/src/opus/celt/celt_encoder.c
Line | Count | Source (jump to first uncovered line) |
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 | | #ifndef M_PI |
56 | | #define M_PI 3.141592653 |
57 | | #endif |
58 | | |
59 | | |
60 | | /** Encoder state |
61 | | @brief Encoder state |
62 | | */ |
63 | | struct OpusCustomEncoder { |
64 | | const OpusCustomMode *mode; /**< Mode used by the encoder */ |
65 | | int channels; |
66 | | int stream_channels; |
67 | | |
68 | | int force_intra; |
69 | | int clip; |
70 | | int disable_pf; |
71 | | int complexity; |
72 | | int upsample; |
73 | | int start, end; |
74 | | |
75 | | opus_int32 bitrate; |
76 | | int vbr; |
77 | | int signalling; |
78 | | int constrained_vbr; /* If zero, VBR can do whatever it likes with the rate */ |
79 | | int loss_rate; |
80 | | int lsb_depth; |
81 | | int lfe; |
82 | | int disable_inv; |
83 | | int arch; |
84 | | |
85 | | /* Everything beyond this point gets cleared on a reset */ |
86 | | #define ENCODER_RESET_START rng |
87 | | |
88 | | opus_uint32 rng; |
89 | | int spread_decision; |
90 | | opus_val32 delayedIntra; |
91 | | int tonal_average; |
92 | | int lastCodedBands; |
93 | | int hf_average; |
94 | | int tapset_decision; |
95 | | |
96 | | int prefilter_period; |
97 | | opus_val16 prefilter_gain; |
98 | | int prefilter_tapset; |
99 | | #ifdef RESYNTH |
100 | | int prefilter_period_old; |
101 | | opus_val16 prefilter_gain_old; |
102 | | int prefilter_tapset_old; |
103 | | #endif |
104 | | int consec_transient; |
105 | | AnalysisInfo analysis; |
106 | | SILKInfo silk_info; |
107 | | |
108 | | opus_val32 preemph_memE[2]; |
109 | | opus_val32 preemph_memD[2]; |
110 | | |
111 | | /* VBR-related parameters */ |
112 | | opus_int32 vbr_reservoir; |
113 | | opus_int32 vbr_drift; |
114 | | opus_int32 vbr_offset; |
115 | | opus_int32 vbr_count; |
116 | | opus_val32 overlap_max; |
117 | | opus_val16 stereo_saving; |
118 | | int intensity; |
119 | | celt_glog *energy_mask; |
120 | | celt_glog spec_avg; |
121 | | |
122 | | #ifdef RESYNTH |
123 | | /* +MAX_PERIOD/2 to make space for overlap */ |
124 | | celt_sig syn_mem[2][2*MAX_PERIOD+MAX_PERIOD/2]; |
125 | | #endif |
126 | | |
127 | | celt_sig in_mem[1]; /* Size = channels*mode->overlap */ |
128 | | /* celt_sig prefilter_mem[], Size = channels*COMBFILTER_MAXPERIOD */ |
129 | | /* celt_glog oldBandE[], Size = channels*mode->nbEBands */ |
130 | | /* celt_glog oldLogE[], Size = channels*mode->nbEBands */ |
131 | | /* celt_glog oldLogE2[], Size = channels*mode->nbEBands */ |
132 | | /* celt_glog energyError[], Size = channels*mode->nbEBands */ |
133 | | }; |
134 | | |
135 | | int celt_encoder_get_size(int channels) |
136 | 0 | { |
137 | 0 | CELTMode *mode = opus_custom_mode_create(48000, 960, NULL); |
138 | 0 | return opus_custom_encoder_get_size(mode, channels); |
139 | 0 | } |
140 | | |
141 | | OPUS_CUSTOM_NOSTATIC int opus_custom_encoder_get_size(const CELTMode *mode, int channels) |
142 | 0 | { |
143 | 0 | int size = sizeof(struct CELTEncoder) |
144 | 0 | + (channels*mode->overlap-1)*sizeof(celt_sig) /* celt_sig in_mem[channels*mode->overlap]; */ |
145 | 0 | + channels*COMBFILTER_MAXPERIOD*sizeof(celt_sig) /* celt_sig prefilter_mem[channels*COMBFILTER_MAXPERIOD]; */ |
146 | 0 | + 4*channels*mode->nbEBands*sizeof(celt_glog); /* celt_glog oldBandE[channels*mode->nbEBands]; */ |
147 | | /* celt_glog oldLogE[channels*mode->nbEBands]; */ |
148 | | /* celt_glog oldLogE2[channels*mode->nbEBands]; */ |
149 | | /* celt_glog energyError[channels*mode->nbEBands]; */ |
150 | 0 | return size; |
151 | 0 | } |
152 | | |
153 | | #ifdef CUSTOM_MODES |
154 | | CELTEncoder *opus_custom_encoder_create(const CELTMode *mode, int channels, int *error) |
155 | | { |
156 | | int ret; |
157 | | CELTEncoder *st = (CELTEncoder *)opus_alloc(opus_custom_encoder_get_size(mode, channels)); |
158 | | /* init will handle the NULL case */ |
159 | | ret = opus_custom_encoder_init(st, mode, channels); |
160 | | if (ret != OPUS_OK) |
161 | | { |
162 | | opus_custom_encoder_destroy(st); |
163 | | st = NULL; |
164 | | } |
165 | | if (error) |
166 | | *error = ret; |
167 | | return st; |
168 | | } |
169 | | #endif /* CUSTOM_MODES */ |
170 | | |
171 | | static int opus_custom_encoder_init_arch(CELTEncoder *st, const CELTMode *mode, |
172 | | int channels, int arch) |
173 | 0 | { |
174 | 0 | if (channels < 0 || channels > 2) |
175 | 0 | return OPUS_BAD_ARG; |
176 | | |
177 | 0 | if (st==NULL || mode==NULL) |
178 | 0 | return OPUS_ALLOC_FAIL; |
179 | | |
180 | 0 | OPUS_CLEAR((char*)st, opus_custom_encoder_get_size(mode, channels)); |
181 | |
|
182 | 0 | st->mode = mode; |
183 | 0 | st->stream_channels = st->channels = channels; |
184 | |
|
185 | 0 | st->upsample = 1; |
186 | 0 | st->start = 0; |
187 | 0 | st->end = st->mode->effEBands; |
188 | 0 | st->signalling = 1; |
189 | 0 | st->arch = arch; |
190 | |
|
191 | 0 | st->constrained_vbr = 1; |
192 | 0 | st->clip = 1; |
193 | |
|
194 | 0 | st->bitrate = OPUS_BITRATE_MAX; |
195 | 0 | st->vbr = 0; |
196 | 0 | st->force_intra = 0; |
197 | 0 | st->complexity = 5; |
198 | 0 | st->lsb_depth=24; |
199 | |
|
200 | 0 | opus_custom_encoder_ctl(st, OPUS_RESET_STATE); |
201 | |
|
202 | 0 | return OPUS_OK; |
203 | 0 | } |
204 | | |
205 | | #ifdef CUSTOM_MODES |
206 | | int opus_custom_encoder_init(CELTEncoder *st, const CELTMode *mode, int channels) |
207 | | { |
208 | | return opus_custom_encoder_init_arch(st, mode, channels, opus_select_arch()); |
209 | | } |
210 | | #endif |
211 | | |
212 | | int celt_encoder_init(CELTEncoder *st, opus_int32 sampling_rate, int channels, |
213 | | int arch) |
214 | 0 | { |
215 | 0 | int ret; |
216 | 0 | ret = opus_custom_encoder_init_arch(st, |
217 | 0 | opus_custom_mode_create(48000, 960, NULL), channels, arch); |
218 | 0 | if (ret != OPUS_OK) |
219 | 0 | return ret; |
220 | 0 | st->upsample = resampling_factor(sampling_rate); |
221 | 0 | return OPUS_OK; |
222 | 0 | } |
223 | | |
224 | | #ifdef CUSTOM_MODES |
225 | | void opus_custom_encoder_destroy(CELTEncoder *st) |
226 | | { |
227 | | opus_free(st); |
228 | | } |
229 | | #endif /* CUSTOM_MODES */ |
230 | | |
231 | | |
232 | | static int transient_analysis(const opus_val32 * OPUS_RESTRICT in, int len, int C, |
233 | | opus_val16 *tf_estimate, int *tf_chan, int allow_weak_transients, |
234 | | int *weak_transient, opus_val16 tone_freq, opus_val32 toneishness) |
235 | 0 | { |
236 | 0 | int i; |
237 | 0 | VARDECL(opus_val16, tmp); |
238 | 0 | opus_val32 mem0,mem1; |
239 | 0 | int is_transient = 0; |
240 | 0 | opus_int32 mask_metric = 0; |
241 | 0 | int c; |
242 | 0 | opus_val16 tf_max; |
243 | 0 | int len2; |
244 | | /* Forward masking: 6.7 dB/ms. */ |
245 | | #ifdef FIXED_POINT |
246 | | int forward_shift = 4; |
247 | | #else |
248 | 0 | opus_val16 forward_decay = QCONST16(.0625f,15); |
249 | 0 | #endif |
250 | | /* Table of 6*64/x, trained on real data to minimize the average error */ |
251 | 0 | static const unsigned char inv_table[128] = { |
252 | 0 | 255,255,156,110, 86, 70, 59, 51, 45, 40, 37, 33, 31, 28, 26, 25, |
253 | 0 | 23, 22, 21, 20, 19, 18, 17, 16, 16, 15, 15, 14, 13, 13, 12, 12, |
254 | 0 | 12, 12, 11, 11, 11, 10, 10, 10, 9, 9, 9, 9, 9, 9, 8, 8, |
255 | 0 | 8, 8, 8, 7, 7, 7, 7, 7, 7, 6, 6, 6, 6, 6, 6, 6, |
256 | 0 | 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5, 5, 5, 5, |
257 | 0 | 5, 5, 5, 5, 5, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, |
258 | 0 | 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 3, 3, |
259 | 0 | 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 2, |
260 | 0 | }; |
261 | 0 | SAVE_STACK; |
262 | | #ifdef FIXED_POINT |
263 | | int in_shift = IMAX(0, celt_ilog2(1+celt_maxabs32(in, C*len))-14); |
264 | | #endif |
265 | 0 | ALLOC(tmp, len, opus_val16); |
266 | |
|
267 | 0 | *weak_transient = 0; |
268 | | /* For lower bitrates, let's be more conservative and have a forward masking |
269 | | decay of 3.3 dB/ms. This avoids having to code transients at very low |
270 | | bitrate (mostly for hybrid), which can result in unstable energy and/or |
271 | | partial collapse. */ |
272 | 0 | if (allow_weak_transients) |
273 | 0 | { |
274 | | #ifdef FIXED_POINT |
275 | | forward_shift = 5; |
276 | | #else |
277 | 0 | forward_decay = QCONST16(.03125f,15); |
278 | 0 | #endif |
279 | 0 | } |
280 | 0 | len2=len/2; |
281 | 0 | for (c=0;c<C;c++) |
282 | 0 | { |
283 | 0 | opus_val32 mean; |
284 | 0 | opus_int32 unmask=0; |
285 | 0 | opus_val32 norm; |
286 | 0 | opus_val16 maxE; |
287 | 0 | mem0=0; |
288 | 0 | mem1=0; |
289 | | /* High-pass filter: (1 - 2*z^-1 + z^-2) / (1 - z^-1 + .5*z^-2) */ |
290 | 0 | for (i=0;i<len;i++) |
291 | 0 | { |
292 | 0 | #ifndef FIXED_POINT |
293 | 0 | float mem00; |
294 | 0 | #endif |
295 | 0 | opus_val32 x,y; |
296 | 0 | x = SHR32(in[i+c*len],in_shift); |
297 | 0 | y = ADD32(mem0, x); |
298 | | #ifdef FIXED_POINT |
299 | | mem0 = mem1 + y - SHL32(x,1); |
300 | | mem1 = x - SHR32(y,1); |
301 | | #else |
302 | | /* Original code: |
303 | | mem0 = mem1 + y - 2*x; |
304 | | mem1 = x - .5f*y; |
305 | | Modified code to shorten dependency chains: */ |
306 | 0 | mem00=mem0; |
307 | 0 | mem0 = mem0 - x + .5f*mem1; |
308 | 0 | mem1 = x - mem00; |
309 | 0 | #endif |
310 | 0 | tmp[i] = SROUND16(y, 2); |
311 | | /*printf("%f ", tmp[i]);*/ |
312 | 0 | } |
313 | | /*printf("\n");*/ |
314 | | /* First few samples are bad because we don't propagate the memory */ |
315 | 0 | OPUS_CLEAR(tmp, 12); |
316 | |
|
317 | | #ifdef FIXED_POINT |
318 | | /* Normalize tmp to max range */ |
319 | | { |
320 | | int shift=0; |
321 | | shift = 14-celt_ilog2(MAX16(1, celt_maxabs16(tmp, len))); |
322 | | if (shift!=0) |
323 | | { |
324 | | for (i=0;i<len;i++) |
325 | | tmp[i] = SHL16(tmp[i], shift); |
326 | | } |
327 | | } |
328 | | #endif |
329 | |
|
330 | 0 | mean=0; |
331 | 0 | mem0=0; |
332 | | /* Grouping by two to reduce complexity */ |
333 | | /* Forward pass to compute the post-echo threshold*/ |
334 | 0 | for (i=0;i<len2;i++) |
335 | 0 | { |
336 | 0 | opus_val32 x2 = PSHR32(MULT16_16(tmp[2*i],tmp[2*i]) + MULT16_16(tmp[2*i+1],tmp[2*i+1]),4); |
337 | 0 | mean += PSHR32(x2, 12); |
338 | | #ifdef FIXED_POINT |
339 | | /* FIXME: Use PSHR16() instead */ |
340 | | mem0 = mem0 + PSHR32(x2-mem0,forward_shift); |
341 | | tmp[i] = PSHR32(mem0, 12); |
342 | | #else |
343 | 0 | mem0 = x2 + (1.f-forward_decay)*mem0; |
344 | 0 | tmp[i] = forward_decay*mem0; |
345 | 0 | #endif |
346 | 0 | } |
347 | |
|
348 | 0 | mem0=0; |
349 | 0 | maxE=0; |
350 | | /* Backward pass to compute the pre-echo threshold */ |
351 | 0 | for (i=len2-1;i>=0;i--) |
352 | 0 | { |
353 | | /* Backward masking: 13.9 dB/ms. */ |
354 | | #ifdef FIXED_POINT |
355 | | /* FIXME: Use PSHR16() instead */ |
356 | | mem0 = mem0 + PSHR32(SHL32(tmp[i],4)-mem0,3); |
357 | | tmp[i] = PSHR32(mem0, 4); |
358 | | maxE = MAX16(maxE, tmp[i]); |
359 | | #else |
360 | 0 | mem0 = tmp[i] + 0.875f*mem0; |
361 | 0 | tmp[i] = 0.125f*mem0; |
362 | 0 | maxE = MAX16(maxE, 0.125f*mem0); |
363 | 0 | #endif |
364 | 0 | } |
365 | | /*for (i=0;i<len2;i++)printf("%f ", tmp[i]/mean);printf("\n");*/ |
366 | | |
367 | | /* Compute the ratio of the "frame energy" over the harmonic mean of the energy. |
368 | | This essentially corresponds to a bitrate-normalized temporal noise-to-mask |
369 | | ratio */ |
370 | | |
371 | | /* As a compromise with the old transient detector, frame energy is the |
372 | | geometric mean of the energy and half the max */ |
373 | | #ifdef FIXED_POINT |
374 | | /* Costs two sqrt() to avoid overflows */ |
375 | | mean = MULT16_16(celt_sqrt(mean), celt_sqrt(MULT16_16(maxE,len2>>1))); |
376 | | #else |
377 | 0 | mean = celt_sqrt(mean * maxE*.5*len2); |
378 | 0 | #endif |
379 | | /* Inverse of the mean energy in Q15+6 */ |
380 | 0 | norm = SHL32(EXTEND32(len2),6+14)/ADD32(EPSILON,SHR32(mean,1)); |
381 | | /* Compute harmonic mean discarding the unreliable boundaries |
382 | | The data is smooth, so we only take 1/4th of the samples */ |
383 | 0 | unmask=0; |
384 | | /* We should never see NaNs here. If we find any, then something really bad happened and we better abort |
385 | | before it does any damage later on. If these asserts are disabled (no hardening), then the table |
386 | | lookup a few lines below (id = ...) is likely to crash dur to an out-of-bounds read. DO NOT FIX |
387 | | that crash on NaN since it could result in a worse issue later on. */ |
388 | 0 | celt_assert(!celt_isnan(tmp[0])); |
389 | 0 | celt_assert(!celt_isnan(norm)); |
390 | 0 | for (i=12;i<len2-5;i+=4) |
391 | 0 | { |
392 | 0 | int id; |
393 | | #ifdef FIXED_POINT |
394 | | id = MAX32(0,MIN32(127,MULT16_32_Q15(tmp[i]+EPSILON,norm))); /* Do not round to nearest */ |
395 | | #else |
396 | 0 | id = (int)MAX32(0,MIN32(127,floor(64*norm*(tmp[i]+EPSILON)))); /* Do not round to nearest */ |
397 | 0 | #endif |
398 | 0 | unmask += inv_table[id]; |
399 | 0 | } |
400 | | /*printf("%d\n", unmask);*/ |
401 | | /* Normalize, compensate for the 1/4th of the sample and the factor of 6 in the inverse table */ |
402 | 0 | unmask = 64*unmask*4/(6*(len2-17)); |
403 | 0 | if (unmask>mask_metric) |
404 | 0 | { |
405 | 0 | *tf_chan = c; |
406 | 0 | mask_metric = unmask; |
407 | 0 | } |
408 | 0 | } |
409 | 0 | is_transient = mask_metric>200; |
410 | | /* Prevent the transient detector from confusing the partial cycle of a |
411 | | very low frequency tone with a transient. */ |
412 | 0 | if (toneishness > QCONST32(.98f, 29) && tone_freq < QCONST16(0.026f, 13)) |
413 | 0 | is_transient = 0; |
414 | | /* For low bitrates, define "weak transients" that need to be |
415 | | handled differently to avoid partial collapse. */ |
416 | 0 | if (allow_weak_transients && is_transient && mask_metric<600) { |
417 | 0 | is_transient = 0; |
418 | 0 | *weak_transient = 1; |
419 | 0 | } |
420 | | /* Arbitrary metric for VBR boost */ |
421 | 0 | tf_max = MAX16(0,celt_sqrt(27*mask_metric)-42); |
422 | | /* *tf_estimate = 1 + MIN16(1, sqrt(MAX16(0, tf_max-30))/20); */ |
423 | 0 | *tf_estimate = celt_sqrt(MAX32(0, SHL32(MULT16_16(QCONST16(0.0069,14),MIN16(163,tf_max)),14)-QCONST32(0.139,28))); |
424 | | /*printf("%d %f\n", tf_max, mask_metric);*/ |
425 | 0 | RESTORE_STACK; |
426 | | #ifdef FUZZING |
427 | | is_transient = rand()&0x1; |
428 | | #endif |
429 | | /*printf("%d %f %d\n", is_transient, (float)*tf_estimate, tf_max);*/ |
430 | 0 | return is_transient; |
431 | 0 | } |
432 | | |
433 | | /* Looks for sudden increases of energy to decide whether we need to patch |
434 | | the transient decision */ |
435 | | static int patch_transient_decision(celt_glog *newE, celt_glog *oldE, int nbEBands, |
436 | | int start, int end, int C) |
437 | 0 | { |
438 | 0 | int i, c; |
439 | 0 | opus_val32 mean_diff=0; |
440 | 0 | celt_glog spread_old[26]; |
441 | | /* Apply an aggressive (-6 dB/Bark) spreading function to the old frame to |
442 | | avoid false detection caused by irrelevant bands */ |
443 | 0 | if (C==1) |
444 | 0 | { |
445 | 0 | spread_old[start] = oldE[start]; |
446 | 0 | for (i=start+1;i<end;i++) |
447 | 0 | spread_old[i] = MAXG(spread_old[i-1]-GCONST(1.0f), oldE[i]); |
448 | 0 | } else { |
449 | 0 | spread_old[start] = MAXG(oldE[start],oldE[start+nbEBands]); |
450 | 0 | for (i=start+1;i<end;i++) |
451 | 0 | spread_old[i] = MAXG(spread_old[i-1]-GCONST(1.0f), |
452 | 0 | MAXG(oldE[i],oldE[i+nbEBands])); |
453 | 0 | } |
454 | 0 | for (i=end-2;i>=start;i--) |
455 | 0 | spread_old[i] = MAXG(spread_old[i], spread_old[i+1]-GCONST(1.0f)); |
456 | | /* Compute mean increase */ |
457 | 0 | c=0; do { |
458 | 0 | for (i=IMAX(2,start);i<end-1;i++) |
459 | 0 | { |
460 | 0 | opus_val16 x1, x2; |
461 | 0 | x1 = MAXG(0, newE[i + c*nbEBands]); |
462 | 0 | x2 = MAXG(0, spread_old[i]); |
463 | 0 | mean_diff = ADD32(mean_diff, MAXG(0, SUB32(x1, x2))); |
464 | 0 | } |
465 | 0 | } while (++c<C); |
466 | 0 | mean_diff = DIV32(mean_diff, C*(end-1-IMAX(2,start))); |
467 | | /*printf("%f %f %d\n", mean_diff, max_diff, count);*/ |
468 | 0 | return mean_diff > GCONST(1.f); |
469 | 0 | } |
470 | | |
471 | | /** Apply window and compute the MDCT for all sub-frames and |
472 | | all channels in a frame */ |
473 | | static void compute_mdcts(const CELTMode *mode, int shortBlocks, celt_sig * OPUS_RESTRICT in, |
474 | | celt_sig * OPUS_RESTRICT out, int C, int CC, int LM, int upsample, |
475 | | int arch) |
476 | 0 | { |
477 | 0 | const int overlap = mode->overlap; |
478 | 0 | int N; |
479 | 0 | int B; |
480 | 0 | int shift; |
481 | 0 | int i, b, c; |
482 | 0 | if (shortBlocks) |
483 | 0 | { |
484 | 0 | B = shortBlocks; |
485 | 0 | N = mode->shortMdctSize; |
486 | 0 | shift = mode->maxLM; |
487 | 0 | } else { |
488 | 0 | B = 1; |
489 | 0 | N = mode->shortMdctSize<<LM; |
490 | 0 | shift = mode->maxLM-LM; |
491 | 0 | } |
492 | 0 | c=0; do { |
493 | 0 | for (b=0;b<B;b++) |
494 | 0 | { |
495 | | /* Interleaving the sub-frames while doing the MDCTs */ |
496 | 0 | clt_mdct_forward(&mode->mdct, in+c*(B*N+overlap)+b*N, |
497 | 0 | &out[b+c*N*B], mode->window, overlap, shift, B, |
498 | 0 | arch); |
499 | 0 | } |
500 | 0 | } while (++c<CC); |
501 | 0 | if (CC==2&&C==1) |
502 | 0 | { |
503 | 0 | for (i=0;i<B*N;i++) |
504 | 0 | out[i] = ADD32(HALF32(out[i]), HALF32(out[B*N+i])); |
505 | 0 | } |
506 | 0 | if (upsample != 1) |
507 | 0 | { |
508 | 0 | c=0; do |
509 | 0 | { |
510 | 0 | int bound = B*N/upsample; |
511 | 0 | for (i=0;i<bound;i++) |
512 | 0 | out[c*B*N+i] *= upsample; |
513 | 0 | OPUS_CLEAR(&out[c*B*N+bound], B*N-bound); |
514 | 0 | } while (++c<C); |
515 | 0 | } |
516 | 0 | } |
517 | | |
518 | | |
519 | | void celt_preemphasis(const opus_res * OPUS_RESTRICT pcmp, celt_sig * OPUS_RESTRICT inp, |
520 | | int N, int CC, int upsample, const opus_val16 *coef, celt_sig *mem, int clip) |
521 | 0 | { |
522 | 0 | int i; |
523 | 0 | opus_val16 coef0; |
524 | 0 | celt_sig m; |
525 | 0 | int Nu; |
526 | |
|
527 | 0 | coef0 = coef[0]; |
528 | 0 | m = *mem; |
529 | | |
530 | | /* Fast path for the normal 48kHz case and no clipping */ |
531 | 0 | if (coef[1] == 0 && upsample == 1 && !clip) |
532 | 0 | { |
533 | 0 | for (i=0;i<N;i++) |
534 | 0 | { |
535 | 0 | celt_sig x; |
536 | 0 | x = RES2SIG(pcmp[CC*i]); |
537 | | /* Apply pre-emphasis */ |
538 | 0 | inp[i] = x - m; |
539 | 0 | m = MULT16_32_Q15(coef0, x); |
540 | 0 | } |
541 | 0 | *mem = m; |
542 | 0 | return; |
543 | 0 | } |
544 | | |
545 | 0 | Nu = N/upsample; |
546 | 0 | if (upsample!=1) |
547 | 0 | { |
548 | 0 | OPUS_CLEAR(inp, N); |
549 | 0 | } |
550 | 0 | for (i=0;i<Nu;i++) |
551 | 0 | inp[i*upsample] = RES2SIG(pcmp[CC*i]); |
552 | |
|
553 | 0 | #ifndef FIXED_POINT |
554 | 0 | if (clip) |
555 | 0 | { |
556 | | /* Clip input to avoid encoding non-portable files */ |
557 | 0 | for (i=0;i<Nu;i++) |
558 | 0 | inp[i*upsample] = MAX32(-65536.f, MIN32(65536.f,inp[i*upsample])); |
559 | 0 | } |
560 | | #else |
561 | | (void)clip; /* Avoids a warning about clip being unused. */ |
562 | | #endif |
563 | | #ifdef CUSTOM_MODES |
564 | | if (coef[1] != 0) |
565 | | { |
566 | | opus_val16 coef1 = coef[1]; |
567 | | opus_val16 coef2 = coef[2]; |
568 | | for (i=0;i<N;i++) |
569 | | { |
570 | | celt_sig x, tmp; |
571 | | x = inp[i]; |
572 | | /* Apply pre-emphasis */ |
573 | | tmp = SHL32(MULT16_32_Q15(coef2, x), 15-SIG_SHIFT); |
574 | | inp[i] = tmp + m; |
575 | | m = MULT16_32_Q15(coef1, inp[i]) - MULT16_32_Q15(coef0, tmp); |
576 | | } |
577 | | } else |
578 | | #endif |
579 | 0 | { |
580 | 0 | for (i=0;i<N;i++) |
581 | 0 | { |
582 | 0 | celt_sig x; |
583 | 0 | x = inp[i]; |
584 | | /* Apply pre-emphasis */ |
585 | 0 | inp[i] = x - m; |
586 | 0 | m = MULT16_32_Q15(coef0, x); |
587 | 0 | } |
588 | 0 | } |
589 | 0 | *mem = m; |
590 | 0 | } |
591 | | |
592 | | |
593 | | |
594 | | static opus_val32 l1_metric(const celt_norm *tmp, int N, int LM, opus_val16 bias) |
595 | 0 | { |
596 | 0 | int i; |
597 | 0 | opus_val32 L1; |
598 | 0 | L1 = 0; |
599 | 0 | for (i=0;i<N;i++) |
600 | 0 | L1 += EXTEND32(ABS16(tmp[i])); |
601 | | /* When in doubt, prefer good freq resolution */ |
602 | 0 | L1 = MAC16_32_Q15(L1, LM*bias, L1); |
603 | 0 | return L1; |
604 | |
|
605 | 0 | } |
606 | | |
607 | | static int tf_analysis(const CELTMode *m, int len, int isTransient, |
608 | | int *tf_res, int lambda, celt_norm *X, int N0, int LM, |
609 | | opus_val16 tf_estimate, int tf_chan, int *importance) |
610 | 0 | { |
611 | 0 | int i; |
612 | 0 | VARDECL(int, metric); |
613 | 0 | int cost0; |
614 | 0 | int cost1; |
615 | 0 | VARDECL(int, path0); |
616 | 0 | VARDECL(int, path1); |
617 | 0 | VARDECL(celt_norm, tmp); |
618 | 0 | VARDECL(celt_norm, tmp_1); |
619 | 0 | int sel; |
620 | 0 | int selcost[2]; |
621 | 0 | int tf_select=0; |
622 | 0 | opus_val16 bias; |
623 | |
|
624 | 0 | SAVE_STACK; |
625 | 0 | bias = MULT16_16_Q14(QCONST16(.04f,15), MAX16(-QCONST16(.25f,14), QCONST16(.5f,14)-tf_estimate)); |
626 | | /*printf("%f ", bias);*/ |
627 | |
|
628 | 0 | ALLOC(metric, len, int); |
629 | 0 | ALLOC(tmp, (m->eBands[len]-m->eBands[len-1])<<LM, celt_norm); |
630 | 0 | ALLOC(tmp_1, (m->eBands[len]-m->eBands[len-1])<<LM, celt_norm); |
631 | 0 | ALLOC(path0, len, int); |
632 | 0 | ALLOC(path1, len, int); |
633 | |
|
634 | 0 | for (i=0;i<len;i++) |
635 | 0 | { |
636 | 0 | int k, N; |
637 | 0 | int narrow; |
638 | 0 | opus_val32 L1, best_L1; |
639 | 0 | int best_level=0; |
640 | 0 | N = (m->eBands[i+1]-m->eBands[i])<<LM; |
641 | | /* band is too narrow to be split down to LM=-1 */ |
642 | 0 | narrow = (m->eBands[i+1]-m->eBands[i])==1; |
643 | 0 | OPUS_COPY(tmp, &X[tf_chan*N0 + (m->eBands[i]<<LM)], N); |
644 | | /* Just add the right channel if we're in stereo */ |
645 | | /*if (C==2) |
646 | | for (j=0;j<N;j++) |
647 | | tmp[j] = ADD16(SHR16(tmp[j], 1),SHR16(X[N0+j+(m->eBands[i]<<LM)], 1));*/ |
648 | 0 | L1 = l1_metric(tmp, N, isTransient ? LM : 0, bias); |
649 | 0 | best_L1 = L1; |
650 | | /* Check the -1 case for transients */ |
651 | 0 | if (isTransient && !narrow) |
652 | 0 | { |
653 | 0 | OPUS_COPY(tmp_1, tmp, N); |
654 | 0 | haar1(tmp_1, N>>LM, 1<<LM); |
655 | 0 | L1 = l1_metric(tmp_1, N, LM+1, bias); |
656 | 0 | if (L1<best_L1) |
657 | 0 | { |
658 | 0 | best_L1 = L1; |
659 | 0 | best_level = -1; |
660 | 0 | } |
661 | 0 | } |
662 | | /*printf ("%f ", L1);*/ |
663 | 0 | for (k=0;k<LM+!(isTransient||narrow);k++) |
664 | 0 | { |
665 | 0 | int B; |
666 | |
|
667 | 0 | if (isTransient) |
668 | 0 | B = (LM-k-1); |
669 | 0 | else |
670 | 0 | B = k+1; |
671 | |
|
672 | 0 | haar1(tmp, N>>k, 1<<k); |
673 | |
|
674 | 0 | L1 = l1_metric(tmp, N, B, bias); |
675 | |
|
676 | 0 | if (L1 < best_L1) |
677 | 0 | { |
678 | 0 | best_L1 = L1; |
679 | 0 | best_level = k+1; |
680 | 0 | } |
681 | 0 | } |
682 | | /*printf ("%d ", isTransient ? LM-best_level : best_level);*/ |
683 | | /* metric is in Q1 to be able to select the mid-point (-0.5) for narrower bands */ |
684 | 0 | if (isTransient) |
685 | 0 | metric[i] = 2*best_level; |
686 | 0 | else |
687 | 0 | metric[i] = -2*best_level; |
688 | | /* For bands that can't be split to -1, set the metric to the half-way point to avoid |
689 | | biasing the decision */ |
690 | 0 | if (narrow && (metric[i]==0 || metric[i]==-2*LM)) |
691 | 0 | metric[i]-=1; |
692 | | /*printf("%d ", metric[i]/2 + (!isTransient)*LM);*/ |
693 | 0 | } |
694 | | /*printf("\n");*/ |
695 | | /* Search for the optimal tf resolution, including tf_select */ |
696 | 0 | tf_select = 0; |
697 | 0 | for (sel=0;sel<2;sel++) |
698 | 0 | { |
699 | 0 | cost0 = importance[0]*abs(metric[0]-2*tf_select_table[LM][4*isTransient+2*sel+0]); |
700 | 0 | cost1 = importance[0]*abs(metric[0]-2*tf_select_table[LM][4*isTransient+2*sel+1]) + (isTransient ? 0 : lambda); |
701 | 0 | for (i=1;i<len;i++) |
702 | 0 | { |
703 | 0 | int curr0, curr1; |
704 | 0 | curr0 = IMIN(cost0, cost1 + lambda); |
705 | 0 | curr1 = IMIN(cost0 + lambda, cost1); |
706 | 0 | cost0 = curr0 + importance[i]*abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*sel+0]); |
707 | 0 | cost1 = curr1 + importance[i]*abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*sel+1]); |
708 | 0 | } |
709 | 0 | cost0 = IMIN(cost0, cost1); |
710 | 0 | selcost[sel]=cost0; |
711 | 0 | } |
712 | | /* For now, we're conservative and only allow tf_select=1 for transients. |
713 | | * If tests confirm it's useful for non-transients, we could allow it. */ |
714 | 0 | if (selcost[1]<selcost[0] && isTransient) |
715 | 0 | tf_select=1; |
716 | 0 | cost0 = importance[0]*abs(metric[0]-2*tf_select_table[LM][4*isTransient+2*tf_select+0]); |
717 | 0 | cost1 = importance[0]*abs(metric[0]-2*tf_select_table[LM][4*isTransient+2*tf_select+1]) + (isTransient ? 0 : lambda); |
718 | | /* Viterbi forward pass */ |
719 | 0 | for (i=1;i<len;i++) |
720 | 0 | { |
721 | 0 | int curr0, curr1; |
722 | 0 | int from0, from1; |
723 | |
|
724 | 0 | from0 = cost0; |
725 | 0 | from1 = cost1 + lambda; |
726 | 0 | if (from0 < from1) |
727 | 0 | { |
728 | 0 | curr0 = from0; |
729 | 0 | path0[i]= 0; |
730 | 0 | } else { |
731 | 0 | curr0 = from1; |
732 | 0 | path0[i]= 1; |
733 | 0 | } |
734 | |
|
735 | 0 | from0 = cost0 + lambda; |
736 | 0 | from1 = cost1; |
737 | 0 | if (from0 < from1) |
738 | 0 | { |
739 | 0 | curr1 = from0; |
740 | 0 | path1[i]= 0; |
741 | 0 | } else { |
742 | 0 | curr1 = from1; |
743 | 0 | path1[i]= 1; |
744 | 0 | } |
745 | 0 | cost0 = curr0 + importance[i]*abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*tf_select+0]); |
746 | 0 | cost1 = curr1 + importance[i]*abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*tf_select+1]); |
747 | 0 | } |
748 | 0 | tf_res[len-1] = cost0 < cost1 ? 0 : 1; |
749 | | /* Viterbi backward pass to check the decisions */ |
750 | 0 | for (i=len-2;i>=0;i--) |
751 | 0 | { |
752 | 0 | if (tf_res[i+1] == 1) |
753 | 0 | tf_res[i] = path1[i+1]; |
754 | 0 | else |
755 | 0 | tf_res[i] = path0[i+1]; |
756 | 0 | } |
757 | | /*printf("%d %f\n", *tf_sum, tf_estimate);*/ |
758 | 0 | RESTORE_STACK; |
759 | | #ifdef FUZZING |
760 | | tf_select = rand()&0x1; |
761 | | tf_res[0] = rand()&0x1; |
762 | | for (i=1;i<len;i++) |
763 | | tf_res[i] = tf_res[i-1] ^ ((rand()&0xF) == 0); |
764 | | #endif |
765 | 0 | return tf_select; |
766 | 0 | } |
767 | | |
768 | | static void tf_encode(int start, int end, int isTransient, int *tf_res, int LM, int tf_select, ec_enc *enc) |
769 | 0 | { |
770 | 0 | int curr, i; |
771 | 0 | int tf_select_rsv; |
772 | 0 | int tf_changed; |
773 | 0 | int logp; |
774 | 0 | opus_uint32 budget; |
775 | 0 | opus_uint32 tell; |
776 | 0 | budget = enc->storage*8; |
777 | 0 | tell = ec_tell(enc); |
778 | 0 | logp = isTransient ? 2 : 4; |
779 | | /* Reserve space to code the tf_select decision. */ |
780 | 0 | tf_select_rsv = LM>0 && tell+logp+1 <= budget; |
781 | 0 | budget -= tf_select_rsv; |
782 | 0 | curr = tf_changed = 0; |
783 | 0 | for (i=start;i<end;i++) |
784 | 0 | { |
785 | 0 | if (tell+logp<=budget) |
786 | 0 | { |
787 | 0 | ec_enc_bit_logp(enc, tf_res[i] ^ curr, logp); |
788 | 0 | tell = ec_tell(enc); |
789 | 0 | curr = tf_res[i]; |
790 | 0 | tf_changed |= curr; |
791 | 0 | } |
792 | 0 | else |
793 | 0 | tf_res[i] = curr; |
794 | 0 | logp = isTransient ? 4 : 5; |
795 | 0 | } |
796 | | /* Only code tf_select if it would actually make a difference. */ |
797 | 0 | if (tf_select_rsv && |
798 | 0 | tf_select_table[LM][4*isTransient+0+tf_changed]!= |
799 | 0 | tf_select_table[LM][4*isTransient+2+tf_changed]) |
800 | 0 | ec_enc_bit_logp(enc, tf_select, 1); |
801 | 0 | else |
802 | 0 | tf_select = 0; |
803 | 0 | for (i=start;i<end;i++) |
804 | 0 | tf_res[i] = tf_select_table[LM][4*isTransient+2*tf_select+tf_res[i]]; |
805 | | /*for(i=0;i<end;i++)printf("%d ", isTransient ? tf_res[i] : LM+tf_res[i]);printf("\n");*/ |
806 | 0 | } |
807 | | |
808 | | |
809 | | static int alloc_trim_analysis(const CELTMode *m, const celt_norm *X, |
810 | | const celt_glog *bandLogE, int end, int LM, int C, int N0, |
811 | | AnalysisInfo *analysis, opus_val16 *stereo_saving, opus_val16 tf_estimate, |
812 | | int intensity, celt_glog surround_trim, opus_int32 equiv_rate, int arch) |
813 | 0 | { |
814 | 0 | int i; |
815 | 0 | opus_val32 diff=0; |
816 | 0 | int c; |
817 | 0 | int trim_index; |
818 | 0 | opus_val16 trim = QCONST16(5.f, 8); |
819 | 0 | opus_val16 logXC, logXC2; |
820 | | /* At low bitrate, reducing the trim seems to help. At higher bitrates, it's less |
821 | | clear what's best, so we're keeping it as it was before, at least for now. */ |
822 | 0 | if (equiv_rate < 64000) { |
823 | 0 | trim = QCONST16(4.f, 8); |
824 | 0 | } else if (equiv_rate < 80000) { |
825 | 0 | opus_int32 frac = (equiv_rate-64000) >> 10; |
826 | 0 | trim = QCONST16(4.f, 8) + QCONST16(1.f/16.f, 8)*frac; |
827 | 0 | } |
828 | 0 | if (C==2) |
829 | 0 | { |
830 | 0 | opus_val16 sum = 0; /* Q10 */ |
831 | 0 | opus_val16 minXC; /* Q10 */ |
832 | | /* Compute inter-channel correlation for low frequencies */ |
833 | 0 | for (i=0;i<8;i++) |
834 | 0 | { |
835 | 0 | opus_val32 partial; |
836 | 0 | partial = celt_inner_prod(&X[m->eBands[i]<<LM], &X[N0+(m->eBands[i]<<LM)], |
837 | 0 | (m->eBands[i+1]-m->eBands[i])<<LM, arch); |
838 | 0 | sum = ADD16(sum, EXTRACT16(SHR32(partial, 18))); |
839 | 0 | } |
840 | 0 | sum = MULT16_16_Q15(QCONST16(1.f/8, 15), sum); |
841 | 0 | sum = MIN16(QCONST16(1.f, 10), ABS16(sum)); |
842 | 0 | minXC = sum; |
843 | 0 | for (i=8;i<intensity;i++) |
844 | 0 | { |
845 | 0 | opus_val32 partial; |
846 | 0 | partial = celt_inner_prod(&X[m->eBands[i]<<LM], &X[N0+(m->eBands[i]<<LM)], |
847 | 0 | (m->eBands[i+1]-m->eBands[i])<<LM, arch); |
848 | 0 | minXC = MIN16(minXC, ABS16(EXTRACT16(SHR32(partial, 18)))); |
849 | 0 | } |
850 | 0 | minXC = MIN16(QCONST16(1.f, 10), ABS16(minXC)); |
851 | | /*printf ("%f\n", sum);*/ |
852 | | /* mid-side savings estimations based on the LF average*/ |
853 | 0 | logXC = celt_log2(QCONST32(1.001f, 20)-MULT16_16(sum, sum)); |
854 | | /* mid-side savings estimations based on min correlation */ |
855 | 0 | logXC2 = MAX16(HALF16(logXC), celt_log2(QCONST32(1.001f, 20)-MULT16_16(minXC, minXC))); |
856 | | #ifdef FIXED_POINT |
857 | | /* Compensate for Q20 vs Q14 input and convert output to Q8 */ |
858 | | logXC = PSHR32(logXC-QCONST16(6.f, 10),10-8); |
859 | | logXC2 = PSHR32(logXC2-QCONST16(6.f, 10),10-8); |
860 | | #endif |
861 | |
|
862 | 0 | trim += MAX16(-QCONST16(4.f, 8), MULT16_16_Q15(QCONST16(.75f,15),logXC)); |
863 | 0 | *stereo_saving = MIN16(*stereo_saving + QCONST16(0.25f, 8), -HALF16(logXC2)); |
864 | 0 | } |
865 | | |
866 | | /* Estimate spectral tilt */ |
867 | 0 | c=0; do { |
868 | 0 | for (i=0;i<end-1;i++) |
869 | 0 | { |
870 | 0 | diff += SHR32(bandLogE[i+c*m->nbEBands], 5)*(opus_int32)(2+2*i-end); |
871 | 0 | } |
872 | 0 | } while (++c<C); |
873 | 0 | diff /= C*(end-1); |
874 | | /*printf("%f\n", diff);*/ |
875 | 0 | trim -= MAX32(-QCONST16(2.f, 8), MIN32(QCONST16(2.f, 8), SHR32(diff+QCONST32(1.f, DB_SHIFT-5),DB_SHIFT-13)/6 )); |
876 | 0 | trim -= SHR16(surround_trim, DB_SHIFT-8); |
877 | 0 | trim -= 2*SHR16(tf_estimate, 14-8); |
878 | 0 | #ifndef DISABLE_FLOAT_API |
879 | 0 | if (analysis->valid) |
880 | 0 | { |
881 | 0 | trim -= MAX16(-QCONST16(2.f, 8), MIN16(QCONST16(2.f, 8), |
882 | 0 | (opus_val16)(QCONST16(2.f, 8)*(analysis->tonality_slope+.05f)))); |
883 | 0 | } |
884 | | #else |
885 | | (void)analysis; |
886 | | #endif |
887 | |
|
888 | | #ifdef FIXED_POINT |
889 | | trim_index = PSHR32(trim, 8); |
890 | | #else |
891 | 0 | trim_index = (int)floor(.5f+trim); |
892 | 0 | #endif |
893 | 0 | trim_index = IMAX(0, IMIN(10, trim_index)); |
894 | | /*printf("%d\n", trim_index);*/ |
895 | | #ifdef FUZZING |
896 | | trim_index = rand()%11; |
897 | | #endif |
898 | 0 | return trim_index; |
899 | 0 | } |
900 | | |
901 | | static int stereo_analysis(const CELTMode *m, const celt_norm *X, |
902 | | int LM, int N0) |
903 | 0 | { |
904 | 0 | int i; |
905 | 0 | int thetas; |
906 | 0 | opus_val32 sumLR = EPSILON, sumMS = EPSILON; |
907 | | |
908 | | /* Use the L1 norm to model the entropy of the L/R signal vs the M/S signal */ |
909 | 0 | for (i=0;i<13;i++) |
910 | 0 | { |
911 | 0 | int j; |
912 | 0 | for (j=m->eBands[i]<<LM;j<m->eBands[i+1]<<LM;j++) |
913 | 0 | { |
914 | 0 | opus_val32 L, R, M, S; |
915 | | /* We cast to 32-bit first because of the -32768 case */ |
916 | 0 | L = EXTEND32(X[j]); |
917 | 0 | R = EXTEND32(X[N0+j]); |
918 | 0 | M = ADD32(L, R); |
919 | 0 | S = SUB32(L, R); |
920 | 0 | sumLR = ADD32(sumLR, ADD32(ABS32(L), ABS32(R))); |
921 | 0 | sumMS = ADD32(sumMS, ADD32(ABS32(M), ABS32(S))); |
922 | 0 | } |
923 | 0 | } |
924 | 0 | sumMS = MULT16_32_Q15(QCONST16(0.707107f, 15), sumMS); |
925 | 0 | thetas = 13; |
926 | | /* We don't need thetas for lower bands with LM<=1 */ |
927 | 0 | if (LM<=1) |
928 | 0 | thetas -= 8; |
929 | 0 | return MULT16_32_Q15((m->eBands[13]<<(LM+1))+thetas, sumMS) |
930 | 0 | > MULT16_32_Q15(m->eBands[13]<<(LM+1), sumLR); |
931 | 0 | } |
932 | | |
933 | 0 | #define MSWAP(a,b) do {celt_glog tmp = a;a=b;b=tmp;} while(0) |
934 | | static celt_glog median_of_5(const celt_glog *x) |
935 | 0 | { |
936 | 0 | celt_glog t0, t1, t2, t3, t4; |
937 | 0 | t2 = x[2]; |
938 | 0 | if (x[0] > x[1]) |
939 | 0 | { |
940 | 0 | t0 = x[1]; |
941 | 0 | t1 = x[0]; |
942 | 0 | } else { |
943 | 0 | t0 = x[0]; |
944 | 0 | t1 = x[1]; |
945 | 0 | } |
946 | 0 | if (x[3] > x[4]) |
947 | 0 | { |
948 | 0 | t3 = x[4]; |
949 | 0 | t4 = x[3]; |
950 | 0 | } else { |
951 | 0 | t3 = x[3]; |
952 | 0 | t4 = x[4]; |
953 | 0 | } |
954 | 0 | if (t0 > t3) |
955 | 0 | { |
956 | 0 | MSWAP(t0, t3); |
957 | 0 | MSWAP(t1, t4); |
958 | 0 | } |
959 | 0 | if (t2 > t1) |
960 | 0 | { |
961 | 0 | if (t1 < t3) |
962 | 0 | return MING(t2, t3); |
963 | 0 | else |
964 | 0 | return MING(t4, t1); |
965 | 0 | } else { |
966 | 0 | if (t2 < t3) |
967 | 0 | return MING(t1, t3); |
968 | 0 | else |
969 | 0 | return MING(t2, t4); |
970 | 0 | } |
971 | 0 | } |
972 | | |
973 | | static celt_glog median_of_3(const celt_glog *x) |
974 | 0 | { |
975 | 0 | celt_glog t0, t1, t2; |
976 | 0 | if (x[0] > x[1]) |
977 | 0 | { |
978 | 0 | t0 = x[1]; |
979 | 0 | t1 = x[0]; |
980 | 0 | } else { |
981 | 0 | t0 = x[0]; |
982 | 0 | t1 = x[1]; |
983 | 0 | } |
984 | 0 | t2 = x[2]; |
985 | 0 | if (t1 < t2) |
986 | 0 | return t1; |
987 | 0 | else if (t0 < t2) |
988 | 0 | return t2; |
989 | 0 | else |
990 | 0 | return t0; |
991 | 0 | } |
992 | | |
993 | | static celt_glog dynalloc_analysis(const celt_glog *bandLogE, const celt_glog *bandLogE2, const celt_glog *oldBandE, |
994 | | int nbEBands, int start, int end, int C, int *offsets, int lsb_depth, const opus_int16 *logN, |
995 | | int isTransient, int vbr, int constrained_vbr, const opus_int16 *eBands, int LM, |
996 | | int effectiveBytes, opus_int32 *tot_boost_, int lfe, celt_glog *surround_dynalloc, |
997 | | AnalysisInfo *analysis, int *importance, int *spread_weight, opus_val16 tone_freq, opus_val32 toneishness) |
998 | 0 | { |
999 | 0 | int i, c; |
1000 | 0 | opus_int32 tot_boost=0; |
1001 | 0 | celt_glog maxDepth; |
1002 | 0 | VARDECL(celt_glog, follower); |
1003 | 0 | VARDECL(celt_glog, noise_floor); |
1004 | 0 | VARDECL(celt_glog, bandLogE3); |
1005 | 0 | SAVE_STACK; |
1006 | 0 | ALLOC(follower, C*nbEBands, celt_glog); |
1007 | 0 | ALLOC(noise_floor, C*nbEBands, celt_glog); |
1008 | 0 | ALLOC(bandLogE3, nbEBands, celt_glog); |
1009 | 0 | OPUS_CLEAR(offsets, nbEBands); |
1010 | | /* Dynamic allocation code */ |
1011 | 0 | maxDepth=-GCONST(31.9f); |
1012 | 0 | for (i=0;i<end;i++) |
1013 | 0 | { |
1014 | | /* Noise floor must take into account eMeans, the depth, the width of the bands |
1015 | | and the preemphasis filter (approx. square of bark band ID) */ |
1016 | 0 | noise_floor[i] = GCONST(0.0625f)*logN[i] |
1017 | 0 | +GCONST(.5f)+SHL32(9-lsb_depth,DB_SHIFT)-SHL32(eMeans[i],DB_SHIFT-4) |
1018 | 0 | +GCONST(.0062f)*(i+5)*(i+5); |
1019 | 0 | } |
1020 | 0 | c=0;do |
1021 | 0 | { |
1022 | 0 | for (i=0;i<end;i++) |
1023 | 0 | maxDepth = MAXG(maxDepth, bandLogE[c*nbEBands+i]-noise_floor[i]); |
1024 | 0 | } while (++c<C); |
1025 | 0 | { |
1026 | | /* Compute a really simple masking model to avoid taking into account completely masked |
1027 | | bands when computing the spreading decision. */ |
1028 | 0 | VARDECL(celt_glog, mask); |
1029 | 0 | VARDECL(celt_glog, sig); |
1030 | 0 | ALLOC(mask, nbEBands, celt_glog); |
1031 | 0 | ALLOC(sig, nbEBands, celt_glog); |
1032 | 0 | for (i=0;i<end;i++) |
1033 | 0 | mask[i] = bandLogE[i]-noise_floor[i]; |
1034 | 0 | if (C==2) |
1035 | 0 | { |
1036 | 0 | for (i=0;i<end;i++) |
1037 | 0 | mask[i] = MAXG(mask[i], bandLogE[nbEBands+i]-noise_floor[i]); |
1038 | 0 | } |
1039 | 0 | OPUS_COPY(sig, mask, end); |
1040 | 0 | for (i=1;i<end;i++) |
1041 | 0 | mask[i] = MAXG(mask[i], mask[i-1] - GCONST(2.f)); |
1042 | 0 | for (i=end-2;i>=0;i--) |
1043 | 0 | mask[i] = MAXG(mask[i], mask[i+1] - GCONST(3.f)); |
1044 | 0 | for (i=0;i<end;i++) |
1045 | 0 | { |
1046 | | /* Compute SMR: Mask is never more than 72 dB below the peak and never below the noise floor.*/ |
1047 | 0 | celt_glog smr = sig[i]-MAXG(MAXG(0, maxDepth-GCONST(12.f)), mask[i]); |
1048 | | /* Clamp SMR to make sure we're not shifting by something negative or too large. */ |
1049 | | #ifdef FIXED_POINT |
1050 | | /* FIXME: Use PSHR16() instead */ |
1051 | | int shift = -PSHR32(MAXG(-GCONST(5.f), MING(0, smr)), DB_SHIFT); |
1052 | | #else |
1053 | 0 | int shift = IMIN(5, IMAX(0, -(int)floor(.5f + smr))); |
1054 | 0 | #endif |
1055 | 0 | spread_weight[i] = 32 >> shift; |
1056 | 0 | } |
1057 | | /*for (i=0;i<end;i++) |
1058 | | printf("%d ", spread_weight[i]); |
1059 | | printf("\n");*/ |
1060 | 0 | } |
1061 | | /* Make sure that dynamic allocation can't make us bust the budget. |
1062 | | We enable the feature starting at 24 kb/s for 20-ms frames |
1063 | | and 96 kb/s for 2.5 ms frames. */ |
1064 | 0 | if (effectiveBytes >= (30 + 5*LM) && !lfe) |
1065 | 0 | { |
1066 | 0 | int last=0; |
1067 | 0 | c=0;do |
1068 | 0 | { |
1069 | 0 | celt_glog offset; |
1070 | 0 | celt_glog tmp; |
1071 | 0 | celt_glog *f; |
1072 | 0 | OPUS_COPY(bandLogE3, &bandLogE2[c*nbEBands], end); |
1073 | 0 | if (LM==0) { |
1074 | | /* For 2.5 ms frames, the first 8 bands have just one bin, so the |
1075 | | energy is highly unreliable (high variance). For that reason, |
1076 | | we take the max with the previous energy so that at least 2 bins |
1077 | | are getting used. */ |
1078 | 0 | for (i=0;i<IMIN(8,end);i++) bandLogE3[i] = MAXG(bandLogE2[c*nbEBands+i], oldBandE[c*nbEBands+i]); |
1079 | 0 | } |
1080 | 0 | f = &follower[c*nbEBands]; |
1081 | 0 | f[0] = bandLogE3[0]; |
1082 | 0 | for (i=1;i<end;i++) |
1083 | 0 | { |
1084 | | /* The last band to be at least 3 dB higher than the previous one |
1085 | | is the last we'll consider. Otherwise, we run into problems on |
1086 | | bandlimited signals. */ |
1087 | 0 | if (bandLogE3[i] > bandLogE3[i-1]+GCONST(.5f)) |
1088 | 0 | last=i; |
1089 | 0 | f[i] = MING(f[i-1]+GCONST(1.5f), bandLogE3[i]); |
1090 | 0 | } |
1091 | 0 | for (i=last-1;i>=0;i--) |
1092 | 0 | f[i] = MING(f[i], MING(f[i+1]+GCONST(2.f), bandLogE3[i])); |
1093 | | |
1094 | | /* Combine with a median filter to avoid dynalloc triggering unnecessarily. |
1095 | | The "offset" value controls how conservative we are -- a higher offset |
1096 | | reduces the impact of the median filter and makes dynalloc use more bits. */ |
1097 | 0 | offset = GCONST(1.f); |
1098 | 0 | for (i=2;i<end-2;i++) |
1099 | 0 | f[i] = MAXG(f[i], median_of_5(&bandLogE3[i-2])-offset); |
1100 | 0 | tmp = median_of_3(&bandLogE3[0])-offset; |
1101 | 0 | f[0] = MAXG(f[0], tmp); |
1102 | 0 | f[1] = MAXG(f[1], tmp); |
1103 | 0 | tmp = median_of_3(&bandLogE3[end-3])-offset; |
1104 | 0 | f[end-2] = MAXG(f[end-2], tmp); |
1105 | 0 | f[end-1] = MAXG(f[end-1], tmp); |
1106 | |
|
1107 | 0 | for (i=0;i<end;i++) |
1108 | 0 | f[i] = MAXG(f[i], noise_floor[i]); |
1109 | 0 | } while (++c<C); |
1110 | 0 | if (C==2) |
1111 | 0 | { |
1112 | 0 | for (i=start;i<end;i++) |
1113 | 0 | { |
1114 | | /* Consider 24 dB "cross-talk" */ |
1115 | 0 | follower[nbEBands+i] = MAXG(follower[nbEBands+i], follower[ i]-GCONST(4.f)); |
1116 | 0 | follower[ i] = MAXG(follower[ i], follower[nbEBands+i]-GCONST(4.f)); |
1117 | 0 | follower[i] = HALF32(MAXG(0, bandLogE[i]-follower[i]) + MAXG(0, bandLogE[nbEBands+i]-follower[nbEBands+i])); |
1118 | 0 | } |
1119 | 0 | } else { |
1120 | 0 | for (i=start;i<end;i++) |
1121 | 0 | { |
1122 | 0 | follower[i] = MAXG(0, bandLogE[i]-follower[i]); |
1123 | 0 | } |
1124 | 0 | } |
1125 | 0 | for (i=start;i<end;i++) |
1126 | 0 | follower[i] = MAXG(follower[i], surround_dynalloc[i]); |
1127 | 0 | for (i=start;i<end;i++) |
1128 | 0 | { |
1129 | | #ifdef FIXED_POINT |
1130 | | importance[i] = PSHR32(13*celt_exp2_db(MING(follower[i], GCONST(4.f))), 16); |
1131 | | #else |
1132 | 0 | importance[i] = (int)floor(.5f+13*celt_exp2_db(MING(follower[i], GCONST(4.f)))); |
1133 | 0 | #endif |
1134 | 0 | } |
1135 | | /* For non-transient CBR/CVBR frames, halve the dynalloc contribution */ |
1136 | 0 | if ((!vbr || constrained_vbr)&&!isTransient) |
1137 | 0 | { |
1138 | 0 | for (i=start;i<end;i++) |
1139 | 0 | follower[i] = HALF32(follower[i]); |
1140 | 0 | } |
1141 | 0 | for (i=start;i<end;i++) |
1142 | 0 | { |
1143 | 0 | if (i<8) |
1144 | 0 | follower[i] *= 2; |
1145 | 0 | if (i>=12) |
1146 | 0 | follower[i] = HALF32(follower[i]); |
1147 | 0 | } |
1148 | | /* Compensate for Opus' under-allocation on tones. */ |
1149 | 0 | if (toneishness > QCONST32(.98f, 29)) { |
1150 | | #ifdef FIXED_POINT |
1151 | | int freq_bin = PSHR32(MULT16_16(tone_freq, QCONST16(120/M_PI, 9)), 13+9); |
1152 | | #else |
1153 | 0 | int freq_bin = (int)floor(.5 + tone_freq*120/M_PI); |
1154 | 0 | #endif |
1155 | 0 | for (i=start;i<end;i++) { |
1156 | 0 | if (freq_bin >= eBands[i] && freq_bin <= eBands[i+1]) follower[i] += GCONST(2.f); |
1157 | 0 | if (freq_bin >= eBands[i]-1 && freq_bin <= eBands[i+1]+1) follower[i] += GCONST(1.f); |
1158 | 0 | if (freq_bin >= eBands[i]-2 && freq_bin <= eBands[i+1]+2) follower[i] += GCONST(1.f); |
1159 | 0 | if (freq_bin >= eBands[i]-3 && freq_bin <= eBands[i+1]+3) follower[i] += GCONST(.5f); |
1160 | 0 | } |
1161 | 0 | } |
1162 | | #ifdef DISABLE_FLOAT_API |
1163 | | (void)analysis; |
1164 | | #else |
1165 | 0 | if (analysis->valid) |
1166 | 0 | { |
1167 | 0 | for (i=start;i<IMIN(LEAK_BANDS, end);i++) |
1168 | 0 | follower[i] = follower[i] + GCONST(1.f/64.f)*analysis->leak_boost[i]; |
1169 | 0 | } |
1170 | 0 | #endif |
1171 | 0 | for (i=start;i<end;i++) |
1172 | 0 | { |
1173 | 0 | int width; |
1174 | 0 | int boost; |
1175 | 0 | int boost_bits; |
1176 | |
|
1177 | 0 | follower[i] = MING(follower[i], GCONST(4)); |
1178 | |
|
1179 | 0 | follower[i] = SHR32(follower[i], 8); |
1180 | 0 | width = C*(eBands[i+1]-eBands[i])<<LM; |
1181 | 0 | if (width<6) |
1182 | 0 | { |
1183 | 0 | boost = (int)SHR32(follower[i],DB_SHIFT-8); |
1184 | 0 | boost_bits = boost*width<<BITRES; |
1185 | 0 | } else if (width > 48) { |
1186 | 0 | boost = (int)SHR32(follower[i]*8,DB_SHIFT-8); |
1187 | 0 | boost_bits = (boost*width<<BITRES)/8; |
1188 | 0 | } else { |
1189 | 0 | boost = (int)SHR32(follower[i]*width/6,DB_SHIFT-8); |
1190 | 0 | boost_bits = boost*6<<BITRES; |
1191 | 0 | } |
1192 | | /* For CBR and non-transient CVBR frames, limit dynalloc to 2/3 of the bits */ |
1193 | 0 | if ((!vbr || (constrained_vbr&&!isTransient)) |
1194 | 0 | && (tot_boost+boost_bits)>>BITRES>>3 > 2*effectiveBytes/3) |
1195 | 0 | { |
1196 | 0 | opus_int32 cap = ((2*effectiveBytes/3)<<BITRES<<3); |
1197 | 0 | offsets[i] = cap-tot_boost; |
1198 | 0 | tot_boost = cap; |
1199 | 0 | break; |
1200 | 0 | } else { |
1201 | 0 | offsets[i] = boost; |
1202 | 0 | tot_boost += boost_bits; |
1203 | 0 | } |
1204 | 0 | } |
1205 | 0 | } else { |
1206 | 0 | for (i=start;i<end;i++) |
1207 | 0 | importance[i] = 13; |
1208 | 0 | } |
1209 | 0 | *tot_boost_ = tot_boost; |
1210 | 0 | RESTORE_STACK; |
1211 | 0 | return maxDepth; |
1212 | 0 | } |
1213 | | |
1214 | | #ifdef FIXED_POINT |
1215 | | void normalize_tone_input(opus_val16 *x, int len) { |
1216 | | opus_val32 ac0=len; |
1217 | | int i; |
1218 | | int shift; |
1219 | | for (i=0;i<len;i++) { |
1220 | | ac0 = ADD32(ac0, SHR32(MULT16_16(x[i], x[i]), 10)); |
1221 | | } |
1222 | | shift = 5 - (28-celt_ilog2(ac0))/2; |
1223 | | if (shift > 0) { |
1224 | | for (i=0;i<len;i++) { |
1225 | | x[i] = PSHR32(x[i], shift); |
1226 | | } |
1227 | | } |
1228 | | } |
1229 | | int acos_approx(opus_val32 x) { |
1230 | | opus_val16 x14; |
1231 | | opus_val32 tmp; |
1232 | | int flip = x<0; |
1233 | | x = abs(x); |
1234 | | x14 = x>>15; |
1235 | | tmp = (762*x14>>14)-3308; |
1236 | | tmp = (tmp*x14>>14)+25726; |
1237 | | tmp = tmp*celt_sqrt(IMAX(0, (1<<30) - (x<<1)))>>16; |
1238 | | if (flip) tmp = 25736 - tmp; |
1239 | | return tmp; |
1240 | | } |
1241 | | #endif |
1242 | | |
1243 | | /* Compute the LPC coefficients using a least-squares fit for both forward and backward prediction. */ |
1244 | 0 | static int tone_lpc(const opus_val16 *x, int len, int delay, opus_val32 *lpc) { |
1245 | 0 | int i; |
1246 | 0 | opus_val32 r00=0, r01=0, r11=0, r02=0, r12=0, r22=0; |
1247 | 0 | opus_val32 edges; |
1248 | 0 | opus_val32 num0, num1, den; |
1249 | 0 | celt_assert(len > 2*delay); |
1250 | | /* Compute correlations as if using the forward prediction covariance method. */ |
1251 | 0 | for (i=0;i<len-2*delay;i++) { |
1252 | 0 | r00 += MULT16_16(x[i],x[i]); |
1253 | 0 | r01 += MULT16_16(x[i],x[i+delay]); |
1254 | 0 | r02 += MULT16_16(x[i],x[i+2*delay]); |
1255 | 0 | } |
1256 | 0 | edges = 0; |
1257 | 0 | for (i=0;i<delay;i++) edges += MULT16_16(x[len+i-2*delay],x[len+i-2*delay]) - MULT16_16(x[i],x[i]); |
1258 | 0 | r11 = r00+edges; |
1259 | 0 | edges = 0; |
1260 | 0 | for (i=0;i<delay;i++) edges += MULT16_16(x[len+i-delay],x[len+i-delay]) - MULT16_16(x[i+delay],x[i+delay]); |
1261 | 0 | r22 = r11+edges; |
1262 | 0 | edges = 0; |
1263 | 0 | for (i=0;i<delay;i++) edges += MULT16_16(x[len+i-2*delay],x[len+i-delay]) - MULT16_16(x[i],x[i+delay]); |
1264 | 0 | r12 = r01+edges; |
1265 | | /* Reverse and sum to get the backward contribution. */ |
1266 | 0 | { |
1267 | 0 | opus_val32 R00, R01, R11, R02, R12, R22; |
1268 | 0 | R00 = r00 + r22; |
1269 | 0 | R01 = r01 + r12; |
1270 | 0 | R11 = 2*r11; |
1271 | 0 | R02 = 2*r02; |
1272 | 0 | R12 = r12 + r01; |
1273 | 0 | R22 = r00 + r22; |
1274 | 0 | r00 = R00; |
1275 | 0 | r01 = R01; |
1276 | 0 | r11 = R11; |
1277 | 0 | r02 = R02; |
1278 | 0 | r12 = R12; |
1279 | 0 | r22 = R22; |
1280 | 0 | } |
1281 | | /* Solve A*x=b, where A=[r00, r01; r01, r11] and b=[r02; r12]. */ |
1282 | 0 | den = MULT32_32_Q31(r00,r11) - MULT32_32_Q31(r01,r01); |
1283 | | #ifdef FIXED_POINT |
1284 | | if (den <= SHR32(MULT32_32_Q31(r00,r11), 10)) return 1; |
1285 | | #else |
1286 | 0 | if (den < .001f*MULT32_32_Q31(r00,r11)) return 1; |
1287 | 0 | #endif |
1288 | 0 | num1 = MULT32_32_Q31(r02,r11) - MULT32_32_Q31(r01,r12); |
1289 | 0 | if (num1 >= den) lpc[1] = QCONST32(1.f, 29); |
1290 | 0 | else if (num1 <= -den) lpc[1] = -QCONST32(1.f, 29); |
1291 | 0 | else lpc[1] = frac_div32_q29(num1, den); |
1292 | 0 | num0 = MULT32_32_Q31(r00,r12) - MULT32_32_Q31(r02,r01); |
1293 | 0 | if (HALF32(num0) >= den) lpc[0] = QCONST32(1.999999f, 29); |
1294 | 0 | else if (HALF32(num0) <= -den) lpc[0] = -QCONST32(1.999999f, 29); |
1295 | 0 | else lpc[0] = frac_div32_q29(num0, den); |
1296 | | /*printf("%f %f\n", lpc[0], lpc[1]);*/ |
1297 | 0 | return 0; |
1298 | 0 | } |
1299 | | |
1300 | | /* Detects pure of nearly pure tones so we can prevent them from causing problems with the encoder. */ |
1301 | 0 | static opus_val16 tone_detect(const celt_sig *in, int CC, int N, opus_val32 *toneishness, opus_int32 Fs) { |
1302 | 0 | int i; |
1303 | 0 | int delay = 1; |
1304 | 0 | int fail; |
1305 | 0 | opus_val32 lpc[2]; |
1306 | 0 | opus_val16 freq; |
1307 | 0 | VARDECL(opus_val16, x); |
1308 | 0 | ALLOC(x, N, opus_val16); |
1309 | | /* Shift by SIG_SHIFT+1 (+2 for stereo) to account for HF gain of the preemphasis filter. */ |
1310 | 0 | if (CC==2) { |
1311 | 0 | for (i=0;i<N;i++) x[i] = PSHR32(ADD32(in[i], in[i+N]), SIG_SHIFT+3); |
1312 | 0 | } else { |
1313 | 0 | for (i=0;i<N;i++) x[i] = PSHR32(in[i], SIG_SHIFT+2); |
1314 | 0 | } |
1315 | | #ifdef FIXED_POINT |
1316 | | normalize_tone_input(x, N); |
1317 | | #endif |
1318 | 0 | fail = tone_lpc(x, N, delay, lpc); |
1319 | | /* If our LPC filter resonates too close to DC, retry the analysis with down-sampling. */ |
1320 | 0 | while (delay <= Fs/3000 && (fail || (lpc[0] > QCONST32(1.f, 29) && lpc[1] < 0))) { |
1321 | 0 | delay *= 2; |
1322 | 0 | fail = tone_lpc(x, N, delay, lpc); |
1323 | 0 | } |
1324 | | /* Check that our filter has complex roots. */ |
1325 | 0 | if (!fail && MULT32_32_Q31(lpc[0],lpc[0]) + MULT32_32_Q31(QCONST32(3.999999, 29), lpc[1]) < 0) { |
1326 | | /* Squared radius of the poles. */ |
1327 | 0 | *toneishness = -lpc[1]; |
1328 | | #ifdef FIXED_POINT |
1329 | | freq = acos_approx(lpc[0]>>1)/delay; |
1330 | | #else |
1331 | 0 | freq = acos(.5f*lpc[0])/delay; |
1332 | 0 | #endif |
1333 | 0 | } else { |
1334 | 0 | freq = -1; |
1335 | 0 | *toneishness=0; |
1336 | 0 | } |
1337 | | /*printf("%f %f %f %f\n", freq, lpc[0], lpc[1], *toneishness);*/ |
1338 | 0 | return freq; |
1339 | 0 | } |
1340 | | |
1341 | | static int run_prefilter(CELTEncoder *st, celt_sig *in, celt_sig *prefilter_mem, int CC, int N, |
1342 | | int prefilter_tapset, int *pitch, opus_val16 *gain, int *qgain, int enabled, opus_val16 tf_estimate, int nbAvailableBytes, AnalysisInfo *analysis, opus_val16 tone_freq, opus_val32 toneishness) |
1343 | 0 | { |
1344 | 0 | int c; |
1345 | 0 | VARDECL(celt_sig, _pre); |
1346 | 0 | celt_sig *pre[2]; |
1347 | 0 | const CELTMode *mode; |
1348 | 0 | int pitch_index; |
1349 | 0 | opus_val16 gain1; |
1350 | 0 | opus_val16 pf_threshold; |
1351 | 0 | int pf_on; |
1352 | 0 | int qg; |
1353 | 0 | int overlap; |
1354 | 0 | opus_val32 before[2]={0}, after[2]={0}; |
1355 | 0 | int cancel_pitch=0; |
1356 | 0 | SAVE_STACK; |
1357 | |
|
1358 | 0 | mode = st->mode; |
1359 | 0 | overlap = mode->overlap; |
1360 | 0 | ALLOC(_pre, CC*(N+COMBFILTER_MAXPERIOD), celt_sig); |
1361 | |
|
1362 | 0 | pre[0] = _pre; |
1363 | 0 | pre[1] = _pre + (N+COMBFILTER_MAXPERIOD); |
1364 | | |
1365 | |
|
1366 | 0 | c=0; do { |
1367 | 0 | OPUS_COPY(pre[c], prefilter_mem+c*COMBFILTER_MAXPERIOD, COMBFILTER_MAXPERIOD); |
1368 | 0 | OPUS_COPY(pre[c]+COMBFILTER_MAXPERIOD, in+c*(N+overlap)+overlap, N); |
1369 | 0 | } while (++c<CC); |
1370 | |
|
1371 | 0 | if (enabled) |
1372 | 0 | { |
1373 | 0 | VARDECL(opus_val16, pitch_buf); |
1374 | 0 | ALLOC(pitch_buf, (COMBFILTER_MAXPERIOD+N)>>1, opus_val16); |
1375 | |
|
1376 | 0 | pitch_downsample(pre, pitch_buf, COMBFILTER_MAXPERIOD+N, CC, st->arch); |
1377 | | /* Don't search for the fir last 1.5 octave of the range because |
1378 | | there's too many false-positives due to short-term correlation */ |
1379 | 0 | pitch_search(pitch_buf+(COMBFILTER_MAXPERIOD>>1), pitch_buf, N, |
1380 | 0 | COMBFILTER_MAXPERIOD-3*COMBFILTER_MINPERIOD, &pitch_index, |
1381 | 0 | st->arch); |
1382 | 0 | pitch_index = COMBFILTER_MAXPERIOD-pitch_index; |
1383 | |
|
1384 | 0 | gain1 = remove_doubling(pitch_buf, COMBFILTER_MAXPERIOD, COMBFILTER_MINPERIOD, |
1385 | 0 | N, &pitch_index, st->prefilter_period, st->prefilter_gain, st->arch); |
1386 | 0 | if (pitch_index > COMBFILTER_MAXPERIOD-2) |
1387 | 0 | pitch_index = COMBFILTER_MAXPERIOD-2; |
1388 | 0 | gain1 = MULT16_16_Q15(QCONST16(.7f,15),gain1); |
1389 | | /* If we detect that the signal is dominated by a single tone, don't rely on the standard pitch |
1390 | | estimator, as it can become unreliable. */ |
1391 | 0 | if (toneishness > QCONST32(.99f, 29)) { |
1392 | | /* If the pitch is too high for our post-filter, apply pitch doubling until |
1393 | | we can get something that fits (not ideal, but better than nothing). */ |
1394 | 0 | while (tone_freq >= QCONST16(0.39f, 13)) tone_freq/=2; |
1395 | 0 | if (tone_freq > QCONST16(0.006148f, 13)) { |
1396 | | #ifdef FIXED_POINT |
1397 | | pitch_index = IMIN(51472/tone_freq, COMBFILTER_MAXPERIOD-2); |
1398 | | #else |
1399 | 0 | pitch_index = IMIN((int)floor(.5+2.f*M_PI/tone_freq), COMBFILTER_MAXPERIOD-2); |
1400 | 0 | #endif |
1401 | 0 | } else { |
1402 | | /* If the pitch is too low, using a very high pitch will actually give us an improvement |
1403 | | due to the DC component of the filter that will be close to our tone. Again, not ideal, |
1404 | | but if we only have a single tone, it's better than nothing. */ |
1405 | 0 | pitch_index = COMBFILTER_MINPERIOD; |
1406 | 0 | } |
1407 | 0 | gain1 = QCONST16(.75f, 15); |
1408 | 0 | } |
1409 | | /*printf("%d %d %f %f\n", pitch_change, pitch_index, gain1, st->analysis.tonality);*/ |
1410 | 0 | if (st->loss_rate>2) |
1411 | 0 | gain1 = HALF32(gain1); |
1412 | 0 | if (st->loss_rate>4) |
1413 | 0 | gain1 = HALF32(gain1); |
1414 | 0 | if (st->loss_rate>8) |
1415 | 0 | gain1 = 0; |
1416 | 0 | } else { |
1417 | 0 | gain1 = 0; |
1418 | 0 | pitch_index = COMBFILTER_MINPERIOD; |
1419 | 0 | } |
1420 | 0 | #ifndef DISABLE_FLOAT_API |
1421 | 0 | if (analysis->valid) |
1422 | 0 | gain1 = (opus_val16)(gain1 * analysis->max_pitch_ratio); |
1423 | | #else |
1424 | | (void)analysis; |
1425 | | #endif |
1426 | | /* Gain threshold for enabling the prefilter/postfilter */ |
1427 | 0 | pf_threshold = QCONST16(.2f,15); |
1428 | | |
1429 | | /* Adjusting the threshold based on rate and continuity */ |
1430 | 0 | if (abs(pitch_index-st->prefilter_period)*10>pitch_index) |
1431 | 0 | { |
1432 | 0 | pf_threshold += QCONST16(.2f,15); |
1433 | | /* Completely disable the prefilter on strong transients without continuity. */ |
1434 | 0 | if (tf_estimate > QCONST16(.98f, 14)) |
1435 | 0 | gain1 = 0; |
1436 | 0 | } |
1437 | 0 | if (nbAvailableBytes<25) |
1438 | 0 | pf_threshold += QCONST16(.1f,15); |
1439 | 0 | if (nbAvailableBytes<35) |
1440 | 0 | pf_threshold += QCONST16(.1f,15); |
1441 | 0 | if (st->prefilter_gain > QCONST16(.4f,15)) |
1442 | 0 | pf_threshold -= QCONST16(.1f,15); |
1443 | 0 | if (st->prefilter_gain > QCONST16(.55f,15)) |
1444 | 0 | pf_threshold -= QCONST16(.1f,15); |
1445 | | |
1446 | | /* Hard threshold at 0.2 */ |
1447 | 0 | pf_threshold = MAX16(pf_threshold, QCONST16(.2f,15)); |
1448 | 0 | if (gain1<pf_threshold) |
1449 | 0 | { |
1450 | 0 | gain1 = 0; |
1451 | 0 | pf_on = 0; |
1452 | 0 | qg = 0; |
1453 | 0 | } else { |
1454 | | /*This block is not gated by a total bits check only because |
1455 | | of the nbAvailableBytes check above.*/ |
1456 | 0 | if (ABS16(gain1-st->prefilter_gain)<QCONST16(.1f,15)) |
1457 | 0 | gain1=st->prefilter_gain; |
1458 | |
|
1459 | | #ifdef FIXED_POINT |
1460 | | qg = ((gain1+1536)>>10)/3-1; |
1461 | | #else |
1462 | 0 | qg = (int)floor(.5f+gain1*32/3)-1; |
1463 | 0 | #endif |
1464 | 0 | qg = IMAX(0, IMIN(7, qg)); |
1465 | 0 | gain1 = QCONST16(0.09375f,15)*(qg+1); |
1466 | 0 | pf_on = 1; |
1467 | 0 | } |
1468 | | /*printf("%d %f\n", pitch_index, gain1);*/ |
1469 | |
|
1470 | 0 | c=0; do { |
1471 | 0 | int i; |
1472 | 0 | int offset = mode->shortMdctSize-overlap; |
1473 | 0 | st->prefilter_period=IMAX(st->prefilter_period, COMBFILTER_MINPERIOD); |
1474 | 0 | OPUS_COPY(in+c*(N+overlap), st->in_mem+c*(overlap), overlap); |
1475 | 0 | for (i=0;i<N;i++) before[c] += ABS32(SHR32(in[c*(N+overlap)+overlap+i], 12)); |
1476 | 0 | if (offset) |
1477 | 0 | comb_filter(in+c*(N+overlap)+overlap, pre[c]+COMBFILTER_MAXPERIOD, |
1478 | 0 | st->prefilter_period, st->prefilter_period, offset, -st->prefilter_gain, -st->prefilter_gain, |
1479 | 0 | st->prefilter_tapset, st->prefilter_tapset, NULL, 0, st->arch); |
1480 | |
|
1481 | 0 | comb_filter(in+c*(N+overlap)+overlap+offset, pre[c]+COMBFILTER_MAXPERIOD+offset, |
1482 | 0 | st->prefilter_period, pitch_index, N-offset, -st->prefilter_gain, -gain1, |
1483 | 0 | st->prefilter_tapset, prefilter_tapset, mode->window, overlap, st->arch); |
1484 | 0 | for (i=0;i<N;i++) after[c] += ABS32(SHR32(in[c*(N+overlap)+overlap+i], 12)); |
1485 | 0 | } while (++c<CC); |
1486 | |
|
1487 | 0 | if (CC==2) { |
1488 | 0 | opus_val16 thresh[2]; |
1489 | 0 | thresh[0] = MULT16_32_Q15(MULT16_16_Q15(QCONST16(.25f, 15), gain1), before[0]) + MULT16_32_Q15(QCONST16(.01f,15), before[1]); |
1490 | 0 | thresh[1] = MULT16_32_Q15(MULT16_16_Q15(QCONST16(.25f, 15), gain1), before[1]) + MULT16_32_Q15(QCONST16(.01f,15), before[0]); |
1491 | | /* Don't use the filter if one channel gets significantly worse. */ |
1492 | 0 | if (after[0]-before[0] > thresh[0] || after[1]-before[1] > thresh[1]) cancel_pitch = 1; |
1493 | | /* Use the filter only if at least one channel gets significantly better. */ |
1494 | 0 | if (before[0]-after[0] < thresh[0] && before[1]-after[1] < thresh[1]) cancel_pitch = 1; |
1495 | 0 | } else { |
1496 | | /* Check that the mono channel actually got better. */ |
1497 | 0 | if (after[0] > before[0]) cancel_pitch = 1; |
1498 | 0 | } |
1499 | | /* If needed, revert to a gain of zero. */ |
1500 | 0 | if (cancel_pitch) { |
1501 | 0 | c=0; do { |
1502 | 0 | int offset = mode->shortMdctSize-overlap; |
1503 | 0 | OPUS_COPY(in+c*(N+overlap)+overlap, pre[c]+COMBFILTER_MAXPERIOD, N); |
1504 | 0 | comb_filter(in+c*(N+overlap)+overlap+offset, pre[c]+COMBFILTER_MAXPERIOD+offset, |
1505 | 0 | st->prefilter_period, pitch_index, overlap, -st->prefilter_gain, -0, |
1506 | 0 | st->prefilter_tapset, prefilter_tapset, mode->window, overlap, st->arch); |
1507 | 0 | } while (++c<CC); |
1508 | 0 | gain1 = 0; |
1509 | 0 | pf_on = 0; |
1510 | 0 | qg = 0; |
1511 | 0 | } |
1512 | |
|
1513 | 0 | c=0; do { |
1514 | 0 | OPUS_COPY(st->in_mem+c*(overlap), in+c*(N+overlap)+N, overlap); |
1515 | |
|
1516 | 0 | if (N>COMBFILTER_MAXPERIOD) |
1517 | 0 | { |
1518 | 0 | OPUS_COPY(prefilter_mem+c*COMBFILTER_MAXPERIOD, pre[c]+N, COMBFILTER_MAXPERIOD); |
1519 | 0 | } else { |
1520 | 0 | OPUS_MOVE(prefilter_mem+c*COMBFILTER_MAXPERIOD, prefilter_mem+c*COMBFILTER_MAXPERIOD+N, COMBFILTER_MAXPERIOD-N); |
1521 | 0 | OPUS_COPY(prefilter_mem+c*COMBFILTER_MAXPERIOD+COMBFILTER_MAXPERIOD-N, pre[c]+COMBFILTER_MAXPERIOD, N); |
1522 | 0 | } |
1523 | 0 | } while (++c<CC); |
1524 | |
|
1525 | 0 | RESTORE_STACK; |
1526 | 0 | *gain = gain1; |
1527 | 0 | *pitch = pitch_index; |
1528 | 0 | *qgain = qg; |
1529 | 0 | return pf_on; |
1530 | 0 | } |
1531 | | |
1532 | | static int compute_vbr(const CELTMode *mode, AnalysisInfo *analysis, opus_int32 base_target, |
1533 | | int LM, opus_int32 bitrate, int lastCodedBands, int C, int intensity, |
1534 | | int constrained_vbr, opus_val16 stereo_saving, int tot_boost, |
1535 | | opus_val16 tf_estimate, int pitch_change, celt_glog maxDepth, |
1536 | | int lfe, int has_surround_mask, celt_glog surround_masking, |
1537 | | celt_glog temporal_vbr) |
1538 | 0 | { |
1539 | | /* The target rate in 8th bits per frame */ |
1540 | 0 | opus_int32 target; |
1541 | 0 | int coded_bins; |
1542 | 0 | int coded_bands; |
1543 | 0 | opus_val16 tf_calibration; |
1544 | 0 | int nbEBands; |
1545 | 0 | const opus_int16 *eBands; |
1546 | |
|
1547 | 0 | nbEBands = mode->nbEBands; |
1548 | 0 | eBands = mode->eBands; |
1549 | |
|
1550 | 0 | coded_bands = lastCodedBands ? lastCodedBands : nbEBands; |
1551 | 0 | coded_bins = eBands[coded_bands]<<LM; |
1552 | 0 | if (C==2) |
1553 | 0 | coded_bins += eBands[IMIN(intensity, coded_bands)]<<LM; |
1554 | |
|
1555 | 0 | target = base_target; |
1556 | | |
1557 | | /*printf("%f %f %f %f %d %d ", st->analysis.activity, st->analysis.tonality, tf_estimate, st->stereo_saving, tot_boost, coded_bands);*/ |
1558 | 0 | #ifndef DISABLE_FLOAT_API |
1559 | 0 | if (analysis->valid && analysis->activity<.4) |
1560 | 0 | target -= (opus_int32)((coded_bins<<BITRES)*(.4f-analysis->activity)); |
1561 | 0 | #endif |
1562 | | /* Stereo savings */ |
1563 | 0 | if (C==2) |
1564 | 0 | { |
1565 | 0 | int coded_stereo_bands; |
1566 | 0 | int coded_stereo_dof; |
1567 | 0 | opus_val16 max_frac; |
1568 | 0 | coded_stereo_bands = IMIN(intensity, coded_bands); |
1569 | 0 | coded_stereo_dof = (eBands[coded_stereo_bands]<<LM)-coded_stereo_bands; |
1570 | | /* Maximum fraction of the bits we can save if the signal is mono. */ |
1571 | 0 | max_frac = DIV32_16(MULT16_16(QCONST16(0.8f, 15), coded_stereo_dof), coded_bins); |
1572 | 0 | stereo_saving = MIN16(stereo_saving, QCONST16(1.f, 8)); |
1573 | | /*printf("%d %d %d ", coded_stereo_dof, coded_bins, tot_boost);*/ |
1574 | 0 | target -= (opus_int32)MIN32(MULT16_32_Q15(max_frac,target), |
1575 | 0 | SHR32(MULT16_16(stereo_saving-QCONST16(0.1f,8),(coded_stereo_dof<<BITRES)),8)); |
1576 | 0 | } |
1577 | | /* Boost the rate according to dynalloc (minus the dynalloc average for calibration). */ |
1578 | 0 | target += tot_boost-(19<<LM); |
1579 | | /* Apply transient boost, compensating for average boost. */ |
1580 | 0 | tf_calibration = QCONST16(0.044f,14); |
1581 | 0 | target += (opus_int32)SHL32(MULT16_32_Q15(tf_estimate-tf_calibration, target),1); |
1582 | |
|
1583 | 0 | #ifndef DISABLE_FLOAT_API |
1584 | | /* Apply tonality boost */ |
1585 | 0 | if (analysis->valid && !lfe) |
1586 | 0 | { |
1587 | 0 | opus_int32 tonal_target; |
1588 | 0 | float tonal; |
1589 | | |
1590 | | /* Tonality boost (compensating for the average). */ |
1591 | 0 | tonal = MAX16(0.f,analysis->tonality-.15f)-0.12f; |
1592 | 0 | tonal_target = target + (opus_int32)((coded_bins<<BITRES)*1.2f*tonal); |
1593 | 0 | if (pitch_change) |
1594 | 0 | tonal_target += (opus_int32)((coded_bins<<BITRES)*.8f); |
1595 | | /*printf("%f %f ", analysis->tonality, tonal);*/ |
1596 | 0 | target = tonal_target; |
1597 | 0 | } |
1598 | | #else |
1599 | | (void)analysis; |
1600 | | (void)pitch_change; |
1601 | | #endif |
1602 | |
|
1603 | 0 | if (has_surround_mask&&!lfe) |
1604 | 0 | { |
1605 | 0 | opus_int32 surround_target = target + (opus_int32)SHR32(MULT16_16(SHR32(surround_masking,DB_SHIFT-10),coded_bins<<BITRES), 10); |
1606 | | /*printf("%f %d %d %d %d %d %d ", surround_masking, coded_bins, st->end, st->intensity, surround_target, target, st->bitrate);*/ |
1607 | 0 | target = IMAX(target/4, surround_target); |
1608 | 0 | } |
1609 | |
|
1610 | 0 | { |
1611 | 0 | opus_int32 floor_depth; |
1612 | 0 | int bins; |
1613 | 0 | bins = eBands[nbEBands-2]<<LM; |
1614 | | /*floor_depth = SHR32(MULT16_16((C*bins<<BITRES),celt_log2(SHL32(MAX16(1,sample_max),13))), DB_SHIFT);*/ |
1615 | 0 | floor_depth = (opus_int32)SHR32(MULT16_32_Q15((C*bins<<BITRES),maxDepth), DB_SHIFT-15); |
1616 | 0 | floor_depth = IMAX(floor_depth, target>>2); |
1617 | 0 | target = IMIN(target, floor_depth); |
1618 | | /*printf("%f %d\n", maxDepth, floor_depth);*/ |
1619 | 0 | } |
1620 | | |
1621 | | /* Make VBR less aggressive for constrained VBR because we can't keep a higher bitrate |
1622 | | for long. Needs tuning. */ |
1623 | 0 | if ((!has_surround_mask||lfe) && constrained_vbr) |
1624 | 0 | { |
1625 | 0 | target = base_target + (opus_int32)MULT16_32_Q15(QCONST16(0.67f, 15), target-base_target); |
1626 | 0 | } |
1627 | |
|
1628 | 0 | if (!has_surround_mask && tf_estimate < QCONST16(.2f, 14)) |
1629 | 0 | { |
1630 | 0 | opus_val16 amount; |
1631 | 0 | opus_val16 tvbr_factor; |
1632 | 0 | amount = MULT16_16_Q15(QCONST16(.0000031f, 30), IMAX(0, IMIN(32000, 96000-bitrate))); |
1633 | 0 | tvbr_factor = SHR32(MULT16_16(SHR32(temporal_vbr, DB_SHIFT-10), amount), 10); |
1634 | 0 | target += (opus_int32)MULT16_32_Q15(tvbr_factor, target); |
1635 | 0 | } |
1636 | | |
1637 | | /* Don't allow more than doubling the rate */ |
1638 | 0 | target = IMIN(2*base_target, target); |
1639 | |
|
1640 | 0 | return target; |
1641 | 0 | } |
1642 | | |
1643 | | int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_res * pcm, int frame_size, unsigned char *compressed, int nbCompressedBytes, ec_enc *enc) |
1644 | 0 | { |
1645 | 0 | int i, c, N; |
1646 | 0 | opus_int32 bits; |
1647 | 0 | ec_enc _enc; |
1648 | 0 | VARDECL(celt_sig, in); |
1649 | 0 | VARDECL(celt_sig, freq); |
1650 | 0 | VARDECL(celt_norm, X); |
1651 | 0 | VARDECL(celt_ener, bandE); |
1652 | 0 | VARDECL(celt_glog, bandLogE); |
1653 | 0 | VARDECL(celt_glog, bandLogE2); |
1654 | 0 | VARDECL(int, fine_quant); |
1655 | 0 | VARDECL(celt_glog, error); |
1656 | 0 | VARDECL(int, pulses); |
1657 | 0 | VARDECL(int, cap); |
1658 | 0 | VARDECL(int, offsets); |
1659 | 0 | VARDECL(int, importance); |
1660 | 0 | VARDECL(int, spread_weight); |
1661 | 0 | VARDECL(int, fine_priority); |
1662 | 0 | VARDECL(int, tf_res); |
1663 | 0 | VARDECL(unsigned char, collapse_masks); |
1664 | 0 | celt_sig *prefilter_mem; |
1665 | 0 | celt_glog *oldBandE, *oldLogE, *oldLogE2, *energyError; |
1666 | 0 | int shortBlocks=0; |
1667 | 0 | int isTransient=0; |
1668 | 0 | const int CC = st->channels; |
1669 | 0 | const int C = st->stream_channels; |
1670 | 0 | int LM, M; |
1671 | 0 | int tf_select; |
1672 | 0 | int nbFilledBytes, nbAvailableBytes; |
1673 | 0 | int start; |
1674 | 0 | int end; |
1675 | 0 | int effEnd; |
1676 | 0 | int codedBands; |
1677 | 0 | int alloc_trim; |
1678 | 0 | int pitch_index=COMBFILTER_MINPERIOD; |
1679 | 0 | opus_val16 gain1 = 0; |
1680 | 0 | int dual_stereo=0; |
1681 | 0 | int effectiveBytes; |
1682 | 0 | int dynalloc_logp; |
1683 | 0 | opus_int32 vbr_rate; |
1684 | 0 | opus_int32 total_bits; |
1685 | 0 | opus_int32 total_boost; |
1686 | 0 | opus_int32 balance; |
1687 | 0 | opus_int32 tell; |
1688 | 0 | opus_int32 tell0_frac; |
1689 | 0 | int prefilter_tapset=0; |
1690 | 0 | int pf_on; |
1691 | 0 | int anti_collapse_rsv; |
1692 | 0 | int anti_collapse_on=0; |
1693 | 0 | int silence=0; |
1694 | 0 | int tf_chan = 0; |
1695 | 0 | opus_val16 tf_estimate; |
1696 | 0 | int pitch_change=0; |
1697 | 0 | opus_int32 tot_boost; |
1698 | 0 | opus_val32 sample_max; |
1699 | 0 | celt_glog maxDepth; |
1700 | 0 | const OpusCustomMode *mode; |
1701 | 0 | int nbEBands; |
1702 | 0 | int overlap; |
1703 | 0 | const opus_int16 *eBands; |
1704 | 0 | int secondMdct; |
1705 | 0 | int signalBandwidth; |
1706 | 0 | int transient_got_disabled=0; |
1707 | 0 | celt_glog surround_masking=0; |
1708 | 0 | celt_glog temporal_vbr=0; |
1709 | 0 | celt_glog surround_trim = 0; |
1710 | 0 | opus_int32 equiv_rate; |
1711 | 0 | int hybrid; |
1712 | 0 | int weak_transient = 0; |
1713 | 0 | int enable_tf_analysis; |
1714 | 0 | opus_val16 tone_freq=-1; |
1715 | 0 | opus_val32 toneishness=0; |
1716 | 0 | VARDECL(celt_glog, surround_dynalloc); |
1717 | 0 | ALLOC_STACK; |
1718 | |
|
1719 | 0 | mode = st->mode; |
1720 | 0 | nbEBands = mode->nbEBands; |
1721 | 0 | overlap = mode->overlap; |
1722 | 0 | eBands = mode->eBands; |
1723 | 0 | start = st->start; |
1724 | 0 | end = st->end; |
1725 | 0 | hybrid = start != 0; |
1726 | 0 | tf_estimate = 0; |
1727 | 0 | if (nbCompressedBytes<2 || pcm==NULL) |
1728 | 0 | { |
1729 | 0 | RESTORE_STACK; |
1730 | 0 | return OPUS_BAD_ARG; |
1731 | 0 | } |
1732 | | |
1733 | 0 | frame_size *= st->upsample; |
1734 | 0 | for (LM=0;LM<=mode->maxLM;LM++) |
1735 | 0 | if (mode->shortMdctSize<<LM==frame_size) |
1736 | 0 | break; |
1737 | 0 | if (LM>mode->maxLM) |
1738 | 0 | { |
1739 | 0 | RESTORE_STACK; |
1740 | 0 | return OPUS_BAD_ARG; |
1741 | 0 | } |
1742 | 0 | M=1<<LM; |
1743 | 0 | N = M*mode->shortMdctSize; |
1744 | |
|
1745 | 0 | prefilter_mem = st->in_mem+CC*(overlap); |
1746 | 0 | oldBandE = (celt_glog*)(st->in_mem+CC*(overlap+COMBFILTER_MAXPERIOD)); |
1747 | 0 | oldLogE = oldBandE + CC*nbEBands; |
1748 | 0 | oldLogE2 = oldLogE + CC*nbEBands; |
1749 | 0 | energyError = oldLogE2 + CC*nbEBands; |
1750 | |
|
1751 | 0 | if (enc==NULL) |
1752 | 0 | { |
1753 | 0 | tell0_frac=tell=1; |
1754 | 0 | nbFilledBytes=0; |
1755 | 0 | } else { |
1756 | 0 | tell0_frac=ec_tell_frac(enc); |
1757 | 0 | tell=ec_tell(enc); |
1758 | 0 | nbFilledBytes=(tell+4)>>3; |
1759 | 0 | } |
1760 | |
|
1761 | | #ifdef CUSTOM_MODES |
1762 | | if (st->signalling && enc==NULL) |
1763 | | { |
1764 | | int tmp = (mode->effEBands-end)>>1; |
1765 | | end = st->end = IMAX(1, mode->effEBands-tmp); |
1766 | | compressed[0] = tmp<<5; |
1767 | | compressed[0] |= LM<<3; |
1768 | | compressed[0] |= (C==2)<<2; |
1769 | | /* Convert "standard mode" to Opus header */ |
1770 | | if (mode->Fs==48000 && mode->shortMdctSize==120) |
1771 | | { |
1772 | | int c0 = toOpus(compressed[0]); |
1773 | | if (c0<0) |
1774 | | { |
1775 | | RESTORE_STACK; |
1776 | | return OPUS_BAD_ARG; |
1777 | | } |
1778 | | compressed[0] = c0; |
1779 | | } |
1780 | | compressed++; |
1781 | | nbCompressedBytes--; |
1782 | | } |
1783 | | #else |
1784 | 0 | celt_assert(st->signalling==0); |
1785 | 0 | #endif |
1786 | | |
1787 | | /* Can't produce more than 1275 output bytes */ |
1788 | 0 | nbCompressedBytes = IMIN(nbCompressedBytes,1275); |
1789 | |
|
1790 | 0 | if (st->vbr && st->bitrate!=OPUS_BITRATE_MAX) |
1791 | 0 | { |
1792 | 0 | opus_int32 den=mode->Fs>>BITRES; |
1793 | 0 | vbr_rate=(st->bitrate*frame_size+(den>>1))/den; |
1794 | | #ifdef CUSTOM_MODES |
1795 | | if (st->signalling) |
1796 | | vbr_rate -= 8<<BITRES; |
1797 | | #endif |
1798 | 0 | effectiveBytes = vbr_rate>>(3+BITRES); |
1799 | 0 | } else { |
1800 | 0 | opus_int32 tmp; |
1801 | 0 | vbr_rate = 0; |
1802 | 0 | tmp = st->bitrate*frame_size; |
1803 | 0 | if (tell>1) |
1804 | 0 | tmp += tell*mode->Fs; |
1805 | 0 | if (st->bitrate!=OPUS_BITRATE_MAX) |
1806 | 0 | { |
1807 | 0 | nbCompressedBytes = IMAX(2, IMIN(nbCompressedBytes, |
1808 | 0 | (tmp+4*mode->Fs)/(8*mode->Fs)-!!st->signalling)); |
1809 | 0 | if (enc != NULL) |
1810 | 0 | ec_enc_shrink(enc, nbCompressedBytes); |
1811 | 0 | } |
1812 | 0 | effectiveBytes = nbCompressedBytes - nbFilledBytes; |
1813 | 0 | } |
1814 | 0 | nbAvailableBytes = nbCompressedBytes - nbFilledBytes; |
1815 | 0 | equiv_rate = ((opus_int32)nbCompressedBytes*8*50 << (3-LM)) - (40*C+20)*((400>>LM) - 50); |
1816 | 0 | if (st->bitrate != OPUS_BITRATE_MAX) |
1817 | 0 | equiv_rate = IMIN(equiv_rate, st->bitrate - (40*C+20)*((400>>LM) - 50)); |
1818 | |
|
1819 | 0 | if (enc==NULL) |
1820 | 0 | { |
1821 | 0 | ec_enc_init(&_enc, compressed, nbCompressedBytes); |
1822 | 0 | enc = &_enc; |
1823 | 0 | } |
1824 | |
|
1825 | 0 | if (vbr_rate>0) |
1826 | 0 | { |
1827 | | /* Computes the max bit-rate allowed in VBR mode to avoid violating the |
1828 | | target rate and buffering. |
1829 | | We must do this up front so that bust-prevention logic triggers |
1830 | | correctly if we don't have enough bits. */ |
1831 | 0 | if (st->constrained_vbr) |
1832 | 0 | { |
1833 | 0 | opus_int32 vbr_bound; |
1834 | 0 | opus_int32 max_allowed; |
1835 | | /* We could use any multiple of vbr_rate as bound (depending on the |
1836 | | delay). |
1837 | | This is clamped to ensure we use at least two bytes if the encoder |
1838 | | was entirely empty, but to allow 0 in hybrid mode. */ |
1839 | 0 | vbr_bound = vbr_rate; |
1840 | 0 | max_allowed = IMIN(IMAX(tell==1?2:0, |
1841 | 0 | (vbr_rate+vbr_bound-st->vbr_reservoir)>>(BITRES+3)), |
1842 | 0 | nbAvailableBytes); |
1843 | 0 | if(max_allowed < nbAvailableBytes) |
1844 | 0 | { |
1845 | 0 | nbCompressedBytes = nbFilledBytes+max_allowed; |
1846 | 0 | nbAvailableBytes = max_allowed; |
1847 | 0 | ec_enc_shrink(enc, nbCompressedBytes); |
1848 | 0 | } |
1849 | 0 | } |
1850 | 0 | } |
1851 | 0 | total_bits = nbCompressedBytes*8; |
1852 | |
|
1853 | 0 | effEnd = end; |
1854 | 0 | if (effEnd > mode->effEBands) |
1855 | 0 | effEnd = mode->effEBands; |
1856 | |
|
1857 | 0 | ALLOC(in, CC*(N+overlap), celt_sig); |
1858 | |
|
1859 | 0 | sample_max=MAX32(st->overlap_max, celt_maxabs_res(pcm, C*(N-overlap)/st->upsample)); |
1860 | 0 | st->overlap_max=celt_maxabs_res(pcm+C*(N-overlap)/st->upsample, C*overlap/st->upsample); |
1861 | 0 | sample_max=MAX32(sample_max, st->overlap_max); |
1862 | | #ifdef FIXED_POINT |
1863 | | silence = (sample_max==0); |
1864 | | #else |
1865 | 0 | silence = (sample_max <= (opus_val16)1/(1<<st->lsb_depth)); |
1866 | 0 | #endif |
1867 | | #ifdef FUZZING |
1868 | | if ((rand()&0x3F)==0) |
1869 | | silence = 1; |
1870 | | #endif |
1871 | 0 | if (tell==1) |
1872 | 0 | ec_enc_bit_logp(enc, silence, 15); |
1873 | 0 | else |
1874 | 0 | silence=0; |
1875 | 0 | if (silence) |
1876 | 0 | { |
1877 | | /*In VBR mode there is no need to send more than the minimum. */ |
1878 | 0 | if (vbr_rate>0) |
1879 | 0 | { |
1880 | 0 | effectiveBytes=nbCompressedBytes=IMIN(nbCompressedBytes, nbFilledBytes+2); |
1881 | 0 | total_bits=nbCompressedBytes*8; |
1882 | 0 | nbAvailableBytes=2; |
1883 | 0 | ec_enc_shrink(enc, nbCompressedBytes); |
1884 | 0 | } |
1885 | | /* Pretend we've filled all the remaining bits with zeros |
1886 | | (that's what the initialiser did anyway) */ |
1887 | 0 | tell = nbCompressedBytes*8; |
1888 | 0 | enc->nbits_total+=tell-ec_tell(enc); |
1889 | 0 | } |
1890 | 0 | c=0; do { |
1891 | 0 | int need_clip=0; |
1892 | 0 | #ifndef FIXED_POINT |
1893 | 0 | need_clip = st->clip && sample_max>65536.f; |
1894 | 0 | #endif |
1895 | 0 | celt_preemphasis(pcm+c, in+c*(N+overlap)+overlap, N, CC, st->upsample, |
1896 | 0 | mode->preemph, st->preemph_memE+c, need_clip); |
1897 | 0 | OPUS_COPY(in+c*(N+overlap), &prefilter_mem[(1+c)*COMBFILTER_MAXPERIOD-overlap], overlap); |
1898 | 0 | } while (++c<CC); |
1899 | | |
1900 | |
|
1901 | 0 | tone_freq = tone_detect(in, CC, N+overlap, &toneishness, mode->Fs); |
1902 | 0 | isTransient = 0; |
1903 | 0 | shortBlocks = 0; |
1904 | 0 | if (st->complexity >= 1 && !st->lfe) |
1905 | 0 | { |
1906 | | /* Reduces the likelihood of energy instability on fricatives at low bitrate |
1907 | | in hybrid mode. It seems like we still want to have real transients on vowels |
1908 | | though (small SILK quantization offset value). */ |
1909 | 0 | int allow_weak_transients = hybrid && effectiveBytes<15 && st->silk_info.signalType != 2; |
1910 | 0 | isTransient = transient_analysis(in, N+overlap, CC, |
1911 | 0 | &tf_estimate, &tf_chan, allow_weak_transients, &weak_transient, tone_freq, toneishness); |
1912 | 0 | } |
1913 | | /* Find pitch period and gain */ |
1914 | 0 | { |
1915 | 0 | int enabled; |
1916 | 0 | int qg; |
1917 | 0 | enabled = ((st->lfe&&nbAvailableBytes>3) || nbAvailableBytes>12*C) && !hybrid && !silence && tell+16<=total_bits && !st->disable_pf |
1918 | 0 | && st->complexity >= 5; |
1919 | |
|
1920 | 0 | prefilter_tapset = st->tapset_decision; |
1921 | 0 | pf_on = run_prefilter(st, in, prefilter_mem, CC, N, prefilter_tapset, &pitch_index, &gain1, &qg, enabled, tf_estimate, nbAvailableBytes, &st->analysis, tone_freq, toneishness); |
1922 | 0 | if ((gain1 > QCONST16(.4f,15) || st->prefilter_gain > QCONST16(.4f,15)) && (!st->analysis.valid || st->analysis.tonality > .3) |
1923 | 0 | && (pitch_index > 1.26*st->prefilter_period || pitch_index < .79*st->prefilter_period)) |
1924 | 0 | pitch_change = 1; |
1925 | 0 | if (pf_on==0) |
1926 | 0 | { |
1927 | 0 | if(!hybrid && tell+16<=total_bits) |
1928 | 0 | ec_enc_bit_logp(enc, 0, 1); |
1929 | 0 | } else { |
1930 | | /*This block is not gated by a total bits check only because |
1931 | | of the nbAvailableBytes check above.*/ |
1932 | 0 | int octave; |
1933 | 0 | ec_enc_bit_logp(enc, 1, 1); |
1934 | 0 | pitch_index += 1; |
1935 | 0 | octave = EC_ILOG(pitch_index)-5; |
1936 | 0 | ec_enc_uint(enc, octave, 6); |
1937 | 0 | ec_enc_bits(enc, pitch_index-(16<<octave), 4+octave); |
1938 | 0 | pitch_index -= 1; |
1939 | 0 | ec_enc_bits(enc, qg, 3); |
1940 | 0 | ec_enc_icdf(enc, prefilter_tapset, tapset_icdf, 2); |
1941 | 0 | } |
1942 | 0 | } |
1943 | 0 | if (LM>0 && ec_tell(enc)+3<=total_bits) |
1944 | 0 | { |
1945 | 0 | if (isTransient) |
1946 | 0 | shortBlocks = M; |
1947 | 0 | } else { |
1948 | 0 | isTransient = 0; |
1949 | 0 | transient_got_disabled=1; |
1950 | 0 | } |
1951 | |
|
1952 | 0 | ALLOC(freq, CC*N, celt_sig); /**< Interleaved signal MDCTs */ |
1953 | 0 | ALLOC(bandE,nbEBands*CC, celt_ener); |
1954 | 0 | ALLOC(bandLogE,nbEBands*CC, celt_glog); |
1955 | |
|
1956 | 0 | secondMdct = shortBlocks && st->complexity>=8; |
1957 | 0 | ALLOC(bandLogE2, C*nbEBands, celt_glog); |
1958 | 0 | if (secondMdct) |
1959 | 0 | { |
1960 | 0 | compute_mdcts(mode, 0, in, freq, C, CC, LM, st->upsample, st->arch); |
1961 | 0 | compute_band_energies(mode, freq, bandE, effEnd, C, LM, st->arch); |
1962 | 0 | amp2Log2(mode, effEnd, end, bandE, bandLogE2, C); |
1963 | 0 | for (c=0;c<C;c++) |
1964 | 0 | { |
1965 | 0 | for (i=0;i<end;i++) |
1966 | 0 | bandLogE2[nbEBands*c+i] += HALF32(SHL32(LM, DB_SHIFT)); |
1967 | 0 | } |
1968 | 0 | } |
1969 | |
|
1970 | 0 | compute_mdcts(mode, shortBlocks, in, freq, C, CC, LM, st->upsample, st->arch); |
1971 | | /* This should catch any NaN in the CELT input. Since we're not supposed to see any (they're filtered |
1972 | | at the Opus layer), just abort. */ |
1973 | 0 | celt_assert(!celt_isnan(freq[0]) && (C==1 || !celt_isnan(freq[N]))); |
1974 | 0 | if (CC==2&&C==1) |
1975 | 0 | tf_chan = 0; |
1976 | 0 | compute_band_energies(mode, freq, bandE, effEnd, C, LM, st->arch); |
1977 | |
|
1978 | 0 | if (st->lfe) |
1979 | 0 | { |
1980 | 0 | for (i=2;i<end;i++) |
1981 | 0 | { |
1982 | 0 | bandE[i] = IMIN(bandE[i], MULT16_32_Q15(QCONST16(1e-4f,15),bandE[0])); |
1983 | 0 | bandE[i] = MAX32(bandE[i], EPSILON); |
1984 | 0 | } |
1985 | 0 | } |
1986 | 0 | amp2Log2(mode, effEnd, end, bandE, bandLogE, C); |
1987 | |
|
1988 | 0 | ALLOC(surround_dynalloc, C*nbEBands, celt_glog); |
1989 | 0 | OPUS_CLEAR(surround_dynalloc, end); |
1990 | | /* This computes how much masking takes place between surround channels */ |
1991 | 0 | if (!hybrid&&st->energy_mask&&!st->lfe) |
1992 | 0 | { |
1993 | 0 | int mask_end; |
1994 | 0 | int midband; |
1995 | 0 | int count_dynalloc; |
1996 | 0 | opus_val32 mask_avg=0; |
1997 | 0 | opus_val32 diff=0; |
1998 | 0 | int count=0; |
1999 | 0 | mask_end = IMAX(2,st->lastCodedBands); |
2000 | 0 | for (c=0;c<C;c++) |
2001 | 0 | { |
2002 | 0 | for(i=0;i<mask_end;i++) |
2003 | 0 | { |
2004 | 0 | celt_glog mask; |
2005 | 0 | opus_val16 mask16; |
2006 | 0 | mask = MAXG(MING(st->energy_mask[nbEBands*c+i], |
2007 | 0 | GCONST(.25f)), -GCONST(2.0f)); |
2008 | 0 | if (mask > 0) |
2009 | 0 | mask = HALF32(mask); |
2010 | 0 | mask16 = SHR32(mask, DB_SHIFT-10); |
2011 | 0 | mask_avg += MULT16_16(mask16, eBands[i+1]-eBands[i]); |
2012 | 0 | count += eBands[i+1]-eBands[i]; |
2013 | 0 | diff += MULT16_16(mask16, 1+2*i-mask_end); |
2014 | 0 | } |
2015 | 0 | } |
2016 | 0 | celt_assert(count>0); |
2017 | 0 | mask_avg = SHL32(DIV32_16(mask_avg,count), DB_SHIFT-10); |
2018 | 0 | mask_avg += GCONST(.2f); |
2019 | 0 | diff = SHL32(diff*6/(C*(mask_end-1)*(mask_end+1)*mask_end), DB_SHIFT-10); |
2020 | | /* Again, being conservative */ |
2021 | 0 | diff = HALF32(diff); |
2022 | 0 | diff = MAX32(MIN32(diff, GCONST(.031f)), -GCONST(.031f)); |
2023 | | /* Find the band that's in the middle of the coded spectrum */ |
2024 | 0 | for (midband=0;eBands[midband+1] < eBands[mask_end]/2;midband++); |
2025 | 0 | count_dynalloc=0; |
2026 | 0 | for(i=0;i<mask_end;i++) |
2027 | 0 | { |
2028 | 0 | opus_val32 lin; |
2029 | 0 | celt_glog unmask; |
2030 | 0 | lin = mask_avg + diff*(i-midband); |
2031 | 0 | if (C==2) |
2032 | 0 | unmask = MAXG(st->energy_mask[i], st->energy_mask[nbEBands+i]); |
2033 | 0 | else |
2034 | 0 | unmask = st->energy_mask[i]; |
2035 | 0 | unmask = MING(unmask, GCONST(.0f)); |
2036 | 0 | unmask -= lin; |
2037 | 0 | if (unmask > GCONST(.25f)) |
2038 | 0 | { |
2039 | 0 | surround_dynalloc[i] = unmask - GCONST(.25f); |
2040 | 0 | count_dynalloc++; |
2041 | 0 | } |
2042 | 0 | } |
2043 | 0 | if (count_dynalloc>=3) |
2044 | 0 | { |
2045 | | /* If we need dynalloc in many bands, it's probably because our |
2046 | | initial masking rate was too low. */ |
2047 | 0 | mask_avg += GCONST(.25f); |
2048 | 0 | if (mask_avg>0) |
2049 | 0 | { |
2050 | | /* Something went really wrong in the original calculations, |
2051 | | disabling masking. */ |
2052 | 0 | mask_avg = 0; |
2053 | 0 | diff = 0; |
2054 | 0 | OPUS_CLEAR(surround_dynalloc, mask_end); |
2055 | 0 | } else { |
2056 | 0 | for(i=0;i<mask_end;i++) |
2057 | 0 | surround_dynalloc[i] = MAXG(0, surround_dynalloc[i]-GCONST(.25f)); |
2058 | 0 | } |
2059 | 0 | } |
2060 | 0 | mask_avg += GCONST(.2f); |
2061 | | /* Convert to 1/64th units used for the trim */ |
2062 | 0 | surround_trim = 64*diff; |
2063 | | /*printf("%d %d ", mask_avg, surround_trim);*/ |
2064 | 0 | surround_masking = mask_avg; |
2065 | 0 | } |
2066 | | /* Temporal VBR (but not for LFE) */ |
2067 | 0 | if (!st->lfe) |
2068 | 0 | { |
2069 | 0 | celt_glog follow=-QCONST32(10.0f, DB_SHIFT-5); |
2070 | 0 | opus_val32 frame_avg=0; |
2071 | 0 | celt_glog offset = shortBlocks?HALF32(SHL32(LM, DB_SHIFT-5)):0; |
2072 | 0 | for(i=start;i<end;i++) |
2073 | 0 | { |
2074 | 0 | follow = MAXG(follow-QCONST32(1.0f, DB_SHIFT-5), SHR32(bandLogE[i],5)-offset); |
2075 | 0 | if (C==2) |
2076 | 0 | follow = MAXG(follow, SHR32(bandLogE[i+nbEBands],5)-offset); |
2077 | 0 | frame_avg += follow; |
2078 | 0 | } |
2079 | 0 | frame_avg /= (end-start); |
2080 | 0 | temporal_vbr = SUB32(SHL32(frame_avg, 5),st->spec_avg); |
2081 | 0 | temporal_vbr = MING(GCONST(3.f), MAXG(-GCONST(1.5f), temporal_vbr)); |
2082 | 0 | st->spec_avg += MULT16_32_Q15(QCONST16(.02f, 15), temporal_vbr); |
2083 | 0 | } |
2084 | | /*for (i=0;i<21;i++) |
2085 | | printf("%f ", bandLogE[i]); |
2086 | | printf("\n");*/ |
2087 | |
|
2088 | 0 | if (!secondMdct) |
2089 | 0 | { |
2090 | 0 | OPUS_COPY(bandLogE2, bandLogE, C*nbEBands); |
2091 | 0 | } |
2092 | | |
2093 | | /* Last chance to catch any transient we might have missed in the |
2094 | | time-domain analysis */ |
2095 | 0 | if (LM>0 && ec_tell(enc)+3<=total_bits && !isTransient && st->complexity>=5 && !st->lfe && !hybrid) |
2096 | 0 | { |
2097 | 0 | if (patch_transient_decision(bandLogE, oldBandE, nbEBands, start, end, C)) |
2098 | 0 | { |
2099 | 0 | isTransient = 1; |
2100 | 0 | shortBlocks = M; |
2101 | 0 | compute_mdcts(mode, shortBlocks, in, freq, C, CC, LM, st->upsample, st->arch); |
2102 | 0 | compute_band_energies(mode, freq, bandE, effEnd, C, LM, st->arch); |
2103 | 0 | amp2Log2(mode, effEnd, end, bandE, bandLogE, C); |
2104 | | /* Compensate for the scaling of short vs long mdcts */ |
2105 | 0 | for (c=0;c<C;c++) |
2106 | 0 | { |
2107 | 0 | for (i=0;i<end;i++) |
2108 | 0 | bandLogE2[nbEBands*c+i] += HALF32(SHL32(LM, DB_SHIFT)); |
2109 | 0 | } |
2110 | 0 | tf_estimate = QCONST16(.2f,14); |
2111 | 0 | } |
2112 | 0 | } |
2113 | |
|
2114 | 0 | if (LM>0 && ec_tell(enc)+3<=total_bits) |
2115 | 0 | ec_enc_bit_logp(enc, isTransient, 3); |
2116 | |
|
2117 | 0 | ALLOC(X, C*N, celt_norm); /**< Interleaved normalised MDCTs */ |
2118 | | |
2119 | | /* Band normalisation */ |
2120 | 0 | normalise_bands(mode, freq, X, bandE, effEnd, C, M); |
2121 | |
|
2122 | 0 | enable_tf_analysis = effectiveBytes>=15*C && !hybrid && st->complexity>=2 && !st->lfe && toneishness < QCONST32(.98f, 29); |
2123 | |
|
2124 | 0 | ALLOC(offsets, nbEBands, int); |
2125 | 0 | ALLOC(importance, nbEBands, int); |
2126 | 0 | ALLOC(spread_weight, nbEBands, int); |
2127 | |
|
2128 | 0 | maxDepth = dynalloc_analysis(bandLogE, bandLogE2, oldBandE, nbEBands, start, end, C, offsets, |
2129 | 0 | st->lsb_depth, mode->logN, isTransient, st->vbr, st->constrained_vbr, |
2130 | 0 | eBands, LM, effectiveBytes, &tot_boost, st->lfe, surround_dynalloc, &st->analysis, importance, spread_weight, tone_freq, toneishness); |
2131 | |
|
2132 | 0 | ALLOC(tf_res, nbEBands, int); |
2133 | | /* Disable variable tf resolution for hybrid and at very low bitrate */ |
2134 | 0 | if (enable_tf_analysis) |
2135 | 0 | { |
2136 | 0 | int lambda; |
2137 | 0 | lambda = IMAX(80, 20480/effectiveBytes + 2); |
2138 | 0 | tf_select = tf_analysis(mode, effEnd, isTransient, tf_res, lambda, X, N, LM, tf_estimate, tf_chan, importance); |
2139 | 0 | for (i=effEnd;i<end;i++) |
2140 | 0 | tf_res[i] = tf_res[effEnd-1]; |
2141 | 0 | } else if (hybrid && weak_transient) |
2142 | 0 | { |
2143 | | /* For weak transients, we rely on the fact that improving time resolution using |
2144 | | TF on a long window is imperfect and will not result in an energy collapse at |
2145 | | low bitrate. */ |
2146 | 0 | for (i=0;i<end;i++) |
2147 | 0 | tf_res[i] = 1; |
2148 | 0 | tf_select=0; |
2149 | 0 | } else if (hybrid && effectiveBytes<15 && st->silk_info.signalType != 2) |
2150 | 0 | { |
2151 | | /* For low bitrate hybrid, we force temporal resolution to 5 ms rather than 2.5 ms. */ |
2152 | 0 | for (i=0;i<end;i++) |
2153 | 0 | tf_res[i] = 0; |
2154 | 0 | tf_select=isTransient; |
2155 | 0 | } else { |
2156 | 0 | for (i=0;i<end;i++) |
2157 | 0 | tf_res[i] = isTransient; |
2158 | 0 | tf_select=0; |
2159 | 0 | } |
2160 | |
|
2161 | 0 | ALLOC(error, C*nbEBands, celt_glog); |
2162 | 0 | c=0; |
2163 | 0 | do { |
2164 | 0 | for (i=start;i<end;i++) |
2165 | 0 | { |
2166 | | /* When the energy is stable, slightly bias energy quantization towards |
2167 | | the previous error to make the gain more stable (a constant offset is |
2168 | | better than fluctuations). */ |
2169 | 0 | if (ABS32(SUB32(bandLogE[i+c*nbEBands], oldBandE[i+c*nbEBands])) < GCONST(2.f)) |
2170 | 0 | { |
2171 | 0 | bandLogE[i+c*nbEBands] -= MULT16_32_Q15(QCONST16(0.25f, 15), energyError[i+c*nbEBands]); |
2172 | 0 | } |
2173 | 0 | } |
2174 | 0 | } while (++c < C); |
2175 | 0 | quant_coarse_energy(mode, start, end, effEnd, bandLogE, |
2176 | 0 | oldBandE, total_bits, error, enc, |
2177 | 0 | C, LM, nbAvailableBytes, st->force_intra, |
2178 | 0 | &st->delayedIntra, st->complexity >= 4, st->loss_rate, st->lfe); |
2179 | |
|
2180 | 0 | tf_encode(start, end, isTransient, tf_res, LM, tf_select, enc); |
2181 | |
|
2182 | 0 | if (ec_tell(enc)+4<=total_bits) |
2183 | 0 | { |
2184 | 0 | if (st->lfe) |
2185 | 0 | { |
2186 | 0 | st->tapset_decision = 0; |
2187 | 0 | st->spread_decision = SPREAD_NORMAL; |
2188 | 0 | } else if (hybrid) |
2189 | 0 | { |
2190 | 0 | if (st->complexity == 0) |
2191 | 0 | st->spread_decision = SPREAD_NONE; |
2192 | 0 | else if (isTransient) |
2193 | 0 | st->spread_decision = SPREAD_NORMAL; |
2194 | 0 | else |
2195 | 0 | st->spread_decision = SPREAD_AGGRESSIVE; |
2196 | 0 | } else if (shortBlocks || st->complexity < 3 || nbAvailableBytes < 10*C) |
2197 | 0 | { |
2198 | 0 | if (st->complexity == 0) |
2199 | 0 | st->spread_decision = SPREAD_NONE; |
2200 | 0 | else |
2201 | 0 | st->spread_decision = SPREAD_NORMAL; |
2202 | 0 | } else { |
2203 | | /* Disable new spreading+tapset estimator until we can show it works |
2204 | | better than the old one. So far it seems like spreading_decision() |
2205 | | works best. */ |
2206 | | #if 0 |
2207 | | if (st->analysis.valid) |
2208 | | { |
2209 | | static const opus_val16 spread_thresholds[3] = {-QCONST16(.6f, 15), -QCONST16(.2f, 15), -QCONST16(.07f, 15)}; |
2210 | | static const opus_val16 spread_histeresis[3] = {QCONST16(.15f, 15), QCONST16(.07f, 15), QCONST16(.02f, 15)}; |
2211 | | static const opus_val16 tapset_thresholds[2] = {QCONST16(.0f, 15), QCONST16(.15f, 15)}; |
2212 | | static const opus_val16 tapset_histeresis[2] = {QCONST16(.1f, 15), QCONST16(.05f, 15)}; |
2213 | | st->spread_decision = hysteresis_decision(-st->analysis.tonality, spread_thresholds, spread_histeresis, 3, st->spread_decision); |
2214 | | st->tapset_decision = hysteresis_decision(st->analysis.tonality_slope, tapset_thresholds, tapset_histeresis, 2, st->tapset_decision); |
2215 | | } else |
2216 | | #endif |
2217 | 0 | { |
2218 | 0 | st->spread_decision = spreading_decision(mode, X, |
2219 | 0 | &st->tonal_average, st->spread_decision, &st->hf_average, |
2220 | 0 | &st->tapset_decision, pf_on&&!shortBlocks, effEnd, C, M, spread_weight); |
2221 | 0 | } |
2222 | | /*printf("%d %d\n", st->tapset_decision, st->spread_decision);*/ |
2223 | | /*printf("%f %d %f %d\n\n", st->analysis.tonality, st->spread_decision, st->analysis.tonality_slope, st->tapset_decision);*/ |
2224 | 0 | } |
2225 | 0 | ec_enc_icdf(enc, st->spread_decision, spread_icdf, 5); |
2226 | 0 | } else { |
2227 | 0 | st->spread_decision = SPREAD_NORMAL; |
2228 | 0 | } |
2229 | | |
2230 | | /* For LFE, everything interesting is in the first band */ |
2231 | 0 | if (st->lfe) |
2232 | 0 | offsets[0] = IMIN(8, effectiveBytes/3); |
2233 | 0 | ALLOC(cap, nbEBands, int); |
2234 | 0 | init_caps(mode,cap,LM,C); |
2235 | |
|
2236 | 0 | dynalloc_logp = 6; |
2237 | 0 | total_bits<<=BITRES; |
2238 | 0 | total_boost = 0; |
2239 | 0 | tell = ec_tell_frac(enc); |
2240 | 0 | for (i=start;i<end;i++) |
2241 | 0 | { |
2242 | 0 | int width, quanta; |
2243 | 0 | int dynalloc_loop_logp; |
2244 | 0 | int boost; |
2245 | 0 | int j; |
2246 | 0 | width = C*(eBands[i+1]-eBands[i])<<LM; |
2247 | | /* quanta is 6 bits, but no more than 1 bit/sample |
2248 | | and no less than 1/8 bit/sample */ |
2249 | 0 | quanta = IMIN(width<<BITRES, IMAX(6<<BITRES, width)); |
2250 | 0 | dynalloc_loop_logp = dynalloc_logp; |
2251 | 0 | boost = 0; |
2252 | 0 | for (j = 0; tell+(dynalloc_loop_logp<<BITRES) < total_bits-total_boost |
2253 | 0 | && boost < cap[i]; j++) |
2254 | 0 | { |
2255 | 0 | int flag; |
2256 | 0 | flag = j<offsets[i]; |
2257 | 0 | ec_enc_bit_logp(enc, flag, dynalloc_loop_logp); |
2258 | 0 | tell = ec_tell_frac(enc); |
2259 | 0 | if (!flag) |
2260 | 0 | break; |
2261 | 0 | boost += quanta; |
2262 | 0 | total_boost += quanta; |
2263 | 0 | dynalloc_loop_logp = 1; |
2264 | 0 | } |
2265 | | /* Making dynalloc more likely */ |
2266 | 0 | if (j) |
2267 | 0 | dynalloc_logp = IMAX(2, dynalloc_logp-1); |
2268 | 0 | offsets[i] = boost; |
2269 | 0 | } |
2270 | |
|
2271 | 0 | if (C==2) |
2272 | 0 | { |
2273 | 0 | static const opus_val16 intensity_thresholds[21]= |
2274 | | /* 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 off*/ |
2275 | 0 | { 1, 2, 3, 4, 5, 6, 7, 8,16,24,36,44,50,56,62,67,72,79,88,106,134}; |
2276 | 0 | static const opus_val16 intensity_histeresis[21]= |
2277 | 0 | { 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 2, 3, 3, 4, 5, 6, 8, 8}; |
2278 | | |
2279 | | /* Always use MS for 2.5 ms frames until we can do a better analysis */ |
2280 | 0 | if (LM!=0) |
2281 | 0 | dual_stereo = stereo_analysis(mode, X, LM, N); |
2282 | |
|
2283 | 0 | st->intensity = hysteresis_decision((opus_val16)(equiv_rate/1000), |
2284 | 0 | intensity_thresholds, intensity_histeresis, 21, st->intensity); |
2285 | 0 | st->intensity = IMIN(end,IMAX(start, st->intensity)); |
2286 | 0 | } |
2287 | |
|
2288 | 0 | alloc_trim = 5; |
2289 | 0 | if (tell+(6<<BITRES) <= total_bits - total_boost) |
2290 | 0 | { |
2291 | 0 | if (start > 0 || st->lfe) |
2292 | 0 | { |
2293 | 0 | st->stereo_saving = 0; |
2294 | 0 | alloc_trim = 5; |
2295 | 0 | } else { |
2296 | 0 | alloc_trim = alloc_trim_analysis(mode, X, bandLogE, |
2297 | 0 | end, LM, C, N, &st->analysis, &st->stereo_saving, tf_estimate, |
2298 | 0 | st->intensity, surround_trim, equiv_rate, st->arch); |
2299 | 0 | } |
2300 | 0 | ec_enc_icdf(enc, alloc_trim, trim_icdf, 7); |
2301 | 0 | tell = ec_tell_frac(enc); |
2302 | 0 | } |
2303 | | |
2304 | | /* Variable bitrate */ |
2305 | 0 | if (vbr_rate>0) |
2306 | 0 | { |
2307 | 0 | opus_val16 alpha; |
2308 | 0 | opus_int32 delta; |
2309 | | /* The target rate in 8th bits per frame */ |
2310 | 0 | opus_int32 target, base_target; |
2311 | 0 | opus_int32 min_allowed; |
2312 | 0 | int lm_diff = mode->maxLM - LM; |
2313 | | |
2314 | | /* Don't attempt to use more than 510 kb/s, even for frames smaller than 20 ms. |
2315 | | The CELT allocator will just not be able to use more than that anyway. */ |
2316 | 0 | nbCompressedBytes = IMIN(nbCompressedBytes,1275>>(3-LM)); |
2317 | 0 | if (!hybrid) |
2318 | 0 | { |
2319 | 0 | base_target = vbr_rate - ((40*C+20)<<BITRES); |
2320 | 0 | } else { |
2321 | 0 | base_target = IMAX(0, vbr_rate - ((9*C+4)<<BITRES)); |
2322 | 0 | } |
2323 | |
|
2324 | 0 | if (st->constrained_vbr) |
2325 | 0 | base_target += (st->vbr_offset>>lm_diff); |
2326 | |
|
2327 | 0 | if (!hybrid) |
2328 | 0 | { |
2329 | 0 | target = compute_vbr(mode, &st->analysis, base_target, LM, equiv_rate, |
2330 | 0 | st->lastCodedBands, C, st->intensity, st->constrained_vbr, |
2331 | 0 | st->stereo_saving, tot_boost, tf_estimate, pitch_change, maxDepth, |
2332 | 0 | st->lfe, st->energy_mask!=NULL, surround_masking, |
2333 | 0 | temporal_vbr); |
2334 | 0 | } else { |
2335 | 0 | target = base_target; |
2336 | | /* Tonal frames (offset<100) need more bits than noisy (offset>100) ones. */ |
2337 | 0 | if (st->silk_info.offset < 100) target += 12 << BITRES >> (3-LM); |
2338 | 0 | if (st->silk_info.offset > 100) target -= 18 << BITRES >> (3-LM); |
2339 | | /* Boosting bitrate on transients and vowels with significant temporal |
2340 | | spikes. */ |
2341 | 0 | target += (opus_int32)MULT16_16_Q14(tf_estimate-QCONST16(.25f,14), (50<<BITRES)); |
2342 | | /* If we have a strong transient, let's make sure it has enough bits to code |
2343 | | the first two bands, so that it can use folding rather than noise. */ |
2344 | 0 | if (tf_estimate > QCONST16(.7f,14)) |
2345 | 0 | target = IMAX(target, 50<<BITRES); |
2346 | 0 | } |
2347 | | /* The current offset is removed from the target and the space used |
2348 | | so far is added*/ |
2349 | 0 | target=target+tell; |
2350 | | /* In VBR mode the frame size must not be reduced so much that it would |
2351 | | result in the encoder running out of bits. |
2352 | | The margin of 2 bytes ensures that none of the bust-prevention logic |
2353 | | in the decoder will have triggered so far. */ |
2354 | 0 | min_allowed = ((tell+total_boost+(1<<(BITRES+3))-1)>>(BITRES+3)) + 2; |
2355 | | /* Take into account the 37 bits we need to have left in the packet to |
2356 | | signal a redundant frame in hybrid mode. Creating a shorter packet would |
2357 | | create an entropy coder desync. */ |
2358 | 0 | if (hybrid) |
2359 | 0 | min_allowed = IMAX(min_allowed, (tell0_frac+(37<<BITRES)+total_boost+(1<<(BITRES+3))-1)>>(BITRES+3)); |
2360 | |
|
2361 | 0 | nbAvailableBytes = (target+(1<<(BITRES+2)))>>(BITRES+3); |
2362 | 0 | nbAvailableBytes = IMAX(min_allowed,nbAvailableBytes); |
2363 | 0 | nbAvailableBytes = IMIN(nbCompressedBytes,nbAvailableBytes); |
2364 | | |
2365 | | /* By how much did we "miss" the target on that frame */ |
2366 | 0 | delta = target - vbr_rate; |
2367 | |
|
2368 | 0 | target=nbAvailableBytes<<(BITRES+3); |
2369 | | |
2370 | | /*If the frame is silent we don't adjust our drift, otherwise |
2371 | | the encoder will shoot to very high rates after hitting a |
2372 | | span of silence, but we do allow the bitres to refill. |
2373 | | This means that we'll undershoot our target in CVBR/VBR modes |
2374 | | on files with lots of silence. */ |
2375 | 0 | if(silence) |
2376 | 0 | { |
2377 | 0 | nbAvailableBytes = 2; |
2378 | 0 | target = 2*8<<BITRES; |
2379 | 0 | delta = 0; |
2380 | 0 | } |
2381 | |
|
2382 | 0 | if (st->vbr_count < 970) |
2383 | 0 | { |
2384 | 0 | st->vbr_count++; |
2385 | 0 | alpha = celt_rcp(SHL32(EXTEND32(st->vbr_count+20),16)); |
2386 | 0 | } else |
2387 | 0 | alpha = QCONST16(.001f,15); |
2388 | | /* How many bits have we used in excess of what we're allowed */ |
2389 | 0 | if (st->constrained_vbr) |
2390 | 0 | st->vbr_reservoir += target - vbr_rate; |
2391 | | /*printf ("%d\n", st->vbr_reservoir);*/ |
2392 | | |
2393 | | /* Compute the offset we need to apply in order to reach the target */ |
2394 | 0 | if (st->constrained_vbr) |
2395 | 0 | { |
2396 | 0 | st->vbr_drift += (opus_int32)MULT16_32_Q15(alpha,(delta*(1<<lm_diff))-st->vbr_offset-st->vbr_drift); |
2397 | 0 | st->vbr_offset = -st->vbr_drift; |
2398 | 0 | } |
2399 | | /*printf ("%d\n", st->vbr_drift);*/ |
2400 | |
|
2401 | 0 | if (st->constrained_vbr && st->vbr_reservoir < 0) |
2402 | 0 | { |
2403 | | /* We're under the min value -- increase rate */ |
2404 | 0 | int adjust = (-st->vbr_reservoir)/(8<<BITRES); |
2405 | | /* Unless we're just coding silence */ |
2406 | 0 | nbAvailableBytes += silence?0:adjust; |
2407 | 0 | st->vbr_reservoir = 0; |
2408 | | /*printf ("+%d\n", adjust);*/ |
2409 | 0 | } |
2410 | 0 | nbCompressedBytes = IMIN(nbCompressedBytes,nbAvailableBytes); |
2411 | | /*printf("%d\n", nbCompressedBytes*50*8);*/ |
2412 | | /* This moves the raw bits to take into account the new compressed size */ |
2413 | 0 | ec_enc_shrink(enc, nbCompressedBytes); |
2414 | 0 | } |
2415 | | |
2416 | | /* Bit allocation */ |
2417 | 0 | ALLOC(fine_quant, nbEBands, int); |
2418 | 0 | ALLOC(pulses, nbEBands, int); |
2419 | 0 | ALLOC(fine_priority, nbEBands, int); |
2420 | | |
2421 | | /* bits = packet size - where we are - safety*/ |
2422 | 0 | bits = (((opus_int32)nbCompressedBytes*8)<<BITRES) - (opus_int32)ec_tell_frac(enc) - 1; |
2423 | 0 | anti_collapse_rsv = isTransient&&LM>=2&&bits>=((LM+2)<<BITRES) ? (1<<BITRES) : 0; |
2424 | 0 | bits -= anti_collapse_rsv; |
2425 | 0 | signalBandwidth = end-1; |
2426 | 0 | #ifndef DISABLE_FLOAT_API |
2427 | 0 | if (st->analysis.valid) |
2428 | 0 | { |
2429 | 0 | int min_bandwidth; |
2430 | 0 | if (equiv_rate < (opus_int32)32000*C) |
2431 | 0 | min_bandwidth = 13; |
2432 | 0 | else if (equiv_rate < (opus_int32)48000*C) |
2433 | 0 | min_bandwidth = 16; |
2434 | 0 | else if (equiv_rate < (opus_int32)60000*C) |
2435 | 0 | min_bandwidth = 18; |
2436 | 0 | else if (equiv_rate < (opus_int32)80000*C) |
2437 | 0 | min_bandwidth = 19; |
2438 | 0 | else |
2439 | 0 | min_bandwidth = 20; |
2440 | 0 | signalBandwidth = IMAX(st->analysis.bandwidth, min_bandwidth); |
2441 | 0 | } |
2442 | 0 | #endif |
2443 | 0 | if (st->lfe) |
2444 | 0 | signalBandwidth = 1; |
2445 | 0 | codedBands = clt_compute_allocation(mode, start, end, offsets, cap, |
2446 | 0 | alloc_trim, &st->intensity, &dual_stereo, bits, &balance, pulses, |
2447 | 0 | fine_quant, fine_priority, C, LM, enc, 1, st->lastCodedBands, signalBandwidth); |
2448 | 0 | if (st->lastCodedBands) |
2449 | 0 | st->lastCodedBands = IMIN(st->lastCodedBands+1,IMAX(st->lastCodedBands-1,codedBands)); |
2450 | 0 | else |
2451 | 0 | st->lastCodedBands = codedBands; |
2452 | |
|
2453 | 0 | quant_fine_energy(mode, start, end, oldBandE, error, fine_quant, enc, C); |
2454 | | |
2455 | | /* Residual quantisation */ |
2456 | 0 | ALLOC(collapse_masks, C*nbEBands, unsigned char); |
2457 | 0 | quant_all_bands(1, mode, start, end, X, C==2 ? X+N : NULL, collapse_masks, |
2458 | 0 | bandE, pulses, shortBlocks, st->spread_decision, |
2459 | 0 | dual_stereo, st->intensity, tf_res, nbCompressedBytes*(8<<BITRES)-anti_collapse_rsv, |
2460 | 0 | balance, enc, LM, codedBands, &st->rng, st->complexity, st->arch, st->disable_inv); |
2461 | |
|
2462 | 0 | if (anti_collapse_rsv > 0) |
2463 | 0 | { |
2464 | 0 | anti_collapse_on = st->consec_transient<2; |
2465 | | #ifdef FUZZING |
2466 | | anti_collapse_on = rand()&0x1; |
2467 | | #endif |
2468 | 0 | ec_enc_bits(enc, anti_collapse_on, 1); |
2469 | 0 | } |
2470 | 0 | quant_energy_finalise(mode, start, end, oldBandE, error, fine_quant, fine_priority, nbCompressedBytes*8-ec_tell(enc), enc, C); |
2471 | 0 | OPUS_CLEAR(energyError, nbEBands*CC); |
2472 | 0 | c=0; |
2473 | 0 | do { |
2474 | 0 | for (i=start;i<end;i++) |
2475 | 0 | { |
2476 | 0 | energyError[i+c*nbEBands] = MAXG(-GCONST(0.5f), MING(GCONST(0.5f), error[i+c*nbEBands])); |
2477 | 0 | } |
2478 | 0 | } while (++c < C); |
2479 | |
|
2480 | 0 | if (silence) |
2481 | 0 | { |
2482 | 0 | for (i=0;i<C*nbEBands;i++) |
2483 | 0 | oldBandE[i] = -GCONST(28.f); |
2484 | 0 | } |
2485 | |
|
2486 | | #ifdef RESYNTH |
2487 | | /* Re-synthesis of the coded audio if required */ |
2488 | | { |
2489 | | celt_sig *out_mem[2]; |
2490 | | |
2491 | | if (anti_collapse_on) |
2492 | | { |
2493 | | anti_collapse(mode, X, collapse_masks, LM, C, N, |
2494 | | start, end, oldBandE, oldLogE, oldLogE2, pulses, st->rng, 1, st->arch); |
2495 | | } |
2496 | | |
2497 | | c=0; do { |
2498 | | OPUS_MOVE(st->syn_mem[c], st->syn_mem[c]+N, 2*MAX_PERIOD-N+overlap/2); |
2499 | | } while (++c<CC); |
2500 | | |
2501 | | c=0; do { |
2502 | | out_mem[c] = st->syn_mem[c]+2*MAX_PERIOD-N; |
2503 | | } while (++c<CC); |
2504 | | |
2505 | | celt_synthesis(mode, X, out_mem, oldBandE, start, effEnd, |
2506 | | C, CC, isTransient, LM, st->upsample, silence, st->arch); |
2507 | | |
2508 | | c=0; do { |
2509 | | st->prefilter_period=IMAX(st->prefilter_period, COMBFILTER_MINPERIOD); |
2510 | | st->prefilter_period_old=IMAX(st->prefilter_period_old, COMBFILTER_MINPERIOD); |
2511 | | comb_filter(out_mem[c], out_mem[c], st->prefilter_period_old, st->prefilter_period, mode->shortMdctSize, |
2512 | | st->prefilter_gain_old, st->prefilter_gain, st->prefilter_tapset_old, st->prefilter_tapset, |
2513 | | mode->window, overlap, st->arch); |
2514 | | if (LM!=0) |
2515 | | comb_filter(out_mem[c]+mode->shortMdctSize, out_mem[c]+mode->shortMdctSize, st->prefilter_period, pitch_index, N-mode->shortMdctSize, |
2516 | | st->prefilter_gain, gain1, st->prefilter_tapset, prefilter_tapset, |
2517 | | mode->window, overlap, st->arch); |
2518 | | } while (++c<CC); |
2519 | | |
2520 | | /* We reuse freq[] as scratch space for the de-emphasis */ |
2521 | | deemphasis(out_mem, (opus_res*)pcm, N, CC, st->upsample, mode->preemph, st->preemph_memD, 0); |
2522 | | st->prefilter_period_old = st->prefilter_period; |
2523 | | st->prefilter_gain_old = st->prefilter_gain; |
2524 | | st->prefilter_tapset_old = st->prefilter_tapset; |
2525 | | } |
2526 | | #endif |
2527 | |
|
2528 | 0 | st->prefilter_period = pitch_index; |
2529 | 0 | st->prefilter_gain = gain1; |
2530 | 0 | st->prefilter_tapset = prefilter_tapset; |
2531 | | #ifdef RESYNTH |
2532 | | if (LM!=0) |
2533 | | { |
2534 | | st->prefilter_period_old = st->prefilter_period; |
2535 | | st->prefilter_gain_old = st->prefilter_gain; |
2536 | | st->prefilter_tapset_old = st->prefilter_tapset; |
2537 | | } |
2538 | | #endif |
2539 | |
|
2540 | 0 | if (CC==2&&C==1) { |
2541 | 0 | OPUS_COPY(&oldBandE[nbEBands], oldBandE, nbEBands); |
2542 | 0 | } |
2543 | |
|
2544 | 0 | if (!isTransient) |
2545 | 0 | { |
2546 | 0 | OPUS_COPY(oldLogE2, oldLogE, CC*nbEBands); |
2547 | 0 | OPUS_COPY(oldLogE, oldBandE, CC*nbEBands); |
2548 | 0 | } else { |
2549 | 0 | for (i=0;i<CC*nbEBands;i++) |
2550 | 0 | oldLogE[i] = MING(oldLogE[i], oldBandE[i]); |
2551 | 0 | } |
2552 | | /* In case start or end were to change */ |
2553 | 0 | c=0; do |
2554 | 0 | { |
2555 | 0 | for (i=0;i<start;i++) |
2556 | 0 | { |
2557 | 0 | oldBandE[c*nbEBands+i]=0; |
2558 | 0 | oldLogE[c*nbEBands+i]=oldLogE2[c*nbEBands+i]=-GCONST(28.f); |
2559 | 0 | } |
2560 | 0 | for (i=end;i<nbEBands;i++) |
2561 | 0 | { |
2562 | 0 | oldBandE[c*nbEBands+i]=0; |
2563 | 0 | oldLogE[c*nbEBands+i]=oldLogE2[c*nbEBands+i]=-GCONST(28.f); |
2564 | 0 | } |
2565 | 0 | } while (++c<CC); |
2566 | |
|
2567 | 0 | if (isTransient || transient_got_disabled) |
2568 | 0 | st->consec_transient++; |
2569 | 0 | else |
2570 | 0 | st->consec_transient=0; |
2571 | 0 | st->rng = enc->rng; |
2572 | | |
2573 | | /* If there's any room left (can only happen for very high rates), |
2574 | | it's already filled with zeros */ |
2575 | 0 | ec_enc_done(enc); |
2576 | |
|
2577 | | #ifdef CUSTOM_MODES |
2578 | | if (st->signalling) |
2579 | | nbCompressedBytes++; |
2580 | | #endif |
2581 | |
|
2582 | 0 | RESTORE_STACK; |
2583 | 0 | if (ec_get_error(enc)) |
2584 | 0 | return OPUS_INTERNAL_ERROR; |
2585 | 0 | else |
2586 | 0 | return nbCompressedBytes; |
2587 | 0 | } |
2588 | | |
2589 | | |
2590 | | #ifdef CUSTOM_MODES |
2591 | | |
2592 | | #if defined(FIXED_POINT) && !defined(ENABLE_RES24) |
2593 | | int opus_custom_encode(CELTEncoder * OPUS_RESTRICT st, const opus_int16 * pcm, int frame_size, unsigned char *compressed, int nbCompressedBytes) |
2594 | | { |
2595 | | return celt_encode_with_ec(st, pcm, frame_size, compressed, nbCompressedBytes, NULL); |
2596 | | } |
2597 | | #else |
2598 | | int opus_custom_encode(CELTEncoder * OPUS_RESTRICT st, const opus_int16 * pcm, int frame_size, unsigned char *compressed, int nbCompressedBytes) |
2599 | | { |
2600 | | int j, ret, C, N; |
2601 | | VARDECL(opus_res, in); |
2602 | | ALLOC_STACK; |
2603 | | |
2604 | | if (pcm==NULL) |
2605 | | return OPUS_BAD_ARG; |
2606 | | |
2607 | | C = st->channels; |
2608 | | N = frame_size; |
2609 | | ALLOC(in, C*N, opus_res); |
2610 | | |
2611 | | for (j=0;j<C*N;j++) |
2612 | | in[j] = INT16TORES(pcm[j]); |
2613 | | |
2614 | | ret=celt_encode_with_ec(st,in,frame_size,compressed,nbCompressedBytes, NULL); |
2615 | | #ifdef RESYNTH |
2616 | | for (j=0;j<C*N;j++) |
2617 | | ((opus_int16*)pcm)[j]=RES2INT16(in[j]); |
2618 | | #endif |
2619 | | RESTORE_STACK; |
2620 | | return ret; |
2621 | | } |
2622 | | #endif |
2623 | | |
2624 | | |
2625 | | #if defined(FIXED_POINT) && defined(ENABLE_RES24) |
2626 | | int opus_custom_encode24(CELTEncoder * OPUS_RESTRICT st, const opus_int32 * pcm, int frame_size, unsigned char *compressed, int nbCompressedBytes) |
2627 | | { |
2628 | | return celt_encode_with_ec(st, pcm, frame_size, compressed, nbCompressedBytes, NULL); |
2629 | | } |
2630 | | #else |
2631 | | int opus_custom_encode24(CELTEncoder * OPUS_RESTRICT st, const opus_int32 * pcm, int frame_size, unsigned char *compressed, int nbCompressedBytes) |
2632 | | { |
2633 | | int j, ret, C, N; |
2634 | | VARDECL(opus_res, in); |
2635 | | ALLOC_STACK; |
2636 | | |
2637 | | if (pcm==NULL) |
2638 | | return OPUS_BAD_ARG; |
2639 | | |
2640 | | C = st->channels; |
2641 | | N = frame_size; |
2642 | | ALLOC(in, C*N, opus_res); |
2643 | | |
2644 | | for (j=0;j<C*N;j++) |
2645 | | in[j] = INT24TORES(pcm[j]); |
2646 | | |
2647 | | ret=celt_encode_with_ec(st,in,frame_size,compressed,nbCompressedBytes, NULL); |
2648 | | #ifdef RESYNTH |
2649 | | for (j=0;j<C*N;j++) |
2650 | | ((opus_int32*)pcm)[j]=RES2INT24(in[j]); |
2651 | | #endif |
2652 | | RESTORE_STACK; |
2653 | | return ret; |
2654 | | } |
2655 | | #endif |
2656 | | |
2657 | | |
2658 | | #ifndef DISABLE_FLOAT_API |
2659 | | |
2660 | | # if !defined(FIXED_POINT) |
2661 | | int opus_custom_encode_float(CELTEncoder * OPUS_RESTRICT st, const float * pcm, int frame_size, unsigned char *compressed, int nbCompressedBytes) |
2662 | | { |
2663 | | return celt_encode_with_ec(st, pcm, frame_size, compressed, nbCompressedBytes, NULL); |
2664 | | } |
2665 | | # else |
2666 | | int opus_custom_encode_float(CELTEncoder * OPUS_RESTRICT st, const float * pcm, int frame_size, unsigned char *compressed, int nbCompressedBytes) |
2667 | | { |
2668 | | int j, ret, C, N; |
2669 | | VARDECL(opus_res, in); |
2670 | | ALLOC_STACK; |
2671 | | |
2672 | | if (pcm==NULL) |
2673 | | return OPUS_BAD_ARG; |
2674 | | |
2675 | | C = st->channels; |
2676 | | N = frame_size; |
2677 | | ALLOC(in, C*N, opus_res); |
2678 | | |
2679 | | for (j=0;j<C*N;j++) |
2680 | | in[j] = FLOAT2RES(pcm[j]); |
2681 | | |
2682 | | ret=celt_encode_with_ec(st,in,frame_size,compressed,nbCompressedBytes, NULL); |
2683 | | #ifdef RESYNTH |
2684 | | for (j=0;j<C*N;j++) |
2685 | | ((float*)pcm)[j]=RES2FLOAT(in[j]); |
2686 | | #endif |
2687 | | RESTORE_STACK; |
2688 | | return ret; |
2689 | | } |
2690 | | # endif |
2691 | | |
2692 | | #endif |
2693 | | |
2694 | | #endif /* CUSTOM_MODES */ |
2695 | | |
2696 | | int opus_custom_encoder_ctl(CELTEncoder * OPUS_RESTRICT st, int request, ...) |
2697 | 0 | { |
2698 | 0 | va_list ap; |
2699 | |
|
2700 | 0 | va_start(ap, request); |
2701 | 0 | switch (request) |
2702 | 0 | { |
2703 | 0 | case OPUS_SET_COMPLEXITY_REQUEST: |
2704 | 0 | { |
2705 | 0 | int value = va_arg(ap, opus_int32); |
2706 | 0 | if (value<0 || value>10) |
2707 | 0 | goto bad_arg; |
2708 | 0 | st->complexity = value; |
2709 | 0 | } |
2710 | 0 | break; |
2711 | 0 | case CELT_SET_START_BAND_REQUEST: |
2712 | 0 | { |
2713 | 0 | opus_int32 value = va_arg(ap, opus_int32); |
2714 | 0 | if (value<0 || value>=st->mode->nbEBands) |
2715 | 0 | goto bad_arg; |
2716 | 0 | st->start = value; |
2717 | 0 | } |
2718 | 0 | break; |
2719 | 0 | case CELT_SET_END_BAND_REQUEST: |
2720 | 0 | { |
2721 | 0 | opus_int32 value = va_arg(ap, opus_int32); |
2722 | 0 | if (value<1 || value>st->mode->nbEBands) |
2723 | 0 | goto bad_arg; |
2724 | 0 | st->end = value; |
2725 | 0 | } |
2726 | 0 | break; |
2727 | 0 | case CELT_SET_PREDICTION_REQUEST: |
2728 | 0 | { |
2729 | 0 | int value = va_arg(ap, opus_int32); |
2730 | 0 | if (value<0 || value>2) |
2731 | 0 | goto bad_arg; |
2732 | 0 | st->disable_pf = value<=1; |
2733 | 0 | st->force_intra = value==0; |
2734 | 0 | } |
2735 | 0 | break; |
2736 | 0 | case OPUS_SET_PACKET_LOSS_PERC_REQUEST: |
2737 | 0 | { |
2738 | 0 | int value = va_arg(ap, opus_int32); |
2739 | 0 | if (value<0 || value>100) |
2740 | 0 | goto bad_arg; |
2741 | 0 | st->loss_rate = value; |
2742 | 0 | } |
2743 | 0 | break; |
2744 | 0 | case OPUS_SET_VBR_CONSTRAINT_REQUEST: |
2745 | 0 | { |
2746 | 0 | opus_int32 value = va_arg(ap, opus_int32); |
2747 | 0 | st->constrained_vbr = value; |
2748 | 0 | } |
2749 | 0 | break; |
2750 | 0 | case OPUS_SET_VBR_REQUEST: |
2751 | 0 | { |
2752 | 0 | opus_int32 value = va_arg(ap, opus_int32); |
2753 | 0 | st->vbr = value; |
2754 | 0 | } |
2755 | 0 | break; |
2756 | 0 | case OPUS_SET_BITRATE_REQUEST: |
2757 | 0 | { |
2758 | 0 | opus_int32 value = va_arg(ap, opus_int32); |
2759 | 0 | if (value<=500 && value!=OPUS_BITRATE_MAX) |
2760 | 0 | goto bad_arg; |
2761 | 0 | value = IMIN(value, 260000*st->channels); |
2762 | 0 | st->bitrate = value; |
2763 | 0 | } |
2764 | 0 | break; |
2765 | 0 | case CELT_SET_CHANNELS_REQUEST: |
2766 | 0 | { |
2767 | 0 | opus_int32 value = va_arg(ap, opus_int32); |
2768 | 0 | if (value<1 || value>2) |
2769 | 0 | goto bad_arg; |
2770 | 0 | st->stream_channels = value; |
2771 | 0 | } |
2772 | 0 | break; |
2773 | 0 | case OPUS_SET_LSB_DEPTH_REQUEST: |
2774 | 0 | { |
2775 | 0 | opus_int32 value = va_arg(ap, opus_int32); |
2776 | 0 | if (value<8 || value>24) |
2777 | 0 | goto bad_arg; |
2778 | 0 | st->lsb_depth=value; |
2779 | 0 | } |
2780 | 0 | break; |
2781 | 0 | case OPUS_GET_LSB_DEPTH_REQUEST: |
2782 | 0 | { |
2783 | 0 | opus_int32 *value = va_arg(ap, opus_int32*); |
2784 | 0 | *value=st->lsb_depth; |
2785 | 0 | } |
2786 | 0 | break; |
2787 | 0 | case OPUS_SET_PHASE_INVERSION_DISABLED_REQUEST: |
2788 | 0 | { |
2789 | 0 | opus_int32 value = va_arg(ap, opus_int32); |
2790 | 0 | if(value<0 || value>1) |
2791 | 0 | { |
2792 | 0 | goto bad_arg; |
2793 | 0 | } |
2794 | 0 | st->disable_inv = value; |
2795 | 0 | } |
2796 | 0 | break; |
2797 | 0 | case OPUS_GET_PHASE_INVERSION_DISABLED_REQUEST: |
2798 | 0 | { |
2799 | 0 | opus_int32 *value = va_arg(ap, opus_int32*); |
2800 | 0 | if (!value) |
2801 | 0 | { |
2802 | 0 | goto bad_arg; |
2803 | 0 | } |
2804 | 0 | *value = st->disable_inv; |
2805 | 0 | } |
2806 | 0 | break; |
2807 | 0 | case OPUS_RESET_STATE: |
2808 | 0 | { |
2809 | 0 | int i; |
2810 | 0 | celt_glog *oldBandE, *oldLogE, *oldLogE2; |
2811 | 0 | oldBandE = (celt_glog*)(st->in_mem+st->channels*(st->mode->overlap+COMBFILTER_MAXPERIOD)); |
2812 | 0 | oldLogE = oldBandE + st->channels*st->mode->nbEBands; |
2813 | 0 | oldLogE2 = oldLogE + st->channels*st->mode->nbEBands; |
2814 | 0 | OPUS_CLEAR((char*)&st->ENCODER_RESET_START, |
2815 | 0 | opus_custom_encoder_get_size(st->mode, st->channels)- |
2816 | 0 | ((char*)&st->ENCODER_RESET_START - (char*)st)); |
2817 | 0 | for (i=0;i<st->channels*st->mode->nbEBands;i++) |
2818 | 0 | oldLogE[i]=oldLogE2[i]=-GCONST(28.f); |
2819 | 0 | st->vbr_offset = 0; |
2820 | 0 | st->delayedIntra = 1; |
2821 | 0 | st->spread_decision = SPREAD_NORMAL; |
2822 | 0 | st->tonal_average = 256; |
2823 | 0 | st->hf_average = 0; |
2824 | 0 | st->tapset_decision = 0; |
2825 | 0 | } |
2826 | 0 | break; |
2827 | | #ifdef CUSTOM_MODES |
2828 | | case CELT_SET_INPUT_CLIPPING_REQUEST: |
2829 | | { |
2830 | | opus_int32 value = va_arg(ap, opus_int32); |
2831 | | st->clip = value; |
2832 | | } |
2833 | | break; |
2834 | | #endif |
2835 | 0 | case CELT_SET_SIGNALLING_REQUEST: |
2836 | 0 | { |
2837 | 0 | opus_int32 value = va_arg(ap, opus_int32); |
2838 | 0 | st->signalling = value; |
2839 | 0 | } |
2840 | 0 | break; |
2841 | 0 | case CELT_SET_ANALYSIS_REQUEST: |
2842 | 0 | { |
2843 | 0 | AnalysisInfo *info = va_arg(ap, AnalysisInfo *); |
2844 | 0 | if (info) |
2845 | 0 | OPUS_COPY(&st->analysis, info, 1); |
2846 | 0 | } |
2847 | 0 | break; |
2848 | 0 | case CELT_SET_SILK_INFO_REQUEST: |
2849 | 0 | { |
2850 | 0 | SILKInfo *info = va_arg(ap, SILKInfo *); |
2851 | 0 | if (info) |
2852 | 0 | OPUS_COPY(&st->silk_info, info, 1); |
2853 | 0 | } |
2854 | 0 | break; |
2855 | 0 | case CELT_GET_MODE_REQUEST: |
2856 | 0 | { |
2857 | 0 | const CELTMode ** value = va_arg(ap, const CELTMode**); |
2858 | 0 | if (value==0) |
2859 | 0 | goto bad_arg; |
2860 | 0 | *value=st->mode; |
2861 | 0 | } |
2862 | 0 | break; |
2863 | 0 | case OPUS_GET_FINAL_RANGE_REQUEST: |
2864 | 0 | { |
2865 | 0 | opus_uint32 * value = va_arg(ap, opus_uint32 *); |
2866 | 0 | if (value==0) |
2867 | 0 | goto bad_arg; |
2868 | 0 | *value=st->rng; |
2869 | 0 | } |
2870 | 0 | break; |
2871 | 0 | case OPUS_SET_LFE_REQUEST: |
2872 | 0 | { |
2873 | 0 | opus_int32 value = va_arg(ap, opus_int32); |
2874 | 0 | st->lfe = value; |
2875 | 0 | } |
2876 | 0 | break; |
2877 | 0 | case OPUS_SET_ENERGY_MASK_REQUEST: |
2878 | 0 | { |
2879 | 0 | celt_glog *value = va_arg(ap, celt_glog*); |
2880 | 0 | st->energy_mask = value; |
2881 | 0 | } |
2882 | 0 | break; |
2883 | 0 | default: |
2884 | 0 | goto bad_request; |
2885 | 0 | } |
2886 | 0 | va_end(ap); |
2887 | 0 | return OPUS_OK; |
2888 | 0 | bad_arg: |
2889 | 0 | va_end(ap); |
2890 | 0 | return OPUS_BAD_ARG; |
2891 | 0 | bad_request: |
2892 | 0 | va_end(ap); |
2893 | 0 | return OPUS_UNIMPLEMENTED; |
2894 | 0 | } |