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"
38 #include "g723_1_data.h"
40 typedef struct g723_1_context {
44 G723_1_Subframe subframe[4];
45 enum FrameType cur_frame_type;
46 enum FrameType past_frame_type;
48 uint8_t lsp_index[LSP_BANDS];
52 int16_t prev_lsp[LPC_ORDER];
53 int16_t prev_excitation[PITCH_MAX];
54 int16_t excitation[PITCH_MAX + FRAME_LEN + 4];
55 int16_t synth_mem[LPC_ORDER];
56 int16_t fir_mem[LPC_ORDER];
57 int iir_mem[LPC_ORDER];
65 int pf_gain; ///< formant postfilter
66 ///< gain scaling unit memory
68 int16_t audio[FRAME_LEN + LPC_ORDER + PITCH_MAX];
69 int16_t prev_data[HALF_FRAME_LEN];
70 int16_t prev_weight_sig[PITCH_MAX];
73 int16_t hpf_fir_mem; ///< highpass filter fir
74 int hpf_iir_mem; ///< and iir memories
75 int16_t perf_fir_mem[LPC_ORDER]; ///< perceptual filter fir
76 int16_t perf_iir_mem[LPC_ORDER]; ///< and iir memories
78 int16_t harmonic_mem[PITCH_MAX];
81 static av_cold int g723_1_decode_init(AVCodecContext *avctx)
83 G723_1_Context *p = avctx->priv_data;
85 avctx->channel_layout = AV_CH_LAYOUT_MONO;
86 avctx->sample_fmt = AV_SAMPLE_FMT_S16;
90 avcodec_get_frame_defaults(&p->frame);
91 avctx->coded_frame = &p->frame;
93 memcpy(p->prev_lsp, dc_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
99 * Unpack the frame into parameters.
101 * @param p the context
102 * @param buf pointer to the input buffer
103 * @param buf_size size of the input buffer
105 static int unpack_bitstream(G723_1_Context *p, const uint8_t *buf,
110 int temp, info_bits, i;
112 init_get_bits(&gb, buf, buf_size * 8);
114 /* Extract frame type and rate info */
115 info_bits = get_bits(&gb, 2);
117 if (info_bits == 3) {
118 p->cur_frame_type = UNTRANSMITTED_FRAME;
122 /* Extract 24 bit lsp indices, 8 bit for each band */
123 p->lsp_index[2] = get_bits(&gb, 8);
124 p->lsp_index[1] = get_bits(&gb, 8);
125 p->lsp_index[0] = get_bits(&gb, 8);
127 if (info_bits == 2) {
128 p->cur_frame_type = SID_FRAME;
129 p->subframe[0].amp_index = get_bits(&gb, 6);
133 /* Extract the info common to both rates */
134 p->cur_rate = info_bits ? RATE_5300 : RATE_6300;
135 p->cur_frame_type = ACTIVE_FRAME;
137 p->pitch_lag[0] = get_bits(&gb, 7);
138 if (p->pitch_lag[0] > 123) /* test if forbidden code */
140 p->pitch_lag[0] += PITCH_MIN;
141 p->subframe[1].ad_cb_lag = get_bits(&gb, 2);
143 p->pitch_lag[1] = get_bits(&gb, 7);
144 if (p->pitch_lag[1] > 123)
146 p->pitch_lag[1] += PITCH_MIN;
147 p->subframe[3].ad_cb_lag = get_bits(&gb, 2);
148 p->subframe[0].ad_cb_lag = 1;
149 p->subframe[2].ad_cb_lag = 1;
151 for (i = 0; i < SUBFRAMES; i++) {
152 /* Extract combined gain */
153 temp = get_bits(&gb, 12);
155 p->subframe[i].dirac_train = 0;
156 if (p->cur_rate == RATE_6300 && p->pitch_lag[i >> 1] < SUBFRAME_LEN - 2) {
157 p->subframe[i].dirac_train = temp >> 11;
161 p->subframe[i].ad_cb_gain = FASTDIV(temp, GAIN_LEVELS);
162 if (p->subframe[i].ad_cb_gain < ad_cb_len) {
163 p->subframe[i].amp_index = temp - p->subframe[i].ad_cb_gain *
170 p->subframe[0].grid_index = get_bits1(&gb);
171 p->subframe[1].grid_index = get_bits1(&gb);
172 p->subframe[2].grid_index = get_bits1(&gb);
173 p->subframe[3].grid_index = get_bits1(&gb);
175 if (p->cur_rate == RATE_6300) {
176 skip_bits1(&gb); /* skip reserved bit */
178 /* Compute pulse_pos index using the 13-bit combined position index */
179 temp = get_bits(&gb, 13);
180 p->subframe[0].pulse_pos = temp / 810;
182 temp -= p->subframe[0].pulse_pos * 810;
183 p->subframe[1].pulse_pos = FASTDIV(temp, 90);
185 temp -= p->subframe[1].pulse_pos * 90;
186 p->subframe[2].pulse_pos = FASTDIV(temp, 9);
187 p->subframe[3].pulse_pos = temp - p->subframe[2].pulse_pos * 9;
189 p->subframe[0].pulse_pos = (p->subframe[0].pulse_pos << 16) +
191 p->subframe[1].pulse_pos = (p->subframe[1].pulse_pos << 14) +
193 p->subframe[2].pulse_pos = (p->subframe[2].pulse_pos << 16) +
195 p->subframe[3].pulse_pos = (p->subframe[3].pulse_pos << 14) +
198 p->subframe[0].pulse_sign = get_bits(&gb, 6);
199 p->subframe[1].pulse_sign = get_bits(&gb, 5);
200 p->subframe[2].pulse_sign = get_bits(&gb, 6);
201 p->subframe[3].pulse_sign = get_bits(&gb, 5);
202 } else { /* 5300 bps */
203 p->subframe[0].pulse_pos = get_bits(&gb, 12);
204 p->subframe[1].pulse_pos = get_bits(&gb, 12);
205 p->subframe[2].pulse_pos = get_bits(&gb, 12);
206 p->subframe[3].pulse_pos = get_bits(&gb, 12);
208 p->subframe[0].pulse_sign = get_bits(&gb, 4);
209 p->subframe[1].pulse_sign = get_bits(&gb, 4);
210 p->subframe[2].pulse_sign = get_bits(&gb, 4);
211 p->subframe[3].pulse_sign = get_bits(&gb, 4);
218 * Bitexact implementation of sqrt(val/2).
220 static int16_t square_root(int val)
222 return (ff_sqrt(val << 1) >> 1) & (~1);
226 * Calculate the number of left-shifts required for normalizing the input.
228 * @param num input number
229 * @param width width of the input, 15 or 31 bits
231 static int normalize_bits(int num, int width)
233 return width - av_log2(num) - 1;
236 #define normalize_bits_int16(num) normalize_bits(num, 15)
237 #define normalize_bits_int32(num) normalize_bits(num, 31)
240 * Scale vector contents based on the largest of their absolutes.
242 static int scale_vector(int16_t *dst, const int16_t *vector, int length)
247 for (i = 0; i < length; i++)
248 max |= FFABS(vector[i]);
250 bits= 14 - av_log2_16bit(max);
251 bits= FFMAX(bits, 0);
253 for (i = 0; i < length; i++)
254 dst[i] = vector[i] << bits >> 3;
260 * Perform inverse quantization of LSP frequencies.
262 * @param cur_lsp the current LSP vector
263 * @param prev_lsp the previous LSP vector
264 * @param lsp_index VQ indices
265 * @param bad_frame bad frame flag
267 static void inverse_quant(int16_t *cur_lsp, int16_t *prev_lsp,
268 uint8_t *lsp_index, int bad_frame)
271 int i, j, temp, stable;
273 /* Check for frame erasure */
280 lsp_index[0] = lsp_index[1] = lsp_index[2] = 0;
283 /* Get the VQ table entry corresponding to the transmitted index */
284 cur_lsp[0] = lsp_band0[lsp_index[0]][0];
285 cur_lsp[1] = lsp_band0[lsp_index[0]][1];
286 cur_lsp[2] = lsp_band0[lsp_index[0]][2];
287 cur_lsp[3] = lsp_band1[lsp_index[1]][0];
288 cur_lsp[4] = lsp_band1[lsp_index[1]][1];
289 cur_lsp[5] = lsp_band1[lsp_index[1]][2];
290 cur_lsp[6] = lsp_band2[lsp_index[2]][0];
291 cur_lsp[7] = lsp_band2[lsp_index[2]][1];
292 cur_lsp[8] = lsp_band2[lsp_index[2]][2];
293 cur_lsp[9] = lsp_band2[lsp_index[2]][3];
295 /* Add predicted vector & DC component to the previously quantized vector */
296 for (i = 0; i < LPC_ORDER; i++) {
297 temp = ((prev_lsp[i] - dc_lsp[i]) * pred + (1 << 14)) >> 15;
298 cur_lsp[i] += dc_lsp[i] + temp;
301 for (i = 0; i < LPC_ORDER; i++) {
302 cur_lsp[0] = FFMAX(cur_lsp[0], 0x180);
303 cur_lsp[LPC_ORDER - 1] = FFMIN(cur_lsp[LPC_ORDER - 1], 0x7e00);
305 /* Stability check */
306 for (j = 1; j < LPC_ORDER; j++) {
307 temp = min_dist + cur_lsp[j - 1] - cur_lsp[j];
310 cur_lsp[j - 1] -= temp;
315 for (j = 1; j < LPC_ORDER; j++) {
316 temp = cur_lsp[j - 1] + min_dist - cur_lsp[j] - 4;
326 memcpy(cur_lsp, prev_lsp, LPC_ORDER * sizeof(*cur_lsp));
330 * Bitexact implementation of 2ab scaled by 1/2^16.
332 * @param a 32 bit multiplicand
333 * @param b 16 bit multiplier
335 #define MULL2(a, b) \
339 * Convert LSP frequencies to LPC coefficients.
341 * @param lpc buffer for LPC coefficients
343 static void lsp2lpc(int16_t *lpc)
345 int f1[LPC_ORDER / 2 + 1];
346 int f2[LPC_ORDER / 2 + 1];
349 /* Calculate negative cosine */
350 for (j = 0; j < LPC_ORDER; j++) {
351 int index = lpc[j] >> 7;
352 int offset = lpc[j] & 0x7f;
353 int temp1 = cos_tab[index] << 16;
354 int temp2 = (cos_tab[index + 1] - cos_tab[index]) *
355 ((offset << 8) + 0x80) << 1;
357 lpc[j] = -(av_sat_dadd32(1 << 15, temp1 + temp2) >> 16);
361 * Compute sum and difference polynomial coefficients
362 * (bitexact alternative to lsp2poly() in lsp.c)
364 /* Initialize with values in Q28 */
366 f1[1] = (lpc[0] << 14) + (lpc[2] << 14);
367 f1[2] = lpc[0] * lpc[2] + (2 << 28);
370 f2[1] = (lpc[1] << 14) + (lpc[3] << 14);
371 f2[2] = lpc[1] * lpc[3] + (2 << 28);
374 * Calculate and scale the coefficients by 1/2 in
375 * each iteration for a final scaling factor of Q25
377 for (i = 2; i < LPC_ORDER / 2; i++) {
378 f1[i + 1] = f1[i - 1] + MULL2(f1[i], lpc[2 * i]);
379 f2[i + 1] = f2[i - 1] + MULL2(f2[i], lpc[2 * i + 1]);
381 for (j = i; j >= 2; j--) {
382 f1[j] = MULL2(f1[j - 1], lpc[2 * i]) +
383 (f1[j] >> 1) + (f1[j - 2] >> 1);
384 f2[j] = MULL2(f2[j - 1], lpc[2 * i + 1]) +
385 (f2[j] >> 1) + (f2[j - 2] >> 1);
390 f1[1] = ((lpc[2 * i] << 16 >> i) + f1[1]) >> 1;
391 f2[1] = ((lpc[2 * i + 1] << 16 >> i) + f2[1]) >> 1;
394 /* Convert polynomial coefficients to LPC coefficients */
395 for (i = 0; i < LPC_ORDER / 2; i++) {
396 int64_t ff1 = f1[i + 1] + f1[i];
397 int64_t ff2 = f2[i + 1] - f2[i];
399 lpc[i] = av_clipl_int32(((ff1 + ff2) << 3) + (1 << 15)) >> 16;
400 lpc[LPC_ORDER - i - 1] = av_clipl_int32(((ff1 - ff2) << 3) +
406 * Quantize LSP frequencies by interpolation and convert them to
407 * the corresponding LPC coefficients.
409 * @param lpc buffer for LPC coefficients
410 * @param cur_lsp the current LSP vector
411 * @param prev_lsp the previous LSP vector
413 static void lsp_interpolate(int16_t *lpc, int16_t *cur_lsp, int16_t *prev_lsp)
416 int16_t *lpc_ptr = lpc;
418 /* cur_lsp * 0.25 + prev_lsp * 0.75 */
419 ff_acelp_weighted_vector_sum(lpc, cur_lsp, prev_lsp,
420 4096, 12288, 1 << 13, 14, LPC_ORDER);
421 ff_acelp_weighted_vector_sum(lpc + LPC_ORDER, cur_lsp, prev_lsp,
422 8192, 8192, 1 << 13, 14, LPC_ORDER);
423 ff_acelp_weighted_vector_sum(lpc + 2 * LPC_ORDER, cur_lsp, prev_lsp,
424 12288, 4096, 1 << 13, 14, LPC_ORDER);
425 memcpy(lpc + 3 * LPC_ORDER, cur_lsp, LPC_ORDER * sizeof(*lpc));
427 for (i = 0; i < SUBFRAMES; i++) {
429 lpc_ptr += LPC_ORDER;
434 * Generate a train of dirac functions with period as pitch lag.
436 static void gen_dirac_train(int16_t *buf, int pitch_lag)
438 int16_t vector[SUBFRAME_LEN];
441 memcpy(vector, buf, SUBFRAME_LEN * sizeof(*vector));
442 for (i = pitch_lag; i < SUBFRAME_LEN; i += pitch_lag) {
443 for (j = 0; j < SUBFRAME_LEN - i; j++)
444 buf[i + j] += vector[j];
449 * Generate fixed codebook excitation vector.
451 * @param vector decoded excitation vector
452 * @param subfrm current subframe
453 * @param cur_rate current bitrate
454 * @param pitch_lag closed loop pitch lag
455 * @param index current subframe index
457 static void gen_fcb_excitation(int16_t *vector, G723_1_Subframe *subfrm,
458 enum Rate cur_rate, int pitch_lag, int index)
462 memset(vector, 0, SUBFRAME_LEN * sizeof(*vector));
464 if (cur_rate == RATE_6300) {
465 if (subfrm->pulse_pos >= max_pos[index])
468 /* Decode amplitudes and positions */
469 j = PULSE_MAX - pulses[index];
470 temp = subfrm->pulse_pos;
471 for (i = 0; i < SUBFRAME_LEN / GRID_SIZE; i++) {
472 temp -= combinatorial_table[j][i];
475 temp += combinatorial_table[j++][i];
476 if (subfrm->pulse_sign & (1 << (PULSE_MAX - j))) {
477 vector[subfrm->grid_index + GRID_SIZE * i] =
478 -fixed_cb_gain[subfrm->amp_index];
480 vector[subfrm->grid_index + GRID_SIZE * i] =
481 fixed_cb_gain[subfrm->amp_index];
486 if (subfrm->dirac_train == 1)
487 gen_dirac_train(vector, pitch_lag);
488 } else { /* 5300 bps */
489 int cb_gain = fixed_cb_gain[subfrm->amp_index];
490 int cb_shift = subfrm->grid_index;
491 int cb_sign = subfrm->pulse_sign;
492 int cb_pos = subfrm->pulse_pos;
493 int offset, beta, lag;
495 for (i = 0; i < 8; i += 2) {
496 offset = ((cb_pos & 7) << 3) + cb_shift + i;
497 vector[offset] = (cb_sign & 1) ? cb_gain : -cb_gain;
502 /* Enhance harmonic components */
503 lag = pitch_contrib[subfrm->ad_cb_gain << 1] + pitch_lag +
504 subfrm->ad_cb_lag - 1;
505 beta = pitch_contrib[(subfrm->ad_cb_gain << 1) + 1];
507 if (lag < SUBFRAME_LEN - 2) {
508 for (i = lag; i < SUBFRAME_LEN; i++)
509 vector[i] += beta * vector[i - lag] >> 15;
515 * Get delayed contribution from the previous excitation vector.
517 static void get_residual(int16_t *residual, int16_t *prev_excitation, int lag)
519 int offset = PITCH_MAX - PITCH_ORDER / 2 - lag;
522 residual[0] = prev_excitation[offset];
523 residual[1] = prev_excitation[offset + 1];
526 for (i = 2; i < SUBFRAME_LEN + PITCH_ORDER - 1; i++)
527 residual[i] = prev_excitation[offset + (i - 2) % lag];
530 static int dot_product(const int16_t *a, const int16_t *b, int length)
532 int sum = ff_dot_product(a,b,length);
533 return av_sat_add32(sum, sum);
537 * Generate adaptive codebook excitation.
539 static void gen_acb_excitation(int16_t *vector, int16_t *prev_excitation,
540 int pitch_lag, G723_1_Subframe *subfrm,
543 int16_t residual[SUBFRAME_LEN + PITCH_ORDER - 1];
544 const int16_t *cb_ptr;
545 int lag = pitch_lag + subfrm->ad_cb_lag - 1;
550 get_residual(residual, prev_excitation, lag);
552 /* Select quantization table */
553 if (cur_rate == RATE_6300 && pitch_lag < SUBFRAME_LEN - 2) {
554 cb_ptr = adaptive_cb_gain85;
556 cb_ptr = adaptive_cb_gain170;
558 /* Calculate adaptive vector */
559 cb_ptr += subfrm->ad_cb_gain * 20;
560 for (i = 0; i < SUBFRAME_LEN; i++) {
561 sum = ff_dot_product(residual + i, cb_ptr, PITCH_ORDER);
562 vector[i] = av_sat_dadd32(1 << 15, av_sat_add32(sum, sum)) >> 16;
567 * Estimate maximum auto-correlation around pitch lag.
569 * @param buf buffer with offset applied
570 * @param offset offset of the excitation vector
571 * @param ccr_max pointer to the maximum auto-correlation
572 * @param pitch_lag decoded pitch lag
573 * @param length length of autocorrelation
574 * @param dir forward lag(1) / backward lag(-1)
576 static int autocorr_max(const int16_t *buf, int offset, int *ccr_max,
577 int pitch_lag, int length, int dir)
579 int limit, ccr, lag = 0;
582 pitch_lag = FFMIN(PITCH_MAX - 3, pitch_lag);
584 limit = FFMIN(FRAME_LEN + PITCH_MAX - offset - length, pitch_lag + 3);
586 limit = pitch_lag + 3;
588 for (i = pitch_lag - 3; i <= limit; i++) {
589 ccr = dot_product(buf, buf + dir * i, length);
591 if (ccr > *ccr_max) {
600 * Calculate pitch postfilter optimal and scaling gains.
602 * @param lag pitch postfilter forward/backward lag
603 * @param ppf pitch postfilter parameters
604 * @param cur_rate current bitrate
605 * @param tgt_eng target energy
606 * @param ccr cross-correlation
607 * @param res_eng residual energy
609 static void comp_ppf_gains(int lag, PPFParam *ppf, enum Rate cur_rate,
610 int tgt_eng, int ccr, int res_eng)
612 int pf_residual; /* square of postfiltered residual */
617 temp1 = tgt_eng * res_eng >> 1;
618 temp2 = ccr * ccr << 1;
621 if (ccr >= res_eng) {
622 ppf->opt_gain = ppf_gain_weight[cur_rate];
624 ppf->opt_gain = (ccr << 15) / res_eng *
625 ppf_gain_weight[cur_rate] >> 15;
627 /* pf_res^2 = tgt_eng + 2*ccr*gain + res_eng*gain^2 */
628 temp1 = (tgt_eng << 15) + (ccr * ppf->opt_gain << 1);
629 temp2 = (ppf->opt_gain * ppf->opt_gain >> 15) * res_eng;
630 pf_residual = av_sat_add32(temp1, temp2 + (1 << 15)) >> 16;
632 if (tgt_eng >= pf_residual << 1) {
635 temp1 = (tgt_eng << 14) / pf_residual;
638 /* scaling_gain = sqrt(tgt_eng/pf_res^2) */
639 ppf->sc_gain = square_root(temp1 << 16);
642 ppf->sc_gain = 0x7fff;
645 ppf->opt_gain = av_clip_int16(ppf->opt_gain * ppf->sc_gain >> 15);
649 * Calculate pitch postfilter parameters.
651 * @param p the context
652 * @param offset offset of the excitation vector
653 * @param pitch_lag decoded pitch lag
654 * @param ppf pitch postfilter parameters
655 * @param cur_rate current bitrate
657 static void comp_ppf_coeff(G723_1_Context *p, int offset, int pitch_lag,
658 PPFParam *ppf, enum Rate cur_rate)
667 * 1 - forward cross-correlation
668 * 2 - forward residual energy
669 * 3 - backward cross-correlation
670 * 4 - backward residual energy
672 int energy[5] = {0, 0, 0, 0, 0};
673 int16_t *buf = p->audio + LPC_ORDER + offset;
674 int fwd_lag = autocorr_max(buf, offset, &energy[1], pitch_lag,
676 int back_lag = autocorr_max(buf, offset, &energy[3], pitch_lag,
681 ppf->sc_gain = 0x7fff;
683 /* Case 0, Section 3.6 */
684 if (!back_lag && !fwd_lag)
687 /* Compute target energy */
688 energy[0] = dot_product(buf, buf, SUBFRAME_LEN);
690 /* Compute forward residual energy */
692 energy[2] = dot_product(buf + fwd_lag, buf + fwd_lag, SUBFRAME_LEN);
694 /* Compute backward residual energy */
696 energy[4] = dot_product(buf - back_lag, buf - back_lag, SUBFRAME_LEN);
698 /* Normalize and shorten */
700 for (i = 0; i < 5; i++)
701 temp1 = FFMAX(energy[i], temp1);
703 scale = normalize_bits(temp1, 31);
704 for (i = 0; i < 5; i++)
705 energy[i] = av_clipl_int32(energy[i] << scale) >> 16;
707 if (fwd_lag && !back_lag) { /* Case 1 */
708 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
710 } else if (!fwd_lag) { /* Case 2 */
711 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
713 } else { /* Case 3 */
716 * Select the largest of energy[1]^2/energy[2]
717 * and energy[3]^2/energy[4]
719 temp1 = energy[4] * ((energy[1] * energy[1] + (1 << 14)) >> 15);
720 temp2 = energy[2] * ((energy[3] * energy[3] + (1 << 14)) >> 15);
721 if (temp1 >= temp2) {
722 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
725 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
732 * Classify frames as voiced/unvoiced.
734 * @param p the context
735 * @param pitch_lag decoded pitch_lag
736 * @param exc_eng excitation energy estimation
737 * @param scale scaling factor of exc_eng
739 * @return residual interpolation index if voiced, 0 otherwise
741 static int comp_interp_index(G723_1_Context *p, int pitch_lag,
742 int *exc_eng, int *scale)
744 int offset = PITCH_MAX + 2 * SUBFRAME_LEN;
745 int16_t *buf = p->audio + LPC_ORDER;
747 int index, ccr, tgt_eng, best_eng, temp;
749 *scale = scale_vector(buf, p->excitation, FRAME_LEN + PITCH_MAX);
752 /* Compute maximum backward cross-correlation */
754 index = autocorr_max(buf, offset, &ccr, pitch_lag, SUBFRAME_LEN * 2, -1);
755 ccr = av_sat_add32(ccr, 1 << 15) >> 16;
757 /* Compute target energy */
758 tgt_eng = dot_product(buf, buf, SUBFRAME_LEN * 2);
759 *exc_eng = av_sat_add32(tgt_eng, 1 << 15) >> 16;
764 /* Compute best energy */
765 best_eng = dot_product(buf - index, buf - index, SUBFRAME_LEN * 2);
766 best_eng = av_sat_add32(best_eng, 1 << 15) >> 16;
768 temp = best_eng * *exc_eng >> 3;
770 if (temp < ccr * ccr) {
777 * Peform residual interpolation based on frame classification.
779 * @param buf decoded excitation vector
780 * @param out output vector
781 * @param lag decoded pitch lag
782 * @param gain interpolated gain
783 * @param rseed seed for random number generator
785 static void residual_interp(int16_t *buf, int16_t *out, int lag,
786 int gain, int *rseed)
789 if (lag) { /* Voiced */
790 int16_t *vector_ptr = buf + PITCH_MAX;
792 for (i = 0; i < lag; i++)
793 out[i] = vector_ptr[i - lag] * 3 >> 2;
794 av_memcpy_backptr((uint8_t*)(out + lag), lag * sizeof(*out),
795 (FRAME_LEN - lag) * sizeof(*out));
796 } else { /* Unvoiced */
797 for (i = 0; i < FRAME_LEN; i++) {
798 *rseed = *rseed * 521 + 259;
799 out[i] = gain * *rseed >> 15;
801 memset(buf, 0, (FRAME_LEN + PITCH_MAX) * sizeof(*buf));
806 * Perform IIR filtering.
808 * @param fir_coef FIR coefficients
809 * @param iir_coef IIR coefficients
810 * @param src source vector
811 * @param dest destination vector
812 * @param width width of the output, 16 bits(0) / 32 bits(1)
814 #define iir_filter(fir_coef, iir_coef, src, dest, width)\
817 int res_shift = 16 & ~-(width);\
818 int in_shift = 16 - res_shift;\
820 for (m = 0; m < SUBFRAME_LEN; m++) {\
822 for (n = 1; n <= LPC_ORDER; n++) {\
823 filter -= (fir_coef)[n - 1] * (src)[m - n] -\
824 (iir_coef)[n - 1] * ((dest)[m - n] >> in_shift);\
827 (dest)[m] = av_clipl_int32(((src)[m] << 16) + (filter << 3) +\
828 (1 << 15)) >> res_shift;\
833 * Adjust gain of postfiltered signal.
835 * @param p the context
836 * @param buf postfiltered output vector
837 * @param energy input energy coefficient
839 static void gain_scale(G723_1_Context *p, int16_t * buf, int energy)
841 int num, denom, gain, bits1, bits2;
846 for (i = 0; i < SUBFRAME_LEN; i++) {
847 int temp = buf[i] >> 2;
849 denom = av_sat_dadd32(denom, temp);
853 bits1 = normalize_bits(num, 31);
854 bits2 = normalize_bits(denom, 31);
855 num = num << bits1 >> 1;
858 bits2 = 5 + bits1 - bits2;
859 bits2 = FFMAX(0, bits2);
861 gain = (num >> 1) / (denom >> 16);
862 gain = square_root(gain << 16 >> bits2);
867 for (i = 0; i < SUBFRAME_LEN; i++) {
868 p->pf_gain = (15 * p->pf_gain + gain + (1 << 3)) >> 4;
869 buf[i] = av_clip_int16((buf[i] * (p->pf_gain + (p->pf_gain >> 4)) +
875 * Perform formant filtering.
877 * @param p the context
878 * @param lpc quantized lpc coefficients
879 * @param buf input buffer
880 * @param dst output buffer
882 static void formant_postfilter(G723_1_Context *p, int16_t *lpc,
883 int16_t *buf, int16_t *dst)
885 int16_t filter_coef[2][LPC_ORDER];
886 int filter_signal[LPC_ORDER + FRAME_LEN], *signal_ptr;
889 memcpy(buf, p->fir_mem, LPC_ORDER * sizeof(*buf));
890 memcpy(filter_signal, p->iir_mem, LPC_ORDER * sizeof(*filter_signal));
892 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
893 for (k = 0; k < LPC_ORDER; k++) {
894 filter_coef[0][k] = (-lpc[k] * postfilter_tbl[0][k] +
896 filter_coef[1][k] = (-lpc[k] * postfilter_tbl[1][k] +
899 iir_filter(filter_coef[0], filter_coef[1], buf + i,
900 filter_signal + i, 1);
904 memcpy(p->fir_mem, buf + FRAME_LEN, LPC_ORDER * sizeof(int16_t));
905 memcpy(p->iir_mem, filter_signal + FRAME_LEN, LPC_ORDER * sizeof(int));
908 signal_ptr = filter_signal + LPC_ORDER;
909 for (i = 0; i < SUBFRAMES; i++) {
915 scale = scale_vector(dst, buf, SUBFRAME_LEN);
917 /* Compute auto correlation coefficients */
918 auto_corr[0] = dot_product(dst, dst + 1, SUBFRAME_LEN - 1);
919 auto_corr[1] = dot_product(dst, dst, SUBFRAME_LEN);
921 /* Compute reflection coefficient */
922 temp = auto_corr[1] >> 16;
924 temp = (auto_corr[0] >> 2) / temp;
926 p->reflection_coef = (3 * p->reflection_coef + temp + 2) >> 2;
927 temp = -p->reflection_coef >> 1 & ~3;
929 /* Compensation filter */
930 for (j = 0; j < SUBFRAME_LEN; j++) {
931 dst[j] = av_sat_dadd32(signal_ptr[j],
932 (signal_ptr[j - 1] >> 16) * temp) >> 16;
935 /* Compute normalized signal energy */
936 temp = 2 * scale + 4;
938 energy = av_clipl_int32((int64_t)auto_corr[1] << -temp);
940 energy = auto_corr[1] >> temp;
942 gain_scale(p, dst, energy);
945 signal_ptr += SUBFRAME_LEN;
950 static int g723_1_decode_frame(AVCodecContext *avctx, void *data,
951 int *got_frame_ptr, AVPacket *avpkt)
953 G723_1_Context *p = avctx->priv_data;
954 const uint8_t *buf = avpkt->data;
955 int buf_size = avpkt->size;
956 int dec_mode = buf[0] & 3;
958 PPFParam ppf[SUBFRAMES];
959 int16_t cur_lsp[LPC_ORDER];
960 int16_t lpc[SUBFRAMES * LPC_ORDER];
961 int16_t acb_vector[SUBFRAME_LEN];
963 int bad_frame = 0, i, j, ret;
964 int16_t *audio = p->audio;
966 if (buf_size < frame_size[dec_mode]) {
968 av_log(avctx, AV_LOG_WARNING,
969 "Expected %d bytes, got %d - skipping packet\n",
970 frame_size[dec_mode], buf_size);
975 if (unpack_bitstream(p, buf, buf_size) < 0) {
977 if (p->past_frame_type == ACTIVE_FRAME)
978 p->cur_frame_type = ACTIVE_FRAME;
980 p->cur_frame_type = UNTRANSMITTED_FRAME;
983 p->frame.nb_samples = FRAME_LEN;
984 if ((ret = avctx->get_buffer(avctx, &p->frame)) < 0) {
985 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
989 out = (int16_t *)p->frame.data[0];
991 if (p->cur_frame_type == ACTIVE_FRAME) {
993 p->erased_frames = 0;
994 else if (p->erased_frames != 3)
997 inverse_quant(cur_lsp, p->prev_lsp, p->lsp_index, bad_frame);
998 lsp_interpolate(lpc, cur_lsp, p->prev_lsp);
1000 /* Save the lsp_vector for the next frame */
1001 memcpy(p->prev_lsp, cur_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
1003 /* Generate the excitation for the frame */
1004 memcpy(p->excitation, p->prev_excitation,
1005 PITCH_MAX * sizeof(*p->excitation));
1006 if (!p->erased_frames) {
1007 int16_t *vector_ptr = p->excitation + PITCH_MAX;
1009 /* Update interpolation gain memory */
1010 p->interp_gain = fixed_cb_gain[(p->subframe[2].amp_index +
1011 p->subframe[3].amp_index) >> 1];
1012 for (i = 0; i < SUBFRAMES; i++) {
1013 gen_fcb_excitation(vector_ptr, &p->subframe[i], p->cur_rate,
1014 p->pitch_lag[i >> 1], i);
1015 gen_acb_excitation(acb_vector, &p->excitation[SUBFRAME_LEN * i],
1016 p->pitch_lag[i >> 1], &p->subframe[i],
1018 /* Get the total excitation */
1019 for (j = 0; j < SUBFRAME_LEN; j++) {
1020 int v = av_clip_int16(vector_ptr[j] << 1);
1021 vector_ptr[j] = av_clip_int16(v + acb_vector[j]);
1023 vector_ptr += SUBFRAME_LEN;
1026 vector_ptr = p->excitation + PITCH_MAX;
1028 p->interp_index = comp_interp_index(p, p->pitch_lag[1],
1029 &p->sid_gain, &p->cur_gain);
1031 /* Peform pitch postfiltering */
1032 if (p->postfilter) {
1034 for (j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1035 comp_ppf_coeff(p, i, p->pitch_lag[j >> 1],
1036 ppf + j, p->cur_rate);
1038 for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1039 ff_acelp_weighted_vector_sum(p->audio + LPC_ORDER + i,
1041 vector_ptr + i + ppf[j].index,
1044 1 << 14, 15, SUBFRAME_LEN);
1046 audio = vector_ptr - LPC_ORDER;
1049 /* Save the excitation for the next frame */
1050 memcpy(p->prev_excitation, p->excitation + FRAME_LEN,
1051 PITCH_MAX * sizeof(*p->excitation));
1053 p->interp_gain = (p->interp_gain * 3 + 2) >> 2;
1054 if (p->erased_frames == 3) {
1056 memset(p->excitation, 0,
1057 (FRAME_LEN + PITCH_MAX) * sizeof(*p->excitation));
1058 memset(p->prev_excitation, 0,
1059 PITCH_MAX * sizeof(*p->excitation));
1060 memset(p->frame.data[0], 0,
1061 (FRAME_LEN + LPC_ORDER) * sizeof(int16_t));
1063 int16_t *buf = p->audio + LPC_ORDER;
1065 /* Regenerate frame */
1066 residual_interp(p->excitation, buf, p->interp_index,
1067 p->interp_gain, &p->random_seed);
1069 /* Save the excitation for the next frame */
1070 memcpy(p->prev_excitation, buf + (FRAME_LEN - PITCH_MAX),
1071 PITCH_MAX * sizeof(*p->excitation));
1075 memset(out, 0, FRAME_LEN * 2);
1076 av_log(avctx, AV_LOG_WARNING,
1077 "G.723.1: Comfort noise generation not supported yet\n");
1080 *(AVFrame *)data = p->frame;
1081 return frame_size[dec_mode];
1084 p->past_frame_type = p->cur_frame_type;
1086 memcpy(p->audio, p->synth_mem, LPC_ORDER * sizeof(*p->audio));
1087 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1088 ff_celp_lp_synthesis_filter(p->audio + i, &lpc[j * LPC_ORDER],
1089 audio + i, SUBFRAME_LEN, LPC_ORDER,
1091 memcpy(p->synth_mem, p->audio + FRAME_LEN, LPC_ORDER * sizeof(*p->audio));
1093 if (p->postfilter) {
1094 formant_postfilter(p, lpc, p->audio, out);
1095 } else { // if output is not postfiltered it should be scaled by 2
1096 for (i = 0; i < FRAME_LEN; i++)
1097 out[i] = av_clip_int16(p->audio[LPC_ORDER + i] << 1);
1101 *(AVFrame *)data = p->frame;
1103 return frame_size[dec_mode];
1106 #define OFFSET(x) offsetof(G723_1_Context, x)
1107 #define AD AV_OPT_FLAG_AUDIO_PARAM | AV_OPT_FLAG_DECODING_PARAM
1109 static const AVOption options[] = {
1110 { "postfilter", "postfilter on/off", OFFSET(postfilter), AV_OPT_TYPE_INT,
1116 static const AVClass g723_1dec_class = {
1117 .class_name = "G.723.1 decoder",
1118 .item_name = av_default_item_name,
1120 .version = LIBAVUTIL_VERSION_INT,
1123 AVCodec ff_g723_1_decoder = {
1125 .type = AVMEDIA_TYPE_AUDIO,
1126 .id = AV_CODEC_ID_G723_1,
1127 .priv_data_size = sizeof(G723_1_Context),
1128 .init = g723_1_decode_init,
1129 .decode = g723_1_decode_frame,
1130 .long_name = NULL_IF_CONFIG_SMALL("G.723.1"),
1131 .capabilities = CODEC_CAP_SUBFRAMES | CODEC_CAP_DR1,
1132 .priv_class = &g723_1dec_class,
1135 #if CONFIG_G723_1_ENCODER
1136 #define BITSTREAM_WRITER_LE
1137 #include "put_bits.h"
1139 static av_cold int g723_1_encode_init(AVCodecContext *avctx)
1141 G723_1_Context *p = avctx->priv_data;
1143 if (avctx->sample_rate != 8000) {
1144 av_log(avctx, AV_LOG_ERROR, "Only 8000Hz sample rate supported\n");
1148 if (avctx->channels != 1) {
1149 av_log(avctx, AV_LOG_ERROR, "Only mono supported\n");
1150 return AVERROR(EINVAL);
1153 if (avctx->bit_rate == 6300) {
1154 p->cur_rate = RATE_6300;
1155 } else if (avctx->bit_rate == 5300) {
1156 av_log(avctx, AV_LOG_ERROR, "Bitrate not supported yet, use 6.3k\n");
1157 return AVERROR_PATCHWELCOME;
1159 av_log(avctx, AV_LOG_ERROR,
1160 "Bitrate not supported, use 6.3k\n");
1161 return AVERROR(EINVAL);
1163 avctx->frame_size = 240;
1164 memcpy(p->prev_lsp, dc_lsp, LPC_ORDER * sizeof(int16_t));
1170 * Remove DC component from the input signal.
1172 * @param buf input signal
1173 * @param fir zero memory
1174 * @param iir pole memory
1176 static void highpass_filter(int16_t *buf, int16_t *fir, int *iir)
1179 for (i = 0; i < FRAME_LEN; i++) {
1180 *iir = (buf[i] << 15) + ((-*fir) << 15) + MULL2(*iir, 0x7f00);
1182 buf[i] = av_clipl_int32((int64_t)*iir + (1 << 15)) >> 16;
1187 * Estimate autocorrelation of the input vector.
1189 * @param buf input buffer
1190 * @param autocorr autocorrelation coefficients vector
1192 static void comp_autocorr(int16_t *buf, int16_t *autocorr)
1195 int16_t vector[LPC_FRAME];
1197 scale_vector(vector, buf, LPC_FRAME);
1199 /* Apply the Hamming window */
1200 for (i = 0; i < LPC_FRAME; i++)
1201 vector[i] = (vector[i] * hamming_window[i] + (1 << 14)) >> 15;
1203 /* Compute the first autocorrelation coefficient */
1204 temp = ff_dot_product(vector, vector, LPC_FRAME);
1206 /* Apply a white noise correlation factor of (1025/1024) */
1210 scale = normalize_bits_int32(temp);
1211 autocorr[0] = av_clipl_int32((int64_t)(temp << scale) +
1214 /* Compute the remaining coefficients */
1216 memset(autocorr + 1, 0, LPC_ORDER * sizeof(int16_t));
1218 for (i = 1; i <= LPC_ORDER; i++) {
1219 temp = ff_dot_product(vector, vector + i, LPC_FRAME - i);
1220 temp = MULL2((temp << scale), binomial_window[i - 1]);
1221 autocorr[i] = av_clipl_int32((int64_t)temp + (1 << 15)) >> 16;
1227 * Use Levinson-Durbin recursion to compute LPC coefficients from
1228 * autocorrelation values.
1230 * @param lpc LPC coefficients vector
1231 * @param autocorr autocorrelation coefficients vector
1232 * @param error prediction error
1234 static void levinson_durbin(int16_t *lpc, int16_t *autocorr, int16_t error)
1236 int16_t vector[LPC_ORDER];
1237 int16_t partial_corr;
1240 memset(lpc, 0, LPC_ORDER * sizeof(int16_t));
1242 for (i = 0; i < LPC_ORDER; i++) {
1243 /* Compute the partial correlation coefficient */
1245 for (j = 0; j < i; j++)
1246 temp -= lpc[j] * autocorr[i - j - 1];
1247 temp = ((autocorr[i] << 13) + temp) << 3;
1249 if (FFABS(temp) >= (error << 16))
1252 partial_corr = temp / (error << 1);
1254 lpc[i] = av_clipl_int32((int64_t)(partial_corr << 14) +
1257 /* Update the prediction error */
1258 temp = MULL2(temp, partial_corr);
1259 error = av_clipl_int32((int64_t)(error << 16) - temp +
1262 memcpy(vector, lpc, i * sizeof(int16_t));
1263 for (j = 0; j < i; j++) {
1264 temp = partial_corr * vector[i - j - 1] << 1;
1265 lpc[j] = av_clipl_int32((int64_t)(lpc[j] << 16) - temp +
1272 * Calculate LPC coefficients for the current frame.
1274 * @param buf current frame
1275 * @param prev_data 2 trailing subframes of the previous frame
1276 * @param lpc LPC coefficients vector
1278 static void comp_lpc_coeff(int16_t *buf, int16_t *lpc)
1280 int16_t autocorr[(LPC_ORDER + 1) * SUBFRAMES];
1281 int16_t *autocorr_ptr = autocorr;
1282 int16_t *lpc_ptr = lpc;
1285 for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
1286 comp_autocorr(buf + i, autocorr_ptr);
1287 levinson_durbin(lpc_ptr, autocorr_ptr + 1, autocorr_ptr[0]);
1289 lpc_ptr += LPC_ORDER;
1290 autocorr_ptr += LPC_ORDER + 1;
1294 static void lpc2lsp(int16_t *lpc, int16_t *prev_lsp, int16_t *lsp)
1296 int f[LPC_ORDER + 2]; ///< coefficients of the sum and difference
1297 ///< polynomials (F1, F2) ordered as
1298 ///< f1[0], f2[0], ...., f1[5], f2[5]
1300 int max, shift, cur_val, prev_val, count, p;
1304 /* Initialize f1[0] and f2[0] to 1 in Q25 */
1305 for (i = 0; i < LPC_ORDER; i++)
1306 lsp[i] = (lpc[i] * bandwidth_expand[i] + (1 << 14)) >> 15;
1308 /* Apply bandwidth expansion on the LPC coefficients */
1309 f[0] = f[1] = 1 << 25;
1311 /* Compute the remaining coefficients */
1312 for (i = 0; i < LPC_ORDER / 2; i++) {
1314 f[2 * i + 2] = -f[2 * i] - ((lsp[i] + lsp[LPC_ORDER - 1 - i]) << 12);
1316 f[2 * i + 3] = f[2 * i + 1] - ((lsp[i] - lsp[LPC_ORDER - 1 - i]) << 12);
1319 /* Divide f1[5] and f2[5] by 2 for use in polynomial evaluation */
1321 f[LPC_ORDER + 1] >>= 1;
1323 /* Normalize and shorten */
1325 for (i = 1; i < LPC_ORDER + 2; i++)
1326 max = FFMAX(max, FFABS(f[i]));
1328 shift = normalize_bits_int32(max);
1330 for (i = 0; i < LPC_ORDER + 2; i++)
1331 f[i] = av_clipl_int32((int64_t)(f[i] << shift) + (1 << 15)) >> 16;
1334 * Evaluate F1 and F2 at uniform intervals of pi/256 along the
1335 * unit circle and check for zero crossings.
1339 for (i = 0; i <= LPC_ORDER / 2; i++)
1340 temp += f[2 * i] * cos_tab[0];
1341 prev_val = av_clipl_int32(temp << 1);
1343 for ( i = 1; i < COS_TBL_SIZE / 2; i++) {
1346 for (j = 0; j <= LPC_ORDER / 2; j++)
1347 temp += f[LPC_ORDER - 2 * j + p] * cos_tab[i * j % COS_TBL_SIZE];
1348 cur_val = av_clipl_int32(temp << 1);
1350 /* Check for sign change, indicating a zero crossing */
1351 if ((cur_val ^ prev_val) < 0) {
1352 int abs_cur = FFABS(cur_val);
1353 int abs_prev = FFABS(prev_val);
1354 int sum = abs_cur + abs_prev;
1356 shift = normalize_bits_int32(sum);
1358 abs_prev = abs_prev << shift >> 8;
1359 lsp[count++] = ((i - 1) << 7) + (abs_prev >> 1) / (sum >> 16);
1361 if (count == LPC_ORDER)
1364 /* Switch between sum and difference polynomials */
1369 for (j = 0; j <= LPC_ORDER / 2; j++){
1370 temp += f[LPC_ORDER - 2 * j + p] *
1371 cos_tab[i * j % COS_TBL_SIZE];
1373 cur_val = av_clipl_int32(temp<<1);
1378 if (count != LPC_ORDER)
1379 memcpy(lsp, prev_lsp, LPC_ORDER * sizeof(int16_t));
1383 * Quantize the current LSP subvector.
1385 * @param num band number
1386 * @param offset offset of the current subvector in an LPC_ORDER vector
1387 * @param size size of the current subvector
1389 #define get_index(num, offset, size) \
1391 int error, max = -1;\
1394 for (i = 0; i < LSP_CB_SIZE; i++) {\
1395 for (j = 0; j < size; j++){\
1396 temp[j] = (weight[j + (offset)] * lsp_band##num[i][j] +\
1399 error = dot_product(lsp + (offset), temp, size) << 1;\
1400 error -= dot_product(lsp_band##num[i], temp, size);\
1403 lsp_index[num] = i;\
1409 * Vector quantize the LSP frequencies.
1411 * @param lsp the current lsp vector
1412 * @param prev_lsp the previous lsp vector
1414 static void lsp_quantize(uint8_t *lsp_index, int16_t *lsp, int16_t *prev_lsp)
1416 int16_t weight[LPC_ORDER];
1420 /* Calculate the VQ weighting vector */
1421 weight[0] = (1 << 20) / (lsp[1] - lsp[0]);
1422 weight[LPC_ORDER - 1] = (1 << 20) /
1423 (lsp[LPC_ORDER - 1] - lsp[LPC_ORDER - 2]);
1425 for (i = 1; i < LPC_ORDER - 1; i++) {
1426 min = FFMIN(lsp[i] - lsp[i - 1], lsp[i + 1] - lsp[i]);
1428 weight[i] = (1 << 20) / min;
1430 weight[i] = INT16_MAX;
1435 for (i = 0; i < LPC_ORDER; i++)
1436 max = FFMAX(weight[i], max);
1438 shift = normalize_bits_int16(max);
1439 for (i = 0; i < LPC_ORDER; i++) {
1440 weight[i] <<= shift;
1443 /* Compute the VQ target vector */
1444 for (i = 0; i < LPC_ORDER; i++) {
1445 lsp[i] -= dc_lsp[i] +
1446 (((prev_lsp[i] - dc_lsp[i]) * 12288 + (1 << 14)) >> 15);
1455 * Apply the formant perceptual weighting filter.
1457 * @param flt_coef filter coefficients
1458 * @param unq_lpc unquantized lpc vector
1460 static void perceptual_filter(G723_1_Context *p, int16_t *flt_coef,
1461 int16_t *unq_lpc, int16_t *buf)
1463 int16_t vector[FRAME_LEN + LPC_ORDER];
1466 memcpy(buf, p->iir_mem, sizeof(int16_t) * LPC_ORDER);
1467 memcpy(vector, p->fir_mem, sizeof(int16_t) * LPC_ORDER);
1468 memcpy(vector + LPC_ORDER, buf + LPC_ORDER, sizeof(int16_t) * FRAME_LEN);
1470 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
1471 for (k = 0; k < LPC_ORDER; k++) {
1472 flt_coef[k + 2 * l] = (unq_lpc[k + l] * percept_flt_tbl[0][k] +
1474 flt_coef[k + 2 * l + LPC_ORDER] = (unq_lpc[k + l] *
1475 percept_flt_tbl[1][k] +
1478 iir_filter(flt_coef + 2 * l, flt_coef + 2 * l + LPC_ORDER, vector + i,
1482 memcpy(p->iir_mem, buf + FRAME_LEN, sizeof(int16_t) * LPC_ORDER);
1483 memcpy(p->fir_mem, vector + FRAME_LEN, sizeof(int16_t) * LPC_ORDER);
1487 * Estimate the open loop pitch period.
1489 * @param buf perceptually weighted speech
1490 * @param start estimation is carried out from this position
1492 static int estimate_pitch(int16_t *buf, int start)
1495 int max_ccr = 0x4000;
1496 int max_eng = 0x7fff;
1497 int index = PITCH_MIN;
1498 int offset = start - PITCH_MIN + 1;
1500 int ccr, eng, orig_eng, ccr_eng, exp;
1505 orig_eng = ff_dot_product(buf + offset, buf + offset, HALF_FRAME_LEN);
1507 for (i = PITCH_MIN; i <= PITCH_MAX - 3; i++) {
1510 /* Update energy and compute correlation */
1511 orig_eng += buf[offset] * buf[offset] -
1512 buf[offset + HALF_FRAME_LEN] * buf[offset + HALF_FRAME_LEN];
1513 ccr = ff_dot_product(buf + start, buf + offset, HALF_FRAME_LEN);
1517 /* Split into mantissa and exponent to maintain precision */
1518 exp = normalize_bits_int32(ccr);
1519 ccr = av_clipl_int32((int64_t)(ccr << exp) + (1 << 15)) >> 16;
1522 temp = normalize_bits_int32(ccr);
1523 ccr = ccr << temp >> 16;
1526 temp = normalize_bits_int32(orig_eng);
1527 eng = av_clipl_int32((int64_t)(orig_eng << temp) + (1 << 15)) >> 16;
1537 if (exp + 1 < max_exp)
1540 /* Equalize exponents before comparison */
1541 if (exp + 1 == max_exp)
1542 temp = max_ccr >> 1;
1545 ccr_eng = ccr * max_eng;
1546 diff = ccr_eng - eng * temp;
1547 if (diff > 0 && (i - index < PITCH_MIN || diff > ccr_eng >> 2)) {
1559 * Compute harmonic noise filter parameters.
1561 * @param buf perceptually weighted speech
1562 * @param pitch_lag open loop pitch period
1563 * @param hf harmonic filter parameters
1565 static void comp_harmonic_coeff(int16_t *buf, int16_t pitch_lag, HFParam *hf)
1567 int ccr, eng, max_ccr, max_eng;
1572 for (i = 0, j = pitch_lag - 3; j <= pitch_lag + 3; i++, j++) {
1573 /* Compute residual energy */
1574 energy[i << 1] = ff_dot_product(buf - j, buf - j, SUBFRAME_LEN);
1575 /* Compute correlation */
1576 energy[(i << 1) + 1] = ff_dot_product(buf, buf - j, SUBFRAME_LEN);
1579 /* Compute target energy */
1580 energy[14] = ff_dot_product(buf, buf, SUBFRAME_LEN);
1584 for (i = 0; i < 15; i++)
1585 max = FFMAX(max, FFABS(energy[i]));
1587 exp = normalize_bits_int32(max);
1588 for (i = 0; i < 15; i++) {
1589 energy[i] = av_clipl_int32((int64_t)(energy[i] << exp) +
1598 for (i = 0; i <= 6; i++) {
1599 eng = energy[i << 1];
1600 ccr = energy[(i << 1) + 1];
1605 ccr = (ccr * ccr + (1 << 14)) >> 15;
1606 diff = ccr * max_eng - eng * max_ccr;
1614 if (hf->index == -1) {
1615 hf->index = pitch_lag;
1619 eng = energy[14] * max_eng;
1620 eng = (eng >> 2) + (eng >> 3);
1621 ccr = energy[(hf->index << 1) + 1] * energy[(hf->index << 1) + 1];
1623 eng = energy[(hf->index << 1) + 1];
1628 hf->gain = ((eng << 15) / max_eng * 0x2800 + (1 << 14)) >> 15;
1630 hf->index += pitch_lag - 3;
1634 * Apply the harmonic noise shaping filter.
1636 * @param hf filter parameters
1638 static void harmonic_filter(HFParam *hf, int16_t *src, int16_t *dest)
1642 for (i = 0; i < SUBFRAME_LEN; i++) {
1643 int64_t temp = hf->gain * src[i - hf->index] << 1;
1644 dest[i] = av_clipl_int32((src[i] << 16) - temp + (1 << 15)) >> 16;
1648 static void harmonic_noise_sub(HFParam *hf, int16_t *src, int16_t *dest)
1651 for (i = 0; i < SUBFRAME_LEN; i++) {
1652 int64_t temp = hf->gain * src[i - hf->index] << 1;
1653 dest[i] = av_clipl_int32(((dest[i] - src[i]) << 16) + temp +
1660 * Combined synthesis and formant perceptual weighting filer.
1662 * @param qnt_lpc quantized lpc coefficients
1663 * @param perf_lpc perceptual filter coefficients
1664 * @param perf_fir perceptual filter fir memory
1665 * @param perf_iir perceptual filter iir memory
1666 * @param scale the filter output will be scaled by 2^scale
1668 static void synth_percept_filter(int16_t *qnt_lpc, int16_t *perf_lpc,
1669 int16_t *perf_fir, int16_t *perf_iir,
1670 int16_t *src, int16_t *dest, int scale)
1673 int16_t buf_16[SUBFRAME_LEN + LPC_ORDER];
1674 int64_t buf[SUBFRAME_LEN];
1676 int16_t *bptr_16 = buf_16 + LPC_ORDER;
1678 memcpy(buf_16, perf_fir, sizeof(int16_t) * LPC_ORDER);
1679 memcpy(dest - LPC_ORDER, perf_iir, sizeof(int16_t) * LPC_ORDER);
1681 for (i = 0; i < SUBFRAME_LEN; i++) {
1683 for (j = 1; j <= LPC_ORDER; j++)
1684 temp -= qnt_lpc[j - 1] * bptr_16[i - j];
1686 buf[i] = (src[i] << 15) + (temp << 3);
1687 bptr_16[i] = av_clipl_int32(buf[i] + (1 << 15)) >> 16;
1690 for (i = 0; i < SUBFRAME_LEN; i++) {
1691 int64_t fir = 0, iir = 0;
1692 for (j = 1; j <= LPC_ORDER; j++) {
1693 fir -= perf_lpc[j - 1] * bptr_16[i - j];
1694 iir += perf_lpc[j + LPC_ORDER - 1] * dest[i - j];
1696 dest[i] = av_clipl_int32(((buf[i] + (fir << 3)) << scale) + (iir << 3) +
1699 memcpy(perf_fir, buf_16 + SUBFRAME_LEN, sizeof(int16_t) * LPC_ORDER);
1700 memcpy(perf_iir, dest + SUBFRAME_LEN - LPC_ORDER,
1701 sizeof(int16_t) * LPC_ORDER);
1705 * Compute the adaptive codebook contribution.
1707 * @param buf input signal
1708 * @param index the current subframe index
1710 static void acb_search(G723_1_Context *p, int16_t *residual,
1711 int16_t *impulse_resp, int16_t *buf,
1715 int16_t flt_buf[PITCH_ORDER][SUBFRAME_LEN];
1717 const int16_t *cb_tbl = adaptive_cb_gain85;
1719 int ccr_buf[PITCH_ORDER * SUBFRAMES << 2];
1721 int pitch_lag = p->pitch_lag[index >> 1];
1724 int odd_frame = index & 1;
1725 int iter = 3 + odd_frame;
1729 int i, j, k, l, max;
1733 if (pitch_lag == PITCH_MIN)
1736 pitch_lag = FFMIN(pitch_lag, PITCH_MAX - 5);
1739 for (i = 0; i < iter; i++) {
1740 get_residual(residual, p->prev_excitation, pitch_lag + i - 1);
1742 for (j = 0; j < SUBFRAME_LEN; j++) {
1744 for (k = 0; k <= j; k++)
1745 temp += residual[PITCH_ORDER - 1 + k] * impulse_resp[j - k];
1746 flt_buf[PITCH_ORDER - 1][j] = av_clipl_int32((temp << 1) +
1750 for (j = PITCH_ORDER - 2; j >= 0; j--) {
1751 flt_buf[j][0] = ((residual[j] << 13) + (1 << 14)) >> 15;
1752 for (k = 1; k < SUBFRAME_LEN; k++) {
1753 temp = (flt_buf[j + 1][k - 1] << 15) +
1754 residual[j] * impulse_resp[k];
1755 flt_buf[j][k] = av_clipl_int32((temp << 1) + (1 << 15)) >> 16;
1759 /* Compute crosscorrelation with the signal */
1760 for (j = 0; j < PITCH_ORDER; j++) {
1761 temp = ff_dot_product(buf, flt_buf[j], SUBFRAME_LEN);
1762 ccr_buf[count++] = av_clipl_int32(temp << 1);
1765 /* Compute energies */
1766 for (j = 0; j < PITCH_ORDER; j++) {
1767 ccr_buf[count++] = dot_product(flt_buf[j], flt_buf[j],
1771 for (j = 1; j < PITCH_ORDER; j++) {
1772 for (k = 0; k < j; k++) {
1773 temp = ff_dot_product(flt_buf[j], flt_buf[k], SUBFRAME_LEN);
1774 ccr_buf[count++] = av_clipl_int32(temp<<2);
1779 /* Normalize and shorten */
1781 for (i = 0; i < 20 * iter; i++)
1782 max = FFMAX(max, FFABS(ccr_buf[i]));
1784 temp = normalize_bits_int32(max);
1786 for (i = 0; i < 20 * iter; i++){
1787 ccr_buf[i] = av_clipl_int32((int64_t)(ccr_buf[i] << temp) +
1792 for (i = 0; i < iter; i++) {
1793 /* Select quantization table */
1794 if (!odd_frame && pitch_lag + i - 1 >= SUBFRAME_LEN - 2 ||
1795 odd_frame && pitch_lag >= SUBFRAME_LEN - 2) {
1796 cb_tbl = adaptive_cb_gain170;
1800 for (j = 0, k = 0; j < tbl_size; j++, k += 20) {
1802 for (l = 0; l < 20; l++)
1803 temp += ccr_buf[20 * i + l] * cb_tbl[k + l];
1804 temp = av_clipl_int32(temp);
1815 pitch_lag += acb_lag - 1;
1819 p->pitch_lag[index >> 1] = pitch_lag;
1820 p->subframe[index].ad_cb_lag = acb_lag;
1821 p->subframe[index].ad_cb_gain = acb_gain;
1825 * Subtract the adaptive codebook contribution from the input
1826 * to obtain the residual.
1828 * @param buf target vector
1830 static void sub_acb_contrib(int16_t *residual, int16_t *impulse_resp,
1834 /* Subtract adaptive CB contribution to obtain the residual */
1835 for (i = 0; i < SUBFRAME_LEN; i++) {
1836 int64_t temp = buf[i] << 14;
1837 for (j = 0; j <= i; j++)
1838 temp -= residual[j] * impulse_resp[i - j];
1840 buf[i] = av_clipl_int32((temp << 2) + (1 << 15)) >> 16;
1845 * Quantize the residual signal using the fixed codebook (MP-MLQ).
1847 * @param optim optimized fixed codebook parameters
1848 * @param buf excitation vector
1850 static void get_fcb_param(FCBParam *optim, int16_t *impulse_resp,
1851 int16_t *buf, int pulse_cnt, int pitch_lag)
1854 int16_t impulse_r[SUBFRAME_LEN];
1855 int16_t temp_corr[SUBFRAME_LEN];
1856 int16_t impulse_corr[SUBFRAME_LEN];
1858 int ccr1[SUBFRAME_LEN];
1859 int ccr2[SUBFRAME_LEN];
1860 int amp, err, max, max_amp_index, min, scale, i, j, k, l;
1864 /* Update impulse response */
1865 memcpy(impulse_r, impulse_resp, sizeof(int16_t) * SUBFRAME_LEN);
1866 param.dirac_train = 0;
1867 if (pitch_lag < SUBFRAME_LEN - 2) {
1868 param.dirac_train = 1;
1869 gen_dirac_train(impulse_r, pitch_lag);
1872 for (i = 0; i < SUBFRAME_LEN; i++)
1873 temp_corr[i] = impulse_r[i] >> 1;
1875 /* Compute impulse response autocorrelation */
1876 temp = dot_product(temp_corr, temp_corr, SUBFRAME_LEN);
1878 scale = normalize_bits_int32(temp);
1879 impulse_corr[0] = av_clipl_int32((temp << scale) + (1 << 15)) >> 16;
1881 for (i = 1; i < SUBFRAME_LEN; i++) {
1882 temp = dot_product(temp_corr + i, temp_corr, SUBFRAME_LEN - i);
1883 impulse_corr[i] = av_clipl_int32((temp << scale) + (1 << 15)) >> 16;
1886 /* Compute crosscorrelation of impulse response with residual signal */
1888 for (i = 0; i < SUBFRAME_LEN; i++){
1889 temp = dot_product(buf + i, impulse_r, SUBFRAME_LEN - i);
1891 ccr1[i] = temp >> -scale;
1893 ccr1[i] = av_clipl_int32(temp << scale);
1897 for (i = 0; i < GRID_SIZE; i++) {
1898 /* Maximize the crosscorrelation */
1900 for (j = i; j < SUBFRAME_LEN; j += GRID_SIZE) {
1901 temp = FFABS(ccr1[j]);
1904 param.pulse_pos[0] = j;
1908 /* Quantize the gain (max crosscorrelation/impulse_corr[0]) */
1911 max_amp_index = GAIN_LEVELS - 2;
1912 for (j = max_amp_index; j >= 2; j--) {
1913 temp = av_clipl_int32((int64_t)fixed_cb_gain[j] *
1914 impulse_corr[0] << 1);
1915 temp = FFABS(temp - amp);
1923 /* Select additional gain values */
1924 for (j = 1; j < 5; j++) {
1925 for (k = i; k < SUBFRAME_LEN; k += GRID_SIZE) {
1929 param.amp_index = max_amp_index + j - 2;
1930 amp = fixed_cb_gain[param.amp_index];
1932 param.pulse_sign[0] = (ccr2[param.pulse_pos[0]] < 0) ? -amp : amp;
1933 temp_corr[param.pulse_pos[0]] = 1;
1935 for (k = 1; k < pulse_cnt; k++) {
1937 for (l = i; l < SUBFRAME_LEN; l += GRID_SIZE) {
1940 temp = impulse_corr[FFABS(l - param.pulse_pos[k - 1])];
1941 temp = av_clipl_int32((int64_t)temp *
1942 param.pulse_sign[k - 1] << 1);
1944 temp = FFABS(ccr2[l]);
1947 param.pulse_pos[k] = l;
1951 param.pulse_sign[k] = (ccr2[param.pulse_pos[k]] < 0) ?
1953 temp_corr[param.pulse_pos[k]] = 1;
1956 /* Create the error vector */
1957 memset(temp_corr, 0, sizeof(int16_t) * SUBFRAME_LEN);
1959 for (k = 0; k < pulse_cnt; k++)
1960 temp_corr[param.pulse_pos[k]] = param.pulse_sign[k];
1962 for (k = SUBFRAME_LEN - 1; k >= 0; k--) {
1964 for (l = 0; l <= k; l++) {
1965 int prod = av_clipl_int32((int64_t)temp_corr[l] *
1966 impulse_r[k - l] << 1);
1967 temp = av_clipl_int32(temp + prod);
1969 temp_corr[k] = temp << 2 >> 16;
1972 /* Compute square of error */
1974 for (k = 0; k < SUBFRAME_LEN; k++) {
1976 prod = av_clipl_int32((int64_t)buf[k] * temp_corr[k] << 1);
1977 err = av_clipl_int32(err - prod);
1978 prod = av_clipl_int32((int64_t)temp_corr[k] * temp_corr[k]);
1979 err = av_clipl_int32(err + prod);
1983 if (err < optim->min_err) {
1984 optim->min_err = err;
1985 optim->grid_index = i;
1986 optim->amp_index = param.amp_index;
1987 optim->dirac_train = param.dirac_train;
1989 for (k = 0; k < pulse_cnt; k++) {
1990 optim->pulse_sign[k] = param.pulse_sign[k];
1991 optim->pulse_pos[k] = param.pulse_pos[k];
1999 * Encode the pulse position and gain of the current subframe.
2001 * @param optim optimized fixed CB parameters
2002 * @param buf excitation vector
2004 static void pack_fcb_param(G723_1_Subframe *subfrm, FCBParam *optim,
2005 int16_t *buf, int pulse_cnt)
2009 j = PULSE_MAX - pulse_cnt;
2011 subfrm->pulse_sign = 0;
2012 subfrm->pulse_pos = 0;
2014 for (i = 0; i < SUBFRAME_LEN >> 1; i++) {
2015 int val = buf[optim->grid_index + (i << 1)];
2017 subfrm->pulse_pos += combinatorial_table[j][i];
2019 subfrm->pulse_sign <<= 1;
2020 if (val < 0) subfrm->pulse_sign++;
2023 if (j == PULSE_MAX) break;
2026 subfrm->amp_index = optim->amp_index;
2027 subfrm->grid_index = optim->grid_index;
2028 subfrm->dirac_train = optim->dirac_train;
2032 * Compute the fixed codebook excitation.
2034 * @param buf target vector
2035 * @param impulse_resp impulse response of the combined filter
2037 static void fcb_search(G723_1_Context *p, int16_t *impulse_resp,
2038 int16_t *buf, int index)
2041 int pulse_cnt = pulses[index];
2044 optim.min_err = 1 << 30;
2045 get_fcb_param(&optim, impulse_resp, buf, pulse_cnt, SUBFRAME_LEN);
2047 if (p->pitch_lag[index >> 1] < SUBFRAME_LEN - 2) {
2048 get_fcb_param(&optim, impulse_resp, buf, pulse_cnt,
2049 p->pitch_lag[index >> 1]);
2052 /* Reconstruct the excitation */
2053 memset(buf, 0, sizeof(int16_t) * SUBFRAME_LEN);
2054 for (i = 0; i < pulse_cnt; i++)
2055 buf[optim.pulse_pos[i]] = optim.pulse_sign[i];
2057 pack_fcb_param(&p->subframe[index], &optim, buf, pulse_cnt);
2059 if (optim.dirac_train)
2060 gen_dirac_train(buf, p->pitch_lag[index >> 1]);
2064 * Pack the frame parameters into output bitstream.
2066 * @param frame output buffer
2067 * @param size size of the buffer
2069 static int pack_bitstream(G723_1_Context *p, unsigned char *frame, int size)
2072 int info_bits, i, temp;
2074 init_put_bits(&pb, frame, size);
2076 if (p->cur_rate == RATE_6300) {
2078 put_bits(&pb, 2, info_bits);
2081 put_bits(&pb, 8, p->lsp_index[2]);
2082 put_bits(&pb, 8, p->lsp_index[1]);
2083 put_bits(&pb, 8, p->lsp_index[0]);
2085 put_bits(&pb, 7, p->pitch_lag[0] - PITCH_MIN);
2086 put_bits(&pb, 2, p->subframe[1].ad_cb_lag);
2087 put_bits(&pb, 7, p->pitch_lag[1] - PITCH_MIN);
2088 put_bits(&pb, 2, p->subframe[3].ad_cb_lag);
2090 /* Write 12 bit combined gain */
2091 for (i = 0; i < SUBFRAMES; i++) {
2092 temp = p->subframe[i].ad_cb_gain * GAIN_LEVELS +
2093 p->subframe[i].amp_index;
2094 if (p->cur_rate == RATE_6300)
2095 temp += p->subframe[i].dirac_train << 11;
2096 put_bits(&pb, 12, temp);
2099 put_bits(&pb, 1, p->subframe[0].grid_index);
2100 put_bits(&pb, 1, p->subframe[1].grid_index);
2101 put_bits(&pb, 1, p->subframe[2].grid_index);
2102 put_bits(&pb, 1, p->subframe[3].grid_index);
2104 if (p->cur_rate == RATE_6300) {
2105 skip_put_bits(&pb, 1); /* reserved bit */
2107 /* Write 13 bit combined position index */
2108 temp = (p->subframe[0].pulse_pos >> 16) * 810 +
2109 (p->subframe[1].pulse_pos >> 14) * 90 +
2110 (p->subframe[2].pulse_pos >> 16) * 9 +
2111 (p->subframe[3].pulse_pos >> 14);
2112 put_bits(&pb, 13, temp);
2114 put_bits(&pb, 16, p->subframe[0].pulse_pos & 0xffff);
2115 put_bits(&pb, 14, p->subframe[1].pulse_pos & 0x3fff);
2116 put_bits(&pb, 16, p->subframe[2].pulse_pos & 0xffff);
2117 put_bits(&pb, 14, p->subframe[3].pulse_pos & 0x3fff);
2119 put_bits(&pb, 6, p->subframe[0].pulse_sign);
2120 put_bits(&pb, 5, p->subframe[1].pulse_sign);
2121 put_bits(&pb, 6, p->subframe[2].pulse_sign);
2122 put_bits(&pb, 5, p->subframe[3].pulse_sign);
2125 flush_put_bits(&pb);
2126 return frame_size[info_bits];
2129 static int g723_1_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
2130 const AVFrame *frame, int *got_packet_ptr)
2132 G723_1_Context *p = avctx->priv_data;
2133 int16_t unq_lpc[LPC_ORDER * SUBFRAMES];
2134 int16_t qnt_lpc[LPC_ORDER * SUBFRAMES];
2135 int16_t cur_lsp[LPC_ORDER];
2136 int16_t weighted_lpc[LPC_ORDER * SUBFRAMES << 1];
2137 int16_t vector[FRAME_LEN + PITCH_MAX];
2139 int16_t *in = (const int16_t *)frame->data[0];
2144 highpass_filter(in, &p->hpf_fir_mem, &p->hpf_iir_mem);
2146 memcpy(vector, p->prev_data, HALF_FRAME_LEN * sizeof(int16_t));
2147 memcpy(vector + HALF_FRAME_LEN, in, FRAME_LEN * sizeof(int16_t));
2149 comp_lpc_coeff(vector, unq_lpc);
2150 lpc2lsp(&unq_lpc[LPC_ORDER * 3], p->prev_lsp, cur_lsp);
2151 lsp_quantize(p->lsp_index, cur_lsp, p->prev_lsp);
2154 memcpy(vector + LPC_ORDER, p->prev_data + SUBFRAME_LEN,
2155 sizeof(int16_t) * SUBFRAME_LEN);
2156 memcpy(vector + LPC_ORDER + SUBFRAME_LEN, in,
2157 sizeof(int16_t) * (HALF_FRAME_LEN + SUBFRAME_LEN));
2158 memcpy(p->prev_data, in + HALF_FRAME_LEN,
2159 sizeof(int16_t) * HALF_FRAME_LEN);
2160 memcpy(in, vector + LPC_ORDER, sizeof(int16_t) * FRAME_LEN);
2162 perceptual_filter(p, weighted_lpc, unq_lpc, vector);
2164 memcpy(in, vector + LPC_ORDER, sizeof(int16_t) * FRAME_LEN);
2165 memcpy(vector, p->prev_weight_sig, sizeof(int16_t) * PITCH_MAX);
2166 memcpy(vector + PITCH_MAX, in, sizeof(int16_t) * FRAME_LEN);
2168 scale_vector(vector, vector, FRAME_LEN + PITCH_MAX);
2170 p->pitch_lag[0] = estimate_pitch(vector, PITCH_MAX);
2171 p->pitch_lag[1] = estimate_pitch(vector, PITCH_MAX + HALF_FRAME_LEN);
2173 for (i = PITCH_MAX, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
2174 comp_harmonic_coeff(vector + i, p->pitch_lag[j >> 1], hf + j);
2176 memcpy(vector, p->prev_weight_sig, sizeof(int16_t) * PITCH_MAX);
2177 memcpy(vector + PITCH_MAX, in, sizeof(int16_t) * FRAME_LEN);
2178 memcpy(p->prev_weight_sig, vector + FRAME_LEN, sizeof(int16_t) * PITCH_MAX);
2180 for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
2181 harmonic_filter(hf + j, vector + PITCH_MAX + i, in + i);
2183 inverse_quant(cur_lsp, p->prev_lsp, p->lsp_index, 0);
2184 lsp_interpolate(qnt_lpc, cur_lsp, p->prev_lsp);
2186 memcpy(p->prev_lsp, cur_lsp, sizeof(int16_t) * LPC_ORDER);
2189 for (i = 0; i < SUBFRAMES; i++) {
2190 int16_t impulse_resp[SUBFRAME_LEN];
2191 int16_t residual[SUBFRAME_LEN + PITCH_ORDER - 1];
2192 int16_t flt_in[SUBFRAME_LEN];
2193 int16_t zero[LPC_ORDER], fir[LPC_ORDER], iir[LPC_ORDER];
2196 * Compute the combined impulse response of the synthesis filter,
2197 * formant perceptual weighting filter and harmonic noise shaping filter
2199 memset(zero, 0, sizeof(int16_t) * LPC_ORDER);
2200 memset(vector, 0, sizeof(int16_t) * PITCH_MAX);
2201 memset(flt_in, 0, sizeof(int16_t) * SUBFRAME_LEN);
2203 flt_in[0] = 1 << 13; /* Unit impulse */
2204 synth_percept_filter(qnt_lpc + offset, weighted_lpc + (offset << 1),
2205 zero, zero, flt_in, vector + PITCH_MAX, 1);
2206 harmonic_filter(hf + i, vector + PITCH_MAX, impulse_resp);
2208 /* Compute the combined zero input response */
2210 memcpy(fir, p->perf_fir_mem, sizeof(int16_t) * LPC_ORDER);
2211 memcpy(iir, p->perf_iir_mem, sizeof(int16_t) * LPC_ORDER);
2213 synth_percept_filter(qnt_lpc + offset, weighted_lpc + (offset << 1),
2214 fir, iir, flt_in, vector + PITCH_MAX, 0);
2215 memcpy(vector, p->harmonic_mem, sizeof(int16_t) * PITCH_MAX);
2216 harmonic_noise_sub(hf + i, vector + PITCH_MAX, in);
2218 acb_search(p, residual, impulse_resp, in, i);
2219 gen_acb_excitation(residual, p->prev_excitation,p->pitch_lag[i >> 1],
2220 &p->subframe[i], p->cur_rate);
2221 sub_acb_contrib(residual, impulse_resp, in);
2223 fcb_search(p, impulse_resp, in, i);
2225 /* Reconstruct the excitation */
2226 gen_acb_excitation(impulse_resp, p->prev_excitation, p->pitch_lag[i >> 1],
2227 &p->subframe[i], RATE_6300);
2229 memmove(p->prev_excitation, p->prev_excitation + SUBFRAME_LEN,
2230 sizeof(int16_t) * (PITCH_MAX - SUBFRAME_LEN));
2231 for (j = 0; j < SUBFRAME_LEN; j++)
2232 in[j] = av_clip_int16((in[j] << 1) + impulse_resp[j]);
2233 memcpy(p->prev_excitation + PITCH_MAX - SUBFRAME_LEN, in,
2234 sizeof(int16_t) * SUBFRAME_LEN);
2236 /* Update filter memories */
2237 synth_percept_filter(qnt_lpc + offset, weighted_lpc + (offset << 1),
2238 p->perf_fir_mem, p->perf_iir_mem,
2239 in, vector + PITCH_MAX, 0);
2240 memmove(p->harmonic_mem, p->harmonic_mem + SUBFRAME_LEN,
2241 sizeof(int16_t) * (PITCH_MAX - SUBFRAME_LEN));
2242 memcpy(p->harmonic_mem + PITCH_MAX - SUBFRAME_LEN, vector + PITCH_MAX,
2243 sizeof(int16_t) * SUBFRAME_LEN);
2246 offset += LPC_ORDER;
2249 if ((ret = ff_alloc_packet2(avctx, avpkt, 24)))
2252 *got_packet_ptr = 1;
2253 avpkt->size = pack_bitstream(p, avpkt->data, avpkt->size);
2257 AVCodec ff_g723_1_encoder = {
2259 .type = AVMEDIA_TYPE_AUDIO,
2260 .id = AV_CODEC_ID_G723_1,
2261 .priv_data_size = sizeof(G723_1_Context),
2262 .init = g723_1_encode_init,
2263 .encode2 = g723_1_encode_frame,
2264 .long_name = NULL_IF_CONFIG_SMALL("G.723.1"),
2265 .sample_fmts = (const enum AVSampleFormat[]){AV_SAMPLE_FMT_S16,
2266 AV_SAMPLE_FMT_NONE},