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
24 #include "libavutil/avassert.h"
25 #include "libavutil/channel_layout.h"
26 #include "libavutil/common.h"
27 #include "libavutil/ffmath.h"
36 #define MAX_CHANNELS 6
37 #define DCA_MAX_FRAME_SIZE 16384
38 #define DCA_HEADER_SIZE 13
39 #define DCA_LFE_SAMPLES 8
41 #define DCAENC_SUBBANDS 32
43 #define SUBSUBFRAMES 2
44 #define SUBBAND_SAMPLES (SUBFRAMES * SUBSUBFRAMES * 8)
47 typedef struct DCAEncContext {
51 int fullband_channels;
57 const int32_t *band_interpolation;
58 const int32_t *band_spectrum;
62 const int8_t *channel_order_tab; ///< channel reordering table, lfe and non lfe
64 int32_t history[MAX_CHANNELS][512]; /* This is a circular buffer */
65 int32_t subband[MAX_CHANNELS][DCAENC_SUBBANDS][SUBBAND_SAMPLES];
66 int32_t quantized[MAX_CHANNELS][DCAENC_SUBBANDS][SUBBAND_SAMPLES];
67 int32_t peak_cb[MAX_CHANNELS][DCAENC_SUBBANDS];
68 int32_t downsampled_lfe[DCA_LFE_SAMPLES];
69 int32_t masking_curve_cb[SUBSUBFRAMES][256];
70 int32_t bit_allocation_sel[MAX_CHANNELS];
71 int abits[MAX_CHANNELS][DCAENC_SUBBANDS];
72 int scale_factor[MAX_CHANNELS][DCAENC_SUBBANDS];
73 softfloat quant[MAX_CHANNELS][DCAENC_SUBBANDS];
74 int32_t quant_index_sel[MAX_CHANNELS][DCA_CODE_BOOKS];
75 int32_t eff_masking_curve_cb[256];
76 int32_t band_masking_cb[32];
77 int32_t worst_quantization_noise;
78 int32_t worst_noise_ever;
82 static int32_t cos_table[2048];
83 static int32_t band_interpolation[2][512];
84 static int32_t band_spectrum[2][8];
85 static int32_t auf[9][AUBANDS][256];
86 static int32_t cb_to_add[256];
87 static int32_t cb_to_level[2048];
88 static int32_t lfe_fir_64i[512];
90 /* Transfer function of outer and middle ear, Hz -> dB */
91 static double hom(double f)
95 return -3.64 * pow(f1, -0.8)
96 + 6.8 * exp(-0.6 * (f1 - 3.4) * (f1 - 3.4))
97 - 6.0 * exp(-0.15 * (f1 - 8.7) * (f1 - 8.7))
98 - 0.0006 * (f1 * f1) * (f1 * f1);
101 static double gammafilter(int i, double f)
103 double h = (f - fc[i]) / erb[i];
107 return 20 * log10(h);
110 static int encode_init(AVCodecContext *avctx)
112 DCAEncContext *c = avctx->priv_data;
113 uint64_t layout = avctx->channel_layout;
114 int i, j, min_frame_bits;
116 c->fullband_channels = c->channels = avctx->channels;
117 c->lfe_channel = (avctx->channels == 3 || avctx->channels == 6);
118 c->band_interpolation = band_interpolation[1];
119 c->band_spectrum = band_spectrum[1];
120 c->worst_quantization_noise = -2047;
121 c->worst_noise_ever = -2047;
124 av_log(avctx, AV_LOG_WARNING, "No channel layout specified. The "
125 "encoder will guess the layout, but it "
126 "might be incorrect.\n");
127 layout = av_get_default_channel_layout(avctx->channels);
130 case AV_CH_LAYOUT_MONO: c->channel_config = 0; break;
131 case AV_CH_LAYOUT_STEREO: c->channel_config = 2; break;
132 case AV_CH_LAYOUT_2_2: c->channel_config = 8; break;
133 case AV_CH_LAYOUT_5POINT0: c->channel_config = 9; break;
134 case AV_CH_LAYOUT_5POINT1: c->channel_config = 9; break;
136 av_log(avctx, AV_LOG_ERROR, "Unsupported channel layout!\n");
137 return AVERROR_PATCHWELCOME;
140 if (c->lfe_channel) {
141 c->fullband_channels--;
142 c->channel_order_tab = channel_reorder_lfe[c->channel_config];
144 c->channel_order_tab = channel_reorder_nolfe[c->channel_config];
147 for (i = 0; i < MAX_CHANNELS; i++) {
148 for (j = 0; j < DCA_CODE_BOOKS; j++) {
149 c->quant_index_sel[i][j] = ff_dca_quant_index_group_size[j];
152 c->bit_allocation_sel[i] = 6;
155 for (i = 0; i < 9; i++) {
156 if (sample_rates[i] == avctx->sample_rate)
160 return AVERROR(EINVAL);
161 c->samplerate_index = i;
163 if (avctx->bit_rate < 32000 || avctx->bit_rate > 3840000) {
164 av_log(avctx, AV_LOG_ERROR, "Bit rate %"PRId64" not supported.", (int64_t)avctx->bit_rate);
165 return AVERROR(EINVAL);
167 for (i = 0; ff_dca_bit_rates[i] < avctx->bit_rate; i++)
169 c->bitrate_index = i;
170 c->frame_bits = FFALIGN((avctx->bit_rate * 512 + avctx->sample_rate - 1) / avctx->sample_rate, 32);
171 min_frame_bits = 132 + (493 + 28 * 32) * c->fullband_channels + c->lfe_channel * 72;
172 if (c->frame_bits < min_frame_bits || c->frame_bits > (DCA_MAX_FRAME_SIZE << 3))
173 return AVERROR(EINVAL);
175 c->frame_size = (c->frame_bits + 7) / 8;
177 avctx->frame_size = 32 * SUBBAND_SAMPLES;
182 cos_table[0] = 0x7fffffff;
184 cos_table[1024] = -cos_table[0];
185 for (i = 1; i < 512; i++) {
186 cos_table[i] = (int32_t)(0x7fffffff * cos(M_PI * i / 1024));
187 cos_table[1024-i] = -cos_table[i];
188 cos_table[1024+i] = -cos_table[i];
189 cos_table[2048-i] = cos_table[i];
191 for (i = 0; i < 2048; i++) {
192 cb_to_level[i] = (int32_t)(0x7fffffff * ff_exp10(-0.005 * i));
195 for (k = 0; k < 32; k++) {
196 for (j = 0; j < 8; j++) {
197 lfe_fir_64i[64 * j + k] = (int32_t)(0xffffff800000ULL * ff_dca_lfe_fir_64[8 * k + j]);
198 lfe_fir_64i[64 * (7-j) + (63 - k)] = (int32_t)(0xffffff800000ULL * ff_dca_lfe_fir_64[8 * k + j]);
202 for (i = 0; i < 512; i++) {
203 band_interpolation[0][i] = (int32_t)(0x1000000000ULL * ff_dca_fir_32bands_perfect[i]);
204 band_interpolation[1][i] = (int32_t)(0x1000000000ULL * ff_dca_fir_32bands_nonperfect[i]);
207 for (i = 0; i < 9; i++) {
208 for (j = 0; j < AUBANDS; j++) {
209 for (k = 0; k < 256; k++) {
210 double freq = sample_rates[i] * (k + 0.5) / 512;
212 auf[i][j][k] = (int32_t)(10 * (hom(freq) + gammafilter(j, freq)));
217 for (i = 0; i < 256; i++) {
218 double add = 1 + ff_exp10(-0.01 * i);
219 cb_to_add[i] = (int32_t)(100 * log10(add));
221 for (j = 0; j < 8; j++) {
223 for (i = 0; i < 512; i++) {
224 double reconst = ff_dca_fir_32bands_perfect[i] * ((i & 64) ? (-1) : 1);
225 accum += reconst * cos(2 * M_PI * (i + 0.5 - 256) * (j + 0.5) / 512);
227 band_spectrum[0][j] = (int32_t)(200 * log10(accum));
229 for (j = 0; j < 8; j++) {
231 for (i = 0; i < 512; i++) {
232 double reconst = ff_dca_fir_32bands_nonperfect[i] * ((i & 64) ? (-1) : 1);
233 accum += reconst * cos(2 * M_PI * (i + 0.5 - 256) * (j + 0.5) / 512);
235 band_spectrum[1][j] = (int32_t)(200 * log10(accum));
241 static inline int32_t cos_t(int x)
243 return cos_table[x & 2047];
246 static inline int32_t sin_t(int x)
248 return cos_t(x - 512);
251 static inline int32_t half32(int32_t a)
256 static inline int32_t mul32(int32_t a, int32_t b)
258 int64_t r = (int64_t)a * b + 0x80000000ULL;
262 static void subband_transform(DCAEncContext *c, const int32_t *input)
264 int ch, subs, i, k, j;
266 for (ch = 0; ch < c->fullband_channels; ch++) {
267 /* History is copied because it is also needed for PSY */
270 const int chi = c->channel_order_tab[ch];
272 memcpy(hist, &c->history[ch][0], 512 * sizeof(int32_t));
274 for (subs = 0; subs < SUBBAND_SAMPLES; subs++) {
279 /* Calculate the convolutions at once */
280 memset(accum, 0, 64 * sizeof(int32_t));
282 for (k = 0, i = hist_start, j = 0;
283 i < 512; k = (k + 1) & 63, i++, j++)
284 accum[k] += mul32(hist[i], c->band_interpolation[j]);
285 for (i = 0; i < hist_start; k = (k + 1) & 63, i++, j++)
286 accum[k] += mul32(hist[i], c->band_interpolation[j]);
288 for (k = 16; k < 32; k++)
289 accum[k] = accum[k] - accum[31 - k];
290 for (k = 32; k < 48; k++)
291 accum[k] = accum[k] + accum[95 - k];
293 for (band = 0; band < 32; band++) {
295 for (i = 16; i < 48; i++) {
296 int s = (2 * band + 1) * (2 * (i + 16) + 1);
297 resp += mul32(accum[i], cos_t(s << 3)) >> 3;
300 c->subband[ch][band][subs] = ((band + 1) & 2) ? -resp : resp;
303 /* Copy in 32 new samples from input */
304 for (i = 0; i < 32; i++)
305 hist[i + hist_start] = input[(subs * 32 + i) * c->channels + chi];
307 hist_start = (hist_start + 32) & 511;
312 static void lfe_downsample(DCAEncContext *c, const int32_t *input)
314 /* FIXME: make 128x LFE downsampling possible */
315 const int lfech = lfe_index[c->channel_config];
321 memcpy(hist, &c->history[c->channels - 1][0], 512 * sizeof(int32_t));
323 for (lfes = 0; lfes < DCA_LFE_SAMPLES; lfes++) {
324 /* Calculate the convolution */
327 for (i = hist_start, j = 0; i < 512; i++, j++)
328 accum += mul32(hist[i], lfe_fir_64i[j]);
329 for (i = 0; i < hist_start; i++, j++)
330 accum += mul32(hist[i], lfe_fir_64i[j]);
332 c->downsampled_lfe[lfes] = accum;
334 /* Copy in 64 new samples from input */
335 for (i = 0; i < 64; i++)
336 hist[i + hist_start] = input[(lfes * 64 + i) * c->channels + lfech];
338 hist_start = (hist_start + 64) & 511;
347 static void fft(const int32_t in[2 * 256], cplx32 out[256])
349 cplx32 buf[256], rin[256], rout[256];
352 /* do two transforms in parallel */
353 for (i = 0; i < 256; i++) {
354 /* Apply the Hann window */
355 rin[i].re = mul32(in[2 * i], 0x3fffffff - (cos_t(8 * i + 2) >> 1));
356 rin[i].im = mul32(in[2 * i + 1], 0x3fffffff - (cos_t(8 * i + 6) >> 1));
359 for (i = 0; i < 256; i++) {
360 buf[i].re = mul32(cos_t(4 * i + 2), rin[i].re)
361 - mul32(sin_t(4 * i + 2), rin[i].im);
362 buf[i].im = mul32(cos_t(4 * i + 2), rin[i].im)
363 + mul32(sin_t(4 * i + 2), rin[i].re);
366 for (j = 256, l = 1; j != 1; j >>= 1, l <<= 1) {
367 for (k = 0; k < 256; k += j) {
368 for (i = k; i < k + j / 2; i++) {
372 sum.re = buf[i].re + buf[i + j / 2].re;
373 sum.im = buf[i].im + buf[i + j / 2].im;
375 diff.re = buf[i].re - buf[i + j / 2].re;
376 diff.im = buf[i].im - buf[i + j / 2].im;
378 buf[i].re = half32(sum.re);
379 buf[i].im = half32(sum.im);
381 buf[i + j / 2].re = mul32(diff.re, cos_t(t))
382 - mul32(diff.im, sin_t(t));
383 buf[i + j / 2].im = mul32(diff.im, cos_t(t))
384 + mul32(diff.re, sin_t(t));
389 for (i = 0; i < 256; i++) {
390 int b = ff_reverse[i];
391 rout[i].re = mul32(buf[b].re, cos_t(4 * i))
392 - mul32(buf[b].im, sin_t(4 * i));
393 rout[i].im = mul32(buf[b].im, cos_t(4 * i))
394 + mul32(buf[b].re, sin_t(4 * i));
396 for (i = 0; i < 256; i++) {
397 /* separate the results of the two transforms */
400 o1.re = rout[i].re - rout[255 - i].re;
401 o1.im = rout[i].im + rout[255 - i].im;
403 o2.re = rout[i].im - rout[255 - i].im;
404 o2.im = -rout[i].re - rout[255 - i].re;
406 /* combine them into one long transform */
407 out[i].re = mul32( o1.re + o2.re, cos_t(2 * i + 1))
408 + mul32( o1.im - o2.im, sin_t(2 * i + 1));
409 out[i].im = mul32( o1.im + o2.im, cos_t(2 * i + 1))
410 + mul32(-o1.re + o2.re, sin_t(2 * i + 1));
414 static int32_t get_cb(int32_t in)
421 for (i = 1024; i > 0; i >>= 1) {
422 if (cb_to_level[i + res] >= in)
428 static int32_t add_cb(int32_t a, int32_t b)
431 FFSWAP(int32_t, a, b);
435 return a + cb_to_add[a - b];
438 static void adjust_jnd(int samplerate_index,
439 const int32_t in[512], int32_t out_cb[256])
443 int32_t out_cb_unnorm[256];
445 const int32_t ca_cb = -1114;
446 const int32_t cs_cb = 928;
451 for (j = 0; j < 256; j++) {
452 power[j] = add_cb(get_cb(out[j].re), get_cb(out[j].im));
453 out_cb_unnorm[j] = -2047; /* and can only grow */
456 for (i = 0; i < AUBANDS; i++) {
457 denom = ca_cb; /* and can only grow */
458 for (j = 0; j < 256; j++)
459 denom = add_cb(denom, power[j] + auf[samplerate_index][i][j]);
460 for (j = 0; j < 256; j++)
461 out_cb_unnorm[j] = add_cb(out_cb_unnorm[j],
462 -denom + auf[samplerate_index][i][j]);
465 for (j = 0; j < 256; j++)
466 out_cb[j] = add_cb(out_cb[j], -out_cb_unnorm[j] - ca_cb - cs_cb);
469 typedef void (*walk_band_t)(DCAEncContext *c, int band1, int band2, int f,
470 int32_t spectrum1, int32_t spectrum2, int channel,
473 static void walk_band_low(DCAEncContext *c, int band, int channel,
474 walk_band_t walk, int32_t *arg)
479 for (f = 0; f < 4; f++)
480 walk(c, 0, 0, f, 0, -2047, channel, arg);
482 for (f = 0; f < 8; f++)
483 walk(c, band, band - 1, 8 * band - 4 + f,
484 c->band_spectrum[7 - f], c->band_spectrum[f], channel, arg);
488 static void walk_band_high(DCAEncContext *c, int band, int channel,
489 walk_band_t walk, int32_t *arg)
494 for (f = 0; f < 4; f++)
495 walk(c, 31, 31, 256 - 4 + f, 0, -2047, channel, arg);
497 for (f = 0; f < 8; f++)
498 walk(c, band, band + 1, 8 * band + 4 + f,
499 c->band_spectrum[f], c->band_spectrum[7 - f], channel, arg);
503 static void update_band_masking(DCAEncContext *c, int band1, int band2,
504 int f, int32_t spectrum1, int32_t spectrum2,
505 int channel, int32_t * arg)
507 int32_t value = c->eff_masking_curve_cb[f] - spectrum1;
509 if (value < c->band_masking_cb[band1])
510 c->band_masking_cb[band1] = value;
513 static void calc_masking(DCAEncContext *c, const int32_t *input)
515 int i, k, band, ch, ssf;
518 for (i = 0; i < 256; i++)
519 for (ssf = 0; ssf < SUBSUBFRAMES; ssf++)
520 c->masking_curve_cb[ssf][i] = -2047;
522 for (ssf = 0; ssf < SUBSUBFRAMES; ssf++)
523 for (ch = 0; ch < c->fullband_channels; ch++) {
524 const int chi = c->channel_order_tab[ch];
526 for (i = 0, k = 128 + 256 * ssf; k < 512; i++, k++)
527 data[i] = c->history[ch][k];
528 for (k -= 512; i < 512; i++, k++)
529 data[i] = input[k * c->channels + chi];
530 adjust_jnd(c->samplerate_index, data, c->masking_curve_cb[ssf]);
532 for (i = 0; i < 256; i++) {
535 for (ssf = 0; ssf < SUBSUBFRAMES; ssf++)
536 if (c->masking_curve_cb[ssf][i] < m)
537 m = c->masking_curve_cb[ssf][i];
538 c->eff_masking_curve_cb[i] = m;
541 for (band = 0; band < 32; band++) {
542 c->band_masking_cb[band] = 2048;
543 walk_band_low(c, band, 0, update_band_masking, NULL);
544 walk_band_high(c, band, 0, update_band_masking, NULL);
548 static void find_peaks(DCAEncContext *c)
552 for (ch = 0; ch < c->fullband_channels; ch++)
553 for (band = 0; band < 32; band++) {
557 for (sample = 0; sample < SUBBAND_SAMPLES; sample++) {
558 int32_t s = abs(c->subband[ch][band][sample]);
562 c->peak_cb[ch][band] = get_cb(m);
565 if (c->lfe_channel) {
569 for (sample = 0; sample < DCA_LFE_SAMPLES; sample++)
570 if (m < abs(c->downsampled_lfe[sample]))
571 m = abs(c->downsampled_lfe[sample]);
572 c->lfe_peak_cb = get_cb(m);
576 static const int snr_fudge = 128;
577 #define USED_1ABITS 1
578 #define USED_NABITS 2
579 #define USED_26ABITS 4
581 static int32_t quantize_value(int32_t value, softfloat quant)
583 int32_t offset = 1 << (quant.e - 1);
585 value = mul32(value, quant.m) + offset;
586 value = value >> quant.e;
590 static int calc_one_scale(int32_t peak_cb, int abits, softfloat *quant)
593 int our_nscale, try_remove;
596 av_assert0(peak_cb <= 0);
597 av_assert0(peak_cb >= -2047);
600 peak = cb_to_level[-peak_cb];
602 for (try_remove = 64; try_remove > 0; try_remove >>= 1) {
603 if (scalefactor_inv[our_nscale - try_remove].e + stepsize_inv[abits].e <= 17)
605 our_quant.m = mul32(scalefactor_inv[our_nscale - try_remove].m, stepsize_inv[abits].m);
606 our_quant.e = scalefactor_inv[our_nscale - try_remove].e + stepsize_inv[abits].e - 17;
607 if ((ff_dca_quant_levels[abits] - 1) / 2 < quantize_value(peak, our_quant))
609 our_nscale -= try_remove;
612 if (our_nscale >= 125)
615 quant->m = mul32(scalefactor_inv[our_nscale].m, stepsize_inv[abits].m);
616 quant->e = scalefactor_inv[our_nscale].e + stepsize_inv[abits].e - 17;
617 av_assert0((ff_dca_quant_levels[abits] - 1) / 2 >= quantize_value(peak, *quant));
622 static void quantize_all(DCAEncContext *c)
624 int sample, band, ch;
626 for (ch = 0; ch < c->fullband_channels; ch++)
627 for (band = 0; band < 32; band++)
628 for (sample = 0; sample < SUBBAND_SAMPLES; sample++)
629 c->quantized[ch][band][sample] = quantize_value(c->subband[ch][band][sample], c->quant[ch][band]);
632 static void accumulate_huff_bit_consumption(int abits, int32_t *quantized, uint32_t *result)
634 uint8_t sel, id = abits - 1;
635 for (sel = 0; sel < ff_dca_quant_index_group_size[id]; sel++)
636 result[sel] += ff_dca_vlc_calc_quant_bits(quantized, SUBBAND_SAMPLES, sel, id);
639 static uint32_t set_best_code(uint32_t vlc_bits[DCA_CODE_BOOKS][7], uint32_t clc_bits[DCA_CODE_BOOKS], int32_t res[DCA_CODE_BOOKS])
642 uint32_t best_sel_bits[DCA_CODE_BOOKS];
643 int32_t best_sel_id[DCA_CODE_BOOKS];
644 uint32_t t, bits = 0;
646 for (i = 0; i < DCA_CODE_BOOKS; i++) {
648 av_assert0(!((!!vlc_bits[i][0]) ^ (!!clc_bits[i])));
649 if (vlc_bits[i][0] == 0) {
650 /* do not transmit adjustment index for empty codebooks */
651 res[i] = ff_dca_quant_index_group_size[i];
656 best_sel_bits[i] = vlc_bits[i][0];
658 for (sel = 0; sel < ff_dca_quant_index_group_size[i]; sel++) {
659 if (best_sel_bits[i] > vlc_bits[i][sel] && vlc_bits[i][sel]) {
660 best_sel_bits[i] = vlc_bits[i][sel];
661 best_sel_id[i] = sel;
665 /* 2 bits to transmit scale factor adjustment index */
666 t = best_sel_bits[i] + 2;
667 if (t < clc_bits[i]) {
668 res[i] = best_sel_id[i];
671 res[i] = ff_dca_quant_index_group_size[i];
678 static uint32_t set_best_abits_code(int abits[DCAENC_SUBBANDS], int bands, int32_t *res)
682 int32_t best_sel = 6;
683 int32_t best_bits = bands * 5;
685 /* Check do we have subband which cannot be encoded by Huffman tables */
686 for (i = 0; i < bands; i++) {
693 for (i = 0; i < DCA_BITALLOC_12_COUNT; i++) {
694 t = ff_dca_vlc_calc_alloc_bits(abits, bands, i);
705 static int init_quantization_noise(DCAEncContext *c, int noise)
707 int ch, band, ret = 0;
708 uint32_t huff_bit_count_accum[MAX_CHANNELS][DCA_CODE_BOOKS][7];
709 uint32_t clc_bit_count_accum[MAX_CHANNELS][DCA_CODE_BOOKS];
710 uint32_t bits_counter = 0;
712 c->consumed_bits = 132 + 333 * c->fullband_channels;
714 c->consumed_bits += 72;
716 /* attempt to guess the bit distribution based on the prevoius frame */
717 for (ch = 0; ch < c->fullband_channels; ch++) {
718 for (band = 0; band < 32; band++) {
719 int snr_cb = c->peak_cb[ch][band] - c->band_masking_cb[band] - noise;
721 if (snr_cb >= 1312) {
722 c->abits[ch][band] = 26;
724 } else if (snr_cb >= 222) {
725 c->abits[ch][band] = 8 + mul32(snr_cb - 222, 69000000);
727 } else if (snr_cb >= 0) {
728 c->abits[ch][band] = 2 + mul32(snr_cb, 106000000);
731 c->abits[ch][band] = 1;
735 c->consumed_bits += set_best_abits_code(c->abits[ch], 32, &c->bit_allocation_sel[ch]);
738 /* Recalc scale_factor each time to get bits consumption in case of Huffman coding.
739 It is suboptimal solution */
740 /* TODO: May be cache scaled values */
741 for (ch = 0; ch < c->fullband_channels; ch++) {
742 for (band = 0; band < 32; band++) {
743 c->scale_factor[ch][band] = calc_one_scale(c->peak_cb[ch][band],
745 &c->quant[ch][band]);
750 memset(huff_bit_count_accum, 0, MAX_CHANNELS * DCA_CODE_BOOKS * 7 * sizeof(uint32_t));
751 memset(clc_bit_count_accum, 0, MAX_CHANNELS * DCA_CODE_BOOKS * sizeof(uint32_t));
752 for (ch = 0; ch < c->fullband_channels; ch++) {
753 for (band = 0; band < 32; band++) {
754 if (c->abits[ch][band] && c->abits[ch][band] <= DCA_CODE_BOOKS) {
755 accumulate_huff_bit_consumption(c->abits[ch][band], c->quantized[ch][band], huff_bit_count_accum[ch][c->abits[ch][band] - 1]);
756 clc_bit_count_accum[ch][c->abits[ch][band] - 1] += bit_consumption[c->abits[ch][band]];
758 bits_counter += bit_consumption[c->abits[ch][band]];
763 for (ch = 0; ch < c->fullband_channels; ch++) {
764 bits_counter += set_best_code(huff_bit_count_accum[ch], clc_bit_count_accum[ch], c->quant_index_sel[ch]);
767 c->consumed_bits += bits_counter;
772 static void assign_bits(DCAEncContext *c)
774 /* Find the bounds where the binary search should work */
778 init_quantization_noise(c, c->worst_quantization_noise);
779 low = high = c->worst_quantization_noise;
780 if (c->consumed_bits > c->frame_bits) {
781 while (c->consumed_bits > c->frame_bits) {
782 av_assert0(used_abits != USED_1ABITS);
785 used_abits = init_quantization_noise(c, high);
788 while (c->consumed_bits <= c->frame_bits) {
790 if (used_abits == USED_26ABITS)
791 goto out; /* The requested bitrate is too high, pad with zeros */
793 used_abits = init_quantization_noise(c, low);
797 /* Now do a binary search between low and high to see what fits */
798 for (down = snr_fudge >> 1; down; down >>= 1) {
799 init_quantization_noise(c, high - down);
800 if (c->consumed_bits <= c->frame_bits)
803 init_quantization_noise(c, high);
805 c->worst_quantization_noise = high;
806 if (high > c->worst_noise_ever)
807 c->worst_noise_ever = high;
810 static void shift_history(DCAEncContext *c, const int32_t *input)
814 for (k = 0; k < 512; k++)
815 for (ch = 0; ch < c->channels; ch++) {
816 const int chi = c->channel_order_tab[ch];
818 c->history[ch][k] = input[k * c->channels + chi];
822 static void calc_lfe_scales(DCAEncContext *c)
825 c->lfe_scale_factor = calc_one_scale(c->lfe_peak_cb, 11, &c->lfe_quant);
828 static void put_frame_header(DCAEncContext *c)
831 put_bits(&c->pb, 16, 0x7ffe);
832 put_bits(&c->pb, 16, 0x8001);
834 /* Frame type: normal */
835 put_bits(&c->pb, 1, 1);
837 /* Deficit sample count: none */
838 put_bits(&c->pb, 5, 31);
840 /* CRC is not present */
841 put_bits(&c->pb, 1, 0);
843 /* Number of PCM sample blocks */
844 put_bits(&c->pb, 7, SUBBAND_SAMPLES - 1);
846 /* Primary frame byte size */
847 put_bits(&c->pb, 14, c->frame_size - 1);
849 /* Audio channel arrangement */
850 put_bits(&c->pb, 6, c->channel_config);
852 /* Core audio sampling frequency */
853 put_bits(&c->pb, 4, bitstream_sfreq[c->samplerate_index]);
855 /* Transmission bit rate */
856 put_bits(&c->pb, 5, c->bitrate_index);
858 /* Embedded down mix: disabled */
859 put_bits(&c->pb, 1, 0);
861 /* Embedded dynamic range flag: not present */
862 put_bits(&c->pb, 1, 0);
864 /* Embedded time stamp flag: not present */
865 put_bits(&c->pb, 1, 0);
867 /* Auxiliary data flag: not present */
868 put_bits(&c->pb, 1, 0);
870 /* HDCD source: no */
871 put_bits(&c->pb, 1, 0);
873 /* Extension audio ID: N/A */
874 put_bits(&c->pb, 3, 0);
876 /* Extended audio data: not present */
877 put_bits(&c->pb, 1, 0);
879 /* Audio sync word insertion flag: after each sub-frame */
880 put_bits(&c->pb, 1, 0);
882 /* Low frequency effects flag: not present or 64x subsampling */
883 put_bits(&c->pb, 2, c->lfe_channel ? 2 : 0);
885 /* Predictor history switch flag: on */
886 put_bits(&c->pb, 1, 1);
889 /* Multirate interpolator switch: non-perfect reconstruction */
890 put_bits(&c->pb, 1, 0);
892 /* Encoder software revision: 7 */
893 put_bits(&c->pb, 4, 7);
895 /* Copy history: 0 */
896 put_bits(&c->pb, 2, 0);
898 /* Source PCM resolution: 16 bits, not DTS ES */
899 put_bits(&c->pb, 3, 0);
901 /* Front sum/difference coding: no */
902 put_bits(&c->pb, 1, 0);
904 /* Surrounds sum/difference coding: no */
905 put_bits(&c->pb, 1, 0);
907 /* Dialog normalization: 0 dB */
908 put_bits(&c->pb, 4, 0);
911 static void put_primary_audio_header(DCAEncContext *c)
914 /* Number of subframes */
915 put_bits(&c->pb, 4, SUBFRAMES - 1);
917 /* Number of primary audio channels */
918 put_bits(&c->pb, 3, c->fullband_channels - 1);
920 /* Subband activity count */
921 for (ch = 0; ch < c->fullband_channels; ch++)
922 put_bits(&c->pb, 5, DCAENC_SUBBANDS - 2);
924 /* High frequency VQ start subband */
925 for (ch = 0; ch < c->fullband_channels; ch++)
926 put_bits(&c->pb, 5, DCAENC_SUBBANDS - 1);
928 /* Joint intensity coding index: 0, 0 */
929 for (ch = 0; ch < c->fullband_channels; ch++)
930 put_bits(&c->pb, 3, 0);
932 /* Transient mode codebook: A4, A4 (arbitrary) */
933 for (ch = 0; ch < c->fullband_channels; ch++)
934 put_bits(&c->pb, 2, 0);
936 /* Scale factor code book: 7 bit linear, 7-bit sqrt table (for each channel) */
937 for (ch = 0; ch < c->fullband_channels; ch++)
938 put_bits(&c->pb, 3, 6);
940 /* Bit allocation quantizer select: linear 5-bit */
941 for (ch = 0; ch < c->fullband_channels; ch++)
942 put_bits(&c->pb, 3, c->bit_allocation_sel[ch]);
944 /* Quantization index codebook select */
945 for (i = 0; i < DCA_CODE_BOOKS; i++)
946 for (ch = 0; ch < c->fullband_channels; ch++)
947 put_bits(&c->pb, ff_dca_quant_index_sel_nbits[i], c->quant_index_sel[ch][i]);
949 /* Scale factor adjustment index: transmitted in case of Huffman coding */
950 for (i = 0; i < DCA_CODE_BOOKS; i++)
951 for (ch = 0; ch < c->fullband_channels; ch++)
952 if (c->quant_index_sel[ch][i] < ff_dca_quant_index_group_size[i])
953 put_bits(&c->pb, 2, 0);
955 /* Audio header CRC check word: not transmitted */
958 static void put_subframe_samples(DCAEncContext *c, int ss, int band, int ch)
960 int i, j, sum, bits, sel;
961 if (c->abits[ch][band] <= DCA_CODE_BOOKS) {
962 av_assert0(c->abits[ch][band] > 0);
963 sel = c->quant_index_sel[ch][c->abits[ch][band] - 1];
965 if (sel < ff_dca_quant_index_group_size[c->abits[ch][band] - 1]) {
966 ff_dca_vlc_enc_quant(&c->pb, &c->quantized[ch][band][ss * 8], 8, sel, c->abits[ch][band] - 1);
971 if (c->abits[ch][band] <= 7) {
972 for (i = 0; i < 8; i += 4) {
974 for (j = 3; j >= 0; j--) {
975 sum *= ff_dca_quant_levels[c->abits[ch][band]];
976 sum += c->quantized[ch][band][ss * 8 + i + j];
977 sum += (ff_dca_quant_levels[c->abits[ch][band]] - 1) / 2;
979 put_bits(&c->pb, bit_consumption[c->abits[ch][band]] / 4, sum);
985 for (i = 0; i < 8; i++) {
986 bits = bit_consumption[c->abits[ch][band]] / 16;
987 put_sbits(&c->pb, bits, c->quantized[ch][band][ss * 8 + i]);
991 static void put_subframe(DCAEncContext *c, int subframe)
995 /* Subsubframes count */
996 put_bits(&c->pb, 2, SUBSUBFRAMES -1);
998 /* Partial subsubframe sample count: dummy */
999 put_bits(&c->pb, 3, 0);
1001 /* Prediction mode: no ADPCM, in each channel and subband */
1002 for (ch = 0; ch < c->fullband_channels; ch++)
1003 for (band = 0; band < DCAENC_SUBBANDS; band++)
1004 put_bits(&c->pb, 1, 0);
1006 /* Prediction VQ address: not transmitted */
1007 /* Bit allocation index */
1008 for (ch = 0; ch < c->fullband_channels; ch++) {
1009 if (c->bit_allocation_sel[ch] == 6) {
1010 for (band = 0; band < DCAENC_SUBBANDS; band++) {
1011 put_bits(&c->pb, 5, c->abits[ch][band]);
1014 ff_dca_vlc_enc_alloc(&c->pb, c->abits[ch], DCAENC_SUBBANDS, c->bit_allocation_sel[ch]);
1018 if (SUBSUBFRAMES > 1) {
1019 /* Transition mode: none for each channel and subband */
1020 for (ch = 0; ch < c->fullband_channels; ch++)
1021 for (band = 0; band < DCAENC_SUBBANDS; band++)
1022 put_bits(&c->pb, 1, 0); /* codebook A4 */
1026 for (ch = 0; ch < c->fullband_channels; ch++)
1027 for (band = 0; band < DCAENC_SUBBANDS; band++)
1028 put_bits(&c->pb, 7, c->scale_factor[ch][band]);
1030 /* Joint subband scale factor codebook select: not transmitted */
1031 /* Scale factors for joint subband coding: not transmitted */
1032 /* Stereo down-mix coefficients: not transmitted */
1033 /* Dynamic range coefficient: not transmitted */
1034 /* Stde information CRC check word: not transmitted */
1035 /* VQ encoded high frequency subbands: not transmitted */
1037 /* LFE data: 8 samples and scalefactor */
1038 if (c->lfe_channel) {
1039 for (i = 0; i < DCA_LFE_SAMPLES; i++)
1040 put_bits(&c->pb, 8, quantize_value(c->downsampled_lfe[i], c->lfe_quant) & 0xff);
1041 put_bits(&c->pb, 8, c->lfe_scale_factor);
1044 /* Audio data (subsubframes) */
1045 for (ss = 0; ss < SUBSUBFRAMES ; ss++)
1046 for (ch = 0; ch < c->fullband_channels; ch++)
1047 for (band = 0; band < DCAENC_SUBBANDS; band++)
1048 put_subframe_samples(c, ss, band, ch);
1051 put_bits(&c->pb, 16, 0xffff);
1054 static int encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
1055 const AVFrame *frame, int *got_packet_ptr)
1057 DCAEncContext *c = avctx->priv_data;
1058 const int32_t *samples;
1061 if ((ret = ff_alloc_packet2(avctx, avpkt, c->frame_size, 0)) < 0)
1064 samples = (const int32_t *)frame->data[0];
1066 subband_transform(c, samples);
1068 lfe_downsample(c, samples);
1070 calc_masking(c, samples);
1074 shift_history(c, samples);
1076 init_put_bits(&c->pb, avpkt->data, avpkt->size);
1077 put_frame_header(c);
1078 put_primary_audio_header(c);
1079 for (i = 0; i < SUBFRAMES; i++)
1083 for (i = put_bits_count(&c->pb); i < 8*c->frame_size; i++)
1084 put_bits(&c->pb, 1, 0);
1086 flush_put_bits(&c->pb);
1088 avpkt->pts = frame->pts;
1089 avpkt->duration = ff_samples_to_time_base(avctx, frame->nb_samples);
1090 avpkt->size = put_bits_count(&c->pb) >> 3;
1091 *got_packet_ptr = 1;
1095 static const AVCodecDefault defaults[] = {
1100 AVCodec ff_dca_encoder = {
1102 .long_name = NULL_IF_CONFIG_SMALL("DCA (DTS Coherent Acoustics)"),
1103 .type = AVMEDIA_TYPE_AUDIO,
1104 .id = AV_CODEC_ID_DTS,
1105 .priv_data_size = sizeof(DCAEncContext),
1106 .init = encode_init,
1107 .encode2 = encode_frame,
1108 .capabilities = AV_CODEC_CAP_EXPERIMENTAL,
1109 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_S32,
1110 AV_SAMPLE_FMT_NONE },
1111 .supported_samplerates = sample_rates,
1112 .channel_layouts = (const uint64_t[]) { AV_CH_LAYOUT_MONO,
1113 AV_CH_LAYOUT_STEREO,
1115 AV_CH_LAYOUT_5POINT0,
1116 AV_CH_LAYOUT_5POINT1,
1118 .defaults = defaults,