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 32#define BITSTREAM_READER_LE 33#include "avcodec.h" 34#include "codec_internal.h" 35#include "get_bits.h" 36#include "internal.h" 37#include "lsp.h" 38#include "acelp_vectors.h" 39#include "acelp_pitch_delay.h" 40#include "acelp_filters.h" 41#include "celp_filters.h" 42 43#define MAX_SUBFRAME_COUNT 5 44 45#include "sipr.h" 46#include "siprdata.h" 47 48typedef struct SiprModeParam { 49 const char *mode_name; 50 uint16_t bits_per_frame; 51 uint8_t subframe_count; 52 uint8_t frames_per_packet; 53 float pitch_sharp_factor; 54 55 /* bitstream parameters */ 56 uint8_t number_of_fc_indexes; 57 uint8_t ma_predictor_bits; ///< size in bits of the switched MA predictor 58 59 /** size in bits of the i-th stage vector of quantizer */ 60 uint8_t vq_indexes_bits[5]; 61 62 /** size in bits of the adaptive-codebook index for every subframe */ 63 uint8_t pitch_delay_bits[5]; 64 65 uint8_t gp_index_bits; 66 uint8_t fc_index_bits[10]; ///< size in bits of the fixed codebook indexes 67 uint8_t gc_index_bits; ///< size in bits of the gain codebook indexes 68} SiprModeParam; 69 70static const SiprModeParam modes[MODE_COUNT] = { 71 [MODE_16k] = { 72 .mode_name = "16k", 73 .bits_per_frame = 160, 74 .subframe_count = SUBFRAME_COUNT_16k, 75 .frames_per_packet = 1, 76 .pitch_sharp_factor = 0.00, 77 78 .number_of_fc_indexes = 10, 79 .ma_predictor_bits = 1, 80 .vq_indexes_bits = {7, 8, 7, 7, 7}, 81 .pitch_delay_bits = {9, 6}, 82 .gp_index_bits = 4, 83 .fc_index_bits = {4, 5, 4, 5, 4, 5, 4, 5, 4, 5}, 84 .gc_index_bits = 5 85 }, 86 87 [MODE_8k5] = { 88 .mode_name = "8k5", 89 .bits_per_frame = 152, 90 .subframe_count = 3, 91 .frames_per_packet = 1, 92 .pitch_sharp_factor = 0.8, 93 94 .number_of_fc_indexes = 3, 95 .ma_predictor_bits = 0, 96 .vq_indexes_bits = {6, 7, 7, 7, 5}, 97 .pitch_delay_bits = {8, 5, 5}, 98 .gp_index_bits = 0, 99 .fc_index_bits = {9, 9, 9}, 100 .gc_index_bits = 7 101 }, 102 103 [MODE_6k5] = { 104 .mode_name = "6k5", 105 .bits_per_frame = 232, 106 .subframe_count = 3, 107 .frames_per_packet = 2, 108 .pitch_sharp_factor = 0.8, 109 110 .number_of_fc_indexes = 3, 111 .ma_predictor_bits = 0, 112 .vq_indexes_bits = {6, 7, 7, 7, 5}, 113 .pitch_delay_bits = {8, 5, 5}, 114 .gp_index_bits = 0, 115 .fc_index_bits = {5, 5, 5}, 116 .gc_index_bits = 7 117 }, 118 119 [MODE_5k0] = { 120 .mode_name = "5k0", 121 .bits_per_frame = 296, 122 .subframe_count = 5, 123 .frames_per_packet = 2, 124 .pitch_sharp_factor = 0.85, 125 126 .number_of_fc_indexes = 1, 127 .ma_predictor_bits = 0, 128 .vq_indexes_bits = {6, 7, 7, 7, 5}, 129 .pitch_delay_bits = {8, 5, 8, 5, 5}, 130 .gp_index_bits = 0, 131 .fc_index_bits = {10}, 132 .gc_index_bits = 7 133 } 134}; 135 136const float ff_pow_0_5[] = { 137 1.0/(1 << 1), 1.0/(1 << 2), 1.0/(1 << 3), 1.0/(1 << 4), 138 1.0/(1 << 5), 1.0/(1 << 6), 1.0/(1 << 7), 1.0/(1 << 8), 139 1.0/(1 << 9), 1.0/(1 << 10), 1.0/(1 << 11), 1.0/(1 << 12), 140 1.0/(1 << 13), 1.0/(1 << 14), 1.0/(1 << 15), 1.0/(1 << 16) 141}; 142 143static void dequant(float *out, const int *idx, const float * const cbs[]) 144{ 145 int i; 146 int stride = 2; 147 int num_vec = 5; 148 149 for (i = 0; i < num_vec; i++) 150 memcpy(out + stride*i, cbs[i] + stride*idx[i], stride*sizeof(float)); 151 152} 153 154static void lsf_decode_fp(float *lsfnew, float *lsf_history, 155 const SiprParameters *parm) 156{ 157 int i; 158 float lsf_tmp[LP_FILTER_ORDER]; 159 160 dequant(lsf_tmp, parm->vq_indexes, lsf_codebooks); 161 162 for (i = 0; i < LP_FILTER_ORDER; i++) 163 lsfnew[i] = lsf_history[i] * 0.33 + lsf_tmp[i] + mean_lsf[i]; 164 165 ff_sort_nearly_sorted_floats(lsfnew, LP_FILTER_ORDER - 1); 166 167 /* Note that a minimum distance is not enforced between the last value and 168 the previous one, contrary to what is done in ff_acelp_reorder_lsf() */ 169 ff_set_min_dist_lsf(lsfnew, LSFQ_DIFF_MIN, LP_FILTER_ORDER - 1); 170 lsfnew[9] = FFMIN(lsfnew[LP_FILTER_ORDER - 1], 1.3 * M_PI); 171 172 memcpy(lsf_history, lsf_tmp, LP_FILTER_ORDER * sizeof(*lsf_history)); 173 174 for (i = 0; i < LP_FILTER_ORDER - 1; i++) 175 lsfnew[i] = cos(lsfnew[i]); 176 lsfnew[LP_FILTER_ORDER - 1] *= 6.153848 / M_PI; 177} 178 179/** Apply pitch lag to the fixed vector (AMR section 6.1.2). */ 180static void pitch_sharpening(int pitch_lag_int, float beta, 181 float *fixed_vector) 182{ 183 int i; 184 185 for (i = pitch_lag_int; i < SUBFR_SIZE; i++) 186 fixed_vector[i] += beta * fixed_vector[i - pitch_lag_int]; 187} 188 189/** 190 * Extract decoding parameters from the input bitstream. 191 * @param parms parameters structure 192 * @param pgb pointer to initialized GetBitContext structure 193 */ 194static void decode_parameters(SiprParameters* parms, GetBitContext *pgb, 195 const SiprModeParam *p) 196{ 197 int i, j; 198 199 if (p->ma_predictor_bits) 200 parms->ma_pred_switch = get_bits(pgb, p->ma_predictor_bits); 201 202 for (i = 0; i < 5; i++) 203 parms->vq_indexes[i] = get_bits(pgb, p->vq_indexes_bits[i]); 204 205 for (i = 0; i < p->subframe_count; i++) { 206 parms->pitch_delay[i] = get_bits(pgb, p->pitch_delay_bits[i]); 207 if (p->gp_index_bits) 208 parms->gp_index[i] = get_bits(pgb, p->gp_index_bits); 209 210 for (j = 0; j < p->number_of_fc_indexes; j++) 211 parms->fc_indexes[i][j] = get_bits(pgb, p->fc_index_bits[j]); 212 213 parms->gc_index[i] = get_bits(pgb, p->gc_index_bits); 214 } 215} 216 217static void sipr_decode_lp(float *lsfnew, const float *lsfold, float *Az, 218 int num_subfr) 219{ 220 double lsfint[LP_FILTER_ORDER]; 221 int i,j; 222 float t, t0 = 1.0 / num_subfr; 223 224 t = t0 * 0.5; 225 for (i = 0; i < num_subfr; i++) { 226 for (j = 0; j < LP_FILTER_ORDER; j++) 227 lsfint[j] = lsfold[j] * (1 - t) + t * lsfnew[j]; 228 229 ff_amrwb_lsp2lpc(lsfint, Az, LP_FILTER_ORDER); 230 Az += LP_FILTER_ORDER; 231 t += t0; 232 } 233} 234 235/** 236 * Evaluate the adaptive impulse response. 237 */ 238static void eval_ir(const float *Az, int pitch_lag, float *freq, 239 float pitch_sharp_factor) 240{ 241 float tmp1[SUBFR_SIZE+1], tmp2[LP_FILTER_ORDER+1]; 242 int i; 243 244 tmp1[0] = 1.0; 245 for (i = 0; i < LP_FILTER_ORDER; i++) { 246 tmp1[i+1] = Az[i] * ff_pow_0_55[i]; 247 tmp2[i ] = Az[i] * ff_pow_0_7 [i]; 248 } 249 memset(tmp1 + 11, 0, 37 * sizeof(float)); 250 251 ff_celp_lp_synthesis_filterf(freq, tmp2, tmp1, SUBFR_SIZE, 252 LP_FILTER_ORDER); 253 254 pitch_sharpening(pitch_lag, pitch_sharp_factor, freq); 255} 256 257/** 258 * Evaluate the convolution of a vector with a sparse vector. 259 */ 260static void convolute_with_sparse(float *out, const AMRFixed *pulses, 261 const float *shape, int length) 262{ 263 int i, j; 264 265 memset(out, 0, length*sizeof(float)); 266 for (i = 0; i < pulses->n; i++) 267 for (j = pulses->x[i]; j < length; j++) 268 out[j] += pulses->y[i] * shape[j - pulses->x[i]]; 269} 270 271/** 272 * Apply postfilter, very similar to AMR one. 273 */ 274static void postfilter_5k0(SiprContext *ctx, const float *lpc, float *samples) 275{ 276 float buf[SUBFR_SIZE + LP_FILTER_ORDER]; 277 float *pole_out = buf + LP_FILTER_ORDER; 278 float lpc_n[LP_FILTER_ORDER]; 279 float lpc_d[LP_FILTER_ORDER]; 280 int i; 281 282 for (i = 0; i < LP_FILTER_ORDER; i++) { 283 lpc_d[i] = lpc[i] * ff_pow_0_75[i]; 284 lpc_n[i] = lpc[i] * ff_pow_0_5 [i]; 285 }; 286 287 memcpy(pole_out - LP_FILTER_ORDER, ctx->postfilter_mem, 288 LP_FILTER_ORDER*sizeof(float)); 289 290 ff_celp_lp_synthesis_filterf(pole_out, lpc_d, samples, SUBFR_SIZE, 291 LP_FILTER_ORDER); 292 293 memcpy(ctx->postfilter_mem, pole_out + SUBFR_SIZE - LP_FILTER_ORDER, 294 LP_FILTER_ORDER*sizeof(float)); 295 296 ff_tilt_compensation(&ctx->tilt_mem, 0.4, pole_out, SUBFR_SIZE); 297 298 memcpy(pole_out - LP_FILTER_ORDER, ctx->postfilter_mem5k0, 299 LP_FILTER_ORDER*sizeof(*pole_out)); 300 301 memcpy(ctx->postfilter_mem5k0, pole_out + SUBFR_SIZE - LP_FILTER_ORDER, 302 LP_FILTER_ORDER*sizeof(*pole_out)); 303 304 ff_celp_lp_zero_synthesis_filterf(samples, lpc_n, pole_out, SUBFR_SIZE, 305 LP_FILTER_ORDER); 306 307} 308 309static void decode_fixed_sparse(AMRFixed *fixed_sparse, const int16_t *pulses, 310 SiprMode mode, int low_gain) 311{ 312 int i; 313 314 switch (mode) { 315 case MODE_6k5: 316 for (i = 0; i < 3; i++) { 317 fixed_sparse->x[i] = 3 * (pulses[i] & 0xf) + i; 318 fixed_sparse->y[i] = pulses[i] & 0x10 ? -1 : 1; 319 } 320 fixed_sparse->n = 3; 321 break; 322 case MODE_8k5: 323 for (i = 0; i < 3; i++) { 324 fixed_sparse->x[2*i ] = 3 * ((pulses[i] >> 4) & 0xf) + i; 325 fixed_sparse->x[2*i + 1] = 3 * ( pulses[i] & 0xf) + i; 326 327 fixed_sparse->y[2*i ] = (pulses[i] & 0x100) ? -1.0: 1.0; 328 329 fixed_sparse->y[2*i + 1] = 330 (fixed_sparse->x[2*i + 1] < fixed_sparse->x[2*i]) ? 331 -fixed_sparse->y[2*i ] : fixed_sparse->y[2*i]; 332 } 333 334 fixed_sparse->n = 6; 335 break; 336 case MODE_5k0: 337 default: 338 if (low_gain) { 339 int offset = (pulses[0] & 0x200) ? 2 : 0; 340 int val = pulses[0]; 341 342 for (i = 0; i < 3; i++) { 343 int index = (val & 0x7) * 6 + 4 - i*2; 344 345 fixed_sparse->y[i] = (offset + index) & 0x3 ? -1 : 1; 346 fixed_sparse->x[i] = index; 347 348 val >>= 3; 349 } 350 fixed_sparse->n = 3; 351 } else { 352 int pulse_subset = (pulses[0] >> 8) & 1; 353 354 fixed_sparse->x[0] = ((pulses[0] >> 4) & 15) * 3 + pulse_subset; 355 fixed_sparse->x[1] = ( pulses[0] & 15) * 3 + pulse_subset + 1; 356 357 fixed_sparse->y[0] = pulses[0] & 0x200 ? -1 : 1; 358 fixed_sparse->y[1] = -fixed_sparse->y[0]; 359 fixed_sparse->n = 2; 360 } 361 break; 362 } 363} 364 365static void decode_frame(SiprContext *ctx, SiprParameters *params, 366 float *out_data) 367{ 368 int i, j; 369 int subframe_count = modes[ctx->mode].subframe_count; 370 int frame_size = subframe_count * SUBFR_SIZE; 371 float Az[LP_FILTER_ORDER * MAX_SUBFRAME_COUNT]; 372 float *excitation; 373 float ir_buf[SUBFR_SIZE + LP_FILTER_ORDER]; 374 float lsf_new[LP_FILTER_ORDER]; 375 float *impulse_response = ir_buf + LP_FILTER_ORDER; 376 float *synth = ctx->synth_buf + 16; // 16 instead of LP_FILTER_ORDER for 377 // memory alignment 378 int t0_first = 0; 379 AMRFixed fixed_cb; 380 381 memset(ir_buf, 0, LP_FILTER_ORDER * sizeof(float)); 382 lsf_decode_fp(lsf_new, ctx->lsf_history, params); 383 384 sipr_decode_lp(lsf_new, ctx->lsp_history, Az, subframe_count); 385 386 memcpy(ctx->lsp_history, lsf_new, LP_FILTER_ORDER * sizeof(float)); 387 388 excitation = ctx->excitation + PITCH_DELAY_MAX + L_INTERPOL; 389 390 for (i = 0; i < subframe_count; i++) { 391 float *pAz = Az + i*LP_FILTER_ORDER; 392 float fixed_vector[SUBFR_SIZE]; 393 int T0,T0_frac; 394 float pitch_gain, gain_code, avg_energy; 395 396 ff_decode_pitch_lag(&T0, &T0_frac, params->pitch_delay[i], t0_first, i, 397 ctx->mode == MODE_5k0, 6); 398 399 if (i == 0 || (i == 2 && ctx->mode == MODE_5k0)) 400 t0_first = T0; 401 402 ff_acelp_interpolatef(excitation, excitation - T0 + (T0_frac <= 0), 403 ff_b60_sinc, 6, 404 2 * ((2 + T0_frac)%3 + 1), LP_FILTER_ORDER, 405 SUBFR_SIZE); 406 407 decode_fixed_sparse(&fixed_cb, params->fc_indexes[i], ctx->mode, 408 ctx->past_pitch_gain < 0.8); 409 410 eval_ir(pAz, T0, impulse_response, modes[ctx->mode].pitch_sharp_factor); 411 412 convolute_with_sparse(fixed_vector, &fixed_cb, impulse_response, 413 SUBFR_SIZE); 414 415 avg_energy = (0.01 + avpriv_scalarproduct_float_c(fixed_vector, 416 fixed_vector, 417 SUBFR_SIZE)) / 418 SUBFR_SIZE; 419 420 ctx->past_pitch_gain = pitch_gain = gain_cb[params->gc_index[i]][0]; 421 422 gain_code = ff_amr_set_fixed_gain(gain_cb[params->gc_index[i]][1], 423 avg_energy, ctx->energy_history, 424 34 - 15.0/(0.05*M_LN10/M_LN2), 425 pred); 426 427 ff_weighted_vector_sumf(excitation, excitation, fixed_vector, 428 pitch_gain, gain_code, SUBFR_SIZE); 429 430 pitch_gain *= 0.5 * pitch_gain; 431 pitch_gain = FFMIN(pitch_gain, 0.4); 432 433 ctx->gain_mem = 0.7 * ctx->gain_mem + 0.3 * pitch_gain; 434 ctx->gain_mem = FFMIN(ctx->gain_mem, pitch_gain); 435 gain_code *= ctx->gain_mem; 436 437 for (j = 0; j < SUBFR_SIZE; j++) 438 fixed_vector[j] = excitation[j] - gain_code * fixed_vector[j]; 439 440 if (ctx->mode == MODE_5k0) { 441 postfilter_5k0(ctx, pAz, fixed_vector); 442 443 ff_celp_lp_synthesis_filterf(ctx->postfilter_syn5k0 + LP_FILTER_ORDER + i*SUBFR_SIZE, 444 pAz, excitation, SUBFR_SIZE, 445 LP_FILTER_ORDER); 446 } 447 448 ff_celp_lp_synthesis_filterf(synth + i*SUBFR_SIZE, pAz, fixed_vector, 449 SUBFR_SIZE, LP_FILTER_ORDER); 450 451 excitation += SUBFR_SIZE; 452 } 453 454 memcpy(synth - LP_FILTER_ORDER, synth + frame_size - LP_FILTER_ORDER, 455 LP_FILTER_ORDER * sizeof(float)); 456 457 if (ctx->mode == MODE_5k0) { 458 for (i = 0; i < subframe_count; i++) { 459 float energy = avpriv_scalarproduct_float_c(ctx->postfilter_syn5k0 + LP_FILTER_ORDER + i * SUBFR_SIZE, 460 ctx->postfilter_syn5k0 + LP_FILTER_ORDER + i * SUBFR_SIZE, 461 SUBFR_SIZE); 462 ff_adaptive_gain_control(&synth[i * SUBFR_SIZE], 463 &synth[i * SUBFR_SIZE], energy, 464 SUBFR_SIZE, 0.9, &ctx->postfilter_agc); 465 } 466 467 memcpy(ctx->postfilter_syn5k0, ctx->postfilter_syn5k0 + frame_size, 468 LP_FILTER_ORDER*sizeof(float)); 469 } 470 memmove(ctx->excitation, excitation - PITCH_DELAY_MAX - L_INTERPOL, 471 (PITCH_DELAY_MAX + L_INTERPOL) * sizeof(float)); 472 473 ff_acelp_apply_order_2_transfer_function(out_data, synth, 474 (const float[2]) {-1.99997 , 1.000000000}, 475 (const float[2]) {-1.93307352, 0.935891986}, 476 0.939805806, 477 ctx->highpass_filt_mem, 478 frame_size); 479} 480 481static av_cold int sipr_decoder_init(AVCodecContext * avctx) 482{ 483 SiprContext *ctx = avctx->priv_data; 484 int i; 485 486 switch (avctx->block_align) { 487 case 20: ctx->mode = MODE_16k; break; 488 case 19: ctx->mode = MODE_8k5; break; 489 case 29: ctx->mode = MODE_6k5; break; 490 case 37: ctx->mode = MODE_5k0; break; 491 default: 492 if (avctx->bit_rate > 12200) ctx->mode = MODE_16k; 493 else if (avctx->bit_rate > 7500 ) ctx->mode = MODE_8k5; 494 else if (avctx->bit_rate > 5750 ) ctx->mode = MODE_6k5; 495 else ctx->mode = MODE_5k0; 496 av_log(avctx, AV_LOG_WARNING, 497 "Invalid block_align: %d. Mode %s guessed based on bitrate: %"PRId64"\n", 498 avctx->block_align, modes[ctx->mode].mode_name, avctx->bit_rate); 499 } 500 501 av_log(avctx, AV_LOG_DEBUG, "Mode: %s\n", modes[ctx->mode].mode_name); 502 503 if (ctx->mode == MODE_16k) { 504 ff_sipr_init_16k(ctx); 505 ctx->decode_frame = ff_sipr_decode_frame_16k; 506 } else { 507 ctx->decode_frame = decode_frame; 508 } 509 510 for (i = 0; i < LP_FILTER_ORDER; i++) 511 ctx->lsp_history[i] = cos((i+1) * M_PI / (LP_FILTER_ORDER + 1)); 512 513 for (i = 0; i < 4; i++) 514 ctx->energy_history[i] = -14; 515 516 av_channel_layout_uninit(&avctx->ch_layout); 517 avctx->ch_layout = (AVChannelLayout)AV_CHANNEL_LAYOUT_MONO; 518 avctx->sample_fmt = AV_SAMPLE_FMT_FLT; 519 520 return 0; 521} 522 523static int sipr_decode_frame(AVCodecContext *avctx, AVFrame *frame, 524 int *got_frame_ptr, AVPacket *avpkt) 525{ 526 SiprContext *ctx = avctx->priv_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 AVERROR_INVALIDDATA; 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 565const FFCodec ff_sipr_decoder = { 566 .p.name = "sipr", 567 .p.long_name = NULL_IF_CONFIG_SMALL("RealAudio SIPR / ACELP.NET"), 568 .p.type = AVMEDIA_TYPE_AUDIO, 569 .p.id = AV_CODEC_ID_SIPR, 570 .priv_data_size = sizeof(SiprContext), 571 .init = sipr_decoder_init, 572 FF_CODEC_DECODE_CB(sipr_decode_frame), 573 .p.capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_CHANNEL_CONF, 574 .caps_internal = FF_CODEC_CAP_INIT_THREADSAFE, 575}; 576