1/* 2 * AMR wideband decoder 3 * Copyright (c) 2010 Marcelo Galvao Povoa 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 * AMR wideband decoder 25 */ 26 27#include "libavutil/channel_layout.h" 28#include "libavutil/common.h" 29#include "libavutil/float_dsp.h" 30#include "libavutil/lfg.h" 31 32#include "avcodec.h" 33#include "lsp.h" 34#include "celp_filters.h" 35#include "celp_math.h" 36#include "acelp_filters.h" 37#include "acelp_vectors.h" 38#include "acelp_pitch_delay.h" 39#include "codec_internal.h" 40#include "internal.h" 41 42#define AMR_USE_16BIT_TABLES 43#include "amr.h" 44 45#include "amrwbdata.h" 46#include "mips/amrwbdec_mips.h" 47 48typedef struct AMRWBContext { 49 AMRWBFrame frame; ///< AMRWB parameters decoded from bitstream 50 enum Mode fr_cur_mode; ///< mode index of current frame 51 uint8_t fr_quality; ///< frame quality index (FQI) 52 float isf_cur[LP_ORDER]; ///< working ISF vector from current frame 53 float isf_q_past[LP_ORDER]; ///< quantized ISF vector of the previous frame 54 float isf_past_final[LP_ORDER]; ///< final processed ISF vector of the previous frame 55 double isp[4][LP_ORDER]; ///< ISP vectors from current frame 56 double isp_sub4_past[LP_ORDER]; ///< ISP vector for the 4th subframe of the previous frame 57 58 float lp_coef[4][LP_ORDER]; ///< Linear Prediction Coefficients from ISP vector 59 60 uint8_t base_pitch_lag; ///< integer part of pitch lag for the next relative subframe 61 uint8_t pitch_lag_int; ///< integer part of pitch lag of the previous subframe 62 63 float excitation_buf[AMRWB_P_DELAY_MAX + LP_ORDER + 2 + AMRWB_SFR_SIZE]; ///< current excitation and all necessary excitation history 64 float *excitation; ///< points to current excitation in excitation_buf[] 65 66 float pitch_vector[AMRWB_SFR_SIZE]; ///< adaptive codebook (pitch) vector for current subframe 67 float fixed_vector[AMRWB_SFR_SIZE]; ///< algebraic codebook (fixed) vector for current subframe 68 69 float prediction_error[4]; ///< quantified prediction errors {20log10(^gamma_gc)} for previous four subframes 70 float pitch_gain[6]; ///< quantified pitch gains for the current and previous five subframes 71 float fixed_gain[2]; ///< quantified fixed gains for the current and previous subframes 72 73 float tilt_coef; ///< {beta_1} related to the voicing of the previous subframe 74 75 float prev_sparse_fixed_gain; ///< previous fixed gain; used by anti-sparseness to determine "onset" 76 uint8_t prev_ir_filter_nr; ///< previous impulse response filter "impNr": 0 - strong, 1 - medium, 2 - none 77 float prev_tr_gain; ///< previous initial gain used by noise enhancer for threshold 78 79 float samples_az[LP_ORDER + AMRWB_SFR_SIZE]; ///< low-band samples and memory from synthesis at 12.8kHz 80 float samples_up[UPS_MEM_SIZE + AMRWB_SFR_SIZE]; ///< low-band samples and memory processed for upsampling 81 float samples_hb[LP_ORDER_16k + AMRWB_SFR_SIZE_16k]; ///< high-band samples and memory from synthesis at 16kHz 82 83 float hpf_31_mem[2], hpf_400_mem[2]; ///< previous values in the high pass filters 84 float demph_mem[1]; ///< previous value in the de-emphasis filter 85 float bpf_6_7_mem[HB_FIR_SIZE]; ///< previous values in the high-band band pass filter 86 float lpf_7_mem[HB_FIR_SIZE]; ///< previous values in the high-band low pass filter 87 88 AVLFG prng; ///< random number generator for white noise excitation 89 uint8_t first_frame; ///< flag active during decoding of the first frame 90 ACELPFContext acelpf_ctx; ///< context for filters for ACELP-based codecs 91 ACELPVContext acelpv_ctx; ///< context for vector operations for ACELP-based codecs 92 CELPFContext celpf_ctx; ///< context for filters for CELP-based codecs 93 CELPMContext celpm_ctx; ///< context for fixed point math operations 94 95} AMRWBContext; 96 97typedef struct AMRWBChannelsContext { 98 AMRWBContext ch[2]; 99} AMRWBChannelsContext; 100 101static av_cold int amrwb_decode_init(AVCodecContext *avctx) 102{ 103 AMRWBChannelsContext *s = avctx->priv_data; 104 int i; 105 106 if (avctx->ch_layout.nb_channels > 2) { 107 avpriv_report_missing_feature(avctx, ">2 channel AMR"); 108 return AVERROR_PATCHWELCOME; 109 } 110 111 if (!avctx->ch_layout.nb_channels) { 112 av_channel_layout_uninit(&avctx->ch_layout); 113 avctx->ch_layout = (AVChannelLayout)AV_CHANNEL_LAYOUT_MONO; 114 } 115 if (!avctx->sample_rate) 116 avctx->sample_rate = 16000; 117 avctx->sample_fmt = AV_SAMPLE_FMT_FLTP; 118 119 for (int ch = 0; ch < avctx->ch_layout.nb_channels; ch++) { 120 AMRWBContext *ctx = &s->ch[ch]; 121 122 av_lfg_init(&ctx->prng, 1); 123 124 ctx->excitation = &ctx->excitation_buf[AMRWB_P_DELAY_MAX + LP_ORDER + 1]; 125 ctx->first_frame = 1; 126 127 for (i = 0; i < LP_ORDER; i++) 128 ctx->isf_past_final[i] = isf_init[i] * (1.0f / (1 << 15)); 129 130 for (i = 0; i < 4; i++) 131 ctx->prediction_error[i] = MIN_ENERGY; 132 133 ff_acelp_filter_init(&ctx->acelpf_ctx); 134 ff_acelp_vectors_init(&ctx->acelpv_ctx); 135 ff_celp_filter_init(&ctx->celpf_ctx); 136 ff_celp_math_init(&ctx->celpm_ctx); 137 } 138 139 return 0; 140} 141 142/** 143 * Decode the frame header in the "MIME/storage" format. This format 144 * is simpler and does not carry the auxiliary frame information. 145 * 146 * @param[in] ctx The Context 147 * @param[in] buf Pointer to the input buffer 148 * 149 * @return The decoded header length in bytes 150 */ 151static int decode_mime_header(AMRWBContext *ctx, const uint8_t *buf) 152{ 153 /* Decode frame header (1st octet) */ 154 ctx->fr_cur_mode = buf[0] >> 3 & 0x0F; 155 ctx->fr_quality = (buf[0] & 0x4) == 0x4; 156 157 return 1; 158} 159 160/** 161 * Decode quantized ISF vectors using 36-bit indexes (6K60 mode only). 162 * 163 * @param[in] ind Array of 5 indexes 164 * @param[out] isf_q Buffer for isf_q[LP_ORDER] 165 */ 166static void decode_isf_indices_36b(uint16_t *ind, float *isf_q) 167{ 168 int i; 169 170 for (i = 0; i < 9; i++) 171 isf_q[i] = dico1_isf[ind[0]][i] * (1.0f / (1 << 15)); 172 173 for (i = 0; i < 7; i++) 174 isf_q[i + 9] = dico2_isf[ind[1]][i] * (1.0f / (1 << 15)); 175 176 for (i = 0; i < 5; i++) 177 isf_q[i] += dico21_isf_36b[ind[2]][i] * (1.0f / (1 << 15)); 178 179 for (i = 0; i < 4; i++) 180 isf_q[i + 5] += dico22_isf_36b[ind[3]][i] * (1.0f / (1 << 15)); 181 182 for (i = 0; i < 7; i++) 183 isf_q[i + 9] += dico23_isf_36b[ind[4]][i] * (1.0f / (1 << 15)); 184} 185 186/** 187 * Decode quantized ISF vectors using 46-bit indexes (except 6K60 mode). 188 * 189 * @param[in] ind Array of 7 indexes 190 * @param[out] isf_q Buffer for isf_q[LP_ORDER] 191 */ 192static void decode_isf_indices_46b(uint16_t *ind, float *isf_q) 193{ 194 int i; 195 196 for (i = 0; i < 9; i++) 197 isf_q[i] = dico1_isf[ind[0]][i] * (1.0f / (1 << 15)); 198 199 for (i = 0; i < 7; i++) 200 isf_q[i + 9] = dico2_isf[ind[1]][i] * (1.0f / (1 << 15)); 201 202 for (i = 0; i < 3; i++) 203 isf_q[i] += dico21_isf[ind[2]][i] * (1.0f / (1 << 15)); 204 205 for (i = 0; i < 3; i++) 206 isf_q[i + 3] += dico22_isf[ind[3]][i] * (1.0f / (1 << 15)); 207 208 for (i = 0; i < 3; i++) 209 isf_q[i + 6] += dico23_isf[ind[4]][i] * (1.0f / (1 << 15)); 210 211 for (i = 0; i < 3; i++) 212 isf_q[i + 9] += dico24_isf[ind[5]][i] * (1.0f / (1 << 15)); 213 214 for (i = 0; i < 4; i++) 215 isf_q[i + 12] += dico25_isf[ind[6]][i] * (1.0f / (1 << 15)); 216} 217 218/** 219 * Apply mean and past ISF values using the prediction factor. 220 * Updates past ISF vector. 221 * 222 * @param[in,out] isf_q Current quantized ISF 223 * @param[in,out] isf_past Past quantized ISF 224 */ 225static void isf_add_mean_and_past(float *isf_q, float *isf_past) 226{ 227 int i; 228 float tmp; 229 230 for (i = 0; i < LP_ORDER; i++) { 231 tmp = isf_q[i]; 232 isf_q[i] += isf_mean[i] * (1.0f / (1 << 15)); 233 isf_q[i] += PRED_FACTOR * isf_past[i]; 234 isf_past[i] = tmp; 235 } 236} 237 238/** 239 * Interpolate the fourth ISP vector from current and past frames 240 * to obtain an ISP vector for each subframe. 241 * 242 * @param[in,out] isp_q ISPs for each subframe 243 * @param[in] isp4_past Past ISP for subframe 4 244 */ 245static void interpolate_isp(double isp_q[4][LP_ORDER], const double *isp4_past) 246{ 247 int i, k; 248 249 for (k = 0; k < 3; k++) { 250 float c = isfp_inter[k]; 251 for (i = 0; i < LP_ORDER; i++) 252 isp_q[k][i] = (1.0 - c) * isp4_past[i] + c * isp_q[3][i]; 253 } 254} 255 256/** 257 * Decode an adaptive codebook index into pitch lag (except 6k60, 8k85 modes). 258 * Calculate integer lag and fractional lag always using 1/4 resolution. 259 * In 1st and 3rd subframes the index is relative to last subframe integer lag. 260 * 261 * @param[out] lag_int Decoded integer pitch lag 262 * @param[out] lag_frac Decoded fractional pitch lag 263 * @param[in] pitch_index Adaptive codebook pitch index 264 * @param[in,out] base_lag_int Base integer lag used in relative subframes 265 * @param[in] subframe Current subframe index (0 to 3) 266 */ 267static void decode_pitch_lag_high(int *lag_int, int *lag_frac, int pitch_index, 268 uint8_t *base_lag_int, int subframe) 269{ 270 if (subframe == 0 || subframe == 2) { 271 if (pitch_index < 376) { 272 *lag_int = (pitch_index + 137) >> 2; 273 *lag_frac = pitch_index - (*lag_int << 2) + 136; 274 } else if (pitch_index < 440) { 275 *lag_int = (pitch_index + 257 - 376) >> 1; 276 *lag_frac = (pitch_index - (*lag_int << 1) + 256 - 376) * 2; 277 /* the actual resolution is 1/2 but expressed as 1/4 */ 278 } else { 279 *lag_int = pitch_index - 280; 280 *lag_frac = 0; 281 } 282 /* minimum lag for next subframe */ 283 *base_lag_int = av_clip(*lag_int - 8 - (*lag_frac < 0), 284 AMRWB_P_DELAY_MIN, AMRWB_P_DELAY_MAX - 15); 285 // XXX: the spec states clearly that *base_lag_int should be 286 // the nearest integer to *lag_int (minus 8), but the ref code 287 // actually always uses its floor, I'm following the latter 288 } else { 289 *lag_int = (pitch_index + 1) >> 2; 290 *lag_frac = pitch_index - (*lag_int << 2); 291 *lag_int += *base_lag_int; 292 } 293} 294 295/** 296 * Decode an adaptive codebook index into pitch lag for 8k85 and 6k60 modes. 297 * The description is analogous to decode_pitch_lag_high, but in 6k60 the 298 * relative index is used for all subframes except the first. 299 */ 300static void decode_pitch_lag_low(int *lag_int, int *lag_frac, int pitch_index, 301 uint8_t *base_lag_int, int subframe, enum Mode mode) 302{ 303 if (subframe == 0 || (subframe == 2 && mode != MODE_6k60)) { 304 if (pitch_index < 116) { 305 *lag_int = (pitch_index + 69) >> 1; 306 *lag_frac = (pitch_index - (*lag_int << 1) + 68) * 2; 307 } else { 308 *lag_int = pitch_index - 24; 309 *lag_frac = 0; 310 } 311 // XXX: same problem as before 312 *base_lag_int = av_clip(*lag_int - 8 - (*lag_frac < 0), 313 AMRWB_P_DELAY_MIN, AMRWB_P_DELAY_MAX - 15); 314 } else { 315 *lag_int = (pitch_index + 1) >> 1; 316 *lag_frac = (pitch_index - (*lag_int << 1)) * 2; 317 *lag_int += *base_lag_int; 318 } 319} 320 321/** 322 * Find the pitch vector by interpolating the past excitation at the 323 * pitch delay, which is obtained in this function. 324 * 325 * @param[in,out] ctx The context 326 * @param[in] amr_subframe Current subframe data 327 * @param[in] subframe Current subframe index (0 to 3) 328 */ 329static void decode_pitch_vector(AMRWBContext *ctx, 330 const AMRWBSubFrame *amr_subframe, 331 const int subframe) 332{ 333 int pitch_lag_int, pitch_lag_frac; 334 int i; 335 float *exc = ctx->excitation; 336 enum Mode mode = ctx->fr_cur_mode; 337 338 if (mode <= MODE_8k85) { 339 decode_pitch_lag_low(&pitch_lag_int, &pitch_lag_frac, amr_subframe->adap, 340 &ctx->base_pitch_lag, subframe, mode); 341 } else 342 decode_pitch_lag_high(&pitch_lag_int, &pitch_lag_frac, amr_subframe->adap, 343 &ctx->base_pitch_lag, subframe); 344 345 ctx->pitch_lag_int = pitch_lag_int; 346 pitch_lag_int += pitch_lag_frac > 0; 347 348 /* Calculate the pitch vector by interpolating the past excitation at the 349 pitch lag using a hamming windowed sinc function */ 350 ctx->acelpf_ctx.acelp_interpolatef(exc, 351 exc + 1 - pitch_lag_int, 352 ac_inter, 4, 353 pitch_lag_frac + (pitch_lag_frac > 0 ? 0 : 4), 354 LP_ORDER, AMRWB_SFR_SIZE + 1); 355 356 /* Check which pitch signal path should be used 357 * 6k60 and 8k85 modes have the ltp flag set to 0 */ 358 if (amr_subframe->ltp) { 359 memcpy(ctx->pitch_vector, exc, AMRWB_SFR_SIZE * sizeof(float)); 360 } else { 361 for (i = 0; i < AMRWB_SFR_SIZE; i++) 362 ctx->pitch_vector[i] = 0.18 * exc[i - 1] + 0.64 * exc[i] + 363 0.18 * exc[i + 1]; 364 memcpy(exc, ctx->pitch_vector, AMRWB_SFR_SIZE * sizeof(float)); 365 } 366} 367 368/** Get x bits in the index interval [lsb,lsb+len-1] inclusive */ 369#define BIT_STR(x,lsb,len) av_mod_uintp2((x) >> (lsb), (len)) 370 371/** Get the bit at specified position */ 372#define BIT_POS(x, p) (((x) >> (p)) & 1) 373 374/** 375 * The next six functions decode_[i]p_track decode exactly i pulses 376 * positions and amplitudes (-1 or 1) in a subframe track using 377 * an encoded pulse indexing (TS 26.190 section 5.8.2). 378 * 379 * The results are given in out[], in which a negative number means 380 * amplitude -1 and vice versa (i.e., ampl(x) = x / abs(x) ). 381 * 382 * @param[out] out Output buffer (writes i elements) 383 * @param[in] code Pulse index (no. of bits varies, see below) 384 * @param[in] m (log2) Number of potential positions 385 * @param[in] off Offset for decoded positions 386 */ 387static inline void decode_1p_track(int *out, int code, int m, int off) 388{ 389 int pos = BIT_STR(code, 0, m) + off; ///code: m+1 bits 390 391 out[0] = BIT_POS(code, m) ? -pos : pos; 392} 393 394static inline void decode_2p_track(int *out, int code, int m, int off) ///code: 2m+1 bits 395{ 396 int pos0 = BIT_STR(code, m, m) + off; 397 int pos1 = BIT_STR(code, 0, m) + off; 398 399 out[0] = BIT_POS(code, 2*m) ? -pos0 : pos0; 400 out[1] = BIT_POS(code, 2*m) ? -pos1 : pos1; 401 out[1] = pos0 > pos1 ? -out[1] : out[1]; 402} 403 404static void decode_3p_track(int *out, int code, int m, int off) ///code: 3m+1 bits 405{ 406 int half_2p = BIT_POS(code, 2*m - 1) << (m - 1); 407 408 decode_2p_track(out, BIT_STR(code, 0, 2*m - 1), 409 m - 1, off + half_2p); 410 decode_1p_track(out + 2, BIT_STR(code, 2*m, m + 1), m, off); 411} 412 413static void decode_4p_track(int *out, int code, int m, int off) ///code: 4m bits 414{ 415 int half_4p, subhalf_2p; 416 int b_offset = 1 << (m - 1); 417 418 switch (BIT_STR(code, 4*m - 2, 2)) { /* case ID (2 bits) */ 419 case 0: /* 0 pulses in A, 4 pulses in B or vice versa */ 420 half_4p = BIT_POS(code, 4*m - 3) << (m - 1); // which has 4 pulses 421 subhalf_2p = BIT_POS(code, 2*m - 3) << (m - 2); 422 423 decode_2p_track(out, BIT_STR(code, 0, 2*m - 3), 424 m - 2, off + half_4p + subhalf_2p); 425 decode_2p_track(out + 2, BIT_STR(code, 2*m - 2, 2*m - 1), 426 m - 1, off + half_4p); 427 break; 428 case 1: /* 1 pulse in A, 3 pulses in B */ 429 decode_1p_track(out, BIT_STR(code, 3*m - 2, m), 430 m - 1, off); 431 decode_3p_track(out + 1, BIT_STR(code, 0, 3*m - 2), 432 m - 1, off + b_offset); 433 break; 434 case 2: /* 2 pulses in each half */ 435 decode_2p_track(out, BIT_STR(code, 2*m - 1, 2*m - 1), 436 m - 1, off); 437 decode_2p_track(out + 2, BIT_STR(code, 0, 2*m - 1), 438 m - 1, off + b_offset); 439 break; 440 case 3: /* 3 pulses in A, 1 pulse in B */ 441 decode_3p_track(out, BIT_STR(code, m, 3*m - 2), 442 m - 1, off); 443 decode_1p_track(out + 3, BIT_STR(code, 0, m), 444 m - 1, off + b_offset); 445 break; 446 } 447} 448 449static void decode_5p_track(int *out, int code, int m, int off) ///code: 5m bits 450{ 451 int half_3p = BIT_POS(code, 5*m - 1) << (m - 1); 452 453 decode_3p_track(out, BIT_STR(code, 2*m + 1, 3*m - 2), 454 m - 1, off + half_3p); 455 456 decode_2p_track(out + 3, BIT_STR(code, 0, 2*m + 1), m, off); 457} 458 459static void decode_6p_track(int *out, int code, int m, int off) ///code: 6m-2 bits 460{ 461 int b_offset = 1 << (m - 1); 462 /* which half has more pulses in cases 0 to 2 */ 463 int half_more = BIT_POS(code, 6*m - 5) << (m - 1); 464 int half_other = b_offset - half_more; 465 466 switch (BIT_STR(code, 6*m - 4, 2)) { /* case ID (2 bits) */ 467 case 0: /* 0 pulses in A, 6 pulses in B or vice versa */ 468 decode_1p_track(out, BIT_STR(code, 0, m), 469 m - 1, off + half_more); 470 decode_5p_track(out + 1, BIT_STR(code, m, 5*m - 5), 471 m - 1, off + half_more); 472 break; 473 case 1: /* 1 pulse in A, 5 pulses in B or vice versa */ 474 decode_1p_track(out, BIT_STR(code, 0, m), 475 m - 1, off + half_other); 476 decode_5p_track(out + 1, BIT_STR(code, m, 5*m - 5), 477 m - 1, off + half_more); 478 break; 479 case 2: /* 2 pulses in A, 4 pulses in B or vice versa */ 480 decode_2p_track(out, BIT_STR(code, 0, 2*m - 1), 481 m - 1, off + half_other); 482 decode_4p_track(out + 2, BIT_STR(code, 2*m - 1, 4*m - 4), 483 m - 1, off + half_more); 484 break; 485 case 3: /* 3 pulses in A, 3 pulses in B */ 486 decode_3p_track(out, BIT_STR(code, 3*m - 2, 3*m - 2), 487 m - 1, off); 488 decode_3p_track(out + 3, BIT_STR(code, 0, 3*m - 2), 489 m - 1, off + b_offset); 490 break; 491 } 492} 493 494/** 495 * Decode the algebraic codebook index to pulse positions and signs, 496 * then construct the algebraic codebook vector. 497 * 498 * @param[out] fixed_vector Buffer for the fixed codebook excitation 499 * @param[in] pulse_hi MSBs part of the pulse index array (higher modes only) 500 * @param[in] pulse_lo LSBs part of the pulse index array 501 * @param[in] mode Mode of the current frame 502 */ 503static void decode_fixed_vector(float *fixed_vector, const uint16_t *pulse_hi, 504 const uint16_t *pulse_lo, const enum Mode mode) 505{ 506 /* sig_pos stores for each track the decoded pulse position indexes 507 * (1-based) multiplied by its corresponding amplitude (+1 or -1) */ 508 int sig_pos[4][6]; 509 int spacing = (mode == MODE_6k60) ? 2 : 4; 510 int i, j; 511 512 switch (mode) { 513 case MODE_6k60: 514 for (i = 0; i < 2; i++) 515 decode_1p_track(sig_pos[i], pulse_lo[i], 5, 1); 516 break; 517 case MODE_8k85: 518 for (i = 0; i < 4; i++) 519 decode_1p_track(sig_pos[i], pulse_lo[i], 4, 1); 520 break; 521 case MODE_12k65: 522 for (i = 0; i < 4; i++) 523 decode_2p_track(sig_pos[i], pulse_lo[i], 4, 1); 524 break; 525 case MODE_14k25: 526 for (i = 0; i < 2; i++) 527 decode_3p_track(sig_pos[i], pulse_lo[i], 4, 1); 528 for (i = 2; i < 4; i++) 529 decode_2p_track(sig_pos[i], pulse_lo[i], 4, 1); 530 break; 531 case MODE_15k85: 532 for (i = 0; i < 4; i++) 533 decode_3p_track(sig_pos[i], pulse_lo[i], 4, 1); 534 break; 535 case MODE_18k25: 536 for (i = 0; i < 4; i++) 537 decode_4p_track(sig_pos[i], (int) pulse_lo[i] + 538 ((int) pulse_hi[i] << 14), 4, 1); 539 break; 540 case MODE_19k85: 541 for (i = 0; i < 2; i++) 542 decode_5p_track(sig_pos[i], (int) pulse_lo[i] + 543 ((int) pulse_hi[i] << 10), 4, 1); 544 for (i = 2; i < 4; i++) 545 decode_4p_track(sig_pos[i], (int) pulse_lo[i] + 546 ((int) pulse_hi[i] << 14), 4, 1); 547 break; 548 case MODE_23k05: 549 case MODE_23k85: 550 for (i = 0; i < 4; i++) 551 decode_6p_track(sig_pos[i], (int) pulse_lo[i] + 552 ((int) pulse_hi[i] << 11), 4, 1); 553 break; 554 } 555 556 memset(fixed_vector, 0, sizeof(float) * AMRWB_SFR_SIZE); 557 558 for (i = 0; i < 4; i++) 559 for (j = 0; j < pulses_nb_per_mode_tr[mode][i]; j++) { 560 int pos = (FFABS(sig_pos[i][j]) - 1) * spacing + i; 561 562 fixed_vector[pos] += sig_pos[i][j] < 0 ? -1.0 : 1.0; 563 } 564} 565 566/** 567 * Decode pitch gain and fixed gain correction factor. 568 * 569 * @param[in] vq_gain Vector-quantized index for gains 570 * @param[in] mode Mode of the current frame 571 * @param[out] fixed_gain_factor Decoded fixed gain correction factor 572 * @param[out] pitch_gain Decoded pitch gain 573 */ 574static void decode_gains(const uint8_t vq_gain, const enum Mode mode, 575 float *fixed_gain_factor, float *pitch_gain) 576{ 577 const int16_t *gains = (mode <= MODE_8k85 ? qua_gain_6b[vq_gain] : 578 qua_gain_7b[vq_gain]); 579 580 *pitch_gain = gains[0] * (1.0f / (1 << 14)); 581 *fixed_gain_factor = gains[1] * (1.0f / (1 << 11)); 582} 583 584/** 585 * Apply pitch sharpening filters to the fixed codebook vector. 586 * 587 * @param[in] ctx The context 588 * @param[in,out] fixed_vector Fixed codebook excitation 589 */ 590// XXX: Spec states this procedure should be applied when the pitch 591// lag is less than 64, but this checking seems absent in reference and AMR-NB 592static void pitch_sharpening(AMRWBContext *ctx, float *fixed_vector) 593{ 594 int i; 595 596 /* Tilt part */ 597 for (i = AMRWB_SFR_SIZE - 1; i != 0; i--) 598 fixed_vector[i] -= fixed_vector[i - 1] * ctx->tilt_coef; 599 600 /* Periodicity enhancement part */ 601 for (i = ctx->pitch_lag_int; i < AMRWB_SFR_SIZE; i++) 602 fixed_vector[i] += fixed_vector[i - ctx->pitch_lag_int] * 0.85; 603} 604 605/** 606 * Calculate the voicing factor (-1.0 = unvoiced to 1.0 = voiced). 607 * 608 * @param[in] p_vector, f_vector Pitch and fixed excitation vectors 609 * @param[in] p_gain, f_gain Pitch and fixed gains 610 * @param[in] ctx The context 611 */ 612// XXX: There is something wrong with the precision here! The magnitudes 613// of the energies are not correct. Please check the reference code carefully 614static float voice_factor(float *p_vector, float p_gain, 615 float *f_vector, float f_gain, 616 CELPMContext *ctx) 617{ 618 double p_ener = (double) ctx->dot_productf(p_vector, p_vector, 619 AMRWB_SFR_SIZE) * 620 p_gain * p_gain; 621 double f_ener = (double) ctx->dot_productf(f_vector, f_vector, 622 AMRWB_SFR_SIZE) * 623 f_gain * f_gain; 624 625 return (p_ener - f_ener) / (p_ener + f_ener + 0.01); 626} 627 628/** 629 * Reduce fixed vector sparseness by smoothing with one of three IR filters, 630 * also known as "adaptive phase dispersion". 631 * 632 * @param[in] ctx The context 633 * @param[in,out] fixed_vector Unfiltered fixed vector 634 * @param[out] buf Space for modified vector if necessary 635 * 636 * @return The potentially overwritten filtered fixed vector address 637 */ 638static float *anti_sparseness(AMRWBContext *ctx, 639 float *fixed_vector, float *buf) 640{ 641 int ir_filter_nr; 642 643 if (ctx->fr_cur_mode > MODE_8k85) // no filtering in higher modes 644 return fixed_vector; 645 646 if (ctx->pitch_gain[0] < 0.6) { 647 ir_filter_nr = 0; // strong filtering 648 } else if (ctx->pitch_gain[0] < 0.9) { 649 ir_filter_nr = 1; // medium filtering 650 } else 651 ir_filter_nr = 2; // no filtering 652 653 /* detect 'onset' */ 654 if (ctx->fixed_gain[0] > 3.0 * ctx->fixed_gain[1]) { 655 if (ir_filter_nr < 2) 656 ir_filter_nr++; 657 } else { 658 int i, count = 0; 659 660 for (i = 0; i < 6; i++) 661 if (ctx->pitch_gain[i] < 0.6) 662 count++; 663 664 if (count > 2) 665 ir_filter_nr = 0; 666 667 if (ir_filter_nr > ctx->prev_ir_filter_nr + 1) 668 ir_filter_nr--; 669 } 670 671 /* update ir filter strength history */ 672 ctx->prev_ir_filter_nr = ir_filter_nr; 673 674 ir_filter_nr += (ctx->fr_cur_mode == MODE_8k85); 675 676 if (ir_filter_nr < 2) { 677 int i; 678 const float *coef = ir_filters_lookup[ir_filter_nr]; 679 680 /* Circular convolution code in the reference 681 * decoder was modified to avoid using one 682 * extra array. The filtered vector is given by: 683 * 684 * c2(n) = sum(i,0,len-1){ c(i) * coef( (n - i + len) % len ) } 685 */ 686 687 memset(buf, 0, sizeof(float) * AMRWB_SFR_SIZE); 688 for (i = 0; i < AMRWB_SFR_SIZE; i++) 689 if (fixed_vector[i]) 690 ff_celp_circ_addf(buf, buf, coef, i, fixed_vector[i], 691 AMRWB_SFR_SIZE); 692 fixed_vector = buf; 693 } 694 695 return fixed_vector; 696} 697 698/** 699 * Calculate a stability factor {teta} based on distance between 700 * current and past isf. A value of 1 shows maximum signal stability. 701 */ 702static float stability_factor(const float *isf, const float *isf_past) 703{ 704 int i; 705 float acc = 0.0; 706 707 for (i = 0; i < LP_ORDER - 1; i++) 708 acc += (isf[i] - isf_past[i]) * (isf[i] - isf_past[i]); 709 710 // XXX: This part is not so clear from the reference code 711 // the result is more accurate changing the "/ 256" to "* 512" 712 return FFMAX(0.0, 1.25 - acc * 0.8 * 512); 713} 714 715/** 716 * Apply a non-linear fixed gain smoothing in order to reduce 717 * fluctuation in the energy of excitation. 718 * 719 * @param[in] fixed_gain Unsmoothed fixed gain 720 * @param[in,out] prev_tr_gain Previous threshold gain (updated) 721 * @param[in] voice_fac Frame voicing factor 722 * @param[in] stab_fac Frame stability factor 723 * 724 * @return The smoothed gain 725 */ 726static float noise_enhancer(float fixed_gain, float *prev_tr_gain, 727 float voice_fac, float stab_fac) 728{ 729 float sm_fac = 0.5 * (1 - voice_fac) * stab_fac; 730 float g0; 731 732 // XXX: the following fixed-point constants used to in(de)crement 733 // gain by 1.5dB were taken from the reference code, maybe it could 734 // be simpler 735 if (fixed_gain < *prev_tr_gain) { 736 g0 = FFMIN(*prev_tr_gain, fixed_gain + fixed_gain * 737 (6226 * (1.0f / (1 << 15)))); // +1.5 dB 738 } else 739 g0 = FFMAX(*prev_tr_gain, fixed_gain * 740 (27536 * (1.0f / (1 << 15)))); // -1.5 dB 741 742 *prev_tr_gain = g0; // update next frame threshold 743 744 return sm_fac * g0 + (1 - sm_fac) * fixed_gain; 745} 746 747/** 748 * Filter the fixed_vector to emphasize the higher frequencies. 749 * 750 * @param[in,out] fixed_vector Fixed codebook vector 751 * @param[in] voice_fac Frame voicing factor 752 */ 753static void pitch_enhancer(float *fixed_vector, float voice_fac) 754{ 755 int i; 756 float cpe = 0.125 * (1 + voice_fac); 757 float last = fixed_vector[0]; // holds c(i - 1) 758 759 fixed_vector[0] -= cpe * fixed_vector[1]; 760 761 for (i = 1; i < AMRWB_SFR_SIZE - 1; i++) { 762 float cur = fixed_vector[i]; 763 764 fixed_vector[i] -= cpe * (last + fixed_vector[i + 1]); 765 last = cur; 766 } 767 768 fixed_vector[AMRWB_SFR_SIZE - 1] -= cpe * last; 769} 770 771/** 772 * Conduct 16th order linear predictive coding synthesis from excitation. 773 * 774 * @param[in] ctx Pointer to the AMRWBContext 775 * @param[in] lpc Pointer to the LPC coefficients 776 * @param[out] excitation Buffer for synthesis final excitation 777 * @param[in] fixed_gain Fixed codebook gain for synthesis 778 * @param[in] fixed_vector Algebraic codebook vector 779 * @param[in,out] samples Pointer to the output samples and memory 780 */ 781static void synthesis(AMRWBContext *ctx, float *lpc, float *excitation, 782 float fixed_gain, const float *fixed_vector, 783 float *samples) 784{ 785 ctx->acelpv_ctx.weighted_vector_sumf(excitation, ctx->pitch_vector, fixed_vector, 786 ctx->pitch_gain[0], fixed_gain, AMRWB_SFR_SIZE); 787 788 /* emphasize pitch vector contribution in low bitrate modes */ 789 if (ctx->pitch_gain[0] > 0.5 && ctx->fr_cur_mode <= MODE_8k85) { 790 int i; 791 float energy = ctx->celpm_ctx.dot_productf(excitation, excitation, 792 AMRWB_SFR_SIZE); 793 794 // XXX: Weird part in both ref code and spec. A unknown parameter 795 // {beta} seems to be identical to the current pitch gain 796 float pitch_factor = 0.25 * ctx->pitch_gain[0] * ctx->pitch_gain[0]; 797 798 for (i = 0; i < AMRWB_SFR_SIZE; i++) 799 excitation[i] += pitch_factor * ctx->pitch_vector[i]; 800 801 ff_scale_vector_to_given_sum_of_squares(excitation, excitation, 802 energy, AMRWB_SFR_SIZE); 803 } 804 805 ctx->celpf_ctx.celp_lp_synthesis_filterf(samples, lpc, excitation, 806 AMRWB_SFR_SIZE, LP_ORDER); 807} 808 809/** 810 * Apply to synthesis a de-emphasis filter of the form: 811 * H(z) = 1 / (1 - m * z^-1) 812 * 813 * @param[out] out Output buffer 814 * @param[in] in Input samples array with in[-1] 815 * @param[in] m Filter coefficient 816 * @param[in,out] mem State from last filtering 817 */ 818static void de_emphasis(float *out, float *in, float m, float mem[1]) 819{ 820 int i; 821 822 out[0] = in[0] + m * mem[0]; 823 824 for (i = 1; i < AMRWB_SFR_SIZE; i++) 825 out[i] = in[i] + out[i - 1] * m; 826 827 mem[0] = out[AMRWB_SFR_SIZE - 1]; 828} 829 830/** 831 * Upsample a signal by 5/4 ratio (from 12.8kHz to 16kHz) using 832 * a FIR interpolation filter. Uses past data from before *in address. 833 * 834 * @param[out] out Buffer for interpolated signal 835 * @param[in] in Current signal data (length 0.8*o_size) 836 * @param[in] o_size Output signal length 837 * @param[in] ctx The context 838 */ 839static void upsample_5_4(float *out, const float *in, int o_size, CELPMContext *ctx) 840{ 841 const float *in0 = in - UPS_FIR_SIZE + 1; 842 int i, j, k; 843 int int_part = 0, frac_part; 844 845 i = 0; 846 for (j = 0; j < o_size / 5; j++) { 847 out[i] = in[int_part]; 848 frac_part = 4; 849 i++; 850 851 for (k = 1; k < 5; k++) { 852 out[i] = ctx->dot_productf(in0 + int_part, 853 upsample_fir[4 - frac_part], 854 UPS_MEM_SIZE); 855 int_part++; 856 frac_part--; 857 i++; 858 } 859 } 860} 861 862/** 863 * Calculate the high-band gain based on encoded index (23k85 mode) or 864 * on the low-band speech signal and the Voice Activity Detection flag. 865 * 866 * @param[in] ctx The context 867 * @param[in] synth LB speech synthesis at 12.8k 868 * @param[in] hb_idx Gain index for mode 23k85 only 869 * @param[in] vad VAD flag for the frame 870 */ 871static float find_hb_gain(AMRWBContext *ctx, const float *synth, 872 uint16_t hb_idx, uint8_t vad) 873{ 874 int wsp = (vad > 0); 875 float tilt; 876 float tmp; 877 878 if (ctx->fr_cur_mode == MODE_23k85) 879 return qua_hb_gain[hb_idx] * (1.0f / (1 << 14)); 880 881 tmp = ctx->celpm_ctx.dot_productf(synth, synth + 1, AMRWB_SFR_SIZE - 1); 882 883 if (tmp > 0) { 884 tilt = tmp / ctx->celpm_ctx.dot_productf(synth, synth, AMRWB_SFR_SIZE); 885 } else 886 tilt = 0; 887 888 /* return gain bounded by [0.1, 1.0] */ 889 return av_clipf((1.0 - tilt) * (1.25 - 0.25 * wsp), 0.1, 1.0); 890} 891 892/** 893 * Generate the high-band excitation with the same energy from the lower 894 * one and scaled by the given gain. 895 * 896 * @param[in] ctx The context 897 * @param[out] hb_exc Buffer for the excitation 898 * @param[in] synth_exc Low-band excitation used for synthesis 899 * @param[in] hb_gain Wanted excitation gain 900 */ 901static void scaled_hb_excitation(AMRWBContext *ctx, float *hb_exc, 902 const float *synth_exc, float hb_gain) 903{ 904 int i; 905 float energy = ctx->celpm_ctx.dot_productf(synth_exc, synth_exc, 906 AMRWB_SFR_SIZE); 907 908 /* Generate a white-noise excitation */ 909 for (i = 0; i < AMRWB_SFR_SIZE_16k; i++) 910 hb_exc[i] = 32768.0 - (uint16_t) av_lfg_get(&ctx->prng); 911 912 ff_scale_vector_to_given_sum_of_squares(hb_exc, hb_exc, 913 energy * hb_gain * hb_gain, 914 AMRWB_SFR_SIZE_16k); 915} 916 917/** 918 * Calculate the auto-correlation for the ISF difference vector. 919 */ 920static float auto_correlation(float *diff_isf, float mean, int lag) 921{ 922 int i; 923 float sum = 0.0; 924 925 for (i = 7; i < LP_ORDER - 2; i++) { 926 float prod = (diff_isf[i] - mean) * (diff_isf[i - lag] - mean); 927 sum += prod * prod; 928 } 929 return sum; 930} 931 932/** 933 * Extrapolate a ISF vector to the 16kHz range (20th order LP) 934 * used at mode 6k60 LP filter for the high frequency band. 935 * 936 * @param[out] isf Buffer for extrapolated isf; contains LP_ORDER 937 * values on input 938 */ 939static void extrapolate_isf(float isf[LP_ORDER_16k]) 940{ 941 float diff_isf[LP_ORDER - 2], diff_mean; 942 float corr_lag[3]; 943 float est, scale; 944 int i, j, i_max_corr; 945 946 isf[LP_ORDER_16k - 1] = isf[LP_ORDER - 1]; 947 948 /* Calculate the difference vector */ 949 for (i = 0; i < LP_ORDER - 2; i++) 950 diff_isf[i] = isf[i + 1] - isf[i]; 951 952 diff_mean = 0.0; 953 for (i = 2; i < LP_ORDER - 2; i++) 954 diff_mean += diff_isf[i] * (1.0f / (LP_ORDER - 4)); 955 956 /* Find which is the maximum autocorrelation */ 957 i_max_corr = 0; 958 for (i = 0; i < 3; i++) { 959 corr_lag[i] = auto_correlation(diff_isf, diff_mean, i + 2); 960 961 if (corr_lag[i] > corr_lag[i_max_corr]) 962 i_max_corr = i; 963 } 964 i_max_corr++; 965 966 for (i = LP_ORDER - 1; i < LP_ORDER_16k - 1; i++) 967 isf[i] = isf[i - 1] + isf[i - 1 - i_max_corr] 968 - isf[i - 2 - i_max_corr]; 969 970 /* Calculate an estimate for ISF(18) and scale ISF based on the error */ 971 est = 7965 + (isf[2] - isf[3] - isf[4]) / 6.0; 972 scale = 0.5 * (FFMIN(est, 7600) - isf[LP_ORDER - 2]) / 973 (isf[LP_ORDER_16k - 2] - isf[LP_ORDER - 2]); 974 975 for (i = LP_ORDER - 1, j = 0; i < LP_ORDER_16k - 1; i++, j++) 976 diff_isf[j] = scale * (isf[i] - isf[i - 1]); 977 978 /* Stability insurance */ 979 for (i = 1; i < LP_ORDER_16k - LP_ORDER; i++) 980 if (diff_isf[i] + diff_isf[i - 1] < 5.0) { 981 if (diff_isf[i] > diff_isf[i - 1]) { 982 diff_isf[i - 1] = 5.0 - diff_isf[i]; 983 } else 984 diff_isf[i] = 5.0 - diff_isf[i - 1]; 985 } 986 987 for (i = LP_ORDER - 1, j = 0; i < LP_ORDER_16k - 1; i++, j++) 988 isf[i] = isf[i - 1] + diff_isf[j] * (1.0f / (1 << 15)); 989 990 /* Scale the ISF vector for 16000 Hz */ 991 for (i = 0; i < LP_ORDER_16k - 1; i++) 992 isf[i] *= 0.8; 993} 994 995/** 996 * Spectral expand the LP coefficients using the equation: 997 * y[i] = x[i] * (gamma ** i) 998 * 999 * @param[out] out Output buffer (may use input array) 1000 * @param[in] lpc LP coefficients array 1001 * @param[in] gamma Weighting factor 1002 * @param[in] size LP array size 1003 */ 1004static void lpc_weighting(float *out, const float *lpc, float gamma, int size) 1005{ 1006 int i; 1007 float fac = gamma; 1008 1009 for (i = 0; i < size; i++) { 1010 out[i] = lpc[i] * fac; 1011 fac *= gamma; 1012 } 1013} 1014 1015/** 1016 * Conduct 20th order linear predictive coding synthesis for the high 1017 * frequency band excitation at 16kHz. 1018 * 1019 * @param[in] ctx The context 1020 * @param[in] subframe Current subframe index (0 to 3) 1021 * @param[in,out] samples Pointer to the output speech samples 1022 * @param[in] exc Generated white-noise scaled excitation 1023 * @param[in] isf Current frame isf vector 1024 * @param[in] isf_past Past frame final isf vector 1025 */ 1026static void hb_synthesis(AMRWBContext *ctx, int subframe, float *samples, 1027 const float *exc, const float *isf, const float *isf_past) 1028{ 1029 float hb_lpc[LP_ORDER_16k]; 1030 enum Mode mode = ctx->fr_cur_mode; 1031 1032 if (mode == MODE_6k60) { 1033 float e_isf[LP_ORDER_16k]; // ISF vector for extrapolation 1034 double e_isp[LP_ORDER_16k]; 1035 1036 ctx->acelpv_ctx.weighted_vector_sumf(e_isf, isf_past, isf, isfp_inter[subframe], 1037 1.0 - isfp_inter[subframe], LP_ORDER); 1038 1039 extrapolate_isf(e_isf); 1040 1041 e_isf[LP_ORDER_16k - 1] *= 2.0; 1042 ff_acelp_lsf2lspd(e_isp, e_isf, LP_ORDER_16k); 1043 ff_amrwb_lsp2lpc(e_isp, hb_lpc, LP_ORDER_16k); 1044 1045 lpc_weighting(hb_lpc, hb_lpc, 0.9, LP_ORDER_16k); 1046 } else { 1047 lpc_weighting(hb_lpc, ctx->lp_coef[subframe], 0.6, LP_ORDER); 1048 } 1049 1050 ctx->celpf_ctx.celp_lp_synthesis_filterf(samples, hb_lpc, exc, AMRWB_SFR_SIZE_16k, 1051 (mode == MODE_6k60) ? LP_ORDER_16k : LP_ORDER); 1052} 1053 1054/** 1055 * Apply a 15th order filter to high-band samples. 1056 * The filter characteristic depends on the given coefficients. 1057 * 1058 * @param[out] out Buffer for filtered output 1059 * @param[in] fir_coef Filter coefficients 1060 * @param[in,out] mem State from last filtering (updated) 1061 * @param[in] in Input speech data (high-band) 1062 * 1063 * @remark It is safe to pass the same array in in and out parameters 1064 */ 1065 1066#ifndef hb_fir_filter 1067static void hb_fir_filter(float *out, const float fir_coef[HB_FIR_SIZE + 1], 1068 float mem[HB_FIR_SIZE], const float *in) 1069{ 1070 int i, j; 1071 float data[AMRWB_SFR_SIZE_16k + HB_FIR_SIZE]; // past and current samples 1072 1073 memcpy(data, mem, HB_FIR_SIZE * sizeof(float)); 1074 memcpy(data + HB_FIR_SIZE, in, AMRWB_SFR_SIZE_16k * sizeof(float)); 1075 1076 for (i = 0; i < AMRWB_SFR_SIZE_16k; i++) { 1077 out[i] = 0.0; 1078 for (j = 0; j <= HB_FIR_SIZE; j++) 1079 out[i] += data[i + j] * fir_coef[j]; 1080 } 1081 1082 memcpy(mem, data + AMRWB_SFR_SIZE_16k, HB_FIR_SIZE * sizeof(float)); 1083} 1084#endif /* hb_fir_filter */ 1085 1086/** 1087 * Update context state before the next subframe. 1088 */ 1089static void update_sub_state(AMRWBContext *ctx) 1090{ 1091 memmove(&ctx->excitation_buf[0], &ctx->excitation_buf[AMRWB_SFR_SIZE], 1092 (AMRWB_P_DELAY_MAX + LP_ORDER + 1) * sizeof(float)); 1093 1094 memmove(&ctx->pitch_gain[1], &ctx->pitch_gain[0], 5 * sizeof(float)); 1095 memmove(&ctx->fixed_gain[1], &ctx->fixed_gain[0], 1 * sizeof(float)); 1096 1097 memmove(&ctx->samples_az[0], &ctx->samples_az[AMRWB_SFR_SIZE], 1098 LP_ORDER * sizeof(float)); 1099 memmove(&ctx->samples_up[0], &ctx->samples_up[AMRWB_SFR_SIZE], 1100 UPS_MEM_SIZE * sizeof(float)); 1101 memmove(&ctx->samples_hb[0], &ctx->samples_hb[AMRWB_SFR_SIZE_16k], 1102 LP_ORDER_16k * sizeof(float)); 1103} 1104 1105static int amrwb_decode_frame(AVCodecContext *avctx, AVFrame *frame, 1106 int *got_frame_ptr, AVPacket *avpkt) 1107{ 1108 AMRWBChannelsContext *s = avctx->priv_data; 1109 const uint8_t *buf = avpkt->data; 1110 int buf_size = avpkt->size; 1111 int sub, i, ret; 1112 1113 /* get output buffer */ 1114 frame->nb_samples = 4 * AMRWB_SFR_SIZE_16k; 1115 if ((ret = ff_get_buffer(avctx, frame, 0)) < 0) 1116 return ret; 1117 1118 for (int ch = 0; ch < avctx->ch_layout.nb_channels; ch++) { 1119 AMRWBContext *ctx = &s->ch[ch]; 1120 AMRWBFrame *cf = &ctx->frame; 1121 int expected_fr_size, header_size; 1122 float spare_vector[AMRWB_SFR_SIZE]; // extra stack space to hold result from anti-sparseness processing 1123 float fixed_gain_factor; // fixed gain correction factor (gamma) 1124 float *synth_fixed_vector; // pointer to the fixed vector that synthesis should use 1125 float synth_fixed_gain; // the fixed gain that synthesis should use 1126 float voice_fac, stab_fac; // parameters used for gain smoothing 1127 float synth_exc[AMRWB_SFR_SIZE]; // post-processed excitation for synthesis 1128 float hb_exc[AMRWB_SFR_SIZE_16k]; // excitation for the high frequency band 1129 float hb_samples[AMRWB_SFR_SIZE_16k]; // filtered high-band samples from synthesis 1130 float hb_gain; 1131 float *buf_out = (float *)frame->extended_data[ch]; 1132 1133 header_size = decode_mime_header(ctx, buf); 1134 expected_fr_size = ((cf_sizes_wb[ctx->fr_cur_mode] + 7) >> 3) + 1; 1135 1136 if (!ctx->fr_quality) 1137 av_log(avctx, AV_LOG_ERROR, "Encountered a bad or corrupted frame\n"); 1138 1139 if (ctx->fr_cur_mode == NO_DATA || !ctx->fr_quality) { 1140 /* The specification suggests a "random signal" and 1141 "a muting technique" to "gradually decrease the output level". */ 1142 av_samples_set_silence(&frame->extended_data[ch], 0, frame->nb_samples, 1, AV_SAMPLE_FMT_FLT); 1143 buf += expected_fr_size; 1144 buf_size -= expected_fr_size; 1145 continue; 1146 } 1147 if (ctx->fr_cur_mode > MODE_SID) { 1148 av_log(avctx, AV_LOG_ERROR, 1149 "Invalid mode %d\n", ctx->fr_cur_mode); 1150 return AVERROR_INVALIDDATA; 1151 } 1152 1153 if (buf_size < expected_fr_size) { 1154 av_log(avctx, AV_LOG_ERROR, 1155 "Frame too small (%d bytes). Truncated file?\n", buf_size); 1156 *got_frame_ptr = 0; 1157 return AVERROR_INVALIDDATA; 1158 } 1159 1160 if (ctx->fr_cur_mode == MODE_SID) { /* Comfort noise frame */ 1161 avpriv_request_sample(avctx, "SID mode"); 1162 return AVERROR_PATCHWELCOME; 1163 } 1164 1165 ff_amr_bit_reorder((uint16_t *) &ctx->frame, sizeof(AMRWBFrame), 1166 buf + header_size, amr_bit_orderings_by_mode[ctx->fr_cur_mode]); 1167 1168 /* Decode the quantized ISF vector */ 1169 if (ctx->fr_cur_mode == MODE_6k60) { 1170 decode_isf_indices_36b(cf->isp_id, ctx->isf_cur); 1171 } else { 1172 decode_isf_indices_46b(cf->isp_id, ctx->isf_cur); 1173 } 1174 1175 isf_add_mean_and_past(ctx->isf_cur, ctx->isf_q_past); 1176 ff_set_min_dist_lsf(ctx->isf_cur, MIN_ISF_SPACING, LP_ORDER - 1); 1177 1178 stab_fac = stability_factor(ctx->isf_cur, ctx->isf_past_final); 1179 1180 ctx->isf_cur[LP_ORDER - 1] *= 2.0; 1181 ff_acelp_lsf2lspd(ctx->isp[3], ctx->isf_cur, LP_ORDER); 1182 1183 /* Generate a ISP vector for each subframe */ 1184 if (ctx->first_frame) { 1185 ctx->first_frame = 0; 1186 memcpy(ctx->isp_sub4_past, ctx->isp[3], LP_ORDER * sizeof(double)); 1187 } 1188 interpolate_isp(ctx->isp, ctx->isp_sub4_past); 1189 1190 for (sub = 0; sub < 4; sub++) 1191 ff_amrwb_lsp2lpc(ctx->isp[sub], ctx->lp_coef[sub], LP_ORDER); 1192 1193 for (sub = 0; sub < 4; sub++) { 1194 const AMRWBSubFrame *cur_subframe = &cf->subframe[sub]; 1195 float *sub_buf = buf_out + sub * AMRWB_SFR_SIZE_16k; 1196 1197 /* Decode adaptive codebook (pitch vector) */ 1198 decode_pitch_vector(ctx, cur_subframe, sub); 1199 /* Decode innovative codebook (fixed vector) */ 1200 decode_fixed_vector(ctx->fixed_vector, cur_subframe->pul_ih, 1201 cur_subframe->pul_il, ctx->fr_cur_mode); 1202 1203 pitch_sharpening(ctx, ctx->fixed_vector); 1204 1205 decode_gains(cur_subframe->vq_gain, ctx->fr_cur_mode, 1206 &fixed_gain_factor, &ctx->pitch_gain[0]); 1207 1208 ctx->fixed_gain[0] = 1209 ff_amr_set_fixed_gain(fixed_gain_factor, 1210 ctx->celpm_ctx.dot_productf(ctx->fixed_vector, 1211 ctx->fixed_vector, 1212 AMRWB_SFR_SIZE) / 1213 AMRWB_SFR_SIZE, 1214 ctx->prediction_error, 1215 ENERGY_MEAN, energy_pred_fac); 1216 1217 /* Calculate voice factor and store tilt for next subframe */ 1218 voice_fac = voice_factor(ctx->pitch_vector, ctx->pitch_gain[0], 1219 ctx->fixed_vector, ctx->fixed_gain[0], 1220 &ctx->celpm_ctx); 1221 ctx->tilt_coef = voice_fac * 0.25 + 0.25; 1222 1223 /* Construct current excitation */ 1224 for (i = 0; i < AMRWB_SFR_SIZE; i++) { 1225 ctx->excitation[i] *= ctx->pitch_gain[0]; 1226 ctx->excitation[i] += ctx->fixed_gain[0] * ctx->fixed_vector[i]; 1227 ctx->excitation[i] = truncf(ctx->excitation[i]); 1228 } 1229 1230 /* Post-processing of excitation elements */ 1231 synth_fixed_gain = noise_enhancer(ctx->fixed_gain[0], &ctx->prev_tr_gain, 1232 voice_fac, stab_fac); 1233 1234 synth_fixed_vector = anti_sparseness(ctx, ctx->fixed_vector, 1235 spare_vector); 1236 1237 pitch_enhancer(synth_fixed_vector, voice_fac); 1238 1239 synthesis(ctx, ctx->lp_coef[sub], synth_exc, synth_fixed_gain, 1240 synth_fixed_vector, &ctx->samples_az[LP_ORDER]); 1241 1242 /* Synthesis speech post-processing */ 1243 de_emphasis(&ctx->samples_up[UPS_MEM_SIZE], 1244 &ctx->samples_az[LP_ORDER], PREEMPH_FAC, ctx->demph_mem); 1245 1246 ctx->acelpf_ctx.acelp_apply_order_2_transfer_function(&ctx->samples_up[UPS_MEM_SIZE], 1247 &ctx->samples_up[UPS_MEM_SIZE], hpf_zeros, hpf_31_poles, 1248 hpf_31_gain, ctx->hpf_31_mem, AMRWB_SFR_SIZE); 1249 1250 upsample_5_4(sub_buf, &ctx->samples_up[UPS_FIR_SIZE], 1251 AMRWB_SFR_SIZE_16k, &ctx->celpm_ctx); 1252 1253 /* High frequency band (6.4 - 7.0 kHz) generation part */ 1254 ctx->acelpf_ctx.acelp_apply_order_2_transfer_function(hb_samples, 1255 &ctx->samples_up[UPS_MEM_SIZE], hpf_zeros, hpf_400_poles, 1256 hpf_400_gain, ctx->hpf_400_mem, AMRWB_SFR_SIZE); 1257 1258 hb_gain = find_hb_gain(ctx, hb_samples, 1259 cur_subframe->hb_gain, cf->vad); 1260 1261 scaled_hb_excitation(ctx, hb_exc, synth_exc, hb_gain); 1262 1263 hb_synthesis(ctx, sub, &ctx->samples_hb[LP_ORDER_16k], 1264 hb_exc, ctx->isf_cur, ctx->isf_past_final); 1265 1266 /* High-band post-processing filters */ 1267 hb_fir_filter(hb_samples, bpf_6_7_coef, ctx->bpf_6_7_mem, 1268 &ctx->samples_hb[LP_ORDER_16k]); 1269 1270 if (ctx->fr_cur_mode == MODE_23k85) 1271 hb_fir_filter(hb_samples, lpf_7_coef, ctx->lpf_7_mem, 1272 hb_samples); 1273 1274 /* Add the low and high frequency bands */ 1275 for (i = 0; i < AMRWB_SFR_SIZE_16k; i++) 1276 sub_buf[i] = (sub_buf[i] + hb_samples[i]) * (1.0f / (1 << 15)); 1277 1278 /* Update buffers and history */ 1279 update_sub_state(ctx); 1280 } 1281 1282 /* update state for next frame */ 1283 memcpy(ctx->isp_sub4_past, ctx->isp[3], LP_ORDER * sizeof(ctx->isp[3][0])); 1284 memcpy(ctx->isf_past_final, ctx->isf_cur, LP_ORDER * sizeof(float)); 1285 1286 buf += expected_fr_size; 1287 buf_size -= expected_fr_size; 1288 } 1289 1290 *got_frame_ptr = 1; 1291 1292 return avpkt->size; 1293} 1294 1295const FFCodec ff_amrwb_decoder = { 1296 .p.name = "amrwb", 1297 .p.long_name = NULL_IF_CONFIG_SMALL("AMR-WB (Adaptive Multi-Rate WideBand)"), 1298 .p.type = AVMEDIA_TYPE_AUDIO, 1299 .p.id = AV_CODEC_ID_AMR_WB, 1300 .priv_data_size = sizeof(AMRWBChannelsContext), 1301 .init = amrwb_decode_init, 1302 FF_CODEC_DECODE_CB(amrwb_decode_frame), 1303 .p.capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_CHANNEL_CONF, 1304 .p.sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLT, 1305 AV_SAMPLE_FMT_NONE }, 1306 .caps_internal = FF_CODEC_CAP_INIT_THREADSAFE, 1307}; 1308