 } |
|
2157 |
|
2158 ret = celt_encode_with_ec(st,in,frame_size,compressed,nbCompressedBytes, NULL); |
|
2159 #ifdef RESYNTH |
|
2160 for (j=0;j<C*N;j++) |
|
2161 ((opus_int16*)pcm)[j] = FLOAT2INT16(in[j]); |
|
2162 #endif |
|
2163 RESTORE_STACK; |
|
2164 return ret; |
|
2165 } |
|
2166 |
|
2167 int opus_custom_encode_float(CELTEncoder * OPUS_RESTRICT st, const float * pcm, int frame_size, unsigned char *compressed, int nbCompressedBytes) |
|
2168 { |
|
2169 return celt_encode_with_ec(st, pcm, frame_size, compressed, nbCompressedBytes, NULL); |
|
2170 } |
|
2171 |
|
2172 #endif |
|
2173 |
|
2174 #endif /* CUSTOM_MODES */ |
|
2175 |
|
2176 int opus_custom_encoder_ctl(CELTEncoder * OPUS_RESTRICT st, int request, ...) |
|
2177 { |
|
2178 va_list ap; |
|
2179 |
|
2180 va_start(ap, request); |
|
2181 switch (request) |
|
2182 { |
|
2183 case OPUS_SET_COMPLEXITY_REQUEST: |
|
2184 { |
|
2185 int value = va_arg(ap, opus_int32); |
|
2186 if (value<0 || value>10) |
|
2187 goto bad_arg; |
|
2188 st->complexity = value; |
|
2189 } |
|
2190 break; |
|
2191 case CELT_SET_START_BAND_REQUEST: |
|
2192 { |
|
2193 opus_int32 value = va_arg(ap, opus_int32); |
|
2194 if (value<0 || value>=st->mode->nbEBands) |
|
2195 goto bad_arg; |
|
2196 st->start = value; |
|
2197 } |
|
2198 break; |
|
2199 case CELT_SET_END_BAND_REQUEST: |
|
2200 { |
|
2201 opus_int32 value = va_arg(ap, opus_int32); |
|
2202 if (value<1 || value>st->mode->nbEBands) |
|
2203 goto bad_arg; |
|
2204 st->end = value; |
|
2205 } |
|
2206 break; |
|
2207 case CELT_SET_PREDICTION_REQUEST: |
|
2208 { |
|
2209 int value = va_arg(ap, opus_int32); |
|
2210 if (value<0 || value>2) |
|
2211 goto bad_arg; |
|
2212 st->disable_pf = value<=1; |
|
2213 st->force_intra = value==0; |
|
2214 } |
|
2215 break; |
|
2216 case OPUS_SET_PACKET_LOSS_PERC_REQUEST: |
|
2217 { |
|
2218 int value = va_arg(ap, opus_int32); |
|
2219 if (value<0 || value>100) |
|
2220 goto bad_arg; |
|
2221 st->loss_rate = value; |
|
2222 } |
|
2223 break; |
|
2224 case OPUS_SET_VBR_CONSTRAINT_REQUEST: |
|
2225 { |
|
2226 opus_int32 value = va_arg(ap, opus_int32); |
|
2227 st->constrained_vbr = value; |
|
2228 } |
|
2229 break; |
|
2230 case OPUS_SET_VBR_REQUEST: |
|
2231 { |
|
2232 opus_int32 value = va_arg(ap, opus_int32); |
|
2233 st->vbr = value; |
|
2234 } |
|
2235 break; |
|
2236 case OPUS_SET_BITRATE_REQUEST: |
|
2237 { |
|
2238 opus_int32 value = va_arg(ap, opus_int32); |
|
2239 if (value<=500 && value!=OPUS_BITRATE_MAX) |
|
2240 goto bad_arg; |
|
2241 value = IMIN(value, 260000*st->channels); |
|
2242 st->bitrate = value; |
|
2243 } |
|
2244 break; |
|
2245 case CELT_SET_CHANNELS_REQUEST: |
|
2246 { |
|
2247 opus_int32 value = va_arg(ap, opus_int32); |
|
2248 if (value<1 || value>2) |
|
2249 goto bad_arg; |
|
2250 st->stream_channels = value; |
|
2251 } |
|
2252 break; |
|
2253 case OPUS_SET_LSB_DEPTH_REQUEST: |
|
2254 { |
|
2255 opus_int32 value = va_arg(ap, opus_int32); |
|
2256 if (value<8 || value>24) |
|
2257 goto bad_arg; |
|
2258 st->lsb_depth=value; |
|
2259 } |
|
2260 break; |
|
2261 case OPUS_GET_LSB_DEPTH_REQUEST: |
|
2262 { |
|
2263 opus_int32 *value = va_arg(ap, opus_int32*); |
|
2264 *value=st->lsb_depth; |
|
2265 } |
|
2266 break; |
|
2267 case OPUS_SET_EXPERT_FRAME_DURATION_REQUEST: |
|
2268 { |
|
2269 opus_int32 value = va_arg(ap, opus_int32); |
|
2270 st->variable_duration = value; |
|
2271 } |
|
2272 break; |
|
2273 case OPUS_RESET_STATE: |
|
2274 { |
|
2275 int i; |
|
2276 opus_val16 *oldBandE, *oldLogE, *oldLogE2; |
|
2277 oldBandE = (opus_val16*)(st->in_mem+st->channels*(st->overlap+COMBFILTER_MAXPERIOD)); |
|
2278 oldLogE = oldBandE + st->channels*st->mode->nbEBands; |
|
2279 oldLogE2 = oldLogE + st->channels*st->mode->nbEBands; |
|
2280 OPUS_CLEAR((char*)&st->ENCODER_RESET_START, |
|
2281 opus_custom_encoder_get_size(st->mode, st->channels)- |
|
2282 ((char*)&st->ENCODER_RESET_START - (char*)st)); |
|
2283 for (i=0;i<st->channels*st->mode->nbEBands;i++) |
|
2284 oldLogE[i]=oldLogE2[i]=-QCONST16(28.f,DB_SHIFT); |
|
2285 st->vbr_offset = 0; |
|
2286 st->delayedIntra = 1; |
|
2287 st->spread_decision = SPREAD_NORMAL; |
|
2288 st->tonal_average = 256; |
|
2289 st->hf_average = 0; |
|
2290 st->tapset_decision = 0; |
|
2291 } |
|
2292 break; |
|
2293 #ifdef CUSTOM_MODES |
|
2294 case CELT_SET_INPUT_CLIPPING_REQUEST: |
|
2295 { |
|
2296 opus_int32 value = va_arg(ap, opus_int32); |
|
2297 st->clip = value; |
|
2298 } |
|
2299 break; |
|
2300 #endif |
|
2301 case CELT_SET_SIGNALLING_REQUEST: |
|
2302 { |
|
2303 opus_int32 value = va_arg(ap, opus_int32); |
|
2304 st->signalling = value; |
|
2305 } |
|
2306 break; |
|
2307 case CELT_SET_ANALYSIS_REQUEST: |
|
2308 { |
|
2309 AnalysisInfo *info = va_arg(ap, AnalysisInfo *); |
|
2310 if (info) |
|
2311 OPUS_COPY(&st->analysis, info, 1); |
|
2312 } |
|
2313 break; |
|
2314 case CELT_GET_MODE_REQUEST: |
|
2315 { |
|
2316 const CELTMode ** value = va_arg(ap, const CELTMode**); |
|
2317 if (value==0) |
|
2318 goto bad_arg; |
|
2319 *value=st->mode; |
|
2320 } |
|
2321 break; |
|
2322 case OPUS_GET_FINAL_RANGE_REQUEST: |
|
2323 { |
|
2324 opus_uint32 * value = va_arg(ap, opus_uint32 *); |
|
2325 if (value==0) |
|
2326 goto bad_arg; |
|
2327 *value=st->rng; |
|
2328 } |
|
2329 break; |
|
2330 case OPUS_SET_LFE_REQUEST: |
|
2331 { |
|
2332 opus_int32 value = va_arg(ap, opus_int32); |
|
2333 st->lfe = value; |
|
2334 } |
|
2335 break; |
|
2336 case OPUS_SET_ENERGY_MASK_REQUEST: |
|
2337 { |
|
2338 opus_val16 *value = va_arg(ap, opus_val16*); |
|
2339 st->energy_mask = value; |
|
2340 } |
|
2341 break; |
|
2342 default: |
|
2343 goto bad_request; |
|
2344 } |
|
2345 va_end(ap); |
|
2346 return OPUS_OK; |
|
2347 bad_arg: |
|
2348 va_end(ap); |
|
2349 return OPUS_BAD_ARG; |
|
2350 bad_request: |
|
2351 va_end(ap); |
|
2352 return OPUS_UNIMPLEMENTED; |
|
2353 } |