aacenc_tns: rework coefficient quantization and filter application
[ffmpeg.git] / libavcodec / aacenc_tns.c
1 /*
2  * AAC encoder TNS
3  * Copyright (C) 2015 Rostislav Pehlivanov
4  *
5  * This file is part of FFmpeg.
6  *
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.
11  *
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.
16  *
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
20  */
21
22 /**
23  * @file
24  * AAC encoder temporal noise shaping
25  * @author Rostislav Pehlivanov ( atomnuker gmail com )
26  */
27
28 #include "aacenc.h"
29 #include "aacenc_tns.h"
30 #include "aactab.h"
31 #include "aacenc_utils.h"
32 #include "aacenc_quantization.h"
33
34 /**
35  * Encode TNS data.
36  * Coefficient compression saves a single bit per coefficient.
37  */
38 void ff_aac_encode_tns_info(AACEncContext *s, SingleChannelElement *sce)
39 {
40     uint8_t u_coef;
41     const uint8_t coef_res = TNS_Q_BITS == 4;
42     int i, w, filt, coef_len, coef_compress = 0;
43     const int is8 = sce->ics.window_sequence[0] == EIGHT_SHORT_SEQUENCE;
44     TemporalNoiseShaping *tns = &sce->tns;
45
46     if (!sce->tns.present)
47         return;
48
49     for (i = 0; i < sce->ics.num_windows; i++) {
50         put_bits(&s->pb, 2 - is8, sce->tns.n_filt[i]);
51         if (tns->n_filt[i]) {
52             put_bits(&s->pb, 1, coef_res);
53             for (filt = 0; filt < tns->n_filt[i]; filt++) {
54                 put_bits(&s->pb, 6 - 2 * is8, tns->length[i][filt]);
55                 put_bits(&s->pb, 5 - 2 * is8, tns->order[i][filt]);
56                 if (tns->order[i][filt]) {
57                     put_bits(&s->pb, 1, !!tns->direction[i][filt]);
58                     put_bits(&s->pb, 1, !!coef_compress);
59                     coef_len = coef_res + 3 - coef_compress;
60                     for (w = 0; w < tns->order[i][filt]; w++) {
61                         u_coef = (tns->coef_idx[i][filt][w])&(~(~0<<coef_len));
62                         put_bits(&s->pb, coef_len, u_coef);
63                     }
64                 }
65             }
66         }
67     }
68 }
69
70 static int quantize_coefs(double *coef, int *idx, float *lpc, int order)
71 {
72     int i;
73     uint8_t u_coef;
74     const float *quant_arr = tns_tmp2_map[TNS_Q_BITS == 4];
75     const double iqfac_p = ((1 << (TNS_Q_BITS-1)) - 0.5)/(M_PI/2.0);
76     const double iqfac_m = ((1 << (TNS_Q_BITS-1)) + 0.5)/(M_PI/2.0);
77     for (i = 0; i < order; i++) {
78         idx[i] = ceilf(asin(coef[i])*((coef[i] >= 0) ? iqfac_p : iqfac_m));
79         u_coef = (idx[i])&(~(~0<<TNS_Q_BITS));
80         lpc[i] = quant_arr[u_coef];
81     }
82     return order;
83 }
84
85 /* Apply TNS filter */
86 void ff_aac_apply_tns(AACEncContext *s, SingleChannelElement *sce)
87 {
88     TemporalNoiseShaping *tns = &sce->tns;
89     IndividualChannelStream *ics = &sce->ics;
90     int w, filt, m, i, top, order, bottom, start, end, size, inc;
91     const int mmm = FFMIN(ics->tns_max_bands, ics->max_sfb);
92     float lpc[TNS_MAX_ORDER];
93
94     for (w = 0; w < ics->num_windows; w++) {
95         bottom = ics->num_swb;
96         for (filt = 0; filt < tns->n_filt[w]; filt++) {
97             top    = bottom;
98             bottom = FFMAX(0, top - tns->length[w][filt]);
99             order  = tns->order[w][filt];
100             if (order == 0)
101                 continue;
102
103             // tns_decode_coef
104             compute_lpc_coefs(tns->coef[w][filt], order, lpc, 0, 0, 0);
105
106             start = ics->swb_offset[FFMIN(bottom, mmm)];
107             end   = ics->swb_offset[FFMIN(   top, mmm)];
108             if ((size = end - start) <= 0)
109                 continue;
110             if (tns->direction[w][filt]) {
111                 inc = -1;
112                 start = end - 1;
113             } else {
114                 inc = 1;
115             }
116             start += w * 128;
117
118             // ar filter
119             for (m = 0; m < size; m++, start += inc)
120                 for (i = 1; i <= FFMIN(m, order); i++)
121                     sce->coeffs[start] += lpc[i-1]*sce->pcoeffs[start - i*inc];
122         }
123     }
124 }
125
126 void ff_aac_search_for_tns(AACEncContext *s, SingleChannelElement *sce)
127 {
128     TemporalNoiseShaping *tns = &sce->tns;
129     int w, w2, g, count = 0;
130     const int mmm = FFMIN(sce->ics.tns_max_bands, sce->ics.max_sfb);
131     const int is8 = sce->ics.window_sequence[0] == EIGHT_SHORT_SEQUENCE;
132     int order = is8 ? 7 : s->profile == FF_PROFILE_AAC_LOW ? 12 : TNS_MAX_ORDER;
133
134     int sfb_start = av_clip(tns_min_sfb[is8][s->samplerate_index], 0, mmm);
135     int sfb_end   = av_clip(sce->ics.num_swb, 0, mmm);
136
137     for (w = 0; w < sce->ics.num_windows; w++) {
138         float en_low = 0.0f, en_high = 0.0f, threshold = 0.0f, spread = 0.0f;
139         double gain = 0.0f, coefs[MAX_LPC_ORDER] = {0};
140
141         int coef_start = w*sce->ics.num_swb + sce->ics.swb_offset[sfb_start];
142         int coef_len = sce->ics.swb_offset[sfb_end] - sce->ics.swb_offset[sfb_start];
143
144         for (g = 0;  g < sce->ics.num_swb; g++) {
145             if (w*16+g < sfb_start || w*16+g > sfb_end)
146                 continue;
147             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
148                 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
149                 if ((w+w2)*16+g > sfb_start + ((sfb_end - sfb_start)/2))
150                     en_high += band->energy;
151                 else
152                     en_low  += band->energy;
153                 threshold += band->threshold;
154                 spread += band->spread;
155             }
156         }
157
158         if (coef_len <= 0 || (sfb_end - sfb_start) <= 0)
159             continue;
160
161         /* LPC */
162         gain = ff_lpc_calc_ref_coefs_f(&s->lpc, &sce->coeffs[coef_start],
163                                        coef_len, order, coefs);
164
165         gain *= s->lambda/110.0f;
166
167         if (gain > TNS_GAIN_THRESHOLD_LOW && gain*0 < TNS_GAIN_THRESHOLD_HIGH &&
168             (en_low+en_high) > TNS_GAIN_THRESHOLD_LOW*threshold &&
169             spread > TNS_SPREAD_THRESHOLD) {
170             tns->n_filt[w] = 1;
171             for (g = 0; g < tns->n_filt[w]; g++) {
172                 tns->length[w][g] = sfb_end - sfb_start;
173                 tns->direction[w][g] = en_low < en_high && TNS_DIRECTION_VARY;
174                 tns->order[w][g] = quantize_coefs(coefs, tns->coef_idx[w][g],
175                                                   tns->coef[w][g], order);
176             }
177             count++;
178         }
179     }
180
181     sce->tns.present = !!count;
182 }