aarch64: Add assembly support for -fsanitize=hwaddress tagged globals.
[libav.git] / libavcodec / qcelpdec.c
CommitLineData
cb377ec5
KG
1/*
2 * QCELP decoder
3 * Copyright (c) 2007 Reynaldo H. Verdejo Pinochet
4 *
2912e87a 5 * This file is part of Libav.
cb377ec5 6 *
2912e87a 7 * Libav is free software; you can redistribute it and/or
cb377ec5
KG
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 *
2912e87a 12 * Libav is distributed in the hope that it will be useful,
cb377ec5
KG
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
2912e87a 18 * License along with Libav; if not, write to the Free Software
cb377ec5
KG
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20 */
1c3ae1ab 21
cb377ec5 22/**
ba87f080 23 * @file
cb377ec5
KG
24 * QCELP decoder
25 * @author Reynaldo H. Verdejo Pinochet
2912e87a 26 * @remark Libav merging spearheaded by Kenan Gillet
9e00c20e 27 * @remark Development mentored by Benjamin Larson
cb377ec5
KG
28 */
29
30#include <stddef.h>
31
a903f8f0 32#include "libavutil/channel_layout.h"
d56668bd 33#include "libavutil/float_dsp.h"
0dabd329 34
cb377ec5 35#include "avcodec.h"
0dabd329 36#include "bitstream.h"
dbbec0c2 37#include "internal.h"
cb377ec5 38#include "qcelpdata.h"
cb377ec5 39#include "celp_filters.h"
805a83d5 40#include "acelp_filters.h"
95e83257 41#include "acelp_vectors.h"
33ae681f 42#include "lsp.h"
cb377ec5
KG
43
44#undef NDEBUG
45#include <assert.h>
46
25c2d76b 47typedef enum {
adbfc605 48 I_F_Q = -1, /**< insufficient frame quality */
061f407e
KG
49 SILENCE,
50 RATE_OCTAVE,
51 RATE_QUARTER,
52 RATE_HALF,
53 RATE_FULL
54} qcelp_packet_rate;
55
7f9f771e 56typedef struct QCELPContext {
0dabd329 57 BitstreamContext bc;
640760da 58 qcelp_packet_rate bitrate;
adbfc605 59 QCELPFrame frame; /**< unpacked data frame */
cf139541
RVP
60
61 uint8_t erasure_count;
adbfc605 62 uint8_t octave_count; /**< count the consecutive RATE_OCTAVE frames */
cf139541 63 float prev_lspf[10];
adbfc605 64 float predictor_lspf[10];/**< LSP predictor for RATE_OCTAVE and I_F_Q */
cf139541
RVP
65 float pitch_synthesis_filter_mem[303];
66 float pitch_pre_filter_mem[303];
67 float rnd_fir_filter_mem[180];
68 float formant_mem[170];
69 float last_codebook_gain;
70 int prev_g1[2];
71 int prev_bitrate;
72 float pitch_gain[4];
73 uint8_t pitch_lag[4];
74 uint16_t first16bits;
5a3e9f2c 75 uint8_t warned_buf_mismatch_bitrate;
805a83d5
RB
76
77 /* postfilter */
78 float postfilter_synth_mem[10];
79 float postfilter_agc_mem;
80 float postfilter_tilt_mem;
640760da
KG
81} QCELPContext;
82
061f407e 83/**
39ded680
KG
84 * Initialize the speech codec according to the specification.
85 *
86 * TIA/EIA/IS-733 2.4.9
87 */
1b321c5c
RVP
88static av_cold int qcelp_decode_init(AVCodecContext *avctx)
89{
39ded680
KG
90 QCELPContext *q = avctx->priv_data;
91 int i;
92
e3d6ab57
JR
93 avctx->channels = 1;
94 avctx->channel_layout = AV_CH_LAYOUT_MONO;
95 avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
39ded680 96
25c2d76b 97 for (i = 0; i < 10; i++)
4a2ef394 98 q->prev_lspf[i] = (i + 1) / 11.0;
39ded680
KG
99
100 return 0;
101}
102
103/**
49bd8e4b
MR
104 * Decode the 10 quantized LSP frequencies from the LSPV/LSP
105 * transmission codes of any bitrate and check for badly received packets.
148c31b9
KG
106 *
107 * @param q the context
108 * @param lspf line spectral pair frequencies
109 *
110 * @return 0 on success, -1 if the packet is badly received
111 *
112 * TIA/EIA/IS-733 2.4.3.2.6.2-2, 2.4.8.7.3
113 */
1b321c5c
RVP
114static int decode_lspf(QCELPContext *q, float *lspf)
115{
148c31b9 116 int i;
cf139541
RVP
117 float tmp_lspf, smooth, erasure_coeff;
118 const float *predictors;
148c31b9 119
dd376b1a 120 if (q->bitrate == RATE_OCTAVE || q->bitrate == I_F_Q) {
2fd29184
DB
121 predictors = q->prev_bitrate != RATE_OCTAVE &&
122 q->prev_bitrate != I_F_Q ? q->prev_lspf
123 : q->predictor_lspf;
148c31b9 124
dd376b1a 125 if (q->bitrate == RATE_OCTAVE) {
148c31b9
KG
126 q->octave_count++;
127
25c2d76b 128 for (i = 0; i < 10; i++) {
148c31b9 129 q->predictor_lspf[i] =
640760da 130 lspf[i] = (q->frame.lspv[i] ? QCELP_LSP_SPREAD_FACTOR
25c2d76b
AL
131 : -QCELP_LSP_SPREAD_FACTOR) +
132 predictors[i] * QCELP_LSP_OCTAVE_PREDICTOR +
133 (i + 1) * ((1 - QCELP_LSP_OCTAVE_PREDICTOR) / 11);
148c31b9 134 }
2fd29184 135 smooth = q->octave_count < 10 ? .875 : 0.1;
dd376b1a 136 } else {
cf139541 137 erasure_coeff = QCELP_LSP_OCTAVE_PREDICTOR;
148c31b9
KG
138
139 assert(q->bitrate == I_F_Q);
140
25c2d76b 141 if (q->erasure_count > 1)
2fd29184 142 erasure_coeff *= q->erasure_count < 4 ? 0.9 : 0.7;
148c31b9 143
25c2d76b 144 for (i = 0; i < 10; i++) {
148c31b9 145 q->predictor_lspf[i] =
25c2d76b
AL
146 lspf[i] = (i + 1) * (1 - erasure_coeff) / 11 +
147 erasure_coeff * predictors[i];
148c31b9
KG
148 }
149 smooth = 0.125;
150 }
151
152 // Check the stability of the LSP frequencies.
153 lspf[0] = FFMAX(lspf[0], QCELP_LSP_SPREAD_FACTOR);
25c2d76b 154 for (i = 1; i < 10; i++)
2fd29184 155 lspf[i] = FFMAX(lspf[i], lspf[i - 1] + QCELP_LSP_SPREAD_FACTOR);
148c31b9 156
2fd29184 157 lspf[9] = FFMIN(lspf[9], 1.0 - QCELP_LSP_SPREAD_FACTOR);
25c2d76b 158 for (i = 9; i > 0; i--)
2fd29184 159 lspf[i - 1] = FFMIN(lspf[i - 1], lspf[i] - QCELP_LSP_SPREAD_FACTOR);
148c31b9
KG
160
161 // Low-pass filter the LSP frequencies.
25c2d76b 162 ff_weighted_vector_sumf(lspf, lspf, q->prev_lspf, smooth, 1.0 - smooth, 10);
dd376b1a 163 } else {
148c31b9
KG
164 q->octave_count = 0;
165
4a2ef394 166 tmp_lspf = 0.0;
dd376b1a 167 for (i = 0; i < 5; i++) {
25c2d76b
AL
168 lspf[2 * i + 0] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][0] * 0.0001;
169 lspf[2 * i + 1] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][1] * 0.0001;
148c31b9
KG
170 }
171
172 // Check for badly received packets.
dd376b1a 173 if (q->bitrate == RATE_QUARTER) {
25c2d76b 174 if (lspf[9] <= .70 || lspf[9] >= .97)
148c31b9 175 return -1;
25c2d76b
AL
176 for (i = 3; i < 10; i++)
177 if (fabs(lspf[i] - lspf[i - 2]) < .08)
148c31b9 178 return -1;
dd376b1a 179 } else {
25c2d76b 180 if (lspf[9] <= .66 || lspf[9] >= .985)
148c31b9 181 return -1;
25c2d76b
AL
182 for (i = 4; i < 10; i++)
183 if (fabs(lspf[i] - lspf[i - 4]) < .0931)
148c31b9
KG
184 return -1;
185 }
186 }
187 return 0;
188}
189
190/**
49bd8e4b 191 * Convert codebook transmission codes to GAIN and INDEX.
640760da
KG
192 *
193 * @param q the context
194 * @param gain array holding the decoded gain
195 *
196 * TIA/EIA/IS-733 2.4.6.2
197 */
25c2d76b
AL
198static void decode_gain_and_index(QCELPContext *q, float *gain)
199{
200 int i, subframes_count, g1[16];
640760da
KG
201 float slope;
202
dd376b1a
JR
203 if (q->bitrate >= RATE_QUARTER) {
204 switch (q->bitrate) {
25c2d76b
AL
205 case RATE_FULL: subframes_count = 16; break;
206 case RATE_HALF: subframes_count = 4; break;
207 default: subframes_count = 5;
640760da 208 }
25c2d76b 209 for (i = 0; i < subframes_count; i++) {
640760da 210 g1[i] = 4 * q->frame.cbgain[i];
25c2d76b
AL
211 if (q->bitrate == RATE_FULL && !((i + 1) & 3)) {
212 g1[i] += av_clip((g1[i - 1] + g1[i - 2] + g1[i - 3]) / 3 - 6, 0, 32);
640760da
KG
213 }
214
215 gain[i] = qcelp_g12ga[g1[i]];
216
dd376b1a 217 if (q->frame.cbsign[i]) {
640760da 218 gain[i] = -gain[i];
25c2d76b 219 q->frame.cindex[i] = (q->frame.cindex[i] - 89) & 127;
640760da
KG
220 }
221 }
222
25c2d76b
AL
223 q->prev_g1[0] = g1[i - 2];
224 q->prev_g1[1] = g1[i - 1];
225 q->last_codebook_gain = qcelp_g12ga[g1[i - 1]];
640760da 226
dd376b1a 227 if (q->bitrate == RATE_QUARTER) {
640760da 228 // Provide smoothing of the unvoiced excitation energy.
25c2d76b
AL
229 gain[7] = gain[4];
230 gain[6] = 0.4 * gain[3] + 0.6 * gain[4];
231 gain[5] = gain[3];
232 gain[4] = 0.8 * gain[2] + 0.2 * gain[3];
233 gain[3] = 0.2 * gain[1] + 0.8 * gain[2];
234 gain[2] = gain[1];
235 gain[1] = 0.6 * gain[0] + 0.4 * gain[1];
640760da 236 }
dd376b1a
JR
237 } else if (q->bitrate != SILENCE) {
238 if (q->bitrate == RATE_OCTAVE) {
25c2d76b
AL
239 g1[0] = 2 * q->frame.cbgain[0] +
240 av_clip((q->prev_g1[0] + q->prev_g1[1]) / 2 - 5, 0, 54);
640760da 241 subframes_count = 8;
dd376b1a 242 } else {
640760da
KG
243 assert(q->bitrate == I_F_Q);
244
245 g1[0] = q->prev_g1[1];
dd376b1a 246 switch (q->erasure_count) {
25c2d76b
AL
247 case 1 : break;
248 case 2 : g1[0] -= 1; break;
249 case 3 : g1[0] -= 2; break;
250 default: g1[0] -= 6;
640760da 251 }
25c2d76b 252 if (g1[0] < 0)
640760da
KG
253 g1[0] = 0;
254 subframes_count = 4;
255 }
256 // This interpolation is done to produce smoother background noise.
25c2d76b
AL
257 slope = 0.5 * (qcelp_g12ga[g1[0]] - q->last_codebook_gain) / subframes_count;
258 for (i = 1; i <= subframes_count; i++)
259 gain[i - 1] = q->last_codebook_gain + slope * i;
640760da 260
25c2d76b
AL
261 q->last_codebook_gain = gain[i - 2];
262 q->prev_g1[0] = q->prev_g1[1];
263 q->prev_g1[1] = g1[0];
640760da
KG
264 }
265}
266
267/**
1b321c5c
RVP
268 * If the received packet is Rate 1/4 a further sanity check is made of the
269 * codebook gain.
148c31b9
KG
270 *
271 * @param cbgain the unpacked cbgain array
272 * @return -1 if the sanity check fails, 0 otherwise
273 *
274 * TIA/EIA/IS-733 2.4.8.7.3
275 */
1b321c5c
RVP
276static int codebook_sanity_check_for_rate_quarter(const uint8_t *cbgain)
277{
25c2d76b 278 int i, diff, prev_diff = 0;
148c31b9 279
25c2d76b 280 for (i = 1; i < 5; i++) {
cf139541 281 diff = cbgain[i] - cbgain[i-1];
25c2d76b 282 if (FFABS(diff) > 10)
1b321c5c 283 return -1;
25c2d76b 284 else if (FFABS(diff - prev_diff) > 12)
1b321c5c
RVP
285 return -1;
286 prev_diff = diff;
8372e3d2
RVP
287 }
288 return 0;
148c31b9
KG
289}
290
291/**
49bd8e4b 292 * Compute the scaled codebook vector Cdn From INDEX and GAIN
39ded680
KG
293 * for all rates.
294 *
295 * The specification lacks some information here.
296 *
297 * TIA/EIA/IS-733 has an omission on the codebook index determination
298 * formula for RATE_FULL and RATE_HALF frames at section 2.4.8.1.1. It says
299 * you have to subtract the decoded index parameter from the given scaled
300 * codebook vector index 'n' to get the desired circular codebook index, but
301 * it does not mention that you have to clamp 'n' to [0-9] in order to get
302 * RI-compliant results.
303 *
304 * The reason for this mistake seems to be the fact they forgot to mention you
305 * have to do these calculations per codebook subframe and adjust given
306 * equation values accordingly.
307 *
308 * @param q the context
309 * @param gain array holding the 4 pitch subframe gain values
310 * @param cdn_vector array for the generated scaled codebook vector
311 */
b12c7627 312static void compute_svector(QCELPContext *q, const float *gain,
1b321c5c
RVP
313 float *cdn_vector)
314{
25c2d76b 315 int i, j, k;
39ded680 316 uint16_t cbseed, cindex;
25c2d76b 317 float *rnd, tmp_gain, fir_filter_value;
39ded680 318
dd376b1a 319 switch (q->bitrate) {
25c2d76b
AL
320 case RATE_FULL:
321 for (i = 0; i < 16; i++) {
322 tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO;
323 cindex = -q->frame.cindex[i];
324 for (j = 0; j < 10; j++)
325 *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cindex++ & 127];
326 }
39ded680 327 break;
25c2d76b
AL
328 case RATE_HALF:
329 for (i = 0; i < 4; i++) {
330 tmp_gain = gain[i] * QCELP_RATE_HALF_CODEBOOK_RATIO;
331 cindex = -q->frame.cindex[i];
332 for (j = 0; j < 40; j++)
39ded680 333 *cdn_vector++ = tmp_gain * qcelp_rate_half_codebook[cindex++ & 127];
25c2d76b 334 }
39ded680 335 break;
25c2d76b
AL
336 case RATE_QUARTER:
337 cbseed = (0x0003 & q->frame.lspv[4]) << 14 |
338 (0x003F & q->frame.lspv[3]) << 8 |
339 (0x0060 & q->frame.lspv[2]) << 1 |
340 (0x0007 & q->frame.lspv[1]) << 3 |
341 (0x0038 & q->frame.lspv[0]) >> 3;
342 rnd = q->rnd_fir_filter_mem + 20;
343 for (i = 0; i < 8; i++) {
344 tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0);
345 for (k = 0; k < 20; k++) {
346 cbseed = 521 * cbseed + 259;
347 *rnd = (int16_t) cbseed;
1b321c5c
RVP
348
349 // FIR filter
25c2d76b
AL
350 fir_filter_value = 0.0;
351 for (j = 0; j < 10; j++)
352 fir_filter_value += qcelp_rnd_fir_coefs[j] *
353 (rnd[-j] + rnd[-20+j]);
354
355 fir_filter_value += qcelp_rnd_fir_coefs[10] * rnd[-10];
356 *cdn_vector++ = tmp_gain * fir_filter_value;
357 rnd++;
39ded680 358 }
25c2d76b
AL
359 }
360 memcpy(q->rnd_fir_filter_mem, q->rnd_fir_filter_mem + 160,
361 20 * sizeof(float));
39ded680 362 break;
25c2d76b
AL
363 case RATE_OCTAVE:
364 cbseed = q->first16bits;
365 for (i = 0; i < 8; i++) {
366 tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0);
367 for (j = 0; j < 20; j++) {
368 cbseed = 521 * cbseed + 259;
369 *cdn_vector++ = tmp_gain * (int16_t) cbseed;
39ded680 370 }
25c2d76b 371 }
39ded680 372 break;
25c2d76b
AL
373 case I_F_Q:
374 cbseed = -44; // random codebook index
375 for (i = 0; i < 4; i++) {
376 tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO;
377 for (j = 0; j < 40; j++)
378 *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cbseed++ & 127];
379 }
39ded680 380 break;
25c2d76b
AL
381 case SILENCE:
382 memset(cdn_vector, 0, 160 * sizeof(float));
4f54fb54 383 break;
39ded680
KG
384 }
385}
386
387/**
809e07c6
KG
388 * Apply generic gain control.
389 *
390 * @param v_out output vector
391 * @param v_in gain-controlled vector
392 * @param v_ref vector to control gain of
393 *
394 * TIA/EIA/IS-733 2.4.8.3, 2.4.8.6
395 */
25c2d76b 396static void apply_gain_ctrl(float *v_out, const float *v_ref, const float *v_in)
1b321c5c 397{
0c50f8e6 398 int i;
39ded680 399
d56668bd
RB
400 for (i = 0; i < 160; i += 40) {
401 float res = avpriv_scalarproduct_float_c(v_ref + i, v_ref + i, 40);
402 ff_scale_vector_to_given_sum_of_squares(v_out + i, v_in + i, res, 40);
403 }
39ded680
KG
404}
405
406/**
cb377ec5
KG
407 * Apply filter in pitch-subframe steps.
408 *
409 * @param memory buffer for the previous state of the filter
410 * - must be able to contain 303 elements
411 * - the 143 first elements are from the previous state
412 * - the next 160 are for output
413 * @param v_in input filter vector
414 * @param gain per-subframe gain array, each element is between 0.0 and 2.0
415 * @param lag per-subframe lag array, each element is
416 * - between 16 and 143 if its corresponding pfrac is 0,
417 * - between 16 and 139 otherwise
1b321c5c
RVP
418 * @param pfrac per-subframe boolean array, 1 if the lag is fractional, 0
419 * otherwise
cb377ec5
KG
420 *
421 * @return filter output vector
422 */
1c3ae1ab
RVP
423static const float *do_pitchfilter(float memory[303], const float v_in[160],
424 const float gain[4], const uint8_t *lag,
425 const uint8_t pfrac[4])
426{
25c2d76b
AL
427 int i, j;
428 float *v_lag, *v_out;
cb377ec5
KG
429 const float *v_len;
430
431 v_out = memory + 143; // Output vector starts at memory[143].
432
dd376b1a
JR
433 for (i = 0; i < 4; i++) {
434 if (gain[i]) {
cb377ec5 435 v_lag = memory + 143 + 40 * i - lag[i];
dd376b1a
JR
436 for (v_len = v_in + 40; v_in < v_len; v_in++) {
437 if (pfrac[i]) { // If it is a fractional lag...
4a2ef394 438 for (j = 0, *v_out = 0.0; j < 4; j++)
25c2d76b
AL
439 *v_out += qcelp_hammsinc_table[j] * (v_lag[j - 4] + v_lag[3 - j]);
440 } else
cb377ec5
KG
441 *v_out = *v_lag;
442
443 *v_out = *v_in + gain[i] * *v_out;
444
445 v_lag++;
446 v_out++;
447 }
dd376b1a 448 } else {
cb377ec5
KG
449 memcpy(v_out, v_in, 40 * sizeof(float));
450 v_in += 40;
451 v_out += 40;
452 }
1c3ae1ab 453 }
cb377ec5
KG
454
455 memmove(memory, memory + 160, 143 * sizeof(float));
456 return memory + 143;
457}
458
200de8c6 459/**
73b458e3 460 * Apply pitch synthesis filter and pitch prefilter to the scaled codebook vector.
4f54fb54 461 * TIA/EIA/IS-733 2.4.5.2, 2.4.8.7.2
73b458e3
KG
462 *
463 * @param q the context
464 * @param cdn_vector the scaled codebook vector
465 */
cf139541
RVP
466static void apply_pitch_filters(QCELPContext *q, float *cdn_vector)
467{
25c2d76b 468 int i;
73b458e3
KG
469 const float *v_synthesis_filtered, *v_pre_filtered;
470
25c2d76b
AL
471 if (q->bitrate >= RATE_HALF || q->bitrate == SILENCE ||
472 (q->bitrate == I_F_Q && (q->prev_bitrate >= RATE_HALF))) {
73b458e3 473
25c2d76b 474 if (q->bitrate >= RATE_HALF) {
73b458e3 475 // Compute gain & lag for the whole frame.
dd376b1a 476 for (i = 0; i < 4; i++) {
73b458e3
KG
477 q->pitch_gain[i] = q->frame.plag[i] ? (q->frame.pgain[i] + 1) * 0.25 : 0.0;
478
479 q->pitch_lag[i] = q->frame.plag[i] + 16;
480 }
dd376b1a 481 } else {
3f16ed15
RVP
482 float max_pitch_gain;
483
dd376b1a 484 if (q->bitrate == I_F_Q) {
a3636fa0
RVP
485 if (q->erasure_count < 3)
486 max_pitch_gain = 0.9 - 0.3 * (q->erasure_count - 1);
487 else
488 max_pitch_gain = 0.0;
dd376b1a 489 } else {
4f54fb54
RVP
490 assert(q->bitrate == SILENCE);
491 max_pitch_gain = 1.0;
492 }
25c2d76b 493 for (i = 0; i < 4; i++)
73b458e3
KG
494 q->pitch_gain[i] = FFMIN(q->pitch_gain[i], max_pitch_gain);
495
496 memset(q->frame.pfrac, 0, sizeof(q->frame.pfrac));
497 }
498
499 // pitch synthesis filter
cf139541
RVP
500 v_synthesis_filtered = do_pitchfilter(q->pitch_synthesis_filter_mem,
501 cdn_vector, q->pitch_gain,
502 q->pitch_lag, q->frame.pfrac);
73b458e3
KG
503
504 // pitch prefilter update
25c2d76b 505 for (i = 0; i < 4; i++)
73b458e3
KG
506 q->pitch_gain[i] = 0.5 * FFMIN(q->pitch_gain[i], 1.0);
507
25c2d76b
AL
508 v_pre_filtered = do_pitchfilter(q->pitch_pre_filter_mem,
509 v_synthesis_filtered,
510 q->pitch_gain, q->pitch_lag,
511 q->frame.pfrac);
73b458e3
KG
512
513 apply_gain_ctrl(cdn_vector, v_synthesis_filtered, v_pre_filtered);
dd376b1a 514 } else {
25c2d76b 515 memcpy(q->pitch_synthesis_filter_mem, cdn_vector + 17, 143 * sizeof(float));
cf139541 516 memcpy(q->pitch_pre_filter_mem, cdn_vector + 17, 143 * sizeof(float));
73b458e3
KG
517 memset(q->pitch_gain, 0, sizeof(q->pitch_gain));
518 memset(q->pitch_lag, 0, sizeof(q->pitch_lag));
519 }
520}
521
522/**
49bd8e4b
MR
523 * Reconstruct LPC coefficients from the line spectral pair frequencies
524 * and perform bandwidth expansion.
87574416
RVP
525 *
526 * @param lspf line spectral pair frequencies
527 * @param lpc linear predictive coding coefficients
528 *
e19f9952 529 * @note: bandwidth_expansion_coeff could be precalculated into a table
87574416
RVP
530 * but it seems to be slower on x86
531 *
532 * TIA/EIA/IS-733 2.4.3.3.5
533 */
3c00556f 534static void lspf2lpc(const float *lspf, float *lpc)
87574416 535{
1e1e02ea 536 double lsp[10];
e19f9952 537 double bandwidth_expansion_coeff = QCELP_BANDWIDTH_EXPANSION_COEFF;
25c2d76b 538 int i;
87574416 539
25c2d76b 540 for (i = 0; i < 10; i++)
1e1e02ea 541 lsp[i] = cos(M_PI * lspf[i]);
87574416 542
e26d1318 543 ff_acelp_lspd2lpc(lsp, lpc, 5);
87574416 544
dd376b1a 545 for (i = 0; i < 10; i++) {
25c2d76b 546 lpc[i] *= bandwidth_expansion_coeff;
e19f9952 547 bandwidth_expansion_coeff *= QCELP_BANDWIDTH_EXPANSION_COEFF;
87574416
RVP
548 }
549}
550
551/**
48966b02 552 * Interpolate LSP frequencies and compute LPC coefficients
148c31b9 553 * for a given bitrate & pitch subframe.
200de8c6 554 *
4f54fb54 555 * TIA/EIA/IS-733 2.4.3.3.4, 2.4.8.7.2
200de8c6
KG
556 *
557 * @param q the context
558 * @param curr_lspf LSP frequencies vector of the current frame
559 * @param lpc float vector for the resulting LPC
560 * @param subframe_num frame number in decoded stream
561 */
0e426b7a
VS
562static void interpolate_lpc(QCELPContext *q, const float *curr_lspf,
563 float *lpc, const int subframe_num)
1c3ae1ab 564{
200de8c6
KG
565 float interpolated_lspf[10];
566 float weight;
567
25c2d76b 568 if (q->bitrate >= RATE_QUARTER)
200de8c6 569 weight = 0.25 * (subframe_num + 1);
25c2d76b 570 else if (q->bitrate == RATE_OCTAVE && !subframe_num)
200de8c6 571 weight = 0.625;
1c3ae1ab 572 else
200de8c6 573 weight = 1.0;
200de8c6 574
dd376b1a 575 if (weight != 1.0) {
95e83257 576 ff_weighted_vector_sumf(interpolated_lspf, curr_lspf, q->prev_lspf,
d2af5697 577 weight, 1.0 - weight, 10);
87574416 578 lspf2lpc(interpolated_lspf, lpc);
25c2d76b
AL
579 } else if (q->bitrate >= RATE_QUARTER ||
580 (q->bitrate == I_F_Q && !subframe_num))
87574416 581 lspf2lpc(curr_lspf, lpc);
25c2d76b 582 else if (q->bitrate == SILENCE && !subframe_num)
87574416 583 lspf2lpc(q->prev_lspf, lpc);
200de8c6
KG
584}
585
6e74619e 586static qcelp_packet_rate buf_size2bitrate(const int buf_size)
1c3ae1ab 587{
dd376b1a 588 switch (buf_size) {
25c2d76b
AL
589 case 35: return RATE_FULL;
590 case 17: return RATE_HALF;
591 case 8: return RATE_QUARTER;
592 case 4: return RATE_OCTAVE;
593 case 1: return SILENCE;
2ae1a9b2 594 }
1c3ae1ab 595
6e74619e 596 return I_F_Q;
2ae1a9b2
KG
597}
598
061f407e
KG
599/**
600 * Determine the bitrate from the frame size and/or the first byte of the frame.
601 *
602 * @param avctx the AV codec context
603 * @param buf_size length of the buffer
41ed7ab4 604 * @param buf the buffer
061f407e
KG
605 *
606 * @return the bitrate on success,
607 * I_F_Q if the bitrate cannot be satisfactorily determined
608 *
609 * TIA/EIA/IS-733 2.4.8.7.1
610 */
25c2d76b
AL
611static qcelp_packet_rate determine_bitrate(AVCodecContext *avctx,
612 const int buf_size,
613 const uint8_t **buf)
cf139541 614{
061f407e
KG
615 qcelp_packet_rate bitrate;
616
dd376b1a
JR
617 if ((bitrate = buf_size2bitrate(buf_size)) >= 0) {
618 if (bitrate > **buf) {
5a3e9f2c 619 QCELPContext *q = avctx->priv_data;
dd376b1a 620 if (!q->warned_buf_mismatch_bitrate) {
cf139541
RVP
621 av_log(avctx, AV_LOG_WARNING,
622 "Claimed bitrate and buffer size mismatch.\n");
5a3e9f2c
KG
623 q->warned_buf_mismatch_bitrate = 1;
624 }
061f407e 625 bitrate = **buf;
dd376b1a 626 } else if (bitrate < **buf) {
cf139541
RVP
627 av_log(avctx, AV_LOG_ERROR,
628 "Buffer is too small for the claimed bitrate.\n");
061f407e
KG
629 return I_F_Q;
630 }
631 (*buf)++;
dd376b1a 632 } else if ((bitrate = buf_size2bitrate(buf_size + 1)) >= 0) {
061f407e
KG
633 av_log(avctx, AV_LOG_WARNING,
634 "Bitrate byte is missing, guessing the bitrate from packet size.\n");
25c2d76b 635 } else
061f407e
KG
636 return I_F_Q;
637
dd376b1a 638 if (bitrate == SILENCE) {
6d97484d
DB
639 // FIXME: Remove this warning when tested with samples.
640 avpriv_request_sample(avctx, "Blank frame handling");
061f407e
KG
641 }
642 return bitrate;
643}
644
cb377ec5 645static void warn_insufficient_frame_quality(AVCodecContext *avctx,
1c3ae1ab
RVP
646 const char *message)
647{
25c2d76b
AL
648 av_log(avctx, AV_LOG_WARNING, "Frame #%d, IFQ: %s\n",
649 avctx->frame_number, message);
cb377ec5 650}
200de8c6 651
805a83d5
RB
652static void postfilter(QCELPContext *q, float *samples, float *lpc)
653{
654 static const float pow_0_775[10] = {
655 0.775000, 0.600625, 0.465484, 0.360750, 0.279582,
656 0.216676, 0.167924, 0.130141, 0.100859, 0.078166
657 }, pow_0_625[10] = {
658 0.625000, 0.390625, 0.244141, 0.152588, 0.095367,
659 0.059605, 0.037253, 0.023283, 0.014552, 0.009095
660 };
661 float lpc_s[10], lpc_p[10], pole_out[170], zero_out[160];
662 int n;
663
664 for (n = 0; n < 10; n++) {
665 lpc_s[n] = lpc[n] * pow_0_625[n];
666 lpc_p[n] = lpc[n] * pow_0_775[n];
667 }
668
669 ff_celp_lp_zero_synthesis_filterf(zero_out, lpc_s,
670 q->formant_mem + 10, 160, 10);
25c2d76b 671 memcpy(pole_out, q->postfilter_synth_mem, sizeof(float) * 10);
805a83d5
RB
672 ff_celp_lp_synthesis_filterf(pole_out + 10, lpc_p, zero_out, 160, 10);
673 memcpy(q->postfilter_synth_mem, pole_out + 160, sizeof(float) * 10);
674
675 ff_tilt_compensation(&q->postfilter_tilt_mem, 0.3, pole_out + 10, 160);
676
677 ff_adaptive_gain_control(samples, pole_out + 10,
d56668bd
RB
678 avpriv_scalarproduct_float_c(q->formant_mem + 10,
679 q->formant_mem + 10,
680 160),
25c2d76b 681 160, 0.9375, &q->postfilter_agc_mem);
805a83d5
RB
682}
683
0eea2129
JR
684static int qcelp_decode_frame(AVCodecContext *avctx, void *data,
685 int *got_frame_ptr, AVPacket *avpkt)
cf139541 686{
7a00bbad 687 const uint8_t *buf = avpkt->data;
25c2d76b
AL
688 int buf_size = avpkt->size;
689 QCELPContext *q = avctx->priv_data;
1b9b6d6e 690 AVFrame *frame = data;
0eea2129
JR
691 float *outbuffer;
692 int i, ret;
cf139541
RVP
693 float quantized_lspf[10], lpc[10];
694 float gain[16];
695 float *formant_mem;
696
0eea2129 697 /* get output buffer */
1b9b6d6e 698 frame->nb_samples = 160;
759001c5 699 if ((ret = ff_get_buffer(avctx, frame, 0)) < 0) {
0eea2129
JR
700 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
701 return ret;
e43dd3d2 702 }
1b9b6d6e 703 outbuffer = (float *)frame->data[0];
e43dd3d2 704
dd376b1a 705 if ((q->bitrate = determine_bitrate(avctx, buf_size, &buf)) == I_F_Q) {
640760da
KG
706 warn_insufficient_frame_quality(avctx, "bitrate cannot be determined.");
707 goto erasure;
708 }
709
25c2d76b
AL
710 if (q->bitrate == RATE_OCTAVE &&
711 (q->first16bits = AV_RB16(buf)) == 0xFFFF) {
640760da
KG
712 warn_insufficient_frame_quality(avctx, "Bitrate is 1/8 and first 16 bits are on.");
713 goto erasure;
714 }
715
dd376b1a 716 if (q->bitrate > SILENCE) {
640760da 717 const QCELPBitmap *bitmaps = qcelp_unpacking_bitmaps_per_rate[q->bitrate];
25c2d76b
AL
718 const QCELPBitmap *bitmaps_end = qcelp_unpacking_bitmaps_per_rate[q->bitrate] +
719 qcelp_unpacking_bitmaps_lengths[q->bitrate];
720 uint8_t *unpacked_data = (uint8_t *)&q->frame;
640760da 721
a25dac97 722 bitstream_init8(&q->bc, buf, buf_size);
640760da
KG
723
724 memset(&q->frame, 0, sizeof(QCELPFrame));
725
25c2d76b 726 for (; bitmaps < bitmaps_end; bitmaps++)
0dabd329 727 unpacked_data[bitmaps->index] |= bitstream_read(&q->bc, bitmaps->bitlen) << bitmaps->bitpos;
640760da
KG
728
729 // Check for erasures/blanks on rates 1, 1/4 and 1/8.
dd376b1a 730 if (q->frame.reserved) {
640760da
KG
731 warn_insufficient_frame_quality(avctx, "Wrong data in reserved frame area.");
732 goto erasure;
733 }
25c2d76b
AL
734 if (q->bitrate == RATE_QUARTER &&
735 codebook_sanity_check_for_rate_quarter(q->frame.cbgain)) {
640760da
KG
736 warn_insufficient_frame_quality(avctx, "Codebook gain sanity check failed.");
737 goto erasure;
738 }
739
dd376b1a
JR
740 if (q->bitrate >= RATE_HALF) {
741 for (i = 0; i < 4; i++) {
742 if (q->frame.pfrac[i] && q->frame.plag[i] >= 124) {
640760da
KG
743 warn_insufficient_frame_quality(avctx, "Cannot initialize pitch filter.");
744 goto erasure;
745 }
746 }
747 }
748 }
749
750 decode_gain_and_index(q, gain);
751 compute_svector(q, gain, outbuffer);
752
dd376b1a 753 if (decode_lspf(q, quantized_lspf) < 0) {
640760da
KG
754 warn_insufficient_frame_quality(avctx, "Badly received packets in frame.");
755 goto erasure;
756 }
757
640760da
KG
758 apply_pitch_filters(q, outbuffer);
759
dd376b1a 760 if (q->bitrate == I_F_Q) {
640760da
KG
761erasure:
762 q->bitrate = I_F_Q;
763 q->erasure_count++;
764 decode_gain_and_index(q, gain);
765 compute_svector(q, gain, outbuffer);
766 decode_lspf(q, quantized_lspf);
767 apply_pitch_filters(q, outbuffer);
25c2d76b 768 } else
640760da
KG
769 q->erasure_count = 0;
770
771 formant_mem = q->formant_mem + 10;
dd376b1a 772 for (i = 0; i < 4; i++) {
640760da 773 interpolate_lpc(q, quantized_lspf, lpc, i);
25c2d76b 774 ff_celp_lp_synthesis_filterf(formant_mem, lpc, outbuffer + i * 40, 40, 10);
640760da
KG
775 formant_mem += 40;
776 }
640760da 777
805a83d5
RB
778 // postfilter, as per TIA/EIA/IS-733 2.4.8.6
779 postfilter(q, outbuffer, lpc);
780
781 memcpy(q->formant_mem, q->formant_mem + 160, 10 * sizeof(float));
640760da 782
640760da 783 memcpy(q->prev_lspf, quantized_lspf, sizeof(q->prev_lspf));
25c2d76b 784 q->prev_bitrate = q->bitrate;
640760da 785
1b9b6d6e 786 *got_frame_ptr = 1;
640760da 787
bde25700 788 return buf_size;
640760da
KG
789}
790
25c2d76b
AL
791AVCodec ff_qcelp_decoder = {
792 .name = "qcelp",
b2bed932 793 .long_name = NULL_IF_CONFIG_SMALL("QCELP / PureVoice"),
25c2d76b 794 .type = AVMEDIA_TYPE_AUDIO,
36ef5369 795 .id = AV_CODEC_ID_QCELP,
25c2d76b
AL
796 .init = qcelp_decode_init,
797 .decode = qcelp_decode_frame,
def97856 798 .capabilities = AV_CODEC_CAP_DR1,
200de8c6 799 .priv_data_size = sizeof(QCELPContext),
200de8c6 800};