3 * Copyright (c) 2006 Justin Ruggles <justin.ruggles@gmail.com>
5 * This file is part of Libav.
7 * Libav 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 * Libav 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 Libav; 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/intmath.h"
24 #include "libavutil/md5.h"
25 #include "libavutil/opt.h"
36 #define FLAC_SUBFRAME_CONSTANT 0
37 #define FLAC_SUBFRAME_VERBATIM 1
38 #define FLAC_SUBFRAME_FIXED 8
39 #define FLAC_SUBFRAME_LPC 32
41 #define MAX_FIXED_ORDER 4
42 #define MAX_PARTITION_ORDER 8
43 #define MAX_PARTITIONS (1 << MAX_PARTITION_ORDER)
44 #define MAX_LPC_PRECISION 15
45 #define MAX_LPC_SHIFT 15
46 #define MAX_RICE_PARAM 14
48 typedef struct CompressionOptions {
49 int compression_level;
51 enum FFLPCType lpc_type;
53 int lpc_coeff_precision;
54 int min_prediction_order;
55 int max_prediction_order;
56 int prediction_order_method;
57 int min_partition_order;
58 int max_partition_order;
62 typedef struct RiceContext {
64 int params[MAX_PARTITIONS];
67 typedef struct FlacSubframe {
73 int32_t coefs[MAX_LPC_ORDER];
76 int32_t samples[FLAC_MAX_BLOCKSIZE];
77 int32_t residual[FLAC_MAX_BLOCKSIZE+1];
80 typedef struct FlacFrame {
81 FlacSubframe subframes[FLAC_MAX_CHANNELS];
89 typedef struct FlacEncodeContext {
98 int max_encoded_framesize;
100 uint64_t sample_count;
103 CompressionOptions options;
104 AVCodecContext *avctx;
106 struct AVMD5 *md5ctx;
108 unsigned int md5_buffer_size;
110 FLACDSPContext flac_dsp;
115 * Write streaminfo metadata block to byte array.
117 static void write_streaminfo(FlacEncodeContext *s, uint8_t *header)
121 memset(header, 0, FLAC_STREAMINFO_SIZE);
122 init_put_bits(&pb, header, FLAC_STREAMINFO_SIZE);
124 /* streaminfo metadata block */
125 put_bits(&pb, 16, s->max_blocksize);
126 put_bits(&pb, 16, s->max_blocksize);
127 put_bits(&pb, 24, s->min_framesize);
128 put_bits(&pb, 24, s->max_framesize);
129 put_bits(&pb, 20, s->samplerate);
130 put_bits(&pb, 3, s->channels-1);
131 put_bits(&pb, 5, 15); /* bits per sample - 1 */
132 /* write 36-bit sample count in 2 put_bits() calls */
133 put_bits(&pb, 24, (s->sample_count & 0xFFFFFF000LL) >> 12);
134 put_bits(&pb, 12, s->sample_count & 0x000000FFFLL);
136 memcpy(&header[18], s->md5sum, 16);
141 * Set blocksize based on samplerate.
142 * Choose the closest predefined blocksize >= BLOCK_TIME_MS milliseconds.
144 static int select_blocksize(int samplerate, int block_time_ms)
150 assert(samplerate > 0);
151 blocksize = ff_flac_blocksize_table[1];
152 target = (samplerate * block_time_ms) / 1000;
153 for (i = 0; i < 16; i++) {
154 if (target >= ff_flac_blocksize_table[i] &&
155 ff_flac_blocksize_table[i] > blocksize) {
156 blocksize = ff_flac_blocksize_table[i];
163 static av_cold void dprint_compression_options(FlacEncodeContext *s)
165 AVCodecContext *avctx = s->avctx;
166 CompressionOptions *opt = &s->options;
168 av_log(avctx, AV_LOG_DEBUG, " compression: %d\n", opt->compression_level);
170 switch (opt->lpc_type) {
171 case FF_LPC_TYPE_NONE:
172 av_log(avctx, AV_LOG_DEBUG, " lpc type: None\n");
174 case FF_LPC_TYPE_FIXED:
175 av_log(avctx, AV_LOG_DEBUG, " lpc type: Fixed pre-defined coefficients\n");
177 case FF_LPC_TYPE_LEVINSON:
178 av_log(avctx, AV_LOG_DEBUG, " lpc type: Levinson-Durbin recursion with Welch window\n");
180 case FF_LPC_TYPE_CHOLESKY:
181 av_log(avctx, AV_LOG_DEBUG, " lpc type: Cholesky factorization, %d pass%s\n",
182 opt->lpc_passes, opt->lpc_passes == 1 ? "" : "es");
186 av_log(avctx, AV_LOG_DEBUG, " prediction order: %d, %d\n",
187 opt->min_prediction_order, opt->max_prediction_order);
189 switch (opt->prediction_order_method) {
190 case ORDER_METHOD_EST:
191 av_log(avctx, AV_LOG_DEBUG, " order method: %s\n", "estimate");
193 case ORDER_METHOD_2LEVEL:
194 av_log(avctx, AV_LOG_DEBUG, " order method: %s\n", "2-level");
196 case ORDER_METHOD_4LEVEL:
197 av_log(avctx, AV_LOG_DEBUG, " order method: %s\n", "4-level");
199 case ORDER_METHOD_8LEVEL:
200 av_log(avctx, AV_LOG_DEBUG, " order method: %s\n", "8-level");
202 case ORDER_METHOD_SEARCH:
203 av_log(avctx, AV_LOG_DEBUG, " order method: %s\n", "full search");
205 case ORDER_METHOD_LOG:
206 av_log(avctx, AV_LOG_DEBUG, " order method: %s\n", "log search");
211 av_log(avctx, AV_LOG_DEBUG, " partition order: %d, %d\n",
212 opt->min_partition_order, opt->max_partition_order);
214 av_log(avctx, AV_LOG_DEBUG, " block size: %d\n", avctx->frame_size);
216 av_log(avctx, AV_LOG_DEBUG, " lpc precision: %d\n",
217 opt->lpc_coeff_precision);
221 static av_cold int flac_encode_init(AVCodecContext *avctx)
223 int freq = avctx->sample_rate;
224 int channels = avctx->channels;
225 FlacEncodeContext *s = avctx->priv_data;
231 if (avctx->sample_fmt != AV_SAMPLE_FMT_S16)
234 if (channels < 1 || channels > FLAC_MAX_CHANNELS)
236 s->channels = channels;
238 /* find samplerate in table */
241 for (i = 4; i < 12; i++) {
242 if (freq == ff_flac_sample_rate_table[i]) {
243 s->samplerate = ff_flac_sample_rate_table[i];
249 /* if not in table, samplerate is non-standard */
251 if (freq % 1000 == 0 && freq < 255000) {
253 s->sr_code[1] = freq / 1000;
254 } else if (freq % 10 == 0 && freq < 655350) {
256 s->sr_code[1] = freq / 10;
257 } else if (freq < 65535) {
259 s->sr_code[1] = freq;
263 s->samplerate = freq;
266 /* set compression option defaults based on avctx->compression_level */
267 if (avctx->compression_level < 0)
268 s->options.compression_level = 5;
270 s->options.compression_level = avctx->compression_level;
272 level = s->options.compression_level;
274 av_log(avctx, AV_LOG_ERROR, "invalid compression level: %d\n",
275 s->options.compression_level);
279 s->options.block_time_ms = ((int[]){ 27, 27, 27,105,105,105,105,105,105,105,105,105,105})[level];
281 if (s->options.lpc_type == FF_LPC_TYPE_DEFAULT)
282 s->options.lpc_type = ((int[]){ FF_LPC_TYPE_FIXED, FF_LPC_TYPE_FIXED, FF_LPC_TYPE_FIXED,
283 FF_LPC_TYPE_LEVINSON, FF_LPC_TYPE_LEVINSON, FF_LPC_TYPE_LEVINSON,
284 FF_LPC_TYPE_LEVINSON, FF_LPC_TYPE_LEVINSON, FF_LPC_TYPE_LEVINSON,
285 FF_LPC_TYPE_LEVINSON, FF_LPC_TYPE_LEVINSON, FF_LPC_TYPE_LEVINSON,
286 FF_LPC_TYPE_LEVINSON})[level];
288 s->options.min_prediction_order = ((int[]){ 2, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1})[level];
289 s->options.max_prediction_order = ((int[]){ 3, 4, 4, 6, 8, 8, 8, 8, 12, 12, 12, 32, 32})[level];
291 if (s->options.prediction_order_method < 0)
292 s->options.prediction_order_method = ((int[]){ ORDER_METHOD_EST, ORDER_METHOD_EST, ORDER_METHOD_EST,
293 ORDER_METHOD_EST, ORDER_METHOD_EST, ORDER_METHOD_EST,
294 ORDER_METHOD_4LEVEL, ORDER_METHOD_LOG, ORDER_METHOD_4LEVEL,
295 ORDER_METHOD_LOG, ORDER_METHOD_SEARCH, ORDER_METHOD_LOG,
296 ORDER_METHOD_SEARCH})[level];
298 if (s->options.min_partition_order > s->options.max_partition_order) {
299 av_log(avctx, AV_LOG_ERROR, "invalid partition orders: min=%d max=%d\n",
300 s->options.min_partition_order, s->options.max_partition_order);
301 return AVERROR(EINVAL);
303 if (s->options.min_partition_order < 0)
304 s->options.min_partition_order = ((int[]){ 2, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0})[level];
305 if (s->options.max_partition_order < 0)
306 s->options.max_partition_order = ((int[]){ 2, 2, 3, 3, 3, 8, 8, 8, 8, 8, 8, 8, 8})[level];
308 if (s->options.lpc_type == FF_LPC_TYPE_NONE) {
309 s->options.min_prediction_order = 0;
310 } else if (avctx->min_prediction_order >= 0) {
311 if (s->options.lpc_type == FF_LPC_TYPE_FIXED) {
312 if (avctx->min_prediction_order > MAX_FIXED_ORDER) {
313 av_log(avctx, AV_LOG_ERROR, "invalid min prediction order: %d\n",
314 avctx->min_prediction_order);
317 } else if (avctx->min_prediction_order < MIN_LPC_ORDER ||
318 avctx->min_prediction_order > MAX_LPC_ORDER) {
319 av_log(avctx, AV_LOG_ERROR, "invalid min prediction order: %d\n",
320 avctx->min_prediction_order);
323 s->options.min_prediction_order = avctx->min_prediction_order;
325 if (s->options.lpc_type == FF_LPC_TYPE_NONE) {
326 s->options.max_prediction_order = 0;
327 } else if (avctx->max_prediction_order >= 0) {
328 if (s->options.lpc_type == FF_LPC_TYPE_FIXED) {
329 if (avctx->max_prediction_order > MAX_FIXED_ORDER) {
330 av_log(avctx, AV_LOG_ERROR, "invalid max prediction order: %d\n",
331 avctx->max_prediction_order);
334 } else if (avctx->max_prediction_order < MIN_LPC_ORDER ||
335 avctx->max_prediction_order > MAX_LPC_ORDER) {
336 av_log(avctx, AV_LOG_ERROR, "invalid max prediction order: %d\n",
337 avctx->max_prediction_order);
340 s->options.max_prediction_order = avctx->max_prediction_order;
342 if (s->options.max_prediction_order < s->options.min_prediction_order) {
343 av_log(avctx, AV_LOG_ERROR, "invalid prediction orders: min=%d max=%d\n",
344 s->options.min_prediction_order, s->options.max_prediction_order);
348 if (avctx->frame_size > 0) {
349 if (avctx->frame_size < FLAC_MIN_BLOCKSIZE ||
350 avctx->frame_size > FLAC_MAX_BLOCKSIZE) {
351 av_log(avctx, AV_LOG_ERROR, "invalid block size: %d\n",
356 s->avctx->frame_size = select_blocksize(s->samplerate, s->options.block_time_ms);
358 s->max_blocksize = s->avctx->frame_size;
360 /* set maximum encoded frame size in verbatim mode */
361 s->max_framesize = ff_flac_get_max_frame_size(s->avctx->frame_size,
364 /* initialize MD5 context */
365 s->md5ctx = av_md5_alloc();
367 return AVERROR(ENOMEM);
368 av_md5_init(s->md5ctx);
370 streaminfo = av_malloc(FLAC_STREAMINFO_SIZE);
372 return AVERROR(ENOMEM);
373 write_streaminfo(s, streaminfo);
374 avctx->extradata = streaminfo;
375 avctx->extradata_size = FLAC_STREAMINFO_SIZE;
378 s->min_framesize = s->max_framesize;
380 #if FF_API_OLD_ENCODE_AUDIO
381 avctx->coded_frame = avcodec_alloc_frame();
382 if (!avctx->coded_frame)
383 return AVERROR(ENOMEM);
386 ret = ff_lpc_init(&s->lpc_ctx, avctx->frame_size,
387 s->options.max_prediction_order, FF_LPC_TYPE_LEVINSON);
389 ff_dsputil_init(&s->dsp, avctx);
390 ff_flacdsp_init(&s->flac_dsp, avctx->sample_fmt, 16);
392 dprint_compression_options(s);
398 static void init_frame(FlacEncodeContext *s, int nb_samples)
405 for (i = 0; i < 16; i++) {
406 if (nb_samples == ff_flac_blocksize_table[i]) {
407 frame->blocksize = ff_flac_blocksize_table[i];
408 frame->bs_code[0] = i;
409 frame->bs_code[1] = 0;
414 frame->blocksize = nb_samples;
415 if (frame->blocksize <= 256) {
416 frame->bs_code[0] = 6;
417 frame->bs_code[1] = frame->blocksize-1;
419 frame->bs_code[0] = 7;
420 frame->bs_code[1] = frame->blocksize-1;
424 for (ch = 0; ch < s->channels; ch++) {
425 frame->subframes[ch].wasted = 0;
426 frame->subframes[ch].obits = 16;
429 frame->verbatim_only = 0;
434 * Copy channel-interleaved input samples into separate subframes.
436 static void copy_samples(FlacEncodeContext *s, const int16_t *samples)
442 for (i = 0, j = 0; i < frame->blocksize; i++)
443 for (ch = 0; ch < s->channels; ch++, j++)
444 frame->subframes[ch].samples[i] = samples[j];
448 static uint64_t rice_count_exact(int32_t *res, int n, int k)
453 for (i = 0; i < n; i++) {
454 int32_t v = -2 * res[i] - 1;
456 count += (v >> k) + 1 + k;
462 static uint64_t subframe_count_exact(FlacEncodeContext *s, FlacSubframe *sub,
465 int p, porder, psize;
469 /* subframe header */
473 if (sub->type == FLAC_SUBFRAME_CONSTANT) {
475 } else if (sub->type == FLAC_SUBFRAME_VERBATIM) {
476 count += s->frame.blocksize * sub->obits;
478 /* warm-up samples */
479 count += pred_order * sub->obits;
481 /* LPC coefficients */
482 if (sub->type == FLAC_SUBFRAME_LPC)
483 count += 4 + 5 + pred_order * s->options.lpc_coeff_precision;
485 /* rice-encoded block */
488 /* partition order */
489 porder = sub->rc.porder;
490 psize = s->frame.blocksize >> porder;
496 for (p = 0; p < 1 << porder; p++) {
497 int k = sub->rc.params[p];
499 count += rice_count_exact(&sub->residual[i], part_end - i, k);
501 part_end = FFMIN(s->frame.blocksize, part_end + psize);
509 #define rice_encode_count(sum, n, k) (((n)*((k)+1))+((sum-(n>>1))>>(k)))
512 * Solve for d/dk(rice_encode_count) = n-((sum-(n>>1))>>(k+1)) = 0.
514 static int find_optimal_param(uint64_t sum, int n)
521 sum2 = sum - (n >> 1);
522 k = av_log2(av_clipl_int32(sum2 / n));
523 return FFMIN(k, MAX_RICE_PARAM);
527 static uint64_t calc_optimal_rice_params(RiceContext *rc, int porder,
528 uint64_t *sums, int n, int pred_order)
534 part = (1 << porder);
537 cnt = (n >> porder) - pred_order;
538 for (i = 0; i < part; i++) {
539 k = find_optimal_param(sums[i], cnt);
541 all_bits += rice_encode_count(sums[i], cnt, k);
551 static void calc_sums(int pmin, int pmax, uint32_t *data, int n, int pred_order,
552 uint64_t sums[][MAX_PARTITIONS])
556 uint32_t *res, *res_end;
558 /* sums for highest level */
560 res = &data[pred_order];
561 res_end = &data[n >> pmax];
562 for (i = 0; i < parts; i++) {
564 while (res < res_end)
567 res_end += n >> pmax;
569 /* sums for lower levels */
570 for (i = pmax - 1; i >= pmin; i--) {
572 for (j = 0; j < parts; j++)
573 sums[i][j] = sums[i+1][2*j] + sums[i+1][2*j+1];
578 static uint64_t calc_rice_params(RiceContext *rc, int pmin, int pmax,
579 int32_t *data, int n, int pred_order)
582 uint64_t bits[MAX_PARTITION_ORDER+1];
586 uint64_t sums[MAX_PARTITION_ORDER+1][MAX_PARTITIONS];
588 assert(pmin >= 0 && pmin <= MAX_PARTITION_ORDER);
589 assert(pmax >= 0 && pmax <= MAX_PARTITION_ORDER);
590 assert(pmin <= pmax);
592 udata = av_malloc(n * sizeof(uint32_t));
593 for (i = 0; i < n; i++)
594 udata[i] = (2*data[i]) ^ (data[i]>>31);
596 calc_sums(pmin, pmax, udata, n, pred_order, sums);
599 bits[pmin] = UINT32_MAX;
600 for (i = pmin; i <= pmax; i++) {
601 bits[i] = calc_optimal_rice_params(&tmp_rc, i, sums[i], n, pred_order);
602 if (bits[i] <= bits[opt_porder]) {
609 return bits[opt_porder];
613 static int get_max_p_order(int max_porder, int n, int order)
615 int porder = FFMIN(max_porder, av_log2(n^(n-1)));
617 porder = FFMIN(porder, av_log2(n/order));
622 static uint64_t find_subframe_rice_params(FlacEncodeContext *s,
623 FlacSubframe *sub, int pred_order)
625 int pmin = get_max_p_order(s->options.min_partition_order,
626 s->frame.blocksize, pred_order);
627 int pmax = get_max_p_order(s->options.max_partition_order,
628 s->frame.blocksize, pred_order);
630 uint64_t bits = 8 + pred_order * sub->obits + 2 + 4;
631 if (sub->type == FLAC_SUBFRAME_LPC)
632 bits += 4 + 5 + pred_order * s->options.lpc_coeff_precision;
633 bits += calc_rice_params(&sub->rc, pmin, pmax, sub->residual,
634 s->frame.blocksize, pred_order);
639 static void encode_residual_fixed(int32_t *res, const int32_t *smp, int n,
644 for (i = 0; i < order; i++)
648 for (i = order; i < n; i++)
650 } else if (order == 1) {
651 for (i = order; i < n; i++)
652 res[i] = smp[i] - smp[i-1];
653 } else if (order == 2) {
654 int a = smp[order-1] - smp[order-2];
655 for (i = order; i < n; i += 2) {
656 int b = smp[i ] - smp[i-1];
658 a = smp[i+1] - smp[i ];
661 } else if (order == 3) {
662 int a = smp[order-1] - smp[order-2];
663 int c = smp[order-1] - 2*smp[order-2] + smp[order-3];
664 for (i = order; i < n; i += 2) {
665 int b = smp[i ] - smp[i-1];
668 a = smp[i+1] - smp[i ];
673 int a = smp[order-1] - smp[order-2];
674 int c = smp[order-1] - 2*smp[order-2] + smp[order-3];
675 int e = smp[order-1] - 3*smp[order-2] + 3*smp[order-3] - smp[order-4];
676 for (i = order; i < n; i += 2) {
677 int b = smp[i ] - smp[i-1];
681 a = smp[i+1] - smp[i ];
690 static int encode_residual_ch(FlacEncodeContext *s, int ch)
693 int min_order, max_order, opt_order, omethod;
696 int32_t coefs[MAX_LPC_ORDER][MAX_LPC_ORDER];
697 int shift[MAX_LPC_ORDER];
701 sub = &frame->subframes[ch];
704 n = frame->blocksize;
707 for (i = 1; i < n; i++)
711 sub->type = sub->type_code = FLAC_SUBFRAME_CONSTANT;
713 return subframe_count_exact(s, sub, 0);
717 if (frame->verbatim_only || n < 5) {
718 sub->type = sub->type_code = FLAC_SUBFRAME_VERBATIM;
719 memcpy(res, smp, n * sizeof(int32_t));
720 return subframe_count_exact(s, sub, 0);
723 min_order = s->options.min_prediction_order;
724 max_order = s->options.max_prediction_order;
725 omethod = s->options.prediction_order_method;
728 sub->type = FLAC_SUBFRAME_FIXED;
729 if (s->options.lpc_type == FF_LPC_TYPE_NONE ||
730 s->options.lpc_type == FF_LPC_TYPE_FIXED || n <= max_order) {
731 uint64_t bits[MAX_FIXED_ORDER+1];
732 if (max_order > MAX_FIXED_ORDER)
733 max_order = MAX_FIXED_ORDER;
735 bits[0] = UINT32_MAX;
736 for (i = min_order; i <= max_order; i++) {
737 encode_residual_fixed(res, smp, n, i);
738 bits[i] = find_subframe_rice_params(s, sub, i);
739 if (bits[i] < bits[opt_order])
742 sub->order = opt_order;
743 sub->type_code = sub->type | sub->order;
744 if (sub->order != max_order) {
745 encode_residual_fixed(res, smp, n, sub->order);
746 find_subframe_rice_params(s, sub, sub->order);
748 return subframe_count_exact(s, sub, sub->order);
752 sub->type = FLAC_SUBFRAME_LPC;
753 opt_order = ff_lpc_calc_coefs(&s->lpc_ctx, smp, n, min_order, max_order,
754 s->options.lpc_coeff_precision, coefs, shift, s->options.lpc_type,
755 s->options.lpc_passes, omethod,
758 if (omethod == ORDER_METHOD_2LEVEL ||
759 omethod == ORDER_METHOD_4LEVEL ||
760 omethod == ORDER_METHOD_8LEVEL) {
761 int levels = 1 << omethod;
762 uint64_t bits[1 << ORDER_METHOD_8LEVEL];
764 int opt_index = levels-1;
765 opt_order = max_order-1;
766 bits[opt_index] = UINT32_MAX;
767 for (i = levels-1; i >= 0; i--) {
768 order = min_order + (((max_order-min_order+1) * (i+1)) / levels)-1;
771 s->flac_dsp.lpc_encode(res, smp, n, order+1, coefs[order],
773 bits[i] = find_subframe_rice_params(s, sub, order+1);
774 if (bits[i] < bits[opt_index]) {
780 } else if (omethod == ORDER_METHOD_SEARCH) {
781 // brute-force optimal order search
782 uint64_t bits[MAX_LPC_ORDER];
784 bits[0] = UINT32_MAX;
785 for (i = min_order-1; i < max_order; i++) {
786 s->flac_dsp.lpc_encode(res, smp, n, i+1, coefs[i], shift[i]);
787 bits[i] = find_subframe_rice_params(s, sub, i+1);
788 if (bits[i] < bits[opt_order])
792 } else if (omethod == ORDER_METHOD_LOG) {
793 uint64_t bits[MAX_LPC_ORDER];
796 opt_order = min_order - 1 + (max_order-min_order)/3;
797 memset(bits, -1, sizeof(bits));
799 for (step = 16; step; step >>= 1) {
800 int last = opt_order;
801 for (i = last-step; i <= last+step; i += step) {
802 if (i < min_order-1 || i >= max_order || bits[i] < UINT32_MAX)
804 s->flac_dsp.lpc_encode(res, smp, n, i+1, coefs[i], shift[i]);
805 bits[i] = find_subframe_rice_params(s, sub, i+1);
806 if (bits[i] < bits[opt_order])
813 sub->order = opt_order;
814 sub->type_code = sub->type | (sub->order-1);
815 sub->shift = shift[sub->order-1];
816 for (i = 0; i < sub->order; i++)
817 sub->coefs[i] = coefs[sub->order-1][i];
819 s->flac_dsp.lpc_encode(res, smp, n, sub->order, sub->coefs, sub->shift);
821 find_subframe_rice_params(s, sub, sub->order);
823 return subframe_count_exact(s, sub, sub->order);
827 static int count_frame_header(FlacEncodeContext *s)
829 uint8_t av_unused tmp;
835 <1> Blocking strategy
836 <4> Block size in inter-channel samples
838 <4> Channel assignment
839 <3> Sample size in bits
844 /* coded frame number */
845 PUT_UTF8(s->frame_count, tmp, count += 8;)
847 /* explicit block size */
848 if (s->frame.bs_code[0] == 6)
850 else if (s->frame.bs_code[0] == 7)
853 /* explicit sample rate */
854 count += ((s->sr_code[0] == 12) + (s->sr_code[0] > 12)) * 8;
856 /* frame header CRC-8 */
863 static int encode_frame(FlacEncodeContext *s)
868 count = count_frame_header(s);
870 for (ch = 0; ch < s->channels; ch++)
871 count += encode_residual_ch(s, ch);
873 count += (8 - (count & 7)) & 7; // byte alignment
874 count += 16; // CRC-16
883 static void remove_wasted_bits(FlacEncodeContext *s)
887 for (ch = 0; ch < s->channels; ch++) {
888 FlacSubframe *sub = &s->frame.subframes[ch];
891 for (i = 0; i < s->frame.blocksize; i++) {
892 v |= sub->samples[i];
900 for (i = 0; i < s->frame.blocksize; i++)
901 sub->samples[i] >>= v;
910 static int estimate_stereo_mode(int32_t *left_ch, int32_t *right_ch, int n)
918 /* calculate sum of 2nd order residual for each channel */
919 sum[0] = sum[1] = sum[2] = sum[3] = 0;
920 for (i = 2; i < n; i++) {
921 lt = left_ch[i] - 2*left_ch[i-1] + left_ch[i-2];
922 rt = right_ch[i] - 2*right_ch[i-1] + right_ch[i-2];
923 sum[2] += FFABS((lt + rt) >> 1);
924 sum[3] += FFABS(lt - rt);
928 /* estimate bit counts */
929 for (i = 0; i < 4; i++) {
930 k = find_optimal_param(2 * sum[i], n);
931 sum[i] = rice_encode_count( 2 * sum[i], n, k);
934 /* calculate score for each mode */
935 score[0] = sum[0] + sum[1];
936 score[1] = sum[0] + sum[3];
937 score[2] = sum[1] + sum[3];
938 score[3] = sum[2] + sum[3];
940 /* return mode with lowest score */
942 for (i = 1; i < 4; i++)
943 if (score[i] < score[best])
951 * Perform stereo channel decorrelation.
953 static void channel_decorrelation(FlacEncodeContext *s)
956 int32_t *left, *right;
960 n = frame->blocksize;
961 left = frame->subframes[0].samples;
962 right = frame->subframes[1].samples;
964 if (s->channels != 2) {
965 frame->ch_mode = FLAC_CHMODE_INDEPENDENT;
969 if (s->options.ch_mode < 0)
970 frame->ch_mode = estimate_stereo_mode(left, right, n);
972 frame->ch_mode = s->options.ch_mode;
974 /* perform decorrelation and adjust bits-per-sample */
975 if (frame->ch_mode == FLAC_CHMODE_INDEPENDENT)
977 if (frame->ch_mode == FLAC_CHMODE_MID_SIDE) {
979 for (i = 0; i < n; i++) {
981 left[i] = (tmp + right[i]) >> 1;
982 right[i] = tmp - right[i];
984 frame->subframes[1].obits++;
985 } else if (frame->ch_mode == FLAC_CHMODE_LEFT_SIDE) {
986 for (i = 0; i < n; i++)
987 right[i] = left[i] - right[i];
988 frame->subframes[1].obits++;
990 for (i = 0; i < n; i++)
992 frame->subframes[0].obits++;
997 static void write_utf8(PutBitContext *pb, uint32_t val)
1000 PUT_UTF8(val, tmp, put_bits(pb, 8, tmp);)
1004 static void write_frame_header(FlacEncodeContext *s)
1011 put_bits(&s->pb, 16, 0xFFF8);
1012 put_bits(&s->pb, 4, frame->bs_code[0]);
1013 put_bits(&s->pb, 4, s->sr_code[0]);
1015 if (frame->ch_mode == FLAC_CHMODE_INDEPENDENT)
1016 put_bits(&s->pb, 4, s->channels-1);
1018 put_bits(&s->pb, 4, frame->ch_mode + FLAC_MAX_CHANNELS - 1);
1020 put_bits(&s->pb, 3, 4); /* bits-per-sample code */
1021 put_bits(&s->pb, 1, 0);
1022 write_utf8(&s->pb, s->frame_count);
1024 if (frame->bs_code[0] == 6)
1025 put_bits(&s->pb, 8, frame->bs_code[1]);
1026 else if (frame->bs_code[0] == 7)
1027 put_bits(&s->pb, 16, frame->bs_code[1]);
1029 if (s->sr_code[0] == 12)
1030 put_bits(&s->pb, 8, s->sr_code[1]);
1031 else if (s->sr_code[0] > 12)
1032 put_bits(&s->pb, 16, s->sr_code[1]);
1034 flush_put_bits(&s->pb);
1035 crc = av_crc(av_crc_get_table(AV_CRC_8_ATM), 0, s->pb.buf,
1036 put_bits_count(&s->pb) >> 3);
1037 put_bits(&s->pb, 8, crc);
1041 static void write_subframes(FlacEncodeContext *s)
1045 for (ch = 0; ch < s->channels; ch++) {
1046 FlacSubframe *sub = &s->frame.subframes[ch];
1047 int i, p, porder, psize;
1049 int32_t *res = sub->residual;
1050 int32_t *frame_end = &sub->residual[s->frame.blocksize];
1052 /* subframe header */
1053 put_bits(&s->pb, 1, 0);
1054 put_bits(&s->pb, 6, sub->type_code);
1055 put_bits(&s->pb, 1, !!sub->wasted);
1057 put_bits(&s->pb, sub->wasted, 1);
1060 if (sub->type == FLAC_SUBFRAME_CONSTANT) {
1061 put_sbits(&s->pb, sub->obits, res[0]);
1062 } else if (sub->type == FLAC_SUBFRAME_VERBATIM) {
1063 while (res < frame_end)
1064 put_sbits(&s->pb, sub->obits, *res++);
1066 /* warm-up samples */
1067 for (i = 0; i < sub->order; i++)
1068 put_sbits(&s->pb, sub->obits, *res++);
1070 /* LPC coefficients */
1071 if (sub->type == FLAC_SUBFRAME_LPC) {
1072 int cbits = s->options.lpc_coeff_precision;
1073 put_bits( &s->pb, 4, cbits-1);
1074 put_sbits(&s->pb, 5, sub->shift);
1075 for (i = 0; i < sub->order; i++)
1076 put_sbits(&s->pb, cbits, sub->coefs[i]);
1079 /* rice-encoded block */
1080 put_bits(&s->pb, 2, 0);
1082 /* partition order */
1083 porder = sub->rc.porder;
1084 psize = s->frame.blocksize >> porder;
1085 put_bits(&s->pb, 4, porder);
1088 part_end = &sub->residual[psize];
1089 for (p = 0; p < 1 << porder; p++) {
1090 int k = sub->rc.params[p];
1091 put_bits(&s->pb, 4, k);
1092 while (res < part_end)
1093 set_sr_golomb_flac(&s->pb, *res++, k, INT32_MAX, 0);
1094 part_end = FFMIN(frame_end, part_end + psize);
1101 static void write_frame_footer(FlacEncodeContext *s)
1104 flush_put_bits(&s->pb);
1105 crc = av_bswap16(av_crc(av_crc_get_table(AV_CRC_16_ANSI), 0, s->pb.buf,
1106 put_bits_count(&s->pb)>>3));
1107 put_bits(&s->pb, 16, crc);
1108 flush_put_bits(&s->pb);
1112 static int write_frame(FlacEncodeContext *s, AVPacket *avpkt)
1114 init_put_bits(&s->pb, avpkt->data, avpkt->size);
1115 write_frame_header(s);
1117 write_frame_footer(s);
1118 return put_bits_count(&s->pb) >> 3;
1122 static int update_md5_sum(FlacEncodeContext *s, const int16_t *samples)
1125 int buf_size = s->frame.blocksize * s->channels * 2;
1127 if (HAVE_BIGENDIAN) {
1128 av_fast_malloc(&s->md5_buffer, &s->md5_buffer_size, buf_size);
1130 return AVERROR(ENOMEM);
1133 buf = (const uint8_t *)samples;
1135 s->dsp.bswap16_buf((uint16_t *)s->md5_buffer,
1136 (const uint16_t *)samples, buf_size / 2);
1137 buf = s->md5_buffer;
1139 av_md5_update(s->md5ctx, buf, buf_size);
1145 static int flac_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
1146 const AVFrame *frame, int *got_packet_ptr)
1148 FlacEncodeContext *s;
1149 const int16_t *samples;
1150 int frame_bytes, out_bytes, ret;
1152 s = avctx->priv_data;
1154 /* when the last block is reached, update the header in extradata */
1156 s->max_framesize = s->max_encoded_framesize;
1157 av_md5_final(s->md5ctx, s->md5sum);
1158 write_streaminfo(s, avctx->extradata);
1161 samples = (const int16_t *)frame->data[0];
1163 /* change max_framesize for small final frame */
1164 if (frame->nb_samples < s->frame.blocksize) {
1165 s->max_framesize = ff_flac_get_max_frame_size(frame->nb_samples,
1169 init_frame(s, frame->nb_samples);
1171 copy_samples(s, samples);
1173 channel_decorrelation(s);
1175 remove_wasted_bits(s);
1177 frame_bytes = encode_frame(s);
1179 /* fallback to verbatim mode if the compressed frame is larger than it
1180 would be if encoded uncompressed. */
1181 if (frame_bytes < 0 || frame_bytes > s->max_framesize) {
1182 s->frame.verbatim_only = 1;
1183 frame_bytes = encode_frame(s);
1184 if (frame_bytes < 0) {
1185 av_log(avctx, AV_LOG_ERROR, "Bad frame count\n");
1190 if ((ret = ff_alloc_packet(avpkt, frame_bytes))) {
1191 av_log(avctx, AV_LOG_ERROR, "Error getting output packet\n");
1195 out_bytes = write_frame(s, avpkt);
1198 s->sample_count += frame->nb_samples;
1199 if ((ret = update_md5_sum(s, samples)) < 0) {
1200 av_log(avctx, AV_LOG_ERROR, "Error updating MD5 checksum\n");
1203 if (out_bytes > s->max_encoded_framesize)
1204 s->max_encoded_framesize = out_bytes;
1205 if (out_bytes < s->min_framesize)
1206 s->min_framesize = out_bytes;
1208 avpkt->pts = frame->pts;
1209 avpkt->duration = ff_samples_to_time_base(avctx, frame->nb_samples);
1210 avpkt->size = out_bytes;
1211 *got_packet_ptr = 1;
1216 static av_cold int flac_encode_close(AVCodecContext *avctx)
1218 if (avctx->priv_data) {
1219 FlacEncodeContext *s = avctx->priv_data;
1220 av_freep(&s->md5ctx);
1221 av_freep(&s->md5_buffer);
1222 ff_lpc_end(&s->lpc_ctx);
1224 av_freep(&avctx->extradata);
1225 avctx->extradata_size = 0;
1226 #if FF_API_OLD_ENCODE_AUDIO
1227 av_freep(&avctx->coded_frame);
1232 #define FLAGS AV_OPT_FLAG_ENCODING_PARAM | AV_OPT_FLAG_AUDIO_PARAM
1233 static const AVOption options[] = {
1234 { "lpc_coeff_precision", "LPC coefficient precision", offsetof(FlacEncodeContext, options.lpc_coeff_precision), AV_OPT_TYPE_INT, {.i64 = 15 }, 0, MAX_LPC_PRECISION, FLAGS },
1235 { "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" },
1236 { "none", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = FF_LPC_TYPE_NONE }, INT_MIN, INT_MAX, FLAGS, "lpc_type" },
1237 { "fixed", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = FF_LPC_TYPE_FIXED }, INT_MIN, INT_MAX, FLAGS, "lpc_type" },
1238 { "levinson", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = FF_LPC_TYPE_LEVINSON }, INT_MIN, INT_MAX, FLAGS, "lpc_type" },
1239 { "cholesky", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = FF_LPC_TYPE_CHOLESKY }, INT_MIN, INT_MAX, FLAGS, "lpc_type" },
1240 { "lpc_passes", "Number of passes to use for Cholesky factorization during LPC analysis", offsetof(FlacEncodeContext, options.lpc_passes), AV_OPT_TYPE_INT, {.i64 = -1 }, INT_MIN, INT_MAX, FLAGS },
1241 { "min_partition_order", NULL, offsetof(FlacEncodeContext, options.min_partition_order), AV_OPT_TYPE_INT, {.i64 = -1 }, -1, MAX_PARTITION_ORDER, FLAGS },
1242 { "max_partition_order", NULL, offsetof(FlacEncodeContext, options.max_partition_order), AV_OPT_TYPE_INT, {.i64 = -1 }, -1, MAX_PARTITION_ORDER, FLAGS },
1243 { "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" },
1244 { "estimation", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = ORDER_METHOD_EST }, INT_MIN, INT_MAX, FLAGS, "predm" },
1245 { "2level", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = ORDER_METHOD_2LEVEL }, INT_MIN, INT_MAX, FLAGS, "predm" },
1246 { "4level", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = ORDER_METHOD_4LEVEL }, INT_MIN, INT_MAX, FLAGS, "predm" },
1247 { "8level", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = ORDER_METHOD_8LEVEL }, INT_MIN, INT_MAX, FLAGS, "predm" },
1248 { "search", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = ORDER_METHOD_SEARCH }, INT_MIN, INT_MAX, FLAGS, "predm" },
1249 { "log", NULL, 0, AV_OPT_TYPE_CONST, {.i64 = ORDER_METHOD_LOG }, INT_MIN, INT_MAX, FLAGS, "predm" },
1250 { "ch_mode", "Stereo decorrelation mode", offsetof(FlacEncodeContext, options.ch_mode), AV_OPT_TYPE_INT, { .i64 = -1 }, -1, FLAC_CHMODE_MID_SIDE, FLAGS, "ch_mode" },
1251 { "auto", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = -1 }, INT_MIN, INT_MAX, FLAGS, "ch_mode" },
1252 { "indep", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FLAC_CHMODE_INDEPENDENT }, INT_MIN, INT_MAX, FLAGS, "ch_mode" },
1253 { "left_side", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FLAC_CHMODE_LEFT_SIDE }, INT_MIN, INT_MAX, FLAGS, "ch_mode" },
1254 { "right_side", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FLAC_CHMODE_RIGHT_SIDE }, INT_MIN, INT_MAX, FLAGS, "ch_mode" },
1255 { "mid_side", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = FLAC_CHMODE_MID_SIDE }, INT_MIN, INT_MAX, FLAGS, "ch_mode" },
1259 static const AVClass flac_encoder_class = {
1261 av_default_item_name,
1263 LIBAVUTIL_VERSION_INT,
1266 AVCodec ff_flac_encoder = {
1268 .type = AVMEDIA_TYPE_AUDIO,
1269 .id = AV_CODEC_ID_FLAC,
1270 .priv_data_size = sizeof(FlacEncodeContext),
1271 .init = flac_encode_init,
1272 .encode2 = flac_encode_frame,
1273 .close = flac_encode_close,
1274 .capabilities = CODEC_CAP_SMALL_LAST_FRAME | CODEC_CAP_DELAY,
1275 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_S16,
1276 AV_SAMPLE_FMT_NONE },
1277 .long_name = NULL_IF_CONFIG_SMALL("FLAC (Free Lossless Audio Codec)"),
1278 .priv_class = &flac_encoder_class,