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->channel_layout = AV_CH_LAYOUT_MONO;
87 avctx->sample_fmt = AV_SAMPLE_FMT_S16;
91 avcodec_get_frame_defaults(&p->frame);
92 avctx->coded_frame = &p->frame;
94 memcpy(p->prev_lsp, dc_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
100 * Unpack the frame into parameters.
102 * @param p the context
103 * @param buf pointer to the input buffer
104 * @param buf_size size of the input buffer
106 static int unpack_bitstream(G723_1_Context *p, const uint8_t *buf,
111 int temp, info_bits, i;
113 init_get_bits(&gb, buf, buf_size * 8);
115 /* Extract frame type and rate info */
116 info_bits = get_bits(&gb, 2);
118 if (info_bits == 3) {
119 p->cur_frame_type = UNTRANSMITTED_FRAME;
123 /* Extract 24 bit lsp indices, 8 bit for each band */
124 p->lsp_index[2] = get_bits(&gb, 8);
125 p->lsp_index[1] = get_bits(&gb, 8);
126 p->lsp_index[0] = get_bits(&gb, 8);
128 if (info_bits == 2) {
129 p->cur_frame_type = SID_FRAME;
130 p->subframe[0].amp_index = get_bits(&gb, 6);
134 /* Extract the info common to both rates */
135 p->cur_rate = info_bits ? RATE_5300 : RATE_6300;
136 p->cur_frame_type = ACTIVE_FRAME;
138 p->pitch_lag[0] = get_bits(&gb, 7);
139 if (p->pitch_lag[0] > 123) /* test if forbidden code */
141 p->pitch_lag[0] += PITCH_MIN;
142 p->subframe[1].ad_cb_lag = get_bits(&gb, 2);
144 p->pitch_lag[1] = get_bits(&gb, 7);
145 if (p->pitch_lag[1] > 123)
147 p->pitch_lag[1] += PITCH_MIN;
148 p->subframe[3].ad_cb_lag = get_bits(&gb, 2);
149 p->subframe[0].ad_cb_lag = 1;
150 p->subframe[2].ad_cb_lag = 1;
152 for (i = 0; i < SUBFRAMES; i++) {
153 /* Extract combined gain */
154 temp = get_bits(&gb, 12);
156 p->subframe[i].dirac_train = 0;
157 if (p->cur_rate == RATE_6300 && p->pitch_lag[i >> 1] < SUBFRAME_LEN - 2) {
158 p->subframe[i].dirac_train = temp >> 11;
162 p->subframe[i].ad_cb_gain = FASTDIV(temp, GAIN_LEVELS);
163 if (p->subframe[i].ad_cb_gain < ad_cb_len) {
164 p->subframe[i].amp_index = temp - p->subframe[i].ad_cb_gain *
171 p->subframe[0].grid_index = get_bits1(&gb);
172 p->subframe[1].grid_index = get_bits1(&gb);
173 p->subframe[2].grid_index = get_bits1(&gb);
174 p->subframe[3].grid_index = get_bits1(&gb);
176 if (p->cur_rate == RATE_6300) {
177 skip_bits1(&gb); /* skip reserved bit */
179 /* Compute pulse_pos index using the 13-bit combined position index */
180 temp = get_bits(&gb, 13);
181 p->subframe[0].pulse_pos = temp / 810;
183 temp -= p->subframe[0].pulse_pos * 810;
184 p->subframe[1].pulse_pos = FASTDIV(temp, 90);
186 temp -= p->subframe[1].pulse_pos * 90;
187 p->subframe[2].pulse_pos = FASTDIV(temp, 9);
188 p->subframe[3].pulse_pos = temp - p->subframe[2].pulse_pos * 9;
190 p->subframe[0].pulse_pos = (p->subframe[0].pulse_pos << 16) +
192 p->subframe[1].pulse_pos = (p->subframe[1].pulse_pos << 14) +
194 p->subframe[2].pulse_pos = (p->subframe[2].pulse_pos << 16) +
196 p->subframe[3].pulse_pos = (p->subframe[3].pulse_pos << 14) +
199 p->subframe[0].pulse_sign = get_bits(&gb, 6);
200 p->subframe[1].pulse_sign = get_bits(&gb, 5);
201 p->subframe[2].pulse_sign = get_bits(&gb, 6);
202 p->subframe[3].pulse_sign = get_bits(&gb, 5);
203 } else { /* 5300 bps */
204 p->subframe[0].pulse_pos = get_bits(&gb, 12);
205 p->subframe[1].pulse_pos = get_bits(&gb, 12);
206 p->subframe[2].pulse_pos = get_bits(&gb, 12);
207 p->subframe[3].pulse_pos = get_bits(&gb, 12);
209 p->subframe[0].pulse_sign = get_bits(&gb, 4);
210 p->subframe[1].pulse_sign = get_bits(&gb, 4);
211 p->subframe[2].pulse_sign = get_bits(&gb, 4);
212 p->subframe[3].pulse_sign = get_bits(&gb, 4);
219 * Bitexact implementation of sqrt(val/2).
221 static int16_t square_root(int val)
223 return (ff_sqrt(val << 1) >> 1) & (~1);
227 * Calculate the number of left-shifts required for normalizing the input.
229 * @param num input number
230 * @param width width of the input, 15 or 31 bits
232 static int normalize_bits(int num, int width)
241 i= width - av_log2(num) - 1;
247 #define normalize_bits_int16(num) normalize_bits(num, 15)
248 #define normalize_bits_int32(num) normalize_bits(num, 31)
249 #define dot_product(a,b,c,d) (ff_dot_product(a,b,c)<<(d))
252 * Scale vector contents based on the largest of their absolutes.
254 static int scale_vector(int16_t *vector, int length)
256 int bits, scale, max = 0;
259 const int16_t shift_table[16] = {
260 0x0001, 0x0002, 0x0004, 0x0008, 0x0010, 0x0020, 0x0040, 0x0080,
261 0x0100, 0x0200, 0x0400, 0x0800, 0x1000, 0x2000, 0x4000, 0x7fff
264 for (i = 0; i < length; i++)
265 max = FFMAX(max, FFABS(vector[i]));
267 bits = normalize_bits(max, 15);
268 scale = shift_table[bits];
270 for (i = 0; i < length; i++)
271 vector[i] = (vector[i] * scale) >> 3;
277 * Perform inverse quantization of LSP frequencies.
279 * @param cur_lsp the current LSP vector
280 * @param prev_lsp the previous LSP vector
281 * @param lsp_index VQ indices
282 * @param bad_frame bad frame flag
284 static void inverse_quant(int16_t *cur_lsp, int16_t *prev_lsp,
285 uint8_t *lsp_index, int bad_frame)
288 int i, j, temp, stable;
290 /* Check for frame erasure */
297 lsp_index[0] = lsp_index[1] = lsp_index[2] = 0;
300 /* Get the VQ table entry corresponding to the transmitted index */
301 cur_lsp[0] = lsp_band0[lsp_index[0]][0];
302 cur_lsp[1] = lsp_band0[lsp_index[0]][1];
303 cur_lsp[2] = lsp_band0[lsp_index[0]][2];
304 cur_lsp[3] = lsp_band1[lsp_index[1]][0];
305 cur_lsp[4] = lsp_band1[lsp_index[1]][1];
306 cur_lsp[5] = lsp_band1[lsp_index[1]][2];
307 cur_lsp[6] = lsp_band2[lsp_index[2]][0];
308 cur_lsp[7] = lsp_band2[lsp_index[2]][1];
309 cur_lsp[8] = lsp_band2[lsp_index[2]][2];
310 cur_lsp[9] = lsp_band2[lsp_index[2]][3];
312 /* Add predicted vector & DC component to the previously quantized vector */
313 for (i = 0; i < LPC_ORDER; i++) {
314 temp = ((prev_lsp[i] - dc_lsp[i]) * pred + (1 << 14)) >> 15;
315 cur_lsp[i] += dc_lsp[i] + temp;
318 for (i = 0; i < LPC_ORDER; i++) {
319 cur_lsp[0] = FFMAX(cur_lsp[0], 0x180);
320 cur_lsp[LPC_ORDER - 1] = FFMIN(cur_lsp[LPC_ORDER - 1], 0x7e00);
322 /* Stability check */
323 for (j = 1; j < LPC_ORDER; j++) {
324 temp = min_dist + cur_lsp[j - 1] - cur_lsp[j];
327 cur_lsp[j - 1] -= temp;
332 for (j = 1; j < LPC_ORDER; j++) {
333 temp = cur_lsp[j - 1] + min_dist - cur_lsp[j] - 4;
343 memcpy(cur_lsp, prev_lsp, LPC_ORDER * sizeof(*cur_lsp));
347 * Bitexact implementation of 2ab scaled by 1/2^16.
349 * @param a 32 bit multiplicand
350 * @param b 16 bit multiplier
352 #define MULL2(a, b) \
356 * Convert LSP frequencies to LPC coefficients.
358 * @param lpc buffer for LPC coefficients
360 static void lsp2lpc(int16_t *lpc)
362 int f1[LPC_ORDER / 2 + 1];
363 int f2[LPC_ORDER / 2 + 1];
366 /* Calculate negative cosine */
367 for (j = 0; j < LPC_ORDER; j++) {
368 int index = lpc[j] >> 7;
369 int offset = lpc[j] & 0x7f;
370 int64_t temp1 = cos_tab[index] << 16;
371 int temp2 = (cos_tab[index + 1] - cos_tab[index]) *
372 ((offset << 8) + 0x80) << 1;
374 lpc[j] = -(av_clipl_int32(((temp1 + temp2) << 1) + (1 << 15)) >> 16);
378 * Compute sum and difference polynomial coefficients
379 * (bitexact alternative to lsp2poly() in lsp.c)
381 /* Initialize with values in Q28 */
383 f1[1] = (lpc[0] << 14) + (lpc[2] << 14);
384 f1[2] = lpc[0] * lpc[2] + (2 << 28);
387 f2[1] = (lpc[1] << 14) + (lpc[3] << 14);
388 f2[2] = lpc[1] * lpc[3] + (2 << 28);
391 * Calculate and scale the coefficients by 1/2 in
392 * each iteration for a final scaling factor of Q25
394 for (i = 2; i < LPC_ORDER / 2; i++) {
395 f1[i + 1] = f1[i - 1] + MULL2(f1[i], lpc[2 * i]);
396 f2[i + 1] = f2[i - 1] + MULL2(f2[i], lpc[2 * i + 1]);
398 for (j = i; j >= 2; j--) {
399 f1[j] = MULL2(f1[j - 1], lpc[2 * i]) +
400 (f1[j] >> 1) + (f1[j - 2] >> 1);
401 f2[j] = MULL2(f2[j - 1], lpc[2 * i + 1]) +
402 (f2[j] >> 1) + (f2[j - 2] >> 1);
407 f1[1] = ((lpc[2 * i] << 16 >> i) + f1[1]) >> 1;
408 f2[1] = ((lpc[2 * i + 1] << 16 >> i) + f2[1]) >> 1;
411 /* Convert polynomial coefficients to LPC coefficients */
412 for (i = 0; i < LPC_ORDER / 2; i++) {
413 int64_t ff1 = f1[i + 1] + f1[i];
414 int64_t ff2 = f2[i + 1] - f2[i];
416 lpc[i] = av_clipl_int32(((ff1 + ff2) << 3) + (1 << 15)) >> 16;
417 lpc[LPC_ORDER - i - 1] = av_clipl_int32(((ff1 - ff2) << 3) +
423 * Quantize LSP frequencies by interpolation and convert them to
424 * the corresponding LPC coefficients.
426 * @param lpc buffer for LPC coefficients
427 * @param cur_lsp the current LSP vector
428 * @param prev_lsp the previous LSP vector
430 static void lsp_interpolate(int16_t *lpc, int16_t *cur_lsp, int16_t *prev_lsp)
433 int16_t *lpc_ptr = lpc;
435 /* cur_lsp * 0.25 + prev_lsp * 0.75 */
436 ff_acelp_weighted_vector_sum(lpc, cur_lsp, prev_lsp,
437 4096, 12288, 1 << 13, 14, LPC_ORDER);
438 ff_acelp_weighted_vector_sum(lpc + LPC_ORDER, cur_lsp, prev_lsp,
439 8192, 8192, 1 << 13, 14, LPC_ORDER);
440 ff_acelp_weighted_vector_sum(lpc + 2 * LPC_ORDER, cur_lsp, prev_lsp,
441 12288, 4096, 1 << 13, 14, LPC_ORDER);
442 memcpy(lpc + 3 * LPC_ORDER, cur_lsp, LPC_ORDER * sizeof(*lpc));
444 for (i = 0; i < SUBFRAMES; i++) {
446 lpc_ptr += LPC_ORDER;
451 * Generate a train of dirac functions with period as pitch lag.
453 static void gen_dirac_train(int16_t *buf, int pitch_lag)
455 int16_t vector[SUBFRAME_LEN];
458 memcpy(vector, buf, SUBFRAME_LEN * sizeof(*vector));
459 for (i = pitch_lag; i < SUBFRAME_LEN; i += pitch_lag) {
460 for (j = 0; j < SUBFRAME_LEN - i; j++)
461 buf[i + j] += vector[j];
466 * Generate fixed codebook excitation vector.
468 * @param vector decoded excitation vector
469 * @param subfrm current subframe
470 * @param cur_rate current bitrate
471 * @param pitch_lag closed loop pitch lag
472 * @param index current subframe index
474 static void gen_fcb_excitation(int16_t *vector, G723_1_Subframe subfrm,
475 enum Rate cur_rate, int pitch_lag, int index)
479 memset(vector, 0, SUBFRAME_LEN * sizeof(*vector));
481 if (cur_rate == RATE_6300) {
482 if (subfrm.pulse_pos >= max_pos[index])
485 /* Decode amplitudes and positions */
486 j = PULSE_MAX - pulses[index];
487 temp = subfrm.pulse_pos;
488 for (i = 0; i < SUBFRAME_LEN / GRID_SIZE; i++) {
489 temp -= combinatorial_table[j][i];
492 temp += combinatorial_table[j++][i];
493 if (subfrm.pulse_sign & (1 << (PULSE_MAX - j))) {
494 vector[subfrm.grid_index + GRID_SIZE * i] =
495 -fixed_cb_gain[subfrm.amp_index];
497 vector[subfrm.grid_index + GRID_SIZE * i] =
498 fixed_cb_gain[subfrm.amp_index];
503 if (subfrm.dirac_train == 1)
504 gen_dirac_train(vector, pitch_lag);
505 } else { /* 5300 bps */
506 int cb_gain = fixed_cb_gain[subfrm.amp_index];
507 int cb_shift = subfrm.grid_index;
508 int cb_sign = subfrm.pulse_sign;
509 int cb_pos = subfrm.pulse_pos;
510 int offset, beta, lag;
512 for (i = 0; i < 8; i += 2) {
513 offset = ((cb_pos & 7) << 3) + cb_shift + i;
514 vector[offset] = (cb_sign & 1) ? cb_gain : -cb_gain;
519 /* Enhance harmonic components */
520 lag = pitch_contrib[subfrm.ad_cb_gain << 1] + pitch_lag +
521 subfrm.ad_cb_lag - 1;
522 beta = pitch_contrib[(subfrm.ad_cb_gain << 1) + 1];
524 if (lag < SUBFRAME_LEN - 2) {
525 for (i = lag; i < SUBFRAME_LEN; i++)
526 vector[i] += beta * vector[i - lag] >> 15;
532 * Get delayed contribution from the previous excitation vector.
534 static void get_residual(int16_t *residual, int16_t *prev_excitation, int lag)
536 int offset = PITCH_MAX - PITCH_ORDER / 2 - lag;
539 residual[0] = prev_excitation[offset];
540 residual[1] = prev_excitation[offset + 1];
543 for (i = 2; i < SUBFRAME_LEN + PITCH_ORDER - 1; i++)
544 residual[i] = prev_excitation[offset + (i - 2) % lag];
548 * Generate adaptive codebook excitation.
550 static void gen_acb_excitation(int16_t *vector, int16_t *prev_excitation,
551 int pitch_lag, G723_1_Subframe subfrm,
554 int16_t residual[SUBFRAME_LEN + PITCH_ORDER - 1];
555 const int16_t *cb_ptr;
556 int lag = pitch_lag + subfrm.ad_cb_lag - 1;
561 get_residual(residual, prev_excitation, lag);
563 /* Select quantization table */
564 if (cur_rate == RATE_6300 && pitch_lag < SUBFRAME_LEN - 2) {
565 cb_ptr = adaptive_cb_gain85;
567 cb_ptr = adaptive_cb_gain170;
569 /* Calculate adaptive vector */
570 cb_ptr += subfrm.ad_cb_gain * 20;
571 for (i = 0; i < SUBFRAME_LEN; i++) {
572 sum = ff_dot_product(residual + i, cb_ptr, PITCH_ORDER);
573 vector[i] = av_clipl_int32((sum << 2) + (1 << 15)) >> 16;
578 * Estimate maximum auto-correlation around pitch lag.
580 * @param p the context
581 * @param offset offset of the excitation vector
582 * @param ccr_max pointer to the maximum auto-correlation
583 * @param pitch_lag decoded pitch lag
584 * @param length length of autocorrelation
585 * @param dir forward lag(1) / backward lag(-1)
587 static int autocorr_max(G723_1_Context *p, int offset, int *ccr_max,
588 int pitch_lag, int length, int dir)
590 int limit, ccr, lag = 0;
591 int16_t *buf = p->excitation + offset;
594 pitch_lag = FFMIN(PITCH_MAX - 3, pitch_lag);
595 limit = FFMIN(FRAME_LEN + PITCH_MAX - offset - length, pitch_lag + 3);
597 for (i = pitch_lag - 3; i <= limit; i++) {
598 ccr = ff_dot_product(buf, buf + dir * i, length)<<1;
600 if (ccr > *ccr_max) {
609 * Calculate pitch postfilter optimal and scaling gains.
611 * @param lag pitch postfilter forward/backward lag
612 * @param ppf pitch postfilter parameters
613 * @param cur_rate current bitrate
614 * @param tgt_eng target energy
615 * @param ccr cross-correlation
616 * @param res_eng residual energy
618 static void comp_ppf_gains(int lag, PPFParam *ppf, enum Rate cur_rate,
619 int tgt_eng, int ccr, int res_eng)
621 int pf_residual; /* square of postfiltered residual */
622 int64_t temp1, temp2;
626 temp1 = tgt_eng * res_eng >> 1;
627 temp2 = ccr * ccr << 1;
630 if (ccr >= res_eng) {
631 ppf->opt_gain = ppf_gain_weight[cur_rate];
633 ppf->opt_gain = (ccr << 15) / res_eng *
634 ppf_gain_weight[cur_rate] >> 15;
636 /* pf_res^2 = tgt_eng + 2*ccr*gain + res_eng*gain^2 */
637 temp1 = (tgt_eng << 15) + (ccr * ppf->opt_gain << 1);
638 temp2 = (ppf->opt_gain * ppf->opt_gain >> 15) * res_eng;
639 pf_residual = av_clipl_int32(temp1 + temp2 + (1 << 15)) >> 16;
641 if (tgt_eng >= pf_residual << 1) {
644 temp1 = (tgt_eng << 14) / pf_residual;
647 /* scaling_gain = sqrt(tgt_eng/pf_res^2) */
648 ppf->sc_gain = square_root(temp1 << 16);
651 ppf->sc_gain = 0x7fff;
654 ppf->opt_gain = av_clip_int16(ppf->opt_gain * ppf->sc_gain >> 15);
658 * Calculate pitch postfilter parameters.
660 * @param p the context
661 * @param offset offset of the excitation vector
662 * @param pitch_lag decoded pitch lag
663 * @param ppf pitch postfilter parameters
664 * @param cur_rate current bitrate
666 static void comp_ppf_coeff(G723_1_Context *p, int offset, int pitch_lag,
667 PPFParam *ppf, enum Rate cur_rate)
672 int64_t temp1, temp2;
676 * 1 - forward cross-correlation
677 * 2 - forward residual energy
678 * 3 - backward cross-correlation
679 * 4 - backward residual energy
681 int energy[5] = {0, 0, 0, 0, 0};
682 int16_t *buf = p->excitation + offset;
683 int fwd_lag = autocorr_max(p, offset, &energy[1], pitch_lag,
685 int back_lag = autocorr_max(p, offset, &energy[3], pitch_lag,
690 ppf->sc_gain = 0x7fff;
692 /* Case 0, Section 3.6 */
693 if (!back_lag && !fwd_lag)
696 /* Compute target energy */
697 energy[0] = ff_dot_product(buf, buf, SUBFRAME_LEN)<<1;
699 /* Compute forward residual energy */
701 energy[2] = ff_dot_product(buf + fwd_lag, buf + fwd_lag,
704 /* Compute backward residual energy */
706 energy[4] = ff_dot_product(buf - back_lag, buf - back_lag,
709 /* Normalize and shorten */
711 for (i = 0; i < 5; i++)
712 temp1 = FFMAX(energy[i], temp1);
714 scale = normalize_bits(temp1, 31);
715 for (i = 0; i < 5; i++)
716 energy[i] = av_clipl_int32(energy[i] << scale) >> 16;
718 if (fwd_lag && !back_lag) { /* Case 1 */
719 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
721 } else if (!fwd_lag) { /* Case 2 */
722 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
724 } else { /* Case 3 */
727 * Select the largest of energy[1]^2/energy[2]
728 * and energy[3]^2/energy[4]
730 temp1 = energy[4] * ((energy[1] * energy[1] + (1 << 14)) >> 15);
731 temp2 = energy[2] * ((energy[3] * energy[3] + (1 << 14)) >> 15);
732 if (temp1 >= temp2) {
733 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
736 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
743 * Classify frames as voiced/unvoiced.
745 * @param p the context
746 * @param pitch_lag decoded pitch_lag
747 * @param exc_eng excitation energy estimation
748 * @param scale scaling factor of exc_eng
750 * @return residual interpolation index if voiced, 0 otherwise
752 static int comp_interp_index(G723_1_Context *p, int pitch_lag,
753 int *exc_eng, int *scale)
755 int offset = PITCH_MAX + 2 * SUBFRAME_LEN;
756 int16_t *buf = p->excitation + offset;
758 int index, ccr, tgt_eng, best_eng, temp;
760 *scale = scale_vector(p->excitation, FRAME_LEN + PITCH_MAX);
762 /* Compute maximum backward cross-correlation */
764 index = autocorr_max(p, offset, &ccr, pitch_lag, SUBFRAME_LEN * 2, -1);
765 ccr = av_clipl_int32((int64_t)ccr + (1 << 15)) >> 16;
767 /* Compute target energy */
768 tgt_eng = ff_dot_product(buf, buf, SUBFRAME_LEN * 2)<<1;
769 *exc_eng = av_clipl_int32(tgt_eng + (1 << 15)) >> 16;
774 /* Compute best energy */
775 best_eng = ff_dot_product(buf - index, buf - index,
776 SUBFRAME_LEN * 2)<<1;
777 best_eng = av_clipl_int32((int64_t)best_eng + (1 << 15)) >> 16;
779 temp = best_eng * *exc_eng >> 3;
781 if (temp < ccr * ccr) {
788 * Peform residual interpolation based on frame classification.
790 * @param buf decoded excitation vector
791 * @param out output vector
792 * @param lag decoded pitch lag
793 * @param gain interpolated gain
794 * @param rseed seed for random number generator
796 static void residual_interp(int16_t *buf, int16_t *out, int lag,
797 int gain, int *rseed)
800 if (lag) { /* Voiced */
801 int16_t *vector_ptr = buf + PITCH_MAX;
803 for (i = 0; i < lag; i++)
804 vector_ptr[i - lag] = vector_ptr[i - lag] * 3 >> 2;
805 av_memcpy_backptr((uint8_t*)vector_ptr, lag * sizeof(*vector_ptr),
806 FRAME_LEN * sizeof(*vector_ptr));
807 memcpy(out, vector_ptr, FRAME_LEN * sizeof(*vector_ptr));
808 } else { /* Unvoiced */
809 for (i = 0; i < FRAME_LEN; i++) {
810 *rseed = *rseed * 521 + 259;
811 out[i] = gain * *rseed >> 15;
813 memset(buf, 0, (FRAME_LEN + PITCH_MAX) * sizeof(*buf));
818 * Perform IIR filtering.
820 * @param fir_coef FIR coefficients
821 * @param iir_coef IIR coefficients
822 * @param src source vector
823 * @param dest destination vector
824 * @param width width of the output, 16 bits(0) / 32 bits(1)
826 #define iir_filter(fir_coef, iir_coef, src, dest, width)\
829 int res_shift = 16 & ~-(width);\
830 int in_shift = 16 - res_shift;\
832 for (m = 0; m < SUBFRAME_LEN; m++) {\
834 for (n = 1; n <= LPC_ORDER; n++) {\
835 filter -= (fir_coef)[n - 1] * (src)[m - n] -\
836 (iir_coef)[n - 1] * ((dest)[m - n] >> in_shift);\
839 (dest)[m] = av_clipl_int32(((src)[m] << 16) + (filter << 3) +\
840 (1 << 15)) >> res_shift;\
845 * Adjust gain of postfiltered signal.
847 * @param p the context
848 * @param buf postfiltered output vector
849 * @param energy input energy coefficient
851 static void gain_scale(G723_1_Context *p, int16_t * buf, int energy)
853 int num, denom, gain, bits1, bits2;
858 for (i = 0; i < SUBFRAME_LEN; i++) {
859 int64_t temp = buf[i] >> 2;
860 temp = av_clipl_int32(MUL64(temp, temp) << 1);
861 denom = av_clipl_int32(denom + temp);
865 bits1 = normalize_bits(num, 31);
866 bits2 = normalize_bits(denom, 31);
867 num = num << bits1 >> 1;
870 bits2 = 5 + bits1 - bits2;
871 bits2 = FFMAX(0, bits2);
873 gain = (num >> 1) / (denom >> 16);
874 gain = square_root(gain << 16 >> bits2);
879 for (i = 0; i < SUBFRAME_LEN; i++) {
880 p->pf_gain = ((p->pf_gain << 4) - p->pf_gain + gain + (1 << 3)) >> 4;
881 buf[i] = av_clip_int16((buf[i] * (p->pf_gain + (p->pf_gain >> 4)) +
887 * Perform formant filtering.
889 * @param p the context
890 * @param lpc quantized lpc coefficients
891 * @param buf output buffer
893 static void formant_postfilter(G723_1_Context *p, int16_t *lpc, int16_t *buf)
895 int16_t filter_coef[2][LPC_ORDER], *buf_ptr;
896 int filter_signal[LPC_ORDER + FRAME_LEN], *signal_ptr;
899 memcpy(buf, p->fir_mem, LPC_ORDER * sizeof(*buf));
900 memcpy(filter_signal, p->iir_mem, LPC_ORDER * sizeof(*filter_signal));
902 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
903 for (k = 0; k < LPC_ORDER; k++) {
904 filter_coef[0][k] = (-lpc[k] * postfilter_tbl[0][k] +
906 filter_coef[1][k] = (-lpc[k] * postfilter_tbl[1][k] +
909 iir_filter(filter_coef[0], filter_coef[1], buf + i,
910 filter_signal + i, 1);
913 memcpy(p->fir_mem, buf + FRAME_LEN, LPC_ORDER * sizeof(int16_t));
914 memcpy(p->iir_mem, filter_signal + FRAME_LEN, LPC_ORDER * sizeof(int));
916 buf_ptr = buf + LPC_ORDER;
917 signal_ptr = filter_signal + LPC_ORDER;
918 for (i = 0; i < SUBFRAMES; i++) {
919 int16_t temp_vector[SUBFRAME_LEN];
925 memcpy(temp_vector, buf_ptr, SUBFRAME_LEN * sizeof(*temp_vector));
926 scale = scale_vector(temp_vector, SUBFRAME_LEN);
928 /* Compute auto correlation coefficients */
929 auto_corr[0] = ff_dot_product(temp_vector, temp_vector + 1,
930 SUBFRAME_LEN - 1)<<1;
931 auto_corr[1] = ff_dot_product(temp_vector, temp_vector,
934 /* Compute reflection coefficient */
935 temp = auto_corr[1] >> 16;
937 temp = (auto_corr[0] >> 2) / temp;
939 p->reflection_coef = ((p->reflection_coef << 2) - p->reflection_coef +
941 temp = (p->reflection_coef * 0xffffc >> 3) & 0xfffc;
943 /* Compensation filter */
944 for (j = 0; j < SUBFRAME_LEN; j++) {
945 buf_ptr[j] = av_clipl_int32(signal_ptr[j] +
946 ((signal_ptr[j - 1] >> 16) *
950 /* Compute normalized signal energy */
951 temp = 2 * scale + 4;
953 energy = av_clipl_int32((int64_t)auto_corr[1] << -temp);
955 energy = auto_corr[1] >> temp;
957 gain_scale(p, buf_ptr, energy);
959 buf_ptr += SUBFRAME_LEN;
960 signal_ptr += SUBFRAME_LEN;
964 static int g723_1_decode_frame(AVCodecContext *avctx, void *data,
965 int *got_frame_ptr, AVPacket *avpkt)
967 G723_1_Context *p = avctx->priv_data;
968 const uint8_t *buf = avpkt->data;
969 int buf_size = avpkt->size;
971 int dec_mode = buf[0] & 3;
973 PPFParam ppf[SUBFRAMES];
974 int16_t cur_lsp[LPC_ORDER];
975 int16_t lpc[SUBFRAMES * LPC_ORDER];
976 int16_t acb_vector[SUBFRAME_LEN];
978 int bad_frame = 0, i, j, ret;
980 if (!buf_size || buf_size < frame_size[dec_mode]) {
985 if (unpack_bitstream(p, buf, buf_size) < 0) {
987 if (p->past_frame_type == ACTIVE_FRAME)
988 p->cur_frame_type = ACTIVE_FRAME;
990 p->cur_frame_type = UNTRANSMITTED_FRAME;
993 p->frame.nb_samples = FRAME_LEN + LPC_ORDER;
994 if ((ret = avctx->get_buffer(avctx, &p->frame)) < 0) {
995 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
998 out= (int16_t*)p->frame.data[0];
1001 if (p->cur_frame_type == ACTIVE_FRAME) {
1003 p->erased_frames = 0;
1004 else if (p->erased_frames != 3)
1007 inverse_quant(cur_lsp, p->prev_lsp, p->lsp_index, bad_frame);
1008 lsp_interpolate(lpc, cur_lsp, p->prev_lsp);
1010 /* Save the lsp_vector for the next frame */
1011 memcpy(p->prev_lsp, cur_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
1013 /* Generate the excitation for the frame */
1014 memcpy(p->excitation, p->prev_excitation,
1015 PITCH_MAX * sizeof(*p->excitation));
1016 vector_ptr = p->excitation + PITCH_MAX;
1017 if (!p->erased_frames) {
1018 /* Update interpolation gain memory */
1019 p->interp_gain = fixed_cb_gain[(p->subframe[2].amp_index +
1020 p->subframe[3].amp_index) >> 1];
1021 for (i = 0; i < SUBFRAMES; i++) {
1022 gen_fcb_excitation(vector_ptr, p->subframe[i], p->cur_rate,
1023 p->pitch_lag[i >> 1], i);
1024 gen_acb_excitation(acb_vector, &p->excitation[SUBFRAME_LEN * i],
1025 p->pitch_lag[i >> 1], p->subframe[i],
1027 /* Get the total excitation */
1028 for (j = 0; j < SUBFRAME_LEN; j++) {
1029 vector_ptr[j] = av_clip_int16(vector_ptr[j] << 1);
1030 vector_ptr[j] = av_clip_int16(vector_ptr[j] +
1033 vector_ptr += SUBFRAME_LEN;
1036 vector_ptr = p->excitation + PITCH_MAX;
1038 /* Save the excitation */
1039 memcpy(out, vector_ptr, FRAME_LEN * sizeof(int16_t));
1041 p->interp_index = comp_interp_index(p, p->pitch_lag[1],
1042 &p->sid_gain, &p->cur_gain);
1044 if (p->postfilter) {
1046 for (j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1047 comp_ppf_coeff(p, i, p->pitch_lag[j >> 1],
1048 ppf + j, p->cur_rate);
1051 /* Restore the original excitation */
1052 memcpy(p->excitation, p->prev_excitation,
1053 PITCH_MAX * sizeof(*p->excitation));
1054 memcpy(vector_ptr, out, FRAME_LEN * sizeof(*vector_ptr));
1056 /* Peform pitch postfiltering */
1058 for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1059 ff_acelp_weighted_vector_sum(out + LPC_ORDER + i,
1061 vector_ptr + i + ppf[j].index,
1064 1 << 14, 15, SUBFRAME_LEN);
1066 p->interp_gain = (p->interp_gain * 3 + 2) >> 2;
1067 if (p->erased_frames == 3) {
1069 memset(p->excitation, 0,
1070 (FRAME_LEN + PITCH_MAX) * sizeof(*p->excitation));
1071 memset(out, 0, (FRAME_LEN + LPC_ORDER) * sizeof(int16_t));
1073 /* Regenerate frame */
1074 residual_interp(p->excitation, out + LPC_ORDER, p->interp_index,
1075 p->interp_gain, &p->random_seed);
1078 /* Save the excitation for the next frame */
1079 memcpy(p->prev_excitation, p->excitation + FRAME_LEN,
1080 PITCH_MAX * sizeof(*p->excitation));
1082 memset(out, 0, sizeof(int16_t)*FRAME_LEN);
1083 av_log(avctx, AV_LOG_WARNING,
1084 "G.723.1: Comfort noise generation not supported yet\n");
1085 return frame_size[dec_mode];
1088 p->past_frame_type = p->cur_frame_type;
1090 memcpy(out, p->synth_mem, LPC_ORDER * sizeof(int16_t));
1091 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1092 ff_celp_lp_synthesis_filter(out + i, &lpc[j * LPC_ORDER],
1093 out + i, SUBFRAME_LEN, LPC_ORDER,
1095 memcpy(p->synth_mem, out + FRAME_LEN, LPC_ORDER * sizeof(int16_t));
1098 formant_postfilter(p, lpc, out);
1100 memmove(out, out + LPC_ORDER, sizeof(int16_t)*FRAME_LEN);
1101 p->frame.nb_samples = FRAME_LEN;
1102 *(AVFrame*)data = p->frame;
1105 return frame_size[dec_mode];
1108 #define OFFSET(x) offsetof(G723_1_Context, x)
1109 #define AD AV_OPT_FLAG_AUDIO_PARAM | AV_OPT_FLAG_DECODING_PARAM
1111 static const AVOption options[] = {
1112 { "postfilter", "postfilter on/off", OFFSET(postfilter), AV_OPT_TYPE_INT,
1117 static const AVClass g723_1dec_class = {
1118 .class_name = "G.723.1 decoder",
1119 .item_name = av_default_item_name,
1121 .version = LIBAVUTIL_VERSION_INT,
1124 AVCodec ff_g723_1_decoder = {
1126 .type = AVMEDIA_TYPE_AUDIO,
1127 .id = CODEC_ID_G723_1,
1128 .priv_data_size = sizeof(G723_1_Context),
1129 .init = g723_1_decode_init,
1130 .decode = g723_1_decode_frame,
1131 .long_name = NULL_IF_CONFIG_SMALL("G.723.1"),
1132 .capabilities = CODEC_CAP_SUBFRAMES | CODEC_CAP_DR1,
1133 .priv_class = &g723_1dec_class,
1136 #if CONFIG_G723_1_ENCODER
1137 #define BITSTREAM_WRITER_LE
1138 #include "put_bits.h"
1140 static av_cold int g723_1_encode_init(AVCodecContext *avctx)
1142 G723_1_Context *p = avctx->priv_data;
1144 if (avctx->sample_rate != 8000) {
1145 av_log(avctx, AV_LOG_ERROR, "Only 8000Hz sample rate supported\n");
1149 if (avctx->channels != 1) {
1150 av_log(avctx, AV_LOG_ERROR, "Only mono supported\n");
1151 return AVERROR(EINVAL);
1154 if (avctx->bit_rate == 6300) {
1155 p->cur_rate = RATE_6300;
1156 } else if (avctx->bit_rate == 5300) {
1157 av_log(avctx, AV_LOG_ERROR, "Bitrate not supported yet, use 6.3k\n");
1158 return AVERROR_PATCHWELCOME;
1160 av_log(avctx, AV_LOG_ERROR,
1161 "Bitrate not supported, use 6.3k\n");
1162 return AVERROR(EINVAL);
1164 avctx->frame_size = 240;
1165 memcpy(p->prev_lsp, dc_lsp, LPC_ORDER * sizeof(int16_t));
1171 * Remove DC component from the input signal.
1173 * @param buf input signal
1174 * @param fir zero memory
1175 * @param iir pole memory
1177 static void highpass_filter(int16_t *buf, int16_t *fir, int *iir)
1180 for (i = 0; i < FRAME_LEN; i++) {
1181 *iir = (buf[i] << 15) + ((-*fir) << 15) + MULL2(*iir, 0x7f00);
1183 buf[i] = av_clipl_int32((int64_t)*iir + (1 << 15)) >> 16;
1188 * Estimate autocorrelation of the input vector.
1190 * @param buf input buffer
1191 * @param autocorr autocorrelation coefficients vector
1193 static void comp_autocorr(int16_t *buf, int16_t *autocorr)
1196 int16_t vector[LPC_FRAME];
1198 memcpy(vector, buf, LPC_FRAME * sizeof(int16_t));
1199 scale_vector(vector, LPC_FRAME);
1201 /* Apply the Hamming window */
1202 for (i = 0; i < LPC_FRAME; i++)
1203 vector[i] = (vector[i] * hamming_window[i] + (1 << 14)) >> 15;
1205 /* Compute the first autocorrelation coefficient */
1206 temp = dot_product(vector, vector, LPC_FRAME, 0);
1208 /* Apply a white noise correlation factor of (1025/1024) */
1212 scale = normalize_bits_int32(temp);
1213 autocorr[0] = av_clipl_int32((int64_t)(temp << scale) +
1216 /* Compute the remaining coefficients */
1218 memset(autocorr + 1, 0, LPC_ORDER * sizeof(int16_t));
1220 for (i = 1; i <= LPC_ORDER; i++) {
1221 temp = dot_product(vector, vector + i, LPC_FRAME - i, 0);
1222 temp = MULL2((temp << scale), binomial_window[i - 1]);
1223 autocorr[i] = av_clipl_int32((int64_t)temp + (1 << 15)) >> 16;
1229 * Use Levinson-Durbin recursion to compute LPC coefficients from
1230 * autocorrelation values.
1232 * @param lpc LPC coefficients vector
1233 * @param autocorr autocorrelation coefficients vector
1234 * @param error prediction error
1236 static void levinson_durbin(int16_t *lpc, int16_t *autocorr, int16_t error)
1238 int16_t vector[LPC_ORDER];
1239 int16_t partial_corr;
1242 memset(lpc, 0, LPC_ORDER * sizeof(int16_t));
1244 for (i = 0; i < LPC_ORDER; i++) {
1245 /* Compute the partial correlation coefficient */
1247 for (j = 0; j < i; j++)
1248 temp -= lpc[j] * autocorr[i - j - 1];
1249 temp = ((autocorr[i] << 13) + temp) << 3;
1251 if (FFABS(temp) >= (error << 16))
1254 partial_corr = temp / (error << 1);
1256 lpc[i] = av_clipl_int32((int64_t)(partial_corr << 14) +
1259 /* Update the prediction error */
1260 temp = MULL2(temp, partial_corr);
1261 error = av_clipl_int32((int64_t)(error << 16) - temp +
1264 memcpy(vector, lpc, i * sizeof(int16_t));
1265 for (j = 0; j < i; j++) {
1266 temp = partial_corr * vector[i - j - 1] << 1;
1267 lpc[j] = av_clipl_int32((int64_t)(lpc[j] << 16) - temp +
1274 * Calculate LPC coefficients for the current frame.
1276 * @param buf current frame
1277 * @param prev_data 2 trailing subframes of the previous frame
1278 * @param lpc LPC coefficients vector
1280 static void comp_lpc_coeff(int16_t *buf, int16_t *lpc)
1282 int16_t autocorr[(LPC_ORDER + 1) * SUBFRAMES];
1283 int16_t *autocorr_ptr = autocorr;
1284 int16_t *lpc_ptr = lpc;
1287 for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
1288 comp_autocorr(buf + i, autocorr_ptr);
1289 levinson_durbin(lpc_ptr, autocorr_ptr + 1, autocorr_ptr[0]);
1291 lpc_ptr += LPC_ORDER;
1292 autocorr_ptr += LPC_ORDER + 1;
1296 static void lpc2lsp(int16_t *lpc, int16_t *prev_lsp, int16_t *lsp)
1298 int f[LPC_ORDER + 2]; ///< coefficients of the sum and difference
1299 ///< polynomials (F1, F2) ordered as
1300 ///< f1[0], f2[0], ...., f1[5], f2[5]
1302 int max, shift, cur_val, prev_val, count, p;
1306 /* Initialize f1[0] and f2[0] to 1 in Q25 */
1307 for (i = 0; i < LPC_ORDER; i++)
1308 lsp[i] = (lpc[i] * bandwidth_expand[i] + (1 << 14)) >> 15;
1310 /* Apply bandwidth expansion on the LPC coefficients */
1311 f[0] = f[1] = 1 << 25;
1313 /* Compute the remaining coefficients */
1314 for (i = 0; i < LPC_ORDER / 2; i++) {
1316 f[2 * i + 2] = -f[2 * i] - ((lsp[i] + lsp[LPC_ORDER - 1 - i]) << 12);
1318 f[2 * i + 3] = f[2 * i + 1] - ((lsp[i] - lsp[LPC_ORDER - 1 - i]) << 12);
1321 /* Divide f1[5] and f2[5] by 2 for use in polynomial evaluation */
1323 f[LPC_ORDER + 1] >>= 1;
1325 /* Normalize and shorten */
1327 for (i = 1; i < LPC_ORDER + 2; i++)
1328 max = FFMAX(max, FFABS(f[i]));
1330 shift = normalize_bits_int32(max);
1332 for (i = 0; i < LPC_ORDER + 2; i++)
1333 f[i] = av_clipl_int32((int64_t)(f[i] << shift) + (1 << 15)) >> 16;
1336 * Evaluate F1 and F2 at uniform intervals of pi/256 along the
1337 * unit circle and check for zero crossings.
1341 for (i = 0; i <= LPC_ORDER / 2; i++)
1342 temp += f[2 * i] * cos_tab[0];
1343 prev_val = av_clipl_int32(temp << 1);
1345 for ( i = 1; i < COS_TBL_SIZE / 2; i++) {
1348 for (j = 0; j <= LPC_ORDER / 2; j++)
1349 temp += f[LPC_ORDER - 2 * j + p] * cos_tab[i * j % COS_TBL_SIZE];
1350 cur_val = av_clipl_int32(temp << 1);
1352 /* Check for sign change, indicating a zero crossing */
1353 if ((cur_val ^ prev_val) < 0) {
1354 int abs_cur = FFABS(cur_val);
1355 int abs_prev = FFABS(prev_val);
1356 int sum = abs_cur + abs_prev;
1358 shift = normalize_bits_int32(sum);
1360 abs_prev = abs_prev << shift >> 8;
1361 lsp[count++] = ((i - 1) << 7) + (abs_prev >> 1) / (sum >> 16);
1363 if (count == LPC_ORDER)
1366 /* Switch between sum and difference polynomials */
1371 for (j = 0; j <= LPC_ORDER / 2; j++){
1372 temp += f[LPC_ORDER - 2 * j + p] *
1373 cos_tab[i * j % COS_TBL_SIZE];
1375 cur_val = av_clipl_int32(temp<<1);
1380 if (count != LPC_ORDER)
1381 memcpy(lsp, prev_lsp, LPC_ORDER * sizeof(int16_t));
1385 * Quantize the current LSP subvector.
1387 * @param num band number
1388 * @param offset offset of the current subvector in an LPC_ORDER vector
1389 * @param size size of the current subvector
1391 #define get_index(num, offset, size) \
1393 int error, max = -1;\
1396 for (i = 0; i < LSP_CB_SIZE; i++) {\
1397 for (j = 0; j < size; j++){\
1398 temp[j] = (weight[j + (offset)] * lsp_band##num[i][j] +\
1401 error = dot_product(lsp + (offset), temp, size, 1) << 1;\
1402 error -= dot_product(lsp_band##num[i], temp, size, 1);\
1405 lsp_index[num] = i;\
1411 * Vector quantize the LSP frequencies.
1413 * @param lsp the current lsp vector
1414 * @param prev_lsp the previous lsp vector
1416 static void lsp_quantize(uint8_t *lsp_index, int16_t *lsp, int16_t *prev_lsp)
1418 int16_t weight[LPC_ORDER];
1422 /* Calculate the VQ weighting vector */
1423 weight[0] = (1 << 20) / (lsp[1] - lsp[0]);
1424 weight[LPC_ORDER - 1] = (1 << 20) /
1425 (lsp[LPC_ORDER - 1] - lsp[LPC_ORDER - 2]);
1427 for (i = 1; i < LPC_ORDER - 1; i++) {
1428 min = FFMIN(lsp[i] - lsp[i - 1], lsp[i + 1] - lsp[i]);
1430 weight[i] = (1 << 20) / min;
1432 weight[i] = INT16_MAX;
1437 for (i = 0; i < LPC_ORDER; i++)
1438 max = FFMAX(weight[i], max);
1440 shift = normalize_bits_int16(max);
1441 for (i = 0; i < LPC_ORDER; i++) {
1442 weight[i] <<= shift;
1445 /* Compute the VQ target vector */
1446 for (i = 0; i < LPC_ORDER; i++) {
1447 lsp[i] -= dc_lsp[i] +
1448 (((prev_lsp[i] - dc_lsp[i]) * 12288 + (1 << 14)) >> 15);
1457 * Apply the formant perceptual weighting filter.
1459 * @param flt_coef filter coefficients
1460 * @param unq_lpc unquantized lpc vector
1462 static void perceptual_filter(G723_1_Context *p, int16_t *flt_coef,
1463 int16_t *unq_lpc, int16_t *buf)
1465 int16_t vector[FRAME_LEN + LPC_ORDER];
1468 memcpy(buf, p->iir_mem, sizeof(int16_t) * LPC_ORDER);
1469 memcpy(vector, p->fir_mem, sizeof(int16_t) * LPC_ORDER);
1470 memcpy(vector + LPC_ORDER, buf + LPC_ORDER, sizeof(int16_t) * FRAME_LEN);
1472 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
1473 for (k = 0; k < LPC_ORDER; k++) {
1474 flt_coef[k + 2 * l] = (unq_lpc[k + l] * percept_flt_tbl[0][k] +
1476 flt_coef[k + 2 * l + LPC_ORDER] = (unq_lpc[k + l] *
1477 percept_flt_tbl[1][k] +
1480 iir_filter(flt_coef + 2 * l, flt_coef + 2 * l + LPC_ORDER, vector + i,
1484 memcpy(p->iir_mem, buf + FRAME_LEN, sizeof(int16_t) * LPC_ORDER);
1485 memcpy(p->fir_mem, vector + FRAME_LEN, sizeof(int16_t) * LPC_ORDER);
1489 * Estimate the open loop pitch period.
1491 * @param buf perceptually weighted speech
1492 * @param start estimation is carried out from this position
1494 static int estimate_pitch(int16_t *buf, int start)
1497 int max_ccr = 0x4000;
1498 int max_eng = 0x7fff;
1499 int index = PITCH_MIN;
1500 int offset = start - PITCH_MIN + 1;
1502 int ccr, eng, orig_eng, ccr_eng, exp;
1507 orig_eng = dot_product(buf + offset, buf + offset, HALF_FRAME_LEN, 0);
1509 for (i = PITCH_MIN; i <= PITCH_MAX - 3; i++) {
1512 /* Update energy and compute correlation */
1513 orig_eng += buf[offset] * buf[offset] -
1514 buf[offset + HALF_FRAME_LEN] * buf[offset + HALF_FRAME_LEN];
1515 ccr = dot_product(buf + start, buf + offset, HALF_FRAME_LEN, 0);
1519 /* Split into mantissa and exponent to maintain precision */
1520 exp = normalize_bits_int32(ccr);
1521 ccr = av_clipl_int32((int64_t)(ccr << exp) + (1 << 15)) >> 16;
1524 temp = normalize_bits_int32(ccr);
1525 ccr = ccr << temp >> 16;
1528 temp = normalize_bits_int32(orig_eng);
1529 eng = av_clipl_int32((int64_t)(orig_eng << temp) + (1 << 15)) >> 16;
1539 if (exp + 1 < max_exp)
1542 /* Equalize exponents before comparison */
1543 if (exp + 1 == max_exp)
1544 temp = max_ccr >> 1;
1547 ccr_eng = ccr * max_eng;
1548 diff = ccr_eng - eng * temp;
1549 if (diff > 0 && (i - index < PITCH_MIN || diff > ccr_eng >> 2)) {
1561 * Compute harmonic noise filter parameters.
1563 * @param buf perceptually weighted speech
1564 * @param pitch_lag open loop pitch period
1565 * @param hf harmonic filter parameters
1567 static void comp_harmonic_coeff(int16_t *buf, int16_t pitch_lag, HFParam *hf)
1569 int ccr, eng, max_ccr, max_eng;
1574 for (i = 0, j = pitch_lag - 3; j <= pitch_lag + 3; i++, j++) {
1575 /* Compute residual energy */
1576 energy[i << 1] = dot_product(buf - j, buf - j, SUBFRAME_LEN, 0);
1577 /* Compute correlation */
1578 energy[(i << 1) + 1] = dot_product(buf, buf - j, SUBFRAME_LEN, 0);
1581 /* Compute target energy */
1582 energy[14] = dot_product(buf, buf, SUBFRAME_LEN, 0);
1586 for (i = 0; i < 15; i++)
1587 max = FFMAX(max, FFABS(energy[i]));
1589 exp = normalize_bits_int32(max);
1590 for (i = 0; i < 15; i++) {
1591 energy[i] = av_clipl_int32((int64_t)(energy[i] << exp) +
1600 for (i = 0; i <= 6; i++) {
1601 eng = energy[i << 1];
1602 ccr = energy[(i << 1) + 1];
1607 ccr = (ccr * ccr + (1 << 14)) >> 15;
1608 diff = ccr * max_eng - eng * max_ccr;
1616 if (hf->index == -1) {
1617 hf->index = pitch_lag;
1621 eng = energy[14] * max_eng;
1622 eng = (eng >> 2) + (eng >> 3);
1623 ccr = energy[(hf->index << 1) + 1] * energy[(hf->index << 1) + 1];
1625 eng = energy[(hf->index << 1) + 1];
1630 hf->gain = ((eng << 15) / max_eng * 0x2800 + (1 << 14)) >> 15;
1632 hf->index += pitch_lag - 3;
1636 * Apply the harmonic noise shaping filter.
1638 * @param hf filter parameters
1640 static void harmonic_filter(HFParam *hf, int16_t *src, int16_t *dest)
1644 for (i = 0; i < SUBFRAME_LEN; i++) {
1645 int64_t temp = hf->gain * src[i - hf->index] << 1;
1646 dest[i] = av_clipl_int32((src[i] << 16) - temp + (1 << 15)) >> 16;
1650 static void harmonic_noise_sub(HFParam *hf, int16_t *src, int16_t *dest)
1653 for (i = 0; i < SUBFRAME_LEN; i++) {
1654 int64_t temp = hf->gain * src[i - hf->index] << 1;
1655 dest[i] = av_clipl_int32(((dest[i] - src[i]) << 16) + temp +
1662 * Combined synthesis and formant perceptual weighting filer.
1664 * @param qnt_lpc quantized lpc coefficients
1665 * @param perf_lpc perceptual filter coefficients
1666 * @param perf_fir perceptual filter fir memory
1667 * @param perf_iir perceptual filter iir memory
1668 * @param scale the filter output will be scaled by 2^scale
1670 static void synth_percept_filter(int16_t *qnt_lpc, int16_t *perf_lpc,
1671 int16_t *perf_fir, int16_t *perf_iir,
1672 int16_t *src, int16_t *dest, int scale)
1675 int16_t buf_16[SUBFRAME_LEN + LPC_ORDER];
1676 int64_t buf[SUBFRAME_LEN];
1678 int16_t *bptr_16 = buf_16 + LPC_ORDER;
1680 memcpy(buf_16, perf_fir, sizeof(int16_t) * LPC_ORDER);
1681 memcpy(dest - LPC_ORDER, perf_iir, sizeof(int16_t) * LPC_ORDER);
1683 for (i = 0; i < SUBFRAME_LEN; i++) {
1685 for (j = 1; j <= LPC_ORDER; j++)
1686 temp -= qnt_lpc[j - 1] * bptr_16[i - j];
1688 buf[i] = (src[i] << 15) + (temp << 3);
1689 bptr_16[i] = av_clipl_int32(buf[i] + (1 << 15)) >> 16;
1692 for (i = 0; i < SUBFRAME_LEN; i++) {
1693 int64_t fir = 0, iir = 0;
1694 for (j = 1; j <= LPC_ORDER; j++) {
1695 fir -= perf_lpc[j - 1] * bptr_16[i - j];
1696 iir += perf_lpc[j + LPC_ORDER - 1] * dest[i - j];
1698 dest[i] = av_clipl_int32(((buf[i] + (fir << 3)) << scale) + (iir << 3) +
1701 memcpy(perf_fir, buf_16 + SUBFRAME_LEN, sizeof(int16_t) * LPC_ORDER);
1702 memcpy(perf_iir, dest + SUBFRAME_LEN - LPC_ORDER,
1703 sizeof(int16_t) * LPC_ORDER);
1707 * Compute the adaptive codebook contribution.
1709 * @param buf input signal
1710 * @param index the current subframe index
1712 static void acb_search(G723_1_Context *p, int16_t *residual,
1713 int16_t *impulse_resp, int16_t *buf,
1717 int16_t flt_buf[PITCH_ORDER][SUBFRAME_LEN];
1719 const int16_t *cb_tbl = adaptive_cb_gain85;
1721 int ccr_buf[PITCH_ORDER * SUBFRAMES << 2];
1723 int pitch_lag = p->pitch_lag[index >> 1];
1726 int odd_frame = index & 1;
1727 int iter = 3 + odd_frame;
1731 int i, j, k, l, max;
1735 if (pitch_lag == PITCH_MIN)
1738 pitch_lag = FFMIN(pitch_lag, PITCH_MAX - 5);
1741 for (i = 0; i < iter; i++) {
1742 get_residual(residual, p->prev_excitation, pitch_lag + i - 1);
1744 for (j = 0; j < SUBFRAME_LEN; j++) {
1746 for (k = 0; k <= j; k++)
1747 temp += residual[PITCH_ORDER - 1 + k] * impulse_resp[j - k];
1748 flt_buf[PITCH_ORDER - 1][j] = av_clipl_int32((temp << 1) +
1752 for (j = PITCH_ORDER - 2; j >= 0; j--) {
1753 flt_buf[j][0] = ((residual[j] << 13) + (1 << 14)) >> 15;
1754 for (k = 1; k < SUBFRAME_LEN; k++) {
1755 temp = (flt_buf[j + 1][k - 1] << 15) +
1756 residual[j] * impulse_resp[k];
1757 flt_buf[j][k] = av_clipl_int32((temp << 1) + (1 << 15)) >> 16;
1761 /* Compute crosscorrelation with the signal */
1762 for (j = 0; j < PITCH_ORDER; j++) {
1763 temp = dot_product(buf, flt_buf[j], SUBFRAME_LEN, 0);
1764 ccr_buf[count++] = av_clipl_int32(temp << 1);
1767 /* Compute energies */
1768 for (j = 0; j < PITCH_ORDER; j++) {
1769 ccr_buf[count++] = dot_product(flt_buf[j], flt_buf[j],
1773 for (j = 1; j < PITCH_ORDER; j++) {
1774 for (k = 0; k < j; k++) {
1775 temp = dot_product(flt_buf[j], flt_buf[k], SUBFRAME_LEN, 0);
1776 ccr_buf[count++] = av_clipl_int32(temp<<2);
1781 /* Normalize and shorten */
1783 for (i = 0; i < 20 * iter; i++)
1784 max = FFMAX(max, FFABS(ccr_buf[i]));
1786 temp = normalize_bits_int32(max);
1788 for (i = 0; i < 20 * iter; i++){
1789 ccr_buf[i] = av_clipl_int32((int64_t)(ccr_buf[i] << temp) +
1794 for (i = 0; i < iter; i++) {
1795 /* Select quantization table */
1796 if (!odd_frame && pitch_lag + i - 1 >= SUBFRAME_LEN - 2 ||
1797 odd_frame && pitch_lag >= SUBFRAME_LEN - 2) {
1798 cb_tbl = adaptive_cb_gain170;
1802 for (j = 0, k = 0; j < tbl_size; j++, k += 20) {
1804 for (l = 0; l < 20; l++)
1805 temp += ccr_buf[20 * i + l] * cb_tbl[k + l];
1806 temp = av_clipl_int32(temp);
1817 pitch_lag += acb_lag - 1;
1821 p->pitch_lag[index >> 1] = pitch_lag;
1822 p->subframe[index].ad_cb_lag = acb_lag;
1823 p->subframe[index].ad_cb_gain = acb_gain;
1827 * Subtract the adaptive codebook contribution from the input
1828 * to obtain the residual.
1830 * @param buf target vector
1832 static void sub_acb_contrib(int16_t *residual, int16_t *impulse_resp,
1836 /* Subtract adaptive CB contribution to obtain the residual */
1837 for (i = 0; i < SUBFRAME_LEN; i++) {
1838 int64_t temp = buf[i] << 14;
1839 for (j = 0; j <= i; j++)
1840 temp -= residual[j] * impulse_resp[i - j];
1842 buf[i] = av_clipl_int32((temp << 2) + (1 << 15)) >> 16;
1847 * Quantize the residual signal using the fixed codebook (MP-MLQ).
1849 * @param optim optimized fixed codebook parameters
1850 * @param buf excitation vector
1852 static void get_fcb_param(FCBParam *optim, int16_t *impulse_resp,
1853 int16_t *buf, int pulse_cnt, int pitch_lag)
1856 int16_t impulse_r[SUBFRAME_LEN];
1857 int16_t temp_corr[SUBFRAME_LEN];
1858 int16_t impulse_corr[SUBFRAME_LEN];
1860 int ccr1[SUBFRAME_LEN];
1861 int ccr2[SUBFRAME_LEN];
1862 int amp, err, max, max_amp_index, min, scale, i, j, k, l;
1866 /* Update impulse response */
1867 memcpy(impulse_r, impulse_resp, sizeof(int16_t) * SUBFRAME_LEN);
1868 param.dirac_train = 0;
1869 if (pitch_lag < SUBFRAME_LEN - 2) {
1870 param.dirac_train = 1;
1871 gen_dirac_train(impulse_r, pitch_lag);
1874 for (i = 0; i < SUBFRAME_LEN; i++)
1875 temp_corr[i] = impulse_r[i] >> 1;
1877 /* Compute impulse response autocorrelation */
1878 temp = dot_product(temp_corr, temp_corr, SUBFRAME_LEN, 1);
1880 scale = normalize_bits_int32(temp);
1881 impulse_corr[0] = av_clipl_int32((temp << scale) + (1 << 15)) >> 16;
1883 for (i = 1; i < SUBFRAME_LEN; i++) {
1884 temp = dot_product(temp_corr + i, temp_corr, SUBFRAME_LEN - i, 1);
1885 impulse_corr[i] = av_clipl_int32((temp << scale) + (1 << 15)) >> 16;
1888 /* Compute crosscorrelation of impulse response with residual signal */
1890 for (i = 0; i < SUBFRAME_LEN; i++){
1891 temp = dot_product(buf + i, impulse_r, SUBFRAME_LEN - i, 1);
1893 ccr1[i] = temp >> -scale;
1895 ccr1[i] = av_clipl_int32(temp << scale);
1899 for (i = 0; i < GRID_SIZE; i++) {
1900 /* Maximize the crosscorrelation */
1902 for (j = i; j < SUBFRAME_LEN; j += GRID_SIZE) {
1903 temp = FFABS(ccr1[j]);
1906 param.pulse_pos[0] = j;
1910 /* Quantize the gain (max crosscorrelation/impulse_corr[0]) */
1913 max_amp_index = GAIN_LEVELS - 2;
1914 for (j = max_amp_index; j >= 2; j--) {
1915 temp = av_clipl_int32((int64_t)fixed_cb_gain[j] *
1916 impulse_corr[0] << 1);
1917 temp = FFABS(temp - amp);
1925 /* Select additional gain values */
1926 for (j = 1; j < 5; j++) {
1927 for (k = i; k < SUBFRAME_LEN; k += GRID_SIZE) {
1931 param.amp_index = max_amp_index + j - 2;
1932 amp = fixed_cb_gain[param.amp_index];
1934 param.pulse_sign[0] = (ccr2[param.pulse_pos[0]] < 0) ? -amp : amp;
1935 temp_corr[param.pulse_pos[0]] = 1;
1937 for (k = 1; k < pulse_cnt; k++) {
1939 for (l = i; l < SUBFRAME_LEN; l += GRID_SIZE) {
1942 temp = impulse_corr[FFABS(l - param.pulse_pos[k - 1])];
1943 temp = av_clipl_int32((int64_t)temp *
1944 param.pulse_sign[k - 1] << 1);
1946 temp = FFABS(ccr2[l]);
1949 param.pulse_pos[k] = l;
1953 param.pulse_sign[k] = (ccr2[param.pulse_pos[k]] < 0) ?
1955 temp_corr[param.pulse_pos[k]] = 1;
1958 /* Create the error vector */
1959 memset(temp_corr, 0, sizeof(int16_t) * SUBFRAME_LEN);
1961 for (k = 0; k < pulse_cnt; k++)
1962 temp_corr[param.pulse_pos[k]] = param.pulse_sign[k];
1964 for (k = SUBFRAME_LEN - 1; k >= 0; k--) {
1966 for (l = 0; l <= k; l++) {
1967 int prod = av_clipl_int32((int64_t)temp_corr[l] *
1968 impulse_r[k - l] << 1);
1969 temp = av_clipl_int32(temp + prod);
1971 temp_corr[k] = temp << 2 >> 16;
1974 /* Compute square of error */
1976 for (k = 0; k < SUBFRAME_LEN; k++) {
1978 prod = av_clipl_int32((int64_t)buf[k] * temp_corr[k] << 1);
1979 err = av_clipl_int32(err - prod);
1980 prod = av_clipl_int32((int64_t)temp_corr[k] * temp_corr[k]);
1981 err = av_clipl_int32(err + prod);
1985 if (err < optim->min_err) {
1986 optim->min_err = err;
1987 optim->grid_index = i;
1988 optim->amp_index = param.amp_index;
1989 optim->dirac_train = param.dirac_train;
1991 for (k = 0; k < pulse_cnt; k++) {
1992 optim->pulse_sign[k] = param.pulse_sign[k];
1993 optim->pulse_pos[k] = param.pulse_pos[k];
2001 * Encode the pulse position and gain of the current subframe.
2003 * @param optim optimized fixed CB parameters
2004 * @param buf excitation vector
2006 static void pack_fcb_param(G723_1_Subframe *subfrm, FCBParam *optim,
2007 int16_t *buf, int pulse_cnt)
2011 j = PULSE_MAX - pulse_cnt;
2013 subfrm->pulse_sign = 0;
2014 subfrm->pulse_pos = 0;
2016 for (i = 0; i < SUBFRAME_LEN >> 1; i++) {
2017 int val = buf[optim->grid_index + (i << 1)];
2019 subfrm->pulse_pos += combinatorial_table[j][i];
2021 subfrm->pulse_sign <<= 1;
2022 if (val < 0) subfrm->pulse_sign++;
2025 if (j == PULSE_MAX) break;
2028 subfrm->amp_index = optim->amp_index;
2029 subfrm->grid_index = optim->grid_index;
2030 subfrm->dirac_train = optim->dirac_train;
2034 * Compute the fixed codebook excitation.
2036 * @param buf target vector
2037 * @param impulse_resp impulse response of the combined filter
2039 static void fcb_search(G723_1_Context *p, int16_t *impulse_resp,
2040 int16_t *buf, int index)
2043 int pulse_cnt = pulses[index];
2046 optim.min_err = 1 << 30;
2047 get_fcb_param(&optim, impulse_resp, buf, pulse_cnt, SUBFRAME_LEN);
2049 if (p->pitch_lag[index >> 1] < SUBFRAME_LEN - 2) {
2050 get_fcb_param(&optim, impulse_resp, buf, pulse_cnt,
2051 p->pitch_lag[index >> 1]);
2054 /* Reconstruct the excitation */
2055 memset(buf, 0, sizeof(int16_t) * SUBFRAME_LEN);
2056 for (i = 0; i < pulse_cnt; i++)
2057 buf[optim.pulse_pos[i]] = optim.pulse_sign[i];
2059 pack_fcb_param(&p->subframe[index], &optim, buf, pulse_cnt);
2061 if (optim.dirac_train)
2062 gen_dirac_train(buf, p->pitch_lag[index >> 1]);
2066 * Pack the frame parameters into output bitstream.
2068 * @param frame output buffer
2069 * @param size size of the buffer
2071 static int pack_bitstream(G723_1_Context *p, unsigned char *frame, int size)
2074 int info_bits, i, temp;
2076 init_put_bits(&pb, frame, size);
2078 if (p->cur_rate == RATE_6300) {
2080 put_bits(&pb, 2, info_bits);
2083 put_bits(&pb, 8, p->lsp_index[2]);
2084 put_bits(&pb, 8, p->lsp_index[1]);
2085 put_bits(&pb, 8, p->lsp_index[0]);
2087 put_bits(&pb, 7, p->pitch_lag[0] - PITCH_MIN);
2088 put_bits(&pb, 2, p->subframe[1].ad_cb_lag);
2089 put_bits(&pb, 7, p->pitch_lag[1] - PITCH_MIN);
2090 put_bits(&pb, 2, p->subframe[3].ad_cb_lag);
2092 /* Write 12 bit combined gain */
2093 for (i = 0; i < SUBFRAMES; i++) {
2094 temp = p->subframe[i].ad_cb_gain * GAIN_LEVELS +
2095 p->subframe[i].amp_index;
2096 if (p->cur_rate == RATE_6300)
2097 temp += p->subframe[i].dirac_train << 11;
2098 put_bits(&pb, 12, temp);
2101 put_bits(&pb, 1, p->subframe[0].grid_index);
2102 put_bits(&pb, 1, p->subframe[1].grid_index);
2103 put_bits(&pb, 1, p->subframe[2].grid_index);
2104 put_bits(&pb, 1, p->subframe[3].grid_index);
2106 if (p->cur_rate == RATE_6300) {
2107 skip_put_bits(&pb, 1); /* reserved bit */
2109 /* Write 13 bit combined position index */
2110 temp = (p->subframe[0].pulse_pos >> 16) * 810 +
2111 (p->subframe[1].pulse_pos >> 14) * 90 +
2112 (p->subframe[2].pulse_pos >> 16) * 9 +
2113 (p->subframe[3].pulse_pos >> 14);
2114 put_bits(&pb, 13, temp);
2116 put_bits(&pb, 16, p->subframe[0].pulse_pos & 0xffff);
2117 put_bits(&pb, 14, p->subframe[1].pulse_pos & 0x3fff);
2118 put_bits(&pb, 16, p->subframe[2].pulse_pos & 0xffff);
2119 put_bits(&pb, 14, p->subframe[3].pulse_pos & 0x3fff);
2121 put_bits(&pb, 6, p->subframe[0].pulse_sign);
2122 put_bits(&pb, 5, p->subframe[1].pulse_sign);
2123 put_bits(&pb, 6, p->subframe[2].pulse_sign);
2124 put_bits(&pb, 5, p->subframe[3].pulse_sign);
2127 flush_put_bits(&pb);
2128 return frame_size[info_bits];
2131 static int g723_1_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
2132 const AVFrame *frame, int *got_packet_ptr)
2134 G723_1_Context *p = avctx->priv_data;
2135 int16_t unq_lpc[LPC_ORDER * SUBFRAMES];
2136 int16_t qnt_lpc[LPC_ORDER * SUBFRAMES];
2137 int16_t cur_lsp[LPC_ORDER];
2138 int16_t weighted_lpc[LPC_ORDER * SUBFRAMES << 1];
2139 int16_t vector[FRAME_LEN + PITCH_MAX];
2141 int16_t *in = (const int16_t *)frame->data[0];
2146 highpass_filter(in, &p->hpf_fir_mem, &p->hpf_iir_mem);
2148 memcpy(vector, p->prev_data, HALF_FRAME_LEN * sizeof(int16_t));
2149 memcpy(vector + HALF_FRAME_LEN, in, FRAME_LEN * sizeof(int16_t));
2151 comp_lpc_coeff(vector, unq_lpc);
2152 lpc2lsp(&unq_lpc[LPC_ORDER * 3], p->prev_lsp, cur_lsp);
2153 lsp_quantize(p->lsp_index, cur_lsp, p->prev_lsp);
2156 memcpy(vector + LPC_ORDER, p->prev_data + SUBFRAME_LEN,
2157 sizeof(int16_t) * SUBFRAME_LEN);
2158 memcpy(vector + LPC_ORDER + SUBFRAME_LEN, in,
2159 sizeof(int16_t) * (HALF_FRAME_LEN + SUBFRAME_LEN));
2160 memcpy(p->prev_data, in + HALF_FRAME_LEN,
2161 sizeof(int16_t) * HALF_FRAME_LEN);
2162 memcpy(in, vector + LPC_ORDER, sizeof(int16_t) * FRAME_LEN);
2164 perceptual_filter(p, weighted_lpc, unq_lpc, vector);
2166 memcpy(in, vector + LPC_ORDER, sizeof(int16_t) * FRAME_LEN);
2167 memcpy(vector, p->prev_weight_sig, sizeof(int16_t) * PITCH_MAX);
2168 memcpy(vector + PITCH_MAX, in, sizeof(int16_t) * FRAME_LEN);
2170 scale_vector(vector, FRAME_LEN + PITCH_MAX);
2172 p->pitch_lag[0] = estimate_pitch(vector, PITCH_MAX);
2173 p->pitch_lag[1] = estimate_pitch(vector, PITCH_MAX + HALF_FRAME_LEN);
2175 for (i = PITCH_MAX, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
2176 comp_harmonic_coeff(vector + i, p->pitch_lag[j >> 1], hf + j);
2178 memcpy(vector, p->prev_weight_sig, sizeof(int16_t) * PITCH_MAX);
2179 memcpy(vector + PITCH_MAX, in, sizeof(int16_t) * FRAME_LEN);
2180 memcpy(p->prev_weight_sig, vector + FRAME_LEN, sizeof(int16_t) * PITCH_MAX);
2182 for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
2183 harmonic_filter(hf + j, vector + PITCH_MAX + i, in + i);
2185 inverse_quant(cur_lsp, p->prev_lsp, p->lsp_index, 0);
2186 lsp_interpolate(qnt_lpc, cur_lsp, p->prev_lsp);
2188 memcpy(p->prev_lsp, cur_lsp, sizeof(int16_t) * LPC_ORDER);
2191 for (i = 0; i < SUBFRAMES; i++) {
2192 int16_t impulse_resp[SUBFRAME_LEN];
2193 int16_t residual[SUBFRAME_LEN + PITCH_ORDER - 1];
2194 int16_t flt_in[SUBFRAME_LEN];
2195 int16_t zero[LPC_ORDER], fir[LPC_ORDER], iir[LPC_ORDER];
2198 * Compute the combined impulse response of the synthesis filter,
2199 * formant perceptual weighting filter and harmonic noise shaping filter
2201 memset(zero, 0, sizeof(int16_t) * LPC_ORDER);
2202 memset(vector, 0, sizeof(int16_t) * PITCH_MAX);
2203 memset(flt_in, 0, sizeof(int16_t) * SUBFRAME_LEN);
2205 flt_in[0] = 1 << 13; /* Unit impulse */
2206 synth_percept_filter(qnt_lpc + offset, weighted_lpc + (offset << 1),
2207 zero, zero, flt_in, vector + PITCH_MAX, 1);
2208 harmonic_filter(hf + i, vector + PITCH_MAX, impulse_resp);
2210 /* Compute the combined zero input response */
2212 memcpy(fir, p->perf_fir_mem, sizeof(int16_t) * LPC_ORDER);
2213 memcpy(iir, p->perf_iir_mem, sizeof(int16_t) * LPC_ORDER);
2215 synth_percept_filter(qnt_lpc + offset, weighted_lpc + (offset << 1),
2216 fir, iir, flt_in, vector + PITCH_MAX, 0);
2217 memcpy(vector, p->harmonic_mem, sizeof(int16_t) * PITCH_MAX);
2218 harmonic_noise_sub(hf + i, vector + PITCH_MAX, in);
2220 acb_search(p, residual, impulse_resp, in, i);
2221 gen_acb_excitation(residual, p->prev_excitation,p->pitch_lag[i >> 1],
2222 p->subframe[i], p->cur_rate);
2223 sub_acb_contrib(residual, impulse_resp, in);
2225 fcb_search(p, impulse_resp, in, i);
2227 /* Reconstruct the excitation */
2228 gen_acb_excitation(impulse_resp, p->prev_excitation, p->pitch_lag[i >> 1],
2229 p->subframe[i], RATE_6300);
2231 memmove(p->prev_excitation, p->prev_excitation + SUBFRAME_LEN,
2232 sizeof(int16_t) * (PITCH_MAX - SUBFRAME_LEN));
2233 for (j = 0; j < SUBFRAME_LEN; j++)
2234 in[j] = av_clip_int16((in[j] << 1) + impulse_resp[j]);
2235 memcpy(p->prev_excitation + PITCH_MAX - SUBFRAME_LEN, in,
2236 sizeof(int16_t) * SUBFRAME_LEN);
2238 /* Update filter memories */
2239 synth_percept_filter(qnt_lpc + offset, weighted_lpc + (offset << 1),
2240 p->perf_fir_mem, p->perf_iir_mem,
2241 in, vector + PITCH_MAX, 0);
2242 memmove(p->harmonic_mem, p->harmonic_mem + SUBFRAME_LEN,
2243 sizeof(int16_t) * (PITCH_MAX - SUBFRAME_LEN));
2244 memcpy(p->harmonic_mem + PITCH_MAX - SUBFRAME_LEN, vector + PITCH_MAX,
2245 sizeof(int16_t) * SUBFRAME_LEN);
2248 offset += LPC_ORDER;
2251 if ((ret = ff_alloc_packet2(avctx, avpkt, 24)))
2254 *got_packet_ptr = 1;
2255 avpkt->size = pack_bitstream(p, avpkt->data, avpkt->size);
2259 AVCodec ff_g723_1_encoder = {
2261 .type = AVMEDIA_TYPE_AUDIO,
2262 .id = CODEC_ID_G723_1,
2263 .priv_data_size = sizeof(G723_1_Context),
2264 .init = g723_1_encode_init,
2265 .encode2 = g723_1_encode_frame,
2266 .long_name = NULL_IF_CONFIG_SMALL("G.723.1"),
2267 .sample_fmts = (const enum AVSampleFormat[]){AV_SAMPLE_FMT_S16,
2268 AV_SAMPLE_FMT_NONE},