OLD | NEW |
(Empty) | |
| 1 /* Copyright (c) 2007-2008 CSIRO |
| 2 Copyright (c) 2007-2010 Xiph.Org Foundation |
| 3 Copyright (c) 2008 Gregory Maxwell |
| 4 Written by Jean-Marc Valin and Gregory Maxwell */ |
| 5 /* |
| 6 Redistribution and use in source and binary forms, with or without |
| 7 modification, are permitted provided that the following conditions |
| 8 are met: |
| 9 |
| 10 - Redistributions of source code must retain the above copyright |
| 11 notice, this list of conditions and the following disclaimer. |
| 12 |
| 13 - Redistributions in binary form must reproduce the above copyright |
| 14 notice, this list of conditions and the following disclaimer in the |
| 15 documentation and/or other materials provided with the distribution. |
| 16 |
| 17 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS |
| 18 ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT |
| 19 LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR |
| 20 A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER |
| 21 OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, |
| 22 EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, |
| 23 PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR |
| 24 PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF |
| 25 LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING |
| 26 NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS |
| 27 SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
| 28 */ |
| 29 |
| 30 #ifdef HAVE_CONFIG_H |
| 31 #include "config.h" |
| 32 #endif |
| 33 |
| 34 #define CELT_ENCODER_C |
| 35 |
| 36 #include "cpu_support.h" |
| 37 #include "os_support.h" |
| 38 #include "mdct.h" |
| 39 #include <math.h> |
| 40 #include "celt.h" |
| 41 #include "pitch.h" |
| 42 #include "bands.h" |
| 43 #include "modes.h" |
| 44 #include "entcode.h" |
| 45 #include "quant_bands.h" |
| 46 #include "rate.h" |
| 47 #include "stack_alloc.h" |
| 48 #include "mathops.h" |
| 49 #include "float_cast.h" |
| 50 #include <stdarg.h> |
| 51 #include "celt_lpc.h" |
| 52 #include "vq.h" |
| 53 |
| 54 |
| 55 /** Encoder state |
| 56 @brief Encoder state |
| 57 */ |
| 58 struct OpusCustomEncoder { |
| 59 const OpusCustomMode *mode; /**< Mode used by the encoder */ |
| 60 int overlap; |
| 61 int channels; |
| 62 int stream_channels; |
| 63 |
| 64 int force_intra; |
| 65 int clip; |
| 66 int disable_pf; |
| 67 int complexity; |
| 68 int upsample; |
| 69 int start, end; |
| 70 |
| 71 opus_int32 bitrate; |
| 72 int vbr; |
| 73 int signalling; |
| 74 int constrained_vbr; /* If zero, VBR can do whatever it likes with the r
ate */ |
| 75 int loss_rate; |
| 76 int lsb_depth; |
| 77 int variable_duration; |
| 78 int lfe; |
| 79 int arch; |
| 80 |
| 81 /* Everything beyond this point gets cleared on a reset */ |
| 82 #define ENCODER_RESET_START rng |
| 83 |
| 84 opus_uint32 rng; |
| 85 int spread_decision; |
| 86 opus_val32 delayedIntra; |
| 87 int tonal_average; |
| 88 int lastCodedBands; |
| 89 int hf_average; |
| 90 int tapset_decision; |
| 91 |
| 92 int prefilter_period; |
| 93 opus_val16 prefilter_gain; |
| 94 int prefilter_tapset; |
| 95 #ifdef RESYNTH |
| 96 int prefilter_period_old; |
| 97 opus_val16 prefilter_gain_old; |
| 98 int prefilter_tapset_old; |
| 99 #endif |
| 100 int consec_transient; |
| 101 AnalysisInfo analysis; |
| 102 |
| 103 opus_val32 preemph_memE[2]; |
| 104 opus_val32 preemph_memD[2]; |
| 105 |
| 106 /* VBR-related parameters */ |
| 107 opus_int32 vbr_reservoir; |
| 108 opus_int32 vbr_drift; |
| 109 opus_int32 vbr_offset; |
| 110 opus_int32 vbr_count; |
| 111 opus_val32 overlap_max; |
| 112 opus_val16 stereo_saving; |
| 113 int intensity; |
| 114 opus_val16 *energy_mask; |
| 115 opus_val16 spec_avg; |
| 116 |
| 117 #ifdef RESYNTH |
| 118 /* +MAX_PERIOD/2 to make space for overlap */ |
| 119 celt_sig syn_mem[2][2*MAX_PERIOD+MAX_PERIOD/2]; |
| 120 #endif |
| 121 |
| 122 celt_sig in_mem[1]; /* Size = channels*mode->overlap */ |
| 123 /* celt_sig prefilter_mem[], Size = channels*COMBFILTER_MAXPERIOD */ |
| 124 /* opus_val16 oldBandE[], Size = channels*mode->nbEBands */ |
| 125 /* opus_val16 oldLogE[], Size = channels*mode->nbEBands */ |
| 126 /* opus_val16 oldLogE2[], Size = channels*mode->nbEBands */ |
| 127 }; |
| 128 |
| 129 int celt_encoder_get_size(int channels) |
| 130 { |
| 131 CELTMode *mode = opus_custom_mode_create(48000, 960, NULL); |
| 132 return opus_custom_encoder_get_size(mode, channels); |
| 133 } |
| 134 |
| 135 OPUS_CUSTOM_NOSTATIC int opus_custom_encoder_get_size(const CELTMode *mode, int
channels) |
| 136 { |
| 137 int size = sizeof(struct CELTEncoder) |
| 138 + (channels*mode->overlap-1)*sizeof(celt_sig) /* celt_sig in_mem[cha
nnels*mode->overlap]; */ |
| 139 + channels*COMBFILTER_MAXPERIOD*sizeof(celt_sig) /* celt_sig prefilter_
mem[channels*COMBFILTER_MAXPERIOD]; */ |
| 140 + 3*channels*mode->nbEBands*sizeof(opus_val16); /* opus_val16 oldBandE
[channels*mode->nbEBands]; */ |
| 141 /* opus_val16 oldLogE[
channels*mode->nbEBands]; */ |
| 142 /* opus_val16 oldLogE2
[channels*mode->nbEBands]; */ |
| 143 return size; |
| 144 } |
| 145 |
| 146 #ifdef CUSTOM_MODES |
| 147 CELTEncoder *opus_custom_encoder_create(const CELTMode *mode, int channels, int
*error) |
| 148 { |
| 149 int ret; |
| 150 CELTEncoder *st = (CELTEncoder *)opus_alloc(opus_custom_encoder_get_size(mode
, channels)); |
| 151 /* init will handle the NULL case */ |
| 152 ret = opus_custom_encoder_init(st, mode, channels); |
| 153 if (ret != OPUS_OK) |
| 154 { |
| 155 opus_custom_encoder_destroy(st); |
| 156 st = NULL; |
| 157 } |
| 158 if (error) |
| 159 *error = ret; |
| 160 return st; |
| 161 } |
| 162 #endif /* CUSTOM_MODES */ |
| 163 |
| 164 int celt_encoder_init(CELTEncoder *st, opus_int32 sampling_rate, int channels) |
| 165 { |
| 166 int ret; |
| 167 ret = opus_custom_encoder_init(st, opus_custom_mode_create(48000, 960, NULL),
channels); |
| 168 if (ret != OPUS_OK) |
| 169 return ret; |
| 170 st->upsample = resampling_factor(sampling_rate); |
| 171 return OPUS_OK; |
| 172 } |
| 173 |
| 174 OPUS_CUSTOM_NOSTATIC int opus_custom_encoder_init(CELTEncoder *st, const CELTMod
e *mode, int channels) |
| 175 { |
| 176 if (channels < 0 || channels > 2) |
| 177 return OPUS_BAD_ARG; |
| 178 |
| 179 if (st==NULL || mode==NULL) |
| 180 return OPUS_ALLOC_FAIL; |
| 181 |
| 182 OPUS_CLEAR((char*)st, opus_custom_encoder_get_size(mode, channels)); |
| 183 |
| 184 st->mode = mode; |
| 185 st->overlap = mode->overlap; |
| 186 st->stream_channels = st->channels = channels; |
| 187 |
| 188 st->upsample = 1; |
| 189 st->start = 0; |
| 190 st->end = st->mode->effEBands; |
| 191 st->signalling = 1; |
| 192 |
| 193 st->arch = opus_select_arch(); |
| 194 |
| 195 st->constrained_vbr = 1; |
| 196 st->clip = 1; |
| 197 |
| 198 st->bitrate = OPUS_BITRATE_MAX; |
| 199 st->vbr = 0; |
| 200 st->force_intra = 0; |
| 201 st->complexity = 5; |
| 202 st->lsb_depth=24; |
| 203 |
| 204 opus_custom_encoder_ctl(st, OPUS_RESET_STATE); |
| 205 |
| 206 return OPUS_OK; |
| 207 } |
| 208 |
| 209 #ifdef CUSTOM_MODES |
| 210 void opus_custom_encoder_destroy(CELTEncoder *st) |
| 211 { |
| 212 opus_free(st); |
| 213 } |
| 214 #endif /* CUSTOM_MODES */ |
| 215 |
| 216 |
| 217 static int transient_analysis(const opus_val32 * OPUS_RESTRICT in, int len, int
C, |
| 218 opus_val16 *tf_estimate, int *tf_chan) |
| 219 { |
| 220 int i; |
| 221 VARDECL(opus_val16, tmp); |
| 222 opus_val32 mem0,mem1; |
| 223 int is_transient = 0; |
| 224 opus_int32 mask_metric = 0; |
| 225 int c; |
| 226 opus_val16 tf_max; |
| 227 int len2; |
| 228 /* Table of 6*64/x, trained on real data to minimize the average error */ |
| 229 static const unsigned char inv_table[128] = { |
| 230 255,255,156,110, 86, 70, 59, 51, 45, 40, 37, 33, 31, 28, 26, 25, |
| 231 23, 22, 21, 20, 19, 18, 17, 16, 16, 15, 15, 14, 13, 13, 12, 12, |
| 232 12, 12, 11, 11, 11, 10, 10, 10, 9, 9, 9, 9, 9, 9, 8, 8, |
| 233 8, 8, 8, 7, 7, 7, 7, 7, 7, 6, 6, 6, 6, 6, 6, 6, |
| 234 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5, 5, 5, 5, |
| 235 5, 5, 5, 5, 5, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, |
| 236 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 3, 3, |
| 237 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 2, |
| 238 }; |
| 239 SAVE_STACK; |
| 240 ALLOC(tmp, len, opus_val16); |
| 241 |
| 242 len2=len/2; |
| 243 tf_max = 0; |
| 244 for (c=0;c<C;c++) |
| 245 { |
| 246 opus_val32 mean; |
| 247 opus_int32 unmask=0; |
| 248 opus_val32 norm; |
| 249 opus_val16 maxE; |
| 250 mem0=0; |
| 251 mem1=0; |
| 252 /* High-pass filter: (1 - 2*z^-1 + z^-2) / (1 - z^-1 + .5*z^-2) */ |
| 253 for (i=0;i<len;i++) |
| 254 { |
| 255 opus_val32 x,y; |
| 256 x = SHR32(in[i+c*len],SIG_SHIFT); |
| 257 y = ADD32(mem0, x); |
| 258 #ifdef FIXED_POINT |
| 259 mem0 = mem1 + y - SHL32(x,1); |
| 260 mem1 = x - SHR32(y,1); |
| 261 #else |
| 262 mem0 = mem1 + y - 2*x; |
| 263 mem1 = x - .5f*y; |
| 264 #endif |
| 265 tmp[i] = EXTRACT16(SHR32(y,2)); |
| 266 /*printf("%f ", tmp[i]);*/ |
| 267 } |
| 268 /*printf("\n");*/ |
| 269 /* First few samples are bad because we don't propagate the memory */ |
| 270 for (i=0;i<12;i++) |
| 271 tmp[i] = 0; |
| 272 |
| 273 #ifdef FIXED_POINT |
| 274 /* Normalize tmp to max range */ |
| 275 { |
| 276 int shift=0; |
| 277 shift = 14-celt_ilog2(1+celt_maxabs16(tmp, len)); |
| 278 if (shift!=0) |
| 279 { |
| 280 for (i=0;i<len;i++) |
| 281 tmp[i] = SHL16(tmp[i], shift); |
| 282 } |
| 283 } |
| 284 #endif |
| 285 |
| 286 mean=0; |
| 287 mem0=0; |
| 288 /* Grouping by two to reduce complexity */ |
| 289 /* Forward pass to compute the post-echo threshold*/ |
| 290 for (i=0;i<len2;i++) |
| 291 { |
| 292 opus_val16 x2 = PSHR32(MULT16_16(tmp[2*i],tmp[2*i]) + MULT16_16(tmp[2*i
+1],tmp[2*i+1]),16); |
| 293 mean += x2; |
| 294 #ifdef FIXED_POINT |
| 295 /* FIXME: Use PSHR16() instead */ |
| 296 tmp[i] = mem0 + PSHR32(x2-mem0,4); |
| 297 #else |
| 298 tmp[i] = mem0 + MULT16_16_P15(QCONST16(.0625f,15),x2-mem0); |
| 299 #endif |
| 300 mem0 = tmp[i]; |
| 301 } |
| 302 |
| 303 mem0=0; |
| 304 maxE=0; |
| 305 /* Backward pass to compute the pre-echo threshold */ |
| 306 for (i=len2-1;i>=0;i--) |
| 307 { |
| 308 #ifdef FIXED_POINT |
| 309 /* FIXME: Use PSHR16() instead */ |
| 310 tmp[i] = mem0 + PSHR32(tmp[i]-mem0,3); |
| 311 #else |
| 312 tmp[i] = mem0 + MULT16_16_P15(QCONST16(0.125f,15),tmp[i]-mem0); |
| 313 #endif |
| 314 mem0 = tmp[i]; |
| 315 maxE = MAX16(maxE, mem0); |
| 316 } |
| 317 /*for (i=0;i<len2;i++)printf("%f ", tmp[i]/mean);printf("\n");*/ |
| 318 |
| 319 /* Compute the ratio of the "frame energy" over the harmonic mean of the e
nergy. |
| 320 This essentially corresponds to a bitrate-normalized temporal noise-to-
mask |
| 321 ratio */ |
| 322 |
| 323 /* As a compromise with the old transient detector, frame energy is the |
| 324 geometric mean of the energy and half the max */ |
| 325 #ifdef FIXED_POINT |
| 326 /* Costs two sqrt() to avoid overflows */ |
| 327 mean = MULT16_16(celt_sqrt(mean), celt_sqrt(MULT16_16(maxE,len2>>1))); |
| 328 #else |
| 329 mean = celt_sqrt(mean * maxE*.5*len2); |
| 330 #endif |
| 331 /* Inverse of the mean energy in Q15+6 */ |
| 332 norm = SHL32(EXTEND32(len2),6+14)/ADD32(EPSILON,SHR32(mean,1)); |
| 333 /* Compute harmonic mean discarding the unreliable boundaries |
| 334 The data is smooth, so we only take 1/4th of the samples */ |
| 335 unmask=0; |
| 336 for (i=12;i<len2-5;i+=4) |
| 337 { |
| 338 int id; |
| 339 #ifdef FIXED_POINT |
| 340 id = IMAX(0,IMIN(127,MULT16_32_Q15(tmp[i],norm))); /* Do not round to n
earest */ |
| 341 #else |
| 342 id = IMAX(0,IMIN(127,(int)floor(64*norm*tmp[i]))); /* Do not round to n
earest */ |
| 343 #endif |
| 344 unmask += inv_table[id]; |
| 345 } |
| 346 /*printf("%d\n", unmask);*/ |
| 347 /* Normalize, compensate for the 1/4th of the sample and the factor of 6 i
n the inverse table */ |
| 348 unmask = 64*unmask*4/(6*(len2-17)); |
| 349 if (unmask>mask_metric) |
| 350 { |
| 351 *tf_chan = c; |
| 352 mask_metric = unmask; |
| 353 } |
| 354 } |
| 355 is_transient = mask_metric>200; |
| 356 |
| 357 /* Arbitrary metric for VBR boost */ |
| 358 tf_max = MAX16(0,celt_sqrt(27*mask_metric)-42); |
| 359 /* *tf_estimate = 1 + MIN16(1, sqrt(MAX16(0, tf_max-30))/20); */ |
| 360 *tf_estimate = celt_sqrt(MAX16(0, SHL32(MULT16_16(QCONST16(0.0069,14),MIN16(1
63,tf_max)),14)-QCONST32(0.139,28))); |
| 361 /*printf("%d %f\n", tf_max, mask_metric);*/ |
| 362 RESTORE_STACK; |
| 363 #ifdef FUZZING |
| 364 is_transient = rand()&0x1; |
| 365 #endif |
| 366 /*printf("%d %f %d\n", is_transient, (float)*tf_estimate, tf_max);*/ |
| 367 return is_transient; |
| 368 } |
| 369 |
| 370 /* Looks for sudden increases of energy to decide whether we need to patch |
| 371 the transient decision */ |
| 372 int patch_transient_decision(opus_val16 *newE, opus_val16 *oldE, int nbEBands, |
| 373 int end, int C) |
| 374 { |
| 375 int i, c; |
| 376 opus_val32 mean_diff=0; |
| 377 opus_val16 spread_old[26]; |
| 378 /* Apply an aggressive (-6 dB/Bark) spreading function to the old frame to |
| 379 avoid false detection caused by irrelevant bands */ |
| 380 if (C==1) |
| 381 { |
| 382 spread_old[0] = oldE[0]; |
| 383 for (i=1;i<end;i++) |
| 384 spread_old[i] = MAX16(spread_old[i-1]-QCONST16(1.0f, DB_SHIFT), oldE[i]
); |
| 385 } else { |
| 386 spread_old[0] = MAX16(oldE[0],oldE[nbEBands]); |
| 387 for (i=1;i<end;i++) |
| 388 spread_old[i] = MAX16(spread_old[i-1]-QCONST16(1.0f, DB_SHIFT), |
| 389 MAX16(oldE[i],oldE[i+nbEBands])); |
| 390 } |
| 391 for (i=end-2;i>=0;i--) |
| 392 spread_old[i] = MAX16(spread_old[i], spread_old[i+1]-QCONST16(1.0f, DB_SHI
FT)); |
| 393 /* Compute mean increase */ |
| 394 c=0; do { |
| 395 for (i=2;i<end-1;i++) |
| 396 { |
| 397 opus_val16 x1, x2; |
| 398 x1 = MAX16(0, newE[i]); |
| 399 x2 = MAX16(0, spread_old[i]); |
| 400 mean_diff = ADD32(mean_diff, EXTEND32(MAX16(0, SUB16(x1, x2)))); |
| 401 } |
| 402 } while (++c<C); |
| 403 mean_diff = DIV32(mean_diff, C*(end-3)); |
| 404 /*printf("%f %f %d\n", mean_diff, max_diff, count);*/ |
| 405 return mean_diff > QCONST16(1.f, DB_SHIFT); |
| 406 } |
| 407 |
| 408 /** Apply window and compute the MDCT for all sub-frames and |
| 409 all channels in a frame */ |
| 410 static void compute_mdcts(const CELTMode *mode, int shortBlocks, celt_sig * OPUS
_RESTRICT in, |
| 411 celt_sig * OPUS_RESTRICT out, int C, int CC, int LM, i
nt upsample) |
| 412 { |
| 413 const int overlap = OVERLAP(mode); |
| 414 int N; |
| 415 int B; |
| 416 int shift; |
| 417 int i, b, c; |
| 418 if (shortBlocks) |
| 419 { |
| 420 B = shortBlocks; |
| 421 N = mode->shortMdctSize; |
| 422 shift = mode->maxLM; |
| 423 } else { |
| 424 B = 1; |
| 425 N = mode->shortMdctSize<<LM; |
| 426 shift = mode->maxLM-LM; |
| 427 } |
| 428 c=0; do { |
| 429 for (b=0;b<B;b++) |
| 430 { |
| 431 /* Interleaving the sub-frames while doing the MDCTs */ |
| 432 clt_mdct_forward(&mode->mdct, in+c*(B*N+overlap)+b*N, &out[b+c*N*B], mo
de->window, overlap, shift, B); |
| 433 } |
| 434 } while (++c<CC); |
| 435 if (CC==2&&C==1) |
| 436 { |
| 437 for (i=0;i<B*N;i++) |
| 438 out[i] = ADD32(HALF32(out[i]), HALF32(out[B*N+i])); |
| 439 } |
| 440 if (upsample != 1) |
| 441 { |
| 442 c=0; do |
| 443 { |
| 444 int bound = B*N/upsample; |
| 445 for (i=0;i<bound;i++) |
| 446 out[c*B*N+i] *= upsample; |
| 447 for (;i<B*N;i++) |
| 448 out[c*B*N+i] = 0; |
| 449 } while (++c<C); |
| 450 } |
| 451 } |
| 452 |
| 453 |
| 454 void preemphasis(const opus_val16 * OPUS_RESTRICT pcmp, celt_sig * OPUS_RESTRICT
inp, |
| 455 int N, int CC, int upsample, const opus_val16 *coef, cel
t_sig *mem, int clip) |
| 456 { |
| 457 int i; |
| 458 opus_val16 coef0; |
| 459 celt_sig m; |
| 460 int Nu; |
| 461 |
| 462 coef0 = coef[0]; |
| 463 |
| 464 |
| 465 Nu = N/upsample; |
| 466 if (upsample!=1) |
| 467 { |
| 468 for (i=0;i<N;i++) |
| 469 inp[i] = 0; |
| 470 } |
| 471 for (i=0;i<Nu;i++) |
| 472 { |
| 473 celt_sig x; |
| 474 |
| 475 x = SCALEIN(pcmp[CC*i]); |
| 476 #ifndef FIXED_POINT |
| 477 /* Replace NaNs with zeros */ |
| 478 if (!(x==x)) |
| 479 x = 0; |
| 480 #endif |
| 481 inp[i*upsample] = x; |
| 482 } |
| 483 |
| 484 #ifndef FIXED_POINT |
| 485 if (clip) |
| 486 { |
| 487 /* Clip input to avoid encoding non-portable files */ |
| 488 for (i=0;i<Nu;i++) |
| 489 inp[i*upsample] = MAX32(-65536.f, MIN32(65536.f,inp[i*upsample])); |
| 490 } |
| 491 #endif |
| 492 m = *mem; |
| 493 #ifdef CUSTOM_MODES |
| 494 if (coef[1] != 0) |
| 495 { |
| 496 opus_val16 coef1 = coef[1]; |
| 497 opus_val16 coef2 = coef[2]; |
| 498 for (i=0;i<N;i++) |
| 499 { |
| 500 opus_val16 x, tmp; |
| 501 x = inp[i]; |
| 502 /* Apply pre-emphasis */ |
| 503 tmp = MULT16_16(coef2, x); |
| 504 inp[i] = tmp + m; |
| 505 m = MULT16_32_Q15(coef1, inp[i]) - MULT16_32_Q15(coef0, tmp); |
| 506 } |
| 507 } else |
| 508 #endif |
| 509 { |
| 510 for (i=0;i<N;i++) |
| 511 { |
| 512 celt_sig x; |
| 513 x = SHL32(inp[i], SIG_SHIFT); |
| 514 /* Apply pre-emphasis */ |
| 515 inp[i] = x + m; |
| 516 m = - MULT16_32_Q15(coef0, x); |
| 517 } |
| 518 } |
| 519 *mem = m; |
| 520 } |
| 521 |
| 522 |
| 523 |
| 524 static opus_val32 l1_metric(const celt_norm *tmp, int N, int LM, opus_val16 bias
) |
| 525 { |
| 526 int i; |
| 527 opus_val32 L1; |
| 528 L1 = 0; |
| 529 for (i=0;i<N;i++) |
| 530 L1 += EXTEND32(ABS16(tmp[i])); |
| 531 /* When in doubt, prefer good freq resolution */ |
| 532 L1 = MAC16_32_Q15(L1, LM*bias, L1); |
| 533 return L1; |
| 534 |
| 535 } |
| 536 |
| 537 static int tf_analysis(const CELTMode *m, int len, int isTransient, |
| 538 int *tf_res, int lambda, celt_norm *X, int N0, int LM, |
| 539 int *tf_sum, opus_val16 tf_estimate, int tf_chan) |
| 540 { |
| 541 int i; |
| 542 VARDECL(int, metric); |
| 543 int cost0; |
| 544 int cost1; |
| 545 VARDECL(int, path0); |
| 546 VARDECL(int, path1); |
| 547 VARDECL(celt_norm, tmp); |
| 548 VARDECL(celt_norm, tmp_1); |
| 549 int sel; |
| 550 int selcost[2]; |
| 551 int tf_select=0; |
| 552 opus_val16 bias; |
| 553 |
| 554 SAVE_STACK; |
| 555 bias = MULT16_16_Q14(QCONST16(.04f,15), MAX16(-QCONST16(.25f,14), QCONST16(.5
f,14)-tf_estimate)); |
| 556 /*printf("%f ", bias);*/ |
| 557 |
| 558 ALLOC(metric, len, int); |
| 559 ALLOC(tmp, (m->eBands[len]-m->eBands[len-1])<<LM, celt_norm); |
| 560 ALLOC(tmp_1, (m->eBands[len]-m->eBands[len-1])<<LM, celt_norm); |
| 561 ALLOC(path0, len, int); |
| 562 ALLOC(path1, len, int); |
| 563 |
| 564 *tf_sum = 0; |
| 565 for (i=0;i<len;i++) |
| 566 { |
| 567 int j, k, N; |
| 568 int narrow; |
| 569 opus_val32 L1, best_L1; |
| 570 int best_level=0; |
| 571 N = (m->eBands[i+1]-m->eBands[i])<<LM; |
| 572 /* band is too narrow to be split down to LM=-1 */ |
| 573 narrow = (m->eBands[i+1]-m->eBands[i])==1; |
| 574 for (j=0;j<N;j++) |
| 575 tmp[j] = X[tf_chan*N0 + j+(m->eBands[i]<<LM)]; |
| 576 /* Just add the right channel if we're in stereo */ |
| 577 /*if (C==2) |
| 578 for (j=0;j<N;j++) |
| 579 tmp[j] = ADD16(SHR16(tmp[j], 1),SHR16(X[N0+j+(m->eBands[i]<<LM)], 1)
);*/ |
| 580 L1 = l1_metric(tmp, N, isTransient ? LM : 0, bias); |
| 581 best_L1 = L1; |
| 582 /* Check the -1 case for transients */ |
| 583 if (isTransient && !narrow) |
| 584 { |
| 585 for (j=0;j<N;j++) |
| 586 tmp_1[j] = tmp[j]; |
| 587 haar1(tmp_1, N>>LM, 1<<LM); |
| 588 L1 = l1_metric(tmp_1, N, LM+1, bias); |
| 589 if (L1<best_L1) |
| 590 { |
| 591 best_L1 = L1; |
| 592 best_level = -1; |
| 593 } |
| 594 } |
| 595 /*printf ("%f ", L1);*/ |
| 596 for (k=0;k<LM+!(isTransient||narrow);k++) |
| 597 { |
| 598 int B; |
| 599 |
| 600 if (isTransient) |
| 601 B = (LM-k-1); |
| 602 else |
| 603 B = k+1; |
| 604 |
| 605 haar1(tmp, N>>k, 1<<k); |
| 606 |
| 607 L1 = l1_metric(tmp, N, B, bias); |
| 608 |
| 609 if (L1 < best_L1) |
| 610 { |
| 611 best_L1 = L1; |
| 612 best_level = k+1; |
| 613 } |
| 614 } |
| 615 /*printf ("%d ", isTransient ? LM-best_level : best_level);*/ |
| 616 /* metric is in Q1 to be able to select the mid-point (-0.5) for narrower
bands */ |
| 617 if (isTransient) |
| 618 metric[i] = 2*best_level; |
| 619 else |
| 620 metric[i] = -2*best_level; |
| 621 *tf_sum += (isTransient ? LM : 0) - metric[i]/2; |
| 622 /* For bands that can't be split to -1, set the metric to the half-way poi
nt to avoid |
| 623 biasing the decision */ |
| 624 if (narrow && (metric[i]==0 || metric[i]==-2*LM)) |
| 625 metric[i]-=1; |
| 626 /*printf("%d ", metric[i]);*/ |
| 627 } |
| 628 /*printf("\n");*/ |
| 629 /* Search for the optimal tf resolution, including tf_select */ |
| 630 tf_select = 0; |
| 631 for (sel=0;sel<2;sel++) |
| 632 { |
| 633 cost0 = 0; |
| 634 cost1 = isTransient ? 0 : lambda; |
| 635 for (i=1;i<len;i++) |
| 636 { |
| 637 int curr0, curr1; |
| 638 curr0 = IMIN(cost0, cost1 + lambda); |
| 639 curr1 = IMIN(cost0 + lambda, cost1); |
| 640 cost0 = curr0 + abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*sel
+0]); |
| 641 cost1 = curr1 + abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*sel
+1]); |
| 642 } |
| 643 cost0 = IMIN(cost0, cost1); |
| 644 selcost[sel]=cost0; |
| 645 } |
| 646 /* For now, we're conservative and only allow tf_select=1 for transients. |
| 647 * If tests confirm it's useful for non-transients, we could allow it. */ |
| 648 if (selcost[1]<selcost[0] && isTransient) |
| 649 tf_select=1; |
| 650 cost0 = 0; |
| 651 cost1 = isTransient ? 0 : lambda; |
| 652 /* Viterbi forward pass */ |
| 653 for (i=1;i<len;i++) |
| 654 { |
| 655 int curr0, curr1; |
| 656 int from0, from1; |
| 657 |
| 658 from0 = cost0; |
| 659 from1 = cost1 + lambda; |
| 660 if (from0 < from1) |
| 661 { |
| 662 curr0 = from0; |
| 663 path0[i]= 0; |
| 664 } else { |
| 665 curr0 = from1; |
| 666 path0[i]= 1; |
| 667 } |
| 668 |
| 669 from0 = cost0 + lambda; |
| 670 from1 = cost1; |
| 671 if (from0 < from1) |
| 672 { |
| 673 curr1 = from0; |
| 674 path1[i]= 0; |
| 675 } else { |
| 676 curr1 = from1; |
| 677 path1[i]= 1; |
| 678 } |
| 679 cost0 = curr0 + abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*tf_sel
ect+0]); |
| 680 cost1 = curr1 + abs(metric[i]-2*tf_select_table[LM][4*isTransient+2*tf_sel
ect+1]); |
| 681 } |
| 682 tf_res[len-1] = cost0 < cost1 ? 0 : 1; |
| 683 /* Viterbi backward pass to check the decisions */ |
| 684 for (i=len-2;i>=0;i--) |
| 685 { |
| 686 if (tf_res[i+1] == 1) |
| 687 tf_res[i] = path1[i+1]; |
| 688 else |
| 689 tf_res[i] = path0[i+1]; |
| 690 } |
| 691 /*printf("%d %f\n", *tf_sum, tf_estimate);*/ |
| 692 RESTORE_STACK; |
| 693 #ifdef FUZZING |
| 694 tf_select = rand()&0x1; |
| 695 tf_res[0] = rand()&0x1; |
| 696 for (i=1;i<len;i++) |
| 697 tf_res[i] = tf_res[i-1] ^ ((rand()&0xF) == 0); |
| 698 #endif |
| 699 return tf_select; |
| 700 } |
| 701 |
| 702 static void tf_encode(int start, int end, int isTransient, int *tf_res, int LM,
int tf_select, ec_enc *enc) |
| 703 { |
| 704 int curr, i; |
| 705 int tf_select_rsv; |
| 706 int tf_changed; |
| 707 int logp; |
| 708 opus_uint32 budget; |
| 709 opus_uint32 tell; |
| 710 budget = enc->storage*8; |
| 711 tell = ec_tell(enc); |
| 712 logp = isTransient ? 2 : 4; |
| 713 /* Reserve space to code the tf_select decision. */ |
| 714 tf_select_rsv = LM>0 && tell+logp+1 <= budget; |
| 715 budget -= tf_select_rsv; |
| 716 curr = tf_changed = 0; |
| 717 for (i=start;i<end;i++) |
| 718 { |
| 719 if (tell+logp<=budget) |
| 720 { |
| 721 ec_enc_bit_logp(enc, tf_res[i] ^ curr, logp); |
| 722 tell = ec_tell(enc); |
| 723 curr = tf_res[i]; |
| 724 tf_changed |= curr; |
| 725 } |
| 726 else |
| 727 tf_res[i] = curr; |
| 728 logp = isTransient ? 4 : 5; |
| 729 } |
| 730 /* Only code tf_select if it would actually make a difference. */ |
| 731 if (tf_select_rsv && |
| 732 tf_select_table[LM][4*isTransient+0+tf_changed]!= |
| 733 tf_select_table[LM][4*isTransient+2+tf_changed]) |
| 734 ec_enc_bit_logp(enc, tf_select, 1); |
| 735 else |
| 736 tf_select = 0; |
| 737 for (i=start;i<end;i++) |
| 738 tf_res[i] = tf_select_table[LM][4*isTransient+2*tf_select+tf_res[i]]; |
| 739 /*for(i=0;i<end;i++)printf("%d ", isTransient ? tf_res[i] : LM+tf_res[i]);pri
ntf("\n");*/ |
| 740 } |
| 741 |
| 742 |
| 743 static int alloc_trim_analysis(const CELTMode *m, const celt_norm *X, |
| 744 const opus_val16 *bandLogE, int end, int LM, int C, int N0, |
| 745 AnalysisInfo *analysis, opus_val16 *stereo_saving, opus_val16 tf_estimate, |
| 746 int intensity, opus_val16 surround_trim) |
| 747 { |
| 748 int i; |
| 749 opus_val32 diff=0; |
| 750 int c; |
| 751 int trim_index = 5; |
| 752 opus_val16 trim = QCONST16(5.f, 8); |
| 753 opus_val16 logXC, logXC2; |
| 754 if (C==2) |
| 755 { |
| 756 opus_val16 sum = 0; /* Q10 */ |
| 757 opus_val16 minXC; /* Q10 */ |
| 758 /* Compute inter-channel correlation for low frequencies */ |
| 759 for (i=0;i<8;i++) |
| 760 { |
| 761 int j; |
| 762 opus_val32 partial = 0; |
| 763 for (j=m->eBands[i]<<LM;j<m->eBands[i+1]<<LM;j++) |
| 764 partial = MAC16_16(partial, X[j], X[N0+j]); |
| 765 sum = ADD16(sum, EXTRACT16(SHR32(partial, 18))); |
| 766 } |
| 767 sum = MULT16_16_Q15(QCONST16(1.f/8, 15), sum); |
| 768 sum = MIN16(QCONST16(1.f, 10), ABS16(sum)); |
| 769 minXC = sum; |
| 770 for (i=8;i<intensity;i++) |
| 771 { |
| 772 int j; |
| 773 opus_val32 partial = 0; |
| 774 for (j=m->eBands[i]<<LM;j<m->eBands[i+1]<<LM;j++) |
| 775 partial = MAC16_16(partial, X[j], X[N0+j]); |
| 776 minXC = MIN16(minXC, ABS16(EXTRACT16(SHR32(partial, 18)))); |
| 777 } |
| 778 minXC = MIN16(QCONST16(1.f, 10), ABS16(minXC)); |
| 779 /*printf ("%f\n", sum);*/ |
| 780 if (sum > QCONST16(.995f,10)) |
| 781 trim_index-=4; |
| 782 else if (sum > QCONST16(.92f,10)) |
| 783 trim_index-=3; |
| 784 else if (sum > QCONST16(.85f,10)) |
| 785 trim_index-=2; |
| 786 else if (sum > QCONST16(.8f,10)) |
| 787 trim_index-=1; |
| 788 /* mid-side savings estimations based on the LF average*/ |
| 789 logXC = celt_log2(QCONST32(1.001f, 20)-MULT16_16(sum, sum)); |
| 790 /* mid-side savings estimations based on min correlation */ |
| 791 logXC2 = MAX16(HALF16(logXC), celt_log2(QCONST32(1.001f, 20)-MULT16_16(min
XC, minXC))); |
| 792 #ifdef FIXED_POINT |
| 793 /* Compensate for Q20 vs Q14 input and convert output to Q8 */ |
| 794 logXC = PSHR32(logXC-QCONST16(6.f, DB_SHIFT),DB_SHIFT-8); |
| 795 logXC2 = PSHR32(logXC2-QCONST16(6.f, DB_SHIFT),DB_SHIFT-8); |
| 796 #endif |
| 797 |
| 798 trim += MAX16(-QCONST16(4.f, 8), MULT16_16_Q15(QCONST16(.75f,15),logXC)); |
| 799 *stereo_saving = MIN16(*stereo_saving + QCONST16(0.25f, 8), -HALF16(logXC2
)); |
| 800 } |
| 801 |
| 802 /* Estimate spectral tilt */ |
| 803 c=0; do { |
| 804 for (i=0;i<end-1;i++) |
| 805 { |
| 806 diff += bandLogE[i+c*m->nbEBands]*(opus_int32)(2+2*i-end); |
| 807 } |
| 808 } while (++c<C); |
| 809 diff /= C*(end-1); |
| 810 /*printf("%f\n", diff);*/ |
| 811 if (diff > QCONST16(2.f, DB_SHIFT)) |
| 812 trim_index--; |
| 813 if (diff > QCONST16(8.f, DB_SHIFT)) |
| 814 trim_index--; |
| 815 if (diff < -QCONST16(4.f, DB_SHIFT)) |
| 816 trim_index++; |
| 817 if (diff < -QCONST16(10.f, DB_SHIFT)) |
| 818 trim_index++; |
| 819 trim -= MAX16(-QCONST16(2.f, 8), MIN16(QCONST16(2.f, 8), SHR16(diff+QCONST16(
1.f, DB_SHIFT),DB_SHIFT-8)/6 )); |
| 820 trim -= SHR16(surround_trim, DB_SHIFT-8); |
| 821 trim -= 2*SHR16(tf_estimate, 14-8); |
| 822 #ifndef DISABLE_FLOAT_API |
| 823 if (analysis->valid) |
| 824 { |
| 825 trim -= MAX16(-QCONST16(2.f, 8), MIN16(QCONST16(2.f, 8), QCONST16(2.f, 8)*
(analysis->tonality_slope+.05f))); |
| 826 } |
| 827 #endif |
| 828 |
| 829 #ifdef FIXED_POINT |
| 830 trim_index = PSHR32(trim, 8); |
| 831 #else |
| 832 trim_index = (int)floor(.5f+trim); |
| 833 #endif |
| 834 if (trim_index<0) |
| 835 trim_index = 0; |
| 836 if (trim_index>10) |
| 837 trim_index = 10; |
| 838 /*printf("%d\n", trim_index);*/ |
| 839 #ifdef FUZZING |
| 840 trim_index = rand()%11; |
| 841 #endif |
| 842 return trim_index; |
| 843 } |
| 844 |
| 845 static int stereo_analysis(const CELTMode *m, const celt_norm *X, |
| 846 int LM, int N0) |
| 847 { |
| 848 int i; |
| 849 int thetas; |
| 850 opus_val32 sumLR = EPSILON, sumMS = EPSILON; |
| 851 |
| 852 /* Use the L1 norm to model the entropy of the L/R signal vs the M/S signal *
/ |
| 853 for (i=0;i<13;i++) |
| 854 { |
| 855 int j; |
| 856 for (j=m->eBands[i]<<LM;j<m->eBands[i+1]<<LM;j++) |
| 857 { |
| 858 opus_val32 L, R, M, S; |
| 859 /* We cast to 32-bit first because of the -32768 case */ |
| 860 L = EXTEND32(X[j]); |
| 861 R = EXTEND32(X[N0+j]); |
| 862 M = ADD32(L, R); |
| 863 S = SUB32(L, R); |
| 864 sumLR = ADD32(sumLR, ADD32(ABS32(L), ABS32(R))); |
| 865 sumMS = ADD32(sumMS, ADD32(ABS32(M), ABS32(S))); |
| 866 } |
| 867 } |
| 868 sumMS = MULT16_32_Q15(QCONST16(0.707107f, 15), sumMS); |
| 869 thetas = 13; |
| 870 /* We don't need thetas for lower bands with LM<=1 */ |
| 871 if (LM<=1) |
| 872 thetas -= 8; |
| 873 return MULT16_32_Q15((m->eBands[13]<<(LM+1))+thetas, sumMS) |
| 874 > MULT16_32_Q15(m->eBands[13]<<(LM+1), sumLR); |
| 875 } |
| 876 |
| 877 static opus_val16 dynalloc_analysis(const opus_val16 *bandLogE, const opus_val16
*bandLogE2, |
| 878 int nbEBands, int start, int end, int C, int *offsets, int lsb_depth, cons
t opus_int16 *logN, |
| 879 int isTransient, int vbr, int constrained_vbr, const opus_int16 *eBands, i
nt LM, |
| 880 int effectiveBytes, opus_int32 *tot_boost_, int lfe, opus_val16 *surround_
dynalloc) |
| 881 { |
| 882 int i, c; |
| 883 opus_int32 tot_boost=0; |
| 884 opus_val16 maxDepth; |
| 885 VARDECL(opus_val16, follower); |
| 886 VARDECL(opus_val16, noise_floor); |
| 887 SAVE_STACK; |
| 888 ALLOC(follower, C*nbEBands, opus_val16); |
| 889 ALLOC(noise_floor, C*nbEBands, opus_val16); |
| 890 for (i=0;i<nbEBands;i++) |
| 891 offsets[i] = 0; |
| 892 /* Dynamic allocation code */ |
| 893 maxDepth=-QCONST16(31.9f, DB_SHIFT); |
| 894 for (i=0;i<end;i++) |
| 895 { |
| 896 /* Noise floor must take into account eMeans, the depth, the width of the
bands |
| 897 and the preemphasis filter (approx. square of bark band ID) */ |
| 898 noise_floor[i] = MULT16_16(QCONST16(0.0625f, DB_SHIFT),logN[i]) |
| 899 +QCONST16(.5f,DB_SHIFT)+SHL16(9-lsb_depth,DB_SHIFT)-SHL16(eMeans[i],
6) |
| 900 +MULT16_16(QCONST16(.0062,DB_SHIFT),(i+5)*(i+5)); |
| 901 } |
| 902 c=0;do |
| 903 { |
| 904 for (i=0;i<end;i++) |
| 905 maxDepth = MAX16(maxDepth, bandLogE[c*nbEBands+i]-noise_floor[i]); |
| 906 } while (++c<C); |
| 907 /* Make sure that dynamic allocation can't make us bust the budget */ |
| 908 if (effectiveBytes > 50 && LM>=1 && !lfe) |
| 909 { |
| 910 int last=0; |
| 911 c=0;do |
| 912 { |
| 913 follower[c*nbEBands] = bandLogE2[c*nbEBands]; |
| 914 for (i=1;i<end;i++) |
| 915 { |
| 916 /* The last band to be at least 3 dB higher than the previous one |
| 917 is the last we'll consider. Otherwise, we run into problems on |
| 918 bandlimited signals. */ |
| 919 if (bandLogE2[c*nbEBands+i] > bandLogE2[c*nbEBands+i-1]+QCONST16(.5f
,DB_SHIFT)) |
| 920 last=i; |
| 921 follower[c*nbEBands+i] = MIN16(follower[c*nbEBands+i-1]+QCONST16(1.5
f,DB_SHIFT), bandLogE2[c*nbEBands+i]); |
| 922 } |
| 923 for (i=last-1;i>=0;i--) |
| 924 follower[c*nbEBands+i] = MIN16(follower[c*nbEBands+i], MIN16(followe
r[c*nbEBands+i+1]+QCONST16(2.f,DB_SHIFT), bandLogE2[c*nbEBands+i])); |
| 925 for (i=0;i<end;i++) |
| 926 follower[c*nbEBands+i] = MAX16(follower[c*nbEBands+i], noise_floor[i
]); |
| 927 } while (++c<C); |
| 928 if (C==2) |
| 929 { |
| 930 for (i=start;i<end;i++) |
| 931 { |
| 932 /* Consider 24 dB "cross-talk" */ |
| 933 follower[nbEBands+i] = MAX16(follower[nbEBands+i], follower[
i]-QCONST16(4.f,DB_SHIFT)); |
| 934 follower[ i] = MAX16(follower[ i], follower[nbEBands
+i]-QCONST16(4.f,DB_SHIFT)); |
| 935 follower[i] = HALF16(MAX16(0, bandLogE[i]-follower[i]) + MAX16(0, ba
ndLogE[nbEBands+i]-follower[nbEBands+i])); |
| 936 } |
| 937 } else { |
| 938 for (i=start;i<end;i++) |
| 939 { |
| 940 follower[i] = MAX16(0, bandLogE[i]-follower[i]); |
| 941 } |
| 942 } |
| 943 for (i=start;i<end;i++) |
| 944 follower[i] = MAX16(follower[i], surround_dynalloc[i]); |
| 945 /* For non-transient CBR/CVBR frames, halve the dynalloc contribution */ |
| 946 if ((!vbr || constrained_vbr)&&!isTransient) |
| 947 { |
| 948 for (i=start;i<end;i++) |
| 949 follower[i] = HALF16(follower[i]); |
| 950 } |
| 951 for (i=start;i<end;i++) |
| 952 { |
| 953 int width; |
| 954 int boost; |
| 955 int boost_bits; |
| 956 |
| 957 if (i<8) |
| 958 follower[i] *= 2; |
| 959 if (i>=12) |
| 960 follower[i] = HALF16(follower[i]); |
| 961 follower[i] = MIN16(follower[i], QCONST16(4, DB_SHIFT)); |
| 962 |
| 963 width = C*(eBands[i+1]-eBands[i])<<LM; |
| 964 if (width<6) |
| 965 { |
| 966 boost = (int)SHR32(EXTEND32(follower[i]),DB_SHIFT); |
| 967 boost_bits = boost*width<<BITRES; |
| 968 } else if (width > 48) { |
| 969 boost = (int)SHR32(EXTEND32(follower[i])*8,DB_SHIFT); |
| 970 boost_bits = (boost*width<<BITRES)/8; |
| 971 } else { |
| 972 boost = (int)SHR32(EXTEND32(follower[i])*width/6,DB_SHIFT); |
| 973 boost_bits = boost*6<<BITRES; |
| 974 } |
| 975 /* For CBR and non-transient CVBR frames, limit dynalloc to 1/4 of the
bits */ |
| 976 if ((!vbr || (constrained_vbr&&!isTransient)) |
| 977 && (tot_boost+boost_bits)>>BITRES>>3 > effectiveBytes/4) |
| 978 { |
| 979 opus_int32 cap = ((effectiveBytes/4)<<BITRES<<3); |
| 980 offsets[i] = cap-tot_boost; |
| 981 tot_boost = cap; |
| 982 break; |
| 983 } else { |
| 984 offsets[i] = boost; |
| 985 tot_boost += boost_bits; |
| 986 } |
| 987 } |
| 988 } |
| 989 *tot_boost_ = tot_boost; |
| 990 RESTORE_STACK; |
| 991 return maxDepth; |
| 992 } |
| 993 |
| 994 |
| 995 static int run_prefilter(CELTEncoder *st, celt_sig *in, celt_sig *prefilter_mem,
int CC, int N, |
| 996 int prefilter_tapset, int *pitch, opus_val16 *gain, int *qgain, int enable
d, int nbAvailableBytes) |
| 997 { |
| 998 int c; |
| 999 VARDECL(celt_sig, _pre); |
| 1000 celt_sig *pre[2]; |
| 1001 const CELTMode *mode; |
| 1002 int pitch_index; |
| 1003 opus_val16 gain1; |
| 1004 opus_val16 pf_threshold; |
| 1005 int pf_on; |
| 1006 int qg; |
| 1007 SAVE_STACK; |
| 1008 |
| 1009 mode = st->mode; |
| 1010 ALLOC(_pre, CC*(N+COMBFILTER_MAXPERIOD), celt_sig); |
| 1011 |
| 1012 pre[0] = _pre; |
| 1013 pre[1] = _pre + (N+COMBFILTER_MAXPERIOD); |
| 1014 |
| 1015 |
| 1016 c=0; do { |
| 1017 OPUS_COPY(pre[c], prefilter_mem+c*COMBFILTER_MAXPERIOD, COMBFILTER_MAXPERI
OD); |
| 1018 OPUS_COPY(pre[c]+COMBFILTER_MAXPERIOD, in+c*(N+st->overlap)+st->overlap, N
); |
| 1019 } while (++c<CC); |
| 1020 |
| 1021 if (enabled) |
| 1022 { |
| 1023 VARDECL(opus_val16, pitch_buf); |
| 1024 ALLOC(pitch_buf, (COMBFILTER_MAXPERIOD+N)>>1, opus_val16); |
| 1025 |
| 1026 pitch_downsample(pre, pitch_buf, COMBFILTER_MAXPERIOD+N, CC); |
| 1027 /* Don't search for the fir last 1.5 octave of the range because |
| 1028 there's too many false-positives due to short-term correlation */ |
| 1029 pitch_search(pitch_buf+(COMBFILTER_MAXPERIOD>>1), pitch_buf, N, |
| 1030 COMBFILTER_MAXPERIOD-3*COMBFILTER_MINPERIOD, &pitch_index); |
| 1031 pitch_index = COMBFILTER_MAXPERIOD-pitch_index; |
| 1032 |
| 1033 gain1 = remove_doubling(pitch_buf, COMBFILTER_MAXPERIOD, COMBFILTER_MINPER
IOD, |
| 1034 N, &pitch_index, st->prefilter_period, st->prefilter_gain); |
| 1035 if (pitch_index > COMBFILTER_MAXPERIOD-2) |
| 1036 pitch_index = COMBFILTER_MAXPERIOD-2; |
| 1037 gain1 = MULT16_16_Q15(QCONST16(.7f,15),gain1); |
| 1038 /*printf("%d %d %f %f\n", pitch_change, pitch_index, gain1, st->analysis.t
onality);*/ |
| 1039 if (st->loss_rate>2) |
| 1040 gain1 = HALF32(gain1); |
| 1041 if (st->loss_rate>4) |
| 1042 gain1 = HALF32(gain1); |
| 1043 if (st->loss_rate>8) |
| 1044 gain1 = 0; |
| 1045 } else { |
| 1046 gain1 = 0; |
| 1047 pitch_index = COMBFILTER_MINPERIOD; |
| 1048 } |
| 1049 |
| 1050 /* Gain threshold for enabling the prefilter/postfilter */ |
| 1051 pf_threshold = QCONST16(.2f,15); |
| 1052 |
| 1053 /* Adjusting the threshold based on rate and continuity */ |
| 1054 if (abs(pitch_index-st->prefilter_period)*10>pitch_index) |
| 1055 pf_threshold += QCONST16(.2f,15); |
| 1056 if (nbAvailableBytes<25) |
| 1057 pf_threshold += QCONST16(.1f,15); |
| 1058 if (nbAvailableBytes<35) |
| 1059 pf_threshold += QCONST16(.1f,15); |
| 1060 if (st->prefilter_gain > QCONST16(.4f,15)) |
| 1061 pf_threshold -= QCONST16(.1f,15); |
| 1062 if (st->prefilter_gain > QCONST16(.55f,15)) |
| 1063 pf_threshold -= QCONST16(.1f,15); |
| 1064 |
| 1065 /* Hard threshold at 0.2 */ |
| 1066 pf_threshold = MAX16(pf_threshold, QCONST16(.2f,15)); |
| 1067 if (gain1<pf_threshold) |
| 1068 { |
| 1069 gain1 = 0; |
| 1070 pf_on = 0; |
| 1071 qg = 0; |
| 1072 } else { |
| 1073 /*This block is not gated by a total bits check only because |
| 1074 of the nbAvailableBytes check above.*/ |
| 1075 if (ABS16(gain1-st->prefilter_gain)<QCONST16(.1f,15)) |
| 1076 gain1=st->prefilter_gain; |
| 1077 |
| 1078 #ifdef FIXED_POINT |
| 1079 qg = ((gain1+1536)>>10)/3-1; |
| 1080 #else |
| 1081 qg = (int)floor(.5f+gain1*32/3)-1; |
| 1082 #endif |
| 1083 qg = IMAX(0, IMIN(7, qg)); |
| 1084 gain1 = QCONST16(0.09375f,15)*(qg+1); |
| 1085 pf_on = 1; |
| 1086 } |
| 1087 /*printf("%d %f\n", pitch_index, gain1);*/ |
| 1088 |
| 1089 c=0; do { |
| 1090 int offset = mode->shortMdctSize-st->overlap; |
| 1091 st->prefilter_period=IMAX(st->prefilter_period, COMBFILTER_MINPERIOD); |
| 1092 OPUS_COPY(in+c*(N+st->overlap), st->in_mem+c*(st->overlap), st->overlap); |
| 1093 if (offset) |
| 1094 comb_filter(in+c*(N+st->overlap)+st->overlap, pre[c]+COMBFILTER_MAXPERI
OD, |
| 1095 st->prefilter_period, st->prefilter_period, offset, -st->prefilte
r_gain, -st->prefilter_gain, |
| 1096 st->prefilter_tapset, st->prefilter_tapset, NULL, 0); |
| 1097 |
| 1098 comb_filter(in+c*(N+st->overlap)+st->overlap+offset, pre[c]+COMBFILTER_MAX
PERIOD+offset, |
| 1099 st->prefilter_period, pitch_index, N-offset, -st->prefilter_gain, -g
ain1, |
| 1100 st->prefilter_tapset, prefilter_tapset, mode->window, st->overlap); |
| 1101 OPUS_COPY(st->in_mem+c*(st->overlap), in+c*(N+st->overlap)+N, st->overlap)
; |
| 1102 |
| 1103 if (N>COMBFILTER_MAXPERIOD) |
| 1104 { |
| 1105 OPUS_MOVE(prefilter_mem+c*COMBFILTER_MAXPERIOD, pre[c]+N, COMBFILTER_MA
XPERIOD); |
| 1106 } else { |
| 1107 OPUS_MOVE(prefilter_mem+c*COMBFILTER_MAXPERIOD, prefilter_mem+c*COMBFIL
TER_MAXPERIOD+N, COMBFILTER_MAXPERIOD-N); |
| 1108 OPUS_MOVE(prefilter_mem+c*COMBFILTER_MAXPERIOD+COMBFILTER_MAXPERIOD-N,
pre[c]+COMBFILTER_MAXPERIOD, N); |
| 1109 } |
| 1110 } while (++c<CC); |
| 1111 |
| 1112 RESTORE_STACK; |
| 1113 *gain = gain1; |
| 1114 *pitch = pitch_index; |
| 1115 *qgain = qg; |
| 1116 return pf_on; |
| 1117 } |
| 1118 |
| 1119 static int compute_vbr(const CELTMode *mode, AnalysisInfo *analysis, opus_int32
base_target, |
| 1120 int LM, opus_int32 bitrate, int lastCodedBands, int C, int intensity, |
| 1121 int constrained_vbr, opus_val16 stereo_saving, int tot_boost, |
| 1122 opus_val16 tf_estimate, int pitch_change, opus_val16 maxDepth, |
| 1123 int variable_duration, int lfe, int has_surround_mask, opus_val16 surround
_masking, |
| 1124 opus_val16 temporal_vbr) |
| 1125 { |
| 1126 /* The target rate in 8th bits per frame */ |
| 1127 opus_int32 target; |
| 1128 int coded_bins; |
| 1129 int coded_bands; |
| 1130 opus_val16 tf_calibration; |
| 1131 int nbEBands; |
| 1132 const opus_int16 *eBands; |
| 1133 |
| 1134 nbEBands = mode->nbEBands; |
| 1135 eBands = mode->eBands; |
| 1136 |
| 1137 coded_bands = lastCodedBands ? lastCodedBands : nbEBands; |
| 1138 coded_bins = eBands[coded_bands]<<LM; |
| 1139 if (C==2) |
| 1140 coded_bins += eBands[IMIN(intensity, coded_bands)]<<LM; |
| 1141 |
| 1142 target = base_target; |
| 1143 |
| 1144 /*printf("%f %f %f %f %d %d ", st->analysis.activity, st->analysis.tonality,
tf_estimate, st->stereo_saving, tot_boost, coded_bands);*/ |
| 1145 #ifndef DISABLE_FLOAT_API |
| 1146 if (analysis->valid && analysis->activity<.4) |
| 1147 target -= (opus_int32)((coded_bins<<BITRES)*(.4f-analysis->activity)); |
| 1148 #endif |
| 1149 /* Stereo savings */ |
| 1150 if (C==2) |
| 1151 { |
| 1152 int coded_stereo_bands; |
| 1153 int coded_stereo_dof; |
| 1154 opus_val16 max_frac; |
| 1155 coded_stereo_bands = IMIN(intensity, coded_bands); |
| 1156 coded_stereo_dof = (eBands[coded_stereo_bands]<<LM)-coded_stereo_bands; |
| 1157 /* Maximum fraction of the bits we can save if the signal is mono. */ |
| 1158 max_frac = DIV32_16(MULT16_16(QCONST16(0.8f, 15), coded_stereo_dof), coded
_bins); |
| 1159 /*printf("%d %d %d ", coded_stereo_dof, coded_bins, tot_boost);*/ |
| 1160 target -= (opus_int32)MIN32(MULT16_32_Q15(max_frac,target), |
| 1161 SHR32(MULT16_16(stereo_saving-QCONST16(0.1f,8),(coded_ster
eo_dof<<BITRES)),8)); |
| 1162 } |
| 1163 /* Boost the rate according to dynalloc (minus the dynalloc average for calib
ration). */ |
| 1164 target += tot_boost-(16<<LM); |
| 1165 /* Apply transient boost, compensating for average boost. */ |
| 1166 tf_calibration = variable_duration==OPUS_FRAMESIZE_VARIABLE ? |
| 1167 QCONST16(0.02f,14) : QCONST16(0.04f,14); |
| 1168 target += (opus_int32)SHL32(MULT16_32_Q15(tf_estimate-tf_calibration, target)
,1); |
| 1169 |
| 1170 #ifndef DISABLE_FLOAT_API |
| 1171 /* Apply tonality boost */ |
| 1172 if (analysis->valid && !lfe) |
| 1173 { |
| 1174 opus_int32 tonal_target; |
| 1175 float tonal; |
| 1176 |
| 1177 /* Tonality boost (compensating for the average). */ |
| 1178 tonal = MAX16(0.f,analysis->tonality-.15f)-0.09f; |
| 1179 tonal_target = target + (opus_int32)((coded_bins<<BITRES)*1.2f*tonal); |
| 1180 if (pitch_change) |
| 1181 tonal_target += (opus_int32)((coded_bins<<BITRES)*.8f); |
| 1182 /*printf("%f %f ", analysis->tonality, tonal);*/ |
| 1183 target = tonal_target; |
| 1184 } |
| 1185 #endif |
| 1186 |
| 1187 if (has_surround_mask&&!lfe) |
| 1188 { |
| 1189 opus_int32 surround_target = target + (opus_int32)SHR32(MULT16_16(surround
_masking,coded_bins<<BITRES), DB_SHIFT); |
| 1190 /*printf("%f %d %d %d %d %d %d ", surround_masking, coded_bins, st->end, s
t->intensity, surround_target, target, st->bitrate);*/ |
| 1191 target = IMAX(target/4, surround_target); |
| 1192 } |
| 1193 |
| 1194 { |
| 1195 opus_int32 floor_depth; |
| 1196 int bins; |
| 1197 bins = eBands[nbEBands-2]<<LM; |
| 1198 /*floor_depth = SHR32(MULT16_16((C*bins<<BITRES),celt_log2(SHL32(MAX16(1,s
ample_max),13))), DB_SHIFT);*/ |
| 1199 floor_depth = (opus_int32)SHR32(MULT16_16((C*bins<<BITRES),maxDepth), DB_S
HIFT); |
| 1200 floor_depth = IMAX(floor_depth, target>>2); |
| 1201 target = IMIN(target, floor_depth); |
| 1202 /*printf("%f %d\n", maxDepth, floor_depth);*/ |
| 1203 } |
| 1204 |
| 1205 if ((!has_surround_mask||lfe) && (constrained_vbr || bitrate<64000)) |
| 1206 { |
| 1207 opus_val16 rate_factor; |
| 1208 #ifdef FIXED_POINT |
| 1209 rate_factor = MAX16(0,(bitrate-32000)); |
| 1210 #else |
| 1211 rate_factor = MAX16(0,(1.f/32768)*(bitrate-32000)); |
| 1212 #endif |
| 1213 if (constrained_vbr) |
| 1214 rate_factor = MIN16(rate_factor, QCONST16(0.67f, 15)); |
| 1215 target = base_target + (opus_int32)MULT16_32_Q15(rate_factor, target-base_
target); |
| 1216 |
| 1217 } |
| 1218 |
| 1219 if (!has_surround_mask && tf_estimate < QCONST16(.2f, 14)) |
| 1220 { |
| 1221 opus_val16 amount; |
| 1222 opus_val16 tvbr_factor; |
| 1223 amount = MULT16_16_Q15(QCONST16(.0000031f, 30), IMAX(0, IMIN(32000, 96000-
bitrate))); |
| 1224 tvbr_factor = SHR32(MULT16_16(temporal_vbr, amount), DB_SHIFT); |
| 1225 target += (opus_int32)MULT16_32_Q15(tvbr_factor, target); |
| 1226 } |
| 1227 |
| 1228 /* Don't allow more than doubling the rate */ |
| 1229 target = IMIN(2*base_target, target); |
| 1230 |
| 1231 return target; |
| 1232 } |
| 1233 |
| 1234 int celt_encode_with_ec(CELTEncoder * OPUS_RESTRICT st, const opus_val16 * pcm,
int frame_size, unsigned char *compressed, int nbCompressedBytes, ec_enc *enc) |
| 1235 { |
| 1236 int i, c, N; |
| 1237 opus_int32 bits; |
| 1238 ec_enc _enc; |
| 1239 VARDECL(celt_sig, in); |
| 1240 VARDECL(celt_sig, freq); |
| 1241 VARDECL(celt_norm, X); |
| 1242 VARDECL(celt_ener, bandE); |
| 1243 VARDECL(opus_val16, bandLogE); |
| 1244 VARDECL(opus_val16, bandLogE2); |
| 1245 VARDECL(int, fine_quant); |
| 1246 VARDECL(opus_val16, error); |
| 1247 VARDECL(int, pulses); |
| 1248 VARDECL(int, cap); |
| 1249 VARDECL(int, offsets); |
| 1250 VARDECL(int, fine_priority); |
| 1251 VARDECL(int, tf_res); |
| 1252 VARDECL(unsigned char, collapse_masks); |
| 1253 celt_sig *prefilter_mem; |
| 1254 opus_val16 *oldBandE, *oldLogE, *oldLogE2; |
| 1255 int shortBlocks=0; |
| 1256 int isTransient=0; |
| 1257 const int CC = st->channels; |
| 1258 const int C = st->stream_channels; |
| 1259 int LM, M; |
| 1260 int tf_select; |
| 1261 int nbFilledBytes, nbAvailableBytes; |
| 1262 int effEnd; |
| 1263 int codedBands; |
| 1264 int tf_sum; |
| 1265 int alloc_trim; |
| 1266 int pitch_index=COMBFILTER_MINPERIOD; |
| 1267 opus_val16 gain1 = 0; |
| 1268 int dual_stereo=0; |
| 1269 int effectiveBytes; |
| 1270 int dynalloc_logp; |
| 1271 opus_int32 vbr_rate; |
| 1272 opus_int32 total_bits; |
| 1273 opus_int32 total_boost; |
| 1274 opus_int32 balance; |
| 1275 opus_int32 tell; |
| 1276 int prefilter_tapset=0; |
| 1277 int pf_on; |
| 1278 int anti_collapse_rsv; |
| 1279 int anti_collapse_on=0; |
| 1280 int silence=0; |
| 1281 int tf_chan = 0; |
| 1282 opus_val16 tf_estimate; |
| 1283 int pitch_change=0; |
| 1284 opus_int32 tot_boost; |
| 1285 opus_val32 sample_max; |
| 1286 opus_val16 maxDepth; |
| 1287 const OpusCustomMode *mode; |
| 1288 int nbEBands; |
| 1289 int overlap; |
| 1290 const opus_int16 *eBands; |
| 1291 int secondMdct; |
| 1292 int signalBandwidth; |
| 1293 int transient_got_disabled=0; |
| 1294 opus_val16 surround_masking=0; |
| 1295 opus_val16 temporal_vbr=0; |
| 1296 opus_val16 surround_trim = 0; |
| 1297 VARDECL(opus_val16, surround_dynalloc); |
| 1298 ALLOC_STACK; |
| 1299 |
| 1300 mode = st->mode; |
| 1301 nbEBands = mode->nbEBands; |
| 1302 overlap = mode->overlap; |
| 1303 eBands = mode->eBands; |
| 1304 tf_estimate = 0; |
| 1305 if (nbCompressedBytes<2 || pcm==NULL) |
| 1306 return OPUS_BAD_ARG; |
| 1307 |
| 1308 frame_size *= st->upsample; |
| 1309 for (LM=0;LM<=mode->maxLM;LM++) |
| 1310 if (mode->shortMdctSize<<LM==frame_size) |
| 1311 break; |
| 1312 if (LM>mode->maxLM) |
| 1313 return OPUS_BAD_ARG; |
| 1314 M=1<<LM; |
| 1315 N = M*mode->shortMdctSize; |
| 1316 |
| 1317 prefilter_mem = st->in_mem+CC*(st->overlap); |
| 1318 oldBandE = (opus_val16*)(st->in_mem+CC*(st->overlap+COMBFILTER_MAXPERIOD)); |
| 1319 oldLogE = oldBandE + CC*nbEBands; |
| 1320 oldLogE2 = oldLogE + CC*nbEBands; |
| 1321 |
| 1322 if (enc==NULL) |
| 1323 { |
| 1324 tell=1; |
| 1325 nbFilledBytes=0; |
| 1326 } else { |
| 1327 tell=ec_tell(enc); |
| 1328 nbFilledBytes=(tell+4)>>3; |
| 1329 } |
| 1330 |
| 1331 #ifdef CUSTOM_MODES |
| 1332 if (st->signalling && enc==NULL) |
| 1333 { |
| 1334 int tmp = (mode->effEBands-st->end)>>1; |
| 1335 st->end = IMAX(1, mode->effEBands-tmp); |
| 1336 compressed[0] = tmp<<5; |
| 1337 compressed[0] |= LM<<3; |
| 1338 compressed[0] |= (C==2)<<2; |
| 1339 /* Convert "standard mode" to Opus header */ |
| 1340 if (mode->Fs==48000 && mode->shortMdctSize==120) |
| 1341 { |
| 1342 int c0 = toOpus(compressed[0]); |
| 1343 if (c0<0) |
| 1344 return OPUS_BAD_ARG; |
| 1345 compressed[0] = c0; |
| 1346 } |
| 1347 compressed++; |
| 1348 nbCompressedBytes--; |
| 1349 } |
| 1350 #else |
| 1351 celt_assert(st->signalling==0); |
| 1352 #endif |
| 1353 |
| 1354 /* Can't produce more than 1275 output bytes */ |
| 1355 nbCompressedBytes = IMIN(nbCompressedBytes,1275); |
| 1356 nbAvailableBytes = nbCompressedBytes - nbFilledBytes; |
| 1357 |
| 1358 if (st->vbr && st->bitrate!=OPUS_BITRATE_MAX) |
| 1359 { |
| 1360 opus_int32 den=mode->Fs>>BITRES; |
| 1361 vbr_rate=(st->bitrate*frame_size+(den>>1))/den; |
| 1362 #ifdef CUSTOM_MODES |
| 1363 if (st->signalling) |
| 1364 vbr_rate -= 8<<BITRES; |
| 1365 #endif |
| 1366 effectiveBytes = vbr_rate>>(3+BITRES); |
| 1367 } else { |
| 1368 opus_int32 tmp; |
| 1369 vbr_rate = 0; |
| 1370 tmp = st->bitrate*frame_size; |
| 1371 if (tell>1) |
| 1372 tmp += tell; |
| 1373 if (st->bitrate!=OPUS_BITRATE_MAX) |
| 1374 nbCompressedBytes = IMAX(2, IMIN(nbCompressedBytes, |
| 1375 (tmp+4*mode->Fs)/(8*mode->Fs)-!!st->signalling)); |
| 1376 effectiveBytes = nbCompressedBytes; |
| 1377 } |
| 1378 |
| 1379 if (enc==NULL) |
| 1380 { |
| 1381 ec_enc_init(&_enc, compressed, nbCompressedBytes); |
| 1382 enc = &_enc; |
| 1383 } |
| 1384 |
| 1385 if (vbr_rate>0) |
| 1386 { |
| 1387 /* Computes the max bit-rate allowed in VBR mode to avoid violating the |
| 1388 target rate and buffering. |
| 1389 We must do this up front so that bust-prevention logic triggers |
| 1390 correctly if we don't have enough bits. */ |
| 1391 if (st->constrained_vbr) |
| 1392 { |
| 1393 opus_int32 vbr_bound; |
| 1394 opus_int32 max_allowed; |
| 1395 /* We could use any multiple of vbr_rate as bound (depending on the |
| 1396 delay). |
| 1397 This is clamped to ensure we use at least two bytes if the encoder |
| 1398 was entirely empty, but to allow 0 in hybrid mode. */ |
| 1399 vbr_bound = vbr_rate; |
| 1400 max_allowed = IMIN(IMAX(tell==1?2:0, |
| 1401 (vbr_rate+vbr_bound-st->vbr_reservoir)>>(BITRES+3)), |
| 1402 nbAvailableBytes); |
| 1403 if(max_allowed < nbAvailableBytes) |
| 1404 { |
| 1405 nbCompressedBytes = nbFilledBytes+max_allowed; |
| 1406 nbAvailableBytes = max_allowed; |
| 1407 ec_enc_shrink(enc, nbCompressedBytes); |
| 1408 } |
| 1409 } |
| 1410 } |
| 1411 total_bits = nbCompressedBytes*8; |
| 1412 |
| 1413 effEnd = st->end; |
| 1414 if (effEnd > mode->effEBands) |
| 1415 effEnd = mode->effEBands; |
| 1416 |
| 1417 ALLOC(in, CC*(N+st->overlap), celt_sig); |
| 1418 |
| 1419 sample_max=MAX32(st->overlap_max, celt_maxabs16(pcm, C*(N-overlap)/st->upsamp
le)); |
| 1420 st->overlap_max=celt_maxabs16(pcm+C*(N-overlap)/st->upsample, C*overlap/st->u
psample); |
| 1421 sample_max=MAX32(sample_max, st->overlap_max); |
| 1422 #ifdef FIXED_POINT |
| 1423 silence = (sample_max==0); |
| 1424 #else |
| 1425 silence = (sample_max <= (opus_val16)1/(1<<st->lsb_depth)); |
| 1426 #endif |
| 1427 #ifdef FUZZING |
| 1428 if ((rand()&0x3F)==0) |
| 1429 silence = 1; |
| 1430 #endif |
| 1431 if (tell==1) |
| 1432 ec_enc_bit_logp(enc, silence, 15); |
| 1433 else |
| 1434 silence=0; |
| 1435 if (silence) |
| 1436 { |
| 1437 /*In VBR mode there is no need to send more than the minimum. */ |
| 1438 if (vbr_rate>0) |
| 1439 { |
| 1440 effectiveBytes=nbCompressedBytes=IMIN(nbCompressedBytes, nbFilledBytes+
2); |
| 1441 total_bits=nbCompressedBytes*8; |
| 1442 nbAvailableBytes=2; |
| 1443 ec_enc_shrink(enc, nbCompressedBytes); |
| 1444 } |
| 1445 /* Pretend we've filled all the remaining bits with zeros |
| 1446 (that's what the initialiser did anyway) */ |
| 1447 tell = nbCompressedBytes*8; |
| 1448 enc->nbits_total+=tell-ec_tell(enc); |
| 1449 } |
| 1450 c=0; do { |
| 1451 preemphasis(pcm+c, in+c*(N+st->overlap)+st->overlap, N, CC, st->upsample, |
| 1452 mode->preemph, st->preemph_memE+c, st->clip); |
| 1453 } while (++c<CC); |
| 1454 |
| 1455 |
| 1456 |
| 1457 /* Find pitch period and gain */ |
| 1458 { |
| 1459 int enabled; |
| 1460 int qg; |
| 1461 enabled = (st->lfe || nbAvailableBytes>12*C) && st->start==0 && !silence &
& !st->disable_pf |
| 1462 && st->complexity >= 5 && !(st->consec_transient && LM!=3 && st->var
iable_duration==OPUS_FRAMESIZE_VARIABLE); |
| 1463 |
| 1464 prefilter_tapset = st->tapset_decision; |
| 1465 pf_on = run_prefilter(st, in, prefilter_mem, CC, N, prefilter_tapset, &pit
ch_index, &gain1, &qg, enabled, nbAvailableBytes); |
| 1466 if ((gain1 > QCONST16(.4f,15) || st->prefilter_gain > QCONST16(.4f,15)) &&
(!st->analysis.valid || st->analysis.tonality > .3) |
| 1467 && (pitch_index > 1.26*st->prefilter_period || pitch_index < .79*st-
>prefilter_period)) |
| 1468 pitch_change = 1; |
| 1469 if (pf_on==0) |
| 1470 { |
| 1471 if(st->start==0 && tell+16<=total_bits) |
| 1472 ec_enc_bit_logp(enc, 0, 1); |
| 1473 } else { |
| 1474 /*This block is not gated by a total bits check only because |
| 1475 of the nbAvailableBytes check above.*/ |
| 1476 int octave; |
| 1477 ec_enc_bit_logp(enc, 1, 1); |
| 1478 pitch_index += 1; |
| 1479 octave = EC_ILOG(pitch_index)-5; |
| 1480 ec_enc_uint(enc, octave, 6); |
| 1481 ec_enc_bits(enc, pitch_index-(16<<octave), 4+octave); |
| 1482 pitch_index -= 1; |
| 1483 ec_enc_bits(enc, qg, 3); |
| 1484 ec_enc_icdf(enc, prefilter_tapset, tapset_icdf, 2); |
| 1485 } |
| 1486 } |
| 1487 |
| 1488 isTransient = 0; |
| 1489 shortBlocks = 0; |
| 1490 if (st->complexity >= 1 && !st->lfe) |
| 1491 { |
| 1492 isTransient = transient_analysis(in, N+st->overlap, CC, |
| 1493 &tf_estimate, &tf_chan); |
| 1494 } |
| 1495 if (LM>0 && ec_tell(enc)+3<=total_bits) |
| 1496 { |
| 1497 if (isTransient) |
| 1498 shortBlocks = M; |
| 1499 } else { |
| 1500 isTransient = 0; |
| 1501 transient_got_disabled=1; |
| 1502 } |
| 1503 |
| 1504 ALLOC(freq, CC*N, celt_sig); /**< Interleaved signal MDCTs */ |
| 1505 ALLOC(bandE,nbEBands*CC, celt_ener); |
| 1506 ALLOC(bandLogE,nbEBands*CC, opus_val16); |
| 1507 |
| 1508 secondMdct = shortBlocks && st->complexity>=8; |
| 1509 ALLOC(bandLogE2, C*nbEBands, opus_val16); |
| 1510 if (secondMdct) |
| 1511 { |
| 1512 compute_mdcts(mode, 0, in, freq, C, CC, LM, st->upsample); |
| 1513 compute_band_energies(mode, freq, bandE, effEnd, C, M); |
| 1514 amp2Log2(mode, effEnd, st->end, bandE, bandLogE2, C); |
| 1515 for (i=0;i<C*nbEBands;i++) |
| 1516 bandLogE2[i] += HALF16(SHL16(LM, DB_SHIFT)); |
| 1517 } |
| 1518 |
| 1519 compute_mdcts(mode, shortBlocks, in, freq, C, CC, LM, st->upsample); |
| 1520 if (CC==2&&C==1) |
| 1521 tf_chan = 0; |
| 1522 compute_band_energies(mode, freq, bandE, effEnd, C, M); |
| 1523 |
| 1524 if (st->lfe) |
| 1525 { |
| 1526 for (i=2;i<st->end;i++) |
| 1527 { |
| 1528 bandE[i] = IMIN(bandE[i], MULT16_32_Q15(QCONST16(1e-4f,15),bandE[0])); |
| 1529 bandE[i] = MAX32(bandE[i], EPSILON); |
| 1530 } |
| 1531 } |
| 1532 amp2Log2(mode, effEnd, st->end, bandE, bandLogE, C); |
| 1533 |
| 1534 ALLOC(surround_dynalloc, C*nbEBands, opus_val16); |
| 1535 for(i=0;i<st->end;i++) |
| 1536 surround_dynalloc[i] = 0; |
| 1537 /* This computes how much masking takes place between surround channels */ |
| 1538 if (st->start==0&&st->energy_mask&&!st->lfe) |
| 1539 { |
| 1540 int mask_end; |
| 1541 int midband; |
| 1542 int count_dynalloc; |
| 1543 opus_val32 mask_avg=0; |
| 1544 opus_val32 diff=0; |
| 1545 int count=0; |
| 1546 mask_end = IMAX(2,st->lastCodedBands); |
| 1547 for (c=0;c<C;c++) |
| 1548 { |
| 1549 for(i=0;i<mask_end;i++) |
| 1550 { |
| 1551 opus_val16 mask; |
| 1552 mask = MAX16(MIN16(st->energy_mask[nbEBands*c+i], |
| 1553 QCONST16(.25f, DB_SHIFT)), -QCONST16(2.0f, DB_SHIFT)); |
| 1554 if (mask > 0) |
| 1555 mask = HALF16(mask); |
| 1556 mask_avg += MULT16_16(mask, eBands[i+1]-eBands[i]); |
| 1557 count += eBands[i+1]-eBands[i]; |
| 1558 diff += MULT16_16(mask, 1+2*i-mask_end); |
| 1559 } |
| 1560 } |
| 1561 mask_avg = DIV32_16(mask_avg,count); |
| 1562 mask_avg += QCONST16(.2f, DB_SHIFT); |
| 1563 diff = diff*6/(C*(mask_end-1)*(mask_end+1)*mask_end); |
| 1564 /* Again, being conservative */ |
| 1565 diff = HALF32(diff); |
| 1566 diff = MAX32(MIN32(diff, QCONST32(.031f, DB_SHIFT)), -QCONST32(.031f, DB_S
HIFT)); |
| 1567 /* Find the band that's in the middle of the coded spectrum */ |
| 1568 for (midband=0;eBands[midband+1] < eBands[mask_end]/2;midband++); |
| 1569 count_dynalloc=0; |
| 1570 for(i=0;i<mask_end;i++) |
| 1571 { |
| 1572 opus_val32 lin; |
| 1573 opus_val16 unmask; |
| 1574 lin = mask_avg + diff*(i-midband); |
| 1575 if (C==2) |
| 1576 unmask = MAX16(st->energy_mask[i], st->energy_mask[nbEBands+i]); |
| 1577 else |
| 1578 unmask = st->energy_mask[i]; |
| 1579 unmask = MIN16(unmask, QCONST16(.0f, DB_SHIFT)); |
| 1580 unmask -= lin; |
| 1581 if (unmask > QCONST16(.25f, DB_SHIFT)) |
| 1582 { |
| 1583 surround_dynalloc[i] = unmask - QCONST16(.25f, DB_SHIFT); |
| 1584 count_dynalloc++; |
| 1585 } |
| 1586 } |
| 1587 if (count_dynalloc>=3) |
| 1588 { |
| 1589 /* If we need dynalloc in many bands, it's probably because our |
| 1590 initial masking rate was too low. */ |
| 1591 mask_avg += QCONST16(.25f, DB_SHIFT); |
| 1592 if (mask_avg>0) |
| 1593 { |
| 1594 /* Something went really wrong in the original calculations, |
| 1595 disabling masking. */ |
| 1596 mask_avg = 0; |
| 1597 diff = 0; |
| 1598 for(i=0;i<mask_end;i++) |
| 1599 surround_dynalloc[i] = 0; |
| 1600 } else { |
| 1601 for(i=0;i<mask_end;i++) |
| 1602 surround_dynalloc[i] = MAX16(0, surround_dynalloc[i]-QCONST16(.25
f, DB_SHIFT)); |
| 1603 } |
| 1604 } |
| 1605 mask_avg += QCONST16(.2f, DB_SHIFT); |
| 1606 /* Convert to 1/64th units used for the trim */ |
| 1607 surround_trim = 64*diff; |
| 1608 /*printf("%d %d ", mask_avg, surround_trim);*/ |
| 1609 surround_masking = mask_avg; |
| 1610 } |
| 1611 /* Temporal VBR (but not for LFE) */ |
| 1612 if (!st->lfe) |
| 1613 { |
| 1614 opus_val16 follow=-QCONST16(10.0f,DB_SHIFT); |
| 1615 float frame_avg=0; |
| 1616 opus_val16 offset = shortBlocks?HALF16(SHL16(LM, DB_SHIFT)):0; |
| 1617 for(i=st->start;i<st->end;i++) |
| 1618 { |
| 1619 follow = MAX16(follow-QCONST16(1.f, DB_SHIFT), bandLogE[i]-offset); |
| 1620 if (C==2) |
| 1621 follow = MAX16(follow, bandLogE[i+nbEBands]-offset); |
| 1622 frame_avg += follow; |
| 1623 } |
| 1624 frame_avg /= (st->end-st->start); |
| 1625 temporal_vbr = SUB16(frame_avg,st->spec_avg); |
| 1626 temporal_vbr = MIN16(QCONST16(3.f, DB_SHIFT), MAX16(-QCONST16(1.5f, DB_SHI
FT), temporal_vbr)); |
| 1627 st->spec_avg += MULT16_16_Q15(QCONST16(.02f, 15), temporal_vbr); |
| 1628 } |
| 1629 /*for (i=0;i<21;i++) |
| 1630 printf("%f ", bandLogE[i]); |
| 1631 printf("\n");*/ |
| 1632 |
| 1633 if (!secondMdct) |
| 1634 { |
| 1635 for (i=0;i<C*nbEBands;i++) |
| 1636 bandLogE2[i] = bandLogE[i]; |
| 1637 } |
| 1638 |
| 1639 /* Last chance to catch any transient we might have missed in the |
| 1640 time-domain analysis */ |
| 1641 if (LM>0 && ec_tell(enc)+3<=total_bits && !isTransient && st->complexity>=5 &
& !st->lfe) |
| 1642 { |
| 1643 if (patch_transient_decision(bandLogE, oldBandE, nbEBands, st->end, C)) |
| 1644 { |
| 1645 isTransient = 1; |
| 1646 shortBlocks = M; |
| 1647 compute_mdcts(mode, shortBlocks, in, freq, C, CC, LM, st->upsample); |
| 1648 compute_band_energies(mode, freq, bandE, effEnd, C, M); |
| 1649 amp2Log2(mode, effEnd, st->end, bandE, bandLogE, C); |
| 1650 /* Compensate for the scaling of short vs long mdcts */ |
| 1651 for (i=0;i<C*nbEBands;i++) |
| 1652 bandLogE2[i] += HALF16(SHL16(LM, DB_SHIFT)); |
| 1653 tf_estimate = QCONST16(.2f,14); |
| 1654 } |
| 1655 } |
| 1656 |
| 1657 if (LM>0 && ec_tell(enc)+3<=total_bits) |
| 1658 ec_enc_bit_logp(enc, isTransient, 3); |
| 1659 |
| 1660 ALLOC(X, C*N, celt_norm); /**< Interleaved normalised MDCTs */ |
| 1661 |
| 1662 /* Band normalisation */ |
| 1663 normalise_bands(mode, freq, X, bandE, effEnd, C, M); |
| 1664 |
| 1665 ALLOC(tf_res, nbEBands, int); |
| 1666 /* Disable variable tf resolution for hybrid and at very low bitrate */ |
| 1667 if (effectiveBytes>=15*C && st->start==0 && st->complexity>=2 && !st->lfe) |
| 1668 { |
| 1669 int lambda; |
| 1670 if (effectiveBytes<40) |
| 1671 lambda = 12; |
| 1672 else if (effectiveBytes<60) |
| 1673 lambda = 6; |
| 1674 else if (effectiveBytes<100) |
| 1675 lambda = 4; |
| 1676 else |
| 1677 lambda = 3; |
| 1678 lambda*=2; |
| 1679 tf_select = tf_analysis(mode, effEnd, isTransient, tf_res, lambda, X, N, L
M, &tf_sum, tf_estimate, tf_chan); |
| 1680 for (i=effEnd;i<st->end;i++) |
| 1681 tf_res[i] = tf_res[effEnd-1]; |
| 1682 } else { |
| 1683 tf_sum = 0; |
| 1684 for (i=0;i<st->end;i++) |
| 1685 tf_res[i] = isTransient; |
| 1686 tf_select=0; |
| 1687 } |
| 1688 |
| 1689 ALLOC(error, C*nbEBands, opus_val16); |
| 1690 quant_coarse_energy(mode, st->start, st->end, effEnd, bandLogE, |
| 1691 oldBandE, total_bits, error, enc, |
| 1692 C, LM, nbAvailableBytes, st->force_intra, |
| 1693 &st->delayedIntra, st->complexity >= 4, st->loss_rate, st->lfe); |
| 1694 |
| 1695 tf_encode(st->start, st->end, isTransient, tf_res, LM, tf_select, enc); |
| 1696 |
| 1697 if (ec_tell(enc)+4<=total_bits) |
| 1698 { |
| 1699 if (st->lfe) |
| 1700 { |
| 1701 st->tapset_decision = 0; |
| 1702 st->spread_decision = SPREAD_NORMAL; |
| 1703 } else if (shortBlocks || st->complexity < 3 || nbAvailableBytes < 10*C ||
st->start != 0) |
| 1704 { |
| 1705 if (st->complexity == 0) |
| 1706 st->spread_decision = SPREAD_NONE; |
| 1707 else |
| 1708 st->spread_decision = SPREAD_NORMAL; |
| 1709 } else { |
| 1710 /* Disable new spreading+tapset estimator until we can show it works |
| 1711 better than the old one. So far it seems like spreading_decision() |
| 1712 works best. */ |
| 1713 if (0&&st->analysis.valid) |
| 1714 { |
| 1715 static const opus_val16 spread_thresholds[3] = {-QCONST16(.6f, 15),
-QCONST16(.2f, 15), -QCONST16(.07f, 15)}; |
| 1716 static const opus_val16 spread_histeresis[3] = {QCONST16(.15f, 15),
QCONST16(.07f, 15), QCONST16(.02f, 15)}; |
| 1717 static const opus_val16 tapset_thresholds[2] = {QCONST16(.0f, 15), Q
CONST16(.15f, 15)}; |
| 1718 static const opus_val16 tapset_histeresis[2] = {QCONST16(.1f, 15), Q
CONST16(.05f, 15)}; |
| 1719 st->spread_decision = hysteresis_decision(-st->analysis.tonality, sp
read_thresholds, spread_histeresis, 3, st->spread_decision); |
| 1720 st->tapset_decision = hysteresis_decision(st->analysis.tonality_slop
e, tapset_thresholds, tapset_histeresis, 2, st->tapset_decision); |
| 1721 } else { |
| 1722 st->spread_decision = spreading_decision(mode, X, |
| 1723 &st->tonal_average, st->spread_decision, &st->hf_average, |
| 1724 &st->tapset_decision, pf_on&&!shortBlocks, effEnd, C, M); |
| 1725 } |
| 1726 /*printf("%d %d\n", st->tapset_decision, st->spread_decision);*/ |
| 1727 /*printf("%f %d %f %d\n\n", st->analysis.tonality, st->spread_decision,
st->analysis.tonality_slope, st->tapset_decision);*/ |
| 1728 } |
| 1729 ec_enc_icdf(enc, st->spread_decision, spread_icdf, 5); |
| 1730 } |
| 1731 |
| 1732 ALLOC(offsets, nbEBands, int); |
| 1733 |
| 1734 maxDepth = dynalloc_analysis(bandLogE, bandLogE2, nbEBands, st->start, st->en
d, C, offsets, |
| 1735 st->lsb_depth, mode->logN, isTransient, st->vbr, st->constrained_vbr, |
| 1736 eBands, LM, effectiveBytes, &tot_boost, st->lfe, surround_dynalloc); |
| 1737 /* For LFE, everything interesting is in the first band */ |
| 1738 if (st->lfe) |
| 1739 offsets[0] = IMIN(8, effectiveBytes/3); |
| 1740 ALLOC(cap, nbEBands, int); |
| 1741 init_caps(mode,cap,LM,C); |
| 1742 |
| 1743 dynalloc_logp = 6; |
| 1744 total_bits<<=BITRES; |
| 1745 total_boost = 0; |
| 1746 tell = ec_tell_frac(enc); |
| 1747 for (i=st->start;i<st->end;i++) |
| 1748 { |
| 1749 int width, quanta; |
| 1750 int dynalloc_loop_logp; |
| 1751 int boost; |
| 1752 int j; |
| 1753 width = C*(eBands[i+1]-eBands[i])<<LM; |
| 1754 /* quanta is 6 bits, but no more than 1 bit/sample |
| 1755 and no less than 1/8 bit/sample */ |
| 1756 quanta = IMIN(width<<BITRES, IMAX(6<<BITRES, width)); |
| 1757 dynalloc_loop_logp = dynalloc_logp; |
| 1758 boost = 0; |
| 1759 for (j = 0; tell+(dynalloc_loop_logp<<BITRES) < total_bits-total_boost |
| 1760 && boost < cap[i]; j++) |
| 1761 { |
| 1762 int flag; |
| 1763 flag = j<offsets[i]; |
| 1764 ec_enc_bit_logp(enc, flag, dynalloc_loop_logp); |
| 1765 tell = ec_tell_frac(enc); |
| 1766 if (!flag) |
| 1767 break; |
| 1768 boost += quanta; |
| 1769 total_boost += quanta; |
| 1770 dynalloc_loop_logp = 1; |
| 1771 } |
| 1772 /* Making dynalloc more likely */ |
| 1773 if (j) |
| 1774 dynalloc_logp = IMAX(2, dynalloc_logp-1); |
| 1775 offsets[i] = boost; |
| 1776 } |
| 1777 |
| 1778 if (C==2) |
| 1779 { |
| 1780 int effectiveRate; |
| 1781 |
| 1782 static const opus_val16 intensity_thresholds[21]= |
| 1783 /* 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 off*/ |
| 1784 { 16,21,23,25,27,29,31,33,35,38,42,46,50,54,58,63,68,75,84,102,130}; |
| 1785 static const opus_val16 intensity_histeresis[21]= |
| 1786 { 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 3, 4, 5, 6, 8, 12}; |
| 1787 |
| 1788 /* Always use MS for 2.5 ms frames until we can do a better analysis */ |
| 1789 if (LM!=0) |
| 1790 dual_stereo = stereo_analysis(mode, X, LM, N); |
| 1791 |
| 1792 /* Account for coarse energy */ |
| 1793 effectiveRate = (8*effectiveBytes - 80)>>LM; |
| 1794 |
| 1795 /* effectiveRate in kb/s */ |
| 1796 effectiveRate = 2*effectiveRate/5; |
| 1797 |
| 1798 st->intensity = hysteresis_decision((opus_val16)effectiveRate, intensity_t
hresholds, intensity_histeresis, 21, st->intensity); |
| 1799 st->intensity = IMIN(st->end,IMAX(st->start, st->intensity)); |
| 1800 } |
| 1801 |
| 1802 alloc_trim = 5; |
| 1803 if (tell+(6<<BITRES) <= total_bits - total_boost) |
| 1804 { |
| 1805 if (st->lfe) |
| 1806 alloc_trim = 5; |
| 1807 else |
| 1808 alloc_trim = alloc_trim_analysis(mode, X, bandLogE, |
| 1809 st->end, LM, C, N, &st->analysis, &st->stereo_saving, tf_estimate, s
t->intensity, surround_trim); |
| 1810 ec_enc_icdf(enc, alloc_trim, trim_icdf, 7); |
| 1811 tell = ec_tell_frac(enc); |
| 1812 } |
| 1813 |
| 1814 /* Variable bitrate */ |
| 1815 if (vbr_rate>0) |
| 1816 { |
| 1817 opus_val16 alpha; |
| 1818 opus_int32 delta; |
| 1819 /* The target rate in 8th bits per frame */ |
| 1820 opus_int32 target, base_target; |
| 1821 opus_int32 min_allowed; |
| 1822 int lm_diff = mode->maxLM - LM; |
| 1823 |
| 1824 /* Don't attempt to use more than 510 kb/s, even for frames smaller than 20
ms. |
| 1825 The CELT allocator will just not be able to use more than that anyway. *
/ |
| 1826 nbCompressedBytes = IMIN(nbCompressedBytes,1275>>(3-LM)); |
| 1827 base_target = vbr_rate - ((40*C+20)<<BITRES); |
| 1828 |
| 1829 if (st->constrained_vbr) |
| 1830 base_target += (st->vbr_offset>>lm_diff); |
| 1831 |
| 1832 target = compute_vbr(mode, &st->analysis, base_target, LM, st->bitrate, |
| 1833 st->lastCodedBands, C, st->intensity, st->constrained_vbr, |
| 1834 st->stereo_saving, tot_boost, tf_estimate, pitch_change, maxDepth, |
| 1835 st->variable_duration, st->lfe, st->energy_mask!=NULL, surround_maski
ng, |
| 1836 temporal_vbr); |
| 1837 |
| 1838 /* The current offset is removed from the target and the space used |
| 1839 so far is added*/ |
| 1840 target=target+tell; |
| 1841 /* In VBR mode the frame size must not be reduced so much that it would |
| 1842 result in the encoder running out of bits. |
| 1843 The margin of 2 bytes ensures that none of the bust-prevention logic |
| 1844 in the decoder will have triggered so far. */ |
| 1845 min_allowed = ((tell+total_boost+(1<<(BITRES+3))-1)>>(BITRES+3)) + 2 - nbFi
lledBytes; |
| 1846 |
| 1847 nbAvailableBytes = (target+(1<<(BITRES+2)))>>(BITRES+3); |
| 1848 nbAvailableBytes = IMAX(min_allowed,nbAvailableBytes); |
| 1849 nbAvailableBytes = IMIN(nbCompressedBytes,nbAvailableBytes+nbFilledBytes) -
nbFilledBytes; |
| 1850 |
| 1851 /* By how much did we "miss" the target on that frame */ |
| 1852 delta = target - vbr_rate; |
| 1853 |
| 1854 target=nbAvailableBytes<<(BITRES+3); |
| 1855 |
| 1856 /*If the frame is silent we don't adjust our drift, otherwise |
| 1857 the encoder will shoot to very high rates after hitting a |
| 1858 span of silence, but we do allow the bitres to refill. |
| 1859 This means that we'll undershoot our target in CVBR/VBR modes |
| 1860 on files with lots of silence. */ |
| 1861 if(silence) |
| 1862 { |
| 1863 nbAvailableBytes = 2; |
| 1864 target = 2*8<<BITRES; |
| 1865 delta = 0; |
| 1866 } |
| 1867 |
| 1868 if (st->vbr_count < 970) |
| 1869 { |
| 1870 st->vbr_count++; |
| 1871 alpha = celt_rcp(SHL32(EXTEND32(st->vbr_count+20),16)); |
| 1872 } else |
| 1873 alpha = QCONST16(.001f,15); |
| 1874 /* How many bits have we used in excess of what we're allowed */ |
| 1875 if (st->constrained_vbr) |
| 1876 st->vbr_reservoir += target - vbr_rate; |
| 1877 /*printf ("%d\n", st->vbr_reservoir);*/ |
| 1878 |
| 1879 /* Compute the offset we need to apply in order to reach the target */ |
| 1880 if (st->constrained_vbr) |
| 1881 { |
| 1882 st->vbr_drift += (opus_int32)MULT16_32_Q15(alpha,(delta*(1<<lm_diff))-st
->vbr_offset-st->vbr_drift); |
| 1883 st->vbr_offset = -st->vbr_drift; |
| 1884 } |
| 1885 /*printf ("%d\n", st->vbr_drift);*/ |
| 1886 |
| 1887 if (st->constrained_vbr && st->vbr_reservoir < 0) |
| 1888 { |
| 1889 /* We're under the min value -- increase rate */ |
| 1890 int adjust = (-st->vbr_reservoir)/(8<<BITRES); |
| 1891 /* Unless we're just coding silence */ |
| 1892 nbAvailableBytes += silence?0:adjust; |
| 1893 st->vbr_reservoir = 0; |
| 1894 /*printf ("+%d\n", adjust);*/ |
| 1895 } |
| 1896 nbCompressedBytes = IMIN(nbCompressedBytes,nbAvailableBytes+nbFilledBytes); |
| 1897 /*printf("%d\n", nbCompressedBytes*50*8);*/ |
| 1898 /* This moves the raw bits to take into account the new compressed size */ |
| 1899 ec_enc_shrink(enc, nbCompressedBytes); |
| 1900 } |
| 1901 |
| 1902 /* Bit allocation */ |
| 1903 ALLOC(fine_quant, nbEBands, int); |
| 1904 ALLOC(pulses, nbEBands, int); |
| 1905 ALLOC(fine_priority, nbEBands, int); |
| 1906 |
| 1907 /* bits = packet size - where we are - safety*/ |
| 1908 bits = (((opus_int32)nbCompressedBytes*8)<<BITRES) - ec_tell_frac(enc) - 1; |
| 1909 anti_collapse_rsv = isTransient&&LM>=2&&bits>=((LM+2)<<BITRES) ? (1<<BITRES)
: 0; |
| 1910 bits -= anti_collapse_rsv; |
| 1911 signalBandwidth = st->end-1; |
| 1912 #ifndef DISABLE_FLOAT_API |
| 1913 if (st->analysis.valid) |
| 1914 { |
| 1915 int min_bandwidth; |
| 1916 if (st->bitrate < (opus_int32)32000*C) |
| 1917 min_bandwidth = 13; |
| 1918 else if (st->bitrate < (opus_int32)48000*C) |
| 1919 min_bandwidth = 16; |
| 1920 else if (st->bitrate < (opus_int32)60000*C) |
| 1921 min_bandwidth = 18; |
| 1922 else if (st->bitrate < (opus_int32)80000*C) |
| 1923 min_bandwidth = 19; |
| 1924 else |
| 1925 min_bandwidth = 20; |
| 1926 signalBandwidth = IMAX(st->analysis.bandwidth, min_bandwidth); |
| 1927 } |
| 1928 #endif |
| 1929 if (st->lfe) |
| 1930 signalBandwidth = 1; |
| 1931 codedBands = compute_allocation(mode, st->start, st->end, offsets, cap, |
| 1932 alloc_trim, &st->intensity, &dual_stereo, bits, &balance, pulses, |
| 1933 fine_quant, fine_priority, C, LM, enc, 1, st->lastCodedBands, signalBan
dwidth); |
| 1934 if (st->lastCodedBands) |
| 1935 st->lastCodedBands = IMIN(st->lastCodedBands+1,IMAX(st->lastCodedBands-1,c
odedBands)); |
| 1936 else |
| 1937 st->lastCodedBands = codedBands; |
| 1938 |
| 1939 quant_fine_energy(mode, st->start, st->end, oldBandE, error, fine_quant, enc,
C); |
| 1940 |
| 1941 /* Residual quantisation */ |
| 1942 ALLOC(collapse_masks, C*nbEBands, unsigned char); |
| 1943 quant_all_bands(1, mode, st->start, st->end, X, C==2 ? X+N : NULL, collapse_m
asks, |
| 1944 bandE, pulses, shortBlocks, st->spread_decision, dual_stereo, st->inten
sity, tf_res, |
| 1945 nbCompressedBytes*(8<<BITRES)-anti_collapse_rsv, balance, enc, LM, code
dBands, &st->rng); |
| 1946 |
| 1947 if (anti_collapse_rsv > 0) |
| 1948 { |
| 1949 anti_collapse_on = st->consec_transient<2; |
| 1950 #ifdef FUZZING |
| 1951 anti_collapse_on = rand()&0x1; |
| 1952 #endif |
| 1953 ec_enc_bits(enc, anti_collapse_on, 1); |
| 1954 } |
| 1955 quant_energy_finalise(mode, st->start, st->end, oldBandE, error, fine_quant,
fine_priority, nbCompressedBytes*8-ec_tell(enc), enc, C); |
| 1956 |
| 1957 if (silence) |
| 1958 { |
| 1959 for (i=0;i<C*nbEBands;i++) |
| 1960 oldBandE[i] = -QCONST16(28.f,DB_SHIFT); |
| 1961 } |
| 1962 |
| 1963 #ifdef RESYNTH |
| 1964 /* Re-synthesis of the coded audio if required */ |
| 1965 { |
| 1966 celt_sig *out_mem[2]; |
| 1967 |
| 1968 if (anti_collapse_on) |
| 1969 { |
| 1970 anti_collapse(mode, X, collapse_masks, LM, C, N, |
| 1971 st->start, st->end, oldBandE, oldLogE, oldLogE2, pulses, st->rng)
; |
| 1972 } |
| 1973 |
| 1974 if (silence) |
| 1975 { |
| 1976 for (i=0;i<C*N;i++) |
| 1977 freq[i] = 0; |
| 1978 } else { |
| 1979 /* Synthesis */ |
| 1980 denormalise_bands(mode, X, freq, oldBandE, st->start, effEnd, C, M); |
| 1981 } |
| 1982 |
| 1983 c=0; do { |
| 1984 OPUS_MOVE(st->syn_mem[c], st->syn_mem[c]+N, 2*MAX_PERIOD-N+overlap/2); |
| 1985 } while (++c<CC); |
| 1986 |
| 1987 if (CC==2&&C==1) |
| 1988 { |
| 1989 for (i=0;i<N;i++) |
| 1990 freq[N+i] = freq[i]; |
| 1991 } |
| 1992 |
| 1993 c=0; do { |
| 1994 out_mem[c] = st->syn_mem[c]+2*MAX_PERIOD-N; |
| 1995 } while (++c<CC); |
| 1996 |
| 1997 compute_inv_mdcts(mode, shortBlocks, freq, out_mem, CC, LM); |
| 1998 |
| 1999 c=0; do { |
| 2000 st->prefilter_period=IMAX(st->prefilter_period, COMBFILTER_MINPERIOD); |
| 2001 st->prefilter_period_old=IMAX(st->prefilter_period_old, COMBFILTER_MINP
ERIOD); |
| 2002 comb_filter(out_mem[c], out_mem[c], st->prefilter_period_old, st->prefi
lter_period, mode->shortMdctSize, |
| 2003 st->prefilter_gain_old, st->prefilter_gain, st->prefilter_tapset_
old, st->prefilter_tapset, |
| 2004 mode->window, st->overlap); |
| 2005 if (LM!=0) |
| 2006 comb_filter(out_mem[c]+mode->shortMdctSize, out_mem[c]+mode->shortMd
ctSize, st->prefilter_period, pitch_index, N-mode->shortMdctSize, |
| 2007 st->prefilter_gain, gain1, st->prefilter_tapset, prefilter_tap
set, |
| 2008 mode->window, overlap); |
| 2009 } while (++c<CC); |
| 2010 |
| 2011 /* We reuse freq[] as scratch space for the de-emphasis */ |
| 2012 deemphasis(out_mem, (opus_val16*)pcm, N, CC, st->upsample, mode->preemph,
st->preemph_memD, freq); |
| 2013 st->prefilter_period_old = st->prefilter_period; |
| 2014 st->prefilter_gain_old = st->prefilter_gain; |
| 2015 st->prefilter_tapset_old = st->prefilter_tapset; |
| 2016 } |
| 2017 #endif |
| 2018 |
| 2019 st->prefilter_period = pitch_index; |
| 2020 st->prefilter_gain = gain1; |
| 2021 st->prefilter_tapset = prefilter_tapset; |
| 2022 #ifdef RESYNTH |
| 2023 if (LM!=0) |
| 2024 { |
| 2025 st->prefilter_period_old = st->prefilter_period; |
| 2026 st->prefilter_gain_old = st->prefilter_gain; |
| 2027 st->prefilter_tapset_old = st->prefilter_tapset; |
| 2028 } |
| 2029 #endif |
| 2030 |
| 2031 if (CC==2&&C==1) { |
| 2032 for (i=0;i<nbEBands;i++) |
| 2033 oldBandE[nbEBands+i]=oldBandE[i]; |
| 2034 } |
| 2035 |
| 2036 if (!isTransient) |
| 2037 { |
| 2038 for (i=0;i<CC*nbEBands;i++) |
| 2039 oldLogE2[i] = oldLogE[i]; |
| 2040 for (i=0;i<CC*nbEBands;i++) |
| 2041 oldLogE[i] = oldBandE[i]; |
| 2042 } else { |
| 2043 for (i=0;i<CC*nbEBands;i++) |
| 2044 oldLogE[i] = MIN16(oldLogE[i], oldBandE[i]); |
| 2045 } |
| 2046 /* In case start or end were to change */ |
| 2047 c=0; do |
| 2048 { |
| 2049 for (i=0;i<st->start;i++) |
| 2050 { |
| 2051 oldBandE[c*nbEBands+i]=0; |
| 2052 oldLogE[c*nbEBands+i]=oldLogE2[c*nbEBands+i]=-QCONST16(28.f,DB_SHIFT); |
| 2053 } |
| 2054 for (i=st->end;i<nbEBands;i++) |
| 2055 { |
| 2056 oldBandE[c*nbEBands+i]=0; |
| 2057 oldLogE[c*nbEBands+i]=oldLogE2[c*nbEBands+i]=-QCONST16(28.f,DB_SHIFT); |
| 2058 } |
| 2059 } while (++c<CC); |
| 2060 |
| 2061 if (isTransient || transient_got_disabled) |
| 2062 st->consec_transient++; |
| 2063 else |
| 2064 st->consec_transient=0; |
| 2065 st->rng = enc->rng; |
| 2066 |
| 2067 /* If there's any room left (can only happen for very high rates), |
| 2068 it's already filled with zeros */ |
| 2069 ec_enc_done(enc); |
| 2070 |
| 2071 #ifdef CUSTOM_MODES |
| 2072 if (st->signalling) |
| 2073 nbCompressedBytes++; |
| 2074 #endif |
| 2075 |
| 2076 RESTORE_STACK; |
| 2077 if (ec_get_error(enc)) |
| 2078 return OPUS_INTERNAL_ERROR; |
| 2079 else |
| 2080 return nbCompressedBytes; |
| 2081 } |
| 2082 |
| 2083 |
| 2084 #ifdef CUSTOM_MODES |
| 2085 |
| 2086 #ifdef FIXED_POINT |
| 2087 int opus_custom_encode(CELTEncoder * OPUS_RESTRICT st, const opus_int16 * pcm, i
nt frame_size, unsigned char *compressed, int nbCompressedBytes) |
| 2088 { |
| 2089 return celt_encode_with_ec(st, pcm, frame_size, compressed, nbCompressedBytes
, NULL); |
| 2090 } |
| 2091 |
| 2092 #ifndef DISABLE_FLOAT_API |
| 2093 int opus_custom_encode_float(CELTEncoder * OPUS_RESTRICT st, const float * pcm,
int frame_size, unsigned char *compressed, int nbCompressedBytes) |
| 2094 { |
| 2095 int j, ret, C, N; |
| 2096 VARDECL(opus_int16, in); |
| 2097 ALLOC_STACK; |
| 2098 |
| 2099 if (pcm==NULL) |
| 2100 return OPUS_BAD_ARG; |
| 2101 |
| 2102 C = st->channels; |
| 2103 N = frame_size; |
| 2104 ALLOC(in, C*N, opus_int16); |
| 2105 |
| 2106 for (j=0;j<C*N;j++) |
| 2107 in[j] = FLOAT2INT16(pcm[j]); |
| 2108 |
| 2109 ret=celt_encode_with_ec(st,in,frame_size,compressed,nbCompressedBytes, NULL); |
| 2110 #ifdef RESYNTH |
| 2111 for (j=0;j<C*N;j++) |
| 2112 ((float*)pcm)[j]=in[j]*(1.f/32768.f); |
| 2113 #endif |
| 2114 RESTORE_STACK; |
| 2115 return ret; |
| 2116 } |
| 2117 #endif /* DISABLE_FLOAT_API */ |
| 2118 #else |
| 2119 |
| 2120 int opus_custom_encode(CELTEncoder * OPUS_RESTRICT st, const opus_int16 * pcm, i
nt frame_size, unsigned char *compressed, int nbCompressedBytes) |
| 2121 { |
| 2122 int j, ret, C, N; |
| 2123 VARDECL(celt_sig, in); |
| 2124 ALLOC_STACK; |
| 2125 |
| 2126 if (pcm==NULL) |
| 2127 return OPUS_BAD_ARG; |
| 2128 |
| 2129 C=st->channels; |
| 2130 N=frame_size; |
| 2131 ALLOC(in, C*N, celt_sig); |
| 2132 for (j=0;j<C*N;j++) { |
| 2133 in[j] = SCALEOUT(pcm[j]); |
| 2134 } |
| 2135 |
| 2136 ret = celt_encode_with_ec(st,in,frame_size,compressed,nbCompressedBytes, NULL
); |
| 2137 #ifdef RESYNTH |
| 2138 for (j=0;j<C*N;j++) |
| 2139 ((opus_int16*)pcm)[j] = FLOAT2INT16(in[j]); |
| 2140 #endif |
| 2141 RESTORE_STACK; |
| 2142 return ret; |
| 2143 } |
| 2144 |
| 2145 int opus_custom_encode_float(CELTEncoder * OPUS_RESTRICT st, const float * pcm,
int frame_size, unsigned char *compressed, int nbCompressedBytes) |
| 2146 { |
| 2147 return celt_encode_with_ec(st, pcm, frame_size, compressed, nbCompressedBytes
, NULL); |
| 2148 } |
| 2149 |
| 2150 #endif |
| 2151 |
| 2152 #endif /* CUSTOM_MODES */ |
| 2153 |
| 2154 int opus_custom_encoder_ctl(CELTEncoder * OPUS_RESTRICT st, int request, ...) |
| 2155 { |
| 2156 va_list ap; |
| 2157 |
| 2158 va_start(ap, request); |
| 2159 switch (request) |
| 2160 { |
| 2161 case OPUS_SET_COMPLEXITY_REQUEST: |
| 2162 { |
| 2163 int value = va_arg(ap, opus_int32); |
| 2164 if (value<0 || value>10) |
| 2165 goto bad_arg; |
| 2166 st->complexity = value; |
| 2167 } |
| 2168 break; |
| 2169 case CELT_SET_START_BAND_REQUEST: |
| 2170 { |
| 2171 opus_int32 value = va_arg(ap, opus_int32); |
| 2172 if (value<0 || value>=st->mode->nbEBands) |
| 2173 goto bad_arg; |
| 2174 st->start = value; |
| 2175 } |
| 2176 break; |
| 2177 case CELT_SET_END_BAND_REQUEST: |
| 2178 { |
| 2179 opus_int32 value = va_arg(ap, opus_int32); |
| 2180 if (value<1 || value>st->mode->nbEBands) |
| 2181 goto bad_arg; |
| 2182 st->end = value; |
| 2183 } |
| 2184 break; |
| 2185 case CELT_SET_PREDICTION_REQUEST: |
| 2186 { |
| 2187 int value = va_arg(ap, opus_int32); |
| 2188 if (value<0 || value>2) |
| 2189 goto bad_arg; |
| 2190 st->disable_pf = value<=1; |
| 2191 st->force_intra = value==0; |
| 2192 } |
| 2193 break; |
| 2194 case OPUS_SET_PACKET_LOSS_PERC_REQUEST: |
| 2195 { |
| 2196 int value = va_arg(ap, opus_int32); |
| 2197 if (value<0 || value>100) |
| 2198 goto bad_arg; |
| 2199 st->loss_rate = value; |
| 2200 } |
| 2201 break; |
| 2202 case OPUS_SET_VBR_CONSTRAINT_REQUEST: |
| 2203 { |
| 2204 opus_int32 value = va_arg(ap, opus_int32); |
| 2205 st->constrained_vbr = value; |
| 2206 } |
| 2207 break; |
| 2208 case OPUS_SET_VBR_REQUEST: |
| 2209 { |
| 2210 opus_int32 value = va_arg(ap, opus_int32); |
| 2211 st->vbr = value; |
| 2212 } |
| 2213 break; |
| 2214 case OPUS_SET_BITRATE_REQUEST: |
| 2215 { |
| 2216 opus_int32 value = va_arg(ap, opus_int32); |
| 2217 if (value<=500 && value!=OPUS_BITRATE_MAX) |
| 2218 goto bad_arg; |
| 2219 value = IMIN(value, 260000*st->channels); |
| 2220 st->bitrate = value; |
| 2221 } |
| 2222 break; |
| 2223 case CELT_SET_CHANNELS_REQUEST: |
| 2224 { |
| 2225 opus_int32 value = va_arg(ap, opus_int32); |
| 2226 if (value<1 || value>2) |
| 2227 goto bad_arg; |
| 2228 st->stream_channels = value; |
| 2229 } |
| 2230 break; |
| 2231 case OPUS_SET_LSB_DEPTH_REQUEST: |
| 2232 { |
| 2233 opus_int32 value = va_arg(ap, opus_int32); |
| 2234 if (value<8 || value>24) |
| 2235 goto bad_arg; |
| 2236 st->lsb_depth=value; |
| 2237 } |
| 2238 break; |
| 2239 case OPUS_GET_LSB_DEPTH_REQUEST: |
| 2240 { |
| 2241 opus_int32 *value = va_arg(ap, opus_int32*); |
| 2242 *value=st->lsb_depth; |
| 2243 } |
| 2244 break; |
| 2245 case OPUS_SET_EXPERT_FRAME_DURATION_REQUEST: |
| 2246 { |
| 2247 opus_int32 value = va_arg(ap, opus_int32); |
| 2248 st->variable_duration = value; |
| 2249 } |
| 2250 break; |
| 2251 case OPUS_RESET_STATE: |
| 2252 { |
| 2253 int i; |
| 2254 opus_val16 *oldBandE, *oldLogE, *oldLogE2; |
| 2255 oldBandE = (opus_val16*)(st->in_mem+st->channels*(st->overlap+COMBFILTE
R_MAXPERIOD)); |
| 2256 oldLogE = oldBandE + st->channels*st->mode->nbEBands; |
| 2257 oldLogE2 = oldLogE + st->channels*st->mode->nbEBands; |
| 2258 OPUS_CLEAR((char*)&st->ENCODER_RESET_START, |
| 2259 opus_custom_encoder_get_size(st->mode, st->channels)- |
| 2260 ((char*)&st->ENCODER_RESET_START - (char*)st)); |
| 2261 for (i=0;i<st->channels*st->mode->nbEBands;i++) |
| 2262 oldLogE[i]=oldLogE2[i]=-QCONST16(28.f,DB_SHIFT); |
| 2263 st->vbr_offset = 0; |
| 2264 st->delayedIntra = 1; |
| 2265 st->spread_decision = SPREAD_NORMAL; |
| 2266 st->tonal_average = 256; |
| 2267 st->hf_average = 0; |
| 2268 st->tapset_decision = 0; |
| 2269 } |
| 2270 break; |
| 2271 #ifdef CUSTOM_MODES |
| 2272 case CELT_SET_INPUT_CLIPPING_REQUEST: |
| 2273 { |
| 2274 opus_int32 value = va_arg(ap, opus_int32); |
| 2275 st->clip = value; |
| 2276 } |
| 2277 break; |
| 2278 #endif |
| 2279 case CELT_SET_SIGNALLING_REQUEST: |
| 2280 { |
| 2281 opus_int32 value = va_arg(ap, opus_int32); |
| 2282 st->signalling = value; |
| 2283 } |
| 2284 break; |
| 2285 case CELT_SET_ANALYSIS_REQUEST: |
| 2286 { |
| 2287 AnalysisInfo *info = va_arg(ap, AnalysisInfo *); |
| 2288 if (info) |
| 2289 OPUS_COPY(&st->analysis, info, 1); |
| 2290 } |
| 2291 break; |
| 2292 case CELT_GET_MODE_REQUEST: |
| 2293 { |
| 2294 const CELTMode ** value = va_arg(ap, const CELTMode**); |
| 2295 if (value==0) |
| 2296 goto bad_arg; |
| 2297 *value=st->mode; |
| 2298 } |
| 2299 break; |
| 2300 case OPUS_GET_FINAL_RANGE_REQUEST: |
| 2301 { |
| 2302 opus_uint32 * value = va_arg(ap, opus_uint32 *); |
| 2303 if (value==0) |
| 2304 goto bad_arg; |
| 2305 *value=st->rng; |
| 2306 } |
| 2307 break; |
| 2308 case OPUS_SET_LFE_REQUEST: |
| 2309 { |
| 2310 opus_int32 value = va_arg(ap, opus_int32); |
| 2311 st->lfe = value; |
| 2312 } |
| 2313 break; |
| 2314 case OPUS_SET_ENERGY_MASK_REQUEST: |
| 2315 { |
| 2316 opus_val16 *value = va_arg(ap, opus_val16*); |
| 2317 st->energy_mask = value; |
| 2318 } |
| 2319 break; |
| 2320 default: |
| 2321 goto bad_request; |
| 2322 } |
| 2323 va_end(ap); |
| 2324 return OPUS_OK; |
| 2325 bad_arg: |
| 2326 va_end(ap); |
| 2327 return OPUS_BAD_ARG; |
| 2328 bad_request: |
| 2329 va_end(ap); |
| 2330 return OPUS_UNIMPLEMENTED; |
| 2331 } |
OLD | NEW |