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
29 #define BITSTREAM_READER_LE
32 #include "acelp_vectors.h"
33 #include "celp_filters.h"
34 #include "celp_math.h"
36 #include "libavutil/lzo.h"
37 #include "g723_1_data.h"
39 typedef struct g723_1_context {
41 G723_1_Subframe subframe[4];
42 FrameType cur_frame_type;
43 FrameType past_frame_type;
45 uint8_t lsp_index[LSP_BANDS];
49 int16_t prev_lsp[LPC_ORDER];
50 int16_t prev_excitation[PITCH_MAX];
51 int16_t excitation[PITCH_MAX + FRAME_LEN];
52 int16_t synth_mem[LPC_ORDER];
53 int16_t fir_mem[LPC_ORDER];
54 int iir_mem[LPC_ORDER];
62 int pf_gain; ///< formant postfilter
63 ///< gain scaling unit memory
65 int16_t prev_data[HALF_FRAME_LEN];
66 int16_t prev_weight_sig[PITCH_MAX];
69 int16_t hpf_fir_mem; ///< highpass filter fir
70 int hpf_iir_mem; ///< and iir memories
71 int16_t perf_fir_mem[LPC_ORDER]; ///< perceptual filter fir
72 int16_t perf_iir_mem[LPC_ORDER]; ///< and iir memories
74 int16_t harmonic_mem[PITCH_MAX];
77 static av_cold int g723_1_decode_init(AVCodecContext *avctx)
79 G723_1_Context *p = avctx->priv_data;
81 avctx->sample_fmt = AV_SAMPLE_FMT_S16;
83 memcpy(p->prev_lsp, dc_lsp, LPC_ORDER * sizeof(int16_t));
85 avcodec_get_frame_defaults(&p->frame);
86 avctx->coded_frame = &p->frame;
92 * Unpack the frame into parameters.
94 * @param p the context
95 * @param buf pointer to the input buffer
96 * @param buf_size size of the input buffer
98 static int unpack_bitstream(G723_1_Context *p, const uint8_t *buf,
103 int temp, info_bits, i;
105 init_get_bits(&gb, buf, buf_size * 8);
107 /* Extract frame type and rate info */
108 info_bits = get_bits(&gb, 2);
110 if (info_bits == 3) {
111 p->cur_frame_type = UntransmittedFrame;
115 /* Extract 24 bit lsp indices, 8 bit for each band */
116 p->lsp_index[2] = get_bits(&gb, 8);
117 p->lsp_index[1] = get_bits(&gb, 8);
118 p->lsp_index[0] = get_bits(&gb, 8);
120 if (info_bits == 2) {
121 p->cur_frame_type = SIDFrame;
122 p->subframe[0].amp_index = get_bits(&gb, 6);
126 /* Extract the info common to both rates */
127 p->cur_rate = info_bits ? Rate5k3 : Rate6k3;
128 p->cur_frame_type = ActiveFrame;
130 p->pitch_lag[0] = get_bits(&gb, 7);
131 if (p->pitch_lag[0] > 123) /* test if forbidden code */
133 p->pitch_lag[0] += PITCH_MIN;
134 p->subframe[1].ad_cb_lag = get_bits(&gb, 2);
136 p->pitch_lag[1] = get_bits(&gb, 7);
137 if (p->pitch_lag[1] > 123)
139 p->pitch_lag[1] += PITCH_MIN;
140 p->subframe[3].ad_cb_lag = get_bits(&gb, 2);
141 p->subframe[0].ad_cb_lag = 1;
142 p->subframe[2].ad_cb_lag = 1;
144 for (i = 0; i < SUBFRAMES; i++) {
145 /* Extract combined gain */
146 temp = get_bits(&gb, 12);
148 p->subframe[i].dirac_train = 0;
149 if (p->cur_rate == Rate6k3 && p->pitch_lag[i >> 1] < SUBFRAME_LEN - 2) {
150 p->subframe[i].dirac_train = temp >> 11;
154 p->subframe[i].ad_cb_gain = FASTDIV(temp, GAIN_LEVELS);
155 if (p->subframe[i].ad_cb_gain < ad_cb_len) {
156 p->subframe[i].amp_index = temp - p->subframe[i].ad_cb_gain *
163 p->subframe[0].grid_index = get_bits1(&gb);
164 p->subframe[1].grid_index = get_bits1(&gb);
165 p->subframe[2].grid_index = get_bits1(&gb);
166 p->subframe[3].grid_index = get_bits1(&gb);
168 if (p->cur_rate == Rate6k3) {
169 skip_bits1(&gb); /* skip reserved bit */
171 /* Compute pulse_pos index using the 13-bit combined position index */
172 temp = get_bits(&gb, 13);
173 p->subframe[0].pulse_pos = temp / 810;
175 temp -= p->subframe[0].pulse_pos * 810;
176 p->subframe[1].pulse_pos = FASTDIV(temp, 90);
178 temp -= p->subframe[1].pulse_pos * 90;
179 p->subframe[2].pulse_pos = FASTDIV(temp, 9);
180 p->subframe[3].pulse_pos = temp - p->subframe[2].pulse_pos * 9;
182 p->subframe[0].pulse_pos = (p->subframe[0].pulse_pos << 16) +
184 p->subframe[1].pulse_pos = (p->subframe[1].pulse_pos << 14) +
186 p->subframe[2].pulse_pos = (p->subframe[2].pulse_pos << 16) +
188 p->subframe[3].pulse_pos = (p->subframe[3].pulse_pos << 14) +
191 p->subframe[0].pulse_sign = get_bits(&gb, 6);
192 p->subframe[1].pulse_sign = get_bits(&gb, 5);
193 p->subframe[2].pulse_sign = get_bits(&gb, 6);
194 p->subframe[3].pulse_sign = get_bits(&gb, 5);
195 } else { /* Rate5k3 */
196 p->subframe[0].pulse_pos = get_bits(&gb, 12);
197 p->subframe[1].pulse_pos = get_bits(&gb, 12);
198 p->subframe[2].pulse_pos = get_bits(&gb, 12);
199 p->subframe[3].pulse_pos = get_bits(&gb, 12);
201 p->subframe[0].pulse_sign = get_bits(&gb, 4);
202 p->subframe[1].pulse_sign = get_bits(&gb, 4);
203 p->subframe[2].pulse_sign = get_bits(&gb, 4);
204 p->subframe[3].pulse_sign = get_bits(&gb, 4);
211 * Bitexact implementation of sqrt(val/2).
213 static int16_t square_root(int val)
215 return (ff_sqrt(val << 1) >> 1) & (~1);
219 * Calculate the number of left-shifts required for normalizing the input.
221 * @param num input number
222 * @param width width of the input, 16 bits(0) / 32 bits(1)
224 static int normalize_bits(int num, int width)
227 int bits = (width) ? 31 : 15;
234 i= bits - av_log2(num) - 1;
240 #define normalize_bits_int16(num) normalize_bits(num, 0)
241 #define normalize_bits_int32(num) normalize_bits(num, 1)
242 #define dot_product(a,b,c,d) (ff_dot_product(a,b,c)<<(d))
245 * Scale vector contents based on the largest of their absolutes.
247 static int scale_vector(int16_t *vector, int length)
249 int bits, scale, max = 0;
252 const int16_t shift_table[16] = {
253 0x0001, 0x0002, 0x0004, 0x0008, 0x0010, 0x0020, 0x0040, 0x0080,
254 0x0100, 0x0200, 0x0400, 0x0800, 0x1000, 0x2000, 0x4000, 0x7fff
257 for (i = 0; i < length; i++)
258 max = FFMAX(max, FFABS(vector[i]));
260 bits = normalize_bits(max, 0);
261 scale = shift_table[bits];
263 for (i = 0; i < length; i++)
264 vector[i] = (vector[i] * scale) >> 3;
270 * Perform inverse quantization of LSP frequencies.
272 * @param cur_lsp the current LSP vector
273 * @param prev_lsp the previous LSP vector
274 * @param lsp_index VQ indices
275 * @param bad_frame bad frame flag
277 static void inverse_quant(int16_t *cur_lsp, int16_t *prev_lsp,
278 uint8_t *lsp_index, int bad_frame)
281 int i, j, temp, stable;
283 /* Check for frame erasure */
290 lsp_index[0] = lsp_index[1] = lsp_index[2] = 0;
293 /* Get the VQ table entry corresponding to the transmitted index */
294 cur_lsp[0] = lsp_band0[lsp_index[0]][0];
295 cur_lsp[1] = lsp_band0[lsp_index[0]][1];
296 cur_lsp[2] = lsp_band0[lsp_index[0]][2];
297 cur_lsp[3] = lsp_band1[lsp_index[1]][0];
298 cur_lsp[4] = lsp_band1[lsp_index[1]][1];
299 cur_lsp[5] = lsp_band1[lsp_index[1]][2];
300 cur_lsp[6] = lsp_band2[lsp_index[2]][0];
301 cur_lsp[7] = lsp_band2[lsp_index[2]][1];
302 cur_lsp[8] = lsp_band2[lsp_index[2]][2];
303 cur_lsp[9] = lsp_band2[lsp_index[2]][3];
305 /* Add predicted vector & DC component to the previously quantized vector */
306 for (i = 0; i < LPC_ORDER; i++) {
307 temp = ((prev_lsp[i] - dc_lsp[i]) * pred + (1 << 14)) >> 15;
308 cur_lsp[i] += dc_lsp[i] + temp;
311 for (i = 0; i < LPC_ORDER; i++) {
312 cur_lsp[0] = FFMAX(cur_lsp[0], 0x180);
313 cur_lsp[LPC_ORDER - 1] = FFMIN(cur_lsp[LPC_ORDER - 1], 0x7e00);
315 /* Stability check */
316 for (j = 1; j < LPC_ORDER; j++) {
317 temp = min_dist + cur_lsp[j - 1] - cur_lsp[j];
320 cur_lsp[j - 1] -= temp;
325 for (j = 1; j < LPC_ORDER; j++) {
326 temp = cur_lsp[j - 1] + min_dist - cur_lsp[j] - 4;
336 memcpy(cur_lsp, prev_lsp, LPC_ORDER * sizeof(int16_t));
340 * Bitexact implementation of 2ab scaled by 1/2^16.
342 * @param a 32 bit multiplicand
343 * @param b 16 bit multiplier
345 #define MULL2(a, b) \
349 * Convert LSP frequencies to LPC coefficients.
351 * @param lpc buffer for LPC coefficients
353 static void lsp2lpc(int16_t *lpc)
355 int f1[LPC_ORDER / 2 + 1];
356 int f2[LPC_ORDER / 2 + 1];
359 /* Calculate negative cosine */
360 for (j = 0; j < LPC_ORDER; j++) {
361 int index = lpc[j] >> 7;
362 int offset = lpc[j] & 0x7f;
363 int64_t temp1 = cos_tab[index] << 16;
364 int temp2 = (cos_tab[index + 1] - cos_tab[index]) *
365 ((offset << 8) + 0x80) << 1;
367 lpc[j] = -(av_clipl_int32(((temp1 + temp2) << 1) + (1 << 15)) >> 16);
371 * Compute sum and difference polynomial coefficients
372 * (bitexact alternative to lsp2poly() in lsp.c)
374 /* Initialize with values in Q28 */
376 f1[1] = (lpc[0] << 14) + (lpc[2] << 14);
377 f1[2] = lpc[0] * lpc[2] + (2 << 28);
380 f2[1] = (lpc[1] << 14) + (lpc[3] << 14);
381 f2[2] = lpc[1] * lpc[3] + (2 << 28);
384 * Calculate and scale the coefficients by 1/2 in
385 * each iteration for a final scaling factor of Q25
387 for (i = 2; i < LPC_ORDER / 2; i++) {
388 f1[i + 1] = f1[i - 1] + MULL2(f1[i], lpc[2 * i]);
389 f2[i + 1] = f2[i - 1] + MULL2(f2[i], lpc[2 * i + 1]);
391 for (j = i; j >= 2; j--) {
392 f1[j] = MULL2(f1[j - 1], lpc[2 * i]) +
393 (f1[j] >> 1) + (f1[j - 2] >> 1);
394 f2[j] = MULL2(f2[j - 1], lpc[2 * i + 1]) +
395 (f2[j] >> 1) + (f2[j - 2] >> 1);
400 f1[1] = ((lpc[2 * i] << 16 >> i) + f1[1]) >> 1;
401 f2[1] = ((lpc[2 * i + 1] << 16 >> i) + f2[1]) >> 1;
404 /* Convert polynomial coefficients to LPC coefficients */
405 for (i = 0; i < LPC_ORDER / 2; i++) {
406 int64_t ff1 = f1[i + 1] + f1[i];
407 int64_t ff2 = f2[i + 1] - f2[i];
409 lpc[i] = av_clipl_int32(((ff1 + ff2) << 3) + (1 << 15)) >> 16;
410 lpc[LPC_ORDER - i - 1] = av_clipl_int32(((ff1 - ff2) << 3) +
416 * Quantize LSP frequencies by interpolation and convert them to
417 * the corresponding LPC coefficients.
419 * @param lpc buffer for LPC coefficients
420 * @param cur_lsp the current LSP vector
421 * @param prev_lsp the previous LSP vector
423 static void lsp_interpolate(int16_t *lpc, int16_t *cur_lsp, int16_t *prev_lsp)
426 int16_t *lpc_ptr = lpc;
428 /* cur_lsp * 0.25 + prev_lsp * 0.75 */
429 ff_acelp_weighted_vector_sum(lpc, cur_lsp, prev_lsp,
430 4096, 12288, 1 << 13, 14, LPC_ORDER);
431 ff_acelp_weighted_vector_sum(lpc + LPC_ORDER, cur_lsp, prev_lsp,
432 8192, 8192, 1 << 13, 14, LPC_ORDER);
433 ff_acelp_weighted_vector_sum(lpc + 2 * LPC_ORDER, cur_lsp, prev_lsp,
434 12288, 4096, 1 << 13, 14, LPC_ORDER);
435 memcpy(lpc + 3 * LPC_ORDER, cur_lsp, LPC_ORDER * sizeof(int16_t));
437 for (i = 0; i < SUBFRAMES; i++) {
439 lpc_ptr += LPC_ORDER;
444 * Generate a train of dirac functions with period as pitch lag.
446 static void gen_dirac_train(int16_t *buf, int pitch_lag)
448 int16_t vector[SUBFRAME_LEN];
451 memcpy(vector, buf, SUBFRAME_LEN * sizeof(int16_t));
452 for (i = pitch_lag; i < SUBFRAME_LEN; i += pitch_lag) {
453 for (j = 0; j < SUBFRAME_LEN - i; j++)
454 buf[i + j] += vector[j];
459 * Generate fixed codebook excitation vector.
461 * @param vector decoded excitation vector
462 * @param subfrm current subframe
463 * @param cur_rate current bitrate
464 * @param pitch_lag closed loop pitch lag
465 * @param index current subframe index
467 static void gen_fcb_excitation(int16_t *vector, G723_1_Subframe subfrm,
468 Rate cur_rate, int pitch_lag, int index)
472 memset(vector, 0, SUBFRAME_LEN * sizeof(int16_t));
474 if (cur_rate == Rate6k3) {
475 if (subfrm.pulse_pos >= max_pos[index])
478 /* Decode amplitudes and positions */
479 j = PULSE_MAX - pulses[index];
480 temp = subfrm.pulse_pos;
481 for (i = 0; i < SUBFRAME_LEN / GRID_SIZE; i++) {
482 temp -= combinatorial_table[j][i];
485 temp += combinatorial_table[j++][i];
486 if (subfrm.pulse_sign & (1 << (PULSE_MAX - j))) {
487 vector[subfrm.grid_index + GRID_SIZE * i] =
488 -fixed_cb_gain[subfrm.amp_index];
490 vector[subfrm.grid_index + GRID_SIZE * i] =
491 fixed_cb_gain[subfrm.amp_index];
496 if (subfrm.dirac_train == 1)
497 gen_dirac_train(vector, pitch_lag);
498 } else { /* Rate5k3 */
499 int cb_gain = fixed_cb_gain[subfrm.amp_index];
500 int cb_shift = subfrm.grid_index;
501 int cb_sign = subfrm.pulse_sign;
502 int cb_pos = subfrm.pulse_pos;
503 int offset, beta, lag;
505 for (i = 0; i < 8; i += 2) {
506 offset = ((cb_pos & 7) << 3) + cb_shift + i;
507 vector[offset] = (cb_sign & 1) ? cb_gain : -cb_gain;
512 /* Enhance harmonic components */
513 lag = pitch_contrib[subfrm.ad_cb_gain << 1] + pitch_lag +
514 subfrm.ad_cb_lag - 1;
515 beta = pitch_contrib[(subfrm.ad_cb_gain << 1) + 1];
517 if (lag < SUBFRAME_LEN - 2) {
518 for (i = lag; i < SUBFRAME_LEN; i++)
519 vector[i] += beta * vector[i - lag] >> 15;
525 * Get delayed contribution from the previous excitation vector.
527 static void get_residual(int16_t *residual, int16_t *prev_excitation, int lag)
529 int offset = PITCH_MAX - PITCH_ORDER / 2 - lag;
532 residual[0] = prev_excitation[offset];
533 residual[1] = prev_excitation[offset + 1];
536 for (i = 2; i < SUBFRAME_LEN + PITCH_ORDER - 1; i++)
537 residual[i] = prev_excitation[offset + (i - 2) % lag];
541 * Generate adaptive codebook excitation.
543 static void gen_acb_excitation(int16_t *vector, int16_t *prev_excitation,
544 int pitch_lag, G723_1_Subframe subfrm,
547 int16_t residual[SUBFRAME_LEN + PITCH_ORDER - 1];
548 const int16_t *cb_ptr;
549 int lag = pitch_lag + subfrm.ad_cb_lag - 1;
554 get_residual(residual, prev_excitation, lag);
556 /* Select quantization table */
557 if (cur_rate == Rate6k3 && pitch_lag < SUBFRAME_LEN - 2) {
558 cb_ptr = adaptive_cb_gain85;
560 cb_ptr = adaptive_cb_gain170;
562 /* Calculate adaptive vector */
563 cb_ptr += subfrm.ad_cb_gain * 20;
564 for (i = 0; i < SUBFRAME_LEN; i++) {
565 sum = ff_dot_product(residual + i, cb_ptr, PITCH_ORDER);
566 vector[i] = av_clipl_int32((sum << 2) + (1 << 15)) >> 16;
571 * Estimate maximum auto-correlation around pitch lag.
573 * @param p the context
574 * @param offset offset of the excitation vector
575 * @param ccr_max pointer to the maximum auto-correlation
576 * @param pitch_lag decoded pitch lag
577 * @param length length of autocorrelation
578 * @param dir forward lag(1) / backward lag(-1)
580 static int autocorr_max(G723_1_Context *p, int offset, int *ccr_max,
581 int pitch_lag, int length, int dir)
583 int limit, ccr, lag = 0;
584 int16_t *buf = p->excitation + offset;
587 pitch_lag = FFMIN(PITCH_MAX - 3, pitch_lag);
588 limit = FFMIN(FRAME_LEN + PITCH_MAX - offset - length, pitch_lag + 3);
590 for (i = pitch_lag - 3; i <= limit; i++) {
591 ccr = ff_dot_product(buf, buf + dir * i, length)<<1;
593 if (ccr > *ccr_max) {
602 * Calculate pitch postfilter optimal and scaling gains.
604 * @param lag pitch postfilter forward/backward lag
605 * @param ppf pitch postfilter parameters
606 * @param cur_rate current bitrate
607 * @param tgt_eng target energy
608 * @param ccr cross-correlation
609 * @param res_eng residual energy
611 static void comp_ppf_gains(int lag, PPFParam *ppf, Rate cur_rate,
612 int tgt_eng, int ccr, int res_eng)
614 int pf_residual; /* square of postfiltered residual */
615 int64_t temp1, temp2;
619 temp1 = tgt_eng * res_eng >> 1;
620 temp2 = ccr * ccr << 1;
623 if (ccr >= res_eng) {
624 ppf->opt_gain = ppf_gain_weight[cur_rate];
626 ppf->opt_gain = (ccr << 15) / res_eng *
627 ppf_gain_weight[cur_rate] >> 15;
629 /* pf_res^2 = tgt_eng + 2*ccr*gain + res_eng*gain^2 */
630 temp1 = (tgt_eng << 15) + (ccr * ppf->opt_gain << 1);
631 temp2 = (ppf->opt_gain * ppf->opt_gain >> 15) * res_eng;
632 pf_residual = av_clipl_int32(temp1 + temp2 + (1 << 15)) >> 16;
634 if (tgt_eng >= pf_residual << 1) {
637 temp1 = (tgt_eng << 14) / pf_residual;
640 /* scaling_gain = sqrt(tgt_eng/pf_res^2) */
641 ppf->sc_gain = square_root(temp1 << 16);
644 ppf->sc_gain = 0x7fff;
647 ppf->opt_gain = av_clip_int16(ppf->opt_gain * ppf->sc_gain >> 15);
651 * Calculate pitch postfilter parameters.
653 * @param p the context
654 * @param offset offset of the excitation vector
655 * @param pitch_lag decoded pitch lag
656 * @param ppf pitch postfilter parameters
657 * @param cur_rate current bitrate
659 static void comp_ppf_coeff(G723_1_Context *p, int offset, int pitch_lag,
660 PPFParam *ppf, Rate cur_rate)
665 int64_t temp1, temp2;
669 * 1 - forward cross-correlation
670 * 2 - forward residual energy
671 * 3 - backward cross-correlation
672 * 4 - backward residual energy
674 int energy[5] = {0, 0, 0, 0, 0};
675 int16_t *buf = p->excitation + offset;
676 int fwd_lag = autocorr_max(p, offset, &energy[1], pitch_lag,
678 int back_lag = autocorr_max(p, offset, &energy[3], pitch_lag,
683 ppf->sc_gain = 0x7fff;
685 /* Case 0, Section 3.6 */
686 if (!back_lag && !fwd_lag)
689 /* Compute target energy */
690 energy[0] = ff_dot_product(buf, buf, SUBFRAME_LEN)<<1;
692 /* Compute forward residual energy */
694 energy[2] = ff_dot_product(buf + fwd_lag, buf + fwd_lag,
697 /* Compute backward residual energy */
699 energy[4] = ff_dot_product(buf - back_lag, buf - back_lag,
702 /* Normalize and shorten */
704 for (i = 0; i < 5; i++)
705 temp1 = FFMAX(energy[i], temp1);
707 scale = normalize_bits(temp1, 1);
708 for (i = 0; i < 5; i++)
709 energy[i] = av_clipl_int32(energy[i] << scale) >> 16;
711 if (fwd_lag && !back_lag) { /* Case 1 */
712 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
714 } else if (!fwd_lag) { /* Case 2 */
715 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
717 } else { /* Case 3 */
720 * Select the largest of energy[1]^2/energy[2]
721 * and energy[3]^2/energy[4]
723 temp1 = energy[4] * ((energy[1] * energy[1] + (1 << 14)) >> 15);
724 temp2 = energy[2] * ((energy[3] * energy[3] + (1 << 14)) >> 15);
725 if (temp1 >= temp2) {
726 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
729 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
736 * Classify frames as voiced/unvoiced.
738 * @param p the context
739 * @param pitch_lag decoded pitch_lag
740 * @param exc_eng excitation energy estimation
741 * @param scale scaling factor of exc_eng
743 * @return residual interpolation index if voiced, 0 otherwise
745 static int comp_interp_index(G723_1_Context *p, int pitch_lag,
746 int *exc_eng, int *scale)
748 int offset = PITCH_MAX + 2 * SUBFRAME_LEN;
749 int16_t *buf = p->excitation + offset;
751 int index, ccr, tgt_eng, best_eng, temp;
753 *scale = scale_vector(p->excitation, FRAME_LEN + PITCH_MAX);
755 /* Compute maximum backward cross-correlation */
757 index = autocorr_max(p, offset, &ccr, pitch_lag, SUBFRAME_LEN * 2, -1);
758 ccr = av_clipl_int32((int64_t)ccr + (1 << 15)) >> 16;
760 /* Compute target energy */
761 tgt_eng = ff_dot_product(buf, buf, SUBFRAME_LEN * 2)<<1;
762 *exc_eng = av_clipl_int32(tgt_eng + (1 << 15)) >> 16;
767 /* Compute best energy */
768 best_eng = ff_dot_product(buf - index, buf - index,
769 SUBFRAME_LEN * 2)<<1;
770 best_eng = av_clipl_int32((int64_t)best_eng + (1 << 15)) >> 16;
772 temp = best_eng * *exc_eng >> 3;
774 if (temp < ccr * ccr) {
781 * Peform residual interpolation based on frame classification.
783 * @param buf decoded excitation vector
784 * @param out output vector
785 * @param lag decoded pitch lag
786 * @param gain interpolated gain
787 * @param rseed seed for random number generator
789 static void residual_interp(int16_t *buf, int16_t *out, int lag,
790 int gain, int *rseed)
793 if (lag) { /* Voiced */
794 int16_t *vector_ptr = buf + PITCH_MAX;
796 for (i = 0; i < lag; i++)
797 vector_ptr[i - lag] = vector_ptr[i - lag] * 3 >> 2;
798 av_memcpy_backptr((uint8_t*)vector_ptr, lag * sizeof(int16_t),
799 FRAME_LEN * sizeof(int16_t));
800 memcpy(out, vector_ptr, FRAME_LEN * sizeof(int16_t));
801 } else { /* Unvoiced */
802 for (i = 0; i < FRAME_LEN; i++) {
803 *rseed = *rseed * 521 + 259;
804 out[i] = gain * *rseed >> 15;
806 memset(buf, 0, (FRAME_LEN + PITCH_MAX) * sizeof(int16_t));
811 * Perform IIR filtering.
813 * @param fir_coef FIR coefficients
814 * @param iir_coef IIR coefficients
815 * @param src source vector
816 * @param dest destination vector
817 * @param width width of the output, 16 bits(0) / 32 bits(1)
819 #define iir_filter(fir_coef, iir_coef, src, dest, width)\
822 int res_shift = 16 & ~-(width);\
823 int in_shift = 16 - res_shift;\
825 for (m = 0; m < SUBFRAME_LEN; m++) {\
827 for (n = 1; n <= LPC_ORDER; n++) {\
828 filter -= (fir_coef)[n - 1] * (src)[m - n] -\
829 (iir_coef)[n - 1] * ((dest)[m - n] >> in_shift);\
832 (dest)[m] = av_clipl_int32(((src)[m] << 16) + (filter << 3) +\
833 (1 << 15)) >> res_shift;\
838 * Adjust gain of postfiltered signal.
840 * @param p the context
841 * @param buf postfiltered output vector
842 * @param energy input energy coefficient
844 static void gain_scale(G723_1_Context *p, int16_t * buf, int energy)
846 int num, denom, gain, bits1, bits2;
851 for (i = 0; i < SUBFRAME_LEN; i++) {
852 int64_t temp = buf[i] >> 2;
853 temp = av_clipl_int32(MUL64(temp, temp) << 1);
854 denom = av_clipl_int32(denom + temp);
858 bits1 = normalize_bits(num, 1);
859 bits2 = normalize_bits(denom, 1);
860 num = num << bits1 >> 1;
863 bits2 = 5 + bits1 - bits2;
864 bits2 = FFMAX(0, bits2);
866 gain = (num >> 1) / (denom >> 16);
867 gain = square_root(gain << 16 >> bits2);
872 for (i = 0; i < SUBFRAME_LEN; i++) {
873 p->pf_gain = ((p->pf_gain << 4) - p->pf_gain + gain + (1 << 3)) >> 4;
874 buf[i] = av_clip_int16((buf[i] * (p->pf_gain + (p->pf_gain >> 4)) +
880 * Perform formant filtering.
882 * @param p the context
883 * @param lpc quantized lpc coefficients
884 * @param buf output buffer
886 static void formant_postfilter(G723_1_Context *p, int16_t *lpc, int16_t *buf)
888 int16_t filter_coef[2][LPC_ORDER], *buf_ptr;
889 int filter_signal[LPC_ORDER + FRAME_LEN], *signal_ptr;
892 memcpy(buf, p->fir_mem, LPC_ORDER * sizeof(int16_t));
893 memcpy(filter_signal, p->iir_mem, LPC_ORDER * sizeof(int));
895 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
896 for (k = 0; k < LPC_ORDER; k++) {
897 filter_coef[0][k] = (-lpc[k] * postfilter_tbl[0][k] +
899 filter_coef[1][k] = (-lpc[k] * postfilter_tbl[1][k] +
902 iir_filter(filter_coef[0], filter_coef[1], buf + i,
903 filter_signal + i, 1);
906 memcpy(p->fir_mem, buf + FRAME_LEN, LPC_ORDER * sizeof(int16_t));
907 memcpy(p->iir_mem, filter_signal + FRAME_LEN, LPC_ORDER * sizeof(int));
909 buf_ptr = buf + LPC_ORDER;
910 signal_ptr = filter_signal + LPC_ORDER;
911 for (i = 0; i < SUBFRAMES; i++) {
912 int16_t temp_vector[SUBFRAME_LEN];
918 memcpy(temp_vector, buf_ptr, SUBFRAME_LEN * sizeof(int16_t));
919 scale = scale_vector(temp_vector, SUBFRAME_LEN);
921 /* Compute auto correlation coefficients */
922 auto_corr[0] = ff_dot_product(temp_vector, temp_vector + 1,
923 SUBFRAME_LEN - 1)<<1;
924 auto_corr[1] = ff_dot_product(temp_vector, temp_vector,
927 /* Compute reflection coefficient */
928 temp = auto_corr[1] >> 16;
930 temp = (auto_corr[0] >> 2) / temp;
932 p->reflection_coef = ((p->reflection_coef << 2) - p->reflection_coef +
934 temp = (p->reflection_coef * 0xffffc >> 3) & 0xfffc;
936 /* Compensation filter */
937 for (j = 0; j < SUBFRAME_LEN; j++) {
938 buf_ptr[j] = av_clipl_int32(signal_ptr[j] +
939 ((signal_ptr[j - 1] >> 16) *
943 /* Compute normalized signal energy */
944 temp = 2 * scale + 4;
946 energy = av_clipl_int32((int64_t)auto_corr[1] << -temp);
948 energy = auto_corr[1] >> temp;
950 gain_scale(p, buf_ptr, energy);
952 buf_ptr += SUBFRAME_LEN;
953 signal_ptr += SUBFRAME_LEN;
957 static int g723_1_decode_frame(AVCodecContext *avctx, void *data,
958 int *got_frame_ptr, AVPacket *avpkt)
960 G723_1_Context *p = avctx->priv_data;
961 const uint8_t *buf = avpkt->data;
962 int buf_size = avpkt->size;
964 int dec_mode = buf[0] & 3;
966 PPFParam ppf[SUBFRAMES];
967 int16_t cur_lsp[LPC_ORDER];
968 int16_t lpc[SUBFRAMES * LPC_ORDER];
969 int16_t acb_vector[SUBFRAME_LEN];
971 int bad_frame = 0, i, j, ret;
973 if (!buf_size || buf_size < frame_size[dec_mode]) {
978 if (unpack_bitstream(p, buf, buf_size) < 0) {
980 p->cur_frame_type = p->past_frame_type == ActiveFrame ?
981 ActiveFrame : UntransmittedFrame;
984 p->frame.nb_samples = FRAME_LEN + LPC_ORDER;
985 if ((ret = avctx->get_buffer(avctx, &p->frame)) < 0) {
986 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
989 out= (int16_t*)p->frame.data[0];
992 if(p->cur_frame_type == ActiveFrame) {
994 p->erased_frames = 0;
995 } else if(p->erased_frames != 3)
998 inverse_quant(cur_lsp, p->prev_lsp, p->lsp_index, bad_frame);
999 lsp_interpolate(lpc, cur_lsp, p->prev_lsp);
1001 /* Save the lsp_vector for the next frame */
1002 memcpy(p->prev_lsp, cur_lsp, LPC_ORDER * sizeof(int16_t));
1004 /* Generate the excitation for the frame */
1005 memcpy(p->excitation, p->prev_excitation, PITCH_MAX * sizeof(int16_t));
1006 vector_ptr = p->excitation + PITCH_MAX;
1007 if (!p->erased_frames) {
1008 /* Update interpolation gain memory */
1009 p->interp_gain = fixed_cb_gain[(p->subframe[2].amp_index +
1010 p->subframe[3].amp_index) >> 1];
1011 for (i = 0; i < SUBFRAMES; i++) {
1012 gen_fcb_excitation(vector_ptr, p->subframe[i], p->cur_rate,
1013 p->pitch_lag[i >> 1], i);
1014 gen_acb_excitation(acb_vector, &p->excitation[SUBFRAME_LEN * i],
1015 p->pitch_lag[i >> 1], p->subframe[i],
1017 /* Get the total excitation */
1018 for (j = 0; j < SUBFRAME_LEN; j++) {
1019 vector_ptr[j] = av_clip_int16(vector_ptr[j] << 1);
1020 vector_ptr[j] = av_clip_int16(vector_ptr[j] +
1023 vector_ptr += SUBFRAME_LEN;
1026 vector_ptr = p->excitation + PITCH_MAX;
1028 /* Save the excitation */
1029 memcpy(out, vector_ptr, FRAME_LEN * sizeof(int16_t));
1031 p->interp_index = comp_interp_index(p, p->pitch_lag[1],
1032 &p->sid_gain, &p->cur_gain);
1034 for (i = PITCH_MAX, 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 /* Restore the original excitation */
1039 memcpy(p->excitation, p->prev_excitation,
1040 PITCH_MAX * sizeof(int16_t));
1041 memcpy(vector_ptr, out, FRAME_LEN * sizeof(int16_t));
1043 /* Peform pitch postfiltering */
1044 for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1045 ff_acelp_weighted_vector_sum(out + LPC_ORDER + i, vector_ptr + i,
1046 vector_ptr + i + ppf[j].index,
1047 ppf[j].sc_gain, ppf[j].opt_gain,
1048 1 << 14, 15, SUBFRAME_LEN);
1050 p->interp_gain = (p->interp_gain * 3 + 2) >> 2;
1051 if (p->erased_frames == 3) {
1053 memset(p->excitation, 0,
1054 (FRAME_LEN + PITCH_MAX) * sizeof(int16_t));
1055 memset(out, 0, (FRAME_LEN + LPC_ORDER) * sizeof(int16_t));
1057 /* Regenerate frame */
1058 residual_interp(p->excitation, out + LPC_ORDER, p->interp_index,
1059 p->interp_gain, &p->random_seed);
1062 /* Save the excitation for the next frame */
1063 memcpy(p->prev_excitation, p->excitation + FRAME_LEN,
1064 PITCH_MAX * sizeof(int16_t));
1066 memset(out, 0, sizeof(int16_t)*FRAME_LEN);
1067 av_log(avctx, AV_LOG_WARNING,
1068 "G.723.1: Comfort noise generation not supported yet\n");
1069 return frame_size[dec_mode];
1072 p->past_frame_type = p->cur_frame_type;
1074 memcpy(out, p->synth_mem, LPC_ORDER * sizeof(int16_t));
1075 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1076 ff_celp_lp_synthesis_filter(out + i, &lpc[j * LPC_ORDER],
1077 out + i, SUBFRAME_LEN, LPC_ORDER,
1079 memcpy(p->synth_mem, out + FRAME_LEN, LPC_ORDER * sizeof(int16_t));
1081 formant_postfilter(p, lpc, out);
1083 memmove(out, out + LPC_ORDER, sizeof(int16_t)*FRAME_LEN);
1084 p->frame.nb_samples = FRAME_LEN;
1085 *(AVFrame*)data = p->frame;
1088 return frame_size[dec_mode];
1091 AVCodec ff_g723_1_decoder = {
1093 .type = AVMEDIA_TYPE_AUDIO,
1094 .id = CODEC_ID_G723_1,
1095 .priv_data_size = sizeof(G723_1_Context),
1096 .init = g723_1_decode_init,
1097 .decode = g723_1_decode_frame,
1098 .long_name = NULL_IF_CONFIG_SMALL("G.723.1"),
1099 .capabilities = CODEC_CAP_SUBFRAMES | CODEC_CAP_DR1,
1102 #if CONFIG_G723_1_ENCODER
1103 #define BITSTREAM_WRITER_LE
1104 #include "put_bits.h"
1106 static av_cold int g723_1_encode_init(AVCodecContext *avctx)
1108 G723_1_Context *p = avctx->priv_data;
1110 if (avctx->sample_rate != 8000) {
1111 av_log(avctx, AV_LOG_ERROR, "Only 8000Hz sample rate supported\n");
1115 if (avctx->channels != 1) {
1116 av_log(avctx, AV_LOG_ERROR, "Only mono supported\n");
1117 return AVERROR(EINVAL);
1120 if (avctx->bit_rate == 6300) {
1121 p->cur_rate = Rate6k3;
1122 } else if (avctx->bit_rate == 5300) {
1123 av_log(avctx, AV_LOG_ERROR, "Bitrate not supported yet, use 6.3k\n");
1124 return AVERROR_PATCHWELCOME;
1126 av_log(avctx, AV_LOG_ERROR,
1127 "Bitrate not supported, use 6.3k\n");
1128 return AVERROR(EINVAL);
1130 avctx->frame_size = 240;
1131 memcpy(p->prev_lsp, dc_lsp, LPC_ORDER * sizeof(int16_t));
1137 * Remove DC component from the input signal.
1139 * @param buf input signal
1140 * @param fir zero memory
1141 * @param iir pole memory
1143 static void highpass_filter(int16_t *buf, int16_t *fir, int *iir)
1146 for (i = 0; i < FRAME_LEN; i++) {
1147 *iir = (buf[i] << 15) + ((-*fir) << 15) + MULL2(*iir, 0x7f00);
1149 buf[i] = av_clipl_int32((int64_t)*iir + (1 << 15)) >> 16;
1154 * Estimate autocorrelation of the input vector.
1156 * @param buf input buffer
1157 * @param autocorr autocorrelation coefficients vector
1159 static void comp_autocorr(int16_t *buf, int16_t *autocorr)
1162 int16_t vector[LPC_FRAME];
1164 memcpy(vector, buf, LPC_FRAME * sizeof(int16_t));
1165 scale_vector(vector, LPC_FRAME);
1167 /* Apply the Hamming window */
1168 for (i = 0; i < LPC_FRAME; i++)
1169 vector[i] = (vector[i] * hamming_window[i] + (1 << 14)) >> 15;
1171 /* Compute the first autocorrelation coefficient */
1172 temp = dot_product(vector, vector, LPC_FRAME, 0);
1174 /* Apply a white noise correlation factor of (1025/1024) */
1178 scale = normalize_bits_int32(temp);
1179 autocorr[0] = av_clipl_int32((int64_t)(temp << scale) +
1182 /* Compute the remaining coefficients */
1184 memset(autocorr + 1, 0, LPC_ORDER * sizeof(int16_t));
1186 for (i = 1; i <= LPC_ORDER; i++) {
1187 temp = dot_product(vector, vector + i, LPC_FRAME - i, 0);
1188 temp = MULL2((temp << scale), binomial_window[i - 1]);
1189 autocorr[i] = av_clipl_int32((int64_t)temp + (1 << 15)) >> 16;
1195 * Use Levinson-Durbin recursion to compute LPC coefficients from
1196 * autocorrelation values.
1198 * @param lpc LPC coefficients vector
1199 * @param autocorr autocorrelation coefficients vector
1200 * @param error prediction error
1202 static void levinson_durbin(int16_t *lpc, int16_t *autocorr, int16_t error)
1204 int16_t vector[LPC_ORDER];
1205 int16_t partial_corr;
1208 memset(lpc, 0, LPC_ORDER * sizeof(int16_t));
1210 for (i = 0; i < LPC_ORDER; i++) {
1211 /* Compute the partial correlation coefficient */
1213 for (j = 0; j < i; j++)
1214 temp -= lpc[j] * autocorr[i - j - 1];
1215 temp = ((autocorr[i] << 13) + temp) << 3;
1217 if (FFABS(temp) >= (error << 16))
1220 partial_corr = temp / (error << 1);
1222 lpc[i] = av_clipl_int32((int64_t)(partial_corr << 14) +
1225 /* Update the prediction error */
1226 temp = MULL2(temp, partial_corr);
1227 error = av_clipl_int32((int64_t)(error << 16) - temp +
1230 memcpy(vector, lpc, i * sizeof(int16_t));
1231 for (j = 0; j < i; j++) {
1232 temp = partial_corr * vector[i - j - 1] << 1;
1233 lpc[j] = av_clipl_int32((int64_t)(lpc[j] << 16) - temp +
1240 * Calculate LPC coefficients for the current frame.
1242 * @param buf current frame
1243 * @param prev_data 2 trailing subframes of the previous frame
1244 * @param lpc LPC coefficients vector
1246 static void comp_lpc_coeff(int16_t *buf, int16_t *lpc)
1248 int16_t autocorr[(LPC_ORDER + 1) * SUBFRAMES];
1249 int16_t *autocorr_ptr = autocorr;
1250 int16_t *lpc_ptr = lpc;
1253 for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
1254 comp_autocorr(buf + i, autocorr_ptr);
1255 levinson_durbin(lpc_ptr, autocorr_ptr + 1, autocorr_ptr[0]);
1257 lpc_ptr += LPC_ORDER;
1258 autocorr_ptr += LPC_ORDER + 1;
1262 static void lpc2lsp(int16_t *lpc, int16_t *prev_lsp, int16_t *lsp)
1264 int f[LPC_ORDER + 2]; ///< coefficients of the sum and difference
1265 ///< polynomials (F1, F2) ordered as
1266 ///< f1[0], f2[0], ...., f1[5], f2[5]
1268 int max, shift, cur_val, prev_val, count, p;
1272 /* Initialize f1[0] and f2[0] to 1 in Q25 */
1273 for (i = 0; i < LPC_ORDER; i++)
1274 lsp[i] = (lpc[i] * bandwidth_expand[i] + (1 << 14)) >> 15;
1276 /* Apply bandwidth expansion on the LPC coefficients */
1277 f[0] = f[1] = 1 << 25;
1279 /* Compute the remaining coefficients */
1280 for (i = 0; i < LPC_ORDER / 2; i++) {
1282 f[2 * i + 2] = -f[2 * i] - ((lsp[i] + lsp[LPC_ORDER - 1 - i]) << 12);
1284 f[2 * i + 3] = f[2 * i + 1] - ((lsp[i] - lsp[LPC_ORDER - 1 - i]) << 12);
1287 /* Divide f1[5] and f2[5] by 2 for use in polynomial evaluation */
1289 f[LPC_ORDER + 1] >>= 1;
1291 /* Normalize and shorten */
1293 for (i = 1; i < LPC_ORDER + 2; i++)
1294 max = FFMAX(max, FFABS(f[i]));
1296 shift = normalize_bits_int32(max);
1298 for (i = 0; i < LPC_ORDER + 2; i++)
1299 f[i] = av_clipl_int32((int64_t)(f[i] << shift) + (1 << 15)) >> 16;
1302 * Evaluate F1 and F2 at uniform intervals of pi/256 along the
1303 * unit circle and check for zero crossings.
1307 for (i = 0; i <= LPC_ORDER / 2; i++)
1308 temp += f[2 * i] * cos_tab[0];
1309 prev_val = av_clipl_int32(temp << 1);
1311 for ( i = 1; i < COS_TBL_SIZE / 2; i++) {
1314 for (j = 0; j <= LPC_ORDER / 2; j++)
1315 temp += f[LPC_ORDER - 2 * j + p] * cos_tab[i * j % COS_TBL_SIZE];
1316 cur_val = av_clipl_int32(temp << 1);
1318 /* Check for sign change, indicating a zero crossing */
1319 if ((cur_val ^ prev_val) < 0) {
1320 int abs_cur = FFABS(cur_val);
1321 int abs_prev = FFABS(prev_val);
1322 int sum = abs_cur + abs_prev;
1324 shift = normalize_bits_int32(sum);
1326 abs_prev = abs_prev << shift >> 8;
1327 lsp[count++] = ((i - 1) << 7) + (abs_prev >> 1) / (sum >> 16);
1329 if (count == LPC_ORDER)
1332 /* Switch between sum and difference polynomials */
1337 for (j = 0; j <= LPC_ORDER / 2; j++){
1338 temp += f[LPC_ORDER - 2 * j + p] *
1339 cos_tab[i * j % COS_TBL_SIZE];
1341 cur_val = av_clipl_int32(temp<<1);
1346 if (count != LPC_ORDER)
1347 memcpy(lsp, prev_lsp, LPC_ORDER * sizeof(int16_t));
1351 * Quantize the current LSP subvector.
1353 * @param num band number
1354 * @param offset offset of the current subvector in an LPC_ORDER vector
1355 * @param size size of the current subvector
1357 #define get_index(num, offset, size) \
1359 int error, max = -1;\
1362 for (i = 0; i < LSP_CB_SIZE; i++) {\
1363 for (j = 0; j < size; j++){\
1364 temp[j] = (weight[j + (offset)] * lsp_band##num[i][j] +\
1367 error = dot_product(lsp + (offset), temp, size, 1) << 1;\
1368 error -= dot_product(lsp_band##num[i], temp, size, 1);\
1371 lsp_index[num] = i;\
1377 * Vector quantize the LSP frequencies.
1379 * @param lsp the current lsp vector
1380 * @param prev_lsp the previous lsp vector
1382 static void lsp_quantize(uint8_t *lsp_index, int16_t *lsp, int16_t *prev_lsp)
1384 int16_t weight[LPC_ORDER];
1388 /* Calculate the VQ weighting vector */
1389 weight[0] = (1 << 20) / (lsp[1] - lsp[0]);
1390 weight[LPC_ORDER - 1] = (1 << 20) /
1391 (lsp[LPC_ORDER - 1] - lsp[LPC_ORDER - 2]);
1393 for (i = 1; i < LPC_ORDER - 1; i++) {
1394 min = FFMIN(lsp[i] - lsp[i - 1], lsp[i + 1] - lsp[i]);
1396 weight[i] = (1 << 20) / min;
1398 weight[i] = INT16_MAX;
1403 for (i = 0; i < LPC_ORDER; i++)
1404 max = FFMAX(weight[i], max);
1406 shift = normalize_bits_int16(max);
1407 for (i = 0; i < LPC_ORDER; i++) {
1408 weight[i] <<= shift;
1411 /* Compute the VQ target vector */
1412 for (i = 0; i < LPC_ORDER; i++) {
1413 lsp[i] -= dc_lsp[i] +
1414 (((prev_lsp[i] - dc_lsp[i]) * 12288 + (1 << 14)) >> 15);
1423 * Apply the formant perceptual weighting filter.
1425 * @param flt_coef filter coefficients
1426 * @param unq_lpc unquantized lpc vector
1428 static void perceptual_filter(G723_1_Context *p, int16_t *flt_coef,
1429 int16_t *unq_lpc, int16_t *buf)
1431 int16_t vector[FRAME_LEN + LPC_ORDER];
1434 memcpy(buf, p->iir_mem, sizeof(int16_t) * LPC_ORDER);
1435 memcpy(vector, p->fir_mem, sizeof(int16_t) * LPC_ORDER);
1436 memcpy(vector + LPC_ORDER, buf + LPC_ORDER, sizeof(int16_t) * FRAME_LEN);
1438 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
1439 for (k = 0; k < LPC_ORDER; k++) {
1440 flt_coef[k + 2 * l] = (unq_lpc[k + l] * percept_flt_tbl[0][k] +
1442 flt_coef[k + 2 * l + LPC_ORDER] = (unq_lpc[k + l] *
1443 percept_flt_tbl[1][k] +
1446 iir_filter(flt_coef + 2 * l, flt_coef + 2 * l + LPC_ORDER, vector + i,
1450 memcpy(p->iir_mem, buf + FRAME_LEN, sizeof(int16_t) * LPC_ORDER);
1451 memcpy(p->fir_mem, vector + FRAME_LEN, sizeof(int16_t) * LPC_ORDER);
1455 * Estimate the open loop pitch period.
1457 * @param buf perceptually weighted speech
1458 * @param start estimation is carried out from this position
1460 static int estimate_pitch(int16_t *buf, int start)
1463 int max_ccr = 0x4000;
1464 int max_eng = 0x7fff;
1465 int index = PITCH_MIN;
1466 int offset = start - PITCH_MIN + 1;
1468 int ccr, eng, orig_eng, ccr_eng, exp;
1473 orig_eng = dot_product(buf + offset, buf + offset, HALF_FRAME_LEN, 0);
1475 for (i = PITCH_MIN; i <= PITCH_MAX - 3; i++) {
1478 /* Update energy and compute correlation */
1479 orig_eng += buf[offset] * buf[offset] -
1480 buf[offset + HALF_FRAME_LEN] * buf[offset + HALF_FRAME_LEN];
1481 ccr = dot_product(buf + start, buf + offset, HALF_FRAME_LEN, 0);
1485 /* Split into mantissa and exponent to maintain precision */
1486 exp = normalize_bits_int32(ccr);
1487 ccr = av_clipl_int32((int64_t)(ccr << exp) + (1 << 15)) >> 16;
1490 temp = normalize_bits_int32(ccr);
1491 ccr = ccr << temp >> 16;
1494 temp = normalize_bits_int32(orig_eng);
1495 eng = av_clipl_int32((int64_t)(orig_eng << temp) + (1 << 15)) >> 16;
1505 if (exp + 1 < max_exp)
1508 /* Equalize exponents before comparison */
1509 if (exp + 1 == max_exp)
1510 temp = max_ccr >> 1;
1513 ccr_eng = ccr * max_eng;
1514 diff = ccr_eng - eng * temp;
1515 if (diff > 0 && (i - index < PITCH_MIN || diff > ccr_eng >> 2)) {
1527 * Compute harmonic noise filter parameters.
1529 * @param buf perceptually weighted speech
1530 * @param pitch_lag open loop pitch period
1531 * @param hf harmonic filter parameters
1533 static void comp_harmonic_coeff(int16_t *buf, int16_t pitch_lag, HFParam *hf)
1535 int ccr, eng, max_ccr, max_eng;
1540 for (i = 0, j = pitch_lag - 3; j <= pitch_lag + 3; i++, j++) {
1541 /* Compute residual energy */
1542 energy[i << 1] = dot_product(buf - j, buf - j, SUBFRAME_LEN, 0);
1543 /* Compute correlation */
1544 energy[(i << 1) + 1] = dot_product(buf, buf - j, SUBFRAME_LEN, 0);
1547 /* Compute target energy */
1548 energy[14] = dot_product(buf, buf, SUBFRAME_LEN, 0);
1552 for (i = 0; i < 15; i++)
1553 max = FFMAX(max, FFABS(energy[i]));
1555 exp = normalize_bits_int32(max);
1556 for (i = 0; i < 15; i++) {
1557 energy[i] = av_clipl_int32((int64_t)(energy[i] << exp) +
1566 for (i = 0; i <= 6; i++) {
1567 eng = energy[i << 1];
1568 ccr = energy[(i << 1) + 1];
1573 ccr = (ccr * ccr + (1 << 14)) >> 15;
1574 diff = ccr * max_eng - eng * max_ccr;
1582 if (hf->index == -1) {
1583 hf->index = pitch_lag;
1587 eng = energy[14] * max_eng;
1588 eng = (eng >> 2) + (eng >> 3);
1589 ccr = energy[(hf->index << 1) + 1] * energy[(hf->index << 1) + 1];
1591 eng = energy[(hf->index << 1) + 1];
1596 hf->gain = ((eng << 15) / max_eng * 0x2800 + (1 << 14)) >> 15;
1598 hf->index += pitch_lag - 3;
1602 * Apply the harmonic noise shaping filter.
1604 * @param hf filter parameters
1606 static void harmonic_filter(HFParam *hf, int16_t *src, int16_t *dest)
1610 for (i = 0; i < SUBFRAME_LEN; i++) {
1611 int64_t temp = hf->gain * src[i - hf->index] << 1;
1612 dest[i] = av_clipl_int32((src[i] << 16) - temp + (1 << 15)) >> 16;
1616 static void harmonic_noise_sub(HFParam *hf, int16_t *src, int16_t *dest)
1619 for (i = 0; i < SUBFRAME_LEN; i++) {
1620 int64_t temp = hf->gain * src[i - hf->index] << 1;
1621 dest[i] = av_clipl_int32(((dest[i] - src[i]) << 16) + temp +
1628 * Combined synthesis and formant perceptual weighting filer.
1630 * @param qnt_lpc quantized lpc coefficients
1631 * @param perf_lpc perceptual filter coefficients
1632 * @param perf_fir perceptual filter fir memory
1633 * @param perf_iir perceptual filter iir memory
1634 * @param scale the filter output will be scaled by 2^scale
1636 static void synth_percept_filter(int16_t *qnt_lpc, int16_t *perf_lpc,
1637 int16_t *perf_fir, int16_t *perf_iir,
1638 int16_t *src, int16_t *dest, int scale)
1641 int16_t buf_16[SUBFRAME_LEN + LPC_ORDER];
1642 int64_t buf[SUBFRAME_LEN];
1644 int16_t *bptr_16 = buf_16 + LPC_ORDER;
1646 memcpy(buf_16, perf_fir, sizeof(int16_t) * LPC_ORDER);
1647 memcpy(dest - LPC_ORDER, perf_iir, sizeof(int16_t) * LPC_ORDER);
1649 for (i = 0; i < SUBFRAME_LEN; i++) {
1651 for (j = 1; j <= LPC_ORDER; j++)
1652 temp -= qnt_lpc[j - 1] * bptr_16[i - j];
1654 buf[i] = (src[i] << 15) + (temp << 3);
1655 bptr_16[i] = av_clipl_int32(buf[i] + (1 << 15)) >> 16;
1658 for (i = 0; i < SUBFRAME_LEN; i++) {
1659 int64_t fir = 0, iir = 0;
1660 for (j = 1; j <= LPC_ORDER; j++) {
1661 fir -= perf_lpc[j - 1] * bptr_16[i - j];
1662 iir += perf_lpc[j + LPC_ORDER - 1] * dest[i - j];
1664 dest[i] = av_clipl_int32(((buf[i] + (fir << 3)) << scale) + (iir << 3) +
1667 memcpy(perf_fir, buf_16 + SUBFRAME_LEN, sizeof(int16_t) * LPC_ORDER);
1668 memcpy(perf_iir, dest + SUBFRAME_LEN - LPC_ORDER,
1669 sizeof(int16_t) * LPC_ORDER);
1673 * Compute the adaptive codebook contribution.
1675 * @param buf input signal
1676 * @param index the current subframe index
1678 static void acb_search(G723_1_Context *p, int16_t *residual,
1679 int16_t *impulse_resp, int16_t *buf,
1683 int16_t flt_buf[PITCH_ORDER][SUBFRAME_LEN];
1685 const int16_t *cb_tbl = adaptive_cb_gain85;
1687 int ccr_buf[PITCH_ORDER * SUBFRAMES << 2];
1689 int pitch_lag = p->pitch_lag[index >> 1];
1692 int odd_frame = index & 1;
1693 int iter = 3 + odd_frame;
1697 int i, j, k, l, max;
1701 if (pitch_lag == PITCH_MIN)
1704 pitch_lag = FFMIN(pitch_lag, PITCH_MAX - 5);
1707 for (i = 0; i < iter; i++) {
1708 get_residual(residual, p->prev_excitation, pitch_lag + i - 1);
1710 for (j = 0; j < SUBFRAME_LEN; j++) {
1712 for (k = 0; k <= j; k++)
1713 temp += residual[PITCH_ORDER - 1 + k] * impulse_resp[j - k];
1714 flt_buf[PITCH_ORDER - 1][j] = av_clipl_int32((temp << 1) +
1718 for (j = PITCH_ORDER - 2; j >= 0; j--) {
1719 flt_buf[j][0] = ((residual[j] << 13) + (1 << 14)) >> 15;
1720 for (k = 1; k < SUBFRAME_LEN; k++) {
1721 temp = (flt_buf[j + 1][k - 1] << 15) +
1722 residual[j] * impulse_resp[k];
1723 flt_buf[j][k] = av_clipl_int32((temp << 1) + (1 << 15)) >> 16;
1727 /* Compute crosscorrelation with the signal */
1728 for (j = 0; j < PITCH_ORDER; j++) {
1729 temp = dot_product(buf, flt_buf[j], SUBFRAME_LEN, 0);
1730 ccr_buf[count++] = av_clipl_int32(temp << 1);
1733 /* Compute energies */
1734 for (j = 0; j < PITCH_ORDER; j++) {
1735 ccr_buf[count++] = dot_product(flt_buf[j], flt_buf[j],
1739 for (j = 1; j < PITCH_ORDER; j++) {
1740 for (k = 0; k < j; k++) {
1741 temp = dot_product(flt_buf[j], flt_buf[k], SUBFRAME_LEN, 0);
1742 ccr_buf[count++] = av_clipl_int32(temp<<2);
1747 /* Normalize and shorten */
1749 for (i = 0; i < 20 * iter; i++)
1750 max = FFMAX(max, FFABS(ccr_buf[i]));
1752 temp = normalize_bits_int32(max);
1754 for (i = 0; i < 20 * iter; i++){
1755 ccr_buf[i] = av_clipl_int32((int64_t)(ccr_buf[i] << temp) +
1760 for (i = 0; i < iter; i++) {
1761 /* Select quantization table */
1762 if (!odd_frame && pitch_lag + i - 1 >= SUBFRAME_LEN - 2 ||
1763 odd_frame && pitch_lag >= SUBFRAME_LEN - 2) {
1764 cb_tbl = adaptive_cb_gain170;
1768 for (j = 0, k = 0; j < tbl_size; j++, k += 20) {
1770 for (l = 0; l < 20; l++)
1771 temp += ccr_buf[20 * i + l] * cb_tbl[k + l];
1772 temp = av_clipl_int32(temp);
1783 pitch_lag += acb_lag - 1;
1787 p->pitch_lag[index >> 1] = pitch_lag;
1788 p->subframe[index].ad_cb_lag = acb_lag;
1789 p->subframe[index].ad_cb_gain = acb_gain;
1793 * Subtract the adaptive codebook contribution from the input
1794 * to obtain the residual.
1796 * @param buf target vector
1798 static void sub_acb_contrib(int16_t *residual, int16_t *impulse_resp,
1802 /* Subtract adaptive CB contribution to obtain the residual */
1803 for (i = 0; i < SUBFRAME_LEN; i++) {
1804 int64_t temp = buf[i] << 14;
1805 for (j = 0; j <= i; j++)
1806 temp -= residual[j] * impulse_resp[i - j];
1808 buf[i] = av_clipl_int32((temp << 2) + (1 << 15)) >> 16;
1813 * Quantize the residual signal using the fixed codebook (MP-MLQ).
1815 * @param optim optimized fixed codebook parameters
1816 * @param buf excitation vector
1818 static void get_fcb_param(FCBParam *optim, int16_t *impulse_resp,
1819 int16_t *buf, int pulse_cnt, int pitch_lag)
1822 int16_t impulse_r[SUBFRAME_LEN];
1823 int16_t temp_corr[SUBFRAME_LEN];
1824 int16_t impulse_corr[SUBFRAME_LEN];
1826 int ccr1[SUBFRAME_LEN];
1827 int ccr2[SUBFRAME_LEN];
1828 int amp, err, max, max_amp_index, min, scale, i, j, k, l;
1832 /* Update impulse response */
1833 memcpy(impulse_r, impulse_resp, sizeof(int16_t) * SUBFRAME_LEN);
1834 param.dirac_train = 0;
1835 if (pitch_lag < SUBFRAME_LEN - 2) {
1836 param.dirac_train = 1;
1837 gen_dirac_train(impulse_r, pitch_lag);
1840 for (i = 0; i < SUBFRAME_LEN; i++)
1841 temp_corr[i] = impulse_r[i] >> 1;
1843 /* Compute impulse response autocorrelation */
1844 temp = dot_product(temp_corr, temp_corr, SUBFRAME_LEN, 1);
1846 scale = normalize_bits_int32(temp);
1847 impulse_corr[0] = av_clipl_int32((temp << scale) + (1 << 15)) >> 16;
1849 for (i = 1; i < SUBFRAME_LEN; i++) {
1850 temp = dot_product(temp_corr + i, temp_corr, SUBFRAME_LEN - i, 1);
1851 impulse_corr[i] = av_clipl_int32((temp << scale) + (1 << 15)) >> 16;
1854 /* Compute crosscorrelation of impulse response with residual signal */
1856 for (i = 0; i < SUBFRAME_LEN; i++){
1857 temp = dot_product(buf + i, impulse_r, SUBFRAME_LEN - i, 1);
1859 ccr1[i] = temp >> -scale;
1861 ccr1[i] = av_clipl_int32(temp << scale);
1865 for (i = 0; i < GRID_SIZE; i++) {
1866 /* Maximize the crosscorrelation */
1868 for (j = i; j < SUBFRAME_LEN; j += GRID_SIZE) {
1869 temp = FFABS(ccr1[j]);
1872 param.pulse_pos[0] = j;
1876 /* Quantize the gain (max crosscorrelation/impulse_corr[0]) */
1879 max_amp_index = GAIN_LEVELS - 2;
1880 for (j = max_amp_index; j >= 2; j--) {
1881 temp = av_clipl_int32((int64_t)fixed_cb_gain[j] *
1882 impulse_corr[0] << 1);
1883 temp = FFABS(temp - amp);
1891 /* Select additional gain values */
1892 for (j = 1; j < 5; j++) {
1893 for (k = i; k < SUBFRAME_LEN; k += GRID_SIZE) {
1897 param.amp_index = max_amp_index + j - 2;
1898 amp = fixed_cb_gain[param.amp_index];
1900 param.pulse_sign[0] = (ccr2[param.pulse_pos[0]] < 0) ? -amp : amp;
1901 temp_corr[param.pulse_pos[0]] = 1;
1903 for (k = 1; k < pulse_cnt; k++) {
1905 for (l = i; l < SUBFRAME_LEN; l += GRID_SIZE) {
1908 temp = impulse_corr[FFABS(l - param.pulse_pos[k - 1])];
1909 temp = av_clipl_int32((int64_t)temp *
1910 param.pulse_sign[k - 1] << 1);
1912 temp = FFABS(ccr2[l]);
1915 param.pulse_pos[k] = l;
1919 param.pulse_sign[k] = (ccr2[param.pulse_pos[k]] < 0) ?
1921 temp_corr[param.pulse_pos[k]] = 1;
1924 /* Create the error vector */
1925 memset(temp_corr, 0, sizeof(int16_t) * SUBFRAME_LEN);
1927 for (k = 0; k < pulse_cnt; k++)
1928 temp_corr[param.pulse_pos[k]] = param.pulse_sign[k];
1930 for (k = SUBFRAME_LEN - 1; k >= 0; k--) {
1932 for (l = 0; l <= k; l++) {
1933 int prod = av_clipl_int32((int64_t)temp_corr[l] *
1934 impulse_r[k - l] << 1);
1935 temp = av_clipl_int32(temp + prod);
1937 temp_corr[k] = temp << 2 >> 16;
1940 /* Compute square of error */
1942 for (k = 0; k < SUBFRAME_LEN; k++) {
1944 prod = av_clipl_int32((int64_t)buf[k] * temp_corr[k] << 1);
1945 err = av_clipl_int32(err - prod);
1946 prod = av_clipl_int32((int64_t)temp_corr[k] * temp_corr[k]);
1947 err = av_clipl_int32(err + prod);
1951 if (err < optim->min_err) {
1952 optim->min_err = err;
1953 optim->grid_index = i;
1954 optim->amp_index = param.amp_index;
1955 optim->dirac_train = param.dirac_train;
1957 for (k = 0; k < pulse_cnt; k++) {
1958 optim->pulse_sign[k] = param.pulse_sign[k];
1959 optim->pulse_pos[k] = param.pulse_pos[k];
1967 * Encode the pulse position and gain of the current subframe.
1969 * @param optim optimized fixed CB parameters
1970 * @param buf excitation vector
1972 static void pack_fcb_param(G723_1_Subframe *subfrm, FCBParam *optim,
1973 int16_t *buf, int pulse_cnt)
1977 j = PULSE_MAX - pulse_cnt;
1979 subfrm->pulse_sign = 0;
1980 subfrm->pulse_pos = 0;
1982 for (i = 0; i < SUBFRAME_LEN >> 1; i++) {
1983 int val = buf[optim->grid_index + (i << 1)];
1985 subfrm->pulse_pos += combinatorial_table[j][i];
1987 subfrm->pulse_sign <<= 1;
1988 if (val < 0) subfrm->pulse_sign++;
1991 if (j == PULSE_MAX) break;
1994 subfrm->amp_index = optim->amp_index;
1995 subfrm->grid_index = optim->grid_index;
1996 subfrm->dirac_train = optim->dirac_train;
2000 * Compute the fixed codebook excitation.
2002 * @param buf target vector
2003 * @param impulse_resp impulse response of the combined filter
2005 static void fcb_search(G723_1_Context *p, int16_t *impulse_resp,
2006 int16_t *buf, int index)
2009 int pulse_cnt = pulses[index];
2012 optim.min_err = 1 << 30;
2013 get_fcb_param(&optim, impulse_resp, buf, pulse_cnt, SUBFRAME_LEN);
2015 if (p->pitch_lag[index >> 1] < SUBFRAME_LEN - 2) {
2016 get_fcb_param(&optim, impulse_resp, buf, pulse_cnt,
2017 p->pitch_lag[index >> 1]);
2020 /* Reconstruct the excitation */
2021 memset(buf, 0, sizeof(int16_t) * SUBFRAME_LEN);
2022 for (i = 0; i < pulse_cnt; i++)
2023 buf[optim.pulse_pos[i]] = optim.pulse_sign[i];
2025 pack_fcb_param(&p->subframe[index], &optim, buf, pulse_cnt);
2027 if (optim.dirac_train)
2028 gen_dirac_train(buf, p->pitch_lag[index >> 1]);
2032 * Pack the frame parameters into output bitstream.
2034 * @param frame output buffer
2035 * @param size size of the buffer
2037 static int pack_bitstream(G723_1_Context *p, unsigned char *frame, int size)
2040 int info_bits, i, temp;
2042 init_put_bits(&pb, frame, size);
2044 if (p->cur_rate == Rate6k3) {
2046 put_bits(&pb, 2, info_bits);
2049 put_bits(&pb, 8, p->lsp_index[2]);
2050 put_bits(&pb, 8, p->lsp_index[1]);
2051 put_bits(&pb, 8, p->lsp_index[0]);
2053 put_bits(&pb, 7, p->pitch_lag[0] - PITCH_MIN);
2054 put_bits(&pb, 2, p->subframe[1].ad_cb_lag);
2055 put_bits(&pb, 7, p->pitch_lag[1] - PITCH_MIN);
2056 put_bits(&pb, 2, p->subframe[3].ad_cb_lag);
2058 /* Write 12 bit combined gain */
2059 for (i = 0; i < SUBFRAMES; i++) {
2060 temp = p->subframe[i].ad_cb_gain * GAIN_LEVELS +
2061 p->subframe[i].amp_index;
2062 if (p->cur_rate == Rate6k3)
2063 temp += p->subframe[i].dirac_train << 11;
2064 put_bits(&pb, 12, temp);
2067 put_bits(&pb, 1, p->subframe[0].grid_index);
2068 put_bits(&pb, 1, p->subframe[1].grid_index);
2069 put_bits(&pb, 1, p->subframe[2].grid_index);
2070 put_bits(&pb, 1, p->subframe[3].grid_index);
2072 if (p->cur_rate == Rate6k3) {
2073 skip_put_bits(&pb, 1); /* reserved bit */
2075 /* Write 13 bit combined position index */
2076 temp = (p->subframe[0].pulse_pos >> 16) * 810 +
2077 (p->subframe[1].pulse_pos >> 14) * 90 +
2078 (p->subframe[2].pulse_pos >> 16) * 9 +
2079 (p->subframe[3].pulse_pos >> 14);
2080 put_bits(&pb, 13, temp);
2082 put_bits(&pb, 16, p->subframe[0].pulse_pos & 0xffff);
2083 put_bits(&pb, 14, p->subframe[1].pulse_pos & 0x3fff);
2084 put_bits(&pb, 16, p->subframe[2].pulse_pos & 0xffff);
2085 put_bits(&pb, 14, p->subframe[3].pulse_pos & 0x3fff);
2087 put_bits(&pb, 6, p->subframe[0].pulse_sign);
2088 put_bits(&pb, 5, p->subframe[1].pulse_sign);
2089 put_bits(&pb, 6, p->subframe[2].pulse_sign);
2090 put_bits(&pb, 5, p->subframe[3].pulse_sign);
2093 flush_put_bits(&pb);
2094 return frame_size[info_bits];
2097 static int g723_1_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
2098 const AVFrame *frame, int *got_packet_ptr)
2100 G723_1_Context *p = avctx->priv_data;
2101 int16_t unq_lpc[LPC_ORDER * SUBFRAMES];
2102 int16_t qnt_lpc[LPC_ORDER * SUBFRAMES];
2103 int16_t cur_lsp[LPC_ORDER];
2104 int16_t weighted_lpc[LPC_ORDER * SUBFRAMES << 1];
2105 int16_t vector[FRAME_LEN + PITCH_MAX];
2107 int16_t *in = (const int16_t *)frame->data[0];
2112 highpass_filter(in, &p->hpf_fir_mem, &p->hpf_iir_mem);
2114 memcpy(vector, p->prev_data, HALF_FRAME_LEN * sizeof(int16_t));
2115 memcpy(vector + HALF_FRAME_LEN, in, FRAME_LEN * sizeof(int16_t));
2117 comp_lpc_coeff(vector, unq_lpc);
2118 lpc2lsp(&unq_lpc[LPC_ORDER * 3], p->prev_lsp, cur_lsp);
2119 lsp_quantize(p->lsp_index, cur_lsp, p->prev_lsp);
2122 memcpy(vector + LPC_ORDER, p->prev_data + SUBFRAME_LEN,
2123 sizeof(int16_t) * SUBFRAME_LEN);
2124 memcpy(vector + LPC_ORDER + SUBFRAME_LEN, in,
2125 sizeof(int16_t) * (HALF_FRAME_LEN + SUBFRAME_LEN));
2126 memcpy(p->prev_data, in + HALF_FRAME_LEN,
2127 sizeof(int16_t) * HALF_FRAME_LEN);
2128 memcpy(in, vector + LPC_ORDER, sizeof(int16_t) * FRAME_LEN);
2130 perceptual_filter(p, weighted_lpc, unq_lpc, vector);
2132 memcpy(in, vector + LPC_ORDER, sizeof(int16_t) * FRAME_LEN);
2133 memcpy(vector, p->prev_weight_sig, sizeof(int16_t) * PITCH_MAX);
2134 memcpy(vector + PITCH_MAX, in, sizeof(int16_t) * FRAME_LEN);
2136 scale_vector(vector, FRAME_LEN + PITCH_MAX);
2138 p->pitch_lag[0] = estimate_pitch(vector, PITCH_MAX);
2139 p->pitch_lag[1] = estimate_pitch(vector, PITCH_MAX + HALF_FRAME_LEN);
2141 for (i = PITCH_MAX, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
2142 comp_harmonic_coeff(vector + i, p->pitch_lag[j >> 1], hf + j);
2144 memcpy(vector, p->prev_weight_sig, sizeof(int16_t) * PITCH_MAX);
2145 memcpy(vector + PITCH_MAX, in, sizeof(int16_t) * FRAME_LEN);
2146 memcpy(p->prev_weight_sig, vector + FRAME_LEN, sizeof(int16_t) * PITCH_MAX);
2148 for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
2149 harmonic_filter(hf + j, vector + PITCH_MAX + i, in + i);
2151 inverse_quant(cur_lsp, p->prev_lsp, p->lsp_index, 0);
2152 lsp_interpolate(qnt_lpc, cur_lsp, p->prev_lsp);
2154 memcpy(p->prev_lsp, cur_lsp, sizeof(int16_t) * LPC_ORDER);
2157 for (i = 0; i < SUBFRAMES; i++) {
2158 int16_t impulse_resp[SUBFRAME_LEN];
2159 int16_t residual[SUBFRAME_LEN + PITCH_ORDER - 1];
2160 int16_t flt_in[SUBFRAME_LEN];
2161 int16_t zero[LPC_ORDER], fir[LPC_ORDER], iir[LPC_ORDER];
2164 * Compute the combined impulse response of the synthesis filter,
2165 * formant perceptual weighting filter and harmonic noise shaping filter
2167 memset(zero, 0, sizeof(int16_t) * LPC_ORDER);
2168 memset(vector, 0, sizeof(int16_t) * PITCH_MAX);
2169 memset(flt_in, 0, sizeof(int16_t) * SUBFRAME_LEN);
2171 flt_in[0] = 1 << 13; /* Unit impulse */
2172 synth_percept_filter(qnt_lpc + offset, weighted_lpc + (offset << 1),
2173 zero, zero, flt_in, vector + PITCH_MAX, 1);
2174 harmonic_filter(hf + i, vector + PITCH_MAX, impulse_resp);
2176 /* Compute the combined zero input response */
2178 memcpy(fir, p->perf_fir_mem, sizeof(int16_t) * LPC_ORDER);
2179 memcpy(iir, p->perf_iir_mem, sizeof(int16_t) * LPC_ORDER);
2181 synth_percept_filter(qnt_lpc + offset, weighted_lpc + (offset << 1),
2182 fir, iir, flt_in, vector + PITCH_MAX, 0);
2183 memcpy(vector, p->harmonic_mem, sizeof(int16_t) * PITCH_MAX);
2184 harmonic_noise_sub(hf + i, vector + PITCH_MAX, in);
2186 acb_search(p, residual, impulse_resp, in, i);
2187 gen_acb_excitation(residual, p->prev_excitation,p->pitch_lag[i >> 1],
2188 p->subframe[i], p->cur_rate);
2189 sub_acb_contrib(residual, impulse_resp, in);
2191 fcb_search(p, impulse_resp, in, i);
2193 /* Reconstruct the excitation */
2194 gen_acb_excitation(impulse_resp, p->prev_excitation, p->pitch_lag[i >> 1],
2195 p->subframe[i], Rate6k3);
2197 memmove(p->prev_excitation, p->prev_excitation + SUBFRAME_LEN,
2198 sizeof(int16_t) * (PITCH_MAX - SUBFRAME_LEN));
2199 for (j = 0; j < SUBFRAME_LEN; j++)
2200 in[j] = av_clip_int16((in[j] << 1) + impulse_resp[j]);
2201 memcpy(p->prev_excitation + PITCH_MAX - SUBFRAME_LEN, in,
2202 sizeof(int16_t) * SUBFRAME_LEN);
2204 /* Update filter memories */
2205 synth_percept_filter(qnt_lpc + offset, weighted_lpc + (offset << 1),
2206 p->perf_fir_mem, p->perf_iir_mem,
2207 in, vector + PITCH_MAX, 0);
2208 memmove(p->harmonic_mem, p->harmonic_mem + SUBFRAME_LEN,
2209 sizeof(int16_t) * (PITCH_MAX - SUBFRAME_LEN));
2210 memcpy(p->harmonic_mem + PITCH_MAX - SUBFRAME_LEN, vector + PITCH_MAX,
2211 sizeof(int16_t) * SUBFRAME_LEN);
2214 offset += LPC_ORDER;
2217 if ((ret = ff_alloc_packet2(avctx, avpkt, 24)))
2220 *got_packet_ptr = 1;
2221 avpkt->size = pack_bitstream(p, avpkt->data, avpkt->size);
2225 AVCodec ff_g723_1_encoder = {
2227 .type = AVMEDIA_TYPE_AUDIO,
2228 .id = CODEC_ID_G723_1,
2229 .priv_data_size = sizeof(G723_1_Context),
2230 .init = g723_1_encode_init,
2231 .encode2 = g723_1_encode_frame,
2232 .long_name = NULL_IF_CONFIG_SMALL("G.723.1"),
2233 .sample_fmts = (const enum AVSampleFormat[]){AV_SAMPLE_FMT_S16,
2234 AV_SAMPLE_FMT_NONE},