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
22 #include "libavutil/crc.h"
23 #include "libavutil/lls.h"
25 #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 {
106 uint32_t frame_count;
108 CompressionOptions options;
109 AVCodecContext *avctx;
113 static const int flac_samplerates[16] = {
115 8000, 16000, 22050, 24000, 32000, 44100, 48000, 96000,
119 static const int flac_blocksizes[16] = {
122 576, 1152, 2304, 4608,
124 256, 512, 1024, 2048, 4096, 8192, 16384, 32768
128 * Writes streaminfo metadata block to byte array
130 static void write_streaminfo(FlacEncodeContext *s, uint8_t *header)
134 memset(header, 0, FLAC_STREAMINFO_SIZE);
135 init_put_bits(&pb, header, FLAC_STREAMINFO_SIZE);
137 /* streaminfo metadata block */
138 put_bits(&pb, 16, s->avctx->frame_size);
139 put_bits(&pb, 16, s->avctx->frame_size);
140 put_bits(&pb, 24, 0);
141 put_bits(&pb, 24, s->max_framesize);
142 put_bits(&pb, 20, s->samplerate);
143 put_bits(&pb, 3, s->channels-1);
144 put_bits(&pb, 5, 15); /* bits per sample - 1 */
146 /* total samples = 0 */
147 /* MD5 signature = 0 */
151 * Sets blocksize based on samplerate
152 * Chooses the closest predefined blocksize >= BLOCK_TIME_MS milliseconds
154 static int select_blocksize(int samplerate, int block_time_ms)
160 assert(samplerate > 0);
161 blocksize = flac_blocksizes[1];
162 target = (samplerate * block_time_ms) / 1000;
163 for(i=0; i<16; i++) {
164 if(target >= flac_blocksizes[i] && flac_blocksizes[i] > blocksize) {
165 blocksize = flac_blocksizes[i];
171 static av_cold int flac_encode_init(AVCodecContext *avctx)
173 int freq = avctx->sample_rate;
174 int channels = avctx->channels;
175 FlacEncodeContext *s = avctx->priv_data;
181 dsputil_init(&s->dsp, avctx);
183 if(avctx->sample_fmt != SAMPLE_FMT_S16) {
187 if(channels < 1 || channels > FLAC_MAX_CH) {
190 s->channels = channels;
191 s->ch_code = s->channels-1;
193 /* find samplerate in table */
196 for(i=4; i<12; i++) {
197 if(freq == flac_samplerates[i]) {
198 s->samplerate = flac_samplerates[i];
204 /* if not in table, samplerate is non-standard */
206 if(freq % 1000 == 0 && freq < 255000) {
208 s->sr_code[1] = freq / 1000;
209 } else if(freq % 10 == 0 && freq < 655350) {
211 s->sr_code[1] = freq / 10;
212 } else if(freq < 65535) {
214 s->sr_code[1] = freq;
218 s->samplerate = freq;
221 /* set compression option defaults based on avctx->compression_level */
222 if(avctx->compression_level < 0) {
223 s->options.compression_level = 5;
225 s->options.compression_level = avctx->compression_level;
227 av_log(avctx, AV_LOG_DEBUG, " compression: %d\n", s->options.compression_level);
229 level= s->options.compression_level;
231 av_log(avctx, AV_LOG_ERROR, "invalid compression level: %d\n",
232 s->options.compression_level);
236 s->options.block_time_ms = ((int[]){ 27, 27, 27,105,105,105,105,105,105,105,105,105,105})[level];
237 s->options.use_lpc = ((int[]){ 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1})[level];
238 s->options.min_prediction_order= ((int[]){ 2, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1})[level];
239 s->options.max_prediction_order= ((int[]){ 3, 4, 4, 6, 8, 8, 8, 8, 12, 12, 12, 32, 32})[level];
240 s->options.prediction_order_method = ((int[]){ ORDER_METHOD_EST, ORDER_METHOD_EST, ORDER_METHOD_EST,
241 ORDER_METHOD_EST, ORDER_METHOD_EST, ORDER_METHOD_EST,
242 ORDER_METHOD_4LEVEL, ORDER_METHOD_LOG, ORDER_METHOD_4LEVEL,
243 ORDER_METHOD_LOG, ORDER_METHOD_SEARCH, ORDER_METHOD_LOG,
244 ORDER_METHOD_SEARCH})[level];
245 s->options.min_partition_order = ((int[]){ 2, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0})[level];
246 s->options.max_partition_order = ((int[]){ 2, 2, 3, 3, 3, 8, 8, 8, 8, 8, 8, 8, 8})[level];
248 /* set compression option overrides from AVCodecContext */
249 if(avctx->use_lpc >= 0) {
250 s->options.use_lpc = av_clip(avctx->use_lpc, 0, 11);
252 if(s->options.use_lpc == 1)
253 av_log(avctx, AV_LOG_DEBUG, " use lpc: Levinson-Durbin recursion with Welch window\n");
254 else if(s->options.use_lpc > 1)
255 av_log(avctx, AV_LOG_DEBUG, " use lpc: Cholesky factorization\n");
257 if(avctx->min_prediction_order >= 0) {
258 if(s->options.use_lpc) {
259 if(avctx->min_prediction_order < MIN_LPC_ORDER ||
260 avctx->min_prediction_order > MAX_LPC_ORDER) {
261 av_log(avctx, AV_LOG_ERROR, "invalid min prediction order: %d\n",
262 avctx->min_prediction_order);
266 if(avctx->min_prediction_order > MAX_FIXED_ORDER) {
267 av_log(avctx, AV_LOG_ERROR, "invalid min prediction order: %d\n",
268 avctx->min_prediction_order);
272 s->options.min_prediction_order = avctx->min_prediction_order;
274 if(avctx->max_prediction_order >= 0) {
275 if(s->options.use_lpc) {
276 if(avctx->max_prediction_order < MIN_LPC_ORDER ||
277 avctx->max_prediction_order > MAX_LPC_ORDER) {
278 av_log(avctx, AV_LOG_ERROR, "invalid max prediction order: %d\n",
279 avctx->max_prediction_order);
283 if(avctx->max_prediction_order > MAX_FIXED_ORDER) {
284 av_log(avctx, AV_LOG_ERROR, "invalid max prediction order: %d\n",
285 avctx->max_prediction_order);
289 s->options.max_prediction_order = avctx->max_prediction_order;
291 if(s->options.max_prediction_order < s->options.min_prediction_order) {
292 av_log(avctx, AV_LOG_ERROR, "invalid prediction orders: min=%d max=%d\n",
293 s->options.min_prediction_order, s->options.max_prediction_order);
296 av_log(avctx, AV_LOG_DEBUG, " prediction order: %d, %d\n",
297 s->options.min_prediction_order, s->options.max_prediction_order);
299 if(avctx->prediction_order_method >= 0) {
300 if(avctx->prediction_order_method > ORDER_METHOD_LOG) {
301 av_log(avctx, AV_LOG_ERROR, "invalid prediction order method: %d\n",
302 avctx->prediction_order_method);
305 s->options.prediction_order_method = avctx->prediction_order_method;
307 switch(s->options.prediction_order_method) {
308 case ORDER_METHOD_EST: av_log(avctx, AV_LOG_DEBUG, " order method: %s\n",
310 case ORDER_METHOD_2LEVEL: av_log(avctx, AV_LOG_DEBUG, " order method: %s\n",
312 case ORDER_METHOD_4LEVEL: av_log(avctx, AV_LOG_DEBUG, " order method: %s\n",
314 case ORDER_METHOD_8LEVEL: av_log(avctx, AV_LOG_DEBUG, " order method: %s\n",
316 case ORDER_METHOD_SEARCH: av_log(avctx, AV_LOG_DEBUG, " order method: %s\n",
317 "full search"); break;
318 case ORDER_METHOD_LOG: av_log(avctx, AV_LOG_DEBUG, " order method: %s\n",
319 "log search"); break;
322 if(avctx->min_partition_order >= 0) {
323 if(avctx->min_partition_order > MAX_PARTITION_ORDER) {
324 av_log(avctx, AV_LOG_ERROR, "invalid min partition order: %d\n",
325 avctx->min_partition_order);
328 s->options.min_partition_order = avctx->min_partition_order;
330 if(avctx->max_partition_order >= 0) {
331 if(avctx->max_partition_order > MAX_PARTITION_ORDER) {
332 av_log(avctx, AV_LOG_ERROR, "invalid max partition order: %d\n",
333 avctx->max_partition_order);
336 s->options.max_partition_order = avctx->max_partition_order;
338 if(s->options.max_partition_order < s->options.min_partition_order) {
339 av_log(avctx, AV_LOG_ERROR, "invalid partition orders: min=%d max=%d\n",
340 s->options.min_partition_order, s->options.max_partition_order);
343 av_log(avctx, AV_LOG_DEBUG, " partition order: %d, %d\n",
344 s->options.min_partition_order, s->options.max_partition_order);
346 if(avctx->frame_size > 0) {
347 if(avctx->frame_size < FLAC_MIN_BLOCKSIZE ||
348 avctx->frame_size > FLAC_MAX_BLOCKSIZE) {
349 av_log(avctx, AV_LOG_ERROR, "invalid block size: %d\n",
354 s->avctx->frame_size = select_blocksize(s->samplerate, s->options.block_time_ms);
356 av_log(avctx, AV_LOG_DEBUG, " block size: %d\n", s->avctx->frame_size);
358 /* set LPC precision */
359 if(avctx->lpc_coeff_precision > 0) {
360 if(avctx->lpc_coeff_precision > MAX_LPC_PRECISION) {
361 av_log(avctx, AV_LOG_ERROR, "invalid lpc coeff precision: %d\n",
362 avctx->lpc_coeff_precision);
365 s->options.lpc_coeff_precision = avctx->lpc_coeff_precision;
367 /* default LPC precision */
368 s->options.lpc_coeff_precision = 15;
370 av_log(avctx, AV_LOG_DEBUG, " lpc precision: %d\n",
371 s->options.lpc_coeff_precision);
373 /* set maximum encoded frame size in verbatim mode */
374 if(s->channels == 2) {
375 s->max_framesize = 14 + ((s->avctx->frame_size * 33 + 7) >> 3);
377 s->max_framesize = 14 + (s->avctx->frame_size * s->channels * 2);
380 streaminfo = av_malloc(FLAC_STREAMINFO_SIZE);
381 write_streaminfo(s, streaminfo);
382 avctx->extradata = streaminfo;
383 avctx->extradata_size = FLAC_STREAMINFO_SIZE;
387 avctx->coded_frame = avcodec_alloc_frame();
388 avctx->coded_frame->key_frame = 1;
393 static void init_frame(FlacEncodeContext *s)
400 for(i=0; i<16; i++) {
401 if(s->avctx->frame_size == flac_blocksizes[i]) {
402 frame->blocksize = flac_blocksizes[i];
403 frame->bs_code[0] = i;
404 frame->bs_code[1] = 0;
409 frame->blocksize = s->avctx->frame_size;
410 if(frame->blocksize <= 256) {
411 frame->bs_code[0] = 6;
412 frame->bs_code[1] = frame->blocksize-1;
414 frame->bs_code[0] = 7;
415 frame->bs_code[1] = frame->blocksize-1;
419 for(ch=0; ch<s->channels; ch++) {
420 frame->subframes[ch].obits = 16;
425 * Copy channel-interleaved input samples into separate subframes
427 static void copy_samples(FlacEncodeContext *s, int16_t *samples)
433 for(i=0,j=0; i<frame->blocksize; i++) {
434 for(ch=0; ch<s->channels; ch++,j++) {
435 frame->subframes[ch].samples[i] = samples[j];
441 #define rice_encode_count(sum, n, k) (((n)*((k)+1))+((sum-(n>>1))>>(k)))
444 * Solve for d/dk(rice_encode_count) = n-((sum-(n>>1))>>(k+1)) = 0
446 static int find_optimal_param(uint32_t sum, int n)
454 k = av_log2(n<256 ? FASTDIV(sum2,n) : sum2/n);
455 return FFMIN(k, MAX_RICE_PARAM);
458 static uint32_t calc_optimal_rice_params(RiceContext *rc, int porder,
459 uint32_t *sums, int n, int pred_order)
465 part = (1 << porder);
468 cnt = (n >> porder) - pred_order;
469 for(i=0; i<part; i++) {
470 k = find_optimal_param(sums[i], cnt);
472 all_bits += rice_encode_count(sums[i], cnt, k);
481 static void calc_sums(int pmin, int pmax, uint32_t *data, int n, int pred_order,
482 uint32_t sums[][MAX_PARTITIONS])
486 uint32_t *res, *res_end;
488 /* sums for highest level */
490 res = &data[pred_order];
491 res_end = &data[n >> pmax];
492 for(i=0; i<parts; i++) {
494 while(res < res_end){
500 /* sums for lower levels */
501 for(i=pmax-1; i>=pmin; i--) {
503 for(j=0; j<parts; j++) {
504 sums[i][j] = sums[i+1][2*j] + sums[i+1][2*j+1];
509 static uint32_t calc_rice_params(RiceContext *rc, int pmin, int pmax,
510 int32_t *data, int n, int pred_order)
513 uint32_t bits[MAX_PARTITION_ORDER+1];
517 uint32_t sums[MAX_PARTITION_ORDER+1][MAX_PARTITIONS];
519 assert(pmin >= 0 && pmin <= MAX_PARTITION_ORDER);
520 assert(pmax >= 0 && pmax <= MAX_PARTITION_ORDER);
521 assert(pmin <= pmax);
523 udata = av_malloc(n * sizeof(uint32_t));
525 udata[i] = (2*data[i]) ^ (data[i]>>31);
528 calc_sums(pmin, pmax, udata, n, pred_order, sums);
531 bits[pmin] = UINT32_MAX;
532 for(i=pmin; i<=pmax; i++) {
533 bits[i] = calc_optimal_rice_params(&tmp_rc, i, sums[i], n, pred_order);
534 if(bits[i] <= bits[opt_porder]) {
541 return bits[opt_porder];
544 static int get_max_p_order(int max_porder, int n, int order)
546 int porder = FFMIN(max_porder, av_log2(n^(n-1)));
548 porder = FFMIN(porder, av_log2(n/order));
552 static uint32_t calc_rice_params_fixed(RiceContext *rc, int pmin, int pmax,
553 int32_t *data, int n, int pred_order,
557 pmin = get_max_p_order(pmin, n, pred_order);
558 pmax = get_max_p_order(pmax, n, pred_order);
559 bits = pred_order*bps + 6;
560 bits += calc_rice_params(rc, pmin, pmax, data, n, pred_order);
564 static uint32_t calc_rice_params_lpc(RiceContext *rc, int pmin, int pmax,
565 int32_t *data, int n, int pred_order,
566 int bps, int precision)
569 pmin = get_max_p_order(pmin, n, pred_order);
570 pmax = get_max_p_order(pmax, n, pred_order);
571 bits = pred_order*bps + 4 + 5 + pred_order*precision + 6;
572 bits += calc_rice_params(rc, pmin, pmax, data, n, pred_order);
577 * Apply Welch window function to audio block
579 static void apply_welch_window(const int32_t *data, int len, double *w_data)
585 assert(!(len&1)); //the optimization in r11881 does not support odd len
586 //if someone wants odd len extend the change in r11881
589 c = 2.0 / (len - 1.0);
593 for(i=0; i<n2; i++) {
596 w_data[-i-1] = data[-i-1] * w;
597 w_data[+i ] = data[+i ] * w;
602 * Calculates autocorrelation data from audio samples
603 * A Welch window function is applied before calculation.
605 void ff_flac_compute_autocorr(const int32_t *data, int len, int lag,
609 double tmp[len + lag + 1];
610 double *data1= tmp + lag;
612 apply_welch_window(data, len, data1);
618 for(j=0; j<lag; j+=2){
619 double sum0 = 1.0, sum1 = 1.0;
620 for(i=0; i<len; i++){
621 sum0 += data1[i] * data1[i-j];
622 sum1 += data1[i] * data1[i-j-1];
630 for(i=0; i<len; i+=2){
631 sum += data1[i ] * data1[i-j ]
632 + data1[i+1] * data1[i-j+1];
639 * Levinson-Durbin recursion.
640 * Produces LPC coefficients from autocorrelation data.
642 static void compute_lpc_coefs(const double *autoc, int max_order,
643 double lpc[][MAX_LPC_ORDER], double *ref)
647 double lpc_tmp[MAX_LPC_ORDER];
649 for(i=0; i<max_order; i++) lpc_tmp[i] = 0;
652 for(i=0; i<max_order; i++) {
655 r -= lpc_tmp[j] * autoc[i-j];
660 err *= 1.0 - (r * r);
664 for(j=0; j<i2; j++) {
666 lpc_tmp[j] += r * lpc_tmp[i-1-j];
667 lpc_tmp[i-1-j] += r * tmp;
670 lpc_tmp[j] += lpc_tmp[j] * r;
673 for(j=0; j<=i; j++) {
674 lpc[i][j] = -lpc_tmp[j];
680 * Quantize LPC coefficients
682 static void quantize_lpc_coefs(double *lpc_in, int order, int precision,
683 int32_t *lpc_out, int *shift)
690 /* define maximum levels */
691 qmax = (1 << (precision - 1)) - 1;
693 /* find maximum coefficient value */
695 for(i=0; i<order; i++) {
696 cmax= FFMAX(cmax, fabs(lpc_in[i]));
699 /* if maximum value quantizes to zero, return all zeros */
700 if(cmax * (1 << MAX_LPC_SHIFT) < 1.0) {
702 memset(lpc_out, 0, sizeof(int32_t) * order);
706 /* calculate level shift which scales max coeff to available bits */
708 while((cmax * (1 << sh) > qmax) && (sh > 0)) {
712 /* since negative shift values are unsupported in decoder, scale down
713 coefficients instead */
714 if(sh == 0 && cmax > qmax) {
715 double scale = ((double)qmax) / cmax;
716 for(i=0; i<order; i++) {
721 /* output quantized coefficients and level shift */
723 for(i=0; i<order; i++) {
724 error += lpc_in[i] * (1 << sh);
725 lpc_out[i] = av_clip(lrintf(error), -qmax, qmax);
731 static int estimate_best_order(double *ref, int max_order)
736 for(i=max_order-1; i>=0; i--) {
746 * Calculate LPC coefficients for multiple orders
748 static int lpc_calc_coefs(FlacEncodeContext *s,
749 const int32_t *samples, int blocksize, int max_order,
750 int precision, int32_t coefs[][MAX_LPC_ORDER],
751 int *shift, int use_lpc, int omethod)
753 double autoc[MAX_LPC_ORDER+1];
754 double ref[MAX_LPC_ORDER];
755 double lpc[MAX_LPC_ORDER][MAX_LPC_ORDER];
759 assert(max_order >= MIN_LPC_ORDER && max_order <= MAX_LPC_ORDER);
762 s->dsp.flac_compute_autocorr(samples, blocksize, max_order, autoc);
764 compute_lpc_coefs(autoc, max_order, lpc, ref);
767 double var[MAX_LPC_ORDER+1], weight;
769 for(pass=0; pass<use_lpc-1; pass++){
770 av_init_lls(&m[pass&1], max_order);
773 for(i=max_order; i<blocksize; i++){
774 for(j=0; j<=max_order; j++)
775 var[j]= samples[i-j];
778 double eval, inv, rinv;
779 eval= av_evaluate_lls(&m[(pass-1)&1], var+1, max_order-1);
780 eval= (512>>pass) + fabs(eval - var[0]);
783 for(j=0; j<=max_order; j++)
789 av_update_lls(&m[pass&1], var, 1.0);
791 av_solve_lls(&m[pass&1], 0.001, 0);
794 for(i=0; i<max_order; i++){
795 for(j=0; j<max_order; j++)
796 lpc[i][j]= m[(pass-1)&1].coeff[i][j];
797 ref[i]= sqrt(m[(pass-1)&1].variance[i] / weight) * (blocksize - max_order) / 4000;
799 for(i=max_order-1; i>0; i--)
800 ref[i] = ref[i-1] - ref[i];
802 opt_order = max_order;
804 if(omethod == ORDER_METHOD_EST) {
805 opt_order = estimate_best_order(ref, max_order);
807 quantize_lpc_coefs(lpc[i], i+1, precision, coefs[i], &shift[i]);
809 for(i=0; i<max_order; i++) {
810 quantize_lpc_coefs(lpc[i], i+1, precision, coefs[i], &shift[i]);
818 static void encode_residual_verbatim(int32_t *res, int32_t *smp, int n)
821 memcpy(res, smp, n * sizeof(int32_t));
824 static void encode_residual_fixed(int32_t *res, const int32_t *smp, int n,
829 for(i=0; i<order; i++) {
834 for(i=order; i<n; i++)
837 for(i=order; i<n; i++)
838 res[i]= smp[i] - smp[i-1];
840 int a = smp[order-1] - smp[order-2];
841 for(i=order; i<n; i+=2) {
842 int b = smp[i] - smp[i-1];
844 a = smp[i+1] - smp[i];
848 int a = smp[order-1] - smp[order-2];
849 int c = smp[order-1] - 2*smp[order-2] + smp[order-3];
850 for(i=order; i<n; i+=2) {
851 int b = smp[i] - smp[i-1];
854 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 int e = smp[order-1] - 3*smp[order-2] + 3*smp[order-3] - smp[order-4];
862 for(i=order; i<n; i+=2) {
863 int b = smp[i] - smp[i-1];
867 a = smp[i+1] - smp[i];
876 int c = coefs[(x)-1];\
882 static av_always_inline void encode_residual_lpc_unrolled(
883 int32_t *res, const int32_t *smp, int n,
884 int order, const int32_t *coefs, int shift, int big)
887 for(i=order; i<n; i+=2) {
888 int s = smp[i-order];
937 res[i ] = smp[i ] - (p0 >> shift);
938 res[i+1] = smp[i+1] - (p1 >> shift);
942 static void encode_residual_lpc(int32_t *res, const int32_t *smp, int n,
943 int order, const int32_t *coefs, int shift)
946 for(i=0; i<order; i++) {
950 for(i=order; i<n; i+=2) {
954 for(j=0; j<order; j++) {
960 res[i ] = smp[i ] - (p0 >> shift);
961 res[i+1] = smp[i+1] - (p1 >> shift);
965 case 1: encode_residual_lpc_unrolled(res, smp, n, 1, coefs, shift, 0); break;
966 case 2: encode_residual_lpc_unrolled(res, smp, n, 2, coefs, shift, 0); break;
967 case 3: encode_residual_lpc_unrolled(res, smp, n, 3, coefs, shift, 0); break;
968 case 4: encode_residual_lpc_unrolled(res, smp, n, 4, coefs, shift, 0); break;
969 case 5: encode_residual_lpc_unrolled(res, smp, n, 5, coefs, shift, 0); break;
970 case 6: encode_residual_lpc_unrolled(res, smp, n, 6, coefs, shift, 0); break;
971 case 7: encode_residual_lpc_unrolled(res, smp, n, 7, coefs, shift, 0); break;
972 case 8: encode_residual_lpc_unrolled(res, smp, n, 8, coefs, shift, 0); break;
973 default: encode_residual_lpc_unrolled(res, smp, n, order, coefs, shift, 1); break;
978 static int encode_residual(FlacEncodeContext *ctx, int ch)
981 int min_order, max_order, opt_order, precision, omethod;
982 int min_porder, max_porder;
985 int32_t coefs[MAX_LPC_ORDER][MAX_LPC_ORDER];
986 int shift[MAX_LPC_ORDER];
990 sub = &frame->subframes[ch];
993 n = frame->blocksize;
997 if(smp[i] != smp[0]) break;
1000 sub->type = sub->type_code = FLAC_SUBFRAME_CONSTANT;
1007 sub->type = sub->type_code = FLAC_SUBFRAME_VERBATIM;
1008 encode_residual_verbatim(res, smp, n);
1009 return sub->obits * n;
1012 min_order = ctx->options.min_prediction_order;
1013 max_order = ctx->options.max_prediction_order;
1014 min_porder = ctx->options.min_partition_order;
1015 max_porder = ctx->options.max_partition_order;
1016 precision = ctx->options.lpc_coeff_precision;
1017 omethod = ctx->options.prediction_order_method;
1020 if(!ctx->options.use_lpc || max_order == 0 || (n <= max_order)) {
1021 uint32_t bits[MAX_FIXED_ORDER+1];
1022 if(max_order > MAX_FIXED_ORDER) max_order = MAX_FIXED_ORDER;
1024 bits[0] = UINT32_MAX;
1025 for(i=min_order; i<=max_order; i++) {
1026 encode_residual_fixed(res, smp, n, i);
1027 bits[i] = calc_rice_params_fixed(&sub->rc, min_porder, max_porder, res,
1029 if(bits[i] < bits[opt_order]) {
1033 sub->order = opt_order;
1034 sub->type = FLAC_SUBFRAME_FIXED;
1035 sub->type_code = sub->type | sub->order;
1036 if(sub->order != max_order) {
1037 encode_residual_fixed(res, smp, n, sub->order);
1038 return calc_rice_params_fixed(&sub->rc, min_porder, max_porder, res, n,
1039 sub->order, sub->obits);
1041 return bits[sub->order];
1045 opt_order = lpc_calc_coefs(ctx, smp, n, max_order, precision, coefs, shift, ctx->options.use_lpc, omethod);
1047 if(omethod == ORDER_METHOD_2LEVEL ||
1048 omethod == ORDER_METHOD_4LEVEL ||
1049 omethod == ORDER_METHOD_8LEVEL) {
1050 int levels = 1 << omethod;
1051 uint32_t bits[levels];
1053 int opt_index = levels-1;
1054 opt_order = max_order-1;
1055 bits[opt_index] = UINT32_MAX;
1056 for(i=levels-1; i>=0; i--) {
1057 order = min_order + (((max_order-min_order+1) * (i+1)) / levels)-1;
1058 if(order < 0) order = 0;
1059 encode_residual_lpc(res, smp, n, order+1, coefs[order], shift[order]);
1060 bits[i] = calc_rice_params_lpc(&sub->rc, min_porder, max_porder,
1061 res, n, order+1, sub->obits, precision);
1062 if(bits[i] < bits[opt_index]) {
1068 } else if(omethod == ORDER_METHOD_SEARCH) {
1069 // brute-force optimal order search
1070 uint32_t bits[MAX_LPC_ORDER];
1072 bits[0] = UINT32_MAX;
1073 for(i=min_order-1; i<max_order; i++) {
1074 encode_residual_lpc(res, smp, n, i+1, coefs[i], shift[i]);
1075 bits[i] = calc_rice_params_lpc(&sub->rc, min_porder, max_porder,
1076 res, n, i+1, sub->obits, precision);
1077 if(bits[i] < bits[opt_order]) {
1082 } else if(omethod == ORDER_METHOD_LOG) {
1083 uint32_t bits[MAX_LPC_ORDER];
1086 opt_order= min_order - 1 + (max_order-min_order)/3;
1087 memset(bits, -1, sizeof(bits));
1089 for(step=16 ;step; step>>=1){
1090 int last= opt_order;
1091 for(i=last-step; i<=last+step; i+= step){
1092 if(i<min_order-1 || i>=max_order || bits[i] < UINT32_MAX)
1094 encode_residual_lpc(res, smp, n, i+1, coefs[i], shift[i]);
1095 bits[i] = calc_rice_params_lpc(&sub->rc, min_porder, max_porder,
1096 res, n, i+1, sub->obits, precision);
1097 if(bits[i] < bits[opt_order])
1104 sub->order = opt_order;
1105 sub->type = FLAC_SUBFRAME_LPC;
1106 sub->type_code = sub->type | (sub->order-1);
1107 sub->shift = shift[sub->order-1];
1108 for(i=0; i<sub->order; i++) {
1109 sub->coefs[i] = coefs[sub->order-1][i];
1111 encode_residual_lpc(res, smp, n, sub->order, sub->coefs, sub->shift);
1112 return calc_rice_params_lpc(&sub->rc, min_porder, max_porder, res, n, sub->order,
1113 sub->obits, precision);
1116 static int encode_residual_v(FlacEncodeContext *ctx, int ch)
1123 frame = &ctx->frame;
1124 sub = &frame->subframes[ch];
1125 res = sub->residual;
1127 n = frame->blocksize;
1130 for(i=1; i<n; i++) {
1131 if(smp[i] != smp[0]) break;
1134 sub->type = sub->type_code = FLAC_SUBFRAME_CONSTANT;
1140 sub->type = sub->type_code = FLAC_SUBFRAME_VERBATIM;
1141 encode_residual_verbatim(res, smp, n);
1142 return sub->obits * n;
1145 static int estimate_stereo_mode(int32_t *left_ch, int32_t *right_ch, int n)
1153 /* calculate sum of 2nd order residual for each channel */
1154 sum[0] = sum[1] = sum[2] = sum[3] = 0;
1155 for(i=2; i<n; i++) {
1156 lt = left_ch[i] - 2*left_ch[i-1] + left_ch[i-2];
1157 rt = right_ch[i] - 2*right_ch[i-1] + right_ch[i-2];
1158 sum[2] += FFABS((lt + rt) >> 1);
1159 sum[3] += FFABS(lt - rt);
1160 sum[0] += FFABS(lt);
1161 sum[1] += FFABS(rt);
1163 /* estimate bit counts */
1164 for(i=0; i<4; i++) {
1165 k = find_optimal_param(2*sum[i], n);
1166 sum[i] = rice_encode_count(2*sum[i], n, k);
1169 /* calculate score for each mode */
1170 score[0] = sum[0] + sum[1];
1171 score[1] = sum[0] + sum[3];
1172 score[2] = sum[1] + sum[3];
1173 score[3] = sum[2] + sum[3];
1175 /* return mode with lowest score */
1177 for(i=1; i<4; i++) {
1178 if(score[i] < score[best]) {
1183 return FLAC_CHMODE_LEFT_RIGHT;
1184 } else if(best == 1) {
1185 return FLAC_CHMODE_LEFT_SIDE;
1186 } else if(best == 2) {
1187 return FLAC_CHMODE_RIGHT_SIDE;
1189 return FLAC_CHMODE_MID_SIDE;
1194 * Perform stereo channel decorrelation
1196 static void channel_decorrelation(FlacEncodeContext *ctx)
1199 int32_t *left, *right;
1202 frame = &ctx->frame;
1203 n = frame->blocksize;
1204 left = frame->subframes[0].samples;
1205 right = frame->subframes[1].samples;
1207 if(ctx->channels != 2) {
1208 frame->ch_mode = FLAC_CHMODE_NOT_STEREO;
1212 frame->ch_mode = estimate_stereo_mode(left, right, n);
1214 /* perform decorrelation and adjust bits-per-sample */
1215 if(frame->ch_mode == FLAC_CHMODE_LEFT_RIGHT) {
1218 if(frame->ch_mode == FLAC_CHMODE_MID_SIDE) {
1220 for(i=0; i<n; i++) {
1222 left[i] = (tmp + right[i]) >> 1;
1223 right[i] = tmp - right[i];
1225 frame->subframes[1].obits++;
1226 } else if(frame->ch_mode == FLAC_CHMODE_LEFT_SIDE) {
1227 for(i=0; i<n; i++) {
1228 right[i] = left[i] - right[i];
1230 frame->subframes[1].obits++;
1232 for(i=0; i<n; i++) {
1233 left[i] -= right[i];
1235 frame->subframes[0].obits++;
1239 static void write_utf8(PutBitContext *pb, uint32_t val)
1242 PUT_UTF8(val, tmp, put_bits(pb, 8, tmp);)
1245 static void output_frame_header(FlacEncodeContext *s)
1252 put_bits(&s->pb, 16, 0xFFF8);
1253 put_bits(&s->pb, 4, frame->bs_code[0]);
1254 put_bits(&s->pb, 4, s->sr_code[0]);
1255 if(frame->ch_mode == FLAC_CHMODE_NOT_STEREO) {
1256 put_bits(&s->pb, 4, s->ch_code);
1258 put_bits(&s->pb, 4, frame->ch_mode);
1260 put_bits(&s->pb, 3, 4); /* bits-per-sample code */
1261 put_bits(&s->pb, 1, 0);
1262 write_utf8(&s->pb, s->frame_count);
1263 if(frame->bs_code[0] == 6) {
1264 put_bits(&s->pb, 8, frame->bs_code[1]);
1265 } else if(frame->bs_code[0] == 7) {
1266 put_bits(&s->pb, 16, frame->bs_code[1]);
1268 if(s->sr_code[0] == 12) {
1269 put_bits(&s->pb, 8, s->sr_code[1]);
1270 } else if(s->sr_code[0] > 12) {
1271 put_bits(&s->pb, 16, s->sr_code[1]);
1273 flush_put_bits(&s->pb);
1274 crc = av_crc(av_crc_get_table(AV_CRC_8_ATM), 0,
1275 s->pb.buf, put_bits_count(&s->pb)>>3);
1276 put_bits(&s->pb, 8, crc);
1279 static void output_subframe_constant(FlacEncodeContext *s, int ch)
1284 sub = &s->frame.subframes[ch];
1285 res = sub->residual[0];
1286 put_sbits(&s->pb, sub->obits, res);
1289 static void output_subframe_verbatim(FlacEncodeContext *s, int ch)
1297 sub = &frame->subframes[ch];
1299 for(i=0; i<frame->blocksize; i++) {
1300 res = sub->residual[i];
1301 put_sbits(&s->pb, sub->obits, res);
1305 static void output_residual(FlacEncodeContext *ctx, int ch)
1307 int i, j, p, n, parts;
1308 int k, porder, psize, res_cnt;
1313 frame = &ctx->frame;
1314 sub = &frame->subframes[ch];
1315 res = sub->residual;
1316 n = frame->blocksize;
1318 /* rice-encoded block */
1319 put_bits(&ctx->pb, 2, 0);
1321 /* partition order */
1322 porder = sub->rc.porder;
1323 psize = n >> porder;
1324 parts = (1 << porder);
1325 put_bits(&ctx->pb, 4, porder);
1326 res_cnt = psize - sub->order;
1330 for(p=0; p<parts; p++) {
1331 k = sub->rc.params[p];
1332 put_bits(&ctx->pb, 4, k);
1333 if(p == 1) res_cnt = psize;
1334 for(i=0; i<res_cnt && j<n; i++, j++) {
1335 set_sr_golomb_flac(&ctx->pb, res[j], k, INT32_MAX, 0);
1340 static void output_subframe_fixed(FlacEncodeContext *ctx, int ch)
1346 frame = &ctx->frame;
1347 sub = &frame->subframes[ch];
1349 /* warm-up samples */
1350 for(i=0; i<sub->order; i++) {
1351 put_sbits(&ctx->pb, sub->obits, sub->residual[i]);
1355 output_residual(ctx, ch);
1358 static void output_subframe_lpc(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]);
1372 /* LPC coefficients */
1373 cbits = ctx->options.lpc_coeff_precision;
1374 put_bits(&ctx->pb, 4, cbits-1);
1375 put_sbits(&ctx->pb, 5, sub->shift);
1376 for(i=0; i<sub->order; i++) {
1377 put_sbits(&ctx->pb, cbits, sub->coefs[i]);
1381 output_residual(ctx, ch);
1384 static void output_subframes(FlacEncodeContext *s)
1392 for(ch=0; ch<s->channels; ch++) {
1393 sub = &frame->subframes[ch];
1395 /* subframe header */
1396 put_bits(&s->pb, 1, 0);
1397 put_bits(&s->pb, 6, sub->type_code);
1398 put_bits(&s->pb, 1, 0); /* no wasted bits */
1401 if(sub->type == FLAC_SUBFRAME_CONSTANT) {
1402 output_subframe_constant(s, ch);
1403 } else if(sub->type == FLAC_SUBFRAME_VERBATIM) {
1404 output_subframe_verbatim(s, ch);
1405 } else if(sub->type == FLAC_SUBFRAME_FIXED) {
1406 output_subframe_fixed(s, ch);
1407 } else if(sub->type == FLAC_SUBFRAME_LPC) {
1408 output_subframe_lpc(s, ch);
1413 static void output_frame_footer(FlacEncodeContext *s)
1416 flush_put_bits(&s->pb);
1417 crc = bswap_16(av_crc(av_crc_get_table(AV_CRC_16_ANSI), 0,
1418 s->pb.buf, put_bits_count(&s->pb)>>3));
1419 put_bits(&s->pb, 16, crc);
1420 flush_put_bits(&s->pb);
1423 static int flac_encode_frame(AVCodecContext *avctx, uint8_t *frame,
1424 int buf_size, void *data)
1427 FlacEncodeContext *s;
1428 int16_t *samples = data;
1431 s = avctx->priv_data;
1435 copy_samples(s, samples);
1437 channel_decorrelation(s);
1439 for(ch=0; ch<s->channels; ch++) {
1440 encode_residual(s, ch);
1442 init_put_bits(&s->pb, frame, buf_size);
1443 output_frame_header(s);
1444 output_subframes(s);
1445 output_frame_footer(s);
1446 out_bytes = put_bits_count(&s->pb) >> 3;
1448 if(out_bytes > s->max_framesize || out_bytes >= buf_size) {
1449 /* frame too large. use verbatim mode */
1450 for(ch=0; ch<s->channels; ch++) {
1451 encode_residual_v(s, ch);
1453 init_put_bits(&s->pb, frame, buf_size);
1454 output_frame_header(s);
1455 output_subframes(s);
1456 output_frame_footer(s);
1457 out_bytes = put_bits_count(&s->pb) >> 3;
1459 if(out_bytes > s->max_framesize || out_bytes >= buf_size) {
1460 /* still too large. must be an error. */
1461 av_log(avctx, AV_LOG_ERROR, "error encoding frame\n");
1470 static av_cold int flac_encode_close(AVCodecContext *avctx)
1472 av_freep(&avctx->extradata);
1473 avctx->extradata_size = 0;
1474 av_freep(&avctx->coded_frame);
1478 AVCodec flac_encoder = {
1482 sizeof(FlacEncodeContext),
1487 .capabilities = CODEC_CAP_SMALL_LAST_FRAME,
1488 .long_name = NULL_IF_CONFIG_SMALL("FLAC (Free Lossless Audio Codec)"),