This commit once again improves the PNS implementation by scaling the thresholds with frequency. The thresholds get looser as the frequency increases since higher frequencies are basically noise to human ears. Also, this introduces quantization error correction for PNS. Should the error be too much, no PNS will be used. The energy_ratio is used to regulate the actual encoded PNS energy: if the generated PNS energy is higher than the energy from the psy system, energy_ratio is used to correct it so that hopefully once requantized and transmitted the value in the decoder will be closer to what the encoder has.
Unified · Split
| @@ -876,36 +876,39 @@ static void search_for_pns(AACEncContext *s, AVCodecContext *avctx, SingleChanne · 875 unmodified lines | ||
| 876 | 876 | for (g = 0; g < sce->ics.num_swb; g++) { |
| 877 | 877 | int noise_sfi, try_pns = 0; |
| 878 | 878 | float dist1 = 0.0f, dist2 = 0.0f, noise_amp; |
| 879 | − float energy = 0.0f, threshold = 0.0f, spread = 0.0f; | |
| 879 | + float pns_energy = 0.0f, energy_ratio, dist_thresh; | |
| 880 | + float sfb_energy = 0.0f, threshold = 0.0f, spread = 0.0f; | |
| 881 | + float freq_boost = FFMAX(0.88f*start*freq_mult/NOISE_LOW_LIMIT, 1.0f); | |
| 880 | 882 | if (start*freq_mult < NOISE_LOW_LIMIT) { |
| 881 | 883 | start += sce->ics.swb_sizes[g]; |
| 882 | 884 | continue; |
| 885 | + } else { | |
| 886 | + dist_thresh = FFMIN(0.008f*(NOISE_LOW_LIMIT/start*freq_mult), 1.11f); | |
| 883 | 887 | } |
| 884 | 888 | for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) { |
| 885 | 889 | band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g]; |
| 886 | − energy += band->energy; | |
| 887 | − spread += band->spread; | |
| 888 | − threshold += band->threshold; | |
| 890 | + sfb_energy += band->energy; | |
| 891 | + spread += band->spread; | |
| 892 | + threshold += band->threshold; | |
| 889 | 893 | } |
| 890 | − sce->pns_ener[w*16+g] = energy; | |
| 891 | 894 | |
| 892 | 895 | if (sce->zeroes[w*16+g]) { |
| 893 | 896 | try_pns = 1; |
| 894 | − } else if (energy < threshold) { | |
| 897 | + } else if (sfb_energy < threshold*freq_boost) { | |
| 895 | 898 | try_pns = 1; |
| 896 | 899 | } else if (spread > spread_threshold) { |
| 897 | 900 | try_pns = 0; |
| 898 | − } else if (energy < threshold*thr_mult) { | |
| 901 | + } else if (sfb_energy < threshold*thr_mult*freq_boost) { | |
| 899 | 902 | try_pns = 1; |
| 900 | 903 | } |
| 901 | 904 | |
| 902 | − if (!try_pns || !energy) { | |
| 905 | + if (!try_pns || !sfb_energy) { | |
| 903 | 906 | start += sce->ics.swb_sizes[g]; |
| 904 | 907 | continue; |
| 905 | 908 | } |
| 906 | 909 | |
| 907 | − noise_sfi = av_clip(roundf(log2f(energy)*2), -100, 155); /* Quantize */ | |
| 908 | − noise_amp = -ff_aac_pow2sf_tab[noise_sfi + POW_SF2_ZERO]; /* Dequantize */ | |
| 910 | + noise_sfi = av_clip(roundf(log2f(sfb_energy)*2), -100, 155); /* Quantize */ | |
| 911 | + noise_amp = -ff_aac_pow2sf_tab[noise_sfi + POW_SF2_ZERO]; /* Dequantize */ | |
| 909 | 912 | for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) { |
| 910 | 913 | float band_energy, scale; |
| 911 | 914 | band = &s->psy.ch[s->cur_channel+0].psy_bands[(w+w2)*16+g]; |
| @@ -914,6 +917,7 @@ static void search_for_pns(AACEncContext *s, AVCodecContext *avctx, SingleChanne · 2 unmodified lines | ||
| 914 | 917 | band_energy = s->fdsp->scalarproduct_float(PNS, PNS, sce->ics.swb_sizes[g]); |
| 915 | 918 | scale = noise_amp/sqrtf(band_energy); |
| 916 | 919 | s->fdsp->vector_fmul_scalar(PNS, PNS, scale, sce->ics.swb_sizes[g]); |
| 920 | + pns_energy += s->fdsp->scalarproduct_float(PNS, PNS, sce->ics.swb_sizes[g]); | |
| 917 | 921 | abs_pow34_v(NOR34, &sce->coeffs[start+(w+w2)*128], sce->ics.swb_sizes[g]); |
| 918 | 922 | abs_pow34_v(PNS34, PNS, sce->ics.swb_sizes[g]); |
| 919 | 923 | dist1 += quantize_band_cost(s, &sce->coeffs[start + (w+w2)*128], |
| @@ -929,7 +933,9 @@ static void search_for_pns(AACEncContext *s, AVCodecContext *avctx, SingleChanne · 9 unmodified lines | ||
| 929 | 933 | NOISE_BT, |
| 930 | 934 | lambda/band->threshold, INFINITY, NULL, 0); |
| 931 | 935 | } |
| 932 | − if (dist2 < dist1) { | |
| 936 | + energy_ratio = sfb_energy/pns_energy; /* Compensates for quantization error */ | |
| 937 | + sce->pns_ener[w*16+g] = energy_ratio*sfb_energy; | |
| 938 | + if (energy_ratio > 0.80f && energy_ratio < 1.20f && dist1/dist2 > dist_thresh) { | |
| 933 | 939 | sce->band_type[w*16+g] = NOISE_BT; |
| 934 | 940 | sce->zeroes[w*16+g] = 0; |
| 935 | 941 | if (sce->band_type[w*16+g-1] != NOISE_BT && /* Prevent holes */ |
| @@ -163,7 +163,7 @@ fate-aac-pns-encode: CMD = enc_dec_pcm adts wav s16le $(TARGET_SAMPLES)/audio-re · 162 unmodified lines | ||
| 163 | 163 | fate-aac-pns-encode: CMP = stddev |
| 164 | 164 | fate-aac-pns-encode: REF = $(SAMPLES)/audio-reference/luckynight_2ch_44kHz_s16.wav |
| 165 | 165 | fate-aac-pns-encode: CMP_SHIFT = -4096 |
| 166 | −fate-aac-pns-encode: CMP_TARGET = 647.11 | |
| 166 | +fate-aac-pns-encode: CMP_TARGET = 637.66 | |
| 167 | 167 | fate-aac-pns-encode: SIZE_TOLERANCE = 3560 |
| 168 | 168 | fate-aac-pns-encode: FUZZ = 5 |
| 169 | 169 | |