3 * Copyright (C) 2015 Rostislav Pehlivanov
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 temporal noise shaping
25 * @author Rostislav Pehlivanov ( atomnuker gmail com )
29 #include "aacenc_tns.h"
31 #include "aacenc_utils.h"
32 #include "aacenc_quantization.h"
36 * Coefficient compression saves a single bit per coefficient.
38 void ff_aac_encode_tns_info(AACEncContext *s, SingleChannelElement *sce)
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;
46 if (!sce->tns.present)
49 for (i = 0; i < sce->ics.num_windows; i++) {
50 put_bits(&s->pb, 2 - is8, sce->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);
70 /* Apply TNS filter */
71 void ff_aac_apply_tns(AACEncContext *s, SingleChannelElement *sce)
73 TemporalNoiseShaping *tns = &sce->tns;
74 IndividualChannelStream *ics = &sce->ics;
75 int w, filt, m, i, top, order, bottom, start, end, size, inc;
76 const int mmm = FFMIN(ics->tns_max_bands, ics->max_sfb);
77 float lpc[TNS_MAX_ORDER];
79 for (w = 0; w < ics->num_windows; w++) {
80 bottom = ics->num_swb;
81 for (filt = 0; filt < tns->n_filt[w]; filt++) {
83 bottom = FFMAX(0, top - tns->length[w][filt]);
84 order = tns->order[w][filt];
89 compute_lpc_coefs(tns->coef[w][filt], order, lpc, 0, 0, 0);
91 start = ics->swb_offset[FFMIN(bottom, mmm)];
92 end = ics->swb_offset[FFMIN( top, mmm)];
93 if ((size = end - start) <= 0)
95 if (tns->direction[w][filt]) {
104 for (m = 0; m < size; m++, start += inc)
105 for (i = 1; i <= FFMIN(m, order); i++)
106 sce->coeffs[start] += lpc[i-1]*sce->pcoeffs[start - i*inc];
112 * c_bits - 1 if 4 bit coefficients, 0 if 3 bit coefficients
114 static inline void quantize_coefs(double *coef, int *idx, float *lpc, int order,
118 const float *quant_arr = tns_tmp2_map[c_bits];
119 for (i = 0; i < order; i++) {
120 idx[i] = quant_array_idx((float)coef[i], quant_arr, c_bits ? 16 : 8);
121 lpc[i] = quant_arr[idx[i]];
126 * 3 bits per coefficient with 8 short windows
128 void ff_aac_search_for_tns(AACEncContext *s, SingleChannelElement *sce)
130 TemporalNoiseShaping *tns = &sce->tns;
131 int w, w2, g, count = 0;
132 const int mmm = FFMIN(sce->ics.tns_max_bands, sce->ics.max_sfb);
133 const int is8 = sce->ics.window_sequence[0] == EIGHT_SHORT_SEQUENCE;
134 const int c_bits = is8 ? TNS_Q_BITS_SHORT == 4 : TNS_Q_BITS == 4;
136 int sfb_start = av_clip(tns_min_sfb[is8][s->samplerate_index], 0, mmm);
137 int sfb_end = av_clip(sce->ics.num_swb, 0, mmm);
139 for (w = 0; w < sce->ics.num_windows; w++) {
141 int order = is8 ? 5 : s->profile == FF_PROFILE_AAC_LOW ? 12 : TNS_MAX_ORDER;
142 int coef_start = w*sce->ics.num_swb + sce->ics.swb_offset[sfb_start];
143 int coef_len = sce->ics.swb_offset[sfb_end] - sce->ics.swb_offset[sfb_start];
144 float e_ratio = 0.0f, threshold = 0.0f, spread = 0.0f, en[2] = {0.0, 0.0f};
145 double gain = 0.0f, coefs[MAX_LPC_ORDER] = {0};
147 for (g = 0; g < sce->ics.num_swb; g++) {
148 if (w*16+g < sfb_start || w*16+g > sfb_end)
150 for (w2 = 0; w2 < sce->ics.group_len[w]; w2++) {
151 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
152 if ((w+w2)*16+g > sfb_start + ((sfb_end - sfb_start)/2))
153 en[1] += band->energy;
155 en[0] += band->energy;
156 threshold += band->threshold;
157 spread += band->spread;
161 if (coef_len <= 0 || (sfb_end - sfb_start) <= 0)
165 gain = ff_lpc_calc_ref_coefs_f(&s->lpc, &sce->coeffs[coef_start],
166 coef_len, order, coefs);
168 if (!order || gain < TNS_GAIN_THRESHOLD_LOW || gain > TNS_GAIN_THRESHOLD_HIGH)
170 else if ((en[0]+en[1]) < TNS_GAIN_THRESHOLD_LOW*threshold || spread < TNS_SPREAD_THRESHOLD)
176 e_ratio = en[0]/en[1];
177 if (is8 || order < 2 || (e_ratio > TNS_E_RATIO_LOW && e_ratio < TNS_E_RATIO_HIGH)) {
179 for (g = 0; g < tns->n_filt[w]; g++) {
180 tns->length[w][g] = sfb_end - sfb_start;
181 tns->direction[w][g] = en[0] < en[1];
182 tns->order[w][g] = order;
183 quantize_coefs(coefs, tns->coef_idx[w][g], tns->coef[w][g],
186 } else { /* 2 filters due to energy disbalance */
188 for (g = 0; g < tns->n_filt[w]; g++) {
189 tns->direction[w][g] = en[g] < en[!g];
190 tns->order[w][g] = !g ? order/2 : order - tns->order[w][g-1];
191 tns->length[w][g] = !g ? (sfb_end - sfb_start)/2 : \
192 (sfb_end - sfb_start) - tns->length[w][g-1];
193 quantize_coefs(&coefs[!g ? 0 : order - tns->order[w][g-1]],
194 tns->coef_idx[w][g], tns->coef[w][g],
195 tns->order[w][g], c_bits);
201 sce->tns.present = !!count;