3 * Copyright (c) 2006 Justin Ruggles <jruggle@earthlink.net>
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
23 #include "bitstream.h"
30 #define FLAC_MIN_BLOCKSIZE 16
31 #define FLAC_MAX_BLOCKSIZE 65535
33 #define FLAC_SUBFRAME_CONSTANT 0
34 #define FLAC_SUBFRAME_VERBATIM 1
35 #define FLAC_SUBFRAME_FIXED 8
36 #define FLAC_SUBFRAME_LPC 32
38 #define FLAC_CHMODE_NOT_STEREO 0
39 #define FLAC_CHMODE_LEFT_RIGHT 1
40 #define FLAC_CHMODE_LEFT_SIDE 8
41 #define FLAC_CHMODE_RIGHT_SIDE 9
42 #define FLAC_CHMODE_MID_SIDE 10
44 #define ORDER_METHOD_EST 0
45 #define ORDER_METHOD_2LEVEL 1
46 #define ORDER_METHOD_4LEVEL 2
47 #define ORDER_METHOD_8LEVEL 3
48 #define ORDER_METHOD_SEARCH 4
49 #define ORDER_METHOD_LOG 5
51 #define FLAC_STREAMINFO_SIZE 34
53 #define MIN_LPC_ORDER 1
54 #define MAX_LPC_ORDER 32
55 #define MAX_FIXED_ORDER 4
56 #define MAX_PARTITION_ORDER 8
57 #define MAX_PARTITIONS (1 << MAX_PARTITION_ORDER)
58 #define MAX_LPC_PRECISION 15
59 #define MAX_LPC_SHIFT 15
60 #define MAX_RICE_PARAM 14
62 typedef struct CompressionOptions {
63 int compression_level;
66 int lpc_coeff_precision;
67 int min_prediction_order;
68 int max_prediction_order;
69 int prediction_order_method;
70 int min_partition_order;
71 int max_partition_order;
74 typedef struct RiceContext {
76 int params[MAX_PARTITIONS];
79 typedef struct FlacSubframe {
84 int32_t coefs[MAX_LPC_ORDER];
87 int32_t samples[FLAC_MAX_BLOCKSIZE];
88 int32_t residual[FLAC_MAX_BLOCKSIZE+1];
91 typedef struct FlacFrame {
92 FlacSubframe subframes[FLAC_MAX_CH];
99 typedef struct FlacEncodeContext {
107 uint32_t frame_count;
109 CompressionOptions options;
110 AVCodecContext *avctx;
114 static const int flac_samplerates[16] = {
116 8000, 16000, 22050, 24000, 32000, 44100, 48000, 96000,
120 static const int flac_blocksizes[16] = {
123 576, 1152, 2304, 4608,
125 256, 512, 1024, 2048, 4096, 8192, 16384, 32768
129 * Writes streaminfo metadata block to byte array
131 static void write_streaminfo(FlacEncodeContext *s, uint8_t *header)
135 memset(header, 0, FLAC_STREAMINFO_SIZE);
136 init_put_bits(&pb, header, FLAC_STREAMINFO_SIZE);
138 /* streaminfo metadata block */
139 put_bits(&pb, 16, s->blocksize);
140 put_bits(&pb, 16, s->blocksize);
141 put_bits(&pb, 24, 0);
142 put_bits(&pb, 24, s->max_framesize);
143 put_bits(&pb, 20, s->samplerate);
144 put_bits(&pb, 3, s->channels-1);
145 put_bits(&pb, 5, 15); /* bits per sample - 1 */
147 /* total samples = 0 */
148 /* MD5 signature = 0 */
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 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",
354 s->blocksize = avctx->frame_size;
356 s->blocksize = select_blocksize(s->samplerate, s->options.block_time_ms);
357 avctx->frame_size = s->blocksize;
359 av_log(avctx, AV_LOG_DEBUG, " block size: %d\n", s->blocksize);
361 /* set LPC precision */
362 if(avctx->lpc_coeff_precision > 0) {
363 if(avctx->lpc_coeff_precision > MAX_LPC_PRECISION) {
364 av_log(avctx, AV_LOG_ERROR, "invalid lpc coeff precision: %d\n",
365 avctx->lpc_coeff_precision);
368 s->options.lpc_coeff_precision = avctx->lpc_coeff_precision;
370 /* select LPC precision based on block size */
371 if( s->blocksize <= 192) s->options.lpc_coeff_precision = 7;
372 else if(s->blocksize <= 384) s->options.lpc_coeff_precision = 8;
373 else if(s->blocksize <= 576) s->options.lpc_coeff_precision = 9;
374 else if(s->blocksize <= 1152) s->options.lpc_coeff_precision = 10;
375 else if(s->blocksize <= 2304) s->options.lpc_coeff_precision = 11;
376 else if(s->blocksize <= 4608) s->options.lpc_coeff_precision = 12;
377 else if(s->blocksize <= 8192) s->options.lpc_coeff_precision = 13;
378 else if(s->blocksize <= 16384) s->options.lpc_coeff_precision = 14;
379 else s->options.lpc_coeff_precision = 15;
381 av_log(avctx, AV_LOG_DEBUG, " lpc precision: %d\n",
382 s->options.lpc_coeff_precision);
384 /* set maximum encoded frame size in verbatim mode */
385 if(s->channels == 2) {
386 s->max_framesize = 14 + ((s->blocksize * 33 + 7) >> 3);
388 s->max_framesize = 14 + (s->blocksize * s->channels * 2);
391 streaminfo = av_malloc(FLAC_STREAMINFO_SIZE);
392 write_streaminfo(s, streaminfo);
393 avctx->extradata = streaminfo;
394 avctx->extradata_size = FLAC_STREAMINFO_SIZE;
398 avctx->coded_frame = avcodec_alloc_frame();
399 avctx->coded_frame->key_frame = 1;
404 static void init_frame(FlacEncodeContext *s)
411 for(i=0; i<16; i++) {
412 if(s->blocksize == flac_blocksizes[i]) {
413 frame->blocksize = flac_blocksizes[i];
414 frame->bs_code[0] = i;
415 frame->bs_code[1] = 0;
420 frame->blocksize = s->blocksize;
421 if(frame->blocksize <= 256) {
422 frame->bs_code[0] = 6;
423 frame->bs_code[1] = frame->blocksize-1;
425 frame->bs_code[0] = 7;
426 frame->bs_code[1] = frame->blocksize-1;
430 for(ch=0; ch<s->channels; ch++) {
431 frame->subframes[ch].obits = 16;
436 * Copy channel-interleaved input samples into separate subframes
438 static void copy_samples(FlacEncodeContext *s, int16_t *samples)
444 for(i=0,j=0; i<frame->blocksize; i++) {
445 for(ch=0; ch<s->channels; ch++,j++) {
446 frame->subframes[ch].samples[i] = samples[j];
452 #define rice_encode_count(sum, n, k) (((n)*((k)+1))+((sum-(n>>1))>>(k)))
455 * Solve for d/dk(rice_encode_count) = n-((sum-(n>>1))>>(k+1)) = 0
457 static int find_optimal_param(uint32_t sum, int n)
465 k = av_log2(n<256 ? FASTDIV(sum2,n) : sum2/n);
466 return FFMIN(k, MAX_RICE_PARAM);
469 static uint32_t calc_optimal_rice_params(RiceContext *rc, int porder,
470 uint32_t *sums, int n, int pred_order)
476 part = (1 << porder);
479 cnt = (n >> porder) - pred_order;
480 for(i=0; i<part; i++) {
481 k = find_optimal_param(sums[i], cnt);
483 all_bits += rice_encode_count(sums[i], cnt, k);
492 static void calc_sums(int pmin, int pmax, uint32_t *data, int n, int pred_order,
493 uint32_t sums[][MAX_PARTITIONS])
497 uint32_t *res, *res_end;
499 /* sums for highest level */
501 res = &data[pred_order];
502 res_end = &data[n >> pmax];
503 for(i=0; i<parts; i++) {
505 while(res < res_end){
511 /* sums for lower levels */
512 for(i=pmax-1; i>=pmin; i--) {
514 for(j=0; j<parts; j++) {
515 sums[i][j] = sums[i+1][2*j] + sums[i+1][2*j+1];
520 static uint32_t calc_rice_params(RiceContext *rc, int pmin, int pmax,
521 int32_t *data, int n, int pred_order)
524 uint32_t bits[MAX_PARTITION_ORDER+1];
528 uint32_t sums[MAX_PARTITION_ORDER+1][MAX_PARTITIONS];
530 assert(pmin >= 0 && pmin <= MAX_PARTITION_ORDER);
531 assert(pmax >= 0 && pmax <= MAX_PARTITION_ORDER);
532 assert(pmin <= pmax);
534 udata = av_malloc(n * sizeof(uint32_t));
536 udata[i] = (2*data[i]) ^ (data[i]>>31);
539 calc_sums(pmin, pmax, udata, n, pred_order, sums);
542 bits[pmin] = UINT32_MAX;
543 for(i=pmin; i<=pmax; i++) {
544 bits[i] = calc_optimal_rice_params(&tmp_rc, i, sums[i], n, pred_order);
545 if(bits[i] <= bits[opt_porder]) {
552 return bits[opt_porder];
555 static int get_max_p_order(int max_porder, int n, int order)
557 int porder = FFMIN(max_porder, av_log2(n^(n-1)));
559 porder = FFMIN(porder, av_log2(n/order));
563 static uint32_t calc_rice_params_fixed(RiceContext *rc, int pmin, int pmax,
564 int32_t *data, int n, int pred_order,
568 pmin = get_max_p_order(pmin, n, pred_order);
569 pmax = get_max_p_order(pmax, n, pred_order);
570 bits = pred_order*bps + 6;
571 bits += calc_rice_params(rc, pmin, pmax, data, n, pred_order);
575 static uint32_t calc_rice_params_lpc(RiceContext *rc, int pmin, int pmax,
576 int32_t *data, int n, int pred_order,
577 int bps, int precision)
580 pmin = get_max_p_order(pmin, n, pred_order);
581 pmax = get_max_p_order(pmax, n, pred_order);
582 bits = pred_order*bps + 4 + 5 + pred_order*precision + 6;
583 bits += calc_rice_params(rc, pmin, pmax, data, n, pred_order);
588 * Apply Welch window function to audio block
590 static void apply_welch_window(const int32_t *data, int len, double *w_data)
596 assert(!(len&1)); //the optimization in r11881 does not support odd len
597 //if someone wants odd len extend the change in r11881
600 c = 2.0 / (len - 1.0);
604 for(i=0; i<n2; i++) {
607 w_data[-i-1] = data[-i-1] * w;
608 w_data[+i ] = data[+i ] * w;
613 * Calculates autocorrelation data from audio samples
614 * A Welch window function is applied before calculation.
616 void ff_flac_compute_autocorr(const int32_t *data, int len, int lag,
620 double tmp[len + lag + 1];
621 double *data1= tmp + lag;
623 apply_welch_window(data, len, data1);
629 for(j=0; j<lag; j+=2){
630 double sum0 = 1.0, sum1 = 1.0;
631 for(i=0; i<len; i++){
632 sum0 += data1[i] * data1[i-j];
633 sum1 += data1[i] * data1[i-j-1];
641 for(i=0; i<len; i+=2){
642 sum += data1[i ] * data1[i-j ]
643 + data1[i+1] * data1[i-j+1];
650 * Levinson-Durbin recursion.
651 * Produces LPC coefficients from autocorrelation data.
653 static void compute_lpc_coefs(const double *autoc, int max_order,
654 double lpc[][MAX_LPC_ORDER], double *ref)
658 double lpc_tmp[MAX_LPC_ORDER];
660 for(i=0; i<max_order; i++) lpc_tmp[i] = 0;
663 for(i=0; i<max_order; i++) {
666 r -= lpc_tmp[j] * autoc[i-j];
671 err *= 1.0 - (r * r);
675 for(j=0; j<i2; j++) {
677 lpc_tmp[j] += r * lpc_tmp[i-1-j];
678 lpc_tmp[i-1-j] += r * tmp;
681 lpc_tmp[j] += lpc_tmp[j] * r;
684 for(j=0; j<=i; j++) {
685 lpc[i][j] = -lpc_tmp[j];
691 * Quantize LPC coefficients
693 static void quantize_lpc_coefs(double *lpc_in, int order, int precision,
694 int32_t *lpc_out, int *shift)
701 /* define maximum levels */
702 qmax = (1 << (precision - 1)) - 1;
704 /* find maximum coefficient value */
706 for(i=0; i<order; i++) {
707 cmax= FFMAX(cmax, fabs(lpc_in[i]));
710 /* if maximum value quantizes to zero, return all zeros */
711 if(cmax * (1 << MAX_LPC_SHIFT) < 1.0) {
713 memset(lpc_out, 0, sizeof(int32_t) * order);
717 /* calculate level shift which scales max coeff to available bits */
719 while((cmax * (1 << sh) > qmax) && (sh > 0)) {
723 /* since negative shift values are unsupported in decoder, scale down
724 coefficients instead */
725 if(sh == 0 && cmax > qmax) {
726 double scale = ((double)qmax) / cmax;
727 for(i=0; i<order; i++) {
732 /* output quantized coefficients and level shift */
734 for(i=0; i<order; i++) {
735 error += lpc_in[i] * (1 << sh);
736 lpc_out[i] = av_clip(lrintf(error), -qmax, qmax);
742 static int estimate_best_order(double *ref, int max_order)
747 for(i=max_order-1; i>=0; i--) {
757 * Calculate LPC coefficients for multiple orders
759 static int lpc_calc_coefs(FlacEncodeContext *s,
760 const int32_t *samples, int blocksize, int max_order,
761 int precision, int32_t coefs[][MAX_LPC_ORDER],
762 int *shift, int use_lpc, int omethod)
764 double autoc[MAX_LPC_ORDER+1];
765 double ref[MAX_LPC_ORDER];
766 double lpc[MAX_LPC_ORDER][MAX_LPC_ORDER];
770 assert(max_order >= MIN_LPC_ORDER && max_order <= MAX_LPC_ORDER);
773 s->dsp.flac_compute_autocorr(samples, blocksize, max_order, autoc);
775 compute_lpc_coefs(autoc, max_order, lpc, ref);
778 double var[MAX_LPC_ORDER+1], weight;
780 for(pass=0; pass<use_lpc-1; pass++){
781 av_init_lls(&m[pass&1], max_order);
784 for(i=max_order; i<blocksize; i++){
785 for(j=0; j<=max_order; j++)
786 var[j]= samples[i-j];
789 double eval, inv, rinv;
790 eval= av_evaluate_lls(&m[(pass-1)&1], var+1, max_order-1);
791 eval= (512>>pass) + fabs(eval - var[0]);
794 for(j=0; j<=max_order; j++)
800 av_update_lls(&m[pass&1], var, 1.0);
802 av_solve_lls(&m[pass&1], 0.001, 0);
805 for(i=0; i<max_order; i++){
806 for(j=0; j<max_order; j++)
807 lpc[i][j]= m[(pass-1)&1].coeff[i][j];
808 ref[i]= sqrt(m[(pass-1)&1].variance[i] / weight) * (blocksize - max_order) / 4000;
810 for(i=max_order-1; i>0; i--)
811 ref[i] = ref[i-1] - ref[i];
813 opt_order = max_order;
815 if(omethod == ORDER_METHOD_EST) {
816 opt_order = estimate_best_order(ref, max_order);
818 quantize_lpc_coefs(lpc[i], i+1, precision, coefs[i], &shift[i]);
820 for(i=0; i<max_order; i++) {
821 quantize_lpc_coefs(lpc[i], i+1, precision, coefs[i], &shift[i]);
829 static void encode_residual_verbatim(int32_t *res, int32_t *smp, int n)
832 memcpy(res, smp, n * sizeof(int32_t));
835 static void encode_residual_fixed(int32_t *res, const int32_t *smp, int n,
840 for(i=0; i<order; i++) {
845 for(i=order; i<n; i++)
848 for(i=order; i<n; i++)
849 res[i]= smp[i] - smp[i-1];
851 int a = smp[order-1] - smp[order-2];
852 for(i=order; i<n; i+=2) {
853 int b = smp[i] - smp[i-1];
855 a = smp[i+1] - smp[i];
859 int a = smp[order-1] - smp[order-2];
860 int c = smp[order-1] - 2*smp[order-2] + smp[order-3];
861 for(i=order; i<n; i+=2) {
862 int b = smp[i] - smp[i-1];
865 a = smp[i+1] - smp[i];
870 int a = smp[order-1] - smp[order-2];
871 int c = smp[order-1] - 2*smp[order-2] + smp[order-3];
872 int e = smp[order-1] - 3*smp[order-2] + 3*smp[order-3] - smp[order-4];
873 for(i=order; i<n; i+=2) {
874 int b = smp[i] - smp[i-1];
878 a = smp[i+1] - smp[i];
887 int c = coefs[(x)-1];\
893 static av_always_inline void encode_residual_lpc_unrolled(
894 int32_t *res, const int32_t *smp, int n,
895 int order, const int32_t *coefs, int shift, int big)
898 for(i=order; i<n; i+=2) {
899 int s = smp[i-order];
948 res[i ] = smp[i ] - (p0 >> shift);
949 res[i+1] = smp[i+1] - (p1 >> shift);
953 static void encode_residual_lpc(int32_t *res, const int32_t *smp, int n,
954 int order, const int32_t *coefs, int shift)
957 for(i=0; i<order; i++) {
961 for(i=order; i<n; i+=2) {
965 for(j=0; j<order; j++) {
971 res[i ] = smp[i ] - (p0 >> shift);
972 res[i+1] = smp[i+1] - (p1 >> shift);
976 case 1: encode_residual_lpc_unrolled(res, smp, n, 1, coefs, shift, 0); break;
977 case 2: encode_residual_lpc_unrolled(res, smp, n, 2, coefs, shift, 0); break;
978 case 3: encode_residual_lpc_unrolled(res, smp, n, 3, coefs, shift, 0); break;
979 case 4: encode_residual_lpc_unrolled(res, smp, n, 4, coefs, shift, 0); break;
980 case 5: encode_residual_lpc_unrolled(res, smp, n, 5, coefs, shift, 0); break;
981 case 6: encode_residual_lpc_unrolled(res, smp, n, 6, coefs, shift, 0); break;
982 case 7: encode_residual_lpc_unrolled(res, smp, n, 7, coefs, shift, 0); break;
983 case 8: encode_residual_lpc_unrolled(res, smp, n, 8, coefs, shift, 0); break;
984 default: encode_residual_lpc_unrolled(res, smp, n, order, coefs, shift, 1); break;
989 static int encode_residual(FlacEncodeContext *ctx, int ch)
992 int min_order, max_order, opt_order, precision, omethod;
993 int min_porder, max_porder;
996 int32_t coefs[MAX_LPC_ORDER][MAX_LPC_ORDER];
997 int shift[MAX_LPC_ORDER];
1000 frame = &ctx->frame;
1001 sub = &frame->subframes[ch];
1002 res = sub->residual;
1004 n = frame->blocksize;
1007 for(i=1; i<n; i++) {
1008 if(smp[i] != smp[0]) break;
1011 sub->type = sub->type_code = FLAC_SUBFRAME_CONSTANT;
1018 sub->type = sub->type_code = FLAC_SUBFRAME_VERBATIM;
1019 encode_residual_verbatim(res, smp, n);
1020 return sub->obits * n;
1023 min_order = ctx->options.min_prediction_order;
1024 max_order = ctx->options.max_prediction_order;
1025 min_porder = ctx->options.min_partition_order;
1026 max_porder = ctx->options.max_partition_order;
1027 precision = ctx->options.lpc_coeff_precision;
1028 omethod = ctx->options.prediction_order_method;
1031 if(!ctx->options.use_lpc || max_order == 0 || (n <= max_order)) {
1032 uint32_t bits[MAX_FIXED_ORDER+1];
1033 if(max_order > MAX_FIXED_ORDER) max_order = MAX_FIXED_ORDER;
1035 bits[0] = UINT32_MAX;
1036 for(i=min_order; i<=max_order; i++) {
1037 encode_residual_fixed(res, smp, n, i);
1038 bits[i] = calc_rice_params_fixed(&sub->rc, min_porder, max_porder, res,
1040 if(bits[i] < bits[opt_order]) {
1044 sub->order = opt_order;
1045 sub->type = FLAC_SUBFRAME_FIXED;
1046 sub->type_code = sub->type | sub->order;
1047 if(sub->order != max_order) {
1048 encode_residual_fixed(res, smp, n, sub->order);
1049 return calc_rice_params_fixed(&sub->rc, min_porder, max_porder, res, n,
1050 sub->order, sub->obits);
1052 return bits[sub->order];
1056 opt_order = lpc_calc_coefs(ctx, smp, n, max_order, precision, coefs, shift, ctx->options.use_lpc, omethod);
1058 if(omethod == ORDER_METHOD_2LEVEL ||
1059 omethod == ORDER_METHOD_4LEVEL ||
1060 omethod == ORDER_METHOD_8LEVEL) {
1061 int levels = 1 << omethod;
1062 uint32_t bits[levels];
1064 int opt_index = levels-1;
1065 opt_order = max_order-1;
1066 bits[opt_index] = UINT32_MAX;
1067 for(i=levels-1; i>=0; i--) {
1068 order = min_order + (((max_order-min_order+1) * (i+1)) / levels)-1;
1069 if(order < 0) order = 0;
1070 encode_residual_lpc(res, smp, n, order+1, coefs[order], shift[order]);
1071 bits[i] = calc_rice_params_lpc(&sub->rc, min_porder, max_porder,
1072 res, n, order+1, sub->obits, precision);
1073 if(bits[i] < bits[opt_index]) {
1079 } else if(omethod == ORDER_METHOD_SEARCH) {
1080 // brute-force optimal order search
1081 uint32_t bits[MAX_LPC_ORDER];
1083 bits[0] = UINT32_MAX;
1084 for(i=min_order-1; i<max_order; i++) {
1085 encode_residual_lpc(res, smp, n, i+1, coefs[i], shift[i]);
1086 bits[i] = calc_rice_params_lpc(&sub->rc, min_porder, max_porder,
1087 res, n, i+1, sub->obits, precision);
1088 if(bits[i] < bits[opt_order]) {
1093 } else if(omethod == ORDER_METHOD_LOG) {
1094 uint32_t bits[MAX_LPC_ORDER];
1097 opt_order= min_order - 1 + (max_order-min_order)/3;
1098 memset(bits, -1, sizeof(bits));
1100 for(step=16 ;step; step>>=1){
1101 int last= opt_order;
1102 for(i=last-step; i<=last+step; i+= step){
1103 if(i<min_order-1 || i>=max_order || bits[i] < UINT32_MAX)
1105 encode_residual_lpc(res, smp, n, i+1, coefs[i], shift[i]);
1106 bits[i] = calc_rice_params_lpc(&sub->rc, min_porder, max_porder,
1107 res, n, i+1, sub->obits, precision);
1108 if(bits[i] < bits[opt_order])
1115 sub->order = opt_order;
1116 sub->type = FLAC_SUBFRAME_LPC;
1117 sub->type_code = sub->type | (sub->order-1);
1118 sub->shift = shift[sub->order-1];
1119 for(i=0; i<sub->order; i++) {
1120 sub->coefs[i] = coefs[sub->order-1][i];
1122 encode_residual_lpc(res, smp, n, sub->order, sub->coefs, sub->shift);
1123 return calc_rice_params_lpc(&sub->rc, min_porder, max_porder, res, n, sub->order,
1124 sub->obits, precision);
1127 static int encode_residual_v(FlacEncodeContext *ctx, int ch)
1134 frame = &ctx->frame;
1135 sub = &frame->subframes[ch];
1136 res = sub->residual;
1138 n = frame->blocksize;
1141 for(i=1; i<n; i++) {
1142 if(smp[i] != smp[0]) break;
1145 sub->type = sub->type_code = FLAC_SUBFRAME_CONSTANT;
1151 sub->type = sub->type_code = FLAC_SUBFRAME_VERBATIM;
1152 encode_residual_verbatim(res, smp, n);
1153 return sub->obits * n;
1156 static int estimate_stereo_mode(int32_t *left_ch, int32_t *right_ch, int n)
1164 /* calculate sum of 2nd order residual for each channel */
1165 sum[0] = sum[1] = sum[2] = sum[3] = 0;
1166 for(i=2; i<n; i++) {
1167 lt = left_ch[i] - 2*left_ch[i-1] + left_ch[i-2];
1168 rt = right_ch[i] - 2*right_ch[i-1] + right_ch[i-2];
1169 sum[2] += FFABS((lt + rt) >> 1);
1170 sum[3] += FFABS(lt - rt);
1171 sum[0] += FFABS(lt);
1172 sum[1] += FFABS(rt);
1174 /* estimate bit counts */
1175 for(i=0; i<4; i++) {
1176 k = find_optimal_param(2*sum[i], n);
1177 sum[i] = rice_encode_count(2*sum[i], n, k);
1180 /* calculate score for each mode */
1181 score[0] = sum[0] + sum[1];
1182 score[1] = sum[0] + sum[3];
1183 score[2] = sum[1] + sum[3];
1184 score[3] = sum[2] + sum[3];
1186 /* return mode with lowest score */
1188 for(i=1; i<4; i++) {
1189 if(score[i] < score[best]) {
1194 return FLAC_CHMODE_LEFT_RIGHT;
1195 } else if(best == 1) {
1196 return FLAC_CHMODE_LEFT_SIDE;
1197 } else if(best == 2) {
1198 return FLAC_CHMODE_RIGHT_SIDE;
1200 return FLAC_CHMODE_MID_SIDE;
1205 * Perform stereo channel decorrelation
1207 static void channel_decorrelation(FlacEncodeContext *ctx)
1210 int32_t *left, *right;
1213 frame = &ctx->frame;
1214 n = frame->blocksize;
1215 left = frame->subframes[0].samples;
1216 right = frame->subframes[1].samples;
1218 if(ctx->channels != 2) {
1219 frame->ch_mode = FLAC_CHMODE_NOT_STEREO;
1223 frame->ch_mode = estimate_stereo_mode(left, right, n);
1225 /* perform decorrelation and adjust bits-per-sample */
1226 if(frame->ch_mode == FLAC_CHMODE_LEFT_RIGHT) {
1229 if(frame->ch_mode == FLAC_CHMODE_MID_SIDE) {
1231 for(i=0; i<n; i++) {
1233 left[i] = (tmp + right[i]) >> 1;
1234 right[i] = tmp - right[i];
1236 frame->subframes[1].obits++;
1237 } else if(frame->ch_mode == FLAC_CHMODE_LEFT_SIDE) {
1238 for(i=0; i<n; i++) {
1239 right[i] = left[i] - right[i];
1241 frame->subframes[1].obits++;
1243 for(i=0; i<n; i++) {
1244 left[i] -= right[i];
1246 frame->subframes[0].obits++;
1250 static void put_sbits(PutBitContext *pb, int bits, int32_t val)
1252 assert(bits >= 0 && bits <= 31);
1254 put_bits(pb, bits, val & ((1<<bits)-1));
1257 static void write_utf8(PutBitContext *pb, uint32_t val)
1260 PUT_UTF8(val, tmp, put_bits(pb, 8, tmp);)
1263 static void output_frame_header(FlacEncodeContext *s)
1270 put_bits(&s->pb, 16, 0xFFF8);
1271 put_bits(&s->pb, 4, frame->bs_code[0]);
1272 put_bits(&s->pb, 4, s->sr_code[0]);
1273 if(frame->ch_mode == FLAC_CHMODE_NOT_STEREO) {
1274 put_bits(&s->pb, 4, s->ch_code);
1276 put_bits(&s->pb, 4, frame->ch_mode);
1278 put_bits(&s->pb, 3, 4); /* bits-per-sample code */
1279 put_bits(&s->pb, 1, 0);
1280 write_utf8(&s->pb, s->frame_count);
1281 if(frame->bs_code[0] == 6) {
1282 put_bits(&s->pb, 8, frame->bs_code[1]);
1283 } else if(frame->bs_code[0] == 7) {
1284 put_bits(&s->pb, 16, frame->bs_code[1]);
1286 if(s->sr_code[0] == 12) {
1287 put_bits(&s->pb, 8, s->sr_code[1]);
1288 } else if(s->sr_code[0] > 12) {
1289 put_bits(&s->pb, 16, s->sr_code[1]);
1291 flush_put_bits(&s->pb);
1292 crc = av_crc(av_crc_get_table(AV_CRC_8_ATM), 0,
1293 s->pb.buf, put_bits_count(&s->pb)>>3);
1294 put_bits(&s->pb, 8, crc);
1297 static void output_subframe_constant(FlacEncodeContext *s, int ch)
1302 sub = &s->frame.subframes[ch];
1303 res = sub->residual[0];
1304 put_sbits(&s->pb, sub->obits, res);
1307 static void output_subframe_verbatim(FlacEncodeContext *s, int ch)
1315 sub = &frame->subframes[ch];
1317 for(i=0; i<frame->blocksize; i++) {
1318 res = sub->residual[i];
1319 put_sbits(&s->pb, sub->obits, res);
1323 static void output_residual(FlacEncodeContext *ctx, int ch)
1325 int i, j, p, n, parts;
1326 int k, porder, psize, res_cnt;
1331 frame = &ctx->frame;
1332 sub = &frame->subframes[ch];
1333 res = sub->residual;
1334 n = frame->blocksize;
1336 /* rice-encoded block */
1337 put_bits(&ctx->pb, 2, 0);
1339 /* partition order */
1340 porder = sub->rc.porder;
1341 psize = n >> porder;
1342 parts = (1 << porder);
1343 put_bits(&ctx->pb, 4, porder);
1344 res_cnt = psize - sub->order;
1348 for(p=0; p<parts; p++) {
1349 k = sub->rc.params[p];
1350 put_bits(&ctx->pb, 4, k);
1351 if(p == 1) res_cnt = psize;
1352 for(i=0; i<res_cnt && j<n; i++, j++) {
1353 set_sr_golomb_flac(&ctx->pb, res[j], k, INT32_MAX, 0);
1358 static void output_subframe_fixed(FlacEncodeContext *ctx, int ch)
1364 frame = &ctx->frame;
1365 sub = &frame->subframes[ch];
1367 /* warm-up samples */
1368 for(i=0; i<sub->order; i++) {
1369 put_sbits(&ctx->pb, sub->obits, sub->residual[i]);
1373 output_residual(ctx, ch);
1376 static void output_subframe_lpc(FlacEncodeContext *ctx, int ch)
1382 frame = &ctx->frame;
1383 sub = &frame->subframes[ch];
1385 /* warm-up samples */
1386 for(i=0; i<sub->order; i++) {
1387 put_sbits(&ctx->pb, sub->obits, sub->residual[i]);
1390 /* LPC coefficients */
1391 cbits = ctx->options.lpc_coeff_precision;
1392 put_bits(&ctx->pb, 4, cbits-1);
1393 put_sbits(&ctx->pb, 5, sub->shift);
1394 for(i=0; i<sub->order; i++) {
1395 put_sbits(&ctx->pb, cbits, sub->coefs[i]);
1399 output_residual(ctx, ch);
1402 static void output_subframes(FlacEncodeContext *s)
1410 for(ch=0; ch<s->channels; ch++) {
1411 sub = &frame->subframes[ch];
1413 /* subframe header */
1414 put_bits(&s->pb, 1, 0);
1415 put_bits(&s->pb, 6, sub->type_code);
1416 put_bits(&s->pb, 1, 0); /* no wasted bits */
1419 if(sub->type == FLAC_SUBFRAME_CONSTANT) {
1420 output_subframe_constant(s, ch);
1421 } else if(sub->type == FLAC_SUBFRAME_VERBATIM) {
1422 output_subframe_verbatim(s, ch);
1423 } else if(sub->type == FLAC_SUBFRAME_FIXED) {
1424 output_subframe_fixed(s, ch);
1425 } else if(sub->type == FLAC_SUBFRAME_LPC) {
1426 output_subframe_lpc(s, ch);
1431 static void output_frame_footer(FlacEncodeContext *s)
1434 flush_put_bits(&s->pb);
1435 crc = bswap_16(av_crc(av_crc_get_table(AV_CRC_16_ANSI), 0,
1436 s->pb.buf, put_bits_count(&s->pb)>>3));
1437 put_bits(&s->pb, 16, crc);
1438 flush_put_bits(&s->pb);
1441 static int flac_encode_frame(AVCodecContext *avctx, uint8_t *frame,
1442 int buf_size, void *data)
1445 FlacEncodeContext *s;
1446 int16_t *samples = data;
1449 s = avctx->priv_data;
1451 s->blocksize = avctx->frame_size;
1454 copy_samples(s, samples);
1456 channel_decorrelation(s);
1458 for(ch=0; ch<s->channels; ch++) {
1459 encode_residual(s, ch);
1461 init_put_bits(&s->pb, frame, buf_size);
1462 output_frame_header(s);
1463 output_subframes(s);
1464 output_frame_footer(s);
1465 out_bytes = put_bits_count(&s->pb) >> 3;
1467 if(out_bytes > s->max_framesize || out_bytes >= buf_size) {
1468 /* frame too large. use verbatim mode */
1469 for(ch=0; ch<s->channels; ch++) {
1470 encode_residual_v(s, ch);
1472 init_put_bits(&s->pb, frame, buf_size);
1473 output_frame_header(s);
1474 output_subframes(s);
1475 output_frame_footer(s);
1476 out_bytes = put_bits_count(&s->pb) >> 3;
1478 if(out_bytes > s->max_framesize || out_bytes >= buf_size) {
1479 /* still too large. must be an error. */
1480 av_log(avctx, AV_LOG_ERROR, "error encoding frame\n");
1489 static int flac_encode_close(AVCodecContext *avctx)
1491 av_freep(&avctx->extradata);
1492 avctx->extradata_size = 0;
1493 av_freep(&avctx->coded_frame);
1497 AVCodec flac_encoder = {
1501 sizeof(FlacEncodeContext),
1506 .capabilities = CODEC_CAP_SMALL_LAST_FRAME,