1/* 2 * SIPR / ACELP.NET decoder 3 * 4 * Copyright (c) 2008 Vladimir Voroshilov 5 * Copyright (c) 2009 Vitor Sessak 6 * 7 * This file is part of FFmpeg. 8 * 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. 13 * 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. 18 * 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 22 */ 23 24#include <math.h> 25#include <stdint.h> 26#include <string.h> 27 28#include "libavutil/channel_layout.h" 29#include "libavutil/float_dsp.h" 30#include "libavutil/mathematics.h" 31#include "avcodec.h" 32#define BITSTREAM_READER_LE 33#include "get_bits.h" 34#include "internal.h" 35 36#include "lsp.h" 37#include "acelp_vectors.h" 38#include "acelp_pitch_delay.h" 39#include "acelp_filters.h" 40#include "celp_filters.h" 41 42#define MAX_SUBFRAME_COUNT 5 43 44#include "sipr.h" 45#include "siprdata.h" 46 47typedef struct { 48 const char *mode_name; 49 uint16_t bits_per_frame; 50 uint8_t subframe_count; 51 uint8_t frames_per_packet; 52 float pitch_sharp_factor; 53 54 /* bitstream parameters */ 55 uint8_t number_of_fc_indexes; 56 uint8_t ma_predictor_bits; ///< size in bits of the switched MA predictor 57 58 /** size in bits of the i-th stage vector of quantizer */ 59 uint8_t vq_indexes_bits[5]; 60 61 /** size in bits of the adaptive-codebook index for every subframe */ 62 uint8_t pitch_delay_bits[5]; 63 64 uint8_t gp_index_bits; 65 uint8_t fc_index_bits[10]; ///< size in bits of the fixed codebook indexes 66 uint8_t gc_index_bits; ///< size in bits of the gain codebook indexes 67} SiprModeParam; 68 69static const SiprModeParam modes[MODE_COUNT] = { 70 [MODE_16k] = { 71 .mode_name = "16k", 72 .bits_per_frame = 160, 73 .subframe_count = SUBFRAME_COUNT_16k, 74 .frames_per_packet = 1, 75 .pitch_sharp_factor = 0.00, 76 77 .number_of_fc_indexes = 10, 78 .ma_predictor_bits = 1, 79 .vq_indexes_bits = {7, 8, 7, 7, 7}, 80 .pitch_delay_bits = {9, 6}, 81 .gp_index_bits = 4, 82 .fc_index_bits = {4, 5, 4, 5, 4, 5, 4, 5, 4, 5}, 83 .gc_index_bits = 5 84 }, 85 86 [MODE_8k5] = { 87 .mode_name = "8k5", 88 .bits_per_frame = 152, 89 .subframe_count = 3, 90 .frames_per_packet = 1, 91 .pitch_sharp_factor = 0.8, 92 93 .number_of_fc_indexes = 3, 94 .ma_predictor_bits = 0, 95 .vq_indexes_bits = {6, 7, 7, 7, 5}, 96 .pitch_delay_bits = {8, 5, 5}, 97 .gp_index_bits = 0, 98 .fc_index_bits = {9, 9, 9}, 99 .gc_index_bits = 7 100 }, 101 102 [MODE_6k5] = { 103 .mode_name = "6k5", 104 .bits_per_frame = 232, 105 .subframe_count = 3, 106 .frames_per_packet = 2, 107 .pitch_sharp_factor = 0.8, 108 109 .number_of_fc_indexes = 3, 110 .ma_predictor_bits = 0, 111 .vq_indexes_bits = {6, 7, 7, 7, 5}, 112 .pitch_delay_bits = {8, 5, 5}, 113 .gp_index_bits = 0, 114 .fc_index_bits = {5, 5, 5}, 115 .gc_index_bits = 7 116 }, 117 118 [MODE_5k0] = { 119 .mode_name = "5k0", 120 .bits_per_frame = 296, 121 .subframe_count = 5, 122 .frames_per_packet = 2, 123 .pitch_sharp_factor = 0.85, 124 125 .number_of_fc_indexes = 1, 126 .ma_predictor_bits = 0, 127 .vq_indexes_bits = {6, 7, 7, 7, 5}, 128 .pitch_delay_bits = {8, 5, 8, 5, 5}, 129 .gp_index_bits = 0, 130 .fc_index_bits = {10}, 131 .gc_index_bits = 7 132 } 133}; 134 135const float ff_pow_0_5[] = { 136 1.0/(1 << 1), 1.0/(1 << 2), 1.0/(1 << 3), 1.0/(1 << 4), 137 1.0/(1 << 5), 1.0/(1 << 6), 1.0/(1 << 7), 1.0/(1 << 8), 138 1.0/(1 << 9), 1.0/(1 << 10), 1.0/(1 << 11), 1.0/(1 << 12), 139 1.0/(1 << 13), 1.0/(1 << 14), 1.0/(1 << 15), 1.0/(1 << 16) 140}; 141 142static void dequant(float *out, const int *idx, const float *cbs[]) 143{ 144 int i; 145 int stride = 2; 146 int num_vec = 5; 147 148 for (i = 0; i < num_vec; i++) 149 memcpy(out + stride*i, cbs[i] + stride*idx[i], stride*sizeof(float)); 150 151} 152 153static void lsf_decode_fp(float *lsfnew, float *lsf_history, 154 const SiprParameters *parm) 155{ 156 int i; 157 float lsf_tmp[LP_FILTER_ORDER]; 158 159 dequant(lsf_tmp, parm->vq_indexes, lsf_codebooks); 160 161 for (i = 0; i < LP_FILTER_ORDER; i++) 162 lsfnew[i] = lsf_history[i] * 0.33 + lsf_tmp[i] + mean_lsf[i]; 163 164 ff_sort_nearly_sorted_floats(lsfnew, LP_FILTER_ORDER - 1); 165 166 /* Note that a minimum distance is not enforced between the last value and 167 the previous one, contrary to what is done in ff_acelp_reorder_lsf() */ 168 ff_set_min_dist_lsf(lsfnew, LSFQ_DIFF_MIN, LP_FILTER_ORDER - 1); 169 lsfnew[9] = FFMIN(lsfnew[LP_FILTER_ORDER - 1], 1.3 * M_PI); 170 171 memcpy(lsf_history, lsf_tmp, LP_FILTER_ORDER * sizeof(*lsf_history)); 172 173 for (i = 0; i < LP_FILTER_ORDER - 1; i++) 174 lsfnew[i] = cos(lsfnew[i]); 175 lsfnew[LP_FILTER_ORDER - 1] *= 6.153848 / M_PI; 176} 177 178/** Apply pitch lag to the fixed vector (AMR section 6.1.2). */ 179static void pitch_sharpening(int pitch_lag_int, float beta, 180 float *fixed_vector) 181{ 182 int i; 183 184 for (i = pitch_lag_int; i < SUBFR_SIZE; i++) 185 fixed_vector[i] += beta * fixed_vector[i - pitch_lag_int]; 186} 187 188/** 189 * Extract decoding parameters from the input bitstream. 190 * @param parms parameters structure 191 * @param pgb pointer to initialized GetBitContext structure 192 */ 193static void decode_parameters(SiprParameters* parms, GetBitContext *pgb, 194 const SiprModeParam *p) 195{ 196 int i, j; 197 198 if (p->ma_predictor_bits) 199 parms->ma_pred_switch = get_bits(pgb, p->ma_predictor_bits); 200 201 for (i = 0; i < 5; i++) 202 parms->vq_indexes[i] = get_bits(pgb, p->vq_indexes_bits[i]); 203 204 for (i = 0; i < p->subframe_count; i++) { 205 parms->pitch_delay[i] = get_bits(pgb, p->pitch_delay_bits[i]); 206 if (p->gp_index_bits) 207 parms->gp_index[i] = get_bits(pgb, p->gp_index_bits); 208 209 for (j = 0; j < p->number_of_fc_indexes; j++) 210 parms->fc_indexes[i][j] = get_bits(pgb, p->fc_index_bits[j]); 211 212 parms->gc_index[i] = get_bits(pgb, p->gc_index_bits); 213 } 214} 215 216static void sipr_decode_lp(float *lsfnew, const float *lsfold, float *Az, 217 int num_subfr) 218{ 219 double lsfint[LP_FILTER_ORDER]; 220 int i,j; 221 float t, t0 = 1.0 / num_subfr; 222 223 t = t0 * 0.5; 224 for (i = 0; i < num_subfr; i++) { 225 for (j = 0; j < LP_FILTER_ORDER; j++) 226 lsfint[j] = lsfold[j] * (1 - t) + t * lsfnew[j]; 227 228 ff_amrwb_lsp2lpc(lsfint, Az, LP_FILTER_ORDER); 229 Az += LP_FILTER_ORDER; 230 t += t0; 231 } 232} 233 234/** 235 * Evaluate the adaptive impulse response. 236 */ 237static void eval_ir(const float *Az, int pitch_lag, float *freq, 238 float pitch_sharp_factor) 239{ 240 float tmp1[SUBFR_SIZE+1], tmp2[LP_FILTER_ORDER+1]; 241 int i; 242 243 tmp1[0] = 1.0; 244 for (i = 0; i < LP_FILTER_ORDER; i++) { 245 tmp1[i+1] = Az[i] * ff_pow_0_55[i]; 246 tmp2[i ] = Az[i] * ff_pow_0_7 [i]; 247 } 248 memset(tmp1 + 11, 0, 37 * sizeof(float)); 249 250 ff_celp_lp_synthesis_filterf(freq, tmp2, tmp1, SUBFR_SIZE, 251 LP_FILTER_ORDER); 252 253 pitch_sharpening(pitch_lag, pitch_sharp_factor, freq); 254} 255 256/** 257 * Evaluate the convolution of a vector with a sparse vector. 258 */ 259static void convolute_with_sparse(float *out, const AMRFixed *pulses, 260 const float *shape, int length) 261{ 262 int i, j; 263 264 memset(out, 0, length*sizeof(float)); 265 for (i = 0; i < pulses->n; i++) 266 for (j = pulses->x[i]; j < length; j++) 267 out[j] += pulses->y[i] * shape[j - pulses->x[i]]; 268} 269 270/** 271 * Apply postfilter, very similar to AMR one. 272 */ 273static void postfilter_5k0(SiprContext *ctx, const float *lpc, float *samples) 274{ 275 float buf[SUBFR_SIZE + LP_FILTER_ORDER]; 276 float *pole_out = buf + LP_FILTER_ORDER; 277 float lpc_n[LP_FILTER_ORDER]; 278 float lpc_d[LP_FILTER_ORDER]; 279 int i; 280 281 for (i = 0; i < LP_FILTER_ORDER; i++) { 282 lpc_d[i] = lpc[i] * ff_pow_0_75[i]; 283 lpc_n[i] = lpc[i] * ff_pow_0_5 [i]; 284 }; 285 286 memcpy(pole_out - LP_FILTER_ORDER, ctx->postfilter_mem, 287 LP_FILTER_ORDER*sizeof(float)); 288 289 ff_celp_lp_synthesis_filterf(pole_out, lpc_d, samples, SUBFR_SIZE, 290 LP_FILTER_ORDER); 291 292 memcpy(ctx->postfilter_mem, pole_out + SUBFR_SIZE - LP_FILTER_ORDER, 293 LP_FILTER_ORDER*sizeof(float)); 294 295 ff_tilt_compensation(&ctx->tilt_mem, 0.4, pole_out, SUBFR_SIZE); 296 297 memcpy(pole_out - LP_FILTER_ORDER, ctx->postfilter_mem5k0, 298 LP_FILTER_ORDER*sizeof(*pole_out)); 299 300 memcpy(ctx->postfilter_mem5k0, pole_out + SUBFR_SIZE - LP_FILTER_ORDER, 301 LP_FILTER_ORDER*sizeof(*pole_out)); 302 303 ff_celp_lp_zero_synthesis_filterf(samples, lpc_n, pole_out, SUBFR_SIZE, 304 LP_FILTER_ORDER); 305 306} 307 308static void decode_fixed_sparse(AMRFixed *fixed_sparse, const int16_t *pulses, 309 SiprMode mode, int low_gain) 310{ 311 int i; 312 313 switch (mode) { 314 case MODE_6k5: 315 for (i = 0; i < 3; i++) { 316 fixed_sparse->x[i] = 3 * (pulses[i] & 0xf) + i; 317 fixed_sparse->y[i] = pulses[i] & 0x10 ? -1 : 1; 318 } 319 fixed_sparse->n = 3; 320 break; 321 case MODE_8k5: 322 for (i = 0; i < 3; i++) { 323 fixed_sparse->x[2*i ] = 3 * ((pulses[i] >> 4) & 0xf) + i; 324 fixed_sparse->x[2*i + 1] = 3 * ( pulses[i] & 0xf) + i; 325 326 fixed_sparse->y[2*i ] = (pulses[i] & 0x100) ? -1.0: 1.0; 327 328 fixed_sparse->y[2*i + 1] = 329 (fixed_sparse->x[2*i + 1] < fixed_sparse->x[2*i]) ? 330 -fixed_sparse->y[2*i ] : fixed_sparse->y[2*i]; 331 } 332 333 fixed_sparse->n = 6; 334 break; 335 case MODE_5k0: 336 default: 337 if (low_gain) { 338 int offset = (pulses[0] & 0x200) ? 2 : 0; 339 int val = pulses[0]; 340 341 for (i = 0; i < 3; i++) { 342 int index = (val & 0x7) * 6 + 4 - i*2; 343 344 fixed_sparse->y[i] = (offset + index) & 0x3 ? -1 : 1; 345 fixed_sparse->x[i] = index; 346 347 val >>= 3; 348 } 349 fixed_sparse->n = 3; 350 } else { 351 int pulse_subset = (pulses[0] >> 8) & 1; 352 353 fixed_sparse->x[0] = ((pulses[0] >> 4) & 15) * 3 + pulse_subset; 354 fixed_sparse->x[1] = ( pulses[0] & 15) * 3 + pulse_subset + 1; 355 356 fixed_sparse->y[0] = pulses[0] & 0x200 ? -1 : 1; 357 fixed_sparse->y[1] = -fixed_sparse->y[0]; 358 fixed_sparse->n = 2; 359 } 360 break; 361 } 362} 363 364static void decode_frame(SiprContext *ctx, SiprParameters *params, 365 float *out_data) 366{ 367 int i, j; 368 int subframe_count = modes[ctx->mode].subframe_count; 369 int frame_size = subframe_count * SUBFR_SIZE; 370 float Az[LP_FILTER_ORDER * MAX_SUBFRAME_COUNT]; 371 float *excitation; 372 float ir_buf[SUBFR_SIZE + LP_FILTER_ORDER]; 373 float lsf_new[LP_FILTER_ORDER]; 374 float *impulse_response = ir_buf + LP_FILTER_ORDER; 375 float *synth = ctx->synth_buf + 16; // 16 instead of LP_FILTER_ORDER for 376 // memory alignment 377 int t0_first = 0; 378 AMRFixed fixed_cb; 379 380 memset(ir_buf, 0, LP_FILTER_ORDER * sizeof(float)); 381 lsf_decode_fp(lsf_new, ctx->lsf_history, params); 382 383 sipr_decode_lp(lsf_new, ctx->lsp_history, Az, subframe_count); 384 385 memcpy(ctx->lsp_history, lsf_new, LP_FILTER_ORDER * sizeof(float)); 386 387 excitation = ctx->excitation + PITCH_DELAY_MAX + L_INTERPOL; 388 389 for (i = 0; i < subframe_count; i++) { 390 float *pAz = Az + i*LP_FILTER_ORDER; 391 float fixed_vector[SUBFR_SIZE]; 392 int T0,T0_frac; 393 float pitch_gain, gain_code, avg_energy; 394 395 ff_decode_pitch_lag(&T0, &T0_frac, params->pitch_delay[i], t0_first, i, 396 ctx->mode == MODE_5k0, 6); 397 398 if (i == 0 || (i == 2 && ctx->mode == MODE_5k0)) 399 t0_first = T0; 400 401 ff_acelp_interpolatef(excitation, excitation - T0 + (T0_frac <= 0), 402 ff_b60_sinc, 6, 403 2 * ((2 + T0_frac)%3 + 1), LP_FILTER_ORDER, 404 SUBFR_SIZE); 405 406 decode_fixed_sparse(&fixed_cb, params->fc_indexes[i], ctx->mode, 407 ctx->past_pitch_gain < 0.8); 408 409 eval_ir(pAz, T0, impulse_response, modes[ctx->mode].pitch_sharp_factor); 410 411 convolute_with_sparse(fixed_vector, &fixed_cb, impulse_response, 412 SUBFR_SIZE); 413 414 avg_energy = (0.01 + avpriv_scalarproduct_float_c(fixed_vector, 415 fixed_vector, 416 SUBFR_SIZE)) / 417 SUBFR_SIZE; 418 419 ctx->past_pitch_gain = pitch_gain = gain_cb[params->gc_index[i]][0]; 420 421 gain_code = ff_amr_set_fixed_gain(gain_cb[params->gc_index[i]][1], 422 avg_energy, ctx->energy_history, 423 34 - 15.0/(0.05*M_LN10/M_LN2), 424 pred); 425 426 ff_weighted_vector_sumf(excitation, excitation, fixed_vector, 427 pitch_gain, gain_code, SUBFR_SIZE); 428 429 pitch_gain *= 0.5 * pitch_gain; 430 pitch_gain = FFMIN(pitch_gain, 0.4); 431 432 ctx->gain_mem = 0.7 * ctx->gain_mem + 0.3 * pitch_gain; 433 ctx->gain_mem = FFMIN(ctx->gain_mem, pitch_gain); 434 gain_code *= ctx->gain_mem; 435 436 for (j = 0; j < SUBFR_SIZE; j++) 437 fixed_vector[j] = excitation[j] - gain_code * fixed_vector[j]; 438 439 if (ctx->mode == MODE_5k0) { 440 postfilter_5k0(ctx, pAz, fixed_vector); 441 442 ff_celp_lp_synthesis_filterf(ctx->postfilter_syn5k0 + LP_FILTER_ORDER + i*SUBFR_SIZE, 443 pAz, excitation, SUBFR_SIZE, 444 LP_FILTER_ORDER); 445 } 446 447 ff_celp_lp_synthesis_filterf(synth + i*SUBFR_SIZE, pAz, fixed_vector, 448 SUBFR_SIZE, LP_FILTER_ORDER); 449 450 excitation += SUBFR_SIZE; 451 } 452 453 memcpy(synth - LP_FILTER_ORDER, synth + frame_size - LP_FILTER_ORDER, 454 LP_FILTER_ORDER * sizeof(float)); 455 456 if (ctx->mode == MODE_5k0) { 457 for (i = 0; i < subframe_count; i++) { 458 float energy = avpriv_scalarproduct_float_c(ctx->postfilter_syn5k0 + LP_FILTER_ORDER + i * SUBFR_SIZE, 459 ctx->postfilter_syn5k0 + LP_FILTER_ORDER + i * SUBFR_SIZE, 460 SUBFR_SIZE); 461 ff_adaptive_gain_control(&synth[i * SUBFR_SIZE], 462 &synth[i * SUBFR_SIZE], energy, 463 SUBFR_SIZE, 0.9, &ctx->postfilter_agc); 464 } 465 466 memcpy(ctx->postfilter_syn5k0, ctx->postfilter_syn5k0 + frame_size, 467 LP_FILTER_ORDER*sizeof(float)); 468 } 469 memmove(ctx->excitation, excitation - PITCH_DELAY_MAX - L_INTERPOL, 470 (PITCH_DELAY_MAX + L_INTERPOL) * sizeof(float)); 471 472 ff_acelp_apply_order_2_transfer_function(out_data, synth, 473 (const float[2]) {-1.99997 , 1.000000000}, 474 (const float[2]) {-1.93307352, 0.935891986}, 475 0.939805806, 476 ctx->highpass_filt_mem, 477 frame_size); 478} 479 480static av_cold int sipr_decoder_init(AVCodecContext * avctx) 481{ 482 SiprContext *ctx = avctx->priv_data; 483 int i; 484 485 switch (avctx->block_align) { 486 case 20: ctx->mode = MODE_16k; break; 487 case 19: ctx->mode = MODE_8k5; break; 488 case 29: ctx->mode = MODE_6k5; break; 489 case 37: ctx->mode = MODE_5k0; break; 490 default: 491 if (avctx->bit_rate > 12200) ctx->mode = MODE_16k; 492 else if (avctx->bit_rate > 7500 ) ctx->mode = MODE_8k5; 493 else if (avctx->bit_rate > 5750 ) ctx->mode = MODE_6k5; 494 else ctx->mode = MODE_5k0; 495 av_log(avctx, AV_LOG_WARNING, 496 "Invalid block_align: %d. Mode %s guessed based on bitrate: %d\n", 497 avctx->block_align, modes[ctx->mode].mode_name, avctx->bit_rate); 498 } 499 500 av_log(avctx, AV_LOG_DEBUG, "Mode: %s\n", modes[ctx->mode].mode_name); 501 502 if (ctx->mode == MODE_16k) { 503 ff_sipr_init_16k(ctx); 504 ctx->decode_frame = ff_sipr_decode_frame_16k; 505 } else { 506 ctx->decode_frame = decode_frame; 507 } 508 509 for (i = 0; i < LP_FILTER_ORDER; i++) 510 ctx->lsp_history[i] = cos((i+1) * M_PI / (LP_FILTER_ORDER + 1)); 511 512 for (i = 0; i < 4; i++) 513 ctx->energy_history[i] = -14; 514 515 avctx->channels = 1; 516 avctx->channel_layout = AV_CH_LAYOUT_MONO; 517 avctx->sample_fmt = AV_SAMPLE_FMT_FLT; 518 519 return 0; 520} 521 522static int sipr_decode_frame(AVCodecContext *avctx, void *data, 523 int *got_frame_ptr, AVPacket *avpkt) 524{ 525 SiprContext *ctx = avctx->priv_data; 526 AVFrame *frame = data; 527 const uint8_t *buf=avpkt->data; 528 SiprParameters parm; 529 const SiprModeParam *mode_par = &modes[ctx->mode]; 530 GetBitContext gb; 531 float *samples; 532 int subframe_size = ctx->mode == MODE_16k ? L_SUBFR_16k : SUBFR_SIZE; 533 int i, ret; 534 535 ctx->avctx = avctx; 536 if (avpkt->size < (mode_par->bits_per_frame >> 3)) { 537 av_log(avctx, AV_LOG_ERROR, 538 "Error processing packet: packet size (%d) too small\n", 539 avpkt->size); 540 return -1; 541 } 542 543 /* get output buffer */ 544 frame->nb_samples = mode_par->frames_per_packet * subframe_size * 545 mode_par->subframe_count; 546 if ((ret = ff_get_buffer(avctx, frame, 0)) < 0) 547 return ret; 548 samples = (float *)frame->data[0]; 549 550 init_get_bits(&gb, buf, mode_par->bits_per_frame); 551 552 for (i = 0; i < mode_par->frames_per_packet; i++) { 553 decode_parameters(&parm, &gb, mode_par); 554 555 ctx->decode_frame(ctx, &parm, samples); 556 557 samples += subframe_size * mode_par->subframe_count; 558 } 559 560 *got_frame_ptr = 1; 561 562 return mode_par->bits_per_frame >> 3; 563} 564 565AVCodec ff_sipr_decoder = { 566 .name = "sipr", 567 .long_name = NULL_IF_CONFIG_SMALL("RealAudio SIPR / ACELP.NET"), 568 .type = AVMEDIA_TYPE_AUDIO, 569 .id = AV_CODEC_ID_SIPR, 570 .priv_data_size = sizeof(SiprContext), 571 .init = sipr_decoder_init, 572 .decode = sipr_decode_frame, 573 .capabilities = CODEC_CAP_DR1, 574}; 575