2 * G.723.1 compatible decoder
3 * Copyright (c) 2006 Benjamin Larsson
4 * Copyright (c) 2010 Mohamed Naufal Basheer
6 * This file is part of FFmpeg.
8 * FFmpeg is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
13 * FFmpeg is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with FFmpeg; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
25 * G.723.1 compatible decoder
28 #include "libavutil/channel_layout.h"
29 #include "libavutil/mem.h"
30 #include "libavutil/opt.h"
32 #define BITSTREAM_READER_LE
33 #include "acelp_vectors.h"
35 #include "celp_filters.h"
36 #include "celp_math.h"
41 #define CNG_RANDOM_SEED 12345
43 static av_cold int g723_1_decode_init(AVCodecContext *avctx)
45 G723_1_Context *p = avctx->priv_data;
47 avctx->channel_layout = AV_CH_LAYOUT_MONO;
48 avctx->sample_fmt = AV_SAMPLE_FMT_S16;
52 memcpy(p->prev_lsp, dc_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
53 memcpy(p->sid_lsp, dc_lsp, LPC_ORDER * sizeof(*p->sid_lsp));
55 p->cng_random_seed = CNG_RANDOM_SEED;
56 p->past_frame_type = SID_FRAME;
62 * Unpack the frame into parameters.
64 * @param p the context
65 * @param buf pointer to the input buffer
66 * @param buf_size size of the input buffer
68 static int unpack_bitstream(G723_1_Context *p, const uint8_t *buf,
73 int temp, info_bits, i;
75 init_get_bits(&gb, buf, buf_size * 8);
77 /* Extract frame type and rate info */
78 info_bits = get_bits(&gb, 2);
81 p->cur_frame_type = UNTRANSMITTED_FRAME;
85 /* Extract 24 bit lsp indices, 8 bit for each band */
86 p->lsp_index[2] = get_bits(&gb, 8);
87 p->lsp_index[1] = get_bits(&gb, 8);
88 p->lsp_index[0] = get_bits(&gb, 8);
91 p->cur_frame_type = SID_FRAME;
92 p->subframe[0].amp_index = get_bits(&gb, 6);
96 /* Extract the info common to both rates */
97 p->cur_rate = info_bits ? RATE_5300 : RATE_6300;
98 p->cur_frame_type = ACTIVE_FRAME;
100 p->pitch_lag[0] = get_bits(&gb, 7);
101 if (p->pitch_lag[0] > 123) /* test if forbidden code */
103 p->pitch_lag[0] += PITCH_MIN;
104 p->subframe[1].ad_cb_lag = get_bits(&gb, 2);
106 p->pitch_lag[1] = get_bits(&gb, 7);
107 if (p->pitch_lag[1] > 123)
109 p->pitch_lag[1] += PITCH_MIN;
110 p->subframe[3].ad_cb_lag = get_bits(&gb, 2);
111 p->subframe[0].ad_cb_lag = 1;
112 p->subframe[2].ad_cb_lag = 1;
114 for (i = 0; i < SUBFRAMES; i++) {
115 /* Extract combined gain */
116 temp = get_bits(&gb, 12);
118 p->subframe[i].dirac_train = 0;
119 if (p->cur_rate == RATE_6300 && p->pitch_lag[i >> 1] < SUBFRAME_LEN - 2) {
120 p->subframe[i].dirac_train = temp >> 11;
124 p->subframe[i].ad_cb_gain = FASTDIV(temp, GAIN_LEVELS);
125 if (p->subframe[i].ad_cb_gain < ad_cb_len) {
126 p->subframe[i].amp_index = temp - p->subframe[i].ad_cb_gain *
133 p->subframe[0].grid_index = get_bits1(&gb);
134 p->subframe[1].grid_index = get_bits1(&gb);
135 p->subframe[2].grid_index = get_bits1(&gb);
136 p->subframe[3].grid_index = get_bits1(&gb);
138 if (p->cur_rate == RATE_6300) {
139 skip_bits1(&gb); /* skip reserved bit */
141 /* Compute pulse_pos index using the 13-bit combined position index */
142 temp = get_bits(&gb, 13);
143 p->subframe[0].pulse_pos = temp / 810;
145 temp -= p->subframe[0].pulse_pos * 810;
146 p->subframe[1].pulse_pos = FASTDIV(temp, 90);
148 temp -= p->subframe[1].pulse_pos * 90;
149 p->subframe[2].pulse_pos = FASTDIV(temp, 9);
150 p->subframe[3].pulse_pos = temp - p->subframe[2].pulse_pos * 9;
152 p->subframe[0].pulse_pos = (p->subframe[0].pulse_pos << 16) +
154 p->subframe[1].pulse_pos = (p->subframe[1].pulse_pos << 14) +
156 p->subframe[2].pulse_pos = (p->subframe[2].pulse_pos << 16) +
158 p->subframe[3].pulse_pos = (p->subframe[3].pulse_pos << 14) +
161 p->subframe[0].pulse_sign = get_bits(&gb, 6);
162 p->subframe[1].pulse_sign = get_bits(&gb, 5);
163 p->subframe[2].pulse_sign = get_bits(&gb, 6);
164 p->subframe[3].pulse_sign = get_bits(&gb, 5);
165 } else { /* 5300 bps */
166 p->subframe[0].pulse_pos = get_bits(&gb, 12);
167 p->subframe[1].pulse_pos = get_bits(&gb, 12);
168 p->subframe[2].pulse_pos = get_bits(&gb, 12);
169 p->subframe[3].pulse_pos = get_bits(&gb, 12);
171 p->subframe[0].pulse_sign = get_bits(&gb, 4);
172 p->subframe[1].pulse_sign = get_bits(&gb, 4);
173 p->subframe[2].pulse_sign = get_bits(&gb, 4);
174 p->subframe[3].pulse_sign = get_bits(&gb, 4);
181 * Bitexact implementation of sqrt(val/2).
183 static int16_t square_root(unsigned val)
185 av_assert2(!(val & 0x80000000));
187 return (ff_sqrt(val << 1) >> 1) & (~1);
191 * Generate fixed codebook excitation vector.
193 * @param vector decoded excitation vector
194 * @param subfrm current subframe
195 * @param cur_rate current bitrate
196 * @param pitch_lag closed loop pitch lag
197 * @param index current subframe index
199 static void gen_fcb_excitation(int16_t *vector, G723_1_Subframe *subfrm,
200 enum Rate cur_rate, int pitch_lag, int index)
204 memset(vector, 0, SUBFRAME_LEN * sizeof(*vector));
206 if (cur_rate == RATE_6300) {
207 if (subfrm->pulse_pos >= max_pos[index])
210 /* Decode amplitudes and positions */
211 j = PULSE_MAX - pulses[index];
212 temp = subfrm->pulse_pos;
213 for (i = 0; i < SUBFRAME_LEN / GRID_SIZE; i++) {
214 temp -= combinatorial_table[j][i];
217 temp += combinatorial_table[j++][i];
218 if (subfrm->pulse_sign & (1 << (PULSE_MAX - j))) {
219 vector[subfrm->grid_index + GRID_SIZE * i] =
220 -fixed_cb_gain[subfrm->amp_index];
222 vector[subfrm->grid_index + GRID_SIZE * i] =
223 fixed_cb_gain[subfrm->amp_index];
228 if (subfrm->dirac_train == 1)
229 ff_g723_1_gen_dirac_train(vector, pitch_lag);
230 } else { /* 5300 bps */
231 int cb_gain = fixed_cb_gain[subfrm->amp_index];
232 int cb_shift = subfrm->grid_index;
233 int cb_sign = subfrm->pulse_sign;
234 int cb_pos = subfrm->pulse_pos;
235 int offset, beta, lag;
237 for (i = 0; i < 8; i += 2) {
238 offset = ((cb_pos & 7) << 3) + cb_shift + i;
239 vector[offset] = (cb_sign & 1) ? cb_gain : -cb_gain;
244 /* Enhance harmonic components */
245 lag = pitch_contrib[subfrm->ad_cb_gain << 1] + pitch_lag +
246 subfrm->ad_cb_lag - 1;
247 beta = pitch_contrib[(subfrm->ad_cb_gain << 1) + 1];
249 if (lag < SUBFRAME_LEN - 2) {
250 for (i = lag; i < SUBFRAME_LEN; i++)
251 vector[i] += beta * vector[i - lag] >> 15;
257 * Estimate maximum auto-correlation around pitch lag.
259 * @param buf buffer with offset applied
260 * @param offset offset of the excitation vector
261 * @param ccr_max pointer to the maximum auto-correlation
262 * @param pitch_lag decoded pitch lag
263 * @param length length of autocorrelation
264 * @param dir forward lag(1) / backward lag(-1)
266 static int autocorr_max(const int16_t *buf, int offset, int *ccr_max,
267 int pitch_lag, int length, int dir)
269 int limit, ccr, lag = 0;
272 pitch_lag = FFMIN(PITCH_MAX - 3, pitch_lag);
274 limit = FFMIN(FRAME_LEN + PITCH_MAX - offset - length, pitch_lag + 3);
276 limit = pitch_lag + 3;
278 for (i = pitch_lag - 3; i <= limit; i++) {
279 ccr = ff_g723_1_dot_product(buf, buf + dir * i, length);
281 if (ccr > *ccr_max) {
290 * Calculate pitch postfilter optimal and scaling gains.
292 * @param lag pitch postfilter forward/backward lag
293 * @param ppf pitch postfilter parameters
294 * @param cur_rate current bitrate
295 * @param tgt_eng target energy
296 * @param ccr cross-correlation
297 * @param res_eng residual energy
299 static void comp_ppf_gains(int lag, PPFParam *ppf, enum Rate cur_rate,
300 int tgt_eng, int ccr, int res_eng)
302 int pf_residual; /* square of postfiltered residual */
307 temp1 = tgt_eng * res_eng >> 1;
308 temp2 = ccr * ccr << 1;
311 if (ccr >= res_eng) {
312 ppf->opt_gain = ppf_gain_weight[cur_rate];
314 ppf->opt_gain = (ccr << 15) / res_eng *
315 ppf_gain_weight[cur_rate] >> 15;
317 /* pf_res^2 = tgt_eng + 2*ccr*gain + res_eng*gain^2 */
318 temp1 = (tgt_eng << 15) + (ccr * ppf->opt_gain << 1);
319 temp2 = (ppf->opt_gain * ppf->opt_gain >> 15) * res_eng;
320 pf_residual = av_sat_add32(temp1, temp2 + (1 << 15)) >> 16;
322 if (tgt_eng >= pf_residual << 1) {
325 temp1 = (tgt_eng << 14) / pf_residual;
328 /* scaling_gain = sqrt(tgt_eng/pf_res^2) */
329 ppf->sc_gain = square_root(temp1 << 16);
332 ppf->sc_gain = 0x7fff;
335 ppf->opt_gain = av_clip_int16(ppf->opt_gain * ppf->sc_gain >> 15);
339 * Calculate pitch postfilter parameters.
341 * @param p the context
342 * @param offset offset of the excitation vector
343 * @param pitch_lag decoded pitch lag
344 * @param ppf pitch postfilter parameters
345 * @param cur_rate current bitrate
347 static void comp_ppf_coeff(G723_1_Context *p, int offset, int pitch_lag,
348 PPFParam *ppf, enum Rate cur_rate)
357 * 1 - forward cross-correlation
358 * 2 - forward residual energy
359 * 3 - backward cross-correlation
360 * 4 - backward residual energy
362 int energy[5] = {0, 0, 0, 0, 0};
363 int16_t *buf = p->audio + LPC_ORDER + offset;
364 int fwd_lag = autocorr_max(buf, offset, &energy[1], pitch_lag,
366 int back_lag = autocorr_max(buf, offset, &energy[3], pitch_lag,
371 ppf->sc_gain = 0x7fff;
373 /* Case 0, Section 3.6 */
374 if (!back_lag && !fwd_lag)
377 /* Compute target energy */
378 energy[0] = ff_g723_1_dot_product(buf, buf, SUBFRAME_LEN);
380 /* Compute forward residual energy */
382 energy[2] = ff_g723_1_dot_product(buf + fwd_lag, buf + fwd_lag,
385 /* Compute backward residual energy */
387 energy[4] = ff_g723_1_dot_product(buf - back_lag, buf - back_lag,
390 /* Normalize and shorten */
392 for (i = 0; i < 5; i++)
393 temp1 = FFMAX(energy[i], temp1);
395 scale = ff_g723_1_normalize_bits(temp1, 31);
396 for (i = 0; i < 5; i++)
397 energy[i] = (energy[i] << scale) >> 16;
399 if (fwd_lag && !back_lag) { /* Case 1 */
400 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
402 } else if (!fwd_lag) { /* Case 2 */
403 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
405 } else { /* Case 3 */
408 * Select the largest of energy[1]^2/energy[2]
409 * and energy[3]^2/energy[4]
411 temp1 = energy[4] * ((energy[1] * energy[1] + (1 << 14)) >> 15);
412 temp2 = energy[2] * ((energy[3] * energy[3] + (1 << 14)) >> 15);
413 if (temp1 >= temp2) {
414 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
417 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
424 * Classify frames as voiced/unvoiced.
426 * @param p the context
427 * @param pitch_lag decoded pitch_lag
428 * @param exc_eng excitation energy estimation
429 * @param scale scaling factor of exc_eng
431 * @return residual interpolation index if voiced, 0 otherwise
433 static int comp_interp_index(G723_1_Context *p, int pitch_lag,
434 int *exc_eng, int *scale)
436 int offset = PITCH_MAX + 2 * SUBFRAME_LEN;
437 int16_t *buf = p->audio + LPC_ORDER;
439 int index, ccr, tgt_eng, best_eng, temp;
441 *scale = ff_g723_1_scale_vector(buf, p->excitation, FRAME_LEN + PITCH_MAX);
444 /* Compute maximum backward cross-correlation */
446 index = autocorr_max(buf, offset, &ccr, pitch_lag, SUBFRAME_LEN * 2, -1);
447 ccr = av_sat_add32(ccr, 1 << 15) >> 16;
449 /* Compute target energy */
450 tgt_eng = ff_g723_1_dot_product(buf, buf, SUBFRAME_LEN * 2);
451 *exc_eng = av_sat_add32(tgt_eng, 1 << 15) >> 16;
456 /* Compute best energy */
457 best_eng = ff_g723_1_dot_product(buf - index, buf - index,
459 best_eng = av_sat_add32(best_eng, 1 << 15) >> 16;
461 temp = best_eng * *exc_eng >> 3;
463 if (temp < ccr * ccr) {
470 * Perform residual interpolation based on frame classification.
472 * @param buf decoded excitation vector
473 * @param out output vector
474 * @param lag decoded pitch lag
475 * @param gain interpolated gain
476 * @param rseed seed for random number generator
478 static void residual_interp(int16_t *buf, int16_t *out, int lag,
479 int gain, int *rseed)
482 if (lag) { /* Voiced */
483 int16_t *vector_ptr = buf + PITCH_MAX;
485 for (i = 0; i < lag; i++)
486 out[i] = vector_ptr[i - lag] * 3 >> 2;
487 av_memcpy_backptr((uint8_t*)(out + lag), lag * sizeof(*out),
488 (FRAME_LEN - lag) * sizeof(*out));
489 } else { /* Unvoiced */
490 for (i = 0; i < FRAME_LEN; i++) {
491 *rseed = *rseed * 521 + 259;
492 out[i] = gain * *rseed >> 15;
494 memset(buf, 0, (FRAME_LEN + PITCH_MAX) * sizeof(*buf));
499 * Perform IIR filtering.
501 * @param fir_coef FIR coefficients
502 * @param iir_coef IIR coefficients
503 * @param src source vector
504 * @param dest destination vector
505 * @param width width of the output, 16 bits(0) / 32 bits(1)
507 #define iir_filter(fir_coef, iir_coef, src, dest, width)\
510 int res_shift = 16 & ~-(width);\
511 int in_shift = 16 - res_shift;\
513 for (m = 0; m < SUBFRAME_LEN; m++) {\
515 for (n = 1; n <= LPC_ORDER; n++) {\
516 filter -= (fir_coef)[n - 1] * (src)[m - n] -\
517 (iir_coef)[n - 1] * ((dest)[m - n] >> in_shift);\
520 (dest)[m] = av_clipl_int32(((src)[m] << 16) + (filter << 3) +\
521 (1 << 15)) >> res_shift;\
526 * Adjust gain of postfiltered signal.
528 * @param p the context
529 * @param buf postfiltered output vector
530 * @param energy input energy coefficient
532 static void gain_scale(G723_1_Context *p, int16_t * buf, int energy)
534 int num, denom, gain, bits1, bits2;
539 for (i = 0; i < SUBFRAME_LEN; i++) {
540 int temp = buf[i] >> 2;
542 denom = av_sat_dadd32(denom, temp);
546 bits1 = ff_g723_1_normalize_bits(num, 31);
547 bits2 = ff_g723_1_normalize_bits(denom, 31);
548 num = num << bits1 >> 1;
551 bits2 = 5 + bits1 - bits2;
552 bits2 = FFMAX(0, bits2);
554 gain = (num >> 1) / (denom >> 16);
555 gain = square_root(gain << 16 >> bits2);
560 for (i = 0; i < SUBFRAME_LEN; i++) {
561 p->pf_gain = (15 * p->pf_gain + gain + (1 << 3)) >> 4;
562 buf[i] = av_clip_int16((buf[i] * (p->pf_gain + (p->pf_gain >> 4)) +
568 * Perform formant filtering.
570 * @param p the context
571 * @param lpc quantized lpc coefficients
572 * @param buf input buffer
573 * @param dst output buffer
575 static void formant_postfilter(G723_1_Context *p, int16_t *lpc,
576 int16_t *buf, int16_t *dst)
578 int16_t filter_coef[2][LPC_ORDER];
579 int filter_signal[LPC_ORDER + FRAME_LEN], *signal_ptr;
582 memcpy(buf, p->fir_mem, LPC_ORDER * sizeof(*buf));
583 memcpy(filter_signal, p->iir_mem, LPC_ORDER * sizeof(*filter_signal));
585 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
586 for (k = 0; k < LPC_ORDER; k++) {
587 filter_coef[0][k] = (-lpc[k] * postfilter_tbl[0][k] +
589 filter_coef[1][k] = (-lpc[k] * postfilter_tbl[1][k] +
592 iir_filter(filter_coef[0], filter_coef[1], buf + i, filter_signal + i, 1);
596 memcpy(p->fir_mem, buf + FRAME_LEN, LPC_ORDER * sizeof(int16_t));
597 memcpy(p->iir_mem, filter_signal + FRAME_LEN, LPC_ORDER * sizeof(int));
600 signal_ptr = filter_signal + LPC_ORDER;
601 for (i = 0; i < SUBFRAMES; i++) {
607 scale = ff_g723_1_scale_vector(dst, buf, SUBFRAME_LEN);
609 /* Compute auto correlation coefficients */
610 auto_corr[0] = ff_g723_1_dot_product(dst, dst + 1, SUBFRAME_LEN - 1);
611 auto_corr[1] = ff_g723_1_dot_product(dst, dst, SUBFRAME_LEN);
613 /* Compute reflection coefficient */
614 temp = auto_corr[1] >> 16;
616 temp = (auto_corr[0] >> 2) / temp;
618 p->reflection_coef = (3 * p->reflection_coef + temp + 2) >> 2;
619 temp = -p->reflection_coef >> 1 & ~3;
621 /* Compensation filter */
622 for (j = 0; j < SUBFRAME_LEN; j++) {
623 dst[j] = av_sat_dadd32(signal_ptr[j],
624 (signal_ptr[j - 1] >> 16) * temp) >> 16;
627 /* Compute normalized signal energy */
628 temp = 2 * scale + 4;
630 energy = av_clipl_int32((int64_t)auto_corr[1] << -temp);
632 energy = auto_corr[1] >> temp;
634 gain_scale(p, dst, energy);
637 signal_ptr += SUBFRAME_LEN;
642 static int sid_gain_to_lsp_index(int gain)
646 else if (gain < 0x20)
647 return gain - 8 << 7;
649 return gain - 20 << 8;
652 static inline int cng_rand(int *state, int base)
654 *state = (*state * 521 + 259) & 0xFFFF;
655 return (*state & 0x7FFF) * base >> 15;
658 static int estimate_sid_gain(G723_1_Context *p)
660 int i, shift, seg, seg2, t, val, val_add, x, y;
662 shift = 16 - p->cur_gain * 2;
664 t = p->sid_gain << shift;
666 t = p->sid_gain >> -shift;
667 x = t * cng_filt[0] >> 16;
669 if (x >= cng_bseg[2])
672 if (x >= cng_bseg[1]) {
677 seg = (x >= cng_bseg[0]);
679 seg2 = FFMIN(seg, 3);
683 for (i = 0; i < shift; i++) {
684 t = seg * 32 + (val << seg2);
693 t = seg * 32 + (val << seg2);
696 t = seg * 32 + (val + 1 << seg2);
698 val = (seg2 - 1 << 4) + val;
702 t = seg * 32 + (val - 1 << seg2);
704 val = (seg2 - 1 << 4) + val;
712 static void generate_noise(G723_1_Context *p)
716 int signs[SUBFRAMES / 2 * 11], pos[SUBFRAMES / 2 * 11];
717 int tmp[SUBFRAME_LEN * 2];
720 int b0, c, delta, x, shift;
722 p->pitch_lag[0] = cng_rand(&p->cng_random_seed, 21) + 123;
723 p->pitch_lag[1] = cng_rand(&p->cng_random_seed, 19) + 123;
725 for (i = 0; i < SUBFRAMES; i++) {
726 p->subframe[i].ad_cb_gain = cng_rand(&p->cng_random_seed, 50) + 1;
727 p->subframe[i].ad_cb_lag = cng_adaptive_cb_lag[i];
730 for (i = 0; i < SUBFRAMES / 2; i++) {
731 t = cng_rand(&p->cng_random_seed, 1 << 13);
733 off[i * 2 + 1] = ((t >> 1) & 1) + SUBFRAME_LEN;
735 for (j = 0; j < 11; j++) {
736 signs[i * 11 + j] = (t & 1) * 2 - 1 << 14;
742 for (i = 0; i < SUBFRAMES; i++) {
743 for (j = 0; j < SUBFRAME_LEN / 2; j++)
745 t = SUBFRAME_LEN / 2;
746 for (j = 0; j < pulses[i]; j++, idx++) {
747 int idx2 = cng_rand(&p->cng_random_seed, t);
749 pos[idx] = tmp[idx2] * 2 + off[i];
750 tmp[idx2] = tmp[--t];
754 vector_ptr = p->audio + LPC_ORDER;
755 memcpy(vector_ptr, p->prev_excitation,
756 PITCH_MAX * sizeof(*p->excitation));
757 for (i = 0; i < SUBFRAMES; i += 2) {
758 ff_g723_1_gen_acb_excitation(vector_ptr, vector_ptr,
759 p->pitch_lag[i >> 1], &p->subframe[i],
761 ff_g723_1_gen_acb_excitation(vector_ptr + SUBFRAME_LEN,
762 vector_ptr + SUBFRAME_LEN,
763 p->pitch_lag[i >> 1], &p->subframe[i + 1],
767 for (j = 0; j < SUBFRAME_LEN * 2; j++)
768 t |= FFABS(vector_ptr[j]);
769 t = FFMIN(t, 0x7FFF);
773 shift = -10 + av_log2(t);
779 for (j = 0; j < SUBFRAME_LEN * 2; j++) {
780 t = vector_ptr[j] << -shift;
785 for (j = 0; j < SUBFRAME_LEN * 2; j++) {
786 t = vector_ptr[j] >> shift;
793 for (j = 0; j < 11; j++)
794 b0 += tmp[pos[(i / 2) * 11 + j]] * signs[(i / 2) * 11 + j];
795 b0 = b0 * 2 * 2979LL + (1 << 29) >> 30; // approximated division by 11
797 c = p->cur_gain * (p->cur_gain * SUBFRAME_LEN >> 5);
798 if (shift * 2 + 3 >= 0)
801 c <<= -(shift * 2 + 3);
802 c = (av_clipl_int32(sum << 1) - c) * 2979LL >> 15;
804 delta = b0 * b0 * 2 - c;
808 delta = square_root(delta);
811 if (FFABS(t) < FFABS(x))
819 x = av_clip(x, -10000, 10000);
821 for (j = 0; j < 11; j++) {
822 idx = (i / 2) * 11 + j;
823 vector_ptr[pos[idx]] = av_clip_int16(vector_ptr[pos[idx]] +
824 (x * signs[idx] >> 15));
827 /* copy decoded data to serve as a history for the next decoded subframes */
828 memcpy(vector_ptr + PITCH_MAX, vector_ptr,
829 sizeof(*vector_ptr) * SUBFRAME_LEN * 2);
830 vector_ptr += SUBFRAME_LEN * 2;
832 /* Save the excitation for the next frame */
833 memcpy(p->prev_excitation, p->audio + LPC_ORDER + FRAME_LEN,
834 PITCH_MAX * sizeof(*p->excitation));
837 static int g723_1_decode_frame(AVCodecContext *avctx, void *data,
838 int *got_frame_ptr, AVPacket *avpkt)
840 G723_1_Context *p = avctx->priv_data;
841 AVFrame *frame = data;
842 const uint8_t *buf = avpkt->data;
843 int buf_size = avpkt->size;
844 int dec_mode = buf[0] & 3;
846 PPFParam ppf[SUBFRAMES];
847 int16_t cur_lsp[LPC_ORDER];
848 int16_t lpc[SUBFRAMES * LPC_ORDER];
849 int16_t acb_vector[SUBFRAME_LEN];
851 int bad_frame = 0, i, j, ret;
852 int16_t *audio = p->audio;
854 if (buf_size < frame_size[dec_mode]) {
856 av_log(avctx, AV_LOG_WARNING,
857 "Expected %d bytes, got %d - skipping packet\n",
858 frame_size[dec_mode], buf_size);
863 if (unpack_bitstream(p, buf, buf_size) < 0) {
865 if (p->past_frame_type == ACTIVE_FRAME)
866 p->cur_frame_type = ACTIVE_FRAME;
868 p->cur_frame_type = UNTRANSMITTED_FRAME;
871 frame->nb_samples = FRAME_LEN;
872 if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
875 out = (int16_t *)frame->data[0];
877 if (p->cur_frame_type == ACTIVE_FRAME) {
879 p->erased_frames = 0;
880 else if (p->erased_frames != 3)
883 ff_g723_1_inverse_quant(cur_lsp, p->prev_lsp, p->lsp_index, bad_frame);
884 ff_g723_1_lsp_interpolate(lpc, cur_lsp, p->prev_lsp);
886 /* Save the lsp_vector for the next frame */
887 memcpy(p->prev_lsp, cur_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
889 /* Generate the excitation for the frame */
890 memcpy(p->excitation, p->prev_excitation,
891 PITCH_MAX * sizeof(*p->excitation));
892 if (!p->erased_frames) {
893 int16_t *vector_ptr = p->excitation + PITCH_MAX;
895 /* Update interpolation gain memory */
896 p->interp_gain = fixed_cb_gain[(p->subframe[2].amp_index +
897 p->subframe[3].amp_index) >> 1];
898 for (i = 0; i < SUBFRAMES; i++) {
899 gen_fcb_excitation(vector_ptr, &p->subframe[i], p->cur_rate,
900 p->pitch_lag[i >> 1], i);
901 ff_g723_1_gen_acb_excitation(acb_vector,
902 &p->excitation[SUBFRAME_LEN * i],
903 p->pitch_lag[i >> 1],
904 &p->subframe[i], p->cur_rate);
905 /* Get the total excitation */
906 for (j = 0; j < SUBFRAME_LEN; j++) {
907 int v = av_clip_int16(vector_ptr[j] << 1);
908 vector_ptr[j] = av_clip_int16(v + acb_vector[j]);
910 vector_ptr += SUBFRAME_LEN;
913 vector_ptr = p->excitation + PITCH_MAX;
915 p->interp_index = comp_interp_index(p, p->pitch_lag[1],
916 &p->sid_gain, &p->cur_gain);
918 /* Perform pitch postfiltering */
921 for (j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
922 comp_ppf_coeff(p, i, p->pitch_lag[j >> 1],
923 ppf + j, p->cur_rate);
925 for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
926 ff_acelp_weighted_vector_sum(p->audio + LPC_ORDER + i,
928 vector_ptr + i + ppf[j].index,
931 1 << 14, 15, SUBFRAME_LEN);
933 audio = vector_ptr - LPC_ORDER;
936 /* Save the excitation for the next frame */
937 memcpy(p->prev_excitation, p->excitation + FRAME_LEN,
938 PITCH_MAX * sizeof(*p->excitation));
940 p->interp_gain = (p->interp_gain * 3 + 2) >> 2;
941 if (p->erased_frames == 3) {
943 memset(p->excitation, 0,
944 (FRAME_LEN + PITCH_MAX) * sizeof(*p->excitation));
945 memset(p->prev_excitation, 0,
946 PITCH_MAX * sizeof(*p->excitation));
947 memset(frame->data[0], 0,
948 (FRAME_LEN + LPC_ORDER) * sizeof(int16_t));
950 int16_t *buf = p->audio + LPC_ORDER;
952 /* Regenerate frame */
953 residual_interp(p->excitation, buf, p->interp_index,
954 p->interp_gain, &p->random_seed);
956 /* Save the excitation for the next frame */
957 memcpy(p->prev_excitation, buf + (FRAME_LEN - PITCH_MAX),
958 PITCH_MAX * sizeof(*p->excitation));
961 p->cng_random_seed = CNG_RANDOM_SEED;
963 if (p->cur_frame_type == SID_FRAME) {
964 p->sid_gain = sid_gain_to_lsp_index(p->subframe[0].amp_index);
965 ff_g723_1_inverse_quant(p->sid_lsp, p->prev_lsp, p->lsp_index, 0);
966 } else if (p->past_frame_type == ACTIVE_FRAME) {
967 p->sid_gain = estimate_sid_gain(p);
970 if (p->past_frame_type == ACTIVE_FRAME)
971 p->cur_gain = p->sid_gain;
973 p->cur_gain = (p->cur_gain * 7 + p->sid_gain) >> 3;
975 ff_g723_1_lsp_interpolate(lpc, p->sid_lsp, p->prev_lsp);
976 /* Save the lsp_vector for the next frame */
977 memcpy(p->prev_lsp, p->sid_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
980 p->past_frame_type = p->cur_frame_type;
982 memcpy(p->audio, p->synth_mem, LPC_ORDER * sizeof(*p->audio));
983 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
984 ff_celp_lp_synthesis_filter(p->audio + i, &lpc[j * LPC_ORDER],
985 audio + i, SUBFRAME_LEN, LPC_ORDER,
987 memcpy(p->synth_mem, p->audio + FRAME_LEN, LPC_ORDER * sizeof(*p->audio));
990 formant_postfilter(p, lpc, p->audio, out);
991 } else { // if output is not postfiltered it should be scaled by 2
992 for (i = 0; i < FRAME_LEN; i++)
993 out[i] = av_clip_int16(p->audio[LPC_ORDER + i] << 1);
998 return frame_size[dec_mode];
1001 #define OFFSET(x) offsetof(G723_1_Context, x)
1002 #define AD AV_OPT_FLAG_AUDIO_PARAM | AV_OPT_FLAG_DECODING_PARAM
1004 static const AVOption options[] = {
1005 { "postfilter", "enable postfilter", OFFSET(postfilter), AV_OPT_TYPE_BOOL,
1006 { .i64 = 1 }, 0, 1, AD },
1011 static const AVClass g723_1dec_class = {
1012 .class_name = "G.723.1 decoder",
1013 .item_name = av_default_item_name,
1015 .version = LIBAVUTIL_VERSION_INT,
1018 AVCodec ff_g723_1_decoder = {
1020 .long_name = NULL_IF_CONFIG_SMALL("G.723.1"),
1021 .type = AVMEDIA_TYPE_AUDIO,
1022 .id = AV_CODEC_ID_G723_1,
1023 .priv_data_size = sizeof(G723_1_Context),
1024 .init = g723_1_decode_init,
1025 .decode = g723_1_decode_frame,
1026 .capabilities = AV_CODEC_CAP_SUBFRAMES | AV_CODEC_CAP_DR1,
1027 .priv_class = &g723_1dec_class,