aacenc_tns: add moving average filter for LTP
[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  * Shifts the values as well if compression is possible.
36  */
37 static inline int compress_coeffs(int *coef, int order, int c_bits)
38 {
39     int i, res = 0;
40     const int low_idx   = c_bits ?  4 : 2;
41     const int shift_val = c_bits ?  8 : 4;
42     const int high_idx  = c_bits ? 11 : 5;
43     for (i = 0; i < order; i++)
44         if (coef[i] < low_idx || coef[i] > high_idx)
45             res++;
46     if (res == order)
47         for (i = 0; i < order; i++)
48             coef[i] -= (coef[i] > high_idx) ? shift_val : 0;
49     return res == order;
50 }
51
52 /**
53  * Encode TNS data.
54  * Coefficient compression saves a single bit per coefficient.
55  */
56 void ff_aac_encode_tns_info(AACEncContext *s, SingleChannelElement *sce)
57 {
58     int i, w, filt, coef_len, coef_compress = 0;
59     const int is8 = sce->ics.window_sequence[0] == EIGHT_SHORT_SEQUENCE;
60     TemporalNoiseShaping *tns = &sce->tns;
61     const int c_bits = is8 ? TNS_Q_BITS_SHORT == 4 : TNS_Q_BITS == 4;
62
63     if (!sce->tns.present)
64         return;
65
66     for (i = 0; i < sce->ics.num_windows; i++) {
67         put_bits(&s->pb, 2 - is8, sce->tns.n_filt[i]);
68         if (tns->n_filt[i]) {
69             put_bits(&s->pb, 1, c_bits);
70             for (filt = 0; filt < tns->n_filt[i]; filt++) {
71                 put_bits(&s->pb, 6 - 2 * is8, tns->length[i][filt]);
72                 put_bits(&s->pb, 5 - 2 * is8, tns->order[i][filt]);
73                 if (tns->order[i][filt]) {
74                     coef_compress = compress_coeffs(tns->coef_idx[i][filt],
75                                                     tns->order[i][filt], c_bits);
76                     put_bits(&s->pb, 1, !!tns->direction[i][filt]);
77                     put_bits(&s->pb, 1, !!coef_compress);
78                     coef_len = c_bits + 3 - coef_compress;
79                     for (w = 0; w < tns->order[i][filt]; w++)
80                         put_bits(&s->pb, coef_len, tns->coef_idx[i][filt][w]);
81                 }
82             }
83         }
84     }
85 }
86
87 /* Apply TNS filter */
88 void ff_aac_apply_tns(AACEncContext *s, SingleChannelElement *sce)
89 {
90     TemporalNoiseShaping *tns = &sce->tns;
91     IndividualChannelStream *ics = &sce->ics;
92     int w, filt, m, i, top, order, bottom, start, end, size, inc;
93     const int mmm = FFMIN(ics->tns_max_bands, ics->max_sfb);
94     float lpc[TNS_MAX_ORDER], tmp[TNS_MAX_ORDER+1];
95
96     for (w = 0; w < ics->num_windows; w++) {
97         bottom = ics->num_swb;
98         for (filt = 0; filt < tns->n_filt[w]; filt++) {
99             top    = bottom;
100             bottom = FFMAX(0, top - tns->length[w][filt]);
101             order  = tns->order[w][filt];
102             if (order == 0)
103                 continue;
104
105             // tns_decode_coef
106             compute_lpc_coefs(tns->coef[w][filt], order, lpc, 0, 0, 0);
107
108             start = ics->swb_offset[FFMIN(bottom, mmm)];
109             end   = ics->swb_offset[FFMIN(   top, mmm)];
110             if ((size = end - start) <= 0)
111                 continue;
112             if (tns->direction[w][filt]) {
113                 inc = -1;
114                 start = end - 1;
115             } else {
116                 inc = 1;
117             }
118             start += w * 128;
119
120             if (!s->options.ltp) {     // ar filter
121                 for (m = 0; m < size; m++, start += inc) {
122                     for (i = 1; i <= FFMIN(m, order); i++) {
123                         sce->coeffs[start] += lpc[i-1]*sce->pcoeffs[start - i*inc];
124                     }
125                 }
126             } else {                   // ma filter
127                 for (m = 0; m < size; m++, start += inc) {
128                     tmp[0] = sce->pcoeffs[start];
129                     for (i = 1; i <= FFMIN(m, order); i++)
130                         sce->coeffs[start] += lpc[i-1]*tmp[i];
131                     for (i = order; i > 0; i--)
132                         tmp[i] = tmp[i - 1];
133                 }
134             }
135         }
136     }
137 }
138
139 /*
140  * c_bits - 1 if 4 bit coefficients, 0 if 3 bit coefficients
141  */
142 static inline void quantize_coefs(double *coef, int *idx, float *lpc, int order,
143                                   int c_bits)
144 {
145     int i;
146     const float *quant_arr = tns_tmp2_map[c_bits];
147     for (i = 0; i < order; i++) {
148         idx[i] = quant_array_idx((float)coef[i], quant_arr, c_bits ? 16 : 8);
149         lpc[i] = quant_arr[idx[i]];
150     }
151 }
152
153 /*
154  * 3 bits per coefficient with 8 short windows
155  */
156 void ff_aac_search_for_tns(AACEncContext *s, SingleChannelElement *sce)
157 {
158     TemporalNoiseShaping *tns = &sce->tns;
159     int w, w2, g, count = 0;
160     const int mmm = FFMIN(sce->ics.tns_max_bands, sce->ics.max_sfb);
161     const int is8 = sce->ics.window_sequence[0] == EIGHT_SHORT_SEQUENCE;
162     const int c_bits = is8 ? TNS_Q_BITS_SHORT == 4 : TNS_Q_BITS == 4;
163
164     int sfb_start = av_clip(tns_min_sfb[is8][s->samplerate_index], 0, mmm);
165     int sfb_end   = av_clip(sce->ics.num_swb, 0, mmm);
166
167     for (w = 0; w < sce->ics.num_windows; w++) {
168         int use_tns;
169         int order = is8 ? 5 : s->profile == FF_PROFILE_AAC_LOW ? 12 : TNS_MAX_ORDER;
170         int coef_start = w*sce->ics.num_swb + sce->ics.swb_offset[sfb_start];
171         int coef_len = sce->ics.swb_offset[sfb_end] - sce->ics.swb_offset[sfb_start];
172         float e_ratio = 0.0f, threshold = 0.0f, spread = 0.0f, en[2] = {0.0, 0.0f};
173         double gain = 0.0f, coefs[MAX_LPC_ORDER] = {0};
174
175         for (g = 0;  g < sce->ics.num_swb; g++) {
176             if (w*16+g < sfb_start || w*16+g > sfb_end)
177                 continue;
178             for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
179                 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
180                 if ((w+w2)*16+g > sfb_start + ((sfb_end - sfb_start)/2))
181                     en[1] += band->energy;
182                 else
183                     en[0] += band->energy;
184                 threshold += band->threshold;
185                 spread    += band->spread;
186             }
187         }
188
189         if (coef_len <= 0 || (sfb_end - sfb_start) <= 0)
190             continue;
191
192         /* LPC */
193         gain = ff_lpc_calc_ref_coefs_f(&s->lpc, &sce->coeffs[coef_start],
194                                        coef_len, order, coefs);
195
196         if (!order || gain < TNS_GAIN_THRESHOLD_LOW || gain > TNS_GAIN_THRESHOLD_HIGH)
197             use_tns = 0;
198         else if ((en[0]+en[1]) < TNS_GAIN_THRESHOLD_LOW*threshold || spread < TNS_SPREAD_THRESHOLD)
199             use_tns = 0;
200         else
201             use_tns = 1;
202
203         if (use_tns) {
204             e_ratio = en[0]/en[1];
205             if (is8 || order < 2 || (e_ratio > TNS_E_RATIO_LOW && e_ratio < TNS_E_RATIO_HIGH)) {
206                 tns->n_filt[w] = 1;
207                 for (g = 0; g < tns->n_filt[w]; g++) {
208                     tns->length[w][g] = sfb_end - sfb_start;
209                     tns->direction[w][g] = en[0] < en[1];
210                     tns->order[w][g] = order;
211                     quantize_coefs(coefs, tns->coef_idx[w][g], tns->coef[w][g],
212                                    order, c_bits);
213                 }
214             } else {  /* 2 filters due to energy disbalance */
215                 tns->n_filt[w] = 2;
216                 for (g = 0; g < tns->n_filt[w]; g++) {
217                     tns->direction[w][g] = en[g] < en[!g];
218                     tns->order[w][g] = !g ? order/2 : order - tns->order[w][g-1];
219                     tns->length[w][g] = !g ? (sfb_end - sfb_start)/2 : \
220                                     (sfb_end - sfb_start) - tns->length[w][g-1];
221                     quantize_coefs(&coefs[!g ? 0 : order - tns->order[w][g-1]],
222                                    tns->coef_idx[w][g], tns->coef[w][g],
223                                    tns->order[w][g], c_bits);
224                 }
225             }
226             count++;
227         }
228     }
229     sce->tns.present = !!count;
230 }