This patch does 4 things, all of which interact and thus it woudln't be possible to commit them separately without causing either quality regressions or assertion failures. Fate comparison targets don't all reflect improvements in quality, yet listening tests show substantially improved quality and stability. 1. Increase SF range utilization. The spec requires SF delta values to be constrained within the range -60..60. The previous code was applying that range to the whole SF array and not only the deltas of consecutive values, because doing so requires smarter code: zeroing or otherwise skipping a band may invalidate lots of SF choices. This patch implements that logic to allow the coders to utilize the full dynamic range of scalefactors, increasing quality quite considerably, and fixing delta-SF-related assertion failures, since now the limitation is enforced rather than asserted. 2. PNS tweaks The previous modification makes big improvements in twoloop's efficiency, and every time that happens PNS logic needs to be tweaked accordingly to avoid it from stepping all over twoloop's decisions. This patch includes modifications of the sort. 3. Account for lowpass cutoff during PSY analysis The closer PSY's allocation is to final allocation the better the quality is, and given these modifications, twoloop is now very efficient at avoiding holes. Thus, to compute accurate thresholds, PSY needs to account for the lowpass applied implicitly during twoloop (by zeroing high bands). This patch makes twoloop set the cutoff in psymodel's context the first time it runs, and makes PSY account for it during threshold computation, making PE and threshold computations closer to the final allocation and thus achieving better subjective quality. 4. Tweaks to RC lambda tracking loop in relation to PNS Without this tweak some corner cases cause quality regressions. Basically, lambda needs to react faster to overall bitrate efficiency changes since now PNS can be quite successful in enforcing maximum bitrates, when PSY allocates too many bits to the lower bands, suppressing the signals RC logic uses to lower lambda in those cases and causing aggressive PNS. This tweak makes PNS much less aggressive, though it can still use some further tweaks. Also update MIPS specializations and adjust fuzz Also in lavc/mips/aacpsy_mips.h: remove trailing whitespace
Showing the changes to libavcodec/aaccoder.c (1 of 11 files). Show all
Unified · Split
| @@ -54,7 +54,7 @@ · 53 unmodified lines | |||
| 54 | 54 | ||
| 55 | /* Parameter of f(x) = a*(lambda/100), defines the maximum fourier spread | 55 | /* Parameter of f(x) = a*(lambda/100), defines the maximum fourier spread |
| 56 | * beyond which no PNS is used (since the SFBs contain tone rather than noise) */ | 56 | * beyond which no PNS is used (since the SFBs contain tone rather than noise) */ |
| 57 | #define NOISE_SPREAD_THRESHOLD 0.5073f | 57 | #define NOISE_SPREAD_THRESHOLD 0.9f |
| 58 | 58 | ||
| 59 | /* Parameter of f(x) = a*(100/lambda), defines how much PNS is allowed to | 59 | /* Parameter of f(x) = a*(100/lambda), defines how much PNS is allowed to |
| 60 | * replace low energy non zero bands */ | 60 | * replace low energy non zero bands */ |
| @@ -591,6 +591,7 @@ static void search_for_pns(AACEncContext *s, AVCodecContext *avctx, SingleChanne · 530 unmodified lines | |||
| 591 | int bandwidth, cutoff; | 591 | int bandwidth, cutoff; |
| 592 | float *PNS = &s->scoefs[0*128], *PNS34 = &s->scoefs[1*128]; | 592 | float *PNS = &s->scoefs[0*128], *PNS34 = &s->scoefs[1*128]; |
| 593 | float *NOR34 = &s->scoefs[3*128]; | 593 | float *NOR34 = &s->scoefs[3*128]; |
| 594 | uint8_t nextband[128]; | ||
| 594 | const float lambda = s->lambda; | 595 | const float lambda = s->lambda; |
| 595 | const float freq_mult = avctx->sample_rate*0.5f/wlen; | 596 | const float freq_mult = avctx->sample_rate*0.5f/wlen; |
| 596 | const float thr_mult = NOISE_LAMBDA_REPLACE*(100.0f/lambda); | 597 | const float thr_mult = NOISE_LAMBDA_REPLACE*(100.0f/lambda); |
| @@ -604,6 +605,7 @@ static void search_for_pns(AACEncContext *s, AVCodecContext *avctx, SingleChanne · 7 unmodified lines | |||
| 604 | 605 | ||
| 605 | /** Keep this in sync with twoloop's cutoff selection */ | 606 | /** Keep this in sync with twoloop's cutoff selection */ |
| 606 | float rate_bandwidth_multiplier = 1.5f; | 607 | float rate_bandwidth_multiplier = 1.5f; |
| 608 | int prev = -1000, prev_sf = -1; | ||
| 607 | int frame_bit_rate = (avctx->flags & CODEC_FLAG_QSCALE) | 609 | int frame_bit_rate = (avctx->flags & CODEC_FLAG_QSCALE) |
| 608 | ? (refbits * rate_bandwidth_multiplier * avctx->sample_rate / 1024) | 610 | ? (refbits * rate_bandwidth_multiplier * avctx->sample_rate / 1024) |
| 609 | : (avctx->bit_rate / avctx->channels); | 611 | : (avctx->bit_rate / avctx->channels); |
| @@ -619,6 +621,7 @@ static void search_for_pns(AACEncContext *s, AVCodecContext *avctx, SingleChanne · 9 unmodified lines | |||
| 619 | cutoff = bandwidth * 2 * wlen / avctx->sample_rate; | 621 | cutoff = bandwidth * 2 * wlen / avctx->sample_rate; |
| 620 | 622 | ||
| 621 | memcpy(sce->band_alt, sce->band_type, sizeof(sce->band_type)); | 623 | memcpy(sce->band_alt, sce->band_type, sizeof(sce->band_type)); |
| 624 | ff_init_nextband_map(sce, nextband); | ||
| 622 | for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) { | 625 | for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) { |
| 623 | int wstart = w*128; | 626 | int wstart = w*128; |
| 624 | for (g = 0; g < sce->ics.num_swb; g++) { | 627 | for (g = 0; g < sce->ics.num_swb; g++) { |
| @@ -655,16 +658,27 @@ static void search_for_pns(AACEncContext *s, AVCodecContext *avctx, SingleChanne · 30 unmodified lines | |||
| 655 | * | 658 | * |
| 656 | * At this stage, point 2 is relaxed for zeroed bands near the noise threshold (hole avoidance is more important) | 659 | * At this stage, point 2 is relaxed for zeroed bands near the noise threshold (hole avoidance is more important) |
| 657 | */ | 660 | */ |
| 658 | if (((sce->zeroes[w*16+g] || !sce->band_alt[w*16+g]) && sfb_energy < threshold*sqrtf(1.5f/freq_boost)) || spread < spread_threshold || | 661 | if ((!sce->zeroes[w*16+g] && !ff_sfdelta_can_remove_band(sce, nextband, prev_sf, w*16+g)) || |
| 662 | ((sce->zeroes[w*16+g] || !sce->band_alt[w*16+g]) && sfb_energy < threshold*sqrtf(1.0f/freq_boost)) || spread < spread_threshold || | ||
| 659 | (!sce->zeroes[w*16+g] && sce->band_alt[w*16+g] && sfb_energy > threshold*thr_mult*freq_boost) || | 663 | (!sce->zeroes[w*16+g] && sce->band_alt[w*16+g] && sfb_energy > threshold*thr_mult*freq_boost) || |
| 660 | min_energy < pns_transient_energy_r * max_energy ) { | 664 | min_energy < pns_transient_energy_r * max_energy ) { |
| 661 | sce->pns_ener[w*16+g] = sfb_energy; | 665 | sce->pns_ener[w*16+g] = sfb_energy; |
| 666 | if (!sce->zeroes[w*16+g]) | ||
| 667 | prev_sf = sce->sf_idx[w*16+g]; | ||
| 662 | continue; | 668 | continue; |
| 663 | } | 669 | } |
| 664 | 670 | ||
| 665 | pns_tgt_energy = sfb_energy*FFMIN(1.0f, spread*spread); | 671 | pns_tgt_energy = sfb_energy*FFMIN(1.0f, spread*spread); |
| 666 | noise_sfi = av_clip(roundf(log2f(pns_tgt_energy)*2), -100, 155); /* Quantize */ | 672 | noise_sfi = av_clip(roundf(log2f(pns_tgt_energy)*2), -100, 155); /* Quantize */ |
| 667 | noise_amp = -ff_aac_pow2sf_tab[noise_sfi + POW_SF2_ZERO]; /* Dequantize */ | 673 | noise_amp = -ff_aac_pow2sf_tab[noise_sfi + POW_SF2_ZERO]; /* Dequantize */ |
| 674 | if (prev != -1000) { | ||
| 675 | int noise_sfdiff = noise_sfi - prev + SCALE_DIFF_ZERO; | ||
| 676 | if (noise_sfdiff < 0 || noise_sfdiff > 2*SCALE_MAX_DIFF) { | ||
| 677 | if (!sce->zeroes[w*16+g]) | ||
| 678 | prev_sf = sce->sf_idx[w*16+g]; | ||
| 679 | continue; | ||
| 680 | } | ||
| 681 | } | ||
| 668 | for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) { | 682 | for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) { |
| 669 | float band_energy, scale, pns_senergy; | 683 | float band_energy, scale, pns_senergy; |
| 670 | const int start_c = (w+w2)*128+sce->ics.swb_offset[g]; | 684 | const int start_c = (w+w2)*128+sce->ics.swb_offset[g]; |
| @@ -697,7 +711,10 @@ static void search_for_pns(AACEncContext *s, AVCodecContext *avctx, SingleChanne · 26 unmodified lines | |||
| 697 | if (sce->zeroes[w*16+g] || !sce->band_alt[w*16+g] || (energy_ratio > 0.85f && energy_ratio < 1.25f && dist2 < dist1)) { | 711 | if (sce->zeroes[w*16+g] || !sce->band_alt[w*16+g] || (energy_ratio > 0.85f && energy_ratio < 1.25f && dist2 < dist1)) { |
| 698 | sce->band_type[w*16+g] = NOISE_BT; | 712 | sce->band_type[w*16+g] = NOISE_BT; |
| 699 | sce->zeroes[w*16+g] = 0; | 713 | sce->zeroes[w*16+g] = 0; |
| 714 | prev = noise_sfi; | ||
| 700 | } | 715 | } |
| 716 | if (!sce->zeroes[w*16+g]) | ||
| 717 | prev_sf = sce->sf_idx[w*16+g]; | ||
| 701 | } | 718 | } |
| 702 | } | 719 | } |
| 703 | } | 720 | } |
| @@ -775,7 +792,8 @@ static void mark_pns(AACEncContext *s, AVCodecContext *avctx, SingleChannelEleme · 71 unmodified lines | |||
| 775 | 792 | ||
| 776 | static void search_for_ms(AACEncContext *s, ChannelElement *cpe) | 793 | static void search_for_ms(AACEncContext *s, ChannelElement *cpe) |
| 777 | { | 794 | { |
| 778 | int start = 0, i, w, w2, g, sid_sf_boost; | 795 | int start = 0, i, w, w2, g, sid_sf_boost, prev_mid, prev_side; |
| 796 | uint8_t nextband0[128], nextband1[128]; | ||
| 779 | float M[128], S[128]; | 797 | float M[128], S[128]; |
| 780 | float *L34 = s->scoefs, *R34 = s->scoefs + 128, *M34 = s->scoefs + 128*2, *S34 = s->scoefs + 128*3; | 798 | float *L34 = s->scoefs, *R34 = s->scoefs + 128, *M34 = s->scoefs + 128*2, *S34 = s->scoefs + 128*3; |
| 781 | const float lambda = s->lambda; | 799 | const float lambda = s->lambda; |
| @@ -784,21 +802,19 @@ static void search_for_ms(AACEncContext *s, ChannelElement *cpe) · 2 unmodified lines | |||
| 784 | SingleChannelElement *sce1 = &cpe->ch[1]; | 802 | SingleChannelElement *sce1 = &cpe->ch[1]; |
| 785 | if (!cpe->common_window) | 803 | if (!cpe->common_window) |
| 786 | return; | 804 | return; |
| 787 | for (w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]) { | ||
| 788 | int min_sf_idx_mid = SCALE_MAX_POS; | ||
| 789 | int min_sf_idx_side = SCALE_MAX_POS; | ||
| 790 | for (g = 0; g < sce0->ics.num_swb; g++) { | ||
| 791 | if (!sce0->zeroes[w*16+g] && sce0->band_type[w*16+g] < RESERVED_BT) | ||
| 792 | min_sf_idx_mid = FFMIN(min_sf_idx_mid, sce0->sf_idx[w*16+g]); | ||
| 793 | if (!sce1->zeroes[w*16+g] && sce1->band_type[w*16+g] < RESERVED_BT) | ||
| 794 | min_sf_idx_side = FFMIN(min_sf_idx_side, sce1->sf_idx[w*16+g]); | ||
| 795 | } | ||
| 796 | 805 | ||
| 806 | /** Scout out next nonzero bands */ | ||
| 807 | ff_init_nextband_map(sce0, nextband0); | ||
| 808 | ff_init_nextband_map(sce1, nextband1); | ||
| 809 | |||
| 810 | prev_mid = sce0->sf_idx[0]; | ||
| 811 | prev_side = sce1->sf_idx[0]; | ||
| 812 | for (w = 0; w < sce0->ics.num_windows; w += sce0->ics.group_len[w]) { | ||
| 797 | start = 0; | 813 | start = 0; |
| 798 | for (g = 0; g < sce0->ics.num_swb; g++) { | 814 | for (g = 0; g < sce0->ics.num_swb; g++) { |
| 799 | float bmax = bval2bmax(g * 17.0f / sce0->ics.num_swb) / 0.0045f; | 815 | float bmax = bval2bmax(g * 17.0f / sce0->ics.num_swb) / 0.0045f; |
| 800 | cpe->ms_mask[w*16+g] = 0; | 816 | cpe->ms_mask[w*16+g] = 0; |
| 801 | if (!cpe->ch[0].zeroes[w*16+g] && !cpe->ch[1].zeroes[w*16+g]) { | 817 | if (!sce0->zeroes[w*16+g] && !sce1->zeroes[w*16+g]) { |
| 802 | float Mmax = 0.0f, Smax = 0.0f; | 818 | float Mmax = 0.0f, Smax = 0.0f; |
| 803 | 819 | ||
| 804 | /* Must compute mid/side SF and book for the whole window group */ | 820 | /* Must compute mid/side SF and book for the whole window group */ |
| @@ -825,16 +841,18 @@ static void search_for_ms(AACEncContext *s, ChannelElement *cpe) · 20 unmodified lines | |||
| 825 | int midcb, sidcb; | 841 | int midcb, sidcb; |
| 826 | 842 | ||
| 827 | minidx = FFMIN(sce0->sf_idx[w*16+g], sce1->sf_idx[w*16+g]); | 843 | minidx = FFMIN(sce0->sf_idx[w*16+g], sce1->sf_idx[w*16+g]); |
| 828 | mididx = av_clip(minidx, min_sf_idx_mid, min_sf_idx_mid + SCALE_MAX_DIFF); | 844 | mididx = av_clip(minidx, 0, SCALE_MAX_POS - SCALE_DIV_512); |
| 829 | sididx = av_clip(minidx - sid_sf_boost * 3, min_sf_idx_side, min_sf_idx_side + SCALE_MAX_DIFF); | 845 | sididx = av_clip(minidx - sid_sf_boost * 3, 0, SCALE_MAX_POS - SCALE_DIV_512); |
| 830 | midcb = find_min_book(Mmax, mididx); | 846 | if (!cpe->is_mask[w*16+g] && sce0->band_type[w*16+g] != NOISE_BT && sce1->band_type[w*16+g] != NOISE_BT |
| 831 | sidcb = find_min_book(Smax, sididx); | 847 | && ( !ff_sfdelta_can_replace(sce0, nextband0, prev_mid, mididx, w*16+g) |
| 832 | 848 | || !ff_sfdelta_can_replace(sce1, nextband1, prev_side, sididx, w*16+g))) { | |
| 833 | if ((mididx > minidx) || (sididx > minidx)) { | ||
| 834 | /* scalefactor range violation, bad stuff, will decrease quality unacceptably */ | 849 | /* scalefactor range violation, bad stuff, will decrease quality unacceptably */ |
| 835 | continue; | 850 | continue; |
| 836 | } | 851 | } |
| 837 | 852 | ||
| 853 | midcb = find_min_book(Mmax, mididx); | ||
| 854 | sidcb = find_min_book(Smax, sididx); | ||
| 855 | |||
| 838 | /* No CB can be zero */ | 856 | /* No CB can be zero */ |
| 839 | midcb = FFMAX(1,midcb); | 857 | midcb = FFMAX(1,midcb); |
| 840 | sidcb = FFMAX(1,sidcb); | 858 | sidcb = FFMAX(1,sidcb); |
| @@ -900,6 +918,10 @@ static void search_for_ms(AACEncContext *s, ChannelElement *cpe) · 59 unmodified lines | |||
| 900 | } | 918 | } |
| 901 | } | 919 | } |
| 902 | } | 920 | } |
| 921 | if (!sce0->zeroes[w*16+g] && sce0->band_type[w*16+g] < RESERVED_BT) | ||
| 922 | prev_mid = sce0->sf_idx[w*16+g]; | ||
| 923 | if (!sce1->zeroes[w*16+g] && !cpe->is_mask[w*16+g] && sce1->band_type[w*16+g] < RESERVED_BT) | ||
| 924 | prev_side = sce1->sf_idx[w*16+g]; | ||
| 903 | start += sce0->ics.swb_sizes[g]; | 925 | start += sce0->ics.swb_sizes[g]; |
| 904 | } | 926 | } |
| 905 | } | 927 | } |