3 * Copyright (c) 2006 Justin Ruggles <justin.ruggles@gmail.com>
5 * This file is part of FFmpeg.
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
22 #include "libavutil/crc.h"
23 #include "libavutil/md5.h"
32 #define FLAC_SUBFRAME_CONSTANT 0
33 #define FLAC_SUBFRAME_VERBATIM 1
34 #define FLAC_SUBFRAME_FIXED 8
35 #define FLAC_SUBFRAME_LPC 32
37 #define MAX_FIXED_ORDER 4
38 #define MAX_PARTITION_ORDER 8
39 #define MAX_PARTITIONS (1 << MAX_PARTITION_ORDER)
40 #define MAX_LPC_PRECISION 15
41 #define MAX_LPC_SHIFT 15
42 #define MAX_RICE_PARAM 14
44 typedef struct CompressionOptions {
45 int compression_level;
48 int lpc_coeff_precision;
49 int min_prediction_order;
50 int max_prediction_order;
51 int prediction_order_method;
52 int min_partition_order;
53 int max_partition_order;
56 typedef struct RiceContext {
58 int params[MAX_PARTITIONS];
61 typedef struct FlacSubframe {
66 int32_t coefs[MAX_LPC_ORDER];
69 int32_t samples[FLAC_MAX_BLOCKSIZE];
70 int32_t residual[FLAC_MAX_BLOCKSIZE+1];
73 typedef struct FlacFrame {
74 FlacSubframe subframes[FLAC_MAX_CHANNELS];
81 typedef struct FlacEncodeContext {
88 int max_encoded_framesize;
90 uint64_t sample_count;
93 CompressionOptions options;
94 AVCodecContext *avctx;
100 * Writes streaminfo metadata block to byte array
102 static void write_streaminfo(FlacEncodeContext *s, uint8_t *header)
106 memset(header, 0, FLAC_STREAMINFO_SIZE);
107 init_put_bits(&pb, header, FLAC_STREAMINFO_SIZE);
109 /* streaminfo metadata block */
110 put_bits(&pb, 16, s->avctx->frame_size);
111 put_bits(&pb, 16, s->avctx->frame_size);
112 put_bits(&pb, 24, s->min_framesize);
113 put_bits(&pb, 24, s->max_framesize);
114 put_bits(&pb, 20, s->samplerate);
115 put_bits(&pb, 3, s->channels-1);
116 put_bits(&pb, 5, 15); /* bits per sample - 1 */
117 /* write 36-bit sample count in 2 put_bits() calls */
118 put_bits(&pb, 24, (s->sample_count & 0xFFFFFF000LL) >> 12);
119 put_bits(&pb, 12, s->sample_count & 0x000000FFFLL);
121 memcpy(&header[18], s->md5sum, 16);
125 * Sets blocksize based on samplerate
126 * Chooses the closest predefined blocksize >= BLOCK_TIME_MS milliseconds
128 static int select_blocksize(int samplerate, int block_time_ms)
134 assert(samplerate > 0);
135 blocksize = ff_flac_blocksize_table[1];
136 target = (samplerate * block_time_ms) / 1000;
137 for(i=0; i<16; i++) {
138 if(target >= ff_flac_blocksize_table[i] && ff_flac_blocksize_table[i] > blocksize) {
139 blocksize = ff_flac_blocksize_table[i];
145 static av_cold int flac_encode_init(AVCodecContext *avctx)
147 int freq = avctx->sample_rate;
148 int channels = avctx->channels;
149 FlacEncodeContext *s = avctx->priv_data;
155 dsputil_init(&s->dsp, avctx);
157 if(avctx->sample_fmt != SAMPLE_FMT_S16) {
161 if(channels < 1 || channels > FLAC_MAX_CHANNELS) {
164 s->channels = channels;
166 /* find samplerate in table */
169 for(i=4; i<12; i++) {
170 if(freq == ff_flac_sample_rate_table[i]) {
171 s->samplerate = ff_flac_sample_rate_table[i];
177 /* if not in table, samplerate is non-standard */
179 if(freq % 1000 == 0 && freq < 255000) {
181 s->sr_code[1] = freq / 1000;
182 } else if(freq % 10 == 0 && freq < 655350) {
184 s->sr_code[1] = freq / 10;
185 } else if(freq < 65535) {
187 s->sr_code[1] = freq;
191 s->samplerate = freq;
194 /* set compression option defaults based on avctx->compression_level */
195 if(avctx->compression_level < 0) {
196 s->options.compression_level = 5;
198 s->options.compression_level = avctx->compression_level;
200 av_log(avctx, AV_LOG_DEBUG, " compression: %d\n", s->options.compression_level);
202 level= s->options.compression_level;
204 av_log(avctx, AV_LOG_ERROR, "invalid compression level: %d\n",
205 s->options.compression_level);
209 s->options.block_time_ms = ((int[]){ 27, 27, 27,105,105,105,105,105,105,105,105,105,105})[level];
210 s->options.use_lpc = ((int[]){ 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1})[level];
211 s->options.min_prediction_order= ((int[]){ 2, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1})[level];
212 s->options.max_prediction_order= ((int[]){ 3, 4, 4, 6, 8, 8, 8, 8, 12, 12, 12, 32, 32})[level];
213 s->options.prediction_order_method = ((int[]){ ORDER_METHOD_EST, ORDER_METHOD_EST, ORDER_METHOD_EST,
214 ORDER_METHOD_EST, ORDER_METHOD_EST, ORDER_METHOD_EST,
215 ORDER_METHOD_4LEVEL, ORDER_METHOD_LOG, ORDER_METHOD_4LEVEL,
216 ORDER_METHOD_LOG, ORDER_METHOD_SEARCH, ORDER_METHOD_LOG,
217 ORDER_METHOD_SEARCH})[level];
218 s->options.min_partition_order = ((int[]){ 2, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0})[level];
219 s->options.max_partition_order = ((int[]){ 2, 2, 3, 3, 3, 8, 8, 8, 8, 8, 8, 8, 8})[level];
221 /* set compression option overrides from AVCodecContext */
222 if(avctx->use_lpc >= 0) {
223 s->options.use_lpc = av_clip(avctx->use_lpc, 0, 11);
225 if(s->options.use_lpc == 1)
226 av_log(avctx, AV_LOG_DEBUG, " use lpc: Levinson-Durbin recursion with Welch window\n");
227 else if(s->options.use_lpc > 1)
228 av_log(avctx, AV_LOG_DEBUG, " use lpc: Cholesky factorization\n");
230 if(avctx->min_prediction_order >= 0) {
231 if(s->options.use_lpc) {
232 if(avctx->min_prediction_order < MIN_LPC_ORDER ||
233 avctx->min_prediction_order > MAX_LPC_ORDER) {
234 av_log(avctx, AV_LOG_ERROR, "invalid min prediction order: %d\n",
235 avctx->min_prediction_order);
239 if(avctx->min_prediction_order > MAX_FIXED_ORDER) {
240 av_log(avctx, AV_LOG_ERROR, "invalid min prediction order: %d\n",
241 avctx->min_prediction_order);
245 s->options.min_prediction_order = avctx->min_prediction_order;
247 if(avctx->max_prediction_order >= 0) {
248 if(s->options.use_lpc) {
249 if(avctx->max_prediction_order < MIN_LPC_ORDER ||
250 avctx->max_prediction_order > MAX_LPC_ORDER) {
251 av_log(avctx, AV_LOG_ERROR, "invalid max prediction order: %d\n",
252 avctx->max_prediction_order);
256 if(avctx->max_prediction_order > MAX_FIXED_ORDER) {
257 av_log(avctx, AV_LOG_ERROR, "invalid max prediction order: %d\n",
258 avctx->max_prediction_order);
262 s->options.max_prediction_order = avctx->max_prediction_order;
264 if(s->options.max_prediction_order < s->options.min_prediction_order) {
265 av_log(avctx, AV_LOG_ERROR, "invalid prediction orders: min=%d max=%d\n",
266 s->options.min_prediction_order, s->options.max_prediction_order);
269 av_log(avctx, AV_LOG_DEBUG, " prediction order: %d, %d\n",
270 s->options.min_prediction_order, s->options.max_prediction_order);
272 if(avctx->prediction_order_method >= 0) {
273 if(avctx->prediction_order_method > ORDER_METHOD_LOG) {
274 av_log(avctx, AV_LOG_ERROR, "invalid prediction order method: %d\n",
275 avctx->prediction_order_method);
278 s->options.prediction_order_method = avctx->prediction_order_method;
280 switch(s->options.prediction_order_method) {
281 case ORDER_METHOD_EST: av_log(avctx, AV_LOG_DEBUG, " order method: %s\n",
283 case ORDER_METHOD_2LEVEL: av_log(avctx, AV_LOG_DEBUG, " order method: %s\n",
285 case ORDER_METHOD_4LEVEL: av_log(avctx, AV_LOG_DEBUG, " order method: %s\n",
287 case ORDER_METHOD_8LEVEL: av_log(avctx, AV_LOG_DEBUG, " order method: %s\n",
289 case ORDER_METHOD_SEARCH: av_log(avctx, AV_LOG_DEBUG, " order method: %s\n",
290 "full search"); break;
291 case ORDER_METHOD_LOG: av_log(avctx, AV_LOG_DEBUG, " order method: %s\n",
292 "log search"); break;
295 if(avctx->min_partition_order >= 0) {
296 if(avctx->min_partition_order > MAX_PARTITION_ORDER) {
297 av_log(avctx, AV_LOG_ERROR, "invalid min partition order: %d\n",
298 avctx->min_partition_order);
301 s->options.min_partition_order = avctx->min_partition_order;
303 if(avctx->max_partition_order >= 0) {
304 if(avctx->max_partition_order > MAX_PARTITION_ORDER) {
305 av_log(avctx, AV_LOG_ERROR, "invalid max partition order: %d\n",
306 avctx->max_partition_order);
309 s->options.max_partition_order = avctx->max_partition_order;
311 if(s->options.max_partition_order < s->options.min_partition_order) {
312 av_log(avctx, AV_LOG_ERROR, "invalid partition orders: min=%d max=%d\n",
313 s->options.min_partition_order, s->options.max_partition_order);
316 av_log(avctx, AV_LOG_DEBUG, " partition order: %d, %d\n",
317 s->options.min_partition_order, s->options.max_partition_order);
319 if(avctx->frame_size > 0) {
320 if(avctx->frame_size < FLAC_MIN_BLOCKSIZE ||
321 avctx->frame_size > FLAC_MAX_BLOCKSIZE) {
322 av_log(avctx, AV_LOG_ERROR, "invalid block size: %d\n",
327 s->avctx->frame_size = select_blocksize(s->samplerate, s->options.block_time_ms);
329 av_log(avctx, AV_LOG_DEBUG, " block size: %d\n", s->avctx->frame_size);
331 /* set LPC precision */
332 if(avctx->lpc_coeff_precision > 0) {
333 if(avctx->lpc_coeff_precision > MAX_LPC_PRECISION) {
334 av_log(avctx, AV_LOG_ERROR, "invalid lpc coeff precision: %d\n",
335 avctx->lpc_coeff_precision);
338 s->options.lpc_coeff_precision = avctx->lpc_coeff_precision;
340 /* default LPC precision */
341 s->options.lpc_coeff_precision = 15;
343 av_log(avctx, AV_LOG_DEBUG, " lpc precision: %d\n",
344 s->options.lpc_coeff_precision);
346 /* set maximum encoded frame size in verbatim mode */
347 s->max_framesize = ff_flac_get_max_frame_size(s->avctx->frame_size,
350 /* initialize MD5 context */
351 s->md5ctx = av_malloc(av_md5_size);
353 return AVERROR_NOMEM;
354 av_md5_init(s->md5ctx);
356 streaminfo = av_malloc(FLAC_STREAMINFO_SIZE);
357 write_streaminfo(s, streaminfo);
358 avctx->extradata = streaminfo;
359 avctx->extradata_size = FLAC_STREAMINFO_SIZE;
362 s->min_framesize = s->max_framesize;
364 avctx->coded_frame = avcodec_alloc_frame();
365 avctx->coded_frame->key_frame = 1;
370 static void init_frame(FlacEncodeContext *s)
377 for(i=0; i<16; i++) {
378 if(s->avctx->frame_size == ff_flac_blocksize_table[i]) {
379 frame->blocksize = ff_flac_blocksize_table[i];
380 frame->bs_code[0] = i;
381 frame->bs_code[1] = 0;
386 frame->blocksize = s->avctx->frame_size;
387 if(frame->blocksize <= 256) {
388 frame->bs_code[0] = 6;
389 frame->bs_code[1] = frame->blocksize-1;
391 frame->bs_code[0] = 7;
392 frame->bs_code[1] = frame->blocksize-1;
396 for(ch=0; ch<s->channels; ch++) {
397 frame->subframes[ch].obits = 16;
402 * Copy channel-interleaved input samples into separate subframes
404 static void copy_samples(FlacEncodeContext *s, int16_t *samples)
410 for(i=0,j=0; i<frame->blocksize; i++) {
411 for(ch=0; ch<s->channels; ch++,j++) {
412 frame->subframes[ch].samples[i] = samples[j];
418 #define rice_encode_count(sum, n, k) (((n)*((k)+1))+((sum-(n>>1))>>(k)))
421 * Solve for d/dk(rice_encode_count) = n-((sum-(n>>1))>>(k+1)) = 0
423 static int find_optimal_param(uint32_t sum, int n)
431 k = av_log2(n<256 ? FASTDIV(sum2,n) : sum2/n);
432 return FFMIN(k, MAX_RICE_PARAM);
435 static uint32_t calc_optimal_rice_params(RiceContext *rc, int porder,
436 uint32_t *sums, int n, int pred_order)
442 part = (1 << porder);
445 cnt = (n >> porder) - pred_order;
446 for(i=0; i<part; i++) {
447 k = find_optimal_param(sums[i], cnt);
449 all_bits += rice_encode_count(sums[i], cnt, k);
458 static void calc_sums(int pmin, int pmax, uint32_t *data, int n, int pred_order,
459 uint32_t sums[][MAX_PARTITIONS])
463 uint32_t *res, *res_end;
465 /* sums for highest level */
467 res = &data[pred_order];
468 res_end = &data[n >> pmax];
469 for(i=0; i<parts; i++) {
471 while(res < res_end){
477 /* sums for lower levels */
478 for(i=pmax-1; i>=pmin; i--) {
480 for(j=0; j<parts; j++) {
481 sums[i][j] = sums[i+1][2*j] + sums[i+1][2*j+1];
486 static uint32_t calc_rice_params(RiceContext *rc, int pmin, int pmax,
487 int32_t *data, int n, int pred_order)
490 uint32_t bits[MAX_PARTITION_ORDER+1];
494 uint32_t sums[MAX_PARTITION_ORDER+1][MAX_PARTITIONS];
496 assert(pmin >= 0 && pmin <= MAX_PARTITION_ORDER);
497 assert(pmax >= 0 && pmax <= MAX_PARTITION_ORDER);
498 assert(pmin <= pmax);
500 udata = av_malloc(n * sizeof(uint32_t));
502 udata[i] = (2*data[i]) ^ (data[i]>>31);
505 calc_sums(pmin, pmax, udata, n, pred_order, sums);
508 bits[pmin] = UINT32_MAX;
509 for(i=pmin; i<=pmax; i++) {
510 bits[i] = calc_optimal_rice_params(&tmp_rc, i, sums[i], n, pred_order);
511 if(bits[i] <= bits[opt_porder]) {
518 return bits[opt_porder];
521 static int get_max_p_order(int max_porder, int n, int order)
523 int porder = FFMIN(max_porder, av_log2(n^(n-1)));
525 porder = FFMIN(porder, av_log2(n/order));
529 static uint32_t calc_rice_params_fixed(RiceContext *rc, int pmin, int pmax,
530 int32_t *data, int n, int pred_order,
534 pmin = get_max_p_order(pmin, n, pred_order);
535 pmax = get_max_p_order(pmax, n, pred_order);
536 bits = pred_order*bps + 6;
537 bits += calc_rice_params(rc, pmin, pmax, data, n, pred_order);
541 static uint32_t calc_rice_params_lpc(RiceContext *rc, int pmin, int pmax,
542 int32_t *data, int n, int pred_order,
543 int bps, int precision)
546 pmin = get_max_p_order(pmin, n, pred_order);
547 pmax = get_max_p_order(pmax, n, pred_order);
548 bits = pred_order*bps + 4 + 5 + pred_order*precision + 6;
549 bits += calc_rice_params(rc, pmin, pmax, data, n, pred_order);
554 * Apply Welch window function to audio block
556 static void apply_welch_window(const int32_t *data, int len, double *w_data)
562 assert(!(len&1)); //the optimization in r11881 does not support odd len
563 //if someone wants odd len extend the change in r11881
566 c = 2.0 / (len - 1.0);
570 for(i=0; i<n2; i++) {
573 w_data[-i-1] = data[-i-1] * w;
574 w_data[+i ] = data[+i ] * w;
579 * Calculates autocorrelation data from audio samples
580 * A Welch window function is applied before calculation.
582 void ff_flac_compute_autocorr(const int32_t *data, int len, int lag,
586 double tmp[len + lag + 1];
587 double *data1= tmp + lag;
589 apply_welch_window(data, len, data1);
595 for(j=0; j<lag; j+=2){
596 double sum0 = 1.0, sum1 = 1.0;
597 for(i=0; i<len; i++){
598 sum0 += data1[i] * data1[i-j];
599 sum1 += data1[i] * data1[i-j-1];
607 for(i=0; i<len; i+=2){
608 sum += data1[i ] * data1[i-j ]
609 + data1[i+1] * data1[i-j+1];
616 static void encode_residual_verbatim(int32_t *res, int32_t *smp, int n)
619 memcpy(res, smp, n * sizeof(int32_t));
622 static void encode_residual_fixed(int32_t *res, const int32_t *smp, int n,
627 for(i=0; i<order; i++) {
632 for(i=order; i<n; i++)
635 for(i=order; i<n; i++)
636 res[i]= smp[i] - smp[i-1];
638 int a = smp[order-1] - smp[order-2];
639 for(i=order; i<n; i+=2) {
640 int b = smp[i] - smp[i-1];
642 a = smp[i+1] - smp[i];
646 int a = smp[order-1] - smp[order-2];
647 int c = smp[order-1] - 2*smp[order-2] + smp[order-3];
648 for(i=order; i<n; i+=2) {
649 int b = smp[i] - smp[i-1];
652 a = smp[i+1] - smp[i];
657 int a = smp[order-1] - smp[order-2];
658 int c = smp[order-1] - 2*smp[order-2] + smp[order-3];
659 int e = smp[order-1] - 3*smp[order-2] + 3*smp[order-3] - smp[order-4];
660 for(i=order; i<n; i+=2) {
661 int b = smp[i] - smp[i-1];
665 a = smp[i+1] - smp[i];
674 int c = coefs[(x)-1];\
680 static av_always_inline void encode_residual_lpc_unrolled(
681 int32_t *res, const int32_t *smp, int n,
682 int order, const int32_t *coefs, int shift, int big)
685 for(i=order; i<n; i+=2) {
686 int s = smp[i-order];
735 res[i ] = smp[i ] - (p0 >> shift);
736 res[i+1] = smp[i+1] - (p1 >> shift);
740 static void encode_residual_lpc(int32_t *res, const int32_t *smp, int n,
741 int order, const int32_t *coefs, int shift)
744 for(i=0; i<order; i++) {
748 for(i=order; i<n; i+=2) {
752 for(j=0; j<order; j++) {
758 res[i ] = smp[i ] - (p0 >> shift);
759 res[i+1] = smp[i+1] - (p1 >> shift);
763 case 1: encode_residual_lpc_unrolled(res, smp, n, 1, coefs, shift, 0); break;
764 case 2: encode_residual_lpc_unrolled(res, smp, n, 2, coefs, shift, 0); break;
765 case 3: encode_residual_lpc_unrolled(res, smp, n, 3, coefs, shift, 0); break;
766 case 4: encode_residual_lpc_unrolled(res, smp, n, 4, coefs, shift, 0); break;
767 case 5: encode_residual_lpc_unrolled(res, smp, n, 5, coefs, shift, 0); break;
768 case 6: encode_residual_lpc_unrolled(res, smp, n, 6, coefs, shift, 0); break;
769 case 7: encode_residual_lpc_unrolled(res, smp, n, 7, coefs, shift, 0); break;
770 case 8: encode_residual_lpc_unrolled(res, smp, n, 8, coefs, shift, 0); break;
771 default: encode_residual_lpc_unrolled(res, smp, n, order, coefs, shift, 1); break;
776 static int encode_residual(FlacEncodeContext *ctx, int ch)
779 int min_order, max_order, opt_order, precision, omethod;
780 int min_porder, max_porder;
783 int32_t coefs[MAX_LPC_ORDER][MAX_LPC_ORDER];
784 int shift[MAX_LPC_ORDER];
788 sub = &frame->subframes[ch];
791 n = frame->blocksize;
795 if(smp[i] != smp[0]) break;
798 sub->type = sub->type_code = FLAC_SUBFRAME_CONSTANT;
805 sub->type = sub->type_code = FLAC_SUBFRAME_VERBATIM;
806 encode_residual_verbatim(res, smp, n);
807 return sub->obits * n;
810 min_order = ctx->options.min_prediction_order;
811 max_order = ctx->options.max_prediction_order;
812 min_porder = ctx->options.min_partition_order;
813 max_porder = ctx->options.max_partition_order;
814 precision = ctx->options.lpc_coeff_precision;
815 omethod = ctx->options.prediction_order_method;
818 if(!ctx->options.use_lpc || max_order == 0 || (n <= max_order)) {
819 uint32_t bits[MAX_FIXED_ORDER+1];
820 if(max_order > MAX_FIXED_ORDER) max_order = MAX_FIXED_ORDER;
822 bits[0] = UINT32_MAX;
823 for(i=min_order; i<=max_order; i++) {
824 encode_residual_fixed(res, smp, n, i);
825 bits[i] = calc_rice_params_fixed(&sub->rc, min_porder, max_porder, res,
827 if(bits[i] < bits[opt_order]) {
831 sub->order = opt_order;
832 sub->type = FLAC_SUBFRAME_FIXED;
833 sub->type_code = sub->type | sub->order;
834 if(sub->order != max_order) {
835 encode_residual_fixed(res, smp, n, sub->order);
836 return calc_rice_params_fixed(&sub->rc, min_porder, max_porder, res, n,
837 sub->order, sub->obits);
839 return bits[sub->order];
843 opt_order = ff_lpc_calc_coefs(&ctx->dsp, smp, n, min_order, max_order,
844 precision, coefs, shift, ctx->options.use_lpc,
845 omethod, MAX_LPC_SHIFT, 0);
847 if(omethod == ORDER_METHOD_2LEVEL ||
848 omethod == ORDER_METHOD_4LEVEL ||
849 omethod == ORDER_METHOD_8LEVEL) {
850 int levels = 1 << omethod;
851 uint32_t bits[levels];
853 int opt_index = levels-1;
854 opt_order = max_order-1;
855 bits[opt_index] = UINT32_MAX;
856 for(i=levels-1; i>=0; i--) {
857 order = min_order + (((max_order-min_order+1) * (i+1)) / levels)-1;
858 if(order < 0) order = 0;
859 encode_residual_lpc(res, smp, n, order+1, coefs[order], shift[order]);
860 bits[i] = calc_rice_params_lpc(&sub->rc, min_porder, max_porder,
861 res, n, order+1, sub->obits, precision);
862 if(bits[i] < bits[opt_index]) {
868 } else if(omethod == ORDER_METHOD_SEARCH) {
869 // brute-force optimal order search
870 uint32_t bits[MAX_LPC_ORDER];
872 bits[0] = UINT32_MAX;
873 for(i=min_order-1; i<max_order; i++) {
874 encode_residual_lpc(res, smp, n, i+1, coefs[i], shift[i]);
875 bits[i] = calc_rice_params_lpc(&sub->rc, min_porder, max_porder,
876 res, n, i+1, sub->obits, precision);
877 if(bits[i] < bits[opt_order]) {
882 } else if(omethod == ORDER_METHOD_LOG) {
883 uint32_t bits[MAX_LPC_ORDER];
886 opt_order= min_order - 1 + (max_order-min_order)/3;
887 memset(bits, -1, sizeof(bits));
889 for(step=16 ;step; step>>=1){
891 for(i=last-step; i<=last+step; i+= step){
892 if(i<min_order-1 || i>=max_order || bits[i] < UINT32_MAX)
894 encode_residual_lpc(res, smp, n, i+1, coefs[i], shift[i]);
895 bits[i] = calc_rice_params_lpc(&sub->rc, min_porder, max_porder,
896 res, n, i+1, sub->obits, precision);
897 if(bits[i] < bits[opt_order])
904 sub->order = opt_order;
905 sub->type = FLAC_SUBFRAME_LPC;
906 sub->type_code = sub->type | (sub->order-1);
907 sub->shift = shift[sub->order-1];
908 for(i=0; i<sub->order; i++) {
909 sub->coefs[i] = coefs[sub->order-1][i];
911 encode_residual_lpc(res, smp, n, sub->order, sub->coefs, sub->shift);
912 return calc_rice_params_lpc(&sub->rc, min_porder, max_porder, res, n, sub->order,
913 sub->obits, precision);
916 static int encode_residual_v(FlacEncodeContext *ctx, int ch)
924 sub = &frame->subframes[ch];
927 n = frame->blocksize;
931 if(smp[i] != smp[0]) break;
934 sub->type = sub->type_code = FLAC_SUBFRAME_CONSTANT;
940 sub->type = sub->type_code = FLAC_SUBFRAME_VERBATIM;
941 encode_residual_verbatim(res, smp, n);
942 return sub->obits * n;
945 static int estimate_stereo_mode(int32_t *left_ch, int32_t *right_ch, int n)
953 /* calculate sum of 2nd order residual for each channel */
954 sum[0] = sum[1] = sum[2] = sum[3] = 0;
956 lt = left_ch[i] - 2*left_ch[i-1] + left_ch[i-2];
957 rt = right_ch[i] - 2*right_ch[i-1] + right_ch[i-2];
958 sum[2] += FFABS((lt + rt) >> 1);
959 sum[3] += FFABS(lt - rt);
963 /* estimate bit counts */
965 k = find_optimal_param(2*sum[i], n);
966 sum[i] = rice_encode_count(2*sum[i], n, k);
969 /* calculate score for each mode */
970 score[0] = sum[0] + sum[1];
971 score[1] = sum[0] + sum[3];
972 score[2] = sum[1] + sum[3];
973 score[3] = sum[2] + sum[3];
975 /* return mode with lowest score */
978 if(score[i] < score[best]) {
983 return FLAC_CHMODE_INDEPENDENT;
984 } else if(best == 1) {
985 return FLAC_CHMODE_LEFT_SIDE;
986 } else if(best == 2) {
987 return FLAC_CHMODE_RIGHT_SIDE;
989 return FLAC_CHMODE_MID_SIDE;
994 * Perform stereo channel decorrelation
996 static void channel_decorrelation(FlacEncodeContext *ctx)
999 int32_t *left, *right;
1002 frame = &ctx->frame;
1003 n = frame->blocksize;
1004 left = frame->subframes[0].samples;
1005 right = frame->subframes[1].samples;
1007 if(ctx->channels != 2) {
1008 frame->ch_mode = FLAC_CHMODE_INDEPENDENT;
1012 frame->ch_mode = estimate_stereo_mode(left, right, n);
1014 /* perform decorrelation and adjust bits-per-sample */
1015 if(frame->ch_mode == FLAC_CHMODE_INDEPENDENT) {
1018 if(frame->ch_mode == FLAC_CHMODE_MID_SIDE) {
1020 for(i=0; i<n; i++) {
1022 left[i] = (tmp + right[i]) >> 1;
1023 right[i] = tmp - right[i];
1025 frame->subframes[1].obits++;
1026 } else if(frame->ch_mode == FLAC_CHMODE_LEFT_SIDE) {
1027 for(i=0; i<n; i++) {
1028 right[i] = left[i] - right[i];
1030 frame->subframes[1].obits++;
1032 for(i=0; i<n; i++) {
1033 left[i] -= right[i];
1035 frame->subframes[0].obits++;
1039 static void write_utf8(PutBitContext *pb, uint32_t val)
1042 PUT_UTF8(val, tmp, put_bits(pb, 8, tmp);)
1045 static void output_frame_header(FlacEncodeContext *s)
1052 put_bits(&s->pb, 16, 0xFFF8);
1053 put_bits(&s->pb, 4, frame->bs_code[0]);
1054 put_bits(&s->pb, 4, s->sr_code[0]);
1055 if(frame->ch_mode == FLAC_CHMODE_INDEPENDENT) {
1056 put_bits(&s->pb, 4, s->channels-1);
1058 put_bits(&s->pb, 4, frame->ch_mode);
1060 put_bits(&s->pb, 3, 4); /* bits-per-sample code */
1061 put_bits(&s->pb, 1, 0);
1062 write_utf8(&s->pb, s->frame_count);
1063 if(frame->bs_code[0] == 6) {
1064 put_bits(&s->pb, 8, frame->bs_code[1]);
1065 } else if(frame->bs_code[0] == 7) {
1066 put_bits(&s->pb, 16, frame->bs_code[1]);
1068 if(s->sr_code[0] == 12) {
1069 put_bits(&s->pb, 8, s->sr_code[1]);
1070 } else if(s->sr_code[0] > 12) {
1071 put_bits(&s->pb, 16, s->sr_code[1]);
1073 flush_put_bits(&s->pb);
1074 crc = av_crc(av_crc_get_table(AV_CRC_8_ATM), 0,
1075 s->pb.buf, put_bits_count(&s->pb)>>3);
1076 put_bits(&s->pb, 8, crc);
1079 static void output_subframe_constant(FlacEncodeContext *s, int ch)
1084 sub = &s->frame.subframes[ch];
1085 res = sub->residual[0];
1086 put_sbits(&s->pb, sub->obits, res);
1089 static void output_subframe_verbatim(FlacEncodeContext *s, int ch)
1097 sub = &frame->subframes[ch];
1099 for(i=0; i<frame->blocksize; i++) {
1100 res = sub->residual[i];
1101 put_sbits(&s->pb, sub->obits, res);
1105 static void output_residual(FlacEncodeContext *ctx, int ch)
1107 int i, j, p, n, parts;
1108 int k, porder, psize, res_cnt;
1113 frame = &ctx->frame;
1114 sub = &frame->subframes[ch];
1115 res = sub->residual;
1116 n = frame->blocksize;
1118 /* rice-encoded block */
1119 put_bits(&ctx->pb, 2, 0);
1121 /* partition order */
1122 porder = sub->rc.porder;
1123 psize = n >> porder;
1124 parts = (1 << porder);
1125 put_bits(&ctx->pb, 4, porder);
1126 res_cnt = psize - sub->order;
1130 for(p=0; p<parts; p++) {
1131 k = sub->rc.params[p];
1132 put_bits(&ctx->pb, 4, k);
1133 if(p == 1) res_cnt = psize;
1134 for(i=0; i<res_cnt && j<n; i++, j++) {
1135 set_sr_golomb_flac(&ctx->pb, res[j], k, INT32_MAX, 0);
1140 static void output_subframe_fixed(FlacEncodeContext *ctx, int ch)
1146 frame = &ctx->frame;
1147 sub = &frame->subframes[ch];
1149 /* warm-up samples */
1150 for(i=0; i<sub->order; i++) {
1151 put_sbits(&ctx->pb, sub->obits, sub->residual[i]);
1155 output_residual(ctx, ch);
1158 static void output_subframe_lpc(FlacEncodeContext *ctx, int ch)
1164 frame = &ctx->frame;
1165 sub = &frame->subframes[ch];
1167 /* warm-up samples */
1168 for(i=0; i<sub->order; i++) {
1169 put_sbits(&ctx->pb, sub->obits, sub->residual[i]);
1172 /* LPC coefficients */
1173 cbits = ctx->options.lpc_coeff_precision;
1174 put_bits(&ctx->pb, 4, cbits-1);
1175 put_sbits(&ctx->pb, 5, sub->shift);
1176 for(i=0; i<sub->order; i++) {
1177 put_sbits(&ctx->pb, cbits, sub->coefs[i]);
1181 output_residual(ctx, ch);
1184 static void output_subframes(FlacEncodeContext *s)
1192 for(ch=0; ch<s->channels; ch++) {
1193 sub = &frame->subframes[ch];
1195 /* subframe header */
1196 put_bits(&s->pb, 1, 0);
1197 put_bits(&s->pb, 6, sub->type_code);
1198 put_bits(&s->pb, 1, 0); /* no wasted bits */
1201 if(sub->type == FLAC_SUBFRAME_CONSTANT) {
1202 output_subframe_constant(s, ch);
1203 } else if(sub->type == FLAC_SUBFRAME_VERBATIM) {
1204 output_subframe_verbatim(s, ch);
1205 } else if(sub->type == FLAC_SUBFRAME_FIXED) {
1206 output_subframe_fixed(s, ch);
1207 } else if(sub->type == FLAC_SUBFRAME_LPC) {
1208 output_subframe_lpc(s, ch);
1213 static void output_frame_footer(FlacEncodeContext *s)
1216 flush_put_bits(&s->pb);
1217 crc = bswap_16(av_crc(av_crc_get_table(AV_CRC_16_ANSI), 0,
1218 s->pb.buf, put_bits_count(&s->pb)>>3));
1219 put_bits(&s->pb, 16, crc);
1220 flush_put_bits(&s->pb);
1223 static void update_md5_sum(FlacEncodeContext *s, int16_t *samples)
1225 #ifdef WORDS_BIGENDIAN
1227 for(i = 0; i < s->frame.blocksize*s->channels; i++) {
1228 int16_t smp = le2me_16(samples[i]);
1229 av_md5_update(s->md5ctx, (uint8_t *)&smp, 2);
1232 av_md5_update(s->md5ctx, (uint8_t *)samples, s->frame.blocksize*s->channels*2);
1236 static int flac_encode_frame(AVCodecContext *avctx, uint8_t *frame,
1237 int buf_size, void *data)
1240 FlacEncodeContext *s;
1241 int16_t *samples = data;
1245 s = avctx->priv_data;
1247 if(buf_size < s->max_framesize*2) {
1248 av_log(avctx, AV_LOG_ERROR, "output buffer too small\n");
1252 /* when the last block is reached, update the header in extradata */
1254 s->max_framesize = s->max_encoded_framesize;
1255 av_md5_final(s->md5ctx, s->md5sum);
1256 write_streaminfo(s, avctx->extradata);
1262 copy_samples(s, samples);
1264 channel_decorrelation(s);
1266 for(ch=0; ch<s->channels; ch++) {
1267 encode_residual(s, ch);
1271 init_put_bits(&s->pb, frame, buf_size);
1272 output_frame_header(s);
1273 output_subframes(s);
1274 output_frame_footer(s);
1275 out_bytes = put_bits_count(&s->pb) >> 3;
1277 if(out_bytes > s->max_framesize) {
1279 /* still too large. must be an error. */
1280 av_log(avctx, AV_LOG_ERROR, "error encoding frame\n");
1284 /* frame too large. use verbatim mode */
1285 for(ch=0; ch<s->channels; ch++) {
1286 encode_residual_v(s, ch);
1293 s->sample_count += avctx->frame_size;
1294 update_md5_sum(s, samples);
1295 if (out_bytes > s->max_encoded_framesize)
1296 s->max_encoded_framesize = out_bytes;
1297 if (out_bytes < s->min_framesize)
1298 s->min_framesize = out_bytes;
1303 static av_cold int flac_encode_close(AVCodecContext *avctx)
1305 if (avctx->priv_data) {
1306 FlacEncodeContext *s = avctx->priv_data;
1307 av_freep(&s->md5ctx);
1309 av_freep(&avctx->extradata);
1310 avctx->extradata_size = 0;
1311 av_freep(&avctx->coded_frame);
1315 AVCodec flac_encoder = {
1319 sizeof(FlacEncodeContext),
1324 .capabilities = CODEC_CAP_SMALL_LAST_FRAME | CODEC_CAP_DELAY,
1325 .sample_fmts = (enum SampleFormat[]){SAMPLE_FMT_S16,SAMPLE_FMT_NONE},
1326 .long_name = NULL_IF_CONFIG_SMALL("FLAC (Free Lossless Audio Codec)"),