2 * Real Audio 1.0 (14.4K) encoder
3 * Copyright (c) 2010 Francesco Lavra <francescolavra@interfree.it>
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
24 * Real Audio 1.0 (14.4K) encoder
25 * @author Francesco Lavra <francescolavra@interfree.it>
31 #include "audio_frame_queue.h"
34 #include "celp_filters.h"
38 static av_cold int ra144_encode_close(AVCodecContext *avctx)
40 RA144Context *ractx = avctx->priv_data;
41 ff_lpc_end(&ractx->lpc_ctx);
42 ff_af_queue_close(&ractx->afq);
43 #if FF_API_OLD_ENCODE_AUDIO
44 av_freep(&avctx->coded_frame);
50 static av_cold int ra144_encode_init(AVCodecContext * avctx)
55 if (avctx->channels != 1) {
56 av_log(avctx, AV_LOG_ERROR, "invalid number of channels: %d\n",
60 avctx->frame_size = NBLOCKS * BLOCKSIZE;
61 avctx->delay = avctx->frame_size;
62 avctx->bit_rate = 8000;
63 ractx = avctx->priv_data;
64 ractx->lpc_coef[0] = ractx->lpc_tables[0];
65 ractx->lpc_coef[1] = ractx->lpc_tables[1];
67 ret = ff_lpc_init(&ractx->lpc_ctx, avctx->frame_size, LPC_ORDER,
68 FF_LPC_TYPE_LEVINSON);
72 ff_af_queue_init(avctx, &ractx->afq);
74 #if FF_API_OLD_ENCODE_AUDIO
75 avctx->coded_frame = avcodec_alloc_frame();
76 if (!avctx->coded_frame) {
77 ret = AVERROR(ENOMEM);
84 ra144_encode_close(avctx);
90 * Quantize a value by searching a sorted table for the element with the
93 * @param value value to quantize
94 * @param table array containing the quantization table
95 * @param size size of the quantization table
96 * @return index of the quantization table corresponding to the element with the
99 static int quantize(int value, const int16_t *table, unsigned int size)
101 unsigned int low = 0, high = size - 1;
104 int index = (low + high) >> 1;
105 int error = table[index] - value;
108 return table[high] + error > value ? low : high;
119 * Orthogonalize a vector to another vector
121 * @param v vector to orthogonalize
122 * @param u vector against which orthogonalization is performed
124 static void orthogonalize(float *v, const float *u)
127 float num = 0, den = 0;
129 for (i = 0; i < BLOCKSIZE; i++) {
134 for (i = 0; i < BLOCKSIZE; i++)
140 * Calculate match score and gain of an LPC-filtered vector with respect to
141 * input data, possibly othogonalizing it to up to 2 other vectors
143 * @param work array used to calculate the filtered vector
144 * @param coefs coefficients of the LPC filter
145 * @param vect original vector
146 * @param ortho1 first vector against which orthogonalization is performed
147 * @param ortho2 second vector against which orthogonalization is performed
148 * @param data input data
149 * @param score pointer to variable where match score is returned
150 * @param gain pointer to variable where gain is returned
152 static void get_match_score(float *work, const float *coefs, float *vect,
153 const float *ortho1, const float *ortho2,
154 const float *data, float *score, float *gain)
159 ff_celp_lp_synthesis_filterf(work, coefs, vect, BLOCKSIZE, LPC_ORDER);
161 orthogonalize(work, ortho1);
163 orthogonalize(work, ortho2);
165 for (i = 0; i < BLOCKSIZE; i++) {
166 g += work[i] * work[i];
167 c += data[i] * work[i];
179 * Create a vector from the adaptive codebook at a given lag value
181 * @param vect array where vector is stored
182 * @param cb adaptive codebook
183 * @param lag lag value
185 static void create_adapt_vect(float *vect, const int16_t *cb, int lag)
189 cb += BUFFERSIZE - lag;
190 for (i = 0; i < FFMIN(BLOCKSIZE, lag); i++)
193 for (i = 0; i < BLOCKSIZE - lag; i++)
194 vect[lag + i] = cb[i];
199 * Search the adaptive codebook for the best entry and gain and remove its
200 * contribution from input data
202 * @param adapt_cb array from which the adaptive codebook is extracted
203 * @param work array used to calculate LPC-filtered vectors
204 * @param coefs coefficients of the LPC filter
205 * @param data input data
206 * @return index of the best entry of the adaptive codebook
208 static int adaptive_cb_search(const int16_t *adapt_cb, float *work,
209 const float *coefs, float *data)
212 float score, gain, best_score, best_gain;
213 float exc[BLOCKSIZE];
215 gain = best_score = 0;
216 for (i = BLOCKSIZE / 2; i <= BUFFERSIZE; i++) {
217 create_adapt_vect(exc, adapt_cb, i);
218 get_match_score(work, coefs, exc, NULL, NULL, data, &score, &gain);
219 if (score > best_score) {
229 * Re-calculate the filtered vector from the vector with maximum match score
230 * and remove its contribution from input data.
232 create_adapt_vect(exc, adapt_cb, best_vect);
233 ff_celp_lp_synthesis_filterf(work, coefs, exc, BLOCKSIZE, LPC_ORDER);
234 for (i = 0; i < BLOCKSIZE; i++)
235 data[i] -= best_gain * work[i];
236 return best_vect - BLOCKSIZE / 2 + 1;
241 * Find the best vector of a fixed codebook by applying an LPC filter to
242 * codebook entries, possibly othogonalizing them to up to 2 other vectors and
243 * matching the results with input data
245 * @param work array used to calculate the filtered vectors
246 * @param coefs coefficients of the LPC filter
247 * @param cb fixed codebook
248 * @param ortho1 first vector against which orthogonalization is performed
249 * @param ortho2 second vector against which orthogonalization is performed
250 * @param data input data
251 * @param idx pointer to variable where the index of the best codebook entry is
253 * @param gain pointer to variable where the gain of the best codebook entry is
256 static void find_best_vect(float *work, const float *coefs,
257 const int8_t cb[][BLOCKSIZE], const float *ortho1,
258 const float *ortho2, float *data, int *idx,
262 float g, score, best_score;
263 float vect[BLOCKSIZE];
265 *idx = *gain = best_score = 0;
266 for (i = 0; i < FIXED_CB_SIZE; i++) {
267 for (j = 0; j < BLOCKSIZE; j++)
269 get_match_score(work, coefs, vect, ortho1, ortho2, data, &score, &g);
270 if (score > best_score) {
280 * Search the two fixed codebooks for the best entry and gain
282 * @param work array used to calculate LPC-filtered vectors
283 * @param coefs coefficients of the LPC filter
284 * @param data input data
285 * @param cba_idx index of the best entry of the adaptive codebook
286 * @param cb1_idx pointer to variable where the index of the best entry of the
287 * first fixed codebook is returned
288 * @param cb2_idx pointer to variable where the index of the best entry of the
289 * second fixed codebook is returned
291 static void fixed_cb_search(float *work, const float *coefs, float *data,
292 int cba_idx, int *cb1_idx, int *cb2_idx)
296 float cba_vect[BLOCKSIZE], cb1_vect[BLOCKSIZE];
297 float vect[BLOCKSIZE];
300 * The filtered vector from the adaptive codebook can be retrieved from
301 * work, because this function is called just after adaptive_cb_search().
304 memcpy(cba_vect, work, sizeof(cba_vect));
306 find_best_vect(work, coefs, ff_cb1_vects, cba_idx ? cba_vect : NULL, NULL,
307 data, cb1_idx, &gain);
310 * Re-calculate the filtered vector from the vector with maximum match score
311 * and remove its contribution from input data.
314 for (i = 0; i < BLOCKSIZE; i++)
315 vect[i] = ff_cb1_vects[*cb1_idx][i];
316 ff_celp_lp_synthesis_filterf(work, coefs, vect, BLOCKSIZE, LPC_ORDER);
318 orthogonalize(work, cba_vect);
319 for (i = 0; i < BLOCKSIZE; i++)
320 data[i] -= gain * work[i];
321 memcpy(cb1_vect, work, sizeof(cb1_vect));
326 find_best_vect(work, coefs, ff_cb2_vects, cba_idx ? cba_vect : NULL,
327 ortho_cb1 ? cb1_vect : NULL, data, cb2_idx, &gain);
332 * Encode a subblock of the current frame
334 * @param ractx encoder context
335 * @param sblock_data input data of the subblock
336 * @param lpc_coefs coefficients of the LPC filter
337 * @param rms RMS of the reflection coefficients
338 * @param pb pointer to PutBitContext of the current frame
340 static void ra144_encode_subblock(RA144Context *ractx,
341 const int16_t *sblock_data,
342 const int16_t *lpc_coefs, unsigned int rms,
345 float data[BLOCKSIZE] = { 0 }, work[LPC_ORDER + BLOCKSIZE];
346 float coefs[LPC_ORDER];
347 float zero[BLOCKSIZE], cba[BLOCKSIZE], cb1[BLOCKSIZE], cb2[BLOCKSIZE];
348 int16_t cba_vect[BLOCKSIZE];
349 int cba_idx, cb1_idx, cb2_idx, gain;
352 float error, best_error;
354 for (i = 0; i < LPC_ORDER; i++) {
355 work[i] = ractx->curr_sblock[BLOCKSIZE + i];
356 coefs[i] = lpc_coefs[i] * (1/4096.0);
360 * Calculate the zero-input response of the LPC filter and subtract it from
363 ff_celp_lp_synthesis_filterf(work + LPC_ORDER, coefs, data, BLOCKSIZE,
365 for (i = 0; i < BLOCKSIZE; i++) {
366 zero[i] = work[LPC_ORDER + i];
367 data[i] = sblock_data[i] - zero[i];
371 * Codebook search is performed without taking into account the contribution
372 * of the previous subblock, since it has been just subtracted from input
375 memset(work, 0, LPC_ORDER * sizeof(*work));
377 cba_idx = adaptive_cb_search(ractx->adapt_cb, work + LPC_ORDER, coefs,
381 * The filtered vector from the adaptive codebook can be retrieved from
382 * work, see implementation of adaptive_cb_search().
384 memcpy(cba, work + LPC_ORDER, sizeof(cba));
386 ff_copy_and_dup(cba_vect, ractx->adapt_cb, cba_idx + BLOCKSIZE / 2 - 1);
387 m[0] = (ff_irms(cba_vect) * rms) >> 12;
389 fixed_cb_search(work + LPC_ORDER, coefs, data, cba_idx, &cb1_idx, &cb2_idx);
390 for (i = 0; i < BLOCKSIZE; i++) {
391 cb1[i] = ff_cb1_vects[cb1_idx][i];
392 cb2[i] = ff_cb2_vects[cb2_idx][i];
394 ff_celp_lp_synthesis_filterf(work + LPC_ORDER, coefs, cb1, BLOCKSIZE,
396 memcpy(cb1, work + LPC_ORDER, sizeof(cb1));
397 m[1] = (ff_cb1_base[cb1_idx] * rms) >> 8;
398 ff_celp_lp_synthesis_filterf(work + LPC_ORDER, coefs, cb2, BLOCKSIZE,
400 memcpy(cb2, work + LPC_ORDER, sizeof(cb2));
401 m[2] = (ff_cb2_base[cb2_idx] * rms) >> 8;
402 best_error = FLT_MAX;
404 for (n = 0; n < 256; n++) {
405 g[1] = ((ff_gain_val_tab[n][1] * m[1]) >> ff_gain_exp_tab[n]) *
407 g[2] = ((ff_gain_val_tab[n][2] * m[2]) >> ff_gain_exp_tab[n]) *
411 g[0] = ((ff_gain_val_tab[n][0] * m[0]) >> ff_gain_exp_tab[n]) *
413 for (i = 0; i < BLOCKSIZE; i++) {
414 data[i] = zero[i] + g[0] * cba[i] + g[1] * cb1[i] +
416 error += (data[i] - sblock_data[i]) *
417 (data[i] - sblock_data[i]);
420 for (i = 0; i < BLOCKSIZE; i++) {
421 data[i] = zero[i] + g[1] * cb1[i] + g[2] * cb2[i];
422 error += (data[i] - sblock_data[i]) *
423 (data[i] - sblock_data[i]);
426 if (error < best_error) {
431 put_bits(pb, 7, cba_idx);
432 put_bits(pb, 8, gain);
433 put_bits(pb, 7, cb1_idx);
434 put_bits(pb, 7, cb2_idx);
435 ff_subblock_synthesis(ractx, lpc_coefs, cba_idx, cb1_idx, cb2_idx, rms,
440 static int ra144_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
441 const AVFrame *frame, int *got_packet_ptr)
443 static const uint8_t sizes[LPC_ORDER] = {64, 32, 32, 16, 16, 8, 8, 8, 8, 4};
444 static const uint8_t bit_sizes[LPC_ORDER] = {6, 5, 5, 4, 4, 3, 3, 3, 3, 2};
445 RA144Context *ractx = avctx->priv_data;
447 int32_t lpc_data[NBLOCKS * BLOCKSIZE];
448 int32_t lpc_coefs[LPC_ORDER][MAX_LPC_ORDER];
449 int shift[LPC_ORDER];
450 int16_t block_coefs[NBLOCKS][LPC_ORDER];
451 int lpc_refl[LPC_ORDER]; /**< reflection coefficients of the frame */
452 unsigned int refl_rms[NBLOCKS]; /**< RMS of the reflection coefficients */
453 const int16_t *samples = frame ? (const int16_t *)frame->data[0] : NULL;
457 if (ractx->last_frame)
460 if ((ret = ff_alloc_packet2(avctx, avpkt, FRAMESIZE)))
464 * Since the LPC coefficients are calculated on a frame centered over the
465 * fourth subframe, to encode a given frame, data from the next frame is
466 * needed. In each call to this function, the previous frame (whose data are
467 * saved in the encoder context) is encoded, and data from the current frame
468 * are saved in the encoder context to be used in the next function call.
470 for (i = 0; i < (2 * BLOCKSIZE + BLOCKSIZE / 2); i++) {
471 lpc_data[i] = ractx->curr_block[BLOCKSIZE + BLOCKSIZE / 2 + i];
472 energy += (lpc_data[i] * lpc_data[i]) >> 4;
476 for (j = 0; j < frame->nb_samples && i < NBLOCKS * BLOCKSIZE; i++, j++) {
477 lpc_data[i] = samples[j] >> 2;
478 energy += (lpc_data[i] * lpc_data[i]) >> 4;
481 if (i < NBLOCKS * BLOCKSIZE)
482 memset(&lpc_data[i], 0, (NBLOCKS * BLOCKSIZE - i) * sizeof(*lpc_data));
483 energy = ff_energy_tab[quantize(ff_t_sqrt(energy >> 5) >> 10, ff_energy_tab,
486 ff_lpc_calc_coefs(&ractx->lpc_ctx, lpc_data, NBLOCKS * BLOCKSIZE, LPC_ORDER,
487 LPC_ORDER, 16, lpc_coefs, shift, FF_LPC_TYPE_LEVINSON,
488 0, ORDER_METHOD_EST, 12, 0);
489 for (i = 0; i < LPC_ORDER; i++)
490 block_coefs[NBLOCKS - 1][i] = -(lpc_coefs[LPC_ORDER - 1][i] <<
491 (12 - shift[LPC_ORDER - 1]));
494 * TODO: apply perceptual weighting of the input speech through bandwidth
495 * expansion of the LPC filter.
498 if (ff_eval_refl(lpc_refl, block_coefs[NBLOCKS - 1], avctx)) {
500 * The filter is unstable: use the coefficients of the previous frame.
502 ff_int_to_int16(block_coefs[NBLOCKS - 1], ractx->lpc_coef[1]);
503 if (ff_eval_refl(lpc_refl, block_coefs[NBLOCKS - 1], avctx)) {
504 /* the filter is still unstable. set reflection coeffs to zero. */
505 memset(lpc_refl, 0, sizeof(lpc_refl));
508 init_put_bits(&pb, avpkt->data, avpkt->size);
509 for (i = 0; i < LPC_ORDER; i++) {
510 idx = quantize(lpc_refl[i], ff_lpc_refl_cb[i], sizes[i]);
511 put_bits(&pb, bit_sizes[i], idx);
512 lpc_refl[i] = ff_lpc_refl_cb[i][idx];
514 ractx->lpc_refl_rms[0] = ff_rms(lpc_refl);
515 ff_eval_coefs(ractx->lpc_coef[0], lpc_refl);
516 refl_rms[0] = ff_interp(ractx, block_coefs[0], 1, 1, ractx->old_energy);
517 refl_rms[1] = ff_interp(ractx, block_coefs[1], 2,
518 energy <= ractx->old_energy,
519 ff_t_sqrt(energy * ractx->old_energy) >> 12);
520 refl_rms[2] = ff_interp(ractx, block_coefs[2], 3, 0, energy);
521 refl_rms[3] = ff_rescale_rms(ractx->lpc_refl_rms[0], energy);
522 ff_int_to_int16(block_coefs[NBLOCKS - 1], ractx->lpc_coef[0]);
523 put_bits(&pb, 5, quantize(energy, ff_energy_tab, 32));
524 for (i = 0; i < NBLOCKS; i++)
525 ra144_encode_subblock(ractx, ractx->curr_block + i * BLOCKSIZE,
526 block_coefs[i], refl_rms[i], &pb);
528 ractx->old_energy = energy;
529 ractx->lpc_refl_rms[1] = ractx->lpc_refl_rms[0];
530 FFSWAP(unsigned int *, ractx->lpc_coef[0], ractx->lpc_coef[1]);
532 /* copy input samples to current block for processing in next call */
535 for (; i < frame->nb_samples; i++)
536 ractx->curr_block[i] = samples[i] >> 2;
538 if ((ret = ff_af_queue_add(&ractx->afq, frame) < 0))
541 ractx->last_frame = 1;
542 memset(&ractx->curr_block[i], 0,
543 (NBLOCKS * BLOCKSIZE - i) * sizeof(*ractx->curr_block));
545 /* Get the next frame pts/duration */
546 ff_af_queue_remove(&ractx->afq, avctx->frame_size, &avpkt->pts,
549 avpkt->size = FRAMESIZE;
555 AVCodec ff_ra_144_encoder = {
557 .type = AVMEDIA_TYPE_AUDIO,
558 .id = CODEC_ID_RA_144,
559 .priv_data_size = sizeof(RA144Context),
560 .init = ra144_encode_init,
561 .encode2 = ra144_encode_frame,
562 .close = ra144_encode_close,
563 .capabilities = CODEC_CAP_DELAY | CODEC_CAP_SMALL_LAST_FRAME,
564 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_S16,
565 AV_SAMPLE_FMT_NONE },
566 .long_name = NULL_IF_CONFIG_SMALL("RealAudio 1.0 (14.4K)"),