| File: | root/firefox-clang/obj-x86_64-pc-linux-gnu/media/libsoundtouch/src/./../../../../media/libsoundtouch/src/TDStretch.cpp |
| Warning: | line 283, column 5 Value stored to 'bestCorr' is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | /////////////////////////////////////////////////////////////////////////////// |
| 2 | /// |
| 3 | /// Sampled sound tempo changer/time stretch algorithm. Changes the sound tempo |
| 4 | /// while maintaining the original pitch by using a time domain WSOLA-like |
| 5 | /// method with several performance-increasing tweaks. |
| 6 | /// |
| 7 | /// Notes : MMX optimized functions reside in a separate, platform-specific |
| 8 | /// file, e.g. 'mmx_win.cpp' or 'mmx_gcc.cpp'. |
| 9 | /// |
| 10 | /// This source file contains OpenMP optimizations that allow speeding up the |
| 11 | /// corss-correlation algorithm by executing it in several threads / CPU cores |
| 12 | /// in parallel. See the following article link for more detailed discussion |
| 13 | /// about SoundTouch OpenMP optimizations: |
| 14 | /// http://www.softwarecoven.com/parallel-computing-in-embedded-mobile-devices |
| 15 | /// |
| 16 | /// Author : Copyright (c) Olli Parviainen |
| 17 | /// Author e-mail : oparviai 'at' iki.fi |
| 18 | /// SoundTouch WWW: http://www.surina.net/soundtouch |
| 19 | /// |
| 20 | //////////////////////////////////////////////////////////////////////////////// |
| 21 | // |
| 22 | // License : |
| 23 | // |
| 24 | // SoundTouch audio processing library |
| 25 | // Copyright (c) Olli Parviainen |
| 26 | // |
| 27 | // This library is free software; you can redistribute it and/or |
| 28 | // modify it under the terms of the GNU Lesser General Public |
| 29 | // License as published by the Free Software Foundation; either |
| 30 | // version 2.1 of the License, or (at your option) any later version. |
| 31 | // |
| 32 | // This library is distributed in the hope that it will be useful, |
| 33 | // but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 34 | // MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 35 | // Lesser General Public License for more details. |
| 36 | // |
| 37 | // You should have received a copy of the GNU Lesser General Public |
| 38 | // License along with this library; if not, write to the Free Software |
| 39 | // Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
| 40 | // |
| 41 | //////////////////////////////////////////////////////////////////////////////// |
| 42 | |
| 43 | #include <string.h> |
| 44 | #include <limits.h> |
| 45 | #include <assert.h> |
| 46 | #include <math.h> |
| 47 | #include <float.h> |
| 48 | |
| 49 | #include "STTypes.h" |
| 50 | #include "cpu_detect.h" |
| 51 | #include "TDStretch.h" |
| 52 | |
| 53 | using namespace soundtouch; |
| 54 | |
| 55 | #define max(x, y)(((x) > (y)) ? (x) : (y)) (((x) > (y)) ? (x) : (y)) |
| 56 | |
| 57 | /***************************************************************************** |
| 58 | * |
| 59 | * Implementation of the class 'TDStretch' |
| 60 | * |
| 61 | *****************************************************************************/ |
| 62 | |
| 63 | |
| 64 | TDStretch::TDStretch() : FIFOProcessor(&outputBuffer) |
| 65 | { |
| 66 | bQuickSeek = false; |
| 67 | channels = 2; |
| 68 | |
| 69 | pMidBuffer = nullptr; |
| 70 | pMidBufferUnaligned = nullptr; |
| 71 | overlapLength = 0; |
| 72 | |
| 73 | bAutoSeqSetting = true; |
| 74 | bAutoSeekSetting = true; |
| 75 | |
| 76 | tempo = 1.0f; |
| 77 | setParameters(44100, DEFAULT_SEQUENCE_MS0, DEFAULT_SEEKWINDOW_MS0, DEFAULT_OVERLAP_MS8); |
| 78 | setTempo(1.0f); |
| 79 | |
| 80 | clear(); |
| 81 | } |
| 82 | |
| 83 | |
| 84 | |
| 85 | TDStretch::~TDStretch() |
| 86 | { |
| 87 | delete[] pMidBufferUnaligned; |
| 88 | } |
| 89 | |
| 90 | |
| 91 | |
| 92 | // Sets routine control parameters. These control are certain time constants |
| 93 | // defining how the sound is stretched to the desired duration. |
| 94 | // |
| 95 | // 'sampleRate' = sample rate of the sound |
| 96 | // 'sequenceMS' = one processing sequence length in milliseconds (default = 82 ms) |
| 97 | // 'seekwindowMS' = seeking window length for scanning the best overlapping |
| 98 | // position (default = 28 ms) |
| 99 | // 'overlapMS' = overlapping length (default = 12 ms) |
| 100 | |
| 101 | void TDStretch::setParameters(int aSampleRate, int aSequenceMS, |
| 102 | int aSeekWindowMS, int aOverlapMS) |
| 103 | { |
| 104 | // accept only positive parameter values - if zero or negative, use old values instead |
| 105 | if (aSampleRate > 0) |
| 106 | { |
| 107 | if (aSampleRate > 192000) ST_THROW_RT_ERROR("Error: Excessive samplerate"){(static_cast <bool> ((const char *)"Error: Excessive samplerate" ) ? void (0) : __assert_fail ("(const char *)\"Error: Excessive samplerate\"" , __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__ ));}; |
| 108 | this->sampleRate = aSampleRate; |
| 109 | } |
| 110 | |
| 111 | if (aOverlapMS > 0) this->overlapMs = aOverlapMS; |
| 112 | |
| 113 | if (aSequenceMS > 0) |
| 114 | { |
| 115 | this->sequenceMs = aSequenceMS; |
| 116 | bAutoSeqSetting = false; |
| 117 | } |
| 118 | else if (aSequenceMS == 0) |
| 119 | { |
| 120 | // if zero, use automatic setting |
| 121 | bAutoSeqSetting = true; |
| 122 | } |
| 123 | |
| 124 | if (aSeekWindowMS > 0) |
| 125 | { |
| 126 | this->seekWindowMs = aSeekWindowMS; |
| 127 | bAutoSeekSetting = false; |
| 128 | } |
| 129 | else if (aSeekWindowMS == 0) |
| 130 | { |
| 131 | // if zero, use automatic setting |
| 132 | bAutoSeekSetting = true; |
| 133 | } |
| 134 | |
| 135 | calcSeqParameters(); |
| 136 | |
| 137 | calculateOverlapLength(overlapMs); |
| 138 | |
| 139 | // set tempo to recalculate 'sampleReq' |
| 140 | setTempo(tempo); |
| 141 | } |
| 142 | |
| 143 | |
| 144 | |
| 145 | /// Get routine control parameters, see setParameters() function. |
| 146 | /// Any of the parameters to this function can be nullptr, in such case corresponding parameter |
| 147 | /// value isn't returned. |
| 148 | void TDStretch::getParameters(int *pSampleRate, int *pSequenceMs, int *pSeekWindowMs, int *pOverlapMs) const |
| 149 | { |
| 150 | if (pSampleRate) |
| 151 | { |
| 152 | *pSampleRate = sampleRate; |
| 153 | } |
| 154 | |
| 155 | if (pSequenceMs) |
| 156 | { |
| 157 | *pSequenceMs = (bAutoSeqSetting) ? (USE_AUTO_SEQUENCE_LEN0) : sequenceMs; |
| 158 | } |
| 159 | |
| 160 | if (pSeekWindowMs) |
| 161 | { |
| 162 | *pSeekWindowMs = (bAutoSeekSetting) ? (USE_AUTO_SEEKWINDOW_LEN0) : seekWindowMs; |
| 163 | } |
| 164 | |
| 165 | if (pOverlapMs) |
| 166 | { |
| 167 | *pOverlapMs = overlapMs; |
| 168 | } |
| 169 | } |
| 170 | |
| 171 | |
| 172 | // Overlaps samples in 'midBuffer' with the samples in 'pInput' |
| 173 | void TDStretch::overlapMono(SAMPLETYPE *pOutput, const SAMPLETYPE *pInput) const |
| 174 | { |
| 175 | int i; |
| 176 | SAMPLETYPE m1, m2; |
| 177 | |
| 178 | m1 = (SAMPLETYPE)0; |
| 179 | m2 = (SAMPLETYPE)overlapLength; |
| 180 | |
| 181 | for (i = 0; i < overlapLength ; i ++) |
| 182 | { |
| 183 | pOutput[i] = (pInput[i] * m1 + pMidBuffer[i] * m2 ) / overlapLength; |
| 184 | m1 += 1; |
| 185 | m2 -= 1; |
| 186 | } |
| 187 | } |
| 188 | |
| 189 | |
| 190 | |
| 191 | void TDStretch::clearMidBuffer() |
| 192 | { |
| 193 | memset(pMidBuffer, 0, channels * sizeof(SAMPLETYPE) * overlapLength); |
| 194 | } |
| 195 | |
| 196 | |
| 197 | void TDStretch::clearInput() |
| 198 | { |
| 199 | inputBuffer.clear(); |
| 200 | clearMidBuffer(); |
| 201 | isBeginning = true; |
| 202 | maxnorm = 0; |
| 203 | maxnormf = 1e8; |
| 204 | skipFract = 0; |
| 205 | } |
| 206 | |
| 207 | |
| 208 | // Clears the sample buffers |
| 209 | void TDStretch::clear() |
| 210 | { |
| 211 | outputBuffer.clear(); |
| 212 | clearInput(); |
| 213 | } |
| 214 | |
| 215 | |
| 216 | |
| 217 | // Enables/disables the quick position seeking algorithm. Zero to disable, nonzero |
| 218 | // to enable |
| 219 | void TDStretch::enableQuickSeek(bool enable) |
| 220 | { |
| 221 | bQuickSeek = enable; |
| 222 | } |
| 223 | |
| 224 | |
| 225 | // Returns nonzero if the quick seeking algorithm is enabled. |
| 226 | bool TDStretch::isQuickSeekEnabled() const |
| 227 | { |
| 228 | return bQuickSeek; |
| 229 | } |
| 230 | |
| 231 | |
| 232 | // Seeks for the optimal overlap-mixing position. |
| 233 | int TDStretch::seekBestOverlapPosition(const SAMPLETYPE *refPos) |
| 234 | { |
| 235 | if (bQuickSeek) |
| 236 | { |
| 237 | return seekBestOverlapPositionQuick(refPos); |
| 238 | } |
| 239 | else |
| 240 | { |
| 241 | return seekBestOverlapPositionFull(refPos); |
| 242 | } |
| 243 | } |
| 244 | |
| 245 | |
| 246 | // Overlaps samples in 'midBuffer' with the samples in 'pInputBuffer' at position |
| 247 | // of 'ovlPos'. |
| 248 | inline void TDStretch::overlap(SAMPLETYPE *pOutput, const SAMPLETYPE *pInput, uint ovlPos) const |
| 249 | { |
| 250 | #ifndef USE_MULTICH_ALWAYS |
| 251 | if (channels == 1) |
| 252 | { |
| 253 | // mono sound. |
| 254 | overlapMono(pOutput, pInput + ovlPos); |
| 255 | } |
| 256 | else if (channels == 2) |
| 257 | { |
| 258 | // stereo sound |
| 259 | overlapStereo(pOutput, pInput + 2 * ovlPos); |
| 260 | } |
| 261 | else |
| 262 | #endif // USE_MULTICH_ALWAYS |
| 263 | { |
| 264 | assert(channels > 0)(static_cast <bool> (channels > 0) ? void (0) : __assert_fail ("channels > 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 265 | overlapMulti(pOutput, pInput + channels * ovlPos); |
| 266 | } |
| 267 | } |
| 268 | |
| 269 | |
| 270 | // Seeks for the optimal overlap-mixing position. The 'stereo' version of the |
| 271 | // routine |
| 272 | // |
| 273 | // The best position is determined as the position where the two overlapped |
| 274 | // sample sequences are 'most alike', in terms of the highest cross-correlation |
| 275 | // value over the overlapping period |
| 276 | int TDStretch::seekBestOverlapPositionFull(const SAMPLETYPE *refPos) |
| 277 | { |
| 278 | int bestOffs; |
| 279 | double bestCorr; |
| 280 | int i; |
| 281 | double norm; |
| 282 | |
| 283 | bestCorr = -FLT_MAX3.40282347e+38F; |
Value stored to 'bestCorr' is never read | |
| 284 | bestOffs = 0; |
| 285 | |
| 286 | // Scans for the best correlation value by testing each possible position |
| 287 | // over the permitted range. |
| 288 | bestCorr = calcCrossCorr(refPos, pMidBuffer, norm); |
| 289 | bestCorr = (bestCorr + 0.1) * 0.75; |
| 290 | |
| 291 | #pragma omp parallel for |
| 292 | for (i = 1; i < seekLength; i ++) |
| 293 | { |
| 294 | double corr; |
| 295 | // Calculates correlation value for the mixing position corresponding to 'i' |
| 296 | #if defined(_OPENMP) || defined(ST_SIMD_AVOID_UNALIGNED) |
| 297 | // in parallel OpenMP mode, can't use norm accumulator version as parallel executor won't |
| 298 | // iterate the loop in sequential order |
| 299 | // in SIMD mode, avoid accumulator version to allow avoiding unaligned positions |
| 300 | corr = calcCrossCorr(refPos + channels * i, pMidBuffer, norm); |
| 301 | #else |
| 302 | // In non-parallel version call "calcCrossCorrAccumulate" that is otherwise same |
| 303 | // as "calcCrossCorr", but saves time by reusing & updating previously stored |
| 304 | // "norm" value |
| 305 | corr = calcCrossCorrAccumulate(refPos + channels * i, pMidBuffer, norm); |
| 306 | #endif |
| 307 | // heuristic rule to slightly favour values close to mid of the range |
| 308 | double tmp = (double)(2 * i - seekLength) / (double)seekLength; |
| 309 | corr = ((corr + 0.1) * (1.0 - 0.25 * tmp * tmp)); |
| 310 | |
| 311 | // Checks for the highest correlation value |
| 312 | if (corr > bestCorr) |
| 313 | { |
| 314 | // For optimal performance, enter critical section only in case that best value found. |
| 315 | // in such case repeat 'if' condition as it's possible that parallel execution may have |
| 316 | // updated the bestCorr value in the mean time |
| 317 | #pragma omp critical |
| 318 | if (corr > bestCorr) |
| 319 | { |
| 320 | bestCorr = corr; |
| 321 | bestOffs = i; |
| 322 | } |
| 323 | } |
| 324 | } |
| 325 | |
| 326 | #ifdef SOUNDTOUCH_INTEGER_SAMPLES |
| 327 | adaptNormalizer(); |
| 328 | #endif |
| 329 | |
| 330 | // clear cross correlation routine state if necessary (is so e.g. in MMX routines). |
| 331 | clearCrossCorrState(); |
| 332 | |
| 333 | return bestOffs; |
| 334 | } |
| 335 | |
| 336 | |
| 337 | // Quick seek algorithm for improved runtime-performance: First roughly scans through the |
| 338 | // correlation area, and then scan surroundings of two best preliminary correlation candidates |
| 339 | // with improved precision |
| 340 | // |
| 341 | // Based on testing: |
| 342 | // - This algorithm gives on average 99% as good match as the full algorithm |
| 343 | // - this quick seek algorithm finds the best match on ~90% of cases |
| 344 | // - on those 10% of cases when this algorithm doesn't find best match, |
| 345 | // it still finds on average ~90% match vs. the best possible match |
| 346 | int TDStretch::seekBestOverlapPositionQuick(const SAMPLETYPE *refPos) |
| 347 | { |
| 348 | #define _MIN(a, b)(((a) < (b)) ? (a) : (b)) (((a) < (b)) ? (a) : (b)) |
| 349 | #define SCANSTEP16 16 |
| 350 | #define SCANWIND8 8 |
| 351 | |
| 352 | int bestOffs; |
| 353 | int i; |
| 354 | int bestOffs2; |
| 355 | float bestCorr, corr; |
| 356 | float bestCorr2; |
| 357 | double norm; |
| 358 | |
| 359 | // note: 'float' types used in this function in case that the platform would need to use software-fp |
| 360 | |
| 361 | bestCorr = |
| 362 | bestCorr2 = -FLT_MAX3.40282347e+38F; |
| 363 | bestOffs = |
| 364 | bestOffs2 = SCANWIND8; |
| 365 | |
| 366 | // Scans for the best correlation value by testing each possible position |
| 367 | // over the permitted range. Look for two best matches on the first pass to |
| 368 | // increase possibility of ideal match. |
| 369 | // |
| 370 | // Begin from "SCANSTEP" instead of SCANWIND to make the calculation |
| 371 | // catch the 'middlepoint' of seekLength vector as that's the a-priori |
| 372 | // expected best match position |
| 373 | // |
| 374 | // Roughly: |
| 375 | // - 15% of cases find best result directly on the first round, |
| 376 | // - 75% cases find better match on 2nd round around the best match from 1st round |
| 377 | // - 10% cases find better match on 2nd round around the 2nd-best-match from 1st round |
| 378 | for (i = SCANSTEP16; i < seekLength - SCANWIND8 - 1; i += SCANSTEP16) |
| 379 | { |
| 380 | // Calculates correlation value for the mixing position corresponding |
| 381 | // to 'i' |
| 382 | corr = (float)calcCrossCorr(refPos + channels*i, pMidBuffer, norm); |
| 383 | // heuristic rule to slightly favour values close to mid of the seek range |
| 384 | float tmp = (float)(2 * i - seekLength - 1) / (float)seekLength; |
| 385 | corr = ((corr + 0.1f) * (1.0f - 0.25f * tmp * tmp)); |
| 386 | |
| 387 | // Checks for the highest correlation value |
| 388 | if (corr > bestCorr) |
| 389 | { |
| 390 | // found new best match. keep the previous best as 2nd best match |
| 391 | bestCorr2 = bestCorr; |
| 392 | bestOffs2 = bestOffs; |
| 393 | bestCorr = corr; |
| 394 | bestOffs = i; |
| 395 | } |
| 396 | else if (corr > bestCorr2) |
| 397 | { |
| 398 | // not new best, but still new 2nd best match |
| 399 | bestCorr2 = corr; |
| 400 | bestOffs2 = i; |
| 401 | } |
| 402 | } |
| 403 | |
| 404 | // Scans surroundings of the found best match with small stepping |
| 405 | int end = _MIN(bestOffs + SCANWIND + 1, seekLength)(((bestOffs + 8 + 1) < (seekLength)) ? (bestOffs + 8 + 1) : (seekLength)); |
| 406 | for (i = bestOffs - SCANWIND8; i < end; i++) |
| 407 | { |
| 408 | if (i == bestOffs) continue; // this offset already calculated, thus skip |
| 409 | |
| 410 | // Calculates correlation value for the mixing position corresponding |
| 411 | // to 'i' |
| 412 | corr = (float)calcCrossCorr(refPos + channels*i, pMidBuffer, norm); |
| 413 | // heuristic rule to slightly favour values close to mid of the range |
| 414 | float tmp = (float)(2 * i - seekLength - 1) / (float)seekLength; |
| 415 | corr = ((corr + 0.1f) * (1.0f - 0.25f * tmp * tmp)); |
| 416 | |
| 417 | // Checks for the highest correlation value |
| 418 | if (corr > bestCorr) |
| 419 | { |
| 420 | bestCorr = corr; |
| 421 | bestOffs = i; |
| 422 | } |
| 423 | } |
| 424 | |
| 425 | // Scans surroundings of the 2nd best match with small stepping |
| 426 | end = _MIN(bestOffs2 + SCANWIND + 1, seekLength)(((bestOffs2 + 8 + 1) < (seekLength)) ? (bestOffs2 + 8 + 1 ) : (seekLength)); |
| 427 | for (i = bestOffs2 - SCANWIND8; i < end; i++) |
| 428 | { |
| 429 | if (i == bestOffs2) continue; // this offset already calculated, thus skip |
| 430 | |
| 431 | // Calculates correlation value for the mixing position corresponding |
| 432 | // to 'i' |
| 433 | corr = (float)calcCrossCorr(refPos + channels*i, pMidBuffer, norm); |
| 434 | // heuristic rule to slightly favour values close to mid of the range |
| 435 | float tmp = (float)(2 * i - seekLength - 1) / (float)seekLength; |
| 436 | corr = ((corr + 0.1f) * (1.0f - 0.25f * tmp * tmp)); |
| 437 | |
| 438 | // Checks for the highest correlation value |
| 439 | if (corr > bestCorr) |
| 440 | { |
| 441 | bestCorr = corr; |
| 442 | bestOffs = i; |
| 443 | } |
| 444 | } |
| 445 | |
| 446 | // clear cross correlation routine state if necessary (is so e.g. in MMX routines). |
| 447 | clearCrossCorrState(); |
| 448 | |
| 449 | #ifdef SOUNDTOUCH_INTEGER_SAMPLES |
| 450 | adaptNormalizer(); |
| 451 | #endif |
| 452 | |
| 453 | return bestOffs; |
| 454 | } |
| 455 | |
| 456 | |
| 457 | |
| 458 | |
| 459 | /// For integer algorithm: adapt normalization factor divider with music so that |
| 460 | /// it'll not be pessimistically restrictive that can degrade quality on quieter sections |
| 461 | /// yet won't cause integer overflows either |
| 462 | void TDStretch::adaptNormalizer() |
| 463 | { |
| 464 | // Do not adapt normalizer over too silent sequences to avoid averaging filter depleting to |
| 465 | // too low values during pauses in music |
| 466 | if ((maxnorm > 1000) || (maxnormf > 40000000)) |
| 467 | { |
| 468 | //norm averaging filter |
| 469 | maxnormf = 0.9f * maxnormf + 0.1f * (float)maxnorm; |
| 470 | |
| 471 | if ((maxnorm > 800000000) && (overlapDividerBitsNorm < 16)) |
| 472 | { |
| 473 | // large values, so increase divider |
| 474 | overlapDividerBitsNorm++; |
| 475 | if (maxnorm > 1600000000) overlapDividerBitsNorm++; // extra large value => extra increase |
| 476 | } |
| 477 | else if ((maxnormf < 1000000) && (overlapDividerBitsNorm > 0)) |
| 478 | { |
| 479 | // extra small values, decrease divider |
| 480 | overlapDividerBitsNorm--; |
| 481 | } |
| 482 | } |
| 483 | |
| 484 | maxnorm = 0; |
| 485 | } |
| 486 | |
| 487 | |
| 488 | /// clear cross correlation routine state if necessary |
| 489 | void TDStretch::clearCrossCorrState() |
| 490 | { |
| 491 | // default implementation is empty. |
| 492 | } |
| 493 | |
| 494 | |
| 495 | /// Calculates processing sequence length according to tempo setting |
| 496 | void TDStretch::calcSeqParameters() |
| 497 | { |
| 498 | // Adjust tempo param according to tempo, so that variating processing sequence length is used |
| 499 | // at various tempo settings, between the given low...top limits |
| 500 | #define AUTOSEQ_TEMPO_LOW0.5 0.5 // auto setting low tempo range (-50%) |
| 501 | #define AUTOSEQ_TEMPO_TOP2.0 2.0 // auto setting top tempo range (+100%) |
| 502 | |
| 503 | // sequence-ms setting values at above low & top tempo |
| 504 | #define AUTOSEQ_AT_MIN90.0 90.0 |
| 505 | #define AUTOSEQ_AT_MAX40.0 40.0 |
| 506 | #define AUTOSEQ_K((40.0 - 90.0) / (2.0 - 0.5)) ((AUTOSEQ_AT_MAX40.0 - AUTOSEQ_AT_MIN90.0) / (AUTOSEQ_TEMPO_TOP2.0 - AUTOSEQ_TEMPO_LOW0.5)) |
| 507 | #define AUTOSEQ_C(90.0 - (((40.0 - 90.0) / (2.0 - 0.5))) * (0.5)) (AUTOSEQ_AT_MIN90.0 - (AUTOSEQ_K((40.0 - 90.0) / (2.0 - 0.5))) * (AUTOSEQ_TEMPO_LOW0.5)) |
| 508 | |
| 509 | // seek-window-ms setting values at above low & top tempoq |
| 510 | #define AUTOSEEK_AT_MIN20.0 20.0 |
| 511 | #define AUTOSEEK_AT_MAX15.0 15.0 |
| 512 | #define AUTOSEEK_K((15.0 - 20.0) / (2.0 - 0.5)) ((AUTOSEEK_AT_MAX15.0 - AUTOSEEK_AT_MIN20.0) / (AUTOSEQ_TEMPO_TOP2.0 - AUTOSEQ_TEMPO_LOW0.5)) |
| 513 | #define AUTOSEEK_C(20.0 - (((15.0 - 20.0) / (2.0 - 0.5))) * (0.5)) (AUTOSEEK_AT_MIN20.0 - (AUTOSEEK_K((15.0 - 20.0) / (2.0 - 0.5))) * (AUTOSEQ_TEMPO_LOW0.5)) |
| 514 | |
| 515 | #define CHECK_LIMITS(x, mi, ma)(((x) < (mi)) ? (mi) : (((x) > (ma)) ? (ma) : (x))) (((x) < (mi)) ? (mi) : (((x) > (ma)) ? (ma) : (x))) |
| 516 | |
| 517 | double seq, seek; |
| 518 | |
| 519 | if (bAutoSeqSetting) |
| 520 | { |
| 521 | seq = AUTOSEQ_C(90.0 - (((40.0 - 90.0) / (2.0 - 0.5))) * (0.5)) + AUTOSEQ_K((40.0 - 90.0) / (2.0 - 0.5)) * tempo; |
| 522 | seq = CHECK_LIMITS(seq, AUTOSEQ_AT_MAX, AUTOSEQ_AT_MIN)(((seq) < (40.0)) ? (40.0) : (((seq) > (90.0)) ? (90.0) : (seq))); |
| 523 | sequenceMs = (int)(seq + 0.5); |
| 524 | } |
| 525 | |
| 526 | if (bAutoSeekSetting) |
| 527 | { |
| 528 | seek = AUTOSEEK_C(20.0 - (((15.0 - 20.0) / (2.0 - 0.5))) * (0.5)) + AUTOSEEK_K((15.0 - 20.0) / (2.0 - 0.5)) * tempo; |
| 529 | seek = CHECK_LIMITS(seek, AUTOSEEK_AT_MAX, AUTOSEEK_AT_MIN)(((seek) < (15.0)) ? (15.0) : (((seek) > (20.0)) ? (20.0 ) : (seek))); |
| 530 | seekWindowMs = (int)(seek + 0.5); |
| 531 | } |
| 532 | |
| 533 | // Update seek window lengths |
| 534 | seekWindowLength = (sampleRate * sequenceMs) / 1000; |
| 535 | if (seekWindowLength < 2 * overlapLength) |
| 536 | { |
| 537 | seekWindowLength = 2 * overlapLength; |
| 538 | } |
| 539 | seekLength = (sampleRate * seekWindowMs) / 1000; |
| 540 | } |
| 541 | |
| 542 | |
| 543 | |
| 544 | // Sets new target tempo. Normal tempo = 'SCALE', smaller values represent slower |
| 545 | // tempo, larger faster tempo. |
| 546 | void TDStretch::setTempo(double newTempo) |
| 547 | { |
| 548 | int intskip; |
| 549 | |
| 550 | tempo = newTempo; |
| 551 | |
| 552 | // Calculate new sequence duration |
| 553 | calcSeqParameters(); |
| 554 | |
| 555 | // Calculate ideal skip length (according to tempo value) |
| 556 | nominalSkip = tempo * (seekWindowLength - overlapLength); |
| 557 | intskip = (int)(nominalSkip + 0.5); |
| 558 | |
| 559 | // Calculate how many samples are needed in the 'inputBuffer' to |
| 560 | // process another batch of samples |
| 561 | //sampleReq = max(intskip + overlapLength, seekWindowLength) + seekLength / 2; |
| 562 | sampleReq = max(intskip + overlapLength, seekWindowLength)(((intskip + overlapLength) > (seekWindowLength)) ? (intskip + overlapLength) : (seekWindowLength)) + seekLength; |
| 563 | } |
| 564 | |
| 565 | |
| 566 | |
| 567 | // Sets the number of channels, 1 = mono, 2 = stereo |
| 568 | void TDStretch::setChannels(int numChannels) |
| 569 | { |
| 570 | if (!verifyNumberOfChannels(numChannels) || |
| 571 | (channels == numChannels)) return; |
| 572 | |
| 573 | channels = numChannels; |
| 574 | inputBuffer.setChannels(channels); |
| 575 | outputBuffer.setChannels(channels); |
| 576 | |
| 577 | // re-init overlap/buffer |
| 578 | overlapLength=0; |
| 579 | setParameters(sampleRate); |
| 580 | } |
| 581 | |
| 582 | |
| 583 | // nominal tempo, no need for processing, just pass the samples through |
| 584 | // to outputBuffer |
| 585 | /* |
| 586 | void TDStretch::processNominalTempo() |
| 587 | { |
| 588 | assert(tempo == 1.0f); |
| 589 | |
| 590 | if (bMidBufferDirty) |
| 591 | { |
| 592 | // If there are samples in pMidBuffer waiting for overlapping, |
| 593 | // do a single sliding overlapping with them in order to prevent a |
| 594 | // clicking distortion in the output sound |
| 595 | if (inputBuffer.numSamples() < overlapLength) |
| 596 | { |
| 597 | // wait until we've got overlapLength input samples |
| 598 | return; |
| 599 | } |
| 600 | // Mix the samples in the beginning of 'inputBuffer' with the |
| 601 | // samples in 'midBuffer' using sliding overlapping |
| 602 | overlap(outputBuffer.ptrEnd(overlapLength), inputBuffer.ptrBegin(), 0); |
| 603 | outputBuffer.putSamples(overlapLength); |
| 604 | inputBuffer.receiveSamples(overlapLength); |
| 605 | clearMidBuffer(); |
| 606 | // now we've caught the nominal sample flow and may switch to |
| 607 | // bypass mode |
| 608 | } |
| 609 | |
| 610 | // Simply bypass samples from input to output |
| 611 | outputBuffer.moveSamples(inputBuffer); |
| 612 | } |
| 613 | */ |
| 614 | |
| 615 | |
| 616 | // Processes as many processing frames of the samples 'inputBuffer', store |
| 617 | // the result into 'outputBuffer' |
| 618 | void TDStretch::processSamples() |
| 619 | { |
| 620 | int ovlSkip; |
| 621 | int offset = 0; |
| 622 | int temp; |
| 623 | |
| 624 | /* Removed this small optimization - can introduce a click to sound when tempo setting |
| 625 | crosses the nominal value |
| 626 | if (tempo == 1.0f) |
| 627 | { |
| 628 | // tempo not changed from the original, so bypass the processing |
| 629 | processNominalTempo(); |
| 630 | return; |
| 631 | } |
| 632 | */ |
| 633 | |
| 634 | // Process samples as long as there are enough samples in 'inputBuffer' |
| 635 | // to form a processing frame. |
| 636 | while ((int)inputBuffer.numSamples() >= sampleReq) |
| 637 | { |
| 638 | if (isBeginning == false) |
| 639 | { |
| 640 | // apart from the very beginning of the track, |
| 641 | // scan for the best overlapping position & do overlap-add |
| 642 | offset = seekBestOverlapPosition(inputBuffer.ptrBegin()); |
| 643 | |
| 644 | // Mix the samples in the 'inputBuffer' at position of 'offset' with the |
| 645 | // samples in 'midBuffer' using sliding overlapping |
| 646 | // ... first partially overlap with the end of the previous sequence |
| 647 | // (that's in 'midBuffer') |
| 648 | overlap(outputBuffer.ptrEnd((uint)overlapLength), inputBuffer.ptrBegin(), (uint)offset); |
| 649 | outputBuffer.putSamples((uint)overlapLength); |
| 650 | offset += overlapLength; |
| 651 | } |
| 652 | else |
| 653 | { |
| 654 | // Adjust processing offset at beginning of track by not perform initial overlapping |
| 655 | // and compensating that in the 'input buffer skip' calculation |
| 656 | isBeginning = false; |
| 657 | int skip = (int)(tempo * overlapLength + 0.5 * seekLength + 0.5); |
| 658 | |
| 659 | #ifdef ST_SIMD_AVOID_UNALIGNED |
| 660 | // in SIMD mode, round the skip amount to value corresponding to aligned memory address |
| 661 | if (channels == 1) |
| 662 | { |
| 663 | skip &= -4; |
| 664 | } |
| 665 | else if (channels == 2) |
| 666 | { |
| 667 | skip &= -2; |
| 668 | } |
| 669 | #endif |
| 670 | skipFract -= skip; |
| 671 | if (skipFract <= -nominalSkip) |
| 672 | { |
| 673 | skipFract = -nominalSkip; |
| 674 | } |
| 675 | } |
| 676 | |
| 677 | // ... then copy sequence samples from 'inputBuffer' to output: |
| 678 | |
| 679 | // crosscheck that we don't have buffer overflow... |
| 680 | if ((int)inputBuffer.numSamples() < (offset + seekWindowLength - overlapLength)) |
| 681 | { |
| 682 | continue; // just in case, shouldn't really happen |
| 683 | } |
| 684 | |
| 685 | // length of sequence |
| 686 | temp = (seekWindowLength - 2 * overlapLength); |
| 687 | outputBuffer.putSamples(inputBuffer.ptrBegin() + channels * offset, (uint)temp); |
| 688 | |
| 689 | // Copies the end of the current sequence from 'inputBuffer' to |
| 690 | // 'midBuffer' for being mixed with the beginning of the next |
| 691 | // processing sequence and so on |
| 692 | assert((offset + temp + overlapLength) <= (int)inputBuffer.numSamples())(static_cast <bool> ((offset + temp + overlapLength) <= (int)inputBuffer.numSamples()) ? void (0) : __assert_fail ("(offset + temp + overlapLength) <= (int)inputBuffer.numSamples()" , __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__ )); |
| 693 | memcpy(pMidBuffer, inputBuffer.ptrBegin() + channels * (offset + temp), |
| 694 | channels * sizeof(SAMPLETYPE) * overlapLength); |
| 695 | |
| 696 | // Remove the processed samples from the input buffer. Update |
| 697 | // the difference between integer & nominal skip step to 'skipFract' |
| 698 | // in order to prevent the error from accumulating over time. |
| 699 | skipFract += nominalSkip; // real skip size |
| 700 | ovlSkip = (int)skipFract; // rounded to integer skip |
| 701 | skipFract -= ovlSkip; // maintain the fraction part, i.e. real vs. integer skip |
| 702 | inputBuffer.receiveSamples((uint)ovlSkip); |
| 703 | } |
| 704 | } |
| 705 | |
| 706 | |
| 707 | // Adds 'numsamples' pcs of samples from the 'samples' memory position into |
| 708 | // the input of the object. |
| 709 | void TDStretch::putSamples(const SAMPLETYPE *samples, uint nSamples) |
| 710 | { |
| 711 | // Add the samples into the input buffer |
| 712 | inputBuffer.putSamples(samples, nSamples); |
| 713 | // Process the samples in input buffer |
| 714 | processSamples(); |
| 715 | } |
| 716 | |
| 717 | |
| 718 | |
| 719 | /// Set new overlap length parameter & reallocate RefMidBuffer if necessary. |
| 720 | void TDStretch::acceptNewOverlapLength(int newOverlapLength) |
| 721 | { |
| 722 | int prevOvl; |
| 723 | |
| 724 | assert(newOverlapLength >= 0)(static_cast <bool> (newOverlapLength >= 0) ? void ( 0) : __assert_fail ("newOverlapLength >= 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 725 | prevOvl = overlapLength; |
| 726 | overlapLength = newOverlapLength; |
| 727 | |
| 728 | if (overlapLength > prevOvl) |
| 729 | { |
| 730 | delete[] pMidBufferUnaligned; |
| 731 | |
| 732 | pMidBufferUnaligned = new SAMPLETYPE[overlapLength * channels + 16 / sizeof(SAMPLETYPE)]; |
| 733 | // ensure that 'pMidBuffer' is aligned to 16 byte boundary for efficiency |
| 734 | pMidBuffer = (SAMPLETYPE *)SOUNDTOUCH_ALIGN_POINTER_16(pMidBufferUnaligned)( ( (ulongptr)(pMidBufferUnaligned) + 15 ) & ~(ulongptr)15 ); |
| 735 | |
| 736 | clearMidBuffer(); |
| 737 | } |
| 738 | } |
| 739 | |
| 740 | |
| 741 | // Operator 'new' is overloaded so that it automatically creates a suitable instance |
| 742 | // depending on if we've a MMX/SSE/etc-capable CPU available or not. |
| 743 | void * TDStretch::operator new(size_t) |
| 744 | { |
| 745 | // Notice! don't use "new TDStretch" directly, use "newInstance" to create a new instance instead! |
| 746 | ST_THROW_RT_ERROR("Error in TDStretch::new: Don't use 'new TDStretch' directly, use 'newInstance' member instead!"){(static_cast <bool> ((const char *)"Error in TDStretch::new: Don't use 'new TDStretch' directly, use 'newInstance' member instead!" ) ? void (0) : __assert_fail ("(const char *)\"Error in TDStretch::new: Don't use 'new TDStretch' directly, use 'newInstance' member instead!\"" , __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__ ));}; |
| 747 | return newInstance(); |
| 748 | } |
| 749 | |
| 750 | |
| 751 | TDStretch * TDStretch::newInstance() |
| 752 | { |
| 753 | #if defined(SOUNDTOUCH_ALLOW_MMX) || defined(SOUNDTOUCH_ALLOW_SSE1) |
| 754 | uint uExtensions; |
| 755 | |
| 756 | uExtensions = detectCPUextensions(); |
| 757 | #endif |
| 758 | |
| 759 | // Check if MMX/SSE instruction set extensions supported by CPU |
| 760 | |
| 761 | #ifdef SOUNDTOUCH_ALLOW_MMX |
| 762 | // MMX routines available only with integer sample types |
| 763 | if (uExtensions & SUPPORT_MMX0x0001) |
| 764 | { |
| 765 | return ::new TDStretchMMX; |
| 766 | } |
| 767 | else |
| 768 | #endif // SOUNDTOUCH_ALLOW_MMX |
| 769 | |
| 770 | |
| 771 | #ifdef SOUNDTOUCH_ALLOW_SSE1 |
| 772 | if (uExtensions & SUPPORT_SSE0x0008) |
| 773 | { |
| 774 | // SSE support |
| 775 | return ::new TDStretchSSE; |
| 776 | } |
| 777 | else |
| 778 | #endif // SOUNDTOUCH_ALLOW_SSE |
| 779 | |
| 780 | { |
| 781 | // ISA optimizations not supported, use plain C version |
| 782 | return ::new TDStretch; |
| 783 | } |
| 784 | } |
| 785 | |
| 786 | |
| 787 | ////////////////////////////////////////////////////////////////////////////// |
| 788 | // |
| 789 | // Integer arithmetic specific algorithm implementations. |
| 790 | // |
| 791 | ////////////////////////////////////////////////////////////////////////////// |
| 792 | |
| 793 | #ifdef SOUNDTOUCH_INTEGER_SAMPLES |
| 794 | |
| 795 | // Overlaps samples in 'midBuffer' with the samples in 'input'. The 'Stereo' |
| 796 | // version of the routine. |
| 797 | void TDStretch::overlapStereo(short *poutput, const short *input) const |
| 798 | { |
| 799 | int i; |
| 800 | short temp; |
| 801 | int cnt2; |
| 802 | |
| 803 | for (i = 0; i < overlapLength ; i ++) |
| 804 | { |
| 805 | temp = (short)(overlapLength - i); |
| 806 | cnt2 = 2 * i; |
| 807 | poutput[cnt2] = (input[cnt2] * i + pMidBuffer[cnt2] * temp ) / overlapLength; |
| 808 | poutput[cnt2 + 1] = (input[cnt2 + 1] * i + pMidBuffer[cnt2 + 1] * temp ) / overlapLength; |
| 809 | } |
| 810 | } |
| 811 | |
| 812 | |
| 813 | // Overlaps samples in 'midBuffer' with the samples in 'input'. The 'Multi' |
| 814 | // version of the routine. |
| 815 | void TDStretch::overlapMulti(short *poutput, const short *input) const |
| 816 | { |
| 817 | short m1; |
| 818 | int i = 0; |
| 819 | |
| 820 | for (m1 = 0; m1 < overlapLength; m1 ++) |
| 821 | { |
| 822 | short m2 = (short)(overlapLength - m1); |
| 823 | for (int c = 0; c < channels; c ++) |
| 824 | { |
| 825 | poutput[i] = (input[i] * m1 + pMidBuffer[i] * m2) / overlapLength; |
| 826 | i++; |
| 827 | } |
| 828 | } |
| 829 | } |
| 830 | |
| 831 | // Calculates the x having the closest 2^x value for the given value |
| 832 | static int _getClosest2Power(double value) |
| 833 | { |
| 834 | return (int)(log(value) / log(2.0) + 0.5); |
| 835 | } |
| 836 | |
| 837 | |
| 838 | /// Calculates overlap period length in samples. |
| 839 | /// Integer version rounds overlap length to closest power of 2 |
| 840 | /// for a divide scaling operation. |
| 841 | void TDStretch::calculateOverlapLength(int aoverlapMs) |
| 842 | { |
| 843 | int newOvl; |
| 844 | |
| 845 | assert(aoverlapMs >= 0)(static_cast <bool> (aoverlapMs >= 0) ? void (0) : __assert_fail ("aoverlapMs >= 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 846 | |
| 847 | // calculate overlap length so that it's power of 2 - thus it's easy to do |
| 848 | // integer division by right-shifting. Term "-1" at end is to account for |
| 849 | // the extra most significatnt bit left unused in result by signed multiplication |
| 850 | overlapDividerBitsPure = _getClosest2Power((sampleRate * aoverlapMs) / 1000.0) - 1; |
| 851 | if (overlapDividerBitsPure > 9) overlapDividerBitsPure = 9; |
| 852 | if (overlapDividerBitsPure < 3) overlapDividerBitsPure = 3; |
| 853 | newOvl = (int)pow(2.0, (int)overlapDividerBitsPure + 1); // +1 => account for -1 above |
| 854 | |
| 855 | acceptNewOverlapLength(newOvl); |
| 856 | |
| 857 | overlapDividerBitsNorm = overlapDividerBitsPure; |
| 858 | |
| 859 | // calculate sloping divider so that crosscorrelation operation won't |
| 860 | // overflow 32-bit register. Max. sum of the crosscorrelation sum without |
| 861 | // divider would be 2^30*(N^3-N)/3, where N = overlap length |
| 862 | slopingDivider = (newOvl * newOvl - 1) / 3; |
| 863 | } |
| 864 | |
| 865 | |
| 866 | double TDStretch::calcCrossCorr(const short *mixingPos, const short *compare, double &norm) |
| 867 | { |
| 868 | long corr; |
| 869 | unsigned long lnorm; |
| 870 | int i; |
| 871 | |
| 872 | #ifdef ST_SIMD_AVOID_UNALIGNED |
| 873 | // in SIMD mode skip 'mixingPos' positions that aren't aligned to 16-byte boundary |
| 874 | if (((ulongptr)mixingPos) & 15) return -1e50; |
| 875 | #endif |
| 876 | |
| 877 | // hint compiler autovectorization that loop length is divisible by 8 |
| 878 | int ilength = (channels * overlapLength) & -8; |
| 879 | |
| 880 | corr = lnorm = 0; |
| 881 | // Same routine for stereo and mono |
| 882 | for (i = 0; i < ilength; i += 2) |
| 883 | { |
| 884 | corr += (mixingPos[i] * compare[i] + |
| 885 | mixingPos[i + 1] * compare[i + 1]) >> overlapDividerBitsNorm; |
| 886 | lnorm += (mixingPos[i] * mixingPos[i] + |
| 887 | mixingPos[i + 1] * mixingPos[i + 1]) >> overlapDividerBitsNorm; |
| 888 | // do intermediate scalings to avoid integer overflow |
| 889 | } |
| 890 | |
| 891 | if (lnorm > maxnorm) |
| 892 | { |
| 893 | // modify 'maxnorm' inside critical section to avoid multi-access conflict if in OpenMP mode |
| 894 | #pragma omp critical |
| 895 | if (lnorm > maxnorm) |
| 896 | { |
| 897 | maxnorm = lnorm; |
| 898 | } |
| 899 | } |
| 900 | // Normalize result by dividing by sqrt(norm) - this step is easiest |
| 901 | // done using floating point operation |
| 902 | norm = (double)lnorm; |
| 903 | return (double)corr / sqrt((norm < 1e-9) ? 1.0 : norm); |
| 904 | } |
| 905 | |
| 906 | |
| 907 | /// Update cross-correlation by accumulating "norm" coefficient by previously calculated value |
| 908 | double TDStretch::calcCrossCorrAccumulate(const short *mixingPos, const short *compare, double &norm) |
| 909 | { |
| 910 | long corr; |
| 911 | long lnorm; |
| 912 | int i; |
| 913 | |
| 914 | // hint compiler autovectorization that loop length is divisible by 8 |
| 915 | int ilength = (channels * overlapLength) & -8; |
| 916 | |
| 917 | // cancel first normalizer tap from previous round |
| 918 | lnorm = 0; |
| 919 | for (i = 1; i <= channels; i ++) |
| 920 | { |
| 921 | lnorm -= (mixingPos[-i] * mixingPos[-i]) >> overlapDividerBitsNorm; |
| 922 | } |
| 923 | |
| 924 | corr = 0; |
| 925 | // Same routine for stereo and mono. |
| 926 | for (i = 0; i < ilength; i += 2) |
| 927 | { |
| 928 | corr += (mixingPos[i] * compare[i] + |
| 929 | mixingPos[i + 1] * compare[i + 1]) >> overlapDividerBitsNorm; |
| 930 | } |
| 931 | |
| 932 | // update normalizer with last samples of this round |
| 933 | for (int j = 0; j < channels; j ++) |
| 934 | { |
| 935 | i --; |
| 936 | lnorm += (mixingPos[i] * mixingPos[i]) >> overlapDividerBitsNorm; |
| 937 | } |
| 938 | |
| 939 | norm += (double)lnorm; |
| 940 | if (norm > maxnorm) |
| 941 | { |
| 942 | maxnorm = (unsigned long)norm; |
| 943 | } |
| 944 | |
| 945 | // Normalize result by dividing by sqrt(norm) - this step is easiest |
| 946 | // done using floating point operation |
| 947 | return (double)corr / sqrt((norm < 1e-9) ? 1.0 : norm); |
| 948 | } |
| 949 | |
| 950 | #endif // SOUNDTOUCH_INTEGER_SAMPLES |
| 951 | |
| 952 | ////////////////////////////////////////////////////////////////////////////// |
| 953 | // |
| 954 | // Floating point arithmetic specific algorithm implementations. |
| 955 | // |
| 956 | |
| 957 | #ifdef SOUNDTOUCH_FLOAT_SAMPLES1 |
| 958 | |
| 959 | // Overlaps samples in 'midBuffer' with the samples in 'pInput' |
| 960 | void TDStretch::overlapStereo(float *pOutput, const float *pInput) const |
| 961 | { |
| 962 | int i; |
| 963 | float fScale; |
| 964 | float f1; |
| 965 | float f2; |
| 966 | |
| 967 | fScale = 1.0f / (float)overlapLength; |
| 968 | |
| 969 | f1 = 0; |
| 970 | f2 = 1.0f; |
| 971 | |
| 972 | for (i = 0; i < 2 * (int)overlapLength ; i += 2) |
| 973 | { |
| 974 | pOutput[i + 0] = pInput[i + 0] * f1 + pMidBuffer[i + 0] * f2; |
| 975 | pOutput[i + 1] = pInput[i + 1] * f1 + pMidBuffer[i + 1] * f2; |
| 976 | |
| 977 | f1 += fScale; |
| 978 | f2 -= fScale; |
| 979 | } |
| 980 | } |
| 981 | |
| 982 | |
| 983 | // Overlaps samples in 'midBuffer' with the samples in 'input'. |
| 984 | void TDStretch::overlapMulti(float *pOutput, const float *pInput) const |
| 985 | { |
| 986 | int i; |
| 987 | float fScale; |
| 988 | float f1; |
| 989 | float f2; |
| 990 | |
| 991 | fScale = 1.0f / (float)overlapLength; |
| 992 | |
| 993 | f1 = 0; |
| 994 | f2 = 1.0f; |
| 995 | |
| 996 | i=0; |
| 997 | for (int i2 = 0; i2 < overlapLength; i2 ++) |
| 998 | { |
| 999 | // note: Could optimize this slightly by taking into account that always channels > 2 |
| 1000 | for (int c = 0; c < channels; c ++) |
| 1001 | { |
| 1002 | pOutput[i] = pInput[i] * f1 + pMidBuffer[i] * f2; |
| 1003 | i++; |
| 1004 | } |
| 1005 | f1 += fScale; |
| 1006 | f2 -= fScale; |
| 1007 | } |
| 1008 | } |
| 1009 | |
| 1010 | |
| 1011 | /// Calculates overlapInMsec period length in samples. |
| 1012 | void TDStretch::calculateOverlapLength(int overlapInMsec) |
| 1013 | { |
| 1014 | int newOvl; |
| 1015 | |
| 1016 | assert(overlapInMsec >= 0)(static_cast <bool> (overlapInMsec >= 0) ? void (0) : __assert_fail ("overlapInMsec >= 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1017 | newOvl = (sampleRate * overlapInMsec) / 1000; |
| 1018 | if (newOvl < 16) newOvl = 16; |
| 1019 | |
| 1020 | // must be divisible by 8 |
| 1021 | newOvl -= newOvl % 8; |
| 1022 | |
| 1023 | acceptNewOverlapLength(newOvl); |
| 1024 | } |
| 1025 | |
| 1026 | |
| 1027 | /// Calculate cross-correlation |
| 1028 | double TDStretch::calcCrossCorr(const float *mixingPos, const float *compare, double &anorm) |
| 1029 | { |
| 1030 | float corr; |
| 1031 | float norm; |
| 1032 | int i; |
| 1033 | |
| 1034 | #ifdef ST_SIMD_AVOID_UNALIGNED |
| 1035 | // in SIMD mode skip 'mixingPos' positions that aren't aligned to 16-byte boundary |
| 1036 | if (((ulongptr)mixingPos) & 15) return -1e50; |
| 1037 | #endif |
| 1038 | |
| 1039 | // hint compiler autovectorization that loop length is divisible by 8 |
| 1040 | int ilength = (channels * overlapLength) & -8; |
| 1041 | |
| 1042 | corr = norm = 0; |
| 1043 | // Same routine for stereo and mono |
| 1044 | for (i = 0; i < ilength; i ++) |
| 1045 | { |
| 1046 | corr += mixingPos[i] * compare[i]; |
| 1047 | norm += mixingPos[i] * mixingPos[i]; |
| 1048 | } |
| 1049 | |
| 1050 | anorm = norm; |
| 1051 | return corr / sqrt((norm < 1e-9 ? 1.0 : norm)); |
| 1052 | } |
| 1053 | |
| 1054 | |
| 1055 | /// Update cross-correlation by accumulating "norm" coefficient by previously calculated value |
| 1056 | double TDStretch::calcCrossCorrAccumulate(const float *mixingPos, const float *compare, double &norm) |
| 1057 | { |
| 1058 | float corr; |
| 1059 | int i; |
| 1060 | |
| 1061 | corr = 0; |
| 1062 | |
| 1063 | // cancel first normalizer tap from previous round |
| 1064 | for (i = 1; i <= channels; i ++) |
| 1065 | { |
| 1066 | norm -= mixingPos[-i] * mixingPos[-i]; |
| 1067 | } |
| 1068 | |
| 1069 | // hint compiler autovectorization that loop length is divisible by 8 |
| 1070 | int ilength = (channels * overlapLength) & -8; |
| 1071 | |
| 1072 | // Same routine for stereo and mono |
| 1073 | for (i = 0; i < ilength; i ++) |
| 1074 | { |
| 1075 | corr += mixingPos[i] * compare[i]; |
| 1076 | } |
| 1077 | |
| 1078 | // update normalizer with last samples of this round |
| 1079 | for (int j = 0; j < channels; j ++) |
| 1080 | { |
| 1081 | i --; |
| 1082 | norm += mixingPos[i] * mixingPos[i]; |
| 1083 | } |
| 1084 | |
| 1085 | return corr / sqrt((norm < 1e-9 ? 1.0 : norm)); |
| 1086 | } |
| 1087 | |
| 1088 | |
| 1089 | #endif // SOUNDTOUCH_FLOAT_SAMPLES |