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>
32 #include "celp_filters.h"
36 static av_cold int ra144_encode_close(AVCodecContext *avctx)
38 RA144Context *ractx = avctx->priv_data;
39 ff_lpc_end(&ractx->lpc_ctx);
40 av_freep(&avctx->coded_frame);
45 static av_cold int ra144_encode_init(AVCodecContext * avctx)
50 if (avctx->channels != 1) {
51 av_log(avctx, AV_LOG_ERROR, "invalid number of channels: %d\n",
55 avctx->frame_size = NBLOCKS * BLOCKSIZE;
56 avctx->delay = avctx->frame_size;
57 avctx->bit_rate = 8000;
58 ractx = avctx->priv_data;
59 ractx->lpc_coef[0] = ractx->lpc_tables[0];
60 ractx->lpc_coef[1] = ractx->lpc_tables[1];
62 ret = ff_lpc_init(&ractx->lpc_ctx, avctx->frame_size, LPC_ORDER,
63 FF_LPC_TYPE_LEVINSON);
67 avctx->coded_frame = avcodec_alloc_frame();
68 if (!avctx->coded_frame) {
69 ret = AVERROR(ENOMEM);
75 ra144_encode_close(avctx);
81 * Quantize a value by searching a sorted table for the element with the
84 * @param value value to quantize
85 * @param table array containing the quantization table
86 * @param size size of the quantization table
87 * @return index of the quantization table corresponding to the element with the
90 static int quantize(int value, const int16_t *table, unsigned int size)
92 unsigned int low = 0, high = size - 1;
95 int index = (low + high) >> 1;
96 int error = table[index] - value;
99 return table[high] + error > value ? low : high;
110 * Orthogonalize a vector to another vector
112 * @param v vector to orthogonalize
113 * @param u vector against which orthogonalization is performed
115 static void orthogonalize(float *v, const float *u)
118 float num = 0, den = 0;
120 for (i = 0; i < BLOCKSIZE; i++) {
125 for (i = 0; i < BLOCKSIZE; i++)
131 * Calculate match score and gain of an LPC-filtered vector with respect to
132 * input data, possibly othogonalizing it to up to 2 other vectors
134 * @param work array used to calculate the filtered vector
135 * @param coefs coefficients of the LPC filter
136 * @param vect original vector
137 * @param ortho1 first vector against which orthogonalization is performed
138 * @param ortho2 second vector against which orthogonalization is performed
139 * @param data input data
140 * @param score pointer to variable where match score is returned
141 * @param gain pointer to variable where gain is returned
143 static void get_match_score(float *work, const float *coefs, float *vect,
144 const float *ortho1, const float *ortho2,
145 const float *data, float *score, float *gain)
150 ff_celp_lp_synthesis_filterf(work, coefs, vect, BLOCKSIZE, LPC_ORDER);
152 orthogonalize(work, ortho1);
154 orthogonalize(work, ortho2);
156 for (i = 0; i < BLOCKSIZE; i++) {
157 g += work[i] * work[i];
158 c += data[i] * work[i];
170 * Create a vector from the adaptive codebook at a given lag value
172 * @param vect array where vector is stored
173 * @param cb adaptive codebook
174 * @param lag lag value
176 static void create_adapt_vect(float *vect, const int16_t *cb, int lag)
180 cb += BUFFERSIZE - lag;
181 for (i = 0; i < FFMIN(BLOCKSIZE, lag); i++)
184 for (i = 0; i < BLOCKSIZE - lag; i++)
185 vect[lag + i] = cb[i];
190 * Search the adaptive codebook for the best entry and gain and remove its
191 * contribution from input data
193 * @param adapt_cb array from which the adaptive codebook is extracted
194 * @param work array used to calculate LPC-filtered vectors
195 * @param coefs coefficients of the LPC filter
196 * @param data input data
197 * @return index of the best entry of the adaptive codebook
199 static int adaptive_cb_search(const int16_t *adapt_cb, float *work,
200 const float *coefs, float *data)
203 float score, gain, best_score, best_gain;
204 float exc[BLOCKSIZE];
206 gain = best_score = 0;
207 for (i = BLOCKSIZE / 2; i <= BUFFERSIZE; i++) {
208 create_adapt_vect(exc, adapt_cb, i);
209 get_match_score(work, coefs, exc, NULL, NULL, data, &score, &gain);
210 if (score > best_score) {
220 * Re-calculate the filtered vector from the vector with maximum match score
221 * and remove its contribution from input data.
223 create_adapt_vect(exc, adapt_cb, best_vect);
224 ff_celp_lp_synthesis_filterf(work, coefs, exc, BLOCKSIZE, LPC_ORDER);
225 for (i = 0; i < BLOCKSIZE; i++)
226 data[i] -= best_gain * work[i];
227 return best_vect - BLOCKSIZE / 2 + 1;
232 * Find the best vector of a fixed codebook by applying an LPC filter to
233 * codebook entries, possibly othogonalizing them to up to 2 other vectors and
234 * matching the results with input data
236 * @param work array used to calculate the filtered vectors
237 * @param coefs coefficients of the LPC filter
238 * @param cb fixed codebook
239 * @param ortho1 first vector against which orthogonalization is performed
240 * @param ortho2 second vector against which orthogonalization is performed
241 * @param data input data
242 * @param idx pointer to variable where the index of the best codebook entry is
244 * @param gain pointer to variable where the gain of the best codebook entry is
247 static void find_best_vect(float *work, const float *coefs,
248 const int8_t cb[][BLOCKSIZE], const float *ortho1,
249 const float *ortho2, float *data, int *idx,
253 float g, score, best_score;
254 float vect[BLOCKSIZE];
256 *idx = *gain = best_score = 0;
257 for (i = 0; i < FIXED_CB_SIZE; i++) {
258 for (j = 0; j < BLOCKSIZE; j++)
260 get_match_score(work, coefs, vect, ortho1, ortho2, data, &score, &g);
261 if (score > best_score) {
271 * Search the two fixed codebooks for the best entry and gain
273 * @param work array used to calculate LPC-filtered vectors
274 * @param coefs coefficients of the LPC filter
275 * @param data input data
276 * @param cba_idx index of the best entry of the adaptive codebook
277 * @param cb1_idx pointer to variable where the index of the best entry of the
278 * first fixed codebook is returned
279 * @param cb2_idx pointer to variable where the index of the best entry of the
280 * second fixed codebook is returned
282 static void fixed_cb_search(float *work, const float *coefs, float *data,
283 int cba_idx, int *cb1_idx, int *cb2_idx)
287 float cba_vect[BLOCKSIZE], cb1_vect[BLOCKSIZE];
288 float vect[BLOCKSIZE];
291 * The filtered vector from the adaptive codebook can be retrieved from
292 * work, because this function is called just after adaptive_cb_search().
295 memcpy(cba_vect, work, sizeof(cba_vect));
297 find_best_vect(work, coefs, ff_cb1_vects, cba_idx ? cba_vect : NULL, NULL,
298 data, cb1_idx, &gain);
301 * Re-calculate the filtered vector from the vector with maximum match score
302 * and remove its contribution from input data.
305 for (i = 0; i < BLOCKSIZE; i++)
306 vect[i] = ff_cb1_vects[*cb1_idx][i];
307 ff_celp_lp_synthesis_filterf(work, coefs, vect, BLOCKSIZE, LPC_ORDER);
309 orthogonalize(work, cba_vect);
310 for (i = 0; i < BLOCKSIZE; i++)
311 data[i] -= gain * work[i];
312 memcpy(cb1_vect, work, sizeof(cb1_vect));
317 find_best_vect(work, coefs, ff_cb2_vects, cba_idx ? cba_vect : NULL,
318 ortho_cb1 ? cb1_vect : NULL, data, cb2_idx, &gain);
323 * Encode a subblock of the current frame
325 * @param ractx encoder context
326 * @param sblock_data input data of the subblock
327 * @param lpc_coefs coefficients of the LPC filter
328 * @param rms RMS of the reflection coefficients
329 * @param pb pointer to PutBitContext of the current frame
331 static void ra144_encode_subblock(RA144Context *ractx,
332 const int16_t *sblock_data,
333 const int16_t *lpc_coefs, unsigned int rms,
336 float data[BLOCKSIZE], work[LPC_ORDER + BLOCKSIZE];
337 float coefs[LPC_ORDER];
338 float zero[BLOCKSIZE], cba[BLOCKSIZE], cb1[BLOCKSIZE], cb2[BLOCKSIZE];
339 int16_t cba_vect[BLOCKSIZE];
340 int cba_idx, cb1_idx, cb2_idx, gain;
343 float error, best_error;
345 for (i = 0; i < LPC_ORDER; i++) {
346 work[i] = ractx->curr_sblock[BLOCKSIZE + i];
347 coefs[i] = lpc_coefs[i] * (1/4096.0);
351 * Calculate the zero-input response of the LPC filter and subtract it from
354 memset(data, 0, sizeof(data));
355 ff_celp_lp_synthesis_filterf(work + LPC_ORDER, coefs, data, BLOCKSIZE,
357 for (i = 0; i < BLOCKSIZE; i++) {
358 zero[i] = work[LPC_ORDER + i];
359 data[i] = sblock_data[i] - zero[i];
363 * Codebook search is performed without taking into account the contribution
364 * of the previous subblock, since it has been just subtracted from input
367 memset(work, 0, LPC_ORDER * sizeof(*work));
369 cba_idx = adaptive_cb_search(ractx->adapt_cb, work + LPC_ORDER, coefs,
373 * The filtered vector from the adaptive codebook can be retrieved from
374 * work, see implementation of adaptive_cb_search().
376 memcpy(cba, work + LPC_ORDER, sizeof(cba));
378 ff_copy_and_dup(cba_vect, ractx->adapt_cb, cba_idx + BLOCKSIZE / 2 - 1);
379 m[0] = (ff_irms(cba_vect) * rms) >> 12;
381 fixed_cb_search(work + LPC_ORDER, coefs, data, cba_idx, &cb1_idx, &cb2_idx);
382 for (i = 0; i < BLOCKSIZE; i++) {
383 cb1[i] = ff_cb1_vects[cb1_idx][i];
384 cb2[i] = ff_cb2_vects[cb2_idx][i];
386 ff_celp_lp_synthesis_filterf(work + LPC_ORDER, coefs, cb1, BLOCKSIZE,
388 memcpy(cb1, work + LPC_ORDER, sizeof(cb1));
389 m[1] = (ff_cb1_base[cb1_idx] * rms) >> 8;
390 ff_celp_lp_synthesis_filterf(work + LPC_ORDER, coefs, cb2, BLOCKSIZE,
392 memcpy(cb2, work + LPC_ORDER, sizeof(cb2));
393 m[2] = (ff_cb2_base[cb2_idx] * rms) >> 8;
394 best_error = FLT_MAX;
396 for (n = 0; n < 256; n++) {
397 g[1] = ((ff_gain_val_tab[n][1] * m[1]) >> ff_gain_exp_tab[n]) *
399 g[2] = ((ff_gain_val_tab[n][2] * m[2]) >> ff_gain_exp_tab[n]) *
403 g[0] = ((ff_gain_val_tab[n][0] * m[0]) >> ff_gain_exp_tab[n]) *
405 for (i = 0; i < BLOCKSIZE; i++) {
406 data[i] = zero[i] + g[0] * cba[i] + g[1] * cb1[i] +
408 error += (data[i] - sblock_data[i]) *
409 (data[i] - sblock_data[i]);
412 for (i = 0; i < BLOCKSIZE; i++) {
413 data[i] = zero[i] + g[1] * cb1[i] + g[2] * cb2[i];
414 error += (data[i] - sblock_data[i]) *
415 (data[i] - sblock_data[i]);
418 if (error < best_error) {
423 put_bits(pb, 7, cba_idx);
424 put_bits(pb, 8, gain);
425 put_bits(pb, 7, cb1_idx);
426 put_bits(pb, 7, cb2_idx);
427 ff_subblock_synthesis(ractx, lpc_coefs, cba_idx, cb1_idx, cb2_idx, rms,
432 static int ra144_encode_frame(AVCodecContext *avctx, uint8_t *frame,
433 int buf_size, void *data)
435 static const uint8_t sizes[LPC_ORDER] = {64, 32, 32, 16, 16, 8, 8, 8, 8, 4};
436 static const uint8_t bit_sizes[LPC_ORDER] = {6, 5, 5, 4, 4, 3, 3, 3, 3, 2};
437 RA144Context *ractx = avctx->priv_data;
439 int32_t lpc_data[NBLOCKS * BLOCKSIZE];
440 int32_t lpc_coefs[LPC_ORDER][MAX_LPC_ORDER];
441 int shift[LPC_ORDER];
442 int16_t block_coefs[NBLOCKS][LPC_ORDER];
443 int lpc_refl[LPC_ORDER]; /**< reflection coefficients of the frame */
444 unsigned int refl_rms[NBLOCKS]; /**< RMS of the reflection coefficients */
445 const int16_t *samples = data;
449 if (ractx->last_frame)
452 if (buf_size < FRAMESIZE) {
453 av_log(avctx, AV_LOG_ERROR, "output buffer too small\n");
458 * Since the LPC coefficients are calculated on a frame centered over the
459 * fourth subframe, to encode a given frame, data from the next frame is
460 * needed. In each call to this function, the previous frame (whose data are
461 * saved in the encoder context) is encoded, and data from the current frame
462 * are saved in the encoder context to be used in the next function call.
464 for (i = 0; i < (2 * BLOCKSIZE + BLOCKSIZE / 2); i++) {
465 lpc_data[i] = ractx->curr_block[BLOCKSIZE + BLOCKSIZE / 2 + i];
466 energy += (lpc_data[i] * lpc_data[i]) >> 4;
470 for (j = 0; j < avctx->frame_size && i < NBLOCKS * BLOCKSIZE; i++, j++) {
471 lpc_data[i] = samples[j] >> 2;
472 energy += (lpc_data[i] * lpc_data[i]) >> 4;
475 if (i < NBLOCKS * BLOCKSIZE)
476 memset(&lpc_data[i], 0, (NBLOCKS * BLOCKSIZE - i) * sizeof(*lpc_data));
477 energy = ff_energy_tab[quantize(ff_t_sqrt(energy >> 5) >> 10, ff_energy_tab,
480 ff_lpc_calc_coefs(&ractx->lpc_ctx, lpc_data, NBLOCKS * BLOCKSIZE, LPC_ORDER,
481 LPC_ORDER, 16, lpc_coefs, shift, FF_LPC_TYPE_LEVINSON,
482 0, ORDER_METHOD_EST, 12, 0);
483 for (i = 0; i < LPC_ORDER; i++)
484 block_coefs[NBLOCKS - 1][i] = -(lpc_coefs[LPC_ORDER - 1][i] <<
485 (12 - shift[LPC_ORDER - 1]));
488 * TODO: apply perceptual weighting of the input speech through bandwidth
489 * expansion of the LPC filter.
492 if (ff_eval_refl(lpc_refl, block_coefs[NBLOCKS - 1], avctx)) {
494 * The filter is unstable: use the coefficients of the previous frame.
496 ff_int_to_int16(block_coefs[NBLOCKS - 1], ractx->lpc_coef[1]);
497 if (ff_eval_refl(lpc_refl, block_coefs[NBLOCKS - 1], avctx)) {
498 /* the filter is still unstable. set reflection coeffs to zero. */
499 memset(lpc_refl, 0, sizeof(lpc_refl));
502 init_put_bits(&pb, frame, buf_size);
503 for (i = 0; i < LPC_ORDER; i++) {
504 idx = quantize(lpc_refl[i], ff_lpc_refl_cb[i], sizes[i]);
505 put_bits(&pb, bit_sizes[i], idx);
506 lpc_refl[i] = ff_lpc_refl_cb[i][idx];
508 ractx->lpc_refl_rms[0] = ff_rms(lpc_refl);
509 ff_eval_coefs(ractx->lpc_coef[0], lpc_refl);
510 refl_rms[0] = ff_interp(ractx, block_coefs[0], 1, 1, ractx->old_energy);
511 refl_rms[1] = ff_interp(ractx, block_coefs[1], 2,
512 energy <= ractx->old_energy,
513 ff_t_sqrt(energy * ractx->old_energy) >> 12);
514 refl_rms[2] = ff_interp(ractx, block_coefs[2], 3, 0, energy);
515 refl_rms[3] = ff_rescale_rms(ractx->lpc_refl_rms[0], energy);
516 ff_int_to_int16(block_coefs[NBLOCKS - 1], ractx->lpc_coef[0]);
517 put_bits(&pb, 5, quantize(energy, ff_energy_tab, 32));
518 for (i = 0; i < NBLOCKS; i++)
519 ra144_encode_subblock(ractx, ractx->curr_block + i * BLOCKSIZE,
520 block_coefs[i], refl_rms[i], &pb);
522 ractx->old_energy = energy;
523 ractx->lpc_refl_rms[1] = ractx->lpc_refl_rms[0];
524 FFSWAP(unsigned int *, ractx->lpc_coef[0], ractx->lpc_coef[1]);
526 /* copy input samples to current block for processing in next call */
529 for (; i < avctx->frame_size; i++)
530 ractx->curr_block[i] = samples[i] >> 2;
532 ractx->last_frame = 1;
533 memset(&ractx->curr_block[i], 0,
534 (NBLOCKS * BLOCKSIZE - i) * sizeof(*ractx->curr_block));
540 AVCodec ff_ra_144_encoder = {
542 .type = AVMEDIA_TYPE_AUDIO,
543 .id = CODEC_ID_RA_144,
544 .priv_data_size = sizeof(RA144Context),
545 .init = ra144_encode_init,
546 .encode = ra144_encode_frame,
547 .close = ra144_encode_close,
548 .capabilities = CODEC_CAP_DELAY | CODEC_CAP_SMALL_LAST_FRAME,
549 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_S16,
550 AV_SAMPLE_FMT_NONE },
551 .long_name = NULL_IF_CONFIG_SMALL("RealAudio 1.0 (14.4K)"),