3 * Copyright (c) 2006 Justin Ruggles <justin.ruggles@gmail.com>
5 * This file is part of FFmpeg.
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.
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.
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
22 #include "libavutil/avassert.h"
23 #include "libavutil/crc.h"
24 #include "libavutil/intmath.h"
25 #include "libavutil/md5.h"
26 #include "libavutil/opt.h"
37 #define FLAC_SUBFRAME_CONSTANT 0
38 #define FLAC_SUBFRAME_VERBATIM 1
39 #define FLAC_SUBFRAME_FIXED 8
40 #define FLAC_SUBFRAME_LPC 32
42 #define MAX_FIXED_ORDER 4
43 #define MAX_PARTITION_ORDER 8
44 #define MAX_PARTITIONS (1 << MAX_PARTITION_ORDER)
45 #define MAX_LPC_PRECISION 15
46 #define MAX_LPC_SHIFT 15
50 CODING_MODE_RICE2 = 5,
53 typedef struct CompressionOptions {
54 int compression_level;
56 enum FFLPCType lpc_type;
58 int lpc_coeff_precision;
59 int min_prediction_order;
60 int max_prediction_order;
61 int prediction_order_method;
62 int min_partition_order;
63 int max_partition_order;
67 typedef struct RiceContext {
68 enum CodingMode coding_mode;
70 int params[MAX_PARTITIONS];
73 typedef struct FlacSubframe {
79 int32_t coefs[MAX_LPC_ORDER];
82 int32_t samples[FLAC_MAX_BLOCKSIZE];
83 int32_t residual[FLAC_MAX_BLOCKSIZE+11];
86 typedef struct FlacFrame {
87 FlacSubframe subframes[FLAC_MAX_CHANNELS];
95 typedef struct FlacEncodeContext {
105 int max_encoded_framesize;
106 uint32_t frame_count;
107 uint64_t sample_count;
110 CompressionOptions options;
111 AVCodecContext *avctx;
113 struct AVMD5 *md5ctx;
115 unsigned int md5_buffer_size;
116 BswapDSPContext bdsp;
117 FLACDSPContext flac_dsp;
125 * Write streaminfo metadata block to byte array.
127 static void write_streaminfo(FlacEncodeContext *s, uint8_t *header)
131 memset(header, 0, FLAC_STREAMINFO_SIZE);
132 init_put_bits(&pb, header, FLAC_STREAMINFO_SIZE);
134 /* streaminfo metadata block */
135 put_bits(&pb, 16, s->max_blocksize);
136 put_bits(&pb, 16, s->max_blocksize);
137 put_bits(&pb, 24, s->min_framesize);
138 put_bits(&pb, 24, s->max_framesize);
139 put_bits(&pb, 20, s->samplerate);
140 put_bits(&pb, 3, s->channels-1);
141 put_bits(&pb, 5, s->avctx->bits_per_raw_sample - 1);
142 /* write 36-bit sample count in 2 put_bits() calls */
143 put_bits(&pb, 24, (s->sample_count & 0xFFFFFF000LL) >> 12);
144 put_bits(&pb, 12, s->sample_count & 0x000000FFFLL);
146 memcpy(&header[18], s->md5sum, 16);
151 * Set blocksize based on samplerate.
152 * Choose the closest predefined blocksize >= BLOCK_TIME_MS milliseconds.
154 static int select_blocksize(int samplerate, int block_time_ms)
160 av_assert0(samplerate > 0);
161 blocksize = ff_flac_blocksize_table[1];
162 target = (samplerate * block_time_ms) / 1000;
163 for (i = 0; i < 16; i++) {
164 if (target >= ff_flac_blocksize_table[i] &&
165 ff_flac_blocksize_table[i] > blocksize) {
166 blocksize = ff_flac_blocksize_table[i];
173 static av_cold void dprint_compression_options(FlacEncodeContext *s)
175 AVCodecContext *avctx = s->avctx;
176 CompressionOptions *opt = &s->options;
178 av_log(avctx, AV_LOG_DEBUG, " compression: %d\n", opt->compression_level);
180 switch (opt->lpc_type) {
181 case FF_LPC_TYPE_NONE:
182 av_log(avctx, AV_LOG_DEBUG, " lpc type: None\n");
184 case FF_LPC_TYPE_FIXED:
185 av_log(avctx, AV_LOG_DEBUG, " lpc type: Fixed pre-defined coefficients\n");
187 case FF_LPC_TYPE_LEVINSON:
188 av_log(avctx, AV_LOG_DEBUG, " lpc type: Levinson-Durbin recursion with Welch window\n");
190 case FF_LPC_TYPE_CHOLESKY:
191 av_log(avctx, AV_LOG_DEBUG, " lpc type: Cholesky factorization, %d pass%s\n",
192 opt->lpc_passes, opt->lpc_passes == 1 ? "" : "es");
196 av_log(avctx, AV_LOG_DEBUG, " prediction order: %d, %d\n",
197 opt->min_prediction_order, opt->max_prediction_order);
199 switch (opt->prediction_order_method) {
200 case ORDER_METHOD_EST:
201 av_log(avctx, AV_LOG_DEBUG, " order method: %s\n", "estimate");
203 case ORDER_METHOD_2LEVEL:
204 av_log(avctx, AV_LOG_DEBUG, " order method: %s\n", "2-level");
206 case ORDER_METHOD_4LEVEL:
207 av_log(avctx, AV_LOG_DEBUG, " order method: %s\n", "4-level");
209 case ORDER_METHOD_8LEVEL:
210 av_log(avctx, AV_LOG_DEBUG, " order method: %s\n", "8-level");
212 case ORDER_METHOD_SEARCH:
213 av_log(avctx, AV_LOG_DEBUG, " order method: %s\n", "full search");
215 case ORDER_METHOD_LOG:
216 av_log(avctx, AV_LOG_DEBUG, " order method: %s\n", "log search");
221 av_log(avctx, AV_LOG_DEBUG, " partition order: %d, %d\n",
222 opt->min_partition_order, opt->max_partition_order);
224 av_log(avctx, AV_LOG_DEBUG, " block size: %d\n", avctx->frame_size);
226 av_log(avctx, AV_LOG_DEBUG, " lpc precision: %d\n",
227 opt->lpc_coeff_precision);
231 static av_cold int flac_encode_init(AVCodecContext *avctx)
233 int freq = avctx->sample_rate;
234 int channels = avctx->channels;
235 FlacEncodeContext *s = avctx->priv_data;
241 switch (avctx->sample_fmt) {
242 case AV_SAMPLE_FMT_S16:
243 avctx->bits_per_raw_sample = 16;
246 case AV_SAMPLE_FMT_S32:
247 if (avctx->bits_per_raw_sample != 24)
248 av_log(avctx, AV_LOG_WARNING, "encoding as 24 bits-per-sample\n");
249 avctx->bits_per_raw_sample = 24;
254 if (channels < 1 || channels > FLAC_MAX_CHANNELS) {
255 av_log(avctx, AV_LOG_ERROR, "%d channels not supported (max %d)\n",
256 channels, FLAC_MAX_CHANNELS);
257 return AVERROR(EINVAL);
259 s->channels = channels;
261 /* find samplerate in table */
264 for (i = 4; i < 12; i++) {
265 if (freq == ff_flac_sample_rate_table[i]) {
266 s->samplerate = ff_flac_sample_rate_table[i];
272 /* if not in table, samplerate is non-standard */
274 if (freq % 1000 == 0 && freq < 255000) {
276 s->sr_code[1] = freq / 1000;
277 } else if (freq % 10 == 0 && freq < 655350) {
279 s->sr_code[1] = freq / 10;
280 } else if (freq < 65535) {
282 s->sr_code[1] = freq;
284 av_log(avctx, AV_LOG_ERROR, "%d Hz not supported\n", freq);
285 return AVERROR(EINVAL);
287 s->samplerate = freq;
290 /* set compression option defaults based on avctx->compression_level */
291 if (avctx->compression_level < 0)
292 s->options.compression_level = 5;
294 s->options.compression_level = avctx->compression_level;
296 level = s->options.compression_level;
298 av_log(avctx, AV_LOG_ERROR, "invalid compression level: %d\n",
299 s->options.compression_level);
300 return AVERROR(EINVAL);
303 s->options.block_time_ms = ((int[]){ 27, 27, 27,105,105,105,105,105,105,105,105,105,105})[level];
305 if (s->options.lpc_type == FF_LPC_TYPE_DEFAULT)
306 s->options.lpc_type = ((int[]){ FF_LPC_TYPE_FIXED, FF_LPC_TYPE_FIXED, FF_LPC_TYPE_FIXED,
307 FF_LPC_TYPE_LEVINSON, FF_LPC_TYPE_LEVINSON, FF_LPC_TYPE_LEVINSON,
308 FF_LPC_TYPE_LEVINSON, FF_LPC_TYPE_LEVINSON, FF_LPC_TYPE_LEVINSON,
309 FF_LPC_TYPE_LEVINSON, FF_LPC_TYPE_LEVINSON, FF_LPC_TYPE_LEVINSON,
310 FF_LPC_TYPE_LEVINSON})[level];
312 s->options.min_prediction_order = ((int[]){ 2, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1})[level];
313 s->options.max_prediction_order = ((int[]){ 3, 4, 4, 6, 8, 8, 8, 8, 12, 12, 12, 32, 32})[level];
315 if (s->options.prediction_order_method < 0)
316 s->options.prediction_order_method = ((int[]){ ORDER_METHOD_EST, ORDER_METHOD_EST, ORDER_METHOD_EST,
317 ORDER_METHOD_EST, ORDER_METHOD_EST, ORDER_METHOD_EST,
318 ORDER_METHOD_4LEVEL, ORDER_METHOD_LOG, ORDER_METHOD_4LEVEL,
319 ORDER_METHOD_LOG, ORDER_METHOD_SEARCH, ORDER_METHOD_LOG,
320 ORDER_METHOD_SEARCH})[level];
322 if (s->options.min_partition_order > s->options.max_partition_order) {
323 av_log(avctx, AV_LOG_ERROR, "invalid partition orders: min=%d max=%d\n",
324 s->options.min_partition_order, s->options.max_partition_order);
325 return AVERROR(EINVAL);
327 if (s->options.min_partition_order < 0)
328 s->options.min_partition_order = ((int[]){ 2, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0})[level];
329 if (s->options.max_partition_order < 0)
330 s->options.max_partition_order = ((int[]){ 2, 2, 3, 3, 3, 8, 8, 8, 8, 8, 8, 8, 8})[level];
332 if (s->options.lpc_type == FF_LPC_TYPE_NONE) {
333 s->options.min_prediction_order = 0;
334 } else if (avctx->min_prediction_order >= 0) {
335 if (s->options.lpc_type == FF_LPC_TYPE_FIXED) {
336 if (avctx->min_prediction_order > MAX_FIXED_ORDER) {
337 av_log(avctx, AV_LOG_ERROR, "invalid min prediction order: %d\n",
338 avctx->min_prediction_order);
339 return AVERROR(EINVAL);
341 } else if (avctx->min_prediction_order < MIN_LPC_ORDER ||
342 avctx->min_prediction_order > MAX_LPC_ORDER) {
343 av_log(avctx, AV_LOG_ERROR, "invalid min prediction order: %d\n",
344 avctx->min_prediction_order);
345 return AVERROR(EINVAL);
347 s->options.min_prediction_order = avctx->min_prediction_order;
349 if (s->options.lpc_type == FF_LPC_TYPE_NONE) {
350 s->options.max_prediction_order = 0;
351 } else if (avctx->max_prediction_order >= 0) {
352 if (s->options.lpc_type == FF_LPC_TYPE_FIXED) {
353 if (avctx->max_prediction_order > MAX_FIXED_ORDER) {
354 av_log(avctx, AV_LOG_ERROR, "invalid max prediction order: %d\n",
355 avctx->max_prediction_order);
356 return AVERROR(EINVAL);
358 } else if (avctx->max_prediction_order < MIN_LPC_ORDER ||
359 avctx->max_prediction_order > MAX_LPC_ORDER) {
360 av_log(avctx, AV_LOG_ERROR, "invalid max prediction order: %d\n",
361 avctx->max_prediction_order);
362 return AVERROR(EINVAL);
364 s->options.max_prediction_order = avctx->max_prediction_order;
366 if (s->options.max_prediction_order < s->options.min_prediction_order) {
367 av_log(avctx, AV_LOG_ERROR, "invalid prediction orders: min=%d max=%d\n",
368 s->options.min_prediction_order, s->options.max_prediction_order);
369 return AVERROR(EINVAL);
372 if (avctx->frame_size > 0) {
373 if (avctx->frame_size < FLAC_MIN_BLOCKSIZE ||
374 avctx->frame_size > FLAC_MAX_BLOCKSIZE) {
375 av_log(avctx, AV_LOG_ERROR, "invalid block size: %d\n",
377 return AVERROR(EINVAL);
380 s->avctx->frame_size = select_blocksize(s->samplerate, s->options.block_time_ms);
382 s->max_blocksize = s->avctx->frame_size;
384 /* set maximum encoded frame size in verbatim mode */
385 s->max_framesize = ff_flac_get_max_frame_size(s->avctx->frame_size,
387 s->avctx->bits_per_raw_sample);
389 /* initialize MD5 context */
390 s->md5ctx = av_md5_alloc();
392 return AVERROR(ENOMEM);
393 av_md5_init(s->md5ctx);
395 streaminfo = av_malloc(FLAC_STREAMINFO_SIZE);
397 return AVERROR(ENOMEM);
398 write_streaminfo(s, streaminfo);
399 avctx->extradata = streaminfo;
400 avctx->extradata_size = FLAC_STREAMINFO_SIZE;
403 s->min_framesize = s->max_framesize;
406 avctx->channel_layout != (AV_CH_LAYOUT_STEREO|AV_CH_FRONT_CENTER) ||
408 avctx->channel_layout != AV_CH_LAYOUT_2_2 &&
409 avctx->channel_layout != AV_CH_LAYOUT_QUAD ||
411 avctx->channel_layout != AV_CH_LAYOUT_5POINT0 &&
412 avctx->channel_layout != AV_CH_LAYOUT_5POINT0_BACK ||
414 avctx->channel_layout != AV_CH_LAYOUT_5POINT1 &&
415 avctx->channel_layout != AV_CH_LAYOUT_5POINT1_BACK) {
416 if (avctx->channel_layout) {
417 av_log(avctx, AV_LOG_ERROR, "Channel layout not supported by Flac, "
418 "output stream will have incorrect "
419 "channel layout.\n");
421 av_log(avctx, AV_LOG_WARNING, "No channel layout specified. The encoder "
422 "will use Flac channel layout for "
423 "%d channels.\n", channels);
427 ret = ff_lpc_init(&s->lpc_ctx, avctx->frame_size,
428 s->options.max_prediction_order, FF_LPC_TYPE_LEVINSON);
430 ff_bswapdsp_init(&s->bdsp);
431 ff_flacdsp_init(&s->flac_dsp, avctx->sample_fmt, channels,
432 avctx->bits_per_raw_sample);
434 dprint_compression_options(s);
440 static void init_frame(FlacEncodeContext *s, int nb_samples)
447 for (i = 0; i < 16; i++) {
448 if (nb_samples == ff_flac_blocksize_table[i]) {
449 frame->blocksize = ff_flac_blocksize_table[i];
450 frame->bs_code[0] = i;
451 frame->bs_code[1] = 0;
456 frame->blocksize = nb_samples;
457 if (frame->blocksize <= 256) {
458 frame->bs_code[0] = 6;
459 frame->bs_code[1] = frame->blocksize-1;
461 frame->bs_code[0] = 7;
462 frame->bs_code[1] = frame->blocksize-1;
466 for (ch = 0; ch < s->channels; ch++) {
467 FlacSubframe *sub = &frame->subframes[ch];
470 sub->obits = s->avctx->bits_per_raw_sample;
473 sub->rc.coding_mode = CODING_MODE_RICE2;
475 sub->rc.coding_mode = CODING_MODE_RICE;
478 frame->verbatim_only = 0;
483 * Copy channel-interleaved input samples into separate subframes.
485 static void copy_samples(FlacEncodeContext *s, const void *samples)
489 int shift = av_get_bytes_per_sample(s->avctx->sample_fmt) * 8 -
490 s->avctx->bits_per_raw_sample;
492 #define COPY_SAMPLES(bits) do { \
493 const int ## bits ## _t *samples0 = samples; \
495 for (i = 0, j = 0; i < frame->blocksize; i++) \
496 for (ch = 0; ch < s->channels; ch++, j++) \
497 frame->subframes[ch].samples[i] = samples0[j] >> shift; \
500 if (s->avctx->sample_fmt == AV_SAMPLE_FMT_S16)
507 static uint64_t rice_count_exact(int32_t *res, int n, int k)
512 for (i = 0; i < n; i++) {
513 int32_t v = -2 * res[i] - 1;
515 count += (v >> k) + 1 + k;
521 static uint64_t subframe_count_exact(FlacEncodeContext *s, FlacSubframe *sub,
524 int p, porder, psize;
528 /* subframe header */
532 if (sub->type == FLAC_SUBFRAME_CONSTANT) {
534 } else if (sub->type == FLAC_SUBFRAME_VERBATIM) {
535 count += s->frame.blocksize * sub->obits;
537 /* warm-up samples */
538 count += pred_order * sub->obits;
540 /* LPC coefficients */
541 if (sub->type == FLAC_SUBFRAME_LPC)
542 count += 4 + 5 + pred_order * s->options.lpc_coeff_precision;
544 /* rice-encoded block */
547 /* partition order */
548 porder = sub->rc.porder;
549 psize = s->frame.blocksize >> porder;
555 for (p = 0; p < 1 << porder; p++) {
556 int k = sub->rc.params[p];
557 count += sub->rc.coding_mode;
558 count += rice_count_exact(&sub->residual[i], part_end - i, k);
560 part_end = FFMIN(s->frame.blocksize, part_end + psize);
568 #define rice_encode_count(sum, n, k) (((n)*((k)+1))+((sum-(n>>1))>>(k)))
571 * Solve for d/dk(rice_encode_count) = n-((sum-(n>>1))>>(k+1)) = 0.
573 static int find_optimal_param(uint64_t sum, int n, int max_param)
580 sum2 = sum - (n >> 1);
581 k = av_log2(av_clipl_int32(sum2 / n));
582 return FFMIN(k, max_param);
586 static uint64_t calc_optimal_rice_params(RiceContext *rc, int porder,
587 uint64_t *sums, int n, int pred_order)
590 int k, cnt, part, max_param;
593 max_param = (1 << rc->coding_mode) - 2;
595 part = (1 << porder);
598 cnt = (n >> porder) - pred_order;
599 for (i = 0; i < part; i++) {
600 k = find_optimal_param(sums[i], cnt, max_param);
602 all_bits += rice_encode_count(sums[i], cnt, k);
612 static void calc_sum_top(int pmax, uint32_t *data, int n, int pred_order,
613 uint64_t sums[MAX_PARTITIONS])
617 uint32_t *res, *res_end;
619 /* sums for highest level */
621 res = &data[pred_order];
622 res_end = &data[n >> pmax];
623 for (i = 0; i < parts; i++) {
625 while (res < res_end)
628 res_end += n >> pmax;
632 static void calc_sum_next(int level, uint64_t sums[MAX_PARTITIONS])
635 int parts = (1 << level);
636 for (i = 0; i < parts; i++)
637 sums[i] = sums[2*i] + sums[2*i+1];
640 static uint64_t calc_rice_params(RiceContext *rc, int pmin, int pmax,
641 int32_t *data, int n, int pred_order)
644 uint64_t bits[MAX_PARTITION_ORDER+1];
648 uint64_t sums[MAX_PARTITIONS];
650 av_assert1(pmin >= 0 && pmin <= MAX_PARTITION_ORDER);
651 av_assert1(pmax >= 0 && pmax <= MAX_PARTITION_ORDER);
652 av_assert1(pmin <= pmax);
654 tmp_rc.coding_mode = rc->coding_mode;
656 udata = av_malloc_array(n, sizeof(uint32_t));
657 for (i = 0; i < n; i++)
658 udata[i] = (2*data[i]) ^ (data[i]>>31);
660 calc_sum_top(pmax, udata, n, pred_order, sums);
663 bits[pmin] = UINT32_MAX;
665 bits[i] = calc_optimal_rice_params(&tmp_rc, i, sums, n, pred_order);
666 if (bits[i] < bits[opt_porder]) {
672 calc_sum_next(--i, sums);
676 return bits[opt_porder];
680 static int get_max_p_order(int max_porder, int n, int order)
682 int porder = FFMIN(max_porder, av_log2(n^(n-1)));
684 porder = FFMIN(porder, av_log2(n/order));
689 static uint64_t find_subframe_rice_params(FlacEncodeContext *s,
690 FlacSubframe *sub, int pred_order)
692 int pmin = get_max_p_order(s->options.min_partition_order,
693 s->frame.blocksize, pred_order);
694 int pmax = get_max_p_order(s->options.max_partition_order,
695 s->frame.blocksize, pred_order);
697 uint64_t bits = 8 + pred_order * sub->obits + 2 + sub->rc.coding_mode;
698 if (sub->type == FLAC_SUBFRAME_LPC)
699 bits += 4 + 5 + pred_order * s->options.lpc_coeff_precision;
700 bits += calc_rice_params(&sub->rc, pmin, pmax, sub->residual,
701 s->frame.blocksize, pred_order);
706 static void encode_residual_fixed(int32_t *res, const int32_t *smp, int n,
711 for (i = 0; i < order; i++)
715 for (i = order; i < n; i++)
717 } else if (order == 1) {
718 for (i = order; i < n; i++)
719 res[i] = smp[i] - smp[i-1];
720 } else if (order == 2) {
721 int a = smp[order-1] - smp[order-2];
722 for (i = order; i < n; i += 2) {
723 int b = smp[i ] - smp[i-1];
725 a = smp[i+1] - smp[i ];
728 } else if (order == 3) {
729 int a = smp[order-1] - smp[order-2];
730 int c = smp[order-1] - 2*smp[order-2] + smp[order-3];
731 for (i = order; i < n; i += 2) {
732 int b = smp[i ] - smp[i-1];
735 a = smp[i+1] - smp[i ];
740 int a = smp[order-1] - smp[order-2];
741 int c = smp[order-1] - 2*smp[order-2] + smp[order-3];
742 int e = smp[order-1] - 3*smp[order-2] + 3*smp[order-3] - smp[order-4];
743 for (i = order; i < n; i += 2) {
744 int b = smp[i ] - smp[i-1];
748 a = smp[i+1] - smp[i ];
757 static int encode_residual_ch(FlacEncodeContext *s, int ch)
760 int min_order, max_order, opt_order, omethod;
763 int32_t coefs[MAX_LPC_ORDER][MAX_LPC_ORDER];
764 int shift[MAX_LPC_ORDER];
768 sub = &frame->subframes[ch];
771 n = frame->blocksize;
774 for (i = 1; i < n; i++)
778 sub->type = sub->type_code = FLAC_SUBFRAME_CONSTANT;
780 return subframe_count_exact(s, sub, 0);
784 if (frame->verbatim_only || n < 5) {
785 sub->type = sub->type_code = FLAC_SUBFRAME_VERBATIM;
786 memcpy(res, smp, n * sizeof(int32_t));
787 return subframe_count_exact(s, sub, 0);
790 min_order = s->options.min_prediction_order;
791 max_order = s->options.max_prediction_order;
792 omethod = s->options.prediction_order_method;
795 sub->type = FLAC_SUBFRAME_FIXED;
796 if (s->options.lpc_type == FF_LPC_TYPE_NONE ||
797 s->options.lpc_type == FF_LPC_TYPE_FIXED || n <= max_order) {
798 uint64_t bits[MAX_FIXED_ORDER+1];
799 if (max_order > MAX_FIXED_ORDER)
800 max_order = MAX_FIXED_ORDER;
802 bits[0] = UINT32_MAX;
803 for (i = min_order; i <= max_order; i++) {
804 encode_residual_fixed(res, smp, n, i);
805 bits[i] = find_subframe_rice_params(s, sub, i);
806 if (bits[i] < bits[opt_order])
809 sub->order = opt_order;
810 sub->type_code = sub->type | sub->order;
811 if (sub->order != max_order) {
812 encode_residual_fixed(res, smp, n, sub->order);
813 find_subframe_rice_params(s, sub, sub->order);
815 return subframe_count_exact(s, sub, sub->order);
819 sub->type = FLAC_SUBFRAME_LPC;
820 opt_order = ff_lpc_calc_coefs(&s->lpc_ctx, smp, n, min_order, max_order,
821 s->options.lpc_coeff_precision, coefs, shift, s->options.lpc_type,
822 s->options.lpc_passes, omethod,
825 if (omethod == ORDER_METHOD_2LEVEL ||
826 omethod == ORDER_METHOD_4LEVEL ||
827 omethod == ORDER_METHOD_8LEVEL) {
828 int levels = 1 << omethod;
829 uint64_t bits[1 << ORDER_METHOD_8LEVEL];
831 int opt_index = levels-1;
832 opt_order = max_order-1;
833 bits[opt_index] = UINT32_MAX;
834 for (i = levels-1; i >= 0; i--) {
835 int last_order = order;
836 order = min_order + (((max_order-min_order+1) * (i+1)) / levels)-1;
837 order = av_clip(order, min_order - 1, max_order - 1);
838 if (order == last_order)
840 s->flac_dsp.lpc_encode(res, smp, n, order+1, coefs[order],
842 bits[i] = find_subframe_rice_params(s, sub, order+1);
843 if (bits[i] < bits[opt_index]) {
849 } else if (omethod == ORDER_METHOD_SEARCH) {
850 // brute-force optimal order search
851 uint64_t bits[MAX_LPC_ORDER];
853 bits[0] = UINT32_MAX;
854 for (i = min_order-1; i < max_order; i++) {
855 s->flac_dsp.lpc_encode(res, smp, n, i+1, coefs[i], shift[i]);
856 bits[i] = find_subframe_rice_params(s, sub, i+1);
857 if (bits[i] < bits[opt_order])
861 } else if (omethod == ORDER_METHOD_LOG) {
862 uint64_t bits[MAX_LPC_ORDER];
865 opt_order = min_order - 1 + (max_order-min_order)/3;
866 memset(bits, -1, sizeof(bits));
868 for (step = 16; step; step >>= 1) {
869 int last = opt_order;
870 for (i = last-step; i <= last+step; i += step) {
871 if (i < min_order-1 || i >= max_order || bits[i] < UINT32_MAX)
873 s->flac_dsp.lpc_encode(res, smp, n, i+1, coefs[i], shift[i]);
874 bits[i] = find_subframe_rice_params(s, sub, i+1);
875 if (bits[i] < bits[opt_order])
882 sub->order = opt_order;
883 sub->type_code = sub->type | (sub->order-1);
884 sub->shift = shift[sub->order-1];
885 for (i = 0; i < sub->order; i++)
886 sub->coefs[i] = coefs[sub->order-1][i];
888 s->flac_dsp.lpc_encode(res, smp, n, sub->order, sub->coefs, sub->shift);
890 find_subframe_rice_params(s, sub, sub->order);
892 return subframe_count_exact(s, sub, sub->order);
896 static int count_frame_header(FlacEncodeContext *s)
898 uint8_t av_unused tmp;
904 <1> Blocking strategy
905 <4> Block size in inter-channel samples
907 <4> Channel assignment
908 <3> Sample size in bits
913 /* coded frame number */
914 PUT_UTF8(s->frame_count, tmp, count += 8;)
916 /* explicit block size */
917 if (s->frame.bs_code[0] == 6)
919 else if (s->frame.bs_code[0] == 7)
922 /* explicit sample rate */
923 count += ((s->sr_code[0] == 12) + (s->sr_code[0] > 12)) * 8;
925 /* frame header CRC-8 */
932 static int encode_frame(FlacEncodeContext *s)
937 count = count_frame_header(s);
939 for (ch = 0; ch < s->channels; ch++)
940 count += encode_residual_ch(s, ch);
942 count += (8 - (count & 7)) & 7; // byte alignment
943 count += 16; // CRC-16
952 static void remove_wasted_bits(FlacEncodeContext *s)
956 for (ch = 0; ch < s->channels; ch++) {
957 FlacSubframe *sub = &s->frame.subframes[ch];
960 for (i = 0; i < s->frame.blocksize; i++) {
961 v |= sub->samples[i];
969 for (i = 0; i < s->frame.blocksize; i++)
970 sub->samples[i] >>= v;
975 /* for 24-bit, check if removing wasted bits makes the range better
976 suited for using RICE instead of RICE2 for entropy coding */
977 if (sub->obits <= 17)
978 sub->rc.coding_mode = CODING_MODE_RICE;
984 static int estimate_stereo_mode(int32_t *left_ch, int32_t *right_ch, int n,
993 /* calculate sum of 2nd order residual for each channel */
994 sum[0] = sum[1] = sum[2] = sum[3] = 0;
995 for (i = 2; i < n; i++) {
996 lt = left_ch[i] - 2*left_ch[i-1] + left_ch[i-2];
997 rt = right_ch[i] - 2*right_ch[i-1] + right_ch[i-2];
998 sum[2] += FFABS((lt + rt) >> 1);
999 sum[3] += FFABS(lt - rt);
1000 sum[0] += FFABS(lt);
1001 sum[1] += FFABS(rt);
1003 /* estimate bit counts */
1004 for (i = 0; i < 4; i++) {
1005 k = find_optimal_param(2 * sum[i], n, max_rice_param);
1006 sum[i] = rice_encode_count( 2 * sum[i], n, k);
1009 /* calculate score for each mode */
1010 score[0] = sum[0] + sum[1];
1011 score[1] = sum[0] + sum[3];
1012 score[2] = sum[1] + sum[3];
1013 score[3] = sum[2] + sum[3];
1015 /* return mode with lowest score */
1017 for (i = 1; i < 4; i++)
1018 if (score[i] < score[best])
1026 * Perform stereo channel decorrelation.
1028 static void channel_decorrelation(FlacEncodeContext *s)
1031 int32_t *left, *right;
1035 n = frame->blocksize;
1036 left = frame->subframes[0].samples;
1037 right = frame->subframes[1].samples;
1039 if (s->channels != 2) {
1040 frame->ch_mode = FLAC_CHMODE_INDEPENDENT;
1044 if (s->options.ch_mode < 0) {
1045 int max_rice_param = (1 << frame->subframes[0].rc.coding_mode) - 2;
1046 frame->ch_mode = estimate_stereo_mode(left, right, n, max_rice_param);
1048 frame->ch_mode = s->options.ch_mode;
1050 /* perform decorrelation and adjust bits-per-sample */
1051 if (frame->ch_mode == FLAC_CHMODE_INDEPENDENT)
1053 if (frame->ch_mode == FLAC_CHMODE_MID_SIDE) {
1055 for (i = 0; i < n; i++) {
1057 left[i] = (tmp + right[i]) >> 1;
1058 right[i] = tmp - right[i];
1060 frame->subframes[1].obits++;
1061 } else if (frame->ch_mode == FLAC_CHMODE_LEFT_SIDE) {
1062 for (i = 0; i < n; i++)
1063 right[i] = left[i] - right[i];
1064 frame->subframes[1].obits++;
1066 for (i = 0; i < n; i++)
1067 left[i] -= right[i];
1068 frame->subframes[0].obits++;
1073 static void write_utf8(PutBitContext *pb, uint32_t val)
1076 PUT_UTF8(val, tmp, put_bits(pb, 8, tmp);)
1080 static void write_frame_header(FlacEncodeContext *s)
1087 put_bits(&s->pb, 16, 0xFFF8);
1088 put_bits(&s->pb, 4, frame->bs_code[0]);
1089 put_bits(&s->pb, 4, s->sr_code[0]);
1091 if (frame->ch_mode == FLAC_CHMODE_INDEPENDENT)
1092 put_bits(&s->pb, 4, s->channels-1);
1094 put_bits(&s->pb, 4, frame->ch_mode + FLAC_MAX_CHANNELS - 1);
1096 put_bits(&s->pb, 3, s->bps_code);
1097 put_bits(&s->pb, 1, 0);
1098 write_utf8(&s->pb, s->frame_count);
1100 if (frame->bs_code[0] == 6)
1101 put_bits(&s->pb, 8, frame->bs_code[1]);
1102 else if (frame->bs_code[0] == 7)
1103 put_bits(&s->pb, 16, frame->bs_code[1]);
1105 if (s->sr_code[0] == 12)
1106 put_bits(&s->pb, 8, s->sr_code[1]);
1107 else if (s->sr_code[0] > 12)
1108 put_bits(&s->pb, 16, s->sr_code[1]);
1110 flush_put_bits(&s->pb);
1111 crc = av_crc(av_crc_get_table(AV_CRC_8_ATM), 0, s->pb.buf,
1112 put_bits_count(&s->pb) >> 3);
1113 put_bits(&s->pb, 8, crc);
1117 static void write_subframes(FlacEncodeContext *s)
1121 for (ch = 0; ch < s->channels; ch++) {
1122 FlacSubframe *sub = &s->frame.subframes[ch];
1123 int i, p, porder, psize;
1125 int32_t *res = sub->residual;
1126 int32_t *frame_end = &sub->residual[s->frame.blocksize];
1128 /* subframe header */
1129 put_bits(&s->pb, 1, 0);
1130 put_bits(&s->pb, 6, sub->type_code);
1131 put_bits(&s->pb, 1, !!sub->wasted);
1133 put_bits(&s->pb, sub->wasted, 1);
1136 if (sub->type == FLAC_SUBFRAME_CONSTANT) {
1137 put_sbits(&s->pb, sub->obits, res[0]);
1138 } else if (sub->type == FLAC_SUBFRAME_VERBATIM) {
1139 while (res < frame_end)
1140 put_sbits(&s->pb, sub->obits, *res++);
1142 /* warm-up samples */
1143 for (i = 0; i < sub->order; i++)
1144 put_sbits(&s->pb, sub->obits, *res++);
1146 /* LPC coefficients */
1147 if (sub->type == FLAC_SUBFRAME_LPC) {
1148 int cbits = s->options.lpc_coeff_precision;
1149 put_bits( &s->pb, 4, cbits-1);
1150 put_sbits(&s->pb, 5, sub->shift);
1151 for (i = 0; i < sub->order; i++)
1152 put_sbits(&s->pb, cbits, sub->coefs[i]);
1155 /* rice-encoded block */
1156 put_bits(&s->pb, 2, sub->rc.coding_mode - 4);
1158 /* partition order */
1159 porder = sub->rc.porder;
1160 psize = s->frame.blocksize >> porder;
1161 put_bits(&s->pb, 4, porder);
1164 part_end = &sub->residual[psize];
1165 for (p = 0; p < 1 << porder; p++) {
1166 int k = sub->rc.params[p];
1167 put_bits(&s->pb, sub->rc.coding_mode, k);
1168 while (res < part_end)
1169 set_sr_golomb_flac(&s->pb, *res++, k, INT32_MAX, 0);
1170 part_end = FFMIN(frame_end, part_end + psize);
1177 static void write_frame_footer(FlacEncodeContext *s)
1180 flush_put_bits(&s->pb);
1181 crc = av_bswap16(av_crc(av_crc_get_table(AV_CRC_16_ANSI), 0, s->pb.buf,
1182 put_bits_count(&s->pb)>>3));
1183 put_bits(&s->pb, 16, crc);
1184 flush_put_bits(&s->pb);
1188 static int write_frame(FlacEncodeContext *s, AVPacket *avpkt)
1190 init_put_bits(&s->pb, avpkt->data, avpkt->size);
1191 write_frame_header(s);
1193 write_frame_footer(s);
1194 return put_bits_count(&s->pb) >> 3;
1198 static int update_md5_sum(FlacEncodeContext *s, const void *samples)
1201 int buf_size = s->frame.blocksize * s->channels *
1202 ((s->avctx->bits_per_raw_sample + 7) / 8);
1204 if (s->avctx->bits_per_raw_sample > 16 || HAVE_BIGENDIAN) {
1205 av_fast_malloc(&s->md5_buffer, &s->md5_buffer_size, buf_size);
1207 return AVERROR(ENOMEM);
1210 if (s->avctx->bits_per_raw_sample <= 16) {
1211 buf = (const uint8_t *)samples;
1213 s->bdsp.bswap16_buf((uint16_t *) s->md5_buffer,
1214 (const uint16_t *) samples, buf_size / 2);
1215 buf = s->md5_buffer;
1219 const int32_t *samples0 = samples;
1220 uint8_t *tmp = s->md5_buffer;
1222 for (i = 0; i < s->frame.blocksize * s->channels; i++) {
1223 int32_t v = samples0[i] >> 8;
1224 *tmp++ = (v ) & 0xFF;
1225 *tmp++ = (v >> 8) & 0xFF;
1226 *tmp++ = (v >> 16) & 0xFF;
1228 buf = s->md5_buffer;
1230 av_md5_update(s->md5ctx, buf, buf_size);
1236 static int flac_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
1237 const AVFrame *frame, int *got_packet_ptr)
1239 FlacEncodeContext *s;
1240 int frame_bytes, out_bytes, ret;
1242 s = avctx->priv_data;
1244 /* when the last block is reached, update the header in extradata */
1246 s->max_framesize = s->max_encoded_framesize;
1247 av_md5_final(s->md5ctx, s->md5sum);
1248 write_streaminfo(s, avctx->extradata);
1250 if (avctx->side_data_only_packets && !s->flushed) {
1251 uint8_t *side_data = av_packet_new_side_data(avpkt, AV_PKT_DATA_NEW_EXTRADATA,
1252 avctx->extradata_size);
1254 return AVERROR(ENOMEM);
1255 memcpy(side_data, avctx->extradata, avctx->extradata_size);
1257 avpkt->pts = s->next_pts;
1259 *got_packet_ptr = 1;
1266 /* change max_framesize for small final frame */
1267 if (frame->nb_samples < s->frame.blocksize) {
1268 s->max_framesize = ff_flac_get_max_frame_size(frame->nb_samples,
1270 avctx->bits_per_raw_sample);
1273 init_frame(s, frame->nb_samples);
1275 copy_samples(s, frame->data[0]);
1277 channel_decorrelation(s);
1279 remove_wasted_bits(s);
1281 frame_bytes = encode_frame(s);
1283 /* Fall back on verbatim mode if the compressed frame is larger than it
1284 would be if encoded uncompressed. */
1285 if (frame_bytes < 0 || frame_bytes > s->max_framesize) {
1286 s->frame.verbatim_only = 1;
1287 frame_bytes = encode_frame(s);
1288 if (frame_bytes < 0) {
1289 av_log(avctx, AV_LOG_ERROR, "Bad frame count\n");
1294 if ((ret = ff_alloc_packet2(avctx, avpkt, frame_bytes)) < 0)
1297 out_bytes = write_frame(s, avpkt);
1300 s->sample_count += frame->nb_samples;
1301 if ((ret = update_md5_sum(s, frame->data[0])) < 0) {
1302 av_log(avctx, AV_LOG_ERROR, "Error updating MD5 checksum\n");
1305 if (out_bytes > s->max_encoded_framesize)
1306 s->max_encoded_framesize = out_bytes;
1307 if (out_bytes < s->min_framesize)
1308 s->min_framesize = out_bytes;
1310 avpkt->pts = frame->pts;
1311 avpkt->duration = ff_samples_to_time_base(avctx, frame->nb_samples);
1312 avpkt->size = out_bytes;
1314 s->next_pts = avpkt->pts + avpkt->duration;
1316 *got_packet_ptr = 1;
1321 static av_cold int flac_encode_close(AVCodecContext *avctx)
1323 if (avctx->priv_data) {
1324 FlacEncodeContext *s = avctx->priv_data;
1325 av_freep(&s->md5ctx);
1326 av_freep(&s->md5_buffer);
1327 ff_lpc_end(&s->lpc_ctx);
1329 av_freep(&avctx->extradata);
1330 avctx->extradata_size = 0;
1334 #define FLAGS AV_OPT_FLAG_ENCODING_PARAM | AV_OPT_FLAG_AUDIO_PARAM
1335 static const AVOption options[] = {
1336 { "lpc_coeff_precision", "LPC coefficient precision", offsetof(FlacEncodeContext, options.lpc_coeff_precision), AV_OPT_TYPE_INT, {.i64 = 15 }, 0, MAX_LPC_PRECISION, FLAGS },
1337 { "lpc_type", "LPC algorithm", offsetof(FlacEncodeContext, options.lpc_type), AV_OPT_TYPE_INT, {.i64 = FF_LPC_TYPE_DEFAULT }, FF_LPC_TYPE_DEFAULT, FF_LPC_TYPE_NB-1, FLAGS, "lpc_type" },
1338 { "none", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = FF_LPC_TYPE_NONE }, INT_MIN, INT_MAX, FLAGS, "lpc_type" },
1339 { "fixed", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = FF_LPC_TYPE_FIXED }, INT_MIN, INT_MAX, FLAGS, "lpc_type" },
1340 { "levinson", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = FF_LPC_TYPE_LEVINSON }, INT_MIN, INT_MAX, FLAGS, "lpc_type" },
1341 { "cholesky", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = FF_LPC_TYPE_CHOLESKY }, INT_MIN, INT_MAX, FLAGS, "lpc_type" },
1342 { "lpc_passes", "Number of passes to use for Cholesky factorization during LPC analysis", offsetof(FlacEncodeContext, options.lpc_passes), AV_OPT_TYPE_INT, {.i64 = 2 }, 1, INT_MAX, FLAGS },
1343 { "min_partition_order", NULL, offsetof(FlacEncodeContext, options.min_partition_order), AV_OPT_TYPE_INT, {.i64 = -1 }, -1, MAX_PARTITION_ORDER, FLAGS },
1344 { "max_partition_order", NULL, offsetof(FlacEncodeContext, options.max_partition_order), AV_OPT_TYPE_INT, {.i64 = -1 }, -1, MAX_PARTITION_ORDER, FLAGS },
1345 { "prediction_order_method", "Search method for selecting prediction order", offsetof(FlacEncodeContext, options.prediction_order_method), AV_OPT_TYPE_INT, {.i64 = -1 }, -1, ORDER_METHOD_LOG, FLAGS, "predm" },
1346 { "estimation", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = ORDER_METHOD_EST }, INT_MIN, INT_MAX, FLAGS, "predm" },
1347 { "2level", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = ORDER_METHOD_2LEVEL }, INT_MIN, INT_MAX, FLAGS, "predm" },
1348 { "4level", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = ORDER_METHOD_4LEVEL }, INT_MIN, INT_MAX, FLAGS, "predm" },
1349 { "8level", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = ORDER_METHOD_8LEVEL }, INT_MIN, INT_MAX, FLAGS, "predm" },
1350 { "search", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = ORDER_METHOD_SEARCH }, INT_MIN, INT_MAX, FLAGS, "predm" },
1351 { "log", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = ORDER_METHOD_LOG }, INT_MIN, INT_MAX, FLAGS, "predm" },
1352 { "ch_mode", "Stereo decorrelation mode", offsetof(FlacEncodeContext, options.ch_mode), AV_OPT_TYPE_INT, { .i64 = -1 }, -1, FLAC_CHMODE_MID_SIDE, FLAGS, "ch_mode" },
1353 { "auto", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = -1 }, INT_MIN, INT_MAX, FLAGS, "ch_mode" },
1354 { "indep", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FLAC_CHMODE_INDEPENDENT }, INT_MIN, INT_MAX, FLAGS, "ch_mode" },
1355 { "left_side", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FLAC_CHMODE_LEFT_SIDE }, INT_MIN, INT_MAX, FLAGS, "ch_mode" },
1356 { "right_side", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FLAC_CHMODE_RIGHT_SIDE }, INT_MIN, INT_MAX, FLAGS, "ch_mode" },
1357 { "mid_side", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FLAC_CHMODE_MID_SIDE }, INT_MIN, INT_MAX, FLAGS, "ch_mode" },
1361 static const AVClass flac_encoder_class = {
1363 av_default_item_name,
1365 LIBAVUTIL_VERSION_INT,
1368 AVCodec ff_flac_encoder = {
1370 .long_name = NULL_IF_CONFIG_SMALL("FLAC (Free Lossless Audio Codec)"),
1371 .type = AVMEDIA_TYPE_AUDIO,
1372 .id = AV_CODEC_ID_FLAC,
1373 .priv_data_size = sizeof(FlacEncodeContext),
1374 .init = flac_encode_init,
1375 .encode2 = flac_encode_frame,
1376 .close = flac_encode_close,
1377 .capabilities = CODEC_CAP_SMALL_LAST_FRAME | CODEC_CAP_DELAY | CODEC_CAP_LOSSLESS,
1378 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_S16,
1380 AV_SAMPLE_FMT_NONE },
1381 .priv_class = &flac_encoder_class,