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 #define BITSTREAM_READER_LE
29 #include "libavutil/audioconvert.h"
30 #include "libavutil/lzo.h"
31 #include "libavutil/opt.h"
35 #include "acelp_vectors.h"
36 #include "celp_filters.h"
37 #include "celp_math.h"
39 #include "g723_1_data.h"
41 typedef struct g723_1_context {
45 G723_1_Subframe subframe[4];
46 enum FrameType cur_frame_type;
47 enum FrameType past_frame_type;
49 uint8_t lsp_index[LSP_BANDS];
53 int16_t prev_lsp[LPC_ORDER];
54 int16_t prev_excitation[PITCH_MAX];
55 int16_t excitation[PITCH_MAX + FRAME_LEN];
56 int16_t synth_mem[LPC_ORDER];
57 int16_t fir_mem[LPC_ORDER];
58 int iir_mem[LPC_ORDER];
66 int pf_gain; ///< formant postfilter
67 ///< gain scaling unit memory
70 int16_t prev_data[HALF_FRAME_LEN];
71 int16_t prev_weight_sig[PITCH_MAX];
74 int16_t hpf_fir_mem; ///< highpass filter fir
75 int hpf_iir_mem; ///< and iir memories
76 int16_t perf_fir_mem[LPC_ORDER]; ///< perceptual filter fir
77 int16_t perf_iir_mem[LPC_ORDER]; ///< and iir memories
79 int16_t harmonic_mem[PITCH_MAX];
82 static av_cold int g723_1_decode_init(AVCodecContext *avctx)
84 G723_1_Context *p = avctx->priv_data;
86 avctx->sample_fmt = AV_SAMPLE_FMT_S16;
89 avcodec_get_frame_defaults(&p->frame);
90 avctx->coded_frame = &p->frame;
92 memcpy(p->prev_lsp, dc_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
98 * Unpack the frame into parameters.
100 * @param p the context
101 * @param buf pointer to the input buffer
102 * @param buf_size size of the input buffer
104 static int unpack_bitstream(G723_1_Context *p, const uint8_t *buf,
109 int temp, info_bits, i;
111 init_get_bits(&gb, buf, buf_size * 8);
113 /* Extract frame type and rate info */
114 info_bits = get_bits(&gb, 2);
116 if (info_bits == 3) {
117 p->cur_frame_type = UNTRANSMITTED_FRAME;
121 /* Extract 24 bit lsp indices, 8 bit for each band */
122 p->lsp_index[2] = get_bits(&gb, 8);
123 p->lsp_index[1] = get_bits(&gb, 8);
124 p->lsp_index[0] = get_bits(&gb, 8);
126 if (info_bits == 2) {
127 p->cur_frame_type = SID_FRAME;
128 p->subframe[0].amp_index = get_bits(&gb, 6);
132 /* Extract the info common to both rates */
133 p->cur_rate = info_bits ? RATE_5300 : RATE_6300;
134 p->cur_frame_type = ACTIVE_FRAME;
136 p->pitch_lag[0] = get_bits(&gb, 7);
137 if (p->pitch_lag[0] > 123) /* test if forbidden code */
139 p->pitch_lag[0] += PITCH_MIN;
140 p->subframe[1].ad_cb_lag = get_bits(&gb, 2);
142 p->pitch_lag[1] = get_bits(&gb, 7);
143 if (p->pitch_lag[1] > 123)
145 p->pitch_lag[1] += PITCH_MIN;
146 p->subframe[3].ad_cb_lag = get_bits(&gb, 2);
147 p->subframe[0].ad_cb_lag = 1;
148 p->subframe[2].ad_cb_lag = 1;
150 for (i = 0; i < SUBFRAMES; i++) {
151 /* Extract combined gain */
152 temp = get_bits(&gb, 12);
154 p->subframe[i].dirac_train = 0;
155 if (p->cur_rate == RATE_6300 && p->pitch_lag[i >> 1] < SUBFRAME_LEN - 2) {
156 p->subframe[i].dirac_train = temp >> 11;
160 p->subframe[i].ad_cb_gain = FASTDIV(temp, GAIN_LEVELS);
161 if (p->subframe[i].ad_cb_gain < ad_cb_len) {
162 p->subframe[i].amp_index = temp - p->subframe[i].ad_cb_gain *
169 p->subframe[0].grid_index = get_bits1(&gb);
170 p->subframe[1].grid_index = get_bits1(&gb);
171 p->subframe[2].grid_index = get_bits1(&gb);
172 p->subframe[3].grid_index = get_bits1(&gb);
174 if (p->cur_rate == RATE_6300) {
175 skip_bits1(&gb); /* skip reserved bit */
177 /* Compute pulse_pos index using the 13-bit combined position index */
178 temp = get_bits(&gb, 13);
179 p->subframe[0].pulse_pos = temp / 810;
181 temp -= p->subframe[0].pulse_pos * 810;
182 p->subframe[1].pulse_pos = FASTDIV(temp, 90);
184 temp -= p->subframe[1].pulse_pos * 90;
185 p->subframe[2].pulse_pos = FASTDIV(temp, 9);
186 p->subframe[3].pulse_pos = temp - p->subframe[2].pulse_pos * 9;
188 p->subframe[0].pulse_pos = (p->subframe[0].pulse_pos << 16) +
190 p->subframe[1].pulse_pos = (p->subframe[1].pulse_pos << 14) +
192 p->subframe[2].pulse_pos = (p->subframe[2].pulse_pos << 16) +
194 p->subframe[3].pulse_pos = (p->subframe[3].pulse_pos << 14) +
197 p->subframe[0].pulse_sign = get_bits(&gb, 6);
198 p->subframe[1].pulse_sign = get_bits(&gb, 5);
199 p->subframe[2].pulse_sign = get_bits(&gb, 6);
200 p->subframe[3].pulse_sign = get_bits(&gb, 5);
201 } else { /* 5300 bps */
202 p->subframe[0].pulse_pos = get_bits(&gb, 12);
203 p->subframe[1].pulse_pos = get_bits(&gb, 12);
204 p->subframe[2].pulse_pos = get_bits(&gb, 12);
205 p->subframe[3].pulse_pos = get_bits(&gb, 12);
207 p->subframe[0].pulse_sign = get_bits(&gb, 4);
208 p->subframe[1].pulse_sign = get_bits(&gb, 4);
209 p->subframe[2].pulse_sign = get_bits(&gb, 4);
210 p->subframe[3].pulse_sign = get_bits(&gb, 4);
217 * Bitexact implementation of sqrt(val/2).
219 static int16_t square_root(int val)
221 return (ff_sqrt(val << 1) >> 1) & (~1);
225 * Calculate the number of left-shifts required for normalizing the input.
227 * @param num input number
228 * @param width width of the input, 16 bits(0) / 32 bits(1)
230 static int normalize_bits(int num, int width)
233 int bits = (width) ? 31 : 15;
240 i= bits - av_log2(num) - 1;
246 #define normalize_bits_int16(num) normalize_bits(num, 0)
247 #define normalize_bits_int32(num) normalize_bits(num, 1)
248 #define dot_product(a,b,c,d) (ff_dot_product(a,b,c)<<(d))
251 * Scale vector contents based on the largest of their absolutes.
253 static int scale_vector(int16_t *vector, int length)
255 int bits, scale, max = 0;
258 const int16_t shift_table[16] = {
259 0x0001, 0x0002, 0x0004, 0x0008, 0x0010, 0x0020, 0x0040, 0x0080,
260 0x0100, 0x0200, 0x0400, 0x0800, 0x1000, 0x2000, 0x4000, 0x7fff
263 for (i = 0; i < length; i++)
264 max = FFMAX(max, FFABS(vector[i]));
266 bits = normalize_bits(max, 0);
267 scale = shift_table[bits];
269 for (i = 0; i < length; i++)
270 vector[i] = (vector[i] * scale) >> 3;
276 * Perform inverse quantization of LSP frequencies.
278 * @param cur_lsp the current LSP vector
279 * @param prev_lsp the previous LSP vector
280 * @param lsp_index VQ indices
281 * @param bad_frame bad frame flag
283 static void inverse_quant(int16_t *cur_lsp, int16_t *prev_lsp,
284 uint8_t *lsp_index, int bad_frame)
287 int i, j, temp, stable;
289 /* Check for frame erasure */
296 lsp_index[0] = lsp_index[1] = lsp_index[2] = 0;
299 /* Get the VQ table entry corresponding to the transmitted index */
300 cur_lsp[0] = lsp_band0[lsp_index[0]][0];
301 cur_lsp[1] = lsp_band0[lsp_index[0]][1];
302 cur_lsp[2] = lsp_band0[lsp_index[0]][2];
303 cur_lsp[3] = lsp_band1[lsp_index[1]][0];
304 cur_lsp[4] = lsp_band1[lsp_index[1]][1];
305 cur_lsp[5] = lsp_band1[lsp_index[1]][2];
306 cur_lsp[6] = lsp_band2[lsp_index[2]][0];
307 cur_lsp[7] = lsp_band2[lsp_index[2]][1];
308 cur_lsp[8] = lsp_band2[lsp_index[2]][2];
309 cur_lsp[9] = lsp_band2[lsp_index[2]][3];
311 /* Add predicted vector & DC component to the previously quantized vector */
312 for (i = 0; i < LPC_ORDER; i++) {
313 temp = ((prev_lsp[i] - dc_lsp[i]) * pred + (1 << 14)) >> 15;
314 cur_lsp[i] += dc_lsp[i] + temp;
317 for (i = 0; i < LPC_ORDER; i++) {
318 cur_lsp[0] = FFMAX(cur_lsp[0], 0x180);
319 cur_lsp[LPC_ORDER - 1] = FFMIN(cur_lsp[LPC_ORDER - 1], 0x7e00);
321 /* Stability check */
322 for (j = 1; j < LPC_ORDER; j++) {
323 temp = min_dist + cur_lsp[j - 1] - cur_lsp[j];
326 cur_lsp[j - 1] -= temp;
331 for (j = 1; j < LPC_ORDER; j++) {
332 temp = cur_lsp[j - 1] + min_dist - cur_lsp[j] - 4;
342 memcpy(cur_lsp, prev_lsp, LPC_ORDER * sizeof(*cur_lsp));
346 * Bitexact implementation of 2ab scaled by 1/2^16.
348 * @param a 32 bit multiplicand
349 * @param b 16 bit multiplier
351 #define MULL2(a, b) \
355 * Convert LSP frequencies to LPC coefficients.
357 * @param lpc buffer for LPC coefficients
359 static void lsp2lpc(int16_t *lpc)
361 int f1[LPC_ORDER / 2 + 1];
362 int f2[LPC_ORDER / 2 + 1];
365 /* Calculate negative cosine */
366 for (j = 0; j < LPC_ORDER; j++) {
367 int index = lpc[j] >> 7;
368 int offset = lpc[j] & 0x7f;
369 int64_t temp1 = cos_tab[index] << 16;
370 int temp2 = (cos_tab[index + 1] - cos_tab[index]) *
371 ((offset << 8) + 0x80) << 1;
373 lpc[j] = -(av_clipl_int32(((temp1 + temp2) << 1) + (1 << 15)) >> 16);
377 * Compute sum and difference polynomial coefficients
378 * (bitexact alternative to lsp2poly() in lsp.c)
380 /* Initialize with values in Q28 */
382 f1[1] = (lpc[0] << 14) + (lpc[2] << 14);
383 f1[2] = lpc[0] * lpc[2] + (2 << 28);
386 f2[1] = (lpc[1] << 14) + (lpc[3] << 14);
387 f2[2] = lpc[1] * lpc[3] + (2 << 28);
390 * Calculate and scale the coefficients by 1/2 in
391 * each iteration for a final scaling factor of Q25
393 for (i = 2; i < LPC_ORDER / 2; i++) {
394 f1[i + 1] = f1[i - 1] + MULL2(f1[i], lpc[2 * i]);
395 f2[i + 1] = f2[i - 1] + MULL2(f2[i], lpc[2 * i + 1]);
397 for (j = i; j >= 2; j--) {
398 f1[j] = MULL2(f1[j - 1], lpc[2 * i]) +
399 (f1[j] >> 1) + (f1[j - 2] >> 1);
400 f2[j] = MULL2(f2[j - 1], lpc[2 * i + 1]) +
401 (f2[j] >> 1) + (f2[j - 2] >> 1);
406 f1[1] = ((lpc[2 * i] << 16 >> i) + f1[1]) >> 1;
407 f2[1] = ((lpc[2 * i + 1] << 16 >> i) + f2[1]) >> 1;
410 /* Convert polynomial coefficients to LPC coefficients */
411 for (i = 0; i < LPC_ORDER / 2; i++) {
412 int64_t ff1 = f1[i + 1] + f1[i];
413 int64_t ff2 = f2[i + 1] - f2[i];
415 lpc[i] = av_clipl_int32(((ff1 + ff2) << 3) + (1 << 15)) >> 16;
416 lpc[LPC_ORDER - i - 1] = av_clipl_int32(((ff1 - ff2) << 3) +
422 * Quantize LSP frequencies by interpolation and convert them to
423 * the corresponding LPC coefficients.
425 * @param lpc buffer for LPC coefficients
426 * @param cur_lsp the current LSP vector
427 * @param prev_lsp the previous LSP vector
429 static void lsp_interpolate(int16_t *lpc, int16_t *cur_lsp, int16_t *prev_lsp)
432 int16_t *lpc_ptr = lpc;
434 /* cur_lsp * 0.25 + prev_lsp * 0.75 */
435 ff_acelp_weighted_vector_sum(lpc, cur_lsp, prev_lsp,
436 4096, 12288, 1 << 13, 14, LPC_ORDER);
437 ff_acelp_weighted_vector_sum(lpc + LPC_ORDER, cur_lsp, prev_lsp,
438 8192, 8192, 1 << 13, 14, LPC_ORDER);
439 ff_acelp_weighted_vector_sum(lpc + 2 * LPC_ORDER, cur_lsp, prev_lsp,
440 12288, 4096, 1 << 13, 14, LPC_ORDER);
441 memcpy(lpc + 3 * LPC_ORDER, cur_lsp, LPC_ORDER * sizeof(*lpc));
443 for (i = 0; i < SUBFRAMES; i++) {
445 lpc_ptr += LPC_ORDER;
450 * Generate a train of dirac functions with period as pitch lag.
452 static void gen_dirac_train(int16_t *buf, int pitch_lag)
454 int16_t vector[SUBFRAME_LEN];
457 memcpy(vector, buf, SUBFRAME_LEN * sizeof(*vector));
458 for (i = pitch_lag; i < SUBFRAME_LEN; i += pitch_lag) {
459 for (j = 0; j < SUBFRAME_LEN - i; j++)
460 buf[i + j] += vector[j];
465 * Generate fixed codebook excitation vector.
467 * @param vector decoded excitation vector
468 * @param subfrm current subframe
469 * @param cur_rate current bitrate
470 * @param pitch_lag closed loop pitch lag
471 * @param index current subframe index
473 static void gen_fcb_excitation(int16_t *vector, G723_1_Subframe subfrm,
474 enum Rate cur_rate, int pitch_lag, int index)
478 memset(vector, 0, SUBFRAME_LEN * sizeof(*vector));
480 if (cur_rate == RATE_6300) {
481 if (subfrm.pulse_pos >= max_pos[index])
484 /* Decode amplitudes and positions */
485 j = PULSE_MAX - pulses[index];
486 temp = subfrm.pulse_pos;
487 for (i = 0; i < SUBFRAME_LEN / GRID_SIZE; i++) {
488 temp -= combinatorial_table[j][i];
491 temp += combinatorial_table[j++][i];
492 if (subfrm.pulse_sign & (1 << (PULSE_MAX - j))) {
493 vector[subfrm.grid_index + GRID_SIZE * i] =
494 -fixed_cb_gain[subfrm.amp_index];
496 vector[subfrm.grid_index + GRID_SIZE * i] =
497 fixed_cb_gain[subfrm.amp_index];
502 if (subfrm.dirac_train == 1)
503 gen_dirac_train(vector, pitch_lag);
504 } else { /* 5300 bps */
505 int cb_gain = fixed_cb_gain[subfrm.amp_index];
506 int cb_shift = subfrm.grid_index;
507 int cb_sign = subfrm.pulse_sign;
508 int cb_pos = subfrm.pulse_pos;
509 int offset, beta, lag;
511 for (i = 0; i < 8; i += 2) {
512 offset = ((cb_pos & 7) << 3) + cb_shift + i;
513 vector[offset] = (cb_sign & 1) ? cb_gain : -cb_gain;
518 /* Enhance harmonic components */
519 lag = pitch_contrib[subfrm.ad_cb_gain << 1] + pitch_lag +
520 subfrm.ad_cb_lag - 1;
521 beta = pitch_contrib[(subfrm.ad_cb_gain << 1) + 1];
523 if (lag < SUBFRAME_LEN - 2) {
524 for (i = lag; i < SUBFRAME_LEN; i++)
525 vector[i] += beta * vector[i - lag] >> 15;
531 * Get delayed contribution from the previous excitation vector.
533 static void get_residual(int16_t *residual, int16_t *prev_excitation, int lag)
535 int offset = PITCH_MAX - PITCH_ORDER / 2 - lag;
538 residual[0] = prev_excitation[offset];
539 residual[1] = prev_excitation[offset + 1];
542 for (i = 2; i < SUBFRAME_LEN + PITCH_ORDER - 1; i++)
543 residual[i] = prev_excitation[offset + (i - 2) % lag];
547 * Generate adaptive codebook excitation.
549 static void gen_acb_excitation(int16_t *vector, int16_t *prev_excitation,
550 int pitch_lag, G723_1_Subframe subfrm,
553 int16_t residual[SUBFRAME_LEN + PITCH_ORDER - 1];
554 const int16_t *cb_ptr;
555 int lag = pitch_lag + subfrm.ad_cb_lag - 1;
560 get_residual(residual, prev_excitation, lag);
562 /* Select quantization table */
563 if (cur_rate == RATE_6300 && pitch_lag < SUBFRAME_LEN - 2) {
564 cb_ptr = adaptive_cb_gain85;
566 cb_ptr = adaptive_cb_gain170;
568 /* Calculate adaptive vector */
569 cb_ptr += subfrm.ad_cb_gain * 20;
570 for (i = 0; i < SUBFRAME_LEN; i++) {
571 sum = ff_dot_product(residual + i, cb_ptr, PITCH_ORDER);
572 vector[i] = av_clipl_int32((sum << 2) + (1 << 15)) >> 16;
577 * Estimate maximum auto-correlation around pitch lag.
579 * @param p the context
580 * @param offset offset of the excitation vector
581 * @param ccr_max pointer to the maximum auto-correlation
582 * @param pitch_lag decoded pitch lag
583 * @param length length of autocorrelation
584 * @param dir forward lag(1) / backward lag(-1)
586 static int autocorr_max(G723_1_Context *p, int offset, int *ccr_max,
587 int pitch_lag, int length, int dir)
589 int limit, ccr, lag = 0;
590 int16_t *buf = p->excitation + offset;
593 pitch_lag = FFMIN(PITCH_MAX - 3, pitch_lag);
594 limit = FFMIN(FRAME_LEN + PITCH_MAX - offset - length, pitch_lag + 3);
596 for (i = pitch_lag - 3; i <= limit; i++) {
597 ccr = ff_dot_product(buf, buf + dir * i, length)<<1;
599 if (ccr > *ccr_max) {
608 * Calculate pitch postfilter optimal and scaling gains.
610 * @param lag pitch postfilter forward/backward lag
611 * @param ppf pitch postfilter parameters
612 * @param cur_rate current bitrate
613 * @param tgt_eng target energy
614 * @param ccr cross-correlation
615 * @param res_eng residual energy
617 static void comp_ppf_gains(int lag, PPFParam *ppf, enum Rate cur_rate,
618 int tgt_eng, int ccr, int res_eng)
620 int pf_residual; /* square of postfiltered residual */
621 int64_t temp1, temp2;
625 temp1 = tgt_eng * res_eng >> 1;
626 temp2 = ccr * ccr << 1;
629 if (ccr >= res_eng) {
630 ppf->opt_gain = ppf_gain_weight[cur_rate];
632 ppf->opt_gain = (ccr << 15) / res_eng *
633 ppf_gain_weight[cur_rate] >> 15;
635 /* pf_res^2 = tgt_eng + 2*ccr*gain + res_eng*gain^2 */
636 temp1 = (tgt_eng << 15) + (ccr * ppf->opt_gain << 1);
637 temp2 = (ppf->opt_gain * ppf->opt_gain >> 15) * res_eng;
638 pf_residual = av_clipl_int32(temp1 + temp2 + (1 << 15)) >> 16;
640 if (tgt_eng >= pf_residual << 1) {
643 temp1 = (tgt_eng << 14) / pf_residual;
646 /* scaling_gain = sqrt(tgt_eng/pf_res^2) */
647 ppf->sc_gain = square_root(temp1 << 16);
650 ppf->sc_gain = 0x7fff;
653 ppf->opt_gain = av_clip_int16(ppf->opt_gain * ppf->sc_gain >> 15);
657 * Calculate pitch postfilter parameters.
659 * @param p the context
660 * @param offset offset of the excitation vector
661 * @param pitch_lag decoded pitch lag
662 * @param ppf pitch postfilter parameters
663 * @param cur_rate current bitrate
665 static void comp_ppf_coeff(G723_1_Context *p, int offset, int pitch_lag,
666 PPFParam *ppf, enum Rate cur_rate)
671 int64_t temp1, temp2;
675 * 1 - forward cross-correlation
676 * 2 - forward residual energy
677 * 3 - backward cross-correlation
678 * 4 - backward residual energy
680 int energy[5] = {0, 0, 0, 0, 0};
681 int16_t *buf = p->excitation + offset;
682 int fwd_lag = autocorr_max(p, offset, &energy[1], pitch_lag,
684 int back_lag = autocorr_max(p, offset, &energy[3], pitch_lag,
689 ppf->sc_gain = 0x7fff;
691 /* Case 0, Section 3.6 */
692 if (!back_lag && !fwd_lag)
695 /* Compute target energy */
696 energy[0] = ff_dot_product(buf, buf, SUBFRAME_LEN)<<1;
698 /* Compute forward residual energy */
700 energy[2] = ff_dot_product(buf + fwd_lag, buf + fwd_lag,
703 /* Compute backward residual energy */
705 energy[4] = ff_dot_product(buf - back_lag, buf - back_lag,
708 /* Normalize and shorten */
710 for (i = 0; i < 5; i++)
711 temp1 = FFMAX(energy[i], temp1);
713 scale = normalize_bits(temp1, 1);
714 for (i = 0; i < 5; i++)
715 energy[i] = av_clipl_int32(energy[i] << scale) >> 16;
717 if (fwd_lag && !back_lag) { /* Case 1 */
718 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
720 } else if (!fwd_lag) { /* Case 2 */
721 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
723 } else { /* Case 3 */
726 * Select the largest of energy[1]^2/energy[2]
727 * and energy[3]^2/energy[4]
729 temp1 = energy[4] * ((energy[1] * energy[1] + (1 << 14)) >> 15);
730 temp2 = energy[2] * ((energy[3] * energy[3] + (1 << 14)) >> 15);
731 if (temp1 >= temp2) {
732 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
735 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
742 * Classify frames as voiced/unvoiced.
744 * @param p the context
745 * @param pitch_lag decoded pitch_lag
746 * @param exc_eng excitation energy estimation
747 * @param scale scaling factor of exc_eng
749 * @return residual interpolation index if voiced, 0 otherwise
751 static int comp_interp_index(G723_1_Context *p, int pitch_lag,
752 int *exc_eng, int *scale)
754 int offset = PITCH_MAX + 2 * SUBFRAME_LEN;
755 int16_t *buf = p->excitation + offset;
757 int index, ccr, tgt_eng, best_eng, temp;
759 *scale = scale_vector(p->excitation, FRAME_LEN + PITCH_MAX);
761 /* Compute maximum backward cross-correlation */
763 index = autocorr_max(p, offset, &ccr, pitch_lag, SUBFRAME_LEN * 2, -1);
764 ccr = av_clipl_int32((int64_t)ccr + (1 << 15)) >> 16;
766 /* Compute target energy */
767 tgt_eng = ff_dot_product(buf, buf, SUBFRAME_LEN * 2)<<1;
768 *exc_eng = av_clipl_int32(tgt_eng + (1 << 15)) >> 16;
773 /* Compute best energy */
774 best_eng = ff_dot_product(buf - index, buf - index,
775 SUBFRAME_LEN * 2)<<1;
776 best_eng = av_clipl_int32((int64_t)best_eng + (1 << 15)) >> 16;
778 temp = best_eng * *exc_eng >> 3;
780 if (temp < ccr * ccr) {
787 * Peform residual interpolation based on frame classification.
789 * @param buf decoded excitation vector
790 * @param out output vector
791 * @param lag decoded pitch lag
792 * @param gain interpolated gain
793 * @param rseed seed for random number generator
795 static void residual_interp(int16_t *buf, int16_t *out, int lag,
796 int gain, int *rseed)
799 if (lag) { /* Voiced */
800 int16_t *vector_ptr = buf + PITCH_MAX;
802 for (i = 0; i < lag; i++)
803 vector_ptr[i - lag] = vector_ptr[i - lag] * 3 >> 2;
804 av_memcpy_backptr((uint8_t*)vector_ptr, lag * sizeof(*vector_ptr),
805 FRAME_LEN * sizeof(*vector_ptr));
806 memcpy(out, vector_ptr, FRAME_LEN * sizeof(*vector_ptr));
807 } else { /* Unvoiced */
808 for (i = 0; i < FRAME_LEN; i++) {
809 *rseed = *rseed * 521 + 259;
810 out[i] = gain * *rseed >> 15;
812 memset(buf, 0, (FRAME_LEN + PITCH_MAX) * sizeof(*buf));
817 * Perform IIR filtering.
819 * @param fir_coef FIR coefficients
820 * @param iir_coef IIR coefficients
821 * @param src source vector
822 * @param dest destination vector
823 * @param width width of the output, 16 bits(0) / 32 bits(1)
825 #define iir_filter(fir_coef, iir_coef, src, dest, width)\
828 int res_shift = 16 & ~-(width);\
829 int in_shift = 16 - res_shift;\
831 for (m = 0; m < SUBFRAME_LEN; m++) {\
833 for (n = 1; n <= LPC_ORDER; n++) {\
834 filter -= (fir_coef)[n - 1] * (src)[m - n] -\
835 (iir_coef)[n - 1] * ((dest)[m - n] >> in_shift);\
838 (dest)[m] = av_clipl_int32(((src)[m] << 16) + (filter << 3) +\
839 (1 << 15)) >> res_shift;\
844 * Adjust gain of postfiltered signal.
846 * @param p the context
847 * @param buf postfiltered output vector
848 * @param energy input energy coefficient
850 static void gain_scale(G723_1_Context *p, int16_t * buf, int energy)
852 int num, denom, gain, bits1, bits2;
857 for (i = 0; i < SUBFRAME_LEN; i++) {
858 int64_t temp = buf[i] >> 2;
859 temp = av_clipl_int32(MUL64(temp, temp) << 1);
860 denom = av_clipl_int32(denom + temp);
864 bits1 = normalize_bits(num, 1);
865 bits2 = normalize_bits(denom, 1);
866 num = num << bits1 >> 1;
869 bits2 = 5 + bits1 - bits2;
870 bits2 = FFMAX(0, bits2);
872 gain = (num >> 1) / (denom >> 16);
873 gain = square_root(gain << 16 >> bits2);
878 for (i = 0; i < SUBFRAME_LEN; i++) {
879 p->pf_gain = ((p->pf_gain << 4) - p->pf_gain + gain + (1 << 3)) >> 4;
880 buf[i] = av_clip_int16((buf[i] * (p->pf_gain + (p->pf_gain >> 4)) +
886 * Perform formant filtering.
888 * @param p the context
889 * @param lpc quantized lpc coefficients
890 * @param buf output buffer
892 static void formant_postfilter(G723_1_Context *p, int16_t *lpc, int16_t *buf)
894 int16_t filter_coef[2][LPC_ORDER], *buf_ptr;
895 int filter_signal[LPC_ORDER + FRAME_LEN], *signal_ptr;
898 memcpy(buf, p->fir_mem, LPC_ORDER * sizeof(*buf));
899 memcpy(filter_signal, p->iir_mem, LPC_ORDER * sizeof(*filter_signal));
901 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
902 for (k = 0; k < LPC_ORDER; k++) {
903 filter_coef[0][k] = (-lpc[k] * postfilter_tbl[0][k] +
905 filter_coef[1][k] = (-lpc[k] * postfilter_tbl[1][k] +
908 iir_filter(filter_coef[0], filter_coef[1], buf + i,
909 filter_signal + i, 1);
912 memcpy(p->fir_mem, buf + FRAME_LEN, LPC_ORDER * sizeof(int16_t));
913 memcpy(p->iir_mem, filter_signal + FRAME_LEN, LPC_ORDER * sizeof(int));
915 buf_ptr = buf + LPC_ORDER;
916 signal_ptr = filter_signal + LPC_ORDER;
917 for (i = 0; i < SUBFRAMES; i++) {
918 int16_t temp_vector[SUBFRAME_LEN];
924 memcpy(temp_vector, buf_ptr, SUBFRAME_LEN * sizeof(int16_t));
925 scale = scale_vector(temp_vector, SUBFRAME_LEN);
927 /* Compute auto correlation coefficients */
928 auto_corr[0] = ff_dot_product(temp_vector, temp_vector + 1,
929 SUBFRAME_LEN - 1)<<1;
930 auto_corr[1] = ff_dot_product(temp_vector, temp_vector,
933 /* Compute reflection coefficient */
934 temp = auto_corr[1] >> 16;
936 temp = (auto_corr[0] >> 2) / temp;
938 p->reflection_coef = ((p->reflection_coef << 2) - p->reflection_coef +
940 temp = (p->reflection_coef * 0xffffc >> 3) & 0xfffc;
942 /* Compensation filter */
943 for (j = 0; j < SUBFRAME_LEN; j++) {
944 buf_ptr[j] = av_clipl_int32(signal_ptr[j] +
945 ((signal_ptr[j - 1] >> 16) *
949 /* Compute normalized signal energy */
950 temp = 2 * scale + 4;
952 energy = av_clipl_int32((int64_t)auto_corr[1] << -temp);
954 energy = auto_corr[1] >> temp;
956 gain_scale(p, buf_ptr, energy);
958 buf_ptr += SUBFRAME_LEN;
959 signal_ptr += SUBFRAME_LEN;
963 static int g723_1_decode_frame(AVCodecContext *avctx, void *data,
964 int *got_frame_ptr, AVPacket *avpkt)
966 G723_1_Context *p = avctx->priv_data;
967 const uint8_t *buf = avpkt->data;
968 int buf_size = avpkt->size;
970 int dec_mode = buf[0] & 3;
972 PPFParam ppf[SUBFRAMES];
973 int16_t cur_lsp[LPC_ORDER];
974 int16_t lpc[SUBFRAMES * LPC_ORDER];
975 int16_t acb_vector[SUBFRAME_LEN];
977 int bad_frame = 0, i, j, ret;
979 if (!buf_size || buf_size < frame_size[dec_mode]) {
984 if (unpack_bitstream(p, buf, buf_size) < 0) {
986 p->cur_frame_type = p->past_frame_type == ACTIVE_FRAME ?
987 ACTIVE_FRAME : UNTRANSMITTED_FRAME;
990 p->frame.nb_samples = FRAME_LEN + LPC_ORDER;
991 if ((ret = avctx->get_buffer(avctx, &p->frame)) < 0) {
992 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
995 out= (int16_t*)p->frame.data[0];
998 if(p->cur_frame_type == ACTIVE_FRAME) {
1000 p->erased_frames = 0;
1001 else if(p->erased_frames != 3)
1004 inverse_quant(cur_lsp, p->prev_lsp, p->lsp_index, bad_frame);
1005 lsp_interpolate(lpc, cur_lsp, p->prev_lsp);
1007 /* Save the lsp_vector for the next frame */
1008 memcpy(p->prev_lsp, cur_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
1010 /* Generate the excitation for the frame */
1011 memcpy(p->excitation, p->prev_excitation,
1012 PITCH_MAX * sizeof(*p->excitation));
1013 vector_ptr = p->excitation + PITCH_MAX;
1014 if (!p->erased_frames) {
1015 /* Update interpolation gain memory */
1016 p->interp_gain = fixed_cb_gain[(p->subframe[2].amp_index +
1017 p->subframe[3].amp_index) >> 1];
1018 for (i = 0; i < SUBFRAMES; i++) {
1019 gen_fcb_excitation(vector_ptr, p->subframe[i], p->cur_rate,
1020 p->pitch_lag[i >> 1], i);
1021 gen_acb_excitation(acb_vector, &p->excitation[SUBFRAME_LEN * i],
1022 p->pitch_lag[i >> 1], p->subframe[i],
1024 /* Get the total excitation */
1025 for (j = 0; j < SUBFRAME_LEN; j++) {
1026 vector_ptr[j] = av_clip_int16(vector_ptr[j] << 1);
1027 vector_ptr[j] = av_clip_int16(vector_ptr[j] +
1030 vector_ptr += SUBFRAME_LEN;
1033 vector_ptr = p->excitation + PITCH_MAX;
1035 /* Save the excitation */
1036 memcpy(out, vector_ptr, FRAME_LEN * sizeof(int16_t));
1038 p->interp_index = comp_interp_index(p, p->pitch_lag[1],
1039 &p->sid_gain, &p->cur_gain);
1041 for (i = PITCH_MAX, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1042 comp_ppf_coeff(p, i, p->pitch_lag[j >> 1],
1043 ppf + j, p->cur_rate);
1045 /* Restore the original excitation */
1046 memcpy(p->excitation, p->prev_excitation,
1047 PITCH_MAX * sizeof(int16_t));
1048 memcpy(vector_ptr, out, FRAME_LEN * sizeof(int16_t));
1050 /* Peform pitch postfiltering */
1051 for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1052 ff_acelp_weighted_vector_sum(out + LPC_ORDER + i, vector_ptr + i,
1053 vector_ptr + i + ppf[j].index,
1054 ppf[j].sc_gain, ppf[j].opt_gain,
1055 1 << 14, 15, SUBFRAME_LEN);
1057 p->interp_gain = (p->interp_gain * 3 + 2) >> 2;
1058 if (p->erased_frames == 3) {
1060 memset(p->excitation, 0,
1061 (FRAME_LEN + PITCH_MAX) * sizeof(int16_t));
1062 memset(out, 0, (FRAME_LEN + LPC_ORDER) * sizeof(int16_t));
1064 /* Regenerate frame */
1065 residual_interp(p->excitation, out + LPC_ORDER, p->interp_index,
1066 p->interp_gain, &p->random_seed);
1069 /* Save the excitation for the next frame */
1070 memcpy(p->prev_excitation, p->excitation + FRAME_LEN,
1071 PITCH_MAX * sizeof(int16_t));
1073 memset(out, 0, sizeof(int16_t)*FRAME_LEN);
1074 av_log(avctx, AV_LOG_WARNING,
1075 "G.723.1: Comfort noise generation not supported yet\n");
1076 return frame_size[dec_mode];
1079 p->past_frame_type = p->cur_frame_type;
1081 memcpy(out, p->synth_mem, LPC_ORDER * sizeof(int16_t));
1082 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1083 ff_celp_lp_synthesis_filter(out + i, &lpc[j * LPC_ORDER],
1084 out + i, SUBFRAME_LEN, LPC_ORDER,
1086 memcpy(p->synth_mem, out + FRAME_LEN, LPC_ORDER * sizeof(int16_t));
1088 formant_postfilter(p, lpc, out);
1090 memmove(out, out + LPC_ORDER, sizeof(int16_t)*FRAME_LEN);
1091 p->frame.nb_samples = FRAME_LEN;
1092 *(AVFrame*)data = p->frame;
1095 return frame_size[dec_mode];
1098 AVCodec ff_g723_1_decoder = {
1100 .type = AVMEDIA_TYPE_AUDIO,
1101 .id = CODEC_ID_G723_1,
1102 .priv_data_size = sizeof(G723_1_Context),
1103 .init = g723_1_decode_init,
1104 .decode = g723_1_decode_frame,
1105 .long_name = NULL_IF_CONFIG_SMALL("G.723.1"),
1106 .capabilities = CODEC_CAP_SUBFRAMES | CODEC_CAP_DR1,
1109 #if CONFIG_G723_1_ENCODER
1110 #define BITSTREAM_WRITER_LE
1111 #include "put_bits.h"
1113 static av_cold int g723_1_encode_init(AVCodecContext *avctx)
1115 G723_1_Context *p = avctx->priv_data;
1117 if (avctx->sample_rate != 8000) {
1118 av_log(avctx, AV_LOG_ERROR, "Only 8000Hz sample rate supported\n");
1122 if (avctx->channels != 1) {
1123 av_log(avctx, AV_LOG_ERROR, "Only mono supported\n");
1124 return AVERROR(EINVAL);
1127 if (avctx->bit_rate == 6300) {
1128 p->cur_rate = RATE_6300;
1129 } else if (avctx->bit_rate == 5300) {
1130 av_log(avctx, AV_LOG_ERROR, "Bitrate not supported yet, use 6.3k\n");
1131 return AVERROR_PATCHWELCOME;
1133 av_log(avctx, AV_LOG_ERROR,
1134 "Bitrate not supported, use 6.3k\n");
1135 return AVERROR(EINVAL);
1137 avctx->frame_size = 240;
1138 memcpy(p->prev_lsp, dc_lsp, LPC_ORDER * sizeof(int16_t));
1144 * Remove DC component from the input signal.
1146 * @param buf input signal
1147 * @param fir zero memory
1148 * @param iir pole memory
1150 static void highpass_filter(int16_t *buf, int16_t *fir, int *iir)
1153 for (i = 0; i < FRAME_LEN; i++) {
1154 *iir = (buf[i] << 15) + ((-*fir) << 15) + MULL2(*iir, 0x7f00);
1156 buf[i] = av_clipl_int32((int64_t)*iir + (1 << 15)) >> 16;
1161 * Estimate autocorrelation of the input vector.
1163 * @param buf input buffer
1164 * @param autocorr autocorrelation coefficients vector
1166 static void comp_autocorr(int16_t *buf, int16_t *autocorr)
1169 int16_t vector[LPC_FRAME];
1171 memcpy(vector, buf, LPC_FRAME * sizeof(int16_t));
1172 scale_vector(vector, LPC_FRAME);
1174 /* Apply the Hamming window */
1175 for (i = 0; i < LPC_FRAME; i++)
1176 vector[i] = (vector[i] * hamming_window[i] + (1 << 14)) >> 15;
1178 /* Compute the first autocorrelation coefficient */
1179 temp = dot_product(vector, vector, LPC_FRAME, 0);
1181 /* Apply a white noise correlation factor of (1025/1024) */
1185 scale = normalize_bits_int32(temp);
1186 autocorr[0] = av_clipl_int32((int64_t)(temp << scale) +
1189 /* Compute the remaining coefficients */
1191 memset(autocorr + 1, 0, LPC_ORDER * sizeof(int16_t));
1193 for (i = 1; i <= LPC_ORDER; i++) {
1194 temp = dot_product(vector, vector + i, LPC_FRAME - i, 0);
1195 temp = MULL2((temp << scale), binomial_window[i - 1]);
1196 autocorr[i] = av_clipl_int32((int64_t)temp + (1 << 15)) >> 16;
1202 * Use Levinson-Durbin recursion to compute LPC coefficients from
1203 * autocorrelation values.
1205 * @param lpc LPC coefficients vector
1206 * @param autocorr autocorrelation coefficients vector
1207 * @param error prediction error
1209 static void levinson_durbin(int16_t *lpc, int16_t *autocorr, int16_t error)
1211 int16_t vector[LPC_ORDER];
1212 int16_t partial_corr;
1215 memset(lpc, 0, LPC_ORDER * sizeof(int16_t));
1217 for (i = 0; i < LPC_ORDER; i++) {
1218 /* Compute the partial correlation coefficient */
1220 for (j = 0; j < i; j++)
1221 temp -= lpc[j] * autocorr[i - j - 1];
1222 temp = ((autocorr[i] << 13) + temp) << 3;
1224 if (FFABS(temp) >= (error << 16))
1227 partial_corr = temp / (error << 1);
1229 lpc[i] = av_clipl_int32((int64_t)(partial_corr << 14) +
1232 /* Update the prediction error */
1233 temp = MULL2(temp, partial_corr);
1234 error = av_clipl_int32((int64_t)(error << 16) - temp +
1237 memcpy(vector, lpc, i * sizeof(int16_t));
1238 for (j = 0; j < i; j++) {
1239 temp = partial_corr * vector[i - j - 1] << 1;
1240 lpc[j] = av_clipl_int32((int64_t)(lpc[j] << 16) - temp +
1247 * Calculate LPC coefficients for the current frame.
1249 * @param buf current frame
1250 * @param prev_data 2 trailing subframes of the previous frame
1251 * @param lpc LPC coefficients vector
1253 static void comp_lpc_coeff(int16_t *buf, int16_t *lpc)
1255 int16_t autocorr[(LPC_ORDER + 1) * SUBFRAMES];
1256 int16_t *autocorr_ptr = autocorr;
1257 int16_t *lpc_ptr = lpc;
1260 for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
1261 comp_autocorr(buf + i, autocorr_ptr);
1262 levinson_durbin(lpc_ptr, autocorr_ptr + 1, autocorr_ptr[0]);
1264 lpc_ptr += LPC_ORDER;
1265 autocorr_ptr += LPC_ORDER + 1;
1269 static void lpc2lsp(int16_t *lpc, int16_t *prev_lsp, int16_t *lsp)
1271 int f[LPC_ORDER + 2]; ///< coefficients of the sum and difference
1272 ///< polynomials (F1, F2) ordered as
1273 ///< f1[0], f2[0], ...., f1[5], f2[5]
1275 int max, shift, cur_val, prev_val, count, p;
1279 /* Initialize f1[0] and f2[0] to 1 in Q25 */
1280 for (i = 0; i < LPC_ORDER; i++)
1281 lsp[i] = (lpc[i] * bandwidth_expand[i] + (1 << 14)) >> 15;
1283 /* Apply bandwidth expansion on the LPC coefficients */
1284 f[0] = f[1] = 1 << 25;
1286 /* Compute the remaining coefficients */
1287 for (i = 0; i < LPC_ORDER / 2; i++) {
1289 f[2 * i + 2] = -f[2 * i] - ((lsp[i] + lsp[LPC_ORDER - 1 - i]) << 12);
1291 f[2 * i + 3] = f[2 * i + 1] - ((lsp[i] - lsp[LPC_ORDER - 1 - i]) << 12);
1294 /* Divide f1[5] and f2[5] by 2 for use in polynomial evaluation */
1296 f[LPC_ORDER + 1] >>= 1;
1298 /* Normalize and shorten */
1300 for (i = 1; i < LPC_ORDER + 2; i++)
1301 max = FFMAX(max, FFABS(f[i]));
1303 shift = normalize_bits_int32(max);
1305 for (i = 0; i < LPC_ORDER + 2; i++)
1306 f[i] = av_clipl_int32((int64_t)(f[i] << shift) + (1 << 15)) >> 16;
1309 * Evaluate F1 and F2 at uniform intervals of pi/256 along the
1310 * unit circle and check for zero crossings.
1314 for (i = 0; i <= LPC_ORDER / 2; i++)
1315 temp += f[2 * i] * cos_tab[0];
1316 prev_val = av_clipl_int32(temp << 1);
1318 for ( i = 1; i < COS_TBL_SIZE / 2; i++) {
1321 for (j = 0; j <= LPC_ORDER / 2; j++)
1322 temp += f[LPC_ORDER - 2 * j + p] * cos_tab[i * j % COS_TBL_SIZE];
1323 cur_val = av_clipl_int32(temp << 1);
1325 /* Check for sign change, indicating a zero crossing */
1326 if ((cur_val ^ prev_val) < 0) {
1327 int abs_cur = FFABS(cur_val);
1328 int abs_prev = FFABS(prev_val);
1329 int sum = abs_cur + abs_prev;
1331 shift = normalize_bits_int32(sum);
1333 abs_prev = abs_prev << shift >> 8;
1334 lsp[count++] = ((i - 1) << 7) + (abs_prev >> 1) / (sum >> 16);
1336 if (count == LPC_ORDER)
1339 /* Switch between sum and difference polynomials */
1344 for (j = 0; j <= LPC_ORDER / 2; j++){
1345 temp += f[LPC_ORDER - 2 * j + p] *
1346 cos_tab[i * j % COS_TBL_SIZE];
1348 cur_val = av_clipl_int32(temp<<1);
1353 if (count != LPC_ORDER)
1354 memcpy(lsp, prev_lsp, LPC_ORDER * sizeof(int16_t));
1358 * Quantize the current LSP subvector.
1360 * @param num band number
1361 * @param offset offset of the current subvector in an LPC_ORDER vector
1362 * @param size size of the current subvector
1364 #define get_index(num, offset, size) \
1366 int error, max = -1;\
1369 for (i = 0; i < LSP_CB_SIZE; i++) {\
1370 for (j = 0; j < size; j++){\
1371 temp[j] = (weight[j + (offset)] * lsp_band##num[i][j] +\
1374 error = dot_product(lsp + (offset), temp, size, 1) << 1;\
1375 error -= dot_product(lsp_band##num[i], temp, size, 1);\
1378 lsp_index[num] = i;\
1384 * Vector quantize the LSP frequencies.
1386 * @param lsp the current lsp vector
1387 * @param prev_lsp the previous lsp vector
1389 static void lsp_quantize(uint8_t *lsp_index, int16_t *lsp, int16_t *prev_lsp)
1391 int16_t weight[LPC_ORDER];
1395 /* Calculate the VQ weighting vector */
1396 weight[0] = (1 << 20) / (lsp[1] - lsp[0]);
1397 weight[LPC_ORDER - 1] = (1 << 20) /
1398 (lsp[LPC_ORDER - 1] - lsp[LPC_ORDER - 2]);
1400 for (i = 1; i < LPC_ORDER - 1; i++) {
1401 min = FFMIN(lsp[i] - lsp[i - 1], lsp[i + 1] - lsp[i]);
1403 weight[i] = (1 << 20) / min;
1405 weight[i] = INT16_MAX;
1410 for (i = 0; i < LPC_ORDER; i++)
1411 max = FFMAX(weight[i], max);
1413 shift = normalize_bits_int16(max);
1414 for (i = 0; i < LPC_ORDER; i++) {
1415 weight[i] <<= shift;
1418 /* Compute the VQ target vector */
1419 for (i = 0; i < LPC_ORDER; i++) {
1420 lsp[i] -= dc_lsp[i] +
1421 (((prev_lsp[i] - dc_lsp[i]) * 12288 + (1 << 14)) >> 15);
1430 * Apply the formant perceptual weighting filter.
1432 * @param flt_coef filter coefficients
1433 * @param unq_lpc unquantized lpc vector
1435 static void perceptual_filter(G723_1_Context *p, int16_t *flt_coef,
1436 int16_t *unq_lpc, int16_t *buf)
1438 int16_t vector[FRAME_LEN + LPC_ORDER];
1441 memcpy(buf, p->iir_mem, sizeof(int16_t) * LPC_ORDER);
1442 memcpy(vector, p->fir_mem, sizeof(int16_t) * LPC_ORDER);
1443 memcpy(vector + LPC_ORDER, buf + LPC_ORDER, sizeof(int16_t) * FRAME_LEN);
1445 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
1446 for (k = 0; k < LPC_ORDER; k++) {
1447 flt_coef[k + 2 * l] = (unq_lpc[k + l] * percept_flt_tbl[0][k] +
1449 flt_coef[k + 2 * l + LPC_ORDER] = (unq_lpc[k + l] *
1450 percept_flt_tbl[1][k] +
1453 iir_filter(flt_coef + 2 * l, flt_coef + 2 * l + LPC_ORDER, vector + i,
1457 memcpy(p->iir_mem, buf + FRAME_LEN, sizeof(int16_t) * LPC_ORDER);
1458 memcpy(p->fir_mem, vector + FRAME_LEN, sizeof(int16_t) * LPC_ORDER);
1462 * Estimate the open loop pitch period.
1464 * @param buf perceptually weighted speech
1465 * @param start estimation is carried out from this position
1467 static int estimate_pitch(int16_t *buf, int start)
1470 int max_ccr = 0x4000;
1471 int max_eng = 0x7fff;
1472 int index = PITCH_MIN;
1473 int offset = start - PITCH_MIN + 1;
1475 int ccr, eng, orig_eng, ccr_eng, exp;
1480 orig_eng = dot_product(buf + offset, buf + offset, HALF_FRAME_LEN, 0);
1482 for (i = PITCH_MIN; i <= PITCH_MAX - 3; i++) {
1485 /* Update energy and compute correlation */
1486 orig_eng += buf[offset] * buf[offset] -
1487 buf[offset + HALF_FRAME_LEN] * buf[offset + HALF_FRAME_LEN];
1488 ccr = dot_product(buf + start, buf + offset, HALF_FRAME_LEN, 0);
1492 /* Split into mantissa and exponent to maintain precision */
1493 exp = normalize_bits_int32(ccr);
1494 ccr = av_clipl_int32((int64_t)(ccr << exp) + (1 << 15)) >> 16;
1497 temp = normalize_bits_int32(ccr);
1498 ccr = ccr << temp >> 16;
1501 temp = normalize_bits_int32(orig_eng);
1502 eng = av_clipl_int32((int64_t)(orig_eng << temp) + (1 << 15)) >> 16;
1512 if (exp + 1 < max_exp)
1515 /* Equalize exponents before comparison */
1516 if (exp + 1 == max_exp)
1517 temp = max_ccr >> 1;
1520 ccr_eng = ccr * max_eng;
1521 diff = ccr_eng - eng * temp;
1522 if (diff > 0 && (i - index < PITCH_MIN || diff > ccr_eng >> 2)) {
1534 * Compute harmonic noise filter parameters.
1536 * @param buf perceptually weighted speech
1537 * @param pitch_lag open loop pitch period
1538 * @param hf harmonic filter parameters
1540 static void comp_harmonic_coeff(int16_t *buf, int16_t pitch_lag, HFParam *hf)
1542 int ccr, eng, max_ccr, max_eng;
1547 for (i = 0, j = pitch_lag - 3; j <= pitch_lag + 3; i++, j++) {
1548 /* Compute residual energy */
1549 energy[i << 1] = dot_product(buf - j, buf - j, SUBFRAME_LEN, 0);
1550 /* Compute correlation */
1551 energy[(i << 1) + 1] = dot_product(buf, buf - j, SUBFRAME_LEN, 0);
1554 /* Compute target energy */
1555 energy[14] = dot_product(buf, buf, SUBFRAME_LEN, 0);
1559 for (i = 0; i < 15; i++)
1560 max = FFMAX(max, FFABS(energy[i]));
1562 exp = normalize_bits_int32(max);
1563 for (i = 0; i < 15; i++) {
1564 energy[i] = av_clipl_int32((int64_t)(energy[i] << exp) +
1573 for (i = 0; i <= 6; i++) {
1574 eng = energy[i << 1];
1575 ccr = energy[(i << 1) + 1];
1580 ccr = (ccr * ccr + (1 << 14)) >> 15;
1581 diff = ccr * max_eng - eng * max_ccr;
1589 if (hf->index == -1) {
1590 hf->index = pitch_lag;
1594 eng = energy[14] * max_eng;
1595 eng = (eng >> 2) + (eng >> 3);
1596 ccr = energy[(hf->index << 1) + 1] * energy[(hf->index << 1) + 1];
1598 eng = energy[(hf->index << 1) + 1];
1603 hf->gain = ((eng << 15) / max_eng * 0x2800 + (1 << 14)) >> 15;
1605 hf->index += pitch_lag - 3;
1609 * Apply the harmonic noise shaping filter.
1611 * @param hf filter parameters
1613 static void harmonic_filter(HFParam *hf, int16_t *src, int16_t *dest)
1617 for (i = 0; i < SUBFRAME_LEN; i++) {
1618 int64_t temp = hf->gain * src[i - hf->index] << 1;
1619 dest[i] = av_clipl_int32((src[i] << 16) - temp + (1 << 15)) >> 16;
1623 static void harmonic_noise_sub(HFParam *hf, int16_t *src, int16_t *dest)
1626 for (i = 0; i < SUBFRAME_LEN; i++) {
1627 int64_t temp = hf->gain * src[i - hf->index] << 1;
1628 dest[i] = av_clipl_int32(((dest[i] - src[i]) << 16) + temp +
1635 * Combined synthesis and formant perceptual weighting filer.
1637 * @param qnt_lpc quantized lpc coefficients
1638 * @param perf_lpc perceptual filter coefficients
1639 * @param perf_fir perceptual filter fir memory
1640 * @param perf_iir perceptual filter iir memory
1641 * @param scale the filter output will be scaled by 2^scale
1643 static void synth_percept_filter(int16_t *qnt_lpc, int16_t *perf_lpc,
1644 int16_t *perf_fir, int16_t *perf_iir,
1645 int16_t *src, int16_t *dest, int scale)
1648 int16_t buf_16[SUBFRAME_LEN + LPC_ORDER];
1649 int64_t buf[SUBFRAME_LEN];
1651 int16_t *bptr_16 = buf_16 + LPC_ORDER;
1653 memcpy(buf_16, perf_fir, sizeof(int16_t) * LPC_ORDER);
1654 memcpy(dest - LPC_ORDER, perf_iir, sizeof(int16_t) * LPC_ORDER);
1656 for (i = 0; i < SUBFRAME_LEN; i++) {
1658 for (j = 1; j <= LPC_ORDER; j++)
1659 temp -= qnt_lpc[j - 1] * bptr_16[i - j];
1661 buf[i] = (src[i] << 15) + (temp << 3);
1662 bptr_16[i] = av_clipl_int32(buf[i] + (1 << 15)) >> 16;
1665 for (i = 0; i < SUBFRAME_LEN; i++) {
1666 int64_t fir = 0, iir = 0;
1667 for (j = 1; j <= LPC_ORDER; j++) {
1668 fir -= perf_lpc[j - 1] * bptr_16[i - j];
1669 iir += perf_lpc[j + LPC_ORDER - 1] * dest[i - j];
1671 dest[i] = av_clipl_int32(((buf[i] + (fir << 3)) << scale) + (iir << 3) +
1674 memcpy(perf_fir, buf_16 + SUBFRAME_LEN, sizeof(int16_t) * LPC_ORDER);
1675 memcpy(perf_iir, dest + SUBFRAME_LEN - LPC_ORDER,
1676 sizeof(int16_t) * LPC_ORDER);
1680 * Compute the adaptive codebook contribution.
1682 * @param buf input signal
1683 * @param index the current subframe index
1685 static void acb_search(G723_1_Context *p, int16_t *residual,
1686 int16_t *impulse_resp, int16_t *buf,
1690 int16_t flt_buf[PITCH_ORDER][SUBFRAME_LEN];
1692 const int16_t *cb_tbl = adaptive_cb_gain85;
1694 int ccr_buf[PITCH_ORDER * SUBFRAMES << 2];
1696 int pitch_lag = p->pitch_lag[index >> 1];
1699 int odd_frame = index & 1;
1700 int iter = 3 + odd_frame;
1704 int i, j, k, l, max;
1708 if (pitch_lag == PITCH_MIN)
1711 pitch_lag = FFMIN(pitch_lag, PITCH_MAX - 5);
1714 for (i = 0; i < iter; i++) {
1715 get_residual(residual, p->prev_excitation, pitch_lag + i - 1);
1717 for (j = 0; j < SUBFRAME_LEN; j++) {
1719 for (k = 0; k <= j; k++)
1720 temp += residual[PITCH_ORDER - 1 + k] * impulse_resp[j - k];
1721 flt_buf[PITCH_ORDER - 1][j] = av_clipl_int32((temp << 1) +
1725 for (j = PITCH_ORDER - 2; j >= 0; j--) {
1726 flt_buf[j][0] = ((residual[j] << 13) + (1 << 14)) >> 15;
1727 for (k = 1; k < SUBFRAME_LEN; k++) {
1728 temp = (flt_buf[j + 1][k - 1] << 15) +
1729 residual[j] * impulse_resp[k];
1730 flt_buf[j][k] = av_clipl_int32((temp << 1) + (1 << 15)) >> 16;
1734 /* Compute crosscorrelation with the signal */
1735 for (j = 0; j < PITCH_ORDER; j++) {
1736 temp = dot_product(buf, flt_buf[j], SUBFRAME_LEN, 0);
1737 ccr_buf[count++] = av_clipl_int32(temp << 1);
1740 /* Compute energies */
1741 for (j = 0; j < PITCH_ORDER; j++) {
1742 ccr_buf[count++] = dot_product(flt_buf[j], flt_buf[j],
1746 for (j = 1; j < PITCH_ORDER; j++) {
1747 for (k = 0; k < j; k++) {
1748 temp = dot_product(flt_buf[j], flt_buf[k], SUBFRAME_LEN, 0);
1749 ccr_buf[count++] = av_clipl_int32(temp<<2);
1754 /* Normalize and shorten */
1756 for (i = 0; i < 20 * iter; i++)
1757 max = FFMAX(max, FFABS(ccr_buf[i]));
1759 temp = normalize_bits_int32(max);
1761 for (i = 0; i < 20 * iter; i++){
1762 ccr_buf[i] = av_clipl_int32((int64_t)(ccr_buf[i] << temp) +
1767 for (i = 0; i < iter; i++) {
1768 /* Select quantization table */
1769 if (!odd_frame && pitch_lag + i - 1 >= SUBFRAME_LEN - 2 ||
1770 odd_frame && pitch_lag >= SUBFRAME_LEN - 2) {
1771 cb_tbl = adaptive_cb_gain170;
1775 for (j = 0, k = 0; j < tbl_size; j++, k += 20) {
1777 for (l = 0; l < 20; l++)
1778 temp += ccr_buf[20 * i + l] * cb_tbl[k + l];
1779 temp = av_clipl_int32(temp);
1790 pitch_lag += acb_lag - 1;
1794 p->pitch_lag[index >> 1] = pitch_lag;
1795 p->subframe[index].ad_cb_lag = acb_lag;
1796 p->subframe[index].ad_cb_gain = acb_gain;
1800 * Subtract the adaptive codebook contribution from the input
1801 * to obtain the residual.
1803 * @param buf target vector
1805 static void sub_acb_contrib(int16_t *residual, int16_t *impulse_resp,
1809 /* Subtract adaptive CB contribution to obtain the residual */
1810 for (i = 0; i < SUBFRAME_LEN; i++) {
1811 int64_t temp = buf[i] << 14;
1812 for (j = 0; j <= i; j++)
1813 temp -= residual[j] * impulse_resp[i - j];
1815 buf[i] = av_clipl_int32((temp << 2) + (1 << 15)) >> 16;
1820 * Quantize the residual signal using the fixed codebook (MP-MLQ).
1822 * @param optim optimized fixed codebook parameters
1823 * @param buf excitation vector
1825 static void get_fcb_param(FCBParam *optim, int16_t *impulse_resp,
1826 int16_t *buf, int pulse_cnt, int pitch_lag)
1829 int16_t impulse_r[SUBFRAME_LEN];
1830 int16_t temp_corr[SUBFRAME_LEN];
1831 int16_t impulse_corr[SUBFRAME_LEN];
1833 int ccr1[SUBFRAME_LEN];
1834 int ccr2[SUBFRAME_LEN];
1835 int amp, err, max, max_amp_index, min, scale, i, j, k, l;
1839 /* Update impulse response */
1840 memcpy(impulse_r, impulse_resp, sizeof(int16_t) * SUBFRAME_LEN);
1841 param.dirac_train = 0;
1842 if (pitch_lag < SUBFRAME_LEN - 2) {
1843 param.dirac_train = 1;
1844 gen_dirac_train(impulse_r, pitch_lag);
1847 for (i = 0; i < SUBFRAME_LEN; i++)
1848 temp_corr[i] = impulse_r[i] >> 1;
1850 /* Compute impulse response autocorrelation */
1851 temp = dot_product(temp_corr, temp_corr, SUBFRAME_LEN, 1);
1853 scale = normalize_bits_int32(temp);
1854 impulse_corr[0] = av_clipl_int32((temp << scale) + (1 << 15)) >> 16;
1856 for (i = 1; i < SUBFRAME_LEN; i++) {
1857 temp = dot_product(temp_corr + i, temp_corr, SUBFRAME_LEN - i, 1);
1858 impulse_corr[i] = av_clipl_int32((temp << scale) + (1 << 15)) >> 16;
1861 /* Compute crosscorrelation of impulse response with residual signal */
1863 for (i = 0; i < SUBFRAME_LEN; i++){
1864 temp = dot_product(buf + i, impulse_r, SUBFRAME_LEN - i, 1);
1866 ccr1[i] = temp >> -scale;
1868 ccr1[i] = av_clipl_int32(temp << scale);
1872 for (i = 0; i < GRID_SIZE; i++) {
1873 /* Maximize the crosscorrelation */
1875 for (j = i; j < SUBFRAME_LEN; j += GRID_SIZE) {
1876 temp = FFABS(ccr1[j]);
1879 param.pulse_pos[0] = j;
1883 /* Quantize the gain (max crosscorrelation/impulse_corr[0]) */
1886 max_amp_index = GAIN_LEVELS - 2;
1887 for (j = max_amp_index; j >= 2; j--) {
1888 temp = av_clipl_int32((int64_t)fixed_cb_gain[j] *
1889 impulse_corr[0] << 1);
1890 temp = FFABS(temp - amp);
1898 /* Select additional gain values */
1899 for (j = 1; j < 5; j++) {
1900 for (k = i; k < SUBFRAME_LEN; k += GRID_SIZE) {
1904 param.amp_index = max_amp_index + j - 2;
1905 amp = fixed_cb_gain[param.amp_index];
1907 param.pulse_sign[0] = (ccr2[param.pulse_pos[0]] < 0) ? -amp : amp;
1908 temp_corr[param.pulse_pos[0]] = 1;
1910 for (k = 1; k < pulse_cnt; k++) {
1912 for (l = i; l < SUBFRAME_LEN; l += GRID_SIZE) {
1915 temp = impulse_corr[FFABS(l - param.pulse_pos[k - 1])];
1916 temp = av_clipl_int32((int64_t)temp *
1917 param.pulse_sign[k - 1] << 1);
1919 temp = FFABS(ccr2[l]);
1922 param.pulse_pos[k] = l;
1926 param.pulse_sign[k] = (ccr2[param.pulse_pos[k]] < 0) ?
1928 temp_corr[param.pulse_pos[k]] = 1;
1931 /* Create the error vector */
1932 memset(temp_corr, 0, sizeof(int16_t) * SUBFRAME_LEN);
1934 for (k = 0; k < pulse_cnt; k++)
1935 temp_corr[param.pulse_pos[k]] = param.pulse_sign[k];
1937 for (k = SUBFRAME_LEN - 1; k >= 0; k--) {
1939 for (l = 0; l <= k; l++) {
1940 int prod = av_clipl_int32((int64_t)temp_corr[l] *
1941 impulse_r[k - l] << 1);
1942 temp = av_clipl_int32(temp + prod);
1944 temp_corr[k] = temp << 2 >> 16;
1947 /* Compute square of error */
1949 for (k = 0; k < SUBFRAME_LEN; k++) {
1951 prod = av_clipl_int32((int64_t)buf[k] * temp_corr[k] << 1);
1952 err = av_clipl_int32(err - prod);
1953 prod = av_clipl_int32((int64_t)temp_corr[k] * temp_corr[k]);
1954 err = av_clipl_int32(err + prod);
1958 if (err < optim->min_err) {
1959 optim->min_err = err;
1960 optim->grid_index = i;
1961 optim->amp_index = param.amp_index;
1962 optim->dirac_train = param.dirac_train;
1964 for (k = 0; k < pulse_cnt; k++) {
1965 optim->pulse_sign[k] = param.pulse_sign[k];
1966 optim->pulse_pos[k] = param.pulse_pos[k];
1974 * Encode the pulse position and gain of the current subframe.
1976 * @param optim optimized fixed CB parameters
1977 * @param buf excitation vector
1979 static void pack_fcb_param(G723_1_Subframe *subfrm, FCBParam *optim,
1980 int16_t *buf, int pulse_cnt)
1984 j = PULSE_MAX - pulse_cnt;
1986 subfrm->pulse_sign = 0;
1987 subfrm->pulse_pos = 0;
1989 for (i = 0; i < SUBFRAME_LEN >> 1; i++) {
1990 int val = buf[optim->grid_index + (i << 1)];
1992 subfrm->pulse_pos += combinatorial_table[j][i];
1994 subfrm->pulse_sign <<= 1;
1995 if (val < 0) subfrm->pulse_sign++;
1998 if (j == PULSE_MAX) break;
2001 subfrm->amp_index = optim->amp_index;
2002 subfrm->grid_index = optim->grid_index;
2003 subfrm->dirac_train = optim->dirac_train;
2007 * Compute the fixed codebook excitation.
2009 * @param buf target vector
2010 * @param impulse_resp impulse response of the combined filter
2012 static void fcb_search(G723_1_Context *p, int16_t *impulse_resp,
2013 int16_t *buf, int index)
2016 int pulse_cnt = pulses[index];
2019 optim.min_err = 1 << 30;
2020 get_fcb_param(&optim, impulse_resp, buf, pulse_cnt, SUBFRAME_LEN);
2022 if (p->pitch_lag[index >> 1] < SUBFRAME_LEN - 2) {
2023 get_fcb_param(&optim, impulse_resp, buf, pulse_cnt,
2024 p->pitch_lag[index >> 1]);
2027 /* Reconstruct the excitation */
2028 memset(buf, 0, sizeof(int16_t) * SUBFRAME_LEN);
2029 for (i = 0; i < pulse_cnt; i++)
2030 buf[optim.pulse_pos[i]] = optim.pulse_sign[i];
2032 pack_fcb_param(&p->subframe[index], &optim, buf, pulse_cnt);
2034 if (optim.dirac_train)
2035 gen_dirac_train(buf, p->pitch_lag[index >> 1]);
2039 * Pack the frame parameters into output bitstream.
2041 * @param frame output buffer
2042 * @param size size of the buffer
2044 static int pack_bitstream(G723_1_Context *p, unsigned char *frame, int size)
2047 int info_bits, i, temp;
2049 init_put_bits(&pb, frame, size);
2051 if (p->cur_rate == RATE_6300) {
2053 put_bits(&pb, 2, info_bits);
2056 put_bits(&pb, 8, p->lsp_index[2]);
2057 put_bits(&pb, 8, p->lsp_index[1]);
2058 put_bits(&pb, 8, p->lsp_index[0]);
2060 put_bits(&pb, 7, p->pitch_lag[0] - PITCH_MIN);
2061 put_bits(&pb, 2, p->subframe[1].ad_cb_lag);
2062 put_bits(&pb, 7, p->pitch_lag[1] - PITCH_MIN);
2063 put_bits(&pb, 2, p->subframe[3].ad_cb_lag);
2065 /* Write 12 bit combined gain */
2066 for (i = 0; i < SUBFRAMES; i++) {
2067 temp = p->subframe[i].ad_cb_gain * GAIN_LEVELS +
2068 p->subframe[i].amp_index;
2069 if (p->cur_rate == RATE_6300)
2070 temp += p->subframe[i].dirac_train << 11;
2071 put_bits(&pb, 12, temp);
2074 put_bits(&pb, 1, p->subframe[0].grid_index);
2075 put_bits(&pb, 1, p->subframe[1].grid_index);
2076 put_bits(&pb, 1, p->subframe[2].grid_index);
2077 put_bits(&pb, 1, p->subframe[3].grid_index);
2079 if (p->cur_rate == RATE_6300) {
2080 skip_put_bits(&pb, 1); /* reserved bit */
2082 /* Write 13 bit combined position index */
2083 temp = (p->subframe[0].pulse_pos >> 16) * 810 +
2084 (p->subframe[1].pulse_pos >> 14) * 90 +
2085 (p->subframe[2].pulse_pos >> 16) * 9 +
2086 (p->subframe[3].pulse_pos >> 14);
2087 put_bits(&pb, 13, temp);
2089 put_bits(&pb, 16, p->subframe[0].pulse_pos & 0xffff);
2090 put_bits(&pb, 14, p->subframe[1].pulse_pos & 0x3fff);
2091 put_bits(&pb, 16, p->subframe[2].pulse_pos & 0xffff);
2092 put_bits(&pb, 14, p->subframe[3].pulse_pos & 0x3fff);
2094 put_bits(&pb, 6, p->subframe[0].pulse_sign);
2095 put_bits(&pb, 5, p->subframe[1].pulse_sign);
2096 put_bits(&pb, 6, p->subframe[2].pulse_sign);
2097 put_bits(&pb, 5, p->subframe[3].pulse_sign);
2100 flush_put_bits(&pb);
2101 return frame_size[info_bits];
2104 static int g723_1_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
2105 const AVFrame *frame, int *got_packet_ptr)
2107 G723_1_Context *p = avctx->priv_data;
2108 int16_t unq_lpc[LPC_ORDER * SUBFRAMES];
2109 int16_t qnt_lpc[LPC_ORDER * SUBFRAMES];
2110 int16_t cur_lsp[LPC_ORDER];
2111 int16_t weighted_lpc[LPC_ORDER * SUBFRAMES << 1];
2112 int16_t vector[FRAME_LEN + PITCH_MAX];
2114 int16_t *in = (const int16_t *)frame->data[0];
2119 highpass_filter(in, &p->hpf_fir_mem, &p->hpf_iir_mem);
2121 memcpy(vector, p->prev_data, HALF_FRAME_LEN * sizeof(int16_t));
2122 memcpy(vector + HALF_FRAME_LEN, in, FRAME_LEN * sizeof(int16_t));
2124 comp_lpc_coeff(vector, unq_lpc);
2125 lpc2lsp(&unq_lpc[LPC_ORDER * 3], p->prev_lsp, cur_lsp);
2126 lsp_quantize(p->lsp_index, cur_lsp, p->prev_lsp);
2129 memcpy(vector + LPC_ORDER, p->prev_data + SUBFRAME_LEN,
2130 sizeof(int16_t) * SUBFRAME_LEN);
2131 memcpy(vector + LPC_ORDER + SUBFRAME_LEN, in,
2132 sizeof(int16_t) * (HALF_FRAME_LEN + SUBFRAME_LEN));
2133 memcpy(p->prev_data, in + HALF_FRAME_LEN,
2134 sizeof(int16_t) * HALF_FRAME_LEN);
2135 memcpy(in, vector + LPC_ORDER, sizeof(int16_t) * FRAME_LEN);
2137 perceptual_filter(p, weighted_lpc, unq_lpc, vector);
2139 memcpy(in, vector + LPC_ORDER, sizeof(int16_t) * FRAME_LEN);
2140 memcpy(vector, p->prev_weight_sig, sizeof(int16_t) * PITCH_MAX);
2141 memcpy(vector + PITCH_MAX, in, sizeof(int16_t) * FRAME_LEN);
2143 scale_vector(vector, FRAME_LEN + PITCH_MAX);
2145 p->pitch_lag[0] = estimate_pitch(vector, PITCH_MAX);
2146 p->pitch_lag[1] = estimate_pitch(vector, PITCH_MAX + HALF_FRAME_LEN);
2148 for (i = PITCH_MAX, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
2149 comp_harmonic_coeff(vector + i, p->pitch_lag[j >> 1], hf + j);
2151 memcpy(vector, p->prev_weight_sig, sizeof(int16_t) * PITCH_MAX);
2152 memcpy(vector + PITCH_MAX, in, sizeof(int16_t) * FRAME_LEN);
2153 memcpy(p->prev_weight_sig, vector + FRAME_LEN, sizeof(int16_t) * PITCH_MAX);
2155 for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
2156 harmonic_filter(hf + j, vector + PITCH_MAX + i, in + i);
2158 inverse_quant(cur_lsp, p->prev_lsp, p->lsp_index, 0);
2159 lsp_interpolate(qnt_lpc, cur_lsp, p->prev_lsp);
2161 memcpy(p->prev_lsp, cur_lsp, sizeof(int16_t) * LPC_ORDER);
2164 for (i = 0; i < SUBFRAMES; i++) {
2165 int16_t impulse_resp[SUBFRAME_LEN];
2166 int16_t residual[SUBFRAME_LEN + PITCH_ORDER - 1];
2167 int16_t flt_in[SUBFRAME_LEN];
2168 int16_t zero[LPC_ORDER], fir[LPC_ORDER], iir[LPC_ORDER];
2171 * Compute the combined impulse response of the synthesis filter,
2172 * formant perceptual weighting filter and harmonic noise shaping filter
2174 memset(zero, 0, sizeof(int16_t) * LPC_ORDER);
2175 memset(vector, 0, sizeof(int16_t) * PITCH_MAX);
2176 memset(flt_in, 0, sizeof(int16_t) * SUBFRAME_LEN);
2178 flt_in[0] = 1 << 13; /* Unit impulse */
2179 synth_percept_filter(qnt_lpc + offset, weighted_lpc + (offset << 1),
2180 zero, zero, flt_in, vector + PITCH_MAX, 1);
2181 harmonic_filter(hf + i, vector + PITCH_MAX, impulse_resp);
2183 /* Compute the combined zero input response */
2185 memcpy(fir, p->perf_fir_mem, sizeof(int16_t) * LPC_ORDER);
2186 memcpy(iir, p->perf_iir_mem, sizeof(int16_t) * LPC_ORDER);
2188 synth_percept_filter(qnt_lpc + offset, weighted_lpc + (offset << 1),
2189 fir, iir, flt_in, vector + PITCH_MAX, 0);
2190 memcpy(vector, p->harmonic_mem, sizeof(int16_t) * PITCH_MAX);
2191 harmonic_noise_sub(hf + i, vector + PITCH_MAX, in);
2193 acb_search(p, residual, impulse_resp, in, i);
2194 gen_acb_excitation(residual, p->prev_excitation,p->pitch_lag[i >> 1],
2195 p->subframe[i], p->cur_rate);
2196 sub_acb_contrib(residual, impulse_resp, in);
2198 fcb_search(p, impulse_resp, in, i);
2200 /* Reconstruct the excitation */
2201 gen_acb_excitation(impulse_resp, p->prev_excitation, p->pitch_lag[i >> 1],
2202 p->subframe[i], RATE_6300);
2204 memmove(p->prev_excitation, p->prev_excitation + SUBFRAME_LEN,
2205 sizeof(int16_t) * (PITCH_MAX - SUBFRAME_LEN));
2206 for (j = 0; j < SUBFRAME_LEN; j++)
2207 in[j] = av_clip_int16((in[j] << 1) + impulse_resp[j]);
2208 memcpy(p->prev_excitation + PITCH_MAX - SUBFRAME_LEN, in,
2209 sizeof(int16_t) * SUBFRAME_LEN);
2211 /* Update filter memories */
2212 synth_percept_filter(qnt_lpc + offset, weighted_lpc + (offset << 1),
2213 p->perf_fir_mem, p->perf_iir_mem,
2214 in, vector + PITCH_MAX, 0);
2215 memmove(p->harmonic_mem, p->harmonic_mem + SUBFRAME_LEN,
2216 sizeof(int16_t) * (PITCH_MAX - SUBFRAME_LEN));
2217 memcpy(p->harmonic_mem + PITCH_MAX - SUBFRAME_LEN, vector + PITCH_MAX,
2218 sizeof(int16_t) * SUBFRAME_LEN);
2221 offset += LPC_ORDER;
2224 if ((ret = ff_alloc_packet2(avctx, avpkt, 24)))
2227 *got_packet_ptr = 1;
2228 avpkt->size = pack_bitstream(p, avpkt->data, avpkt->size);
2232 AVCodec ff_g723_1_encoder = {
2234 .type = AVMEDIA_TYPE_AUDIO,
2235 .id = CODEC_ID_G723_1,
2236 .priv_data_size = sizeof(G723_1_Context),
2237 .init = g723_1_encode_init,
2238 .encode2 = g723_1_encode_frame,
2239 .long_name = NULL_IF_CONFIG_SMALL("G.723.1"),
2240 .sample_fmts = (const enum AVSampleFormat[]){AV_SAMPLE_FMT_S16,
2241 AV_SAMPLE_FMT_NONE},