1 /*
2 * DCA ADPCM engine
3 * Copyright (C) 2017 Daniil Cherednik
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 #include "dcaadpcm.h"
24 #include "dcaenc.h"
25 #include "dca_core.h"
26 #include "mathops.h"
27
28 typedef int32_t premultiplied_coeffs[10];
29
30 //assume we have DCA_ADPCM_COEFFS values before x
calc_corr(const int32_t * x,int len,int j,int k)31 static inline int64_t calc_corr(const int32_t *x, int len, int j, int k)
32 {
33 int n;
34 int64_t s = 0;
35 for (n = 0; n < len; n++)
36 s += MUL64(x[n-j], x[n-k]);
37 return s;
38 }
39
apply_filter(const int16_t a[DCA_ADPCM_COEFFS],const int64_t corr[15],const int32_t aa[10])40 static inline int64_t apply_filter(const int16_t a[DCA_ADPCM_COEFFS], const int64_t corr[15], const int32_t aa[10])
41 {
42 int64_t err = 0;
43 int64_t tmp = 0;
44
45 err = corr[0];
46
47 tmp += MUL64(a[0], corr[1]);
48 tmp += MUL64(a[1], corr[2]);
49 tmp += MUL64(a[2], corr[3]);
50 tmp += MUL64(a[3], corr[4]);
51
52 tmp = norm__(tmp, 13);
53 tmp += tmp;
54
55 err -= tmp;
56 tmp = 0;
57
58 tmp += MUL64(corr[5], aa[0]);
59 tmp += MUL64(corr[6], aa[1]);
60 tmp += MUL64(corr[7], aa[2]);
61 tmp += MUL64(corr[8], aa[3]);
62
63 tmp += MUL64(corr[9], aa[4]);
64 tmp += MUL64(corr[10], aa[5]);
65 tmp += MUL64(corr[11], aa[6]);
66
67 tmp += MUL64(corr[12], aa[7]);
68 tmp += MUL64(corr[13], aa[8]);
69
70 tmp += MUL64(corr[14], aa[9]);
71
72 tmp = norm__(tmp, 26);
73
74 err += tmp;
75
76 return llabs(err);
77 }
78
find_best_filter(const DCAADPCMEncContext * s,const int32_t * in,int len)79 static int64_t find_best_filter(const DCAADPCMEncContext *s, const int32_t *in, int len)
80 {
81 const premultiplied_coeffs *precalc_data = s->private_data;
82 int i, j, k = 0;
83 int vq = -1;
84 int64_t err;
85 int64_t min_err = 1ll << 62;
86 int64_t corr[15];
87
88 for (i = 0; i <= DCA_ADPCM_COEFFS; i++)
89 for (j = i; j <= DCA_ADPCM_COEFFS; j++)
90 corr[k++] = calc_corr(in+4, len, i, j);
91
92 for (i = 0; i < DCA_ADPCM_VQCODEBOOK_SZ; i++) {
93 err = apply_filter(ff_dca_adpcm_vb[i], corr, *precalc_data);
94 if (err < min_err) {
95 min_err = err;
96 vq = i;
97 }
98 precalc_data++;
99 }
100
101 return vq;
102 }
103
calc_prediction_gain(int pred_vq,const int32_t * in,int32_t * out,int len)104 static inline int64_t calc_prediction_gain(int pred_vq, const int32_t *in, int32_t *out, int len)
105 {
106 int i;
107 int32_t error;
108
109 int64_t signal_energy = 0;
110 int64_t error_energy = 0;
111
112 for (i = 0; i < len; i++) {
113 error = in[DCA_ADPCM_COEFFS + i] - ff_dcaadpcm_predict(pred_vq, in + i);
114 out[i] = error;
115 signal_energy += MUL64(in[DCA_ADPCM_COEFFS + i], in[DCA_ADPCM_COEFFS + i]);
116 error_energy += MUL64(error, error);
117 }
118
119 if (!error_energy)
120 return -1;
121
122 return signal_energy / error_energy;
123 }
124
ff_dcaadpcm_subband_analysis(const DCAADPCMEncContext * s,const int32_t * in,int len,int * diff)125 int ff_dcaadpcm_subband_analysis(const DCAADPCMEncContext *s, const int32_t *in, int len, int *diff)
126 {
127 int pred_vq, i;
128 int32_t input_buffer[16 + DCA_ADPCM_COEFFS];
129 int32_t input_buffer2[16 + DCA_ADPCM_COEFFS];
130
131 int32_t max = 0;
132 int shift_bits;
133 uint64_t pg = 0;
134
135 for (i = 0; i < len + DCA_ADPCM_COEFFS; i++)
136 max |= FFABS(in[i]);
137
138 // normalize input to simplify apply_filter
139 shift_bits = av_log2(max) - 11;
140
141 for (i = 0; i < len + DCA_ADPCM_COEFFS; i++) {
142 input_buffer[i] = norm__(in[i], 7);
143 input_buffer2[i] = norm__(in[i], shift_bits);
144 }
145
146 pred_vq = find_best_filter(s, input_buffer2, len);
147
148 if (pred_vq < 0)
149 return -1;
150
151 pg = calc_prediction_gain(pred_vq, input_buffer, diff, len);
152
153 // Greater than 10db (10*log(10)) prediction gain to use ADPCM.
154 // TODO: Tune it.
155 if (pg < 10)
156 return -1;
157
158 for (i = 0; i < len; i++)
159 diff[i] <<= 7;
160
161 return pred_vq;
162 }
163
precalc(premultiplied_coeffs * data)164 static void precalc(premultiplied_coeffs *data)
165 {
166 int i, j, k;
167
168 for (i = 0; i < DCA_ADPCM_VQCODEBOOK_SZ; i++) {
169 int id = 0;
170 int32_t t = 0;
171 for (j = 0; j < DCA_ADPCM_COEFFS; j++) {
172 for (k = j; k < DCA_ADPCM_COEFFS; k++) {
173 t = (int32_t)ff_dca_adpcm_vb[i][j] * (int32_t)ff_dca_adpcm_vb[i][k];
174 if (j != k)
175 t *= 2;
176 (*data)[id++] = t;
177 }
178 }
179 data++;
180 }
181 }
182
ff_dcaadpcm_do_real(int pred_vq_index,softfloat quant,int32_t scale_factor,int32_t step_size,const int32_t * prev_hist,const int32_t * in,int32_t * next_hist,int32_t * out,int len,int32_t peak)183 int ff_dcaadpcm_do_real(int pred_vq_index,
184 softfloat quant, int32_t scale_factor, int32_t step_size,
185 const int32_t *prev_hist, const int32_t *in, int32_t *next_hist, int32_t *out,
186 int len, int32_t peak)
187 {
188 int i;
189 int64_t delta;
190 int32_t dequant_delta;
191 int32_t work_bufer[16 + DCA_ADPCM_COEFFS];
192
193 memcpy(work_bufer, prev_hist, sizeof(int32_t) * DCA_ADPCM_COEFFS);
194
195 for (i = 0; i < len; i++) {
196 work_bufer[DCA_ADPCM_COEFFS + i] = ff_dcaadpcm_predict(pred_vq_index, &work_bufer[i]);
197
198 delta = (int64_t)in[i] - ((int64_t)work_bufer[DCA_ADPCM_COEFFS + i] << 7);
199
200 out[i] = quantize_value(av_clip64(delta, -peak, peak), quant);
201
202 ff_dca_core_dequantize(&dequant_delta, &out[i], step_size, scale_factor, 0, 1);
203
204 work_bufer[DCA_ADPCM_COEFFS+i] += dequant_delta;
205 }
206
207 memcpy(next_hist, &work_bufer[len], sizeof(int32_t) * DCA_ADPCM_COEFFS);
208
209 return 0;
210 }
211
ff_dcaadpcm_init(DCAADPCMEncContext * s)212 av_cold int ff_dcaadpcm_init(DCAADPCMEncContext *s)
213 {
214 if (!s)
215 return -1;
216
217 s->private_data = av_malloc(sizeof(premultiplied_coeffs) * DCA_ADPCM_VQCODEBOOK_SZ);
218 if (!s->private_data)
219 return AVERROR(ENOMEM);
220
221 precalc(s->private_data);
222 return 0;
223 }
224
ff_dcaadpcm_free(DCAADPCMEncContext * s)225 av_cold void ff_dcaadpcm_free(DCAADPCMEncContext *s)
226 {
227 if (!s)
228 return;
229
230 av_freep(&s->private_data);
231 }
232