add workaround for lack of lround() on MSVC
[platform/upstream/flac.git] / src / libFLAC / lpc.c
1 /* libFLAC - Free Lossless Audio Codec library
2  * Copyright (C) 2000,2001,2002,2003,2004,2005,2006,2007  Josh Coalson
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  *
8  * - Redistributions of source code must retain the above copyright
9  * notice, this list of conditions and the following disclaimer.
10  *
11  * - Redistributions in binary form must reproduce the above copyright
12  * notice, this list of conditions and the following disclaimer in the
13  * documentation and/or other materials provided with the distribution.
14  *
15  * - Neither the name of the Xiph.org Foundation nor the names of its
16  * contributors may be used to endorse or promote products derived from
17  * this software without specific prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
21  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
22  * A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR
23  * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
24  * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
25  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
26  * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
27  * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
28  * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
29  * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
30  */
31
32 #if HAVE_CONFIG_H
33 #  include <config.h>
34 #endif
35
36 #include <math.h>
37 #include "FLAC/assert.h"
38 #include "FLAC/format.h"
39 #include "private/bitmath.h"
40 #include "private/lpc.h"
41 #if defined DEBUG || defined FLAC__OVERFLOW_DETECT || defined FLAC__OVERFLOW_DETECT_VERBOSE
42 #include <stdio.h>
43 #endif
44
45 #ifndef FLAC__INTEGER_ONLY_LIBRARY
46
47 #ifndef M_LN2
48 /* math.h in VC++ doesn't seem to have this (how Microsoft is that?) */
49 #define M_LN2 0.69314718055994530942
50 #endif
51
52 void FLAC__lpc_window_data(const FLAC__real in[], const FLAC__real window[], FLAC__real out[], unsigned data_len)
53 {
54         unsigned i;
55         for(i = 0; i < data_len; i++)
56                 out[i] = in[i] * window[i];
57 }
58
59 void FLAC__lpc_compute_autocorrelation(const FLAC__real data[], unsigned data_len, unsigned lag, FLAC__real autoc[])
60 {
61         /* a readable, but slower, version */
62 #if 0
63         FLAC__real d;
64         unsigned i;
65
66         FLAC__ASSERT(lag > 0);
67         FLAC__ASSERT(lag <= data_len);
68
69         /*
70          * Technically we should subtract the mean first like so:
71          *   for(i = 0; i < data_len; i++)
72          *     data[i] -= mean;
73          * but it appears not to make enough of a difference to matter, and
74          * most signals are already closely centered around zero
75          */
76         while(lag--) {
77                 for(i = lag, d = 0.0; i < data_len; i++)
78                         d += data[i] * data[i - lag];
79                 autoc[lag] = d;
80         }
81 #endif
82
83         /*
84          * this version tends to run faster because of better data locality
85          * ('data_len' is usually much larger than 'lag')
86          */
87         FLAC__real d;
88         unsigned sample, coeff;
89         const unsigned limit = data_len - lag;
90
91         FLAC__ASSERT(lag > 0);
92         FLAC__ASSERT(lag <= data_len);
93
94         for(coeff = 0; coeff < lag; coeff++)
95                 autoc[coeff] = 0.0;
96         for(sample = 0; sample <= limit; sample++) {
97                 d = data[sample];
98                 for(coeff = 0; coeff < lag; coeff++)
99                         autoc[coeff] += d * data[sample+coeff];
100         }
101         for(; sample < data_len; sample++) {
102                 d = data[sample];
103                 for(coeff = 0; coeff < data_len - sample; coeff++)
104                         autoc[coeff] += d * data[sample+coeff];
105         }
106 }
107
108 void FLAC__lpc_compute_lp_coefficients(const FLAC__real autoc[], unsigned *max_order, FLAC__real lp_coeff[][FLAC__MAX_LPC_ORDER], FLAC__double error[])
109 {
110         unsigned i, j;
111         FLAC__double r, err, ref[FLAC__MAX_LPC_ORDER], lpc[FLAC__MAX_LPC_ORDER];
112
113         FLAC__ASSERT(0 != max_order);
114         FLAC__ASSERT(0 < *max_order);
115         FLAC__ASSERT(*max_order <= FLAC__MAX_LPC_ORDER);
116         FLAC__ASSERT(autoc[0] != 0.0);
117
118         err = autoc[0];
119
120         for(i = 0; i < *max_order; i++) {
121                 /* Sum up this iteration's reflection coefficient. */
122                 r = -autoc[i+1];
123                 for(j = 0; j < i; j++)
124                         r -= lpc[j] * autoc[i-j];
125                 ref[i] = (r/=err);
126
127                 /* Update LPC coefficients and total error. */
128                 lpc[i]=r;
129                 for(j = 0; j < (i>>1); j++) {
130                         FLAC__double tmp = lpc[j];
131                         lpc[j] += r * lpc[i-1-j];
132                         lpc[i-1-j] += r * tmp;
133                 }
134                 if(i & 1)
135                         lpc[j] += lpc[j] * r;
136
137                 err *= (1.0 - r * r);
138
139                 /* save this order */
140                 for(j = 0; j <= i; j++)
141                         lp_coeff[i][j] = (FLAC__real)(-lpc[j]); /* negate FIR filter coeff to get predictor coeff */
142                 error[i] = err;
143
144                 /* see SF bug #1601812 http://sourceforge.net/tracker/index.php?func=detail&aid=1601812&group_id=13478&atid=113478 */
145                 if(err == 0.0) {
146                         *max_order = i+1;
147                         return;
148                 }
149         }
150 }
151
152 int FLAC__lpc_quantize_coefficients(const FLAC__real lp_coeff[], unsigned order, unsigned precision, FLAC__int32 qlp_coeff[], int *shift)
153 {
154         unsigned i;
155         FLAC__double cmax;
156         FLAC__int32 qmax, qmin;
157
158         FLAC__ASSERT(precision > 0);
159         FLAC__ASSERT(precision >= FLAC__MIN_QLP_COEFF_PRECISION);
160
161         /* drop one bit for the sign; from here on out we consider only |lp_coeff[i]| */
162         precision--;
163         qmax = 1 << precision;
164         qmin = -qmax;
165         qmax--;
166
167         /* calc cmax = max( |lp_coeff[i]| ) */
168         cmax = 0.0;
169         for(i = 0; i < order; i++) {
170                 const FLAC__double d = fabs(lp_coeff[i]);
171                 if(d > cmax)
172                         cmax = d;
173         }
174
175         if(cmax <= 0.0) {
176                 /* => coefficients are all 0, which means our constant-detect didn't work */
177                 return 2;
178         }
179         else {
180                 const int max_shiftlimit = (1 << (FLAC__SUBFRAME_LPC_QLP_SHIFT_LEN-1)) - 1;
181                 const int min_shiftlimit = -max_shiftlimit - 1;
182                 int log2cmax;
183
184                 (void)frexp(cmax, &log2cmax);
185                 log2cmax--;
186                 *shift = (int)precision - log2cmax - 1;
187
188                 if(*shift < min_shiftlimit || *shift > max_shiftlimit) {
189 #ifdef FLAC__OVERFLOW_DETECT
190                         fprintf(stderr,"FLAC__lpc_quantize_coefficients: shift out of limit: shift=%d cmax=%f precision=%u\n",*shift,cmax,precision+1);
191 #endif
192 #if 0
193                         /*@@@ this does not seem to help at all, but was not extensively tested either: */
194                         if(*shift > max_shiftlimit)
195                                 *shift = max_shiftlimit;
196                         else
197 #endif
198                                 return 1;
199                 }
200         }
201
202         if(*shift >= 0) {
203                 FLAC__double error = 0.0;
204                 FLAC__int32 q;
205                 for(i = 0; i < order; i++) {
206                         error += lp_coeff[i] * (1 << *shift);
207 #if defined _MSC_VER
208                         if(error >= 0.0)
209                                 q = (FLAC__int32)(error + 0.5);
210                         else
211                                 q = (FLAC__int32)(error - 0.5);
212 #else
213                         q = lround(error); /* round() is also suitable */
214 #endif
215 #ifdef FLAC__OVERFLOW_DETECT
216                         if(q > qmax+1) /* we expect q==qmax+1 occasionally due to lround() */
217                                 fprintf(stderr,"FLAC__lpc_quantize_coefficients: quantizer overflow: q>qmax %d>%d shift=%d cmax=%f precision=%u lpc[%u]=%f\n",q,qmax,*shift,cmax,precision+1,i,lp_coeff[i]);
218                         else if(q < qmin)
219                                 fprintf(stderr,"FLAC__lpc_quantize_coefficients: quantizer overflow: q<qmin %d<%d shift=%d cmax=%f precision=%u lpc[%u]=%f\n",q,qmin,*shift,cmax,precision+1,i,lp_coeff[i]);
220 #endif
221                         if(q > qmax)
222                                 q = qmax;
223                         else if(q < qmin)
224                                 q = qmin;
225                         error -= q;
226                         qlp_coeff[i] = q;
227                 }
228         }
229         /* negative shift is very rare but due to design flaw, negative shift is
230          * a NOP in the decoder, so it must be handled specially by scaling down
231          * coeffs
232          */
233         else {
234                 const int nshift = -(*shift);
235                 FLAC__double error = 0.0;
236                 FLAC__int32 q;
237 #ifdef DEBUG
238                 fprintf(stderr,"FLAC__lpc_quantize_coefficients: negative shift=%d order=%u cmax=%f\n", *shift, order, cmax);
239 #endif
240                 for(i = 0; i < order; i++) {
241                         error += lp_coeff[i] / (1 << nshift);
242 #if defined _MSC_VER
243                         if(error >= 0.0)
244                                 q = (FLAC__int32)(error + 0.5);
245                         else
246                                 q = (FLAC__int32)(error - 0.5);
247 #else
248                         q = lround(error); /* round() is also suitable */
249 #endif
250 #ifdef FLAC__OVERFLOW_DETECT
251                         if(q > qmax+1) /* we expect q==qmax+1 occasionally due to lround() */
252                                 fprintf(stderr,"FLAC__lpc_quantize_coefficients: quantizer overflow: q>qmax %d>%d shift=%d cmax=%f precision=%u lpc[%u]=%f\n",q,qmax,*shift,cmax,precision+1,i,lp_coeff[i]);
253                         else if(q < qmin)
254                                 fprintf(stderr,"FLAC__lpc_quantize_coefficients: quantizer overflow: q<qmin %d<%d shift=%d cmax=%f precision=%u lpc[%u]=%f\n",q,qmin,*shift,cmax,precision+1,i,lp_coeff[i]);
255 #endif
256                         if(q > qmax)
257                                 q = qmax;
258                         else if(q < qmin)
259                                 q = qmin;
260                         error -= q;
261                         qlp_coeff[i] = q;
262                 }
263                 *shift = 0;
264         }
265
266         return 0;
267 }
268
269 void FLAC__lpc_compute_residual_from_qlp_coefficients(const FLAC__int32 *data, unsigned data_len, const FLAC__int32 qlp_coeff[], unsigned order, int lp_quantization, FLAC__int32 residual[])
270 {
271 #ifdef FLAC__OVERFLOW_DETECT
272         FLAC__int64 sumo;
273 #endif
274         unsigned i, j;
275         FLAC__int32 sum;
276         const FLAC__int32 *history;
277
278 #ifdef FLAC__OVERFLOW_DETECT_VERBOSE
279         fprintf(stderr,"FLAC__lpc_compute_residual_from_qlp_coefficients: data_len=%d, order=%u, lpq=%d",data_len,order,lp_quantization);
280         for(i=0;i<order;i++)
281                 fprintf(stderr,", q[%u]=%d",i,qlp_coeff[i]);
282         fprintf(stderr,"\n");
283 #endif
284         FLAC__ASSERT(order > 0);
285
286         for(i = 0; i < data_len; i++) {
287 #ifdef FLAC__OVERFLOW_DETECT
288                 sumo = 0;
289 #endif
290                 sum = 0;
291                 history = data;
292                 for(j = 0; j < order; j++) {
293                         sum += qlp_coeff[j] * (*(--history));
294 #ifdef FLAC__OVERFLOW_DETECT
295                         sumo += (FLAC__int64)qlp_coeff[j] * (FLAC__int64)(*history);
296 #if defined _MSC_VER
297                         if(sumo > 2147483647I64 || sumo < -2147483648I64)
298                                 fprintf(stderr,"FLAC__lpc_compute_residual_from_qlp_coefficients: OVERFLOW, i=%u, j=%u, c=%d, d=%d, sumo=%I64d\n",i,j,qlp_coeff[j],*history,sumo);
299 #else
300                         if(sumo > 2147483647ll || sumo < -2147483648ll)
301                                 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,(long long)sumo);
302 #endif
303 #endif
304                 }
305                 *(residual++) = *(data++) - (sum >> lp_quantization);
306         }
307
308         /* Here's a slower but clearer version:
309         for(i = 0; i < data_len; i++) {
310                 sum = 0;
311                 for(j = 0; j < order; j++)
312                         sum += qlp_coeff[j] * data[i-j-1];
313                 residual[i] = data[i] - (sum >> lp_quantization);
314         }
315         */
316 }
317
318 void FLAC__lpc_compute_residual_from_qlp_coefficients_wide(const FLAC__int32 *data, unsigned data_len, const FLAC__int32 qlp_coeff[], unsigned order, int lp_quantization, FLAC__int32 residual[])
319 {
320         unsigned i, j;
321         FLAC__int64 sum;
322         const FLAC__int32 *history;
323
324 #ifdef FLAC__OVERFLOW_DETECT_VERBOSE
325         fprintf(stderr,"FLAC__lpc_compute_residual_from_qlp_coefficients_wide: data_len=%d, order=%u, lpq=%d",data_len,order,lp_quantization);
326         for(i=0;i<order;i++)
327                 fprintf(stderr,", q[%u]=%d",i,qlp_coeff[i]);
328         fprintf(stderr,"\n");
329 #endif
330         FLAC__ASSERT(order > 0);
331
332         for(i = 0; i < data_len; i++) {
333                 sum = 0;
334                 history = data;
335                 for(j = 0; j < order; j++)
336                         sum += (FLAC__int64)qlp_coeff[j] * (FLAC__int64)(*(--history));
337 #ifdef FLAC__OVERFLOW_DETECT
338                 if(FLAC__bitmath_silog2_wide(sum >> lp_quantization) > 32) {
339                         fprintf(stderr,"FLAC__lpc_compute_residual_from_qlp_coefficients_wide: OVERFLOW, i=%u, sum=%lld\n", i, (long long)(sum >> lp_quantization));
340                         break;
341                 }
342                 if(FLAC__bitmath_silog2_wide((FLAC__int64)(*data) - (sum >> lp_quantization)) > 32) {
343                         fprintf(stderr,"FLAC__lpc_compute_residual_from_qlp_coefficients_wide: OVERFLOW, i=%u, data=%d, sum=%lld, residual=%lld\n", i, *data, (long long)(sum >> lp_quantization), (long long)((FLAC__int64)(*data) - (sum >> lp_quantization)));
344                         break;
345                 }
346 #endif
347                 *(residual++) = *(data++) - (FLAC__int32)(sum >> lp_quantization);
348         }
349 }
350
351 #endif /* !defined FLAC__INTEGER_ONLY_LIBRARY */
352
353 void FLAC__lpc_restore_signal(const FLAC__int32 residual[], unsigned data_len, const FLAC__int32 qlp_coeff[], unsigned order, int lp_quantization, FLAC__int32 data[])
354 {
355 #ifdef FLAC__OVERFLOW_DETECT
356         FLAC__int64 sumo;
357 #endif
358         unsigned i, j;
359         FLAC__int32 sum;
360         const FLAC__int32 *r = residual, *history;
361
362 #ifdef FLAC__OVERFLOW_DETECT_VERBOSE
363         fprintf(stderr,"FLAC__lpc_restore_signal: data_len=%d, order=%u, lpq=%d",data_len,order,lp_quantization);
364         for(i=0;i<order;i++)
365                 fprintf(stderr,", q[%u]=%d",i,qlp_coeff[i]);
366         fprintf(stderr,"\n");
367 #endif
368         FLAC__ASSERT(order > 0);
369
370         for(i = 0; i < data_len; i++) {
371 #ifdef FLAC__OVERFLOW_DETECT
372                 sumo = 0;
373 #endif
374                 sum = 0;
375                 history = data;
376                 for(j = 0; j < order; j++) {
377                         sum += qlp_coeff[j] * (*(--history));
378 #ifdef FLAC__OVERFLOW_DETECT
379                         sumo += (FLAC__int64)qlp_coeff[j] * (FLAC__int64)(*history);
380 #if defined _MSC_VER
381                         if(sumo > 2147483647I64 || sumo < -2147483648I64)
382                                 fprintf(stderr,"FLAC__lpc_restore_signal: OVERFLOW, i=%u, j=%u, c=%d, d=%d, sumo=%I64d\n",i,j,qlp_coeff[j],*history,sumo);
383 #else
384                         if(sumo > 2147483647ll || sumo < -2147483648ll)
385                                 fprintf(stderr,"FLAC__lpc_restore_signal: OVERFLOW, i=%u, j=%u, c=%d, d=%d, sumo=%lld\n",i,j,qlp_coeff[j],*history,(long long)sumo);
386 #endif
387 #endif
388                 }
389                 *(data++) = *(r++) + (sum >> lp_quantization);
390         }
391
392         /* Here's a slower but clearer version:
393         for(i = 0; i < data_len; i++) {
394                 sum = 0;
395                 for(j = 0; j < order; j++)
396                         sum += qlp_coeff[j] * data[i-j-1];
397                 data[i] = residual[i] + (sum >> lp_quantization);
398         }
399         */
400 }
401
402 void FLAC__lpc_restore_signal_wide(const FLAC__int32 residual[], unsigned data_len, const FLAC__int32 qlp_coeff[], unsigned order, int lp_quantization, FLAC__int32 data[])
403 {
404         unsigned i, j;
405         FLAC__int64 sum;
406         const FLAC__int32 *r = residual, *history;
407
408 #ifdef FLAC__OVERFLOW_DETECT_VERBOSE
409         fprintf(stderr,"FLAC__lpc_restore_signal_wide: data_len=%d, order=%u, lpq=%d",data_len,order,lp_quantization);
410         for(i=0;i<order;i++)
411                 fprintf(stderr,", q[%u]=%d",i,qlp_coeff[i]);
412         fprintf(stderr,"\n");
413 #endif
414         FLAC__ASSERT(order > 0);
415
416         for(i = 0; i < data_len; i++) {
417                 sum = 0;
418                 history = data;
419                 for(j = 0; j < order; j++)
420                         sum += (FLAC__int64)qlp_coeff[j] * (FLAC__int64)(*(--history));
421 #ifdef FLAC__OVERFLOW_DETECT
422                 if(FLAC__bitmath_silog2_wide(sum >> lp_quantization) > 32) {
423                         fprintf(stderr,"FLAC__lpc_restore_signal_wide: OVERFLOW, i=%u, sum=%lld\n", i, (long long)(sum >> lp_quantization));
424                         break;
425                 }
426                 if(FLAC__bitmath_silog2_wide((FLAC__int64)(*r) + (sum >> lp_quantization)) > 32) {
427                         fprintf(stderr,"FLAC__lpc_restore_signal_wide: OVERFLOW, i=%u, residual=%d, sum=%lld, data=%lld\n", i, *r, (long long)(sum >> lp_quantization), (long long)((FLAC__int64)(*r) + (sum >> lp_quantization)));
428                         break;
429                 }
430 #endif
431                 *(data++) = *(r++) + (FLAC__int32)(sum >> lp_quantization);
432         }
433 }
434
435 #ifndef FLAC__INTEGER_ONLY_LIBRARY
436
437 FLAC__double FLAC__lpc_compute_expected_bits_per_residual_sample(FLAC__double lpc_error, unsigned total_samples)
438 {
439         FLAC__double error_scale;
440
441         FLAC__ASSERT(total_samples > 0);
442
443         error_scale = 0.5 * M_LN2 * M_LN2 / (FLAC__double)total_samples;
444
445         return FLAC__lpc_compute_expected_bits_per_residual_sample_with_error_scale(lpc_error, error_scale);
446 }
447
448 FLAC__double FLAC__lpc_compute_expected_bits_per_residual_sample_with_error_scale(FLAC__double lpc_error, FLAC__double error_scale)
449 {
450         if(lpc_error > 0.0) {
451                 FLAC__double bps = (FLAC__double)0.5 * log(error_scale * lpc_error) / M_LN2;
452                 if(bps >= 0.0)
453                         return bps;
454                 else
455                         return 0.0;
456         }
457         else if(lpc_error < 0.0) { /* error should not be negative but can happen due to inadequate floating-point resolution */
458                 return 1e32;
459         }
460         else {
461                 return 0.0;
462         }
463 }
464
465 unsigned FLAC__lpc_compute_best_order(const FLAC__double lpc_error[], unsigned max_order, unsigned total_samples, unsigned overhead_bits_per_order)
466 {
467         unsigned order, index, best_index; /* 'index' the index into lpc_error; index==order-1 since lpc_error[0] is for order==1, lpc_error[1] is for order==2, etc */
468         FLAC__double bits, best_bits, error_scale;
469
470         FLAC__ASSERT(max_order > 0);
471         FLAC__ASSERT(total_samples > 0);
472
473         error_scale = 0.5 * M_LN2 * M_LN2 / (FLAC__double)total_samples;
474
475         best_index = 0;
476         best_bits = (unsigned)(-1);
477
478         for(index = 0, order = 1; index < max_order; index++, order++) {
479                 bits = FLAC__lpc_compute_expected_bits_per_residual_sample_with_error_scale(lpc_error[index], error_scale) * (FLAC__double)(total_samples - order) + (FLAC__double)(order * overhead_bits_per_order);
480                 if(bits < best_bits) {
481                         best_index = index;
482                         best_bits = bits;
483                 }
484         }
485
486         return best_index+1; /* +1 since index of lpc_error[] is order-1 */
487 }
488
489 #endif /* !defined FLAC__INTEGER_ONLY_LIBRARY */