2 * SIPR / ACELP.NET decoder
4 * Copyright (c) 2008 Vladimir Voroshilov
5 * Copyright (c) 2009 Vitor Sessak
7 * This file is part of FFmpeg.
9 * FFmpeg is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License as published by the Free Software Foundation; either
12 * version 2.1 of the License, or (at your option) any later version.
14 * FFmpeg is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with FFmpeg; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
27 #include "libavutil/mathematics.h"
29 #define ALT_BITSTREAM_READER_LE
34 #include "celp_math.h"
35 #include "acelp_vectors.h"
36 #include "acelp_pitch_delay.h"
37 #include "acelp_filters.h"
38 #include "celp_filters.h"
40 #define MAX_SUBFRAME_COUNT 5
46 const char *mode_name;
47 uint16_t bits_per_frame;
48 uint8_t subframe_count;
49 uint8_t frames_per_packet;
50 float pitch_sharp_factor;
52 /* bitstream parameters */
53 uint8_t number_of_fc_indexes;
54 uint8_t ma_predictor_bits; ///< size in bits of the switched MA predictor
56 /** size in bits of the i-th stage vector of quantizer */
57 uint8_t vq_indexes_bits[5];
59 /** size in bits of the adaptive-codebook index for every subframe */
60 uint8_t pitch_delay_bits[5];
62 uint8_t gp_index_bits;
63 uint8_t fc_index_bits[10]; ///< size in bits of the fixed codebook indexes
64 uint8_t gc_index_bits; ///< size in bits of the gain codebook indexes
67 static const SiprModeParam modes[MODE_COUNT] = {
70 .bits_per_frame = 160,
71 .subframe_count = SUBFRAME_COUNT_16k,
72 .frames_per_packet = 1,
73 .pitch_sharp_factor = 0.00,
75 .number_of_fc_indexes = 10,
76 .ma_predictor_bits = 1,
77 .vq_indexes_bits = {7, 8, 7, 7, 7},
78 .pitch_delay_bits = {9, 6},
80 .fc_index_bits = {4, 5, 4, 5, 4, 5, 4, 5, 4, 5},
86 .bits_per_frame = 152,
88 .frames_per_packet = 1,
89 .pitch_sharp_factor = 0.8,
91 .number_of_fc_indexes = 3,
92 .ma_predictor_bits = 0,
93 .vq_indexes_bits = {6, 7, 7, 7, 5},
94 .pitch_delay_bits = {8, 5, 5},
96 .fc_index_bits = {9, 9, 9},
102 .bits_per_frame = 232,
104 .frames_per_packet = 2,
105 .pitch_sharp_factor = 0.8,
107 .number_of_fc_indexes = 3,
108 .ma_predictor_bits = 0,
109 .vq_indexes_bits = {6, 7, 7, 7, 5},
110 .pitch_delay_bits = {8, 5, 5},
112 .fc_index_bits = {5, 5, 5},
118 .bits_per_frame = 296,
120 .frames_per_packet = 2,
121 .pitch_sharp_factor = 0.85,
123 .number_of_fc_indexes = 1,
124 .ma_predictor_bits = 0,
125 .vq_indexes_bits = {6, 7, 7, 7, 5},
126 .pitch_delay_bits = {8, 5, 8, 5, 5},
128 .fc_index_bits = {10},
133 const float ff_pow_0_5[] = {
134 1.0/(1 << 1), 1.0/(1 << 2), 1.0/(1 << 3), 1.0/(1 << 4),
135 1.0/(1 << 5), 1.0/(1 << 6), 1.0/(1 << 7), 1.0/(1 << 8),
136 1.0/(1 << 9), 1.0/(1 << 10), 1.0/(1 << 11), 1.0/(1 << 12),
137 1.0/(1 << 13), 1.0/(1 << 14), 1.0/(1 << 15), 1.0/(1 << 16)
140 static void dequant(float *out, const int *idx, const float *cbs[])
146 for (i = 0; i < num_vec; i++)
147 memcpy(out + stride*i, cbs[i] + stride*idx[i], stride*sizeof(float));
151 static void lsf_decode_fp(float *lsfnew, float *lsf_history,
152 const SiprParameters *parm)
155 float lsf_tmp[LP_FILTER_ORDER];
157 dequant(lsf_tmp, parm->vq_indexes, lsf_codebooks);
159 for (i = 0; i < LP_FILTER_ORDER; i++)
160 lsfnew[i] = lsf_history[i] * 0.33 + lsf_tmp[i] + mean_lsf[i];
162 ff_sort_nearly_sorted_floats(lsfnew, LP_FILTER_ORDER - 1);
164 /* Note that a minimum distance is not enforced between the last value and
165 the previous one, contrary to what is done in ff_acelp_reorder_lsf() */
166 ff_set_min_dist_lsf(lsfnew, LSFQ_DIFF_MIN, LP_FILTER_ORDER - 1);
167 lsfnew[9] = FFMIN(lsfnew[LP_FILTER_ORDER - 1], 1.3 * M_PI);
169 memcpy(lsf_history, lsf_tmp, LP_FILTER_ORDER * sizeof(*lsf_history));
171 for (i = 0; i < LP_FILTER_ORDER - 1; i++)
172 lsfnew[i] = cos(lsfnew[i]);
173 lsfnew[LP_FILTER_ORDER - 1] *= 6.153848 / M_PI;
176 /** Apply pitch lag to the fixed vector (AMR section 6.1.2). */
177 static void pitch_sharpening(int pitch_lag_int, float beta,
182 for (i = pitch_lag_int; i < SUBFR_SIZE; i++)
183 fixed_vector[i] += beta * fixed_vector[i - pitch_lag_int];
187 * Extract decoding parameters from the input bitstream.
188 * @param parms parameters structure
189 * @param pgb pointer to initialized GetBitContext structure
191 static void decode_parameters(SiprParameters* parms, GetBitContext *pgb,
192 const SiprModeParam *p)
196 parms->ma_pred_switch = get_bits(pgb, p->ma_predictor_bits);
198 for (i = 0; i < 5; i++)
199 parms->vq_indexes[i] = get_bits(pgb, p->vq_indexes_bits[i]);
201 for (i = 0; i < p->subframe_count; i++) {
202 parms->pitch_delay[i] = get_bits(pgb, p->pitch_delay_bits[i]);
203 parms->gp_index[i] = get_bits(pgb, p->gp_index_bits);
205 for (j = 0; j < p->number_of_fc_indexes; j++)
206 parms->fc_indexes[i][j] = get_bits(pgb, p->fc_index_bits[j]);
208 parms->gc_index[i] = get_bits(pgb, p->gc_index_bits);
212 static void lsp2lpc_sipr(const double *lsp, float *Az)
214 int lp_half_order = LP_FILTER_ORDER >> 1;
215 double buf[(LP_FILTER_ORDER >> 1) + 1];
216 double pa[(LP_FILTER_ORDER >> 1) + 1];
217 double *qa = buf + 1;
222 ff_lsp2polyf(lsp , pa, lp_half_order );
223 ff_lsp2polyf(lsp + 1, qa, lp_half_order - 1);
225 for (i = 1, j = LP_FILTER_ORDER - 1; i < lp_half_order; i++, j--) {
226 double paf = pa[i] * (1 + lsp[LP_FILTER_ORDER - 1]);
227 double qaf = (qa[i] - qa[i-2]) * (1 - lsp[LP_FILTER_ORDER - 1]);
228 Az[i-1] = (paf + qaf) * 0.5;
229 Az[j-1] = (paf - qaf) * 0.5;
232 Az[lp_half_order - 1] = (1.0 + lsp[LP_FILTER_ORDER - 1]) *
233 pa[lp_half_order] * 0.5;
235 Az[LP_FILTER_ORDER - 1] = lsp[LP_FILTER_ORDER - 1];
238 static void sipr_decode_lp(float *lsfnew, const float *lsfold, float *Az,
241 double lsfint[LP_FILTER_ORDER];
243 float t, t0 = 1.0 / num_subfr;
246 for (i = 0; i < num_subfr; i++) {
247 for (j = 0; j < LP_FILTER_ORDER; j++)
248 lsfint[j] = lsfold[j] * (1 - t) + t * lsfnew[j];
250 lsp2lpc_sipr(lsfint, Az);
251 Az += LP_FILTER_ORDER;
257 * Evaluate the adaptive impulse response.
259 static void eval_ir(const float *Az, int pitch_lag, float *freq,
260 float pitch_sharp_factor)
262 float tmp1[SUBFR_SIZE+1], tmp2[LP_FILTER_ORDER+1];
266 for (i = 0; i < LP_FILTER_ORDER; i++) {
267 tmp1[i+1] = Az[i] * ff_pow_0_55[i];
268 tmp2[i ] = Az[i] * ff_pow_0_7 [i];
270 memset(tmp1 + 11, 0, 37 * sizeof(float));
272 ff_celp_lp_synthesis_filterf(freq, tmp2, tmp1, SUBFR_SIZE,
275 pitch_sharpening(pitch_lag, pitch_sharp_factor, freq);
279 * Evaluate the convolution of a vector with a sparse vector.
281 static void convolute_with_sparse(float *out, const AMRFixed *pulses,
282 const float *shape, int length)
286 memset(out, 0, length*sizeof(float));
287 for (i = 0; i < pulses->n; i++)
288 for (j = pulses->x[i]; j < length; j++)
289 out[j] += pulses->y[i] * shape[j - pulses->x[i]];
293 * Apply postfilter, very similar to AMR one.
295 static void postfilter_5k0(SiprContext *ctx, const float *lpc, float *samples)
297 float buf[SUBFR_SIZE + LP_FILTER_ORDER];
298 float *pole_out = buf + LP_FILTER_ORDER;
299 float lpc_n[LP_FILTER_ORDER];
300 float lpc_d[LP_FILTER_ORDER];
303 for (i = 0; i < LP_FILTER_ORDER; i++) {
304 lpc_d[i] = lpc[i] * ff_pow_0_75[i];
305 lpc_n[i] = lpc[i] * ff_pow_0_5 [i];
308 memcpy(pole_out - LP_FILTER_ORDER, ctx->postfilter_mem,
309 LP_FILTER_ORDER*sizeof(float));
311 ff_celp_lp_synthesis_filterf(pole_out, lpc_d, samples, SUBFR_SIZE,
314 memcpy(ctx->postfilter_mem, pole_out + SUBFR_SIZE - LP_FILTER_ORDER,
315 LP_FILTER_ORDER*sizeof(float));
317 ff_tilt_compensation(&ctx->tilt_mem, 0.4, pole_out, SUBFR_SIZE);
319 memcpy(pole_out - LP_FILTER_ORDER, ctx->postfilter_mem5k0,
320 LP_FILTER_ORDER*sizeof(*pole_out));
322 memcpy(ctx->postfilter_mem5k0, pole_out + SUBFR_SIZE - LP_FILTER_ORDER,
323 LP_FILTER_ORDER*sizeof(*pole_out));
325 ff_celp_lp_zero_synthesis_filterf(samples, lpc_n, pole_out, SUBFR_SIZE,
330 static void decode_fixed_sparse(AMRFixed *fixed_sparse, const int16_t *pulses,
331 SiprMode mode, int low_gain)
337 for (i = 0; i < 3; i++) {
338 fixed_sparse->x[i] = 3 * (pulses[i] & 0xf) + i;
339 fixed_sparse->y[i] = pulses[i] & 0x10 ? -1 : 1;
344 for (i = 0; i < 3; i++) {
345 fixed_sparse->x[2*i ] = 3 * ((pulses[i] >> 4) & 0xf) + i;
346 fixed_sparse->x[2*i + 1] = 3 * ( pulses[i] & 0xf) + i;
348 fixed_sparse->y[2*i ] = (pulses[i] & 0x100) ? -1.0: 1.0;
350 fixed_sparse->y[2*i + 1] =
351 (fixed_sparse->x[2*i + 1] < fixed_sparse->x[2*i]) ?
352 -fixed_sparse->y[2*i ] : fixed_sparse->y[2*i];
360 int offset = (pulses[0] & 0x200) ? 2 : 0;
363 for (i = 0; i < 3; i++) {
364 int index = (val & 0x7) * 6 + 4 - i*2;
366 fixed_sparse->y[i] = (offset + index) & 0x3 ? -1 : 1;
367 fixed_sparse->x[i] = index;
373 int pulse_subset = (pulses[0] >> 8) & 1;
375 fixed_sparse->x[0] = ((pulses[0] >> 4) & 15) * 3 + pulse_subset;
376 fixed_sparse->x[1] = ( pulses[0] & 15) * 3 + pulse_subset + 1;
378 fixed_sparse->y[0] = pulses[0] & 0x200 ? -1 : 1;
379 fixed_sparse->y[1] = -fixed_sparse->y[0];
386 static void decode_frame(SiprContext *ctx, SiprParameters *params,
390 int subframe_count = modes[ctx->mode].subframe_count;
391 int frame_size = subframe_count * SUBFR_SIZE;
392 float Az[LP_FILTER_ORDER * MAX_SUBFRAME_COUNT];
394 float ir_buf[SUBFR_SIZE + LP_FILTER_ORDER];
395 float lsf_new[LP_FILTER_ORDER];
396 float *impulse_response = ir_buf + LP_FILTER_ORDER;
397 float *synth = ctx->synth_buf + 16; // 16 instead of LP_FILTER_ORDER for
402 memset(ir_buf, 0, LP_FILTER_ORDER * sizeof(float));
403 lsf_decode_fp(lsf_new, ctx->lsf_history, params);
405 sipr_decode_lp(lsf_new, ctx->lsp_history, Az, subframe_count);
407 memcpy(ctx->lsp_history, lsf_new, LP_FILTER_ORDER * sizeof(float));
409 excitation = ctx->excitation + PITCH_DELAY_MAX + L_INTERPOL;
411 for (i = 0; i < subframe_count; i++) {
412 float *pAz = Az + i*LP_FILTER_ORDER;
413 float fixed_vector[SUBFR_SIZE];
415 float pitch_gain, gain_code, avg_energy;
417 ff_decode_pitch_lag(&T0, &T0_frac, params->pitch_delay[i], t0_first, i,
418 ctx->mode == MODE_5k0, 6);
420 if (i == 0 || (i == 2 && ctx->mode == MODE_5k0))
423 ff_acelp_interpolatef(excitation, excitation - T0 + (T0_frac <= 0),
425 2 * ((2 + T0_frac)%3 + 1), LP_FILTER_ORDER,
428 decode_fixed_sparse(&fixed_cb, params->fc_indexes[i], ctx->mode,
429 ctx->past_pitch_gain < 0.8);
431 eval_ir(pAz, T0, impulse_response, modes[ctx->mode].pitch_sharp_factor);
433 convolute_with_sparse(fixed_vector, &fixed_cb, impulse_response,
437 (0.01 + ff_dot_productf(fixed_vector, fixed_vector, SUBFR_SIZE))/
440 ctx->past_pitch_gain = pitch_gain = gain_cb[params->gc_index[i]][0];
442 gain_code = ff_amr_set_fixed_gain(gain_cb[params->gc_index[i]][1],
443 avg_energy, ctx->energy_history,
444 34 - 15.0/(0.05*M_LN10/M_LN2),
447 ff_weighted_vector_sumf(excitation, excitation, fixed_vector,
448 pitch_gain, gain_code, SUBFR_SIZE);
450 pitch_gain *= 0.5 * pitch_gain;
451 pitch_gain = FFMIN(pitch_gain, 0.4);
453 ctx->gain_mem = 0.7 * ctx->gain_mem + 0.3 * pitch_gain;
454 ctx->gain_mem = FFMIN(ctx->gain_mem, pitch_gain);
455 gain_code *= ctx->gain_mem;
457 for (j = 0; j < SUBFR_SIZE; j++)
458 fixed_vector[j] = excitation[j] - gain_code * fixed_vector[j];
460 if (ctx->mode == MODE_5k0) {
461 postfilter_5k0(ctx, pAz, fixed_vector);
463 ff_celp_lp_synthesis_filterf(ctx->postfilter_syn5k0 + LP_FILTER_ORDER + i*SUBFR_SIZE,
464 pAz, excitation, SUBFR_SIZE,
468 ff_celp_lp_synthesis_filterf(synth + i*SUBFR_SIZE, pAz, fixed_vector,
469 SUBFR_SIZE, LP_FILTER_ORDER);
471 excitation += SUBFR_SIZE;
474 memcpy(synth - LP_FILTER_ORDER, synth + frame_size - LP_FILTER_ORDER,
475 LP_FILTER_ORDER * sizeof(float));
477 if (ctx->mode == MODE_5k0) {
478 for (i = 0; i < subframe_count; i++) {
479 float energy = ff_dot_productf(ctx->postfilter_syn5k0 + LP_FILTER_ORDER + i*SUBFR_SIZE,
480 ctx->postfilter_syn5k0 + LP_FILTER_ORDER + i*SUBFR_SIZE,
482 ff_adaptive_gain_control(&synth[i * SUBFR_SIZE],
483 &synth[i * SUBFR_SIZE], energy,
484 SUBFR_SIZE, 0.9, &ctx->postfilter_agc);
487 memcpy(ctx->postfilter_syn5k0, ctx->postfilter_syn5k0 + frame_size,
488 LP_FILTER_ORDER*sizeof(float));
490 memcpy(ctx->excitation, excitation - PITCH_DELAY_MAX - L_INTERPOL,
491 (PITCH_DELAY_MAX + L_INTERPOL) * sizeof(float));
493 ff_acelp_apply_order_2_transfer_function(out_data, synth,
494 (const float[2]) {-1.99997 , 1.000000000},
495 (const float[2]) {-1.93307352, 0.935891986},
497 ctx->highpass_filt_mem,
501 static av_cold int sipr_decoder_init(AVCodecContext * avctx)
503 SiprContext *ctx = avctx->priv_data;
506 if (avctx->bit_rate > 12200) ctx->mode = MODE_16k;
507 else if (avctx->bit_rate > 7500 ) ctx->mode = MODE_8k5;
508 else if (avctx->bit_rate > 5750 ) ctx->mode = MODE_6k5;
509 else ctx->mode = MODE_5k0;
511 av_log(avctx, AV_LOG_DEBUG, "Mode: %s\n", modes[ctx->mode].mode_name);
513 if (ctx->mode == MODE_16k)
514 ff_sipr_init_16k(ctx);
516 for (i = 0; i < LP_FILTER_ORDER; i++)
517 ctx->lsp_history[i] = cos((i+1) * M_PI / (LP_FILTER_ORDER + 1));
519 for (i = 0; i < 4; i++)
520 ctx->energy_history[i] = -14;
522 avctx->sample_fmt = SAMPLE_FMT_FLT;
524 dsputil_init(&ctx->dsp, avctx);
529 static int sipr_decode_frame(AVCodecContext *avctx, void *datap,
530 int *data_size, AVPacket *avpkt)
532 SiprContext *ctx = avctx->priv_data;
533 const uint8_t *buf=avpkt->data;
535 const SiprModeParam *mode_par = &modes[ctx->mode];
538 int subframe_size = ctx->mode == MODE_16k ? L_SUBFR_16k : SUBFR_SIZE;
542 if (avpkt->size < (mode_par->bits_per_frame >> 3)) {
543 av_log(avctx, AV_LOG_ERROR,
544 "Error processing packet: packet size (%d) too small\n",
550 if (*data_size < subframe_size * mode_par->subframe_count * sizeof(float)) {
551 av_log(avctx, AV_LOG_ERROR,
552 "Error processing packet: output buffer (%d) too small\n",
559 init_get_bits(&gb, buf, mode_par->bits_per_frame);
561 for (i = 0; i < mode_par->frames_per_packet; i++) {
562 decode_parameters(&parm, &gb, mode_par);
564 if (ctx->mode == MODE_16k)
565 ff_sipr_decode_frame_16k(ctx, &parm, data);
567 decode_frame(ctx, &parm, data);
569 data += subframe_size * mode_par->subframe_count;
572 *data_size = mode_par->frames_per_packet * subframe_size *
573 mode_par->subframe_count * sizeof(float);
575 return mode_par->bits_per_frame >> 3;
578 AVCodec sipr_decoder = {
587 .long_name = NULL_IF_CONFIG_SMALL("RealAudio SIPR / ACELP.NET"),