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
31 #include "acelp_vectors.h"
32 #include "celp_filters.h"
33 #include "celp_math.h"
35 #include "libavutil/lzo.h"
36 #include "g723_1_data.h"
38 typedef struct g723_1_context {
40 G723_1_Subframe subframe[4];
41 FrameType cur_frame_type;
42 FrameType past_frame_type;
44 uint8_t lsp_index[LSP_BANDS];
48 int16_t prev_lsp[LPC_ORDER];
49 int16_t prev_excitation[PITCH_MAX];
50 int16_t excitation[PITCH_MAX + FRAME_LEN];
51 int16_t synth_mem[LPC_ORDER];
52 int16_t fir_mem[LPC_ORDER];
53 int iir_mem[LPC_ORDER];
61 int pf_gain; ///< formant postfilter
62 ///< gain scaling unit memory
64 int16_t prev_data[HALF_FRAME_LEN];
65 int16_t prev_weight_sig[PITCH_MAX];
68 int16_t hpf_fir_mem; ///< highpass filter fir
69 int hpf_iir_mem; ///< and iir memories
70 int16_t perf_fir_mem[LPC_ORDER]; ///< perceptual filter fir
71 int16_t perf_iir_mem[LPC_ORDER]; ///< and iir memories
73 int16_t harmonic_mem[PITCH_MAX];
76 static av_cold int g723_1_decode_init(AVCodecContext *avctx)
78 G723_1_Context *p = avctx->priv_data;
80 avctx->sample_fmt = AV_SAMPLE_FMT_S16;
82 memcpy(p->prev_lsp, dc_lsp, LPC_ORDER * sizeof(int16_t));
84 avcodec_get_frame_defaults(&p->frame);
85 avctx->coded_frame = &p->frame;
91 * Unpack the frame into parameters.
93 * @param p the context
94 * @param buf pointer to the input buffer
95 * @param buf_size size of the input buffer
97 static int unpack_bitstream(G723_1_Context *p, const uint8_t *buf,
102 int temp, info_bits, i;
104 init_get_bits(&gb, buf, buf_size * 8);
106 /* Extract frame type and rate info */
107 info_bits = get_bits(&gb, 2);
109 if (info_bits == 3) {
110 p->cur_frame_type = UntransmittedFrame;
114 /* Extract 24 bit lsp indices, 8 bit for each band */
115 p->lsp_index[2] = get_bits(&gb, 8);
116 p->lsp_index[1] = get_bits(&gb, 8);
117 p->lsp_index[0] = get_bits(&gb, 8);
119 if (info_bits == 2) {
120 p->cur_frame_type = SIDFrame;
121 p->subframe[0].amp_index = get_bits(&gb, 6);
125 /* Extract the info common to both rates */
126 p->cur_rate = info_bits ? Rate5k3 : Rate6k3;
127 p->cur_frame_type = ActiveFrame;
129 p->pitch_lag[0] = get_bits(&gb, 7);
130 if (p->pitch_lag[0] > 123) /* test if forbidden code */
132 p->pitch_lag[0] += PITCH_MIN;
133 p->subframe[1].ad_cb_lag = get_bits(&gb, 2);
135 p->pitch_lag[1] = get_bits(&gb, 7);
136 if (p->pitch_lag[1] > 123)
138 p->pitch_lag[1] += PITCH_MIN;
139 p->subframe[3].ad_cb_lag = get_bits(&gb, 2);
140 p->subframe[0].ad_cb_lag = 1;
141 p->subframe[2].ad_cb_lag = 1;
143 for (i = 0; i < SUBFRAMES; i++) {
144 /* Extract combined gain */
145 temp = get_bits(&gb, 12);
147 p->subframe[i].dirac_train = 0;
148 if (p->cur_rate == Rate6k3 && p->pitch_lag[i >> 1] < SUBFRAME_LEN - 2) {
149 p->subframe[i].dirac_train = temp >> 11;
153 p->subframe[i].ad_cb_gain = FASTDIV(temp, GAIN_LEVELS);
154 if (p->subframe[i].ad_cb_gain < ad_cb_len) {
155 p->subframe[i].amp_index = temp - p->subframe[i].ad_cb_gain *
162 p->subframe[0].grid_index = get_bits1(&gb);
163 p->subframe[1].grid_index = get_bits1(&gb);
164 p->subframe[2].grid_index = get_bits1(&gb);
165 p->subframe[3].grid_index = get_bits1(&gb);
167 if (p->cur_rate == Rate6k3) {
168 skip_bits1(&gb); /* skip reserved bit */
170 /* Compute pulse_pos index using the 13-bit combined position index */
171 temp = get_bits(&gb, 13);
172 p->subframe[0].pulse_pos = temp / 810;
174 temp -= p->subframe[0].pulse_pos * 810;
175 p->subframe[1].pulse_pos = FASTDIV(temp, 90);
177 temp -= p->subframe[1].pulse_pos * 90;
178 p->subframe[2].pulse_pos = FASTDIV(temp, 9);
179 p->subframe[3].pulse_pos = temp - p->subframe[2].pulse_pos * 9;
181 p->subframe[0].pulse_pos = (p->subframe[0].pulse_pos << 16) +
183 p->subframe[1].pulse_pos = (p->subframe[1].pulse_pos << 14) +
185 p->subframe[2].pulse_pos = (p->subframe[2].pulse_pos << 16) +
187 p->subframe[3].pulse_pos = (p->subframe[3].pulse_pos << 14) +
190 p->subframe[0].pulse_sign = get_bits(&gb, 6);
191 p->subframe[1].pulse_sign = get_bits(&gb, 5);
192 p->subframe[2].pulse_sign = get_bits(&gb, 6);
193 p->subframe[3].pulse_sign = get_bits(&gb, 5);
194 } else { /* Rate5k3 */
195 p->subframe[0].pulse_pos = get_bits(&gb, 12);
196 p->subframe[1].pulse_pos = get_bits(&gb, 12);
197 p->subframe[2].pulse_pos = get_bits(&gb, 12);
198 p->subframe[3].pulse_pos = get_bits(&gb, 12);
200 p->subframe[0].pulse_sign = get_bits(&gb, 4);
201 p->subframe[1].pulse_sign = get_bits(&gb, 4);
202 p->subframe[2].pulse_sign = get_bits(&gb, 4);
203 p->subframe[3].pulse_sign = get_bits(&gb, 4);
210 * Bitexact implementation of sqrt(val/2).
212 static int16_t square_root(int val)
214 return (ff_sqrt(val << 1) >> 1) & (~1);
218 * Calculate the number of left-shifts required for normalizing the input.
220 * @param num input number
221 * @param width width of the input, 16 bits(0) / 32 bits(1)
223 static int normalize_bits(int num, int width)
226 int bits = (width) ? 31 : 15;
233 i= bits - av_log2(num) - 1;
239 #define normalize_bits_int16(num) normalize_bits(num, 0)
240 #define normalize_bits_int32(num) normalize_bits(num, 1)
241 #define dot_product(a,b,c,d) (ff_dot_product(a,b,c)<<(d))
244 * Scale vector contents based on the largest of their absolutes.
246 static int scale_vector(int16_t *vector, int length)
248 int bits, scale, max = 0;
251 const int16_t shift_table[16] = {
252 0x0001, 0x0002, 0x0004, 0x0008, 0x0010, 0x0020, 0x0040, 0x0080,
253 0x0100, 0x0200, 0x0400, 0x0800, 0x1000, 0x2000, 0x4000, 0x7fff
256 for (i = 0; i < length; i++)
257 max = FFMAX(max, FFABS(vector[i]));
259 bits = normalize_bits(max, 0);
260 scale = shift_table[bits];
262 for (i = 0; i < length; i++)
263 vector[i] = (vector[i] * scale) >> 3;
269 * Perform inverse quantization of LSP frequencies.
271 * @param cur_lsp the current LSP vector
272 * @param prev_lsp the previous LSP vector
273 * @param lsp_index VQ indices
274 * @param bad_frame bad frame flag
276 static void inverse_quant(int16_t *cur_lsp, int16_t *prev_lsp,
277 uint8_t *lsp_index, int bad_frame)
280 int i, j, temp, stable;
282 /* Check for frame erasure */
289 lsp_index[0] = lsp_index[1] = lsp_index[2] = 0;
292 /* Get the VQ table entry corresponding to the transmitted index */
293 cur_lsp[0] = lsp_band0[lsp_index[0]][0];
294 cur_lsp[1] = lsp_band0[lsp_index[0]][1];
295 cur_lsp[2] = lsp_band0[lsp_index[0]][2];
296 cur_lsp[3] = lsp_band1[lsp_index[1]][0];
297 cur_lsp[4] = lsp_band1[lsp_index[1]][1];
298 cur_lsp[5] = lsp_band1[lsp_index[1]][2];
299 cur_lsp[6] = lsp_band2[lsp_index[2]][0];
300 cur_lsp[7] = lsp_band2[lsp_index[2]][1];
301 cur_lsp[8] = lsp_band2[lsp_index[2]][2];
302 cur_lsp[9] = lsp_band2[lsp_index[2]][3];
304 /* Add predicted vector & DC component to the previously quantized vector */
305 for (i = 0; i < LPC_ORDER; i++) {
306 temp = ((prev_lsp[i] - dc_lsp[i]) * pred + (1 << 14)) >> 15;
307 cur_lsp[i] += dc_lsp[i] + temp;
310 for (i = 0; i < LPC_ORDER; i++) {
311 cur_lsp[0] = FFMAX(cur_lsp[0], 0x180);
312 cur_lsp[LPC_ORDER - 1] = FFMIN(cur_lsp[LPC_ORDER - 1], 0x7e00);
314 /* Stability check */
315 for (j = 1; j < LPC_ORDER; j++) {
316 temp = min_dist + cur_lsp[j - 1] - cur_lsp[j];
319 cur_lsp[j - 1] -= temp;
324 for (j = 1; j < LPC_ORDER; j++) {
325 temp = cur_lsp[j - 1] + min_dist - cur_lsp[j] - 4;
335 memcpy(cur_lsp, prev_lsp, LPC_ORDER * sizeof(int16_t));
339 * Bitexact implementation of 2ab scaled by 1/2^16.
341 * @param a 32 bit multiplicand
342 * @param b 16 bit multiplier
344 #define MULL2(a, b) \
348 * Convert LSP frequencies to LPC coefficients.
350 * @param lpc buffer for LPC coefficients
352 static void lsp2lpc(int16_t *lpc)
354 int f1[LPC_ORDER / 2 + 1];
355 int f2[LPC_ORDER / 2 + 1];
358 /* Calculate negative cosine */
359 for (j = 0; j < LPC_ORDER; j++) {
360 int index = lpc[j] >> 7;
361 int offset = lpc[j] & 0x7f;
362 int64_t temp1 = cos_tab[index] << 16;
363 int temp2 = (cos_tab[index + 1] - cos_tab[index]) *
364 ((offset << 8) + 0x80) << 1;
366 lpc[j] = -(av_clipl_int32(((temp1 + temp2) << 1) + (1 << 15)) >> 16);
370 * Compute sum and difference polynomial coefficients
371 * (bitexact alternative to lsp2poly() in lsp.c)
373 /* Initialize with values in Q28 */
375 f1[1] = (lpc[0] << 14) + (lpc[2] << 14);
376 f1[2] = lpc[0] * lpc[2] + (2 << 28);
379 f2[1] = (lpc[1] << 14) + (lpc[3] << 14);
380 f2[2] = lpc[1] * lpc[3] + (2 << 28);
383 * Calculate and scale the coefficients by 1/2 in
384 * each iteration for a final scaling factor of Q25
386 for (i = 2; i < LPC_ORDER / 2; i++) {
387 f1[i + 1] = f1[i - 1] + MULL2(f1[i], lpc[2 * i]);
388 f2[i + 1] = f2[i - 1] + MULL2(f2[i], lpc[2 * i + 1]);
390 for (j = i; j >= 2; j--) {
391 f1[j] = MULL2(f1[j - 1], lpc[2 * i]) +
392 (f1[j] >> 1) + (f1[j - 2] >> 1);
393 f2[j] = MULL2(f2[j - 1], lpc[2 * i + 1]) +
394 (f2[j] >> 1) + (f2[j - 2] >> 1);
399 f1[1] = ((lpc[2 * i] << 16 >> i) + f1[1]) >> 1;
400 f2[1] = ((lpc[2 * i + 1] << 16 >> i) + f2[1]) >> 1;
403 /* Convert polynomial coefficients to LPC coefficients */
404 for (i = 0; i < LPC_ORDER / 2; i++) {
405 int64_t ff1 = f1[i + 1] + f1[i];
406 int64_t ff2 = f2[i + 1] - f2[i];
408 lpc[i] = av_clipl_int32(((ff1 + ff2) << 3) + (1 << 15)) >> 16;
409 lpc[LPC_ORDER - i - 1] = av_clipl_int32(((ff1 - ff2) << 3) +
415 * Quantize LSP frequencies by interpolation and convert them to
416 * the corresponding LPC coefficients.
418 * @param lpc buffer for LPC coefficients
419 * @param cur_lsp the current LSP vector
420 * @param prev_lsp the previous LSP vector
422 static void lsp_interpolate(int16_t *lpc, int16_t *cur_lsp, int16_t *prev_lsp)
425 int16_t *lpc_ptr = lpc;
427 /* cur_lsp * 0.25 + prev_lsp * 0.75 */
428 ff_acelp_weighted_vector_sum(lpc, cur_lsp, prev_lsp,
429 4096, 12288, 1 << 13, 14, LPC_ORDER);
430 ff_acelp_weighted_vector_sum(lpc + LPC_ORDER, cur_lsp, prev_lsp,
431 8192, 8192, 1 << 13, 14, LPC_ORDER);
432 ff_acelp_weighted_vector_sum(lpc + 2 * LPC_ORDER, cur_lsp, prev_lsp,
433 12288, 4096, 1 << 13, 14, LPC_ORDER);
434 memcpy(lpc + 3 * LPC_ORDER, cur_lsp, LPC_ORDER * sizeof(int16_t));
436 for (i = 0; i < SUBFRAMES; i++) {
438 lpc_ptr += LPC_ORDER;
443 * Generate a train of dirac functions with period as pitch lag.
445 static void gen_dirac_train(int16_t *buf, int pitch_lag)
447 int16_t vector[SUBFRAME_LEN];
450 memcpy(vector, buf, SUBFRAME_LEN * sizeof(int16_t));
451 for (i = pitch_lag; i < SUBFRAME_LEN; i += pitch_lag) {
452 for (j = 0; j < SUBFRAME_LEN - i; j++)
453 buf[i + j] += vector[j];
458 * Generate fixed codebook excitation vector.
460 * @param vector decoded excitation vector
461 * @param subfrm current subframe
462 * @param cur_rate current bitrate
463 * @param pitch_lag closed loop pitch lag
464 * @param index current subframe index
466 static void gen_fcb_excitation(int16_t *vector, G723_1_Subframe subfrm,
467 Rate cur_rate, int pitch_lag, int index)
471 memset(vector, 0, SUBFRAME_LEN * sizeof(int16_t));
473 if (cur_rate == Rate6k3) {
474 if (subfrm.pulse_pos >= max_pos[index])
477 /* Decode amplitudes and positions */
478 j = PULSE_MAX - pulses[index];
479 temp = subfrm.pulse_pos;
480 for (i = 0; i < SUBFRAME_LEN / GRID_SIZE; i++) {
481 temp -= combinatorial_table[j][i];
484 temp += combinatorial_table[j++][i];
485 if (subfrm.pulse_sign & (1 << (PULSE_MAX - j))) {
486 vector[subfrm.grid_index + GRID_SIZE * i] =
487 -fixed_cb_gain[subfrm.amp_index];
489 vector[subfrm.grid_index + GRID_SIZE * i] =
490 fixed_cb_gain[subfrm.amp_index];
495 if (subfrm.dirac_train == 1)
496 gen_dirac_train(vector, pitch_lag);
497 } else { /* Rate5k3 */
498 int cb_gain = fixed_cb_gain[subfrm.amp_index];
499 int cb_shift = subfrm.grid_index;
500 int cb_sign = subfrm.pulse_sign;
501 int cb_pos = subfrm.pulse_pos;
502 int offset, beta, lag;
504 for (i = 0; i < 8; i += 2) {
505 offset = ((cb_pos & 7) << 3) + cb_shift + i;
506 vector[offset] = (cb_sign & 1) ? cb_gain : -cb_gain;
511 /* Enhance harmonic components */
512 lag = pitch_contrib[subfrm.ad_cb_gain << 1] + pitch_lag +
513 subfrm.ad_cb_lag - 1;
514 beta = pitch_contrib[(subfrm.ad_cb_gain << 1) + 1];
516 if (lag < SUBFRAME_LEN - 2) {
517 for (i = lag; i < SUBFRAME_LEN; i++)
518 vector[i] += beta * vector[i - lag] >> 15;
524 * Get delayed contribution from the previous excitation vector.
526 static void get_residual(int16_t *residual, int16_t *prev_excitation, int lag)
528 int offset = PITCH_MAX - PITCH_ORDER / 2 - lag;
531 residual[0] = prev_excitation[offset];
532 residual[1] = prev_excitation[offset + 1];
535 for (i = 2; i < SUBFRAME_LEN + PITCH_ORDER - 1; i++)
536 residual[i] = prev_excitation[offset + (i - 2) % lag];
540 * Generate adaptive codebook excitation.
542 static void gen_acb_excitation(int16_t *vector, int16_t *prev_excitation,
543 int pitch_lag, G723_1_Subframe subfrm,
546 int16_t residual[SUBFRAME_LEN + PITCH_ORDER - 1];
547 const int16_t *cb_ptr;
548 int lag = pitch_lag + subfrm.ad_cb_lag - 1;
553 get_residual(residual, prev_excitation, lag);
555 /* Select quantization table */
556 if (cur_rate == Rate6k3 && pitch_lag < SUBFRAME_LEN - 2) {
557 cb_ptr = adaptive_cb_gain85;
559 cb_ptr = adaptive_cb_gain170;
561 /* Calculate adaptive vector */
562 cb_ptr += subfrm.ad_cb_gain * 20;
563 for (i = 0; i < SUBFRAME_LEN; i++) {
564 sum = ff_dot_product(residual + i, cb_ptr, PITCH_ORDER);
565 vector[i] = av_clipl_int32((sum << 2) + (1 << 15)) >> 16;
570 * Estimate maximum auto-correlation around pitch lag.
572 * @param p the context
573 * @param offset offset of the excitation vector
574 * @param ccr_max pointer to the maximum auto-correlation
575 * @param pitch_lag decoded pitch lag
576 * @param length length of autocorrelation
577 * @param dir forward lag(1) / backward lag(-1)
579 static int autocorr_max(G723_1_Context *p, int offset, int *ccr_max,
580 int pitch_lag, int length, int dir)
582 int limit, ccr, lag = 0;
583 int16_t *buf = p->excitation + offset;
586 pitch_lag = FFMIN(PITCH_MAX - 3, pitch_lag);
587 limit = FFMIN(FRAME_LEN + PITCH_MAX - offset - length, pitch_lag + 3);
589 for (i = pitch_lag - 3; i <= limit; i++) {
590 ccr = ff_dot_product(buf, buf + dir * i, length)<<1;
592 if (ccr > *ccr_max) {
601 * Calculate pitch postfilter optimal and scaling gains.
603 * @param lag pitch postfilter forward/backward lag
604 * @param ppf pitch postfilter parameters
605 * @param cur_rate current bitrate
606 * @param tgt_eng target energy
607 * @param ccr cross-correlation
608 * @param res_eng residual energy
610 static void comp_ppf_gains(int lag, PPFParam *ppf, Rate cur_rate,
611 int tgt_eng, int ccr, int res_eng)
613 int pf_residual; /* square of postfiltered residual */
614 int64_t temp1, temp2;
618 temp1 = tgt_eng * res_eng >> 1;
619 temp2 = ccr * ccr << 1;
622 if (ccr >= res_eng) {
623 ppf->opt_gain = ppf_gain_weight[cur_rate];
625 ppf->opt_gain = (ccr << 15) / res_eng *
626 ppf_gain_weight[cur_rate] >> 15;
628 /* pf_res^2 = tgt_eng + 2*ccr*gain + res_eng*gain^2 */
629 temp1 = (tgt_eng << 15) + (ccr * ppf->opt_gain << 1);
630 temp2 = (ppf->opt_gain * ppf->opt_gain >> 15) * res_eng;
631 pf_residual = av_clipl_int32(temp1 + temp2 + (1 << 15)) >> 16;
633 if (tgt_eng >= pf_residual << 1) {
636 temp1 = (tgt_eng << 14) / pf_residual;
639 /* scaling_gain = sqrt(tgt_eng/pf_res^2) */
640 ppf->sc_gain = square_root(temp1 << 16);
643 ppf->sc_gain = 0x7fff;
646 ppf->opt_gain = av_clip_int16(ppf->opt_gain * ppf->sc_gain >> 15);
650 * Calculate pitch postfilter parameters.
652 * @param p the context
653 * @param offset offset of the excitation vector
654 * @param pitch_lag decoded pitch lag
655 * @param ppf pitch postfilter parameters
656 * @param cur_rate current bitrate
658 static void comp_ppf_coeff(G723_1_Context *p, int offset, int pitch_lag,
659 PPFParam *ppf, Rate cur_rate)
664 int64_t temp1, temp2;
668 * 1 - forward cross-correlation
669 * 2 - forward residual energy
670 * 3 - backward cross-correlation
671 * 4 - backward residual energy
673 int energy[5] = {0, 0, 0, 0, 0};
674 int16_t *buf = p->excitation + offset;
675 int fwd_lag = autocorr_max(p, offset, &energy[1], pitch_lag,
677 int back_lag = autocorr_max(p, offset, &energy[3], pitch_lag,
682 ppf->sc_gain = 0x7fff;
684 /* Case 0, Section 3.6 */
685 if (!back_lag && !fwd_lag)
688 /* Compute target energy */
689 energy[0] = ff_dot_product(buf, buf, SUBFRAME_LEN)<<1;
691 /* Compute forward residual energy */
693 energy[2] = ff_dot_product(buf + fwd_lag, buf + fwd_lag,
696 /* Compute backward residual energy */
698 energy[4] = ff_dot_product(buf - back_lag, buf - back_lag,
701 /* Normalize and shorten */
703 for (i = 0; i < 5; i++)
704 temp1 = FFMAX(energy[i], temp1);
706 scale = normalize_bits(temp1, 1);
707 for (i = 0; i < 5; i++)
708 energy[i] = av_clipl_int32(energy[i] << scale) >> 16;
710 if (fwd_lag && !back_lag) { /* Case 1 */
711 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
713 } else if (!fwd_lag) { /* Case 2 */
714 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
716 } else { /* Case 3 */
719 * Select the largest of energy[1]^2/energy[2]
720 * and energy[3]^2/energy[4]
722 temp1 = energy[4] * ((energy[1] * energy[1] + (1 << 14)) >> 15);
723 temp2 = energy[2] * ((energy[3] * energy[3] + (1 << 14)) >> 15);
724 if (temp1 >= temp2) {
725 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
728 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
735 * Classify frames as voiced/unvoiced.
737 * @param p the context
738 * @param pitch_lag decoded pitch_lag
739 * @param exc_eng excitation energy estimation
740 * @param scale scaling factor of exc_eng
742 * @return residual interpolation index if voiced, 0 otherwise
744 static int comp_interp_index(G723_1_Context *p, int pitch_lag,
745 int *exc_eng, int *scale)
747 int offset = PITCH_MAX + 2 * SUBFRAME_LEN;
748 int16_t *buf = p->excitation + offset;
750 int index, ccr, tgt_eng, best_eng, temp;
752 *scale = scale_vector(p->excitation, FRAME_LEN + PITCH_MAX);
754 /* Compute maximum backward cross-correlation */
756 index = autocorr_max(p, offset, &ccr, pitch_lag, SUBFRAME_LEN * 2, -1);
757 ccr = av_clipl_int32((int64_t)ccr + (1 << 15)) >> 16;
759 /* Compute target energy */
760 tgt_eng = ff_dot_product(buf, buf, SUBFRAME_LEN * 2)<<1;
761 *exc_eng = av_clipl_int32(tgt_eng + (1 << 15)) >> 16;
766 /* Compute best energy */
767 best_eng = ff_dot_product(buf - index, buf - index,
768 SUBFRAME_LEN * 2)<<1;
769 best_eng = av_clipl_int32((int64_t)best_eng + (1 << 15)) >> 16;
771 temp = best_eng * *exc_eng >> 3;
773 if (temp < ccr * ccr) {
780 * Peform residual interpolation based on frame classification.
782 * @param buf decoded excitation vector
783 * @param out output vector
784 * @param lag decoded pitch lag
785 * @param gain interpolated gain
786 * @param rseed seed for random number generator
788 static void residual_interp(int16_t *buf, int16_t *out, int lag,
789 int gain, int *rseed)
792 if (lag) { /* Voiced */
793 int16_t *vector_ptr = buf + PITCH_MAX;
795 for (i = 0; i < lag; i++)
796 vector_ptr[i - lag] = vector_ptr[i - lag] * 3 >> 2;
797 av_memcpy_backptr((uint8_t*)vector_ptr, lag * sizeof(int16_t),
798 FRAME_LEN * sizeof(int16_t));
799 memcpy(out, vector_ptr, FRAME_LEN * sizeof(int16_t));
800 } else { /* Unvoiced */
801 for (i = 0; i < FRAME_LEN; i++) {
802 *rseed = *rseed * 521 + 259;
803 out[i] = gain * *rseed >> 15;
805 memset(buf, 0, (FRAME_LEN + PITCH_MAX) * sizeof(int16_t));
810 * Perform IIR filtering.
812 * @param fir_coef FIR coefficients
813 * @param iir_coef IIR coefficients
814 * @param src source vector
815 * @param dest destination vector
816 * @param width width of the output, 16 bits(0) / 32 bits(1)
818 #define iir_filter(fir_coef, iir_coef, src, dest, width)\
821 int res_shift = 16 & ~-(width);\
822 int in_shift = 16 - res_shift;\
824 for (m = 0; m < SUBFRAME_LEN; m++) {\
826 for (n = 1; n <= LPC_ORDER; n++) {\
827 filter -= (fir_coef)[n - 1] * (src)[m - n] -\
828 (iir_coef)[n - 1] * ((dest)[m - n] >> in_shift);\
831 (dest)[m] = av_clipl_int32(((src)[m] << 16) + (filter << 3) +\
832 (1 << 15)) >> res_shift;\
837 * Adjust gain of postfiltered signal.
839 * @param p the context
840 * @param buf postfiltered output vector
841 * @param energy input energy coefficient
843 static void gain_scale(G723_1_Context *p, int16_t * buf, int energy)
845 int num, denom, gain, bits1, bits2;
850 for (i = 0; i < SUBFRAME_LEN; i++) {
851 int64_t temp = buf[i] >> 2;
852 temp = av_clipl_int32(MUL64(temp, temp) << 1);
853 denom = av_clipl_int32(denom + temp);
857 bits1 = normalize_bits(num, 1);
858 bits2 = normalize_bits(denom, 1);
859 num = num << bits1 >> 1;
862 bits2 = 5 + bits1 - bits2;
863 bits2 = FFMAX(0, bits2);
865 gain = (num >> 1) / (denom >> 16);
866 gain = square_root(gain << 16 >> bits2);
871 for (i = 0; i < SUBFRAME_LEN; i++) {
872 p->pf_gain = ((p->pf_gain << 4) - p->pf_gain + gain + (1 << 3)) >> 4;
873 buf[i] = av_clip_int16((buf[i] * (p->pf_gain + (p->pf_gain >> 4)) +
879 * Perform formant filtering.
881 * @param p the context
882 * @param lpc quantized lpc coefficients
883 * @param buf output buffer
885 static void formant_postfilter(G723_1_Context *p, int16_t *lpc, int16_t *buf)
887 int16_t filter_coef[2][LPC_ORDER], *buf_ptr;
888 int filter_signal[LPC_ORDER + FRAME_LEN], *signal_ptr;
891 memcpy(buf, p->fir_mem, LPC_ORDER * sizeof(int16_t));
892 memcpy(filter_signal, p->iir_mem, LPC_ORDER * sizeof(int));
894 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
895 for (k = 0; k < LPC_ORDER; k++) {
896 filter_coef[0][k] = (-lpc[k] * postfilter_tbl[0][k] +
898 filter_coef[1][k] = (-lpc[k] * postfilter_tbl[1][k] +
901 iir_filter(filter_coef[0], filter_coef[1], buf + i,
902 filter_signal + i, 1);
905 memcpy(p->fir_mem, buf + FRAME_LEN, LPC_ORDER * sizeof(int16_t));
906 memcpy(p->iir_mem, filter_signal + FRAME_LEN, LPC_ORDER * sizeof(int));
908 buf_ptr = buf + LPC_ORDER;
909 signal_ptr = filter_signal + LPC_ORDER;
910 for (i = 0; i < SUBFRAMES; i++) {
911 int16_t temp_vector[SUBFRAME_LEN];
917 memcpy(temp_vector, buf_ptr, SUBFRAME_LEN * sizeof(int16_t));
918 scale = scale_vector(temp_vector, SUBFRAME_LEN);
920 /* Compute auto correlation coefficients */
921 auto_corr[0] = ff_dot_product(temp_vector, temp_vector + 1,
922 SUBFRAME_LEN - 1)<<1;
923 auto_corr[1] = ff_dot_product(temp_vector, temp_vector,
926 /* Compute reflection coefficient */
927 temp = auto_corr[1] >> 16;
929 temp = (auto_corr[0] >> 2) / temp;
931 p->reflection_coef = ((p->reflection_coef << 2) - p->reflection_coef +
933 temp = (p->reflection_coef * 0xffffc >> 3) & 0xfffc;
935 /* Compensation filter */
936 for (j = 0; j < SUBFRAME_LEN; j++) {
937 buf_ptr[j] = av_clipl_int32(signal_ptr[j] +
938 ((signal_ptr[j - 1] >> 16) *
942 /* Compute normalized signal energy */
943 temp = 2 * scale + 4;
945 energy = av_clipl_int32((int64_t)auto_corr[1] << -temp);
947 energy = auto_corr[1] >> temp;
949 gain_scale(p, buf_ptr, energy);
951 buf_ptr += SUBFRAME_LEN;
952 signal_ptr += SUBFRAME_LEN;
956 static int g723_1_decode_frame(AVCodecContext *avctx, void *data,
957 int *got_frame_ptr, AVPacket *avpkt)
959 G723_1_Context *p = avctx->priv_data;
960 const uint8_t *buf = avpkt->data;
961 int buf_size = avpkt->size;
963 int dec_mode = buf[0] & 3;
965 PPFParam ppf[SUBFRAMES];
966 int16_t cur_lsp[LPC_ORDER];
967 int16_t lpc[SUBFRAMES * LPC_ORDER];
968 int16_t acb_vector[SUBFRAME_LEN];
970 int bad_frame = 0, i, j, ret;
972 if (!buf_size || buf_size < frame_size[dec_mode]) {
977 if (unpack_bitstream(p, buf, buf_size) < 0) {
979 p->cur_frame_type = p->past_frame_type == ActiveFrame ?
980 ActiveFrame : UntransmittedFrame;
983 p->frame.nb_samples = FRAME_LEN + LPC_ORDER;
984 if ((ret = avctx->get_buffer(avctx, &p->frame)) < 0) {
985 av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
988 out= (int16_t*)p->frame.data[0];
991 if(p->cur_frame_type == ActiveFrame) {
993 p->erased_frames = 0;
994 } else if(p->erased_frames != 3)
997 inverse_quant(cur_lsp, p->prev_lsp, p->lsp_index, bad_frame);
998 lsp_interpolate(lpc, cur_lsp, p->prev_lsp);
1000 /* Save the lsp_vector for the next frame */
1001 memcpy(p->prev_lsp, cur_lsp, LPC_ORDER * sizeof(int16_t));
1003 /* Generate the excitation for the frame */
1004 memcpy(p->excitation, p->prev_excitation, PITCH_MAX * sizeof(int16_t));
1005 vector_ptr = p->excitation + PITCH_MAX;
1006 if (!p->erased_frames) {
1007 /* Update interpolation gain memory */
1008 p->interp_gain = fixed_cb_gain[(p->subframe[2].amp_index +
1009 p->subframe[3].amp_index) >> 1];
1010 for (i = 0; i < SUBFRAMES; i++) {
1011 gen_fcb_excitation(vector_ptr, p->subframe[i], p->cur_rate,
1012 p->pitch_lag[i >> 1], i);
1013 gen_acb_excitation(acb_vector, &p->excitation[SUBFRAME_LEN * i],
1014 p->pitch_lag[i >> 1], p->subframe[i],
1016 /* Get the total excitation */
1017 for (j = 0; j < SUBFRAME_LEN; j++) {
1018 vector_ptr[j] = av_clip_int16(vector_ptr[j] << 1);
1019 vector_ptr[j] = av_clip_int16(vector_ptr[j] +
1022 vector_ptr += SUBFRAME_LEN;
1025 vector_ptr = p->excitation + PITCH_MAX;
1027 /* Save the excitation */
1028 memcpy(out, vector_ptr, FRAME_LEN * sizeof(int16_t));
1030 p->interp_index = comp_interp_index(p, p->pitch_lag[1],
1031 &p->sid_gain, &p->cur_gain);
1033 for (i = PITCH_MAX, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1034 comp_ppf_coeff(p, i, p->pitch_lag[j >> 1],
1035 ppf + j, p->cur_rate);
1037 /* Restore the original excitation */
1038 memcpy(p->excitation, p->prev_excitation,
1039 PITCH_MAX * sizeof(int16_t));
1040 memcpy(vector_ptr, out, FRAME_LEN * sizeof(int16_t));
1042 /* Peform pitch postfiltering */
1043 for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1044 ff_acelp_weighted_vector_sum(out + LPC_ORDER + i, vector_ptr + i,
1045 vector_ptr + i + ppf[j].index,
1046 ppf[j].sc_gain, ppf[j].opt_gain,
1047 1 << 14, 15, SUBFRAME_LEN);
1049 p->interp_gain = (p->interp_gain * 3 + 2) >> 2;
1050 if (p->erased_frames == 3) {
1052 memset(p->excitation, 0,
1053 (FRAME_LEN + PITCH_MAX) * sizeof(int16_t));
1054 memset(out, 0, (FRAME_LEN + LPC_ORDER) * sizeof(int16_t));
1056 /* Regenerate frame */
1057 residual_interp(p->excitation, out + LPC_ORDER, p->interp_index,
1058 p->interp_gain, &p->random_seed);
1061 /* Save the excitation for the next frame */
1062 memcpy(p->prev_excitation, p->excitation + FRAME_LEN,
1063 PITCH_MAX * sizeof(int16_t));
1065 memset(out, 0, sizeof(int16_t)*FRAME_LEN);
1066 av_log(avctx, AV_LOG_WARNING,
1067 "G.723.1: Comfort noise generation not supported yet\n");
1068 return frame_size[dec_mode];
1071 p->past_frame_type = p->cur_frame_type;
1073 memcpy(out, p->synth_mem, LPC_ORDER * sizeof(int16_t));
1074 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1075 ff_celp_lp_synthesis_filter(out + i, &lpc[j * LPC_ORDER],
1076 out + i, SUBFRAME_LEN, LPC_ORDER,
1078 memcpy(p->synth_mem, out + FRAME_LEN, LPC_ORDER * sizeof(int16_t));
1080 formant_postfilter(p, lpc, out);
1082 memmove(out, out + LPC_ORDER, sizeof(int16_t)*FRAME_LEN);
1083 p->frame.nb_samples = FRAME_LEN;
1084 *(AVFrame*)data = p->frame;
1087 return frame_size[dec_mode];
1090 AVCodec ff_g723_1_decoder = {
1092 .type = AVMEDIA_TYPE_AUDIO,
1093 .id = CODEC_ID_G723_1,
1094 .priv_data_size = sizeof(G723_1_Context),
1095 .init = g723_1_decode_init,
1096 .decode = g723_1_decode_frame,
1097 .long_name = NULL_IF_CONFIG_SMALL("G.723.1"),
1098 .capabilities = CODEC_CAP_SUBFRAMES,
1101 #if CONFIG_G723_1_ENCODER
1102 #define BITSTREAM_WRITER_LE
1103 #include "put_bits.h"
1105 static av_cold int g723_1_encode_init(AVCodecContext *avctx)
1107 G723_1_Context *p = avctx->priv_data;
1109 if (avctx->sample_rate != 8000) {
1110 av_log(avctx, AV_LOG_ERROR, "Only 8000Hz sample rate supported\n");
1114 if (avctx->channels != 1) {
1115 av_log(avctx, AV_LOG_ERROR, "Only mono supported\n");
1116 return AVERROR(EINVAL);
1119 if (avctx->bit_rate == 6300) {
1120 p->cur_rate = Rate6k3;
1121 } else if (avctx->bit_rate == 5300) {
1122 av_log(avctx, AV_LOG_ERROR, "Bitrate not supported yet, use 6.3k\n");
1123 return AVERROR_PATCHWELCOME;
1125 av_log(avctx, AV_LOG_ERROR,
1126 "Bitrate not supported, use 6.3k\n");
1127 return AVERROR(EINVAL);
1129 avctx->frame_size = 240;
1130 memcpy(p->prev_lsp, dc_lsp, LPC_ORDER * sizeof(int16_t));
1136 * Remove DC component from the input signal.
1138 * @param buf input signal
1139 * @param fir zero memory
1140 * @param iir pole memory
1142 static void highpass_filter(int16_t *buf, int16_t *fir, int *iir)
1145 for (i = 0; i < FRAME_LEN; i++) {
1146 *iir = (buf[i] << 15) + ((-*fir) << 15) + MULL2(*iir, 0x7f00);
1148 buf[i] = av_clipl_int32((int64_t)*iir + (1 << 15)) >> 16;
1153 * Estimate autocorrelation of the input vector.
1155 * @param buf input buffer
1156 * @param autocorr autocorrelation coefficients vector
1158 static void comp_autocorr(int16_t *buf, int16_t *autocorr)
1161 int16_t vector[LPC_FRAME];
1163 memcpy(vector, buf, LPC_FRAME * sizeof(int16_t));
1164 scale_vector(vector, LPC_FRAME);
1166 /* Apply the Hamming window */
1167 for (i = 0; i < LPC_FRAME; i++)
1168 vector[i] = (vector[i] * hamming_window[i] + (1 << 14)) >> 15;
1170 /* Compute the first autocorrelation coefficient */
1171 temp = dot_product(vector, vector, LPC_FRAME, 0);
1173 /* Apply a white noise correlation factor of (1025/1024) */
1177 scale = normalize_bits_int32(temp);
1178 autocorr[0] = av_clipl_int32((int64_t)(temp << scale) +
1181 /* Compute the remaining coefficients */
1183 memset(autocorr + 1, 0, LPC_ORDER * sizeof(int16_t));
1185 for (i = 1; i <= LPC_ORDER; i++) {
1186 temp = dot_product(vector, vector + i, LPC_FRAME - i, 0);
1187 temp = MULL2((temp << scale), binomial_window[i - 1]);
1188 autocorr[i] = av_clipl_int32((int64_t)temp + (1 << 15)) >> 16;
1194 * Use Levinson-Durbin recursion to compute LPC coefficients from
1195 * autocorrelation values.
1197 * @param lpc LPC coefficients vector
1198 * @param autocorr autocorrelation coefficients vector
1199 * @param error prediction error
1201 static void levinson_durbin(int16_t *lpc, int16_t *autocorr, int16_t error)
1203 int16_t vector[LPC_ORDER];
1204 int16_t partial_corr;
1207 memset(lpc, 0, LPC_ORDER * sizeof(int16_t));
1209 for (i = 0; i < LPC_ORDER; i++) {
1210 /* Compute the partial correlation coefficient */
1212 for (j = 0; j < i; j++)
1213 temp -= lpc[j] * autocorr[i - j - 1];
1214 temp = ((autocorr[i] << 13) + temp) << 3;
1216 if (FFABS(temp) >= (error << 16))
1219 partial_corr = temp / (error << 1);
1221 lpc[i] = av_clipl_int32((int64_t)(partial_corr << 14) +
1224 /* Update the prediction error */
1225 temp = MULL2(temp, partial_corr);
1226 error = av_clipl_int32((int64_t)(error << 16) - temp +
1229 memcpy(vector, lpc, i * sizeof(int16_t));
1230 for (j = 0; j < i; j++) {
1231 temp = partial_corr * vector[i - j - 1] << 1;
1232 lpc[j] = av_clipl_int32((int64_t)(lpc[j] << 16) - temp +
1239 * Calculate LPC coefficients for the current frame.
1241 * @param buf current frame
1242 * @param prev_data 2 trailing subframes of the previous frame
1243 * @param lpc LPC coefficients vector
1245 static void comp_lpc_coeff(int16_t *buf, int16_t *lpc)
1247 int16_t autocorr[(LPC_ORDER + 1) * SUBFRAMES];
1248 int16_t *autocorr_ptr = autocorr;
1249 int16_t *lpc_ptr = lpc;
1252 for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
1253 comp_autocorr(buf + i, autocorr_ptr);
1254 levinson_durbin(lpc_ptr, autocorr_ptr + 1, autocorr_ptr[0]);
1256 lpc_ptr += LPC_ORDER;
1257 autocorr_ptr += LPC_ORDER + 1;
1261 static void lpc2lsp(int16_t *lpc, int16_t *prev_lsp, int16_t *lsp)
1263 int f[LPC_ORDER + 2]; ///< coefficients of the sum and difference
1264 ///< polynomials (F1, F2) ordered as
1265 ///< f1[0], f2[0], ...., f1[5], f2[5]
1267 int max, shift, cur_val, prev_val, count, p;
1271 /* Initialize f1[0] and f2[0] to 1 in Q25 */
1272 for (i = 0; i < LPC_ORDER; i++)
1273 lsp[i] = (lpc[i] * bandwidth_expand[i] + (1 << 14)) >> 15;
1275 /* Apply bandwidth expansion on the LPC coefficients */
1276 f[0] = f[1] = 1 << 25;
1278 /* Compute the remaining coefficients */
1279 for (i = 0; i < LPC_ORDER / 2; i++) {
1281 f[2 * i + 2] = -f[2 * i] - ((lsp[i] + lsp[LPC_ORDER - 1 - i]) << 12);
1283 f[2 * i + 3] = f[2 * i + 1] - ((lsp[i] - lsp[LPC_ORDER - 1 - i]) << 12);
1286 /* Divide f1[5] and f2[5] by 2 for use in polynomial evaluation */
1288 f[LPC_ORDER + 1] >>= 1;
1290 /* Normalize and shorten */
1292 for (i = 1; i < LPC_ORDER + 2; i++)
1293 max = FFMAX(max, FFABS(f[i]));
1295 shift = normalize_bits_int32(max);
1297 for (i = 0; i < LPC_ORDER + 2; i++)
1298 f[i] = av_clipl_int32((int64_t)(f[i] << shift) + (1 << 15)) >> 16;
1301 * Evaluate F1 and F2 at uniform intervals of pi/256 along the
1302 * unit circle and check for zero crossings.
1306 for (i = 0; i <= LPC_ORDER / 2; i++)
1307 temp += f[2 * i] * cos_tab[0];
1308 prev_val = av_clipl_int32(temp << 1);
1310 for ( i = 1; i < COS_TBL_SIZE / 2; i++) {
1313 for (j = 0; j <= LPC_ORDER / 2; j++)
1314 temp += f[LPC_ORDER - 2 * j + p] * cos_tab[i * j % COS_TBL_SIZE];
1315 cur_val = av_clipl_int32(temp << 1);
1317 /* Check for sign change, indicating a zero crossing */
1318 if ((cur_val ^ prev_val) < 0) {
1319 int abs_cur = FFABS(cur_val);
1320 int abs_prev = FFABS(prev_val);
1321 int sum = abs_cur + abs_prev;
1323 shift = normalize_bits_int32(sum);
1325 abs_prev = abs_prev << shift >> 8;
1326 lsp[count++] = ((i - 1) << 7) + (abs_prev >> 1) / (sum >> 16);
1328 if (count == LPC_ORDER)
1331 /* Switch between sum and difference polynomials */
1336 for (j = 0; j <= LPC_ORDER / 2; j++){
1337 temp += f[LPC_ORDER - 2 * j + p] *
1338 cos_tab[i * j % COS_TBL_SIZE];
1340 cur_val = av_clipl_int32(temp<<1);
1345 if (count != LPC_ORDER)
1346 memcpy(lsp, prev_lsp, LPC_ORDER * sizeof(int16_t));
1350 * Quantize the current LSP subvector.
1352 * @param num band number
1353 * @param offset offset of the current subvector in an LPC_ORDER vector
1354 * @param size size of the current subvector
1356 #define get_index(num, offset, size) \
1358 int error, max = -1;\
1361 for (i = 0; i < LSP_CB_SIZE; i++) {\
1362 for (j = 0; j < size; j++){\
1363 temp[j] = (weight[j + (offset)] * lsp_band##num[i][j] +\
1366 error = dot_product(lsp + (offset), temp, size, 1) << 1;\
1367 error -= dot_product(lsp_band##num[i], temp, size, 1);\
1370 lsp_index[num] = i;\
1376 * Vector quantize the LSP frequencies.
1378 * @param lsp the current lsp vector
1379 * @param prev_lsp the previous lsp vector
1381 static void lsp_quantize(uint8_t *lsp_index, int16_t *lsp, int16_t *prev_lsp)
1383 int16_t weight[LPC_ORDER];
1387 /* Calculate the VQ weighting vector */
1388 weight[0] = (1 << 20) / (lsp[1] - lsp[0]);
1389 weight[LPC_ORDER - 1] = (1 << 20) /
1390 (lsp[LPC_ORDER - 1] - lsp[LPC_ORDER - 2]);
1392 for (i = 1; i < LPC_ORDER - 1; i++) {
1393 min = FFMIN(lsp[i] - lsp[i - 1], lsp[i + 1] - lsp[i]);
1395 weight[i] = (1 << 20) / min;
1397 weight[i] = INT16_MAX;
1402 for (i = 0; i < LPC_ORDER; i++)
1403 max = FFMAX(weight[i], max);
1405 shift = normalize_bits_int16(max);
1406 for (i = 0; i < LPC_ORDER; i++) {
1407 weight[i] <<= shift;
1410 /* Compute the VQ target vector */
1411 for (i = 0; i < LPC_ORDER; i++) {
1412 lsp[i] -= dc_lsp[i] +
1413 (((prev_lsp[i] - dc_lsp[i]) * 12288 + (1 << 14)) >> 15);
1422 * Apply the formant perceptual weighting filter.
1424 * @param flt_coef filter coefficients
1425 * @param unq_lpc unquantized lpc vector
1427 static void perceptual_filter(G723_1_Context *p, int16_t *flt_coef,
1428 int16_t *unq_lpc, int16_t *buf)
1430 int16_t vector[FRAME_LEN + LPC_ORDER];
1433 memcpy(buf, p->iir_mem, sizeof(int16_t) * LPC_ORDER);
1434 memcpy(vector, p->fir_mem, sizeof(int16_t) * LPC_ORDER);
1435 memcpy(vector + LPC_ORDER, buf + LPC_ORDER, sizeof(int16_t) * FRAME_LEN);
1437 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
1438 for (k = 0; k < LPC_ORDER; k++) {
1439 flt_coef[k + 2 * l] = (unq_lpc[k + l] * percept_flt_tbl[0][k] +
1441 flt_coef[k + 2 * l + LPC_ORDER] = (unq_lpc[k + l] *
1442 percept_flt_tbl[1][k] +
1445 iir_filter(flt_coef + 2 * l, flt_coef + 2 * l + LPC_ORDER, vector + i,
1449 memcpy(p->iir_mem, buf + FRAME_LEN, sizeof(int16_t) * LPC_ORDER);
1450 memcpy(p->fir_mem, vector + FRAME_LEN, sizeof(int16_t) * LPC_ORDER);
1454 * Estimate the open loop pitch period.
1456 * @param buf perceptually weighted speech
1457 * @param start estimation is carried out from this position
1459 static int estimate_pitch(int16_t *buf, int start)
1462 int max_ccr = 0x4000;
1463 int max_eng = 0x7fff;
1464 int index = PITCH_MIN;
1465 int offset = start - PITCH_MIN + 1;
1467 int ccr, eng, orig_eng, ccr_eng, exp;
1472 orig_eng = dot_product(buf + offset, buf + offset, HALF_FRAME_LEN, 0);
1474 for (i = PITCH_MIN; i <= PITCH_MAX - 3; i++) {
1477 /* Update energy and compute correlation */
1478 orig_eng += buf[offset] * buf[offset] -
1479 buf[offset + HALF_FRAME_LEN] * buf[offset + HALF_FRAME_LEN];
1480 ccr = dot_product(buf + start, buf + offset, HALF_FRAME_LEN, 0);
1484 /* Split into mantissa and exponent to maintain precision */
1485 exp = normalize_bits_int32(ccr);
1486 ccr = av_clipl_int32((int64_t)(ccr << exp) + (1 << 15)) >> 16;
1489 temp = normalize_bits_int32(ccr);
1490 ccr = ccr << temp >> 16;
1493 temp = normalize_bits_int32(orig_eng);
1494 eng = av_clipl_int32((int64_t)(orig_eng << temp) + (1 << 15)) >> 16;
1504 if (exp + 1 < max_exp)
1507 /* Equalize exponents before comparison */
1508 if (exp + 1 == max_exp)
1509 temp = max_ccr >> 1;
1512 ccr_eng = ccr * max_eng;
1513 diff = ccr_eng - eng * temp;
1514 if (diff > 0 && (i - index < PITCH_MIN || diff > ccr_eng >> 2)) {
1526 * Compute harmonic noise filter parameters.
1528 * @param buf perceptually weighted speech
1529 * @param pitch_lag open loop pitch period
1530 * @param hf harmonic filter parameters
1532 static void comp_harmonic_coeff(int16_t *buf, int16_t pitch_lag, HFParam *hf)
1534 int ccr, eng, max_ccr, max_eng;
1539 for (i = 0, j = pitch_lag - 3; j <= pitch_lag + 3; i++, j++) {
1540 /* Compute residual energy */
1541 energy[i << 1] = dot_product(buf - j, buf - j, SUBFRAME_LEN, 0);
1542 /* Compute correlation */
1543 energy[(i << 1) + 1] = dot_product(buf, buf - j, SUBFRAME_LEN, 0);
1546 /* Compute target energy */
1547 energy[14] = dot_product(buf, buf, SUBFRAME_LEN, 0);
1551 for (i = 0; i < 15; i++)
1552 max = FFMAX(max, FFABS(energy[i]));
1554 exp = normalize_bits_int32(max);
1555 for (i = 0; i < 15; i++) {
1556 energy[i] = av_clipl_int32((int64_t)(energy[i] << exp) +
1565 for (i = 0; i <= 6; i++) {
1566 eng = energy[i << 1];
1567 ccr = energy[(i << 1) + 1];
1572 ccr = (ccr * ccr + (1 << 14)) >> 15;
1573 diff = ccr * max_eng - eng * max_ccr;
1581 if (hf->index == -1) {
1582 hf->index = pitch_lag;
1586 eng = energy[14] * max_eng;
1587 eng = (eng >> 2) + (eng >> 3);
1588 ccr = energy[(hf->index << 1) + 1] * energy[(hf->index << 1) + 1];
1590 eng = energy[(hf->index << 1) + 1];
1595 hf->gain = ((eng << 15) / max_eng * 0x2800 + (1 << 14)) >> 15;
1597 hf->index += pitch_lag - 3;
1601 * Apply the harmonic noise shaping filter.
1603 * @param hf filter parameters
1605 static void harmonic_filter(HFParam *hf, int16_t *src, int16_t *dest)
1609 for (i = 0; i < SUBFRAME_LEN; i++) {
1610 int64_t temp = hf->gain * src[i - hf->index] << 1;
1611 dest[i] = av_clipl_int32((src[i] << 16) - temp + (1 << 15)) >> 16;
1615 static void harmonic_noise_sub(HFParam *hf, int16_t *src, int16_t *dest)
1618 for (i = 0; i < SUBFRAME_LEN; i++) {
1619 int64_t temp = hf->gain * src[i - hf->index] << 1;
1620 dest[i] = av_clipl_int32(((dest[i] - src[i]) << 16) + temp +
1627 * Combined synthesis and formant perceptual weighting filer.
1629 * @param qnt_lpc quantized lpc coefficients
1630 * @param perf_lpc perceptual filter coefficients
1631 * @param perf_fir perceptual filter fir memory
1632 * @param perf_iir perceptual filter iir memory
1633 * @param scale the filter output will be scaled by 2^scale
1635 static void synth_percept_filter(int16_t *qnt_lpc, int16_t *perf_lpc,
1636 int16_t *perf_fir, int16_t *perf_iir,
1637 int16_t *src, int16_t *dest, int scale)
1640 int16_t buf_16[SUBFRAME_LEN + LPC_ORDER];
1641 int64_t buf[SUBFRAME_LEN];
1643 int16_t *bptr_16 = buf_16 + LPC_ORDER;
1645 memcpy(buf_16, perf_fir, sizeof(int16_t) * LPC_ORDER);
1646 memcpy(dest - LPC_ORDER, perf_iir, sizeof(int16_t) * LPC_ORDER);
1648 for (i = 0; i < SUBFRAME_LEN; i++) {
1650 for (j = 1; j <= LPC_ORDER; j++)
1651 temp -= qnt_lpc[j - 1] * bptr_16[i - j];
1653 buf[i] = (src[i] << 15) + (temp << 3);
1654 bptr_16[i] = av_clipl_int32(buf[i] + (1 << 15)) >> 16;
1657 for (i = 0; i < SUBFRAME_LEN; i++) {
1658 int64_t fir = 0, iir = 0;
1659 for (j = 1; j <= LPC_ORDER; j++) {
1660 fir -= perf_lpc[j - 1] * bptr_16[i - j];
1661 iir += perf_lpc[j + LPC_ORDER - 1] * dest[i - j];
1663 dest[i] = av_clipl_int32(((buf[i] + (fir << 3)) << scale) + (iir << 3) +
1666 memcpy(perf_fir, buf_16 + SUBFRAME_LEN, sizeof(int16_t) * LPC_ORDER);
1667 memcpy(perf_iir, dest + SUBFRAME_LEN - LPC_ORDER,
1668 sizeof(int16_t) * LPC_ORDER);
1672 * Compute the adaptive codebook contribution.
1674 * @param buf input signal
1675 * @param index the current subframe index
1677 static void acb_search(G723_1_Context *p, int16_t *residual,
1678 int16_t *impulse_resp, int16_t *buf,
1682 int16_t flt_buf[PITCH_ORDER][SUBFRAME_LEN];
1684 const int16_t *cb_tbl = adaptive_cb_gain85;
1686 int ccr_buf[PITCH_ORDER * SUBFRAMES << 2];
1688 int pitch_lag = p->pitch_lag[index >> 1];
1691 int odd_frame = index & 1;
1692 int iter = 3 + odd_frame;
1696 int i, j, k, l, max;
1700 if (pitch_lag == PITCH_MIN)
1703 pitch_lag = FFMIN(pitch_lag, PITCH_MAX - 5);
1706 for (i = 0; i < iter; i++) {
1707 get_residual(residual, p->prev_excitation, pitch_lag + i - 1);
1709 for (j = 0; j < SUBFRAME_LEN; j++) {
1711 for (k = 0; k <= j; k++)
1712 temp += residual[PITCH_ORDER - 1 + k] * impulse_resp[j - k];
1713 flt_buf[PITCH_ORDER - 1][j] = av_clipl_int32((temp << 1) +
1717 for (j = PITCH_ORDER - 2; j >= 0; j--) {
1718 flt_buf[j][0] = ((residual[j] << 13) + (1 << 14)) >> 15;
1719 for (k = 1; k < SUBFRAME_LEN; k++) {
1720 temp = (flt_buf[j + 1][k - 1] << 15) +
1721 residual[j] * impulse_resp[k];
1722 flt_buf[j][k] = av_clipl_int32((temp << 1) + (1 << 15)) >> 16;
1726 /* Compute crosscorrelation with the signal */
1727 for (j = 0; j < PITCH_ORDER; j++) {
1728 temp = dot_product(buf, flt_buf[j], SUBFRAME_LEN, 0);
1729 ccr_buf[count++] = av_clipl_int32(temp << 1);
1732 /* Compute energies */
1733 for (j = 0; j < PITCH_ORDER; j++) {
1734 ccr_buf[count++] = dot_product(flt_buf[j], flt_buf[j],
1738 for (j = 1; j < PITCH_ORDER; j++) {
1739 for (k = 0; k < j; k++) {
1740 temp = dot_product(flt_buf[j], flt_buf[k], SUBFRAME_LEN, 0);
1741 ccr_buf[count++] = av_clipl_int32(temp<<2);
1746 /* Normalize and shorten */
1748 for (i = 0; i < 20 * iter; i++)
1749 max = FFMAX(max, FFABS(ccr_buf[i]));
1751 temp = normalize_bits_int32(max);
1753 for (i = 0; i < 20 * iter; i++){
1754 ccr_buf[i] = av_clipl_int32((int64_t)(ccr_buf[i] << temp) +
1759 for (i = 0; i < iter; i++) {
1760 /* Select quantization table */
1761 if (!odd_frame && pitch_lag + i - 1 >= SUBFRAME_LEN - 2 ||
1762 odd_frame && pitch_lag >= SUBFRAME_LEN - 2) {
1763 cb_tbl = adaptive_cb_gain170;
1767 for (j = 0, k = 0; j < tbl_size; j++, k += 20) {
1769 for (l = 0; l < 20; l++)
1770 temp += ccr_buf[20 * i + l] * cb_tbl[k + l];
1771 temp = av_clipl_int32(temp);
1782 pitch_lag += acb_lag - 1;
1786 p->pitch_lag[index >> 1] = pitch_lag;
1787 p->subframe[index].ad_cb_lag = acb_lag;
1788 p->subframe[index].ad_cb_gain = acb_gain;
1792 * Subtract the adaptive codebook contribution from the input
1793 * to obtain the residual.
1795 * @param buf target vector
1797 static void sub_acb_contrib(int16_t *residual, int16_t *impulse_resp,
1801 /* Subtract adaptive CB contribution to obtain the residual */
1802 for (i = 0; i < SUBFRAME_LEN; i++) {
1803 int64_t temp = buf[i] << 14;
1804 for (j = 0; j <= i; j++)
1805 temp -= residual[j] * impulse_resp[i - j];
1807 buf[i] = av_clipl_int32((temp << 2) + (1 << 15)) >> 16;
1812 * Quantize the residual signal using the fixed codebook (MP-MLQ).
1814 * @param optim optimized fixed codebook parameters
1815 * @param buf excitation vector
1817 static void get_fcb_param(FCBParam *optim, int16_t *impulse_resp,
1818 int16_t *buf, int pulse_cnt, int pitch_lag)
1821 int16_t impulse_r[SUBFRAME_LEN];
1822 int16_t temp_corr[SUBFRAME_LEN];
1823 int16_t impulse_corr[SUBFRAME_LEN];
1825 int ccr1[SUBFRAME_LEN];
1826 int ccr2[SUBFRAME_LEN];
1827 int amp, err, max, max_amp_index, min, scale, i, j, k, l;
1831 /* Update impulse response */
1832 memcpy(impulse_r, impulse_resp, sizeof(int16_t) * SUBFRAME_LEN);
1833 param.dirac_train = 0;
1834 if (pitch_lag < SUBFRAME_LEN - 2) {
1835 param.dirac_train = 1;
1836 gen_dirac_train(impulse_r, pitch_lag);
1839 for (i = 0; i < SUBFRAME_LEN; i++)
1840 temp_corr[i] = impulse_r[i] >> 1;
1842 /* Compute impulse response autocorrelation */
1843 temp = dot_product(temp_corr, temp_corr, SUBFRAME_LEN, 1);
1845 scale = normalize_bits_int32(temp);
1846 impulse_corr[0] = av_clipl_int32((temp << scale) + (1 << 15)) >> 16;
1848 for (i = 1; i < SUBFRAME_LEN; i++) {
1849 temp = dot_product(temp_corr + i, temp_corr, SUBFRAME_LEN - i, 1);
1850 impulse_corr[i] = av_clipl_int32((temp << scale) + (1 << 15)) >> 16;
1853 /* Compute crosscorrelation of impulse response with residual signal */
1855 for (i = 0; i < SUBFRAME_LEN; i++){
1856 temp = dot_product(buf + i, impulse_r, SUBFRAME_LEN - i, 1);
1858 ccr1[i] = temp >> -scale;
1860 ccr1[i] = av_clipl_int32(temp << scale);
1864 for (i = 0; i < GRID_SIZE; i++) {
1865 /* Maximize the crosscorrelation */
1867 for (j = i; j < SUBFRAME_LEN; j += GRID_SIZE) {
1868 temp = FFABS(ccr1[j]);
1871 param.pulse_pos[0] = j;
1875 /* Quantize the gain (max crosscorrelation/impulse_corr[0]) */
1878 max_amp_index = GAIN_LEVELS - 2;
1879 for (j = max_amp_index; j >= 2; j--) {
1880 temp = av_clipl_int32((int64_t)fixed_cb_gain[j] *
1881 impulse_corr[0] << 1);
1882 temp = FFABS(temp - amp);
1890 /* Select additional gain values */
1891 for (j = 1; j < 5; j++) {
1892 for (k = i; k < SUBFRAME_LEN; k += GRID_SIZE) {
1896 param.amp_index = max_amp_index + j - 2;
1897 amp = fixed_cb_gain[param.amp_index];
1899 param.pulse_sign[0] = (ccr2[param.pulse_pos[0]] < 0) ? -amp : amp;
1900 temp_corr[param.pulse_pos[0]] = 1;
1902 for (k = 1; k < pulse_cnt; k++) {
1904 for (l = i; l < SUBFRAME_LEN; l += GRID_SIZE) {
1907 temp = impulse_corr[FFABS(l - param.pulse_pos[k - 1])];
1908 temp = av_clipl_int32((int64_t)temp *
1909 param.pulse_sign[k - 1] << 1);
1911 temp = FFABS(ccr2[l]);
1914 param.pulse_pos[k] = l;
1918 param.pulse_sign[k] = (ccr2[param.pulse_pos[k]] < 0) ?
1920 temp_corr[param.pulse_pos[k]] = 1;
1923 /* Create the error vector */
1924 memset(temp_corr, 0, sizeof(int16_t) * SUBFRAME_LEN);
1926 for (k = 0; k < pulse_cnt; k++)
1927 temp_corr[param.pulse_pos[k]] = param.pulse_sign[k];
1929 for (k = SUBFRAME_LEN - 1; k >= 0; k--) {
1931 for (l = 0; l <= k; l++) {
1932 int prod = av_clipl_int32((int64_t)temp_corr[l] *
1933 impulse_r[k - l] << 1);
1934 temp = av_clipl_int32(temp + prod);
1936 temp_corr[k] = temp << 2 >> 16;
1939 /* Compute square of error */
1941 for (k = 0; k < SUBFRAME_LEN; k++) {
1943 prod = av_clipl_int32((int64_t)buf[k] * temp_corr[k] << 1);
1944 err = av_clipl_int32(err - prod);
1945 prod = av_clipl_int32((int64_t)temp_corr[k] * temp_corr[k]);
1946 err = av_clipl_int32(err + prod);
1950 if (err < optim->min_err) {
1951 optim->min_err = err;
1952 optim->grid_index = i;
1953 optim->amp_index = param.amp_index;
1954 optim->dirac_train = param.dirac_train;
1956 for (k = 0; k < pulse_cnt; k++) {
1957 optim->pulse_sign[k] = param.pulse_sign[k];
1958 optim->pulse_pos[k] = param.pulse_pos[k];
1966 * Encode the pulse position and gain of the current subframe.
1968 * @param optim optimized fixed CB parameters
1969 * @param buf excitation vector
1971 static void pack_fcb_param(G723_1_Subframe *subfrm, FCBParam *optim,
1972 int16_t *buf, int pulse_cnt)
1976 j = PULSE_MAX - pulse_cnt;
1978 subfrm->pulse_sign = 0;
1979 subfrm->pulse_pos = 0;
1981 for (i = 0; i < SUBFRAME_LEN >> 1; i++) {
1982 int val = buf[optim->grid_index + (i << 1)];
1984 subfrm->pulse_pos += combinatorial_table[j][i];
1986 subfrm->pulse_sign <<= 1;
1987 if (val < 0) subfrm->pulse_sign++;
1990 if (j == PULSE_MAX) break;
1993 subfrm->amp_index = optim->amp_index;
1994 subfrm->grid_index = optim->grid_index;
1995 subfrm->dirac_train = optim->dirac_train;
1999 * Compute the fixed codebook excitation.
2001 * @param buf target vector
2002 * @param impulse_resp impulse response of the combined filter
2004 static void fcb_search(G723_1_Context *p, int16_t *impulse_resp,
2005 int16_t *buf, int index)
2008 int pulse_cnt = pulses[index];
2011 optim.min_err = 1 << 30;
2012 get_fcb_param(&optim, impulse_resp, buf, pulse_cnt, SUBFRAME_LEN);
2014 if (p->pitch_lag[index >> 1] < SUBFRAME_LEN - 2) {
2015 get_fcb_param(&optim, impulse_resp, buf, pulse_cnt,
2016 p->pitch_lag[index >> 1]);
2019 /* Reconstruct the excitation */
2020 memset(buf, 0, sizeof(int16_t) * SUBFRAME_LEN);
2021 for (i = 0; i < pulse_cnt; i++)
2022 buf[optim.pulse_pos[i]] = optim.pulse_sign[i];
2024 pack_fcb_param(&p->subframe[index], &optim, buf, pulse_cnt);
2026 if (optim.dirac_train)
2027 gen_dirac_train(buf, p->pitch_lag[index >> 1]);
2031 * Pack the frame parameters into output bitstream.
2033 * @param frame output buffer
2034 * @param size size of the buffer
2036 static int pack_bitstream(G723_1_Context *p, unsigned char *frame, int size)
2039 int info_bits, i, temp;
2041 init_put_bits(&pb, frame, size);
2043 if (p->cur_rate == Rate6k3) {
2045 put_bits(&pb, 2, info_bits);
2048 put_bits(&pb, 8, p->lsp_index[2]);
2049 put_bits(&pb, 8, p->lsp_index[1]);
2050 put_bits(&pb, 8, p->lsp_index[0]);
2052 put_bits(&pb, 7, p->pitch_lag[0] - PITCH_MIN);
2053 put_bits(&pb, 2, p->subframe[1].ad_cb_lag);
2054 put_bits(&pb, 7, p->pitch_lag[1] - PITCH_MIN);
2055 put_bits(&pb, 2, p->subframe[3].ad_cb_lag);
2057 /* Write 12 bit combined gain */
2058 for (i = 0; i < SUBFRAMES; i++) {
2059 temp = p->subframe[i].ad_cb_gain * GAIN_LEVELS +
2060 p->subframe[i].amp_index;
2061 if (p->cur_rate == Rate6k3)
2062 temp += p->subframe[i].dirac_train << 11;
2063 put_bits(&pb, 12, temp);
2066 put_bits(&pb, 1, p->subframe[0].grid_index);
2067 put_bits(&pb, 1, p->subframe[1].grid_index);
2068 put_bits(&pb, 1, p->subframe[2].grid_index);
2069 put_bits(&pb, 1, p->subframe[3].grid_index);
2071 if (p->cur_rate == Rate6k3) {
2072 skip_put_bits(&pb, 1); /* reserved bit */
2074 /* Write 13 bit combined position index */
2075 temp = (p->subframe[0].pulse_pos >> 16) * 810 +
2076 (p->subframe[1].pulse_pos >> 14) * 90 +
2077 (p->subframe[2].pulse_pos >> 16) * 9 +
2078 (p->subframe[3].pulse_pos >> 14);
2079 put_bits(&pb, 13, temp);
2081 put_bits(&pb, 16, p->subframe[0].pulse_pos & 0xffff);
2082 put_bits(&pb, 14, p->subframe[1].pulse_pos & 0x3fff);
2083 put_bits(&pb, 16, p->subframe[2].pulse_pos & 0xffff);
2084 put_bits(&pb, 14, p->subframe[3].pulse_pos & 0x3fff);
2086 put_bits(&pb, 6, p->subframe[0].pulse_sign);
2087 put_bits(&pb, 5, p->subframe[1].pulse_sign);
2088 put_bits(&pb, 6, p->subframe[2].pulse_sign);
2089 put_bits(&pb, 5, p->subframe[3].pulse_sign);
2092 flush_put_bits(&pb);
2093 return frame_size[info_bits];
2096 static int g723_1_encode_frame(AVCodecContext *avctx, unsigned char *buf,
2097 int buf_size, void *data)
2099 G723_1_Context *p = avctx->priv_data;
2100 int16_t unq_lpc[LPC_ORDER * SUBFRAMES];
2101 int16_t qnt_lpc[LPC_ORDER * SUBFRAMES];
2102 int16_t cur_lsp[LPC_ORDER];
2103 int16_t weighted_lpc[LPC_ORDER * SUBFRAMES << 1];
2104 int16_t vector[FRAME_LEN + PITCH_MAX];
2111 highpass_filter(in, &p->hpf_fir_mem, &p->hpf_iir_mem);
2113 memcpy(vector, p->prev_data, HALF_FRAME_LEN * sizeof(int16_t));
2114 memcpy(vector + HALF_FRAME_LEN, in, FRAME_LEN * sizeof(int16_t));
2116 comp_lpc_coeff(vector, unq_lpc);
2117 lpc2lsp(&unq_lpc[LPC_ORDER * 3], p->prev_lsp, cur_lsp);
2118 lsp_quantize(p->lsp_index, cur_lsp, p->prev_lsp);
2121 memcpy(vector + LPC_ORDER, p->prev_data + SUBFRAME_LEN,
2122 sizeof(int16_t) * SUBFRAME_LEN);
2123 memcpy(vector + LPC_ORDER + SUBFRAME_LEN, in,
2124 sizeof(int16_t) * (HALF_FRAME_LEN + SUBFRAME_LEN));
2125 memcpy(p->prev_data, in + HALF_FRAME_LEN,
2126 sizeof(int16_t) * HALF_FRAME_LEN);
2127 memcpy(in, vector + LPC_ORDER, sizeof(int16_t) * FRAME_LEN);
2129 perceptual_filter(p, weighted_lpc, unq_lpc, vector);
2131 memcpy(in, vector + LPC_ORDER, sizeof(int16_t) * FRAME_LEN);
2132 memcpy(vector, p->prev_weight_sig, sizeof(int16_t) * PITCH_MAX);
2133 memcpy(vector + PITCH_MAX, in, sizeof(int16_t) * FRAME_LEN);
2135 scale_vector(vector, FRAME_LEN + PITCH_MAX);
2137 p->pitch_lag[0] = estimate_pitch(vector, PITCH_MAX);
2138 p->pitch_lag[1] = estimate_pitch(vector, PITCH_MAX + HALF_FRAME_LEN);
2140 for (i = PITCH_MAX, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
2141 comp_harmonic_coeff(vector + i, p->pitch_lag[j >> 1], hf + j);
2143 memcpy(vector, p->prev_weight_sig, sizeof(int16_t) * PITCH_MAX);
2144 memcpy(vector + PITCH_MAX, in, sizeof(int16_t) * FRAME_LEN);
2145 memcpy(p->prev_weight_sig, vector + FRAME_LEN, sizeof(int16_t) * PITCH_MAX);
2147 for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
2148 harmonic_filter(hf + j, vector + PITCH_MAX + i, in + i);
2150 inverse_quant(cur_lsp, p->prev_lsp, p->lsp_index, 0);
2151 lsp_interpolate(qnt_lpc, cur_lsp, p->prev_lsp);
2153 memcpy(p->prev_lsp, cur_lsp, sizeof(int16_t) * LPC_ORDER);
2156 for (i = 0; i < SUBFRAMES; i++) {
2157 int16_t impulse_resp[SUBFRAME_LEN];
2158 int16_t residual[SUBFRAME_LEN + PITCH_ORDER - 1];
2159 int16_t flt_in[SUBFRAME_LEN];
2160 int16_t zero[LPC_ORDER], fir[LPC_ORDER], iir[LPC_ORDER];
2163 * Compute the combined impulse response of the synthesis filter,
2164 * formant perceptual weighting filter and harmonic noise shaping filter
2166 memset(zero, 0, sizeof(int16_t) * LPC_ORDER);
2167 memset(vector, 0, sizeof(int16_t) * PITCH_MAX);
2168 memset(flt_in, 0, sizeof(int16_t) * SUBFRAME_LEN);
2170 flt_in[0] = 1 << 13; /* Unit impulse */
2171 synth_percept_filter(qnt_lpc + offset, weighted_lpc + (offset << 1),
2172 zero, zero, flt_in, vector + PITCH_MAX, 1);
2173 harmonic_filter(hf + i, vector + PITCH_MAX, impulse_resp);
2175 /* Compute the combined zero input response */
2177 memcpy(fir, p->perf_fir_mem, sizeof(int16_t) * LPC_ORDER);
2178 memcpy(iir, p->perf_iir_mem, sizeof(int16_t) * LPC_ORDER);
2180 synth_percept_filter(qnt_lpc + offset, weighted_lpc + (offset << 1),
2181 fir, iir, flt_in, vector + PITCH_MAX, 0);
2182 memcpy(vector, p->harmonic_mem, sizeof(int16_t) * PITCH_MAX);
2183 harmonic_noise_sub(hf + i, vector + PITCH_MAX, in);
2185 acb_search(p, residual, impulse_resp, in, i);
2186 gen_acb_excitation(residual, p->prev_excitation,p->pitch_lag[i >> 1],
2187 p->subframe[i], p->cur_rate);
2188 sub_acb_contrib(residual, impulse_resp, in);
2190 fcb_search(p, impulse_resp, in, i);
2192 /* Reconstruct the excitation */
2193 gen_acb_excitation(impulse_resp, p->prev_excitation, p->pitch_lag[i >> 1],
2194 p->subframe[i], Rate6k3);
2196 memmove(p->prev_excitation, p->prev_excitation + SUBFRAME_LEN,
2197 sizeof(int16_t) * (PITCH_MAX - SUBFRAME_LEN));
2198 for (j = 0; j < SUBFRAME_LEN; j++)
2199 in[j] = av_clip_int16((in[j] << 1) + impulse_resp[j]);
2200 memcpy(p->prev_excitation + PITCH_MAX - SUBFRAME_LEN, in,
2201 sizeof(int16_t) * SUBFRAME_LEN);
2203 /* Update filter memories */
2204 synth_percept_filter(qnt_lpc + offset, weighted_lpc + (offset << 1),
2205 p->perf_fir_mem, p->perf_iir_mem,
2206 in, vector + PITCH_MAX, 0);
2207 memmove(p->harmonic_mem, p->harmonic_mem + SUBFRAME_LEN,
2208 sizeof(int16_t) * (PITCH_MAX - SUBFRAME_LEN));
2209 memcpy(p->harmonic_mem + PITCH_MAX - SUBFRAME_LEN, vector + PITCH_MAX,
2210 sizeof(int16_t) * SUBFRAME_LEN);
2213 offset += LPC_ORDER;
2216 return pack_bitstream(p, buf, buf_size);
2219 AVCodec ff_g723_1_encoder = {
2221 .type = AVMEDIA_TYPE_AUDIO,
2222 .id = CODEC_ID_G723_1,
2223 .priv_data_size = sizeof(G723_1_Context),
2224 .init = g723_1_encode_init,
2225 .encode = g723_1_encode_frame,
2226 .long_name = NULL_IF_CONFIG_SMALL("G.723.1"),
2227 .sample_fmts = (const enum AVSampleFormat[]){AV_SAMPLE_FMT_S16,
2228 AV_SAMPLE_FMT_NONE},