src/libFLAC/lpc.c

   1 /* libFLAC - Free Lossless Audio Codec library
   2  * Copyright (C) 2000,2001,2002,2003,2004,2005,2006,2007,2008,2009  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 <inttypes.h>
  38 #include "FLAC/assert.h"
  39 #include "FLAC/format.h"
  40 #include "private/bitmath.h"
  41 #include "private/lpc.h"
  42 #include "private/macros.h"
  43 #if defined DEBUG || defined FLAC__OVERFLOW_DETECT || defined FLAC__OVERFLOW_DETECT_VERBOSE
  44 #include <stdio.h>
  45 #endif
  46
  47 /* OPT: #undef'ing this may improve the speed on some architectures */
  48 #define FLAC__LPC_UNROLLED_FILTER_LOOPS
  49
  50 #ifndef FLAC__INTEGER_ONLY_LIBRARY
  51
  52 #ifndef M_LN2
  53 /* math.h in VC++ doesn't seem to have this (how Microsoft is that?) */
  54 #define M_LN2 0.69314718055994530942
  55 #endif
  56
  57 #if !defined(HAVE_LROUND)
  58 #if defined(_MSC_VER)
  59 #include <float.h>
  60 #define copysign _copysign
  61 #elif defined(__GNUC__)
  62 #define copysign __builtin_copysign
  63 #endif
  64 static inline long int lround(double x) {
  65     return (long)(x + copysign (0.5, x));
  66 }
  67 //If this fails, we are in the precence of a mid 90's compiler..move along...
  68 #endif
  69
  70 void FLAC__lpc_window_data(const FLAC__int32 in[], const FLAC__real window[], FLAC__real out[], unsigned data_len)
  71 {
  72         unsigned i;
  73         for(i = 0; i < data_len; i++)
  74                 out[i] = in[i] * window[i];
  75 }
  76
  77 void FLAC__lpc_compute_autocorrelation(const FLAC__real data[], unsigned data_len, unsigned lag, FLAC__real autoc[])
  78 {
  79         /* a readable, but slower, version */
  80 #if 0
  81         FLAC__real d;
  82         unsigned i;
  83
  84         FLAC__ASSERT(lag > 0);
  85         FLAC__ASSERT(lag <= data_len);
  86
  87         /*
  88          * Technically we should subtract the mean first like so:
  89          *   for(i = 0; i < data_len; i++)
  90          *     data[i] -= mean;
  91          * but it appears not to make enough of a difference to matter, and
  92          * most signals are already closely centered around zero
  93          */
  94         while(lag--) {
  95                 for(i = lag, d = 0.0; i < data_len; i++)
  96                         d += data[i] * data[i - lag];
  97                 autoc[lag] = d;
  98         }
  99 #endif
 100
 101         /*
 102          * this version tends to run faster because of better data locality
 103          * ('data_len' is usually much larger than 'lag')
 104          */
 105         FLAC__real d;
 106         unsigned sample, coeff;
 107         const unsigned limit = data_len - lag;
 108
 109         FLAC__ASSERT(lag > 0);
 110         FLAC__ASSERT(lag <= data_len);
 111
 112         for(coeff = 0; coeff < lag; coeff++)
 113                 autoc[coeff] = 0.0;
 114         for(sample = 0; sample <= limit; sample++) {
 115                 d = data[sample];
 116                 for(coeff = 0; coeff < lag; coeff++)
 117                         autoc[coeff] += d * data[sample+coeff];
 118         }
 119         for(; sample < data_len; sample++) {
 120                 d = data[sample];
 121                 for(coeff = 0; coeff < data_len - sample; coeff++)
 122                         autoc[coeff] += d * data[sample+coeff];
 123         }
 124 }
 125
 126 void FLAC__lpc_compute_lp_coefficients(const FLAC__real autoc[], unsigned *max_order, FLAC__real lp_coeff[][FLAC__MAX_LPC_ORDER], FLAC__double error[])
 127 {
 128         unsigned i, j;
 129         FLAC__double r, err, lpc[FLAC__MAX_LPC_ORDER];
 130
 131         FLAC__ASSERT(0 != max_order);
 132         FLAC__ASSERT(0 < *max_order);
 133         FLAC__ASSERT(*max_order <= FLAC__MAX_LPC_ORDER);
 134         FLAC__ASSERT(autoc[0] != 0.0);
 135
 136         err = autoc[0];
 137
 138         for(i = 0; i < *max_order; i++) {
 139                 /* Sum up this iteration's reflection coefficient. */
 140                 r = -autoc[i+1];
 141                 for(j = 0; j < i; j++)
 142                         r -= lpc[j] * autoc[i-j];
 143
 144                 /* Update LPC coefficients and total error. */
 145                 lpc[i]=r;
 146                 for(j = 0; j < (i>>1); j++) {
 147                         FLAC__double tmp = lpc[j];
 148                         lpc[j] += r * lpc[i-1-j];
 149                         lpc[i-1-j] += r * tmp;
 150                 }
 151                 if(i & 1)
 152                         lpc[j] += lpc[j] * r;
 153
 154                 err *= (1.0 - r * r);
 155
 156                 /* save this order */
 157                 for(j = 0; j <= i; j++)
 158                         lp_coeff[i][j] = (FLAC__real)(-lpc[j]); /* negate FIR filter coeff to get predictor coeff */
 159                 error[i] = err;
 160
 161                 /* see SF bug #1601812 http://sourceforge.net/tracker/index.php?func=detail&aid=1601812&group_id=13478&atid=113478 */
 162                 if(err == 0.0) {
 163                         *max_order = i+1;
 164                         return;
 165                 }
 166         }
 167 }
 168
 169 int FLAC__lpc_quantize_coefficients(const FLAC__real lp_coeff[], unsigned order, unsigned precision, FLAC__int32 qlp_coeff[], int *shift)
 170 {
 171         unsigned i;
 172         FLAC__double cmax;
 173         FLAC__int32 qmax, qmin;
 174
 175         FLAC__ASSERT(precision > 0);
 176         FLAC__ASSERT(precision >= FLAC__MIN_QLP_COEFF_PRECISION);
 177
 178         /* drop one bit for the sign; from here on out we consider only |lp_coeff[i]| */
 179         precision--;
 180         qmax = 1 << precision;
 181         qmin = -qmax;
 182         qmax--;
 183
 184         /* calc cmax = max( |lp_coeff[i]| ) */
 185         cmax = 0.0;
 186         for(i = 0; i < order; i++) {
 187                 const FLAC__double d = fabs(lp_coeff[i]);
 188                 if(d > cmax)
 189                         cmax = d;
 190         }
 191
 192         if(cmax <= 0.0) {
 193                 /* => coefficients are all 0, which means our constant-detect didn't work */
 194                 return 2;
 195         }
 196         else {
 197                 const int max_shiftlimit = (1 << (FLAC__SUBFRAME_LPC_QLP_SHIFT_LEN-1)) - 1;
 198                 const int min_shiftlimit = -max_shiftlimit - 1;
 199                 int log2cmax;
 200
 201                 (void)frexp(cmax, &log2cmax);
 202                 log2cmax--;
 203                 *shift = (int)precision - log2cmax - 1;
 204
 205                 if(*shift > max_shiftlimit)
 206                         *shift = max_shiftlimit;
 207                 else if(*shift < min_shiftlimit)
 208                         return 1;
 209         }
 210
 211         if(*shift >= 0) {
 212                 FLAC__double error = 0.0;
 213                 FLAC__int32 q;
 214                 for(i = 0; i < order; i++) {
 215                         error += lp_coeff[i] * (1 << *shift);
 216                         q = lround(error);
 217
 218 #ifdef FLAC__OVERFLOW_DETECT
 219                         if(q > qmax+1) /* we expect q==qmax+1 occasionally due to rounding */
 220                                 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]);
 221                         else if(q < qmin)
 222                                 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]);
 223 #endif
 224                         if(q > qmax)
 225                                 q = qmax;
 226                         else if(q < qmin)
 227                                 q = qmin;
 228                         error -= q;
 229                         qlp_coeff[i] = q;
 230                 }
 231         }
 232         /* negative shift is very rare but due to design flaw, negative shift is
 233          * a NOP in the decoder, so it must be handled specially by scaling down
 234          * coeffs
 235          */
 236         else {
 237                 const int nshift = -(*shift);
 238                 FLAC__double error = 0.0;
 239                 FLAC__int32 q;
 240 #ifdef DEBUG
 241                 fprintf(stderr,"FLAC__lpc_quantize_coefficients: negative shift=%d order=%u cmax=%f\n", *shift, order, cmax);
 242 #endif
 243                 for(i = 0; i < order; i++) {
 244                         error += lp_coeff[i] / (1 << nshift);
 245                         q = lround(error);
 246 #ifdef FLAC__OVERFLOW_DETECT
 247                         if(q > qmax+1) /* we expect q==qmax+1 occasionally due to rounding */
 248                                 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]);
 249                         else if(q < qmin)
 250                                 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]);
 251 #endif
 252                         if(q > qmax)
 253                                 q = qmax;
 254                         else if(q < qmin)
 255                                 q = qmin;
 256                         error -= q;
 257                         qlp_coeff[i] = q;
 258                 }
 259                 *shift = 0;
 260         }
 261
 262         return 0;
 263 }
 264
 265 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[])
 266 #if defined(FLAC__OVERFLOW_DETECT) || !defined(FLAC__LPC_UNROLLED_FILTER_LOOPS)
 267 {
 268         FLAC__int64 sumo;
 269         unsigned i, j;
 270         FLAC__int32 sum;
 271         const FLAC__int32 *history;
 272
 273 #ifdef FLAC__OVERFLOW_DETECT_VERBOSE
 274         fprintf(stderr,"FLAC__lpc_compute_residual_from_qlp_coefficients: data_len=%d, order=%u, lpq=%d",data_len,order,lp_quantization);
 275         for(i=0;i<order;i++)
 276                 fprintf(stderr,", q[%u]=%d",i,qlp_coeff[i]);
 277         fprintf(stderr,"\n");
 278 #endif
 279         FLAC__ASSERT(order > 0);
 280
 281         for(i = 0; i < data_len; i++) {
 282                 sumo = 0;
 283                 sum = 0;
 284                 history = data;
 285                 for(j = 0; j < order; j++) {
 286                         sum += qlp_coeff[j] * (*(--history));
 287                         sumo += (FLAC__int64)qlp_coeff[j] * (FLAC__int64)(*history);
 288                                 fprintf(stderr,"FLAC__lpc_compute_residual_from_qlp_coefficients: OVERFLOW, i=%u, j=%u, c=%d, d=%d, sumo=%" PRId64 "\n",i,j,qlp_coeff[j],*history,sumo);
 289                 }
 290                 *(residual++) = *(data++) - (sum >> lp_quantization);
 291         }
 292
 293         /* Here's a slower but clearer version:
 294         for(i = 0; i < data_len; i++) {
 295                 sum = 0;
 296                 for(j = 0; j < order; j++)
 297                         sum += qlp_coeff[j] * data[i-j-1];
 298                 residual[i] = data[i] - (sum >> lp_quantization);
 299         }
 300         */
 301 }
 302 #else /* fully unrolled version for normal use */
 303 {
 304         int i;
 305         FLAC__int32 sum;
 306
 307         FLAC__ASSERT(order > 0);
 308         FLAC__ASSERT(order <= 32);
 309
 310         /*
 311          * We do unique versions up to 12th order since that's the subset limit.
 312          * Also they are roughly ordered to match frequency of occurrence to
 313          * minimize branching.
 314          */
 315         if(order <= 12) {
 316                 if(order > 8) {
 317                         if(order > 10) {
 318                                 if(order == 12) {
 319                                         for(i = 0; i < (int)data_len; i++) {
 320                                                 sum = 0;
 321                                                 sum += qlp_coeff[11] * data[i-12];
 322                                                 sum += qlp_coeff[10] * data[i-11];
 323                                                 sum += qlp_coeff[9] * data[i-10];
 324                                                 sum += qlp_coeff[8] * data[i-9];
 325                                                 sum += qlp_coeff[7] * data[i-8];
 326                                                 sum += qlp_coeff[6] * data[i-7];
 327                                                 sum += qlp_coeff[5] * data[i-6];
 328                                                 sum += qlp_coeff[4] * data[i-5];
 329                                                 sum += qlp_coeff[3] * data[i-4];
 330                                                 sum += qlp_coeff[2] * data[i-3];
 331                                                 sum += qlp_coeff[1] * data[i-2];
 332                                                 sum += qlp_coeff[0] * data[i-1];
 333                                                 residual[i] = data[i] - (sum >> lp_quantization);
 334                                         }
 335                                 }
 336                                 else { /* order == 11 */
 337                                         for(i = 0; i < (int)data_len; i++) {
 338                                                 sum = 0;
 339                                                 sum += qlp_coeff[10] * data[i-11];
 340                                                 sum += qlp_coeff[9] * data[i-10];
 341                                                 sum += qlp_coeff[8] * data[i-9];
 342                                                 sum += qlp_coeff[7] * data[i-8];
 343                                                 sum += qlp_coeff[6] * data[i-7];
 344                                                 sum += qlp_coeff[5] * data[i-6];
 345                                                 sum += qlp_coeff[4] * data[i-5];
 346                                                 sum += qlp_coeff[3] * data[i-4];
 347                                                 sum += qlp_coeff[2] * data[i-3];
 348                                                 sum += qlp_coeff[1] * data[i-2];
 349                                                 sum += qlp_coeff[0] * data[i-1];
 350                                                 residual[i] = data[i] - (sum >> lp_quantization);
 351                                         }
 352                                 }
 353                         }
 354                         else {
 355                                 if(order == 10) {
 356                                         for(i = 0; i < (int)data_len; i++) {
 357                                                 sum = 0;
 358                                                 sum += qlp_coeff[9] * data[i-10];
 359                                                 sum += qlp_coeff[8] * data[i-9];
 360                                                 sum += qlp_coeff[7] * data[i-8];
 361                                                 sum += qlp_coeff[6] * data[i-7];
 362                                                 sum += qlp_coeff[5] * data[i-6];
 363                                                 sum += qlp_coeff[4] * data[i-5];
 364                                                 sum += qlp_coeff[3] * data[i-4];
 365                                                 sum += qlp_coeff[2] * data[i-3];
 366                                                 sum += qlp_coeff[1] * data[i-2];
 367                                                 sum += qlp_coeff[0] * data[i-1];
 368                                                 residual[i] = data[i] - (sum >> lp_quantization);
 369                                         }
 370                                 }
 371                                 else { /* order == 9 */
 372                                         for(i = 0; i < (int)data_len; i++) {
 373                                                 sum = 0;
 374                                                 sum += qlp_coeff[8] * data[i-9];
 375                                                 sum += qlp_coeff[7] * data[i-8];
 376                                                 sum += qlp_coeff[6] * data[i-7];
 377                                                 sum += qlp_coeff[5] * data[i-6];
 378                                                 sum += qlp_coeff[4] * data[i-5];
 379                                                 sum += qlp_coeff[3] * data[i-4];
 380                                                 sum += qlp_coeff[2] * data[i-3];
 381                                                 sum += qlp_coeff[1] * data[i-2];
 382                                                 sum += qlp_coeff[0] * data[i-1];
 383                                                 residual[i] = data[i] - (sum >> lp_quantization);
 384                                         }
 385                                 }
 386                         }
 387                 }
 388                 else if(order > 4) {
 389                         if(order > 6) {
 390                                 if(order == 8) {
 391                                         for(i = 0; i < (int)data_len; i++) {
 392                                                 sum = 0;
 393                                                 sum += qlp_coeff[7] * data[i-8];
 394                                                 sum += qlp_coeff[6] * data[i-7];
 395                                                 sum += qlp_coeff[5] * data[i-6];
 396                                                 sum += qlp_coeff[4] * data[i-5];
 397                                                 sum += qlp_coeff[3] * data[i-4];
 398                                                 sum += qlp_coeff[2] * data[i-3];
 399                                                 sum += qlp_coeff[1] * data[i-2];
 400                                                 sum += qlp_coeff[0] * data[i-1];
 401                                                 residual[i] = data[i] - (sum >> lp_quantization);
 402                                         }
 403                                 }
 404                                 else { /* order == 7 */
 405                                         for(i = 0; i < (int)data_len; i++) {
 406                                                 sum = 0;
 407                                                 sum += qlp_coeff[6] * data[i-7];
 408                                                 sum += qlp_coeff[5] * data[i-6];
 409                                                 sum += qlp_coeff[4] * data[i-5];
 410                                                 sum += qlp_coeff[3] * data[i-4];
 411                                                 sum += qlp_coeff[2] * data[i-3];
 412                                                 sum += qlp_coeff[1] * data[i-2];
 413                                                 sum += qlp_coeff[0] * data[i-1];
 414                                                 residual[i] = data[i] - (sum >> lp_quantization);
 415                                         }
 416                                 }
 417                         }
 418                         else {
 419                                 if(order == 6) {
 420                                         for(i = 0; i < (int)data_len; i++) {
 421                                                 sum = 0;
 422                                                 sum += qlp_coeff[5] * data[i-6];
 423                                                 sum += qlp_coeff[4] * data[i-5];
 424                                                 sum += qlp_coeff[3] * data[i-4];
 425                                                 sum += qlp_coeff[2] * data[i-3];
 426                                                 sum += qlp_coeff[1] * data[i-2];
 427                                                 sum += qlp_coeff[0] * data[i-1];
 428                                                 residual[i] = data[i] - (sum >> lp_quantization);
 429                                         }
 430                                 }
 431                                 else { /* order == 5 */
 432                                         for(i = 0; i < (int)data_len; i++) {
 433                                                 sum = 0;
 434                                                 sum += qlp_coeff[4] * data[i-5];
 435                                                 sum += qlp_coeff[3] * data[i-4];
 436                                                 sum += qlp_coeff[2] * data[i-3];
 437                                                 sum += qlp_coeff[1] * data[i-2];
 438                                                 sum += qlp_coeff[0] * data[i-1];
 439                                                 residual[i] = data[i] - (sum >> lp_quantization);
 440                                         }
 441                                 }
 442                         }
 443                 }
 444                 else {
 445                         if(order > 2) {
 446                                 if(order == 4) {
 447                                         for(i = 0; i < (int)data_len; i++) {
 448                                                 sum = 0;
 449                                                 sum += qlp_coeff[3] * data[i-4];
 450                                                 sum += qlp_coeff[2] * data[i-3];
 451                                                 sum += qlp_coeff[1] * data[i-2];
 452                                                 sum += qlp_coeff[0] * data[i-1];
 453                                                 residual[i] = data[i] - (sum >> lp_quantization);
 454                                         }
 455                                 }
 456                                 else { /* order == 3 */
 457                                         for(i = 0; i < (int)data_len; i++) {
 458                                                 sum = 0;
 459                                                 sum += qlp_coeff[2] * data[i-3];
 460                                                 sum += qlp_coeff[1] * data[i-2];
 461                                                 sum += qlp_coeff[0] * data[i-1];
 462                                                 residual[i] = data[i] - (sum >> lp_quantization);
 463                                         }
 464                                 }
 465                         }
 466                         else {
 467                                 if(order == 2) {
 468                                         for(i = 0; i < (int)data_len; i++) {
 469                                                 sum = 0;
 470                                                 sum += qlp_coeff[1] * data[i-2];
 471                                                 sum += qlp_coeff[0] * data[i-1];
 472                                                 residual[i] = data[i] - (sum >> lp_quantization);
 473                                         }
 474                                 }
 475                                 else { /* order == 1 */
 476                                         for(i = 0; i < (int)data_len; i++)
 477                                                 residual[i] = data[i] - ((qlp_coeff[0] * data[i-1]) >> lp_quantization);
 478                                 }
 479                         }
 480                 }
 481         }
 482         else { /* order > 12 */
 483                 for(i = 0; i < (int)data_len; i++) {
 484                         sum = 0;
 485                         switch(order) {
 486                                 case 32: sum += qlp_coeff[31] * data[i-32];
 487                                 case 31: sum += qlp_coeff[30] * data[i-31];
 488                                 case 30: sum += qlp_coeff[29] * data[i-30];
 489                                 case 29: sum += qlp_coeff[28] * data[i-29];
 490                                 case 28: sum += qlp_coeff[27] * data[i-28];
 491                                 case 27: sum += qlp_coeff[26] * data[i-27];
 492                                 case 26: sum += qlp_coeff[25] * data[i-26];
 493                                 case 25: sum += qlp_coeff[24] * data[i-25];
 494                                 case 24: sum += qlp_coeff[23] * data[i-24];
 495                                 case 23: sum += qlp_coeff[22] * data[i-23];
 496                                 case 22: sum += qlp_coeff[21] * data[i-22];
 497                                 case 21: sum += qlp_coeff[20] * data[i-21];
 498                                 case 20: sum += qlp_coeff[19] * data[i-20];
 499                                 case 19: sum += qlp_coeff[18] * data[i-19];
 500                                 case 18: sum += qlp_coeff[17] * data[i-18];
 501                                 case 17: sum += qlp_coeff[16] * data[i-17];
 502                                 case 16: sum += qlp_coeff[15] * data[i-16];
 503                                 case 15: sum += qlp_coeff[14] * data[i-15];
 504                                 case 14: sum += qlp_coeff[13] * data[i-14];
 505                                 case 13: sum += qlp_coeff[12] * data[i-13];
 506                                          sum += qlp_coeff[11] * data[i-12];
 507                                          sum += qlp_coeff[10] * data[i-11];
 508                                          sum += qlp_coeff[ 9] * data[i-10];
 509                                          sum += qlp_coeff[ 8] * data[i- 9];
 510                                          sum += qlp_coeff[ 7] * data[i- 8];
 511                                          sum += qlp_coeff[ 6] * data[i- 7];
 512                                          sum += qlp_coeff[ 5] * data[i- 6];
 513                                          sum += qlp_coeff[ 4] * data[i- 5];
 514                                          sum += qlp_coeff[ 3] * data[i- 4];
 515                                          sum += qlp_coeff[ 2] * data[i- 3];
 516                                          sum += qlp_coeff[ 1] * data[i- 2];
 517                                          sum += qlp_coeff[ 0] * data[i- 1];
 518                         }
 519                         residual[i] = data[i] - (sum >> lp_quantization);
 520                 }
 521         }
 522 }
 523 #endif
 524
 525 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[])
 526 #if defined(FLAC__OVERFLOW_DETECT) || !defined(FLAC__LPC_UNROLLED_FILTER_LOOPS)
 527 {
 528         unsigned i, j;
 529         FLAC__int64 sum;
 530         const FLAC__int32 *history;
 531
 532 #ifdef FLAC__OVERFLOW_DETECT_VERBOSE
 533         fprintf(stderr,"FLAC__lpc_compute_residual_from_qlp_coefficients_wide: data_len=%d, order=%u, lpq=%d",data_len,order,lp_quantization);
 534         for(i=0;i<order;i++)
 535                 fprintf(stderr,", q[%u]=%d",i,qlp_coeff[i]);
 536         fprintf(stderr,"\n");
 537 #endif
 538         FLAC__ASSERT(order > 0);
 539
 540         for(i = 0; i < data_len; i++) {
 541                 sum = 0;
 542                 history = data;
 543                 for(j = 0; j < order; j++)
 544                         sum += (FLAC__int64)qlp_coeff[j] * (FLAC__int64)(*(--history));
 545                 if(FLAC__bitmath_silog2_wide(sum >> lp_quantization) > 32) {
 546                         fprintf(stderr,"FLAC__lpc_compute_residual_from_qlp_coefficients_wide: OVERFLOW, i=%u, sum=%" PRId64 "\n", i, (sum >> lp_quantization));
 547                         break;
 548                 }
 549                 if(FLAC__bitmath_silog2_wide((FLAC__int64)(*data) - (sum >> lp_quantization)) > 32) {
 550                         fprintf(stderr,"FLAC__lpc_compute_residual_from_qlp_coefficients_wide: OVERFLOW, i=%u, data=%d, sum=%" PRId64 ", residual=%" PRId64 "\n", i, *data, (long long)(sum >> lp_quantization), ((FLAC__int64)(*data) - (sum >> lp_quantization)));
 551                         break;
 552                 }
 553                 *(residual++) = *(data++) - (FLAC__int32)(sum >> lp_quantization);
 554         }
 555 }
 556 #else /* fully unrolled version for normal use */
 557 {
 558         int i;
 559         FLAC__int64 sum;
 560
 561         FLAC__ASSERT(order > 0);
 562         FLAC__ASSERT(order <= 32);
 563
 564         /*
 565          * We do unique versions up to 12th order since that's the subset limit.
 566          * Also they are roughly ordered to match frequency of occurrence to
 567          * minimize branching.
 568          */
 569         if(order <= 12) {
 570                 if(order > 8) {
 571                         if(order > 10) {
 572                                 if(order == 12) {
 573                                         for(i = 0; i < (int)data_len; i++) {
 574                                                 sum = 0;
 575                                                 sum += qlp_coeff[11] * (FLAC__int64)data[i-12];
 576                                                 sum += qlp_coeff[10] * (FLAC__int64)data[i-11];
 577                                                 sum += qlp_coeff[9] * (FLAC__int64)data[i-10];
 578                                                 sum += qlp_coeff[8] * (FLAC__int64)data[i-9];
 579                                                 sum += qlp_coeff[7] * (FLAC__int64)data[i-8];
 580                                                 sum += qlp_coeff[6] * (FLAC__int64)data[i-7];
 581                                                 sum += qlp_coeff[5] * (FLAC__int64)data[i-6];
 582                                                 sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
 583                                                 sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
 584                                                 sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
 585                                                 sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
 586                                                 sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
 587                                                 residual[i] = data[i] - (FLAC__int32)(sum >> lp_quantization);
 588                                         }
 589                                 }
 590                                 else { /* order == 11 */
 591                                         for(i = 0; i < (int)data_len; i++) {
 592                                                 sum = 0;
 593                                                 sum += qlp_coeff[10] * (FLAC__int64)data[i-11];
 594                                                 sum += qlp_coeff[9] * (FLAC__int64)data[i-10];
 595                                                 sum += qlp_coeff[8] * (FLAC__int64)data[i-9];
 596                                                 sum += qlp_coeff[7] * (FLAC__int64)data[i-8];
 597                                                 sum += qlp_coeff[6] * (FLAC__int64)data[i-7];
 598                                                 sum += qlp_coeff[5] * (FLAC__int64)data[i-6];
 599                                                 sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
 600                                                 sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
 601                                                 sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
 602                                                 sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
 603                                                 sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
 604                                                 residual[i] = data[i] - (FLAC__int32)(sum >> lp_quantization);
 605                                         }
 606                                 }
 607                         }
 608                         else {
 609                                 if(order == 10) {
 610                                         for(i = 0; i < (int)data_len; i++) {
 611                                                 sum = 0;
 612                                                 sum += qlp_coeff[9] * (FLAC__int64)data[i-10];
 613                                                 sum += qlp_coeff[8] * (FLAC__int64)data[i-9];
 614                                                 sum += qlp_coeff[7] * (FLAC__int64)data[i-8];
 615                                                 sum += qlp_coeff[6] * (FLAC__int64)data[i-7];
 616                                                 sum += qlp_coeff[5] * (FLAC__int64)data[i-6];
 617                                                 sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
 618                                                 sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
 619                                                 sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
 620                                                 sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
 621                                                 sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
 622                                                 residual[i] = data[i] - (FLAC__int32)(sum >> lp_quantization);
 623                                         }
 624                                 }
 625                                 else { /* order == 9 */
 626                                         for(i = 0; i < (int)data_len; i++) {
 627                                                 sum = 0;
 628                                                 sum += qlp_coeff[8] * (FLAC__int64)data[i-9];
 629                                                 sum += qlp_coeff[7] * (FLAC__int64)data[i-8];
 630                                                 sum += qlp_coeff[6] * (FLAC__int64)data[i-7];
 631                                                 sum += qlp_coeff[5] * (FLAC__int64)data[i-6];
 632                                                 sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
 633                                                 sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
 634                                                 sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
 635                                                 sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
 636                                                 sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
 637                                                 residual[i] = data[i] - (FLAC__int32)(sum >> lp_quantization);
 638                                         }
 639                                 }
 640                         }
 641                 }
 642                 else if(order > 4) {
 643                         if(order > 6) {
 644                                 if(order == 8) {
 645                                         for(i = 0; i < (int)data_len; i++) {
 646                                                 sum = 0;
 647                                                 sum += qlp_coeff[7] * (FLAC__int64)data[i-8];
 648                                                 sum += qlp_coeff[6] * (FLAC__int64)data[i-7];
 649                                                 sum += qlp_coeff[5] * (FLAC__int64)data[i-6];
 650                                                 sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
 651                                                 sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
 652                                                 sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
 653                                                 sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
 654                                                 sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
 655                                                 residual[i] = data[i] - (FLAC__int32)(sum >> lp_quantization);
 656                                         }
 657                                 }
 658                                 else { /* order == 7 */
 659                                         for(i = 0; i < (int)data_len; i++) {
 660                                                 sum = 0;
 661                                                 sum += qlp_coeff[6] * (FLAC__int64)data[i-7];
 662                                                 sum += qlp_coeff[5] * (FLAC__int64)data[i-6];
 663                                                 sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
 664                                                 sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
 665                                                 sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
 666                                                 sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
 667                                                 sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
 668                                                 residual[i] = data[i] - (FLAC__int32)(sum >> lp_quantization);
 669                                         }
 670                                 }
 671                         }
 672                         else {
 673                                 if(order == 6) {
 674                                         for(i = 0; i < (int)data_len; i++) {
 675                                                 sum = 0;
 676                                                 sum += qlp_coeff[5] * (FLAC__int64)data[i-6];
 677                                                 sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
 678                                                 sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
 679                                                 sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
 680                                                 sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
 681                                                 sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
 682                                                 residual[i] = data[i] - (FLAC__int32)(sum >> lp_quantization);
 683                                         }
 684                                 }
 685                                 else { /* order == 5 */
 686                                         for(i = 0; i < (int)data_len; i++) {
 687                                                 sum = 0;
 688                                                 sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
 689                                                 sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
 690                                                 sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
 691                                                 sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
 692                                                 sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
 693                                                 residual[i] = data[i] - (FLAC__int32)(sum >> lp_quantization);
 694                                         }
 695                                 }
 696                         }
 697                 }
 698                 else {
 699                         if(order > 2) {
 700                                 if(order == 4) {
 701                                         for(i = 0; i < (int)data_len; i++) {
 702                                                 sum = 0;
 703                                                 sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
 704                                                 sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
 705                                                 sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
 706                                                 sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
 707                                                 residual[i] = data[i] - (FLAC__int32)(sum >> lp_quantization);
 708                                         }
 709                                 }
 710                                 else { /* order == 3 */
 711                                         for(i = 0; i < (int)data_len; i++) {
 712                                                 sum = 0;
 713                                                 sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
 714                                                 sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
 715                                                 sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
 716                                                 residual[i] = data[i] - (FLAC__int32)(sum >> lp_quantization);
 717                                         }
 718                                 }
 719                         }
 720                         else {
 721                                 if(order == 2) {
 722                                         for(i = 0; i < (int)data_len; i++) {
 723                                                 sum = 0;
 724                                                 sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
 725                                                 sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
 726                                                 residual[i] = data[i] - (FLAC__int32)(sum >> lp_quantization);
 727                                         }
 728                                 }
 729                                 else { /* order == 1 */
 730                                         for(i = 0; i < (int)data_len; i++)
 731                                                 residual[i] = data[i] - (FLAC__int32)((qlp_coeff[0] * (FLAC__int64)data[i-1]) >> lp_quantization);
 732                                 }
 733                         }
 734                 }
 735         }
 736         else { /* order > 12 */
 737                 for(i = 0; i < (int)data_len; i++) {
 738                         sum = 0;
 739                         switch(order) {
 740                                 case 32: sum += qlp_coeff[31] * (FLAC__int64)data[i-32];
 741                                 case 31: sum += qlp_coeff[30] * (FLAC__int64)data[i-31];
 742                                 case 30: sum += qlp_coeff[29] * (FLAC__int64)data[i-30];
 743                                 case 29: sum += qlp_coeff[28] * (FLAC__int64)data[i-29];
 744                                 case 28: sum += qlp_coeff[27] * (FLAC__int64)data[i-28];
 745                                 case 27: sum += qlp_coeff[26] * (FLAC__int64)data[i-27];
 746                                 case 26: sum += qlp_coeff[25] * (FLAC__int64)data[i-26];
 747                                 case 25: sum += qlp_coeff[24] * (FLAC__int64)data[i-25];
 748                                 case 24: sum += qlp_coeff[23] * (FLAC__int64)data[i-24];
 749                                 case 23: sum += qlp_coeff[22] * (FLAC__int64)data[i-23];
 750                                 case 22: sum += qlp_coeff[21] * (FLAC__int64)data[i-22];
 751                                 case 21: sum += qlp_coeff[20] * (FLAC__int64)data[i-21];
 752                                 case 20: sum += qlp_coeff[19] * (FLAC__int64)data[i-20];
 753                                 case 19: sum += qlp_coeff[18] * (FLAC__int64)data[i-19];
 754                                 case 18: sum += qlp_coeff[17] * (FLAC__int64)data[i-18];
 755                                 case 17: sum += qlp_coeff[16] * (FLAC__int64)data[i-17];
 756                                 case 16: sum += qlp_coeff[15] * (FLAC__int64)data[i-16];
 757                                 case 15: sum += qlp_coeff[14] * (FLAC__int64)data[i-15];
 758                                 case 14: sum += qlp_coeff[13] * (FLAC__int64)data[i-14];
 759                                 case 13: sum += qlp_coeff[12] * (FLAC__int64)data[i-13];
 760                                          sum += qlp_coeff[11] * (FLAC__int64)data[i-12];
 761                                          sum += qlp_coeff[10] * (FLAC__int64)data[i-11];
 762                                          sum += qlp_coeff[ 9] * (FLAC__int64)data[i-10];
 763                                          sum += qlp_coeff[ 8] * (FLAC__int64)data[i- 9];
 764                                          sum += qlp_coeff[ 7] * (FLAC__int64)data[i- 8];
 765                                          sum += qlp_coeff[ 6] * (FLAC__int64)data[i- 7];
 766                                          sum += qlp_coeff[ 5] * (FLAC__int64)data[i- 6];
 767                                          sum += qlp_coeff[ 4] * (FLAC__int64)data[i- 5];
 768                                          sum += qlp_coeff[ 3] * (FLAC__int64)data[i- 4];
 769                                          sum += qlp_coeff[ 2] * (FLAC__int64)data[i- 3];
 770                                          sum += qlp_coeff[ 1] * (FLAC__int64)data[i- 2];
 771                                          sum += qlp_coeff[ 0] * (FLAC__int64)data[i- 1];
 772                         }
 773                         residual[i] = data[i] - (FLAC__int32)(sum >> lp_quantization);
 774                 }
 775         }
 776 }
 777 #endif
 778
 779 #endif /* !defined FLAC__INTEGER_ONLY_LIBRARY */
 780
 781 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[])
 782 #if defined(FLAC__OVERFLOW_DETECT) || !defined(FLAC__LPC_UNROLLED_FILTER_LOOPS)
 783 {
 784         FLAC__int64 sumo;
 785         unsigned i, j;
 786         FLAC__int32 sum;
 787         const FLAC__int32 *r = residual, *history;
 788
 789 #ifdef FLAC__OVERFLOW_DETECT_VERBOSE
 790         fprintf(stderr,"FLAC__lpc_restore_signal: data_len=%d, order=%u, lpq=%d",data_len,order,lp_quantization);
 791         for(i=0;i<order;i++)
 792                 fprintf(stderr,", q[%u]=%d",i,qlp_coeff[i]);
 793         fprintf(stderr,"\n");
 794 #endif
 795         FLAC__ASSERT(order > 0);
 796
 797         for(i = 0; i < data_len; i++) {
 798                 sumo = 0;
 799                 sum = 0;
 800                 history = data;
 801                 for(j = 0; j < order; j++) {
 802                         sum += qlp_coeff[j] * (*(--history));
 803                         sumo += (FLAC__int64)qlp_coeff[j] * (FLAC__int64)(*history);
 804                         if(sumo > 2147483647ll || sumo < -2147483648ll)
 805                                 fprintf(stderr,"FLAC__lpc_restore_signal: OVERFLOW, i=%u, j=%u, c=%d, d=%d, sumo=%" PRId64 "\n",i,j,qlp_coeff[j],*history,sumo);
 806                 }
 807                 *(data++) = *(r++) + (sum >> lp_quantization);
 808         }
 809
 810         /* Here's a slower but clearer version:
 811         for(i = 0; i < data_len; i++) {
 812                 sum = 0;
 813                 for(j = 0; j < order; j++)
 814                         sum += qlp_coeff[j] * data[i-j-1];
 815                 data[i] = residual[i] + (sum >> lp_quantization);
 816         }
 817         */
 818 }
 819 #else /* fully unrolled version for normal use */
 820 {
 821         int i;
 822         FLAC__int32 sum;
 823
 824         FLAC__ASSERT(order > 0);
 825         FLAC__ASSERT(order <= 32);
 826
 827         /*
 828          * We do unique versions up to 12th order since that's the subset limit.
 829          * Also they are roughly ordered to match frequency of occurrence to
 830          * minimize branching.
 831          */
 832         if(order <= 12) {
 833                 if(order > 8) {
 834                         if(order > 10) {
 835                                 if(order == 12) {
 836                                         for(i = 0; i < (int)data_len; i++) {
 837                                                 sum = 0;
 838                                                 sum += qlp_coeff[11] * data[i-12];
 839                                                 sum += qlp_coeff[10] * data[i-11];
 840                                                 sum += qlp_coeff[9] * data[i-10];
 841                                                 sum += qlp_coeff[8] * data[i-9];
 842                                                 sum += qlp_coeff[7] * data[i-8];
 843                                                 sum += qlp_coeff[6] * data[i-7];
 844                                                 sum += qlp_coeff[5] * data[i-6];
 845                                                 sum += qlp_coeff[4] * data[i-5];
 846                                                 sum += qlp_coeff[3] * data[i-4];
 847                                                 sum += qlp_coeff[2] * data[i-3];
 848                                                 sum += qlp_coeff[1] * data[i-2];
 849                                                 sum += qlp_coeff[0] * data[i-1];
 850                                                 data[i] = residual[i] + (sum >> lp_quantization);
 851                                         }
 852                                 }
 853                                 else { /* order == 11 */
 854                                         for(i = 0; i < (int)data_len; i++) {
 855                                                 sum = 0;
 856                                                 sum += qlp_coeff[10] * data[i-11];
 857                                                 sum += qlp_coeff[9] * data[i-10];
 858                                                 sum += qlp_coeff[8] * data[i-9];
 859                                                 sum += qlp_coeff[7] * data[i-8];
 860                                                 sum += qlp_coeff[6] * data[i-7];
 861                                                 sum += qlp_coeff[5] * data[i-6];
 862                                                 sum += qlp_coeff[4] * data[i-5];
 863                                                 sum += qlp_coeff[3] * data[i-4];
 864                                                 sum += qlp_coeff[2] * data[i-3];
 865                                                 sum += qlp_coeff[1] * data[i-2];
 866                                                 sum += qlp_coeff[0] * data[i-1];
 867                                                 data[i] = residual[i] + (sum >> lp_quantization);
 868                                         }
 869                                 }
 870                         }
 871                         else {
 872                                 if(order == 10) {
 873                                         for(i = 0; i < (int)data_len; i++) {
 874                                                 sum = 0;
 875                                                 sum += qlp_coeff[9] * data[i-10];
 876                                                 sum += qlp_coeff[8] * data[i-9];
 877                                                 sum += qlp_coeff[7] * data[i-8];
 878                                                 sum += qlp_coeff[6] * data[i-7];
 879                                                 sum += qlp_coeff[5] * data[i-6];
 880                                                 sum += qlp_coeff[4] * data[i-5];
 881                                                 sum += qlp_coeff[3] * data[i-4];
 882                                                 sum += qlp_coeff[2] * data[i-3];
 883                                                 sum += qlp_coeff[1] * data[i-2];
 884                                                 sum += qlp_coeff[0] * data[i-1];
 885                                                 data[i] = residual[i] + (sum >> lp_quantization);
 886                                         }
 887                                 }
 888                                 else { /* order == 9 */
 889                                         for(i = 0; i < (int)data_len; i++) {
 890                                                 sum = 0;
 891                                                 sum += qlp_coeff[8] * data[i-9];
 892                                                 sum += qlp_coeff[7] * data[i-8];
 893                                                 sum += qlp_coeff[6] * data[i-7];
 894                                                 sum += qlp_coeff[5] * data[i-6];
 895                                                 sum += qlp_coeff[4] * data[i-5];
 896                                                 sum += qlp_coeff[3] * data[i-4];
 897                                                 sum += qlp_coeff[2] * data[i-3];
 898                                                 sum += qlp_coeff[1] * data[i-2];
 899                                                 sum += qlp_coeff[0] * data[i-1];
 900                                                 data[i] = residual[i] + (sum >> lp_quantization);
 901                                         }
 902                                 }
 903                         }
 904                 }
 905                 else if(order > 4) {
 906                         if(order > 6) {
 907                                 if(order == 8) {
 908                                         for(i = 0; i < (int)data_len; i++) {
 909                                                 sum = 0;
 910                                                 sum += qlp_coeff[7] * data[i-8];
 911                                                 sum += qlp_coeff[6] * data[i-7];
 912                                                 sum += qlp_coeff[5] * data[i-6];
 913                                                 sum += qlp_coeff[4] * data[i-5];
 914                                                 sum += qlp_coeff[3] * data[i-4];
 915                                                 sum += qlp_coeff[2] * data[i-3];
 916                                                 sum += qlp_coeff[1] * data[i-2];
 917                                                 sum += qlp_coeff[0] * data[i-1];
 918                                                 data[i] = residual[i] + (sum >> lp_quantization);
 919                                         }
 920                                 }
 921                                 else { /* order == 7 */
 922                                         for(i = 0; i < (int)data_len; i++) {
 923                                                 sum = 0;
 924                                                 sum += qlp_coeff[6] * data[i-7];
 925                                                 sum += qlp_coeff[5] * data[i-6];
 926                                                 sum += qlp_coeff[4] * data[i-5];
 927                                                 sum += qlp_coeff[3] * data[i-4];
 928                                                 sum += qlp_coeff[2] * data[i-3];
 929                                                 sum += qlp_coeff[1] * data[i-2];
 930                                                 sum += qlp_coeff[0] * data[i-1];
 931                                                 data[i] = residual[i] + (sum >> lp_quantization);
 932                                         }
 933                                 }
 934                         }
 935                         else {
 936                                 if(order == 6) {
 937                                         for(i = 0; i < (int)data_len; i++) {
 938                                                 sum = 0;
 939                                                 sum += qlp_coeff[5] * data[i-6];
 940                                                 sum += qlp_coeff[4] * data[i-5];
 941                                                 sum += qlp_coeff[3] * data[i-4];
 942                                                 sum += qlp_coeff[2] * data[i-3];
 943                                                 sum += qlp_coeff[1] * data[i-2];
 944                                                 sum += qlp_coeff[0] * data[i-1];
 945                                                 data[i] = residual[i] + (sum >> lp_quantization);
 946                                         }
 947                                 }
 948                                 else { /* order == 5 */
 949                                         for(i = 0; i < (int)data_len; i++) {
 950                                                 sum = 0;
 951                                                 sum += qlp_coeff[4] * data[i-5];
 952                                                 sum += qlp_coeff[3] * data[i-4];
 953                                                 sum += qlp_coeff[2] * data[i-3];
 954                                                 sum += qlp_coeff[1] * data[i-2];
 955                                                 sum += qlp_coeff[0] * data[i-1];
 956                                                 data[i] = residual[i] + (sum >> lp_quantization);
 957                                         }
 958                                 }
 959                         }
 960                 }
 961                 else {
 962                         if(order > 2) {
 963                                 if(order == 4) {
 964                                         for(i = 0; i < (int)data_len; i++) {
 965                                                 sum = 0;
 966                                                 sum += qlp_coeff[3] * data[i-4];
 967                                                 sum += qlp_coeff[2] * data[i-3];
 968                                                 sum += qlp_coeff[1] * data[i-2];
 969                                                 sum += qlp_coeff[0] * data[i-1];
 970                                                 data[i] = residual[i] + (sum >> lp_quantization);
 971                                         }
 972                                 }
 973                                 else { /* order == 3 */
 974                                         for(i = 0; i < (int)data_len; i++) {
 975                                                 sum = 0;
 976                                                 sum += qlp_coeff[2] * data[i-3];
 977                                                 sum += qlp_coeff[1] * data[i-2];
 978                                                 sum += qlp_coeff[0] * data[i-1];
 979                                                 data[i] = residual[i] + (sum >> lp_quantization);
 980                                         }
 981                                 }
 982                         }
 983                         else {
 984                                 if(order == 2) {
 985                                         for(i = 0; i < (int)data_len; i++) {
 986                                                 sum = 0;
 987                                                 sum += qlp_coeff[1] * data[i-2];
 988                                                 sum += qlp_coeff[0] * data[i-1];
 989                                                 data[i] = residual[i] + (sum >> lp_quantization);
 990                                         }
 991                                 }
 992                                 else { /* order == 1 */
 993                                         for(i = 0; i < (int)data_len; i++)
 994                                                 data[i] = residual[i] + ((qlp_coeff[0] * data[i-1]) >> lp_quantization);
 995                                 }
 996                         }
 997                 }
 998         }
 999         else { /* order > 12 */
1000                 for(i = 0; i < (int)data_len; i++) {
1001                         sum = 0;
1002                         switch(order) {
1003                                 case 32: sum += qlp_coeff[31] * data[i-32];
1004                                 case 31: sum += qlp_coeff[30] * data[i-31];
1005                                 case 30: sum += qlp_coeff[29] * data[i-30];
1006                                 case 29: sum += qlp_coeff[28] * data[i-29];
1007                                 case 28: sum += qlp_coeff[27] * data[i-28];
1008                                 case 27: sum += qlp_coeff[26] * data[i-27];
1009                                 case 26: sum += qlp_coeff[25] * data[i-26];
1010                                 case 25: sum += qlp_coeff[24] * data[i-25];
1011                                 case 24: sum += qlp_coeff[23] * data[i-24];
1012                                 case 23: sum += qlp_coeff[22] * data[i-23];
1013                                 case 22: sum += qlp_coeff[21] * data[i-22];
1014                                 case 21: sum += qlp_coeff[20] * data[i-21];
1015                                 case 20: sum += qlp_coeff[19] * data[i-20];
1016                                 case 19: sum += qlp_coeff[18] * data[i-19];
1017                                 case 18: sum += qlp_coeff[17] * data[i-18];
1018                                 case 17: sum += qlp_coeff[16] * data[i-17];
1019                                 case 16: sum += qlp_coeff[15] * data[i-16];
1020                                 case 15: sum += qlp_coeff[14] * data[i-15];
1021                                 case 14: sum += qlp_coeff[13] * data[i-14];
1022                                 case 13: sum += qlp_coeff[12] * data[i-13];
1023                                          sum += qlp_coeff[11] * data[i-12];
1024                                          sum += qlp_coeff[10] * data[i-11];
1025                                          sum += qlp_coeff[ 9] * data[i-10];
1026                                          sum += qlp_coeff[ 8] * data[i- 9];
1027                                          sum += qlp_coeff[ 7] * data[i- 8];
1028                                          sum += qlp_coeff[ 6] * data[i- 7];
1029                                          sum += qlp_coeff[ 5] * data[i- 6];
1030                                          sum += qlp_coeff[ 4] * data[i- 5];
1031                                          sum += qlp_coeff[ 3] * data[i- 4];
1032                                          sum += qlp_coeff[ 2] * data[i- 3];
1033                                          sum += qlp_coeff[ 1] * data[i- 2];
1034                                          sum += qlp_coeff[ 0] * data[i- 1];
1035                         }
1036                         data[i] = residual[i] + (sum >> lp_quantization);
1037                 }
1038         }
1039 }
1040 #endif
1041
1042 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[])
1043 #if defined(FLAC__OVERFLOW_DETECT) || !defined(FLAC__LPC_UNROLLED_FILTER_LOOPS)
1044 {
1045         unsigned i, j;
1046         FLAC__int64 sum;
1047         const FLAC__int32 *r = residual, *history;
1048
1049 #ifdef FLAC__OVERFLOW_DETECT_VERBOSE
1050         fprintf(stderr,"FLAC__lpc_restore_signal_wide: data_len=%d, order=%u, lpq=%d",data_len,order,lp_quantization);
1051         for(i=0;i<order;i++)
1052                 fprintf(stderr,", q[%u]=%d",i,qlp_coeff[i]);
1053         fprintf(stderr,"\n");
1054 #endif
1055         FLAC__ASSERT(order > 0);
1056
1057         for(i = 0; i < data_len; i++) {
1058                 sum = 0;
1059                 history = data;
1060                 for(j = 0; j < order; j++)
1061                         sum += (FLAC__int64)qlp_coeff[j] * (FLAC__int64)(*(--history));
1062                 if(FLAC__bitmath_silog2_wide(sum >> lp_quantization) > 32) {
1063                         fprintf(stderr,"FLAC__lpc_restore_signal_wide: OVERFLOW, i=%u, sum=%" PRId64 "\n", i, (sum >> lp_quantization));
1064                         break;
1065                 }
1066                 if(FLAC__bitmath_silog2_wide((FLAC__int64)(*r) + (sum >> lp_quantization)) > 32) {
1067                         fprintf(stderr,"FLAC__lpc_restore_signal_wide: OVERFLOW, i=%u, residual=%d, sum=%" PRId64 ", data=%" PRId64 "\n", i, *r, (sum >> lp_quantization), ((FLAC__int64)(*r) + (sum >> lp_quantization)));
1068                         break;
1069                 }
1070                 *(data++) = *(r++) + (FLAC__int32)(sum >> lp_quantization);
1071         }
1072 }
1073 #else /* fully unrolled version for normal use */
1074 {
1075         int i;
1076         FLAC__int64 sum;
1077
1078         FLAC__ASSERT(order > 0);
1079         FLAC__ASSERT(order <= 32);
1080
1081         /*
1082          * We do unique versions up to 12th order since that's the subset limit.
1083          * Also they are roughly ordered to match frequency of occurrence to
1084          * minimize branching.
1085          */
1086         if(order <= 12) {
1087                 if(order > 8) {
1088                         if(order > 10) {
1089                                 if(order == 12) {
1090                                         for(i = 0; i < (int)data_len; i++) {
1091                                                 sum = 0;
1092                                                 sum += qlp_coeff[11] * (FLAC__int64)data[i-12];
1093                                                 sum += qlp_coeff[10] * (FLAC__int64)data[i-11];
1094                                                 sum += qlp_coeff[9] * (FLAC__int64)data[i-10];
1095                                                 sum += qlp_coeff[8] * (FLAC__int64)data[i-9];
1096                                                 sum += qlp_coeff[7] * (FLAC__int64)data[i-8];
1097                                                 sum += qlp_coeff[6] * (FLAC__int64)data[i-7];
1098                                                 sum += qlp_coeff[5] * (FLAC__int64)data[i-6];
1099                                                 sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
1100                                                 sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
1101                                                 sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
1102                                                 sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
1103                                                 sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
1104                                                 data[i] = residual[i] + (FLAC__int32)(sum >> lp_quantization);
1105                                         }
1106                                 }
1107                                 else { /* order == 11 */
1108                                         for(i = 0; i < (int)data_len; i++) {
1109                                                 sum = 0;
1110                                                 sum += qlp_coeff[10] * (FLAC__int64)data[i-11];
1111                                                 sum += qlp_coeff[9] * (FLAC__int64)data[i-10];
1112                                                 sum += qlp_coeff[8] * (FLAC__int64)data[i-9];
1113                                                 sum += qlp_coeff[7] * (FLAC__int64)data[i-8];
1114                                                 sum += qlp_coeff[6] * (FLAC__int64)data[i-7];
1115                                                 sum += qlp_coeff[5] * (FLAC__int64)data[i-6];
1116                                                 sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
1117                                                 sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
1118                                                 sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
1119                                                 sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
1120                                                 sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
1121                                                 data[i] = residual[i] + (FLAC__int32)(sum >> lp_quantization);
1122                                         }
1123                                 }
1124                         }
1125                         else {
1126                                 if(order == 10) {
1127                                         for(i = 0; i < (int)data_len; i++) {
1128                                                 sum = 0;
1129                                                 sum += qlp_coeff[9] * (FLAC__int64)data[i-10];
1130                                                 sum += qlp_coeff[8] * (FLAC__int64)data[i-9];
1131                                                 sum += qlp_coeff[7] * (FLAC__int64)data[i-8];
1132                                                 sum += qlp_coeff[6] * (FLAC__int64)data[i-7];
1133                                                 sum += qlp_coeff[5] * (FLAC__int64)data[i-6];
1134                                                 sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
1135                                                 sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
1136                                                 sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
1137                                                 sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
1138                                                 sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
1139                                                 data[i] = residual[i] + (FLAC__int32)(sum >> lp_quantization);
1140                                         }
1141                                 }
1142                                 else { /* order == 9 */
1143                                         for(i = 0; i < (int)data_len; i++) {
1144                                                 sum = 0;
1145                                                 sum += qlp_coeff[8] * (FLAC__int64)data[i-9];
1146                                                 sum += qlp_coeff[7] * (FLAC__int64)data[i-8];
1147                                                 sum += qlp_coeff[6] * (FLAC__int64)data[i-7];
1148                                                 sum += qlp_coeff[5] * (FLAC__int64)data[i-6];
1149                                                 sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
1150                                                 sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
1151                                                 sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
1152                                                 sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
1153                                                 sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
1154                                                 data[i] = residual[i] + (FLAC__int32)(sum >> lp_quantization);
1155                                         }
1156                                 }
1157                         }
1158                 }
1159                 else if(order > 4) {
1160                         if(order > 6) {
1161                                 if(order == 8) {
1162                                         for(i = 0; i < (int)data_len; i++) {
1163                                                 sum = 0;
1164                                                 sum += qlp_coeff[7] * (FLAC__int64)data[i-8];
1165                                                 sum += qlp_coeff[6] * (FLAC__int64)data[i-7];
1166                                                 sum += qlp_coeff[5] * (FLAC__int64)data[i-6];
1167                                                 sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
1168                                                 sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
1169                                                 sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
1170                                                 sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
1171                                                 sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
1172                                                 data[i] = residual[i] + (FLAC__int32)(sum >> lp_quantization);
1173                                         }
1174                                 }
1175                                 else { /* order == 7 */
1176                                         for(i = 0; i < (int)data_len; i++) {
1177                                                 sum = 0;
1178                                                 sum += qlp_coeff[6] * (FLAC__int64)data[i-7];
1179                                                 sum += qlp_coeff[5] * (FLAC__int64)data[i-6];
1180                                                 sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
1181                                                 sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
1182                                                 sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
1183                                                 sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
1184                                                 sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
1185                                                 data[i] = residual[i] + (FLAC__int32)(sum >> lp_quantization);
1186                                         }
1187                                 }
1188                         }
1189                         else {
1190                                 if(order == 6) {
1191                                         for(i = 0; i < (int)data_len; i++) {
1192                                                 sum = 0;
1193                                                 sum += qlp_coeff[5] * (FLAC__int64)data[i-6];
1194                                                 sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
1195                                                 sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
1196                                                 sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
1197                                                 sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
1198                                                 sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
1199                                                 data[i] = residual[i] + (FLAC__int32)(sum >> lp_quantization);
1200                                         }
1201                                 }
1202                                 else { /* order == 5 */
1203                                         for(i = 0; i < (int)data_len; i++) {
1204                                                 sum = 0;
1205                                                 sum += qlp_coeff[4] * (FLAC__int64)data[i-5];
1206                                                 sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
1207                                                 sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
1208                                                 sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
1209                                                 sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
1210                                                 data[i] = residual[i] + (FLAC__int32)(sum >> lp_quantization);
1211                                         }
1212                                 }
1213                         }
1214                 }
1215                 else {
1216                         if(order > 2) {
1217                                 if(order == 4) {
1218                                         for(i = 0; i < (int)data_len; i++) {
1219                                                 sum = 0;
1220                                                 sum += qlp_coeff[3] * (FLAC__int64)data[i-4];
1221                                                 sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
1222                                                 sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
1223                                                 sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
1224                                                 data[i] = residual[i] + (FLAC__int32)(sum >> lp_quantization);
1225                                         }
1226                                 }
1227                                 else { /* order == 3 */
1228                                         for(i = 0; i < (int)data_len; i++) {
1229                                                 sum = 0;
1230                                                 sum += qlp_coeff[2] * (FLAC__int64)data[i-3];
1231                                                 sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
1232                                                 sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
1233                                                 data[i] = residual[i] + (FLAC__int32)(sum >> lp_quantization);
1234                                         }
1235                                 }
1236                         }
1237                         else {
1238                                 if(order == 2) {
1239                                         for(i = 0; i < (int)data_len; i++) {
1240                                                 sum = 0;
1241                                                 sum += qlp_coeff[1] * (FLAC__int64)data[i-2];
1242                                                 sum += qlp_coeff[0] * (FLAC__int64)data[i-1];
1243                                                 data[i] = residual[i] + (FLAC__int32)(sum >> lp_quantization);
1244                                         }
1245                                 }
1246                                 else { /* order == 1 */
1247                                         for(i = 0; i < (int)data_len; i++)
1248                                                 data[i] = residual[i] + (FLAC__int32)((qlp_coeff[0] * (FLAC__int64)data[i-1]) >> lp_quantization);
1249                                 }
1250                         }
1251                 }
1252         }
1253         else { /* order > 12 */
1254                 for(i = 0; i < (int)data_len; i++) {
1255                         sum = 0;
1256                         switch(order) {
1257                                 case 32: sum += qlp_coeff[31] * (FLAC__int64)data[i-32];
1258                                 case 31: sum += qlp_coeff[30] * (FLAC__int64)data[i-31];
1259                                 case 30: sum += qlp_coeff[29] * (FLAC__int64)data[i-30];
1260                                 case 29: sum += qlp_coeff[28] * (FLAC__int64)data[i-29];
1261                                 case 28: sum += qlp_coeff[27] * (FLAC__int64)data[i-28];
1262                                 case 27: sum += qlp_coeff[26] * (FLAC__int64)data[i-27];
1263                                 case 26: sum += qlp_coeff[25] * (FLAC__int64)data[i-26];
1264                                 case 25: sum += qlp_coeff[24] * (FLAC__int64)data[i-25];
1265                                 case 24: sum += qlp_coeff[23] * (FLAC__int64)data[i-24];
1266                                 case 23: sum += qlp_coeff[22] * (FLAC__int64)data[i-23];
1267                                 case 22: sum += qlp_coeff[21] * (FLAC__int64)data[i-22];
1268                                 case 21: sum += qlp_coeff[20] * (FLAC__int64)data[i-21];
1269                                 case 20: sum += qlp_coeff[19] * (FLAC__int64)data[i-20];
1270                                 case 19: sum += qlp_coeff[18] * (FLAC__int64)data[i-19];
1271                                 case 18: sum += qlp_coeff[17] * (FLAC__int64)data[i-18];
1272                                 case 17: sum += qlp_coeff[16] * (FLAC__int64)data[i-17];
1273                                 case 16: sum += qlp_coeff[15] * (FLAC__int64)data[i-16];
1274                                 case 15: sum += qlp_coeff[14] * (FLAC__int64)data[i-15];
1275                                 case 14: sum += qlp_coeff[13] * (FLAC__int64)data[i-14];
1276                                 case 13: sum += qlp_coeff[12] * (FLAC__int64)data[i-13];
1277                                          sum += qlp_coeff[11] * (FLAC__int64)data[i-12];
1278                                          sum += qlp_coeff[10] * (FLAC__int64)data[i-11];
1279                                          sum += qlp_coeff[ 9] * (FLAC__int64)data[i-10];
1280                                          sum += qlp_coeff[ 8] * (FLAC__int64)data[i- 9];
1281                                          sum += qlp_coeff[ 7] * (FLAC__int64)data[i- 8];
1282                                          sum += qlp_coeff[ 6] * (FLAC__int64)data[i- 7];
1283                                          sum += qlp_coeff[ 5] * (FLAC__int64)data[i- 6];
1284                                          sum += qlp_coeff[ 4] * (FLAC__int64)data[i- 5];
1285                                          sum += qlp_coeff[ 3] * (FLAC__int64)data[i- 4];
1286                                          sum += qlp_coeff[ 2] * (FLAC__int64)data[i- 3];
1287                                          sum += qlp_coeff[ 1] * (FLAC__int64)data[i- 2];
1288                                          sum += qlp_coeff[ 0] * (FLAC__int64)data[i- 1];
1289                         }
1290                         data[i] = residual[i] + (FLAC__int32)(sum >> lp_quantization);
1291                 }
1292         }
1293 }
1294 #endif
1295
1296 #ifndef FLAC__INTEGER_ONLY_LIBRARY
1297
1298 FLAC__double FLAC__lpc_compute_expected_bits_per_residual_sample(FLAC__double lpc_error, unsigned total_samples)
1299 {
1300         FLAC__double error_scale;
1301
1302         FLAC__ASSERT(total_samples > 0);
1303
1304         error_scale = 0.5 * M_LN2 * M_LN2 / (FLAC__double)total_samples;
1305
1306         return FLAC__lpc_compute_expected_bits_per_residual_sample_with_error_scale(lpc_error, error_scale);
1307 }
1308
1309 FLAC__double FLAC__lpc_compute_expected_bits_per_residual_sample_with_error_scale(FLAC__double lpc_error, FLAC__double error_scale)
1310 {
1311         if(lpc_error > 0.0) {
1312                 FLAC__double bps = (FLAC__double)0.5 * log(error_scale * lpc_error) / M_LN2;
1313                 if(bps >= 0.0)
1314                         return bps;
1315                 else
1316                         return 0.0;
1317         }
1318         else if(lpc_error < 0.0) { /* error should not be negative but can happen due to inadequate floating-point resolution */
1319                 return 1e32;
1320         }
1321         else {
1322                 return 0.0;
1323         }
1324 }
1325
1326 unsigned FLAC__lpc_compute_best_order(const FLAC__double lpc_error[], unsigned max_order, unsigned total_samples, unsigned overhead_bits_per_order)
1327 {
1328         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 */
1329         FLAC__double bits, best_bits, error_scale;
1330
1331         FLAC__ASSERT(max_order > 0);
1332         FLAC__ASSERT(total_samples > 0);
1333
1334         error_scale = 0.5 * M_LN2 * M_LN2 / (FLAC__double)total_samples;
1335
1336         best_index = 0;
1337         best_bits = (unsigned)(-1);
1338
1339         for(index = 0, order = 1; index < max_order; index++, order++) {
1340                 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);
1341                 if(bits < best_bits) {
1342                         best_index = index;
1343                         best_bits = bits;
1344                 }
1345         }
1346
1347         return best_index+1; /* +1 since index of lpc_error[] is order-1 */
1348 }
1349
1350 #endif /* !defined FLAC__INTEGER_ONLY_LIBRARY */