2 * G.723.1 compatible decoder
3 * Copyright (c) 2006 Benjamin Larsson
4 * Copyright (c) 2010 Mohamed Naufal Basheer
6 * This file is part of FFmpeg.
8 * FFmpeg is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
13 * FFmpeg is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with FFmpeg; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
25 * G.723.1 compatible decoder
28 #define BITSTREAM_READER_LE
29 #include "libavutil/channel_layout.h"
30 #include "libavutil/mem.h"
31 #include "libavutil/opt.h"
35 #include "acelp_vectors.h"
36 #include "celp_filters.h"
37 #include "celp_math.h"
38 #include "g723_1_data.h"
41 #define CNG_RANDOM_SEED 12345
43 typedef struct g723_1_context {
46 G723_1_Subframe subframe[4];
47 enum FrameType cur_frame_type;
48 enum FrameType past_frame_type;
50 uint8_t lsp_index[LSP_BANDS];
54 int16_t prev_lsp[LPC_ORDER];
55 int16_t sid_lsp[LPC_ORDER];
56 int16_t prev_excitation[PITCH_MAX];
57 int16_t excitation[PITCH_MAX + FRAME_LEN + 4];
58 int16_t synth_mem[LPC_ORDER];
59 int16_t fir_mem[LPC_ORDER];
60 int iir_mem[LPC_ORDER];
69 int pf_gain; ///< formant postfilter
70 ///< gain scaling unit memory
73 int16_t audio[FRAME_LEN + LPC_ORDER + PITCH_MAX + 4];
74 int16_t prev_data[HALF_FRAME_LEN];
75 int16_t prev_weight_sig[PITCH_MAX];
78 int16_t hpf_fir_mem; ///< highpass filter fir
79 int hpf_iir_mem; ///< and iir memories
80 int16_t perf_fir_mem[LPC_ORDER]; ///< perceptual filter fir
81 int16_t perf_iir_mem[LPC_ORDER]; ///< and iir memories
83 int16_t harmonic_mem[PITCH_MAX];
86 static av_cold int g723_1_decode_init(AVCodecContext *avctx)
88 G723_1_Context *p = avctx->priv_data;
90 avctx->channel_layout = AV_CH_LAYOUT_MONO;
91 avctx->sample_fmt = AV_SAMPLE_FMT_S16;
95 memcpy(p->prev_lsp, dc_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
96 memcpy(p->sid_lsp, dc_lsp, LPC_ORDER * sizeof(*p->sid_lsp));
98 p->cng_random_seed = CNG_RANDOM_SEED;
99 p->past_frame_type = SID_FRAME;
105 * Unpack the frame into parameters.
107 * @param p the context
108 * @param buf pointer to the input buffer
109 * @param buf_size size of the input buffer
111 static int unpack_bitstream(G723_1_Context *p, const uint8_t *buf,
116 int temp, info_bits, i;
118 init_get_bits(&gb, buf, buf_size * 8);
120 /* Extract frame type and rate info */
121 info_bits = get_bits(&gb, 2);
123 if (info_bits == 3) {
124 p->cur_frame_type = UNTRANSMITTED_FRAME;
128 /* Extract 24 bit lsp indices, 8 bit for each band */
129 p->lsp_index[2] = get_bits(&gb, 8);
130 p->lsp_index[1] = get_bits(&gb, 8);
131 p->lsp_index[0] = get_bits(&gb, 8);
133 if (info_bits == 2) {
134 p->cur_frame_type = SID_FRAME;
135 p->subframe[0].amp_index = get_bits(&gb, 6);
139 /* Extract the info common to both rates */
140 p->cur_rate = info_bits ? RATE_5300 : RATE_6300;
141 p->cur_frame_type = ACTIVE_FRAME;
143 p->pitch_lag[0] = get_bits(&gb, 7);
144 if (p->pitch_lag[0] > 123) /* test if forbidden code */
146 p->pitch_lag[0] += PITCH_MIN;
147 p->subframe[1].ad_cb_lag = get_bits(&gb, 2);
149 p->pitch_lag[1] = get_bits(&gb, 7);
150 if (p->pitch_lag[1] > 123)
152 p->pitch_lag[1] += PITCH_MIN;
153 p->subframe[3].ad_cb_lag = get_bits(&gb, 2);
154 p->subframe[0].ad_cb_lag = 1;
155 p->subframe[2].ad_cb_lag = 1;
157 for (i = 0; i < SUBFRAMES; i++) {
158 /* Extract combined gain */
159 temp = get_bits(&gb, 12);
161 p->subframe[i].dirac_train = 0;
162 if (p->cur_rate == RATE_6300 && p->pitch_lag[i >> 1] < SUBFRAME_LEN - 2) {
163 p->subframe[i].dirac_train = temp >> 11;
167 p->subframe[i].ad_cb_gain = FASTDIV(temp, GAIN_LEVELS);
168 if (p->subframe[i].ad_cb_gain < ad_cb_len) {
169 p->subframe[i].amp_index = temp - p->subframe[i].ad_cb_gain *
176 p->subframe[0].grid_index = get_bits1(&gb);
177 p->subframe[1].grid_index = get_bits1(&gb);
178 p->subframe[2].grid_index = get_bits1(&gb);
179 p->subframe[3].grid_index = get_bits1(&gb);
181 if (p->cur_rate == RATE_6300) {
182 skip_bits1(&gb); /* skip reserved bit */
184 /* Compute pulse_pos index using the 13-bit combined position index */
185 temp = get_bits(&gb, 13);
186 p->subframe[0].pulse_pos = temp / 810;
188 temp -= p->subframe[0].pulse_pos * 810;
189 p->subframe[1].pulse_pos = FASTDIV(temp, 90);
191 temp -= p->subframe[1].pulse_pos * 90;
192 p->subframe[2].pulse_pos = FASTDIV(temp, 9);
193 p->subframe[3].pulse_pos = temp - p->subframe[2].pulse_pos * 9;
195 p->subframe[0].pulse_pos = (p->subframe[0].pulse_pos << 16) +
197 p->subframe[1].pulse_pos = (p->subframe[1].pulse_pos << 14) +
199 p->subframe[2].pulse_pos = (p->subframe[2].pulse_pos << 16) +
201 p->subframe[3].pulse_pos = (p->subframe[3].pulse_pos << 14) +
204 p->subframe[0].pulse_sign = get_bits(&gb, 6);
205 p->subframe[1].pulse_sign = get_bits(&gb, 5);
206 p->subframe[2].pulse_sign = get_bits(&gb, 6);
207 p->subframe[3].pulse_sign = get_bits(&gb, 5);
208 } else { /* 5300 bps */
209 p->subframe[0].pulse_pos = get_bits(&gb, 12);
210 p->subframe[1].pulse_pos = get_bits(&gb, 12);
211 p->subframe[2].pulse_pos = get_bits(&gb, 12);
212 p->subframe[3].pulse_pos = get_bits(&gb, 12);
214 p->subframe[0].pulse_sign = get_bits(&gb, 4);
215 p->subframe[1].pulse_sign = get_bits(&gb, 4);
216 p->subframe[2].pulse_sign = get_bits(&gb, 4);
217 p->subframe[3].pulse_sign = get_bits(&gb, 4);
224 * Bitexact implementation of sqrt(val/2).
226 static int16_t square_root(unsigned val)
228 av_assert2(!(val & 0x80000000));
230 return (ff_sqrt(val << 1) >> 1) & (~1);
234 * Calculate the number of left-shifts required for normalizing the input.
236 * @param num input number
237 * @param width width of the input, 15 or 31 bits
239 static int normalize_bits(int num, int width)
241 return width - av_log2(num) - 1;
244 #define normalize_bits_int16(num) normalize_bits(num, 15)
245 #define normalize_bits_int32(num) normalize_bits(num, 31)
248 * Scale vector contents based on the largest of their absolutes.
250 static int scale_vector(int16_t *dst, const int16_t *vector, int length)
255 for (i = 0; i < length; i++)
256 max |= FFABS(vector[i]);
258 bits= 14 - av_log2_16bit(max);
259 bits= FFMAX(bits, 0);
261 for (i = 0; i < length; i++)
262 dst[i] = vector[i] << bits >> 3;
268 * Perform inverse quantization of LSP frequencies.
270 * @param cur_lsp the current LSP vector
271 * @param prev_lsp the previous LSP vector
272 * @param lsp_index VQ indices
273 * @param bad_frame bad frame flag
275 static void inverse_quant(int16_t *cur_lsp, int16_t *prev_lsp,
276 uint8_t *lsp_index, int bad_frame)
279 int i, j, temp, stable;
281 /* Check for frame erasure */
288 lsp_index[0] = lsp_index[1] = lsp_index[2] = 0;
291 /* Get the VQ table entry corresponding to the transmitted index */
292 cur_lsp[0] = lsp_band0[lsp_index[0]][0];
293 cur_lsp[1] = lsp_band0[lsp_index[0]][1];
294 cur_lsp[2] = lsp_band0[lsp_index[0]][2];
295 cur_lsp[3] = lsp_band1[lsp_index[1]][0];
296 cur_lsp[4] = lsp_band1[lsp_index[1]][1];
297 cur_lsp[5] = lsp_band1[lsp_index[1]][2];
298 cur_lsp[6] = lsp_band2[lsp_index[2]][0];
299 cur_lsp[7] = lsp_band2[lsp_index[2]][1];
300 cur_lsp[8] = lsp_band2[lsp_index[2]][2];
301 cur_lsp[9] = lsp_band2[lsp_index[2]][3];
303 /* Add predicted vector & DC component to the previously quantized vector */
304 for (i = 0; i < LPC_ORDER; i++) {
305 temp = ((prev_lsp[i] - dc_lsp[i]) * pred + (1 << 14)) >> 15;
306 cur_lsp[i] += dc_lsp[i] + temp;
309 for (i = 0; i < LPC_ORDER; i++) {
310 cur_lsp[0] = FFMAX(cur_lsp[0], 0x180);
311 cur_lsp[LPC_ORDER - 1] = FFMIN(cur_lsp[LPC_ORDER - 1], 0x7e00);
313 /* Stability check */
314 for (j = 1; j < LPC_ORDER; j++) {
315 temp = min_dist + cur_lsp[j - 1] - cur_lsp[j];
318 cur_lsp[j - 1] -= temp;
323 for (j = 1; j < LPC_ORDER; j++) {
324 temp = cur_lsp[j - 1] + min_dist - cur_lsp[j] - 4;
334 memcpy(cur_lsp, prev_lsp, LPC_ORDER * sizeof(*cur_lsp));
338 * Bitexact implementation of 2ab scaled by 1/2^16.
340 * @param a 32 bit multiplicand
341 * @param b 16 bit multiplier
343 #define MULL2(a, b) \
347 * Convert LSP frequencies to LPC coefficients.
349 * @param lpc buffer for LPC coefficients
351 static void lsp2lpc(int16_t *lpc)
353 int f1[LPC_ORDER / 2 + 1];
354 int f2[LPC_ORDER / 2 + 1];
357 /* Calculate negative cosine */
358 for (j = 0; j < LPC_ORDER; j++) {
359 int index = (lpc[j] >> 7) & 0x1FF;
360 int offset = lpc[j] & 0x7f;
361 int temp1 = cos_tab[index] << 16;
362 int temp2 = (cos_tab[index + 1] - cos_tab[index]) *
363 ((offset << 8) + 0x80) << 1;
365 lpc[j] = -(av_sat_dadd32(1 << 15, temp1 + temp2) >> 16);
369 * Compute sum and difference polynomial coefficients
370 * (bitexact alternative to lsp2poly() in lsp.c)
372 /* Initialize with values in Q28 */
374 f1[1] = (lpc[0] << 14) + (lpc[2] << 14);
375 f1[2] = lpc[0] * lpc[2] + (2 << 28);
378 f2[1] = (lpc[1] << 14) + (lpc[3] << 14);
379 f2[2] = lpc[1] * lpc[3] + (2 << 28);
382 * Calculate and scale the coefficients by 1/2 in
383 * each iteration for a final scaling factor of Q25
385 for (i = 2; i < LPC_ORDER / 2; i++) {
386 f1[i + 1] = f1[i - 1] + MULL2(f1[i], lpc[2 * i]);
387 f2[i + 1] = f2[i - 1] + MULL2(f2[i], lpc[2 * i + 1]);
389 for (j = i; j >= 2; j--) {
390 f1[j] = MULL2(f1[j - 1], lpc[2 * i]) +
391 (f1[j] >> 1) + (f1[j - 2] >> 1);
392 f2[j] = MULL2(f2[j - 1], lpc[2 * i + 1]) +
393 (f2[j] >> 1) + (f2[j - 2] >> 1);
398 f1[1] = ((lpc[2 * i] << 16 >> i) + f1[1]) >> 1;
399 f2[1] = ((lpc[2 * i + 1] << 16 >> i) + f2[1]) >> 1;
402 /* Convert polynomial coefficients to LPC coefficients */
403 for (i = 0; i < LPC_ORDER / 2; i++) {
404 int64_t ff1 = f1[i + 1] + f1[i];
405 int64_t ff2 = f2[i + 1] - f2[i];
407 lpc[i] = av_clipl_int32(((ff1 + ff2) << 3) + (1 << 15)) >> 16;
408 lpc[LPC_ORDER - i - 1] = av_clipl_int32(((ff1 - ff2) << 3) +
414 * Quantize LSP frequencies by interpolation and convert them to
415 * the corresponding LPC coefficients.
417 * @param lpc buffer for LPC coefficients
418 * @param cur_lsp the current LSP vector
419 * @param prev_lsp the previous LSP vector
421 static void lsp_interpolate(int16_t *lpc, int16_t *cur_lsp, int16_t *prev_lsp)
424 int16_t *lpc_ptr = lpc;
426 /* cur_lsp * 0.25 + prev_lsp * 0.75 */
427 ff_acelp_weighted_vector_sum(lpc, cur_lsp, prev_lsp,
428 4096, 12288, 1 << 13, 14, LPC_ORDER);
429 ff_acelp_weighted_vector_sum(lpc + LPC_ORDER, cur_lsp, prev_lsp,
430 8192, 8192, 1 << 13, 14, LPC_ORDER);
431 ff_acelp_weighted_vector_sum(lpc + 2 * LPC_ORDER, cur_lsp, prev_lsp,
432 12288, 4096, 1 << 13, 14, LPC_ORDER);
433 memcpy(lpc + 3 * LPC_ORDER, cur_lsp, LPC_ORDER * sizeof(*lpc));
435 for (i = 0; i < SUBFRAMES; i++) {
437 lpc_ptr += LPC_ORDER;
442 * Generate a train of dirac functions with period as pitch lag.
444 static void gen_dirac_train(int16_t *buf, int pitch_lag)
446 int16_t vector[SUBFRAME_LEN];
449 memcpy(vector, buf, SUBFRAME_LEN * sizeof(*vector));
450 for (i = pitch_lag; i < SUBFRAME_LEN; i += pitch_lag) {
451 for (j = 0; j < SUBFRAME_LEN - i; j++)
452 buf[i + j] += vector[j];
457 * Generate fixed codebook excitation vector.
459 * @param vector decoded excitation vector
460 * @param subfrm current subframe
461 * @param cur_rate current bitrate
462 * @param pitch_lag closed loop pitch lag
463 * @param index current subframe index
465 static void gen_fcb_excitation(int16_t *vector, G723_1_Subframe *subfrm,
466 enum Rate cur_rate, int pitch_lag, int index)
470 memset(vector, 0, SUBFRAME_LEN * sizeof(*vector));
472 if (cur_rate == RATE_6300) {
473 if (subfrm->pulse_pos >= max_pos[index])
476 /* Decode amplitudes and positions */
477 j = PULSE_MAX - pulses[index];
478 temp = subfrm->pulse_pos;
479 for (i = 0; i < SUBFRAME_LEN / GRID_SIZE; i++) {
480 temp -= combinatorial_table[j][i];
483 temp += combinatorial_table[j++][i];
484 if (subfrm->pulse_sign & (1 << (PULSE_MAX - j))) {
485 vector[subfrm->grid_index + GRID_SIZE * i] =
486 -fixed_cb_gain[subfrm->amp_index];
488 vector[subfrm->grid_index + GRID_SIZE * i] =
489 fixed_cb_gain[subfrm->amp_index];
494 if (subfrm->dirac_train == 1)
495 gen_dirac_train(vector, pitch_lag);
496 } else { /* 5300 bps */
497 int cb_gain = fixed_cb_gain[subfrm->amp_index];
498 int cb_shift = subfrm->grid_index;
499 int cb_sign = subfrm->pulse_sign;
500 int cb_pos = subfrm->pulse_pos;
501 int offset, beta, lag;
503 for (i = 0; i < 8; i += 2) {
504 offset = ((cb_pos & 7) << 3) + cb_shift + i;
505 vector[offset] = (cb_sign & 1) ? cb_gain : -cb_gain;
510 /* Enhance harmonic components */
511 lag = pitch_contrib[subfrm->ad_cb_gain << 1] + pitch_lag +
512 subfrm->ad_cb_lag - 1;
513 beta = pitch_contrib[(subfrm->ad_cb_gain << 1) + 1];
515 if (lag < SUBFRAME_LEN - 2) {
516 for (i = lag; i < SUBFRAME_LEN; i++)
517 vector[i] += beta * vector[i - lag] >> 15;
523 * Get delayed contribution from the previous excitation vector.
525 static void get_residual(int16_t *residual, int16_t *prev_excitation, int lag)
527 int offset = PITCH_MAX - PITCH_ORDER / 2 - lag;
530 residual[0] = prev_excitation[offset];
531 residual[1] = prev_excitation[offset + 1];
534 for (i = 2; i < SUBFRAME_LEN + PITCH_ORDER - 1; i++)
535 residual[i] = prev_excitation[offset + (i - 2) % lag];
538 static int dot_product(const int16_t *a, const int16_t *b, int length)
540 int sum = ff_dot_product(a,b,length);
541 return av_sat_add32(sum, sum);
545 * Generate adaptive codebook excitation.
547 static void gen_acb_excitation(int16_t *vector, int16_t *prev_excitation,
548 int pitch_lag, G723_1_Subframe *subfrm,
551 int16_t residual[SUBFRAME_LEN + PITCH_ORDER - 1];
552 const int16_t *cb_ptr;
553 int lag = pitch_lag + subfrm->ad_cb_lag - 1;
558 get_residual(residual, prev_excitation, lag);
560 /* Select quantization table */
561 if (cur_rate == RATE_6300 && pitch_lag < SUBFRAME_LEN - 2) {
562 cb_ptr = adaptive_cb_gain85;
564 cb_ptr = adaptive_cb_gain170;
566 /* Calculate adaptive vector */
567 cb_ptr += subfrm->ad_cb_gain * 20;
568 for (i = 0; i < SUBFRAME_LEN; i++) {
569 sum = ff_dot_product(residual + i, cb_ptr, PITCH_ORDER);
570 vector[i] = av_sat_dadd32(1 << 15, av_sat_add32(sum, sum)) >> 16;
575 * Estimate maximum auto-correlation around pitch lag.
577 * @param buf buffer with offset applied
578 * @param offset offset of the excitation vector
579 * @param ccr_max pointer to the maximum auto-correlation
580 * @param pitch_lag decoded pitch lag
581 * @param length length of autocorrelation
582 * @param dir forward lag(1) / backward lag(-1)
584 static int autocorr_max(const int16_t *buf, int offset, int *ccr_max,
585 int pitch_lag, int length, int dir)
587 int limit, ccr, lag = 0;
590 pitch_lag = FFMIN(PITCH_MAX - 3, pitch_lag);
592 limit = FFMIN(FRAME_LEN + PITCH_MAX - offset - length, pitch_lag + 3);
594 limit = pitch_lag + 3;
596 for (i = pitch_lag - 3; i <= limit; i++) {
597 ccr = dot_product(buf, buf + dir * i, length);
599 if (ccr > *ccr_max) {
608 * Calculate pitch postfilter optimal and scaling gains.
610 * @param lag pitch postfilter forward/backward lag
611 * @param ppf pitch postfilter parameters
612 * @param cur_rate current bitrate
613 * @param tgt_eng target energy
614 * @param ccr cross-correlation
615 * @param res_eng residual energy
617 static void comp_ppf_gains(int lag, PPFParam *ppf, enum Rate cur_rate,
618 int tgt_eng, int ccr, int res_eng)
620 int pf_residual; /* square of postfiltered residual */
625 temp1 = tgt_eng * res_eng >> 1;
626 temp2 = ccr * ccr << 1;
629 if (ccr >= res_eng) {
630 ppf->opt_gain = ppf_gain_weight[cur_rate];
632 ppf->opt_gain = (ccr << 15) / res_eng *
633 ppf_gain_weight[cur_rate] >> 15;
635 /* pf_res^2 = tgt_eng + 2*ccr*gain + res_eng*gain^2 */
636 temp1 = (tgt_eng << 15) + (ccr * ppf->opt_gain << 1);
637 temp2 = (ppf->opt_gain * ppf->opt_gain >> 15) * res_eng;
638 pf_residual = av_sat_add32(temp1, temp2 + (1 << 15)) >> 16;
640 if (tgt_eng >= pf_residual << 1) {
643 temp1 = (tgt_eng << 14) / pf_residual;
646 /* scaling_gain = sqrt(tgt_eng/pf_res^2) */
647 ppf->sc_gain = square_root(temp1 << 16);
650 ppf->sc_gain = 0x7fff;
653 ppf->opt_gain = av_clip_int16(ppf->opt_gain * ppf->sc_gain >> 15);
657 * Calculate pitch postfilter parameters.
659 * @param p the context
660 * @param offset offset of the excitation vector
661 * @param pitch_lag decoded pitch lag
662 * @param ppf pitch postfilter parameters
663 * @param cur_rate current bitrate
665 static void comp_ppf_coeff(G723_1_Context *p, int offset, int pitch_lag,
666 PPFParam *ppf, enum Rate cur_rate)
675 * 1 - forward cross-correlation
676 * 2 - forward residual energy
677 * 3 - backward cross-correlation
678 * 4 - backward residual energy
680 int energy[5] = {0, 0, 0, 0, 0};
681 int16_t *buf = p->audio + LPC_ORDER + offset;
682 int fwd_lag = autocorr_max(buf, offset, &energy[1], pitch_lag,
684 int back_lag = autocorr_max(buf, offset, &energy[3], pitch_lag,
689 ppf->sc_gain = 0x7fff;
691 /* Case 0, Section 3.6 */
692 if (!back_lag && !fwd_lag)
695 /* Compute target energy */
696 energy[0] = dot_product(buf, buf, SUBFRAME_LEN);
698 /* Compute forward residual energy */
700 energy[2] = dot_product(buf + fwd_lag, buf + fwd_lag, SUBFRAME_LEN);
702 /* Compute backward residual energy */
704 energy[4] = dot_product(buf - back_lag, buf - back_lag, SUBFRAME_LEN);
706 /* Normalize and shorten */
708 for (i = 0; i < 5; i++)
709 temp1 = FFMAX(energy[i], temp1);
711 scale = normalize_bits(temp1, 31);
712 for (i = 0; i < 5; i++)
713 energy[i] = (energy[i] << scale) >> 16;
715 if (fwd_lag && !back_lag) { /* Case 1 */
716 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
718 } else if (!fwd_lag) { /* Case 2 */
719 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
721 } else { /* Case 3 */
724 * Select the largest of energy[1]^2/energy[2]
725 * and energy[3]^2/energy[4]
727 temp1 = energy[4] * ((energy[1] * energy[1] + (1 << 14)) >> 15);
728 temp2 = energy[2] * ((energy[3] * energy[3] + (1 << 14)) >> 15);
729 if (temp1 >= temp2) {
730 comp_ppf_gains(fwd_lag, ppf, cur_rate, energy[0], energy[1],
733 comp_ppf_gains(-back_lag, ppf, cur_rate, energy[0], energy[3],
740 * Classify frames as voiced/unvoiced.
742 * @param p the context
743 * @param pitch_lag decoded pitch_lag
744 * @param exc_eng excitation energy estimation
745 * @param scale scaling factor of exc_eng
747 * @return residual interpolation index if voiced, 0 otherwise
749 static int comp_interp_index(G723_1_Context *p, int pitch_lag,
750 int *exc_eng, int *scale)
752 int offset = PITCH_MAX + 2 * SUBFRAME_LEN;
753 int16_t *buf = p->audio + LPC_ORDER;
755 int index, ccr, tgt_eng, best_eng, temp;
757 *scale = scale_vector(buf, p->excitation, FRAME_LEN + PITCH_MAX);
760 /* Compute maximum backward cross-correlation */
762 index = autocorr_max(buf, offset, &ccr, pitch_lag, SUBFRAME_LEN * 2, -1);
763 ccr = av_sat_add32(ccr, 1 << 15) >> 16;
765 /* Compute target energy */
766 tgt_eng = dot_product(buf, buf, SUBFRAME_LEN * 2);
767 *exc_eng = av_sat_add32(tgt_eng, 1 << 15) >> 16;
772 /* Compute best energy */
773 best_eng = dot_product(buf - index, buf - index, SUBFRAME_LEN * 2);
774 best_eng = av_sat_add32(best_eng, 1 << 15) >> 16;
776 temp = best_eng * *exc_eng >> 3;
778 if (temp < ccr * ccr) {
785 * Peform residual interpolation based on frame classification.
787 * @param buf decoded excitation vector
788 * @param out output vector
789 * @param lag decoded pitch lag
790 * @param gain interpolated gain
791 * @param rseed seed for random number generator
793 static void residual_interp(int16_t *buf, int16_t *out, int lag,
794 int gain, int *rseed)
797 if (lag) { /* Voiced */
798 int16_t *vector_ptr = buf + PITCH_MAX;
800 for (i = 0; i < lag; i++)
801 out[i] = vector_ptr[i - lag] * 3 >> 2;
802 av_memcpy_backptr((uint8_t*)(out + lag), lag * sizeof(*out),
803 (FRAME_LEN - lag) * sizeof(*out));
804 } else { /* Unvoiced */
805 for (i = 0; i < FRAME_LEN; i++) {
806 *rseed = *rseed * 521 + 259;
807 out[i] = gain * *rseed >> 15;
809 memset(buf, 0, (FRAME_LEN + PITCH_MAX) * sizeof(*buf));
814 * Perform IIR filtering.
816 * @param fir_coef FIR coefficients
817 * @param iir_coef IIR coefficients
818 * @param src source vector
819 * @param dest destination vector
820 * @param width width of the output, 16 bits(0) / 32 bits(1)
822 #define iir_filter(fir_coef, iir_coef, src, dest, width)\
825 int res_shift = 16 & ~-(width);\
826 int in_shift = 16 - res_shift;\
828 for (m = 0; m < SUBFRAME_LEN; m++) {\
830 for (n = 1; n <= LPC_ORDER; n++) {\
831 filter -= (fir_coef)[n - 1] * (src)[m - n] -\
832 (iir_coef)[n - 1] * ((dest)[m - n] >> in_shift);\
835 (dest)[m] = av_clipl_int32(((src)[m] << 16) + (filter << 3) +\
836 (1 << 15)) >> res_shift;\
841 * Adjust gain of postfiltered signal.
843 * @param p the context
844 * @param buf postfiltered output vector
845 * @param energy input energy coefficient
847 static void gain_scale(G723_1_Context *p, int16_t * buf, int energy)
849 int num, denom, gain, bits1, bits2;
854 for (i = 0; i < SUBFRAME_LEN; i++) {
855 int temp = buf[i] >> 2;
857 denom = av_sat_dadd32(denom, temp);
861 bits1 = normalize_bits(num, 31);
862 bits2 = normalize_bits(denom, 31);
863 num = num << bits1 >> 1;
866 bits2 = 5 + bits1 - bits2;
867 bits2 = FFMAX(0, bits2);
869 gain = (num >> 1) / (denom >> 16);
870 gain = square_root(gain << 16 >> bits2);
875 for (i = 0; i < SUBFRAME_LEN; i++) {
876 p->pf_gain = (15 * p->pf_gain + gain + (1 << 3)) >> 4;
877 buf[i] = av_clip_int16((buf[i] * (p->pf_gain + (p->pf_gain >> 4)) +
883 * Perform formant filtering.
885 * @param p the context
886 * @param lpc quantized lpc coefficients
887 * @param buf input buffer
888 * @param dst output buffer
890 static void formant_postfilter(G723_1_Context *p, int16_t *lpc,
891 int16_t *buf, int16_t *dst)
893 int16_t filter_coef[2][LPC_ORDER];
894 int filter_signal[LPC_ORDER + FRAME_LEN], *signal_ptr;
897 memcpy(buf, p->fir_mem, LPC_ORDER * sizeof(*buf));
898 memcpy(filter_signal, p->iir_mem, LPC_ORDER * sizeof(*filter_signal));
900 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
901 for (k = 0; k < LPC_ORDER; k++) {
902 filter_coef[0][k] = (-lpc[k] * postfilter_tbl[0][k] +
904 filter_coef[1][k] = (-lpc[k] * postfilter_tbl[1][k] +
907 iir_filter(filter_coef[0], filter_coef[1], buf + i,
908 filter_signal + i, 1);
912 memcpy(p->fir_mem, buf + FRAME_LEN, LPC_ORDER * sizeof(int16_t));
913 memcpy(p->iir_mem, filter_signal + FRAME_LEN, LPC_ORDER * sizeof(int));
916 signal_ptr = filter_signal + LPC_ORDER;
917 for (i = 0; i < SUBFRAMES; i++) {
923 scale = scale_vector(dst, buf, SUBFRAME_LEN);
925 /* Compute auto correlation coefficients */
926 auto_corr[0] = dot_product(dst, dst + 1, SUBFRAME_LEN - 1);
927 auto_corr[1] = dot_product(dst, dst, SUBFRAME_LEN);
929 /* Compute reflection coefficient */
930 temp = auto_corr[1] >> 16;
932 temp = (auto_corr[0] >> 2) / temp;
934 p->reflection_coef = (3 * p->reflection_coef + temp + 2) >> 2;
935 temp = -p->reflection_coef >> 1 & ~3;
937 /* Compensation filter */
938 for (j = 0; j < SUBFRAME_LEN; j++) {
939 dst[j] = av_sat_dadd32(signal_ptr[j],
940 (signal_ptr[j - 1] >> 16) * temp) >> 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, dst, energy);
953 signal_ptr += SUBFRAME_LEN;
958 static int sid_gain_to_lsp_index(int gain)
962 else if (gain < 0x20)
963 return gain - 8 << 7;
965 return gain - 20 << 8;
968 static inline int cng_rand(int *state, int base)
970 *state = (*state * 521 + 259) & 0xFFFF;
971 return (*state & 0x7FFF) * base >> 15;
974 static int estimate_sid_gain(G723_1_Context *p)
976 int i, shift, seg, seg2, t, val, val_add, x, y;
978 shift = 16 - p->cur_gain * 2;
980 t = p->sid_gain << shift;
982 t = p->sid_gain >> -shift;
983 x = t * cng_filt[0] >> 16;
985 if (x >= cng_bseg[2])
988 if (x >= cng_bseg[1]) {
993 seg = (x >= cng_bseg[0]);
995 seg2 = FFMIN(seg, 3);
999 for (i = 0; i < shift; i++) {
1000 t = seg * 32 + (val << seg2);
1009 t = seg * 32 + (val << seg2);
1012 t = seg * 32 + (val + 1 << seg2);
1014 val = (seg2 - 1 << 4) + val;
1018 t = seg * 32 + (val - 1 << seg2);
1020 val = (seg2 - 1 << 4) + val;
1028 static void generate_noise(G723_1_Context *p)
1032 int signs[SUBFRAMES / 2 * 11], pos[SUBFRAMES / 2 * 11];
1033 int tmp[SUBFRAME_LEN * 2];
1034 int16_t *vector_ptr;
1036 int b0, c, delta, x, shift;
1038 p->pitch_lag[0] = cng_rand(&p->cng_random_seed, 21) + 123;
1039 p->pitch_lag[1] = cng_rand(&p->cng_random_seed, 19) + 123;
1041 for (i = 0; i < SUBFRAMES; i++) {
1042 p->subframe[i].ad_cb_gain = cng_rand(&p->cng_random_seed, 50) + 1;
1043 p->subframe[i].ad_cb_lag = cng_adaptive_cb_lag[i];
1046 for (i = 0; i < SUBFRAMES / 2; i++) {
1047 t = cng_rand(&p->cng_random_seed, 1 << 13);
1049 off[i * 2 + 1] = ((t >> 1) & 1) + SUBFRAME_LEN;
1051 for (j = 0; j < 11; j++) {
1052 signs[i * 11 + j] = (t & 1) * 2 - 1 << 14;
1058 for (i = 0; i < SUBFRAMES; i++) {
1059 for (j = 0; j < SUBFRAME_LEN / 2; j++)
1061 t = SUBFRAME_LEN / 2;
1062 for (j = 0; j < pulses[i]; j++, idx++) {
1063 int idx2 = cng_rand(&p->cng_random_seed, t);
1065 pos[idx] = tmp[idx2] * 2 + off[i];
1066 tmp[idx2] = tmp[--t];
1070 vector_ptr = p->audio + LPC_ORDER;
1071 memcpy(vector_ptr, p->prev_excitation,
1072 PITCH_MAX * sizeof(*p->excitation));
1073 for (i = 0; i < SUBFRAMES; i += 2) {
1074 gen_acb_excitation(vector_ptr, vector_ptr,
1075 p->pitch_lag[i >> 1], &p->subframe[i],
1077 gen_acb_excitation(vector_ptr + SUBFRAME_LEN,
1078 vector_ptr + SUBFRAME_LEN,
1079 p->pitch_lag[i >> 1], &p->subframe[i + 1],
1083 for (j = 0; j < SUBFRAME_LEN * 2; j++)
1084 t |= FFABS(vector_ptr[j]);
1085 t = FFMIN(t, 0x7FFF);
1089 shift = -10 + av_log2(t);
1095 for (j = 0; j < SUBFRAME_LEN * 2; j++) {
1096 t = vector_ptr[j] << -shift;
1101 for (j = 0; j < SUBFRAME_LEN * 2; j++) {
1102 t = vector_ptr[j] >> shift;
1109 for (j = 0; j < 11; j++)
1110 b0 += tmp[pos[(i / 2) * 11 + j]] * signs[(i / 2) * 11 + j];
1111 b0 = b0 * 2 * 2979LL + (1 << 29) >> 30; // approximated division by 11
1113 c = p->cur_gain * (p->cur_gain * SUBFRAME_LEN >> 5);
1114 if (shift * 2 + 3 >= 0)
1115 c >>= shift * 2 + 3;
1117 c <<= -(shift * 2 + 3);
1118 c = (av_clipl_int32(sum << 1) - c) * 2979LL >> 15;
1120 delta = b0 * b0 * 2 - c;
1124 delta = square_root(delta);
1127 if (FFABS(t) < FFABS(x))
1135 x = av_clip(x, -10000, 10000);
1137 for (j = 0; j < 11; j++) {
1138 idx = (i / 2) * 11 + j;
1139 vector_ptr[pos[idx]] = av_clip_int16(vector_ptr[pos[idx]] +
1140 (x * signs[idx] >> 15));
1143 /* copy decoded data to serve as a history for the next decoded subframes */
1144 memcpy(vector_ptr + PITCH_MAX, vector_ptr,
1145 sizeof(*vector_ptr) * SUBFRAME_LEN * 2);
1146 vector_ptr += SUBFRAME_LEN * 2;
1148 /* Save the excitation for the next frame */
1149 memcpy(p->prev_excitation, p->audio + LPC_ORDER + FRAME_LEN,
1150 PITCH_MAX * sizeof(*p->excitation));
1153 static int g723_1_decode_frame(AVCodecContext *avctx, void *data,
1154 int *got_frame_ptr, AVPacket *avpkt)
1156 G723_1_Context *p = avctx->priv_data;
1157 AVFrame *frame = data;
1158 const uint8_t *buf = avpkt->data;
1159 int buf_size = avpkt->size;
1160 int dec_mode = buf[0] & 3;
1162 PPFParam ppf[SUBFRAMES];
1163 int16_t cur_lsp[LPC_ORDER];
1164 int16_t lpc[SUBFRAMES * LPC_ORDER];
1165 int16_t acb_vector[SUBFRAME_LEN];
1167 int bad_frame = 0, i, j, ret;
1168 int16_t *audio = p->audio;
1170 if (buf_size < frame_size[dec_mode]) {
1172 av_log(avctx, AV_LOG_WARNING,
1173 "Expected %d bytes, got %d - skipping packet\n",
1174 frame_size[dec_mode], buf_size);
1179 if (unpack_bitstream(p, buf, buf_size) < 0) {
1181 if (p->past_frame_type == ACTIVE_FRAME)
1182 p->cur_frame_type = ACTIVE_FRAME;
1184 p->cur_frame_type = UNTRANSMITTED_FRAME;
1187 frame->nb_samples = FRAME_LEN;
1188 if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
1191 out = (int16_t *)frame->data[0];
1193 if (p->cur_frame_type == ACTIVE_FRAME) {
1195 p->erased_frames = 0;
1196 else if (p->erased_frames != 3)
1199 inverse_quant(cur_lsp, p->prev_lsp, p->lsp_index, bad_frame);
1200 lsp_interpolate(lpc, cur_lsp, p->prev_lsp);
1202 /* Save the lsp_vector for the next frame */
1203 memcpy(p->prev_lsp, cur_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
1205 /* Generate the excitation for the frame */
1206 memcpy(p->excitation, p->prev_excitation,
1207 PITCH_MAX * sizeof(*p->excitation));
1208 if (!p->erased_frames) {
1209 int16_t *vector_ptr = p->excitation + PITCH_MAX;
1211 /* Update interpolation gain memory */
1212 p->interp_gain = fixed_cb_gain[(p->subframe[2].amp_index +
1213 p->subframe[3].amp_index) >> 1];
1214 for (i = 0; i < SUBFRAMES; i++) {
1215 gen_fcb_excitation(vector_ptr, &p->subframe[i], p->cur_rate,
1216 p->pitch_lag[i >> 1], i);
1217 gen_acb_excitation(acb_vector, &p->excitation[SUBFRAME_LEN * i],
1218 p->pitch_lag[i >> 1], &p->subframe[i],
1220 /* Get the total excitation */
1221 for (j = 0; j < SUBFRAME_LEN; j++) {
1222 int v = av_clip_int16(vector_ptr[j] << 1);
1223 vector_ptr[j] = av_clip_int16(v + acb_vector[j]);
1225 vector_ptr += SUBFRAME_LEN;
1228 vector_ptr = p->excitation + PITCH_MAX;
1230 p->interp_index = comp_interp_index(p, p->pitch_lag[1],
1231 &p->sid_gain, &p->cur_gain);
1233 /* Peform pitch postfiltering */
1234 if (p->postfilter) {
1236 for (j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1237 comp_ppf_coeff(p, i, p->pitch_lag[j >> 1],
1238 ppf + j, p->cur_rate);
1240 for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1241 ff_acelp_weighted_vector_sum(p->audio + LPC_ORDER + i,
1243 vector_ptr + i + ppf[j].index,
1246 1 << 14, 15, SUBFRAME_LEN);
1248 audio = vector_ptr - LPC_ORDER;
1251 /* Save the excitation for the next frame */
1252 memcpy(p->prev_excitation, p->excitation + FRAME_LEN,
1253 PITCH_MAX * sizeof(*p->excitation));
1255 p->interp_gain = (p->interp_gain * 3 + 2) >> 2;
1256 if (p->erased_frames == 3) {
1258 memset(p->excitation, 0,
1259 (FRAME_LEN + PITCH_MAX) * sizeof(*p->excitation));
1260 memset(p->prev_excitation, 0,
1261 PITCH_MAX * sizeof(*p->excitation));
1262 memset(frame->data[0], 0,
1263 (FRAME_LEN + LPC_ORDER) * sizeof(int16_t));
1265 int16_t *buf = p->audio + LPC_ORDER;
1267 /* Regenerate frame */
1268 residual_interp(p->excitation, buf, p->interp_index,
1269 p->interp_gain, &p->random_seed);
1271 /* Save the excitation for the next frame */
1272 memcpy(p->prev_excitation, buf + (FRAME_LEN - PITCH_MAX),
1273 PITCH_MAX * sizeof(*p->excitation));
1276 p->cng_random_seed = CNG_RANDOM_SEED;
1278 if (p->cur_frame_type == SID_FRAME) {
1279 p->sid_gain = sid_gain_to_lsp_index(p->subframe[0].amp_index);
1280 inverse_quant(p->sid_lsp, p->prev_lsp, p->lsp_index, 0);
1281 } else if (p->past_frame_type == ACTIVE_FRAME) {
1282 p->sid_gain = estimate_sid_gain(p);
1285 if (p->past_frame_type == ACTIVE_FRAME)
1286 p->cur_gain = p->sid_gain;
1288 p->cur_gain = (p->cur_gain * 7 + p->sid_gain) >> 3;
1290 lsp_interpolate(lpc, p->sid_lsp, p->prev_lsp);
1291 /* Save the lsp_vector for the next frame */
1292 memcpy(p->prev_lsp, p->sid_lsp, LPC_ORDER * sizeof(*p->prev_lsp));
1295 p->past_frame_type = p->cur_frame_type;
1297 memcpy(p->audio, p->synth_mem, LPC_ORDER * sizeof(*p->audio));
1298 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
1299 ff_celp_lp_synthesis_filter(p->audio + i, &lpc[j * LPC_ORDER],
1300 audio + i, SUBFRAME_LEN, LPC_ORDER,
1302 memcpy(p->synth_mem, p->audio + FRAME_LEN, LPC_ORDER * sizeof(*p->audio));
1304 if (p->postfilter) {
1305 formant_postfilter(p, lpc, p->audio, out);
1306 } else { // if output is not postfiltered it should be scaled by 2
1307 for (i = 0; i < FRAME_LEN; i++)
1308 out[i] = av_clip_int16(p->audio[LPC_ORDER + i] << 1);
1313 return frame_size[dec_mode];
1316 #define OFFSET(x) offsetof(G723_1_Context, x)
1317 #define AD AV_OPT_FLAG_AUDIO_PARAM | AV_OPT_FLAG_DECODING_PARAM
1319 static const AVOption options[] = {
1320 { "postfilter", "postfilter on/off", OFFSET(postfilter), AV_OPT_TYPE_INT,
1321 { .i64 = 1 }, 0, 1, AD },
1326 static const AVClass g723_1dec_class = {
1327 .class_name = "G.723.1 decoder",
1328 .item_name = av_default_item_name,
1330 .version = LIBAVUTIL_VERSION_INT,
1333 AVCodec ff_g723_1_decoder = {
1335 .type = AVMEDIA_TYPE_AUDIO,
1336 .id = AV_CODEC_ID_G723_1,
1337 .priv_data_size = sizeof(G723_1_Context),
1338 .init = g723_1_decode_init,
1339 .decode = g723_1_decode_frame,
1340 .long_name = NULL_IF_CONFIG_SMALL("G.723.1"),
1341 .capabilities = CODEC_CAP_SUBFRAMES | CODEC_CAP_DR1,
1342 .priv_class = &g723_1dec_class,
1345 #if CONFIG_G723_1_ENCODER
1346 #define BITSTREAM_WRITER_LE
1347 #include "put_bits.h"
1349 static av_cold int g723_1_encode_init(AVCodecContext *avctx)
1351 G723_1_Context *p = avctx->priv_data;
1353 if (avctx->sample_rate != 8000) {
1354 av_log(avctx, AV_LOG_ERROR, "Only 8000Hz sample rate supported\n");
1358 if (avctx->channels != 1) {
1359 av_log(avctx, AV_LOG_ERROR, "Only mono supported\n");
1360 return AVERROR(EINVAL);
1363 if (avctx->bit_rate == 6300) {
1364 p->cur_rate = RATE_6300;
1365 } else if (avctx->bit_rate == 5300) {
1366 av_log(avctx, AV_LOG_ERROR, "Bitrate not supported yet, use 6.3k\n");
1367 return AVERROR_PATCHWELCOME;
1369 av_log(avctx, AV_LOG_ERROR,
1370 "Bitrate not supported, use 6.3k\n");
1371 return AVERROR(EINVAL);
1373 avctx->frame_size = 240;
1374 memcpy(p->prev_lsp, dc_lsp, LPC_ORDER * sizeof(int16_t));
1380 * Remove DC component from the input signal.
1382 * @param buf input signal
1383 * @param fir zero memory
1384 * @param iir pole memory
1386 static void highpass_filter(int16_t *buf, int16_t *fir, int *iir)
1389 for (i = 0; i < FRAME_LEN; i++) {
1390 *iir = (buf[i] << 15) + ((-*fir) << 15) + MULL2(*iir, 0x7f00);
1392 buf[i] = av_clipl_int32((int64_t)*iir + (1 << 15)) >> 16;
1397 * Estimate autocorrelation of the input vector.
1399 * @param buf input buffer
1400 * @param autocorr autocorrelation coefficients vector
1402 static void comp_autocorr(int16_t *buf, int16_t *autocorr)
1405 int16_t vector[LPC_FRAME];
1407 scale_vector(vector, buf, LPC_FRAME);
1409 /* Apply the Hamming window */
1410 for (i = 0; i < LPC_FRAME; i++)
1411 vector[i] = (vector[i] * hamming_window[i] + (1 << 14)) >> 15;
1413 /* Compute the first autocorrelation coefficient */
1414 temp = ff_dot_product(vector, vector, LPC_FRAME);
1416 /* Apply a white noise correlation factor of (1025/1024) */
1420 scale = normalize_bits_int32(temp);
1421 autocorr[0] = av_clipl_int32((int64_t)(temp << scale) +
1424 /* Compute the remaining coefficients */
1426 memset(autocorr + 1, 0, LPC_ORDER * sizeof(int16_t));
1428 for (i = 1; i <= LPC_ORDER; i++) {
1429 temp = ff_dot_product(vector, vector + i, LPC_FRAME - i);
1430 temp = MULL2((temp << scale), binomial_window[i - 1]);
1431 autocorr[i] = av_clipl_int32((int64_t)temp + (1 << 15)) >> 16;
1437 * Use Levinson-Durbin recursion to compute LPC coefficients from
1438 * autocorrelation values.
1440 * @param lpc LPC coefficients vector
1441 * @param autocorr autocorrelation coefficients vector
1442 * @param error prediction error
1444 static void levinson_durbin(int16_t *lpc, int16_t *autocorr, int16_t error)
1446 int16_t vector[LPC_ORDER];
1447 int16_t partial_corr;
1450 memset(lpc, 0, LPC_ORDER * sizeof(int16_t));
1452 for (i = 0; i < LPC_ORDER; i++) {
1453 /* Compute the partial correlation coefficient */
1455 for (j = 0; j < i; j++)
1456 temp -= lpc[j] * autocorr[i - j - 1];
1457 temp = ((autocorr[i] << 13) + temp) << 3;
1459 if (FFABS(temp) >= (error << 16))
1462 partial_corr = temp / (error << 1);
1464 lpc[i] = av_clipl_int32((int64_t)(partial_corr << 14) +
1467 /* Update the prediction error */
1468 temp = MULL2(temp, partial_corr);
1469 error = av_clipl_int32((int64_t)(error << 16) - temp +
1472 memcpy(vector, lpc, i * sizeof(int16_t));
1473 for (j = 0; j < i; j++) {
1474 temp = partial_corr * vector[i - j - 1] << 1;
1475 lpc[j] = av_clipl_int32((int64_t)(lpc[j] << 16) - temp +
1482 * Calculate LPC coefficients for the current frame.
1484 * @param buf current frame
1485 * @param prev_data 2 trailing subframes of the previous frame
1486 * @param lpc LPC coefficients vector
1488 static void comp_lpc_coeff(int16_t *buf, int16_t *lpc)
1490 int16_t autocorr[(LPC_ORDER + 1) * SUBFRAMES];
1491 int16_t *autocorr_ptr = autocorr;
1492 int16_t *lpc_ptr = lpc;
1495 for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
1496 comp_autocorr(buf + i, autocorr_ptr);
1497 levinson_durbin(lpc_ptr, autocorr_ptr + 1, autocorr_ptr[0]);
1499 lpc_ptr += LPC_ORDER;
1500 autocorr_ptr += LPC_ORDER + 1;
1504 static void lpc2lsp(int16_t *lpc, int16_t *prev_lsp, int16_t *lsp)
1506 int f[LPC_ORDER + 2]; ///< coefficients of the sum and difference
1507 ///< polynomials (F1, F2) ordered as
1508 ///< f1[0], f2[0], ...., f1[5], f2[5]
1510 int max, shift, cur_val, prev_val, count, p;
1514 /* Initialize f1[0] and f2[0] to 1 in Q25 */
1515 for (i = 0; i < LPC_ORDER; i++)
1516 lsp[i] = (lpc[i] * bandwidth_expand[i] + (1 << 14)) >> 15;
1518 /* Apply bandwidth expansion on the LPC coefficients */
1519 f[0] = f[1] = 1 << 25;
1521 /* Compute the remaining coefficients */
1522 for (i = 0; i < LPC_ORDER / 2; i++) {
1524 f[2 * i + 2] = -f[2 * i] - ((lsp[i] + lsp[LPC_ORDER - 1 - i]) << 12);
1526 f[2 * i + 3] = f[2 * i + 1] - ((lsp[i] - lsp[LPC_ORDER - 1 - i]) << 12);
1529 /* Divide f1[5] and f2[5] by 2 for use in polynomial evaluation */
1531 f[LPC_ORDER + 1] >>= 1;
1533 /* Normalize and shorten */
1535 for (i = 1; i < LPC_ORDER + 2; i++)
1536 max = FFMAX(max, FFABS(f[i]));
1538 shift = normalize_bits_int32(max);
1540 for (i = 0; i < LPC_ORDER + 2; i++)
1541 f[i] = av_clipl_int32((int64_t)(f[i] << shift) + (1 << 15)) >> 16;
1544 * Evaluate F1 and F2 at uniform intervals of pi/256 along the
1545 * unit circle and check for zero crossings.
1549 for (i = 0; i <= LPC_ORDER / 2; i++)
1550 temp += f[2 * i] * cos_tab[0];
1551 prev_val = av_clipl_int32(temp << 1);
1553 for ( i = 1; i < COS_TBL_SIZE / 2; i++) {
1556 for (j = 0; j <= LPC_ORDER / 2; j++)
1557 temp += f[LPC_ORDER - 2 * j + p] * cos_tab[i * j % COS_TBL_SIZE];
1558 cur_val = av_clipl_int32(temp << 1);
1560 /* Check for sign change, indicating a zero crossing */
1561 if ((cur_val ^ prev_val) < 0) {
1562 int abs_cur = FFABS(cur_val);
1563 int abs_prev = FFABS(prev_val);
1564 int sum = abs_cur + abs_prev;
1566 shift = normalize_bits_int32(sum);
1568 abs_prev = abs_prev << shift >> 8;
1569 lsp[count++] = ((i - 1) << 7) + (abs_prev >> 1) / (sum >> 16);
1571 if (count == LPC_ORDER)
1574 /* Switch between sum and difference polynomials */
1579 for (j = 0; j <= LPC_ORDER / 2; j++){
1580 temp += f[LPC_ORDER - 2 * j + p] *
1581 cos_tab[i * j % COS_TBL_SIZE];
1583 cur_val = av_clipl_int32(temp<<1);
1588 if (count != LPC_ORDER)
1589 memcpy(lsp, prev_lsp, LPC_ORDER * sizeof(int16_t));
1593 * Quantize the current LSP subvector.
1595 * @param num band number
1596 * @param offset offset of the current subvector in an LPC_ORDER vector
1597 * @param size size of the current subvector
1599 #define get_index(num, offset, size) \
1601 int error, max = -1;\
1604 for (i = 0; i < LSP_CB_SIZE; i++) {\
1605 for (j = 0; j < size; j++){\
1606 temp[j] = (weight[j + (offset)] * lsp_band##num[i][j] +\
1609 error = dot_product(lsp + (offset), temp, size) << 1;\
1610 error -= dot_product(lsp_band##num[i], temp, size);\
1613 lsp_index[num] = i;\
1619 * Vector quantize the LSP frequencies.
1621 * @param lsp the current lsp vector
1622 * @param prev_lsp the previous lsp vector
1624 static void lsp_quantize(uint8_t *lsp_index, int16_t *lsp, int16_t *prev_lsp)
1626 int16_t weight[LPC_ORDER];
1630 /* Calculate the VQ weighting vector */
1631 weight[0] = (1 << 20) / (lsp[1] - lsp[0]);
1632 weight[LPC_ORDER - 1] = (1 << 20) /
1633 (lsp[LPC_ORDER - 1] - lsp[LPC_ORDER - 2]);
1635 for (i = 1; i < LPC_ORDER - 1; i++) {
1636 min = FFMIN(lsp[i] - lsp[i - 1], lsp[i + 1] - lsp[i]);
1638 weight[i] = (1 << 20) / min;
1640 weight[i] = INT16_MAX;
1645 for (i = 0; i < LPC_ORDER; i++)
1646 max = FFMAX(weight[i], max);
1648 shift = normalize_bits_int16(max);
1649 for (i = 0; i < LPC_ORDER; i++) {
1650 weight[i] <<= shift;
1653 /* Compute the VQ target vector */
1654 for (i = 0; i < LPC_ORDER; i++) {
1655 lsp[i] -= dc_lsp[i] +
1656 (((prev_lsp[i] - dc_lsp[i]) * 12288 + (1 << 14)) >> 15);
1665 * Apply the formant perceptual weighting filter.
1667 * @param flt_coef filter coefficients
1668 * @param unq_lpc unquantized lpc vector
1670 static void perceptual_filter(G723_1_Context *p, int16_t *flt_coef,
1671 int16_t *unq_lpc, int16_t *buf)
1673 int16_t vector[FRAME_LEN + LPC_ORDER];
1676 memcpy(buf, p->iir_mem, sizeof(int16_t) * LPC_ORDER);
1677 memcpy(vector, p->fir_mem, sizeof(int16_t) * LPC_ORDER);
1678 memcpy(vector + LPC_ORDER, buf + LPC_ORDER, sizeof(int16_t) * FRAME_LEN);
1680 for (i = LPC_ORDER, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++) {
1681 for (k = 0; k < LPC_ORDER; k++) {
1682 flt_coef[k + 2 * l] = (unq_lpc[k + l] * percept_flt_tbl[0][k] +
1684 flt_coef[k + 2 * l + LPC_ORDER] = (unq_lpc[k + l] *
1685 percept_flt_tbl[1][k] +
1688 iir_filter(flt_coef + 2 * l, flt_coef + 2 * l + LPC_ORDER, vector + i,
1692 memcpy(p->iir_mem, buf + FRAME_LEN, sizeof(int16_t) * LPC_ORDER);
1693 memcpy(p->fir_mem, vector + FRAME_LEN, sizeof(int16_t) * LPC_ORDER);
1697 * Estimate the open loop pitch period.
1699 * @param buf perceptually weighted speech
1700 * @param start estimation is carried out from this position
1702 static int estimate_pitch(int16_t *buf, int start)
1705 int max_ccr = 0x4000;
1706 int max_eng = 0x7fff;
1707 int index = PITCH_MIN;
1708 int offset = start - PITCH_MIN + 1;
1710 int ccr, eng, orig_eng, ccr_eng, exp;
1715 orig_eng = ff_dot_product(buf + offset, buf + offset, HALF_FRAME_LEN);
1717 for (i = PITCH_MIN; i <= PITCH_MAX - 3; i++) {
1720 /* Update energy and compute correlation */
1721 orig_eng += buf[offset] * buf[offset] -
1722 buf[offset + HALF_FRAME_LEN] * buf[offset + HALF_FRAME_LEN];
1723 ccr = ff_dot_product(buf + start, buf + offset, HALF_FRAME_LEN);
1727 /* Split into mantissa and exponent to maintain precision */
1728 exp = normalize_bits_int32(ccr);
1729 ccr = av_clipl_int32((int64_t)(ccr << exp) + (1 << 15)) >> 16;
1732 temp = normalize_bits_int32(ccr);
1733 ccr = ccr << temp >> 16;
1736 temp = normalize_bits_int32(orig_eng);
1737 eng = av_clipl_int32((int64_t)(orig_eng << temp) + (1 << 15)) >> 16;
1747 if (exp + 1 < max_exp)
1750 /* Equalize exponents before comparison */
1751 if (exp + 1 == max_exp)
1752 temp = max_ccr >> 1;
1755 ccr_eng = ccr * max_eng;
1756 diff = ccr_eng - eng * temp;
1757 if (diff > 0 && (i - index < PITCH_MIN || diff > ccr_eng >> 2)) {
1769 * Compute harmonic noise filter parameters.
1771 * @param buf perceptually weighted speech
1772 * @param pitch_lag open loop pitch period
1773 * @param hf harmonic filter parameters
1775 static void comp_harmonic_coeff(int16_t *buf, int16_t pitch_lag, HFParam *hf)
1777 int ccr, eng, max_ccr, max_eng;
1782 for (i = 0, j = pitch_lag - 3; j <= pitch_lag + 3; i++, j++) {
1783 /* Compute residual energy */
1784 energy[i << 1] = ff_dot_product(buf - j, buf - j, SUBFRAME_LEN);
1785 /* Compute correlation */
1786 energy[(i << 1) + 1] = ff_dot_product(buf, buf - j, SUBFRAME_LEN);
1789 /* Compute target energy */
1790 energy[14] = ff_dot_product(buf, buf, SUBFRAME_LEN);
1794 for (i = 0; i < 15; i++)
1795 max = FFMAX(max, FFABS(energy[i]));
1797 exp = normalize_bits_int32(max);
1798 for (i = 0; i < 15; i++) {
1799 energy[i] = av_clipl_int32((int64_t)(energy[i] << exp) +
1808 for (i = 0; i <= 6; i++) {
1809 eng = energy[i << 1];
1810 ccr = energy[(i << 1) + 1];
1815 ccr = (ccr * ccr + (1 << 14)) >> 15;
1816 diff = ccr * max_eng - eng * max_ccr;
1824 if (hf->index == -1) {
1825 hf->index = pitch_lag;
1829 eng = energy[14] * max_eng;
1830 eng = (eng >> 2) + (eng >> 3);
1831 ccr = energy[(hf->index << 1) + 1] * energy[(hf->index << 1) + 1];
1833 eng = energy[(hf->index << 1) + 1];
1838 hf->gain = ((eng << 15) / max_eng * 0x2800 + (1 << 14)) >> 15;
1840 hf->index += pitch_lag - 3;
1844 * Apply the harmonic noise shaping filter.
1846 * @param hf filter parameters
1848 static void harmonic_filter(HFParam *hf, const int16_t *src, int16_t *dest)
1852 for (i = 0; i < SUBFRAME_LEN; i++) {
1853 int64_t temp = hf->gain * src[i - hf->index] << 1;
1854 dest[i] = av_clipl_int32((src[i] << 16) - temp + (1 << 15)) >> 16;
1858 static void harmonic_noise_sub(HFParam *hf, const int16_t *src, int16_t *dest)
1861 for (i = 0; i < SUBFRAME_LEN; i++) {
1862 int64_t temp = hf->gain * src[i - hf->index] << 1;
1863 dest[i] = av_clipl_int32(((dest[i] - src[i]) << 16) + temp +
1870 * Combined synthesis and formant perceptual weighting filer.
1872 * @param qnt_lpc quantized lpc coefficients
1873 * @param perf_lpc perceptual filter coefficients
1874 * @param perf_fir perceptual filter fir memory
1875 * @param perf_iir perceptual filter iir memory
1876 * @param scale the filter output will be scaled by 2^scale
1878 static void synth_percept_filter(int16_t *qnt_lpc, int16_t *perf_lpc,
1879 int16_t *perf_fir, int16_t *perf_iir,
1880 const int16_t *src, int16_t *dest, int scale)
1883 int16_t buf_16[SUBFRAME_LEN + LPC_ORDER];
1884 int64_t buf[SUBFRAME_LEN];
1886 int16_t *bptr_16 = buf_16 + LPC_ORDER;
1888 memcpy(buf_16, perf_fir, sizeof(int16_t) * LPC_ORDER);
1889 memcpy(dest - LPC_ORDER, perf_iir, sizeof(int16_t) * LPC_ORDER);
1891 for (i = 0; i < SUBFRAME_LEN; i++) {
1893 for (j = 1; j <= LPC_ORDER; j++)
1894 temp -= qnt_lpc[j - 1] * bptr_16[i - j];
1896 buf[i] = (src[i] << 15) + (temp << 3);
1897 bptr_16[i] = av_clipl_int32(buf[i] + (1 << 15)) >> 16;
1900 for (i = 0; i < SUBFRAME_LEN; i++) {
1901 int64_t fir = 0, iir = 0;
1902 for (j = 1; j <= LPC_ORDER; j++) {
1903 fir -= perf_lpc[j - 1] * bptr_16[i - j];
1904 iir += perf_lpc[j + LPC_ORDER - 1] * dest[i - j];
1906 dest[i] = av_clipl_int32(((buf[i] + (fir << 3)) << scale) + (iir << 3) +
1909 memcpy(perf_fir, buf_16 + SUBFRAME_LEN, sizeof(int16_t) * LPC_ORDER);
1910 memcpy(perf_iir, dest + SUBFRAME_LEN - LPC_ORDER,
1911 sizeof(int16_t) * LPC_ORDER);
1915 * Compute the adaptive codebook contribution.
1917 * @param buf input signal
1918 * @param index the current subframe index
1920 static void acb_search(G723_1_Context *p, int16_t *residual,
1921 int16_t *impulse_resp, const int16_t *buf,
1925 int16_t flt_buf[PITCH_ORDER][SUBFRAME_LEN];
1927 const int16_t *cb_tbl = adaptive_cb_gain85;
1929 int ccr_buf[PITCH_ORDER * SUBFRAMES << 2];
1931 int pitch_lag = p->pitch_lag[index >> 1];
1934 int odd_frame = index & 1;
1935 int iter = 3 + odd_frame;
1939 int i, j, k, l, max;
1943 if (pitch_lag == PITCH_MIN)
1946 pitch_lag = FFMIN(pitch_lag, PITCH_MAX - 5);
1949 for (i = 0; i < iter; i++) {
1950 get_residual(residual, p->prev_excitation, pitch_lag + i - 1);
1952 for (j = 0; j < SUBFRAME_LEN; j++) {
1954 for (k = 0; k <= j; k++)
1955 temp += residual[PITCH_ORDER - 1 + k] * impulse_resp[j - k];
1956 flt_buf[PITCH_ORDER - 1][j] = av_clipl_int32((temp << 1) +
1960 for (j = PITCH_ORDER - 2; j >= 0; j--) {
1961 flt_buf[j][0] = ((residual[j] << 13) + (1 << 14)) >> 15;
1962 for (k = 1; k < SUBFRAME_LEN; k++) {
1963 temp = (flt_buf[j + 1][k - 1] << 15) +
1964 residual[j] * impulse_resp[k];
1965 flt_buf[j][k] = av_clipl_int32((temp << 1) + (1 << 15)) >> 16;
1969 /* Compute crosscorrelation with the signal */
1970 for (j = 0; j < PITCH_ORDER; j++) {
1971 temp = ff_dot_product(buf, flt_buf[j], SUBFRAME_LEN);
1972 ccr_buf[count++] = av_clipl_int32(temp << 1);
1975 /* Compute energies */
1976 for (j = 0; j < PITCH_ORDER; j++) {
1977 ccr_buf[count++] = dot_product(flt_buf[j], flt_buf[j],
1981 for (j = 1; j < PITCH_ORDER; j++) {
1982 for (k = 0; k < j; k++) {
1983 temp = ff_dot_product(flt_buf[j], flt_buf[k], SUBFRAME_LEN);
1984 ccr_buf[count++] = av_clipl_int32(temp<<2);
1989 /* Normalize and shorten */
1991 for (i = 0; i < 20 * iter; i++)
1992 max = FFMAX(max, FFABS(ccr_buf[i]));
1994 temp = normalize_bits_int32(max);
1996 for (i = 0; i < 20 * iter; i++){
1997 ccr_buf[i] = av_clipl_int32((int64_t)(ccr_buf[i] << temp) +
2002 for (i = 0; i < iter; i++) {
2003 /* Select quantization table */
2004 if (!odd_frame && pitch_lag + i - 1 >= SUBFRAME_LEN - 2 ||
2005 odd_frame && pitch_lag >= SUBFRAME_LEN - 2) {
2006 cb_tbl = adaptive_cb_gain170;
2010 for (j = 0, k = 0; j < tbl_size; j++, k += 20) {
2012 for (l = 0; l < 20; l++)
2013 temp += ccr_buf[20 * i + l] * cb_tbl[k + l];
2014 temp = av_clipl_int32(temp);
2025 pitch_lag += acb_lag - 1;
2029 p->pitch_lag[index >> 1] = pitch_lag;
2030 p->subframe[index].ad_cb_lag = acb_lag;
2031 p->subframe[index].ad_cb_gain = acb_gain;
2035 * Subtract the adaptive codebook contribution from the input
2036 * to obtain the residual.
2038 * @param buf target vector
2040 static void sub_acb_contrib(const int16_t *residual, const int16_t *impulse_resp,
2044 /* Subtract adaptive CB contribution to obtain the residual */
2045 for (i = 0; i < SUBFRAME_LEN; i++) {
2046 int64_t temp = buf[i] << 14;
2047 for (j = 0; j <= i; j++)
2048 temp -= residual[j] * impulse_resp[i - j];
2050 buf[i] = av_clipl_int32((temp << 2) + (1 << 15)) >> 16;
2055 * Quantize the residual signal using the fixed codebook (MP-MLQ).
2057 * @param optim optimized fixed codebook parameters
2058 * @param buf excitation vector
2060 static void get_fcb_param(FCBParam *optim, int16_t *impulse_resp,
2061 int16_t *buf, int pulse_cnt, int pitch_lag)
2064 int16_t impulse_r[SUBFRAME_LEN];
2065 int16_t temp_corr[SUBFRAME_LEN];
2066 int16_t impulse_corr[SUBFRAME_LEN];
2068 int ccr1[SUBFRAME_LEN];
2069 int ccr2[SUBFRAME_LEN];
2070 int amp, err, max, max_amp_index, min, scale, i, j, k, l;
2074 /* Update impulse response */
2075 memcpy(impulse_r, impulse_resp, sizeof(int16_t) * SUBFRAME_LEN);
2076 param.dirac_train = 0;
2077 if (pitch_lag < SUBFRAME_LEN - 2) {
2078 param.dirac_train = 1;
2079 gen_dirac_train(impulse_r, pitch_lag);
2082 for (i = 0; i < SUBFRAME_LEN; i++)
2083 temp_corr[i] = impulse_r[i] >> 1;
2085 /* Compute impulse response autocorrelation */
2086 temp = dot_product(temp_corr, temp_corr, SUBFRAME_LEN);
2088 scale = normalize_bits_int32(temp);
2089 impulse_corr[0] = av_clipl_int32((temp << scale) + (1 << 15)) >> 16;
2091 for (i = 1; i < SUBFRAME_LEN; i++) {
2092 temp = dot_product(temp_corr + i, temp_corr, SUBFRAME_LEN - i);
2093 impulse_corr[i] = av_clipl_int32((temp << scale) + (1 << 15)) >> 16;
2096 /* Compute crosscorrelation of impulse response with residual signal */
2098 for (i = 0; i < SUBFRAME_LEN; i++){
2099 temp = dot_product(buf + i, impulse_r, SUBFRAME_LEN - i);
2101 ccr1[i] = temp >> -scale;
2103 ccr1[i] = av_clipl_int32(temp << scale);
2107 for (i = 0; i < GRID_SIZE; i++) {
2108 /* Maximize the crosscorrelation */
2110 for (j = i; j < SUBFRAME_LEN; j += GRID_SIZE) {
2111 temp = FFABS(ccr1[j]);
2114 param.pulse_pos[0] = j;
2118 /* Quantize the gain (max crosscorrelation/impulse_corr[0]) */
2121 max_amp_index = GAIN_LEVELS - 2;
2122 for (j = max_amp_index; j >= 2; j--) {
2123 temp = av_clipl_int32((int64_t)fixed_cb_gain[j] *
2124 impulse_corr[0] << 1);
2125 temp = FFABS(temp - amp);
2133 /* Select additional gain values */
2134 for (j = 1; j < 5; j++) {
2135 for (k = i; k < SUBFRAME_LEN; k += GRID_SIZE) {
2139 param.amp_index = max_amp_index + j - 2;
2140 amp = fixed_cb_gain[param.amp_index];
2142 param.pulse_sign[0] = (ccr2[param.pulse_pos[0]] < 0) ? -amp : amp;
2143 temp_corr[param.pulse_pos[0]] = 1;
2145 for (k = 1; k < pulse_cnt; k++) {
2147 for (l = i; l < SUBFRAME_LEN; l += GRID_SIZE) {
2150 temp = impulse_corr[FFABS(l - param.pulse_pos[k - 1])];
2151 temp = av_clipl_int32((int64_t)temp *
2152 param.pulse_sign[k - 1] << 1);
2154 temp = FFABS(ccr2[l]);
2157 param.pulse_pos[k] = l;
2161 param.pulse_sign[k] = (ccr2[param.pulse_pos[k]] < 0) ?
2163 temp_corr[param.pulse_pos[k]] = 1;
2166 /* Create the error vector */
2167 memset(temp_corr, 0, sizeof(int16_t) * SUBFRAME_LEN);
2169 for (k = 0; k < pulse_cnt; k++)
2170 temp_corr[param.pulse_pos[k]] = param.pulse_sign[k];
2172 for (k = SUBFRAME_LEN - 1; k >= 0; k--) {
2174 for (l = 0; l <= k; l++) {
2175 int prod = av_clipl_int32((int64_t)temp_corr[l] *
2176 impulse_r[k - l] << 1);
2177 temp = av_clipl_int32(temp + prod);
2179 temp_corr[k] = temp << 2 >> 16;
2182 /* Compute square of error */
2184 for (k = 0; k < SUBFRAME_LEN; k++) {
2186 prod = av_clipl_int32((int64_t)buf[k] * temp_corr[k] << 1);
2187 err = av_clipl_int32(err - prod);
2188 prod = av_clipl_int32((int64_t)temp_corr[k] * temp_corr[k]);
2189 err = av_clipl_int32(err + prod);
2193 if (err < optim->min_err) {
2194 optim->min_err = err;
2195 optim->grid_index = i;
2196 optim->amp_index = param.amp_index;
2197 optim->dirac_train = param.dirac_train;
2199 for (k = 0; k < pulse_cnt; k++) {
2200 optim->pulse_sign[k] = param.pulse_sign[k];
2201 optim->pulse_pos[k] = param.pulse_pos[k];
2209 * Encode the pulse position and gain of the current subframe.
2211 * @param optim optimized fixed CB parameters
2212 * @param buf excitation vector
2214 static void pack_fcb_param(G723_1_Subframe *subfrm, FCBParam *optim,
2215 int16_t *buf, int pulse_cnt)
2219 j = PULSE_MAX - pulse_cnt;
2221 subfrm->pulse_sign = 0;
2222 subfrm->pulse_pos = 0;
2224 for (i = 0; i < SUBFRAME_LEN >> 1; i++) {
2225 int val = buf[optim->grid_index + (i << 1)];
2227 subfrm->pulse_pos += combinatorial_table[j][i];
2229 subfrm->pulse_sign <<= 1;
2230 if (val < 0) subfrm->pulse_sign++;
2233 if (j == PULSE_MAX) break;
2236 subfrm->amp_index = optim->amp_index;
2237 subfrm->grid_index = optim->grid_index;
2238 subfrm->dirac_train = optim->dirac_train;
2242 * Compute the fixed codebook excitation.
2244 * @param buf target vector
2245 * @param impulse_resp impulse response of the combined filter
2247 static void fcb_search(G723_1_Context *p, int16_t *impulse_resp,
2248 int16_t *buf, int index)
2251 int pulse_cnt = pulses[index];
2254 optim.min_err = 1 << 30;
2255 get_fcb_param(&optim, impulse_resp, buf, pulse_cnt, SUBFRAME_LEN);
2257 if (p->pitch_lag[index >> 1] < SUBFRAME_LEN - 2) {
2258 get_fcb_param(&optim, impulse_resp, buf, pulse_cnt,
2259 p->pitch_lag[index >> 1]);
2262 /* Reconstruct the excitation */
2263 memset(buf, 0, sizeof(int16_t) * SUBFRAME_LEN);
2264 for (i = 0; i < pulse_cnt; i++)
2265 buf[optim.pulse_pos[i]] = optim.pulse_sign[i];
2267 pack_fcb_param(&p->subframe[index], &optim, buf, pulse_cnt);
2269 if (optim.dirac_train)
2270 gen_dirac_train(buf, p->pitch_lag[index >> 1]);
2274 * Pack the frame parameters into output bitstream.
2276 * @param frame output buffer
2277 * @param size size of the buffer
2279 static int pack_bitstream(G723_1_Context *p, unsigned char *frame, int size)
2282 int info_bits, i, temp;
2284 init_put_bits(&pb, frame, size);
2286 if (p->cur_rate == RATE_6300) {
2288 put_bits(&pb, 2, info_bits);
2291 put_bits(&pb, 8, p->lsp_index[2]);
2292 put_bits(&pb, 8, p->lsp_index[1]);
2293 put_bits(&pb, 8, p->lsp_index[0]);
2295 put_bits(&pb, 7, p->pitch_lag[0] - PITCH_MIN);
2296 put_bits(&pb, 2, p->subframe[1].ad_cb_lag);
2297 put_bits(&pb, 7, p->pitch_lag[1] - PITCH_MIN);
2298 put_bits(&pb, 2, p->subframe[3].ad_cb_lag);
2300 /* Write 12 bit combined gain */
2301 for (i = 0; i < SUBFRAMES; i++) {
2302 temp = p->subframe[i].ad_cb_gain * GAIN_LEVELS +
2303 p->subframe[i].amp_index;
2304 if (p->cur_rate == RATE_6300)
2305 temp += p->subframe[i].dirac_train << 11;
2306 put_bits(&pb, 12, temp);
2309 put_bits(&pb, 1, p->subframe[0].grid_index);
2310 put_bits(&pb, 1, p->subframe[1].grid_index);
2311 put_bits(&pb, 1, p->subframe[2].grid_index);
2312 put_bits(&pb, 1, p->subframe[3].grid_index);
2314 if (p->cur_rate == RATE_6300) {
2315 skip_put_bits(&pb, 1); /* reserved bit */
2317 /* Write 13 bit combined position index */
2318 temp = (p->subframe[0].pulse_pos >> 16) * 810 +
2319 (p->subframe[1].pulse_pos >> 14) * 90 +
2320 (p->subframe[2].pulse_pos >> 16) * 9 +
2321 (p->subframe[3].pulse_pos >> 14);
2322 put_bits(&pb, 13, temp);
2324 put_bits(&pb, 16, p->subframe[0].pulse_pos & 0xffff);
2325 put_bits(&pb, 14, p->subframe[1].pulse_pos & 0x3fff);
2326 put_bits(&pb, 16, p->subframe[2].pulse_pos & 0xffff);
2327 put_bits(&pb, 14, p->subframe[3].pulse_pos & 0x3fff);
2329 put_bits(&pb, 6, p->subframe[0].pulse_sign);
2330 put_bits(&pb, 5, p->subframe[1].pulse_sign);
2331 put_bits(&pb, 6, p->subframe[2].pulse_sign);
2332 put_bits(&pb, 5, p->subframe[3].pulse_sign);
2335 flush_put_bits(&pb);
2336 return frame_size[info_bits];
2339 static int g723_1_encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
2340 const AVFrame *frame, int *got_packet_ptr)
2342 G723_1_Context *p = avctx->priv_data;
2343 int16_t unq_lpc[LPC_ORDER * SUBFRAMES];
2344 int16_t qnt_lpc[LPC_ORDER * SUBFRAMES];
2345 int16_t cur_lsp[LPC_ORDER];
2346 int16_t weighted_lpc[LPC_ORDER * SUBFRAMES << 1];
2347 int16_t vector[FRAME_LEN + PITCH_MAX];
2349 int16_t *in = (const int16_t *)frame->data[0];
2354 highpass_filter(in, &p->hpf_fir_mem, &p->hpf_iir_mem);
2356 memcpy(vector, p->prev_data, HALF_FRAME_LEN * sizeof(int16_t));
2357 memcpy(vector + HALF_FRAME_LEN, in, FRAME_LEN * sizeof(int16_t));
2359 comp_lpc_coeff(vector, unq_lpc);
2360 lpc2lsp(&unq_lpc[LPC_ORDER * 3], p->prev_lsp, cur_lsp);
2361 lsp_quantize(p->lsp_index, cur_lsp, p->prev_lsp);
2364 memcpy(vector + LPC_ORDER, p->prev_data + SUBFRAME_LEN,
2365 sizeof(int16_t) * SUBFRAME_LEN);
2366 memcpy(vector + LPC_ORDER + SUBFRAME_LEN, in,
2367 sizeof(int16_t) * (HALF_FRAME_LEN + SUBFRAME_LEN));
2368 memcpy(p->prev_data, in + HALF_FRAME_LEN,
2369 sizeof(int16_t) * HALF_FRAME_LEN);
2370 memcpy(in, vector + LPC_ORDER, sizeof(int16_t) * FRAME_LEN);
2372 perceptual_filter(p, weighted_lpc, unq_lpc, vector);
2374 memcpy(in, vector + LPC_ORDER, sizeof(int16_t) * FRAME_LEN);
2375 memcpy(vector, p->prev_weight_sig, sizeof(int16_t) * PITCH_MAX);
2376 memcpy(vector + PITCH_MAX, in, sizeof(int16_t) * FRAME_LEN);
2378 scale_vector(vector, vector, FRAME_LEN + PITCH_MAX);
2380 p->pitch_lag[0] = estimate_pitch(vector, PITCH_MAX);
2381 p->pitch_lag[1] = estimate_pitch(vector, PITCH_MAX + HALF_FRAME_LEN);
2383 for (i = PITCH_MAX, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
2384 comp_harmonic_coeff(vector + i, p->pitch_lag[j >> 1], hf + j);
2386 memcpy(vector, p->prev_weight_sig, sizeof(int16_t) * PITCH_MAX);
2387 memcpy(vector + PITCH_MAX, in, sizeof(int16_t) * FRAME_LEN);
2388 memcpy(p->prev_weight_sig, vector + FRAME_LEN, sizeof(int16_t) * PITCH_MAX);
2390 for (i = 0, j = 0; j < SUBFRAMES; i += SUBFRAME_LEN, j++)
2391 harmonic_filter(hf + j, vector + PITCH_MAX + i, in + i);
2393 inverse_quant(cur_lsp, p->prev_lsp, p->lsp_index, 0);
2394 lsp_interpolate(qnt_lpc, cur_lsp, p->prev_lsp);
2396 memcpy(p->prev_lsp, cur_lsp, sizeof(int16_t) * LPC_ORDER);
2399 for (i = 0; i < SUBFRAMES; i++) {
2400 int16_t impulse_resp[SUBFRAME_LEN];
2401 int16_t residual[SUBFRAME_LEN + PITCH_ORDER - 1];
2402 int16_t flt_in[SUBFRAME_LEN];
2403 int16_t zero[LPC_ORDER], fir[LPC_ORDER], iir[LPC_ORDER];
2406 * Compute the combined impulse response of the synthesis filter,
2407 * formant perceptual weighting filter and harmonic noise shaping filter
2409 memset(zero, 0, sizeof(int16_t) * LPC_ORDER);
2410 memset(vector, 0, sizeof(int16_t) * PITCH_MAX);
2411 memset(flt_in, 0, sizeof(int16_t) * SUBFRAME_LEN);
2413 flt_in[0] = 1 << 13; /* Unit impulse */
2414 synth_percept_filter(qnt_lpc + offset, weighted_lpc + (offset << 1),
2415 zero, zero, flt_in, vector + PITCH_MAX, 1);
2416 harmonic_filter(hf + i, vector + PITCH_MAX, impulse_resp);
2418 /* Compute the combined zero input response */
2420 memcpy(fir, p->perf_fir_mem, sizeof(int16_t) * LPC_ORDER);
2421 memcpy(iir, p->perf_iir_mem, sizeof(int16_t) * LPC_ORDER);
2423 synth_percept_filter(qnt_lpc + offset, weighted_lpc + (offset << 1),
2424 fir, iir, flt_in, vector + PITCH_MAX, 0);
2425 memcpy(vector, p->harmonic_mem, sizeof(int16_t) * PITCH_MAX);
2426 harmonic_noise_sub(hf + i, vector + PITCH_MAX, in);
2428 acb_search(p, residual, impulse_resp, in, i);
2429 gen_acb_excitation(residual, p->prev_excitation,p->pitch_lag[i >> 1],
2430 &p->subframe[i], p->cur_rate);
2431 sub_acb_contrib(residual, impulse_resp, in);
2433 fcb_search(p, impulse_resp, in, i);
2435 /* Reconstruct the excitation */
2436 gen_acb_excitation(impulse_resp, p->prev_excitation, p->pitch_lag[i >> 1],
2437 &p->subframe[i], RATE_6300);
2439 memmove(p->prev_excitation, p->prev_excitation + SUBFRAME_LEN,
2440 sizeof(int16_t) * (PITCH_MAX - SUBFRAME_LEN));
2441 for (j = 0; j < SUBFRAME_LEN; j++)
2442 in[j] = av_clip_int16((in[j] << 1) + impulse_resp[j]);
2443 memcpy(p->prev_excitation + PITCH_MAX - SUBFRAME_LEN, in,
2444 sizeof(int16_t) * SUBFRAME_LEN);
2446 /* Update filter memories */
2447 synth_percept_filter(qnt_lpc + offset, weighted_lpc + (offset << 1),
2448 p->perf_fir_mem, p->perf_iir_mem,
2449 in, vector + PITCH_MAX, 0);
2450 memmove(p->harmonic_mem, p->harmonic_mem + SUBFRAME_LEN,
2451 sizeof(int16_t) * (PITCH_MAX - SUBFRAME_LEN));
2452 memcpy(p->harmonic_mem + PITCH_MAX - SUBFRAME_LEN, vector + PITCH_MAX,
2453 sizeof(int16_t) * SUBFRAME_LEN);
2456 offset += LPC_ORDER;
2459 if ((ret = ff_alloc_packet2(avctx, avpkt, 24)) < 0)
2462 *got_packet_ptr = 1;
2463 avpkt->size = pack_bitstream(p, avpkt->data, avpkt->size);
2467 AVCodec ff_g723_1_encoder = {
2469 .type = AVMEDIA_TYPE_AUDIO,
2470 .id = AV_CODEC_ID_G723_1,
2471 .priv_data_size = sizeof(G723_1_Context),
2472 .init = g723_1_encode_init,
2473 .encode2 = g723_1_encode_frame,
2474 .long_name = NULL_IF_CONFIG_SMALL("G.723.1"),
2475 .sample_fmts = (const enum AVSampleFormat[]){AV_SAMPLE_FMT_S16,
2476 AV_SAMPLE_FMT_NONE},