celt_encoder.c 79.4 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33
/* Copyright (c) 2007-2008 CSIRO
   Copyright (c) 2007-2010 Xiph.Org Foundation
   Copyright (c) 2008 Gregory Maxwell
   Written by Jean-Marc Valin and Gregory Maxwell */
/*
   Redistribution and use in source and binary forms, with or without
   modification, are permitted provided that the following conditions
   are met:

   - Redistributions of source code must retain the above copyright
   notice, this list of conditions and the following disclaimer.

   - Redistributions in binary form must reproduce the above copyright
   notice, this list of conditions and the following disclaimer in the
   documentation and/or other materials provided with the distribution.

   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
   ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER
   OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
   EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
   PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
   PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
   LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
   NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
   SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/

#ifdef HAVE_CONFIG_H
#include "config.h"
#endif

34
#define CELT_ENCODER_C
35

36
#include "cpu_support.h"
37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75
#include "os_support.h"
#include "mdct.h"
#include <math.h>
#include "celt.h"
#include "pitch.h"
#include "bands.h"
#include "modes.h"
#include "entcode.h"
#include "quant_bands.h"
#include "rate.h"
#include "stack_alloc.h"
#include "mathops.h"
#include "float_cast.h"
#include <stdarg.h>
#include "celt_lpc.h"
#include "vq.h"


/** Encoder state
 @brief Encoder state
 */
struct OpusCustomEncoder {
   const OpusCustomMode *mode;     /**< Mode used by the encoder */
   int channels;
   int stream_channels;

   int force_intra;
   int clip;
   int disable_pf;
   int complexity;
   int upsample;
   int start, end;

   opus_int32 bitrate;
   int vbr;
   int signalling;
   int constrained_vbr;      /* If zero, VBR can do whatever it likes with the rate */
   int loss_rate;
   int lsb_depth;
76
   int variable_duration;
77
   int lfe;
78
   int disable_inv;
79
   int arch;
80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101

   /* Everything beyond this point gets cleared on a reset */
#define ENCODER_RESET_START rng

   opus_uint32 rng;
   int spread_decision;
   opus_val32 delayedIntra;
   int tonal_average;
   int lastCodedBands;
   int hf_average;
   int tapset_decision;

   int prefilter_period;
   opus_val16 prefilter_gain;
   int prefilter_tapset;
#ifdef RESYNTH
   int prefilter_period_old;
   opus_val16 prefilter_gain_old;
   int prefilter_tapset_old;
#endif
   int consec_transient;
   AnalysisInfo analysis;
102
   SILKInfo silk_info;
103 104 105 106 107 108 109 110 111

   opus_val32 preemph_memE[2];
   opus_val32 preemph_memD[2];

   /* VBR-related parameters */
   opus_int32 vbr_reservoir;
   opus_int32 vbr_drift;
   opus_int32 vbr_offset;
   opus_int32 vbr_count;
112
   opus_val32 overlap_max;
113 114
   opus_val16 stereo_saving;
   int intensity;
115
   opus_val16 *energy_mask;
116
   opus_val16 spec_avg;
117 118 119 120 121 122 123 124 125 126 127

#ifdef RESYNTH
   /* +MAX_PERIOD/2 to make space for overlap */
   celt_sig syn_mem[2][2*MAX_PERIOD+MAX_PERIOD/2];
#endif

   celt_sig in_mem[1]; /* Size = channels*mode->overlap */
   /* celt_sig prefilter_mem[],  Size = channels*COMBFILTER_MAXPERIOD */
   /* opus_val16 oldBandE[],     Size = channels*mode->nbEBands */
   /* opus_val16 oldLogE[],      Size = channels*mode->nbEBands */
   /* opus_val16 oldLogE2[],     Size = channels*mode->nbEBands */
128
   /* opus_val16 energyError[],  Size = channels*mode->nbEBands */
129 130 131 132 133 134 135 136 137 138 139 140 141
};

int celt_encoder_get_size(int channels)
{
   CELTMode *mode = opus_custom_mode_create(48000, 960, NULL);
   return opus_custom_encoder_get_size(mode, channels);
}

OPUS_CUSTOM_NOSTATIC int opus_custom_encoder_get_size(const CELTMode *mode, int channels)
{
   int size = sizeof(struct CELTEncoder)
         + (channels*mode->overlap-1)*sizeof(celt_sig)    /* celt_sig in_mem[channels*mode->overlap]; */
         + channels*COMBFILTER_MAXPERIOD*sizeof(celt_sig) /* celt_sig prefilter_mem[channels*COMBFILTER_MAXPERIOD]; */
142
         + 4*channels*mode->nbEBands*sizeof(opus_val16);  /* opus_val16 oldBandE[channels*mode->nbEBands]; */
143 144
                                                          /* opus_val16 oldLogE[channels*mode->nbEBands]; */
                                                          /* opus_val16 oldLogE2[channels*mode->nbEBands]; */
145
                                                          /* opus_val16 energyError[channels*mode->nbEBands]; */
146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166
   return size;
}

#ifdef CUSTOM_MODES
CELTEncoder *opus_custom_encoder_create(const CELTMode *mode, int channels, int *error)
{
   int ret;
   CELTEncoder *st = (CELTEncoder *)opus_alloc(opus_custom_encoder_get_size(mode, channels));
   /* init will handle the NULL case */
   ret = opus_custom_encoder_init(st, mode, channels);
   if (ret != OPUS_OK)
   {
      opus_custom_encoder_destroy(st);
      st = NULL;
   }
   if (error)
      *error = ret;
   return st;
}
#endif /* CUSTOM_MODES */

167 168
static int opus_custom_encoder_init_arch(CELTEncoder *st, const CELTMode *mode,
                                         int channels, int arch)
169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184
{
   if (channels < 0 || channels > 2)
      return OPUS_BAD_ARG;

   if (st==NULL || mode==NULL)
      return OPUS_ALLOC_FAIL;

   OPUS_CLEAR((char*)st, opus_custom_encoder_get_size(mode, channels));

   st->mode = mode;
   st->stream_channels = st->channels = channels;

   st->upsample = 1;
   st->start = 0;
   st->end = st->mode->effEBands;
   st->signalling = 1;
185
   st->arch = arch;
186

187 188 189 190 191 192 193 194 195 196 197 198 199 200
   st->constrained_vbr = 1;
   st->clip = 1;

   st->bitrate = OPUS_BITRATE_MAX;
   st->vbr = 0;
   st->force_intra  = 0;
   st->complexity = 5;
   st->lsb_depth=24;

   opus_custom_encoder_ctl(st, OPUS_RESET_STATE);

   return OPUS_OK;
}

201 202
#ifdef CUSTOM_MODES
int opus_custom_encoder_init(CELTEncoder *st, const CELTMode *mode, int channels)
203 204 205
{
   return opus_custom_encoder_init_arch(st, mode, channels, opus_select_arch());
}
206
#endif
207 208 209 210 211 212 213 214 215 216 217 218 219

int celt_encoder_init(CELTEncoder *st, opus_int32 sampling_rate, int channels,
                      int arch)
{
   int ret;
   ret = opus_custom_encoder_init_arch(st,
           opus_custom_mode_create(48000, 960, NULL), channels, arch);
   if (ret != OPUS_OK)
      return ret;
   st->upsample = resampling_factor(sampling_rate);
   return OPUS_OK;
}

220 221 222 223 224 225 226 227 228
#ifdef CUSTOM_MODES
void opus_custom_encoder_destroy(CELTEncoder *st)
{
   opus_free(st);
}
#endif /* CUSTOM_MODES */


static int transient_analysis(const opus_val32 * OPUS_RESTRICT in, int len, int C,
229 230
                              opus_val16 *tf_estimate, int *tf_chan, int low_rate,
                              int *weak_transient)
231 232 233 234 235 236 237
{
   int i;
   VARDECL(opus_val16, tmp);
   opus_val32 mem0,mem1;
   int is_transient = 0;
   opus_int32 mask_metric = 0;
   int c;
238
   opus_val16 tf_max;
239
   int len2;
240 241 242 243 244 245
   /* Forward masking: 6.7 dB/ms. */
#ifdef FIXED_POINT
   int forward_shift = 4;
#else
   opus_val16 forward_decay = QCONST16(.0625f,15);
#endif
246 247 248 249 250 251 252 253 254 255 256 257 258 259
   /* Table of 6*64/x, trained on real data to minimize the average error */
   static const unsigned char inv_table[128] = {
         255,255,156,110, 86, 70, 59, 51, 45, 40, 37, 33, 31, 28, 26, 25,
          23, 22, 21, 20, 19, 18, 17, 16, 16, 15, 15, 14, 13, 13, 12, 12,
          12, 12, 11, 11, 11, 10, 10, 10,  9,  9,  9,  9,  9,  9,  8,  8,
           8,  8,  8,  7,  7,  7,  7,  7,  7,  6,  6,  6,  6,  6,  6,  6,
           6,  6,  6,  6,  6,  6,  6,  6,  6,  5,  5,  5,  5,  5,  5,  5,
           5,  5,  5,  5,  5,  4,  4,  4,  4,  4,  4,  4,  4,  4,  4,  4,
           4,  4,  4,  4,  4,  4,  4,  4,  4,  4,  4,  4,  4,  4,  3,  3,
           3,  3,  3,  3,  3,  3,  3,  3,  3,  3,  3,  3,  3,  3,  3,  2,
   };
   SAVE_STACK;
   ALLOC(tmp, len, opus_val16);

260
   *weak_transient = 0;
261 262 263 264 265 266 267 268 269 270 271 272
   /* For lower bitrates, let's be more conservative and have a forward masking
      decay of 3.3 dB/ms. This avoids having to code transients at very low
      bitrate (mostly for hybrid), which can result in unstable energy and/or
      partial collapse. */
   if (low_rate)
   {
#ifdef FIXED_POINT
      forward_shift = 5;
#else
      forward_decay = QCONST16(.03125f,15);
#endif
   }
273
   len2=len/2;
274 275 276 277 278
   for (c=0;c<C;c++)
   {
      opus_val32 mean;
      opus_int32 unmask=0;
      opus_val32 norm;
279
      opus_val16 maxE;
280 281 282 283 284 285 286 287 288 289 290 291 292 293 294
      mem0=0;
      mem1=0;
      /* High-pass filter: (1 - 2*z^-1 + z^-2) / (1 - z^-1 + .5*z^-2) */
      for (i=0;i<len;i++)
      {
         opus_val32 x,y;
         x = SHR32(in[i+c*len],SIG_SHIFT);
         y = ADD32(mem0, x);
#ifdef FIXED_POINT
         mem0 = mem1 + y - SHL32(x,1);
         mem1 = x - SHR32(y,1);
#else
         mem0 = mem1 + y - 2*x;
         mem1 = x - .5f*y;
#endif
295
         tmp[i] = SROUND16(y, 2);
296 297 298 299
         /*printf("%f ", tmp[i]);*/
      }
      /*printf("\n");*/
      /* First few samples are bad because we don't propagate the memory */
300
      OPUS_CLEAR(tmp, 12);
301 302 303 304 305

#ifdef FIXED_POINT
      /* Normalize tmp to max range */
      {
         int shift=0;
306
         shift = 14-celt_ilog2(MAX16(1, celt_maxabs16(tmp, len)));
307 308 309 310 311 312 313 314 315 316
         if (shift!=0)
         {
            for (i=0;i<len;i++)
               tmp[i] = SHL16(tmp[i], shift);
         }
      }
#endif

      mean=0;
      mem0=0;
317
      /* Grouping by two to reduce complexity */
318
      /* Forward pass to compute the post-echo threshold*/
319
      for (i=0;i<len2;i++)
320 321 322 323 324
      {
         opus_val16 x2 = PSHR32(MULT16_16(tmp[2*i],tmp[2*i]) + MULT16_16(tmp[2*i+1],tmp[2*i+1]),16);
         mean += x2;
#ifdef FIXED_POINT
         /* FIXME: Use PSHR16() instead */
325
         tmp[i] = mem0 + PSHR32(x2-mem0,forward_shift);
326
#else
327
         tmp[i] = mem0 + MULT16_16_P15(forward_decay,x2-mem0);
328 329 330 331 332
#endif
         mem0 = tmp[i];
      }

      mem0=0;
333
      maxE=0;
334
      /* Backward pass to compute the pre-echo threshold */
335
      for (i=len2-1;i>=0;i--)
336
      {
337
         /* Backward masking: 13.9 dB/ms. */
338 339 340 341 342 343 344
#ifdef FIXED_POINT
         /* FIXME: Use PSHR16() instead */
         tmp[i] = mem0 + PSHR32(tmp[i]-mem0,3);
#else
         tmp[i] = mem0 + MULT16_16_P15(QCONST16(0.125f,15),tmp[i]-mem0);
#endif
         mem0 = tmp[i];
345
         maxE = MAX16(maxE, mem0);
346
      }
347
      /*for (i=0;i<len2;i++)printf("%f ", tmp[i]/mean);printf("\n");*/
348

349
      /* Compute the ratio of the "frame energy" over the harmonic mean of the energy.
350 351 352
         This essentially corresponds to a bitrate-normalized temporal noise-to-mask
         ratio */

353 354 355 356 357 358
      /* As a compromise with the old transient detector, frame energy is the
         geometric mean of the energy and half the max */
#ifdef FIXED_POINT
      /* Costs two sqrt() to avoid overflows */
      mean = MULT16_16(celt_sqrt(mean), celt_sqrt(MULT16_16(maxE,len2>>1)));
#else
359
      mean = celt_sqrt(mean * maxE*.5*len2);
360
#endif
361
      /* Inverse of the mean energy in Q15+6 */
362
      norm = SHL32(EXTEND32(len2),6+14)/ADD32(EPSILON,SHR32(mean,1));
363 364 365
      /* Compute harmonic mean discarding the unreliable boundaries
         The data is smooth, so we only take 1/4th of the samples */
      unmask=0;
366
      for (i=12;i<len2-5;i+=4)
367 368 369
      {
         int id;
#ifdef FIXED_POINT
370
         id = MAX32(0,MIN32(127,MULT16_32_Q15(tmp[i]+EPSILON,norm))); /* Do not round to nearest */
371
#else
372
         id = (int)MAX32(0,MIN32(127,floor(64*norm*(tmp[i]+EPSILON)))); /* Do not round to nearest */
373 374 375 376 377
#endif
         unmask += inv_table[id];
      }
      /*printf("%d\n", unmask);*/
      /* Normalize, compensate for the 1/4th of the sample and the factor of 6 in the inverse table */
378
      unmask = 64*unmask*4/(6*(len2-17));
379 380 381 382 383 384
      if (unmask>mask_metric)
      {
         *tf_chan = c;
         mask_metric = unmask;
      }
   }
385
   is_transient = mask_metric>200;
386 387 388 389 390 391
   /* For low bitrates, define "weak transients" that need to be
      handled differently to avoid partial collapse. */
   if (low_rate && is_transient && mask_metric<600) {
      is_transient = 0;
      *weak_transient = 1;
   }
392
   /* Arbitrary metric for VBR boost */
393
   tf_max = MAX16(0,celt_sqrt(27*mask_metric)-42);
394
   /* *tf_estimate = 1 + MIN16(1, sqrt(MAX16(0, tf_max-30))/20); */
395
   *tf_estimate = celt_sqrt(MAX32(0, SHL32(MULT16_16(QCONST16(0.0069,14),MIN16(163,tf_max)),14)-QCONST32(0.139,28)));
396 397 398 399 400 401 402 403 404
   /*printf("%d %f\n", tf_max, mask_metric);*/
   RESTORE_STACK;
#ifdef FUZZING
   is_transient = rand()&0x1;
#endif
   /*printf("%d %f %d\n", is_transient, (float)*tf_estimate, tf_max);*/
   return is_transient;
}

405 406
/* Looks for sudden increases of energy to decide whether we need to patch
   the transient decision */
407 408
static int patch_transient_decision(opus_val16 *newE, opus_val16 *oldE, int nbEBands,
      int start, int end, int C)
409 410 411 412 413 414 415 416
{
   int i, c;
   opus_val32 mean_diff=0;
   opus_val16 spread_old[26];
   /* Apply an aggressive (-6 dB/Bark) spreading function to the old frame to
      avoid false detection caused by irrelevant bands */
   if (C==1)
   {
417 418
      spread_old[start] = oldE[start];
      for (i=start+1;i<end;i++)
Stefan Hacker's avatar
Stefan Hacker committed
419
         spread_old[i] = MAX16(spread_old[i-1]-QCONST16(1.0f, DB_SHIFT), oldE[i]);
420
   } else {
421 422
      spread_old[start] = MAX16(oldE[start],oldE[start+nbEBands]);
      for (i=start+1;i<end;i++)
423
         spread_old[i] = MAX16(spread_old[i-1]-QCONST16(1.0f, DB_SHIFT),
Stefan Hacker's avatar
Stefan Hacker committed
424
                               MAX16(oldE[i],oldE[i+nbEBands]));
425
   }
426
   for (i=end-2;i>=start;i--)
427 428 429
      spread_old[i] = MAX16(spread_old[i], spread_old[i+1]-QCONST16(1.0f, DB_SHIFT));
   /* Compute mean increase */
   c=0; do {
430
      for (i=IMAX(2,start);i<end-1;i++)
431 432
      {
         opus_val16 x1, x2;
433
         x1 = MAX16(0, newE[i + c*nbEBands]);
434 435 436 437
         x2 = MAX16(0, spread_old[i]);
         mean_diff = ADD32(mean_diff, EXTEND32(MAX16(0, SUB16(x1, x2))));
      }
   } while (++c<C);
438
   mean_diff = DIV32(mean_diff, C*(end-1-IMAX(2,start)));
439 440 441 442
   /*printf("%f %f %d\n", mean_diff, max_diff, count);*/
   return mean_diff > QCONST16(1.f, DB_SHIFT);
}

443 444
/** Apply window and compute the MDCT for all sub-frames and
    all channels in a frame */
445
static void compute_mdcts(const CELTMode *mode, int shortBlocks, celt_sig * OPUS_RESTRICT in,
446 447
                          celt_sig * OPUS_RESTRICT out, int C, int CC, int LM, int upsample,
                          int arch)
448
{
449
   const int overlap = mode->overlap;
450 451 452
   int N;
   int B;
   int shift;
453
   int i, b, c;
454 455 456 457 458 459 460 461 462 463 464 465 466 467
   if (shortBlocks)
   {
      B = shortBlocks;
      N = mode->shortMdctSize;
      shift = mode->maxLM;
   } else {
      B = 1;
      N = mode->shortMdctSize<<LM;
      shift = mode->maxLM-LM;
   }
   c=0; do {
      for (b=0;b<B;b++)
      {
         /* Interleaving the sub-frames while doing the MDCTs */
468 469 470
         clt_mdct_forward(&mode->mdct, in+c*(B*N+overlap)+b*N,
                          &out[b+c*N*B], mode->window, overlap, shift, B,
                          arch);
471
      }
472 473 474 475 476 477 478 479 480 481 482 483 484
   } while (++c<CC);
   if (CC==2&&C==1)
   {
      for (i=0;i<B*N;i++)
         out[i] = ADD32(HALF32(out[i]), HALF32(out[B*N+i]));
   }
   if (upsample != 1)
   {
      c=0; do
      {
         int bound = B*N/upsample;
         for (i=0;i<bound;i++)
            out[c*B*N+i] *= upsample;
485
         OPUS_CLEAR(&out[c*B*N+bound], B*N-bound);
486 487
      } while (++c<C);
   }
488 489 490
}


491
void celt_preemphasis(const opus_val16 * OPUS_RESTRICT pcmp, celt_sig * OPUS_RESTRICT inp,
492 493 494
                        int N, int CC, int upsample, const opus_val16 *coef, celt_sig *mem, int clip)
{
   int i;
495
   opus_val16 coef0;
496 497 498 499
   celt_sig m;
   int Nu;

   coef0 = coef[0];
500
   m = *mem;
501

502 503
   /* Fast path for the normal 48kHz case and no clipping */
   if (coef[1] == 0 && upsample == 1 && !clip)
504 505 506 507 508 509 510 511 512 513 514 515
   {
      for (i=0;i<N;i++)
      {
         opus_val16 x;
         x = SCALEIN(pcmp[CC*i]);
         /* Apply pre-emphasis */
         inp[i] = SHL32(x, SIG_SHIFT) - m;
         m = SHR32(MULT16_16(coef0, x), 15-SIG_SHIFT);
      }
      *mem = m;
      return;
   }
516 517 518 519

   Nu = N/upsample;
   if (upsample!=1)
   {
520
      OPUS_CLEAR(inp, N);
521 522
   }
   for (i=0;i<Nu;i++)
523
      inp[i*upsample] = SCALEIN(pcmp[CC*i]);
524 525 526 527 528 529 530 531

#ifndef FIXED_POINT
   if (clip)
   {
      /* Clip input to avoid encoding non-portable files */
      for (i=0;i<Nu;i++)
         inp[i*upsample] = MAX32(-65536.f, MIN32(65536.f,inp[i*upsample]));
   }
532 533
#else
   (void)clip; /* Avoids a warning about clip being unused. */
534
#endif
535 536
#ifdef CUSTOM_MODES
   if (coef[1] != 0)
537
   {
538
      opus_val16 coef1 = coef[1];
539 540 541
      opus_val16 coef2 = coef[2];
      for (i=0;i<N;i++)
      {
542
         celt_sig x, tmp;
543 544 545 546 547 548
         x = inp[i];
         /* Apply pre-emphasis */
         tmp = MULT16_16(coef2, x);
         inp[i] = tmp + m;
         m = MULT16_32_Q15(coef1, inp[i]) - MULT16_32_Q15(coef0, tmp);
      }
549
   } else
550
#endif
551 552 553
   {
      for (i=0;i<N;i++)
      {
554 555
         opus_val16 x;
         x = inp[i];
556
         /* Apply pre-emphasis */
557 558
         inp[i] = SHL32(x, SIG_SHIFT) - m;
         m = SHR32(MULT16_16(coef0, x), 15-SIG_SHIFT);
559
      }
560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578
   }
   *mem = m;
}



static opus_val32 l1_metric(const celt_norm *tmp, int N, int LM, opus_val16 bias)
{
   int i;
   opus_val32 L1;
   L1 = 0;
   for (i=0;i<N;i++)
      L1 += EXTEND32(ABS16(tmp[i]));
   /* When in doubt, prefer good freq resolution */
   L1 = MAC16_32_Q15(L1, LM*bias, L1);
   return L1;

}

579 580
static int tf_analysis(const CELTMode *m, int len, int isTransient,
      int *tf_res, int lambda, celt_norm *X, int N0, int LM,
581
      opus_val16 tf_estimate, int tf_chan)
582 583 584 585 586 587 588 589 590 591 592 593 594 595 596
{
   int i;
   VARDECL(int, metric);
   int cost0;
   int cost1;
   VARDECL(int, path0);
   VARDECL(int, path1);
   VARDECL(celt_norm, tmp);
   VARDECL(celt_norm, tmp_1);
   int sel;
   int selcost[2];
   int tf_select=0;
   opus_val16 bias;

   SAVE_STACK;
597
   bias = MULT16_16_Q14(QCONST16(.04f,15), MAX16(-QCONST16(.25f,14), QCONST16(.5f,14)-tf_estimate));
598 599 600 601 602 603 604 605 606 607
   /*printf("%f ", bias);*/

   ALLOC(metric, len, int);
   ALLOC(tmp, (m->eBands[len]-m->eBands[len-1])<<LM, celt_norm);
   ALLOC(tmp_1, (m->eBands[len]-m->eBands[len-1])<<LM, celt_norm);
   ALLOC(path0, len, int);
   ALLOC(path1, len, int);

   for (i=0;i<len;i++)
   {
608
      int k, N;
609 610 611 612 613 614
      int narrow;
      opus_val32 L1, best_L1;
      int best_level=0;
      N = (m->eBands[i+1]-m->eBands[i])<<LM;
      /* band is too narrow to be split down to LM=-1 */
      narrow = (m->eBands[i+1]-m->eBands[i])==1;
615
      OPUS_COPY(tmp, &X[tf_chan*N0 + (m->eBands[i]<<LM)], N);
616 617 618 619 620 621 622 623 624
      /* Just add the right channel if we're in stereo */
      /*if (C==2)
         for (j=0;j<N;j++)
            tmp[j] = ADD16(SHR16(tmp[j], 1),SHR16(X[N0+j+(m->eBands[i]<<LM)], 1));*/
      L1 = l1_metric(tmp, N, isTransient ? LM : 0, bias);
      best_L1 = L1;
      /* Check the -1 case for transients */
      if (isTransient && !narrow)
      {
625
         OPUS_COPY(tmp_1, tmp, N);
626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783
         haar1(tmp_1, N>>LM, 1<<LM);
         L1 = l1_metric(tmp_1, N, LM+1, bias);
         if (L1<best_L1)
         {
            best_L1 = L1;
            best_level = -1;
         }
      }
      /*printf ("%f ", L1);*/
      for (k=0;k<LM+!(isTransient||narrow);k++)
      {
         int B;

         if (isTransient)
            B = (LM-k-1);
         else
            B = k+1;

         haar1(tmp, N>>k, 1<<k);

         L1 = l1_metric(tmp, N, B, bias);

         if (L1 < best_L1)
         {
            best_L1 = L1;
            best_level = k+1;
         }
      }
      /*printf ("%d ", isTransient ? LM-best_level : best_level);*/
      /* metric is in Q1 to be able to select the mid-point (-0.5) for narrower bands */
      if (isTransient)
         metric[i] = 2*best_level;
      else
         metric[i] = -2*best_level;
      /* For bands that can't be split to -1, set the metric to the half-way point to avoid
         biasing the decision */
      if (narrow && (metric[i]==0 || metric[i]==-2*LM))
         metric[i]-=1;
      /*printf("%d ", metric[i]);*/
   }
   /*printf("\n");*/
   /* Search for the optimal tf resolution, including tf_select */
   tf_select = 0;
   for (sel=0;sel<2;sel++)
   {
      cost0 = 0;
      cost1 = isTransient ? 0 : lambda;
      for (i=1;i<len;i++)
      {
         int curr0, curr1;
         curr0 = IMIN(cost0, cost1 + lambda);
         curr1 = IMIN(cost0 + lambda, cost1);
         cost0 = curr0 + abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*sel+0]);
         cost1 = curr1 + abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*sel+1]);
      }
      cost0 = IMIN(cost0, cost1);
      selcost[sel]=cost0;
   }
   /* For now, we're conservative and only allow tf_select=1 for transients.
    * If tests confirm it's useful for non-transients, we could allow it. */
   if (selcost[1]<selcost[0] && isTransient)
      tf_select=1;
   cost0 = 0;
   cost1 = isTransient ? 0 : lambda;
   /* Viterbi forward pass */
   for (i=1;i<len;i++)
   {
      int curr0, curr1;
      int from0, from1;

      from0 = cost0;
      from1 = cost1 + lambda;
      if (from0 < from1)
      {
         curr0 = from0;
         path0[i]= 0;
      } else {
         curr0 = from1;
         path0[i]= 1;
      }

      from0 = cost0 + lambda;
      from1 = cost1;
      if (from0 < from1)
      {
         curr1 = from0;
         path1[i]= 0;
      } else {
         curr1 = from1;
         path1[i]= 1;
      }
      cost0 = curr0 + abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*tf_select+0]);
      cost1 = curr1 + abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*tf_select+1]);
   }
   tf_res[len-1] = cost0 < cost1 ? 0 : 1;
   /* Viterbi backward pass to check the decisions */
   for (i=len-2;i>=0;i--)
   {
      if (tf_res[i+1] == 1)
         tf_res[i] = path1[i+1];
      else
         tf_res[i] = path0[i+1];
   }
   /*printf("%d %f\n", *tf_sum, tf_estimate);*/
   RESTORE_STACK;
#ifdef FUZZING
   tf_select = rand()&0x1;
   tf_res[0] = rand()&0x1;
   for (i=1;i<len;i++)
      tf_res[i] = tf_res[i-1] ^ ((rand()&0xF) == 0);
#endif
   return tf_select;
}

static void tf_encode(int start, int end, int isTransient, int *tf_res, int LM, int tf_select, ec_enc *enc)
{
   int curr, i;
   int tf_select_rsv;
   int tf_changed;
   int logp;
   opus_uint32 budget;
   opus_uint32 tell;
   budget = enc->storage*8;
   tell = ec_tell(enc);
   logp = isTransient ? 2 : 4;
   /* Reserve space to code the tf_select decision. */
   tf_select_rsv = LM>0 && tell+logp+1 <= budget;
   budget -= tf_select_rsv;
   curr = tf_changed = 0;
   for (i=start;i<end;i++)
   {
      if (tell+logp<=budget)
      {
         ec_enc_bit_logp(enc, tf_res[i] ^ curr, logp);
         tell = ec_tell(enc);
         curr = tf_res[i];
         tf_changed |= curr;
      }
      else
         tf_res[i] = curr;
      logp = isTransient ? 4 : 5;
   }
   /* Only code tf_select if it would actually make a difference. */
   if (tf_select_rsv &&
         tf_select_table[LM][4*isTransient+0+tf_changed]!=
         tf_select_table[LM][4*isTransient+2+tf_changed])
      ec_enc_bit_logp(enc, tf_select, 1);
   else
      tf_select = 0;
   for (i=start;i<end;i++)
      tf_res[i] = tf_select_table[LM][4*isTransient+2*tf_select+tf_res[i]];
   /*for(i=0;i<end;i++)printf("%d ", isTransient ? tf_res[i] : LM+tf_res[i]);printf("\n");*/
}


static int alloc_trim_analysis(const CELTMode *m, const celt_norm *X,
      const opus_val16 *bandLogE, int end, int LM, int C, int N0,
      AnalysisInfo *analysis, opus_val16 *stereo_saving, opus_val16 tf_estimate,
784
      int intensity, opus_val16 surround_trim, int arch)
785 786 787 788
{
   int i;
   opus_val32 diff=0;
   int c;
789
   int trim_index;
790 791 792 793 794 795 796 797 798
   opus_val16 trim = QCONST16(5.f, 8);
   opus_val16 logXC, logXC2;
   if (C==2)
   {
      opus_val16 sum = 0; /* Q10 */
      opus_val16 minXC; /* Q10 */
      /* Compute inter-channel correlation for low frequencies */
      for (i=0;i<8;i++)
      {
799
         opus_val32 partial;
800 801
         partial = celt_inner_prod(&X[m->eBands[i]<<LM], &X[N0+(m->eBands[i]<<LM)],
               (m->eBands[i+1]-m->eBands[i])<<LM, arch);
802 803 804 805 806 807 808
         sum = ADD16(sum, EXTRACT16(SHR32(partial, 18)));
      }
      sum = MULT16_16_Q15(QCONST16(1.f/8, 15), sum);
      sum = MIN16(QCONST16(1.f, 10), ABS16(sum));
      minXC = sum;
      for (i=8;i<intensity;i++)
      {
809
         opus_val32 partial;
810 811
         partial = celt_inner_prod(&X[m->eBands[i]<<LM], &X[N0+(m->eBands[i]<<LM)],
               (m->eBands[i+1]-m->eBands[i])<<LM, arch);
812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839
         minXC = MIN16(minXC, ABS16(EXTRACT16(SHR32(partial, 18))));
      }
      minXC = MIN16(QCONST16(1.f, 10), ABS16(minXC));
      /*printf ("%f\n", sum);*/
      /* mid-side savings estimations based on the LF average*/
      logXC = celt_log2(QCONST32(1.001f, 20)-MULT16_16(sum, sum));
      /* mid-side savings estimations based on min correlation */
      logXC2 = MAX16(HALF16(logXC), celt_log2(QCONST32(1.001f, 20)-MULT16_16(minXC, minXC)));
#ifdef FIXED_POINT
      /* Compensate for Q20 vs Q14 input and convert output to Q8 */
      logXC = PSHR32(logXC-QCONST16(6.f, DB_SHIFT),DB_SHIFT-8);
      logXC2 = PSHR32(logXC2-QCONST16(6.f, DB_SHIFT),DB_SHIFT-8);
#endif

      trim += MAX16(-QCONST16(4.f, 8), MULT16_16_Q15(QCONST16(.75f,15),logXC));
      *stereo_saving = MIN16(*stereo_saving + QCONST16(0.25f, 8), -HALF16(logXC2));
   }

   /* Estimate spectral tilt */
   c=0; do {
      for (i=0;i<end-1;i++)
      {
         diff += bandLogE[i+c*m->nbEBands]*(opus_int32)(2+2*i-end);
      }
   } while (++c<C);
   diff /= C*(end-1);
   /*printf("%f\n", diff);*/
   trim -= MAX16(-QCONST16(2.f, 8), MIN16(QCONST16(2.f, 8), SHR16(diff+QCONST16(1.f, DB_SHIFT),DB_SHIFT-8)/6 ));
840
   trim -= SHR16(surround_trim, DB_SHIFT-8);
841
   trim -= 2*SHR16(tf_estimate, 14-8);
842
#ifndef DISABLE_FLOAT_API
843 844
   if (analysis->valid)
   {
845 846
      trim -= MAX16(-QCONST16(2.f, 8), MIN16(QCONST16(2.f, 8),
            (opus_val16)(QCONST16(2.f, 8)*(analysis->tonality_slope+.05f))));
847
   }
848 849
#else
   (void)analysis;
850 851 852 853 854
#endif

#ifdef FIXED_POINT
   trim_index = PSHR32(trim, 8);
#else
855
   trim_index = (int)floor(.5f+trim);
856
#endif
857
   trim_index = IMAX(0, IMIN(10, trim_index));
858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896
   /*printf("%d\n", trim_index);*/
#ifdef FUZZING
   trim_index = rand()%11;
#endif
   return trim_index;
}

static int stereo_analysis(const CELTMode *m, const celt_norm *X,
      int LM, int N0)
{
   int i;
   int thetas;
   opus_val32 sumLR = EPSILON, sumMS = EPSILON;

   /* Use the L1 norm to model the entropy of the L/R signal vs the M/S signal */
   for (i=0;i<13;i++)
   {
      int j;
      for (j=m->eBands[i]<<LM;j<m->eBands[i+1]<<LM;j++)
      {
         opus_val32 L, R, M, S;
         /* We cast to 32-bit first because of the -32768 case */
         L = EXTEND32(X[j]);
         R = EXTEND32(X[N0+j]);
         M = ADD32(L, R);
         S = SUB32(L, R);
         sumLR = ADD32(sumLR, ADD32(ABS32(L), ABS32(R)));
         sumMS = ADD32(sumMS, ADD32(ABS32(M), ABS32(S)));
      }
   }
   sumMS = MULT16_32_Q15(QCONST16(0.707107f, 15), sumMS);
   thetas = 13;
   /* We don't need thetas for lower bands with LM<=1 */
   if (LM<=1)
      thetas -= 8;
   return MULT16_32_Q15((m->eBands[13]<<(LM+1))+thetas, sumMS)
         > MULT16_32_Q15(m->eBands[13]<<(LM+1), sumLR);
}

897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956
#define MSWAP(a,b) do {opus_val16 tmp = a;a=b;b=tmp;} while(0)
static opus_val16 median_of_5(const opus_val16 *x)
{
   opus_val16 t0, t1, t2, t3, t4;
   t2 = x[2];
   if (x[0] > x[1])
   {
      t0 = x[1];
      t1 = x[0];
   } else {
      t0 = x[0];
      t1 = x[1];
   }
   if (x[3] > x[4])
   {
      t3 = x[4];
      t4 = x[3];
   } else {
      t3 = x[3];
      t4 = x[4];
   }
   if (t0 > t3)
   {
      MSWAP(t0, t3);
      MSWAP(t1, t4);
   }
   if (t2 > t1)
   {
      if (t1 < t3)
         return MIN16(t2, t3);
      else
         return MIN16(t4, t1);
   } else {
      if (t2 < t3)
         return MIN16(t1, t3);
      else
         return MIN16(t2, t4);
   }
}

static opus_val16 median_of_3(const opus_val16 *x)
{
   opus_val16 t0, t1, t2;
   if (x[0] > x[1])
   {
      t0 = x[1];
      t1 = x[0];
   } else {
      t0 = x[0];
      t1 = x[1];
   }
   t2 = x[2];
   if (t1 < t2)
      return t1;
   else if (t0 < t2)
      return t2;
   else
      return t0;
}

957
static opus_val16 dynalloc_analysis(const opus_val16 *bandLogE, const opus_val16 *bandLogE2,
958 959
      int nbEBands, int start, int end, int C, int *offsets, int lsb_depth, const opus_int16 *logN,
      int isTransient, int vbr, int constrained_vbr, const opus_int16 *eBands, int LM,
960
      int effectiveBytes, opus_int32 *tot_boost_, int lfe, opus_val16 *surround_dynalloc)
961 962 963 964 965 966 967 968 969
{
   int i, c;
   opus_int32 tot_boost=0;
   opus_val16 maxDepth;
   VARDECL(opus_val16, follower);
   VARDECL(opus_val16, noise_floor);
   SAVE_STACK;
   ALLOC(follower, C*nbEBands, opus_val16);
   ALLOC(noise_floor, C*nbEBands, opus_val16);
970
   OPUS_CLEAR(offsets, nbEBands);
971
   /* Dynamic allocation code */
972
   maxDepth=-QCONST16(31.9f, DB_SHIFT);
973 974 975 976 977 978 979 980 981 982 983 984 985 986
   for (i=0;i<end;i++)
   {
      /* Noise floor must take into account eMeans, the depth, the width of the bands
         and the preemphasis filter (approx. square of bark band ID) */
      noise_floor[i] = MULT16_16(QCONST16(0.0625f, DB_SHIFT),logN[i])
            +QCONST16(.5f,DB_SHIFT)+SHL16(9-lsb_depth,DB_SHIFT)-SHL16(eMeans[i],6)
            +MULT16_16(QCONST16(.0062,DB_SHIFT),(i+5)*(i+5));
   }
   c=0;do
   {
      for (i=0;i<end;i++)
         maxDepth = MAX16(maxDepth, bandLogE[c*nbEBands+i]-noise_floor[i]);
   } while (++c<C);
   /* Make sure that dynamic allocation can't make us bust the budget */
987
   if (effectiveBytes > 50 && LM>=1 && !lfe)
988 989 990 991
   {
      int last=0;
      c=0;do
      {
992 993
         opus_val16 offset;
         opus_val16 tmp;
994 995 996
         opus_val16 *f;
         f = &follower[c*nbEBands];
         f[0] = bandLogE2[c*nbEBands];
997 998 999 1000 1001 1002 1003
         for (i=1;i<end;i++)
         {
            /* The last band to be at least 3 dB higher than the previous one
               is the last we'll consider. Otherwise, we run into problems on
               bandlimited signals. */
            if (bandLogE2[c*nbEBands+i] > bandLogE2[c*nbEBands+i-1]+QCONST16(.5f,DB_SHIFT))
               last=i;
1004
            f[i] = MIN16(f[i-1]+QCONST16(1.5f,DB_SHIFT), bandLogE2[c*nbEBands+i]);
1005 1006
         }
         for (i=last-1;i>=0;i--)
1007
            f[i] = MIN16(f[i], MIN16(f[i+1]+QCONST16(2.f,DB_SHIFT), bandLogE2[c*nbEBands+i]));
1008 1009 1010 1011 1012 1013

         /* Combine with a median filter to avoid dynalloc triggering unnecessarily.
            The "offset" value controls how conservative we are -- a higher offset
            reduces the impact of the median filter and makes dynalloc use more bits. */
         offset = QCONST16(1.f, DB_SHIFT);
         for (i=2;i<end-2;i++)
1014
            f[i] = MAX16(f[i], median_of_5(&bandLogE2[c*nbEBands+i-2])-offset);
1015
         tmp = median_of_3(&bandLogE2[c*nbEBands])-offset;
1016 1017
         f[0] = MAX16(f[0], tmp);
         f[1] = MAX16(f[1], tmp);
1018 1019 1020
         tmp = median_of_3(&bandLogE2[c*nbEBands+end-3])-offset;
         f[end-2] = MAX16(f[end-2], tmp);
         f[end-1] = MAX16(f[end-1], tmp);
1021

1022
         for (i=0;i<end;i++)
1023
            f[i] = MAX16(f[i], noise_floor[i]);
1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039
      } while (++c<C);
      if (C==2)
      {
         for (i=start;i<end;i++)
         {
            /* Consider 24 dB "cross-talk" */
            follower[nbEBands+i] = MAX16(follower[nbEBands+i], follower[         i]-QCONST16(4.f,DB_SHIFT));
            follower[         i] = MAX16(follower[         i], follower[nbEBands+i]-QCONST16(4.f,DB_SHIFT));
            follower[i] = HALF16(MAX16(0, bandLogE[i]-follower[i]) + MAX16(0, bandLogE[nbEBands+i]-follower[nbEBands+i]));
         }
      } else {
         for (i=start;i<end;i++)
         {
            follower[i] = MAX16(0, bandLogE[i]-follower[i]);
         }
      }
1040 1041
      for (i=start;i<end;i++)
         follower[i] = MAX16(follower[i], surround_dynalloc[i]);
1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062
      /* For non-transient CBR/CVBR frames, halve the dynalloc contribution */
      if ((!vbr || constrained_vbr)&&!isTransient)
      {
         for (i=start;i<end;i++)
            follower[i] = HALF16(follower[i]);
      }
      for (i=start;i<end;i++)
      {
         int width;
         int boost;
         int boost_bits;

         if (i<8)
            follower[i] *= 2;
         if (i>=12)
            follower[i] = HALF16(follower[i]);
         follower[i] = MIN16(follower[i], QCONST16(4, DB_SHIFT));

         width = C*(eBands[i+1]-eBands[i])<<LM;
         if (width<6)
         {
1063
            boost = (int)SHR32(EXTEND32(follower[i]),DB_SHIFT);
1064 1065
            boost_bits = boost*width<<BITRES;
         } else if (width > 48) {
1066
            boost = (int)SHR32(EXTEND32(follower[i])*8,DB_SHIFT);
1067 1068
            boost_bits = (boost*width<<BITRES)/8;
         } else {
1069
            boost = (int)SHR32(EXTEND32(follower[i])*width/6,DB_SHIFT);
1070 1071 1072 1073 1074 1075
            boost_bits = boost*6<<BITRES;
         }
         /* For CBR and non-transient CVBR frames, limit dynalloc to 1/4 of the bits */
         if ((!vbr || (constrained_vbr&&!isTransient))
               && (tot_boost+boost_bits)>>BITRES>>3 > effectiveBytes/4)
         {
1076 1077 1078
            opus_int32 cap = ((effectiveBytes/4)<<BITRES<<3);
            offsets[i] = cap-tot_boost;
            tot_boost = cap;
1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091
            break;
         } else {
            offsets[i] = boost;
            tot_boost += boost_bits;
         }
      }
   }
   *tot_boost_ = tot_boost;
   RESTORE_STACK;
   return maxDepth;
}


1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103
static int run_prefilter(CELTEncoder *st, celt_sig *in, celt_sig *prefilter_mem, int CC, int N,
      int prefilter_tapset, int *pitch, opus_val16 *gain, int *qgain, int enabled, int nbAvailableBytes)
{
   int c;
   VARDECL(celt_sig, _pre);
   celt_sig *pre[2];
   const CELTMode *mode;
   int pitch_index;
   opus_val16 gain1;
   opus_val16 pf_threshold;
   int pf_on;
   int qg;
Jean-Marc Valin's avatar
Jean-Marc Valin committed
1104
   int overlap;
1105 1106 1107
   SAVE_STACK;

   mode = st->mode;
Jean-Marc Valin's avatar
Jean-Marc Valin committed
1108
   overlap = mode->overlap;
1109 1110 1111 1112 1113 1114 1115 1116
   ALLOC(_pre, CC*(N+COMBFILTER_MAXPERIOD), celt_sig);

   pre[0] = _pre;
   pre[1] = _pre + (N+COMBFILTER_MAXPERIOD);


   c=0; do {
      OPUS_COPY(pre[c], prefilter_mem+c*COMBFILTER_MAXPERIOD, COMBFILTER_MAXPERIOD);
Jean-Marc Valin's avatar
Jean-Marc Valin committed
1117
      OPUS_COPY(pre[c]+COMBFILTER_MAXPERIOD, in+c*(N+overlap)+overlap, N);
1118 1119 1120 1121 1122 1123 1124
   } while (++c<CC);

   if (enabled)
   {
      VARDECL(opus_val16, pitch_buf);
      ALLOC(pitch_buf, (COMBFILTER_MAXPERIOD+N)>>1, opus_val16);

1125
      pitch_downsample(pre, pitch_buf, COMBFILTER_MAXPERIOD+N, CC, st->arch);
1126 1127 1128
      /* Don't search for the fir last 1.5 octave of the range because
         there's too many false-positives due to short-term correlation */
      pitch_search(pitch_buf+(COMBFILTER_MAXPERIOD>>1), pitch_buf, N,
1129 1130
            COMBFILTER_MAXPERIOD-3*COMBFILTER_MINPERIOD, &pitch_index,
            st->arch);
1131 1132 1133
      pitch_index = COMBFILTER_MAXPERIOD-pitch_index;

      gain1 = remove_doubling(pitch_buf, COMBFILTER_MAXPERIOD, COMBFILTER_MINPERIOD,
1134
            N, &pitch_index, st->prefilter_period, st->prefilter_gain, st->arch);
1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189
      if (pitch_index > COMBFILTER_MAXPERIOD-2)
         pitch_index = COMBFILTER_MAXPERIOD-2;
      gain1 = MULT16_16_Q15(QCONST16(.7f,15),gain1);
      /*printf("%d %d %f %f\n", pitch_change, pitch_index, gain1, st->analysis.tonality);*/
      if (st->loss_rate>2)
         gain1 = HALF32(gain1);
      if (st->loss_rate>4)
         gain1 = HALF32(gain1);
      if (st->loss_rate>8)
         gain1 = 0;
   } else {
      gain1 = 0;
      pitch_index = COMBFILTER_MINPERIOD;
   }

   /* Gain threshold for enabling the prefilter/postfilter */
   pf_threshold = QCONST16(.2f,15);

   /* Adjusting the threshold based on rate and continuity */
   if (abs(pitch_index-st->prefilter_period)*10>pitch_index)
      pf_threshold += QCONST16(.2f,15);
   if (nbAvailableBytes<25)
      pf_threshold += QCONST16(.1f,15);
   if (nbAvailableBytes<35)
      pf_threshold += QCONST16(.1f,15);
   if (st->prefilter_gain > QCONST16(.4f,15))
      pf_threshold -= QCONST16(.1f,15);
   if (st->prefilter_gain > QCONST16(.55f,15))
      pf_threshold -= QCONST16(.1f,15);

   /* Hard threshold at 0.2 */
   pf_threshold = MAX16(pf_threshold, QCONST16(.2f,15));
   if (gain1<pf_threshold)
   {
      gain1 = 0;
      pf_on = 0;
      qg = 0;
   } else {
      /*This block is not gated by a total bits check only because
        of the nbAvailableBytes check above.*/
      if (ABS16(gain1-st->prefilter_gain)<QCONST16(.1f,15))
         gain1=st->prefilter_gain;

#ifdef FIXED_POINT
      qg = ((gain1+1536)>>10)/3-1;
#else
      qg = (int)floor(.5f+gain1*32/3)-1;
#endif
      qg = IMAX(0, IMIN(7, qg));
      gain1 = QCONST16(0.09375f,15)*(qg+1);
      pf_on = 1;
   }
   /*printf("%d %f\n", pitch_index, gain1);*/

   c=0; do {
Jean-Marc Valin's avatar
Jean-Marc Valin committed
1190
      int offset = mode->shortMdctSize-overlap;
1191
      st->prefilter_period=IMAX(st->prefilter_period, COMBFILTER_MINPERIOD);
Jean-Marc Valin's avatar
Jean-Marc Valin committed
1192
      OPUS_COPY(in+c*(N+overlap), st->in_mem+c*(overlap), overlap);
1193
      if (offset)
Jean-Marc Valin's avatar
Jean-Marc Valin committed
1194
         comb_filter(in+c*(N+overlap)+overlap, pre[c]+COMBFILTER_MAXPERIOD,
1195
               st->prefilter_period, st->prefilter_period, offset, -st->prefilter_gain, -st->prefilter_gain,
1196
               st->prefilter_tapset, st->prefilter_tapset, NULL, 0, st->arch);
1197

Jean-Marc Valin's avatar
Jean-Marc Valin committed
1198
      comb_filter(in+c*(N+overlap)+overlap+offset, pre[c]+COMBFILTER_MAXPERIOD+offset,
1199
            st->prefilter_period, pitch_index, N-offset, -st->prefilter_gain, -gain1,
1200
            st->prefilter_tapset, prefilter_tapset, mode->window, overlap, st->arch);
Jean-Marc Valin's avatar
Jean-Marc Valin committed
1201
      OPUS_COPY(st->in_mem+c*(overlap), in+c*(N+overlap)+N, overlap);
1202 1203 1204

      if (N>COMBFILTER_MAXPERIOD)
      {
1205
         OPUS_COPY(prefilter_mem+c*COMBFILTER_MAXPERIOD, pre[c]+N, COMBFILTER_MAXPERIOD);
1206 1207
      } else {
         OPUS_MOVE(prefilter_mem+c*COMBFILTER_MAXPERIOD, prefilter_mem+c*COMBFILTER_MAXPERIOD+N, COMBFILTER_MAXPERIOD-N);
1208
         OPUS_COPY(prefilter_mem+c*COMBFILTER_MAXPERIOD+COMBFILTER_MAXPERIOD-N, pre[c]+COMBFILTER_MAXPERIOD, N);
1209 1210 1211 1212 1213 1214 1215 1216 1217 1218
      }
   } while (++c<CC);

   RESTORE_STACK;
   *gain = gain1;
   *pitch = pitch_index;
   *qgain = qg;
   return pf_on;
}

1219 1220 1221 1222
static int compute_vbr(const CELTMode *mode, AnalysisInfo *analysis, opus_int32 base_target,
      int LM, opus_int32 bitrate, int lastCodedBands, int C, int intensity,
      int constrained_vbr, opus_val16 stereo_saving, int tot_boost,
      opus_val16 tf_estimate, int pitch_change, opus_val16 maxDepth,
1223 1224
      int variable_duration, int lfe, int has_surround_mask, opus_val16 surround_masking,
      opus_val16 temporal_vbr)
1225 1226 1227 1228 1229
{
   /* The target rate in 8th bits per frame */
   opus_int32 target;
   int coded_bins;
   int coded_bands;
1230
   opus_val16 tf_calibration;
1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244
   int nbEBands;
   const opus_int16 *eBands;

   nbEBands = mode->nbEBands;
   eBands = mode->eBands;

   coded_bands = lastCodedBands ? lastCodedBands : nbEBands;
   coded_bins = eBands[coded_bands]<<LM;
   if (C==2)
      coded_bins += eBands[IMIN(intensity, coded_bands)]<<LM;

   target = base_target;

   /*printf("%f %f %f %f %d %d ", st->analysis.activity, st->analysis.tonality, tf_estimate, st->stereo_saving, tot_boost, coded_bands);*/
1245
#ifndef DISABLE_FLOAT_API
1246
   if (analysis->valid && analysis->activity<.4)
1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258
      target -= (opus_int32)((coded_bins<<BITRES)*(.4f-analysis->activity));
#endif
   /* Stereo savings */
   if (C==2)
   {
      int coded_stereo_bands;
      int coded_stereo_dof;
      opus_val16 max_frac;
      coded_stereo_bands = IMIN(intensity, coded_bands);
      coded_stereo_dof = (eBands[coded_stereo_bands]<<LM)-coded_stereo_bands;
      /* Maximum fraction of the bits we can save if the signal is mono. */
      max_frac = DIV32_16(MULT16_16(QCONST16(0.8f, 15), coded_stereo_dof), coded_bins);
1259
      stereo_saving = MIN16(stereo_saving, QCONST16(1.f, 8));
1260 1261 1262 1263 1264 1265 1266
      /*printf("%d %d %d ", coded_stereo_dof, coded_bins, tot_boost);*/
      target -= (opus_int32)MIN32(MULT16_32_Q15(max_frac,target),
                      SHR32(MULT16_16(stereo_saving-QCONST16(0.1f,8),(coded_stereo_dof<<BITRES)),8));
   }
   /* Boost the rate according to dynalloc (minus the dynalloc average for calibration). */
   target += tot_boost-(16<<LM);
   /* Apply transient boost, compensating for average boost. */
1267 1268
   tf_calibration = variable_duration==OPUS_FRAMESIZE_VARIABLE ?
                    QCONST16(0.02f,14) : QCONST16(0.04f,14);
1269 1270
   target += (opus_int32)SHL32(MULT16_32_Q15(tf_estimate-tf_calibration, target),1);

1271
#ifndef DISABLE_FLOAT_API
1272
   /* Apply tonality boost */
Jean-Marc Valin's avatar
Jean-Marc Valin committed
1273 1274
   if (analysis->valid && !lfe)
   {
1275 1276 1277 1278 1279 1280 1281 1282
      opus_int32 tonal_target;
      float tonal;

      /* Tonality boost (compensating for the average). */
      tonal = MAX16(0.f,analysis->tonality-.15f)-0.09f;
      tonal_target = target + (opus_int32)((coded_bins<<BITRES)*1.2f*tonal);
      if (pitch_change)
         tonal_target +=  (opus_int32)((coded_bins<<BITRES)*.8f);
1283
      /*printf("%f %f ", analysis->tonality, tonal);*/
1284 1285
      target = tonal_target;
   }
1286 1287 1288
#else
   (void)analysis;
   (void)pitch_change;
1289 1290 1291 1292
#endif

   if (has_surround_mask&&!lfe)
   {
Jean-Marc Valin's avatar
Jean-Marc Valin committed
1293
      opus_int32 surround_target = target + (opus_int32)SHR32(MULT16_16(surround_masking,coded_bins<<BITRES), DB_SHIFT);
1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308
      /*printf("%f %d %d %d %d %d %d ", surround_masking, coded_bins, st->end, st->intensity, surround_target, target, st->bitrate);*/
      target = IMAX(target/4, surround_target);
   }

   {
      opus_int32 floor_depth;
      int bins;
      bins = eBands[nbEBands-2]<<LM;
      /*floor_depth = SHR32(MULT16_16((C*bins<<BITRES),celt_log2(SHL32(MAX16(1,sample_max),13))), DB_SHIFT);*/
      floor_depth = (opus_int32)SHR32(MULT16_16((C*bins<<BITRES),maxDepth), DB_SHIFT);
      floor_depth = IMAX(floor_depth, target>>2);
      target = IMIN(target, floor_depth);
      /*printf("%f %d\n", maxDepth, floor_depth);*/
   }

1309 1310 1311
   /* Make VBR less aggressive for constrained VBR because we can't keep a higher bitrate
      for long. Needs tuning. */
   if ((!has_surround_mask||lfe) && constrained_vbr)
1312
   {
1313
      target = base_target + (opus_int32)MULT16_32_Q15(QCONST16(0.67f, 15), target-base_target);
1314
   }