2 * AAC encoder twoloop coder
3 * Copyright (C) 2008-2009 Konstantin Shishkov
5 * This file is part of FFmpeg.
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
24 * AAC encoder twoloop coder
25 * @author Konstantin Shishkov, Claudio Freire
29 * This file contains a template for the twoloop coder function.
30 * It needs to be provided, externally, as an already included declaration,
31 * the following functions from aacenc_quantization/util.h. They're not included
32 * explicitly here to make it possible to provide alternative implementations:
33 * - quantize_band_cost
40 #ifndef AVCODEC_AACCODER_TWOLOOP_H
41 #define AVCODEC_AACCODER_TWOLOOP_H
44 #include "libavutil/mathematics.h"
51 #include "aacenctab.h"
53 /** Frequency in Hz for lower limit of noise substitution **/
54 #define NOISE_LOW_LIMIT 4000
56 #define sclip(x) av_clip(x,60,218)
58 /* Reflects the cost to change codebooks */
59 static inline int ff_pns_bits(SingleChannelElement *sce, int w, int g)
61 return (!g || !sce->zeroes[w*16+g-1] || !sce->can_pns[w*16+g-1]) ? 9 : 5;
65 * two-loop quantizers search taken from ISO 13818-7 Appendix C
67 static void search_for_quantizers_twoloop(AVCodecContext *avctx,
69 SingleChannelElement *sce,
72 int start = 0, i, w, w2, g, recomprd;
73 int destbits = avctx->bit_rate * 1024.0 / avctx->sample_rate
74 / ((avctx->flags & CODEC_FLAG_QSCALE) ? 2.0f : avctx->channels)
76 int refbits = destbits;
77 int toomanybits, toofewbits;
79 uint8_t nextband[128];
81 float dists[128] = { 0 }, qenergies[128] = { 0 }, uplims[128], euplims[128], energies[128];
82 float maxvals[128], spread_thr_r[128];
83 float min_spread_thr_r, max_spread_thr_r;
86 * rdlambda controls the maximum tolerated distortion. Twoloop
87 * will keep iterating until it fails to lower it or it reaches
88 * ulimit * rdlambda. Keeping it low increases quality on difficult
89 * signals, but lower it too much, and bits will be taken from weak
90 * signals, creating "holes". A balance is necesary.
91 * rdmax and rdmin specify the relative deviation from rdlambda
92 * allowed for tonality compensation
94 float rdlambda = av_clipf(2.0f * 120.f / lambda, 0.0625f, 16.0f);
95 const float nzslope = 1.5f;
96 float rdmin = 0.03125f;
100 * sfoffs controls an offset of optmium allocation that will be
101 * applied based on lambda. Keep it real and modest, the loop
102 * will take care of the rest, this just accelerates convergence
104 float sfoffs = av_clipf(log2f(120.0f / lambda) * 4.0f, -5, 10);
106 int fflag, minscaler, maxscaler, nminscaler;
116 * zeroscale controls a multiplier of the threshold, if band energy
117 * is below this, a zero is forced. Keep it lower than 1, unless
118 * low lambda is used, because energy < threshold doesn't mean there's
119 * no audible signal outright, it's just energy. Also make it rise
120 * slower than rdlambda, as rdscale has due compensation with
121 * noisy band depriorization below, whereas zeroing logic is rather dumb
124 if (lambda > 120.f) {
125 zeroscale = av_clipf(powf(120.f / lambda, 0.25f), 0.0625f, 1.0f);
130 if (s->psy.bitres.alloc >= 0) {
132 * Psy granted us extra bits to use, from the reservoire
133 * adjust for lambda except what psy already did
135 destbits = s->psy.bitres.alloc
136 * (lambda / (avctx->global_quality ? avctx->global_quality : 120));
139 if (avctx->flags & CODEC_FLAG_QSCALE) {
141 * Constant Q-scale doesn't compensate MS coding on its own
142 * No need to be overly precise, this only controls RD
143 * adjustment CB limits when going overboard
145 if (s->options.mid_side && s->cur_type == TYPE_CPE)
149 * When using a constant Q-scale, don't adjust bits, just use RD
150 * Don't let it go overboard, though... 8x psy target is enough
153 toofewbits = destbits / 16;
155 /** Don't offset scalers, just RD */
156 sfoffs = sce->ics.num_windows - 1;
157 rdlambda = sqrtf(rdlambda);
159 /** search further */
162 /* When using ABR, be strict, but a reasonable leeway is
163 * critical to allow RC to smoothly track desired bitrate
164 * without sudden quality drops that cause audible artifacts.
165 * Symmetry is also desirable, to avoid systematic bias.
167 toomanybits = destbits + destbits/8;
168 toofewbits = destbits - destbits/8;
171 rdlambda = sqrtf(rdlambda);
174 /** and zero out above cutoff frequency */
176 int wlen = 1024 / sce->ics.num_windows;
180 * Scale, psy gives us constant quality, this LP only scales
181 * bitrate by lambda, so we save bits on subjectively unimportant HF
182 * rather than increase quantization noise. Adjust nominal bitrate
183 * to effective bitrate according to encoding parameters,
184 * AAC_CUTOFF_FROM_BITRATE is calibrated for effective bitrate.
186 float rate_bandwidth_multiplier = 1.5f;
187 int frame_bit_rate = (avctx->flags & CODEC_FLAG_QSCALE)
188 ? (refbits * rate_bandwidth_multiplier * avctx->sample_rate / 1024)
189 : (avctx->bit_rate / avctx->channels);
191 /** Compensate for extensions that increase efficiency */
192 if (s->options.pns || s->options.intensity_stereo)
193 frame_bit_rate *= 1.15f;
195 if (avctx->cutoff > 0) {
196 bandwidth = avctx->cutoff;
198 bandwidth = FFMAX(3000, AAC_CUTOFF_FROM_BITRATE(frame_bit_rate, 1, avctx->sample_rate));
199 s->psy.cutoff = bandwidth;
202 cutoff = bandwidth * 2 * wlen / avctx->sample_rate;
203 pns_start_pos = NOISE_LOW_LIMIT * 2 * wlen / avctx->sample_rate;
207 * for values above this the decoder might end up in an endless loop
208 * due to always having more bits than what can be encoded.
210 destbits = FFMIN(destbits, 5800);
211 toomanybits = FFMIN(toomanybits, 5800);
212 toofewbits = FFMIN(toofewbits, 5800);
214 * XXX: some heuristic to determine initial quantizers will reduce search time
215 * determine zero bands and upper distortion limits
217 min_spread_thr_r = -1;
218 max_spread_thr_r = -1;
219 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
220 for (g = start = 0; g < sce->ics.num_swb; start += sce->ics.swb_sizes[g++]) {
222 float uplim = 0.0f, energy = 0.0f, spread = 0.0f;
223 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
224 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
225 if (start >= cutoff || band->energy <= (band->threshold * zeroscale) || band->threshold == 0.0f) {
226 sce->zeroes[(w+w2)*16+g] = 1;
235 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
236 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
237 if (band->energy <= (band->threshold * zeroscale) || band->threshold == 0.0f)
239 uplim += band->threshold;
240 energy += band->energy;
241 spread += band->spread;
245 uplims[w*16+g] = uplim;
246 energies[w*16+g] = energy;
248 sce->zeroes[w*16+g] = !nz;
250 if (nz && sce->can_pns[w*16+g]) {
251 spread_thr_r[w*16+g] = energy * nz / (uplim * spread);
252 if (min_spread_thr_r < 0) {
253 min_spread_thr_r = max_spread_thr_r = spread_thr_r[w*16+g];
255 min_spread_thr_r = FFMIN(min_spread_thr_r, spread_thr_r[w*16+g]);
256 max_spread_thr_r = FFMAX(max_spread_thr_r, spread_thr_r[w*16+g]);
262 /** Compute initial scalers */
264 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
265 for (g = 0; g < sce->ics.num_swb; g++) {
266 if (sce->zeroes[w*16+g]) {
267 sce->sf_idx[w*16+g] = SCALE_ONE_POS;
271 * log2f-to-distortion ratio is, technically, 2 (1.5db = 4, but it's power vs level so it's 2).
272 * But, as offsets are applied, low-frequency signals are too sensitive to the induced distortion,
273 * so we make scaling more conservative by choosing a lower log2f-to-distortion ratio, and thus
276 sce->sf_idx[w*16+g] = av_clip(
278 + 1.75*log2f(FFMAX(0.00125f,uplims[w*16+g]) / sce->ics.swb_sizes[g])
281 minscaler = FFMIN(minscaler, sce->sf_idx[w*16+g]);
286 minscaler = av_clip(minscaler, SCALE_ONE_POS - SCALE_DIV_512, SCALE_MAX_POS - SCALE_DIV_512);
287 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w])
288 for (g = 0; g < sce->ics.num_swb; g++)
289 if (!sce->zeroes[w*16+g])
290 sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], minscaler, minscaler + SCALE_MAX_DIFF - 1);
294 abs_pow34_v(s->scoefs, sce->coeffs, 1024);
295 ff_quantize_band_cost_cache_init(s);
297 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
299 for (g = 0; g < sce->ics.num_swb; g++) {
300 const float *scaled = s->scoefs + start;
301 maxvals[w*16+g] = find_max_val(sce->ics.group_len[w], sce->ics.swb_sizes[g], scaled);
302 start += sce->ics.swb_sizes[g];
307 * Scale uplims to match rate distortion to quality
308 * bu applying noisy band depriorization and tonal band priorization.
309 * Maxval-energy ratio gives us an idea of how noisy/tonal the band is.
310 * If maxval^2 ~ energy, then that band is mostly noise, and we can relax
311 * rate distortion requirements.
313 memcpy(euplims, uplims, sizeof(euplims));
314 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
315 /** psy already priorizes transients to some extent */
316 float de_psy_factor = (sce->ics.num_windows > 1) ? 8.0f / sce->ics.group_len[w] : 1.0f;
318 for (g = 0; g < sce->ics.num_swb; g++) {
320 float cleanup_factor = ff_sqrf(av_clipf(start / (cutoff * 0.75f), 1.0f, 2.0f));
321 float energy2uplim = find_form_factor(
322 sce->ics.group_len[w], sce->ics.swb_sizes[g],
323 uplims[w*16+g] / (nzs[g] * sce->ics.swb_sizes[w]),
325 nzslope * cleanup_factor);
326 energy2uplim *= de_psy_factor;
327 if (!(avctx->flags & CODEC_FLAG_QSCALE)) {
328 /** In ABR, we need to priorize less and let rate control do its thing */
329 energy2uplim = sqrtf(energy2uplim);
331 energy2uplim = FFMAX(0.015625f, FFMIN(1.0f, energy2uplim));
332 uplims[w*16+g] *= av_clipf(rdlambda * energy2uplim, rdmin, rdmax)
333 * sce->ics.group_len[w];
335 energy2uplim = find_form_factor(
336 sce->ics.group_len[w], sce->ics.swb_sizes[g],
337 uplims[w*16+g] / (nzs[g] * sce->ics.swb_sizes[w]),
340 energy2uplim *= de_psy_factor;
341 if (!(avctx->flags & CODEC_FLAG_QSCALE)) {
342 /** In ABR, we need to priorize less and let rate control do its thing */
343 energy2uplim = sqrtf(energy2uplim);
345 energy2uplim = FFMAX(0.015625f, FFMIN(1.0f, energy2uplim));
346 euplims[w*16+g] *= av_clipf(rdlambda * energy2uplim * sce->ics.group_len[w],
349 start += sce->ics.swb_sizes[g];
353 for (i = 0; i < sizeof(maxsf) / sizeof(maxsf[0]); ++i)
354 maxsf[i] = SCALE_MAX_POS;
356 //perform two-loop search
357 //outer loop - improve quality
359 //inner loop - quantize spectrum to fit into given number of bits
361 int qstep = its ? 1 : 32;
367 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
369 for (g = 0; g < sce->ics.num_swb; g++) {
370 const float *coefs = &sce->coeffs[start];
371 const float *scaled = &s->scoefs[start];
375 float qenergy = 0.0f;
377 if (sce->zeroes[w*16+g] || sce->sf_idx[w*16+g] >= 218) {
378 start += sce->ics.swb_sizes[g];
379 if (sce->can_pns[w*16+g]) {
380 /** PNS isn't free */
381 tbits += ff_pns_bits(sce, w, g);
385 cb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
386 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
389 dist += quantize_band_cost_cached(s, w + w2, g, coefs + w2*128,
391 sce->ics.swb_sizes[g],
401 dists[w*16+g] = dist - bits;
402 qenergies[w*16+g] = qenergy;
404 int sfdiff = av_clip(sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO, 0, 2*SCALE_MAX_DIFF);
405 bits += ff_aac_scalefactor_bits[sfdiff];
408 start += sce->ics.swb_sizes[g];
409 prev = sce->sf_idx[w*16+g];
412 if (tbits > toomanybits) {
414 for (i = 0; i < 128; i++) {
415 if (sce->sf_idx[i] < (SCALE_MAX_POS - SCALE_DIV_512)) {
416 int maxsf_i = (tbits > 5800) ? SCALE_MAX_POS : maxsf[i];
417 int new_sf = FFMIN(maxsf_i, sce->sf_idx[i] + qstep);
418 if (new_sf != sce->sf_idx[i]) {
419 sce->sf_idx[i] = new_sf;
424 } else if (tbits < toofewbits) {
426 for (i = 0; i < 128; i++) {
427 if (sce->sf_idx[i] > SCALE_ONE_POS) {
428 int new_sf = FFMAX(SCALE_ONE_POS, sce->sf_idx[i] - qstep);
429 if (new_sf != sce->sf_idx[i]) {
430 sce->sf_idx[i] = new_sf;
437 if (!qstep && tbits > toomanybits && sce->sf_idx[0] < 217 && changed)
442 fflag = tbits < toofewbits;
443 for (i = 0; i < 2 && (overdist || recomprd); ++i) {
445 /** Must recompute distortion */
448 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
450 for (g = 0; g < sce->ics.num_swb; g++) {
451 const float *coefs = sce->coeffs + start;
452 const float *scaled = s->scoefs + start;
456 float qenergy = 0.0f;
458 if (sce->zeroes[w*16+g] || sce->sf_idx[w*16+g] >= 218) {
459 start += sce->ics.swb_sizes[g];
460 if (sce->can_pns[w*16+g]) {
461 /** PNS isn't free */
462 tbits += ff_pns_bits(sce, w, g);
466 cb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
467 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
470 dist += quantize_band_cost_cached(s, w + w2, g, coefs + w2*128,
472 sce->ics.swb_sizes[g],
482 dists[w*16+g] = dist - bits;
483 qenergies[w*16+g] = qenergy;
485 int sfdiff = av_clip(sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO, 0, 2*SCALE_MAX_DIFF);
486 bits += ff_aac_scalefactor_bits[sfdiff];
489 start += sce->ics.swb_sizes[g];
490 prev = sce->sf_idx[w*16+g];
494 if (!i && s->options.pns && its > maxits/2 && tbits > toofewbits) {
495 float maxoverdist = 0.0f;
496 float ovrfactor = 1.f+(maxits-its)*16.f/maxits;
497 overdist = recomprd = 0;
498 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
499 for (g = start = 0; g < sce->ics.num_swb; start += sce->ics.swb_sizes[g++]) {
500 if (!sce->zeroes[w*16+g] && sce->sf_idx[w*16+g] > SCALE_ONE_POS && dists[w*16+g] > uplims[w*16+g]*ovrfactor) {
501 float ovrdist = dists[w*16+g] / FFMAX(uplims[w*16+g],euplims[w*16+g]);
502 maxoverdist = FFMAX(maxoverdist, ovrdist);
508 /* We have overdistorted bands, trade for zeroes (that can be noise)
509 * Zero the bands in the lowest 1.25% spread-energy-threshold ranking
511 float minspread = max_spread_thr_r;
512 float maxspread = min_spread_thr_r;
516 int maxzeroed, zloop;
517 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
518 for (g = start = 0; g < sce->ics.num_swb; start += sce->ics.swb_sizes[g++]) {
519 if (start >= pns_start_pos && !sce->zeroes[w*16+g] && sce->can_pns[w*16+g]) {
520 minspread = FFMIN(minspread, spread_thr_r[w*16+g]);
521 maxspread = FFMAX(maxspread, spread_thr_r[w*16+g]);
526 zspread = (maxspread-minspread) * 0.0125f + minspread;
527 /* Don't PNS everything even if allowed. It suppresses bit starvation signals from RC,
528 * and forced the hand of the later search_for_pns step.
529 * Instead, PNS a fraction of the spread_thr_r range depending on how starved for bits we are,
530 * and leave further PNSing to search_for_pns if worthwhile.
532 zspread = FFMIN3(min_spread_thr_r * 8.f, zspread,
533 ((toomanybits - tbits) * min_spread_thr_r + (tbits - toofewbits) * max_spread_thr_r) / (toomanybits - toofewbits + 1));
534 maxzeroed = FFMIN(zeroable, FFMAX(1, (zeroable * its + maxits - 1) / (2 * maxits)));
535 for (zloop = 0; zloop < 2; zloop++) {
536 /* Two passes: first distorted stuff - two birds in one shot and all that,
537 * then anything viable. Viable means not zero, but either CB=zero-able
538 * (too high SF), not SF <= 1 (that means we'd be operating at very high
539 * quality, we don't want PNS when doing VHQ), PNS allowed, and within
540 * the lowest ranking percentile.
542 float loopovrfactor = (zloop) ? 1.0f : ovrfactor;
543 int loopminsf = (zloop) ? (SCALE_ONE_POS - SCALE_DIV_512) : SCALE_ONE_POS;
545 for (g = sce->ics.num_swb-1; g > 0 && zeroed < maxzeroed; g--) {
546 if (sce->ics.swb_offset[g] < pns_start_pos)
548 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
549 if (!sce->zeroes[w*16+g] && sce->can_pns[w*16+g] && spread_thr_r[w*16+g] <= zspread
550 && sce->sf_idx[w*16+g] > loopminsf
551 && (dists[w*16+g] > loopovrfactor*uplims[w*16+g] || !(mcb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]))
552 || (mcb <= 1 && dists[w*16+g] > FFMIN(uplims[w*16+g], euplims[w*16+g]))) ) {
553 sce->zeroes[w*16+g] = 1;
554 sce->band_type[w*16+g] = 0;
561 recomprd = fflag = 1;
568 minscaler = SCALE_MAX_POS;
570 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
571 for (g = 0; g < sce->ics.num_swb; g++) {
572 if (!sce->zeroes[w*16+g]) {
573 minscaler = FFMIN(minscaler, sce->sf_idx[w*16+g]);
574 maxscaler = FFMAX(maxscaler, sce->sf_idx[w*16+g]);
579 minscaler = nminscaler = av_clip(minscaler, SCALE_ONE_POS - SCALE_DIV_512, SCALE_MAX_POS - SCALE_DIV_512);
581 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
582 /** Start with big steps, end up fine-tunning */
583 int depth = (its > maxits/2) ? ((its > maxits*2/3) ? 1 : 3) : 10;
584 int edepth = depth+2;
585 float uplmax = its / (maxits*0.25f) + 1.0f;
586 uplmax *= (tbits > destbits) ? FFMIN(2.0f, tbits / (float)FFMAX(1,destbits)) : 1.0f;
588 for (g = 0; g < sce->ics.num_swb; g++) {
589 int prevsc = sce->sf_idx[w*16+g];
590 if (prev < 0 && !sce->zeroes[w*16+g])
591 prev = sce->sf_idx[0];
592 if (!sce->zeroes[w*16+g]) {
593 const float *coefs = sce->coeffs + start;
594 const float *scaled = s->scoefs + start;
595 int cmb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
596 int mindeltasf = FFMAX(0, prev - SCALE_MAX_DIFF);
597 int maxdeltasf = FFMIN(SCALE_MAX_POS - SCALE_DIV_512, prev + SCALE_MAX_DIFF);
598 if ((!cmb || dists[w*16+g] > uplims[w*16+g]) && sce->sf_idx[w*16+g] > mindeltasf) {
599 /* Try to make sure there is some energy in every nonzero band
600 * NOTE: This algorithm must be forcibly imbalanced, pushing harder
601 * on holes or more distorted bands at first, otherwise there's
602 * no net gain (since the next iteration will offset all bands
603 * on the opposite direction to compensate for extra bits)
605 for (i = 0; i < edepth && sce->sf_idx[w*16+g] > mindeltasf; ++i) {
608 int mb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]-1);
609 cb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
610 dist = qenergy = 0.f;
613 maxsf[w*16+g] = FFMIN(sce->sf_idx[w*16+g]-1, maxsf[w*16+g]);
614 } else if (i >= depth && dists[w*16+g] < euplims[w*16+g]) {
617 /* !g is the DC band, it's important, since quantization error here
618 * applies to less than a cycle, it creates horrible intermodulation
619 * distortion if it doesn't stick to what psy requests
621 if (!g && sce->ics.num_windows > 1 && dists[w*16+g] >= euplims[w*16+g])
622 maxsf[w*16+g] = FFMIN(sce->sf_idx[w*16+g], maxsf[w*16+g]);
623 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
626 dist += quantize_band_cost_cached(s, w + w2, g, coefs + w2*128,
628 sce->ics.swb_sizes[g],
629 sce->sf_idx[w*16+g]-1,
638 sce->sf_idx[w*16+g]--;
639 dists[w*16+g] = dist - bits;
640 qenergies[w*16+g] = qenergy;
641 if (mb && (sce->sf_idx[w*16+g] < mindeltasf || (
642 (dists[w*16+g] < FFMIN(uplmax*uplims[w*16+g], euplims[w*16+g]))
643 && (fabsf(qenergies[w*16+g]-energies[w*16+g]) < euplims[w*16+g])
648 } else if (tbits > toofewbits && sce->sf_idx[w*16+g] < FFMIN(maxdeltasf, maxsf[w*16+g])
649 && (dists[w*16+g] < FFMIN(euplims[w*16+g], uplims[w*16+g]))
650 && (fabsf(qenergies[w*16+g]-energies[w*16+g]) < euplims[w*16+g])
652 /** Um... over target. Save bits for more important stuff. */
653 for (i = 0; i < depth && sce->sf_idx[w*16+g] < maxdeltasf; ++i) {
656 cb = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]+1);
658 dist = qenergy = 0.f;
660 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
663 dist += quantize_band_cost_cached(s, w + w2, g, coefs + w2*128,
665 sce->ics.swb_sizes[g],
666 sce->sf_idx[w*16+g]+1,
676 if (dist < FFMIN(euplims[w*16+g], uplims[w*16+g])) {
677 sce->sf_idx[w*16+g]++;
678 dists[w*16+g] = dist;
679 qenergies[w*16+g] = qenergy;
684 maxsf[w*16+g] = FFMIN(sce->sf_idx[w*16+g], maxsf[w*16+g]);
689 prev = sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], mindeltasf, maxdeltasf);
690 if (sce->sf_idx[w*16+g] != prevsc)
692 nminscaler = FFMIN(nminscaler, sce->sf_idx[w*16+g]);
693 sce->band_type[w*16+g] = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
695 start += sce->ics.swb_sizes[g];
699 /** SF difference limit violation risk. Must re-clamp. */
701 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
702 for (g = 0; g < sce->ics.num_swb; g++) {
703 if (!sce->zeroes[w*16+g]) {
704 int prevsf = sce->sf_idx[w*16+g];
707 sce->sf_idx[w*16+g] = av_clip(sce->sf_idx[w*16+g], prev - SCALE_MAX_DIFF, prev + SCALE_MAX_DIFF);
708 sce->band_type[w*16+g] = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
709 prev = sce->sf_idx[w*16+g];
710 if (!fflag && prevsf != sce->sf_idx[w*16+g])
717 } while (fflag && its < maxits);
719 /** Scout out next nonzero bands */
720 ff_init_nextband_map(sce, nextband);
723 for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
724 /** Make sure proper codebooks are set */
725 for (g = 0; g < sce->ics.num_swb; g++) {
726 if (!sce->zeroes[w*16+g]) {
727 sce->band_type[w*16+g] = find_min_book(maxvals[w*16+g], sce->sf_idx[w*16+g]);
728 if (sce->band_type[w*16+g] <= 0) {
729 if (!ff_sfdelta_can_remove_band(sce, nextband, prev, w*16+g)) {
730 /** Cannot zero out, make sure it's not attempted */
731 sce->band_type[w*16+g] = 1;
733 sce->zeroes[w*16+g] = 1;
734 sce->band_type[w*16+g] = 0;
738 sce->band_type[w*16+g] = 0;
740 /** Check that there's no SF delta range violations */
741 if (!sce->zeroes[w*16+g]) {
743 av_unused int sfdiff = sce->sf_idx[w*16+g] - prev + SCALE_DIFF_ZERO;
744 av_assert1(sfdiff >= 0 && sfdiff <= 2*SCALE_MAX_DIFF);
745 } else if (sce->zeroes[0]) {
746 /** Set global gain to something useful */
747 sce->sf_idx[0] = sce->sf_idx[w*16+g];
749 prev = sce->sf_idx[w*16+g];
755 #endif /* AVCODEC_AACCODER_TWOLOOP_H */