1/* 2 * Copyright (c) 2012 Andrew D'Addesio 3 * Copyright (c) 2013-2014 Mozilla Corporation 4 * 5 * This file is part of FFmpeg. 6 * 7 * FFmpeg is free software; you can redistribute it and/or 8 * modify it under the terms of the GNU Lesser General Public 9 * License as published by the Free Software Foundation; either 10 * version 2.1 of the License, or (at your option) any later version. 11 * 12 * FFmpeg is distributed in the hope that it will be useful, 13 * but WITHOUT ANY WARRANTY; without even the implied warranty of 14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 15 * Lesser General Public License for more details. 16 * 17 * You should have received a copy of the GNU Lesser General Public 18 * License along with FFmpeg; if not, write to the Free Software 19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA 20 */ 21 22/** 23 * @file 24 * Opus decoder/parser shared code 25 */ 26 27#include <stdint.h> 28 29#include "libavutil/channel_layout.h" 30#include "libavutil/error.h" 31#include "libavutil/ffmath.h" 32 33#include "opus_celt.h" 34#include "opustab.h" 35#include "internal.h" 36#include "vorbis.h" 37 38static const uint16_t opus_frame_duration[32] = { 39 480, 960, 1920, 2880, 40 480, 960, 1920, 2880, 41 480, 960, 1920, 2880, 42 480, 960, 43 480, 960, 44 120, 240, 480, 960, 45 120, 240, 480, 960, 46 120, 240, 480, 960, 47 120, 240, 480, 960, 48}; 49 50/** 51 * Read a 1- or 2-byte frame length 52 */ 53static inline int xiph_lacing_16bit(const uint8_t **ptr, const uint8_t *end) 54{ 55 int val; 56 57 if (*ptr >= end) 58 return AVERROR_INVALIDDATA; 59 val = *(*ptr)++; 60 if (val >= 252) { 61 if (*ptr >= end) 62 return AVERROR_INVALIDDATA; 63 val += 4 * *(*ptr)++; 64 } 65 return val; 66} 67 68/** 69 * Read a multi-byte length (used for code 3 packet padding size) 70 */ 71static inline int xiph_lacing_full(const uint8_t **ptr, const uint8_t *end) 72{ 73 int val = 0; 74 int next; 75 76 while (1) { 77 if (*ptr >= end || val > INT_MAX - 254) 78 return AVERROR_INVALIDDATA; 79 next = *(*ptr)++; 80 val += next; 81 if (next < 255) 82 break; 83 else 84 val--; 85 } 86 return val; 87} 88 89/** 90 * Parse Opus packet info from raw packet data 91 */ 92int ff_opus_parse_packet(OpusPacket *pkt, const uint8_t *buf, int buf_size, 93 int self_delimiting) 94{ 95 const uint8_t *ptr = buf; 96 const uint8_t *end = buf + buf_size; 97 int padding = 0; 98 int frame_bytes, i; 99 100 if (buf_size < 1) 101 goto fail; 102 103 /* TOC byte */ 104 i = *ptr++; 105 pkt->code = (i ) & 0x3; 106 pkt->stereo = (i >> 2) & 0x1; 107 pkt->config = (i >> 3) & 0x1F; 108 109 /* code 2 and code 3 packets have at least 1 byte after the TOC */ 110 if (pkt->code >= 2 && buf_size < 2) 111 goto fail; 112 113 switch (pkt->code) { 114 case 0: 115 /* 1 frame */ 116 pkt->frame_count = 1; 117 pkt->vbr = 0; 118 119 if (self_delimiting) { 120 int len = xiph_lacing_16bit(&ptr, end); 121 if (len < 0 || len > end - ptr) 122 goto fail; 123 end = ptr + len; 124 buf_size = end - buf; 125 } 126 127 frame_bytes = end - ptr; 128 if (frame_bytes > MAX_FRAME_SIZE) 129 goto fail; 130 pkt->frame_offset[0] = ptr - buf; 131 pkt->frame_size[0] = frame_bytes; 132 break; 133 case 1: 134 /* 2 frames, equal size */ 135 pkt->frame_count = 2; 136 pkt->vbr = 0; 137 138 if (self_delimiting) { 139 int len = xiph_lacing_16bit(&ptr, end); 140 if (len < 0 || 2 * len > end - ptr) 141 goto fail; 142 end = ptr + 2 * len; 143 buf_size = end - buf; 144 } 145 146 frame_bytes = end - ptr; 147 if (frame_bytes & 1 || frame_bytes >> 1 > MAX_FRAME_SIZE) 148 goto fail; 149 pkt->frame_offset[0] = ptr - buf; 150 pkt->frame_size[0] = frame_bytes >> 1; 151 pkt->frame_offset[1] = pkt->frame_offset[0] + pkt->frame_size[0]; 152 pkt->frame_size[1] = frame_bytes >> 1; 153 break; 154 case 2: 155 /* 2 frames, different sizes */ 156 pkt->frame_count = 2; 157 pkt->vbr = 1; 158 159 /* read 1st frame size */ 160 frame_bytes = xiph_lacing_16bit(&ptr, end); 161 if (frame_bytes < 0) 162 goto fail; 163 164 if (self_delimiting) { 165 int len = xiph_lacing_16bit(&ptr, end); 166 if (len < 0 || len + frame_bytes > end - ptr) 167 goto fail; 168 end = ptr + frame_bytes + len; 169 buf_size = end - buf; 170 } 171 172 pkt->frame_offset[0] = ptr - buf; 173 pkt->frame_size[0] = frame_bytes; 174 175 /* calculate 2nd frame size */ 176 frame_bytes = end - ptr - pkt->frame_size[0]; 177 if (frame_bytes < 0 || frame_bytes > MAX_FRAME_SIZE) 178 goto fail; 179 pkt->frame_offset[1] = pkt->frame_offset[0] + pkt->frame_size[0]; 180 pkt->frame_size[1] = frame_bytes; 181 break; 182 case 3: 183 /* 1 to 48 frames, can be different sizes */ 184 i = *ptr++; 185 pkt->frame_count = (i ) & 0x3F; 186 padding = (i >> 6) & 0x01; 187 pkt->vbr = (i >> 7) & 0x01; 188 189 if (pkt->frame_count == 0 || pkt->frame_count > MAX_FRAMES) 190 goto fail; 191 192 /* read padding size */ 193 if (padding) { 194 padding = xiph_lacing_full(&ptr, end); 195 if (padding < 0) 196 goto fail; 197 } 198 199 /* read frame sizes */ 200 if (pkt->vbr) { 201 /* for VBR, all frames except the final one have their size coded 202 in the bitstream. the last frame size is implicit. */ 203 int total_bytes = 0; 204 for (i = 0; i < pkt->frame_count - 1; i++) { 205 frame_bytes = xiph_lacing_16bit(&ptr, end); 206 if (frame_bytes < 0) 207 goto fail; 208 pkt->frame_size[i] = frame_bytes; 209 total_bytes += frame_bytes; 210 } 211 212 if (self_delimiting) { 213 int len = xiph_lacing_16bit(&ptr, end); 214 if (len < 0 || len + total_bytes + padding > end - ptr) 215 goto fail; 216 end = ptr + total_bytes + len + padding; 217 buf_size = end - buf; 218 } 219 220 frame_bytes = end - ptr - padding; 221 if (total_bytes > frame_bytes) 222 goto fail; 223 pkt->frame_offset[0] = ptr - buf; 224 for (i = 1; i < pkt->frame_count; i++) 225 pkt->frame_offset[i] = pkt->frame_offset[i-1] + pkt->frame_size[i-1]; 226 pkt->frame_size[pkt->frame_count-1] = frame_bytes - total_bytes; 227 } else { 228 /* for CBR, the remaining packet bytes are divided evenly between 229 the frames */ 230 if (self_delimiting) { 231 frame_bytes = xiph_lacing_16bit(&ptr, end); 232 if (frame_bytes < 0 || pkt->frame_count * frame_bytes + padding > end - ptr) 233 goto fail; 234 end = ptr + pkt->frame_count * frame_bytes + padding; 235 buf_size = end - buf; 236 } else { 237 frame_bytes = end - ptr - padding; 238 if (frame_bytes % pkt->frame_count || 239 frame_bytes / pkt->frame_count > MAX_FRAME_SIZE) 240 goto fail; 241 frame_bytes /= pkt->frame_count; 242 } 243 244 pkt->frame_offset[0] = ptr - buf; 245 pkt->frame_size[0] = frame_bytes; 246 for (i = 1; i < pkt->frame_count; i++) { 247 pkt->frame_offset[i] = pkt->frame_offset[i-1] + pkt->frame_size[i-1]; 248 pkt->frame_size[i] = frame_bytes; 249 } 250 } 251 } 252 253 pkt->packet_size = buf_size; 254 pkt->data_size = pkt->packet_size - padding; 255 256 /* total packet duration cannot be larger than 120ms */ 257 pkt->frame_duration = opus_frame_duration[pkt->config]; 258 if (pkt->frame_duration * pkt->frame_count > MAX_PACKET_DUR) 259 goto fail; 260 261 /* set mode and bandwidth */ 262 if (pkt->config < 12) { 263 pkt->mode = OPUS_MODE_SILK; 264 pkt->bandwidth = pkt->config >> 2; 265 } else if (pkt->config < 16) { 266 pkt->mode = OPUS_MODE_HYBRID; 267 pkt->bandwidth = OPUS_BANDWIDTH_SUPERWIDEBAND + (pkt->config >= 14); 268 } else { 269 pkt->mode = OPUS_MODE_CELT; 270 pkt->bandwidth = (pkt->config - 16) >> 2; 271 /* skip medium band */ 272 if (pkt->bandwidth) 273 pkt->bandwidth++; 274 } 275 276 return 0; 277 278fail: 279 memset(pkt, 0, sizeof(*pkt)); 280 return AVERROR_INVALIDDATA; 281} 282 283static int channel_reorder_vorbis(int nb_channels, int channel_idx) 284{ 285 return ff_vorbis_channel_layout_offsets[nb_channels - 1][channel_idx]; 286} 287 288static int channel_reorder_unknown(int nb_channels, int channel_idx) 289{ 290 return channel_idx; 291} 292 293av_cold int ff_opus_parse_extradata(AVCodecContext *avctx, 294 OpusContext *s) 295{ 296 static const uint8_t default_channel_map[2] = { 0, 1 }; 297 298 int (*channel_reorder)(int, int) = channel_reorder_unknown; 299 int channels = avctx->ch_layout.nb_channels; 300 301 const uint8_t *extradata, *channel_map; 302 int extradata_size; 303 int version, map_type, streams, stereo_streams, i, j, ret; 304 AVChannelLayout layout = { 0 }; 305 306 if (!avctx->extradata) { 307 if (channels > 2) { 308 av_log(avctx, AV_LOG_ERROR, 309 "Multichannel configuration without extradata.\n"); 310 return AVERROR(EINVAL); 311 } 312 extradata = opus_default_extradata; 313 extradata_size = sizeof(opus_default_extradata); 314 } else { 315 extradata = avctx->extradata; 316 extradata_size = avctx->extradata_size; 317 } 318 319 if (extradata_size < 19) { 320 av_log(avctx, AV_LOG_ERROR, "Invalid extradata size: %d\n", 321 extradata_size); 322 return AVERROR_INVALIDDATA; 323 } 324 325 version = extradata[8]; 326 if (version > 15) { 327 avpriv_request_sample(avctx, "Extradata version %d", version); 328 return AVERROR_PATCHWELCOME; 329 } 330 331 avctx->delay = AV_RL16(extradata + 10); 332 if (avctx->internal) 333 avctx->internal->skip_samples = avctx->delay; 334 335 channels = avctx->extradata ? extradata[9] : (channels == 1) ? 1 : 2; 336 if (!channels) { 337 av_log(avctx, AV_LOG_ERROR, "Zero channel count specified in the extradata\n"); 338 return AVERROR_INVALIDDATA; 339 } 340 341 s->gain_i = AV_RL16(extradata + 16); 342 if (s->gain_i) 343 s->gain = ff_exp10(s->gain_i / (20.0 * 256)); 344 345 map_type = extradata[18]; 346 if (!map_type) { 347 if (channels > 2) { 348 av_log(avctx, AV_LOG_ERROR, 349 "Channel mapping 0 is only specified for up to 2 channels\n"); 350 ret = AVERROR_INVALIDDATA; 351 goto fail; 352 } 353 layout = (channels == 1) ? (AVChannelLayout)AV_CHANNEL_LAYOUT_MONO : 354 (AVChannelLayout)AV_CHANNEL_LAYOUT_STEREO; 355 streams = 1; 356 stereo_streams = channels - 1; 357 channel_map = default_channel_map; 358 } else if (map_type == 1 || map_type == 2 || map_type == 255) { 359 if (extradata_size < 21 + channels) { 360 av_log(avctx, AV_LOG_ERROR, "Invalid extradata size: %d\n", 361 extradata_size); 362 ret = AVERROR_INVALIDDATA; 363 goto fail; 364 } 365 366 streams = extradata[19]; 367 stereo_streams = extradata[20]; 368 if (!streams || stereo_streams > streams || 369 streams + stereo_streams > 255) { 370 av_log(avctx, AV_LOG_ERROR, 371 "Invalid stream/stereo stream count: %d/%d\n", streams, stereo_streams); 372 ret = AVERROR_INVALIDDATA; 373 goto fail; 374 } 375 376 if (map_type == 1) { 377 if (channels > 8) { 378 av_log(avctx, AV_LOG_ERROR, 379 "Channel mapping 1 is only specified for up to 8 channels\n"); 380 ret = AVERROR_INVALIDDATA; 381 goto fail; 382 } 383 av_channel_layout_copy(&layout, &ff_vorbis_ch_layouts[channels - 1]); 384 channel_reorder = channel_reorder_vorbis; 385 } else if (map_type == 2) { 386 int ambisonic_order = ff_sqrt(channels) - 1; 387 if (channels != ((ambisonic_order + 1) * (ambisonic_order + 1)) && 388 channels != ((ambisonic_order + 1) * (ambisonic_order + 1) + 2)) { 389 av_log(avctx, AV_LOG_ERROR, 390 "Channel mapping 2 is only specified for channel counts" 391 " which can be written as (n + 1)^2 or (n + 1)^2 + 2" 392 " for nonnegative integer n\n"); 393 ret = AVERROR_INVALIDDATA; 394 goto fail; 395 } 396 if (channels > 227) { 397 av_log(avctx, AV_LOG_ERROR, "Too many channels\n"); 398 ret = AVERROR_INVALIDDATA; 399 goto fail; 400 } 401 402 layout.order = AV_CHANNEL_ORDER_AMBISONIC; 403 layout.nb_channels = channels; 404 if (channels != ((ambisonic_order + 1) * (ambisonic_order + 1))) 405 layout.u.mask = AV_CH_LAYOUT_STEREO; 406 } else { 407 layout.order = AV_CHANNEL_ORDER_UNSPEC; 408 layout.nb_channels = channels; 409 } 410 411 channel_map = extradata + 21; 412 } else { 413 avpriv_request_sample(avctx, "Mapping type %d", map_type); 414 return AVERROR_PATCHWELCOME; 415 } 416 417 s->channel_maps = av_calloc(channels, sizeof(*s->channel_maps)); 418 if (!s->channel_maps) { 419 ret = AVERROR(ENOMEM); 420 goto fail; 421 } 422 423 for (i = 0; i < channels; i++) { 424 ChannelMap *map = &s->channel_maps[i]; 425 uint8_t idx = channel_map[channel_reorder(channels, i)]; 426 427 if (idx == 255) { 428 map->silence = 1; 429 continue; 430 } else if (idx >= streams + stereo_streams) { 431 av_log(avctx, AV_LOG_ERROR, 432 "Invalid channel map for output channel %d: %d\n", i, idx); 433 av_freep(&s->channel_maps); 434 ret = AVERROR_INVALIDDATA; 435 goto fail; 436 } 437 438 /* check that we did not see this index yet */ 439 map->copy = 0; 440 for (j = 0; j < i; j++) 441 if (channel_map[channel_reorder(channels, j)] == idx) { 442 map->copy = 1; 443 map->copy_idx = j; 444 break; 445 } 446 447 if (idx < 2 * stereo_streams) { 448 map->stream_idx = idx / 2; 449 map->channel_idx = idx & 1; 450 } else { 451 map->stream_idx = idx - stereo_streams; 452 map->channel_idx = 0; 453 } 454 } 455 456 ret = av_channel_layout_copy(&avctx->ch_layout, &layout); 457 if (ret < 0) 458 goto fail; 459 460 s->nb_streams = streams; 461 s->nb_stereo_streams = stereo_streams; 462 463 return 0; 464fail: 465 av_channel_layout_uninit(&layout); 466 return ret; 467} 468 469void ff_celt_quant_bands(CeltFrame *f, OpusRangeCoder *rc) 470{ 471 float lowband_scratch[8 * 22]; 472 float norm1[2 * 8 * 100]; 473 float *norm2 = norm1 + 8 * 100; 474 475 int totalbits = (f->framebits << 3) - f->anticollapse_needed; 476 477 int update_lowband = 1; 478 int lowband_offset = 0; 479 480 int i, j; 481 482 for (i = f->start_band; i < f->end_band; i++) { 483 uint32_t cm[2] = { (1 << f->blocks) - 1, (1 << f->blocks) - 1 }; 484 int band_offset = ff_celt_freq_bands[i] << f->size; 485 int band_size = ff_celt_freq_range[i] << f->size; 486 float *X = f->block[0].coeffs + band_offset; 487 float *Y = (f->channels == 2) ? f->block[1].coeffs + band_offset : NULL; 488 float *norm_loc1, *norm_loc2; 489 490 int consumed = opus_rc_tell_frac(rc); 491 int effective_lowband = -1; 492 int b = 0; 493 494 /* Compute how many bits we want to allocate to this band */ 495 if (i != f->start_band) 496 f->remaining -= consumed; 497 f->remaining2 = totalbits - consumed - 1; 498 if (i <= f->coded_bands - 1) { 499 int curr_balance = f->remaining / FFMIN(3, f->coded_bands-i); 500 b = av_clip_uintp2(FFMIN(f->remaining2 + 1, f->pulses[i] + curr_balance), 14); 501 } 502 503 if ((ff_celt_freq_bands[i] - ff_celt_freq_range[i] >= ff_celt_freq_bands[f->start_band] || 504 i == f->start_band + 1) && (update_lowband || lowband_offset == 0)) 505 lowband_offset = i; 506 507 if (i == f->start_band + 1) { 508 /* Special Hybrid Folding (RFC 8251 section 9). Copy the first band into 509 the second to ensure the second band never has to use the LCG. */ 510 int count = (ff_celt_freq_range[i] - ff_celt_freq_range[i-1]) << f->size; 511 512 memcpy(&norm1[band_offset], &norm1[band_offset - count], count * sizeof(float)); 513 514 if (f->channels == 2) 515 memcpy(&norm2[band_offset], &norm2[band_offset - count], count * sizeof(float)); 516 } 517 518 /* Get a conservative estimate of the collapse_mask's for the bands we're 519 going to be folding from. */ 520 if (lowband_offset != 0 && (f->spread != CELT_SPREAD_AGGRESSIVE || 521 f->blocks > 1 || f->tf_change[i] < 0)) { 522 int foldstart, foldend; 523 524 /* This ensures we never repeat spectral content within one band */ 525 effective_lowband = FFMAX(ff_celt_freq_bands[f->start_band], 526 ff_celt_freq_bands[lowband_offset] - ff_celt_freq_range[i]); 527 foldstart = lowband_offset; 528 while (ff_celt_freq_bands[--foldstart] > effective_lowband); 529 foldend = lowband_offset - 1; 530 while (++foldend < i && ff_celt_freq_bands[foldend] < effective_lowband + ff_celt_freq_range[i]); 531 532 cm[0] = cm[1] = 0; 533 for (j = foldstart; j < foldend; j++) { 534 cm[0] |= f->block[0].collapse_masks[j]; 535 cm[1] |= f->block[f->channels - 1].collapse_masks[j]; 536 } 537 } 538 539 if (f->dual_stereo && i == f->intensity_stereo) { 540 /* Switch off dual stereo to do intensity */ 541 f->dual_stereo = 0; 542 for (j = ff_celt_freq_bands[f->start_band] << f->size; j < band_offset; j++) 543 norm1[j] = (norm1[j] + norm2[j]) / 2; 544 } 545 546 norm_loc1 = effective_lowband != -1 ? norm1 + (effective_lowband << f->size) : NULL; 547 norm_loc2 = effective_lowband != -1 ? norm2 + (effective_lowband << f->size) : NULL; 548 549 if (f->dual_stereo) { 550 cm[0] = f->pvq->quant_band(f->pvq, f, rc, i, X, NULL, band_size, b >> 1, 551 f->blocks, norm_loc1, f->size, 552 norm1 + band_offset, 0, 1.0f, 553 lowband_scratch, cm[0]); 554 555 cm[1] = f->pvq->quant_band(f->pvq, f, rc, i, Y, NULL, band_size, b >> 1, 556 f->blocks, norm_loc2, f->size, 557 norm2 + band_offset, 0, 1.0f, 558 lowband_scratch, cm[1]); 559 } else { 560 cm[0] = f->pvq->quant_band(f->pvq, f, rc, i, X, Y, band_size, b >> 0, 561 f->blocks, norm_loc1, f->size, 562 norm1 + band_offset, 0, 1.0f, 563 lowband_scratch, cm[0] | cm[1]); 564 cm[1] = cm[0]; 565 } 566 567 f->block[0].collapse_masks[i] = (uint8_t)cm[0]; 568 f->block[f->channels - 1].collapse_masks[i] = (uint8_t)cm[1]; 569 f->remaining += f->pulses[i] + consumed; 570 571 /* Update the folding position only as long as we have 1 bit/sample depth */ 572 update_lowband = (b > band_size << 3); 573 } 574} 575 576#define NORMC(bits) ((bits) << (f->channels - 1) << f->size >> 2) 577 578void ff_celt_bitalloc(CeltFrame *f, OpusRangeCoder *rc, int encode) 579{ 580 int i, j, low, high, total, done, bandbits, remaining, tbits_8ths; 581 int skip_startband = f->start_band; 582 int skip_bit = 0; 583 int intensitystereo_bit = 0; 584 int dualstereo_bit = 0; 585 int dynalloc = 6; 586 int extrabits = 0; 587 588 int boost[CELT_MAX_BANDS] = { 0 }; 589 int trim_offset[CELT_MAX_BANDS]; 590 int threshold[CELT_MAX_BANDS]; 591 int bits1[CELT_MAX_BANDS]; 592 int bits2[CELT_MAX_BANDS]; 593 594 /* Spread */ 595 if (opus_rc_tell(rc) + 4 <= f->framebits) { 596 if (encode) 597 ff_opus_rc_enc_cdf(rc, f->spread, ff_celt_model_spread); 598 else 599 f->spread = ff_opus_rc_dec_cdf(rc, ff_celt_model_spread); 600 } else { 601 f->spread = CELT_SPREAD_NORMAL; 602 } 603 604 /* Initialize static allocation caps */ 605 for (i = 0; i < CELT_MAX_BANDS; i++) 606 f->caps[i] = NORMC((ff_celt_static_caps[f->size][f->channels - 1][i] + 64) * ff_celt_freq_range[i]); 607 608 /* Band boosts */ 609 tbits_8ths = f->framebits << 3; 610 for (i = f->start_band; i < f->end_band; i++) { 611 int quanta = ff_celt_freq_range[i] << (f->channels - 1) << f->size; 612 int b_dynalloc = dynalloc; 613 int boost_amount = f->alloc_boost[i]; 614 quanta = FFMIN(quanta << 3, FFMAX(6 << 3, quanta)); 615 616 while (opus_rc_tell_frac(rc) + (b_dynalloc << 3) < tbits_8ths && boost[i] < f->caps[i]) { 617 int is_boost; 618 if (encode) { 619 is_boost = boost_amount--; 620 ff_opus_rc_enc_log(rc, is_boost, b_dynalloc); 621 } else { 622 is_boost = ff_opus_rc_dec_log(rc, b_dynalloc); 623 } 624 625 if (!is_boost) 626 break; 627 628 boost[i] += quanta; 629 tbits_8ths -= quanta; 630 631 b_dynalloc = 1; 632 } 633 634 if (boost[i]) 635 dynalloc = FFMAX(dynalloc - 1, 2); 636 } 637 638 /* Allocation trim */ 639 if (!encode) 640 f->alloc_trim = 5; 641 if (opus_rc_tell_frac(rc) + (6 << 3) <= tbits_8ths) 642 if (encode) 643 ff_opus_rc_enc_cdf(rc, f->alloc_trim, ff_celt_model_alloc_trim); 644 else 645 f->alloc_trim = ff_opus_rc_dec_cdf(rc, ff_celt_model_alloc_trim); 646 647 /* Anti-collapse bit reservation */ 648 tbits_8ths = (f->framebits << 3) - opus_rc_tell_frac(rc) - 1; 649 f->anticollapse_needed = 0; 650 if (f->transient && f->size >= 2 && tbits_8ths >= ((f->size + 2) << 3)) 651 f->anticollapse_needed = 1 << 3; 652 tbits_8ths -= f->anticollapse_needed; 653 654 /* Band skip bit reservation */ 655 if (tbits_8ths >= 1 << 3) 656 skip_bit = 1 << 3; 657 tbits_8ths -= skip_bit; 658 659 /* Intensity/dual stereo bit reservation */ 660 if (f->channels == 2) { 661 intensitystereo_bit = ff_celt_log2_frac[f->end_band - f->start_band]; 662 if (intensitystereo_bit <= tbits_8ths) { 663 tbits_8ths -= intensitystereo_bit; 664 if (tbits_8ths >= 1 << 3) { 665 dualstereo_bit = 1 << 3; 666 tbits_8ths -= 1 << 3; 667 } 668 } else { 669 intensitystereo_bit = 0; 670 } 671 } 672 673 /* Trim offsets */ 674 for (i = f->start_band; i < f->end_band; i++) { 675 int trim = f->alloc_trim - 5 - f->size; 676 int band = ff_celt_freq_range[i] * (f->end_band - i - 1); 677 int duration = f->size + 3; 678 int scale = duration + f->channels - 1; 679 680 /* PVQ minimum allocation threshold, below this value the band is 681 * skipped */ 682 threshold[i] = FFMAX(3 * ff_celt_freq_range[i] << duration >> 4, 683 f->channels << 3); 684 685 trim_offset[i] = trim * (band << scale) >> 6; 686 687 if (ff_celt_freq_range[i] << f->size == 1) 688 trim_offset[i] -= f->channels << 3; 689 } 690 691 /* Bisection */ 692 low = 1; 693 high = CELT_VECTORS - 1; 694 while (low <= high) { 695 int center = (low + high) >> 1; 696 done = total = 0; 697 698 for (i = f->end_band - 1; i >= f->start_band; i--) { 699 bandbits = NORMC(ff_celt_freq_range[i] * ff_celt_static_alloc[center][i]); 700 701 if (bandbits) 702 bandbits = FFMAX(bandbits + trim_offset[i], 0); 703 bandbits += boost[i]; 704 705 if (bandbits >= threshold[i] || done) { 706 done = 1; 707 total += FFMIN(bandbits, f->caps[i]); 708 } else if (bandbits >= f->channels << 3) { 709 total += f->channels << 3; 710 } 711 } 712 713 if (total > tbits_8ths) 714 high = center - 1; 715 else 716 low = center + 1; 717 } 718 high = low--; 719 720 /* Bisection */ 721 for (i = f->start_band; i < f->end_band; i++) { 722 bits1[i] = NORMC(ff_celt_freq_range[i] * ff_celt_static_alloc[low][i]); 723 bits2[i] = high >= CELT_VECTORS ? f->caps[i] : 724 NORMC(ff_celt_freq_range[i] * ff_celt_static_alloc[high][i]); 725 726 if (bits1[i]) 727 bits1[i] = FFMAX(bits1[i] + trim_offset[i], 0); 728 if (bits2[i]) 729 bits2[i] = FFMAX(bits2[i] + trim_offset[i], 0); 730 731 if (low) 732 bits1[i] += boost[i]; 733 bits2[i] += boost[i]; 734 735 if (boost[i]) 736 skip_startband = i; 737 bits2[i] = FFMAX(bits2[i] - bits1[i], 0); 738 } 739 740 /* Bisection */ 741 low = 0; 742 high = 1 << CELT_ALLOC_STEPS; 743 for (i = 0; i < CELT_ALLOC_STEPS; i++) { 744 int center = (low + high) >> 1; 745 done = total = 0; 746 747 for (j = f->end_band - 1; j >= f->start_band; j--) { 748 bandbits = bits1[j] + (center * bits2[j] >> CELT_ALLOC_STEPS); 749 750 if (bandbits >= threshold[j] || done) { 751 done = 1; 752 total += FFMIN(bandbits, f->caps[j]); 753 } else if (bandbits >= f->channels << 3) 754 total += f->channels << 3; 755 } 756 if (total > tbits_8ths) 757 high = center; 758 else 759 low = center; 760 } 761 762 /* Bisection */ 763 done = total = 0; 764 for (i = f->end_band - 1; i >= f->start_band; i--) { 765 bandbits = bits1[i] + (low * bits2[i] >> CELT_ALLOC_STEPS); 766 767 if (bandbits >= threshold[i] || done) 768 done = 1; 769 else 770 bandbits = (bandbits >= f->channels << 3) ? 771 f->channels << 3 : 0; 772 773 bandbits = FFMIN(bandbits, f->caps[i]); 774 f->pulses[i] = bandbits; 775 total += bandbits; 776 } 777 778 /* Band skipping */ 779 for (f->coded_bands = f->end_band; ; f->coded_bands--) { 780 int allocation; 781 j = f->coded_bands - 1; 782 783 if (j == skip_startband) { 784 /* all remaining bands are not skipped */ 785 tbits_8ths += skip_bit; 786 break; 787 } 788 789 /* determine the number of bits available for coding "do not skip" markers */ 790 remaining = tbits_8ths - total; 791 bandbits = remaining / (ff_celt_freq_bands[j+1] - ff_celt_freq_bands[f->start_band]); 792 remaining -= bandbits * (ff_celt_freq_bands[j+1] - ff_celt_freq_bands[f->start_band]); 793 allocation = f->pulses[j] + bandbits * ff_celt_freq_range[j]; 794 allocation += FFMAX(remaining - (ff_celt_freq_bands[j] - ff_celt_freq_bands[f->start_band]), 0); 795 796 /* a "do not skip" marker is only coded if the allocation is 797 * above the chosen threshold */ 798 if (allocation >= FFMAX(threshold[j], (f->channels + 1) << 3)) { 799 int do_not_skip; 800 if (encode) { 801 do_not_skip = f->coded_bands <= f->skip_band_floor; 802 ff_opus_rc_enc_log(rc, do_not_skip, 1); 803 } else { 804 do_not_skip = ff_opus_rc_dec_log(rc, 1); 805 } 806 807 if (do_not_skip) 808 break; 809 810 total += 1 << 3; 811 allocation -= 1 << 3; 812 } 813 814 /* the band is skipped, so reclaim its bits */ 815 total -= f->pulses[j]; 816 if (intensitystereo_bit) { 817 total -= intensitystereo_bit; 818 intensitystereo_bit = ff_celt_log2_frac[j - f->start_band]; 819 total += intensitystereo_bit; 820 } 821 822 total += f->pulses[j] = (allocation >= f->channels << 3) ? f->channels << 3 : 0; 823 } 824 825 /* IS start band */ 826 if (encode) { 827 if (intensitystereo_bit) { 828 f->intensity_stereo = FFMIN(f->intensity_stereo, f->coded_bands); 829 ff_opus_rc_enc_uint(rc, f->intensity_stereo, f->coded_bands + 1 - f->start_band); 830 } 831 } else { 832 f->intensity_stereo = f->dual_stereo = 0; 833 if (intensitystereo_bit) 834 f->intensity_stereo = f->start_band + ff_opus_rc_dec_uint(rc, f->coded_bands + 1 - f->start_band); 835 } 836 837 /* DS flag */ 838 if (f->intensity_stereo <= f->start_band) 839 tbits_8ths += dualstereo_bit; /* no intensity stereo means no dual stereo */ 840 else if (dualstereo_bit) 841 if (encode) 842 ff_opus_rc_enc_log(rc, f->dual_stereo, 1); 843 else 844 f->dual_stereo = ff_opus_rc_dec_log(rc, 1); 845 846 /* Supply the remaining bits in this frame to lower bands */ 847 remaining = tbits_8ths - total; 848 bandbits = remaining / (ff_celt_freq_bands[f->coded_bands] - ff_celt_freq_bands[f->start_band]); 849 remaining -= bandbits * (ff_celt_freq_bands[f->coded_bands] - ff_celt_freq_bands[f->start_band]); 850 for (i = f->start_band; i < f->coded_bands; i++) { 851 const int bits = FFMIN(remaining, ff_celt_freq_range[i]); 852 f->pulses[i] += bits + bandbits * ff_celt_freq_range[i]; 853 remaining -= bits; 854 } 855 856 /* Finally determine the allocation */ 857 for (i = f->start_band; i < f->coded_bands; i++) { 858 int N = ff_celt_freq_range[i] << f->size; 859 int prev_extra = extrabits; 860 f->pulses[i] += extrabits; 861 862 if (N > 1) { 863 int dof; /* degrees of freedom */ 864 int temp; /* dof * channels * log(dof) */ 865 int fine_bits; 866 int max_bits; 867 int offset; /* fine energy quantization offset, i.e. 868 * extra bits assigned over the standard 869 * totalbits/dof */ 870 871 extrabits = FFMAX(f->pulses[i] - f->caps[i], 0); 872 f->pulses[i] -= extrabits; 873 874 /* intensity stereo makes use of an extra degree of freedom */ 875 dof = N * f->channels + (f->channels == 2 && N > 2 && !f->dual_stereo && i < f->intensity_stereo); 876 temp = dof * (ff_celt_log_freq_range[i] + (f->size << 3)); 877 offset = (temp >> 1) - dof * CELT_FINE_OFFSET; 878 if (N == 2) /* dof=2 is the only case that doesn't fit the model */ 879 offset += dof << 1; 880 881 /* grant an additional bias for the first and second pulses */ 882 if (f->pulses[i] + offset < 2 * (dof << 3)) 883 offset += temp >> 2; 884 else if (f->pulses[i] + offset < 3 * (dof << 3)) 885 offset += temp >> 3; 886 887 fine_bits = (f->pulses[i] + offset + (dof << 2)) / (dof << 3); 888 max_bits = FFMIN((f->pulses[i] >> 3) >> (f->channels - 1), CELT_MAX_FINE_BITS); 889 max_bits = FFMAX(max_bits, 0); 890 f->fine_bits[i] = av_clip(fine_bits, 0, max_bits); 891 892 /* If fine_bits was rounded down or capped, 893 * give priority for the final fine energy pass */ 894 f->fine_priority[i] = (f->fine_bits[i] * (dof << 3) >= f->pulses[i] + offset); 895 896 /* the remaining bits are assigned to PVQ */ 897 f->pulses[i] -= f->fine_bits[i] << (f->channels - 1) << 3; 898 } else { 899 /* all bits go to fine energy except for the sign bit */ 900 extrabits = FFMAX(f->pulses[i] - (f->channels << 3), 0); 901 f->pulses[i] -= extrabits; 902 f->fine_bits[i] = 0; 903 f->fine_priority[i] = 1; 904 } 905 906 /* hand back a limited number of extra fine energy bits to this band */ 907 if (extrabits > 0) { 908 int fineextra = FFMIN(extrabits >> (f->channels + 2), 909 CELT_MAX_FINE_BITS - f->fine_bits[i]); 910 f->fine_bits[i] += fineextra; 911 912 fineextra <<= f->channels + 2; 913 f->fine_priority[i] = (fineextra >= extrabits - prev_extra); 914 extrabits -= fineextra; 915 } 916 } 917 f->remaining = extrabits; 918 919 /* skipped bands dedicate all of their bits for fine energy */ 920 for (; i < f->end_band; i++) { 921 f->fine_bits[i] = f->pulses[i] >> (f->channels - 1) >> 3; 922 f->pulses[i] = 0; 923 f->fine_priority[i] = f->fine_bits[i] < 1; 924 } 925} 926