xref: /third_party/ffmpeg/libavcodec/ra144enc.c (revision cabdff1a)
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 "libavutil/channel_layout.h"
31#include "avcodec.h"
32#include "audio_frame_queue.h"
33#include "celp_filters.h"
34#include "codec_internal.h"
35#include "encode.h"
36#include "mathops.h"
37#include "put_bits.h"
38#include "ra144.h"
39
40static av_cold int ra144_encode_close(AVCodecContext *avctx)
41{
42    RA144Context *ractx = avctx->priv_data;
43    ff_lpc_end(&ractx->lpc_ctx);
44    ff_af_queue_close(&ractx->afq);
45    return 0;
46}
47
48
49static av_cold int ra144_encode_init(AVCodecContext * avctx)
50{
51    RA144Context *ractx;
52    int ret;
53
54    avctx->frame_size = NBLOCKS * BLOCKSIZE;
55    avctx->initial_padding = avctx->frame_size;
56    avctx->bit_rate = 8000;
57    ractx = avctx->priv_data;
58    ractx->lpc_coef[0] = ractx->lpc_tables[0];
59    ractx->lpc_coef[1] = ractx->lpc_tables[1];
60    ractx->avctx = avctx;
61    ff_audiodsp_init(&ractx->adsp);
62    ret = ff_lpc_init(&ractx->lpc_ctx, avctx->frame_size, LPC_ORDER,
63                      FF_LPC_TYPE_LEVINSON);
64    if (ret < 0)
65        return ret;
66
67    ff_af_queue_init(avctx, &ractx->afq);
68
69    return 0;
70}
71
72
73/**
74 * Quantize a value by searching a sorted table for the element with the
75 * nearest value
76 *
77 * @param value value to quantize
78 * @param table array containing the quantization table
79 * @param size size of the quantization table
80 * @return index of the quantization table corresponding to the element with the
81 *         nearest value
82 */
83static int quantize(int value, const int16_t *table, unsigned int size)
84{
85    unsigned int low = 0, high = size - 1;
86
87    while (1) {
88        int index = (low + high) >> 1;
89        int error = table[index] - value;
90
91        if (index == low)
92            return table[high] + error > value ? low : high;
93        if (error > 0) {
94            high = index;
95        } else {
96            low = index;
97        }
98    }
99}
100
101
102/**
103 * Orthogonalize a vector to another vector
104 *
105 * @param v vector to orthogonalize
106 * @param u vector against which orthogonalization is performed
107 */
108static void orthogonalize(float *v, const float *u)
109{
110    int i;
111    float num = 0, den = 0;
112
113    for (i = 0; i < BLOCKSIZE; i++) {
114        num += v[i] * u[i];
115        den += u[i] * u[i];
116    }
117    num /= den;
118    for (i = 0; i < BLOCKSIZE; i++)
119        v[i] -= num * u[i];
120}
121
122
123/**
124 * Calculate match score and gain of an LPC-filtered vector with respect to
125 * input data, possibly orthogonalizing it to up to two other vectors.
126 *
127 * @param work array used to calculate the filtered vector
128 * @param coefs coefficients of the LPC filter
129 * @param vect original vector
130 * @param ortho1 first vector against which orthogonalization is performed
131 * @param ortho2 second vector against which orthogonalization is performed
132 * @param data input data
133 * @param score pointer to variable where match score is returned
134 * @param gain pointer to variable where gain is returned
135 */
136static void get_match_score(float *work, const float *coefs, float *vect,
137                            const float *ortho1, const float *ortho2,
138                            const float *data, float *score, float *gain)
139{
140    float c, g;
141    int i;
142
143    ff_celp_lp_synthesis_filterf(work, coefs, vect, BLOCKSIZE, LPC_ORDER);
144    if (ortho1)
145        orthogonalize(work, ortho1);
146    if (ortho2)
147        orthogonalize(work, ortho2);
148    c = g = 0;
149    for (i = 0; i < BLOCKSIZE; i++) {
150        g += work[i] * work[i];
151        c += data[i] * work[i];
152    }
153    if (c <= 0) {
154        *score = 0;
155        return;
156    }
157    *gain = c / g;
158    *score = *gain * c;
159}
160
161
162/**
163 * Create a vector from the adaptive codebook at a given lag value
164 *
165 * @param vect array where vector is stored
166 * @param cb adaptive codebook
167 * @param lag lag value
168 */
169static void create_adapt_vect(float *vect, const int16_t *cb, int lag)
170{
171    int i;
172
173    cb += BUFFERSIZE - lag;
174    for (i = 0; i < FFMIN(BLOCKSIZE, lag); i++)
175        vect[i] = cb[i];
176    if (lag < BLOCKSIZE)
177        for (i = 0; i < BLOCKSIZE - lag; i++)
178            vect[lag + i] = cb[i];
179}
180
181
182/**
183 * Search the adaptive codebook for the best entry and gain and remove its
184 * contribution from input data
185 *
186 * @param adapt_cb array from which the adaptive codebook is extracted
187 * @param work array used to calculate LPC-filtered vectors
188 * @param coefs coefficients of the LPC filter
189 * @param data input data
190 * @return index of the best entry of the adaptive codebook
191 */
192static int adaptive_cb_search(const int16_t *adapt_cb, float *work,
193                              const float *coefs, float *data)
194{
195    int i, av_uninit(best_vect);
196    float score, gain, best_score, av_uninit(best_gain);
197    float exc[BLOCKSIZE];
198
199    gain = best_score = 0;
200    for (i = BLOCKSIZE / 2; i <= BUFFERSIZE; i++) {
201        create_adapt_vect(exc, adapt_cb, i);
202        get_match_score(work, coefs, exc, NULL, NULL, data, &score, &gain);
203        if (score > best_score) {
204            best_score = score;
205            best_vect = i;
206            best_gain = gain;
207        }
208    }
209    if (!best_score)
210        return 0;
211
212    /**
213     * Re-calculate the filtered vector from the vector with maximum match score
214     * and remove its contribution from input data.
215     */
216    create_adapt_vect(exc, adapt_cb, best_vect);
217    ff_celp_lp_synthesis_filterf(work, coefs, exc, BLOCKSIZE, LPC_ORDER);
218    for (i = 0; i < BLOCKSIZE; i++)
219        data[i] -= best_gain * work[i];
220    return best_vect - BLOCKSIZE / 2 + 1;
221}
222
223
224/**
225 * Find the best vector of a fixed codebook by applying an LPC filter to
226 * codebook entries, possibly orthogonalizing them to up to two other vectors
227 * and matching the results with input data.
228 *
229 * @param work array used to calculate the filtered vectors
230 * @param coefs coefficients of the LPC filter
231 * @param cb fixed codebook
232 * @param ortho1 first vector against which orthogonalization is performed
233 * @param ortho2 second vector against which orthogonalization is performed
234 * @param data input data
235 * @param idx pointer to variable where the index of the best codebook entry is
236 *        returned
237 * @param gain pointer to variable where the gain of the best codebook entry is
238 *        returned
239 */
240static void find_best_vect(float *work, const float *coefs,
241                           const int8_t cb[][BLOCKSIZE], const float *ortho1,
242                           const float *ortho2, float *data, int *idx,
243                           float *gain)
244{
245    int i, j;
246    float g, score, best_score;
247    float vect[BLOCKSIZE];
248
249    *idx = *gain = best_score = 0;
250    for (i = 0; i < FIXED_CB_SIZE; i++) {
251        for (j = 0; j < BLOCKSIZE; j++)
252            vect[j] = cb[i][j];
253        get_match_score(work, coefs, vect, ortho1, ortho2, data, &score, &g);
254        if (score > best_score) {
255            best_score = score;
256            *idx = i;
257            *gain = g;
258        }
259    }
260}
261
262
263/**
264 * Search the two fixed codebooks for the best entry and gain
265 *
266 * @param work array used to calculate LPC-filtered vectors
267 * @param coefs coefficients of the LPC filter
268 * @param data input data
269 * @param cba_idx index of the best entry of the adaptive codebook
270 * @param cb1_idx pointer to variable where the index of the best entry of the
271 *        first fixed codebook is returned
272 * @param cb2_idx pointer to variable where the index of the best entry of the
273 *        second fixed codebook is returned
274 */
275static void fixed_cb_search(float *work, const float *coefs, float *data,
276                            int cba_idx, int *cb1_idx, int *cb2_idx)
277{
278    int i, ortho_cb1;
279    float gain;
280    float cba_vect[BLOCKSIZE], cb1_vect[BLOCKSIZE];
281    float vect[BLOCKSIZE];
282
283    /**
284     * The filtered vector from the adaptive codebook can be retrieved from
285     * work, because this function is called just after adaptive_cb_search().
286     */
287    if (cba_idx)
288        memcpy(cba_vect, work, sizeof(cba_vect));
289
290    find_best_vect(work, coefs, ff_cb1_vects, cba_idx ? cba_vect : NULL, NULL,
291                   data, cb1_idx, &gain);
292
293    /**
294     * Re-calculate the filtered vector from the vector with maximum match score
295     * and remove its contribution from input data.
296     */
297    if (gain) {
298        for (i = 0; i < BLOCKSIZE; i++)
299            vect[i] = ff_cb1_vects[*cb1_idx][i];
300        ff_celp_lp_synthesis_filterf(work, coefs, vect, BLOCKSIZE, LPC_ORDER);
301        if (cba_idx)
302            orthogonalize(work, cba_vect);
303        for (i = 0; i < BLOCKSIZE; i++)
304            data[i] -= gain * work[i];
305        memcpy(cb1_vect, work, sizeof(cb1_vect));
306        ortho_cb1 = 1;
307    } else
308        ortho_cb1 = 0;
309
310    find_best_vect(work, coefs, ff_cb2_vects, cba_idx ? cba_vect : NULL,
311                   ortho_cb1 ? cb1_vect : NULL, data, cb2_idx, &gain);
312}
313
314
315/**
316 * Encode a subblock of the current frame
317 *
318 * @param ractx encoder context
319 * @param sblock_data input data of the subblock
320 * @param lpc_coefs coefficients of the LPC filter
321 * @param rms RMS of the reflection coefficients
322 * @param pb pointer to PutBitContext of the current frame
323 */
324static void ra144_encode_subblock(RA144Context *ractx,
325                                  const int16_t *sblock_data,
326                                  const int16_t *lpc_coefs, unsigned int rms,
327                                  PutBitContext *pb)
328{
329    float data[BLOCKSIZE] = { 0 }, work[LPC_ORDER + BLOCKSIZE];
330    float coefs[LPC_ORDER];
331    float zero[BLOCKSIZE], cba[BLOCKSIZE], cb1[BLOCKSIZE], cb2[BLOCKSIZE];
332    int cba_idx, cb1_idx, cb2_idx, gain;
333    int i, n;
334    unsigned m[3];
335    float g[3];
336    float error, best_error;
337
338    for (i = 0; i < LPC_ORDER; i++) {
339        work[i] = ractx->curr_sblock[BLOCKSIZE + i];
340        coefs[i] = lpc_coefs[i] * (1/4096.0);
341    }
342
343    /**
344     * Calculate the zero-input response of the LPC filter and subtract it from
345     * input data.
346     */
347    ff_celp_lp_synthesis_filterf(work + LPC_ORDER, coefs, data, BLOCKSIZE,
348                                 LPC_ORDER);
349    for (i = 0; i < BLOCKSIZE; i++) {
350        zero[i] = work[LPC_ORDER + i];
351        data[i] = sblock_data[i] - zero[i];
352    }
353
354    /**
355     * Codebook search is performed without taking into account the contribution
356     * of the previous subblock, since it has been just subtracted from input
357     * data.
358     */
359    memset(work, 0, LPC_ORDER * sizeof(*work));
360
361    cba_idx = adaptive_cb_search(ractx->adapt_cb, work + LPC_ORDER, coefs,
362                                 data);
363    if (cba_idx) {
364        /**
365         * The filtered vector from the adaptive codebook can be retrieved from
366         * work, see implementation of adaptive_cb_search().
367         */
368        memcpy(cba, work + LPC_ORDER, sizeof(cba));
369
370        ff_copy_and_dup(ractx->buffer_a, ractx->adapt_cb, cba_idx + BLOCKSIZE / 2 - 1);
371        m[0] = (ff_irms(&ractx->adsp, ractx->buffer_a) * rms) >> 12;
372    }
373    fixed_cb_search(work + LPC_ORDER, coefs, data, cba_idx, &cb1_idx, &cb2_idx);
374    for (i = 0; i < BLOCKSIZE; i++) {
375        cb1[i] = ff_cb1_vects[cb1_idx][i];
376        cb2[i] = ff_cb2_vects[cb2_idx][i];
377    }
378    ff_celp_lp_synthesis_filterf(work + LPC_ORDER, coefs, cb1, BLOCKSIZE,
379                                 LPC_ORDER);
380    memcpy(cb1, work + LPC_ORDER, sizeof(cb1));
381    m[1] = (ff_cb1_base[cb1_idx] * rms) >> 8;
382    ff_celp_lp_synthesis_filterf(work + LPC_ORDER, coefs, cb2, BLOCKSIZE,
383                                 LPC_ORDER);
384    memcpy(cb2, work + LPC_ORDER, sizeof(cb2));
385    m[2] = (ff_cb2_base[cb2_idx] * rms) >> 8;
386    best_error = FLT_MAX;
387    gain = 0;
388    for (n = 0; n < 256; n++) {
389        g[1] = ((ff_gain_val_tab[n][1] * m[1]) >> ff_gain_exp_tab[n]) *
390               (1/4096.0);
391        g[2] = ((ff_gain_val_tab[n][2] * m[2]) >> ff_gain_exp_tab[n]) *
392               (1/4096.0);
393        error = 0;
394        if (cba_idx) {
395            g[0] = ((ff_gain_val_tab[n][0] * m[0]) >> ff_gain_exp_tab[n]) *
396                   (1/4096.0);
397            for (i = 0; i < BLOCKSIZE; i++) {
398                data[i] = zero[i] + g[0] * cba[i] + g[1] * cb1[i] +
399                          g[2] * cb2[i];
400                error += (data[i] - sblock_data[i]) *
401                         (data[i] - sblock_data[i]);
402            }
403        } else {
404            for (i = 0; i < BLOCKSIZE; i++) {
405                data[i] = zero[i] + g[1] * cb1[i] + g[2] * cb2[i];
406                error += (data[i] - sblock_data[i]) *
407                         (data[i] - sblock_data[i]);
408            }
409        }
410        if (error < best_error) {
411            best_error = error;
412            gain = n;
413        }
414    }
415    put_bits(pb, 7, cba_idx);
416    put_bits(pb, 8, gain);
417    put_bits(pb, 7, cb1_idx);
418    put_bits(pb, 7, cb2_idx);
419    ff_subblock_synthesis(ractx, lpc_coefs, cba_idx, cb1_idx, cb2_idx, rms,
420                          gain);
421}
422
423
424static int ra144_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
425                              const AVFrame *frame, int *got_packet_ptr)
426{
427    static const uint8_t sizes[LPC_ORDER] = {64, 32, 32, 16, 16, 8, 8, 8, 8, 4};
428    static const uint8_t bit_sizes[LPC_ORDER] = {6, 5, 5, 4, 4, 3, 3, 3, 3, 2};
429    RA144Context *ractx = avctx->priv_data;
430    PutBitContext pb;
431    int32_t lpc_data[NBLOCKS * BLOCKSIZE];
432    int32_t lpc_coefs[LPC_ORDER][MAX_LPC_ORDER];
433    int shift[LPC_ORDER];
434    int16_t block_coefs[NBLOCKS][LPC_ORDER];
435    int lpc_refl[LPC_ORDER];    /**< reflection coefficients of the frame */
436    unsigned int refl_rms[NBLOCKS]; /**< RMS of the reflection coefficients */
437    const int16_t *samples = frame ? (const int16_t *)frame->data[0] : NULL;
438    int energy = 0;
439    int i, idx, ret;
440
441    if (ractx->last_frame)
442        return 0;
443
444    if ((ret = ff_get_encode_buffer(avctx, avpkt, FRAME_SIZE, 0)) < 0)
445        return ret;
446
447    /**
448     * Since the LPC coefficients are calculated on a frame centered over the
449     * fourth subframe, to encode a given frame, data from the next frame is
450     * needed. In each call to this function, the previous frame (whose data are
451     * saved in the encoder context) is encoded, and data from the current frame
452     * are saved in the encoder context to be used in the next function call.
453     */
454    for (i = 0; i < (2 * BLOCKSIZE + BLOCKSIZE / 2); i++) {
455        lpc_data[i] = ractx->curr_block[BLOCKSIZE + BLOCKSIZE / 2 + i];
456        energy += (lpc_data[i] * lpc_data[i]) >> 4;
457    }
458    if (frame) {
459        int j;
460        for (j = 0; j < frame->nb_samples && i < NBLOCKS * BLOCKSIZE; i++, j++) {
461            lpc_data[i] = samples[j] >> 2;
462            energy += (lpc_data[i] * lpc_data[i]) >> 4;
463        }
464    }
465    if (i < NBLOCKS * BLOCKSIZE)
466        memset(&lpc_data[i], 0, (NBLOCKS * BLOCKSIZE - i) * sizeof(*lpc_data));
467    energy = ff_energy_tab[quantize(ff_t_sqrt(energy >> 5) >> 10, ff_energy_tab,
468                                    32)];
469
470    ff_lpc_calc_coefs(&ractx->lpc_ctx, lpc_data, NBLOCKS * BLOCKSIZE, LPC_ORDER,
471                      LPC_ORDER, 16, lpc_coefs, shift, FF_LPC_TYPE_LEVINSON,
472                      0, ORDER_METHOD_EST, 0, 12, 0);
473    for (i = 0; i < LPC_ORDER; i++)
474        block_coefs[NBLOCKS - 1][i] = -lpc_coefs[LPC_ORDER - 1][i]
475                                       * (1 << (12 - shift[LPC_ORDER - 1]));
476
477    /**
478     * TODO: apply perceptual weighting of the input speech through bandwidth
479     * expansion of the LPC filter.
480     */
481
482    if (ff_eval_refl(lpc_refl, block_coefs[NBLOCKS - 1], avctx)) {
483        /**
484         * The filter is unstable: use the coefficients of the previous frame.
485         */
486        ff_int_to_int16(block_coefs[NBLOCKS - 1], ractx->lpc_coef[1]);
487        if (ff_eval_refl(lpc_refl, block_coefs[NBLOCKS - 1], avctx)) {
488            /* the filter is still unstable. set reflection coeffs to zero. */
489            memset(lpc_refl, 0, sizeof(lpc_refl));
490        }
491    }
492    init_put_bits(&pb, avpkt->data, avpkt->size);
493    for (i = 0; i < LPC_ORDER; i++) {
494        idx = quantize(lpc_refl[i], ff_lpc_refl_cb[i], sizes[i]);
495        put_bits(&pb, bit_sizes[i], idx);
496        lpc_refl[i] = ff_lpc_refl_cb[i][idx];
497    }
498    ractx->lpc_refl_rms[0] = ff_rms(lpc_refl);
499    ff_eval_coefs(ractx->lpc_coef[0], lpc_refl);
500    refl_rms[0] = ff_interp(ractx, block_coefs[0], 1, 1, ractx->old_energy);
501    refl_rms[1] = ff_interp(ractx, block_coefs[1], 2,
502                            energy <= ractx->old_energy,
503                            ff_t_sqrt(energy * ractx->old_energy) >> 12);
504    refl_rms[2] = ff_interp(ractx, block_coefs[2], 3, 0, energy);
505    refl_rms[3] = ff_rescale_rms(ractx->lpc_refl_rms[0], energy);
506    ff_int_to_int16(block_coefs[NBLOCKS - 1], ractx->lpc_coef[0]);
507    put_bits(&pb, 5, quantize(energy, ff_energy_tab, 32));
508    for (i = 0; i < NBLOCKS; i++)
509        ra144_encode_subblock(ractx, ractx->curr_block + i * BLOCKSIZE,
510                              block_coefs[i], refl_rms[i], &pb);
511    flush_put_bits(&pb);
512    ractx->old_energy = energy;
513    ractx->lpc_refl_rms[1] = ractx->lpc_refl_rms[0];
514    FFSWAP(unsigned int *, ractx->lpc_coef[0], ractx->lpc_coef[1]);
515
516    /* copy input samples to current block for processing in next call */
517    i = 0;
518    if (frame) {
519        for (; i < frame->nb_samples; i++)
520            ractx->curr_block[i] = samples[i] >> 2;
521
522        if ((ret = ff_af_queue_add(&ractx->afq, frame)) < 0)
523            return ret;
524    } else
525        ractx->last_frame = 1;
526    memset(&ractx->curr_block[i], 0,
527           (NBLOCKS * BLOCKSIZE - i) * sizeof(*ractx->curr_block));
528
529    /* Get the next frame pts/duration */
530    ff_af_queue_remove(&ractx->afq, avctx->frame_size, &avpkt->pts,
531                       &avpkt->duration);
532
533    *got_packet_ptr = 1;
534    return 0;
535}
536
537
538const FFCodec ff_ra_144_encoder = {
539    .p.name         = "real_144",
540    .p.long_name    = NULL_IF_CONFIG_SMALL("RealAudio 1.0 (14.4K)"),
541    .p.type         = AVMEDIA_TYPE_AUDIO,
542    .p.id           = AV_CODEC_ID_RA_144,
543    .p.capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_DELAY |
544                      AV_CODEC_CAP_SMALL_LAST_FRAME,
545    .priv_data_size = sizeof(RA144Context),
546    .init           = ra144_encode_init,
547    FF_CODEC_ENCODE_CB(ra144_encode_frame),
548    .close          = ra144_encode_close,
549    .p.sample_fmts  = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_S16,
550                                                     AV_SAMPLE_FMT_NONE },
551    .p.supported_samplerates = (const int[]){ 8000, 0 },
552#if FF_API_OLD_CHANNEL_LAYOUT
553    .p.channel_layouts = (const uint64_t[]) { AV_CH_LAYOUT_MONO, 0 },
554#endif
555    .p.ch_layouts   = (const AVChannelLayout[]){ AV_CHANNEL_LAYOUT_MONO, { 0 } },
556    .caps_internal  = FF_CODEC_CAP_INIT_THREADSAFE,
557};
558