fix LOCAL_FABS bug
[flac.git] / src / libFLAC / lpc.c
1 /* libFLAC - Free Lossless Audio Codec library
2  * Copyright (C) 2000,2001  Josh Coalson
3  *
4  * This library is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU Library General Public
6  * License as published by the Free Software Foundation; either
7  * version 2 of the License, or (at your option) any later version.
8  *
9  * This library is distributed in the hope that it will be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
12  * Library General Public License for more details.
13  *
14  * You should have received a copy of the GNU Library General Public
15  * License along with this library; if not, write to the
16  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
17  * Boston, MA  02111-1307, USA.
18  */
19
20 #include <assert.h>
21 #include <math.h>
22 #include <stdio.h>
23 #include "FLAC/format.h"
24 #include "private/lpc.h"
25
26 #ifndef M_LN2
27 /* math.h in VC++ doesn't seem to have this (how Microsoft is that?) */
28 #define M_LN2 0.69314718055994530942
29 #endif
30
31 #define LOCAL_FABS(x) ((x)<0.0? -(x):(x))
32
33 void FLAC__lpc_compute_autocorrelation(const real data[], unsigned data_len, unsigned lag, real autoc[])
34 {
35         /* a readable, but slower, version */
36 #if 0
37         real d;
38         unsigned i;
39
40         assert(lag > 0);
41         assert(lag <= data_len);
42
43         while(lag--) {
44                 for(i = lag, d = 0.0; i < data_len; i++)
45                         d += data[i] * data[i - lag];
46                 autoc[lag] = d;
47         }
48 #endif
49
50         /*
51          * this version tends to run faster because of better data locality
52          * ('data_len' is usually much larger than 'lag')
53          */
54         real d;
55         unsigned sample, coeff;
56         const unsigned limit = data_len - lag;
57
58         assert(lag > 0);
59         assert(lag <= data_len);
60
61         for(coeff = 0; coeff < lag; coeff++)
62                 autoc[coeff] = 0.0;
63         for(sample = 0; sample <= limit; sample++) {
64                 d = data[sample];
65                 for(coeff = 0; coeff < lag; coeff++)
66                         autoc[coeff] += d * data[sample+coeff];
67         }
68         for(; sample < data_len; sample++) {
69                 d = data[sample];
70                 for(coeff = 0; coeff < data_len - sample; coeff++)
71                         autoc[coeff] += d * data[sample+coeff];
72         }
73 }
74
75 void FLAC__lpc_compute_lp_coefficients(const real autoc[], unsigned max_order, real lp_coeff[][FLAC__MAX_LPC_ORDER], real error[])
76 {
77         unsigned i, j;
78         real r, err, ref[FLAC__MAX_LPC_ORDER], lpc[FLAC__MAX_LPC_ORDER];
79
80         assert(0 < max_order);
81         assert(max_order <= FLAC__MAX_LPC_ORDER);
82         assert(autoc[0] != 0.0);
83
84         err = autoc[0];
85
86         for(i = 0; i < max_order; i++) {
87                 /* Sum up this iteration's reflection coefficient. */
88                 r =- autoc[i+1];
89                 for(j = 0; j < i; j++)
90                         r -= lpc[j] * autoc[i-j];
91                 ref[i] = (r/=err);
92
93                 /* Update LPC coefficients and total error. */
94                 lpc[i]=r;
95                 for(j = 0; j < (i>>1); j++) {
96                         real tmp = lpc[j];
97                         lpc[j] += r * lpc[i-1-j];
98                         lpc[i-1-j] += r * tmp;
99                 }
100                 if(i & 1)
101                         lpc[j] += lpc[j] * r;
102
103                 err *= (1.0 - r * r);
104
105                 /* save this order */
106                 for(j = 0; j <= i; j++)
107                         lp_coeff[i][j] = -lpc[j]; /* negate FIR filter coeff to get predictor coeff */
108                 error[i] = err;
109         }
110 }
111
112 int FLAC__lpc_quantize_coefficients(const real lp_coeff[], unsigned order, unsigned precision, unsigned bits_per_sample, int32 qlp_coeff[], int *shift)
113 {
114         unsigned i;
115         real d, cmax = -1e99;
116
117         assert(bits_per_sample > 0);
118         assert(bits_per_sample <= sizeof(int32)*8);
119         assert(precision > 0);
120         assert(precision >= FLAC__MIN_QLP_COEFF_PRECISION);
121         assert(precision + bits_per_sample < sizeof(int32)*8);
122 #ifdef NDEBUG
123         (void)bits_per_sample; /* silence compiler warning about unused parameter */
124 #endif
125
126         /* drop one bit for the sign; from here on out we consider only |lp_coeff[i]| */
127         precision--;
128
129         for(i = 0; i < order; i++) {
130                 if(lp_coeff[i] == 0.0)
131                         continue;
132                 d = LOCAL_FABS(lp_coeff[i]);
133                 if(d > cmax)
134                         cmax = d;
135         }
136         if(cmax < 0.0) {
137                 /* => coefficients are all 0, which means our constant-detect didn't work */
138                 return 2;
139         }
140         else {
141                 const int maxshift = (int)precision - (int)floor(log(cmax) / M_LN2) - 1;
142                 const int max_shiftlimit = (1 << (FLAC__SUBFRAME_LPC_QLP_SHIFT_LEN-1)) - 1;
143                 const int min_shiftlimit = -max_shiftlimit - 1;
144
145                 *shift = maxshift;
146
147                 if(*shift < min_shiftlimit || *shift > max_shiftlimit) {
148                         return 1;
149                 }
150         }
151
152         if(*shift != 0) { /* just to avoid wasting time... */
153                 for(i = 0; i < order; i++)
154                         qlp_coeff[i] = (int32)floor(lp_coeff[i] * (real)(1 << *shift));
155         }
156         return 0;
157 }
158
159 void FLAC__lpc_compute_residual_from_qlp_coefficients(const int32 data[], unsigned data_len, const int32 qlp_coeff[], unsigned order, int lp_quantization, int32 residual[])
160 {
161 #ifdef FLAC__OVERFLOW_DETECT
162         int64 sumo;
163 #endif
164         unsigned i, j;
165         int32 sum;
166         const int32 *history;
167
168 #ifdef FLAC__OVERFLOW_DETECT_VERBOSE
169         fprintf(stderr,"FLAC__lpc_compute_residual_from_qlp_coefficients: data_len=%d, order=%u, lpq=%d",data_len,order,lp_quantization);
170         for(i=0;i<order;i++)
171                 fprintf(stderr,", q[%u]=%d",i,qlp_coeff[i]);
172         fprintf(stderr,"\n");
173 #endif
174         assert(order > 0);
175
176         for(i = 0; i < data_len; i++) {
177 #ifdef FLAC__OVERFLOW_DETECT
178                 sumo = 0;
179 #endif
180                 sum = 0;
181                 history = data;
182                 for(j = 0; j < order; j++) {
183                         sum += qlp_coeff[j] * (*(--history));
184 #ifdef FLAC__OVERFLOW_DETECT
185                         sumo += (int64)qlp_coeff[j] * (int64)(*history);
186                         if(sumo > 2147483647ll || sumo < -2147483648ll) {
187                                 fprintf(stderr,"FLAC__lpc_compute_residual_from_qlp_coefficients: OVERFLOW, i=%u, j=%u, c=%d, d=%d, sumo=%lld\n",i,j,qlp_coeff[j],*history,sumo);
188                         }
189 #endif
190                 }
191                 *(residual++) = *(data++) - (sum >> lp_quantization);
192         }
193
194         /* Here's a slower but clearer version:
195         for(i = 0; i < data_len; i++) {
196                 sum = 0;
197                 for(j = 0; j < order; j++)
198                         sum += qlp_coeff[j] * data[i-j-1];
199                 residual[i] = data[i] - (sum >> lp_quantization);
200         }
201         */
202 }
203
204 void FLAC__lpc_restore_signal(const int32 residual[], unsigned data_len, const int32 qlp_coeff[], unsigned order, int lp_quantization, int32 data[])
205 {
206 #ifdef FLAC__OVERFLOW_DETECT
207         int64 sumo;
208 #endif
209         unsigned i, j;
210         int32 sum;
211         const int32 *history;
212
213 #ifdef FLAC__OVERFLOW_DETECT_VERBOSE
214         fprintf(stderr,"FLAC__lpc_restore_signal: data_len=%d, order=%u, lpq=%d",data_len,order,lp_quantization);
215         for(i=0;i<order;i++)
216                 fprintf(stderr,", q[%u]=%d",i,qlp_coeff[i]);
217         fprintf(stderr,"\n");
218 #endif
219         assert(order > 0);
220
221         for(i = 0; i < data_len; i++) {
222 #ifdef FLAC__OVERFLOW_DETECT
223                 sumo = 0;
224 #endif
225                 sum = 0;
226                 history = data;
227                 for(j = 0; j < order; j++) {
228                         sum += qlp_coeff[j] * (*(--history));
229 #ifdef FLAC__OVERFLOW_DETECT
230                         sumo += (int64)qlp_coeff[j] * (int64)(*history);
231                         if(sumo > 2147483647ll || sumo < -2147483648ll) {
232                                 fprintf(stderr,"FLAC__lpc_restore_signal: OVERFLOW, i=%u, j=%u, c=%d, d=%d, sumo=%lld\n",i,j,qlp_coeff[j],*history,sumo);
233                         }
234 #endif
235                 }
236                 *(data++) = *(residual++) + (sum >> lp_quantization);
237         }
238
239         /* Here's a slower but clearer version:
240         for(i = 0; i < data_len; i++) {
241                 sum = 0;
242                 for(j = 0; j < order; j++)
243                         sum += qlp_coeff[j] * data[i-j-1];
244                 data[i] = residual[i] + (sum >> lp_quantization);
245         }
246         */
247 }
248
249 real FLAC__lpc_compute_expected_bits_per_residual_sample(real lpc_error, unsigned total_samples)
250 {
251         real escale;
252
253         assert(lpc_error >= 0.0); /* the error can never be negative */
254         assert(total_samples > 0);
255
256         escale = 0.5 * M_LN2 * M_LN2 / (real)total_samples;
257
258         if(lpc_error > 0.0) {
259                 real bps = 0.5 * log(escale * lpc_error) / M_LN2;
260                 if(bps >= 0.0)
261                         return bps;
262                 else
263                         return 0.0;
264         }
265         else {
266                 return 0.0;
267         }
268 }
269
270 unsigned FLAC__lpc_compute_best_order(const real lpc_error[], unsigned max_order, unsigned total_samples, unsigned bits_per_signal_sample)
271 {
272         unsigned order, best_order;
273         real best_bits, tmp_bits;
274
275         assert(max_order > 0);
276
277         best_order = 0;
278         best_bits = FLAC__lpc_compute_expected_bits_per_residual_sample(lpc_error[0], total_samples) * (real)total_samples;
279
280         for(order = 1; order < max_order; order++) {
281                 tmp_bits = FLAC__lpc_compute_expected_bits_per_residual_sample(lpc_error[order], total_samples) * (real)(total_samples - order) + (real)(order * bits_per_signal_sample);
282                 if(tmp_bits < best_bits) {
283                         best_order = order;
284                         best_bits = tmp_bits;
285                 }
286         }
287
288         return best_order+1; /* +1 since index of lpc_error[] is order-1 */
289 }