3 * Copyright (c) 2006 Justin Ruggles <justin.ruggles@gmail.com>
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
22 #include "libavutil/lls.h"
25 #define LPC_USE_DOUBLE
30 * Apply Welch window function to audio block
32 static void apply_welch_window(const int32_t *data, int len, double *w_data)
38 assert(!(len&1)); //the optimization in r11881 does not support odd len
39 //if someone wants odd len extend the change in r11881
42 c = 2.0 / (len - 1.0);
49 w_data[-i-1] = data[-i-1] * w;
50 w_data[+i ] = data[+i ] * w;
55 * Calculates autocorrelation data from audio samples
56 * A Welch window function is applied before calculation.
58 void ff_lpc_compute_autocorr(const int32_t *data, int len, int lag,
62 double tmp[len + lag + 1];
63 double *data1= tmp + lag;
65 apply_welch_window(data, len, data1);
71 for(j=0; j<lag; j+=2){
72 double sum0 = 1.0, sum1 = 1.0;
74 sum0 += data1[i] * data1[i-j];
75 sum1 += data1[i] * data1[i-j-1];
83 for(i=j-1; i<len; i+=2){
84 sum += data1[i ] * data1[i-j ]
85 + data1[i+1] * data1[i-j+1];
92 * Quantize LPC coefficients
94 static void quantize_lpc_coefs(double *lpc_in, int order, int precision,
95 int32_t *lpc_out, int *shift, int max_shift, int zero_shift)
102 /* define maximum levels */
103 qmax = (1 << (precision - 1)) - 1;
105 /* find maximum coefficient value */
107 for(i=0; i<order; i++) {
108 cmax= FFMAX(cmax, fabs(lpc_in[i]));
111 /* if maximum value quantizes to zero, return all zeros */
112 if(cmax * (1 << max_shift) < 1.0) {
114 memset(lpc_out, 0, sizeof(int32_t) * order);
118 /* calculate level shift which scales max coeff to available bits */
120 while((cmax * (1 << sh) > qmax) && (sh > 0)) {
124 /* since negative shift values are unsupported in decoder, scale down
125 coefficients instead */
126 if(sh == 0 && cmax > qmax) {
127 double scale = ((double)qmax) / cmax;
128 for(i=0; i<order; i++) {
133 /* output quantized coefficients and level shift */
135 for(i=0; i<order; i++) {
136 error -= lpc_in[i] * (1 << sh);
137 lpc_out[i] = av_clip(lrintf(error), -qmax, qmax);
143 static int estimate_best_order(double *ref, int min_order, int max_order)
148 for(i=max_order-1; i>=min_order-1; i--) {
158 * Calculate LPC coefficients for multiple orders
160 * @param use_lpc LPC method for determining coefficients
161 * 0 = LPC with fixed pre-defined coeffs
162 * 1 = LPC with coeffs determined by Levinson-Durbin recursion
163 * 2+ = LPC with coeffs determined by Cholesky factorization using (use_lpc-1) passes.
165 int ff_lpc_calc_coefs(DSPContext *s,
166 const int32_t *samples, int blocksize, int min_order,
167 int max_order, int precision,
168 int32_t coefs[][MAX_LPC_ORDER], int *shift, int use_lpc,
169 int omethod, int max_shift, int zero_shift)
171 double autoc[MAX_LPC_ORDER+1];
172 double ref[MAX_LPC_ORDER];
173 double lpc[MAX_LPC_ORDER][MAX_LPC_ORDER];
177 assert(max_order >= MIN_LPC_ORDER && max_order <= MAX_LPC_ORDER && use_lpc > 0);
180 s->lpc_compute_autocorr(samples, blocksize, max_order, autoc);
182 compute_lpc_coefs(autoc, max_order, &lpc[0][0], MAX_LPC_ORDER, 0, 1);
184 for(i=0; i<max_order; i++)
185 ref[i] = fabs(lpc[i][i]);
188 double var[MAX_LPC_ORDER+1], av_uninit(weight);
190 for(pass=0; pass<use_lpc-1; pass++){
191 av_init_lls(&m[pass&1], max_order);
194 for(i=max_order; i<blocksize; i++){
195 for(j=0; j<=max_order; j++)
196 var[j]= samples[i-j];
199 double eval, inv, rinv;
200 eval= av_evaluate_lls(&m[(pass-1)&1], var+1, max_order-1);
201 eval= (512>>pass) + fabs(eval - var[0]);
204 for(j=0; j<=max_order; j++)
210 av_update_lls(&m[pass&1], var, 1.0);
212 av_solve_lls(&m[pass&1], 0.001, 0);
215 for(i=0; i<max_order; i++){
216 for(j=0; j<max_order; j++)
217 lpc[i][j]=-m[(pass-1)&1].coeff[i][j];
218 ref[i]= sqrt(m[(pass-1)&1].variance[i] / weight) * (blocksize - max_order) / 4000;
220 for(i=max_order-1; i>0; i--)
221 ref[i] = ref[i-1] - ref[i];
223 opt_order = max_order;
225 if(omethod == ORDER_METHOD_EST) {
226 opt_order = estimate_best_order(ref, min_order, max_order);
228 quantize_lpc_coefs(lpc[i], i+1, precision, coefs[i], &shift[i], max_shift, zero_shift);
230 for(i=min_order-1; i<max_order; i++) {
231 quantize_lpc_coefs(lpc[i], i+1, precision, coefs[i], &shift[i], max_shift, zero_shift);