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_SUBFRAME_CONSTANT 0
33 #define FLAC_SUBFRAME_VERBATIM 1
34 #define FLAC_SUBFRAME_FIXED 8
35 #define FLAC_SUBFRAME_LPC 32
37 #define FLAC_CHMODE_NOT_STEREO 0
38 #define FLAC_CHMODE_LEFT_RIGHT 1
39 #define FLAC_CHMODE_LEFT_SIDE 8
40 #define FLAC_CHMODE_RIGHT_SIDE 9
41 #define FLAC_CHMODE_MID_SIDE 10
43 #define MAX_FIXED_ORDER 4
44 #define MAX_PARTITION_ORDER 8
45 #define MAX_PARTITIONS (1 << MAX_PARTITION_ORDER)
46 #define MAX_LPC_PRECISION 15
47 #define MAX_LPC_SHIFT 15
48 #define MAX_RICE_PARAM 14
50 typedef struct CompressionOptions {
51 int compression_level;
54 int lpc_coeff_precision;
55 int min_prediction_order;
56 int max_prediction_order;
57 int prediction_order_method;
58 int min_partition_order;
59 int max_partition_order;
62 typedef struct RiceContext {
64 int params[MAX_PARTITIONS];
67 typedef struct FlacSubframe {
72 int32_t coefs[MAX_LPC_ORDER];
75 int32_t samples[FLAC_MAX_BLOCKSIZE];
76 int32_t residual[FLAC_MAX_BLOCKSIZE+1];
79 typedef struct FlacFrame {
80 FlacSubframe subframes[FLAC_MAX_CHANNELS];
87 typedef struct FlacEncodeContext {
94 int min_encoded_framesize;
96 int max_encoded_framesize;
98 uint64_t sample_count;
101 CompressionOptions options;
102 AVCodecContext *avctx;
104 struct AVMD5 *md5ctx;
107 static const int flac_samplerates[16] = {
109 8000, 16000, 22050, 24000, 32000, 44100, 48000, 96000,
113 static const int flac_blocksizes[16] = {
116 576, 1152, 2304, 4608,
118 256, 512, 1024, 2048, 4096, 8192, 16384, 32768
122 * Writes streaminfo metadata block to byte array
124 static void write_streaminfo(FlacEncodeContext *s, uint8_t *header)
128 memset(header, 0, FLAC_STREAMINFO_SIZE);
129 init_put_bits(&pb, header, FLAC_STREAMINFO_SIZE);
131 /* streaminfo metadata block */
132 put_bits(&pb, 16, s->avctx->frame_size);
133 put_bits(&pb, 16, s->avctx->frame_size);
134 put_bits(&pb, 24, s->min_framesize);
135 put_bits(&pb, 24, s->max_framesize);
136 put_bits(&pb, 20, s->samplerate);
137 put_bits(&pb, 3, s->channels-1);
138 put_bits(&pb, 5, 15); /* bits per sample - 1 */
139 /* write 36-bit sample count in 2 put_bits() calls */
140 put_bits(&pb, 24, (s->sample_count & 0xFFFFFF000LL) >> 12);
141 put_bits(&pb, 12, s->sample_count & 0x000000FFFLL);
143 memcpy(&header[18], s->md5sum, 16);
147 * Sets blocksize based on samplerate
148 * Chooses the closest predefined blocksize >= BLOCK_TIME_MS milliseconds
150 static int select_blocksize(int samplerate, int block_time_ms)
156 assert(samplerate > 0);
157 blocksize = flac_blocksizes[1];
158 target = (samplerate * block_time_ms) / 1000;
159 for(i=0; i<16; i++) {
160 if(target >= flac_blocksizes[i] && flac_blocksizes[i] > blocksize) {
161 blocksize = flac_blocksizes[i];
167 static av_cold int flac_encode_init(AVCodecContext *avctx)
169 int freq = avctx->sample_rate;
170 int channels = avctx->channels;
171 FlacEncodeContext *s = avctx->priv_data;
177 dsputil_init(&s->dsp, avctx);
179 if(avctx->sample_fmt != SAMPLE_FMT_S16) {
183 if(channels < 1 || channels > FLAC_MAX_CHANNELS) {
186 s->channels = channels;
187 s->ch_code = s->channels-1;
189 /* find samplerate in table */
192 for(i=4; i<12; i++) {
193 if(freq == flac_samplerates[i]) {
194 s->samplerate = flac_samplerates[i];
200 /* if not in table, samplerate is non-standard */
202 if(freq % 1000 == 0 && freq < 255000) {
204 s->sr_code[1] = freq / 1000;
205 } else if(freq % 10 == 0 && freq < 655350) {
207 s->sr_code[1] = freq / 10;
208 } else if(freq < 65535) {
210 s->sr_code[1] = freq;
214 s->samplerate = freq;
217 /* set compression option defaults based on avctx->compression_level */
218 if(avctx->compression_level < 0) {
219 s->options.compression_level = 5;
221 s->options.compression_level = avctx->compression_level;
223 av_log(avctx, AV_LOG_DEBUG, " compression: %d\n", s->options.compression_level);
225 level= s->options.compression_level;
227 av_log(avctx, AV_LOG_ERROR, "invalid compression level: %d\n",
228 s->options.compression_level);
232 s->options.block_time_ms = ((int[]){ 27, 27, 27,105,105,105,105,105,105,105,105,105,105})[level];
233 s->options.use_lpc = ((int[]){ 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1})[level];
234 s->options.min_prediction_order= ((int[]){ 2, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1})[level];
235 s->options.max_prediction_order= ((int[]){ 3, 4, 4, 6, 8, 8, 8, 8, 12, 12, 12, 32, 32})[level];
236 s->options.prediction_order_method = ((int[]){ ORDER_METHOD_EST, ORDER_METHOD_EST, ORDER_METHOD_EST,
237 ORDER_METHOD_EST, ORDER_METHOD_EST, ORDER_METHOD_EST,
238 ORDER_METHOD_4LEVEL, ORDER_METHOD_LOG, ORDER_METHOD_4LEVEL,
239 ORDER_METHOD_LOG, ORDER_METHOD_SEARCH, ORDER_METHOD_LOG,
240 ORDER_METHOD_SEARCH})[level];
241 s->options.min_partition_order = ((int[]){ 2, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0})[level];
242 s->options.max_partition_order = ((int[]){ 2, 2, 3, 3, 3, 8, 8, 8, 8, 8, 8, 8, 8})[level];
244 /* set compression option overrides from AVCodecContext */
245 if(avctx->use_lpc >= 0) {
246 s->options.use_lpc = av_clip(avctx->use_lpc, 0, 11);
248 if(s->options.use_lpc == 1)
249 av_log(avctx, AV_LOG_DEBUG, " use lpc: Levinson-Durbin recursion with Welch window\n");
250 else if(s->options.use_lpc > 1)
251 av_log(avctx, AV_LOG_DEBUG, " use lpc: Cholesky factorization\n");
253 if(avctx->min_prediction_order >= 0) {
254 if(s->options.use_lpc) {
255 if(avctx->min_prediction_order < MIN_LPC_ORDER ||
256 avctx->min_prediction_order > MAX_LPC_ORDER) {
257 av_log(avctx, AV_LOG_ERROR, "invalid min prediction order: %d\n",
258 avctx->min_prediction_order);
262 if(avctx->min_prediction_order > MAX_FIXED_ORDER) {
263 av_log(avctx, AV_LOG_ERROR, "invalid min prediction order: %d\n",
264 avctx->min_prediction_order);
268 s->options.min_prediction_order = avctx->min_prediction_order;
270 if(avctx->max_prediction_order >= 0) {
271 if(s->options.use_lpc) {
272 if(avctx->max_prediction_order < MIN_LPC_ORDER ||
273 avctx->max_prediction_order > MAX_LPC_ORDER) {
274 av_log(avctx, AV_LOG_ERROR, "invalid max prediction order: %d\n",
275 avctx->max_prediction_order);
279 if(avctx->max_prediction_order > MAX_FIXED_ORDER) {
280 av_log(avctx, AV_LOG_ERROR, "invalid max prediction order: %d\n",
281 avctx->max_prediction_order);
285 s->options.max_prediction_order = avctx->max_prediction_order;
287 if(s->options.max_prediction_order < s->options.min_prediction_order) {
288 av_log(avctx, AV_LOG_ERROR, "invalid prediction orders: min=%d max=%d\n",
289 s->options.min_prediction_order, s->options.max_prediction_order);
292 av_log(avctx, AV_LOG_DEBUG, " prediction order: %d, %d\n",
293 s->options.min_prediction_order, s->options.max_prediction_order);
295 if(avctx->prediction_order_method >= 0) {
296 if(avctx->prediction_order_method > ORDER_METHOD_LOG) {
297 av_log(avctx, AV_LOG_ERROR, "invalid prediction order method: %d\n",
298 avctx->prediction_order_method);
301 s->options.prediction_order_method = avctx->prediction_order_method;
303 switch(s->options.prediction_order_method) {
304 case ORDER_METHOD_EST: av_log(avctx, AV_LOG_DEBUG, " order method: %s\n",
306 case ORDER_METHOD_2LEVEL: av_log(avctx, AV_LOG_DEBUG, " order method: %s\n",
308 case ORDER_METHOD_4LEVEL: av_log(avctx, AV_LOG_DEBUG, " order method: %s\n",
310 case ORDER_METHOD_8LEVEL: av_log(avctx, AV_LOG_DEBUG, " order method: %s\n",
312 case ORDER_METHOD_SEARCH: av_log(avctx, AV_LOG_DEBUG, " order method: %s\n",
313 "full search"); break;
314 case ORDER_METHOD_LOG: av_log(avctx, AV_LOG_DEBUG, " order method: %s\n",
315 "log search"); break;
318 if(avctx->min_partition_order >= 0) {
319 if(avctx->min_partition_order > MAX_PARTITION_ORDER) {
320 av_log(avctx, AV_LOG_ERROR, "invalid min partition order: %d\n",
321 avctx->min_partition_order);
324 s->options.min_partition_order = avctx->min_partition_order;
326 if(avctx->max_partition_order >= 0) {
327 if(avctx->max_partition_order > MAX_PARTITION_ORDER) {
328 av_log(avctx, AV_LOG_ERROR, "invalid max partition order: %d\n",
329 avctx->max_partition_order);
332 s->options.max_partition_order = avctx->max_partition_order;
334 if(s->options.max_partition_order < s->options.min_partition_order) {
335 av_log(avctx, AV_LOG_ERROR, "invalid partition orders: min=%d max=%d\n",
336 s->options.min_partition_order, s->options.max_partition_order);
339 av_log(avctx, AV_LOG_DEBUG, " partition order: %d, %d\n",
340 s->options.min_partition_order, s->options.max_partition_order);
342 if(avctx->frame_size > 0) {
343 if(avctx->frame_size < FLAC_MIN_BLOCKSIZE ||
344 avctx->frame_size > FLAC_MAX_BLOCKSIZE) {
345 av_log(avctx, AV_LOG_ERROR, "invalid block size: %d\n",
350 s->avctx->frame_size = select_blocksize(s->samplerate, s->options.block_time_ms);
352 av_log(avctx, AV_LOG_DEBUG, " block size: %d\n", s->avctx->frame_size);
354 /* set LPC precision */
355 if(avctx->lpc_coeff_precision > 0) {
356 if(avctx->lpc_coeff_precision > MAX_LPC_PRECISION) {
357 av_log(avctx, AV_LOG_ERROR, "invalid lpc coeff precision: %d\n",
358 avctx->lpc_coeff_precision);
361 s->options.lpc_coeff_precision = avctx->lpc_coeff_precision;
363 /* default LPC precision */
364 s->options.lpc_coeff_precision = 15;
366 av_log(avctx, AV_LOG_DEBUG, " lpc precision: %d\n",
367 s->options.lpc_coeff_precision);
369 /* set maximum encoded frame size in verbatim mode */
370 if(s->channels == 2) {
371 s->max_framesize = 14 + ((s->avctx->frame_size * 33 + 7) >> 3);
373 s->max_framesize = 14 + (s->avctx->frame_size * s->channels * 2);
375 s->min_encoded_framesize = 0xFFFFFF;
377 /* initialize MD5 context */
378 s->md5ctx = av_malloc(av_md5_size);
380 return AVERROR_NOMEM;
381 av_md5_init(s->md5ctx);
383 streaminfo = av_malloc(FLAC_STREAMINFO_SIZE);
384 write_streaminfo(s, streaminfo);
385 avctx->extradata = streaminfo;
386 avctx->extradata_size = FLAC_STREAMINFO_SIZE;
390 avctx->coded_frame = avcodec_alloc_frame();
391 avctx->coded_frame->key_frame = 1;
396 static void init_frame(FlacEncodeContext *s)
403 for(i=0; i<16; i++) {
404 if(s->avctx->frame_size == flac_blocksizes[i]) {
405 frame->blocksize = flac_blocksizes[i];
406 frame->bs_code[0] = i;
407 frame->bs_code[1] = 0;
412 frame->blocksize = s->avctx->frame_size;
413 if(frame->blocksize <= 256) {
414 frame->bs_code[0] = 6;
415 frame->bs_code[1] = frame->blocksize-1;
417 frame->bs_code[0] = 7;
418 frame->bs_code[1] = frame->blocksize-1;
422 for(ch=0; ch<s->channels; ch++) {
423 frame->subframes[ch].obits = 16;
428 * Copy channel-interleaved input samples into separate subframes
430 static void copy_samples(FlacEncodeContext *s, int16_t *samples)
436 for(i=0,j=0; i<frame->blocksize; i++) {
437 for(ch=0; ch<s->channels; ch++,j++) {
438 frame->subframes[ch].samples[i] = samples[j];
444 #define rice_encode_count(sum, n, k) (((n)*((k)+1))+((sum-(n>>1))>>(k)))
447 * Solve for d/dk(rice_encode_count) = n-((sum-(n>>1))>>(k+1)) = 0
449 static int find_optimal_param(uint32_t sum, int n)
457 k = av_log2(n<256 ? FASTDIV(sum2,n) : sum2/n);
458 return FFMIN(k, MAX_RICE_PARAM);
461 static uint32_t calc_optimal_rice_params(RiceContext *rc, int porder,
462 uint32_t *sums, int n, int pred_order)
468 part = (1 << porder);
471 cnt = (n >> porder) - pred_order;
472 for(i=0; i<part; i++) {
473 k = find_optimal_param(sums[i], cnt);
475 all_bits += rice_encode_count(sums[i], cnt, k);
484 static void calc_sums(int pmin, int pmax, uint32_t *data, int n, int pred_order,
485 uint32_t sums[][MAX_PARTITIONS])
489 uint32_t *res, *res_end;
491 /* sums for highest level */
493 res = &data[pred_order];
494 res_end = &data[n >> pmax];
495 for(i=0; i<parts; i++) {
497 while(res < res_end){
503 /* sums for lower levels */
504 for(i=pmax-1; i>=pmin; i--) {
506 for(j=0; j<parts; j++) {
507 sums[i][j] = sums[i+1][2*j] + sums[i+1][2*j+1];
512 static uint32_t calc_rice_params(RiceContext *rc, int pmin, int pmax,
513 int32_t *data, int n, int pred_order)
516 uint32_t bits[MAX_PARTITION_ORDER+1];
520 uint32_t sums[MAX_PARTITION_ORDER+1][MAX_PARTITIONS];
522 assert(pmin >= 0 && pmin <= MAX_PARTITION_ORDER);
523 assert(pmax >= 0 && pmax <= MAX_PARTITION_ORDER);
524 assert(pmin <= pmax);
526 udata = av_malloc(n * sizeof(uint32_t));
528 udata[i] = (2*data[i]) ^ (data[i]>>31);
531 calc_sums(pmin, pmax, udata, n, pred_order, sums);
534 bits[pmin] = UINT32_MAX;
535 for(i=pmin; i<=pmax; i++) {
536 bits[i] = calc_optimal_rice_params(&tmp_rc, i, sums[i], n, pred_order);
537 if(bits[i] <= bits[opt_porder]) {
544 return bits[opt_porder];
547 static int get_max_p_order(int max_porder, int n, int order)
549 int porder = FFMIN(max_porder, av_log2(n^(n-1)));
551 porder = FFMIN(porder, av_log2(n/order));
555 static uint32_t calc_rice_params_fixed(RiceContext *rc, int pmin, int pmax,
556 int32_t *data, int n, int pred_order,
560 pmin = get_max_p_order(pmin, n, pred_order);
561 pmax = get_max_p_order(pmax, n, pred_order);
562 bits = pred_order*bps + 6;
563 bits += calc_rice_params(rc, pmin, pmax, data, n, pred_order);
567 static uint32_t calc_rice_params_lpc(RiceContext *rc, int pmin, int pmax,
568 int32_t *data, int n, int pred_order,
569 int bps, int precision)
572 pmin = get_max_p_order(pmin, n, pred_order);
573 pmax = get_max_p_order(pmax, n, pred_order);
574 bits = pred_order*bps + 4 + 5 + pred_order*precision + 6;
575 bits += calc_rice_params(rc, pmin, pmax, data, n, pred_order);
580 * Apply Welch window function to audio block
582 static void apply_welch_window(const int32_t *data, int len, double *w_data)
588 assert(!(len&1)); //the optimization in r11881 does not support odd len
589 //if someone wants odd len extend the change in r11881
592 c = 2.0 / (len - 1.0);
596 for(i=0; i<n2; i++) {
599 w_data[-i-1] = data[-i-1] * w;
600 w_data[+i ] = data[+i ] * w;
605 * Calculates autocorrelation data from audio samples
606 * A Welch window function is applied before calculation.
608 void ff_flac_compute_autocorr(const int32_t *data, int len, int lag,
612 double tmp[len + lag + 1];
613 double *data1= tmp + lag;
615 apply_welch_window(data, len, data1);
621 for(j=0; j<lag; j+=2){
622 double sum0 = 1.0, sum1 = 1.0;
623 for(i=0; i<len; i++){
624 sum0 += data1[i] * data1[i-j];
625 sum1 += data1[i] * data1[i-j-1];
633 for(i=0; i<len; i+=2){
634 sum += data1[i ] * data1[i-j ]
635 + data1[i+1] * data1[i-j+1];
642 static void encode_residual_verbatim(int32_t *res, int32_t *smp, int n)
645 memcpy(res, smp, n * sizeof(int32_t));
648 static void encode_residual_fixed(int32_t *res, const int32_t *smp, int n,
653 for(i=0; i<order; i++) {
658 for(i=order; i<n; i++)
661 for(i=order; i<n; i++)
662 res[i]= smp[i] - smp[i-1];
664 int a = smp[order-1] - smp[order-2];
665 for(i=order; i<n; i+=2) {
666 int b = smp[i] - smp[i-1];
668 a = smp[i+1] - smp[i];
672 int a = smp[order-1] - smp[order-2];
673 int c = smp[order-1] - 2*smp[order-2] + smp[order-3];
674 for(i=order; i<n; i+=2) {
675 int b = smp[i] - smp[i-1];
678 a = smp[i+1] - smp[i];
683 int a = smp[order-1] - smp[order-2];
684 int c = smp[order-1] - 2*smp[order-2] + smp[order-3];
685 int e = smp[order-1] - 3*smp[order-2] + 3*smp[order-3] - smp[order-4];
686 for(i=order; i<n; i+=2) {
687 int b = smp[i] - smp[i-1];
691 a = smp[i+1] - smp[i];
700 int c = coefs[(x)-1];\
706 static av_always_inline void encode_residual_lpc_unrolled(
707 int32_t *res, const int32_t *smp, int n,
708 int order, const int32_t *coefs, int shift, int big)
711 for(i=order; i<n; i+=2) {
712 int s = smp[i-order];
761 res[i ] = smp[i ] - (p0 >> shift);
762 res[i+1] = smp[i+1] - (p1 >> shift);
766 static void encode_residual_lpc(int32_t *res, const int32_t *smp, int n,
767 int order, const int32_t *coefs, int shift)
770 for(i=0; i<order; i++) {
774 for(i=order; i<n; i+=2) {
778 for(j=0; j<order; j++) {
784 res[i ] = smp[i ] - (p0 >> shift);
785 res[i+1] = smp[i+1] - (p1 >> shift);
789 case 1: encode_residual_lpc_unrolled(res, smp, n, 1, coefs, shift, 0); break;
790 case 2: encode_residual_lpc_unrolled(res, smp, n, 2, coefs, shift, 0); break;
791 case 3: encode_residual_lpc_unrolled(res, smp, n, 3, coefs, shift, 0); break;
792 case 4: encode_residual_lpc_unrolled(res, smp, n, 4, coefs, shift, 0); break;
793 case 5: encode_residual_lpc_unrolled(res, smp, n, 5, coefs, shift, 0); break;
794 case 6: encode_residual_lpc_unrolled(res, smp, n, 6, coefs, shift, 0); break;
795 case 7: encode_residual_lpc_unrolled(res, smp, n, 7, coefs, shift, 0); break;
796 case 8: encode_residual_lpc_unrolled(res, smp, n, 8, coefs, shift, 0); break;
797 default: encode_residual_lpc_unrolled(res, smp, n, order, coefs, shift, 1); break;
802 static int encode_residual(FlacEncodeContext *ctx, int ch)
805 int min_order, max_order, opt_order, precision, omethod;
806 int min_porder, max_porder;
809 int32_t coefs[MAX_LPC_ORDER][MAX_LPC_ORDER];
810 int shift[MAX_LPC_ORDER];
814 sub = &frame->subframes[ch];
817 n = frame->blocksize;
821 if(smp[i] != smp[0]) break;
824 sub->type = sub->type_code = FLAC_SUBFRAME_CONSTANT;
831 sub->type = sub->type_code = FLAC_SUBFRAME_VERBATIM;
832 encode_residual_verbatim(res, smp, n);
833 return sub->obits * n;
836 min_order = ctx->options.min_prediction_order;
837 max_order = ctx->options.max_prediction_order;
838 min_porder = ctx->options.min_partition_order;
839 max_porder = ctx->options.max_partition_order;
840 precision = ctx->options.lpc_coeff_precision;
841 omethod = ctx->options.prediction_order_method;
844 if(!ctx->options.use_lpc || max_order == 0 || (n <= max_order)) {
845 uint32_t bits[MAX_FIXED_ORDER+1];
846 if(max_order > MAX_FIXED_ORDER) max_order = MAX_FIXED_ORDER;
848 bits[0] = UINT32_MAX;
849 for(i=min_order; i<=max_order; i++) {
850 encode_residual_fixed(res, smp, n, i);
851 bits[i] = calc_rice_params_fixed(&sub->rc, min_porder, max_porder, res,
853 if(bits[i] < bits[opt_order]) {
857 sub->order = opt_order;
858 sub->type = FLAC_SUBFRAME_FIXED;
859 sub->type_code = sub->type | sub->order;
860 if(sub->order != max_order) {
861 encode_residual_fixed(res, smp, n, sub->order);
862 return calc_rice_params_fixed(&sub->rc, min_porder, max_porder, res, n,
863 sub->order, sub->obits);
865 return bits[sub->order];
869 opt_order = ff_lpc_calc_coefs(&ctx->dsp, smp, n, min_order, max_order,
870 precision, coefs, shift, ctx->options.use_lpc,
871 omethod, MAX_LPC_SHIFT, 0);
873 if(omethod == ORDER_METHOD_2LEVEL ||
874 omethod == ORDER_METHOD_4LEVEL ||
875 omethod == ORDER_METHOD_8LEVEL) {
876 int levels = 1 << omethod;
877 uint32_t bits[levels];
879 int opt_index = levels-1;
880 opt_order = max_order-1;
881 bits[opt_index] = UINT32_MAX;
882 for(i=levels-1; i>=0; i--) {
883 order = min_order + (((max_order-min_order+1) * (i+1)) / levels)-1;
884 if(order < 0) order = 0;
885 encode_residual_lpc(res, smp, n, order+1, coefs[order], shift[order]);
886 bits[i] = calc_rice_params_lpc(&sub->rc, min_porder, max_porder,
887 res, n, order+1, sub->obits, precision);
888 if(bits[i] < bits[opt_index]) {
894 } else if(omethod == ORDER_METHOD_SEARCH) {
895 // brute-force optimal order search
896 uint32_t bits[MAX_LPC_ORDER];
898 bits[0] = UINT32_MAX;
899 for(i=min_order-1; i<max_order; i++) {
900 encode_residual_lpc(res, smp, n, i+1, coefs[i], shift[i]);
901 bits[i] = calc_rice_params_lpc(&sub->rc, min_porder, max_porder,
902 res, n, i+1, sub->obits, precision);
903 if(bits[i] < bits[opt_order]) {
908 } else if(omethod == ORDER_METHOD_LOG) {
909 uint32_t bits[MAX_LPC_ORDER];
912 opt_order= min_order - 1 + (max_order-min_order)/3;
913 memset(bits, -1, sizeof(bits));
915 for(step=16 ;step; step>>=1){
917 for(i=last-step; i<=last+step; i+= step){
918 if(i<min_order-1 || i>=max_order || bits[i] < UINT32_MAX)
920 encode_residual_lpc(res, smp, n, i+1, coefs[i], shift[i]);
921 bits[i] = calc_rice_params_lpc(&sub->rc, min_porder, max_porder,
922 res, n, i+1, sub->obits, precision);
923 if(bits[i] < bits[opt_order])
930 sub->order = opt_order;
931 sub->type = FLAC_SUBFRAME_LPC;
932 sub->type_code = sub->type | (sub->order-1);
933 sub->shift = shift[sub->order-1];
934 for(i=0; i<sub->order; i++) {
935 sub->coefs[i] = coefs[sub->order-1][i];
937 encode_residual_lpc(res, smp, n, sub->order, sub->coefs, sub->shift);
938 return calc_rice_params_lpc(&sub->rc, min_porder, max_porder, res, n, sub->order,
939 sub->obits, precision);
942 static int encode_residual_v(FlacEncodeContext *ctx, int ch)
950 sub = &frame->subframes[ch];
953 n = frame->blocksize;
957 if(smp[i] != smp[0]) break;
960 sub->type = sub->type_code = FLAC_SUBFRAME_CONSTANT;
966 sub->type = sub->type_code = FLAC_SUBFRAME_VERBATIM;
967 encode_residual_verbatim(res, smp, n);
968 return sub->obits * n;
971 static int estimate_stereo_mode(int32_t *left_ch, int32_t *right_ch, int n)
979 /* calculate sum of 2nd order residual for each channel */
980 sum[0] = sum[1] = sum[2] = sum[3] = 0;
982 lt = left_ch[i] - 2*left_ch[i-1] + left_ch[i-2];
983 rt = right_ch[i] - 2*right_ch[i-1] + right_ch[i-2];
984 sum[2] += FFABS((lt + rt) >> 1);
985 sum[3] += FFABS(lt - rt);
989 /* estimate bit counts */
991 k = find_optimal_param(2*sum[i], n);
992 sum[i] = rice_encode_count(2*sum[i], n, k);
995 /* calculate score for each mode */
996 score[0] = sum[0] + sum[1];
997 score[1] = sum[0] + sum[3];
998 score[2] = sum[1] + sum[3];
999 score[3] = sum[2] + sum[3];
1001 /* return mode with lowest score */
1003 for(i=1; i<4; i++) {
1004 if(score[i] < score[best]) {
1009 return FLAC_CHMODE_LEFT_RIGHT;
1010 } else if(best == 1) {
1011 return FLAC_CHMODE_LEFT_SIDE;
1012 } else if(best == 2) {
1013 return FLAC_CHMODE_RIGHT_SIDE;
1015 return FLAC_CHMODE_MID_SIDE;
1020 * Perform stereo channel decorrelation
1022 static void channel_decorrelation(FlacEncodeContext *ctx)
1025 int32_t *left, *right;
1028 frame = &ctx->frame;
1029 n = frame->blocksize;
1030 left = frame->subframes[0].samples;
1031 right = frame->subframes[1].samples;
1033 if(ctx->channels != 2) {
1034 frame->ch_mode = FLAC_CHMODE_NOT_STEREO;
1038 frame->ch_mode = estimate_stereo_mode(left, right, n);
1040 /* perform decorrelation and adjust bits-per-sample */
1041 if(frame->ch_mode == FLAC_CHMODE_LEFT_RIGHT) {
1044 if(frame->ch_mode == FLAC_CHMODE_MID_SIDE) {
1046 for(i=0; i<n; i++) {
1048 left[i] = (tmp + right[i]) >> 1;
1049 right[i] = tmp - right[i];
1051 frame->subframes[1].obits++;
1052 } else if(frame->ch_mode == FLAC_CHMODE_LEFT_SIDE) {
1053 for(i=0; i<n; i++) {
1054 right[i] = left[i] - right[i];
1056 frame->subframes[1].obits++;
1058 for(i=0; i<n; i++) {
1059 left[i] -= right[i];
1061 frame->subframes[0].obits++;
1065 static void write_utf8(PutBitContext *pb, uint32_t val)
1068 PUT_UTF8(val, tmp, put_bits(pb, 8, tmp);)
1071 static void output_frame_header(FlacEncodeContext *s)
1078 put_bits(&s->pb, 16, 0xFFF8);
1079 put_bits(&s->pb, 4, frame->bs_code[0]);
1080 put_bits(&s->pb, 4, s->sr_code[0]);
1081 if(frame->ch_mode == FLAC_CHMODE_NOT_STEREO) {
1082 put_bits(&s->pb, 4, s->ch_code);
1084 put_bits(&s->pb, 4, frame->ch_mode);
1086 put_bits(&s->pb, 3, 4); /* bits-per-sample code */
1087 put_bits(&s->pb, 1, 0);
1088 write_utf8(&s->pb, s->frame_count);
1089 if(frame->bs_code[0] == 6) {
1090 put_bits(&s->pb, 8, frame->bs_code[1]);
1091 } else if(frame->bs_code[0] == 7) {
1092 put_bits(&s->pb, 16, frame->bs_code[1]);
1094 if(s->sr_code[0] == 12) {
1095 put_bits(&s->pb, 8, s->sr_code[1]);
1096 } else if(s->sr_code[0] > 12) {
1097 put_bits(&s->pb, 16, s->sr_code[1]);
1099 flush_put_bits(&s->pb);
1100 crc = av_crc(av_crc_get_table(AV_CRC_8_ATM), 0,
1101 s->pb.buf, put_bits_count(&s->pb)>>3);
1102 put_bits(&s->pb, 8, crc);
1105 static void output_subframe_constant(FlacEncodeContext *s, int ch)
1110 sub = &s->frame.subframes[ch];
1111 res = sub->residual[0];
1112 put_sbits(&s->pb, sub->obits, res);
1115 static void output_subframe_verbatim(FlacEncodeContext *s, int ch)
1123 sub = &frame->subframes[ch];
1125 for(i=0; i<frame->blocksize; i++) {
1126 res = sub->residual[i];
1127 put_sbits(&s->pb, sub->obits, res);
1131 static void output_residual(FlacEncodeContext *ctx, int ch)
1133 int i, j, p, n, parts;
1134 int k, porder, psize, res_cnt;
1139 frame = &ctx->frame;
1140 sub = &frame->subframes[ch];
1141 res = sub->residual;
1142 n = frame->blocksize;
1144 /* rice-encoded block */
1145 put_bits(&ctx->pb, 2, 0);
1147 /* partition order */
1148 porder = sub->rc.porder;
1149 psize = n >> porder;
1150 parts = (1 << porder);
1151 put_bits(&ctx->pb, 4, porder);
1152 res_cnt = psize - sub->order;
1156 for(p=0; p<parts; p++) {
1157 k = sub->rc.params[p];
1158 put_bits(&ctx->pb, 4, k);
1159 if(p == 1) res_cnt = psize;
1160 for(i=0; i<res_cnt && j<n; i++, j++) {
1161 set_sr_golomb_flac(&ctx->pb, res[j], k, INT32_MAX, 0);
1166 static void output_subframe_fixed(FlacEncodeContext *ctx, int ch)
1172 frame = &ctx->frame;
1173 sub = &frame->subframes[ch];
1175 /* warm-up samples */
1176 for(i=0; i<sub->order; i++) {
1177 put_sbits(&ctx->pb, sub->obits, sub->residual[i]);
1181 output_residual(ctx, ch);
1184 static void output_subframe_lpc(FlacEncodeContext *ctx, int ch)
1190 frame = &ctx->frame;
1191 sub = &frame->subframes[ch];
1193 /* warm-up samples */
1194 for(i=0; i<sub->order; i++) {
1195 put_sbits(&ctx->pb, sub->obits, sub->residual[i]);
1198 /* LPC coefficients */
1199 cbits = ctx->options.lpc_coeff_precision;
1200 put_bits(&ctx->pb, 4, cbits-1);
1201 put_sbits(&ctx->pb, 5, sub->shift);
1202 for(i=0; i<sub->order; i++) {
1203 put_sbits(&ctx->pb, cbits, sub->coefs[i]);
1207 output_residual(ctx, ch);
1210 static void output_subframes(FlacEncodeContext *s)
1218 for(ch=0; ch<s->channels; ch++) {
1219 sub = &frame->subframes[ch];
1221 /* subframe header */
1222 put_bits(&s->pb, 1, 0);
1223 put_bits(&s->pb, 6, sub->type_code);
1224 put_bits(&s->pb, 1, 0); /* no wasted bits */
1227 if(sub->type == FLAC_SUBFRAME_CONSTANT) {
1228 output_subframe_constant(s, ch);
1229 } else if(sub->type == FLAC_SUBFRAME_VERBATIM) {
1230 output_subframe_verbatim(s, ch);
1231 } else if(sub->type == FLAC_SUBFRAME_FIXED) {
1232 output_subframe_fixed(s, ch);
1233 } else if(sub->type == FLAC_SUBFRAME_LPC) {
1234 output_subframe_lpc(s, ch);
1239 static void output_frame_footer(FlacEncodeContext *s)
1242 flush_put_bits(&s->pb);
1243 crc = bswap_16(av_crc(av_crc_get_table(AV_CRC_16_ANSI), 0,
1244 s->pb.buf, put_bits_count(&s->pb)>>3));
1245 put_bits(&s->pb, 16, crc);
1246 flush_put_bits(&s->pb);
1249 static void update_md5_sum(FlacEncodeContext *s, int16_t *samples)
1251 #ifdef WORDS_BIGENDIAN
1253 for(i = 0; i < s->frame.blocksize*s->channels; i++) {
1254 int16_t smp = le2me_16(samples[i]);
1255 av_md5_update(s->md5ctx, (uint8_t *)&smp, 2);
1258 av_md5_update(s->md5ctx, (uint8_t *)samples, s->frame.blocksize*s->channels*2);
1262 static int flac_encode_frame(AVCodecContext *avctx, uint8_t *frame,
1263 int buf_size, void *data)
1266 FlacEncodeContext *s;
1267 int16_t *samples = data;
1271 s = avctx->priv_data;
1273 if(buf_size < s->max_framesize*2) {
1274 av_log(avctx, AV_LOG_ERROR, "output buffer too small\n");
1278 /* when the last block is reached, update the header in extradata */
1280 s->min_framesize = s->min_encoded_framesize;
1281 s->max_framesize = s->max_encoded_framesize;
1282 av_md5_final(s->md5ctx, s->md5sum);
1283 write_streaminfo(s, avctx->extradata);
1289 copy_samples(s, samples);
1291 channel_decorrelation(s);
1293 for(ch=0; ch<s->channels; ch++) {
1294 encode_residual(s, ch);
1298 init_put_bits(&s->pb, frame, buf_size);
1299 output_frame_header(s);
1300 output_subframes(s);
1301 output_frame_footer(s);
1302 out_bytes = put_bits_count(&s->pb) >> 3;
1304 if(out_bytes > s->max_framesize) {
1306 /* still too large. must be an error. */
1307 av_log(avctx, AV_LOG_ERROR, "error encoding frame\n");
1311 /* frame too large. use verbatim mode */
1312 for(ch=0; ch<s->channels; ch++) {
1313 encode_residual_v(s, ch);
1320 s->sample_count += avctx->frame_size;
1321 update_md5_sum(s, samples);
1322 if (out_bytes > s->max_encoded_framesize)
1323 s->max_encoded_framesize = out_bytes;
1324 if (out_bytes < s->min_encoded_framesize)
1325 s->min_encoded_framesize = out_bytes;
1330 static av_cold int flac_encode_close(AVCodecContext *avctx)
1332 if (avctx->priv_data) {
1333 FlacEncodeContext *s = avctx->priv_data;
1334 av_freep(&s->md5ctx);
1336 av_freep(&avctx->extradata);
1337 avctx->extradata_size = 0;
1338 av_freep(&avctx->coded_frame);
1342 AVCodec flac_encoder = {
1346 sizeof(FlacEncodeContext),
1351 .capabilities = CODEC_CAP_SMALL_LAST_FRAME | CODEC_CAP_DELAY,
1352 .sample_fmts = (enum SampleFormat[]){SAMPLE_FMT_S16,SAMPLE_FMT_NONE},
1353 .long_name = NULL_IF_CONFIG_SMALL("FLAC (Free Lossless Audio Codec)"),