Replace Subversion revisions in comments by Git hashes.
[libav.git] / libavcodec / lpc.c
1 /**
2 * LPC utility code
3 * Copyright (c) 2006 Justin Ruggles <justin.ruggles@gmail.com>
4 *
5 * This file is part of Libav.
6 *
7 * Libav 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 * Libav 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 Libav; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20 */
21
22 #include "libavutil/lls.h"
23
24 #define LPC_USE_DOUBLE
25 #include "lpc.h"
26
27
28 /**
29 * Apply Welch window function to audio block
30 */
31 static void lpc_apply_welch_window_c(const int32_t *data, int len,
32 double *w_data)
33 {
34 int i, n2;
35 double w;
36 double c;
37
38 /* The optimization in commit fa4ed8c does not support odd len.
39 * If someone wants odd len extend that change. */
40 assert(!(len & 1));
41
42 n2 = (len >> 1);
43 c = 2.0 / (len - 1.0);
44
45 w_data+=n2;
46 data+=n2;
47 for(i=0; i<n2; i++) {
48 w = c - n2 + i;
49 w = 1.0 - (w * w);
50 w_data[-i-1] = data[-i-1] * w;
51 w_data[+i ] = data[+i ] * w;
52 }
53 }
54
55 /**
56 * Calculate autocorrelation data from audio samples
57 * A Welch window function is applied before calculation.
58 */
59 static void lpc_compute_autocorr_c(const double *data, int len, int lag,
60 double *autoc)
61 {
62 int i, j;
63
64 for(j=0; j<lag; j+=2){
65 double sum0 = 1.0, sum1 = 1.0;
66 for(i=j; i<len; i++){
67 sum0 += data[i] * data[i-j];
68 sum1 += data[i] * data[i-j-1];
69 }
70 autoc[j ] = sum0;
71 autoc[j+1] = sum1;
72 }
73
74 if(j==lag){
75 double sum = 1.0;
76 for(i=j-1; i<len; i+=2){
77 sum += data[i ] * data[i-j ]
78 + data[i+1] * data[i-j+1];
79 }
80 autoc[j] = sum;
81 }
82 }
83
84 /**
85 * Quantize LPC coefficients
86 */
87 static void quantize_lpc_coefs(double *lpc_in, int order, int precision,
88 int32_t *lpc_out, int *shift, int max_shift, int zero_shift)
89 {
90 int i;
91 double cmax, error;
92 int32_t qmax;
93 int sh;
94
95 /* define maximum levels */
96 qmax = (1 << (precision - 1)) - 1;
97
98 /* find maximum coefficient value */
99 cmax = 0.0;
100 for(i=0; i<order; i++) {
101 cmax= FFMAX(cmax, fabs(lpc_in[i]));
102 }
103
104 /* if maximum value quantizes to zero, return all zeros */
105 if(cmax * (1 << max_shift) < 1.0) {
106 *shift = zero_shift;
107 memset(lpc_out, 0, sizeof(int32_t) * order);
108 return;
109 }
110
111 /* calculate level shift which scales max coeff to available bits */
112 sh = max_shift;
113 while((cmax * (1 << sh) > qmax) && (sh > 0)) {
114 sh--;
115 }
116
117 /* since negative shift values are unsupported in decoder, scale down
118 coefficients instead */
119 if(sh == 0 && cmax > qmax) {
120 double scale = ((double)qmax) / cmax;
121 for(i=0; i<order; i++) {
122 lpc_in[i] *= scale;
123 }
124 }
125
126 /* output quantized coefficients and level shift */
127 error=0;
128 for(i=0; i<order; i++) {
129 error -= lpc_in[i] * (1 << sh);
130 lpc_out[i] = av_clip(lrintf(error), -qmax, qmax);
131 error -= lpc_out[i];
132 }
133 *shift = sh;
134 }
135
136 static int estimate_best_order(double *ref, int min_order, int max_order)
137 {
138 int i, est;
139
140 est = min_order;
141 for(i=max_order-1; i>=min_order-1; i--) {
142 if(ref[i] > 0.10) {
143 est = i+1;
144 break;
145 }
146 }
147 return est;
148 }
149
150 /**
151 * Calculate LPC coefficients for multiple orders
152 *
153 * @param lpc_type LPC method for determining coefficients,
154 * see #FFLPCType for details
155 */
156 int ff_lpc_calc_coefs(LPCContext *s,
157 const int32_t *samples, int blocksize, int min_order,
158 int max_order, int precision,
159 int32_t coefs[][MAX_LPC_ORDER], int *shift,
160 enum FFLPCType lpc_type, int lpc_passes,
161 int omethod, int max_shift, int zero_shift)
162 {
163 double autoc[MAX_LPC_ORDER+1];
164 double ref[MAX_LPC_ORDER];
165 double lpc[MAX_LPC_ORDER][MAX_LPC_ORDER];
166 int i, j, pass;
167 int opt_order;
168
169 assert(max_order >= MIN_LPC_ORDER && max_order <= MAX_LPC_ORDER &&
170 lpc_type > FF_LPC_TYPE_FIXED);
171
172 /* reinit LPC context if parameters have changed */
173 if (blocksize != s->blocksize || max_order != s->max_order ||
174 lpc_type != s->lpc_type) {
175 ff_lpc_end(s);
176 ff_lpc_init(s, blocksize, max_order, lpc_type);
177 }
178
179 if (lpc_type == FF_LPC_TYPE_LEVINSON) {
180 double *windowed_samples = s->windowed_samples + max_order;
181
182 s->lpc_apply_welch_window(samples, blocksize, windowed_samples);
183
184 s->lpc_compute_autocorr(windowed_samples, blocksize, max_order, autoc);
185
186 compute_lpc_coefs(autoc, max_order, &lpc[0][0], MAX_LPC_ORDER, 0, 1);
187
188 for(i=0; i<max_order; i++)
189 ref[i] = fabs(lpc[i][i]);
190 } else if (lpc_type == FF_LPC_TYPE_CHOLESKY) {
191 LLSModel m[2];
192 double var[MAX_LPC_ORDER+1], av_uninit(weight);
193
194 for(pass=0; pass<lpc_passes; pass++){
195 av_init_lls(&m[pass&1], max_order);
196
197 weight=0;
198 for(i=max_order; i<blocksize; i++){
199 for(j=0; j<=max_order; j++)
200 var[j]= samples[i-j];
201
202 if(pass){
203 double eval, inv, rinv;
204 eval= av_evaluate_lls(&m[(pass-1)&1], var+1, max_order-1);
205 eval= (512>>pass) + fabs(eval - var[0]);
206 inv = 1/eval;
207 rinv = sqrt(inv);
208 for(j=0; j<=max_order; j++)
209 var[j] *= rinv;
210 weight += inv;
211 }else
212 weight++;
213
214 av_update_lls(&m[pass&1], var, 1.0);
215 }
216 av_solve_lls(&m[pass&1], 0.001, 0);
217 }
218
219 for(i=0; i<max_order; i++){
220 for(j=0; j<max_order; j++)
221 lpc[i][j]=-m[(pass-1)&1].coeff[i][j];
222 ref[i]= sqrt(m[(pass-1)&1].variance[i] / weight) * (blocksize - max_order) / 4000;
223 }
224 for(i=max_order-1; i>0; i--)
225 ref[i] = ref[i-1] - ref[i];
226 }
227 opt_order = max_order;
228
229 if(omethod == ORDER_METHOD_EST) {
230 opt_order = estimate_best_order(ref, min_order, max_order);
231 i = opt_order-1;
232 quantize_lpc_coefs(lpc[i], i+1, precision, coefs[i], &shift[i], max_shift, zero_shift);
233 } else {
234 for(i=min_order-1; i<max_order; i++) {
235 quantize_lpc_coefs(lpc[i], i+1, precision, coefs[i], &shift[i], max_shift, zero_shift);
236 }
237 }
238
239 return opt_order;
240 }
241
242 av_cold int ff_lpc_init(LPCContext *s, int blocksize, int max_order,
243 enum FFLPCType lpc_type)
244 {
245 s->blocksize = blocksize;
246 s->max_order = max_order;
247 s->lpc_type = lpc_type;
248
249 if (lpc_type == FF_LPC_TYPE_LEVINSON) {
250 s->windowed_samples = av_mallocz((blocksize + max_order + 2) *
251 sizeof(*s->windowed_samples));
252 if (!s->windowed_samples)
253 return AVERROR(ENOMEM);
254 } else {
255 s->windowed_samples = NULL;
256 }
257
258 s->lpc_apply_welch_window = lpc_apply_welch_window_c;
259 s->lpc_compute_autocorr = lpc_compute_autocorr_c;
260
261 if (HAVE_MMX)
262 ff_lpc_init_x86(s);
263
264 return 0;
265 }
266
267 av_cold void ff_lpc_end(LPCContext *s)
268 {
269 av_freep(&s->windowed_samples);
270 }