• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Real Audio 1.0 (14.4K) encoder
3  * Copyright (c) 2010 Francesco Lavra <francescolavra@interfree.it>
4  *
5  * This file is part of FFmpeg.
6  *
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.
11  *
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.
16  *
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
20  */
21 
22 /**
23  * @file
24  * Real Audio 1.0 (14.4K) encoder
25  * @author Francesco Lavra <francescolavra@interfree.it>
26  */
27 
28 #include <float.h>
29 
30 #include "avcodec.h"
31 #include "audio_frame_queue.h"
32 #include "celp_filters.h"
33 #include "internal.h"
34 #include "mathops.h"
35 #include "put_bits.h"
36 #include "ra144.h"
37 
ra144_encode_close(AVCodecContext * avctx)38 static av_cold int ra144_encode_close(AVCodecContext *avctx)
39 {
40     RA144Context *ractx = avctx->priv_data;
41     ff_lpc_end(&ractx->lpc_ctx);
42     ff_af_queue_close(&ractx->afq);
43     return 0;
44 }
45 
46 
ra144_encode_init(AVCodecContext * avctx)47 static av_cold int ra144_encode_init(AVCodecContext * avctx)
48 {
49     RA144Context *ractx;
50     int ret;
51 
52     if (avctx->channels != 1) {
53         av_log(avctx, AV_LOG_ERROR, "invalid number of channels: %d\n",
54                avctx->channels);
55         return -1;
56     }
57     avctx->frame_size = NBLOCKS * BLOCKSIZE;
58     avctx->initial_padding = 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];
63     ractx->avctx = avctx;
64     ff_audiodsp_init(&ractx->adsp);
65     ret = ff_lpc_init(&ractx->lpc_ctx, avctx->frame_size, LPC_ORDER,
66                       FF_LPC_TYPE_LEVINSON);
67     if (ret < 0)
68         return ret;
69 
70     ff_af_queue_init(avctx, &ractx->afq);
71 
72     return 0;
73 }
74 
75 
76 /**
77  * Quantize a value by searching a sorted table for the element with the
78  * nearest value
79  *
80  * @param value value to quantize
81  * @param table array containing the quantization table
82  * @param size size of the quantization table
83  * @return index of the quantization table corresponding to the element with the
84  *         nearest value
85  */
quantize(int value,const int16_t * table,unsigned int size)86 static int quantize(int value, const int16_t *table, unsigned int size)
87 {
88     unsigned int low = 0, high = size - 1;
89 
90     while (1) {
91         int index = (low + high) >> 1;
92         int error = table[index] - value;
93 
94         if (index == low)
95             return table[high] + error > value ? low : high;
96         if (error > 0) {
97             high = index;
98         } else {
99             low = index;
100         }
101     }
102 }
103 
104 
105 /**
106  * Orthogonalize a vector to another vector
107  *
108  * @param v vector to orthogonalize
109  * @param u vector against which orthogonalization is performed
110  */
orthogonalize(float * v,const float * u)111 static void orthogonalize(float *v, const float *u)
112 {
113     int i;
114     float num = 0, den = 0;
115 
116     for (i = 0; i < BLOCKSIZE; i++) {
117         num += v[i] * u[i];
118         den += u[i] * u[i];
119     }
120     num /= den;
121     for (i = 0; i < BLOCKSIZE; i++)
122         v[i] -= num * u[i];
123 }
124 
125 
126 /**
127  * Calculate match score and gain of an LPC-filtered vector with respect to
128  * input data, possibly orthogonalizing it to up to two other vectors.
129  *
130  * @param work array used to calculate the filtered vector
131  * @param coefs coefficients of the LPC filter
132  * @param vect original vector
133  * @param ortho1 first vector against which orthogonalization is performed
134  * @param ortho2 second vector against which orthogonalization is performed
135  * @param data input data
136  * @param score pointer to variable where match score is returned
137  * @param gain pointer to variable where gain is returned
138  */
get_match_score(float * work,const float * coefs,float * vect,const float * ortho1,const float * ortho2,const float * data,float * score,float * gain)139 static void get_match_score(float *work, const float *coefs, float *vect,
140                             const float *ortho1, const float *ortho2,
141                             const float *data, float *score, float *gain)
142 {
143     float c, g;
144     int i;
145 
146     ff_celp_lp_synthesis_filterf(work, coefs, vect, BLOCKSIZE, LPC_ORDER);
147     if (ortho1)
148         orthogonalize(work, ortho1);
149     if (ortho2)
150         orthogonalize(work, ortho2);
151     c = g = 0;
152     for (i = 0; i < BLOCKSIZE; i++) {
153         g += work[i] * work[i];
154         c += data[i] * work[i];
155     }
156     if (c <= 0) {
157         *score = 0;
158         return;
159     }
160     *gain = c / g;
161     *score = *gain * c;
162 }
163 
164 
165 /**
166  * Create a vector from the adaptive codebook at a given lag value
167  *
168  * @param vect array where vector is stored
169  * @param cb adaptive codebook
170  * @param lag lag value
171  */
create_adapt_vect(float * vect,const int16_t * cb,int lag)172 static void create_adapt_vect(float *vect, const int16_t *cb, int lag)
173 {
174     int i;
175 
176     cb += BUFFERSIZE - lag;
177     for (i = 0; i < FFMIN(BLOCKSIZE, lag); i++)
178         vect[i] = cb[i];
179     if (lag < BLOCKSIZE)
180         for (i = 0; i < BLOCKSIZE - lag; i++)
181             vect[lag + i] = cb[i];
182 }
183 
184 
185 /**
186  * Search the adaptive codebook for the best entry and gain and remove its
187  * contribution from input data
188  *
189  * @param adapt_cb array from which the adaptive codebook is extracted
190  * @param work array used to calculate LPC-filtered vectors
191  * @param coefs coefficients of the LPC filter
192  * @param data input data
193  * @return index of the best entry of the adaptive codebook
194  */
adaptive_cb_search(const int16_t * adapt_cb,float * work,const float * coefs,float * data)195 static int adaptive_cb_search(const int16_t *adapt_cb, float *work,
196                               const float *coefs, float *data)
197 {
198     int i, av_uninit(best_vect);
199     float score, gain, best_score, av_uninit(best_gain);
200     float exc[BLOCKSIZE];
201 
202     gain = best_score = 0;
203     for (i = BLOCKSIZE / 2; i <= BUFFERSIZE; i++) {
204         create_adapt_vect(exc, adapt_cb, i);
205         get_match_score(work, coefs, exc, NULL, NULL, data, &score, &gain);
206         if (score > best_score) {
207             best_score = score;
208             best_vect = i;
209             best_gain = gain;
210         }
211     }
212     if (!best_score)
213         return 0;
214 
215     /**
216      * Re-calculate the filtered vector from the vector with maximum match score
217      * and remove its contribution from input data.
218      */
219     create_adapt_vect(exc, adapt_cb, best_vect);
220     ff_celp_lp_synthesis_filterf(work, coefs, exc, BLOCKSIZE, LPC_ORDER);
221     for (i = 0; i < BLOCKSIZE; i++)
222         data[i] -= best_gain * work[i];
223     return best_vect - BLOCKSIZE / 2 + 1;
224 }
225 
226 
227 /**
228  * Find the best vector of a fixed codebook by applying an LPC filter to
229  * codebook entries, possibly orthogonalizing them to up to two other vectors
230  * and matching the results with input data.
231  *
232  * @param work array used to calculate the filtered vectors
233  * @param coefs coefficients of the LPC filter
234  * @param cb fixed codebook
235  * @param ortho1 first vector against which orthogonalization is performed
236  * @param ortho2 second vector against which orthogonalization is performed
237  * @param data input data
238  * @param idx pointer to variable where the index of the best codebook entry is
239  *        returned
240  * @param gain pointer to variable where the gain of the best codebook entry is
241  *        returned
242  */
find_best_vect(float * work,const float * coefs,const int8_t cb[][BLOCKSIZE],const float * ortho1,const float * ortho2,float * data,int * idx,float * gain)243 static void find_best_vect(float *work, const float *coefs,
244                            const int8_t cb[][BLOCKSIZE], const float *ortho1,
245                            const float *ortho2, float *data, int *idx,
246                            float *gain)
247 {
248     int i, j;
249     float g, score, best_score;
250     float vect[BLOCKSIZE];
251 
252     *idx = *gain = best_score = 0;
253     for (i = 0; i < FIXED_CB_SIZE; i++) {
254         for (j = 0; j < BLOCKSIZE; j++)
255             vect[j] = cb[i][j];
256         get_match_score(work, coefs, vect, ortho1, ortho2, data, &score, &g);
257         if (score > best_score) {
258             best_score = score;
259             *idx = i;
260             *gain = g;
261         }
262     }
263 }
264 
265 
266 /**
267  * Search the two fixed codebooks for the best entry and gain
268  *
269  * @param work array used to calculate LPC-filtered vectors
270  * @param coefs coefficients of the LPC filter
271  * @param data input data
272  * @param cba_idx index of the best entry of the adaptive codebook
273  * @param cb1_idx pointer to variable where the index of the best entry of the
274  *        first fixed codebook is returned
275  * @param cb2_idx pointer to variable where the index of the best entry of the
276  *        second fixed codebook is returned
277  */
fixed_cb_search(float * work,const float * coefs,float * data,int cba_idx,int * cb1_idx,int * cb2_idx)278 static void fixed_cb_search(float *work, const float *coefs, float *data,
279                             int cba_idx, int *cb1_idx, int *cb2_idx)
280 {
281     int i, ortho_cb1;
282     float gain;
283     float cba_vect[BLOCKSIZE], cb1_vect[BLOCKSIZE];
284     float vect[BLOCKSIZE];
285 
286     /**
287      * The filtered vector from the adaptive codebook can be retrieved from
288      * work, because this function is called just after adaptive_cb_search().
289      */
290     if (cba_idx)
291         memcpy(cba_vect, work, sizeof(cba_vect));
292 
293     find_best_vect(work, coefs, ff_cb1_vects, cba_idx ? cba_vect : NULL, NULL,
294                    data, cb1_idx, &gain);
295 
296     /**
297      * Re-calculate the filtered vector from the vector with maximum match score
298      * and remove its contribution from input data.
299      */
300     if (gain) {
301         for (i = 0; i < BLOCKSIZE; i++)
302             vect[i] = ff_cb1_vects[*cb1_idx][i];
303         ff_celp_lp_synthesis_filterf(work, coefs, vect, BLOCKSIZE, LPC_ORDER);
304         if (cba_idx)
305             orthogonalize(work, cba_vect);
306         for (i = 0; i < BLOCKSIZE; i++)
307             data[i] -= gain * work[i];
308         memcpy(cb1_vect, work, sizeof(cb1_vect));
309         ortho_cb1 = 1;
310     } else
311         ortho_cb1 = 0;
312 
313     find_best_vect(work, coefs, ff_cb2_vects, cba_idx ? cba_vect : NULL,
314                    ortho_cb1 ? cb1_vect : NULL, data, cb2_idx, &gain);
315 }
316 
317 
318 /**
319  * Encode a subblock of the current frame
320  *
321  * @param ractx encoder context
322  * @param sblock_data input data of the subblock
323  * @param lpc_coefs coefficients of the LPC filter
324  * @param rms RMS of the reflection coefficients
325  * @param pb pointer to PutBitContext of the current frame
326  */
ra144_encode_subblock(RA144Context * ractx,const int16_t * sblock_data,const int16_t * lpc_coefs,unsigned int rms,PutBitContext * pb)327 static void ra144_encode_subblock(RA144Context *ractx,
328                                   const int16_t *sblock_data,
329                                   const int16_t *lpc_coefs, unsigned int rms,
330                                   PutBitContext *pb)
331 {
332     float data[BLOCKSIZE] = { 0 }, work[LPC_ORDER + BLOCKSIZE];
333     float coefs[LPC_ORDER];
334     float zero[BLOCKSIZE], cba[BLOCKSIZE], cb1[BLOCKSIZE], cb2[BLOCKSIZE];
335     int cba_idx, cb1_idx, cb2_idx, gain;
336     int i, n;
337     unsigned m[3];
338     float g[3];
339     float error, best_error;
340 
341     for (i = 0; i < LPC_ORDER; i++) {
342         work[i] = ractx->curr_sblock[BLOCKSIZE + i];
343         coefs[i] = lpc_coefs[i] * (1/4096.0);
344     }
345 
346     /**
347      * Calculate the zero-input response of the LPC filter and subtract it from
348      * input data.
349      */
350     ff_celp_lp_synthesis_filterf(work + LPC_ORDER, coefs, data, BLOCKSIZE,
351                                  LPC_ORDER);
352     for (i = 0; i < BLOCKSIZE; i++) {
353         zero[i] = work[LPC_ORDER + i];
354         data[i] = sblock_data[i] - zero[i];
355     }
356 
357     /**
358      * Codebook search is performed without taking into account the contribution
359      * of the previous subblock, since it has been just subtracted from input
360      * data.
361      */
362     memset(work, 0, LPC_ORDER * sizeof(*work));
363 
364     cba_idx = adaptive_cb_search(ractx->adapt_cb, work + LPC_ORDER, coefs,
365                                  data);
366     if (cba_idx) {
367         /**
368          * The filtered vector from the adaptive codebook can be retrieved from
369          * work, see implementation of adaptive_cb_search().
370          */
371         memcpy(cba, work + LPC_ORDER, sizeof(cba));
372 
373         ff_copy_and_dup(ractx->buffer_a, ractx->adapt_cb, cba_idx + BLOCKSIZE / 2 - 1);
374         m[0] = (ff_irms(&ractx->adsp, ractx->buffer_a) * rms) >> 12;
375     }
376     fixed_cb_search(work + LPC_ORDER, coefs, data, cba_idx, &cb1_idx, &cb2_idx);
377     for (i = 0; i < BLOCKSIZE; i++) {
378         cb1[i] = ff_cb1_vects[cb1_idx][i];
379         cb2[i] = ff_cb2_vects[cb2_idx][i];
380     }
381     ff_celp_lp_synthesis_filterf(work + LPC_ORDER, coefs, cb1, BLOCKSIZE,
382                                  LPC_ORDER);
383     memcpy(cb1, work + LPC_ORDER, sizeof(cb1));
384     m[1] = (ff_cb1_base[cb1_idx] * rms) >> 8;
385     ff_celp_lp_synthesis_filterf(work + LPC_ORDER, coefs, cb2, BLOCKSIZE,
386                                  LPC_ORDER);
387     memcpy(cb2, work + LPC_ORDER, sizeof(cb2));
388     m[2] = (ff_cb2_base[cb2_idx] * rms) >> 8;
389     best_error = FLT_MAX;
390     gain = 0;
391     for (n = 0; n < 256; n++) {
392         g[1] = ((ff_gain_val_tab[n][1] * m[1]) >> ff_gain_exp_tab[n]) *
393                (1/4096.0);
394         g[2] = ((ff_gain_val_tab[n][2] * m[2]) >> ff_gain_exp_tab[n]) *
395                (1/4096.0);
396         error = 0;
397         if (cba_idx) {
398             g[0] = ((ff_gain_val_tab[n][0] * m[0]) >> ff_gain_exp_tab[n]) *
399                    (1/4096.0);
400             for (i = 0; i < BLOCKSIZE; i++) {
401                 data[i] = zero[i] + g[0] * cba[i] + g[1] * cb1[i] +
402                           g[2] * cb2[i];
403                 error += (data[i] - sblock_data[i]) *
404                          (data[i] - sblock_data[i]);
405             }
406         } else {
407             for (i = 0; i < BLOCKSIZE; i++) {
408                 data[i] = zero[i] + g[1] * cb1[i] + g[2] * cb2[i];
409                 error += (data[i] - sblock_data[i]) *
410                          (data[i] - sblock_data[i]);
411             }
412         }
413         if (error < best_error) {
414             best_error = error;
415             gain = n;
416         }
417     }
418     put_bits(pb, 7, cba_idx);
419     put_bits(pb, 8, gain);
420     put_bits(pb, 7, cb1_idx);
421     put_bits(pb, 7, cb2_idx);
422     ff_subblock_synthesis(ractx, lpc_coefs, cba_idx, cb1_idx, cb2_idx, rms,
423                           gain);
424 }
425 
426 
ra144_encode_frame(AVCodecContext * avctx,AVPacket * avpkt,const AVFrame * frame,int * got_packet_ptr)427 static int ra144_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
428                               const AVFrame *frame, int *got_packet_ptr)
429 {
430     static const uint8_t sizes[LPC_ORDER] = {64, 32, 32, 16, 16, 8, 8, 8, 8, 4};
431     static const uint8_t bit_sizes[LPC_ORDER] = {6, 5, 5, 4, 4, 3, 3, 3, 3, 2};
432     RA144Context *ractx = avctx->priv_data;
433     PutBitContext pb;
434     int32_t lpc_data[NBLOCKS * BLOCKSIZE];
435     int32_t lpc_coefs[LPC_ORDER][MAX_LPC_ORDER];
436     int shift[LPC_ORDER];
437     int16_t block_coefs[NBLOCKS][LPC_ORDER];
438     int lpc_refl[LPC_ORDER];    /**< reflection coefficients of the frame */
439     unsigned int refl_rms[NBLOCKS]; /**< RMS of the reflection coefficients */
440     const int16_t *samples = frame ? (const int16_t *)frame->data[0] : NULL;
441     int energy = 0;
442     int i, idx, ret;
443 
444     if (ractx->last_frame)
445         return 0;
446 
447     if ((ret = ff_alloc_packet2(avctx, avpkt, FRAME_SIZE, 0)) < 0)
448         return ret;
449 
450     /**
451      * Since the LPC coefficients are calculated on a frame centered over the
452      * fourth subframe, to encode a given frame, data from the next frame is
453      * needed. In each call to this function, the previous frame (whose data are
454      * saved in the encoder context) is encoded, and data from the current frame
455      * are saved in the encoder context to be used in the next function call.
456      */
457     for (i = 0; i < (2 * BLOCKSIZE + BLOCKSIZE / 2); i++) {
458         lpc_data[i] = ractx->curr_block[BLOCKSIZE + BLOCKSIZE / 2 + i];
459         energy += (lpc_data[i] * lpc_data[i]) >> 4;
460     }
461     if (frame) {
462         int j;
463         for (j = 0; j < frame->nb_samples && i < NBLOCKS * BLOCKSIZE; i++, j++) {
464             lpc_data[i] = samples[j] >> 2;
465             energy += (lpc_data[i] * lpc_data[i]) >> 4;
466         }
467     }
468     if (i < NBLOCKS * BLOCKSIZE)
469         memset(&lpc_data[i], 0, (NBLOCKS * BLOCKSIZE - i) * sizeof(*lpc_data));
470     energy = ff_energy_tab[quantize(ff_t_sqrt(energy >> 5) >> 10, ff_energy_tab,
471                                     32)];
472 
473     ff_lpc_calc_coefs(&ractx->lpc_ctx, lpc_data, NBLOCKS * BLOCKSIZE, LPC_ORDER,
474                       LPC_ORDER, 16, lpc_coefs, shift, FF_LPC_TYPE_LEVINSON,
475                       0, ORDER_METHOD_EST, 0, 12, 0);
476     for (i = 0; i < LPC_ORDER; i++)
477         block_coefs[NBLOCKS - 1][i] = -lpc_coefs[LPC_ORDER - 1][i]
478                                        * (1 << (12 - shift[LPC_ORDER - 1]));
479 
480     /**
481      * TODO: apply perceptual weighting of the input speech through bandwidth
482      * expansion of the LPC filter.
483      */
484 
485     if (ff_eval_refl(lpc_refl, block_coefs[NBLOCKS - 1], avctx)) {
486         /**
487          * The filter is unstable: use the coefficients of the previous frame.
488          */
489         ff_int_to_int16(block_coefs[NBLOCKS - 1], ractx->lpc_coef[1]);
490         if (ff_eval_refl(lpc_refl, block_coefs[NBLOCKS - 1], avctx)) {
491             /* the filter is still unstable. set reflection coeffs to zero. */
492             memset(lpc_refl, 0, sizeof(lpc_refl));
493         }
494     }
495     init_put_bits(&pb, avpkt->data, avpkt->size);
496     for (i = 0; i < LPC_ORDER; i++) {
497         idx = quantize(lpc_refl[i], ff_lpc_refl_cb[i], sizes[i]);
498         put_bits(&pb, bit_sizes[i], idx);
499         lpc_refl[i] = ff_lpc_refl_cb[i][idx];
500     }
501     ractx->lpc_refl_rms[0] = ff_rms(lpc_refl);
502     ff_eval_coefs(ractx->lpc_coef[0], lpc_refl);
503     refl_rms[0] = ff_interp(ractx, block_coefs[0], 1, 1, ractx->old_energy);
504     refl_rms[1] = ff_interp(ractx, block_coefs[1], 2,
505                             energy <= ractx->old_energy,
506                             ff_t_sqrt(energy * ractx->old_energy) >> 12);
507     refl_rms[2] = ff_interp(ractx, block_coefs[2], 3, 0, energy);
508     refl_rms[3] = ff_rescale_rms(ractx->lpc_refl_rms[0], energy);
509     ff_int_to_int16(block_coefs[NBLOCKS - 1], ractx->lpc_coef[0]);
510     put_bits(&pb, 5, quantize(energy, ff_energy_tab, 32));
511     for (i = 0; i < NBLOCKS; i++)
512         ra144_encode_subblock(ractx, ractx->curr_block + i * BLOCKSIZE,
513                               block_coefs[i], refl_rms[i], &pb);
514     flush_put_bits(&pb);
515     ractx->old_energy = energy;
516     ractx->lpc_refl_rms[1] = ractx->lpc_refl_rms[0];
517     FFSWAP(unsigned int *, ractx->lpc_coef[0], ractx->lpc_coef[1]);
518 
519     /* copy input samples to current block for processing in next call */
520     i = 0;
521     if (frame) {
522         for (; i < frame->nb_samples; i++)
523             ractx->curr_block[i] = samples[i] >> 2;
524 
525         if ((ret = ff_af_queue_add(&ractx->afq, frame)) < 0)
526             return ret;
527     } else
528         ractx->last_frame = 1;
529     memset(&ractx->curr_block[i], 0,
530            (NBLOCKS * BLOCKSIZE - i) * sizeof(*ractx->curr_block));
531 
532     /* Get the next frame pts/duration */
533     ff_af_queue_remove(&ractx->afq, avctx->frame_size, &avpkt->pts,
534                        &avpkt->duration);
535 
536     avpkt->size = FRAME_SIZE;
537     *got_packet_ptr = 1;
538     return 0;
539 }
540 
541 
542 AVCodec ff_ra_144_encoder = {
543     .name           = "real_144",
544     .long_name      = NULL_IF_CONFIG_SMALL("RealAudio 1.0 (14.4K)"),
545     .type           = AVMEDIA_TYPE_AUDIO,
546     .id             = AV_CODEC_ID_RA_144,
547     .priv_data_size = sizeof(RA144Context),
548     .init           = ra144_encode_init,
549     .encode2        = ra144_encode_frame,
550     .close          = ra144_encode_close,
551     .capabilities   = AV_CODEC_CAP_DELAY | AV_CODEC_CAP_SMALL_LAST_FRAME,
552     .sample_fmts    = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_S16,
553                                                      AV_SAMPLE_FMT_NONE },
554     .supported_samplerates = (const int[]){ 8000, 0 },
555     .channel_layouts = (const uint64_t[]) { AV_CH_LAYOUT_MONO, 0 },
556 };
557