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/opt.h"
44 #define MAX_CHANNELS 6
45 #define DCA_MAX_FRAME_SIZE 16384
46 #define DCA_HEADER_SIZE 13
47 #define DCA_LFE_SAMPLES 8
49 #define DCAENC_SUBBANDS 32
51 #define SUBSUBFRAMES 2
52 #define SUBBAND_SAMPLES (SUBFRAMES * SUBSUBFRAMES * 8)
55 #define COS_T(x) (c->cos_table[(x) & 2047])
57 typedef struct CompressionOptions {
61 typedef struct DCAEncContext {
64 DCAADPCMEncContext adpcm_ctx;
66 CompressionOptions options;
69 int fullband_channels;
75 const int32_t *band_interpolation;
76 const int32_t *band_spectrum;
80 const int8_t *channel_order_tab; ///< channel reordering table, lfe and non lfe
82 int32_t prediction_mode[MAX_CHANNELS][DCAENC_SUBBANDS];
83 int32_t adpcm_history[MAX_CHANNELS][DCAENC_SUBBANDS][DCA_ADPCM_COEFFS * 2];
84 int32_t history[MAX_CHANNELS][512]; /* This is a circular buffer */
85 int32_t *subband[MAX_CHANNELS][DCAENC_SUBBANDS];
86 int32_t quantized[MAX_CHANNELS][DCAENC_SUBBANDS][SUBBAND_SAMPLES];
87 int32_t peak_cb[MAX_CHANNELS][DCAENC_SUBBANDS];
88 int32_t diff_peak_cb[MAX_CHANNELS][DCAENC_SUBBANDS]; ///< expected peak of residual signal
89 int32_t downsampled_lfe[DCA_LFE_SAMPLES];
90 int32_t masking_curve_cb[SUBSUBFRAMES][256];
91 int32_t bit_allocation_sel[MAX_CHANNELS];
92 int abits[MAX_CHANNELS][DCAENC_SUBBANDS];
93 int scale_factor[MAX_CHANNELS][DCAENC_SUBBANDS];
94 softfloat quant[MAX_CHANNELS][DCAENC_SUBBANDS];
95 int32_t quant_index_sel[MAX_CHANNELS][DCA_CODE_BOOKS];
96 int32_t eff_masking_curve_cb[256];
97 int32_t band_masking_cb[32];
98 int32_t worst_quantization_noise;
99 int32_t worst_noise_ever;
101 int consumed_adpcm_bits; ///< Number of bits to transmit ADPCM related info
103 int32_t cos_table[2048];
104 int32_t band_interpolation_tab[2][512];
105 int32_t band_spectrum_tab[2][8];
106 int32_t auf[9][AUBANDS][256];
107 int32_t cb_to_add[256];
108 int32_t cb_to_level[2048];
109 int32_t lfe_fir_64i[512];
112 /* Transfer function of outer and middle ear, Hz -> dB */
113 static double hom(double f)
115 double f1 = f / 1000;
117 return -3.64 * pow(f1, -0.8)
118 + 6.8 * exp(-0.6 * (f1 - 3.4) * (f1 - 3.4))
119 - 6.0 * exp(-0.15 * (f1 - 8.7) * (f1 - 8.7))
120 - 0.0006 * (f1 * f1) * (f1 * f1);
123 static double gammafilter(int i, double f)
125 double h = (f - fc[i]) / erb[i];
129 return 20 * log10(h);
132 static int subband_bufer_alloc(DCAEncContext *c)
135 int32_t *bufer = av_calloc(MAX_CHANNELS * DCAENC_SUBBANDS *
136 (SUBBAND_SAMPLES + DCA_ADPCM_COEFFS),
139 return AVERROR(ENOMEM);
141 /* we need a place for DCA_ADPCM_COEFF samples from previous frame
142 * to calc prediction coefficients for each subband */
143 for (ch = 0; ch < MAX_CHANNELS; ch++) {
144 for (band = 0; band < DCAENC_SUBBANDS; band++) {
145 c->subband[ch][band] = bufer +
146 ch * DCAENC_SUBBANDS * (SUBBAND_SAMPLES + DCA_ADPCM_COEFFS) +
147 band * (SUBBAND_SAMPLES + DCA_ADPCM_COEFFS) + DCA_ADPCM_COEFFS;
153 static void subband_bufer_free(DCAEncContext *c)
155 if (c->subband[0][0]) {
156 int32_t *bufer = c->subband[0][0] - DCA_ADPCM_COEFFS;
158 c->subband[0][0] = NULL;
162 static int encode_init(AVCodecContext *avctx)
164 DCAEncContext *c = avctx->priv_data;
165 uint64_t layout = avctx->channel_layout;
166 int i, j, k, min_frame_bits;
169 if ((ret = subband_bufer_alloc(c)) < 0)
172 c->fullband_channels = c->channels = avctx->channels;
173 c->lfe_channel = (avctx->channels == 3 || avctx->channels == 6);
174 c->band_interpolation = c->band_interpolation_tab[1];
175 c->band_spectrum = c->band_spectrum_tab[1];
176 c->worst_quantization_noise = -2047;
177 c->worst_noise_ever = -2047;
178 c->consumed_adpcm_bits = 0;
180 if (ff_dcaadpcm_init(&c->adpcm_ctx))
181 return AVERROR(ENOMEM);
184 av_log(avctx, AV_LOG_WARNING, "No channel layout specified. The "
185 "encoder will guess the layout, but it "
186 "might be incorrect.\n");
187 layout = av_get_default_channel_layout(avctx->channels);
190 case AV_CH_LAYOUT_MONO: c->channel_config = 0; break;
191 case AV_CH_LAYOUT_STEREO: c->channel_config = 2; break;
192 case AV_CH_LAYOUT_2_2: c->channel_config = 8; break;
193 case AV_CH_LAYOUT_5POINT0: c->channel_config = 9; break;
194 case AV_CH_LAYOUT_5POINT1: c->channel_config = 9; break;
196 av_log(avctx, AV_LOG_ERROR, "Unsupported channel layout!\n");
197 return AVERROR_PATCHWELCOME;
200 if (c->lfe_channel) {
201 c->fullband_channels--;
202 c->channel_order_tab = channel_reorder_lfe[c->channel_config];
204 c->channel_order_tab = channel_reorder_nolfe[c->channel_config];
207 for (i = 0; i < MAX_CHANNELS; i++) {
208 for (j = 0; j < DCA_CODE_BOOKS; j++) {
209 c->quant_index_sel[i][j] = ff_dca_quant_index_group_size[j];
212 c->bit_allocation_sel[i] = 6;
214 for (j = 0; j < DCAENC_SUBBANDS; j++) {
216 c->prediction_mode[i][j] = -1;
217 memset(c->adpcm_history[i][j], 0, sizeof(int32_t)*DCA_ADPCM_COEFFS);
221 for (i = 0; i < 9; i++) {
222 if (sample_rates[i] == avctx->sample_rate)
226 return AVERROR(EINVAL);
227 c->samplerate_index = i;
229 if (avctx->bit_rate < 32000 || avctx->bit_rate > 3840000) {
230 av_log(avctx, AV_LOG_ERROR, "Bit rate %"PRId64" not supported.", avctx->bit_rate);
231 return AVERROR(EINVAL);
233 for (i = 0; ff_dca_bit_rates[i] < avctx->bit_rate; i++)
235 c->bitrate_index = i;
236 c->frame_bits = FFALIGN((avctx->bit_rate * 512 + avctx->sample_rate - 1) / avctx->sample_rate, 32);
237 min_frame_bits = 132 + (493 + 28 * 32) * c->fullband_channels + c->lfe_channel * 72;
238 if (c->frame_bits < min_frame_bits || c->frame_bits > (DCA_MAX_FRAME_SIZE << 3))
239 return AVERROR(EINVAL);
241 c->frame_size = (c->frame_bits + 7) / 8;
243 avctx->frame_size = 32 * SUBBAND_SAMPLES;
245 if ((ret = ff_mdct_init(&c->mdct, 9, 0, 1.0)) < 0)
248 /* Init all tables */
249 c->cos_table[0] = 0x7fffffff;
250 c->cos_table[512] = 0;
251 c->cos_table[1024] = -c->cos_table[0];
252 for (i = 1; i < 512; i++) {
253 c->cos_table[i] = (int32_t)(0x7fffffff * cos(M_PI * i / 1024));
254 c->cos_table[1024-i] = -c->cos_table[i];
255 c->cos_table[1024+i] = -c->cos_table[i];
256 c->cos_table[2048-i] = +c->cos_table[i];
259 for (i = 0; i < 2048; i++)
260 c->cb_to_level[i] = (int32_t)(0x7fffffff * ff_exp10(-0.005 * i));
262 for (k = 0; k < 32; k++) {
263 for (j = 0; j < 8; j++) {
264 c->lfe_fir_64i[64 * j + k] = (int32_t)(0xffffff800000ULL * ff_dca_lfe_fir_64[8 * k + j]);
265 c->lfe_fir_64i[64 * (7-j) + (63 - k)] = (int32_t)(0xffffff800000ULL * ff_dca_lfe_fir_64[8 * k + j]);
269 for (i = 0; i < 512; i++) {
270 c->band_interpolation_tab[0][i] = (int32_t)(0x1000000000ULL * ff_dca_fir_32bands_perfect[i]);
271 c->band_interpolation_tab[1][i] = (int32_t)(0x1000000000ULL * ff_dca_fir_32bands_nonperfect[i]);
274 for (i = 0; i < 9; i++) {
275 for (j = 0; j < AUBANDS; j++) {
276 for (k = 0; k < 256; k++) {
277 double freq = sample_rates[i] * (k + 0.5) / 512;
279 c->auf[i][j][k] = (int32_t)(10 * (hom(freq) + gammafilter(j, freq)));
284 for (i = 0; i < 256; i++) {
285 double add = 1 + ff_exp10(-0.01 * i);
286 c->cb_to_add[i] = (int32_t)(100 * log10(add));
288 for (j = 0; j < 8; j++) {
290 for (i = 0; i < 512; i++) {
291 double reconst = ff_dca_fir_32bands_perfect[i] * ((i & 64) ? (-1) : 1);
292 accum += reconst * cos(2 * M_PI * (i + 0.5 - 256) * (j + 0.5) / 512);
294 c->band_spectrum_tab[0][j] = (int32_t)(200 * log10(accum));
296 for (j = 0; j < 8; j++) {
298 for (i = 0; i < 512; i++) {
299 double reconst = ff_dca_fir_32bands_nonperfect[i] * ((i & 64) ? (-1) : 1);
300 accum += reconst * cos(2 * M_PI * (i + 0.5 - 256) * (j + 0.5) / 512);
302 c->band_spectrum_tab[1][j] = (int32_t)(200 * log10(accum));
308 static av_cold int encode_close(AVCodecContext *avctx)
310 DCAEncContext *c = avctx->priv_data;
311 ff_mdct_end(&c->mdct);
312 subband_bufer_free(c);
313 ff_dcaadpcm_free(&c->adpcm_ctx);
318 static void subband_transform(DCAEncContext *c, const int32_t *input)
320 int ch, subs, i, k, j;
322 for (ch = 0; ch < c->fullband_channels; ch++) {
323 /* History is copied because it is also needed for PSY */
326 const int chi = c->channel_order_tab[ch];
328 memcpy(hist, &c->history[ch][0], 512 * sizeof(int32_t));
330 for (subs = 0; subs < SUBBAND_SAMPLES; subs++) {
335 /* Calculate the convolutions at once */
336 memset(accum, 0, 64 * sizeof(int32_t));
338 for (k = 0, i = hist_start, j = 0;
339 i < 512; k = (k + 1) & 63, i++, j++)
340 accum[k] += mul32(hist[i], c->band_interpolation[j]);
341 for (i = 0; i < hist_start; k = (k + 1) & 63, i++, j++)
342 accum[k] += mul32(hist[i], c->band_interpolation[j]);
344 for (k = 16; k < 32; k++)
345 accum[k] = accum[k] - accum[31 - k];
346 for (k = 32; k < 48; k++)
347 accum[k] = accum[k] + accum[95 - k];
349 for (band = 0; band < 32; band++) {
351 for (i = 16; i < 48; i++) {
352 int s = (2 * band + 1) * (2 * (i + 16) + 1);
353 resp += mul32(accum[i], COS_T(s << 3)) >> 3;
356 c->subband[ch][band][subs] = ((band + 1) & 2) ? -resp : resp;
359 /* Copy in 32 new samples from input */
360 for (i = 0; i < 32; i++)
361 hist[i + hist_start] = input[(subs * 32 + i) * c->channels + chi];
363 hist_start = (hist_start + 32) & 511;
368 static void lfe_downsample(DCAEncContext *c, const int32_t *input)
370 /* FIXME: make 128x LFE downsampling possible */
371 const int lfech = lfe_index[c->channel_config];
377 memcpy(hist, &c->history[c->channels - 1][0], 512 * sizeof(int32_t));
379 for (lfes = 0; lfes < DCA_LFE_SAMPLES; lfes++) {
380 /* Calculate the convolution */
383 for (i = hist_start, j = 0; i < 512; i++, j++)
384 accum += mul32(hist[i], c->lfe_fir_64i[j]);
385 for (i = 0; i < hist_start; i++, j++)
386 accum += mul32(hist[i], c->lfe_fir_64i[j]);
388 c->downsampled_lfe[lfes] = accum;
390 /* Copy in 64 new samples from input */
391 for (i = 0; i < 64; i++)
392 hist[i + hist_start] = input[(lfes * 64 + i) * c->channels + lfech];
394 hist_start = (hist_start + 64) & 511;
398 static int32_t get_cb(DCAEncContext *c, int32_t in)
403 for (i = 1024; i > 0; i >>= 1) {
404 if (c->cb_to_level[i + res] >= in)
410 static int32_t add_cb(DCAEncContext *c, int32_t a, int32_t b)
413 FFSWAP(int32_t, a, b);
417 return a + c->cb_to_add[a - b];
420 static void calc_power(DCAEncContext *c,
421 const int32_t in[2 * 256], int32_t power[256])
424 LOCAL_ALIGNED_32(int32_t, data, [512]);
425 LOCAL_ALIGNED_32(int32_t, coeff, [256]);
427 for (i = 0; i < 512; i++)
428 data[i] = norm__(mul32(in[i], 0x3fffffff - (COS_T(4 * i + 2) >> 1)), 4);
430 c->mdct.mdct_calc(&c->mdct, coeff, data);
431 for (i = 0; i < 256; i++) {
432 const int32_t cb = get_cb(c, coeff[i]);
433 power[i] = add_cb(c, cb, cb);
437 static void adjust_jnd(DCAEncContext *c,
438 const int32_t in[512], int32_t out_cb[256])
441 int32_t out_cb_unnorm[256];
443 const int32_t ca_cb = -1114;
444 const int32_t cs_cb = 928;
445 const int samplerate_index = c->samplerate_index;
448 calc_power(c, in, power);
450 for (j = 0; j < 256; j++)
451 out_cb_unnorm[j] = -2047; /* and can only grow */
453 for (i = 0; i < AUBANDS; i++) {
454 denom = ca_cb; /* and can only grow */
455 for (j = 0; j < 256; j++)
456 denom = add_cb(c, denom, power[j] + c->auf[samplerate_index][i][j]);
457 for (j = 0; j < 256; j++)
458 out_cb_unnorm[j] = add_cb(c, out_cb_unnorm[j],
459 -denom + c->auf[samplerate_index][i][j]);
462 for (j = 0; j < 256; j++)
463 out_cb[j] = add_cb(c, out_cb[j], -out_cb_unnorm[j] - ca_cb - cs_cb);
466 typedef void (*walk_band_t)(DCAEncContext *c, int band1, int band2, int f,
467 int32_t spectrum1, int32_t spectrum2, int channel,
470 static void walk_band_low(DCAEncContext *c, int band, int channel,
471 walk_band_t walk, int32_t *arg)
476 for (f = 0; f < 4; f++)
477 walk(c, 0, 0, f, 0, -2047, channel, arg);
479 for (f = 0; f < 8; f++)
480 walk(c, band, band - 1, 8 * band - 4 + f,
481 c->band_spectrum[7 - f], c->band_spectrum[f], channel, arg);
485 static void walk_band_high(DCAEncContext *c, int band, int channel,
486 walk_band_t walk, int32_t *arg)
491 for (f = 0; f < 4; f++)
492 walk(c, 31, 31, 256 - 4 + f, 0, -2047, channel, arg);
494 for (f = 0; f < 8; f++)
495 walk(c, band, band + 1, 8 * band + 4 + f,
496 c->band_spectrum[f], c->band_spectrum[7 - f], channel, arg);
500 static void update_band_masking(DCAEncContext *c, int band1, int band2,
501 int f, int32_t spectrum1, int32_t spectrum2,
502 int channel, int32_t * arg)
504 int32_t value = c->eff_masking_curve_cb[f] - spectrum1;
506 if (value < c->band_masking_cb[band1])
507 c->band_masking_cb[band1] = value;
510 static void calc_masking(DCAEncContext *c, const int32_t *input)
512 int i, k, band, ch, ssf;
515 for (i = 0; i < 256; i++)
516 for (ssf = 0; ssf < SUBSUBFRAMES; ssf++)
517 c->masking_curve_cb[ssf][i] = -2047;
519 for (ssf = 0; ssf < SUBSUBFRAMES; ssf++)
520 for (ch = 0; ch < c->fullband_channels; ch++) {
521 const int chi = c->channel_order_tab[ch];
523 for (i = 0, k = 128 + 256 * ssf; k < 512; i++, k++)
524 data[i] = c->history[ch][k];
525 for (k -= 512; i < 512; i++, k++)
526 data[i] = input[k * c->channels + chi];
527 adjust_jnd(c, data, c->masking_curve_cb[ssf]);
529 for (i = 0; i < 256; i++) {
532 for (ssf = 0; ssf < SUBSUBFRAMES; ssf++)
533 if (c->masking_curve_cb[ssf][i] < m)
534 m = c->masking_curve_cb[ssf][i];
535 c->eff_masking_curve_cb[i] = m;
538 for (band = 0; band < 32; band++) {
539 c->band_masking_cb[band] = 2048;
540 walk_band_low(c, band, 0, update_band_masking, NULL);
541 walk_band_high(c, band, 0, update_band_masking, NULL);
545 static inline int32_t find_peak(DCAEncContext *c, const int32_t *in, int len)
549 for (sample = 0; sample < len; sample++) {
550 int32_t s = abs(in[sample]);
557 static void find_peaks(DCAEncContext *c)
561 for (ch = 0; ch < c->fullband_channels; ch++) {
562 for (band = 0; band < 32; band++)
563 c->peak_cb[ch][band] = find_peak(c, c->subband[ch][band],
568 c->lfe_peak_cb = find_peak(c, c->downsampled_lfe, DCA_LFE_SAMPLES);
571 static void adpcm_analysis(DCAEncContext *c)
576 int32_t estimated_diff[SUBBAND_SAMPLES];
578 c->consumed_adpcm_bits = 0;
579 for (ch = 0; ch < c->fullband_channels; ch++) {
580 for (band = 0; band < 32; band++) {
581 samples = c->subband[ch][band] - DCA_ADPCM_COEFFS;
582 pred_vq_id = ff_dcaadpcm_subband_analysis(&c->adpcm_ctx, samples,
583 SUBBAND_SAMPLES, estimated_diff);
584 if (pred_vq_id >= 0) {
585 c->prediction_mode[ch][band] = pred_vq_id;
586 c->consumed_adpcm_bits += 12; //12 bits to transmit prediction vq index
587 c->diff_peak_cb[ch][band] = find_peak(c, estimated_diff, 16);
589 c->prediction_mode[ch][band] = -1;
595 static const int snr_fudge = 128;
596 #define USED_1ABITS 1
597 #define USED_26ABITS 4
599 static inline int32_t get_step_size(DCAEncContext *c, int ch, int band)
603 if (c->bitrate_index == 3)
604 step_size = ff_dca_lossless_quant[c->abits[ch][band]];
606 step_size = ff_dca_lossy_quant[c->abits[ch][band]];
611 static int calc_one_scale(DCAEncContext *c, int32_t peak_cb, int abits,
615 int our_nscale, try_remove;
618 av_assert0(peak_cb <= 0);
619 av_assert0(peak_cb >= -2047);
622 peak = c->cb_to_level[-peak_cb];
624 for (try_remove = 64; try_remove > 0; try_remove >>= 1) {
625 if (scalefactor_inv[our_nscale - try_remove].e + stepsize_inv[abits].e <= 17)
627 our_quant.m = mul32(scalefactor_inv[our_nscale - try_remove].m, stepsize_inv[abits].m);
628 our_quant.e = scalefactor_inv[our_nscale - try_remove].e + stepsize_inv[abits].e - 17;
629 if ((ff_dca_quant_levels[abits] - 1) / 2 < quantize_value(peak, our_quant))
631 our_nscale -= try_remove;
634 if (our_nscale >= 125)
637 quant->m = mul32(scalefactor_inv[our_nscale].m, stepsize_inv[abits].m);
638 quant->e = scalefactor_inv[our_nscale].e + stepsize_inv[abits].e - 17;
639 av_assert0((ff_dca_quant_levels[abits] - 1) / 2 >= quantize_value(peak, *quant));
644 static inline void quantize_adpcm_subband(DCAEncContext *c, int ch, int band)
647 int32_t diff_peak_cb = c->diff_peak_cb[ch][band];
648 c->scale_factor[ch][band] = calc_one_scale(c, diff_peak_cb,
650 &c->quant[ch][band]);
652 step_size = get_step_size(c, ch, band);
653 ff_dcaadpcm_do_real(c->prediction_mode[ch][band],
655 ff_dca_scale_factor_quant7[c->scale_factor[ch][band]],
656 step_size, c->adpcm_history[ch][band], c->subband[ch][band],
657 c->adpcm_history[ch][band] + 4, c->quantized[ch][band],
658 SUBBAND_SAMPLES, c->cb_to_level[-diff_peak_cb]);
661 static void quantize_adpcm(DCAEncContext *c)
665 for (ch = 0; ch < c->fullband_channels; ch++)
666 for (band = 0; band < 32; band++)
667 if (c->prediction_mode[ch][band] >= 0)
668 quantize_adpcm_subband(c, ch, band);
671 static void quantize_pcm(DCAEncContext *c)
673 int sample, band, ch;
675 for (ch = 0; ch < c->fullband_channels; ch++) {
676 for (band = 0; band < 32; band++) {
677 if (c->prediction_mode[ch][band] == -1) {
678 for (sample = 0; sample < SUBBAND_SAMPLES; sample++) {
679 int32_t val = quantize_value(c->subband[ch][band][sample],
681 c->quantized[ch][band][sample] = val;
688 static void accumulate_huff_bit_consumption(int abits, int32_t *quantized,
691 uint8_t sel, id = abits - 1;
692 for (sel = 0; sel < ff_dca_quant_index_group_size[id]; sel++)
693 result[sel] += ff_dca_vlc_calc_quant_bits(quantized, SUBBAND_SAMPLES,
697 static uint32_t set_best_code(uint32_t vlc_bits[DCA_CODE_BOOKS][7],
698 uint32_t clc_bits[DCA_CODE_BOOKS],
699 int32_t res[DCA_CODE_BOOKS])
702 uint32_t best_sel_bits[DCA_CODE_BOOKS];
703 int32_t best_sel_id[DCA_CODE_BOOKS];
704 uint32_t t, bits = 0;
706 for (i = 0; i < DCA_CODE_BOOKS; i++) {
708 av_assert0(!((!!vlc_bits[i][0]) ^ (!!clc_bits[i])));
709 if (vlc_bits[i][0] == 0) {
710 /* do not transmit adjustment index for empty codebooks */
711 res[i] = ff_dca_quant_index_group_size[i];
716 best_sel_bits[i] = vlc_bits[i][0];
718 for (sel = 0; sel < ff_dca_quant_index_group_size[i]; sel++) {
719 if (best_sel_bits[i] > vlc_bits[i][sel] && vlc_bits[i][sel]) {
720 best_sel_bits[i] = vlc_bits[i][sel];
721 best_sel_id[i] = sel;
725 /* 2 bits to transmit scale factor adjustment index */
726 t = best_sel_bits[i] + 2;
727 if (t < clc_bits[i]) {
728 res[i] = best_sel_id[i];
731 res[i] = ff_dca_quant_index_group_size[i];
738 static uint32_t set_best_abits_code(int abits[DCAENC_SUBBANDS], int bands,
743 int32_t best_sel = 6;
744 int32_t best_bits = bands * 5;
746 /* Check do we have subband which cannot be encoded by Huffman tables */
747 for (i = 0; i < bands; i++) {
748 if (abits[i] > 12 || abits[i] == 0) {
754 for (i = 0; i < DCA_BITALLOC_12_COUNT; i++) {
755 t = ff_dca_vlc_calc_alloc_bits(abits, bands, i);
766 static int init_quantization_noise(DCAEncContext *c, int noise, int forbid_zero)
768 int ch, band, ret = USED_26ABITS | USED_1ABITS;
769 uint32_t huff_bit_count_accum[MAX_CHANNELS][DCA_CODE_BOOKS][7];
770 uint32_t clc_bit_count_accum[MAX_CHANNELS][DCA_CODE_BOOKS];
771 uint32_t bits_counter = 0;
773 c->consumed_bits = 132 + 333 * c->fullband_channels;
774 c->consumed_bits += c->consumed_adpcm_bits;
776 c->consumed_bits += 72;
778 /* attempt to guess the bit distribution based on the prevoius frame */
779 for (ch = 0; ch < c->fullband_channels; ch++) {
780 for (band = 0; band < 32; band++) {
781 int snr_cb = c->peak_cb[ch][band] - c->band_masking_cb[band] - noise;
783 if (snr_cb >= 1312) {
784 c->abits[ch][band] = 26;
786 } else if (snr_cb >= 222) {
787 c->abits[ch][band] = 8 + mul32(snr_cb - 222, 69000000);
788 ret &= ~(USED_26ABITS | USED_1ABITS);
789 } else if (snr_cb >= 0) {
790 c->abits[ch][band] = 2 + mul32(snr_cb, 106000000);
791 ret &= ~(USED_26ABITS | USED_1ABITS);
792 } else if (forbid_zero || snr_cb >= -140) {
793 c->abits[ch][band] = 1;
794 ret &= ~USED_26ABITS;
796 c->abits[ch][band] = 0;
797 ret &= ~(USED_26ABITS | USED_1ABITS);
800 c->consumed_bits += set_best_abits_code(c->abits[ch], 32,
801 &c->bit_allocation_sel[ch]);
804 /* Recalc scale_factor each time to get bits consumption in case of Huffman coding.
805 It is suboptimal solution */
806 /* TODO: May be cache scaled values */
807 for (ch = 0; ch < c->fullband_channels; ch++) {
808 for (band = 0; band < 32; band++) {
809 if (c->prediction_mode[ch][band] == -1) {
810 c->scale_factor[ch][band] = calc_one_scale(c, c->peak_cb[ch][band],
812 &c->quant[ch][band]);
819 memset(huff_bit_count_accum, 0, MAX_CHANNELS * DCA_CODE_BOOKS * 7 * sizeof(uint32_t));
820 memset(clc_bit_count_accum, 0, MAX_CHANNELS * DCA_CODE_BOOKS * sizeof(uint32_t));
821 for (ch = 0; ch < c->fullband_channels; ch++) {
822 for (band = 0; band < 32; band++) {
823 if (c->abits[ch][band] && c->abits[ch][band] <= DCA_CODE_BOOKS) {
824 accumulate_huff_bit_consumption(c->abits[ch][band],
825 c->quantized[ch][band],
826 huff_bit_count_accum[ch][c->abits[ch][band] - 1]);
827 clc_bit_count_accum[ch][c->abits[ch][band] - 1] += bit_consumption[c->abits[ch][band]];
829 bits_counter += bit_consumption[c->abits[ch][band]];
834 for (ch = 0; ch < c->fullband_channels; ch++) {
835 bits_counter += set_best_code(huff_bit_count_accum[ch],
836 clc_bit_count_accum[ch],
837 c->quant_index_sel[ch]);
840 c->consumed_bits += bits_counter;
845 static void assign_bits(DCAEncContext *c)
847 /* Find the bounds where the binary search should work */
852 init_quantization_noise(c, c->worst_quantization_noise, forbid_zero);
853 low = high = c->worst_quantization_noise;
854 if (c->consumed_bits > c->frame_bits) {
855 while (c->consumed_bits > c->frame_bits) {
856 if (used_abits == USED_1ABITS && forbid_zero) {
862 used_abits = init_quantization_noise(c, high, forbid_zero);
865 while (c->consumed_bits <= c->frame_bits) {
867 if (used_abits == USED_26ABITS)
868 goto out; /* The requested bitrate is too high, pad with zeros */
870 used_abits = init_quantization_noise(c, low, forbid_zero);
874 /* Now do a binary search between low and high to see what fits */
875 for (down = snr_fudge >> 1; down; down >>= 1) {
876 init_quantization_noise(c, high - down, forbid_zero);
877 if (c->consumed_bits <= c->frame_bits)
880 init_quantization_noise(c, high, forbid_zero);
882 c->worst_quantization_noise = high;
883 if (high > c->worst_noise_ever)
884 c->worst_noise_ever = high;
887 static void shift_history(DCAEncContext *c, const int32_t *input)
891 for (k = 0; k < 512; k++)
892 for (ch = 0; ch < c->channels; ch++) {
893 const int chi = c->channel_order_tab[ch];
895 c->history[ch][k] = input[k * c->channels + chi];
899 static void fill_in_adpcm_bufer(DCAEncContext *c)
903 /* We fill in ADPCM work buffer for subbands which hasn't been ADPCM coded
904 * in current frame - we need this data if subband of next frame is
907 for (ch = 0; ch < c->channels; ch++) {
908 for (band = 0; band < 32; band++) {
909 int32_t *samples = c->subband[ch][band] - DCA_ADPCM_COEFFS;
910 if (c->prediction_mode[ch][band] == -1) {
911 step_size = get_step_size(c, ch, band);
913 ff_dca_core_dequantize(c->adpcm_history[ch][band],
914 c->quantized[ch][band]+12, step_size,
915 ff_dca_scale_factor_quant7[c->scale_factor[ch][band]], 0, 4);
917 AV_COPY128U(c->adpcm_history[ch][band], c->adpcm_history[ch][band]+4);
919 /* Copy dequantized values for LPC analysis.
920 * It reduces artifacts in case of extreme quantization,
921 * example: in current frame abits is 1 and has no prediction flag,
922 * but end of this frame is sine like signal. In this case, if LPC analysis uses
923 * original values, likely LPC analysis returns good prediction gain, and sets prediction flag.
924 * But there are no proper value in decoder history, so likely result will be no good.
925 * Bitstream has "Predictor history flag switch", but this flag disables history for all subbands
927 samples[0] = c->adpcm_history[ch][band][0] << 7;
928 samples[1] = c->adpcm_history[ch][band][1] << 7;
929 samples[2] = c->adpcm_history[ch][band][2] << 7;
930 samples[3] = c->adpcm_history[ch][band][3] << 7;
935 static void calc_lfe_scales(DCAEncContext *c)
938 c->lfe_scale_factor = calc_one_scale(c, c->lfe_peak_cb, 11, &c->lfe_quant);
941 static void put_frame_header(DCAEncContext *c)
944 put_bits(&c->pb, 16, 0x7ffe);
945 put_bits(&c->pb, 16, 0x8001);
947 /* Frame type: normal */
948 put_bits(&c->pb, 1, 1);
950 /* Deficit sample count: none */
951 put_bits(&c->pb, 5, 31);
953 /* CRC is not present */
954 put_bits(&c->pb, 1, 0);
956 /* Number of PCM sample blocks */
957 put_bits(&c->pb, 7, SUBBAND_SAMPLES - 1);
959 /* Primary frame byte size */
960 put_bits(&c->pb, 14, c->frame_size - 1);
962 /* Audio channel arrangement */
963 put_bits(&c->pb, 6, c->channel_config);
965 /* Core audio sampling frequency */
966 put_bits(&c->pb, 4, bitstream_sfreq[c->samplerate_index]);
968 /* Transmission bit rate */
969 put_bits(&c->pb, 5, c->bitrate_index);
971 /* Embedded down mix: disabled */
972 put_bits(&c->pb, 1, 0);
974 /* Embedded dynamic range flag: not present */
975 put_bits(&c->pb, 1, 0);
977 /* Embedded time stamp flag: not present */
978 put_bits(&c->pb, 1, 0);
980 /* Auxiliary data flag: not present */
981 put_bits(&c->pb, 1, 0);
983 /* HDCD source: no */
984 put_bits(&c->pb, 1, 0);
986 /* Extension audio ID: N/A */
987 put_bits(&c->pb, 3, 0);
989 /* Extended audio data: not present */
990 put_bits(&c->pb, 1, 0);
992 /* Audio sync word insertion flag: after each sub-frame */
993 put_bits(&c->pb, 1, 0);
995 /* Low frequency effects flag: not present or 64x subsampling */
996 put_bits(&c->pb, 2, c->lfe_channel ? 2 : 0);
998 /* Predictor history switch flag: on */
999 put_bits(&c->pb, 1, 1);
1002 /* Multirate interpolator switch: non-perfect reconstruction */
1003 put_bits(&c->pb, 1, 0);
1005 /* Encoder software revision: 7 */
1006 put_bits(&c->pb, 4, 7);
1008 /* Copy history: 0 */
1009 put_bits(&c->pb, 2, 0);
1011 /* Source PCM resolution: 16 bits, not DTS ES */
1012 put_bits(&c->pb, 3, 0);
1014 /* Front sum/difference coding: no */
1015 put_bits(&c->pb, 1, 0);
1017 /* Surrounds sum/difference coding: no */
1018 put_bits(&c->pb, 1, 0);
1020 /* Dialog normalization: 0 dB */
1021 put_bits(&c->pb, 4, 0);
1024 static void put_primary_audio_header(DCAEncContext *c)
1027 /* Number of subframes */
1028 put_bits(&c->pb, 4, SUBFRAMES - 1);
1030 /* Number of primary audio channels */
1031 put_bits(&c->pb, 3, c->fullband_channels - 1);
1033 /* Subband activity count */
1034 for (ch = 0; ch < c->fullband_channels; ch++)
1035 put_bits(&c->pb, 5, DCAENC_SUBBANDS - 2);
1037 /* High frequency VQ start subband */
1038 for (ch = 0; ch < c->fullband_channels; ch++)
1039 put_bits(&c->pb, 5, DCAENC_SUBBANDS - 1);
1041 /* Joint intensity coding index: 0, 0 */
1042 for (ch = 0; ch < c->fullband_channels; ch++)
1043 put_bits(&c->pb, 3, 0);
1045 /* Transient mode codebook: A4, A4 (arbitrary) */
1046 for (ch = 0; ch < c->fullband_channels; ch++)
1047 put_bits(&c->pb, 2, 0);
1049 /* Scale factor code book: 7 bit linear, 7-bit sqrt table (for each channel) */
1050 for (ch = 0; ch < c->fullband_channels; ch++)
1051 put_bits(&c->pb, 3, 6);
1053 /* Bit allocation quantizer select: linear 5-bit */
1054 for (ch = 0; ch < c->fullband_channels; ch++)
1055 put_bits(&c->pb, 3, c->bit_allocation_sel[ch]);
1057 /* Quantization index codebook select */
1058 for (i = 0; i < DCA_CODE_BOOKS; i++)
1059 for (ch = 0; ch < c->fullband_channels; ch++)
1060 put_bits(&c->pb, ff_dca_quant_index_sel_nbits[i], c->quant_index_sel[ch][i]);
1062 /* Scale factor adjustment index: transmitted in case of Huffman coding */
1063 for (i = 0; i < DCA_CODE_BOOKS; i++)
1064 for (ch = 0; ch < c->fullband_channels; ch++)
1065 if (c->quant_index_sel[ch][i] < ff_dca_quant_index_group_size[i])
1066 put_bits(&c->pb, 2, 0);
1068 /* Audio header CRC check word: not transmitted */
1071 static void put_subframe_samples(DCAEncContext *c, int ss, int band, int ch)
1073 int i, j, sum, bits, sel;
1074 if (c->abits[ch][band] <= DCA_CODE_BOOKS) {
1075 av_assert0(c->abits[ch][band] > 0);
1076 sel = c->quant_index_sel[ch][c->abits[ch][band] - 1];
1078 if (sel < ff_dca_quant_index_group_size[c->abits[ch][band] - 1]) {
1079 ff_dca_vlc_enc_quant(&c->pb, &c->quantized[ch][band][ss * 8], 8,
1080 sel, c->abits[ch][band] - 1);
1085 if (c->abits[ch][band] <= 7) {
1086 for (i = 0; i < 8; i += 4) {
1088 for (j = 3; j >= 0; j--) {
1089 sum *= ff_dca_quant_levels[c->abits[ch][band]];
1090 sum += c->quantized[ch][band][ss * 8 + i + j];
1091 sum += (ff_dca_quant_levels[c->abits[ch][band]] - 1) / 2;
1093 put_bits(&c->pb, bit_consumption[c->abits[ch][band]] / 4, sum);
1099 for (i = 0; i < 8; i++) {
1100 bits = bit_consumption[c->abits[ch][band]] / 16;
1101 put_sbits(&c->pb, bits, c->quantized[ch][band][ss * 8 + i]);
1105 static void put_subframe(DCAEncContext *c, int subframe)
1107 int i, band, ss, ch;
1109 /* Subsubframes count */
1110 put_bits(&c->pb, 2, SUBSUBFRAMES -1);
1112 /* Partial subsubframe sample count: dummy */
1113 put_bits(&c->pb, 3, 0);
1115 /* Prediction mode: no ADPCM, in each channel and subband */
1116 for (ch = 0; ch < c->fullband_channels; ch++)
1117 for (band = 0; band < DCAENC_SUBBANDS; band++)
1118 put_bits(&c->pb, 1, !(c->prediction_mode[ch][band] == -1));
1120 /* Prediction VQ address */
1121 for (ch = 0; ch < c->fullband_channels; ch++)
1122 for (band = 0; band < DCAENC_SUBBANDS; band++)
1123 if (c->prediction_mode[ch][band] >= 0)
1124 put_bits(&c->pb, 12, c->prediction_mode[ch][band]);
1126 /* Bit allocation index */
1127 for (ch = 0; ch < c->fullband_channels; ch++) {
1128 if (c->bit_allocation_sel[ch] == 6) {
1129 for (band = 0; band < DCAENC_SUBBANDS; band++) {
1130 put_bits(&c->pb, 5, c->abits[ch][band]);
1133 ff_dca_vlc_enc_alloc(&c->pb, c->abits[ch], DCAENC_SUBBANDS,
1134 c->bit_allocation_sel[ch]);
1138 if (SUBSUBFRAMES > 1) {
1139 /* Transition mode: none for each channel and subband */
1140 for (ch = 0; ch < c->fullband_channels; ch++)
1141 for (band = 0; band < DCAENC_SUBBANDS; band++)
1142 if (c->abits[ch][band])
1143 put_bits(&c->pb, 1, 0); /* codebook A4 */
1147 for (ch = 0; ch < c->fullband_channels; ch++)
1148 for (band = 0; band < DCAENC_SUBBANDS; band++)
1149 if (c->abits[ch][band])
1150 put_bits(&c->pb, 7, c->scale_factor[ch][band]);
1152 /* Joint subband scale factor codebook select: not transmitted */
1153 /* Scale factors for joint subband coding: not transmitted */
1154 /* Stereo down-mix coefficients: not transmitted */
1155 /* Dynamic range coefficient: not transmitted */
1156 /* Stde information CRC check word: not transmitted */
1157 /* VQ encoded high frequency subbands: not transmitted */
1159 /* LFE data: 8 samples and scalefactor */
1160 if (c->lfe_channel) {
1161 for (i = 0; i < DCA_LFE_SAMPLES; i++)
1162 put_bits(&c->pb, 8, quantize_value(c->downsampled_lfe[i], c->lfe_quant) & 0xff);
1163 put_bits(&c->pb, 8, c->lfe_scale_factor);
1166 /* Audio data (subsubframes) */
1167 for (ss = 0; ss < SUBSUBFRAMES ; ss++)
1168 for (ch = 0; ch < c->fullband_channels; ch++)
1169 for (band = 0; band < DCAENC_SUBBANDS; band++)
1170 if (c->abits[ch][band])
1171 put_subframe_samples(c, ss, band, ch);
1174 put_bits(&c->pb, 16, 0xffff);
1177 static int encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
1178 const AVFrame *frame, int *got_packet_ptr)
1180 DCAEncContext *c = avctx->priv_data;
1181 const int32_t *samples;
1184 if ((ret = ff_alloc_packet2(avctx, avpkt, c->frame_size, 0)) < 0)
1187 samples = (const int32_t *)frame->data[0];
1189 subband_transform(c, samples);
1191 lfe_downsample(c, samples);
1193 calc_masking(c, samples);
1194 if (c->options.adpcm_mode)
1199 shift_history(c, samples);
1201 init_put_bits(&c->pb, avpkt->data, avpkt->size);
1202 fill_in_adpcm_bufer(c);
1203 put_frame_header(c);
1204 put_primary_audio_header(c);
1205 for (i = 0; i < SUBFRAMES; i++)
1209 for (i = put_bits_count(&c->pb); i < 8*c->frame_size; i++)
1210 put_bits(&c->pb, 1, 0);
1212 flush_put_bits(&c->pb);
1214 avpkt->pts = frame->pts;
1215 avpkt->duration = ff_samples_to_time_base(avctx, frame->nb_samples);
1216 avpkt->size = put_bits_count(&c->pb) >> 3;
1217 *got_packet_ptr = 1;
1221 #define DCAENC_FLAGS AV_OPT_FLAG_ENCODING_PARAM | AV_OPT_FLAG_AUDIO_PARAM
1223 static const AVOption options[] = {
1224 { "dca_adpcm", "Use ADPCM encoding", offsetof(DCAEncContext, options.adpcm_mode), AV_OPT_TYPE_BOOL, {.i64 = 0}, 0, 1, DCAENC_FLAGS },
1228 static const AVClass dcaenc_class = {
1229 .class_name = "DCA (DTS Coherent Acoustics)",
1230 .item_name = av_default_item_name,
1232 .version = LIBAVUTIL_VERSION_INT,
1235 static const AVCodecDefault defaults[] = {
1240 AVCodec ff_dca_encoder = {
1242 .long_name = NULL_IF_CONFIG_SMALL("DCA (DTS Coherent Acoustics)"),
1243 .type = AVMEDIA_TYPE_AUDIO,
1244 .id = AV_CODEC_ID_DTS,
1245 .priv_data_size = sizeof(DCAEncContext),
1246 .init = encode_init,
1247 .close = encode_close,
1248 .encode2 = encode_frame,
1249 .capabilities = AV_CODEC_CAP_EXPERIMENTAL,
1250 .caps_internal = FF_CODEC_CAP_INIT_THREADSAFE | FF_CODEC_CAP_INIT_CLEANUP,
1251 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_S32,
1252 AV_SAMPLE_FMT_NONE },
1253 .supported_samplerates = sample_rates,
1254 .channel_layouts = (const uint64_t[]) { AV_CH_LAYOUT_MONO,
1255 AV_CH_LAYOUT_STEREO,
1257 AV_CH_LAYOUT_5POINT0,
1258 AV_CH_LAYOUT_5POINT1,
1260 .defaults = defaults,
1261 .priv_class = &dcaenc_class,