libavcodec/adpcmenc.c

   1 /*
   2  * Copyright (c) 2001-2003 The FFmpeg project
   3  *
   4  * first version by Francois Revol (revol@free.fr)
   5  * fringe ADPCM codecs (e.g., DK3, DK4, Westwood)
   6  *   by Mike Melanson (melanson@pcisys.net)
   7  *
   8  * This file is part of FFmpeg.
   9  *
  10  * FFmpeg is free software; you can redistribute it and/or
  11  * modify it under the terms of the GNU Lesser General Public
  12  * License as published by the Free Software Foundation; either
  13  * version 2.1 of the License, or (at your option) any later version.
  14  *
  15  * FFmpeg is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18  * Lesser General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU Lesser General Public
  21  * License along with FFmpeg; if not, write to the Free Software
  22  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  23  */
  24
  25 #include "config_components.h"
  26
  27 #include "libavutil/mem.h"
  28 #include "libavutil/opt.h"
  29
  30 #include "avcodec.h"
  31 #include "put_bits.h"
  32 #include "bytestream.h"
  33 #include "adpcm.h"
  34 #include "adpcm_data.h"
  35 #include "codec_internal.h"
  36 #include "encode.h"
  37
  38 /**
  39  * @file
  40  * ADPCM encoders
  41  * See ADPCM decoder reference documents for codec information.
  42  */
  43
  44 #define CASE_0(codec_id, ...)
  45 #define CASE_1(codec_id, ...) \
  46     case codec_id:            \
  47     { __VA_ARGS__ }           \
  48     break;
  49 #define CASE_2(enabled, codec_id, ...) \
  50         CASE_ ## enabled(codec_id, __VA_ARGS__)
  51 #define CASE_3(config, codec_id, ...) \
  52         CASE_2(config, codec_id, __VA_ARGS__)
  53 #define CASE(codec, ...) \
  54         CASE_3(CONFIG_ ## codec ## _ENCODER, AV_CODEC_ID_ ## codec, __VA_ARGS__)
  55
  56 typedef struct TrellisPath {
  57     int nibble;
  58     int prev;
  59 } TrellisPath;
  60
  61 typedef struct TrellisNode {
  62     uint32_t ssd;
  63     int path;
  64     int sample1;
  65     int sample2;
  66     int step;
  67 } TrellisNode;
  68
  69 typedef struct ADPCMEncodeContext {
  70     AVClass *class;
  71     int block_size;
  72
  73     ADPCMChannelStatus status[6];
  74     TrellisPath *paths;
  75     TrellisNode *node_buf;
  76     TrellisNode **nodep_buf;
  77     uint8_t *trellis_hash;
  78 } ADPCMEncodeContext;
  79
  80 #define FREEZE_INTERVAL 128
  81
  82 static av_cold int adpcm_encode_init(AVCodecContext *avctx)
  83 {
  84     ADPCMEncodeContext *s = avctx->priv_data;
  85     int channels = avctx->ch_layout.nb_channels;
  86
  87     /*
  88      * AMV's block size has to match that of the corresponding video
  89      * stream. Relax the POT requirement.
  90      */
  91     if (avctx->codec->id != AV_CODEC_ID_ADPCM_IMA_AMV &&
  92         (s->block_size & (s->block_size - 1))) {
  93         av_log(avctx, AV_LOG_ERROR, "block size must be power of 2\n");
  94         return AVERROR(EINVAL);
  95     }
  96
  97     if (avctx->trellis) {
  98         int frontier, max_paths;
  99
 100         if ((unsigned)avctx->trellis > 16U) {
 101             av_log(avctx, AV_LOG_ERROR, "invalid trellis size\n");
 102             return AVERROR(EINVAL);
 103         }
 104
 105         if (avctx->codec->id == AV_CODEC_ID_ADPCM_IMA_SSI ||
 106             avctx->codec->id == AV_CODEC_ID_ADPCM_IMA_APM ||
 107             avctx->codec->id == AV_CODEC_ID_ADPCM_ARGO    ||
 108             avctx->codec->id == AV_CODEC_ID_ADPCM_IMA_WS) {
 109             /*
 110              * The current trellis implementation doesn't work for extended
 111              * runs of samples without periodic resets. Disallow it.
 112              */
 113             av_log(avctx, AV_LOG_ERROR, "trellis not supported\n");
 114             return AVERROR_PATCHWELCOME;
 115         }
 116
 117         frontier  = 1 << avctx->trellis;
 118         max_paths =  frontier * FREEZE_INTERVAL;
 119         if (!FF_ALLOC_TYPED_ARRAY(s->paths,        max_paths)    ||
 120             !FF_ALLOC_TYPED_ARRAY(s->node_buf,     2 * frontier) ||
 121             !FF_ALLOC_TYPED_ARRAY(s->nodep_buf,    2 * frontier) ||
 122             !FF_ALLOC_TYPED_ARRAY(s->trellis_hash, 65536))
 123             return AVERROR(ENOMEM);
 124     }
 125
 126     avctx->bits_per_coded_sample = av_get_bits_per_sample(avctx->codec->id);
 127
 128     switch (avctx->codec->id) {
 129     CASE(ADPCM_IMA_WAV,
 130         /* each 16 bits sample gives one nibble
 131            and we have 4 bytes per channel overhead */
 132         avctx->frame_size = (s->block_size - 4 * channels) * 8 /
 133                             (4 * channels) + 1;
 134         /* seems frame_size isn't taken into account...
 135            have to buffer the samples :-( */
 136         avctx->block_align = s->block_size;
 137         avctx->bits_per_coded_sample = 4;
 138         ) /* End of CASE */
 139     CASE(ADPCM_IMA_QT,
 140         avctx->frame_size  = 64;
 141         avctx->block_align = 34 * channels;
 142         ) /* End of CASE */
 143     CASE(ADPCM_MS,
 144         uint8_t *extradata;
 145         /* each 16 bits sample gives one nibble
 146            and we have 7 bytes per channel overhead */
 147         avctx->frame_size = (s->block_size - 7 * channels) * 2 / channels + 2;
 148         avctx->bits_per_coded_sample = 4;
 149         avctx->block_align     = s->block_size;
 150         if (!(avctx->extradata = av_malloc(32 + AV_INPUT_BUFFER_PADDING_SIZE)))
 151             return AVERROR(ENOMEM);
 152         avctx->extradata_size = 32;
 153         extradata = avctx->extradata;
 154         bytestream_put_le16(&extradata, avctx->frame_size);
 155         bytestream_put_le16(&extradata, 7); /* wNumCoef */
 156         for (int i = 0; i < 7; i++) {
 157             bytestream_put_le16(&extradata, ff_adpcm_AdaptCoeff1[i] * 4);
 158             bytestream_put_le16(&extradata, ff_adpcm_AdaptCoeff2[i] * 4);
 159         }
 160         ) /* End of CASE */
 161     CASE(ADPCM_YAMAHA,
 162         avctx->frame_size  = s->block_size * 2 / channels;
 163         avctx->block_align = s->block_size;
 164         ) /* End of CASE */
 165     CASE(ADPCM_SWF,
 166         if (avctx->sample_rate != 11025 &&
 167             avctx->sample_rate != 22050 &&
 168             avctx->sample_rate != 44100) {
 169             av_log(avctx, AV_LOG_ERROR, "Sample rate must be 11025, "
 170                    "22050 or 44100\n");
 171             return AVERROR(EINVAL);
 172         }
 173         avctx->frame_size  = 4096; /* Hardcoded according to the SWF spec. */
 174         avctx->block_align = (2 + channels * (22 + 4 * (avctx->frame_size - 1)) + 7) / 8;
 175         ) /* End of CASE */
 176     case AV_CODEC_ID_ADPCM_IMA_SSI:
 177     case AV_CODEC_ID_ADPCM_IMA_ALP:
 178         avctx->frame_size  = s->block_size * 2 / channels;
 179         avctx->block_align = s->block_size;
 180         break;
 181     CASE(ADPCM_IMA_AMV,
 182         if (avctx->sample_rate != 22050) {
 183             av_log(avctx, AV_LOG_ERROR, "Sample rate must be 22050\n");
 184             return AVERROR(EINVAL);
 185         }
 186
 187         if (channels != 1) {
 188             av_log(avctx, AV_LOG_ERROR, "Only mono is supported\n");
 189             return AVERROR(EINVAL);
 190         }
 191
 192         avctx->frame_size  = s->block_size;
 193         avctx->block_align = 8 + (FFALIGN(avctx->frame_size, 2) / 2);
 194         ) /* End of CASE */
 195     CASE(ADPCM_IMA_APM,
 196         avctx->frame_size  = s->block_size * 2 / channels;
 197         avctx->block_align = s->block_size;
 198
 199         if (!(avctx->extradata = av_mallocz(28 + AV_INPUT_BUFFER_PADDING_SIZE)))
 200             return AVERROR(ENOMEM);
 201         avctx->extradata_size = 28;
 202         ) /* End of CASE */
 203     CASE(ADPCM_ARGO,
 204         avctx->frame_size = 32;
 205         avctx->block_align = 17 * channels;
 206         ) /* End of CASE */
 207     CASE(ADPCM_IMA_WS,
 208         /* each 16 bits sample gives one nibble */
 209         avctx->frame_size = s->block_size * 2 / channels;
 210         avctx->block_align = s->block_size;
 211         ) /* End of CASE */
 212     default:
 213         return AVERROR(EINVAL);
 214     }
 215
 216     return 0;
 217 }
 218
 219 static av_cold int adpcm_encode_close(AVCodecContext *avctx)
 220 {
 221     ADPCMEncodeContext *s = avctx->priv_data;
 222     av_freep(&s->paths);
 223     av_freep(&s->node_buf);
 224     av_freep(&s->nodep_buf);
 225     av_freep(&s->trellis_hash);
 226
 227     return 0;
 228 }
 229
 230
 231 static inline uint8_t adpcm_ima_compress_sample(ADPCMChannelStatus *c,
 232                                                 int16_t sample)
 233 {
 234     int delta  = sample - c->prev_sample;
 235     int nibble = FFMIN(7, abs(delta) * 4 /
 236                        ff_adpcm_step_table[c->step_index]) + (delta < 0) * 8;
 237     c->prev_sample += ((ff_adpcm_step_table[c->step_index] *
 238                         ff_adpcm_yamaha_difflookup[nibble]) / 8);
 239     c->prev_sample = av_clip_int16(c->prev_sample);
 240     c->step_index  = av_clip(c->step_index + ff_adpcm_index_table[nibble], 0, 88);
 241     return nibble;
 242 }
 243
 244 static inline uint8_t adpcm_ima_alp_compress_sample(ADPCMChannelStatus *c, int16_t sample)
 245 {
 246     const int delta  = sample - c->prev_sample;
 247     const int step   = ff_adpcm_step_table[c->step_index];
 248     const int sign   = (delta < 0) * 8;
 249
 250     int nibble = FFMIN(abs(delta) * 4 / step, 7);
 251     int diff   = (step * nibble) >> 2;
 252     if (sign)
 253         diff = -diff;
 254
 255     nibble = sign | nibble;
 256
 257     c->prev_sample += diff;
 258     c->prev_sample  = av_clip_int16(c->prev_sample);
 259     c->step_index   = av_clip(c->step_index + ff_adpcm_index_table[nibble], 0, 88);
 260     return nibble;
 261 }
 262
 263 static inline uint8_t adpcm_ima_qt_compress_sample(ADPCMChannelStatus *c,
 264                                                    int16_t sample)
 265 {
 266     int delta  = sample - c->prev_sample;
 267     int diff, step = ff_adpcm_step_table[c->step_index];
 268     int nibble = 8*(delta < 0);
 269
 270     delta= abs(delta);
 271     diff = delta + (step >> 3);
 272
 273     if (delta >= step) {
 274         nibble |= 4;
 275         delta  -= step;
 276     }
 277     step >>= 1;
 278     if (delta >= step) {
 279         nibble |= 2;
 280         delta  -= step;
 281     }
 282     step >>= 1;
 283     if (delta >= step) {
 284         nibble |= 1;
 285         delta  -= step;
 286     }
 287     diff -= delta;
 288
 289     if (nibble & 8)
 290         c->prev_sample -= diff;
 291     else
 292         c->prev_sample += diff;
 293
 294     c->prev_sample = av_clip_int16(c->prev_sample);
 295     c->step_index  = av_clip(c->step_index + ff_adpcm_index_table[nibble], 0, 88);
 296
 297     return nibble;
 298 }
 299
 300 static inline uint8_t adpcm_ms_compress_sample(ADPCMChannelStatus *c,
 301                                                int16_t sample)
 302 {
 303     int predictor, nibble, bias;
 304
 305     predictor = (((c->sample1) * (c->coeff1)) +
 306                 (( c->sample2) * (c->coeff2))) / 64;
 307
 308     nibble = sample - predictor;
 309     if (nibble >= 0)
 310         bias =  c->idelta / 2;
 311     else
 312         bias = -c->idelta / 2;
 313
 314     nibble = (nibble + bias) / c->idelta;
 315     nibble = av_clip_intp2(nibble, 3) & 0x0F;
 316
 317     predictor += ((nibble & 0x08) ? (nibble - 0x10) : nibble) * c->idelta;
 318
 319     c->sample2 = c->sample1;
 320     c->sample1 = av_clip_int16(predictor);
 321
 322     c->idelta = (ff_adpcm_AdaptationTable[nibble] * c->idelta) >> 8;
 323     if (c->idelta < 16)
 324         c->idelta = 16;
 325
 326     return nibble;
 327 }
 328
 329 static inline uint8_t adpcm_yamaha_compress_sample(ADPCMChannelStatus *c,
 330                                                    int16_t sample)
 331 {
 332     int nibble, delta;
 333
 334     if (!c->step) {
 335         c->predictor = 0;
 336         c->step      = 127;
 337     }
 338
 339     delta = sample - c->predictor;
 340
 341     nibble = FFMIN(7, abs(delta) * 4 / c->step) + (delta < 0) * 8;
 342
 343     c->predictor += ((c->step * ff_adpcm_yamaha_difflookup[nibble]) / 8);
 344     c->predictor = av_clip_int16(c->predictor);
 345     c->step = (c->step * ff_adpcm_yamaha_indexscale[nibble]) >> 8;
 346     c->step = av_clip(c->step, 127, 24576);
 347
 348     return nibble;
 349 }
 350
 351 static void adpcm_compress_trellis(AVCodecContext *avctx,
 352                                    const int16_t *samples, uint8_t *dst,
 353                                    ADPCMChannelStatus *c, int n, int stride)
 354 {
 355     //FIXME 6% faster if frontier is a compile-time constant
 356     ADPCMEncodeContext *s = avctx->priv_data;
 357     const int frontier = 1 << avctx->trellis;
 358     const int version  = avctx->codec->id;
 359     TrellisPath *paths       = s->paths, *p;
 360     TrellisNode *node_buf    = s->node_buf;
 361     TrellisNode **nodep_buf  = s->nodep_buf;
 362     TrellisNode **nodes      = nodep_buf; // nodes[] is always sorted by .ssd
 363     TrellisNode **nodes_next = nodep_buf + frontier;
 364     int pathn = 0, froze = -1, i, j, k, generation = 0;
 365     uint8_t *hash = s->trellis_hash;
 366     memset(hash, 0xff, 65536 * sizeof(*hash));
 367
 368     memset(nodep_buf, 0, 2 * frontier * sizeof(*nodep_buf));
 369     nodes[0]          = node_buf + frontier;
 370     nodes[0]->ssd     = 0;
 371     nodes[0]->path    = 0;
 372     nodes[0]->step    = c->step_index;
 373     nodes[0]->sample1 = c->sample1;
 374     nodes[0]->sample2 = c->sample2;
 375     if (version == AV_CODEC_ID_ADPCM_IMA_WAV ||
 376         version == AV_CODEC_ID_ADPCM_IMA_QT  ||
 377         version == AV_CODEC_ID_ADPCM_IMA_AMV ||
 378         version == AV_CODEC_ID_ADPCM_SWF)
 379         nodes[0]->sample1 = c->prev_sample;
 380     if (version == AV_CODEC_ID_ADPCM_MS)
 381         nodes[0]->step = c->idelta;
 382     if (version == AV_CODEC_ID_ADPCM_YAMAHA) {
 383         if (c->step == 0) {
 384             nodes[0]->step    = 127;
 385             nodes[0]->sample1 = 0;
 386         } else {
 387             nodes[0]->step    = c->step;
 388             nodes[0]->sample1 = c->predictor;
 389         }
 390     }
 391
 392     for (i = 0; i < n; i++) {
 393         TrellisNode *t = node_buf + frontier*(i&1);
 394         TrellisNode **u;
 395         int sample   = samples[i * stride];
 396         int heap_pos = 0;
 397         memset(nodes_next, 0, frontier * sizeof(TrellisNode*));
 398         for (j = 0; j < frontier && nodes[j]; j++) {
 399             // higher j have higher ssd already, so they're likely
 400             // to yield a suboptimal next sample too
 401             const int range = (j < frontier / 2) ? 1 : 0;
 402             const int step  = nodes[j]->step;
 403             int nidx;
 404             if (version == AV_CODEC_ID_ADPCM_MS) {
 405                 const int predictor = ((nodes[j]->sample1 * c->coeff1) +
 406                                        (nodes[j]->sample2 * c->coeff2)) / 64;
 407                 const int div  = (sample - predictor) / step;
 408                 const int nmin = av_clip(div-range, -8, 6);
 409                 const int nmax = av_clip(div+range, -7, 7);
 410                 for (nidx = nmin; nidx <= nmax; nidx++) {
 411                     const int nibble = nidx & 0xf;
 412                     int dec_sample   = predictor + nidx * step;
 413 #define STORE_NODE(NAME, STEP_INDEX)\
 414                     int d;\
 415                     uint32_t ssd;\
 416                     int pos;\
 417                     TrellisNode *u;\
 418                     uint8_t *h;\
 419                     dec_sample = av_clip_int16(dec_sample);\
 420                     d = sample - dec_sample;\
 421                     ssd = nodes[j]->ssd + d*(unsigned)d;\
 422                     /* Check for wraparound, skip such samples completely. \
 423                      * Note, changing ssd to a 64 bit variable would be \
 424                      * simpler, avoiding this check, but it's slower on \
 425                      * x86 32 bit at the moment. */\
 426                     if (ssd < nodes[j]->ssd)\
 427                         goto next_##NAME;\
 428                     /* Collapse any two states with the same previous sample value. \
 429                      * One could also distinguish states by step and by 2nd to last
 430                      * sample, but the effects of that are negligible.
 431                      * Since nodes in the previous generation are iterated
 432                      * through a heap, they're roughly ordered from better to
 433                      * worse, but not strictly ordered. Therefore, an earlier
 434                      * node with the same sample value is better in most cases
 435                      * (and thus the current is skipped), but not strictly
 436                      * in all cases. Only skipping samples where ssd >=
 437                      * ssd of the earlier node with the same sample gives
 438                      * slightly worse quality, though, for some reason. */ \
 439                     h = &hash[(uint16_t) dec_sample];\
 440                     if (*h == generation)\
 441                         goto next_##NAME;\
 442                     if (heap_pos < frontier) {\
 443                         pos = heap_pos++;\
 444                     } else {\
 445                         /* Try to replace one of the leaf nodes with the new \
 446                          * one, but try a different slot each time. */\
 447                         pos = (frontier >> 1) +\
 448                               (heap_pos & ((frontier >> 1) - 1));\
 449                         if (ssd > nodes_next[pos]->ssd)\
 450                             goto next_##NAME;\
 451                         heap_pos++;\
 452                     }\
 453                     *h = generation;\
 454                     u  = nodes_next[pos];\
 455                     if (!u) {\
 456                         av_assert1(pathn < FREEZE_INTERVAL << avctx->trellis);\
 457                         u = t++;\
 458                         nodes_next[pos] = u;\
 459                         u->path = pathn++;\
 460                     }\
 461                     u->ssd  = ssd;\
 462                     u->step = STEP_INDEX;\
 463                     u->sample2 = nodes[j]->sample1;\
 464                     u->sample1 = dec_sample;\
 465                     paths[u->path].nibble = nibble;\
 466                     paths[u->path].prev   = nodes[j]->path;\
 467                     /* Sift the newly inserted node up in the heap to \
 468                      * restore the heap property. */\
 469                     while (pos > 0) {\
 470                         int parent = (pos - 1) >> 1;\
 471                         if (nodes_next[parent]->ssd <= ssd)\
 472                             break;\
 473                         FFSWAP(TrellisNode*, nodes_next[parent], nodes_next[pos]);\
 474                         pos = parent;\
 475                     }\
 476                     next_##NAME:;
 477                     STORE_NODE(ms, FFMAX(16,
 478                                (ff_adpcm_AdaptationTable[nibble] * step) >> 8));
 479                 }
 480             } else if (version == AV_CODEC_ID_ADPCM_IMA_WAV ||
 481                        version == AV_CODEC_ID_ADPCM_IMA_QT  ||
 482                        version == AV_CODEC_ID_ADPCM_IMA_AMV ||
 483                        version == AV_CODEC_ID_ADPCM_SWF) {
 484 #define LOOP_NODES(NAME, STEP_TABLE, STEP_INDEX)\
 485                 const int predictor = nodes[j]->sample1;\
 486                 const int div = (sample - predictor) * 4 / STEP_TABLE;\
 487                 int nmin = av_clip(div - range, -7, 6);\
 488                 int nmax = av_clip(div + range, -6, 7);\
 489                 if (nmin <= 0)\
 490                     nmin--; /* distinguish -0 from +0 */\
 491                 if (nmax < 0)\
 492                     nmax--;\
 493                 for (nidx = nmin; nidx <= nmax; nidx++) {\
 494                     const int nibble = nidx < 0 ? 7 - nidx : nidx;\
 495                     int dec_sample = predictor +\
 496                                     (STEP_TABLE *\
 497                                      ff_adpcm_yamaha_difflookup[nibble]) / 8;\
 498                     STORE_NODE(NAME, STEP_INDEX);\
 499                 }
 500                 LOOP_NODES(ima, ff_adpcm_step_table[step],
 501                            av_clip(step + ff_adpcm_index_table[nibble], 0, 88));
 502             } else { //AV_CODEC_ID_ADPCM_YAMAHA
 503                 LOOP_NODES(yamaha, step,
 504                            av_clip((step * ff_adpcm_yamaha_indexscale[nibble]) >> 8,
 505                                    127, 24576));
 506 #undef LOOP_NODES
 507 #undef STORE_NODE
 508             }
 509         }
 510
 511         u = nodes;
 512         nodes = nodes_next;
 513         nodes_next = u;
 514
 515         generation++;
 516         if (generation == 255) {
 517             memset(hash, 0xff, 65536 * sizeof(*hash));
 518             generation = 0;
 519         }
 520
 521         // prevent overflow
 522         if (nodes[0]->ssd > (1 << 28)) {
 523             for (j = 1; j < frontier && nodes[j]; j++)
 524                 nodes[j]->ssd -= nodes[0]->ssd;
 525             nodes[0]->ssd = 0;
 526         }
 527
 528         // merge old paths to save memory
 529         if (i == froze + FREEZE_INTERVAL) {
 530             p = &paths[nodes[0]->path];
 531             for (k = i; k > froze; k--) {
 532                 dst[k] = p->nibble;
 533                 p = &paths[p->prev];
 534             }
 535             froze = i;
 536             pathn = 0;
 537             // other nodes might use paths that don't coincide with the frozen one.
 538             // checking which nodes do so is too slow, so just kill them all.
 539             // this also slightly improves quality, but I don't know why.
 540             memset(nodes + 1, 0, (frontier - 1) * sizeof(TrellisNode*));
 541         }
 542     }
 543
 544     p = &paths[nodes[0]->path];
 545     for (i = n - 1; i > froze; i--) {
 546         dst[i] = p->nibble;
 547         p = &paths[p->prev];
 548     }
 549
 550     c->predictor  = nodes[0]->sample1;
 551     c->sample1    = nodes[0]->sample1;
 552     c->sample2    = nodes[0]->sample2;
 553     c->step_index = nodes[0]->step;
 554     c->step       = nodes[0]->step;
 555     c->idelta     = nodes[0]->step;
 556 }
 557
 558 #if CONFIG_ADPCM_ARGO_ENCODER
 559 static inline int adpcm_argo_compress_nibble(const ADPCMChannelStatus *cs, int16_t s,
 560                                              int shift, int flag)
 561 {
 562     int nibble;
 563
 564     if (flag)
 565         nibble = 4 * s - 8 * cs->sample1 + 4 * cs->sample2;
 566     else
 567         nibble = 4 * s - 4 * cs->sample1;
 568
 569     return (nibble >> shift) & 0x0F;
 570 }
 571
 572 static int64_t adpcm_argo_compress_block(ADPCMChannelStatus *cs, PutBitContext *pb,
 573                                          const int16_t *samples, int nsamples,
 574                                          int shift, int flag)
 575 {
 576     int64_t error = 0;
 577
 578     if (pb) {
 579         put_bits(pb, 4, shift - 2);
 580         put_bits(pb, 1, 0);
 581         put_bits(pb, 1, !!flag);
 582         put_bits(pb, 2, 0);
 583     }
 584
 585     for (int n = 0; n < nsamples; n++) {
 586         /* Compress the nibble, then expand it to see how much precision we've lost. */
 587         int nibble = adpcm_argo_compress_nibble(cs, samples[n], shift, flag);
 588         int16_t sample = ff_adpcm_argo_expand_nibble(cs, nibble, shift, flag);
 589
 590         error += abs(samples[n] - sample);
 591
 592         if (pb)
 593             put_bits(pb, 4, nibble);
 594     }
 595
 596     return error;
 597 }
 598 #endif
 599
 600 static int adpcm_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
 601                               const AVFrame *frame, int *got_packet_ptr)
 602 {
 603     int st, pkt_size, ret;
 604     const int16_t *samples;
 605     const int16_t *const *samples_p;
 606     uint8_t *dst;
 607     ADPCMEncodeContext *c = avctx->priv_data;
 608     int channels = avctx->ch_layout.nb_channels;
 609
 610     samples = (const int16_t *)frame->data[0];
 611     samples_p = (const int16_t *const *)frame->extended_data;
 612     st = channels == 2;
 613
 614     if (avctx->codec_id == AV_CODEC_ID_ADPCM_IMA_SSI ||
 615         avctx->codec_id == AV_CODEC_ID_ADPCM_IMA_ALP ||
 616         avctx->codec_id == AV_CODEC_ID_ADPCM_IMA_APM ||
 617         avctx->codec_id == AV_CODEC_ID_ADPCM_IMA_WS)
 618         pkt_size = (frame->nb_samples * channels + 1) / 2;
 619     else
 620         pkt_size = avctx->block_align;
 621     if ((ret = ff_get_encode_buffer(avctx, avpkt, pkt_size, 0)) < 0)
 622         return ret;
 623     dst = avpkt->data;
 624
 625     switch(avctx->codec->id) {
 626     CASE(ADPCM_IMA_WAV,
 627         int blocks = (frame->nb_samples - 1) / 8;
 628
 629         for (int ch = 0; ch < channels; ch++) {
 630             ADPCMChannelStatus *status = &c->status[ch];
 631             status->prev_sample = samples_p[ch][0];
 632             /* status->step_index = 0;
 633                XXX: not sure how to init the state machine */
 634             bytestream_put_le16(&dst, status->prev_sample);
 635             *dst++ = status->step_index;
 636             *dst++ = 0; /* unknown */
 637         }
 638
 639         /* stereo: 4 bytes (8 samples) for left, 4 bytes for right */
 640         if (avctx->trellis > 0) {
 641             uint8_t *buf;
 642             if (!FF_ALLOC_TYPED_ARRAY(buf, channels * blocks * 8))
 643                 return AVERROR(ENOMEM);
 644             for (int ch = 0; ch < channels; ch++) {
 645                 adpcm_compress_trellis(avctx, &samples_p[ch][1],
 646                                        buf + ch * blocks * 8, &c->status[ch],
 647                                        blocks * 8, 1);
 648             }
 649             for (int i = 0; i < blocks; i++) {
 650                 for (int ch = 0; ch < channels; ch++) {
 651                     uint8_t *buf1 = buf + ch * blocks * 8 + i * 8;
 652                     for (int j = 0; j < 8; j += 2)
 653                         *dst++ = buf1[j] | (buf1[j + 1] << 4);
 654                 }
 655             }
 656             av_free(buf);
 657         } else {
 658             for (int i = 0; i < blocks; i++) {
 659                 for (int ch = 0; ch < channels; ch++) {
 660                     ADPCMChannelStatus *status = &c->status[ch];
 661                     const int16_t *smp = &samples_p[ch][1 + i * 8];
 662                     for (int j = 0; j < 8; j += 2) {
 663                         uint8_t v = adpcm_ima_compress_sample(status, smp[j    ]);
 664                         v        |= adpcm_ima_compress_sample(status, smp[j + 1]) << 4;
 665                         *dst++ = v;
 666                     }
 667                 }
 668             }
 669         }
 670         ) /* End of CASE */
 671     CASE(ADPCM_IMA_QT,
 672         PutBitContext pb;
 673         init_put_bits(&pb, dst, pkt_size);
 674
 675         for (int ch = 0; ch < channels; ch++) {
 676             ADPCMChannelStatus *status = &c->status[ch];
 677             put_bits(&pb, 9, (status->prev_sample & 0xFFFF) >> 7);
 678             put_bits(&pb, 7,  status->step_index);
 679             if (avctx->trellis > 0) {
 680                 uint8_t buf[64];
 681                 adpcm_compress_trellis(avctx, &samples_p[ch][0], buf, status,
 682                                        64, 1);
 683                 for (int i = 0; i < 64; i++)
 684                     put_bits(&pb, 4, buf[i ^ 1]);
 685                 status->prev_sample = status->predictor;
 686             } else {
 687                 for (int i = 0; i < 64; i += 2) {
 688                     int t1, t2;
 689                     t1 = adpcm_ima_qt_compress_sample(status, samples_p[ch][i    ]);
 690                     t2 = adpcm_ima_qt_compress_sample(status, samples_p[ch][i + 1]);
 691                     put_bits(&pb, 4, t2);
 692                     put_bits(&pb, 4, t1);
 693                 }
 694             }
 695         }
 696
 697         flush_put_bits(&pb);
 698         ) /* End of CASE */
 699     CASE(ADPCM_IMA_SSI,
 700         PutBitContext pb;
 701         init_put_bits(&pb, dst, pkt_size);
 702
 703         av_assert0(avctx->trellis == 0);
 704
 705         for (int i = 0; i < frame->nb_samples; i++) {
 706             for (int ch = 0; ch < channels; ch++) {
 707                 put_bits(&pb, 4, adpcm_ima_qt_compress_sample(c->status + ch, *samples++));
 708             }
 709         }
 710
 711         flush_put_bits(&pb);
 712         ) /* End of CASE */
 713     CASE(ADPCM_IMA_ALP,
 714         PutBitContext pb;
 715         init_put_bits(&pb, dst, pkt_size);
 716
 717         av_assert0(avctx->trellis == 0);
 718
 719         for (int n = frame->nb_samples / 2; n > 0; n--) {
 720             for (int ch = 0; ch < channels; ch++) {
 721                 put_bits(&pb, 4, adpcm_ima_alp_compress_sample(c->status + ch, *samples++));
 722                 put_bits(&pb, 4, adpcm_ima_alp_compress_sample(c->status + ch, samples[st]));
 723             }
 724             samples += channels;
 725         }
 726
 727         flush_put_bits(&pb);
 728         ) /* End of CASE */
 729     CASE(ADPCM_SWF,
 730         const int n = frame->nb_samples - 1;
 731         PutBitContext pb;
 732         init_put_bits(&pb, dst, pkt_size);
 733
 734         /* NB: This is safe as we don't have AV_CODEC_CAP_SMALL_LAST_FRAME. */
 735         av_assert0(n == 4095);
 736
 737         // store AdpcmCodeSize
 738         put_bits(&pb, 2, 2);    // set 4-bit flash adpcm format
 739
 740         // init the encoder state
 741         for (int i = 0; i < channels; i++) {
 742             // clip step so it fits 6 bits
 743             c->status[i].step_index = av_clip_uintp2(c->status[i].step_index, 6);
 744             put_sbits(&pb, 16, samples[i]);
 745             put_bits(&pb, 6, c->status[i].step_index);
 746             c->status[i].prev_sample = samples[i];
 747         }
 748
 749         if (avctx->trellis > 0) {
 750             uint8_t buf[8190 /* = 2 * n */];
 751             adpcm_compress_trellis(avctx, samples + channels, buf,
 752                                    &c->status[0], n, channels);
 753             if (channels == 2)
 754                 adpcm_compress_trellis(avctx, samples + channels + 1,
 755                                        buf + n, &c->status[1], n,
 756                                        channels);
 757             for (int i = 0; i < n; i++) {
 758                 put_bits(&pb, 4, buf[i]);
 759                 if (channels == 2)
 760                     put_bits(&pb, 4, buf[n + i]);
 761             }
 762         } else {
 763             for (int i = 1; i < frame->nb_samples; i++) {
 764                 put_bits(&pb, 4, adpcm_ima_compress_sample(&c->status[0],
 765                          samples[channels * i]));
 766                 if (channels == 2)
 767                     put_bits(&pb, 4, adpcm_ima_compress_sample(&c->status[1],
 768                              samples[2 * i + 1]));
 769             }
 770         }
 771         flush_put_bits(&pb);
 772         ) /* End of CASE */
 773     CASE(ADPCM_MS,
 774         for (int i = 0; i < channels; i++) {
 775             int predictor = 0;
 776             *dst++ = predictor;
 777             c->status[i].coeff1 = ff_adpcm_AdaptCoeff1[predictor];
 778             c->status[i].coeff2 = ff_adpcm_AdaptCoeff2[predictor];
 779         }
 780         for (int i = 0; i < channels; i++) {
 781             if (c->status[i].idelta < 16)
 782                 c->status[i].idelta = 16;
 783             bytestream_put_le16(&dst, c->status[i].idelta);
 784         }
 785         for (int i = 0; i < channels; i++)
 786             c->status[i].sample2= *samples++;
 787         for (int i = 0; i < channels; i++) {
 788             c->status[i].sample1 = *samples++;
 789             bytestream_put_le16(&dst, c->status[i].sample1);
 790         }
 791         for (int i = 0; i < channels; i++)
 792             bytestream_put_le16(&dst, c->status[i].sample2);
 793
 794         if (avctx->trellis > 0) {
 795             const int n  = avctx->block_align - 7 * channels;
 796             uint8_t *buf = av_malloc(2 * n);
 797             if (!buf)
 798                 return AVERROR(ENOMEM);
 799             if (channels == 1) {
 800                 adpcm_compress_trellis(avctx, samples, buf, &c->status[0], n,
 801                                        channels);
 802                 for (int i = 0; i < n; i += 2)
 803                     *dst++ = (buf[i] << 4) | buf[i + 1];
 804             } else {
 805                 adpcm_compress_trellis(avctx, samples,     buf,
 806                                        &c->status[0], n, channels);
 807                 adpcm_compress_trellis(avctx, samples + 1, buf + n,
 808                                        &c->status[1], n, channels);
 809                 for (int i = 0; i < n; i++)
 810                     *dst++ = (buf[i] << 4) | buf[n + i];
 811             }
 812             av_free(buf);
 813         } else {
 814             for (int i = 7 * channels; i < avctx->block_align; i++) {
 815                 int nibble;
 816                 nibble  = adpcm_ms_compress_sample(&c->status[ 0], *samples++) << 4;
 817                 nibble |= adpcm_ms_compress_sample(&c->status[st], *samples++);
 818                 *dst++  = nibble;
 819             }
 820         }
 821         ) /* End of CASE */
 822     CASE(ADPCM_YAMAHA,
 823         int n = frame->nb_samples / 2;
 824         if (avctx->trellis > 0) {
 825             uint8_t *buf = av_malloc(2 * n * 2);
 826             if (!buf)
 827                 return AVERROR(ENOMEM);
 828             n *= 2;
 829             if (channels == 1) {
 830                 adpcm_compress_trellis(avctx, samples, buf, &c->status[0], n,
 831                                        channels);
 832                 for (int i = 0; i < n; i += 2)
 833                     *dst++ = buf[i] | (buf[i + 1] << 4);
 834             } else {
 835                 adpcm_compress_trellis(avctx, samples,     buf,
 836                                        &c->status[0], n, channels);
 837                 adpcm_compress_trellis(avctx, samples + 1, buf + n,
 838                                        &c->status[1], n, channels);
 839                 for (int i = 0; i < n; i++)
 840                     *dst++ = buf[i] | (buf[n + i] << 4);
 841             }
 842             av_free(buf);
 843         } else
 844             for (n *= channels; n > 0; n--) {
 845                 int nibble;
 846                 nibble  = adpcm_yamaha_compress_sample(&c->status[ 0], *samples++);
 847                 nibble |= adpcm_yamaha_compress_sample(&c->status[st], *samples++) << 4;
 848                 *dst++  = nibble;
 849             }
 850         ) /* End of CASE */
 851     CASE(ADPCM_IMA_APM,
 852         PutBitContext pb;
 853         init_put_bits(&pb, dst, pkt_size);
 854
 855         av_assert0(avctx->trellis == 0);
 856
 857         for (int n = frame->nb_samples / 2; n > 0; n--) {
 858             for (int ch = 0; ch < channels; ch++) {
 859                 put_bits(&pb, 4, adpcm_ima_qt_compress_sample(c->status + ch, *samples++));
 860                 put_bits(&pb, 4, adpcm_ima_qt_compress_sample(c->status + ch, samples[st]));
 861             }
 862             samples += channels;
 863         }
 864
 865         flush_put_bits(&pb);
 866         ) /* End of CASE */
 867     CASE(ADPCM_IMA_AMV,
 868         av_assert0(channels == 1);
 869
 870         c->status[0].prev_sample = *samples;
 871         bytestream_put_le16(&dst, c->status[0].prev_sample);
 872         bytestream_put_byte(&dst, c->status[0].step_index);
 873         bytestream_put_byte(&dst, 0);
 874         bytestream_put_le32(&dst, avctx->frame_size);
 875
 876         if (avctx->trellis > 0) {
 877             const int n  = frame->nb_samples >> 1;
 878             uint8_t *buf = av_malloc(2 * n);
 879
 880             if (!buf)
 881                 return AVERROR(ENOMEM);
 882
 883             adpcm_compress_trellis(avctx, samples, buf, &c->status[0], 2 * n, channels);
 884             for (int i = 0; i < n; i++)
 885                 bytestream_put_byte(&dst, (buf[2 * i] << 4) | buf[2 * i + 1]);
 886
 887             samples += 2 * n;
 888             av_free(buf);
 889         } else for (int n = frame->nb_samples >> 1; n > 0; n--) {
 890             int nibble;
 891             nibble  = adpcm_ima_compress_sample(&c->status[0], *samples++) << 4;
 892             nibble |= adpcm_ima_compress_sample(&c->status[0], *samples++) & 0x0F;
 893             bytestream_put_byte(&dst, nibble);
 894         }
 895
 896         if (avctx->frame_size & 1) {
 897             int nibble = adpcm_ima_compress_sample(&c->status[0], *samples++) << 4;
 898             bytestream_put_byte(&dst, nibble);
 899         }
 900         ) /* End of CASE */
 901     CASE(ADPCM_ARGO,
 902         PutBitContext pb;
 903         init_put_bits(&pb, dst, pkt_size);
 904
 905         av_assert0(frame->nb_samples == 32);
 906
 907         for (int ch = 0; ch < channels; ch++) {
 908             int64_t error  = INT64_MAX, tmperr = INT64_MAX;
 909             int     shift  = 2, flag = 0;
 910             int     saved1 = c->status[ch].sample1;
 911             int     saved2 = c->status[ch].sample2;
 912
 913             /* Find the optimal coefficients, bail early if we find a perfect result. */
 914             for (int s = 2; s < 18 && tmperr != 0; s++) {
 915                 for (int f = 0; f < 2 && tmperr != 0; f++) {
 916                     c->status[ch].sample1 = saved1;
 917                     c->status[ch].sample2 = saved2;
 918                     tmperr = adpcm_argo_compress_block(c->status + ch, NULL, samples_p[ch],
 919                                                        frame->nb_samples, s, f);
 920                     if (tmperr < error) {
 921                         shift = s;
 922                         flag  = f;
 923                         error = tmperr;
 924                     }
 925                 }
 926             }
 927
 928             /* Now actually do the encode. */
 929             c->status[ch].sample1 = saved1;
 930             c->status[ch].sample2 = saved2;
 931             adpcm_argo_compress_block(c->status + ch, &pb, samples_p[ch],
 932                                       frame->nb_samples, shift, flag);
 933         }
 934
 935         flush_put_bits(&pb);
 936         ) /* End of CASE */
 937     CASE(ADPCM_IMA_WS,
 938         PutBitContext pb;
 939         init_put_bits(&pb, dst, pkt_size);
 940
 941         av_assert0(avctx->trellis == 0);
 942         for (int n = frame->nb_samples / 2; n > 0; n--) {
 943             /* stereo: 1 byte (2 samples) for left, 1 byte for right */
 944             for (int ch = 0; ch < channels; ch++) {
 945                 int t1, t2;
 946                 t1 = adpcm_ima_compress_sample(&c->status[ch], *samples++);
 947                 t2 = adpcm_ima_compress_sample(&c->status[ch], samples[st]);
 948                 put_bits(&pb, 4, t2);
 949                 put_bits(&pb, 4, t1);
 950             }
 951             samples += channels;
 952         }
 953         flush_put_bits(&pb);
 954         ) /* End of CASE */
 955     default:
 956         return AVERROR(EINVAL);
 957     }
 958
 959     *got_packet_ptr = 1;
 960     return 0;
 961 }
 962
 963 static const enum AVSampleFormat sample_fmts[] = {
 964     AV_SAMPLE_FMT_S16, AV_SAMPLE_FMT_NONE
 965 };
 966
 967 static const enum AVSampleFormat sample_fmts_p[] = {
 968     AV_SAMPLE_FMT_S16P, AV_SAMPLE_FMT_NONE
 969 };
 970
 971 static const AVChannelLayout ch_layouts[] = {
 972     AV_CHANNEL_LAYOUT_MONO,
 973     AV_CHANNEL_LAYOUT_STEREO,
 974     { 0 },
 975 };
 976
 977 static const AVOption options[] = {
 978     {
 979         .name        = "block_size",
 980         .help        = "set the block size",
 981         .offset      = offsetof(ADPCMEncodeContext, block_size),
 982         .type        = AV_OPT_TYPE_INT,
 983         .default_val = {.i64 = 1024},
 984         .min         = 32,
 985         .max         = 8192, /* Is this a reasonable upper limit? */
 986         .flags       = AV_OPT_FLAG_ENCODING_PARAM | AV_OPT_FLAG_AUDIO_PARAM
 987     },
 988     { NULL }
 989 };
 990
 991 static const AVClass adpcm_encoder_class = {
 992     .class_name = "ADPCM encoder",
 993     .item_name  = av_default_item_name,
 994     .option     = options,
 995     .version    = LIBAVUTIL_VERSION_INT,
 996 };
 997
 998 #define ADPCM_ENCODER_0(id_, name_, sample_fmts_, capabilities_, long_name_)
 999 #define ADPCM_ENCODER_1(id_, name_, sample_fmts_, capabilities_, long_name_) \
1000 const FFCodec ff_ ## name_ ## _encoder = {                                 \
1001     .p.name         = #name_,                                              \
1002     CODEC_LONG_NAME(long_name_),                                           \
1003     .p.type         = AVMEDIA_TYPE_AUDIO,                                  \
1004     .p.id           = id_,                                                 \
1005     .p.sample_fmts  = sample_fmts_,                                        \
1006     .p.ch_layouts   = ch_layouts,                                          \
1007     .p.capabilities = capabilities_ | AV_CODEC_CAP_DR1 |                   \
1008                       AV_CODEC_CAP_ENCODER_REORDERED_OPAQUE,               \
1009     .p.priv_class   = &adpcm_encoder_class,                                \
1010     .priv_data_size = sizeof(ADPCMEncodeContext),                          \
1011     .init           = adpcm_encode_init,                                   \
1012     FF_CODEC_ENCODE_CB(adpcm_encode_frame),                                \
1013     .close          = adpcm_encode_close,                                  \
1014     .caps_internal  = FF_CODEC_CAP_INIT_CLEANUP,                           \
1015 };
1016 #define ADPCM_ENCODER_2(enabled, codec_id, name, sample_fmts, capabilities, long_name) \
1017     ADPCM_ENCODER_ ## enabled(codec_id, name, sample_fmts, capabilities, long_name)
1018 #define ADPCM_ENCODER_3(config, codec_id, name, sample_fmts, capabilities, long_name) \
1019     ADPCM_ENCODER_2(config, codec_id, name, sample_fmts, capabilities, long_name)
1020 #define ADPCM_ENCODER(codec, name, sample_fmts, capabilities, long_name) \
1021     ADPCM_ENCODER_3(CONFIG_ ## codec ## _ENCODER, AV_CODEC_ID_ ## codec, \
1022                     name, sample_fmts, capabilities, long_name)
1023
1024 ADPCM_ENCODER(ADPCM_ARGO,    adpcm_argo,    sample_fmts_p, 0,                             "ADPCM Argonaut Games")
1025 ADPCM_ENCODER(ADPCM_IMA_AMV, adpcm_ima_amv, sample_fmts,   0,                             "ADPCM IMA AMV")
1026 ADPCM_ENCODER(ADPCM_IMA_APM, adpcm_ima_apm, sample_fmts,   AV_CODEC_CAP_SMALL_LAST_FRAME, "ADPCM IMA Ubisoft APM")
1027 ADPCM_ENCODER(ADPCM_IMA_ALP, adpcm_ima_alp, sample_fmts,   AV_CODEC_CAP_SMALL_LAST_FRAME, "ADPCM IMA High Voltage Software ALP")
1028 ADPCM_ENCODER(ADPCM_IMA_QT,  adpcm_ima_qt,  sample_fmts_p, 0,                             "ADPCM IMA QuickTime")
1029 ADPCM_ENCODER(ADPCM_IMA_SSI, adpcm_ima_ssi, sample_fmts,   AV_CODEC_CAP_SMALL_LAST_FRAME, "ADPCM IMA Simon & Schuster Interactive")
1030 ADPCM_ENCODER(ADPCM_IMA_WAV, adpcm_ima_wav, sample_fmts_p, 0,                             "ADPCM IMA WAV")
1031 ADPCM_ENCODER(ADPCM_IMA_WS,  adpcm_ima_ws,  sample_fmts,   AV_CODEC_CAP_SMALL_LAST_FRAME, "ADPCM IMA Westwood")
1032 ADPCM_ENCODER(ADPCM_MS,      adpcm_ms,      sample_fmts,   0,                             "ADPCM Microsoft")
1033 ADPCM_ENCODER(ADPCM_SWF,     adpcm_swf,     sample_fmts,   0,                             "ADPCM Shockwave Flash")
1034 ADPCM_ENCODER(ADPCM_YAMAHA,  adpcm_yamaha,  sample_fmts,   0,                             "ADPCM Yamaha")