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/crc.h"
23 #include "libavutil/md5.h"
24 #include "libavutil/opt.h"
32 #define FLAC_SUBFRAME_CONSTANT 0
33 #define FLAC_SUBFRAME_VERBATIM 1
34 #define FLAC_SUBFRAME_FIXED 8
35 #define FLAC_SUBFRAME_LPC 32
37 #define MAX_FIXED_ORDER 4
38 #define MAX_PARTITION_ORDER 8
39 #define MAX_PARTITIONS (1 << MAX_PARTITION_ORDER)
40 #define MAX_LPC_PRECISION 15
41 #define MAX_LPC_SHIFT 15
42 #define MAX_RICE_PARAM 14
44 typedef struct CompressionOptions {
45 int compression_level;
47 enum FFLPCType lpc_type;
49 int lpc_coeff_precision;
50 int min_prediction_order;
51 int max_prediction_order;
52 int prediction_order_method;
53 int min_partition_order;
54 int max_partition_order;
57 typedef struct RiceContext {
59 int params[MAX_PARTITIONS];
62 typedef struct FlacSubframe {
67 int32_t coefs[MAX_LPC_ORDER];
70 int32_t samples[FLAC_MAX_BLOCKSIZE];
71 int32_t residual[FLAC_MAX_BLOCKSIZE+1];
74 typedef struct FlacFrame {
75 FlacSubframe subframes[FLAC_MAX_CHANNELS];
83 typedef struct FlacEncodeContext {
92 int max_encoded_framesize;
94 uint64_t sample_count;
97 CompressionOptions options;
98 AVCodecContext *avctx;
100 struct AVMD5 *md5ctx;
105 * Write streaminfo metadata block to byte array.
107 static void write_streaminfo(FlacEncodeContext *s, uint8_t *header)
111 memset(header, 0, FLAC_STREAMINFO_SIZE);
112 init_put_bits(&pb, header, FLAC_STREAMINFO_SIZE);
114 /* streaminfo metadata block */
115 put_bits(&pb, 16, s->max_blocksize);
116 put_bits(&pb, 16, s->max_blocksize);
117 put_bits(&pb, 24, s->min_framesize);
118 put_bits(&pb, 24, s->max_framesize);
119 put_bits(&pb, 20, s->samplerate);
120 put_bits(&pb, 3, s->channels-1);
121 put_bits(&pb, 5, 15); /* bits per sample - 1 */
122 /* write 36-bit sample count in 2 put_bits() calls */
123 put_bits(&pb, 24, (s->sample_count & 0xFFFFFF000LL) >> 12);
124 put_bits(&pb, 12, s->sample_count & 0x000000FFFLL);
126 memcpy(&header[18], s->md5sum, 16);
131 * Set blocksize based on samplerate.
132 * Choose the closest predefined blocksize >= BLOCK_TIME_MS milliseconds.
134 static int select_blocksize(int samplerate, int block_time_ms)
140 assert(samplerate > 0);
141 blocksize = ff_flac_blocksize_table[1];
142 target = (samplerate * block_time_ms) / 1000;
143 for (i = 0; i < 16; i++) {
144 if (target >= ff_flac_blocksize_table[i] &&
145 ff_flac_blocksize_table[i] > blocksize) {
146 blocksize = ff_flac_blocksize_table[i];
153 static av_cold void dprint_compression_options(FlacEncodeContext *s)
155 AVCodecContext *avctx = s->avctx;
156 CompressionOptions *opt = &s->options;
158 av_log(avctx, AV_LOG_DEBUG, " compression: %d\n", opt->compression_level);
160 switch (opt->lpc_type) {
161 case FF_LPC_TYPE_NONE:
162 av_log(avctx, AV_LOG_DEBUG, " lpc type: None\n");
164 case FF_LPC_TYPE_FIXED:
165 av_log(avctx, AV_LOG_DEBUG, " lpc type: Fixed pre-defined coefficients\n");
167 case FF_LPC_TYPE_LEVINSON:
168 av_log(avctx, AV_LOG_DEBUG, " lpc type: Levinson-Durbin recursion with Welch window\n");
170 case FF_LPC_TYPE_CHOLESKY:
171 av_log(avctx, AV_LOG_DEBUG, " lpc type: Cholesky factorization, %d pass%s\n",
172 opt->lpc_passes, opt->lpc_passes == 1 ? "" : "es");
176 av_log(avctx, AV_LOG_DEBUG, " prediction order: %d, %d\n",
177 opt->min_prediction_order, opt->max_prediction_order);
179 switch (opt->prediction_order_method) {
180 case ORDER_METHOD_EST:
181 av_log(avctx, AV_LOG_DEBUG, " order method: %s\n", "estimate");
183 case ORDER_METHOD_2LEVEL:
184 av_log(avctx, AV_LOG_DEBUG, " order method: %s\n", "2-level");
186 case ORDER_METHOD_4LEVEL:
187 av_log(avctx, AV_LOG_DEBUG, " order method: %s\n", "4-level");
189 case ORDER_METHOD_8LEVEL:
190 av_log(avctx, AV_LOG_DEBUG, " order method: %s\n", "8-level");
192 case ORDER_METHOD_SEARCH:
193 av_log(avctx, AV_LOG_DEBUG, " order method: %s\n", "full search");
195 case ORDER_METHOD_LOG:
196 av_log(avctx, AV_LOG_DEBUG, " order method: %s\n", "log search");
201 av_log(avctx, AV_LOG_DEBUG, " partition order: %d, %d\n",
202 opt->min_partition_order, opt->max_partition_order);
204 av_log(avctx, AV_LOG_DEBUG, " block size: %d\n", avctx->frame_size);
206 av_log(avctx, AV_LOG_DEBUG, " lpc precision: %d\n",
207 opt->lpc_coeff_precision);
211 static av_cold int flac_encode_init(AVCodecContext *avctx)
213 int freq = avctx->sample_rate;
214 int channels = avctx->channels;
215 FlacEncodeContext *s = avctx->priv_data;
221 if (avctx->sample_fmt != AV_SAMPLE_FMT_S16)
224 if (channels < 1 || channels > FLAC_MAX_CHANNELS)
226 s->channels = channels;
228 /* find samplerate in table */
231 for (i = 4; i < 12; i++) {
232 if (freq == ff_flac_sample_rate_table[i]) {
233 s->samplerate = ff_flac_sample_rate_table[i];
239 /* if not in table, samplerate is non-standard */
241 if (freq % 1000 == 0 && freq < 255000) {
243 s->sr_code[1] = freq / 1000;
244 } else if (freq % 10 == 0 && freq < 655350) {
246 s->sr_code[1] = freq / 10;
247 } else if (freq < 65535) {
249 s->sr_code[1] = freq;
253 s->samplerate = freq;
256 /* set compression option defaults based on avctx->compression_level */
257 if (avctx->compression_level < 0)
258 s->options.compression_level = 5;
260 s->options.compression_level = avctx->compression_level;
262 level = s->options.compression_level;
264 av_log(avctx, AV_LOG_ERROR, "invalid compression level: %d\n",
265 s->options.compression_level);
269 s->options.block_time_ms = ((int[]){ 27, 27, 27,105,105,105,105,105,105,105,105,105,105})[level];
271 if (s->options.lpc_type == FF_LPC_TYPE_DEFAULT)
272 s->options.lpc_type = ((int[]){ FF_LPC_TYPE_FIXED, FF_LPC_TYPE_FIXED, FF_LPC_TYPE_FIXED,
273 FF_LPC_TYPE_LEVINSON, FF_LPC_TYPE_LEVINSON, FF_LPC_TYPE_LEVINSON,
274 FF_LPC_TYPE_LEVINSON, FF_LPC_TYPE_LEVINSON, FF_LPC_TYPE_LEVINSON,
275 FF_LPC_TYPE_LEVINSON, FF_LPC_TYPE_LEVINSON, FF_LPC_TYPE_LEVINSON,
276 FF_LPC_TYPE_LEVINSON})[level];
278 s->options.min_prediction_order = ((int[]){ 2, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1})[level];
279 s->options.max_prediction_order = ((int[]){ 3, 4, 4, 6, 8, 8, 8, 8, 12, 12, 12, 32, 32})[level];
281 if (s->options.prediction_order_method < 0)
282 s->options.prediction_order_method = ((int[]){ ORDER_METHOD_EST, ORDER_METHOD_EST, ORDER_METHOD_EST,
283 ORDER_METHOD_EST, ORDER_METHOD_EST, ORDER_METHOD_EST,
284 ORDER_METHOD_4LEVEL, ORDER_METHOD_LOG, ORDER_METHOD_4LEVEL,
285 ORDER_METHOD_LOG, ORDER_METHOD_SEARCH, ORDER_METHOD_LOG,
286 ORDER_METHOD_SEARCH})[level];
288 if (s->options.min_partition_order > s->options.max_partition_order) {
289 av_log(avctx, AV_LOG_ERROR, "invalid partition orders: min=%d max=%d\n",
290 s->options.min_partition_order, s->options.max_partition_order);
291 return AVERROR(EINVAL);
293 if (s->options.min_partition_order < 0)
294 s->options.min_partition_order = ((int[]){ 2, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0})[level];
295 if (s->options.max_partition_order < 0)
296 s->options.max_partition_order = ((int[]){ 2, 2, 3, 3, 3, 8, 8, 8, 8, 8, 8, 8, 8})[level];
298 if (s->options.lpc_type == FF_LPC_TYPE_NONE) {
299 s->options.min_prediction_order = 0;
300 } else if (avctx->min_prediction_order >= 0) {
301 if (s->options.lpc_type == FF_LPC_TYPE_FIXED) {
302 if (avctx->min_prediction_order > MAX_FIXED_ORDER) {
303 av_log(avctx, AV_LOG_ERROR, "invalid min prediction order: %d\n",
304 avctx->min_prediction_order);
307 } else if (avctx->min_prediction_order < MIN_LPC_ORDER ||
308 avctx->min_prediction_order > MAX_LPC_ORDER) {
309 av_log(avctx, AV_LOG_ERROR, "invalid min prediction order: %d\n",
310 avctx->min_prediction_order);
313 s->options.min_prediction_order = avctx->min_prediction_order;
315 if (s->options.lpc_type == FF_LPC_TYPE_NONE) {
316 s->options.max_prediction_order = 0;
317 } else if (avctx->max_prediction_order >= 0) {
318 if (s->options.lpc_type == FF_LPC_TYPE_FIXED) {
319 if (avctx->max_prediction_order > MAX_FIXED_ORDER) {
320 av_log(avctx, AV_LOG_ERROR, "invalid max prediction order: %d\n",
321 avctx->max_prediction_order);
324 } else if (avctx->max_prediction_order < MIN_LPC_ORDER ||
325 avctx->max_prediction_order > MAX_LPC_ORDER) {
326 av_log(avctx, AV_LOG_ERROR, "invalid max prediction order: %d\n",
327 avctx->max_prediction_order);
330 s->options.max_prediction_order = avctx->max_prediction_order;
332 if (s->options.max_prediction_order < s->options.min_prediction_order) {
333 av_log(avctx, AV_LOG_ERROR, "invalid prediction orders: min=%d max=%d\n",
334 s->options.min_prediction_order, s->options.max_prediction_order);
338 if (avctx->frame_size > 0) {
339 if (avctx->frame_size < FLAC_MIN_BLOCKSIZE ||
340 avctx->frame_size > FLAC_MAX_BLOCKSIZE) {
341 av_log(avctx, AV_LOG_ERROR, "invalid block size: %d\n",
346 s->avctx->frame_size = select_blocksize(s->samplerate, s->options.block_time_ms);
348 s->max_blocksize = s->avctx->frame_size;
350 /* set maximum encoded frame size in verbatim mode */
351 s->max_framesize = ff_flac_get_max_frame_size(s->avctx->frame_size,
354 /* initialize MD5 context */
355 s->md5ctx = av_malloc(av_md5_size);
357 return AVERROR(ENOMEM);
358 av_md5_init(s->md5ctx);
360 streaminfo = av_malloc(FLAC_STREAMINFO_SIZE);
362 return AVERROR(ENOMEM);
363 write_streaminfo(s, streaminfo);
364 avctx->extradata = streaminfo;
365 avctx->extradata_size = FLAC_STREAMINFO_SIZE;
368 s->min_framesize = s->max_framesize;
370 avctx->coded_frame = avcodec_alloc_frame();
371 if (!avctx->coded_frame)
372 return AVERROR(ENOMEM);
375 avctx->channel_layout != (AV_CH_LAYOUT_STEREO|AV_CH_FRONT_CENTER) ||
377 avctx->channel_layout != AV_CH_LAYOUT_2_2 &&
378 avctx->channel_layout != AV_CH_LAYOUT_QUAD ||
380 avctx->channel_layout != AV_CH_LAYOUT_5POINT0 &&
381 avctx->channel_layout != AV_CH_LAYOUT_5POINT0_BACK ||
383 avctx->channel_layout != AV_CH_LAYOUT_5POINT1 &&
384 avctx->channel_layout != AV_CH_LAYOUT_5POINT1_BACK) {
385 if (avctx->channel_layout) {
386 av_log(avctx, AV_LOG_ERROR, "Channel layout not supported by Flac, "
387 "output stream will have incorrect "
388 "channel layout.\n");
390 av_log(avctx, AV_LOG_WARNING, "No channel layout specified. The encoder "
391 "will use Flac channel layout for "
392 "%d channels.\n", channels);
396 ret = ff_lpc_init(&s->lpc_ctx, avctx->frame_size,
397 s->options.max_prediction_order, FF_LPC_TYPE_LEVINSON);
399 dprint_compression_options(s);
405 static void init_frame(FlacEncodeContext *s)
412 for (i = 0; i < 16; i++) {
413 if (s->avctx->frame_size == ff_flac_blocksize_table[i]) {
414 frame->blocksize = ff_flac_blocksize_table[i];
415 frame->bs_code[0] = i;
416 frame->bs_code[1] = 0;
421 frame->blocksize = s->avctx->frame_size;
422 if (frame->blocksize <= 256) {
423 frame->bs_code[0] = 6;
424 frame->bs_code[1] = frame->blocksize-1;
426 frame->bs_code[0] = 7;
427 frame->bs_code[1] = frame->blocksize-1;
431 for (ch = 0; ch < s->channels; ch++)
432 frame->subframes[ch].obits = 16;
434 frame->verbatim_only = 0;
439 * Copy channel-interleaved input samples into separate subframes.
441 static void copy_samples(FlacEncodeContext *s, const int16_t *samples)
447 for (i = 0, j = 0; i < frame->blocksize; i++)
448 for (ch = 0; ch < s->channels; ch++, j++)
449 frame->subframes[ch].samples[i] = samples[j];
453 static int rice_count_exact(int32_t *res, int n, int k)
458 for (i = 0; i < n; i++) {
459 int32_t v = -2 * res[i] - 1;
461 count += (v >> k) + 1 + k;
467 static int subframe_count_exact(FlacEncodeContext *s, FlacSubframe *sub,
470 int p, porder, psize;
474 /* subframe header */
478 if (sub->type == FLAC_SUBFRAME_CONSTANT) {
480 } else if (sub->type == FLAC_SUBFRAME_VERBATIM) {
481 count += s->frame.blocksize * sub->obits;
483 /* warm-up samples */
484 count += pred_order * sub->obits;
486 /* LPC coefficients */
487 if (sub->type == FLAC_SUBFRAME_LPC)
488 count += 4 + 5 + pred_order * s->options.lpc_coeff_precision;
490 /* rice-encoded block */
493 /* partition order */
494 porder = sub->rc.porder;
495 psize = s->frame.blocksize >> porder;
501 for (p = 0; p < 1 << porder; p++) {
502 int k = sub->rc.params[p];
504 count += rice_count_exact(&sub->residual[i], part_end - i, k);
506 part_end = FFMIN(s->frame.blocksize, part_end + psize);
514 #define rice_encode_count(sum, n, k) (((n)*((k)+1))+((sum-(n>>1))>>(k)))
517 * Solve for d/dk(rice_encode_count) = n-((sum-(n>>1))>>(k+1)) = 0.
519 static int find_optimal_param(uint32_t sum, int n)
526 sum2 = sum - (n >> 1);
527 k = av_log2(n < 256 ? FASTDIV(sum2, n) : sum2 / n);
528 return FFMIN(k, MAX_RICE_PARAM);
532 static uint32_t calc_optimal_rice_params(RiceContext *rc, int porder,
533 uint32_t *sums, int n, int pred_order)
539 part = (1 << porder);
542 cnt = (n >> porder) - pred_order;
543 for (i = 0; i < part; i++) {
544 k = find_optimal_param(sums[i], cnt);
546 all_bits += rice_encode_count(sums[i], cnt, k);
556 static void calc_sums(int pmin, int pmax, uint32_t *data, int n, int pred_order,
557 uint32_t sums[][MAX_PARTITIONS])
561 uint32_t *res, *res_end;
563 /* sums for highest level */
565 res = &data[pred_order];
566 res_end = &data[n >> pmax];
567 for (i = 0; i < parts; i++) {
569 while (res < res_end)
572 res_end += n >> pmax;
574 /* sums for lower levels */
575 for (i = pmax - 1; i >= pmin; i--) {
577 for (j = 0; j < parts; j++)
578 sums[i][j] = sums[i+1][2*j] + sums[i+1][2*j+1];
583 static uint32_t calc_rice_params(RiceContext *rc, int pmin, int pmax,
584 int32_t *data, int n, int pred_order)
587 uint32_t bits[MAX_PARTITION_ORDER+1];
591 uint32_t sums[MAX_PARTITION_ORDER+1][MAX_PARTITIONS];
593 assert(pmin >= 0 && pmin <= MAX_PARTITION_ORDER);
594 assert(pmax >= 0 && pmax <= MAX_PARTITION_ORDER);
595 assert(pmin <= pmax);
597 udata = av_malloc(n * sizeof(uint32_t));
598 for (i = 0; i < n; i++)
599 udata[i] = (2*data[i]) ^ (data[i]>>31);
601 calc_sums(pmin, pmax, udata, n, pred_order, sums);
604 bits[pmin] = UINT32_MAX;
605 for (i = pmin; i <= pmax; i++) {
606 bits[i] = calc_optimal_rice_params(&tmp_rc, i, sums[i], n, pred_order);
607 if (bits[i] <= bits[opt_porder]) {
614 return bits[opt_porder];
618 static int get_max_p_order(int max_porder, int n, int order)
620 int porder = FFMIN(max_porder, av_log2(n^(n-1)));
622 porder = FFMIN(porder, av_log2(n/order));
627 static uint32_t find_subframe_rice_params(FlacEncodeContext *s,
628 FlacSubframe *sub, int pred_order)
630 int pmin = get_max_p_order(s->options.min_partition_order,
631 s->frame.blocksize, pred_order);
632 int pmax = get_max_p_order(s->options.max_partition_order,
633 s->frame.blocksize, pred_order);
635 uint32_t bits = 8 + pred_order * sub->obits + 2 + 4;
636 if (sub->type == FLAC_SUBFRAME_LPC)
637 bits += 4 + 5 + pred_order * s->options.lpc_coeff_precision;
638 bits += calc_rice_params(&sub->rc, pmin, pmax, sub->residual,
639 s->frame.blocksize, pred_order);
644 static void encode_residual_fixed(int32_t *res, const int32_t *smp, int n,
649 for (i = 0; i < order; i++)
653 for (i = order; i < n; i++)
655 } else if (order == 1) {
656 for (i = order; i < n; i++)
657 res[i] = smp[i] - smp[i-1];
658 } else if (order == 2) {
659 int a = smp[order-1] - smp[order-2];
660 for (i = order; i < n; i += 2) {
661 int b = smp[i ] - smp[i-1];
663 a = smp[i+1] - smp[i ];
666 } else if (order == 3) {
667 int a = smp[order-1] - smp[order-2];
668 int c = smp[order-1] - 2*smp[order-2] + smp[order-3];
669 for (i = order; i < n; i += 2) {
670 int b = smp[i ] - smp[i-1];
673 a = smp[i+1] - smp[i ];
678 int a = smp[order-1] - smp[order-2];
679 int c = smp[order-1] - 2*smp[order-2] + smp[order-3];
680 int e = smp[order-1] - 3*smp[order-2] + 3*smp[order-3] - smp[order-4];
681 for (i = order; i < n; i += 2) {
682 int b = smp[i ] - smp[i-1];
686 a = smp[i+1] - smp[i ];
696 int c = coefs[(x)-1];\
702 static av_always_inline void encode_residual_lpc_unrolled(int32_t *res,
703 const int32_t *smp, int n, int order,
704 const int32_t *coefs, int shift, int big)
707 for (i = order; i < n; i += 2) {
708 int s = smp[i-order];
757 res[i ] = smp[i ] - (p0 >> shift);
758 res[i+1] = smp[i+1] - (p1 >> shift);
763 static void encode_residual_lpc(int32_t *res, const int32_t *smp, int n,
764 int order, const int32_t *coefs, int shift)
767 for (i = 0; i < order; i++)
770 for (i = order; i < n; i += 2) {
774 for (j = 0; j < order; j++) {
780 res[i ] = smp[i ] - (p0 >> shift);
781 res[i+1] = smp[i+1] - (p1 >> shift);
785 case 1: encode_residual_lpc_unrolled(res, smp, n, 1, coefs, shift, 0); break;
786 case 2: encode_residual_lpc_unrolled(res, smp, n, 2, coefs, shift, 0); break;
787 case 3: encode_residual_lpc_unrolled(res, smp, n, 3, coefs, shift, 0); break;
788 case 4: encode_residual_lpc_unrolled(res, smp, n, 4, coefs, shift, 0); break;
789 case 5: encode_residual_lpc_unrolled(res, smp, n, 5, coefs, shift, 0); break;
790 case 6: encode_residual_lpc_unrolled(res, smp, n, 6, coefs, shift, 0); break;
791 case 7: encode_residual_lpc_unrolled(res, smp, n, 7, coefs, shift, 0); break;
792 case 8: encode_residual_lpc_unrolled(res, smp, n, 8, coefs, shift, 0); break;
793 default: encode_residual_lpc_unrolled(res, smp, n, order, coefs, shift, 1); break;
799 static int encode_residual_ch(FlacEncodeContext *s, int ch)
802 int min_order, max_order, opt_order, omethod;
805 int32_t coefs[MAX_LPC_ORDER][MAX_LPC_ORDER];
806 int shift[MAX_LPC_ORDER];
810 sub = &frame->subframes[ch];
813 n = frame->blocksize;
816 for (i = 1; i < n; i++)
820 sub->type = sub->type_code = FLAC_SUBFRAME_CONSTANT;
822 return subframe_count_exact(s, sub, 0);
826 if (frame->verbatim_only || n < 5) {
827 sub->type = sub->type_code = FLAC_SUBFRAME_VERBATIM;
828 memcpy(res, smp, n * sizeof(int32_t));
829 return subframe_count_exact(s, sub, 0);
832 min_order = s->options.min_prediction_order;
833 max_order = s->options.max_prediction_order;
834 omethod = s->options.prediction_order_method;
837 sub->type = FLAC_SUBFRAME_FIXED;
838 if (s->options.lpc_type == FF_LPC_TYPE_NONE ||
839 s->options.lpc_type == FF_LPC_TYPE_FIXED || n <= max_order) {
840 uint32_t bits[MAX_FIXED_ORDER+1];
841 if (max_order > MAX_FIXED_ORDER)
842 max_order = MAX_FIXED_ORDER;
844 bits[0] = UINT32_MAX;
845 for (i = min_order; i <= max_order; i++) {
846 encode_residual_fixed(res, smp, n, i);
847 bits[i] = find_subframe_rice_params(s, sub, i);
848 if (bits[i] < bits[opt_order])
851 sub->order = opt_order;
852 sub->type_code = sub->type | sub->order;
853 if (sub->order != max_order) {
854 encode_residual_fixed(res, smp, n, sub->order);
855 find_subframe_rice_params(s, sub, sub->order);
857 return subframe_count_exact(s, sub, sub->order);
861 sub->type = FLAC_SUBFRAME_LPC;
862 opt_order = ff_lpc_calc_coefs(&s->lpc_ctx, smp, n, min_order, max_order,
863 s->options.lpc_coeff_precision, coefs, shift, s->options.lpc_type,
864 s->options.lpc_passes, omethod,
867 if (omethod == ORDER_METHOD_2LEVEL ||
868 omethod == ORDER_METHOD_4LEVEL ||
869 omethod == ORDER_METHOD_8LEVEL) {
870 int levels = 1 << omethod;
871 uint32_t bits[1 << ORDER_METHOD_8LEVEL];
873 int opt_index = levels-1;
874 opt_order = max_order-1;
875 bits[opt_index] = UINT32_MAX;
876 for (i = levels-1; i >= 0; i--) {
877 order = min_order + (((max_order-min_order+1) * (i+1)) / levels)-1;
880 encode_residual_lpc(res, smp, n, order+1, coefs[order], shift[order]);
881 bits[i] = find_subframe_rice_params(s, sub, order+1);
882 if (bits[i] < bits[opt_index]) {
888 } else if (omethod == ORDER_METHOD_SEARCH) {
889 // brute-force optimal order search
890 uint32_t bits[MAX_LPC_ORDER];
892 bits[0] = UINT32_MAX;
893 for (i = min_order-1; i < max_order; i++) {
894 encode_residual_lpc(res, smp, n, i+1, coefs[i], shift[i]);
895 bits[i] = find_subframe_rice_params(s, sub, i+1);
896 if (bits[i] < bits[opt_order])
900 } else if (omethod == ORDER_METHOD_LOG) {
901 uint32_t bits[MAX_LPC_ORDER];
904 opt_order = min_order - 1 + (max_order-min_order)/3;
905 memset(bits, -1, sizeof(bits));
907 for (step = 16; step; step >>= 1) {
908 int last = opt_order;
909 for (i = last-step; i <= last+step; i += step) {
910 if (i < min_order-1 || i >= max_order || bits[i] < UINT32_MAX)
912 encode_residual_lpc(res, smp, n, i+1, coefs[i], shift[i]);
913 bits[i] = find_subframe_rice_params(s, sub, i+1);
914 if (bits[i] < bits[opt_order])
921 sub->order = opt_order;
922 sub->type_code = sub->type | (sub->order-1);
923 sub->shift = shift[sub->order-1];
924 for (i = 0; i < sub->order; i++)
925 sub->coefs[i] = coefs[sub->order-1][i];
927 encode_residual_lpc(res, smp, n, sub->order, sub->coefs, sub->shift);
929 find_subframe_rice_params(s, sub, sub->order);
931 return subframe_count_exact(s, sub, sub->order);
935 static int count_frame_header(FlacEncodeContext *s)
937 uint8_t av_unused tmp;
943 <1> Blocking strategy
944 <4> Block size in inter-channel samples
946 <4> Channel assignment
947 <3> Sample size in bits
952 /* coded frame number */
953 PUT_UTF8(s->frame_count, tmp, count += 8;)
955 /* explicit block size */
956 if (s->frame.bs_code[0] == 6)
958 else if (s->frame.bs_code[0] == 7)
961 /* explicit sample rate */
962 count += ((s->sr_code[0] == 12) + (s->sr_code[0] > 12)) * 8;
964 /* frame header CRC-8 */
971 static int encode_frame(FlacEncodeContext *s)
975 count = count_frame_header(s);
977 for (ch = 0; ch < s->channels; ch++)
978 count += encode_residual_ch(s, ch);
980 count += (8 - (count & 7)) & 7; // byte alignment
981 count += 16; // CRC-16
987 static int estimate_stereo_mode(int32_t *left_ch, int32_t *right_ch, int n)
995 /* calculate sum of 2nd order residual for each channel */
996 sum[0] = sum[1] = sum[2] = sum[3] = 0;
997 for (i = 2; i < n; i++) {
998 lt = left_ch[i] - 2*left_ch[i-1] + left_ch[i-2];
999 rt = right_ch[i] - 2*right_ch[i-1] + right_ch[i-2];
1000 sum[2] += FFABS((lt + rt) >> 1);
1001 sum[3] += FFABS(lt - rt);
1002 sum[0] += FFABS(lt);
1003 sum[1] += FFABS(rt);
1005 /* estimate bit counts */
1006 for (i = 0; i < 4; i++) {
1007 k = find_optimal_param(2 * sum[i], n);
1008 sum[i] = rice_encode_count( 2 * sum[i], n, k);
1011 /* calculate score for each mode */
1012 score[0] = sum[0] + sum[1];
1013 score[1] = sum[0] + sum[3];
1014 score[2] = sum[1] + sum[3];
1015 score[3] = sum[2] + sum[3];
1017 /* return mode with lowest score */
1019 for (i = 1; i < 4; i++)
1020 if (score[i] < score[best])
1023 return FLAC_CHMODE_INDEPENDENT;
1024 } else if (best == 1) {
1025 return FLAC_CHMODE_LEFT_SIDE;
1026 } else if (best == 2) {
1027 return FLAC_CHMODE_RIGHT_SIDE;
1029 return FLAC_CHMODE_MID_SIDE;
1035 * Perform stereo channel decorrelation.
1037 static void channel_decorrelation(FlacEncodeContext *s)
1040 int32_t *left, *right;
1044 n = frame->blocksize;
1045 left = frame->subframes[0].samples;
1046 right = frame->subframes[1].samples;
1048 if (s->channels != 2) {
1049 frame->ch_mode = FLAC_CHMODE_INDEPENDENT;
1053 frame->ch_mode = estimate_stereo_mode(left, right, n);
1055 /* perform decorrelation and adjust bits-per-sample */
1056 if (frame->ch_mode == FLAC_CHMODE_INDEPENDENT)
1058 if (frame->ch_mode == FLAC_CHMODE_MID_SIDE) {
1060 for (i = 0; i < n; i++) {
1062 left[i] = (tmp + right[i]) >> 1;
1063 right[i] = tmp - right[i];
1065 frame->subframes[1].obits++;
1066 } else if (frame->ch_mode == FLAC_CHMODE_LEFT_SIDE) {
1067 for (i = 0; i < n; i++)
1068 right[i] = left[i] - right[i];
1069 frame->subframes[1].obits++;
1071 for (i = 0; i < n; i++)
1072 left[i] -= right[i];
1073 frame->subframes[0].obits++;
1078 static void write_utf8(PutBitContext *pb, uint32_t val)
1081 PUT_UTF8(val, tmp, put_bits(pb, 8, tmp);)
1085 static void write_frame_header(FlacEncodeContext *s)
1092 put_bits(&s->pb, 16, 0xFFF8);
1093 put_bits(&s->pb, 4, frame->bs_code[0]);
1094 put_bits(&s->pb, 4, s->sr_code[0]);
1096 if (frame->ch_mode == FLAC_CHMODE_INDEPENDENT)
1097 put_bits(&s->pb, 4, s->channels-1);
1099 put_bits(&s->pb, 4, frame->ch_mode);
1101 put_bits(&s->pb, 3, 4); /* bits-per-sample code */
1102 put_bits(&s->pb, 1, 0);
1103 write_utf8(&s->pb, s->frame_count);
1105 if (frame->bs_code[0] == 6)
1106 put_bits(&s->pb, 8, frame->bs_code[1]);
1107 else if (frame->bs_code[0] == 7)
1108 put_bits(&s->pb, 16, frame->bs_code[1]);
1110 if (s->sr_code[0] == 12)
1111 put_bits(&s->pb, 8, s->sr_code[1]);
1112 else if (s->sr_code[0] > 12)
1113 put_bits(&s->pb, 16, s->sr_code[1]);
1115 flush_put_bits(&s->pb);
1116 crc = av_crc(av_crc_get_table(AV_CRC_8_ATM), 0, s->pb.buf,
1117 put_bits_count(&s->pb) >> 3);
1118 put_bits(&s->pb, 8, crc);
1122 static void write_subframes(FlacEncodeContext *s)
1126 for (ch = 0; ch < s->channels; ch++) {
1127 FlacSubframe *sub = &s->frame.subframes[ch];
1128 int i, p, porder, psize;
1130 int32_t *res = sub->residual;
1131 int32_t *frame_end = &sub->residual[s->frame.blocksize];
1133 /* subframe header */
1134 put_bits(&s->pb, 1, 0);
1135 put_bits(&s->pb, 6, sub->type_code);
1136 put_bits(&s->pb, 1, 0); /* no wasted bits */
1139 if (sub->type == FLAC_SUBFRAME_CONSTANT) {
1140 put_sbits(&s->pb, sub->obits, res[0]);
1141 } else if (sub->type == FLAC_SUBFRAME_VERBATIM) {
1142 while (res < frame_end)
1143 put_sbits(&s->pb, sub->obits, *res++);
1145 /* warm-up samples */
1146 for (i = 0; i < sub->order; i++)
1147 put_sbits(&s->pb, sub->obits, *res++);
1149 /* LPC coefficients */
1150 if (sub->type == FLAC_SUBFRAME_LPC) {
1151 int cbits = s->options.lpc_coeff_precision;
1152 put_bits( &s->pb, 4, cbits-1);
1153 put_sbits(&s->pb, 5, sub->shift);
1154 for (i = 0; i < sub->order; i++)
1155 put_sbits(&s->pb, cbits, sub->coefs[i]);
1158 /* rice-encoded block */
1159 put_bits(&s->pb, 2, 0);
1161 /* partition order */
1162 porder = sub->rc.porder;
1163 psize = s->frame.blocksize >> porder;
1164 put_bits(&s->pb, 4, porder);
1167 part_end = &sub->residual[psize];
1168 for (p = 0; p < 1 << porder; p++) {
1169 int k = sub->rc.params[p];
1170 put_bits(&s->pb, 4, k);
1171 while (res < part_end)
1172 set_sr_golomb_flac(&s->pb, *res++, k, INT32_MAX, 0);
1173 part_end = FFMIN(frame_end, part_end + psize);
1180 static void write_frame_footer(FlacEncodeContext *s)
1183 flush_put_bits(&s->pb);
1184 crc = av_bswap16(av_crc(av_crc_get_table(AV_CRC_16_ANSI), 0, s->pb.buf,
1185 put_bits_count(&s->pb)>>3));
1186 put_bits(&s->pb, 16, crc);
1187 flush_put_bits(&s->pb);
1191 static int write_frame(FlacEncodeContext *s, uint8_t *frame, int buf_size)
1193 init_put_bits(&s->pb, frame, buf_size);
1194 write_frame_header(s);
1196 write_frame_footer(s);
1197 return put_bits_count(&s->pb) >> 3;
1201 static void update_md5_sum(FlacEncodeContext *s, const int16_t *samples)
1205 for (i = 0; i < s->frame.blocksize * s->channels; i++) {
1206 int16_t smp = av_le2ne16(samples[i]);
1207 av_md5_update(s->md5ctx, (uint8_t *)&smp, 2);
1210 av_md5_update(s->md5ctx, (const uint8_t *)samples, s->frame.blocksize*s->channels*2);
1215 static int flac_encode_frame(AVCodecContext *avctx, uint8_t *frame,
1216 int buf_size, void *data)
1218 FlacEncodeContext *s;
1219 const int16_t *samples = data;
1220 int frame_bytes, out_bytes;
1222 s = avctx->priv_data;
1224 /* when the last block is reached, update the header in extradata */
1226 s->max_framesize = s->max_encoded_framesize;
1227 av_md5_final(s->md5ctx, s->md5sum);
1228 write_streaminfo(s, avctx->extradata);
1232 /* change max_framesize for small final frame */
1233 if (avctx->frame_size < s->frame.blocksize) {
1234 s->max_framesize = ff_flac_get_max_frame_size(avctx->frame_size,
1240 copy_samples(s, samples);
1242 channel_decorrelation(s);
1244 frame_bytes = encode_frame(s);
1246 /* fallback to verbatim mode if the compressed frame is larger than it
1247 would be if encoded uncompressed. */
1248 if (frame_bytes > s->max_framesize) {
1249 s->frame.verbatim_only = 1;
1250 frame_bytes = encode_frame(s);
1253 if (buf_size < frame_bytes) {
1254 av_log(avctx, AV_LOG_ERROR, "output buffer too small\n");
1257 out_bytes = write_frame(s, frame, buf_size);
1260 avctx->coded_frame->pts = s->sample_count;
1261 s->sample_count += avctx->frame_size;
1262 update_md5_sum(s, samples);
1263 if (out_bytes > s->max_encoded_framesize)
1264 s->max_encoded_framesize = out_bytes;
1265 if (out_bytes < s->min_framesize)
1266 s->min_framesize = out_bytes;
1272 static av_cold int flac_encode_close(AVCodecContext *avctx)
1274 if (avctx->priv_data) {
1275 FlacEncodeContext *s = avctx->priv_data;
1276 av_freep(&s->md5ctx);
1277 ff_lpc_end(&s->lpc_ctx);
1279 av_freep(&avctx->extradata);
1280 avctx->extradata_size = 0;
1281 av_freep(&avctx->coded_frame);
1285 #define FLAGS AV_OPT_FLAG_ENCODING_PARAM | AV_OPT_FLAG_AUDIO_PARAM
1286 static const AVOption options[] = {
1287 { "lpc_coeff_precision", "LPC coefficient precision", offsetof(FlacEncodeContext, options.lpc_coeff_precision), AV_OPT_TYPE_INT, {.dbl = 15 }, 0, MAX_LPC_PRECISION, FLAGS },
1288 { "lpc_type", "LPC algorithm", offsetof(FlacEncodeContext, options.lpc_type), AV_OPT_TYPE_INT, {.dbl = FF_LPC_TYPE_DEFAULT }, FF_LPC_TYPE_DEFAULT, FF_LPC_TYPE_NB-1, FLAGS, "lpc_type" },
1289 { "none", NULL, 0, AV_OPT_TYPE_CONST, {.dbl = FF_LPC_TYPE_NONE }, INT_MIN, INT_MAX, FLAGS, "lpc_type" },
1290 { "fixed", NULL, 0, AV_OPT_TYPE_CONST, {.dbl = FF_LPC_TYPE_FIXED }, INT_MIN, INT_MAX, FLAGS, "lpc_type" },
1291 { "levinson", NULL, 0, AV_OPT_TYPE_CONST, {.dbl = FF_LPC_TYPE_LEVINSON }, INT_MIN, INT_MAX, FLAGS, "lpc_type" },
1292 { "cholesky", NULL, 0, AV_OPT_TYPE_CONST, {.dbl = FF_LPC_TYPE_CHOLESKY }, INT_MIN, INT_MAX, FLAGS, "lpc_type" },
1293 { "lpc_passes", "Number of passes to use for Cholesky factorization during LPC analysis", offsetof(FlacEncodeContext, options.lpc_passes), AV_OPT_TYPE_INT, {.dbl = -1 }, INT_MIN, INT_MAX, FLAGS },
1294 { "min_partition_order", NULL, offsetof(FlacEncodeContext, options.min_partition_order), AV_OPT_TYPE_INT, {.dbl = -1 }, -1, MAX_PARTITION_ORDER, FLAGS },
1295 { "max_partition_order", NULL, offsetof(FlacEncodeContext, options.max_partition_order), AV_OPT_TYPE_INT, {.dbl = -1 }, -1, MAX_PARTITION_ORDER, FLAGS },
1296 { "prediction_order_method", "Search method for selecting prediction order", offsetof(FlacEncodeContext, options.prediction_order_method), AV_OPT_TYPE_INT, {.dbl = -1 }, -1, ORDER_METHOD_LOG, FLAGS, "predm" },
1297 { "estimation", NULL, 0, AV_OPT_TYPE_CONST, {.dbl = ORDER_METHOD_EST }, INT_MIN, INT_MAX, FLAGS, "predm" },
1298 { "2level", NULL, 0, AV_OPT_TYPE_CONST, {.dbl = ORDER_METHOD_2LEVEL }, INT_MIN, INT_MAX, FLAGS, "predm" },
1299 { "4level", NULL, 0, AV_OPT_TYPE_CONST, {.dbl = ORDER_METHOD_4LEVEL }, INT_MIN, INT_MAX, FLAGS, "predm" },
1300 { "8level", NULL, 0, AV_OPT_TYPE_CONST, {.dbl = ORDER_METHOD_8LEVEL }, INT_MIN, INT_MAX, FLAGS, "predm" },
1301 { "search", NULL, 0, AV_OPT_TYPE_CONST, {.dbl = ORDER_METHOD_SEARCH }, INT_MIN, INT_MAX, FLAGS, "predm" },
1302 { "log", NULL, 0, AV_OPT_TYPE_CONST, {.dbl = ORDER_METHOD_LOG }, INT_MIN, INT_MAX, FLAGS, "predm" },
1306 static const AVClass flac_encoder_class = {
1308 av_default_item_name,
1310 LIBAVUTIL_VERSION_INT,
1313 AVCodec ff_flac_encoder = {
1315 .type = AVMEDIA_TYPE_AUDIO,
1316 .id = CODEC_ID_FLAC,
1317 .priv_data_size = sizeof(FlacEncodeContext),
1318 .init = flac_encode_init,
1319 .encode = flac_encode_frame,
1320 .close = flac_encode_close,
1321 .capabilities = CODEC_CAP_SMALL_LAST_FRAME | CODEC_CAP_DELAY | CODEC_CAP_LOSSLESS,
1322 .sample_fmts = (const enum AVSampleFormat[]){AV_SAMPLE_FMT_S16,AV_SAMPLE_FMT_NONE},
1323 .long_name = NULL_IF_CONFIG_SMALL("FLAC (Free Lossless Audio Codec)"),
1324 .priv_class = &flac_encoder_class,