1 /* libFLAC - Free Lossless Audio Codec library
2 * Copyright (C) 2000,2001,2002,2003,2004,2005,2006 Josh Coalson
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
8 * - Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
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.
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.
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.
37 #include "private/bitmath.h"
38 #include "private/fixed.h"
39 #include "FLAC/assert.h"
42 /* math.h in VC++ doesn't seem to have this (how Microsoft is that?) */
43 #define M_LN2 0.69314718055994530942
49 #define min(x,y) ((x) < (y)? (x) : (y))
54 #define local_abs(x) ((unsigned)((x)<0? -(x) : (x)))
56 #ifdef FLAC__INTEGER_ONLY_LIBRARY
57 /* rbps stands for residual bits per sample
60 * rbps = log (-----------)
63 static FLAC__fixedpoint local__compute_rbps_integerized(FLAC__uint32 err, FLAC__uint32 n)
66 unsigned bits; /* the number of bits required to represent a number */
67 int fracbits; /* the number of bits of rbps that comprise the fractional part */
69 FLAC__ASSERT(sizeof(rbps) == sizeof(FLAC__fixedpoint));
70 FLAC__ASSERT(err > 0);
73 FLAC__ASSERT(n <= FLAC__MAX_BLOCK_SIZE);
77 * The above two things tell us 1) n fits in 16 bits; 2) err/n > 1.
78 * These allow us later to know we won't lose too much precision in the
79 * fixed-point division (err<<fracbits)/n.
82 fracbits = (8*sizeof(err)) - (FLAC__bitmath_ilog2(err)+1);
86 /* err now holds err/n with fracbits fractional bits */
89 * Whittle err down to 16 bits max. 16 significant bits is enough for
92 FLAC__ASSERT(err > 0);
93 bits = FLAC__bitmath_ilog2(err)+1;
96 fracbits -= (bits-16);
98 rbps = (FLAC__uint32)err;
100 /* Multiply by fixed-point version of ln(2), with 16 fractional bits */
101 rbps *= FLAC__FP_LN2;
103 FLAC__ASSERT(fracbits >= 0);
105 /* FLAC__fixedpoint_log2 requires fracbits%4 to be 0 */
107 const int f = fracbits & 3;
114 rbps = FLAC__fixedpoint_log2(rbps, fracbits, (unsigned)(-1));
120 * The return value must have 16 fractional bits. Since the whole part
121 * of the base-2 log of a 32 bit number must fit in 5 bits, and fracbits
122 * must be >= -3, these assertion allows us to be able to shift rbps
123 * left if necessary to get 16 fracbits without losing any bits of the
124 * whole part of rbps.
126 * There is a slight chance due to accumulated error that the whole part
127 * will require 6 bits, so we use 6 in the assertion. Really though as
128 * long as it fits in 13 bits (32 - (16 - (-3))) we are fine.
130 FLAC__ASSERT((int)FLAC__bitmath_ilog2(rbps)+1 <= fracbits + 6);
131 FLAC__ASSERT(fracbits >= -3);
133 /* now shift the decimal point into place */
135 return rbps << (16-fracbits);
136 else if(fracbits > 16)
137 return rbps >> (fracbits-16);
142 static FLAC__fixedpoint local__compute_rbps_wide_integerized(FLAC__uint64 err, FLAC__uint32 n)
145 unsigned bits; /* the number of bits required to represent a number */
146 int fracbits; /* the number of bits of rbps that comprise the fractional part */
148 FLAC__ASSERT(sizeof(rbps) == sizeof(FLAC__fixedpoint));
149 FLAC__ASSERT(err > 0);
152 FLAC__ASSERT(n <= FLAC__MAX_BLOCK_SIZE);
156 * The above two things tell us 1) n fits in 16 bits; 2) err/n > 1.
157 * These allow us later to know we won't lose too much precision in the
158 * fixed-point division (err<<fracbits)/n.
161 fracbits = (8*sizeof(err)) - (FLAC__bitmath_ilog2_wide(err)+1);
165 /* err now holds err/n with fracbits fractional bits */
168 * Whittle err down to 16 bits max. 16 significant bits is enough for
171 FLAC__ASSERT(err > 0);
172 bits = FLAC__bitmath_ilog2_wide(err)+1;
175 fracbits -= (bits-16);
177 rbps = (FLAC__uint32)err;
179 /* Multiply by fixed-point version of ln(2), with 16 fractional bits */
180 rbps *= FLAC__FP_LN2;
182 FLAC__ASSERT(fracbits >= 0);
184 /* FLAC__fixedpoint_log2 requires fracbits%4 to be 0 */
186 const int f = fracbits & 3;
193 rbps = FLAC__fixedpoint_log2(rbps, fracbits, (unsigned)(-1));
199 * The return value must have 16 fractional bits. Since the whole part
200 * of the base-2 log of a 32 bit number must fit in 5 bits, and fracbits
201 * must be >= -3, these assertion allows us to be able to shift rbps
202 * left if necessary to get 16 fracbits without losing any bits of the
203 * whole part of rbps.
205 * There is a slight chance due to accumulated error that the whole part
206 * will require 6 bits, so we use 6 in the assertion. Really though as
207 * long as it fits in 13 bits (32 - (16 - (-3))) we are fine.
209 FLAC__ASSERT((int)FLAC__bitmath_ilog2(rbps)+1 <= fracbits + 6);
210 FLAC__ASSERT(fracbits >= -3);
212 /* now shift the decimal point into place */
214 return rbps << (16-fracbits);
215 else if(fracbits > 16)
216 return rbps >> (fracbits-16);
222 #ifndef FLAC__INTEGER_ONLY_LIBRARY
223 unsigned FLAC__fixed_compute_best_predictor(const FLAC__int32 data[], unsigned data_len, FLAC__float residual_bits_per_sample[FLAC__MAX_FIXED_ORDER+1])
225 unsigned FLAC__fixed_compute_best_predictor(const FLAC__int32 data[], unsigned data_len, FLAC__fixedpoint residual_bits_per_sample[FLAC__MAX_FIXED_ORDER+1])
228 FLAC__int32 last_error_0 = data[-1];
229 FLAC__int32 last_error_1 = data[-1] - data[-2];
230 FLAC__int32 last_error_2 = last_error_1 - (data[-2] - data[-3]);
231 FLAC__int32 last_error_3 = last_error_2 - (data[-2] - 2*data[-3] + data[-4]);
232 FLAC__int32 error, save;
233 FLAC__uint32 total_error_0 = 0, total_error_1 = 0, total_error_2 = 0, total_error_3 = 0, total_error_4 = 0;
236 for(i = 0; i < data_len; i++) {
237 error = data[i] ; total_error_0 += local_abs(error); save = error;
238 error -= last_error_0; total_error_1 += local_abs(error); last_error_0 = save; save = error;
239 error -= last_error_1; total_error_2 += local_abs(error); last_error_1 = save; save = error;
240 error -= last_error_2; total_error_3 += local_abs(error); last_error_2 = save; save = error;
241 error -= last_error_3; total_error_4 += local_abs(error); last_error_3 = save;
244 if(total_error_0 < min(min(min(total_error_1, total_error_2), total_error_3), total_error_4))
246 else if(total_error_1 < min(min(total_error_2, total_error_3), total_error_4))
248 else if(total_error_2 < min(total_error_3, total_error_4))
250 else if(total_error_3 < total_error_4)
255 /* Estimate the expected number of bits per residual signal sample. */
256 /* 'total_error*' is linearly related to the variance of the residual */
257 /* signal, so we use it directly to compute E(|x|) */
258 FLAC__ASSERT(data_len > 0 || total_error_0 == 0);
259 FLAC__ASSERT(data_len > 0 || total_error_1 == 0);
260 FLAC__ASSERT(data_len > 0 || total_error_2 == 0);
261 FLAC__ASSERT(data_len > 0 || total_error_3 == 0);
262 FLAC__ASSERT(data_len > 0 || total_error_4 == 0);
263 #ifndef FLAC__INTEGER_ONLY_LIBRARY
264 residual_bits_per_sample[0] = (FLAC__float)((total_error_0 > 0) ? log(M_LN2 * (FLAC__double)total_error_0 / (FLAC__double)data_len) / M_LN2 : 0.0);
265 residual_bits_per_sample[1] = (FLAC__float)((total_error_1 > 0) ? log(M_LN2 * (FLAC__double)total_error_1 / (FLAC__double)data_len) / M_LN2 : 0.0);
266 residual_bits_per_sample[2] = (FLAC__float)((total_error_2 > 0) ? log(M_LN2 * (FLAC__double)total_error_2 / (FLAC__double)data_len) / M_LN2 : 0.0);
267 residual_bits_per_sample[3] = (FLAC__float)((total_error_3 > 0) ? log(M_LN2 * (FLAC__double)total_error_3 / (FLAC__double)data_len) / M_LN2 : 0.0);
268 residual_bits_per_sample[4] = (FLAC__float)((total_error_4 > 0) ? log(M_LN2 * (FLAC__double)total_error_4 / (FLAC__double)data_len) / M_LN2 : 0.0);
270 residual_bits_per_sample[0] = (total_error_0 > 0) ? local__compute_rbps_integerized(total_error_0, data_len) : 0;
271 residual_bits_per_sample[1] = (total_error_1 > 0) ? local__compute_rbps_integerized(total_error_1, data_len) : 0;
272 residual_bits_per_sample[2] = (total_error_2 > 0) ? local__compute_rbps_integerized(total_error_2, data_len) : 0;
273 residual_bits_per_sample[3] = (total_error_3 > 0) ? local__compute_rbps_integerized(total_error_3, data_len) : 0;
274 residual_bits_per_sample[4] = (total_error_4 > 0) ? local__compute_rbps_integerized(total_error_4, data_len) : 0;
280 #ifndef FLAC__INTEGER_ONLY_LIBRARY
281 unsigned FLAC__fixed_compute_best_predictor_wide(const FLAC__int32 data[], unsigned data_len, FLAC__float residual_bits_per_sample[FLAC__MAX_FIXED_ORDER+1])
283 unsigned FLAC__fixed_compute_best_predictor_wide(const FLAC__int32 data[], unsigned data_len, FLAC__fixedpoint residual_bits_per_sample[FLAC__MAX_FIXED_ORDER+1])
286 FLAC__int32 last_error_0 = data[-1];
287 FLAC__int32 last_error_1 = data[-1] - data[-2];
288 FLAC__int32 last_error_2 = last_error_1 - (data[-2] - data[-3]);
289 FLAC__int32 last_error_3 = last_error_2 - (data[-2] - 2*data[-3] + data[-4]);
290 FLAC__int32 error, save;
291 /* total_error_* are 64-bits to avoid overflow when encoding
292 * erratic signals when the bits-per-sample and blocksize are
295 FLAC__uint64 total_error_0 = 0, total_error_1 = 0, total_error_2 = 0, total_error_3 = 0, total_error_4 = 0;
298 for(i = 0; i < data_len; i++) {
299 error = data[i] ; total_error_0 += local_abs(error); save = error;
300 error -= last_error_0; total_error_1 += local_abs(error); last_error_0 = save; save = error;
301 error -= last_error_1; total_error_2 += local_abs(error); last_error_1 = save; save = error;
302 error -= last_error_2; total_error_3 += local_abs(error); last_error_2 = save; save = error;
303 error -= last_error_3; total_error_4 += local_abs(error); last_error_3 = save;
306 if(total_error_0 < min(min(min(total_error_1, total_error_2), total_error_3), total_error_4))
308 else if(total_error_1 < min(min(total_error_2, total_error_3), total_error_4))
310 else if(total_error_2 < min(total_error_3, total_error_4))
312 else if(total_error_3 < total_error_4)
317 /* Estimate the expected number of bits per residual signal sample. */
318 /* 'total_error*' is linearly related to the variance of the residual */
319 /* signal, so we use it directly to compute E(|x|) */
320 FLAC__ASSERT(data_len > 0 || total_error_0 == 0);
321 FLAC__ASSERT(data_len > 0 || total_error_1 == 0);
322 FLAC__ASSERT(data_len > 0 || total_error_2 == 0);
323 FLAC__ASSERT(data_len > 0 || total_error_3 == 0);
324 FLAC__ASSERT(data_len > 0 || total_error_4 == 0);
325 #ifndef FLAC__INTEGER_ONLY_LIBRARY
326 #if defined _MSC_VER || defined __MINGW32__
327 /* with MSVC you have to spoon feed it the casting */
328 residual_bits_per_sample[0] = (FLAC__float)((total_error_0 > 0) ? log(M_LN2 * (FLAC__double)(FLAC__int64)total_error_0 / (FLAC__double)data_len) / M_LN2 : 0.0);
329 residual_bits_per_sample[1] = (FLAC__float)((total_error_1 > 0) ? log(M_LN2 * (FLAC__double)(FLAC__int64)total_error_1 / (FLAC__double)data_len) / M_LN2 : 0.0);
330 residual_bits_per_sample[2] = (FLAC__float)((total_error_2 > 0) ? log(M_LN2 * (FLAC__double)(FLAC__int64)total_error_2 / (FLAC__double)data_len) / M_LN2 : 0.0);
331 residual_bits_per_sample[3] = (FLAC__float)((total_error_3 > 0) ? log(M_LN2 * (FLAC__double)(FLAC__int64)total_error_3 / (FLAC__double)data_len) / M_LN2 : 0.0);
332 residual_bits_per_sample[4] = (FLAC__float)((total_error_4 > 0) ? log(M_LN2 * (FLAC__double)(FLAC__int64)total_error_4 / (FLAC__double)data_len) / M_LN2 : 0.0);
334 residual_bits_per_sample[0] = (FLAC__float)((total_error_0 > 0) ? log(M_LN2 * (FLAC__double)total_error_0 / (FLAC__double)data_len) / M_LN2 : 0.0);
335 residual_bits_per_sample[1] = (FLAC__float)((total_error_1 > 0) ? log(M_LN2 * (FLAC__double)total_error_1 / (FLAC__double)data_len) / M_LN2 : 0.0);
336 residual_bits_per_sample[2] = (FLAC__float)((total_error_2 > 0) ? log(M_LN2 * (FLAC__double)total_error_2 / (FLAC__double)data_len) / M_LN2 : 0.0);
337 residual_bits_per_sample[3] = (FLAC__float)((total_error_3 > 0) ? log(M_LN2 * (FLAC__double)total_error_3 / (FLAC__double)data_len) / M_LN2 : 0.0);
338 residual_bits_per_sample[4] = (FLAC__float)((total_error_4 > 0) ? log(M_LN2 * (FLAC__double)total_error_4 / (FLAC__double)data_len) / M_LN2 : 0.0);
341 residual_bits_per_sample[0] = (total_error_0 > 0) ? local__compute_rbps_wide_integerized(total_error_0, data_len) : 0;
342 residual_bits_per_sample[1] = (total_error_1 > 0) ? local__compute_rbps_wide_integerized(total_error_1, data_len) : 0;
343 residual_bits_per_sample[2] = (total_error_2 > 0) ? local__compute_rbps_wide_integerized(total_error_2, data_len) : 0;
344 residual_bits_per_sample[3] = (total_error_3 > 0) ? local__compute_rbps_wide_integerized(total_error_3, data_len) : 0;
345 residual_bits_per_sample[4] = (total_error_4 > 0) ? local__compute_rbps_wide_integerized(total_error_4, data_len) : 0;
351 void FLAC__fixed_compute_residual(const FLAC__int32 data[], unsigned data_len, unsigned order, FLAC__int32 residual[])
353 const int idata_len = (int)data_len;
358 for(i = 0; i < idata_len; i++) {
359 residual[i] = data[i];
363 for(i = 0; i < idata_len; i++) {
364 residual[i] = data[i] - data[i-1];
368 for(i = 0; i < idata_len; i++) {
369 /* == data[i] - 2*data[i-1] + data[i-2] */
370 residual[i] = data[i] - (data[i-1] << 1) + data[i-2];
374 for(i = 0; i < idata_len; i++) {
375 /* == data[i] - 3*data[i-1] + 3*data[i-2] - data[i-3] */
376 residual[i] = data[i] - (((data[i-1]-data[i-2])<<1) + (data[i-1]-data[i-2])) - data[i-3];
380 for(i = 0; i < idata_len; i++) {
381 /* == data[i] - 4*data[i-1] + 6*data[i-2] - 4*data[i-3] + data[i-4] */
382 residual[i] = data[i] - ((data[i-1]+data[i-3])<<2) + ((data[i-2]<<2) + (data[i-2]<<1)) + data[i-4];
390 void FLAC__fixed_restore_signal(const FLAC__int32 residual[], unsigned data_len, unsigned order, FLAC__int32 data[])
392 int i, idata_len = (int)data_len;
396 for(i = 0; i < idata_len; i++) {
397 data[i] = residual[i];
401 for(i = 0; i < idata_len; i++) {
402 data[i] = residual[i] + data[i-1];
406 for(i = 0; i < idata_len; i++) {
407 /* == residual[i] + 2*data[i-1] - data[i-2] */
408 data[i] = residual[i] + (data[i-1]<<1) - data[i-2];
412 for(i = 0; i < idata_len; i++) {
413 /* residual[i] + 3*data[i-1] - 3*data[i-2]) + data[i-3] */
414 data[i] = residual[i] + (((data[i-1]-data[i-2])<<1) + (data[i-1]-data[i-2])) + data[i-3];
418 for(i = 0; i < idata_len; i++) {
419 /* == residual[i] + 4*data[i-1] - 6*data[i-2] + 4*data[i-3] - data[i-4] */
420 data[i] = residual[i] + ((data[i-1]+data[i-3])<<2) - ((data[i-2]<<2) + (data[i-2]<<1)) - data[i-4];