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 typedef struct CompressionOptions {
59 typedef struct DCAEncContext {
62 DCAADPCMEncContext adpcm_ctx;
64 CompressionOptions options;
67 int fullband_channels;
73 const int32_t *band_interpolation;
74 const int32_t *band_spectrum;
78 const int8_t *channel_order_tab; ///< channel reordering table, lfe and non lfe
80 int32_t prediction_mode[MAX_CHANNELS][DCAENC_SUBBANDS];
81 int32_t adpcm_history[MAX_CHANNELS][DCAENC_SUBBANDS][DCA_ADPCM_COEFFS * 2];
82 int32_t history[MAX_CHANNELS][512]; /* This is a circular buffer */
83 int32_t *subband[MAX_CHANNELS][DCAENC_SUBBANDS];
84 int32_t quantized[MAX_CHANNELS][DCAENC_SUBBANDS][SUBBAND_SAMPLES];
85 int32_t peak_cb[MAX_CHANNELS][DCAENC_SUBBANDS];
86 int32_t diff_peak_cb[MAX_CHANNELS][DCAENC_SUBBANDS]; ///< expected peak of residual signal
87 int32_t downsampled_lfe[DCA_LFE_SAMPLES];
88 int32_t masking_curve_cb[SUBSUBFRAMES][256];
89 int32_t bit_allocation_sel[MAX_CHANNELS];
90 int abits[MAX_CHANNELS][DCAENC_SUBBANDS];
91 int scale_factor[MAX_CHANNELS][DCAENC_SUBBANDS];
92 softfloat quant[MAX_CHANNELS][DCAENC_SUBBANDS];
93 int32_t quant_index_sel[MAX_CHANNELS][DCA_CODE_BOOKS];
94 int32_t eff_masking_curve_cb[256];
95 int32_t band_masking_cb[32];
96 int32_t worst_quantization_noise;
97 int32_t worst_noise_ever;
99 int consumed_adpcm_bits; ///< Number of bits to transmit ADPCM related info
102 static int32_t cos_table[2048];
103 static int32_t band_interpolation[2][512];
104 static int32_t band_spectrum[2][8];
105 static int32_t auf[9][AUBANDS][256];
106 static int32_t cb_to_add[256];
107 static int32_t cb_to_level[2048];
108 static int32_t lfe_fir_64i[512];
110 /* Transfer function of outer and middle ear, Hz -> dB */
111 static double hom(double f)
113 double f1 = f / 1000;
115 return -3.64 * pow(f1, -0.8)
116 + 6.8 * exp(-0.6 * (f1 - 3.4) * (f1 - 3.4))
117 - 6.0 * exp(-0.15 * (f1 - 8.7) * (f1 - 8.7))
118 - 0.0006 * (f1 * f1) * (f1 * f1);
121 static double gammafilter(int i, double f)
123 double h = (f - fc[i]) / erb[i];
127 return 20 * log10(h);
130 static int subband_bufer_alloc(DCAEncContext *c)
133 int32_t *bufer = av_calloc(MAX_CHANNELS * DCAENC_SUBBANDS *
134 (SUBBAND_SAMPLES + DCA_ADPCM_COEFFS),
139 /* we need a place for DCA_ADPCM_COEFF samples from previous frame
140 * to calc prediction coefficients for each subband */
141 for (ch = 0; ch < MAX_CHANNELS; ch++) {
142 for (band = 0; band < DCAENC_SUBBANDS; band++) {
143 c->subband[ch][band] = bufer +
144 ch * DCAENC_SUBBANDS * (SUBBAND_SAMPLES + DCA_ADPCM_COEFFS) +
145 band * (SUBBAND_SAMPLES + DCA_ADPCM_COEFFS) + DCA_ADPCM_COEFFS;
151 static void subband_bufer_free(DCAEncContext *c)
153 int32_t *bufer = c->subband[0][0] - DCA_ADPCM_COEFFS;
157 static int encode_init(AVCodecContext *avctx)
159 DCAEncContext *c = avctx->priv_data;
160 uint64_t layout = avctx->channel_layout;
161 int i, j, min_frame_bits;
164 if (subband_bufer_alloc(c))
165 return AVERROR(ENOMEM);
167 c->fullband_channels = c->channels = avctx->channels;
168 c->lfe_channel = (avctx->channels == 3 || avctx->channels == 6);
169 c->band_interpolation = band_interpolation[1];
170 c->band_spectrum = band_spectrum[1];
171 c->worst_quantization_noise = -2047;
172 c->worst_noise_ever = -2047;
173 c->consumed_adpcm_bits = 0;
175 if (ff_dcaadpcm_init(&c->adpcm_ctx))
176 return AVERROR(ENOMEM);
179 av_log(avctx, AV_LOG_WARNING, "No channel layout specified. The "
180 "encoder will guess the layout, but it "
181 "might be incorrect.\n");
182 layout = av_get_default_channel_layout(avctx->channels);
185 case AV_CH_LAYOUT_MONO: c->channel_config = 0; break;
186 case AV_CH_LAYOUT_STEREO: c->channel_config = 2; break;
187 case AV_CH_LAYOUT_2_2: c->channel_config = 8; break;
188 case AV_CH_LAYOUT_5POINT0: c->channel_config = 9; break;
189 case AV_CH_LAYOUT_5POINT1: c->channel_config = 9; break;
191 av_log(avctx, AV_LOG_ERROR, "Unsupported channel layout!\n");
192 return AVERROR_PATCHWELCOME;
195 if (c->lfe_channel) {
196 c->fullband_channels--;
197 c->channel_order_tab = channel_reorder_lfe[c->channel_config];
199 c->channel_order_tab = channel_reorder_nolfe[c->channel_config];
202 for (i = 0; i < MAX_CHANNELS; i++) {
203 for (j = 0; j < DCA_CODE_BOOKS; j++) {
204 c->quant_index_sel[i][j] = ff_dca_quant_index_group_size[j];
207 c->bit_allocation_sel[i] = 6;
209 for (j = 0; j < DCAENC_SUBBANDS; j++) {
211 c->prediction_mode[i][j] = -1;
212 memset(c->adpcm_history[i][j], 0, sizeof(int32_t)*DCA_ADPCM_COEFFS);
216 for (i = 0; i < 9; i++) {
217 if (sample_rates[i] == avctx->sample_rate)
221 return AVERROR(EINVAL);
222 c->samplerate_index = i;
224 if (avctx->bit_rate < 32000 || avctx->bit_rate > 3840000) {
225 av_log(avctx, AV_LOG_ERROR, "Bit rate %"PRId64" not supported.", avctx->bit_rate);
226 return AVERROR(EINVAL);
228 for (i = 0; ff_dca_bit_rates[i] < avctx->bit_rate; i++)
230 c->bitrate_index = i;
231 c->frame_bits = FFALIGN((avctx->bit_rate * 512 + avctx->sample_rate - 1) / avctx->sample_rate, 32);
232 min_frame_bits = 132 + (493 + 28 * 32) * c->fullband_channels + c->lfe_channel * 72;
233 if (c->frame_bits < min_frame_bits || c->frame_bits > (DCA_MAX_FRAME_SIZE << 3))
234 return AVERROR(EINVAL);
236 c->frame_size = (c->frame_bits + 7) / 8;
238 avctx->frame_size = 32 * SUBBAND_SAMPLES;
240 if ((ret = ff_mdct_init(&c->mdct, 9, 0, 1.0)) < 0)
246 cos_table[0] = 0x7fffffff;
248 cos_table[1024] = -cos_table[0];
249 for (i = 1; i < 512; i++) {
250 cos_table[i] = (int32_t)(0x7fffffff * cos(M_PI * i / 1024));
251 cos_table[1024-i] = -cos_table[i];
252 cos_table[1024+i] = -cos_table[i];
253 cos_table[2048-i] = cos_table[i];
255 for (i = 0; i < 2048; i++) {
256 cb_to_level[i] = (int32_t)(0x7fffffff * ff_exp10(-0.005 * i));
259 for (k = 0; k < 32; k++) {
260 for (j = 0; j < 8; j++) {
261 lfe_fir_64i[64 * j + k] = (int32_t)(0xffffff800000ULL * ff_dca_lfe_fir_64[8 * k + j]);
262 lfe_fir_64i[64 * (7-j) + (63 - k)] = (int32_t)(0xffffff800000ULL * ff_dca_lfe_fir_64[8 * k + j]);
266 for (i = 0; i < 512; i++) {
267 band_interpolation[0][i] = (int32_t)(0x1000000000ULL * ff_dca_fir_32bands_perfect[i]);
268 band_interpolation[1][i] = (int32_t)(0x1000000000ULL * ff_dca_fir_32bands_nonperfect[i]);
271 for (i = 0; i < 9; i++) {
272 for (j = 0; j < AUBANDS; j++) {
273 for (k = 0; k < 256; k++) {
274 double freq = sample_rates[i] * (k + 0.5) / 512;
276 auf[i][j][k] = (int32_t)(10 * (hom(freq) + gammafilter(j, freq)));
281 for (i = 0; i < 256; i++) {
282 double add = 1 + ff_exp10(-0.01 * i);
283 cb_to_add[i] = (int32_t)(100 * log10(add));
285 for (j = 0; j < 8; j++) {
287 for (i = 0; i < 512; i++) {
288 double reconst = ff_dca_fir_32bands_perfect[i] * ((i & 64) ? (-1) : 1);
289 accum += reconst * cos(2 * M_PI * (i + 0.5 - 256) * (j + 0.5) / 512);
291 band_spectrum[0][j] = (int32_t)(200 * log10(accum));
293 for (j = 0; j < 8; j++) {
295 for (i = 0; i < 512; i++) {
296 double reconst = ff_dca_fir_32bands_nonperfect[i] * ((i & 64) ? (-1) : 1);
297 accum += reconst * cos(2 * M_PI * (i + 0.5 - 256) * (j + 0.5) / 512);
299 band_spectrum[1][j] = (int32_t)(200 * log10(accum));
305 static av_cold int encode_close(AVCodecContext *avctx)
307 DCAEncContext *c = avctx->priv_data;
308 ff_mdct_end(&c->mdct);
309 subband_bufer_free(c);
310 ff_dcaadpcm_free(&c->adpcm_ctx);
315 static inline int32_t cos_t(int x)
317 return cos_table[x & 2047];
320 static void subband_transform(DCAEncContext *c, const int32_t *input)
322 int ch, subs, i, k, j;
324 for (ch = 0; ch < c->fullband_channels; ch++) {
325 /* History is copied because it is also needed for PSY */
328 const int chi = c->channel_order_tab[ch];
330 memcpy(hist, &c->history[ch][0], 512 * sizeof(int32_t));
332 for (subs = 0; subs < SUBBAND_SAMPLES; subs++) {
337 /* Calculate the convolutions at once */
338 memset(accum, 0, 64 * sizeof(int32_t));
340 for (k = 0, i = hist_start, j = 0;
341 i < 512; k = (k + 1) & 63, i++, j++)
342 accum[k] += mul32(hist[i], c->band_interpolation[j]);
343 for (i = 0; i < hist_start; k = (k + 1) & 63, i++, j++)
344 accum[k] += mul32(hist[i], c->band_interpolation[j]);
346 for (k = 16; k < 32; k++)
347 accum[k] = accum[k] - accum[31 - k];
348 for (k = 32; k < 48; k++)
349 accum[k] = accum[k] + accum[95 - k];
351 for (band = 0; band < 32; band++) {
353 for (i = 16; i < 48; i++) {
354 int s = (2 * band + 1) * (2 * (i + 16) + 1);
355 resp += mul32(accum[i], cos_t(s << 3)) >> 3;
358 c->subband[ch][band][subs] = ((band + 1) & 2) ? -resp : resp;
361 /* Copy in 32 new samples from input */
362 for (i = 0; i < 32; i++)
363 hist[i + hist_start] = input[(subs * 32 + i) * c->channels + chi];
365 hist_start = (hist_start + 32) & 511;
370 static void lfe_downsample(DCAEncContext *c, const int32_t *input)
372 /* FIXME: make 128x LFE downsampling possible */
373 const int lfech = lfe_index[c->channel_config];
379 memcpy(hist, &c->history[c->channels - 1][0], 512 * sizeof(int32_t));
381 for (lfes = 0; lfes < DCA_LFE_SAMPLES; lfes++) {
382 /* Calculate the convolution */
385 for (i = hist_start, j = 0; i < 512; i++, j++)
386 accum += mul32(hist[i], lfe_fir_64i[j]);
387 for (i = 0; i < hist_start; i++, j++)
388 accum += mul32(hist[i], lfe_fir_64i[j]);
390 c->downsampled_lfe[lfes] = accum;
392 /* Copy in 64 new samples from input */
393 for (i = 0; i < 64; i++)
394 hist[i + hist_start] = input[(lfes * 64 + i) * c->channels + lfech];
396 hist_start = (hist_start + 64) & 511;
400 static int32_t get_cb(int32_t in)
407 for (i = 1024; i > 0; i >>= 1) {
408 if (cb_to_level[i + res] >= in)
414 static int32_t add_cb(int32_t a, int32_t b)
417 FFSWAP(int32_t, a, b);
421 return a + cb_to_add[a - b];
424 static void calc_power(DCAEncContext *c,
425 const int32_t in[2 * 256], int32_t power[256])
428 LOCAL_ALIGNED_32(int32_t, data, [512]);
429 LOCAL_ALIGNED_32(int32_t, coeff, [256]);
431 for (i = 0; i < 512; i++) {
432 data[i] = norm__(mul32(in[i], 0x3fffffff - (cos_t(4 * i + 2) >> 1)), 4);
434 c->mdct.mdct_calc(&c->mdct, coeff, data);
435 for (i = 0; i < 256; i++) {
436 const int32_t cb = get_cb(coeff[i]);
437 power[i] = add_cb(cb, cb);
441 static void adjust_jnd(DCAEncContext *c,
442 const int32_t in[512], int32_t out_cb[256])
445 int32_t out_cb_unnorm[256];
447 const int32_t ca_cb = -1114;
448 const int32_t cs_cb = 928;
449 const int samplerate_index = c->samplerate_index;
452 calc_power(c, in, power);
454 for (j = 0; j < 256; j++) {
455 out_cb_unnorm[j] = -2047; /* and can only grow */
458 for (i = 0; i < AUBANDS; i++) {
459 denom = ca_cb; /* and can only grow */
460 for (j = 0; j < 256; j++)
461 denom = add_cb(denom, power[j] + auf[samplerate_index][i][j]);
462 for (j = 0; j < 256; j++)
463 out_cb_unnorm[j] = add_cb(out_cb_unnorm[j],
464 -denom + auf[samplerate_index][i][j]);
467 for (j = 0; j < 256; j++)
468 out_cb[j] = add_cb(out_cb[j], -out_cb_unnorm[j] - ca_cb - cs_cb);
471 typedef void (*walk_band_t)(DCAEncContext *c, int band1, int band2, int f,
472 int32_t spectrum1, int32_t spectrum2, int channel,
475 static void walk_band_low(DCAEncContext *c, int band, int channel,
476 walk_band_t walk, int32_t *arg)
481 for (f = 0; f < 4; f++)
482 walk(c, 0, 0, f, 0, -2047, channel, arg);
484 for (f = 0; f < 8; f++)
485 walk(c, band, band - 1, 8 * band - 4 + f,
486 c->band_spectrum[7 - f], c->band_spectrum[f], channel, arg);
490 static void walk_band_high(DCAEncContext *c, int band, int channel,
491 walk_band_t walk, int32_t *arg)
496 for (f = 0; f < 4; f++)
497 walk(c, 31, 31, 256 - 4 + f, 0, -2047, channel, arg);
499 for (f = 0; f < 8; f++)
500 walk(c, band, band + 1, 8 * band + 4 + f,
501 c->band_spectrum[f], c->band_spectrum[7 - f], channel, arg);
505 static void update_band_masking(DCAEncContext *c, int band1, int band2,
506 int f, int32_t spectrum1, int32_t spectrum2,
507 int channel, int32_t * arg)
509 int32_t value = c->eff_masking_curve_cb[f] - spectrum1;
511 if (value < c->band_masking_cb[band1])
512 c->band_masking_cb[band1] = value;
515 static void calc_masking(DCAEncContext *c, const int32_t *input)
517 int i, k, band, ch, ssf;
520 for (i = 0; i < 256; i++)
521 for (ssf = 0; ssf < SUBSUBFRAMES; ssf++)
522 c->masking_curve_cb[ssf][i] = -2047;
524 for (ssf = 0; ssf < SUBSUBFRAMES; ssf++)
525 for (ch = 0; ch < c->fullband_channels; ch++) {
526 const int chi = c->channel_order_tab[ch];
528 for (i = 0, k = 128 + 256 * ssf; k < 512; i++, k++)
529 data[i] = c->history[ch][k];
530 for (k -= 512; i < 512; i++, k++)
531 data[i] = input[k * c->channels + chi];
532 adjust_jnd(c, data, c->masking_curve_cb[ssf]);
534 for (i = 0; i < 256; i++) {
537 for (ssf = 0; ssf < SUBSUBFRAMES; ssf++)
538 if (c->masking_curve_cb[ssf][i] < m)
539 m = c->masking_curve_cb[ssf][i];
540 c->eff_masking_curve_cb[i] = m;
543 for (band = 0; band < 32; band++) {
544 c->band_masking_cb[band] = 2048;
545 walk_band_low(c, band, 0, update_band_masking, NULL);
546 walk_band_high(c, band, 0, update_band_masking, NULL);
550 static inline int32_t find_peak(const int32_t *in, int len) {
553 for (sample = 0; sample < len; sample++) {
554 int32_t s = abs(in[sample]);
562 static void find_peaks(DCAEncContext *c)
566 for (ch = 0; ch < c->fullband_channels; ch++) {
567 for (band = 0; band < 32; band++) {
568 c->peak_cb[ch][band] = find_peak(c->subband[ch][band], SUBBAND_SAMPLES);
572 if (c->lfe_channel) {
573 c->lfe_peak_cb = find_peak(c->downsampled_lfe, DCA_LFE_SAMPLES);
577 static void adpcm_analysis(DCAEncContext *c)
582 int32_t estimated_diff[SUBBAND_SAMPLES];
584 c->consumed_adpcm_bits = 0;
585 for (ch = 0; ch < c->fullband_channels; ch++) {
586 for (band = 0; band < 32; band++) {
587 samples = c->subband[ch][band] - DCA_ADPCM_COEFFS;
588 pred_vq_id = ff_dcaadpcm_subband_analysis(&c->adpcm_ctx, samples, SUBBAND_SAMPLES, estimated_diff);
589 if (pred_vq_id >= 0) {
590 c->prediction_mode[ch][band] = pred_vq_id;
591 c->consumed_adpcm_bits += 12; //12 bits to transmit prediction vq index
592 c->diff_peak_cb[ch][band] = find_peak(estimated_diff, 16);
594 c->prediction_mode[ch][band] = -1;
600 static const int snr_fudge = 128;
601 #define USED_1ABITS 1
602 #define USED_26ABITS 4
604 static inline int32_t get_step_size(const DCAEncContext *c, int ch, int band)
608 if (c->bitrate_index == 3)
609 step_size = ff_dca_lossless_quant[c->abits[ch][band]];
611 step_size = ff_dca_lossy_quant[c->abits[ch][band]];
616 static int calc_one_scale(int32_t peak_cb, int abits, softfloat *quant)
619 int our_nscale, try_remove;
622 av_assert0(peak_cb <= 0);
623 av_assert0(peak_cb >= -2047);
626 peak = cb_to_level[-peak_cb];
628 for (try_remove = 64; try_remove > 0; try_remove >>= 1) {
629 if (scalefactor_inv[our_nscale - try_remove].e + stepsize_inv[abits].e <= 17)
631 our_quant.m = mul32(scalefactor_inv[our_nscale - try_remove].m, stepsize_inv[abits].m);
632 our_quant.e = scalefactor_inv[our_nscale - try_remove].e + stepsize_inv[abits].e - 17;
633 if ((ff_dca_quant_levels[abits] - 1) / 2 < quantize_value(peak, our_quant))
635 our_nscale -= try_remove;
638 if (our_nscale >= 125)
641 quant->m = mul32(scalefactor_inv[our_nscale].m, stepsize_inv[abits].m);
642 quant->e = scalefactor_inv[our_nscale].e + stepsize_inv[abits].e - 17;
643 av_assert0((ff_dca_quant_levels[abits] - 1) / 2 >= quantize_value(peak, *quant));
648 static inline void quantize_adpcm_subband(DCAEncContext *c, int ch, int band)
651 int32_t diff_peak_cb = c->diff_peak_cb[ch][band];
652 c->scale_factor[ch][band] = calc_one_scale(diff_peak_cb,
654 &c->quant[ch][band]);
656 step_size = get_step_size(c, ch, band);
657 ff_dcaadpcm_do_real(c->prediction_mode[ch][band],
658 c->quant[ch][band], ff_dca_scale_factor_quant7[c->scale_factor[ch][band]], step_size,
659 c->adpcm_history[ch][band], c->subband[ch][band], c->adpcm_history[ch][band]+4, c->quantized[ch][band],
660 SUBBAND_SAMPLES, cb_to_level[-diff_peak_cb]);
663 static void quantize_adpcm(DCAEncContext *c)
667 for (ch = 0; ch < c->fullband_channels; ch++)
668 for (band = 0; band < 32; band++)
669 if (c->prediction_mode[ch][band] >= 0)
670 quantize_adpcm_subband(c, ch, band);
673 static void quantize_pcm(DCAEncContext *c)
675 int sample, band, ch;
677 for (ch = 0; ch < c->fullband_channels; ch++)
678 for (band = 0; band < 32; band++)
679 if (c->prediction_mode[ch][band] == -1)
680 for (sample = 0; sample < SUBBAND_SAMPLES; sample++)
681 c->quantized[ch][band][sample] = quantize_value(c->subband[ch][band][sample], c->quant[ch][band]);
684 static void accumulate_huff_bit_consumption(int abits, int32_t *quantized, uint32_t *result)
686 uint8_t sel, id = abits - 1;
687 for (sel = 0; sel < ff_dca_quant_index_group_size[id]; sel++)
688 result[sel] += ff_dca_vlc_calc_quant_bits(quantized, SUBBAND_SAMPLES, sel, id);
691 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])
694 uint32_t best_sel_bits[DCA_CODE_BOOKS];
695 int32_t best_sel_id[DCA_CODE_BOOKS];
696 uint32_t t, bits = 0;
698 for (i = 0; i < DCA_CODE_BOOKS; i++) {
700 av_assert0(!((!!vlc_bits[i][0]) ^ (!!clc_bits[i])));
701 if (vlc_bits[i][0] == 0) {
702 /* do not transmit adjustment index for empty codebooks */
703 res[i] = ff_dca_quant_index_group_size[i];
708 best_sel_bits[i] = vlc_bits[i][0];
710 for (sel = 0; sel < ff_dca_quant_index_group_size[i]; sel++) {
711 if (best_sel_bits[i] > vlc_bits[i][sel] && vlc_bits[i][sel]) {
712 best_sel_bits[i] = vlc_bits[i][sel];
713 best_sel_id[i] = sel;
717 /* 2 bits to transmit scale factor adjustment index */
718 t = best_sel_bits[i] + 2;
719 if (t < clc_bits[i]) {
720 res[i] = best_sel_id[i];
723 res[i] = ff_dca_quant_index_group_size[i];
730 static uint32_t set_best_abits_code(int abits[DCAENC_SUBBANDS], int bands, int32_t *res)
734 int32_t best_sel = 6;
735 int32_t best_bits = bands * 5;
737 /* Check do we have subband which cannot be encoded by Huffman tables */
738 for (i = 0; i < bands; i++) {
739 if (abits[i] > 12 || abits[i] == 0) {
745 for (i = 0; i < DCA_BITALLOC_12_COUNT; i++) {
746 t = ff_dca_vlc_calc_alloc_bits(abits, bands, i);
757 static int init_quantization_noise(DCAEncContext *c, int noise, int forbid_zero)
759 int ch, band, ret = USED_26ABITS | USED_1ABITS;
760 uint32_t huff_bit_count_accum[MAX_CHANNELS][DCA_CODE_BOOKS][7];
761 uint32_t clc_bit_count_accum[MAX_CHANNELS][DCA_CODE_BOOKS];
762 uint32_t bits_counter = 0;
764 c->consumed_bits = 132 + 333 * c->fullband_channels;
765 c->consumed_bits += c->consumed_adpcm_bits;
767 c->consumed_bits += 72;
769 /* attempt to guess the bit distribution based on the prevoius frame */
770 for (ch = 0; ch < c->fullband_channels; ch++) {
771 for (band = 0; band < 32; band++) {
772 int snr_cb = c->peak_cb[ch][band] - c->band_masking_cb[band] - noise;
774 if (snr_cb >= 1312) {
775 c->abits[ch][band] = 26;
777 } else if (snr_cb >= 222) {
778 c->abits[ch][band] = 8 + mul32(snr_cb - 222, 69000000);
779 ret &= ~(USED_26ABITS | USED_1ABITS);
780 } else if (snr_cb >= 0) {
781 c->abits[ch][band] = 2 + mul32(snr_cb, 106000000);
782 ret &= ~(USED_26ABITS | USED_1ABITS);
783 } else if (forbid_zero || snr_cb >= -140) {
784 c->abits[ch][band] = 1;
785 ret &= ~USED_26ABITS;
787 c->abits[ch][band] = 0;
788 ret &= ~(USED_26ABITS | USED_1ABITS);
791 c->consumed_bits += set_best_abits_code(c->abits[ch], 32, &c->bit_allocation_sel[ch]);
794 /* Recalc scale_factor each time to get bits consumption in case of Huffman coding.
795 It is suboptimal solution */
796 /* TODO: May be cache scaled values */
797 for (ch = 0; ch < c->fullband_channels; ch++) {
798 for (band = 0; band < 32; band++) {
799 if (c->prediction_mode[ch][band] == -1) {
800 c->scale_factor[ch][band] = calc_one_scale(c->peak_cb[ch][band],
802 &c->quant[ch][band]);
809 memset(huff_bit_count_accum, 0, MAX_CHANNELS * DCA_CODE_BOOKS * 7 * sizeof(uint32_t));
810 memset(clc_bit_count_accum, 0, MAX_CHANNELS * DCA_CODE_BOOKS * sizeof(uint32_t));
811 for (ch = 0; ch < c->fullband_channels; ch++) {
812 for (band = 0; band < 32; band++) {
813 if (c->abits[ch][band] && c->abits[ch][band] <= DCA_CODE_BOOKS) {
814 accumulate_huff_bit_consumption(c->abits[ch][band], c->quantized[ch][band], huff_bit_count_accum[ch][c->abits[ch][band] - 1]);
815 clc_bit_count_accum[ch][c->abits[ch][band] - 1] += bit_consumption[c->abits[ch][band]];
817 bits_counter += bit_consumption[c->abits[ch][band]];
822 for (ch = 0; ch < c->fullband_channels; ch++) {
823 bits_counter += set_best_code(huff_bit_count_accum[ch], clc_bit_count_accum[ch], c->quant_index_sel[ch]);
826 c->consumed_bits += bits_counter;
831 static void assign_bits(DCAEncContext *c)
833 /* Find the bounds where the binary search should work */
838 init_quantization_noise(c, c->worst_quantization_noise, forbid_zero);
839 low = high = c->worst_quantization_noise;
840 if (c->consumed_bits > c->frame_bits) {
841 while (c->consumed_bits > c->frame_bits) {
842 if (used_abits == USED_1ABITS && forbid_zero) {
848 used_abits = init_quantization_noise(c, high, forbid_zero);
851 while (c->consumed_bits <= c->frame_bits) {
853 if (used_abits == USED_26ABITS)
854 goto out; /* The requested bitrate is too high, pad with zeros */
856 used_abits = init_quantization_noise(c, low, forbid_zero);
860 /* Now do a binary search between low and high to see what fits */
861 for (down = snr_fudge >> 1; down; down >>= 1) {
862 init_quantization_noise(c, high - down, forbid_zero);
863 if (c->consumed_bits <= c->frame_bits)
866 init_quantization_noise(c, high, forbid_zero);
868 c->worst_quantization_noise = high;
869 if (high > c->worst_noise_ever)
870 c->worst_noise_ever = high;
873 static void shift_history(DCAEncContext *c, const int32_t *input)
877 for (k = 0; k < 512; k++)
878 for (ch = 0; ch < c->channels; ch++) {
879 const int chi = c->channel_order_tab[ch];
881 c->history[ch][k] = input[k * c->channels + chi];
885 static void fill_in_adpcm_bufer(DCAEncContext *c)
889 /* We fill in ADPCM work buffer for subbands which hasn't been ADPCM coded
890 * in current frame - we need this data if subband of next frame is
893 for (ch = 0; ch < c->channels; ch++) {
894 for (band = 0; band < 32; band++) {
895 int32_t *samples = c->subband[ch][band] - DCA_ADPCM_COEFFS;
896 if (c->prediction_mode[ch][band] == -1) {
897 step_size = get_step_size(c, ch, band);
899 ff_dca_core_dequantize(c->adpcm_history[ch][band],
900 c->quantized[ch][band]+12, step_size, ff_dca_scale_factor_quant7[c->scale_factor[ch][band]], 0, 4);
902 AV_COPY128U(c->adpcm_history[ch][band], c->adpcm_history[ch][band]+4);
904 /* Copy dequantized values for LPC analysis.
905 * It reduces artifacts in case of extreme quantization,
906 * example: in current frame abits is 1 and has no prediction flag,
907 * but end of this frame is sine like signal. In this case, if LPC analysis uses
908 * original values, likely LPC analysis returns good prediction gain, and sets prediction flag.
909 * But there are no proper value in decoder history, so likely result will be no good.
910 * Bitstream has "Predictor history flag switch", but this flag disables history for all subbands
912 samples[0] = c->adpcm_history[ch][band][0] << 7;
913 samples[1] = c->adpcm_history[ch][band][1] << 7;
914 samples[2] = c->adpcm_history[ch][band][2] << 7;
915 samples[3] = c->adpcm_history[ch][band][3] << 7;
920 static void calc_lfe_scales(DCAEncContext *c)
923 c->lfe_scale_factor = calc_one_scale(c->lfe_peak_cb, 11, &c->lfe_quant);
926 static void put_frame_header(DCAEncContext *c)
929 put_bits(&c->pb, 16, 0x7ffe);
930 put_bits(&c->pb, 16, 0x8001);
932 /* Frame type: normal */
933 put_bits(&c->pb, 1, 1);
935 /* Deficit sample count: none */
936 put_bits(&c->pb, 5, 31);
938 /* CRC is not present */
939 put_bits(&c->pb, 1, 0);
941 /* Number of PCM sample blocks */
942 put_bits(&c->pb, 7, SUBBAND_SAMPLES - 1);
944 /* Primary frame byte size */
945 put_bits(&c->pb, 14, c->frame_size - 1);
947 /* Audio channel arrangement */
948 put_bits(&c->pb, 6, c->channel_config);
950 /* Core audio sampling frequency */
951 put_bits(&c->pb, 4, bitstream_sfreq[c->samplerate_index]);
953 /* Transmission bit rate */
954 put_bits(&c->pb, 5, c->bitrate_index);
956 /* Embedded down mix: disabled */
957 put_bits(&c->pb, 1, 0);
959 /* Embedded dynamic range flag: not present */
960 put_bits(&c->pb, 1, 0);
962 /* Embedded time stamp flag: not present */
963 put_bits(&c->pb, 1, 0);
965 /* Auxiliary data flag: not present */
966 put_bits(&c->pb, 1, 0);
968 /* HDCD source: no */
969 put_bits(&c->pb, 1, 0);
971 /* Extension audio ID: N/A */
972 put_bits(&c->pb, 3, 0);
974 /* Extended audio data: not present */
975 put_bits(&c->pb, 1, 0);
977 /* Audio sync word insertion flag: after each sub-frame */
978 put_bits(&c->pb, 1, 0);
980 /* Low frequency effects flag: not present or 64x subsampling */
981 put_bits(&c->pb, 2, c->lfe_channel ? 2 : 0);
983 /* Predictor history switch flag: on */
984 put_bits(&c->pb, 1, 1);
987 /* Multirate interpolator switch: non-perfect reconstruction */
988 put_bits(&c->pb, 1, 0);
990 /* Encoder software revision: 7 */
991 put_bits(&c->pb, 4, 7);
993 /* Copy history: 0 */
994 put_bits(&c->pb, 2, 0);
996 /* Source PCM resolution: 16 bits, not DTS ES */
997 put_bits(&c->pb, 3, 0);
999 /* Front sum/difference coding: no */
1000 put_bits(&c->pb, 1, 0);
1002 /* Surrounds sum/difference coding: no */
1003 put_bits(&c->pb, 1, 0);
1005 /* Dialog normalization: 0 dB */
1006 put_bits(&c->pb, 4, 0);
1009 static void put_primary_audio_header(DCAEncContext *c)
1012 /* Number of subframes */
1013 put_bits(&c->pb, 4, SUBFRAMES - 1);
1015 /* Number of primary audio channels */
1016 put_bits(&c->pb, 3, c->fullband_channels - 1);
1018 /* Subband activity count */
1019 for (ch = 0; ch < c->fullband_channels; ch++)
1020 put_bits(&c->pb, 5, DCAENC_SUBBANDS - 2);
1022 /* High frequency VQ start subband */
1023 for (ch = 0; ch < c->fullband_channels; ch++)
1024 put_bits(&c->pb, 5, DCAENC_SUBBANDS - 1);
1026 /* Joint intensity coding index: 0, 0 */
1027 for (ch = 0; ch < c->fullband_channels; ch++)
1028 put_bits(&c->pb, 3, 0);
1030 /* Transient mode codebook: A4, A4 (arbitrary) */
1031 for (ch = 0; ch < c->fullband_channels; ch++)
1032 put_bits(&c->pb, 2, 0);
1034 /* Scale factor code book: 7 bit linear, 7-bit sqrt table (for each channel) */
1035 for (ch = 0; ch < c->fullband_channels; ch++)
1036 put_bits(&c->pb, 3, 6);
1038 /* Bit allocation quantizer select: linear 5-bit */
1039 for (ch = 0; ch < c->fullband_channels; ch++)
1040 put_bits(&c->pb, 3, c->bit_allocation_sel[ch]);
1042 /* Quantization index codebook select */
1043 for (i = 0; i < DCA_CODE_BOOKS; i++)
1044 for (ch = 0; ch < c->fullband_channels; ch++)
1045 put_bits(&c->pb, ff_dca_quant_index_sel_nbits[i], c->quant_index_sel[ch][i]);
1047 /* Scale factor adjustment index: transmitted in case of Huffman coding */
1048 for (i = 0; i < DCA_CODE_BOOKS; i++)
1049 for (ch = 0; ch < c->fullband_channels; ch++)
1050 if (c->quant_index_sel[ch][i] < ff_dca_quant_index_group_size[i])
1051 put_bits(&c->pb, 2, 0);
1053 /* Audio header CRC check word: not transmitted */
1056 static void put_subframe_samples(DCAEncContext *c, int ss, int band, int ch)
1058 int i, j, sum, bits, sel;
1059 if (c->abits[ch][band] <= DCA_CODE_BOOKS) {
1060 av_assert0(c->abits[ch][band] > 0);
1061 sel = c->quant_index_sel[ch][c->abits[ch][band] - 1];
1063 if (sel < ff_dca_quant_index_group_size[c->abits[ch][band] - 1]) {
1064 ff_dca_vlc_enc_quant(&c->pb, &c->quantized[ch][band][ss * 8], 8, sel, c->abits[ch][band] - 1);
1069 if (c->abits[ch][band] <= 7) {
1070 for (i = 0; i < 8; i += 4) {
1072 for (j = 3; j >= 0; j--) {
1073 sum *= ff_dca_quant_levels[c->abits[ch][band]];
1074 sum += c->quantized[ch][band][ss * 8 + i + j];
1075 sum += (ff_dca_quant_levels[c->abits[ch][band]] - 1) / 2;
1077 put_bits(&c->pb, bit_consumption[c->abits[ch][band]] / 4, sum);
1083 for (i = 0; i < 8; i++) {
1084 bits = bit_consumption[c->abits[ch][band]] / 16;
1085 put_sbits(&c->pb, bits, c->quantized[ch][band][ss * 8 + i]);
1089 static void put_subframe(DCAEncContext *c, int subframe)
1091 int i, band, ss, ch;
1093 /* Subsubframes count */
1094 put_bits(&c->pb, 2, SUBSUBFRAMES -1);
1096 /* Partial subsubframe sample count: dummy */
1097 put_bits(&c->pb, 3, 0);
1099 /* Prediction mode: no ADPCM, in each channel and subband */
1100 for (ch = 0; ch < c->fullband_channels; ch++)
1101 for (band = 0; band < DCAENC_SUBBANDS; band++)
1102 put_bits(&c->pb, 1, !(c->prediction_mode[ch][band] == -1));
1104 /* Prediction VQ address */
1105 for (ch = 0; ch < c->fullband_channels; ch++)
1106 for (band = 0; band < DCAENC_SUBBANDS; band++)
1107 if (c->prediction_mode[ch][band] >= 0)
1108 put_bits(&c->pb, 12, c->prediction_mode[ch][band]);
1110 /* Bit allocation index */
1111 for (ch = 0; ch < c->fullband_channels; ch++) {
1112 if (c->bit_allocation_sel[ch] == 6) {
1113 for (band = 0; band < DCAENC_SUBBANDS; band++) {
1114 put_bits(&c->pb, 5, c->abits[ch][band]);
1117 ff_dca_vlc_enc_alloc(&c->pb, c->abits[ch], DCAENC_SUBBANDS, c->bit_allocation_sel[ch]);
1121 if (SUBSUBFRAMES > 1) {
1122 /* Transition mode: none for each channel and subband */
1123 for (ch = 0; ch < c->fullband_channels; ch++)
1124 for (band = 0; band < DCAENC_SUBBANDS; band++)
1125 if (c->abits[ch][band])
1126 put_bits(&c->pb, 1, 0); /* codebook A4 */
1130 for (ch = 0; ch < c->fullband_channels; ch++)
1131 for (band = 0; band < DCAENC_SUBBANDS; band++)
1132 if (c->abits[ch][band])
1133 put_bits(&c->pb, 7, c->scale_factor[ch][band]);
1135 /* Joint subband scale factor codebook select: not transmitted */
1136 /* Scale factors for joint subband coding: not transmitted */
1137 /* Stereo down-mix coefficients: not transmitted */
1138 /* Dynamic range coefficient: not transmitted */
1139 /* Stde information CRC check word: not transmitted */
1140 /* VQ encoded high frequency subbands: not transmitted */
1142 /* LFE data: 8 samples and scalefactor */
1143 if (c->lfe_channel) {
1144 for (i = 0; i < DCA_LFE_SAMPLES; i++)
1145 put_bits(&c->pb, 8, quantize_value(c->downsampled_lfe[i], c->lfe_quant) & 0xff);
1146 put_bits(&c->pb, 8, c->lfe_scale_factor);
1149 /* Audio data (subsubframes) */
1150 for (ss = 0; ss < SUBSUBFRAMES ; ss++)
1151 for (ch = 0; ch < c->fullband_channels; ch++)
1152 for (band = 0; band < DCAENC_SUBBANDS; band++)
1153 if (c->abits[ch][band])
1154 put_subframe_samples(c, ss, band, ch);
1157 put_bits(&c->pb, 16, 0xffff);
1160 static int encode_frame(AVCodecContext *avctx, AVPacket *avpkt,
1161 const AVFrame *frame, int *got_packet_ptr)
1163 DCAEncContext *c = avctx->priv_data;
1164 const int32_t *samples;
1167 if ((ret = ff_alloc_packet2(avctx, avpkt, c->frame_size, 0)) < 0)
1170 samples = (const int32_t *)frame->data[0];
1172 subband_transform(c, samples);
1174 lfe_downsample(c, samples);
1176 calc_masking(c, samples);
1177 if (c->options.adpcm_mode)
1182 shift_history(c, samples);
1184 init_put_bits(&c->pb, avpkt->data, avpkt->size);
1185 fill_in_adpcm_bufer(c);
1186 put_frame_header(c);
1187 put_primary_audio_header(c);
1188 for (i = 0; i < SUBFRAMES; i++)
1192 for (i = put_bits_count(&c->pb); i < 8*c->frame_size; i++)
1193 put_bits(&c->pb, 1, 0);
1195 flush_put_bits(&c->pb);
1197 avpkt->pts = frame->pts;
1198 avpkt->duration = ff_samples_to_time_base(avctx, frame->nb_samples);
1199 avpkt->size = put_bits_count(&c->pb) >> 3;
1200 *got_packet_ptr = 1;
1204 #define DCAENC_FLAGS AV_OPT_FLAG_ENCODING_PARAM | AV_OPT_FLAG_AUDIO_PARAM
1206 static const AVOption options[] = {
1207 { "dca_adpcm", "Use ADPCM encoding", offsetof(DCAEncContext, options.adpcm_mode), AV_OPT_TYPE_BOOL, {.i64 = 0}, 0, 1, DCAENC_FLAGS },
1211 static const AVClass dcaenc_class = {
1212 .class_name = "DCA (DTS Coherent Acoustics)",
1213 .item_name = av_default_item_name,
1215 .version = LIBAVUTIL_VERSION_INT,
1218 static const AVCodecDefault defaults[] = {
1223 AVCodec ff_dca_encoder = {
1225 .long_name = NULL_IF_CONFIG_SMALL("DCA (DTS Coherent Acoustics)"),
1226 .type = AVMEDIA_TYPE_AUDIO,
1227 .id = AV_CODEC_ID_DTS,
1228 .priv_data_size = sizeof(DCAEncContext),
1229 .init = encode_init,
1230 .close = encode_close,
1231 .encode2 = encode_frame,
1232 .capabilities = AV_CODEC_CAP_EXPERIMENTAL,
1233 .caps_internal = FF_CODEC_CAP_INIT_THREADSAFE | FF_CODEC_CAP_INIT_CLEANUP,
1234 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_S32,
1235 AV_SAMPLE_FMT_NONE },
1236 .supported_samplerates = sample_rates,
1237 .channel_layouts = (const uint64_t[]) { AV_CH_LAYOUT_MONO,
1238 AV_CH_LAYOUT_STEREO,
1240 AV_CH_LAYOUT_5POINT0,
1241 AV_CH_LAYOUT_5POINT1,
1243 .defaults = defaults,
1244 .priv_class = &dcaenc_class,