3 * Copyright (C) 2008-2012 Alexander E. Patrakov
4 * 2010 Benjamin Larsson
7 * This file is part of FFmpeg.
9 * FFmpeg is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License as published by the Free Software Foundation; either
12 * version 2.1 of the License, or (at your option) any later version.
14 * FFmpeg is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with FFmpeg; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
25 #define FFT_FIXED_32 1
27 #include "libavutil/avassert.h"
28 #include "libavutil/channel_layout.h"
29 #include "libavutil/common.h"
30 #include "libavutil/ffmath.h"
31 #include "libavutil/mem_internal.h"
32 #include "libavutil/opt.h"
45 #define MAX_CHANNELS 6
46 #define DCA_MAX_FRAME_SIZE 16384
47 #define DCA_HEADER_SIZE 13
48 #define DCA_LFE_SAMPLES 8
50 #define DCAENC_SUBBANDS 32
52 #define SUBSUBFRAMES 2
53 #define SUBBAND_SAMPLES (SUBFRAMES * SUBSUBFRAMES * 8)
56 #define COS_T(x) (c->cos_table[(x) & 2047])
58 typedef struct CompressionOptions {
62 typedef struct DCAEncContext {
65 DCAADPCMEncContext adpcm_ctx;
67 CompressionOptions options;
70 int fullband_channels;
76 const int32_t *band_interpolation;
77 const int32_t *band_spectrum;
81 const int8_t *channel_order_tab; ///< channel reordering table, lfe and non lfe
83 int32_t prediction_mode[MAX_CHANNELS][DCAENC_SUBBANDS];
84 int32_t adpcm_history[MAX_CHANNELS][DCAENC_SUBBANDS][DCA_ADPCM_COEFFS * 2];
85 int32_t history[MAX_CHANNELS][512]; /* This is a circular buffer */
86 int32_t *subband[MAX_CHANNELS][DCAENC_SUBBANDS];
87 int32_t quantized[MAX_CHANNELS][DCAENC_SUBBANDS][SUBBAND_SAMPLES];
88 int32_t peak_cb[MAX_CHANNELS][DCAENC_SUBBANDS];
89 int32_t diff_peak_cb[MAX_CHANNELS][DCAENC_SUBBANDS]; ///< expected peak of residual signal
90 int32_t downsampled_lfe[DCA_LFE_SAMPLES];
91 int32_t masking_curve_cb[SUBSUBFRAMES][256];
92 int32_t bit_allocation_sel[MAX_CHANNELS];
93 int abits[MAX_CHANNELS][DCAENC_SUBBANDS];
94 int scale_factor[MAX_CHANNELS][DCAENC_SUBBANDS];
95 softfloat quant[MAX_CHANNELS][DCAENC_SUBBANDS];
96 int32_t quant_index_sel[MAX_CHANNELS][DCA_CODE_BOOKS];
97 int32_t eff_masking_curve_cb[256];
98 int32_t band_masking_cb[32];
99 int32_t worst_quantization_noise;
100 int32_t worst_noise_ever;
102 int consumed_adpcm_bits; ///< Number of bits to transmit ADPCM related info
104 int32_t cos_table[2048];
105 int32_t band_interpolation_tab[2][512];
106 int32_t band_spectrum_tab[2][8];
107 int32_t auf[9][AUBANDS][256];
108 int32_t cb_to_add[256];
109 int32_t cb_to_level[2048];
110 int32_t lfe_fir_64i[512];
113 /* Transfer function of outer and middle ear, Hz -> dB */
114 static double hom(double f)
116 double f1 = f / 1000;
118 return -3.64 * pow(f1, -0.8)
119 + 6.8 * exp(-0.6 * (f1 - 3.4) * (f1 - 3.4))
120 - 6.0 * exp(-0.15 * (f1 - 8.7) * (f1 - 8.7))
121 - 0.0006 * (f1 * f1) * (f1 * f1);
124 static double gammafilter(int i, double f)
126 double h = (f - fc[i]) / erb[i];
130 return 20 * log10(h);
133 static int subband_bufer_alloc(DCAEncContext *c)
136 int32_t *bufer = av_calloc(MAX_CHANNELS * DCAENC_SUBBANDS *
137 (SUBBAND_SAMPLES + DCA_ADPCM_COEFFS),
140 return AVERROR(ENOMEM);
142 /* we need a place for DCA_ADPCM_COEFF samples from previous frame
143 * to calc prediction coefficients for each subband */
144 for (ch = 0; ch < MAX_CHANNELS; ch++) {
145 for (band = 0; band < DCAENC_SUBBANDS; band++) {
146 c->subband[ch][band] = bufer +
147 ch * DCAENC_SUBBANDS * (SUBBAND_SAMPLES + DCA_ADPCM_COEFFS) +
148 band * (SUBBAND_SAMPLES + DCA_ADPCM_COEFFS) + DCA_ADPCM_COEFFS;
154 static void subband_bufer_free(DCAEncContext *c)
156 if (c->subband[0][0]) {
157 int32_t *bufer = c->subband[0][0] - DCA_ADPCM_COEFFS;
159 c->subband[0][0] = NULL;
163 static int encode_init(AVCodecContext *avctx)
165 DCAEncContext *c = avctx->priv_data;
166 uint64_t layout = avctx->channel_layout;
167 int i, j, k, min_frame_bits;
170 if ((ret = subband_bufer_alloc(c)) < 0)
173 c->fullband_channels = c->channels = avctx->channels;
174 c->lfe_channel = (avctx->channels == 3 || avctx->channels == 6);
175 c->band_interpolation = c->band_interpolation_tab[1];
176 c->band_spectrum = c->band_spectrum_tab[1];
177 c->worst_quantization_noise = -2047;
178 c->worst_noise_ever = -2047;
179 c->consumed_adpcm_bits = 0;
181 if (ff_dcaadpcm_init(&c->adpcm_ctx))
182 return AVERROR(ENOMEM);
185 av_log(avctx, AV_LOG_WARNING, "No channel layout specified. The "
186 "encoder will guess the layout, but it "
187 "might be incorrect.\n");
188 layout = av_get_default_channel_layout(avctx->channels);
191 case AV_CH_LAYOUT_MONO: c->channel_config = 0; break;
192 case AV_CH_LAYOUT_STEREO: c->channel_config = 2; break;
193 case AV_CH_LAYOUT_2_2: c->channel_config = 8; break;
194 case AV_CH_LAYOUT_5POINT0: c->channel_config = 9; break;
195 case AV_CH_LAYOUT_5POINT1: c->channel_config = 9; break;
197 av_log(avctx, AV_LOG_ERROR, "Unsupported channel layout!\n");
198 return AVERROR_PATCHWELCOME;
201 if (c->lfe_channel) {
202 c->fullband_channels--;
203 c->channel_order_tab = channel_reorder_lfe[c->channel_config];
205 c->channel_order_tab = channel_reorder_nolfe[c->channel_config];
208 for (i = 0; i < MAX_CHANNELS; i++) {
209 for (j = 0; j < DCA_CODE_BOOKS; j++) {
210 c->quant_index_sel[i][j] = ff_dca_quant_index_group_size[j];
213 c->bit_allocation_sel[i] = 6;
215 for (j = 0; j < DCAENC_SUBBANDS; j++) {
217 c->prediction_mode[i][j] = -1;
218 memset(c->adpcm_history[i][j], 0, sizeof(int32_t)*DCA_ADPCM_COEFFS);
222 for (i = 0; i < 9; i++) {
223 if (sample_rates[i] == avctx->sample_rate)
227 return AVERROR(EINVAL);
228 c->samplerate_index = i;
230 if (avctx->bit_rate < 32000 || avctx->bit_rate > 3840000) {
231 av_log(avctx, AV_LOG_ERROR, "Bit rate %"PRId64" not supported.", avctx->bit_rate);
232 return AVERROR(EINVAL);
234 for (i = 0; ff_dca_bit_rates[i] < avctx->bit_rate; i++)
236 c->bitrate_index = i;
237 c->frame_bits = FFALIGN((avctx->bit_rate * 512 + avctx->sample_rate - 1) / avctx->sample_rate, 32);
238 min_frame_bits = 132 + (493 + 28 * 32) * c->fullband_channels + c->lfe_channel * 72;
239 if (c->frame_bits < min_frame_bits || c->frame_bits > (DCA_MAX_FRAME_SIZE << 3))
240 return AVERROR(EINVAL);
242 c->frame_size = (c->frame_bits + 7) / 8;
244 avctx->frame_size = 32 * SUBBAND_SAMPLES;
246 if ((ret = ff_mdct_init(&c->mdct, 9, 0, 1.0)) < 0)
249 /* Init all tables */
250 c->cos_table[0] = 0x7fffffff;
251 c->cos_table[512] = 0;
252 c->cos_table[1024] = -c->cos_table[0];
253 for (i = 1; i < 512; i++) {
254 c->cos_table[i] = (int32_t)(0x7fffffff * cos(M_PI * i / 1024));
255 c->cos_table[1024-i] = -c->cos_table[i];
256 c->cos_table[1024+i] = -c->cos_table[i];
257 c->cos_table[2048-i] = +c->cos_table[i];
260 for (i = 0; i < 2048; i++)
261 c->cb_to_level[i] = (int32_t)(0x7fffffff * ff_exp10(-0.005 * i));
263 for (k = 0; k < 32; k++) {
264 for (j = 0; j < 8; j++) {
265 c->lfe_fir_64i[64 * j + k] = (int32_t)(0xffffff800000ULL * ff_dca_lfe_fir_64[8 * k + j]);
266 c->lfe_fir_64i[64 * (7-j) + (63 - k)] = (int32_t)(0xffffff800000ULL * ff_dca_lfe_fir_64[8 * k + j]);
270 for (i = 0; i < 512; i++) {
271 c->band_interpolation_tab[0][i] = (int32_t)(0x1000000000ULL * ff_dca_fir_32bands_perfect[i]);
272 c->band_interpolation_tab[1][i] = (int32_t)(0x1000000000ULL * ff_dca_fir_32bands_nonperfect[i]);
275 for (i = 0; i < 9; i++) {
276 for (j = 0; j < AUBANDS; j++) {
277 for (k = 0; k < 256; k++) {
278 double freq = sample_rates[i] * (k + 0.5) / 512;
280 c->auf[i][j][k] = (int32_t)(10 * (hom(freq) + gammafilter(j, freq)));
285 for (i = 0; i < 256; i++) {
286 double add = 1 + ff_exp10(-0.01 * i);
287 c->cb_to_add[i] = (int32_t)(100 * log10(add));
289 for (j = 0; j < 8; j++) {
291 for (i = 0; i < 512; i++) {
292 double reconst = ff_dca_fir_32bands_perfect[i] * ((i & 64) ? (-1) : 1);
293 accum += reconst * cos(2 * M_PI * (i + 0.5 - 256) * (j + 0.5) / 512);
295 c->band_spectrum_tab[0][j] = (int32_t)(200 * log10(accum));
297 for (j = 0; j < 8; j++) {
299 for (i = 0; i < 512; i++) {
300 double reconst = ff_dca_fir_32bands_nonperfect[i] * ((i & 64) ? (-1) : 1);
301 accum += reconst * cos(2 * M_PI * (i + 0.5 - 256) * (j + 0.5) / 512);
303 c->band_spectrum_tab[1][j] = (int32_t)(200 * log10(accum));
309 static av_cold int encode_close(AVCodecContext *avctx)
311 DCAEncContext *c = avctx->priv_data;
312 ff_mdct_end(&c->mdct);
313 subband_bufer_free(c);
314 ff_dcaadpcm_free(&c->adpcm_ctx);
319 static void subband_transform(DCAEncContext *c, const int32_t *input)
321 int ch, subs, i, k, j;
323 for (ch = 0; ch < c->fullband_channels; ch++) {
324 /* History is copied because it is also needed for PSY */
327 const int chi = c->channel_order_tab[ch];
329 memcpy(hist, &c->history[ch][0], 512 * sizeof(int32_t));
331 for (subs = 0; subs < SUBBAND_SAMPLES; subs++) {
336 /* Calculate the convolutions at once */
337 memset(accum, 0, 64 * sizeof(int32_t));
339 for (k = 0, i = hist_start, j = 0;
340 i < 512; k = (k + 1) & 63, i++, j++)
341 accum[k] += mul32(hist[i], c->band_interpolation[j]);
342 for (i = 0; i < hist_start; k = (k + 1) & 63, i++, j++)
343 accum[k] += mul32(hist[i], c->band_interpolation[j]);
345 for (k = 16; k < 32; k++)
346 accum[k] = accum[k] - accum[31 - k];
347 for (k = 32; k < 48; k++)
348 accum[k] = accum[k] + accum[95 - k];
350 for (band = 0; band < 32; band++) {
352 for (i = 16; i < 48; i++) {
353 int s = (2 * band + 1) * (2 * (i + 16) + 1);
354 resp += mul32(accum[i], COS_T(s << 3)) >> 3;
357 c->subband[ch][band][subs] = ((band + 1) & 2) ? -resp : resp;
360 /* Copy in 32 new samples from input */
361 for (i = 0; i < 32; i++)
362 hist[i + hist_start] = input[(subs * 32 + i) * c->channels + chi];
364 hist_start = (hist_start + 32) & 511;
369 static void lfe_downsample(DCAEncContext *c, const int32_t *input)
371 /* FIXME: make 128x LFE downsampling possible */
372 const int lfech = lfe_index[c->channel_config];
378 memcpy(hist, &c->history[c->channels - 1][0], 512 * sizeof(int32_t));
380 for (lfes = 0; lfes < DCA_LFE_SAMPLES; lfes++) {
381 /* Calculate the convolution */
384 for (i = hist_start, j = 0; i < 512; i++, j++)
385 accum += mul32(hist[i], c->lfe_fir_64i[j]);
386 for (i = 0; i < hist_start; i++, j++)
387 accum += mul32(hist[i], c->lfe_fir_64i[j]);
389 c->downsampled_lfe[lfes] = accum;
391 /* Copy in 64 new samples from input */
392 for (i = 0; i < 64; i++)
393 hist[i + hist_start] = input[(lfes * 64 + i) * c->channels + lfech];
395 hist_start = (hist_start + 64) & 511;
399 static int32_t get_cb(DCAEncContext *c, int32_t in)
404 for (i = 1024; i > 0; i >>= 1) {
405 if (c->cb_to_level[i + res] >= in)
411 static int32_t add_cb(DCAEncContext *c, int32_t a, int32_t b)
414 FFSWAP(int32_t, a, b);
418 return a + c->cb_to_add[a - b];
421 static void calc_power(DCAEncContext *c,
422 const int32_t in[2 * 256], int32_t power[256])
425 LOCAL_ALIGNED_32(int32_t, data, [512]);
426 LOCAL_ALIGNED_32(int32_t, coeff, [256]);
428 for (i = 0; i < 512; i++)
429 data[i] = norm__(mul32(in[i], 0x3fffffff - (COS_T(4 * i + 2) >> 1)), 4);
431 c->mdct.mdct_calc(&c->mdct, coeff, data);
432 for (i = 0; i < 256; i++) {
433 const int32_t cb = get_cb(c, coeff[i]);
434 power[i] = add_cb(c, cb, cb);
438 static void adjust_jnd(DCAEncContext *c,
439 const int32_t in[512], int32_t out_cb[256])
442 int32_t out_cb_unnorm[256];
444 const int32_t ca_cb = -1114;
445 const int32_t cs_cb = 928;
446 const int samplerate_index = c->samplerate_index;
449 calc_power(c, in, power);
451 for (j = 0; j < 256; j++)
452 out_cb_unnorm[j] = -2047; /* and can only grow */
454 for (i = 0; i < AUBANDS; i++) {
455 denom = ca_cb; /* and can only grow */
456 for (j = 0; j < 256; j++)
457 denom = add_cb(c, denom, power[j] + c->auf[samplerate_index][i][j]);
458 for (j = 0; j < 256; j++)
459 out_cb_unnorm[j] = add_cb(c, out_cb_unnorm[j],
460 -denom + c->auf[samplerate_index][i][j]);
463 for (j = 0; j < 256; j++)
464 out_cb[j] = add_cb(c, out_cb[j], -out_cb_unnorm[j] - ca_cb - cs_cb);
467 typedef void (*walk_band_t)(DCAEncContext *c, int band1, int band2, int f,
468 int32_t spectrum1, int32_t spectrum2, int channel,
471 static void walk_band_low(DCAEncContext *c, int band, int channel,
472 walk_band_t walk, int32_t *arg)
477 for (f = 0; f < 4; f++)
478 walk(c, 0, 0, f, 0, -2047, channel, arg);
480 for (f = 0; f < 8; f++)
481 walk(c, band, band - 1, 8 * band - 4 + f,
482 c->band_spectrum[7 - f], c->band_spectrum[f], channel, arg);
486 static void walk_band_high(DCAEncContext *c, int band, int channel,
487 walk_band_t walk, int32_t *arg)
492 for (f = 0; f < 4; f++)
493 walk(c, 31, 31, 256 - 4 + f, 0, -2047, channel, arg);
495 for (f = 0; f < 8; f++)
496 walk(c, band, band + 1, 8 * band + 4 + f,
497 c->band_spectrum[f], c->band_spectrum[7 - f], channel, arg);
501 static void update_band_masking(DCAEncContext *c, int band1, int band2,
502 int f, int32_t spectrum1, int32_t spectrum2,
503 int channel, int32_t * arg)
505 int32_t value = c->eff_masking_curve_cb[f] - spectrum1;
507 if (value < c->band_masking_cb[band1])
508 c->band_masking_cb[band1] = value;
511 static void calc_masking(DCAEncContext *c, const int32_t *input)
513 int i, k, band, ch, ssf;
516 for (i = 0; i < 256; i++)
517 for (ssf = 0; ssf < SUBSUBFRAMES; ssf++)
518 c->masking_curve_cb[ssf][i] = -2047;
520 for (ssf = 0; ssf < SUBSUBFRAMES; ssf++)
521 for (ch = 0; ch < c->fullband_channels; ch++) {
522 const int chi = c->channel_order_tab[ch];
524 for (i = 0, k = 128 + 256 * ssf; k < 512; i++, k++)
525 data[i] = c->history[ch][k];
526 for (k -= 512; i < 512; i++, k++)
527 data[i] = input[k * c->channels + chi];
528 adjust_jnd(c, data, c->masking_curve_cb[ssf]);
530 for (i = 0; i < 256; i++) {
533 for (ssf = 0; ssf < SUBSUBFRAMES; ssf++)
534 if (c->masking_curve_cb[ssf][i] < m)
535 m = c->masking_curve_cb[ssf][i];
536 c->eff_masking_curve_cb[i] = m;
539 for (band = 0; band < 32; band++) {
540 c->band_masking_cb[band] = 2048;
541 walk_band_low(c, band, 0, update_band_masking, NULL);
542 walk_band_high(c, band, 0, update_band_masking, NULL);
546 static inline int32_t find_peak(DCAEncContext *c, const int32_t *in, int len)
550 for (sample = 0; sample < len; sample++) {
551 int32_t s = abs(in[sample]);
558 static void find_peaks(DCAEncContext *c)
562 for (ch = 0; ch < c->fullband_channels; ch++) {
563 for (band = 0; band < 32; band++)
564 c->peak_cb[ch][band] = find_peak(c, c->subband[ch][band],
569 c->lfe_peak_cb = find_peak(c, c->downsampled_lfe, DCA_LFE_SAMPLES);
572 static void adpcm_analysis(DCAEncContext *c)
577 int32_t estimated_diff[SUBBAND_SAMPLES];
579 c->consumed_adpcm_bits = 0;
580 for (ch = 0; ch < c->fullband_channels; ch++) {
581 for (band = 0; band < 32; band++) {
582 samples = c->subband[ch][band] - DCA_ADPCM_COEFFS;
583 pred_vq_id = ff_dcaadpcm_subband_analysis(&c->adpcm_ctx, samples,
584 SUBBAND_SAMPLES, estimated_diff);
585 if (pred_vq_id >= 0) {
586 c->prediction_mode[ch][band] = pred_vq_id;
587 c->consumed_adpcm_bits += 12; //12 bits to transmit prediction vq index
588 c->diff_peak_cb[ch][band] = find_peak(c, estimated_diff, 16);
590 c->prediction_mode[ch][band] = -1;
596 static const int snr_fudge = 128;
597 #define USED_1ABITS 1
598 #define USED_26ABITS 4
600 static inline int32_t get_step_size(DCAEncContext *c, int ch, int band)
604 if (c->bitrate_index == 3)
605 step_size = ff_dca_lossless_quant[c->abits[ch][band]];
607 step_size = ff_dca_lossy_quant[c->abits[ch][band]];
612 static int calc_one_scale(DCAEncContext *c, int32_t peak_cb, int abits,
616 int our_nscale, try_remove;
619 av_assert0(peak_cb <= 0);
620 av_assert0(peak_cb >= -2047);
623 peak = c->cb_to_level[-peak_cb];
625 for (try_remove = 64; try_remove > 0; try_remove >>= 1) {
626 if (scalefactor_inv[our_nscale - try_remove].e + stepsize_inv[abits].e <= 17)
628 our_quant.m = mul32(scalefactor_inv[our_nscale - try_remove].m, stepsize_inv[abits].m);
629 our_quant.e = scalefactor_inv[our_nscale - try_remove].e + stepsize_inv[abits].e - 17;
630 if ((ff_dca_quant_levels[abits] - 1) / 2 < quantize_value(peak, our_quant))
632 our_nscale -= try_remove;
635 if (our_nscale >= 125)
638 quant->m = mul32(scalefactor_inv[our_nscale].m, stepsize_inv[abits].m);
639 quant->e = scalefactor_inv[our_nscale].e + stepsize_inv[abits].e - 17;
640 av_assert0((ff_dca_quant_levels[abits] - 1) / 2 >= quantize_value(peak, *quant));
645 static inline void quantize_adpcm_subband(DCAEncContext *c, int ch, int band)
648 int32_t diff_peak_cb = c->diff_peak_cb[ch][band];
649 c->scale_factor[ch][band] = calc_one_scale(c, diff_peak_cb,
651 &c->quant[ch][band]);
653 step_size = get_step_size(c, ch, band);
654 ff_dcaadpcm_do_real(c->prediction_mode[ch][band],
656 ff_dca_scale_factor_quant7[c->scale_factor[ch][band]],
657 step_size, c->adpcm_history[ch][band], c->subband[ch][band],
658 c->adpcm_history[ch][band] + 4, c->quantized[ch][band],
659 SUBBAND_SAMPLES, c->cb_to_level[-diff_peak_cb]);
662 static void quantize_adpcm(DCAEncContext *c)
666 for (ch = 0; ch < c->fullband_channels; ch++)
667 for (band = 0; band < 32; band++)
668 if (c->prediction_mode[ch][band] >= 0)
669 quantize_adpcm_subband(c, ch, band);
672 static void quantize_pcm(DCAEncContext *c)
674 int sample, band, ch;
676 for (ch = 0; ch < c->fullband_channels; ch++) {
677 for (band = 0; band < 32; band++) {
678 if (c->prediction_mode[ch][band] == -1) {
679 for (sample = 0; sample < SUBBAND_SAMPLES; sample++) {
680 int32_t val = quantize_value(c->subband[ch][band][sample],
682 c->quantized[ch][band][sample] = val;
689 static void accumulate_huff_bit_consumption(int abits, int32_t *quantized,
692 uint8_t sel, id = abits - 1;
693 for (sel = 0; sel < ff_dca_quant_index_group_size[id]; sel++)
694 result[sel] += ff_dca_vlc_calc_quant_bits(quantized, SUBBAND_SAMPLES,
698 static uint32_t set_best_code(uint32_t vlc_bits[DCA_CODE_BOOKS][7],
699 uint32_t clc_bits[DCA_CODE_BOOKS],
700 int32_t res[DCA_CODE_BOOKS])
703 uint32_t best_sel_bits[DCA_CODE_BOOKS];
704 int32_t best_sel_id[DCA_CODE_BOOKS];
705 uint32_t t, bits = 0;
707 for (i = 0; i < DCA_CODE_BOOKS; i++) {
709 av_assert0(!((!!vlc_bits[i][0]) ^ (!!clc_bits[i])));
710 if (vlc_bits[i][0] == 0) {
711 /* do not transmit adjustment index for empty codebooks */
712 res[i] = ff_dca_quant_index_group_size[i];
717 best_sel_bits[i] = vlc_bits[i][0];
719 for (sel = 0; sel < ff_dca_quant_index_group_size[i]; sel++) {
720 if (best_sel_bits[i] > vlc_bits[i][sel] && vlc_bits[i][sel]) {
721 best_sel_bits[i] = vlc_bits[i][sel];
722 best_sel_id[i] = sel;
726 /* 2 bits to transmit scale factor adjustment index */
727 t = best_sel_bits[i] + 2;
728 if (t < clc_bits[i]) {
729 res[i] = best_sel_id[i];
732 res[i] = ff_dca_quant_index_group_size[i];
739 static uint32_t set_best_abits_code(int abits[DCAENC_SUBBANDS], int bands,
744 int32_t best_sel = 6;
745 int32_t best_bits = bands * 5;
747 /* Check do we have subband which cannot be encoded by Huffman tables */
748 for (i = 0; i < bands; i++) {
749 if (abits[i] > 12 || abits[i] == 0) {
755 for (i = 0; i < DCA_BITALLOC_12_COUNT; i++) {
756 t = ff_dca_vlc_calc_alloc_bits(abits, bands, i);
767 static int init_quantization_noise(DCAEncContext *c, int noise, int forbid_zero)
769 int ch, band, ret = USED_26ABITS | USED_1ABITS;
770 uint32_t huff_bit_count_accum[MAX_CHANNELS][DCA_CODE_BOOKS][7];
771 uint32_t clc_bit_count_accum[MAX_CHANNELS][DCA_CODE_BOOKS];
772 uint32_t bits_counter = 0;
774 c->consumed_bits = 132 + 333 * c->fullband_channels;
775 c->consumed_bits += c->consumed_adpcm_bits;
777 c->consumed_bits += 72;
779 /* attempt to guess the bit distribution based on the prevoius frame */
780 for (ch = 0; ch < c->fullband_channels; ch++) {
781 for (band = 0; band < 32; band++) {
782 int snr_cb = c->peak_cb[ch][band] - c->band_masking_cb[band] - noise;
784 if (snr_cb >= 1312) {
785 c->abits[ch][band] = 26;
787 } else if (snr_cb >= 222) {
788 c->abits[ch][band] = 8 + mul32(snr_cb - 222, 69000000);
789 ret &= ~(USED_26ABITS | USED_1ABITS);
790 } else if (snr_cb >= 0) {
791 c->abits[ch][band] = 2 + mul32(snr_cb, 106000000);
792 ret &= ~(USED_26ABITS | USED_1ABITS);
793 } else if (forbid_zero || snr_cb >= -140) {
794 c->abits[ch][band] = 1;
795 ret &= ~USED_26ABITS;
797 c->abits[ch][band] = 0;
798 ret &= ~(USED_26ABITS | USED_1ABITS);
801 c->consumed_bits += set_best_abits_code(c->abits[ch], 32,
802 &c->bit_allocation_sel[ch]);
805 /* Recalc scale_factor each time to get bits consumption in case of Huffman coding.
806 It is suboptimal solution */
807 /* TODO: May be cache scaled values */
808 for (ch = 0; ch < c->fullband_channels; ch++) {
809 for (band = 0; band < 32; band++) {
810 if (c->prediction_mode[ch][band] == -1) {
811 c->scale_factor[ch][band] = calc_one_scale(c, c->peak_cb[ch][band],
813 &c->quant[ch][band]);
820 memset(huff_bit_count_accum, 0, MAX_CHANNELS * DCA_CODE_BOOKS * 7 * sizeof(uint32_t));
821 memset(clc_bit_count_accum, 0, MAX_CHANNELS * DCA_CODE_BOOKS * sizeof(uint32_t));
822 for (ch = 0; ch < c->fullband_channels; ch++) {
823 for (band = 0; band < 32; band++) {
824 if (c->abits[ch][band] && c->abits[ch][band] <= DCA_CODE_BOOKS) {
825 accumulate_huff_bit_consumption(c->abits[ch][band],
826 c->quantized[ch][band],
827 huff_bit_count_accum[ch][c->abits[ch][band] - 1]);
828 clc_bit_count_accum[ch][c->abits[ch][band] - 1] += bit_consumption[c->abits[ch][band]];
830 bits_counter += bit_consumption[c->abits[ch][band]];
835 for (ch = 0; ch < c->fullband_channels; ch++) {
836 bits_counter += set_best_code(huff_bit_count_accum[ch],
837 clc_bit_count_accum[ch],
838 c->quant_index_sel[ch]);
841 c->consumed_bits += bits_counter;
846 static void assign_bits(DCAEncContext *c)
848 /* Find the bounds where the binary search should work */
853 init_quantization_noise(c, c->worst_quantization_noise, forbid_zero);
854 low = high = c->worst_quantization_noise;
855 if (c->consumed_bits > c->frame_bits) {
856 while (c->consumed_bits > c->frame_bits) {
857 if (used_abits == USED_1ABITS && forbid_zero) {
863 used_abits = init_quantization_noise(c, high, forbid_zero);
866 while (c->consumed_bits <= c->frame_bits) {
868 if (used_abits == USED_26ABITS)
869 goto out; /* The requested bitrate is too high, pad with zeros */
871 used_abits = init_quantization_noise(c, low, forbid_zero);
875 /* Now do a binary search between low and high to see what fits */
876 for (down = snr_fudge >> 1; down; down >>= 1) {
877 init_quantization_noise(c, high - down, forbid_zero);
878 if (c->consumed_bits <= c->frame_bits)
881 init_quantization_noise(c, high, forbid_zero);
883 c->worst_quantization_noise = high;
884 if (high > c->worst_noise_ever)
885 c->worst_noise_ever = high;
888 static void shift_history(DCAEncContext *c, const int32_t *input)
892 for (k = 0; k < 512; k++)
893 for (ch = 0; ch < c->channels; ch++) {
894 const int chi = c->channel_order_tab[ch];
896 c->history[ch][k] = input[k * c->channels + chi];
900 static void fill_in_adpcm_bufer(DCAEncContext *c)
904 /* We fill in ADPCM work buffer for subbands which hasn't been ADPCM coded
905 * in current frame - we need this data if subband of next frame is
908 for (ch = 0; ch < c->channels; ch++) {
909 for (band = 0; band < 32; band++) {
910 int32_t *samples = c->subband[ch][band] - DCA_ADPCM_COEFFS;
911 if (c->prediction_mode[ch][band] == -1) {
912 step_size = get_step_size(c, ch, band);
914 ff_dca_core_dequantize(c->adpcm_history[ch][band],
915 c->quantized[ch][band]+12, step_size,
916 ff_dca_scale_factor_quant7[c->scale_factor[ch][band]], 0, 4);
918 AV_COPY128U(c->adpcm_history[ch][band], c->adpcm_history[ch][band]+4);
920 /* Copy dequantized values for LPC analysis.
921 * It reduces artifacts in case of extreme quantization,
922 * example: in current frame abits is 1 and has no prediction flag,
923 * but end of this frame is sine like signal. In this case, if LPC analysis uses
924 * original values, likely LPC analysis returns good prediction gain, and sets prediction flag.
925 * But there are no proper value in decoder history, so likely result will be no good.
926 * Bitstream has "Predictor history flag switch", but this flag disables history for all subbands
928 samples[0] = c->adpcm_history[ch][band][0] << 7;
929 samples[1] = c->adpcm_history[ch][band][1] << 7;
930 samples[2] = c->adpcm_history[ch][band][2] << 7;
931 samples[3] = c->adpcm_history[ch][band][3] << 7;
936 static void calc_lfe_scales(DCAEncContext *c)
939 c->lfe_scale_factor = calc_one_scale(c, c->lfe_peak_cb, 11, &c->lfe_quant);
942 static void put_frame_header(DCAEncContext *c)
945 put_bits(&c->pb, 16, 0x7ffe);
946 put_bits(&c->pb, 16, 0x8001);
948 /* Frame type: normal */
949 put_bits(&c->pb, 1, 1);
951 /* Deficit sample count: none */
952 put_bits(&c->pb, 5, 31);
954 /* CRC is not present */
955 put_bits(&c->pb, 1, 0);
957 /* Number of PCM sample blocks */
958 put_bits(&c->pb, 7, SUBBAND_SAMPLES - 1);
960 /* Primary frame byte size */
961 put_bits(&c->pb, 14, c->frame_size - 1);
963 /* Audio channel arrangement */
964 put_bits(&c->pb, 6, c->channel_config);
966 /* Core audio sampling frequency */
967 put_bits(&c->pb, 4, bitstream_sfreq[c->samplerate_index]);
969 /* Transmission bit rate */
970 put_bits(&c->pb, 5, c->bitrate_index);
972 /* Embedded down mix: disabled */
973 put_bits(&c->pb, 1, 0);
975 /* Embedded dynamic range flag: not present */
976 put_bits(&c->pb, 1, 0);
978 /* Embedded time stamp flag: not present */
979 put_bits(&c->pb, 1, 0);
981 /* Auxiliary data flag: not present */
982 put_bits(&c->pb, 1, 0);
984 /* HDCD source: no */
985 put_bits(&c->pb, 1, 0);
987 /* Extension audio ID: N/A */
988 put_bits(&c->pb, 3, 0);
990 /* Extended audio data: not present */
991 put_bits(&c->pb, 1, 0);
993 /* Audio sync word insertion flag: after each sub-frame */
994 put_bits(&c->pb, 1, 0);
996 /* Low frequency effects flag: not present or 64x subsampling */
997 put_bits(&c->pb, 2, c->lfe_channel ? 2 : 0);
999 /* Predictor history switch flag: on */
1000 put_bits(&c->pb, 1, 1);
1003 /* Multirate interpolator switch: non-perfect reconstruction */
1004 put_bits(&c->pb, 1, 0);
1006 /* Encoder software revision: 7 */
1007 put_bits(&c->pb, 4, 7);
1009 /* Copy history: 0 */
1010 put_bits(&c->pb, 2, 0);
1012 /* Source PCM resolution: 16 bits, not DTS ES */
1013 put_bits(&c->pb, 3, 0);
1015 /* Front sum/difference coding: no */
1016 put_bits(&c->pb, 1, 0);
1018 /* Surrounds sum/difference coding: no */
1019 put_bits(&c->pb, 1, 0);
1021 /* Dialog normalization: 0 dB */
1022 put_bits(&c->pb, 4, 0);
1025 static void put_primary_audio_header(DCAEncContext *c)
1028 /* Number of subframes */
1029 put_bits(&c->pb, 4, SUBFRAMES - 1);
1031 /* Number of primary audio channels */
1032 put_bits(&c->pb, 3, c->fullband_channels - 1);
1034 /* Subband activity count */
1035 for (ch = 0; ch < c->fullband_channels; ch++)
1036 put_bits(&c->pb, 5, DCAENC_SUBBANDS - 2);
1038 /* High frequency VQ start subband */
1039 for (ch = 0; ch < c->fullband_channels; ch++)
1040 put_bits(&c->pb, 5, DCAENC_SUBBANDS - 1);
1042 /* Joint intensity coding index: 0, 0 */
1043 for (ch = 0; ch < c->fullband_channels; ch++)
1044 put_bits(&c->pb, 3, 0);
1046 /* Transient mode codebook: A4, A4 (arbitrary) */
1047 for (ch = 0; ch < c->fullband_channels; ch++)
1048 put_bits(&c->pb, 2, 0);
1050 /* Scale factor code book: 7 bit linear, 7-bit sqrt table (for each channel) */
1051 for (ch = 0; ch < c->fullband_channels; ch++)
1052 put_bits(&c->pb, 3, 6);
1054 /* Bit allocation quantizer select: linear 5-bit */
1055 for (ch = 0; ch < c->fullband_channels; ch++)
1056 put_bits(&c->pb, 3, c->bit_allocation_sel[ch]);
1058 /* Quantization index codebook select */
1059 for (i = 0; i < DCA_CODE_BOOKS; i++)
1060 for (ch = 0; ch < c->fullband_channels; ch++)
1061 put_bits(&c->pb, ff_dca_quant_index_sel_nbits[i], c->quant_index_sel[ch][i]);
1063 /* Scale factor adjustment index: transmitted in case of Huffman coding */
1064 for (i = 0; i < DCA_CODE_BOOKS; i++)
1065 for (ch = 0; ch < c->fullband_channels; ch++)
1066 if (c->quant_index_sel[ch][i] < ff_dca_quant_index_group_size[i])
1067 put_bits(&c->pb, 2, 0);
1069 /* Audio header CRC check word: not transmitted */
1072 static void put_subframe_samples(DCAEncContext *c, int ss, int band, int ch)
1074 int i, j, sum, bits, sel;
1075 if (c->abits[ch][band] <= DCA_CODE_BOOKS) {
1076 av_assert0(c->abits[ch][band] > 0);
1077 sel = c->quant_index_sel[ch][c->abits[ch][band] - 1];
1079 if (sel < ff_dca_quant_index_group_size[c->abits[ch][band] - 1]) {
1080 ff_dca_vlc_enc_quant(&c->pb, &c->quantized[ch][band][ss * 8], 8,
1081 sel, c->abits[ch][band] - 1);
1086 if (c->abits[ch][band] <= 7) {
1087 for (i = 0; i < 8; i += 4) {
1089 for (j = 3; j >= 0; j--) {
1090 sum *= ff_dca_quant_levels[c->abits[ch][band]];
1091 sum += c->quantized[ch][band][ss * 8 + i + j];
1092 sum += (ff_dca_quant_levels[c->abits[ch][band]] - 1) / 2;
1094 put_bits(&c->pb, bit_consumption[c->abits[ch][band]] / 4, sum);
1100 for (i = 0; i < 8; i++) {
1101 bits = bit_consumption[c->abits[ch][band]] / 16;
1102 put_sbits(&c->pb, bits, c->quantized[ch][band][ss * 8 + i]);
1106 static void put_subframe(DCAEncContext *c, int subframe)
1108 int i, band, ss, ch;
1110 /* Subsubframes count */
1111 put_bits(&c->pb, 2, SUBSUBFRAMES -1);
1113 /* Partial subsubframe sample count: dummy */
1114 put_bits(&c->pb, 3, 0);
1116 /* Prediction mode: no ADPCM, in each channel and subband */
1117 for (ch = 0; ch < c->fullband_channels; ch++)
1118 for (band = 0; band < DCAENC_SUBBANDS; band++)
1119 put_bits(&c->pb, 1, !(c->prediction_mode[ch][band] == -1));
1121 /* Prediction VQ address */
1122 for (ch = 0; ch < c->fullband_channels; ch++)
1123 for (band = 0; band < DCAENC_SUBBANDS; band++)
1124 if (c->prediction_mode[ch][band] >= 0)
1125 put_bits(&c->pb, 12, c->prediction_mode[ch][band]);
1127 /* Bit allocation index */
1128 for (ch = 0; ch < c->fullband_channels; ch++) {
1129 if (c->bit_allocation_sel[ch] == 6) {
1130 for (band = 0; band < DCAENC_SUBBANDS; band++) {
1131 put_bits(&c->pb, 5, c->abits[ch][band]);
1134 ff_dca_vlc_enc_alloc(&c->pb, c->abits[ch], DCAENC_SUBBANDS,
1135 c->bit_allocation_sel[ch]);
1139 if (SUBSUBFRAMES > 1) {
1140 /* Transition mode: none for each channel and subband */
1141 for (ch = 0; ch < c->fullband_channels; ch++)
1142 for (band = 0; band < DCAENC_SUBBANDS; band++)
1143 if (c->abits[ch][band])
1144 put_bits(&c->pb, 1, 0); /* codebook A4 */
1148 for (ch = 0; ch < c->fullband_channels; ch++)
1149 for (band = 0; band < DCAENC_SUBBANDS; band++)
1150 if (c->abits[ch][band])
1151 put_bits(&c->pb, 7, c->scale_factor[ch][band]);
1153 /* Joint subband scale factor codebook select: not transmitted */
1154 /* Scale factors for joint subband coding: not transmitted */
1155 /* Stereo down-mix coefficients: not transmitted */
1156 /* Dynamic range coefficient: not transmitted */
1157 /* Stde information CRC check word: not transmitted */
1158 /* VQ encoded high frequency subbands: not transmitted */
1160 /* LFE data: 8 samples and scalefactor */
1161 if (c->lfe_channel) {
1162 for (i = 0; i < DCA_LFE_SAMPLES; i++)
1163 put_bits(&c->pb, 8, quantize_value(c->downsampled_lfe[i], c->lfe_quant) & 0xff);
1164 put_bits(&c->pb, 8, c->lfe_scale_factor);
1167 /* Audio data (subsubframes) */
1168 for (ss = 0; ss < SUBSUBFRAMES ; ss++)
1169 for (ch = 0; ch < c->fullband_channels; ch++)
1170 for (band = 0; band < DCAENC_SUBBANDS; band++)
1171 if (c->abits[ch][band])
1172 put_subframe_samples(c, ss, band, ch);
1175 put_bits(&c->pb, 16, 0xffff);
1178 static int encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
1179 const AVFrame *frame, int *got_packet_ptr)
1181 DCAEncContext *c = avctx->priv_data;
1182 const int32_t *samples;
1185 if ((ret = ff_alloc_packet2(avctx, avpkt, c->frame_size, 0)) < 0)
1188 samples = (const int32_t *)frame->data[0];
1190 subband_transform(c, samples);
1192 lfe_downsample(c, samples);
1194 calc_masking(c, samples);
1195 if (c->options.adpcm_mode)
1200 shift_history(c, samples);
1202 init_put_bits(&c->pb, avpkt->data, avpkt->size);
1203 fill_in_adpcm_bufer(c);
1204 put_frame_header(c);
1205 put_primary_audio_header(c);
1206 for (i = 0; i < SUBFRAMES; i++)
1210 for (i = put_bits_count(&c->pb); i < 8*c->frame_size; i++)
1211 put_bits(&c->pb, 1, 0);
1213 flush_put_bits(&c->pb);
1215 avpkt->pts = frame->pts;
1216 avpkt->duration = ff_samples_to_time_base(avctx, frame->nb_samples);
1217 avpkt->size = put_bits_count(&c->pb) >> 3;
1218 *got_packet_ptr = 1;
1222 #define DCAENC_FLAGS AV_OPT_FLAG_ENCODING_PARAM | AV_OPT_FLAG_AUDIO_PARAM
1224 static const AVOption options[] = {
1225 { "dca_adpcm", "Use ADPCM encoding", offsetof(DCAEncContext, options.adpcm_mode), AV_OPT_TYPE_BOOL, {.i64 = 0}, 0, 1, DCAENC_FLAGS },
1229 static const AVClass dcaenc_class = {
1230 .class_name = "DCA (DTS Coherent Acoustics)",
1231 .item_name = av_default_item_name,
1233 .version = LIBAVUTIL_VERSION_INT,
1236 static const AVCodecDefault defaults[] = {
1241 AVCodec ff_dca_encoder = {
1243 .long_name = NULL_IF_CONFIG_SMALL("DCA (DTS Coherent Acoustics)"),
1244 .type = AVMEDIA_TYPE_AUDIO,
1245 .id = AV_CODEC_ID_DTS,
1246 .priv_data_size = sizeof(DCAEncContext),
1247 .init = encode_init,
1248 .close = encode_close,
1249 .encode2 = encode_frame,
1250 .capabilities = AV_CODEC_CAP_EXPERIMENTAL,
1251 .caps_internal = FF_CODEC_CAP_INIT_THREADSAFE | FF_CODEC_CAP_INIT_CLEANUP,
1252 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_S32,
1253 AV_SAMPLE_FMT_NONE },
1254 .supported_samplerates = sample_rates,
1255 .channel_layouts = (const uint64_t[]) { AV_CH_LAYOUT_MONO,
1256 AV_CH_LAYOUT_STEREO,
1258 AV_CH_LAYOUT_5POINT0,
1259 AV_CH_LAYOUT_5POINT1,
1261 .defaults = defaults,
1262 .priv_class = &dcaenc_class,