FFmpeg: libavcodec/amrwbdec.c Source File

FFmpeg

[フレーム]

libavcodec

amrwbdec.c

Go to the documentation of this file.

1 /*

2 * AMR wideband decoder

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 */

22 /**

23 * @file

24 * AMR wideband decoder

25 */

27 #include "libavutil/channel_layout.h"

28 #include "libavutil/common.h"

29 #include "libavutil/lfg.h"

31 #include "avcodec.h"

32 #include "dsputil.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 "internal.h"

41 #define AMR_USE_16BIT_TABLES

42 #include "amr.h"

44 #include "amrwbdata.h"

45 #include "mips/amrwbdec_mips.h"

47 typedef struct {

48 AVFrame avframe; ///< AVFrame for decoded samples

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

58 float lp_coef[4][LP_ORDER]; ///< Linear Prediction Coefficients from ISP vector

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

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[]

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

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

73 float tilt_coef; ///< {beta_1} related to the voicing of the previous subframe

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

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

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

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

95 } AMRWBContext;

97 static av_cold int amrwb_decode_init(AVCodecContext *avctx)

98 {

99 AMRWBContext *ctx = avctx->priv_data;

100 int i;

101

102 if (avctx->channels > 1) {

103 av_log_missing_feature(avctx, "multi-channel AMR", 0);

104 return AVERROR_PATCHWELCOME;

105 }

106

107 avctx->channels = 1;

108 avctx->channel_layout = AV_CH_LAYOUT_MONO;

109 if (!avctx->sample_rate)

110 avctx->sample_rate = 16000;

111 avctx->sample_fmt = AV_SAMPLE_FMT_FLT;

112

113 av_lfg_init(&ctx->prng, 1);

114

115 ctx->excitation = &ctx->excitation_buf[AMRWB_P_DELAY_MAX + LP_ORDER + 1];

116 ctx->first_frame = 1;

117

118 for (i = 0; i < LP_ORDER; i++)

119 ctx->isf_past_final[i] = isf_init[i] * (1.0f / (1 << 15));

120

121 for (i = 0; i < 4; i++)

122 ctx->prediction_error[i] = MIN_ENERGY;

123

124 avcodec_get_frame_defaults(&ctx->avframe);

125 avctx->coded_frame = &ctx->avframe;

126

127 ff_acelp_filter_init(&ctx->acelpf_ctx);

128 ff_acelp_vectors_init(&ctx->acelpv_ctx);

129 ff_celp_filter_init(&ctx->celpf_ctx);

130 ff_celp_math_init(&ctx->celpm_ctx);

131

132 return 0;

133 }

134

135 /**

136 * Decode the frame header in the "MIME/storage" format. This format

137 * is simpler and does not carry the auxiliary frame information.

138 *

139 * @param[in] ctx The Context

140 * @param[in] buf Pointer to the input buffer

141 *

142 * @return The decoded header length in bytes

143 */

144 static int decode_mime_header(AMRWBContext *ctx, const uint8_t *buf)

145 {

146 /* Decode frame header (1st octet) */

147 ctx->fr_cur_mode = buf[0] >> 3 & 0x0F;

148 ctx->fr_quality = (buf[0] & 0x4) == 0x4;

149

150 return 1;

151 }

152

153 /**

154 * Decode quantized ISF vectors using 36-bit indexes (6K60 mode only).

155 *

156 * @param[in] ind Array of 5 indexes

157 * @param[out] isf_q Buffer for isf_q[LP_ORDER]

158 *

159 */

160 static void decode_isf_indices_36b(uint16_t *ind, float *isf_q)

161 {

162 int i;

163

164 for (i = 0; i < 9; i++)

165 isf_q[i] = dico1_isf[ind[0]][i] * (1.0f / (1 << 15));

166

167 for (i = 0; i < 7; i++)

168 isf_q[i + 9] = dico2_isf[ind[1]][i] * (1.0f / (1 << 15));

169

170 for (i = 0; i < 5; i++)

171 isf_q[i] += dico21_isf_36b[ind[2]][i] * (1.0f / (1 << 15));

172

173 for (i = 0; i < 4; i++)

174 isf_q[i + 5] += dico22_isf_36b[ind[3]][i] * (1.0f / (1 << 15));

175

176 for (i = 0; i < 7; i++)

177 isf_q[i + 9] += dico23_isf_36b[ind[4]][i] * (1.0f / (1 << 15));

178 }

179

180 /**

181 * Decode quantized ISF vectors using 46-bit indexes (except 6K60 mode).

182 *

183 * @param[in] ind Array of 7 indexes

184 * @param[out] isf_q Buffer for isf_q[LP_ORDER]

185 *

186 */

187 static void decode_isf_indices_46b(uint16_t *ind, float *isf_q)

188 {

189 int i;

190

191 for (i = 0; i < 9; i++)

192 isf_q[i] = dico1_isf[ind[0]][i] * (1.0f / (1 << 15));

193

194 for (i = 0; i < 7; i++)

195 isf_q[i + 9] = dico2_isf[ind[1]][i] * (1.0f / (1 << 15));

196

197 for (i = 0; i < 3; i++)

198 isf_q[i] += dico21_isf[ind[2]][i] * (1.0f / (1 << 15));

199

200 for (i = 0; i < 3; i++)

201 isf_q[i + 3] += dico22_isf[ind[3]][i] * (1.0f / (1 << 15));

202

203 for (i = 0; i < 3; i++)

204 isf_q[i + 6] += dico23_isf[ind[4]][i] * (1.0f / (1 << 15));

205

206 for (i = 0; i < 3; i++)

207 isf_q[i + 9] += dico24_isf[ind[5]][i] * (1.0f / (1 << 15));

208

209 for (i = 0; i < 4; i++)

210 isf_q[i + 12] += dico25_isf[ind[6]][i] * (1.0f / (1 << 15));

211 }

212

213 /**

214 * Apply mean and past ISF values using the prediction factor.

215 * Updates past ISF vector.

216 *

217 * @param[in,out] isf_q Current quantized ISF

218 * @param[in,out] isf_past Past quantized ISF

219 *

220 */

221 static void isf_add_mean_and_past(float *isf_q, float *isf_past)

222 {

223 int i;

224 float tmp;

225

226 for (i = 0; i < LP_ORDER; i++) {

227 tmp = isf_q[i];

228 isf_q[i] += isf_mean[i] * (1.0f / (1 << 15));

229 isf_q[i] += PRED_FACTOR * isf_past[i];

230 isf_past[i] = tmp;

231 }

232 }

233

234 /**

235 * Interpolate the fourth ISP vector from current and past frames

236 * to obtain an ISP vector for each subframe.

237 *

238 * @param[in,out] isp_q ISPs for each subframe

239 * @param[in] isp4_past Past ISP for subframe 4

240 */

241 static void interpolate_isp(double isp_q[4][LP_ORDER], const double *isp4_past)

242 {

243 int i, k;

244

245 for (k = 0; k < 3; k++) {

246 float c = isfp_inter[k];

247 for (i = 0; i < LP_ORDER; i++)

248 isp_q[k][i] = (1.0 - c) * isp4_past[i] + c * isp_q[3][i];

249 }

250 }

251

252 /**

253 * Decode an adaptive codebook index into pitch lag (except 6k60, 8k85 modes).

254 * Calculate integer lag and fractional lag always using 1/4 resolution.

255 * In 1st and 3rd subframes the index is relative to last subframe integer lag.

256 *

257 * @param[out] lag_int Decoded integer pitch lag

258 * @param[out] lag_frac Decoded fractional pitch lag

259 * @param[in] pitch_index Adaptive codebook pitch index

260 * @param[in,out] base_lag_int Base integer lag used in relative subframes

261 * @param[in] subframe Current subframe index (0 to 3)

262 */

263 static void decode_pitch_lag_high(int *lag_int, int *lag_frac, int pitch_index,

264 uint8_t *base_lag_int, int subframe)

265 {

266 if (subframe == 0 || subframe == 2) {

267 if (pitch_index < 376) {

268 *lag_int = (pitch_index + 137) >> 2;

269 *lag_frac = pitch_index - (*lag_int << 2) + 136;

270 } else if (pitch_index < 440) {

271 *lag_int = (pitch_index + 257 - 376) >> 1;

272 *lag_frac = (pitch_index - (*lag_int << 1) + 256 - 376) << 1;

273 /* the actual resolution is 1/2 but expressed as 1/4 */

274 } else {

275 *lag_int = pitch_index - 280;

276 *lag_frac = 0;

277 }

278 /* minimum lag for next subframe */

279 *base_lag_int = av_clip(*lag_int - 8 - (*lag_frac < 0),

280 AMRWB_P_DELAY_MIN, AMRWB_P_DELAY_MAX - 15);

281 // XXX: the spec states clearly that *base_lag_int should be

282 // the nearest integer to *lag_int (minus 8), but the ref code

283 // actually always uses its floor, I'm following the latter

284 } else {

285 *lag_int = (pitch_index + 1) >> 2;

286 *lag_frac = pitch_index - (*lag_int << 2);

287 *lag_int += *base_lag_int;

288 }

289 }

290

291 /**

292 * Decode an adaptive codebook index into pitch lag for 8k85 and 6k60 modes.

293 * The description is analogous to decode_pitch_lag_high, but in 6k60 the

294 * relative index is used for all subframes except the first.

295 */

296 static void decode_pitch_lag_low(int *lag_int, int *lag_frac, int pitch_index,

297 uint8_t *base_lag_int, int subframe, enum Mode mode)

298 {

299 if (subframe == 0 || (subframe == 2 && mode != MODE_6k60)) {

300 if (pitch_index < 116) {

301 *lag_int = (pitch_index + 69) >> 1;

302 *lag_frac = (pitch_index - (*lag_int << 1) + 68) << 1;

303 } else {

304 *lag_int = pitch_index - 24;

305 *lag_frac = 0;

306 }

307 // XXX: same problem as before

308 *base_lag_int = av_clip(*lag_int - 8 - (*lag_frac < 0),

309 AMRWB_P_DELAY_MIN, AMRWB_P_DELAY_MAX - 15);

310 } else {

311 *lag_int = (pitch_index + 1) >> 1;

312 *lag_frac = (pitch_index - (*lag_int << 1)) << 1;

313 *lag_int += *base_lag_int;

314 }

315 }

316

317 /**

318 * Find the pitch vector by interpolating the past excitation at the

319 * pitch delay, which is obtained in this function.

320 *

321 * @param[in,out] ctx The context

322 * @param[in] amr_subframe Current subframe data

323 * @param[in] subframe Current subframe index (0 to 3)

324 */

325 static void decode_pitch_vector(AMRWBContext *ctx,

326 const AMRWBSubFrame *amr_subframe,

327 const int subframe)

328 {

329 int pitch_lag_int, pitch_lag_frac;

330 int i;

331 float *exc = ctx->excitation;

332 enum Mode mode = ctx->fr_cur_mode;

333

334 if (mode <= MODE_8k85) {

335 decode_pitch_lag_low(&pitch_lag_int, &pitch_lag_frac, amr_subframe->adap,

336 &ctx->base_pitch_lag, subframe, mode);

337 } else

338 decode_pitch_lag_high(&pitch_lag_int, &pitch_lag_frac, amr_subframe->adap,

339 &ctx->base_pitch_lag, subframe);

340

341 ctx->pitch_lag_int = pitch_lag_int;

342 pitch_lag_int += pitch_lag_frac > 0;

343

344 /* Calculate the pitch vector by interpolating the past excitation at the

345 pitch lag using a hamming windowed sinc function */

346 ctx->acelpf_ctx.acelp_interpolatef(exc,

347 exc + 1 - pitch_lag_int,

348 ac_inter, 4,

349 pitch_lag_frac + (pitch_lag_frac > 0 ? 0 : 4),

350 LP_ORDER, AMRWB_SFR_SIZE + 1);

351

352 /* Check which pitch signal path should be used

353 * 6k60 and 8k85 modes have the ltp flag set to 0 */

354 if (amr_subframe->ltp) {

355 memcpy(ctx->pitch_vector, exc, AMRWB_SFR_SIZE * sizeof(float));

356 } else {

357 for (i = 0; i < AMRWB_SFR_SIZE; i++)

358 ctx->pitch_vector[i] = 0.18 * exc[i - 1] + 0.64 * exc[i] +

359 0.18 * exc[i + 1];

360 memcpy(exc, ctx->pitch_vector, AMRWB_SFR_SIZE * sizeof(float));

361 }

362 }

363

364 /** Get x bits in the index interval [lsb,lsb+len-1] inclusive */

365 #define BIT_STR(x,lsb,len) (((x) >> (lsb)) & ((1 << (len)) - 1))

366

367 /** Get the bit at specified position */

368 #define BIT_POS(x, p) (((x) >> (p)) & 1)

369

370 /**

371 * The next six functions decode_[i]p_track decode exactly i pulses

372 * positions and amplitudes (-1 or 1) in a subframe track using

373 * an encoded pulse indexing (TS 26.190 section 5.8.2).

374 *

375 * The results are given in out[], in which a negative number means

376 * amplitude -1 and vice versa (i.e., ampl(x) = x / abs(x) ).

377 *

378 * @param[out] out Output buffer (writes i elements)

379 * @param[in] code Pulse index (no. of bits varies, see below)

380 * @param[in] m (log2) Number of potential positions

381 * @param[in] off Offset for decoded positions

382 */

383 static inline void decode_1p_track(int *out, int code, int m, int off)

384 {

385 int pos = BIT_STR(code, 0, m) + off; ///code: m+1 bits

386

387 out[0] = BIT_POS(code, m) ? -pos : pos;

388 }

389

390 static inline void decode_2p_track(int *out, int code, int m, int off) ///code: 2m+1 bits

391 {

392 int pos0 = BIT_STR(code, m, m) + off;

393 int pos1 = BIT_STR(code, 0, m) + off;

394

395 out[0] = BIT_POS(code, 2*m) ? -pos0 : pos0;

396 out[1] = BIT_POS(code, 2*m) ? -pos1 : pos1;

397 out[1] = pos0 > pos1 ? -out[1] : out[1];

398 }

399

400 static void decode_3p_track(int *out, int code, int m, int off) ///code: 3m+1 bits

401 {

402 int half_2p = BIT_POS(code, 2*m - 1) << (m - 1);

403

404 decode_2p_track(out, BIT_STR(code, 0, 2*m - 1),

405 m - 1, off + half_2p);

406 decode_1p_track(out + 2, BIT_STR(code, 2*m, m + 1), m, off);

407 }

408

409 static void decode_4p_track(int *out, int code, int m, int off) ///code: 4m bits

410 {

411 int half_4p, subhalf_2p;

412 int b_offset = 1 << (m - 1);

413

414 switch (BIT_STR(code, 4*m - 2, 2)) { /* case ID (2 bits) */

415 case 0: /* 0 pulses in A, 4 pulses in B or vice versa */

416 half_4p = BIT_POS(code, 4*m - 3) << (m - 1); // which has 4 pulses

417 subhalf_2p = BIT_POS(code, 2*m - 3) << (m - 2);

418

419 decode_2p_track(out, BIT_STR(code, 0, 2*m - 3),

420 m - 2, off + half_4p + subhalf_2p);

421 decode_2p_track(out + 2, BIT_STR(code, 2*m - 2, 2*m - 1),

422 m - 1, off + half_4p);

423 break;

424 case 1: /* 1 pulse in A, 3 pulses in B */

425 decode_1p_track(out, BIT_STR(code, 3*m - 2, m),

426 m - 1, off);

427 decode_3p_track(out + 1, BIT_STR(code, 0, 3*m - 2),

428 m - 1, off + b_offset);

429 break;

430 case 2: /* 2 pulses in each half */

431 decode_2p_track(out, BIT_STR(code, 2*m - 1, 2*m - 1),

432 m - 1, off);

433 decode_2p_track(out + 2, BIT_STR(code, 0, 2*m - 1),

434 m - 1, off + b_offset);

435 break;

436 case 3: /* 3 pulses in A, 1 pulse in B */

437 decode_3p_track(out, BIT_STR(code, m, 3*m - 2),

438 m - 1, off);

439 decode_1p_track(out + 3, BIT_STR(code, 0, m),

440 m - 1, off + b_offset);

441 break;

442 }

443 }

444

445 static void decode_5p_track(int *out, int code, int m, int off) ///code: 5m bits

446 {

447 int half_3p = BIT_POS(code, 5*m - 1) << (m - 1);

448

449 decode_3p_track(out, BIT_STR(code, 2*m + 1, 3*m - 2),

450 m - 1, off + half_3p);

451

452 decode_2p_track(out + 3, BIT_STR(code, 0, 2*m + 1), m, off);

453 }

454

455 static void decode_6p_track(int *out, int code, int m, int off) ///code: 6m-2 bits

456 {

457 int b_offset = 1 << (m - 1);

458 /* which half has more pulses in cases 0 to 2 */

459 int half_more = BIT_POS(code, 6*m - 5) << (m - 1);

460 int half_other = b_offset - half_more;

461

462 switch (BIT_STR(code, 6*m - 4, 2)) { /* case ID (2 bits) */

463 case 0: /* 0 pulses in A, 6 pulses in B or vice versa */

464 decode_1p_track(out, BIT_STR(code, 0, m),

465 m - 1, off + half_more);

466 decode_5p_track(out + 1, BIT_STR(code, m, 5*m - 5),

467 m - 1, off + half_more);

468 break;

469 case 1: /* 1 pulse in A, 5 pulses in B or vice versa */

470 decode_1p_track(out, BIT_STR(code, 0, m),

471 m - 1, off + half_other);

472 decode_5p_track(out + 1, BIT_STR(code, m, 5*m - 5),

473 m - 1, off + half_more);

474 break;

475 case 2: /* 2 pulses in A, 4 pulses in B or vice versa */

476 decode_2p_track(out, BIT_STR(code, 0, 2*m - 1),

477 m - 1, off + half_other);

478 decode_4p_track(out + 2, BIT_STR(code, 2*m - 1, 4*m - 4),

479 m - 1, off + half_more);

480 break;

481 case 3: /* 3 pulses in A, 3 pulses in B */

482 decode_3p_track(out, BIT_STR(code, 3*m - 2, 3*m - 2),

483 m - 1, off);

484 decode_3p_track(out + 3, BIT_STR(code, 0, 3*m - 2),

485 m - 1, off + b_offset);

486 break;

487 }

488 }

489

490 /**

491 * Decode the algebraic codebook index to pulse positions and signs,

492 * then construct the algebraic codebook vector.

493 *

494 * @param[out] fixed_vector Buffer for the fixed codebook excitation

495 * @param[in] pulse_hi MSBs part of the pulse index array (higher modes only)

496 * @param[in] pulse_lo LSBs part of the pulse index array

497 * @param[in] mode Mode of the current frame

498 */

499 static void decode_fixed_vector(float *fixed_vector, const uint16_t *pulse_hi,

500 const uint16_t *pulse_lo, const enum Mode mode)

501 {

502 /* sig_pos stores for each track the decoded pulse position indexes

503 * (1-based) multiplied by its corresponding amplitude (+1 or -1) */

504 int sig_pos[4][6];

505 int spacing = (mode == MODE_6k60) ? 2 : 4;

506 int i, j;

507

508 switch (mode) {

509 case MODE_6k60:

510 for (i = 0; i < 2; i++)

511 decode_1p_track(sig_pos[i], pulse_lo[i], 5, 1);

512 break;

513 case MODE_8k85:

514 for (i = 0; i < 4; i++)

515 decode_1p_track(sig_pos[i], pulse_lo[i], 4, 1);

516 break;

517 case MODE_12k65:

518 for (i = 0; i < 4; i++)

519 decode_2p_track(sig_pos[i], pulse_lo[i], 4, 1);

520 break;

521 case MODE_14k25:

522 for (i = 0; i < 2; i++)

523 decode_3p_track(sig_pos[i], pulse_lo[i], 4, 1);

524 for (i = 2; i < 4; i++)

525 decode_2p_track(sig_pos[i], pulse_lo[i], 4, 1);

526 break;

527 case MODE_15k85:

528 for (i = 0; i < 4; i++)

529 decode_3p_track(sig_pos[i], pulse_lo[i], 4, 1);

530 break;

531 case MODE_18k25:

532 for (i = 0; i < 4; i++)

533 decode_4p_track(sig_pos[i], (int) pulse_lo[i] +

534 ((int) pulse_hi[i] << 14), 4, 1);

535 break;

536 case MODE_19k85:

537 for (i = 0; i < 2; i++)

538 decode_5p_track(sig_pos[i], (int) pulse_lo[i] +

539 ((int) pulse_hi[i] << 10), 4, 1);

540 for (i = 2; i < 4; i++)

541 decode_4p_track(sig_pos[i], (int) pulse_lo[i] +

542 ((int) pulse_hi[i] << 14), 4, 1);

543 break;

544 case MODE_23k05:

545 case MODE_23k85:

546 for (i = 0; i < 4; i++)

547 decode_6p_track(sig_pos[i], (int) pulse_lo[i] +

548 ((int) pulse_hi[i] << 11), 4, 1);

549 break;

550 }

551

552 memset(fixed_vector, 0, sizeof(float) * AMRWB_SFR_SIZE);

553

554 for (i = 0; i < 4; i++)

555 for (j = 0; j < pulses_nb_per_mode_tr[mode][i]; j++) {

556 int pos = (FFABS(sig_pos[i][j]) - 1) * spacing + i;

557

558 fixed_vector[pos] += sig_pos[i][j] < 0 ? -1.0 : 1.0;

559 }

560 }

561

562 /**

563 * Decode pitch gain and fixed gain correction factor.

564 *

565 * @param[in] vq_gain Vector-quantized index for gains

566 * @param[in] mode Mode of the current frame

567 * @param[out] fixed_gain_factor Decoded fixed gain correction factor

568 * @param[out] pitch_gain Decoded pitch gain

569 */

570 static void decode_gains(const uint8_t vq_gain, const enum Mode mode,

571 float *fixed_gain_factor, float *pitch_gain)

572 {

573 const int16_t *gains = (mode <= MODE_8k85 ? qua_gain_6b[vq_gain] :

574 qua_gain_7b[vq_gain]);

575

576 *pitch_gain = gains[0] * (1.0f / (1 << 14));

577 *fixed_gain_factor = gains[1] * (1.0f / (1 << 11));

578 }

579

580 /**

581 * Apply pitch sharpening filters to the fixed codebook vector.

582 *

583 * @param[in] ctx The context

584 * @param[in,out] fixed_vector Fixed codebook excitation

585 */

586 // XXX: Spec states this procedure should be applied when the pitch

587 // lag is less than 64, but this checking seems absent in reference and AMR-NB

588 static void pitch_sharpening(AMRWBContext *ctx, float *fixed_vector)

589 {

590 int i;

591

592 /* Tilt part */

593 for (i = AMRWB_SFR_SIZE - 1; i != 0; i--)

594 fixed_vector[i] -= fixed_vector[i - 1] * ctx->tilt_coef;

595

596 /* Periodicity enhancement part */

597 for (i = ctx->pitch_lag_int; i < AMRWB_SFR_SIZE; i++)

598 fixed_vector[i] += fixed_vector[i - ctx->pitch_lag_int] * 0.85;

599 }

600

601 /**

602 * Calculate the voicing factor (-1.0 = unvoiced to 1.0 = voiced).

603 *

604 * @param[in] p_vector, f_vector Pitch and fixed excitation vectors

605 * @param[in] p_gain, f_gain Pitch and fixed gains

606 * @param[in] ctx The context

607 */

608 // XXX: There is something wrong with the precision here! The magnitudes

609 // of the energies are not correct. Please check the reference code carefully

610 static float voice_factor(float *p_vector, float p_gain,

611 float *f_vector, float f_gain,

612 CELPMContext *ctx)

613 {

614 double p_ener = (double) ctx->dot_productf(p_vector, p_vector,

615 AMRWB_SFR_SIZE) *

616 p_gain * p_gain;

617 double f_ener = (double) ctx->dot_productf(f_vector, f_vector,

618 AMRWB_SFR_SIZE) *

619 f_gain * f_gain;

620

621 return (p_ener - f_ener) / (p_ener + f_ener);

622 }

623

624 /**

625 * Reduce fixed vector sparseness by smoothing with one of three IR filters,

626 * also known as "adaptive phase dispersion".

627 *

628 * @param[in] ctx The context

629 * @param[in,out] fixed_vector Unfiltered fixed vector

630 * @param[out] buf Space for modified vector if necessary

631 *

632 * @return The potentially overwritten filtered fixed vector address

633 */

634 static float *anti_sparseness(AMRWBContext *ctx,

635 float *fixed_vector, float *buf)

636 {

637 int ir_filter_nr;

638

639 if (ctx->fr_cur_mode > MODE_8k85) // no filtering in higher modes

640 return fixed_vector;

641

642 if (ctx->pitch_gain[0] < 0.6) {

643 ir_filter_nr = 0; // strong filtering

644 } else if (ctx->pitch_gain[0] < 0.9) {

645 ir_filter_nr = 1; // medium filtering

646 } else

647 ir_filter_nr = 2; // no filtering

648

649 /* detect 'onset' */

650 if (ctx->fixed_gain[0] > 3.0 * ctx->fixed_gain[1]) {

651 if (ir_filter_nr < 2)

652 ir_filter_nr++;

653 } else {

654 int i, count = 0;

655

656 for (i = 0; i < 6; i++)

657 if (ctx->pitch_gain[i] < 0.6)

658 count++;

659

660 if (count > 2)

661 ir_filter_nr = 0;

662

663 if (ir_filter_nr > ctx->prev_ir_filter_nr + 1)

664 ir_filter_nr--;

665 }

666

667 /* update ir filter strength history */

668 ctx->prev_ir_filter_nr = ir_filter_nr;

669

670 ir_filter_nr += (ctx->fr_cur_mode == MODE_8k85);

671

672 if (ir_filter_nr < 2) {

673 int i;

674 const float *coef = ir_filters_lookup[ir_filter_nr];

675

676 /* Circular convolution code in the reference

677 * decoder was modified to avoid using one

678 * extra array. The filtered vector is given by:

679 *

680 * c2(n) = sum(i,0,len-1){ c(i) * coef( (n - i + len) % len ) }

681 */

682

683 memset(buf, 0, sizeof(float) * AMRWB_SFR_SIZE);

684 for (i = 0; i < AMRWB_SFR_SIZE; i++)

685 if (fixed_vector[i])

686 ff_celp_circ_addf(buf, buf, coef, i, fixed_vector[i],

687 AMRWB_SFR_SIZE);

688 fixed_vector = buf;

689 }

690

691 return fixed_vector;

692 }

693

694 /**

695 * Calculate a stability factor {teta} based on distance between

696 * current and past isf. A value of 1 shows maximum signal stability.

697 */

698 static float stability_factor(const float *isf, const float *isf_past)

699 {

700 int i;

701 float acc = 0.0;

702

703 for (i = 0; i < LP_ORDER - 1; i++)

704 acc += (isf[i] - isf_past[i]) * (isf[i] - isf_past[i]);

705

706 // XXX: This part is not so clear from the reference code

707 // the result is more accurate changing the "/ 256" to "* 512"

708 return FFMAX(0.0, 1.25 - acc * 0.8 * 512);

709 }

710

711 /**

712 * Apply a non-linear fixed gain smoothing in order to reduce

713 * fluctuation in the energy of excitation.

714 *

715 * @param[in] fixed_gain Unsmoothed fixed gain

716 * @param[in,out] prev_tr_gain Previous threshold gain (updated)

717 * @param[in] voice_fac Frame voicing factor

718 * @param[in] stab_fac Frame stability factor

719 *

720 * @return The smoothed gain

721 */

722 static float noise_enhancer(float fixed_gain, float *prev_tr_gain,

723 float voice_fac, float stab_fac)

724 {

725 float sm_fac = 0.5 * (1 - voice_fac) * stab_fac;

726 float g0;

727

728 // XXX: the following fixed-point constants used to in(de)crement

729 // gain by 1.5dB were taken from the reference code, maybe it could

730 // be simpler

731 if (fixed_gain < *prev_tr_gain) {

732 g0 = FFMIN(*prev_tr_gain, fixed_gain + fixed_gain *

733 (6226 * (1.0f / (1 << 15)))); // +1.5 dB

734 } else

735 g0 = FFMAX(*prev_tr_gain, fixed_gain *

736 (27536 * (1.0f / (1 << 15)))); // -1.5 dB

737

738 *prev_tr_gain = g0; // update next frame threshold

739

740 return sm_fac * g0 + (1 - sm_fac) * fixed_gain;

741 }

742

743 /**

744 * Filter the fixed_vector to emphasize the higher frequencies.

745 *

746 * @param[in,out] fixed_vector Fixed codebook vector

747 * @param[in] voice_fac Frame voicing factor

748 */

749 static void pitch_enhancer(float *fixed_vector, float voice_fac)

750 {

751 int i;

752 float cpe = 0.125 * (1 + voice_fac);

753 float last = fixed_vector[0]; // holds c(i - 1)

754

755 fixed_vector[0] -= cpe * fixed_vector[1];

756

757 for (i = 1; i < AMRWB_SFR_SIZE - 1; i++) {

758 float cur = fixed_vector[i];

759

760 fixed_vector[i] -= cpe * (last + fixed_vector[i + 1]);

761 last = cur;

762 }

763

764 fixed_vector[AMRWB_SFR_SIZE - 1] -= cpe * last;

765 }

766

767 /**

768 * Conduct 16th order linear predictive coding synthesis from excitation.

769 *

770 * @param[in] ctx Pointer to the AMRWBContext

771 * @param[in] lpc Pointer to the LPC coefficients

772 * @param[out] excitation Buffer for synthesis final excitation

773 * @param[in] fixed_gain Fixed codebook gain for synthesis

774 * @param[in] fixed_vector Algebraic codebook vector

775 * @param[in,out] samples Pointer to the output samples and memory

776 */

777 static void synthesis(AMRWBContext *ctx, float *lpc, float *excitation,

778 float fixed_gain, const float *fixed_vector,

779 float *samples)

780 {

781 ctx->acelpv_ctx.weighted_vector_sumf(excitation, ctx->pitch_vector, fixed_vector,

782 ctx->pitch_gain[0], fixed_gain, AMRWB_SFR_SIZE);

783

784 /* emphasize pitch vector contribution in low bitrate modes */

785 if (ctx->pitch_gain[0] > 0.5 && ctx->fr_cur_mode <= MODE_8k85) {

786 int i;

787 float energy = ctx->celpm_ctx.dot_productf(excitation, excitation,

788 AMRWB_SFR_SIZE);

789

790 // XXX: Weird part in both ref code and spec. A unknown parameter

791 // {beta} seems to be identical to the current pitch gain

792 float pitch_factor = 0.25 * ctx->pitch_gain[0] * ctx->pitch_gain[0];

793

794 for (i = 0; i < AMRWB_SFR_SIZE; i++)

795 excitation[i] += pitch_factor * ctx->pitch_vector[i];

796

797 ff_scale_vector_to_given_sum_of_squares(excitation, excitation,

798 energy, AMRWB_SFR_SIZE);

799 }

800

801 ctx->celpf_ctx.celp_lp_synthesis_filterf(samples, lpc, excitation,

802 AMRWB_SFR_SIZE, LP_ORDER);

803 }

804

805 /**

806 * Apply to synthesis a de-emphasis filter of the form:

807 * H(z) = 1 / (1 - m * z^-1)

808 *

809 * @param[out] out Output buffer

810 * @param[in] in Input samples array with in[-1]

811 * @param[in] m Filter coefficient

812 * @param[in,out] mem State from last filtering

813 */

814 static void de_emphasis(float *out, float *in, float m, float mem[1])

815 {

816 int i;

817

818 out[0] = in[0] + m * mem[0];

819

820 for (i = 1; i < AMRWB_SFR_SIZE; i++)

821 out[i] = in[i] + out[i - 1] * m;

822

823 mem[0] = out[AMRWB_SFR_SIZE - 1];

824 }

825

826 /**

827 * Upsample a signal by 5/4 ratio (from 12.8kHz to 16kHz) using

828 * a FIR interpolation filter. Uses past data from before *in address.

829 *

830 * @param[out] out Buffer for interpolated signal

831 * @param[in] in Current signal data (length 0.8*o_size)

832 * @param[in] o_size Output signal length

833 * @param[in] ctx The context

834 */

835 static void upsample_5_4(float *out, const float *in, int o_size, CELPMContext *ctx)

836 {

837 const float *in0 = in - UPS_FIR_SIZE + 1;

838 int i, j, k;

839 int int_part = 0, frac_part;

840

841 i = 0;

842 for (j = 0; j < o_size / 5; j++) {

843 out[i] = in[int_part];

844 frac_part = 4;

845 i++;

846

847 for (k = 1; k < 5; k++) {

848 out[i] = ctx->dot_productf(in0 + int_part,

849 upsample_fir[4 - frac_part],

850 UPS_MEM_SIZE);

851 int_part++;

852 frac_part--;

853 i++;

854 }

855 }

856 }

857

858 /**

859 * Calculate the high-band gain based on encoded index (23k85 mode) or

860 * on the low-band speech signal and the Voice Activity Detection flag.

861 *

862 * @param[in] ctx The context

863 * @param[in] synth LB speech synthesis at 12.8k

864 * @param[in] hb_idx Gain index for mode 23k85 only

865 * @param[in] vad VAD flag for the frame

866 */

867 static float find_hb_gain(AMRWBContext *ctx, const float *synth,

868 uint16_t hb_idx, uint8_t vad)

869 {

870 int wsp = (vad > 0);

871 float tilt;

872

873 if (ctx->fr_cur_mode == MODE_23k85)

874 return qua_hb_gain[hb_idx] * (1.0f / (1 << 14));

875

876 tilt = ctx->celpm_ctx.dot_productf(synth, synth + 1, AMRWB_SFR_SIZE - 1) /

877 ctx->celpm_ctx.dot_productf(synth, synth, AMRWB_SFR_SIZE);

878

879 /* return gain bounded by [0.1, 1.0] */

880 return av_clipf((1.0 - FFMAX(0.0, tilt)) * (1.25 - 0.25 * wsp), 0.1, 1.0);

881 }

882

883 /**

884 * Generate the high-band excitation with the same energy from the lower

885 * one and scaled by the given gain.

886 *

887 * @param[in] ctx The context

888 * @param[out] hb_exc Buffer for the excitation

889 * @param[in] synth_exc Low-band excitation used for synthesis

890 * @param[in] hb_gain Wanted excitation gain

891 */

892 static void scaled_hb_excitation(AMRWBContext *ctx, float *hb_exc,

893 const float *synth_exc, float hb_gain)

894 {

895 int i;

896 float energy = ctx->celpm_ctx.dot_productf(synth_exc, synth_exc, AMRWB_SFR_SIZE);

897

898 /* Generate a white-noise excitation */

899 for (i = 0; i < AMRWB_SFR_SIZE_16k; i++)

900 hb_exc[i] = 32768.0 - (uint16_t) av_lfg_get(&ctx->prng);

901

902 ff_scale_vector_to_given_sum_of_squares(hb_exc, hb_exc,

903 energy * hb_gain * hb_gain,

904 AMRWB_SFR_SIZE_16k);

905 }

906

907 /**

908 * Calculate the auto-correlation for the ISF difference vector.

909 */

910 static float auto_correlation(float *diff_isf, float mean, int lag)

911 {

912 int i;

913 float sum = 0.0;

914

915 for (i = 7; i < LP_ORDER - 2; i++) {

916 float prod = (diff_isf[i] - mean) * (diff_isf[i - lag] - mean);

917 sum += prod * prod;

918 }

919 return sum;

920 }

921

922 /**

923 * Extrapolate a ISF vector to the 16kHz range (20th order LP)

924 * used at mode 6k60 LP filter for the high frequency band.

925 *

926 * @param[out] isf Buffer for extrapolated isf; contains LP_ORDER

927 * values on input

928 */

929 static void extrapolate_isf(float isf[LP_ORDER_16k])

930 {

931 float diff_isf[LP_ORDER - 2], diff_mean;

932 float corr_lag[3];

933 float est, scale;

934 int i, j, i_max_corr;

935

936 isf[LP_ORDER_16k - 1] = isf[LP_ORDER - 1];

937

938 /* Calculate the difference vector */

939 for (i = 0; i < LP_ORDER - 2; i++)

940 diff_isf[i] = isf[i + 1] - isf[i];

941

942 diff_mean = 0.0;

943 for (i = 2; i < LP_ORDER - 2; i++)

944 diff_mean += diff_isf[i] * (1.0f / (LP_ORDER - 4));

945

946 /* Find which is the maximum autocorrelation */

947 i_max_corr = 0;

948 for (i = 0; i < 3; i++) {

949 corr_lag[i] = auto_correlation(diff_isf, diff_mean, i + 2);

950

951 if (corr_lag[i] > corr_lag[i_max_corr])

952 i_max_corr = i;

953 }

954 i_max_corr++;

955

956 for (i = LP_ORDER - 1; i < LP_ORDER_16k - 1; i++)

957 isf[i] = isf[i - 1] + isf[i - 1 - i_max_corr]

958 - isf[i - 2 - i_max_corr];

959

960 /* Calculate an estimate for ISF(18) and scale ISF based on the error */

961 est = 7965 + (isf[2] - isf[3] - isf[4]) / 6.0;

962 scale = 0.5 * (FFMIN(est, 7600) - isf[LP_ORDER - 2]) /

963 (isf[LP_ORDER_16k - 2] - isf[LP_ORDER - 2]);

964

965 for (i = LP_ORDER - 1, j = 0; i < LP_ORDER_16k - 1; i++, j++)

966 diff_isf[j] = scale * (isf[i] - isf[i - 1]);

967

968 /* Stability insurance */

969 for (i = 1; i < LP_ORDER_16k - LP_ORDER; i++)

970 if (diff_isf[i] + diff_isf[i - 1] < 5.0) {

971 if (diff_isf[i] > diff_isf[i - 1]) {

972 diff_isf[i - 1] = 5.0 - diff_isf[i];

973 } else

974 diff_isf[i] = 5.0 - diff_isf[i - 1];

975 }

976

977 for (i = LP_ORDER - 1, j = 0; i < LP_ORDER_16k - 1; i++, j++)

978 isf[i] = isf[i - 1] + diff_isf[j] * (1.0f / (1 << 15));

979

980 /* Scale the ISF vector for 16000 Hz */

981 for (i = 0; i < LP_ORDER_16k - 1; i++)

982 isf[i] *= 0.8;

983 }

984

985 /**

986 * Spectral expand the LP coefficients using the equation:

987 * y[i] = x[i] * (gamma ** i)

988 *

989 * @param[out] out Output buffer (may use input array)

990 * @param[in] lpc LP coefficients array

991 * @param[in] gamma Weighting factor

992 * @param[in] size LP array size

993 */

994 static void lpc_weighting(float *out, const float *lpc, float gamma, int size)

995 {

996 int i;

997 float fac = gamma;

998

999 for (i = 0; i < size; i++) {

1000 out[i] = lpc[i] * fac;

1001 fac *= gamma;

1002 }

1003 }

1004

1005 /**

1006 * Conduct 20th order linear predictive coding synthesis for the high

1007 * frequency band excitation at 16kHz.

1008 *

1009 * @param[in] ctx The context

1010 * @param[in] subframe Current subframe index (0 to 3)

1011 * @param[in,out] samples Pointer to the output speech samples

1012 * @param[in] exc Generated white-noise scaled excitation

1013 * @param[in] isf Current frame isf vector

1014 * @param[in] isf_past Past frame final isf vector

1015 */

1016 static void hb_synthesis(AMRWBContext *ctx, int subframe, float *samples,

1017 const float *exc, const float *isf, const float *isf_past)

1018 {

1019 float hb_lpc[LP_ORDER_16k];

1020 enum Mode mode = ctx->fr_cur_mode;

1021

1022 if (mode == MODE_6k60) {

1023 float e_isf[LP_ORDER_16k]; // ISF vector for extrapolation

1024 double e_isp[LP_ORDER_16k];

1025

1026 ctx->acelpv_ctx.weighted_vector_sumf(e_isf, isf_past, isf, isfp_inter[subframe],

1027 1.0 - isfp_inter[subframe], LP_ORDER);

1028

1029 extrapolate_isf(e_isf);

1030

1031 e_isf[LP_ORDER_16k - 1] *= 2.0;

1032 ff_acelp_lsf2lspd(e_isp, e_isf, LP_ORDER_16k);

1033 ff_amrwb_lsp2lpc(e_isp, hb_lpc, LP_ORDER_16k);

1034

1035 lpc_weighting(hb_lpc, hb_lpc, 0.9, LP_ORDER_16k);

1036 } else {

1037 lpc_weighting(hb_lpc, ctx->lp_coef[subframe], 0.6, LP_ORDER);

1038 }

1039

1040 ctx->celpf_ctx.celp_lp_synthesis_filterf(samples, hb_lpc, exc, AMRWB_SFR_SIZE_16k,

1041 (mode == MODE_6k60) ? LP_ORDER_16k : LP_ORDER);

1042 }

1043

1044 /**

1045 * Apply a 15th order filter to high-band samples.

1046 * The filter characteristic depends on the given coefficients.

1047 *

1048 * @param[out] out Buffer for filtered output

1049 * @param[in] fir_coef Filter coefficients

1050 * @param[in,out] mem State from last filtering (updated)

1051 * @param[in] in Input speech data (high-band)

1052 *

1053 * @remark It is safe to pass the same array in in and out parameters

1054 */

1055

1056 #ifndef hb_fir_filter

1057 static void hb_fir_filter(float *out, const float fir_coef[HB_FIR_SIZE + 1],

1058 float mem[HB_FIR_SIZE], const float *in)

1059 {

1060 int i, j;

1061 float data[AMRWB_SFR_SIZE_16k + HB_FIR_SIZE]; // past and current samples

1062

1063 memcpy(data, mem, HB_FIR_SIZE * sizeof(float));

1064 memcpy(data + HB_FIR_SIZE, in, AMRWB_SFR_SIZE_16k * sizeof(float));

1065

1066 for (i = 0; i < AMRWB_SFR_SIZE_16k; i++) {

1067 out[i] = 0.0;

1068 for (j = 0; j <= HB_FIR_SIZE; j++)

1069 out[i] += data[i + j] * fir_coef[j];

1070 }

1071

1072 memcpy(mem, data + AMRWB_SFR_SIZE_16k, HB_FIR_SIZE * sizeof(float));

1073 }

1074 #endif /* hb_fir_filter */

1075

1076 /**

1077 * Update context state before the next subframe.

1078 */

1079 static void update_sub_state(AMRWBContext *ctx)

1080 {

1081 memmove(&ctx->excitation_buf[0], &ctx->excitation_buf[AMRWB_SFR_SIZE],

1082 (AMRWB_P_DELAY_MAX + LP_ORDER + 1) * sizeof(float));

1083

1084 memmove(&ctx->pitch_gain[1], &ctx->pitch_gain[0], 5 * sizeof(float));

1085 memmove(&ctx->fixed_gain[1], &ctx->fixed_gain[0], 1 * sizeof(float));

1086

1087 memmove(&ctx->samples_az[0], &ctx->samples_az[AMRWB_SFR_SIZE],

1088 LP_ORDER * sizeof(float));

1089 memmove(&ctx->samples_up[0], &ctx->samples_up[AMRWB_SFR_SIZE],

1090 UPS_MEM_SIZE * sizeof(float));

1091 memmove(&ctx->samples_hb[0], &ctx->samples_hb[AMRWB_SFR_SIZE_16k],

1092 LP_ORDER_16k * sizeof(float));

1093 }

1094

1095 static int amrwb_decode_frame(AVCodecContext *avctx, void *data,

1096 int *got_frame_ptr, AVPacket *avpkt)

1097 {

1098 AMRWBContext *ctx = avctx->priv_data;

1099 AMRWBFrame *cf = &ctx->frame;

1100 const uint8_t *buf = avpkt->data;

1101 int buf_size = avpkt->size;

1102 int expected_fr_size, header_size;

1103 float *buf_out;

1104 float spare_vector[AMRWB_SFR_SIZE]; // extra stack space to hold result from anti-sparseness processing

1105 float fixed_gain_factor; // fixed gain correction factor (gamma)

1106 float *synth_fixed_vector; // pointer to the fixed vector that synthesis should use

1107 float synth_fixed_gain; // the fixed gain that synthesis should use

1108 float voice_fac, stab_fac; // parameters used for gain smoothing

1109 float synth_exc[AMRWB_SFR_SIZE]; // post-processed excitation for synthesis

1110 float hb_exc[AMRWB_SFR_SIZE_16k]; // excitation for the high frequency band

1111 float hb_samples[AMRWB_SFR_SIZE_16k]; // filtered high-band samples from synthesis

1112 float hb_gain;

1113 int sub, i, ret;

1114

1115 /* get output buffer */

1116 ctx->avframe.nb_samples = 4 * AMRWB_SFR_SIZE_16k;

1117 if ((ret = ff_get_buffer(avctx, &ctx->avframe)) < 0) {

1118 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");

1119 return ret;

1120 }

1121 buf_out = (float *)ctx->avframe.data[0];

1122

1123 header_size = decode_mime_header(ctx, buf);

1124 if (ctx->fr_cur_mode > MODE_SID) {

1125 av_log(avctx, AV_LOG_ERROR,

1126 "Invalid mode %d\n", ctx->fr_cur_mode);

1127 return AVERROR_INVALIDDATA;

1128 }

1129 expected_fr_size = ((cf_sizes_wb[ctx->fr_cur_mode] + 7) >> 3) + 1;

1130

1131 if (buf_size < expected_fr_size) {

1132 av_log(avctx, AV_LOG_ERROR,

1133 "Frame too small (%d bytes). Truncated file?\n", buf_size);

1134 *got_frame_ptr = 0;

1135 return AVERROR_INVALIDDATA;

1136 }

1137

1138 if (!ctx->fr_quality || ctx->fr_cur_mode > MODE_SID)

1139 av_log(avctx, AV_LOG_ERROR, "Encountered a bad or corrupted frame\n");

1140

1141 if (ctx->fr_cur_mode == MODE_SID) { /* Comfort noise frame */

1142 av_log_missing_feature(avctx, "SID mode", 1);

1143 return AVERROR_PATCHWELCOME;

1144 }

1145

1146 ff_amr_bit_reorder((uint16_t *) &ctx->frame, sizeof(AMRWBFrame),

1147 buf + header_size, amr_bit_orderings_by_mode[ctx->fr_cur_mode]);

1148

1149 /* Decode the quantized ISF vector */

1150 if (ctx->fr_cur_mode == MODE_6k60) {

1151 decode_isf_indices_36b(cf->isp_id, ctx->isf_cur);

1152 } else {

1153 decode_isf_indices_46b(cf->isp_id, ctx->isf_cur);

1154 }

1155

1156 isf_add_mean_and_past(ctx->isf_cur, ctx->isf_q_past);

1157 ff_set_min_dist_lsf(ctx->isf_cur, MIN_ISF_SPACING, LP_ORDER - 1);

1158

1159 stab_fac = stability_factor(ctx->isf_cur, ctx->isf_past_final);

1160

1161 ctx->isf_cur[LP_ORDER - 1] *= 2.0;

1162 ff_acelp_lsf2lspd(ctx->isp[3], ctx->isf_cur, LP_ORDER);

1163

1164 /* Generate a ISP vector for each subframe */

1165 if (ctx->first_frame) {

1166 ctx->first_frame = 0;

1167 memcpy(ctx->isp_sub4_past, ctx->isp[3], LP_ORDER * sizeof(double));

1168 }

1169 interpolate_isp(ctx->isp, ctx->isp_sub4_past);

1170

1171 for (sub = 0; sub < 4; sub++)

1172 ff_amrwb_lsp2lpc(ctx->isp[sub], ctx->lp_coef[sub], LP_ORDER);

1173

1174 for (sub = 0; sub < 4; sub++) {

1175 const AMRWBSubFrame *cur_subframe = &cf->subframe[sub];

1176 float *sub_buf = buf_out + sub * AMRWB_SFR_SIZE_16k;

1177

1178 /* Decode adaptive codebook (pitch vector) */

1179 decode_pitch_vector(ctx, cur_subframe, sub);

1180 /* Decode innovative codebook (fixed vector) */

1181 decode_fixed_vector(ctx->fixed_vector, cur_subframe->pul_ih,

1182 cur_subframe->pul_il, ctx->fr_cur_mode);

1183

1184 pitch_sharpening(ctx, ctx->fixed_vector);

1185

1186 decode_gains(cur_subframe->vq_gain, ctx->fr_cur_mode,

1187 &fixed_gain_factor, &ctx->pitch_gain[0]);

1188

1189 ctx->fixed_gain[0] =

1190 ff_amr_set_fixed_gain(fixed_gain_factor,

1191 ctx->celpm_ctx.dot_productf(ctx->fixed_vector,

1192 ctx->fixed_vector,

1193 AMRWB_SFR_SIZE) /

1194 AMRWB_SFR_SIZE,

1195 ctx->prediction_error,

1196 ENERGY_MEAN, energy_pred_fac);

1197

1198 /* Calculate voice factor and store tilt for next subframe */

1199 voice_fac = voice_factor(ctx->pitch_vector, ctx->pitch_gain[0],

1200 ctx->fixed_vector, ctx->fixed_gain[0],

1201 &ctx->celpm_ctx);

1202 ctx->tilt_coef = voice_fac * 0.25 + 0.25;

1203

1204 /* Construct current excitation */

1205 for (i = 0; i < AMRWB_SFR_SIZE; i++) {

1206 ctx->excitation[i] *= ctx->pitch_gain[0];

1207 ctx->excitation[i] += ctx->fixed_gain[0] * ctx->fixed_vector[i];

1208 ctx->excitation[i] = truncf(ctx->excitation[i]);

1209 }

1210

1211 /* Post-processing of excitation elements */

1212 synth_fixed_gain = noise_enhancer(ctx->fixed_gain[0], &ctx->prev_tr_gain,

1213 voice_fac, stab_fac);

1214

1215 synth_fixed_vector = anti_sparseness(ctx, ctx->fixed_vector,

1216 spare_vector);

1217

1218 pitch_enhancer(synth_fixed_vector, voice_fac);

1219

1220 synthesis(ctx, ctx->lp_coef[sub], synth_exc, synth_fixed_gain,

1221 synth_fixed_vector, &ctx->samples_az[LP_ORDER]);

1222

1223 /* Synthesis speech post-processing */

1224 de_emphasis(&ctx->samples_up[UPS_MEM_SIZE],

1225 &ctx->samples_az[LP_ORDER], PREEMPH_FAC, ctx->demph_mem);

1226

1227 ctx->acelpf_ctx.acelp_apply_order_2_transfer_function(&ctx->samples_up[UPS_MEM_SIZE],

1228 &ctx->samples_up[UPS_MEM_SIZE], hpf_zeros, hpf_31_poles,

1229 hpf_31_gain, ctx->hpf_31_mem, AMRWB_SFR_SIZE);

1230

1231 upsample_5_4(sub_buf, &ctx->samples_up[UPS_FIR_SIZE],

1232 AMRWB_SFR_SIZE_16k, &ctx->celpm_ctx);

1233

1234 /* High frequency band (6.4 - 7.0 kHz) generation part */

1235 ctx->acelpf_ctx.acelp_apply_order_2_transfer_function(hb_samples,

1236 &ctx->samples_up[UPS_MEM_SIZE], hpf_zeros, hpf_400_poles,

1237 hpf_400_gain, ctx->hpf_400_mem, AMRWB_SFR_SIZE);

1238

1239 hb_gain = find_hb_gain(ctx, hb_samples,

1240 cur_subframe->hb_gain, cf->vad);

1241

1242 scaled_hb_excitation(ctx, hb_exc, synth_exc, hb_gain);

1243

1244 hb_synthesis(ctx, sub, &ctx->samples_hb[LP_ORDER_16k],

1245 hb_exc, ctx->isf_cur, ctx->isf_past_final);

1246

1247 /* High-band post-processing filters */

1248 hb_fir_filter(hb_samples, bpf_6_7_coef, ctx->bpf_6_7_mem,

1249 &ctx->samples_hb[LP_ORDER_16k]);

1250

1251 if (ctx->fr_cur_mode == MODE_23k85)

1252 hb_fir_filter(hb_samples, lpf_7_coef, ctx->lpf_7_mem,

1253 hb_samples);

1254

1255 /* Add the low and high frequency bands */

1256 for (i = 0; i < AMRWB_SFR_SIZE_16k; i++)

1257 sub_buf[i] = (sub_buf[i] + hb_samples[i]) * (1.0f / (1 << 15));

1258

1259 /* Update buffers and history */

1260 update_sub_state(ctx);

1261 }

1262

1263 /* update state for next frame */

1264 memcpy(ctx->isp_sub4_past, ctx->isp[3], LP_ORDER * sizeof(ctx->isp[3][0]));

1265 memcpy(ctx->isf_past_final, ctx->isf_cur, LP_ORDER * sizeof(float));

1266

1267 *got_frame_ptr = 1;

1268 *(AVFrame *)data = ctx->avframe;

1269

1270 return expected_fr_size;

1271 }

1272

1273 AVCodec ff_amrwb_decoder = {

1274 .name = "amrwb",

1275 .type = AVMEDIA_TYPE_AUDIO,

1276 .id = AV_CODEC_ID_AMR_WB,

1277 .priv_data_size = sizeof(AMRWBContext),

1278 .init = amrwb_decode_init,

1279 .decode = amrwb_decode_frame,

1280 .capabilities = CODEC_CAP_DR1,

1281 .long_name = NULL_IF_CONFIG_SMALL("AMR-WB (Adaptive Multi-Rate WideBand)"),

1282 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_FLT,

1283 AV_SAMPLE_FMT_NONE },

1284 };

Generated on Sat May 25 2013 03:58:31 for FFmpeg by doxygen 1.8.2