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/lls.h"
24 #include "libavutil/md5.h"
26 #include "bitstream.h"
32 #define FLAC_MIN_BLOCKSIZE 16
33 #define FLAC_MAX_BLOCKSIZE 65535
35 #define FLAC_SUBFRAME_CONSTANT 0
36 #define FLAC_SUBFRAME_VERBATIM 1
37 #define FLAC_SUBFRAME_FIXED 8
38 #define FLAC_SUBFRAME_LPC 32
40 #define FLAC_CHMODE_NOT_STEREO 0
41 #define FLAC_CHMODE_LEFT_RIGHT 1
42 #define FLAC_CHMODE_LEFT_SIDE 8
43 #define FLAC_CHMODE_RIGHT_SIDE 9
44 #define FLAC_CHMODE_MID_SIDE 10
46 #define FLAC_STREAMINFO_SIZE 34
48 #define MAX_FIXED_ORDER 4
49 #define MAX_PARTITION_ORDER 8
50 #define MAX_PARTITIONS (1 << MAX_PARTITION_ORDER)
51 #define MAX_LPC_PRECISION 15
52 #define MAX_LPC_SHIFT 15
53 #define MAX_RICE_PARAM 14
55 typedef struct CompressionOptions {
56 int compression_level;
59 int lpc_coeff_precision;
60 int min_prediction_order;
61 int max_prediction_order;
62 int prediction_order_method;
63 int min_partition_order;
64 int max_partition_order;
67 typedef struct RiceContext {
69 int params[MAX_PARTITIONS];
72 typedef struct FlacSubframe {
77 int32_t coefs[MAX_LPC_ORDER];
80 int32_t samples[FLAC_MAX_BLOCKSIZE];
81 int32_t residual[FLAC_MAX_BLOCKSIZE+1];
84 typedef struct FlacFrame {
85 FlacSubframe subframes[FLAC_MAX_CH];
92 typedef struct FlacEncodeContext {
99 int min_encoded_framesize;
101 int max_encoded_framesize;
102 uint32_t frame_count;
103 uint64_t sample_count;
106 CompressionOptions options;
107 AVCodecContext *avctx;
109 struct AVMD5 *md5ctx;
112 static const int flac_samplerates[16] = {
114 8000, 16000, 22050, 24000, 32000, 44100, 48000, 96000,
118 static const int flac_blocksizes[16] = {
121 576, 1152, 2304, 4608,
123 256, 512, 1024, 2048, 4096, 8192, 16384, 32768
127 * Writes streaminfo metadata block to byte array
129 static void write_streaminfo(FlacEncodeContext *s, uint8_t *header)
133 memset(header, 0, FLAC_STREAMINFO_SIZE);
134 init_put_bits(&pb, header, FLAC_STREAMINFO_SIZE);
136 /* streaminfo metadata block */
137 put_bits(&pb, 16, s->avctx->frame_size);
138 put_bits(&pb, 16, s->avctx->frame_size);
139 put_bits(&pb, 24, s->min_framesize);
140 put_bits(&pb, 24, s->max_framesize);
141 put_bits(&pb, 20, s->samplerate);
142 put_bits(&pb, 3, s->channels-1);
143 put_bits(&pb, 5, 15); /* bits per sample - 1 */
144 /* write 36-bit sample count in 2 put_bits() calls */
145 put_bits(&pb, 24, (s->sample_count & 0xFFFFFF000LL) >> 12);
146 put_bits(&pb, 12, s->sample_count & 0x000000FFFLL);
148 memcpy(&header[18], s->md5sum, 16);
152 * Sets blocksize based on samplerate
153 * Chooses the closest predefined blocksize >= BLOCK_TIME_MS milliseconds
155 static int select_blocksize(int samplerate, int block_time_ms)
161 assert(samplerate > 0);
162 blocksize = flac_blocksizes[1];
163 target = (samplerate * block_time_ms) / 1000;
164 for(i=0; i<16; i++) {
165 if(target >= flac_blocksizes[i] && flac_blocksizes[i] > blocksize) {
166 blocksize = flac_blocksizes[i];
172 static av_cold int flac_encode_init(AVCodecContext *avctx)
174 int freq = avctx->sample_rate;
175 int channels = avctx->channels;
176 FlacEncodeContext *s = avctx->priv_data;
182 dsputil_init(&s->dsp, avctx);
184 if(avctx->sample_fmt != SAMPLE_FMT_S16) {
188 if(channels < 1 || channels > FLAC_MAX_CH) {
191 s->channels = channels;
192 s->ch_code = s->channels-1;
194 /* find samplerate in table */
197 for(i=4; i<12; i++) {
198 if(freq == flac_samplerates[i]) {
199 s->samplerate = flac_samplerates[i];
205 /* if not in table, samplerate is non-standard */
207 if(freq % 1000 == 0 && freq < 255000) {
209 s->sr_code[1] = freq / 1000;
210 } else if(freq % 10 == 0 && freq < 655350) {
212 s->sr_code[1] = freq / 10;
213 } else if(freq < 65535) {
215 s->sr_code[1] = freq;
219 s->samplerate = freq;
222 /* set compression option defaults based on avctx->compression_level */
223 if(avctx->compression_level < 0) {
224 s->options.compression_level = 5;
226 s->options.compression_level = avctx->compression_level;
228 av_log(avctx, AV_LOG_DEBUG, " compression: %d\n", s->options.compression_level);
230 level= s->options.compression_level;
232 av_log(avctx, AV_LOG_ERROR, "invalid compression level: %d\n",
233 s->options.compression_level);
237 s->options.block_time_ms = ((int[]){ 27, 27, 27,105,105,105,105,105,105,105,105,105,105})[level];
238 s->options.use_lpc = ((int[]){ 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1})[level];
239 s->options.min_prediction_order= ((int[]){ 2, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1})[level];
240 s->options.max_prediction_order= ((int[]){ 3, 4, 4, 6, 8, 8, 8, 8, 12, 12, 12, 32, 32})[level];
241 s->options.prediction_order_method = ((int[]){ ORDER_METHOD_EST, ORDER_METHOD_EST, ORDER_METHOD_EST,
242 ORDER_METHOD_EST, ORDER_METHOD_EST, ORDER_METHOD_EST,
243 ORDER_METHOD_4LEVEL, ORDER_METHOD_LOG, ORDER_METHOD_4LEVEL,
244 ORDER_METHOD_LOG, ORDER_METHOD_SEARCH, ORDER_METHOD_LOG,
245 ORDER_METHOD_SEARCH})[level];
246 s->options.min_partition_order = ((int[]){ 2, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0})[level];
247 s->options.max_partition_order = ((int[]){ 2, 2, 3, 3, 3, 8, 8, 8, 8, 8, 8, 8, 8})[level];
249 /* set compression option overrides from AVCodecContext */
250 if(avctx->use_lpc >= 0) {
251 s->options.use_lpc = av_clip(avctx->use_lpc, 0, 11);
253 if(s->options.use_lpc == 1)
254 av_log(avctx, AV_LOG_DEBUG, " use lpc: Levinson-Durbin recursion with Welch window\n");
255 else if(s->options.use_lpc > 1)
256 av_log(avctx, AV_LOG_DEBUG, " use lpc: Cholesky factorization\n");
258 if(avctx->min_prediction_order >= 0) {
259 if(s->options.use_lpc) {
260 if(avctx->min_prediction_order < MIN_LPC_ORDER ||
261 avctx->min_prediction_order > MAX_LPC_ORDER) {
262 av_log(avctx, AV_LOG_ERROR, "invalid min prediction order: %d\n",
263 avctx->min_prediction_order);
267 if(avctx->min_prediction_order > MAX_FIXED_ORDER) {
268 av_log(avctx, AV_LOG_ERROR, "invalid min prediction order: %d\n",
269 avctx->min_prediction_order);
273 s->options.min_prediction_order = avctx->min_prediction_order;
275 if(avctx->max_prediction_order >= 0) {
276 if(s->options.use_lpc) {
277 if(avctx->max_prediction_order < MIN_LPC_ORDER ||
278 avctx->max_prediction_order > MAX_LPC_ORDER) {
279 av_log(avctx, AV_LOG_ERROR, "invalid max prediction order: %d\n",
280 avctx->max_prediction_order);
284 if(avctx->max_prediction_order > MAX_FIXED_ORDER) {
285 av_log(avctx, AV_LOG_ERROR, "invalid max prediction order: %d\n",
286 avctx->max_prediction_order);
290 s->options.max_prediction_order = avctx->max_prediction_order;
292 if(s->options.max_prediction_order < s->options.min_prediction_order) {
293 av_log(avctx, AV_LOG_ERROR, "invalid prediction orders: min=%d max=%d\n",
294 s->options.min_prediction_order, s->options.max_prediction_order);
297 av_log(avctx, AV_LOG_DEBUG, " prediction order: %d, %d\n",
298 s->options.min_prediction_order, s->options.max_prediction_order);
300 if(avctx->prediction_order_method >= 0) {
301 if(avctx->prediction_order_method > ORDER_METHOD_LOG) {
302 av_log(avctx, AV_LOG_ERROR, "invalid prediction order method: %d\n",
303 avctx->prediction_order_method);
306 s->options.prediction_order_method = avctx->prediction_order_method;
308 switch(s->options.prediction_order_method) {
309 case ORDER_METHOD_EST: av_log(avctx, AV_LOG_DEBUG, " order method: %s\n",
311 case ORDER_METHOD_2LEVEL: av_log(avctx, AV_LOG_DEBUG, " order method: %s\n",
313 case ORDER_METHOD_4LEVEL: av_log(avctx, AV_LOG_DEBUG, " order method: %s\n",
315 case ORDER_METHOD_8LEVEL: av_log(avctx, AV_LOG_DEBUG, " order method: %s\n",
317 case ORDER_METHOD_SEARCH: av_log(avctx, AV_LOG_DEBUG, " order method: %s\n",
318 "full search"); break;
319 case ORDER_METHOD_LOG: av_log(avctx, AV_LOG_DEBUG, " order method: %s\n",
320 "log search"); break;
323 if(avctx->min_partition_order >= 0) {
324 if(avctx->min_partition_order > MAX_PARTITION_ORDER) {
325 av_log(avctx, AV_LOG_ERROR, "invalid min partition order: %d\n",
326 avctx->min_partition_order);
329 s->options.min_partition_order = avctx->min_partition_order;
331 if(avctx->max_partition_order >= 0) {
332 if(avctx->max_partition_order > MAX_PARTITION_ORDER) {
333 av_log(avctx, AV_LOG_ERROR, "invalid max partition order: %d\n",
334 avctx->max_partition_order);
337 s->options.max_partition_order = avctx->max_partition_order;
339 if(s->options.max_partition_order < s->options.min_partition_order) {
340 av_log(avctx, AV_LOG_ERROR, "invalid partition orders: min=%d max=%d\n",
341 s->options.min_partition_order, s->options.max_partition_order);
344 av_log(avctx, AV_LOG_DEBUG, " partition order: %d, %d\n",
345 s->options.min_partition_order, s->options.max_partition_order);
347 if(avctx->frame_size > 0) {
348 if(avctx->frame_size < FLAC_MIN_BLOCKSIZE ||
349 avctx->frame_size > FLAC_MAX_BLOCKSIZE) {
350 av_log(avctx, AV_LOG_ERROR, "invalid block size: %d\n",
355 s->avctx->frame_size = select_blocksize(s->samplerate, s->options.block_time_ms);
357 av_log(avctx, AV_LOG_DEBUG, " block size: %d\n", s->avctx->frame_size);
359 /* set LPC precision */
360 if(avctx->lpc_coeff_precision > 0) {
361 if(avctx->lpc_coeff_precision > MAX_LPC_PRECISION) {
362 av_log(avctx, AV_LOG_ERROR, "invalid lpc coeff precision: %d\n",
363 avctx->lpc_coeff_precision);
366 s->options.lpc_coeff_precision = avctx->lpc_coeff_precision;
368 /* default LPC precision */
369 s->options.lpc_coeff_precision = 15;
371 av_log(avctx, AV_LOG_DEBUG, " lpc precision: %d\n",
372 s->options.lpc_coeff_precision);
374 /* set maximum encoded frame size in verbatim mode */
375 if(s->channels == 2) {
376 s->max_framesize = 14 + ((s->avctx->frame_size * 33 + 7) >> 3);
378 s->max_framesize = 14 + (s->avctx->frame_size * s->channels * 2);
380 s->min_encoded_framesize = 0xFFFFFF;
382 /* initialize MD5 context */
383 s->md5ctx = av_malloc(av_md5_size);
385 return AVERROR_NOMEM;
386 av_md5_init(s->md5ctx);
388 streaminfo = av_malloc(FLAC_STREAMINFO_SIZE);
389 write_streaminfo(s, streaminfo);
390 avctx->extradata = streaminfo;
391 avctx->extradata_size = FLAC_STREAMINFO_SIZE;
395 avctx->coded_frame = avcodec_alloc_frame();
396 avctx->coded_frame->key_frame = 1;
401 static void init_frame(FlacEncodeContext *s)
408 for(i=0; i<16; i++) {
409 if(s->avctx->frame_size == flac_blocksizes[i]) {
410 frame->blocksize = flac_blocksizes[i];
411 frame->bs_code[0] = i;
412 frame->bs_code[1] = 0;
417 frame->blocksize = s->avctx->frame_size;
418 if(frame->blocksize <= 256) {
419 frame->bs_code[0] = 6;
420 frame->bs_code[1] = frame->blocksize-1;
422 frame->bs_code[0] = 7;
423 frame->bs_code[1] = frame->blocksize-1;
427 for(ch=0; ch<s->channels; ch++) {
428 frame->subframes[ch].obits = 16;
433 * Copy channel-interleaved input samples into separate subframes
435 static void copy_samples(FlacEncodeContext *s, int16_t *samples)
441 for(i=0,j=0; i<frame->blocksize; i++) {
442 for(ch=0; ch<s->channels; ch++,j++) {
443 frame->subframes[ch].samples[i] = samples[j];
449 #define rice_encode_count(sum, n, k) (((n)*((k)+1))+((sum-(n>>1))>>(k)))
452 * Solve for d/dk(rice_encode_count) = n-((sum-(n>>1))>>(k+1)) = 0
454 static int find_optimal_param(uint32_t sum, int n)
462 k = av_log2(n<256 ? FASTDIV(sum2,n) : sum2/n);
463 return FFMIN(k, MAX_RICE_PARAM);
466 static uint32_t calc_optimal_rice_params(RiceContext *rc, int porder,
467 uint32_t *sums, int n, int pred_order)
473 part = (1 << porder);
476 cnt = (n >> porder) - pred_order;
477 for(i=0; i<part; i++) {
478 k = find_optimal_param(sums[i], cnt);
480 all_bits += rice_encode_count(sums[i], cnt, k);
489 static void calc_sums(int pmin, int pmax, uint32_t *data, int n, int pred_order,
490 uint32_t sums[][MAX_PARTITIONS])
494 uint32_t *res, *res_end;
496 /* sums for highest level */
498 res = &data[pred_order];
499 res_end = &data[n >> pmax];
500 for(i=0; i<parts; i++) {
502 while(res < res_end){
508 /* sums for lower levels */
509 for(i=pmax-1; i>=pmin; i--) {
511 for(j=0; j<parts; j++) {
512 sums[i][j] = sums[i+1][2*j] + sums[i+1][2*j+1];
517 static uint32_t calc_rice_params(RiceContext *rc, int pmin, int pmax,
518 int32_t *data, int n, int pred_order)
521 uint32_t bits[MAX_PARTITION_ORDER+1];
525 uint32_t sums[MAX_PARTITION_ORDER+1][MAX_PARTITIONS];
527 assert(pmin >= 0 && pmin <= MAX_PARTITION_ORDER);
528 assert(pmax >= 0 && pmax <= MAX_PARTITION_ORDER);
529 assert(pmin <= pmax);
531 udata = av_malloc(n * sizeof(uint32_t));
533 udata[i] = (2*data[i]) ^ (data[i]>>31);
536 calc_sums(pmin, pmax, udata, n, pred_order, sums);
539 bits[pmin] = UINT32_MAX;
540 for(i=pmin; i<=pmax; i++) {
541 bits[i] = calc_optimal_rice_params(&tmp_rc, i, sums[i], n, pred_order);
542 if(bits[i] <= bits[opt_porder]) {
549 return bits[opt_porder];
552 static int get_max_p_order(int max_porder, int n, int order)
554 int porder = FFMIN(max_porder, av_log2(n^(n-1)));
556 porder = FFMIN(porder, av_log2(n/order));
560 static uint32_t calc_rice_params_fixed(RiceContext *rc, int pmin, int pmax,
561 int32_t *data, int n, int pred_order,
565 pmin = get_max_p_order(pmin, n, pred_order);
566 pmax = get_max_p_order(pmax, n, pred_order);
567 bits = pred_order*bps + 6;
568 bits += calc_rice_params(rc, pmin, pmax, data, n, pred_order);
572 static uint32_t calc_rice_params_lpc(RiceContext *rc, int pmin, int pmax,
573 int32_t *data, int n, int pred_order,
574 int bps, int precision)
577 pmin = get_max_p_order(pmin, n, pred_order);
578 pmax = get_max_p_order(pmax, n, pred_order);
579 bits = pred_order*bps + 4 + 5 + pred_order*precision + 6;
580 bits += calc_rice_params(rc, pmin, pmax, data, n, pred_order);
585 * Apply Welch window function to audio block
587 static void apply_welch_window(const int32_t *data, int len, double *w_data)
593 assert(!(len&1)); //the optimization in r11881 does not support odd len
594 //if someone wants odd len extend the change in r11881
597 c = 2.0 / (len - 1.0);
601 for(i=0; i<n2; i++) {
604 w_data[-i-1] = data[-i-1] * w;
605 w_data[+i ] = data[+i ] * w;
610 * Calculates autocorrelation data from audio samples
611 * A Welch window function is applied before calculation.
613 void ff_flac_compute_autocorr(const int32_t *data, int len, int lag,
617 double tmp[len + lag + 1];
618 double *data1= tmp + lag;
620 apply_welch_window(data, len, data1);
626 for(j=0; j<lag; j+=2){
627 double sum0 = 1.0, sum1 = 1.0;
628 for(i=0; i<len; i++){
629 sum0 += data1[i] * data1[i-j];
630 sum1 += data1[i] * data1[i-j-1];
638 for(i=0; i<len; i+=2){
639 sum += data1[i ] * data1[i-j ]
640 + data1[i+1] * data1[i-j+1];
647 static void encode_residual_verbatim(int32_t *res, int32_t *smp, int n)
650 memcpy(res, smp, n * sizeof(int32_t));
653 static void encode_residual_fixed(int32_t *res, const int32_t *smp, int n,
658 for(i=0; i<order; i++) {
663 for(i=order; i<n; i++)
666 for(i=order; i<n; i++)
667 res[i]= smp[i] - smp[i-1];
669 int a = smp[order-1] - smp[order-2];
670 for(i=order; i<n; i+=2) {
671 int b = smp[i] - smp[i-1];
673 a = smp[i+1] - smp[i];
677 int a = smp[order-1] - smp[order-2];
678 int c = smp[order-1] - 2*smp[order-2] + smp[order-3];
679 for(i=order; i<n; i+=2) {
680 int b = smp[i] - smp[i-1];
683 a = smp[i+1] - smp[i];
688 int a = smp[order-1] - smp[order-2];
689 int c = smp[order-1] - 2*smp[order-2] + smp[order-3];
690 int e = smp[order-1] - 3*smp[order-2] + 3*smp[order-3] - smp[order-4];
691 for(i=order; i<n; i+=2) {
692 int b = smp[i] - smp[i-1];
696 a = smp[i+1] - smp[i];
705 int c = coefs[(x)-1];\
711 static av_always_inline void encode_residual_lpc_unrolled(
712 int32_t *res, const int32_t *smp, int n,
713 int order, const int32_t *coefs, int shift, int big)
716 for(i=order; i<n; i+=2) {
717 int s = smp[i-order];
766 res[i ] = smp[i ] - (p0 >> shift);
767 res[i+1] = smp[i+1] - (p1 >> shift);
771 static void encode_residual_lpc(int32_t *res, const int32_t *smp, int n,
772 int order, const int32_t *coefs, int shift)
775 for(i=0; i<order; i++) {
779 for(i=order; i<n; i+=2) {
783 for(j=0; j<order; j++) {
789 res[i ] = smp[i ] - (p0 >> shift);
790 res[i+1] = smp[i+1] - (p1 >> shift);
794 case 1: encode_residual_lpc_unrolled(res, smp, n, 1, coefs, shift, 0); break;
795 case 2: encode_residual_lpc_unrolled(res, smp, n, 2, coefs, shift, 0); break;
796 case 3: encode_residual_lpc_unrolled(res, smp, n, 3, coefs, shift, 0); break;
797 case 4: encode_residual_lpc_unrolled(res, smp, n, 4, coefs, shift, 0); break;
798 case 5: encode_residual_lpc_unrolled(res, smp, n, 5, coefs, shift, 0); break;
799 case 6: encode_residual_lpc_unrolled(res, smp, n, 6, coefs, shift, 0); break;
800 case 7: encode_residual_lpc_unrolled(res, smp, n, 7, coefs, shift, 0); break;
801 case 8: encode_residual_lpc_unrolled(res, smp, n, 8, coefs, shift, 0); break;
802 default: encode_residual_lpc_unrolled(res, smp, n, order, coefs, shift, 1); break;
807 static int encode_residual(FlacEncodeContext *ctx, int ch)
810 int min_order, max_order, opt_order, precision, omethod;
811 int min_porder, max_porder;
814 int32_t coefs[MAX_LPC_ORDER][MAX_LPC_ORDER];
815 int shift[MAX_LPC_ORDER];
819 sub = &frame->subframes[ch];
822 n = frame->blocksize;
826 if(smp[i] != smp[0]) break;
829 sub->type = sub->type_code = FLAC_SUBFRAME_CONSTANT;
836 sub->type = sub->type_code = FLAC_SUBFRAME_VERBATIM;
837 encode_residual_verbatim(res, smp, n);
838 return sub->obits * n;
841 min_order = ctx->options.min_prediction_order;
842 max_order = ctx->options.max_prediction_order;
843 min_porder = ctx->options.min_partition_order;
844 max_porder = ctx->options.max_partition_order;
845 precision = ctx->options.lpc_coeff_precision;
846 omethod = ctx->options.prediction_order_method;
849 if(!ctx->options.use_lpc || max_order == 0 || (n <= max_order)) {
850 uint32_t bits[MAX_FIXED_ORDER+1];
851 if(max_order > MAX_FIXED_ORDER) max_order = MAX_FIXED_ORDER;
853 bits[0] = UINT32_MAX;
854 for(i=min_order; i<=max_order; i++) {
855 encode_residual_fixed(res, smp, n, i);
856 bits[i] = calc_rice_params_fixed(&sub->rc, min_porder, max_porder, res,
858 if(bits[i] < bits[opt_order]) {
862 sub->order = opt_order;
863 sub->type = FLAC_SUBFRAME_FIXED;
864 sub->type_code = sub->type | sub->order;
865 if(sub->order != max_order) {
866 encode_residual_fixed(res, smp, n, sub->order);
867 return calc_rice_params_fixed(&sub->rc, min_porder, max_porder, res, n,
868 sub->order, sub->obits);
870 return bits[sub->order];
874 opt_order = ff_lpc_calc_coefs(&ctx->dsp, smp, n, min_order, max_order,
875 precision, coefs, shift, ctx->options.use_lpc,
876 omethod, MAX_LPC_SHIFT, 0);
878 if(omethod == ORDER_METHOD_2LEVEL ||
879 omethod == ORDER_METHOD_4LEVEL ||
880 omethod == ORDER_METHOD_8LEVEL) {
881 int levels = 1 << omethod;
882 uint32_t bits[levels];
884 int opt_index = levels-1;
885 opt_order = max_order-1;
886 bits[opt_index] = UINT32_MAX;
887 for(i=levels-1; i>=0; i--) {
888 order = min_order + (((max_order-min_order+1) * (i+1)) / levels)-1;
889 if(order < 0) order = 0;
890 encode_residual_lpc(res, smp, n, order+1, coefs[order], shift[order]);
891 bits[i] = calc_rice_params_lpc(&sub->rc, min_porder, max_porder,
892 res, n, order+1, sub->obits, precision);
893 if(bits[i] < bits[opt_index]) {
899 } else if(omethod == ORDER_METHOD_SEARCH) {
900 // brute-force optimal order search
901 uint32_t bits[MAX_LPC_ORDER];
903 bits[0] = UINT32_MAX;
904 for(i=min_order-1; i<max_order; i++) {
905 encode_residual_lpc(res, smp, n, i+1, coefs[i], shift[i]);
906 bits[i] = calc_rice_params_lpc(&sub->rc, min_porder, max_porder,
907 res, n, i+1, sub->obits, precision);
908 if(bits[i] < bits[opt_order]) {
913 } else if(omethod == ORDER_METHOD_LOG) {
914 uint32_t bits[MAX_LPC_ORDER];
917 opt_order= min_order - 1 + (max_order-min_order)/3;
918 memset(bits, -1, sizeof(bits));
920 for(step=16 ;step; step>>=1){
922 for(i=last-step; i<=last+step; i+= step){
923 if(i<min_order-1 || i>=max_order || bits[i] < UINT32_MAX)
925 encode_residual_lpc(res, smp, n, i+1, coefs[i], shift[i]);
926 bits[i] = calc_rice_params_lpc(&sub->rc, min_porder, max_porder,
927 res, n, i+1, sub->obits, precision);
928 if(bits[i] < bits[opt_order])
935 sub->order = opt_order;
936 sub->type = FLAC_SUBFRAME_LPC;
937 sub->type_code = sub->type | (sub->order-1);
938 sub->shift = shift[sub->order-1];
939 for(i=0; i<sub->order; i++) {
940 sub->coefs[i] = coefs[sub->order-1][i];
942 encode_residual_lpc(res, smp, n, sub->order, sub->coefs, sub->shift);
943 return calc_rice_params_lpc(&sub->rc, min_porder, max_porder, res, n, sub->order,
944 sub->obits, precision);
947 static int encode_residual_v(FlacEncodeContext *ctx, int ch)
955 sub = &frame->subframes[ch];
958 n = frame->blocksize;
962 if(smp[i] != smp[0]) break;
965 sub->type = sub->type_code = FLAC_SUBFRAME_CONSTANT;
971 sub->type = sub->type_code = FLAC_SUBFRAME_VERBATIM;
972 encode_residual_verbatim(res, smp, n);
973 return sub->obits * n;
976 static int estimate_stereo_mode(int32_t *left_ch, int32_t *right_ch, int n)
984 /* calculate sum of 2nd order residual for each channel */
985 sum[0] = sum[1] = sum[2] = sum[3] = 0;
987 lt = left_ch[i] - 2*left_ch[i-1] + left_ch[i-2];
988 rt = right_ch[i] - 2*right_ch[i-1] + right_ch[i-2];
989 sum[2] += FFABS((lt + rt) >> 1);
990 sum[3] += FFABS(lt - rt);
994 /* estimate bit counts */
996 k = find_optimal_param(2*sum[i], n);
997 sum[i] = rice_encode_count(2*sum[i], n, k);
1000 /* calculate score for each mode */
1001 score[0] = sum[0] + sum[1];
1002 score[1] = sum[0] + sum[3];
1003 score[2] = sum[1] + sum[3];
1004 score[3] = sum[2] + sum[3];
1006 /* return mode with lowest score */
1008 for(i=1; i<4; i++) {
1009 if(score[i] < score[best]) {
1014 return FLAC_CHMODE_LEFT_RIGHT;
1015 } else if(best == 1) {
1016 return FLAC_CHMODE_LEFT_SIDE;
1017 } else if(best == 2) {
1018 return FLAC_CHMODE_RIGHT_SIDE;
1020 return FLAC_CHMODE_MID_SIDE;
1025 * Perform stereo channel decorrelation
1027 static void channel_decorrelation(FlacEncodeContext *ctx)
1030 int32_t *left, *right;
1033 frame = &ctx->frame;
1034 n = frame->blocksize;
1035 left = frame->subframes[0].samples;
1036 right = frame->subframes[1].samples;
1038 if(ctx->channels != 2) {
1039 frame->ch_mode = FLAC_CHMODE_NOT_STEREO;
1043 frame->ch_mode = estimate_stereo_mode(left, right, n);
1045 /* perform decorrelation and adjust bits-per-sample */
1046 if(frame->ch_mode == FLAC_CHMODE_LEFT_RIGHT) {
1049 if(frame->ch_mode == FLAC_CHMODE_MID_SIDE) {
1051 for(i=0; i<n; i++) {
1053 left[i] = (tmp + right[i]) >> 1;
1054 right[i] = tmp - right[i];
1056 frame->subframes[1].obits++;
1057 } else if(frame->ch_mode == FLAC_CHMODE_LEFT_SIDE) {
1058 for(i=0; i<n; i++) {
1059 right[i] = left[i] - right[i];
1061 frame->subframes[1].obits++;
1063 for(i=0; i<n; i++) {
1064 left[i] -= right[i];
1066 frame->subframes[0].obits++;
1070 static void write_utf8(PutBitContext *pb, uint32_t val)
1073 PUT_UTF8(val, tmp, put_bits(pb, 8, tmp);)
1076 static void output_frame_header(FlacEncodeContext *s)
1083 put_bits(&s->pb, 16, 0xFFF8);
1084 put_bits(&s->pb, 4, frame->bs_code[0]);
1085 put_bits(&s->pb, 4, s->sr_code[0]);
1086 if(frame->ch_mode == FLAC_CHMODE_NOT_STEREO) {
1087 put_bits(&s->pb, 4, s->ch_code);
1089 put_bits(&s->pb, 4, frame->ch_mode);
1091 put_bits(&s->pb, 3, 4); /* bits-per-sample code */
1092 put_bits(&s->pb, 1, 0);
1093 write_utf8(&s->pb, s->frame_count);
1094 if(frame->bs_code[0] == 6) {
1095 put_bits(&s->pb, 8, frame->bs_code[1]);
1096 } else if(frame->bs_code[0] == 7) {
1097 put_bits(&s->pb, 16, frame->bs_code[1]);
1099 if(s->sr_code[0] == 12) {
1100 put_bits(&s->pb, 8, s->sr_code[1]);
1101 } else if(s->sr_code[0] > 12) {
1102 put_bits(&s->pb, 16, s->sr_code[1]);
1104 flush_put_bits(&s->pb);
1105 crc = av_crc(av_crc_get_table(AV_CRC_8_ATM), 0,
1106 s->pb.buf, put_bits_count(&s->pb)>>3);
1107 put_bits(&s->pb, 8, crc);
1110 static void output_subframe_constant(FlacEncodeContext *s, int ch)
1115 sub = &s->frame.subframes[ch];
1116 res = sub->residual[0];
1117 put_sbits(&s->pb, sub->obits, res);
1120 static void output_subframe_verbatim(FlacEncodeContext *s, int ch)
1128 sub = &frame->subframes[ch];
1130 for(i=0; i<frame->blocksize; i++) {
1131 res = sub->residual[i];
1132 put_sbits(&s->pb, sub->obits, res);
1136 static void output_residual(FlacEncodeContext *ctx, int ch)
1138 int i, j, p, n, parts;
1139 int k, porder, psize, res_cnt;
1144 frame = &ctx->frame;
1145 sub = &frame->subframes[ch];
1146 res = sub->residual;
1147 n = frame->blocksize;
1149 /* rice-encoded block */
1150 put_bits(&ctx->pb, 2, 0);
1152 /* partition order */
1153 porder = sub->rc.porder;
1154 psize = n >> porder;
1155 parts = (1 << porder);
1156 put_bits(&ctx->pb, 4, porder);
1157 res_cnt = psize - sub->order;
1161 for(p=0; p<parts; p++) {
1162 k = sub->rc.params[p];
1163 put_bits(&ctx->pb, 4, k);
1164 if(p == 1) res_cnt = psize;
1165 for(i=0; i<res_cnt && j<n; i++, j++) {
1166 set_sr_golomb_flac(&ctx->pb, res[j], k, INT32_MAX, 0);
1171 static void output_subframe_fixed(FlacEncodeContext *ctx, int ch)
1177 frame = &ctx->frame;
1178 sub = &frame->subframes[ch];
1180 /* warm-up samples */
1181 for(i=0; i<sub->order; i++) {
1182 put_sbits(&ctx->pb, sub->obits, sub->residual[i]);
1186 output_residual(ctx, ch);
1189 static void output_subframe_lpc(FlacEncodeContext *ctx, int ch)
1195 frame = &ctx->frame;
1196 sub = &frame->subframes[ch];
1198 /* warm-up samples */
1199 for(i=0; i<sub->order; i++) {
1200 put_sbits(&ctx->pb, sub->obits, sub->residual[i]);
1203 /* LPC coefficients */
1204 cbits = ctx->options.lpc_coeff_precision;
1205 put_bits(&ctx->pb, 4, cbits-1);
1206 put_sbits(&ctx->pb, 5, sub->shift);
1207 for(i=0; i<sub->order; i++) {
1208 put_sbits(&ctx->pb, cbits, sub->coefs[i]);
1212 output_residual(ctx, ch);
1215 static void output_subframes(FlacEncodeContext *s)
1223 for(ch=0; ch<s->channels; ch++) {
1224 sub = &frame->subframes[ch];
1226 /* subframe header */
1227 put_bits(&s->pb, 1, 0);
1228 put_bits(&s->pb, 6, sub->type_code);
1229 put_bits(&s->pb, 1, 0); /* no wasted bits */
1232 if(sub->type == FLAC_SUBFRAME_CONSTANT) {
1233 output_subframe_constant(s, ch);
1234 } else if(sub->type == FLAC_SUBFRAME_VERBATIM) {
1235 output_subframe_verbatim(s, ch);
1236 } else if(sub->type == FLAC_SUBFRAME_FIXED) {
1237 output_subframe_fixed(s, ch);
1238 } else if(sub->type == FLAC_SUBFRAME_LPC) {
1239 output_subframe_lpc(s, ch);
1244 static void output_frame_footer(FlacEncodeContext *s)
1247 flush_put_bits(&s->pb);
1248 crc = bswap_16(av_crc(av_crc_get_table(AV_CRC_16_ANSI), 0,
1249 s->pb.buf, put_bits_count(&s->pb)>>3));
1250 put_bits(&s->pb, 16, crc);
1251 flush_put_bits(&s->pb);
1254 static void update_md5_sum(FlacEncodeContext *s, int16_t *samples)
1256 #ifdef WORDS_BIGENDIAN
1258 for(i = 0; i < s->frame.blocksize*s->channels; i++) {
1259 int16_t smp = le2me_16(samples[i]);
1260 av_md5_update(s->md5ctx, (uint8_t *)&smp, 2);
1263 av_md5_update(s->md5ctx, (uint8_t *)samples, s->frame.blocksize*s->channels*2);
1267 static int flac_encode_frame(AVCodecContext *avctx, uint8_t *frame,
1268 int buf_size, void *data)
1271 FlacEncodeContext *s;
1272 int16_t *samples = data;
1276 s = avctx->priv_data;
1278 if(buf_size < s->max_framesize*2) {
1279 av_log(avctx, AV_LOG_ERROR, "output buffer too small\n");
1283 /* when the last block is reached, update the header in extradata */
1285 s->min_framesize = s->min_encoded_framesize;
1286 s->max_framesize = s->max_encoded_framesize;
1287 av_md5_final(s->md5ctx, s->md5sum);
1288 write_streaminfo(s, avctx->extradata);
1294 copy_samples(s, samples);
1296 channel_decorrelation(s);
1298 for(ch=0; ch<s->channels; ch++) {
1299 encode_residual(s, ch);
1303 init_put_bits(&s->pb, frame, buf_size);
1304 output_frame_header(s);
1305 output_subframes(s);
1306 output_frame_footer(s);
1307 out_bytes = put_bits_count(&s->pb) >> 3;
1309 if(out_bytes > s->max_framesize) {
1311 /* still too large. must be an error. */
1312 av_log(avctx, AV_LOG_ERROR, "error encoding frame\n");
1316 /* frame too large. use verbatim mode */
1317 for(ch=0; ch<s->channels; ch++) {
1318 encode_residual_v(s, ch);
1325 s->sample_count += avctx->frame_size;
1326 update_md5_sum(s, samples);
1327 if (out_bytes > s->max_encoded_framesize)
1328 s->max_encoded_framesize = out_bytes;
1329 if (out_bytes < s->min_encoded_framesize)
1330 s->min_encoded_framesize = out_bytes;
1335 static av_cold int flac_encode_close(AVCodecContext *avctx)
1337 if (avctx->priv_data) {
1338 FlacEncodeContext *s = avctx->priv_data;
1339 av_freep(&s->md5ctx);
1341 av_freep(&avctx->extradata);
1342 avctx->extradata_size = 0;
1343 av_freep(&avctx->coded_frame);
1347 AVCodec flac_encoder = {
1351 sizeof(FlacEncodeContext),
1356 .capabilities = CODEC_CAP_SMALL_LAST_FRAME | CODEC_CAP_DELAY,
1357 .sample_fmts = (enum SampleFormat[]){SAMPLE_FMT_S16,SAMPLE_FMT_NONE},
1358 .long_name = NULL_IF_CONFIG_SMALL("FLAC (Free Lossless Audio Codec)"),