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 {
100 uint64_t sample_count;
103 CompressionOptions options;
104 AVCodecContext *avctx;
106 struct AVMD5 *md5ctx;
109 static const int flac_samplerates[16] = {
111 8000, 16000, 22050, 24000, 32000, 44100, 48000, 96000,
115 static const int flac_blocksizes[16] = {
118 576, 1152, 2304, 4608,
120 256, 512, 1024, 2048, 4096, 8192, 16384, 32768
124 * Writes streaminfo metadata block to byte array
126 static void write_streaminfo(FlacEncodeContext *s, uint8_t *header)
130 memset(header, 0, FLAC_STREAMINFO_SIZE);
131 init_put_bits(&pb, header, FLAC_STREAMINFO_SIZE);
133 /* streaminfo metadata block */
134 put_bits(&pb, 16, s->avctx->frame_size);
135 put_bits(&pb, 16, s->avctx->frame_size);
136 put_bits(&pb, 24, 0);
137 put_bits(&pb, 24, s->max_framesize);
138 put_bits(&pb, 20, s->samplerate);
139 put_bits(&pb, 3, s->channels-1);
140 put_bits(&pb, 5, 15); /* bits per sample - 1 */
141 /* write 36-bit sample count in 2 put_bits() calls */
142 put_bits(&pb, 24, (s->sample_count & 0xFFFFFF000LL) >> 12);
143 put_bits(&pb, 12, s->sample_count & 0x000000FFFLL);
145 memcpy(&header[18], s->md5sum, 16);
149 * Sets blocksize based on samplerate
150 * Chooses the closest predefined blocksize >= BLOCK_TIME_MS milliseconds
152 static int select_blocksize(int samplerate, int block_time_ms)
158 assert(samplerate > 0);
159 blocksize = flac_blocksizes[1];
160 target = (samplerate * block_time_ms) / 1000;
161 for(i=0; i<16; i++) {
162 if(target >= flac_blocksizes[i] && flac_blocksizes[i] > blocksize) {
163 blocksize = flac_blocksizes[i];
169 static av_cold int flac_encode_init(AVCodecContext *avctx)
171 int freq = avctx->sample_rate;
172 int channels = avctx->channels;
173 FlacEncodeContext *s = avctx->priv_data;
179 dsputil_init(&s->dsp, avctx);
181 if(avctx->sample_fmt != SAMPLE_FMT_S16) {
185 if(channels < 1 || channels > FLAC_MAX_CH) {
188 s->channels = channels;
189 s->ch_code = s->channels-1;
191 /* find samplerate in table */
194 for(i=4; i<12; i++) {
195 if(freq == flac_samplerates[i]) {
196 s->samplerate = flac_samplerates[i];
202 /* if not in table, samplerate is non-standard */
204 if(freq % 1000 == 0 && freq < 255000) {
206 s->sr_code[1] = freq / 1000;
207 } else if(freq % 10 == 0 && freq < 655350) {
209 s->sr_code[1] = freq / 10;
210 } else if(freq < 65535) {
212 s->sr_code[1] = freq;
216 s->samplerate = freq;
219 /* set compression option defaults based on avctx->compression_level */
220 if(avctx->compression_level < 0) {
221 s->options.compression_level = 5;
223 s->options.compression_level = avctx->compression_level;
225 av_log(avctx, AV_LOG_DEBUG, " compression: %d\n", s->options.compression_level);
227 level= s->options.compression_level;
229 av_log(avctx, AV_LOG_ERROR, "invalid compression level: %d\n",
230 s->options.compression_level);
234 s->options.block_time_ms = ((int[]){ 27, 27, 27,105,105,105,105,105,105,105,105,105,105})[level];
235 s->options.use_lpc = ((int[]){ 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1})[level];
236 s->options.min_prediction_order= ((int[]){ 2, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1})[level];
237 s->options.max_prediction_order= ((int[]){ 3, 4, 4, 6, 8, 8, 8, 8, 12, 12, 12, 32, 32})[level];
238 s->options.prediction_order_method = ((int[]){ ORDER_METHOD_EST, ORDER_METHOD_EST, ORDER_METHOD_EST,
239 ORDER_METHOD_EST, ORDER_METHOD_EST, ORDER_METHOD_EST,
240 ORDER_METHOD_4LEVEL, ORDER_METHOD_LOG, ORDER_METHOD_4LEVEL,
241 ORDER_METHOD_LOG, ORDER_METHOD_SEARCH, ORDER_METHOD_LOG,
242 ORDER_METHOD_SEARCH})[level];
243 s->options.min_partition_order = ((int[]){ 2, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0})[level];
244 s->options.max_partition_order = ((int[]){ 2, 2, 3, 3, 3, 8, 8, 8, 8, 8, 8, 8, 8})[level];
246 /* set compression option overrides from AVCodecContext */
247 if(avctx->use_lpc >= 0) {
248 s->options.use_lpc = av_clip(avctx->use_lpc, 0, 11);
250 if(s->options.use_lpc == 1)
251 av_log(avctx, AV_LOG_DEBUG, " use lpc: Levinson-Durbin recursion with Welch window\n");
252 else if(s->options.use_lpc > 1)
253 av_log(avctx, AV_LOG_DEBUG, " use lpc: Cholesky factorization\n");
255 if(avctx->min_prediction_order >= 0) {
256 if(s->options.use_lpc) {
257 if(avctx->min_prediction_order < MIN_LPC_ORDER ||
258 avctx->min_prediction_order > MAX_LPC_ORDER) {
259 av_log(avctx, AV_LOG_ERROR, "invalid min prediction order: %d\n",
260 avctx->min_prediction_order);
264 if(avctx->min_prediction_order > MAX_FIXED_ORDER) {
265 av_log(avctx, AV_LOG_ERROR, "invalid min prediction order: %d\n",
266 avctx->min_prediction_order);
270 s->options.min_prediction_order = avctx->min_prediction_order;
272 if(avctx->max_prediction_order >= 0) {
273 if(s->options.use_lpc) {
274 if(avctx->max_prediction_order < MIN_LPC_ORDER ||
275 avctx->max_prediction_order > MAX_LPC_ORDER) {
276 av_log(avctx, AV_LOG_ERROR, "invalid max prediction order: %d\n",
277 avctx->max_prediction_order);
281 if(avctx->max_prediction_order > MAX_FIXED_ORDER) {
282 av_log(avctx, AV_LOG_ERROR, "invalid max prediction order: %d\n",
283 avctx->max_prediction_order);
287 s->options.max_prediction_order = avctx->max_prediction_order;
289 if(s->options.max_prediction_order < s->options.min_prediction_order) {
290 av_log(avctx, AV_LOG_ERROR, "invalid prediction orders: min=%d max=%d\n",
291 s->options.min_prediction_order, s->options.max_prediction_order);
294 av_log(avctx, AV_LOG_DEBUG, " prediction order: %d, %d\n",
295 s->options.min_prediction_order, s->options.max_prediction_order);
297 if(avctx->prediction_order_method >= 0) {
298 if(avctx->prediction_order_method > ORDER_METHOD_LOG) {
299 av_log(avctx, AV_LOG_ERROR, "invalid prediction order method: %d\n",
300 avctx->prediction_order_method);
303 s->options.prediction_order_method = avctx->prediction_order_method;
305 switch(s->options.prediction_order_method) {
306 case ORDER_METHOD_EST: av_log(avctx, AV_LOG_DEBUG, " order method: %s\n",
308 case ORDER_METHOD_2LEVEL: av_log(avctx, AV_LOG_DEBUG, " order method: %s\n",
310 case ORDER_METHOD_4LEVEL: av_log(avctx, AV_LOG_DEBUG, " order method: %s\n",
312 case ORDER_METHOD_8LEVEL: av_log(avctx, AV_LOG_DEBUG, " order method: %s\n",
314 case ORDER_METHOD_SEARCH: av_log(avctx, AV_LOG_DEBUG, " order method: %s\n",
315 "full search"); break;
316 case ORDER_METHOD_LOG: av_log(avctx, AV_LOG_DEBUG, " order method: %s\n",
317 "log search"); break;
320 if(avctx->min_partition_order >= 0) {
321 if(avctx->min_partition_order > MAX_PARTITION_ORDER) {
322 av_log(avctx, AV_LOG_ERROR, "invalid min partition order: %d\n",
323 avctx->min_partition_order);
326 s->options.min_partition_order = avctx->min_partition_order;
328 if(avctx->max_partition_order >= 0) {
329 if(avctx->max_partition_order > MAX_PARTITION_ORDER) {
330 av_log(avctx, AV_LOG_ERROR, "invalid max partition order: %d\n",
331 avctx->max_partition_order);
334 s->options.max_partition_order = avctx->max_partition_order;
336 if(s->options.max_partition_order < s->options.min_partition_order) {
337 av_log(avctx, AV_LOG_ERROR, "invalid partition orders: min=%d max=%d\n",
338 s->options.min_partition_order, s->options.max_partition_order);
341 av_log(avctx, AV_LOG_DEBUG, " partition order: %d, %d\n",
342 s->options.min_partition_order, s->options.max_partition_order);
344 if(avctx->frame_size > 0) {
345 if(avctx->frame_size < FLAC_MIN_BLOCKSIZE ||
346 avctx->frame_size > FLAC_MAX_BLOCKSIZE) {
347 av_log(avctx, AV_LOG_ERROR, "invalid block size: %d\n",
352 s->avctx->frame_size = select_blocksize(s->samplerate, s->options.block_time_ms);
354 av_log(avctx, AV_LOG_DEBUG, " block size: %d\n", s->avctx->frame_size);
356 /* set LPC precision */
357 if(avctx->lpc_coeff_precision > 0) {
358 if(avctx->lpc_coeff_precision > MAX_LPC_PRECISION) {
359 av_log(avctx, AV_LOG_ERROR, "invalid lpc coeff precision: %d\n",
360 avctx->lpc_coeff_precision);
363 s->options.lpc_coeff_precision = avctx->lpc_coeff_precision;
365 /* default LPC precision */
366 s->options.lpc_coeff_precision = 15;
368 av_log(avctx, AV_LOG_DEBUG, " lpc precision: %d\n",
369 s->options.lpc_coeff_precision);
371 /* set maximum encoded frame size in verbatim mode */
372 if(s->channels == 2) {
373 s->max_framesize = 14 + ((s->avctx->frame_size * 33 + 7) >> 3);
375 s->max_framesize = 14 + (s->avctx->frame_size * s->channels * 2);
378 /* initialize MD5 context */
379 s->md5ctx = av_malloc(av_md5_size);
381 return AVERROR_NOMEM;
382 av_md5_init(s->md5ctx);
384 streaminfo = av_malloc(FLAC_STREAMINFO_SIZE);
385 write_streaminfo(s, streaminfo);
386 avctx->extradata = streaminfo;
387 avctx->extradata_size = FLAC_STREAMINFO_SIZE;
391 avctx->coded_frame = avcodec_alloc_frame();
392 avctx->coded_frame->key_frame = 1;
397 static void init_frame(FlacEncodeContext *s)
404 for(i=0; i<16; i++) {
405 if(s->avctx->frame_size == flac_blocksizes[i]) {
406 frame->blocksize = flac_blocksizes[i];
407 frame->bs_code[0] = i;
408 frame->bs_code[1] = 0;
413 frame->blocksize = s->avctx->frame_size;
414 if(frame->blocksize <= 256) {
415 frame->bs_code[0] = 6;
416 frame->bs_code[1] = frame->blocksize-1;
418 frame->bs_code[0] = 7;
419 frame->bs_code[1] = frame->blocksize-1;
423 for(ch=0; ch<s->channels; ch++) {
424 frame->subframes[ch].obits = 16;
429 * Copy channel-interleaved input samples into separate subframes
431 static void copy_samples(FlacEncodeContext *s, int16_t *samples)
437 for(i=0,j=0; i<frame->blocksize; i++) {
438 for(ch=0; ch<s->channels; ch++,j++) {
439 frame->subframes[ch].samples[i] = samples[j];
445 #define rice_encode_count(sum, n, k) (((n)*((k)+1))+((sum-(n>>1))>>(k)))
448 * Solve for d/dk(rice_encode_count) = n-((sum-(n>>1))>>(k+1)) = 0
450 static int find_optimal_param(uint32_t sum, int n)
458 k = av_log2(n<256 ? FASTDIV(sum2,n) : sum2/n);
459 return FFMIN(k, MAX_RICE_PARAM);
462 static uint32_t calc_optimal_rice_params(RiceContext *rc, int porder,
463 uint32_t *sums, int n, int pred_order)
469 part = (1 << porder);
472 cnt = (n >> porder) - pred_order;
473 for(i=0; i<part; i++) {
474 k = find_optimal_param(sums[i], cnt);
476 all_bits += rice_encode_count(sums[i], cnt, k);
485 static void calc_sums(int pmin, int pmax, uint32_t *data, int n, int pred_order,
486 uint32_t sums[][MAX_PARTITIONS])
490 uint32_t *res, *res_end;
492 /* sums for highest level */
494 res = &data[pred_order];
495 res_end = &data[n >> pmax];
496 for(i=0; i<parts; i++) {
498 while(res < res_end){
504 /* sums for lower levels */
505 for(i=pmax-1; i>=pmin; i--) {
507 for(j=0; j<parts; j++) {
508 sums[i][j] = sums[i+1][2*j] + sums[i+1][2*j+1];
513 static uint32_t calc_rice_params(RiceContext *rc, int pmin, int pmax,
514 int32_t *data, int n, int pred_order)
517 uint32_t bits[MAX_PARTITION_ORDER+1];
521 uint32_t sums[MAX_PARTITION_ORDER+1][MAX_PARTITIONS];
523 assert(pmin >= 0 && pmin <= MAX_PARTITION_ORDER);
524 assert(pmax >= 0 && pmax <= MAX_PARTITION_ORDER);
525 assert(pmin <= pmax);
527 udata = av_malloc(n * sizeof(uint32_t));
529 udata[i] = (2*data[i]) ^ (data[i]>>31);
532 calc_sums(pmin, pmax, udata, n, pred_order, sums);
535 bits[pmin] = UINT32_MAX;
536 for(i=pmin; i<=pmax; i++) {
537 bits[i] = calc_optimal_rice_params(&tmp_rc, i, sums[i], n, pred_order);
538 if(bits[i] <= bits[opt_porder]) {
545 return bits[opt_porder];
548 static int get_max_p_order(int max_porder, int n, int order)
550 int porder = FFMIN(max_porder, av_log2(n^(n-1)));
552 porder = FFMIN(porder, av_log2(n/order));
556 static uint32_t calc_rice_params_fixed(RiceContext *rc, int pmin, int pmax,
557 int32_t *data, int n, int pred_order,
561 pmin = get_max_p_order(pmin, n, pred_order);
562 pmax = get_max_p_order(pmax, n, pred_order);
563 bits = pred_order*bps + 6;
564 bits += calc_rice_params(rc, pmin, pmax, data, n, pred_order);
568 static uint32_t calc_rice_params_lpc(RiceContext *rc, int pmin, int pmax,
569 int32_t *data, int n, int pred_order,
570 int bps, int precision)
573 pmin = get_max_p_order(pmin, n, pred_order);
574 pmax = get_max_p_order(pmax, n, pred_order);
575 bits = pred_order*bps + 4 + 5 + pred_order*precision + 6;
576 bits += calc_rice_params(rc, pmin, pmax, data, n, pred_order);
581 * Apply Welch window function to audio block
583 static void apply_welch_window(const int32_t *data, int len, double *w_data)
589 assert(!(len&1)); //the optimization in r11881 does not support odd len
590 //if someone wants odd len extend the change in r11881
593 c = 2.0 / (len - 1.0);
597 for(i=0; i<n2; i++) {
600 w_data[-i-1] = data[-i-1] * w;
601 w_data[+i ] = data[+i ] * w;
606 * Calculates autocorrelation data from audio samples
607 * A Welch window function is applied before calculation.
609 void ff_flac_compute_autocorr(const int32_t *data, int len, int lag,
613 double tmp[len + lag + 1];
614 double *data1= tmp + lag;
616 apply_welch_window(data, len, data1);
622 for(j=0; j<lag; j+=2){
623 double sum0 = 1.0, sum1 = 1.0;
624 for(i=0; i<len; i++){
625 sum0 += data1[i] * data1[i-j];
626 sum1 += data1[i] * data1[i-j-1];
634 for(i=0; i<len; i+=2){
635 sum += data1[i ] * data1[i-j ]
636 + data1[i+1] * data1[i-j+1];
643 static void encode_residual_verbatim(int32_t *res, int32_t *smp, int n)
646 memcpy(res, smp, n * sizeof(int32_t));
649 static void encode_residual_fixed(int32_t *res, const int32_t *smp, int n,
654 for(i=0; i<order; i++) {
659 for(i=order; i<n; i++)
662 for(i=order; i<n; i++)
663 res[i]= smp[i] - smp[i-1];
665 int a = smp[order-1] - smp[order-2];
666 for(i=order; i<n; i+=2) {
667 int b = smp[i] - smp[i-1];
669 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 for(i=order; i<n; i+=2) {
676 int b = smp[i] - smp[i-1];
679 a = smp[i+1] - smp[i];
684 int a = smp[order-1] - smp[order-2];
685 int c = smp[order-1] - 2*smp[order-2] + smp[order-3];
686 int e = smp[order-1] - 3*smp[order-2] + 3*smp[order-3] - smp[order-4];
687 for(i=order; i<n; i+=2) {
688 int b = smp[i] - smp[i-1];
692 a = smp[i+1] - smp[i];
701 int c = coefs[(x)-1];\
707 static av_always_inline void encode_residual_lpc_unrolled(
708 int32_t *res, const int32_t *smp, int n,
709 int order, const int32_t *coefs, int shift, int big)
712 for(i=order; i<n; i+=2) {
713 int s = smp[i-order];
762 res[i ] = smp[i ] - (p0 >> shift);
763 res[i+1] = smp[i+1] - (p1 >> shift);
767 static void encode_residual_lpc(int32_t *res, const int32_t *smp, int n,
768 int order, const int32_t *coefs, int shift)
771 for(i=0; i<order; i++) {
775 for(i=order; i<n; i+=2) {
779 for(j=0; j<order; j++) {
785 res[i ] = smp[i ] - (p0 >> shift);
786 res[i+1] = smp[i+1] - (p1 >> shift);
790 case 1: encode_residual_lpc_unrolled(res, smp, n, 1, coefs, shift, 0); break;
791 case 2: encode_residual_lpc_unrolled(res, smp, n, 2, coefs, shift, 0); break;
792 case 3: encode_residual_lpc_unrolled(res, smp, n, 3, coefs, shift, 0); break;
793 case 4: encode_residual_lpc_unrolled(res, smp, n, 4, coefs, shift, 0); break;
794 case 5: encode_residual_lpc_unrolled(res, smp, n, 5, coefs, shift, 0); break;
795 case 6: encode_residual_lpc_unrolled(res, smp, n, 6, coefs, shift, 0); break;
796 case 7: encode_residual_lpc_unrolled(res, smp, n, 7, coefs, shift, 0); break;
797 case 8: encode_residual_lpc_unrolled(res, smp, n, 8, coefs, shift, 0); break;
798 default: encode_residual_lpc_unrolled(res, smp, n, order, coefs, shift, 1); break;
803 static int encode_residual(FlacEncodeContext *ctx, int ch)
806 int min_order, max_order, opt_order, precision, omethod;
807 int min_porder, max_porder;
810 int32_t coefs[MAX_LPC_ORDER][MAX_LPC_ORDER];
811 int shift[MAX_LPC_ORDER];
815 sub = &frame->subframes[ch];
818 n = frame->blocksize;
822 if(smp[i] != smp[0]) break;
825 sub->type = sub->type_code = FLAC_SUBFRAME_CONSTANT;
832 sub->type = sub->type_code = FLAC_SUBFRAME_VERBATIM;
833 encode_residual_verbatim(res, smp, n);
834 return sub->obits * n;
837 min_order = ctx->options.min_prediction_order;
838 max_order = ctx->options.max_prediction_order;
839 min_porder = ctx->options.min_partition_order;
840 max_porder = ctx->options.max_partition_order;
841 precision = ctx->options.lpc_coeff_precision;
842 omethod = ctx->options.prediction_order_method;
845 if(!ctx->options.use_lpc || max_order == 0 || (n <= max_order)) {
846 uint32_t bits[MAX_FIXED_ORDER+1];
847 if(max_order > MAX_FIXED_ORDER) max_order = MAX_FIXED_ORDER;
849 bits[0] = UINT32_MAX;
850 for(i=min_order; i<=max_order; i++) {
851 encode_residual_fixed(res, smp, n, i);
852 bits[i] = calc_rice_params_fixed(&sub->rc, min_porder, max_porder, res,
854 if(bits[i] < bits[opt_order]) {
858 sub->order = opt_order;
859 sub->type = FLAC_SUBFRAME_FIXED;
860 sub->type_code = sub->type | sub->order;
861 if(sub->order != max_order) {
862 encode_residual_fixed(res, smp, n, sub->order);
863 return calc_rice_params_fixed(&sub->rc, min_porder, max_porder, res, n,
864 sub->order, sub->obits);
866 return bits[sub->order];
870 opt_order = ff_lpc_calc_coefs(&ctx->dsp, smp, n, min_order, max_order,
871 precision, coefs, shift, ctx->options.use_lpc,
872 omethod, MAX_LPC_SHIFT, 0);
874 if(omethod == ORDER_METHOD_2LEVEL ||
875 omethod == ORDER_METHOD_4LEVEL ||
876 omethod == ORDER_METHOD_8LEVEL) {
877 int levels = 1 << omethod;
878 uint32_t bits[levels];
880 int opt_index = levels-1;
881 opt_order = max_order-1;
882 bits[opt_index] = UINT32_MAX;
883 for(i=levels-1; i>=0; i--) {
884 order = min_order + (((max_order-min_order+1) * (i+1)) / levels)-1;
885 if(order < 0) order = 0;
886 encode_residual_lpc(res, smp, n, order+1, coefs[order], shift[order]);
887 bits[i] = calc_rice_params_lpc(&sub->rc, min_porder, max_porder,
888 res, n, order+1, sub->obits, precision);
889 if(bits[i] < bits[opt_index]) {
895 } else if(omethod == ORDER_METHOD_SEARCH) {
896 // brute-force optimal order search
897 uint32_t bits[MAX_LPC_ORDER];
899 bits[0] = UINT32_MAX;
900 for(i=min_order-1; i<max_order; i++) {
901 encode_residual_lpc(res, smp, n, i+1, coefs[i], shift[i]);
902 bits[i] = calc_rice_params_lpc(&sub->rc, min_porder, max_porder,
903 res, n, i+1, sub->obits, precision);
904 if(bits[i] < bits[opt_order]) {
909 } else if(omethod == ORDER_METHOD_LOG) {
910 uint32_t bits[MAX_LPC_ORDER];
913 opt_order= min_order - 1 + (max_order-min_order)/3;
914 memset(bits, -1, sizeof(bits));
916 for(step=16 ;step; step>>=1){
918 for(i=last-step; i<=last+step; i+= step){
919 if(i<min_order-1 || i>=max_order || bits[i] < UINT32_MAX)
921 encode_residual_lpc(res, smp, n, i+1, coefs[i], shift[i]);
922 bits[i] = calc_rice_params_lpc(&sub->rc, min_porder, max_porder,
923 res, n, i+1, sub->obits, precision);
924 if(bits[i] < bits[opt_order])
931 sub->order = opt_order;
932 sub->type = FLAC_SUBFRAME_LPC;
933 sub->type_code = sub->type | (sub->order-1);
934 sub->shift = shift[sub->order-1];
935 for(i=0; i<sub->order; i++) {
936 sub->coefs[i] = coefs[sub->order-1][i];
938 encode_residual_lpc(res, smp, n, sub->order, sub->coefs, sub->shift);
939 return calc_rice_params_lpc(&sub->rc, min_porder, max_porder, res, n, sub->order,
940 sub->obits, precision);
943 static int encode_residual_v(FlacEncodeContext *ctx, int ch)
951 sub = &frame->subframes[ch];
954 n = frame->blocksize;
958 if(smp[i] != smp[0]) break;
961 sub->type = sub->type_code = FLAC_SUBFRAME_CONSTANT;
967 sub->type = sub->type_code = FLAC_SUBFRAME_VERBATIM;
968 encode_residual_verbatim(res, smp, n);
969 return sub->obits * n;
972 static int estimate_stereo_mode(int32_t *left_ch, int32_t *right_ch, int n)
980 /* calculate sum of 2nd order residual for each channel */
981 sum[0] = sum[1] = sum[2] = sum[3] = 0;
983 lt = left_ch[i] - 2*left_ch[i-1] + left_ch[i-2];
984 rt = right_ch[i] - 2*right_ch[i-1] + right_ch[i-2];
985 sum[2] += FFABS((lt + rt) >> 1);
986 sum[3] += FFABS(lt - rt);
990 /* estimate bit counts */
992 k = find_optimal_param(2*sum[i], n);
993 sum[i] = rice_encode_count(2*sum[i], n, k);
996 /* calculate score for each mode */
997 score[0] = sum[0] + sum[1];
998 score[1] = sum[0] + sum[3];
999 score[2] = sum[1] + sum[3];
1000 score[3] = sum[2] + sum[3];
1002 /* return mode with lowest score */
1004 for(i=1; i<4; i++) {
1005 if(score[i] < score[best]) {
1010 return FLAC_CHMODE_LEFT_RIGHT;
1011 } else if(best == 1) {
1012 return FLAC_CHMODE_LEFT_SIDE;
1013 } else if(best == 2) {
1014 return FLAC_CHMODE_RIGHT_SIDE;
1016 return FLAC_CHMODE_MID_SIDE;
1021 * Perform stereo channel decorrelation
1023 static void channel_decorrelation(FlacEncodeContext *ctx)
1026 int32_t *left, *right;
1029 frame = &ctx->frame;
1030 n = frame->blocksize;
1031 left = frame->subframes[0].samples;
1032 right = frame->subframes[1].samples;
1034 if(ctx->channels != 2) {
1035 frame->ch_mode = FLAC_CHMODE_NOT_STEREO;
1039 frame->ch_mode = estimate_stereo_mode(left, right, n);
1041 /* perform decorrelation and adjust bits-per-sample */
1042 if(frame->ch_mode == FLAC_CHMODE_LEFT_RIGHT) {
1045 if(frame->ch_mode == FLAC_CHMODE_MID_SIDE) {
1047 for(i=0; i<n; i++) {
1049 left[i] = (tmp + right[i]) >> 1;
1050 right[i] = tmp - right[i];
1052 frame->subframes[1].obits++;
1053 } else if(frame->ch_mode == FLAC_CHMODE_LEFT_SIDE) {
1054 for(i=0; i<n; i++) {
1055 right[i] = left[i] - right[i];
1057 frame->subframes[1].obits++;
1059 for(i=0; i<n; i++) {
1060 left[i] -= right[i];
1062 frame->subframes[0].obits++;
1066 static void write_utf8(PutBitContext *pb, uint32_t val)
1069 PUT_UTF8(val, tmp, put_bits(pb, 8, tmp);)
1072 static void output_frame_header(FlacEncodeContext *s)
1079 put_bits(&s->pb, 16, 0xFFF8);
1080 put_bits(&s->pb, 4, frame->bs_code[0]);
1081 put_bits(&s->pb, 4, s->sr_code[0]);
1082 if(frame->ch_mode == FLAC_CHMODE_NOT_STEREO) {
1083 put_bits(&s->pb, 4, s->ch_code);
1085 put_bits(&s->pb, 4, frame->ch_mode);
1087 put_bits(&s->pb, 3, 4); /* bits-per-sample code */
1088 put_bits(&s->pb, 1, 0);
1089 write_utf8(&s->pb, s->frame_count);
1090 if(frame->bs_code[0] == 6) {
1091 put_bits(&s->pb, 8, frame->bs_code[1]);
1092 } else if(frame->bs_code[0] == 7) {
1093 put_bits(&s->pb, 16, frame->bs_code[1]);
1095 if(s->sr_code[0] == 12) {
1096 put_bits(&s->pb, 8, s->sr_code[1]);
1097 } else if(s->sr_code[0] > 12) {
1098 put_bits(&s->pb, 16, s->sr_code[1]);
1100 flush_put_bits(&s->pb);
1101 crc = av_crc(av_crc_get_table(AV_CRC_8_ATM), 0,
1102 s->pb.buf, put_bits_count(&s->pb)>>3);
1103 put_bits(&s->pb, 8, crc);
1106 static void output_subframe_constant(FlacEncodeContext *s, int ch)
1111 sub = &s->frame.subframes[ch];
1112 res = sub->residual[0];
1113 put_sbits(&s->pb, sub->obits, res);
1116 static void output_subframe_verbatim(FlacEncodeContext *s, int ch)
1124 sub = &frame->subframes[ch];
1126 for(i=0; i<frame->blocksize; i++) {
1127 res = sub->residual[i];
1128 put_sbits(&s->pb, sub->obits, res);
1132 static void output_residual(FlacEncodeContext *ctx, int ch)
1134 int i, j, p, n, parts;
1135 int k, porder, psize, res_cnt;
1140 frame = &ctx->frame;
1141 sub = &frame->subframes[ch];
1142 res = sub->residual;
1143 n = frame->blocksize;
1145 /* rice-encoded block */
1146 put_bits(&ctx->pb, 2, 0);
1148 /* partition order */
1149 porder = sub->rc.porder;
1150 psize = n >> porder;
1151 parts = (1 << porder);
1152 put_bits(&ctx->pb, 4, porder);
1153 res_cnt = psize - sub->order;
1157 for(p=0; p<parts; p++) {
1158 k = sub->rc.params[p];
1159 put_bits(&ctx->pb, 4, k);
1160 if(p == 1) res_cnt = psize;
1161 for(i=0; i<res_cnt && j<n; i++, j++) {
1162 set_sr_golomb_flac(&ctx->pb, res[j], k, INT32_MAX, 0);
1167 static void output_subframe_fixed(FlacEncodeContext *ctx, int ch)
1173 frame = &ctx->frame;
1174 sub = &frame->subframes[ch];
1176 /* warm-up samples */
1177 for(i=0; i<sub->order; i++) {
1178 put_sbits(&ctx->pb, sub->obits, sub->residual[i]);
1182 output_residual(ctx, ch);
1185 static void output_subframe_lpc(FlacEncodeContext *ctx, int ch)
1191 frame = &ctx->frame;
1192 sub = &frame->subframes[ch];
1194 /* warm-up samples */
1195 for(i=0; i<sub->order; i++) {
1196 put_sbits(&ctx->pb, sub->obits, sub->residual[i]);
1199 /* LPC coefficients */
1200 cbits = ctx->options.lpc_coeff_precision;
1201 put_bits(&ctx->pb, 4, cbits-1);
1202 put_sbits(&ctx->pb, 5, sub->shift);
1203 for(i=0; i<sub->order; i++) {
1204 put_sbits(&ctx->pb, cbits, sub->coefs[i]);
1208 output_residual(ctx, ch);
1211 static void output_subframes(FlacEncodeContext *s)
1219 for(ch=0; ch<s->channels; ch++) {
1220 sub = &frame->subframes[ch];
1222 /* subframe header */
1223 put_bits(&s->pb, 1, 0);
1224 put_bits(&s->pb, 6, sub->type_code);
1225 put_bits(&s->pb, 1, 0); /* no wasted bits */
1228 if(sub->type == FLAC_SUBFRAME_CONSTANT) {
1229 output_subframe_constant(s, ch);
1230 } else if(sub->type == FLAC_SUBFRAME_VERBATIM) {
1231 output_subframe_verbatim(s, ch);
1232 } else if(sub->type == FLAC_SUBFRAME_FIXED) {
1233 output_subframe_fixed(s, ch);
1234 } else if(sub->type == FLAC_SUBFRAME_LPC) {
1235 output_subframe_lpc(s, ch);
1240 static void output_frame_footer(FlacEncodeContext *s)
1243 flush_put_bits(&s->pb);
1244 crc = bswap_16(av_crc(av_crc_get_table(AV_CRC_16_ANSI), 0,
1245 s->pb.buf, put_bits_count(&s->pb)>>3));
1246 put_bits(&s->pb, 16, crc);
1247 flush_put_bits(&s->pb);
1250 static void update_md5_sum(FlacEncodeContext *s, int16_t *samples)
1252 #ifdef WORDS_BIGENDIAN
1254 for(i = 0; i < s->frame.blocksize*s->channels; i++) {
1255 int16_t smp = le2me_16(samples[i]);
1256 av_md5_update(s->md5ctx, (uint8_t *)&smp, 2);
1259 av_md5_update(s->md5ctx, (uint8_t *)samples, s->frame.blocksize*s->channels*2);
1263 static int flac_encode_frame(AVCodecContext *avctx, uint8_t *frame,
1264 int buf_size, void *data)
1267 FlacEncodeContext *s;
1268 int16_t *samples = data;
1272 s = avctx->priv_data;
1274 if(buf_size < s->max_framesize*2) {
1275 av_log(avctx, AV_LOG_ERROR, "output buffer too small\n");
1279 /* when the last block is reached, update the header in extradata */
1281 av_md5_final(s->md5ctx, s->md5sum);
1282 write_streaminfo(s, avctx->extradata);
1288 copy_samples(s, samples);
1290 channel_decorrelation(s);
1292 for(ch=0; ch<s->channels; ch++) {
1293 encode_residual(s, ch);
1297 init_put_bits(&s->pb, frame, buf_size);
1298 output_frame_header(s);
1299 output_subframes(s);
1300 output_frame_footer(s);
1301 out_bytes = put_bits_count(&s->pb) >> 3;
1303 if(out_bytes > s->max_framesize) {
1305 /* still too large. must be an error. */
1306 av_log(avctx, AV_LOG_ERROR, "error encoding frame\n");
1310 /* frame too large. use verbatim mode */
1311 for(ch=0; ch<s->channels; ch++) {
1312 encode_residual_v(s, ch);
1319 s->sample_count += avctx->frame_size;
1320 update_md5_sum(s, samples);
1325 static av_cold int flac_encode_close(AVCodecContext *avctx)
1327 if (avctx->priv_data) {
1328 FlacEncodeContext *s = avctx->priv_data;
1329 av_freep(&s->md5ctx);
1331 av_freep(&avctx->extradata);
1332 avctx->extradata_size = 0;
1333 av_freep(&avctx->coded_frame);
1337 AVCodec flac_encoder = {
1341 sizeof(FlacEncodeContext),
1346 .capabilities = CODEC_CAP_SMALL_LAST_FRAME | CODEC_CAP_DELAY,
1347 .sample_fmts = (enum SampleFormat[]){SAMPLE_FMT_S16,SAMPLE_FMT_NONE},
1348 .long_name = NULL_IF_CONFIG_SMALL("FLAC (Free Lossless Audio Codec)"),