2 * Real Audio 1.0 (14.4K) encoder
3 * Copyright (c) 2010 Francesco Lavra <francescolavra@interfree.it>
5 * This file is part of Libav.
7 * Libav 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 * Libav 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 Libav; 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);
47 static av_cold int ra144_encode_init(AVCodecContext * avctx)
52 if (avctx->channels != 1) {
53 av_log(avctx, AV_LOG_ERROR, "invalid number of channels: %d\n",
57 avctx->frame_size = NBLOCKS * BLOCKSIZE;
58 avctx->delay = avctx->frame_size;
59 avctx->bit_rate = 8000;
60 ractx = avctx->priv_data;
61 ractx->lpc_coef[0] = ractx->lpc_tables[0];
62 ractx->lpc_coef[1] = ractx->lpc_tables[1];
64 ret = ff_lpc_init(&ractx->lpc_ctx, avctx->frame_size, LPC_ORDER,
65 FF_LPC_TYPE_LEVINSON);
69 ff_af_queue_init(avctx, &ractx->afq);
73 ra144_encode_close(avctx);
79 * Quantize a value by searching a sorted table for the element with the
82 * @param value value to quantize
83 * @param table array containing the quantization table
84 * @param size size of the quantization table
85 * @return index of the quantization table corresponding to the element with the
88 static int quantize(int value, const int16_t *table, unsigned int size)
90 unsigned int low = 0, high = size - 1;
93 int index = (low + high) >> 1;
94 int error = table[index] - value;
97 return table[high] + error > value ? low : high;
108 * Orthogonalize a vector to another vector
110 * @param v vector to orthogonalize
111 * @param u vector against which orthogonalization is performed
113 static void orthogonalize(float *v, const float *u)
116 float num = 0, den = 0;
118 for (i = 0; i < BLOCKSIZE; i++) {
123 for (i = 0; i < BLOCKSIZE; i++)
129 * Calculate match score and gain of an LPC-filtered vector with respect to
130 * input data, possibly othogonalizing it to up to 2 other vectors
132 * @param work array used to calculate the filtered vector
133 * @param coefs coefficients of the LPC filter
134 * @param vect original vector
135 * @param ortho1 first vector against which orthogonalization is performed
136 * @param ortho2 second vector against which orthogonalization is performed
137 * @param data input data
138 * @param score pointer to variable where match score is returned
139 * @param gain pointer to variable where gain is returned
141 static void get_match_score(float *work, const float *coefs, float *vect,
142 const float *ortho1, const float *ortho2,
143 const float *data, float *score, float *gain)
148 ff_celp_lp_synthesis_filterf(work, coefs, vect, BLOCKSIZE, LPC_ORDER);
150 orthogonalize(work, ortho1);
152 orthogonalize(work, ortho2);
154 for (i = 0; i < BLOCKSIZE; i++) {
155 g += work[i] * work[i];
156 c += data[i] * work[i];
168 * Create a vector from the adaptive codebook at a given lag value
170 * @param vect array where vector is stored
171 * @param cb adaptive codebook
172 * @param lag lag value
174 static void create_adapt_vect(float *vect, const int16_t *cb, int lag)
178 cb += BUFFERSIZE - lag;
179 for (i = 0; i < FFMIN(BLOCKSIZE, lag); i++)
182 for (i = 0; i < BLOCKSIZE - lag; i++)
183 vect[lag + i] = cb[i];
188 * Search the adaptive codebook for the best entry and gain and remove its
189 * contribution from input data
191 * @param adapt_cb array from which the adaptive codebook is extracted
192 * @param work array used to calculate LPC-filtered vectors
193 * @param coefs coefficients of the LPC filter
194 * @param data input data
195 * @return index of the best entry of the adaptive codebook
197 static int adaptive_cb_search(const int16_t *adapt_cb, float *work,
198 const float *coefs, float *data)
201 float score, gain, best_score, best_gain;
202 float exc[BLOCKSIZE];
204 gain = best_score = 0;
205 for (i = BLOCKSIZE / 2; i <= BUFFERSIZE; i++) {
206 create_adapt_vect(exc, adapt_cb, i);
207 get_match_score(work, coefs, exc, NULL, NULL, data, &score, &gain);
208 if (score > best_score) {
218 * Re-calculate the filtered vector from the vector with maximum match score
219 * and remove its contribution from input data.
221 create_adapt_vect(exc, adapt_cb, best_vect);
222 ff_celp_lp_synthesis_filterf(work, coefs, exc, BLOCKSIZE, LPC_ORDER);
223 for (i = 0; i < BLOCKSIZE; i++)
224 data[i] -= best_gain * work[i];
225 return best_vect - BLOCKSIZE / 2 + 1;
230 * Find the best vector of a fixed codebook by applying an LPC filter to
231 * codebook entries, possibly othogonalizing them to up to 2 other vectors and
232 * matching the results with input data
234 * @param work array used to calculate the filtered vectors
235 * @param coefs coefficients of the LPC filter
236 * @param cb fixed codebook
237 * @param ortho1 first vector against which orthogonalization is performed
238 * @param ortho2 second vector against which orthogonalization is performed
239 * @param data input data
240 * @param idx pointer to variable where the index of the best codebook entry is
242 * @param gain pointer to variable where the gain of the best codebook entry is
245 static void find_best_vect(float *work, const float *coefs,
246 const int8_t cb[][BLOCKSIZE], const float *ortho1,
247 const float *ortho2, float *data, int *idx,
251 float g, score, best_score;
252 float vect[BLOCKSIZE];
254 *idx = *gain = best_score = 0;
255 for (i = 0; i < FIXED_CB_SIZE; i++) {
256 for (j = 0; j < BLOCKSIZE; j++)
258 get_match_score(work, coefs, vect, ortho1, ortho2, data, &score, &g);
259 if (score > best_score) {
269 * Search the two fixed codebooks for the best entry and gain
271 * @param work array used to calculate LPC-filtered vectors
272 * @param coefs coefficients of the LPC filter
273 * @param data input data
274 * @param cba_idx index of the best entry of the adaptive codebook
275 * @param cb1_idx pointer to variable where the index of the best entry of the
276 * first fixed codebook is returned
277 * @param cb2_idx pointer to variable where the index of the best entry of the
278 * second fixed codebook is returned
280 static void fixed_cb_search(float *work, const float *coefs, float *data,
281 int cba_idx, int *cb1_idx, int *cb2_idx)
285 float cba_vect[BLOCKSIZE], cb1_vect[BLOCKSIZE];
286 float vect[BLOCKSIZE];
289 * The filtered vector from the adaptive codebook can be retrieved from
290 * work, because this function is called just after adaptive_cb_search().
293 memcpy(cba_vect, work, sizeof(cba_vect));
295 find_best_vect(work, coefs, ff_cb1_vects, cba_idx ? cba_vect : NULL, NULL,
296 data, cb1_idx, &gain);
299 * Re-calculate the filtered vector from the vector with maximum match score
300 * and remove its contribution from input data.
303 for (i = 0; i < BLOCKSIZE; i++)
304 vect[i] = ff_cb1_vects[*cb1_idx][i];
305 ff_celp_lp_synthesis_filterf(work, coefs, vect, BLOCKSIZE, LPC_ORDER);
307 orthogonalize(work, cba_vect);
308 for (i = 0; i < BLOCKSIZE; i++)
309 data[i] -= gain * work[i];
310 memcpy(cb1_vect, work, sizeof(cb1_vect));
315 find_best_vect(work, coefs, ff_cb2_vects, cba_idx ? cba_vect : NULL,
316 ortho_cb1 ? cb1_vect : NULL, data, cb2_idx, &gain);
321 * Encode a subblock of the current frame
323 * @param ractx encoder context
324 * @param sblock_data input data of the subblock
325 * @param lpc_coefs coefficients of the LPC filter
326 * @param rms RMS of the reflection coefficients
327 * @param pb pointer to PutBitContext of the current frame
329 static void ra144_encode_subblock(RA144Context *ractx,
330 const int16_t *sblock_data,
331 const int16_t *lpc_coefs, unsigned int rms,
334 float data[BLOCKSIZE] = { 0 }, work[LPC_ORDER + BLOCKSIZE];
335 float coefs[LPC_ORDER];
336 float zero[BLOCKSIZE], cba[BLOCKSIZE], cb1[BLOCKSIZE], cb2[BLOCKSIZE];
337 int16_t cba_vect[BLOCKSIZE];
338 int cba_idx, cb1_idx, cb2_idx, gain;
341 float error, best_error;
343 for (i = 0; i < LPC_ORDER; i++) {
344 work[i] = ractx->curr_sblock[BLOCKSIZE + i];
345 coefs[i] = lpc_coefs[i] * (1/4096.0);
349 * Calculate the zero-input response of the LPC filter and subtract it from
352 ff_celp_lp_synthesis_filterf(work + LPC_ORDER, coefs, data, BLOCKSIZE,
354 for (i = 0; i < BLOCKSIZE; i++) {
355 zero[i] = work[LPC_ORDER + i];
356 data[i] = sblock_data[i] - zero[i];
360 * Codebook search is performed without taking into account the contribution
361 * of the previous subblock, since it has been just subtracted from input
364 memset(work, 0, LPC_ORDER * sizeof(*work));
366 cba_idx = adaptive_cb_search(ractx->adapt_cb, work + LPC_ORDER, coefs,
370 * The filtered vector from the adaptive codebook can be retrieved from
371 * work, see implementation of adaptive_cb_search().
373 memcpy(cba, work + LPC_ORDER, sizeof(cba));
375 ff_copy_and_dup(cba_vect, ractx->adapt_cb, cba_idx + BLOCKSIZE / 2 - 1);
376 m[0] = (ff_irms(cba_vect) * rms) >> 12;
378 fixed_cb_search(work + LPC_ORDER, coefs, data, cba_idx, &cb1_idx, &cb2_idx);
379 for (i = 0; i < BLOCKSIZE; i++) {
380 cb1[i] = ff_cb1_vects[cb1_idx][i];
381 cb2[i] = ff_cb2_vects[cb2_idx][i];
383 ff_celp_lp_synthesis_filterf(work + LPC_ORDER, coefs, cb1, BLOCKSIZE,
385 memcpy(cb1, work + LPC_ORDER, sizeof(cb1));
386 m[1] = (ff_cb1_base[cb1_idx] * rms) >> 8;
387 ff_celp_lp_synthesis_filterf(work + LPC_ORDER, coefs, cb2, BLOCKSIZE,
389 memcpy(cb2, work + LPC_ORDER, sizeof(cb2));
390 m[2] = (ff_cb2_base[cb2_idx] * rms) >> 8;
391 best_error = FLT_MAX;
393 for (n = 0; n < 256; n++) {
394 g[1] = ((ff_gain_val_tab[n][1] * m[1]) >> ff_gain_exp_tab[n]) *
396 g[2] = ((ff_gain_val_tab[n][2] * m[2]) >> ff_gain_exp_tab[n]) *
400 g[0] = ((ff_gain_val_tab[n][0] * m[0]) >> ff_gain_exp_tab[n]) *
402 for (i = 0; i < BLOCKSIZE; i++) {
403 data[i] = zero[i] + g[0] * cba[i] + g[1] * cb1[i] +
405 error += (data[i] - sblock_data[i]) *
406 (data[i] - sblock_data[i]);
409 for (i = 0; i < BLOCKSIZE; i++) {
410 data[i] = zero[i] + g[1] * cb1[i] + g[2] * cb2[i];
411 error += (data[i] - sblock_data[i]) *
412 (data[i] - sblock_data[i]);
415 if (error < best_error) {
420 put_bits(pb, 7, cba_idx);
421 put_bits(pb, 8, gain);
422 put_bits(pb, 7, cb1_idx);
423 put_bits(pb, 7, cb2_idx);
424 ff_subblock_synthesis(ractx, lpc_coefs, cba_idx, cb1_idx, cb2_idx, rms,
429 static int ra144_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
430 const AVFrame *frame, int *got_packet_ptr)
432 static const uint8_t sizes[LPC_ORDER] = {64, 32, 32, 16, 16, 8, 8, 8, 8, 4};
433 static const uint8_t bit_sizes[LPC_ORDER] = {6, 5, 5, 4, 4, 3, 3, 3, 3, 2};
434 RA144Context *ractx = avctx->priv_data;
436 int32_t lpc_data[NBLOCKS * BLOCKSIZE];
437 int32_t lpc_coefs[LPC_ORDER][MAX_LPC_ORDER];
438 int shift[LPC_ORDER];
439 int16_t block_coefs[NBLOCKS][LPC_ORDER];
440 int lpc_refl[LPC_ORDER]; /**< reflection coefficients of the frame */
441 unsigned int refl_rms[NBLOCKS]; /**< RMS of the reflection coefficients */
442 const int16_t *samples = frame ? (const int16_t *)frame->data[0] : NULL;
446 if (ractx->last_frame)
449 if ((ret = ff_alloc_packet(avpkt, FRAMESIZE))) {
450 av_log(avctx, AV_LOG_ERROR, "Error getting output packet\n");
455 * Since the LPC coefficients are calculated on a frame centered over the
456 * fourth subframe, to encode a given frame, data from the next frame is
457 * needed. In each call to this function, the previous frame (whose data are
458 * saved in the encoder context) is encoded, and data from the current frame
459 * are saved in the encoder context to be used in the next function call.
461 for (i = 0; i < (2 * BLOCKSIZE + BLOCKSIZE / 2); i++) {
462 lpc_data[i] = ractx->curr_block[BLOCKSIZE + BLOCKSIZE / 2 + i];
463 energy += (lpc_data[i] * lpc_data[i]) >> 4;
467 for (j = 0; j < frame->nb_samples && i < NBLOCKS * BLOCKSIZE; i++, j++) {
468 lpc_data[i] = samples[j] >> 2;
469 energy += (lpc_data[i] * lpc_data[i]) >> 4;
472 if (i < NBLOCKS * BLOCKSIZE)
473 memset(&lpc_data[i], 0, (NBLOCKS * BLOCKSIZE - i) * sizeof(*lpc_data));
474 energy = ff_energy_tab[quantize(ff_t_sqrt(energy >> 5) >> 10, ff_energy_tab,
477 ff_lpc_calc_coefs(&ractx->lpc_ctx, lpc_data, NBLOCKS * BLOCKSIZE, LPC_ORDER,
478 LPC_ORDER, 16, lpc_coefs, shift, FF_LPC_TYPE_LEVINSON,
479 0, ORDER_METHOD_EST, 12, 0);
480 for (i = 0; i < LPC_ORDER; i++)
481 block_coefs[NBLOCKS - 1][i] = -(lpc_coefs[LPC_ORDER - 1][i] <<
482 (12 - shift[LPC_ORDER - 1]));
485 * TODO: apply perceptual weighting of the input speech through bandwidth
486 * expansion of the LPC filter.
489 if (ff_eval_refl(lpc_refl, block_coefs[NBLOCKS - 1], avctx)) {
491 * The filter is unstable: use the coefficients of the previous frame.
493 ff_int_to_int16(block_coefs[NBLOCKS - 1], ractx->lpc_coef[1]);
494 if (ff_eval_refl(lpc_refl, block_coefs[NBLOCKS - 1], avctx)) {
495 /* the filter is still unstable. set reflection coeffs to zero. */
496 memset(lpc_refl, 0, sizeof(lpc_refl));
499 init_put_bits(&pb, avpkt->data, avpkt->size);
500 for (i = 0; i < LPC_ORDER; i++) {
501 idx = quantize(lpc_refl[i], ff_lpc_refl_cb[i], sizes[i]);
502 put_bits(&pb, bit_sizes[i], idx);
503 lpc_refl[i] = ff_lpc_refl_cb[i][idx];
505 ractx->lpc_refl_rms[0] = ff_rms(lpc_refl);
506 ff_eval_coefs(ractx->lpc_coef[0], lpc_refl);
507 refl_rms[0] = ff_interp(ractx, block_coefs[0], 1, 1, ractx->old_energy);
508 refl_rms[1] = ff_interp(ractx, block_coefs[1], 2,
509 energy <= ractx->old_energy,
510 ff_t_sqrt(energy * ractx->old_energy) >> 12);
511 refl_rms[2] = ff_interp(ractx, block_coefs[2], 3, 0, energy);
512 refl_rms[3] = ff_rescale_rms(ractx->lpc_refl_rms[0], energy);
513 ff_int_to_int16(block_coefs[NBLOCKS - 1], ractx->lpc_coef[0]);
514 put_bits(&pb, 5, quantize(energy, ff_energy_tab, 32));
515 for (i = 0; i < NBLOCKS; i++)
516 ra144_encode_subblock(ractx, ractx->curr_block + i * BLOCKSIZE,
517 block_coefs[i], refl_rms[i], &pb);
519 ractx->old_energy = energy;
520 ractx->lpc_refl_rms[1] = ractx->lpc_refl_rms[0];
521 FFSWAP(unsigned int *, ractx->lpc_coef[0], ractx->lpc_coef[1]);
523 /* copy input samples to current block for processing in next call */
526 for (; i < frame->nb_samples; i++)
527 ractx->curr_block[i] = samples[i] >> 2;
529 if ((ret = ff_af_queue_add(&ractx->afq, frame)) < 0)
532 ractx->last_frame = 1;
533 memset(&ractx->curr_block[i], 0,
534 (NBLOCKS * BLOCKSIZE - i) * sizeof(*ractx->curr_block));
536 /* Get the next frame pts/duration */
537 ff_af_queue_remove(&ractx->afq, avctx->frame_size, &avpkt->pts,
540 avpkt->size = FRAMESIZE;
546 AVCodec ff_ra_144_encoder = {
548 .type = AVMEDIA_TYPE_AUDIO,
549 .id = AV_CODEC_ID_RA_144,
550 .priv_data_size = sizeof(RA144Context),
551 .init = ra144_encode_init,
552 .encode2 = ra144_encode_frame,
553 .close = ra144_encode_close,
554 .capabilities = CODEC_CAP_DELAY | CODEC_CAP_SMALL_LAST_FRAME,
555 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_S16,
556 AV_SAMPLE_FMT_NONE },
557 .long_name = NULL_IF_CONFIG_SMALL("RealAudio 1.0 (14.4K)"),