1 //--------------------------------------------------------------------------------- 2 // 3 // Little Color Management System 4 // Copyright (c) 1998-2016 Marti Maria Saguer 5 // 6 // Permission is hereby granted, free of charge, to any person obtaining 7 // a copy of this software and associated documentation files (the "Software"), 8 // to deal in the Software without restriction, including without limitation 9 // the rights to use, copy, modify, merge, publish, distribute, sublicense, 10 // and/or sell copies of the Software, and to permit persons to whom the Software 11 // is furnished to do so, subject to the following conditions: 12 // 13 // The above copyright notice and this permission notice shall be included in 14 // all copies or substantial portions of the Software. 15 // 16 // THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 17 // EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO 18 // THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 19 // NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE 20 // LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION 21 // OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION 22 // WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. 23 // 24 //--------------------------------------------------------------------------------- 25 // 26 27 #include "lcms2_internal.h" 28 29 30 //---------------------------------------------------------------------------------- 31 32 // Optimization for 8 bits, Shaper-CLUT (3 inputs only) 33 typedef struct { 34 35 cmsContext ContextID; 36 37 const cmsInterpParams* p; // Tetrahedrical interpolation parameters. This is a not-owned pointer. 38 39 cmsUInt16Number rx[256], ry[256], rz[256]; 40 cmsUInt32Number X0[256], Y0[256], Z0[256]; // Precomputed nodes and offsets for 8-bit input data 41 42 43 } Prelin8Data; 44 45 46 // Generic optimization for 16 bits Shaper-CLUT-Shaper (any inputs) 47 typedef struct { 48 49 cmsContext ContextID; 50 51 // Number of channels 52 int nInputs; 53 int nOutputs; 54 55 _cmsInterpFn16 EvalCurveIn16[MAX_INPUT_DIMENSIONS]; // The maximum number of input channels is known in advance 56 cmsInterpParams* ParamsCurveIn16[MAX_INPUT_DIMENSIONS]; 57 58 _cmsInterpFn16 EvalCLUT; // The evaluator for 3D grid 59 const cmsInterpParams* CLUTparams; // (not-owned pointer) 60 61 62 _cmsInterpFn16* EvalCurveOut16; // Points to an array of curve evaluators in 16 bits (not-owned pointer) 63 cmsInterpParams** ParamsCurveOut16; // Points to an array of references to interpolation params (not-owned pointer) 64 65 66 } Prelin16Data; 67 68 69 // Optimization for matrix-shaper in 8 bits. Numbers are operated in n.14 signed, tables are stored in 1.14 fixed 70 71 typedef cmsInt32Number cmsS1Fixed14Number; // Note that this may hold more than 16 bits! 72 73 #define DOUBLE_TO_1FIXED14(x) ((cmsS1Fixed14Number) floor((x) * 16384.0 + 0.5)) 74 75 typedef struct { 76 77 cmsContext ContextID; 78 79 cmsS1Fixed14Number Shaper1R[256]; // from 0..255 to 1.14 (0.0...1.0) 80 cmsS1Fixed14Number Shaper1G[256]; 81 cmsS1Fixed14Number Shaper1B[256]; 82 83 cmsS1Fixed14Number Mat[3][3]; // n.14 to n.14 (needs a saturation after that) 84 cmsS1Fixed14Number Off[3]; 85 86 cmsUInt16Number Shaper2R[16385]; // 1.14 to 0..255 87 cmsUInt16Number Shaper2G[16385]; 88 cmsUInt16Number Shaper2B[16385]; 89 90 } MatShaper8Data; 91 92 // Curves, optimization is shared between 8 and 16 bits 93 typedef struct { 94 95 cmsContext ContextID; 96 97 int nCurves; // Number of curves 98 int nElements; // Elements in curves 99 cmsUInt16Number** Curves; // Points to a dynamically allocated array 100 101 } Curves16Data; 102 103 104 // Simple optimizations ---------------------------------------------------------------------------------------------------------- 105 106 107 // Remove an element in linked chain 108 static 109 void _RemoveElement(cmsStage** head) 110 { 111 cmsStage* mpe = *head; 112 cmsStage* next = mpe ->Next; 113 *head = next; 114 cmsStageFree(mpe); 115 } 116 117 // Remove all identities in chain. Note that pt actually is a double pointer to the element that holds the pointer. 118 static 119 cmsBool _Remove1Op(cmsPipeline* Lut, cmsStageSignature UnaryOp) 120 { 121 cmsStage** pt = &Lut ->Elements; 122 cmsBool AnyOpt = FALSE; 123 124 while (*pt != NULL) { 125 126 if ((*pt) ->Implements == UnaryOp) { 127 _RemoveElement(pt); 128 AnyOpt = TRUE; 129 } 130 else 131 pt = &((*pt) -> Next); 132 } 133 134 return AnyOpt; 135 } 136 137 // Same, but only if two adjacent elements are found 138 static 139 cmsBool _Remove2Op(cmsPipeline* Lut, cmsStageSignature Op1, cmsStageSignature Op2) 140 { 141 cmsStage** pt1; 142 cmsStage** pt2; 143 cmsBool AnyOpt = FALSE; 144 145 pt1 = &Lut ->Elements; 146 if (*pt1 == NULL) return AnyOpt; 147 148 while (*pt1 != NULL) { 149 150 pt2 = &((*pt1) -> Next); 151 if (*pt2 == NULL) return AnyOpt; 152 153 if ((*pt1) ->Implements == Op1 && (*pt2) ->Implements == Op2) { 154 _RemoveElement(pt2); 155 _RemoveElement(pt1); 156 AnyOpt = TRUE; 157 } 158 else 159 pt1 = &((*pt1) -> Next); 160 } 161 162 return AnyOpt; 163 } 164 165 166 static 167 cmsBool CloseEnoughFloat(cmsFloat64Number a, cmsFloat64Number b) 168 { 169 return fabs(b - a) < 0.00001f; 170 } 171 172 static 173 cmsBool isFloatMatrixIdentity(const cmsMAT3* a) 174 { 175 cmsMAT3 Identity; 176 int i, j; 177 178 _cmsMAT3identity(&Identity); 179 180 for (i = 0; i < 3; i++) 181 for (j = 0; j < 3; j++) 182 if (!CloseEnoughFloat(a->v[i].n[j], Identity.v[i].n[j])) return FALSE; 183 184 return TRUE; 185 } 186 // if two adjacent matrices are found, multiply them. 187 static 188 cmsBool _MultiplyMatrix(cmsPipeline* Lut) 189 { 190 cmsStage** pt1; 191 cmsStage** pt2; 192 cmsStage* chain; 193 cmsBool AnyOpt = FALSE; 194 195 pt1 = &Lut->Elements; 196 if (*pt1 == NULL) return AnyOpt; 197 198 while (*pt1 != NULL) { 199 200 pt2 = &((*pt1)->Next); 201 if (*pt2 == NULL) return AnyOpt; 202 203 if ((*pt1)->Implements == cmsSigMatrixElemType && (*pt2)->Implements == cmsSigMatrixElemType) { 204 205 // Get both matrices 206 _cmsStageMatrixData* m1 = (_cmsStageMatrixData*) cmsStageData(*pt1); 207 _cmsStageMatrixData* m2 = (_cmsStageMatrixData*) cmsStageData(*pt2); 208 cmsMAT3 res; 209 210 // Input offset and output offset should be zero to use this optimization 211 if (m1->Offset != NULL || m2 ->Offset != NULL || 212 cmsStageInputChannels(*pt1) != 3 || cmsStageOutputChannels(*pt1) != 3 || 213 cmsStageInputChannels(*pt2) != 3 || cmsStageOutputChannels(*pt2) != 3) 214 return FALSE; 215 216 // Multiply both matrices to get the result 217 _cmsMAT3per(&res, (cmsMAT3*)m2->Double, (cmsMAT3*)m1->Double); 218 219 // Get the next in chain afer the matrices 220 chain = (*pt2)->Next; 221 222 // Remove both matrices 223 _RemoveElement(pt2); 224 _RemoveElement(pt1); 225 226 // Now what if the result is a plain identity? 227 if (!isFloatMatrixIdentity(&res)) { 228 229 // We can not get rid of full matrix 230 cmsStage* Multmat = cmsStageAllocMatrix(Lut->ContextID, 3, 3, (const cmsFloat64Number*) &res, NULL); 231 if (Multmat == NULL) return FALSE; // Should never happen 232 233 // Recover the chain 234 Multmat->Next = chain; 235 *pt1 = Multmat; 236 } 237 238 AnyOpt = TRUE; 239 } 240 else 241 pt1 = &((*pt1)->Next); 242 } 243 244 return AnyOpt; 245 } 246 247 248 // Preoptimize just gets rif of no-ops coming paired. Conversion from v2 to v4 followed 249 // by a v4 to v2 and vice-versa. The elements are then discarded. 250 static 251 cmsBool PreOptimize(cmsPipeline* Lut) 252 { 253 cmsBool AnyOpt = FALSE, Opt; 254 255 do { 256 257 Opt = FALSE; 258 259 // Remove all identities 260 Opt |= _Remove1Op(Lut, cmsSigIdentityElemType); 261 262 // Remove XYZ2Lab followed by Lab2XYZ 263 Opt |= _Remove2Op(Lut, cmsSigXYZ2LabElemType, cmsSigLab2XYZElemType); 264 265 // Remove Lab2XYZ followed by XYZ2Lab 266 Opt |= _Remove2Op(Lut, cmsSigLab2XYZElemType, cmsSigXYZ2LabElemType); 267 268 // Remove V4 to V2 followed by V2 to V4 269 Opt |= _Remove2Op(Lut, cmsSigLabV4toV2, cmsSigLabV2toV4); 270 271 // Remove V2 to V4 followed by V4 to V2 272 Opt |= _Remove2Op(Lut, cmsSigLabV2toV4, cmsSigLabV4toV2); 273 274 // Remove float pcs Lab conversions 275 Opt |= _Remove2Op(Lut, cmsSigLab2FloatPCS, cmsSigFloatPCS2Lab); 276 277 // Remove float pcs Lab conversions 278 Opt |= _Remove2Op(Lut, cmsSigXYZ2FloatPCS, cmsSigFloatPCS2XYZ); 279 280 // Simplify matrix. 281 Opt |= _MultiplyMatrix(Lut); 282 283 if (Opt) AnyOpt = TRUE; 284 285 } while (Opt); 286 287 return AnyOpt; 288 } 289 290 static 291 void Eval16nop1D(register const cmsUInt16Number Input[], 292 register cmsUInt16Number Output[], 293 register const struct _cms_interp_struc* p) 294 { 295 Output[0] = Input[0]; 296 297 cmsUNUSED_PARAMETER(p); 298 } 299 300 static 301 void PrelinEval16(register const cmsUInt16Number Input[], 302 register cmsUInt16Number Output[], 303 register const void* D) 304 { 305 Prelin16Data* p16 = (Prelin16Data*) D; 306 cmsUInt16Number StageABC[MAX_INPUT_DIMENSIONS]; 307 cmsUInt16Number StageDEF[cmsMAXCHANNELS]; 308 int i; 309 310 for (i=0; i < p16 ->nInputs; i++) { 311 312 p16 ->EvalCurveIn16[i](&Input[i], &StageABC[i], p16 ->ParamsCurveIn16[i]); 313 } 314 315 p16 ->EvalCLUT(StageABC, StageDEF, p16 ->CLUTparams); 316 317 for (i=0; i < p16 ->nOutputs; i++) { 318 319 p16 ->EvalCurveOut16[i](&StageDEF[i], &Output[i], p16 ->ParamsCurveOut16[i]); 320 } 321 } 322 323 324 static 325 void PrelinOpt16free(cmsContext ContextID, void* ptr) 326 { 327 Prelin16Data* p16 = (Prelin16Data*) ptr; 328 329 _cmsFree(ContextID, p16 ->EvalCurveOut16); 330 _cmsFree(ContextID, p16 ->ParamsCurveOut16); 331 332 _cmsFree(ContextID, p16); 333 } 334 335 static 336 void* Prelin16dup(cmsContext ContextID, const void* ptr) 337 { 338 Prelin16Data* p16 = (Prelin16Data*) ptr; 339 Prelin16Data* Duped = (Prelin16Data*) _cmsDupMem(ContextID, p16, sizeof(Prelin16Data)); 340 341 if (Duped == NULL) return NULL; 342 343 Duped->EvalCurveOut16 = (_cmsInterpFn16*) _cmsDupMem(ContextID, p16->EvalCurveOut16, p16->nOutputs * sizeof(_cmsInterpFn16)); 344 Duped->ParamsCurveOut16 = (cmsInterpParams**)_cmsDupMem(ContextID, p16->ParamsCurveOut16, p16->nOutputs * sizeof(cmsInterpParams*)); 345 346 return Duped; 347 } 348 349 350 static 351 Prelin16Data* PrelinOpt16alloc(cmsContext ContextID, 352 const cmsInterpParams* ColorMap, 353 int nInputs, cmsToneCurve** In, 354 int nOutputs, cmsToneCurve** Out ) 355 { 356 int i; 357 Prelin16Data* p16 = (Prelin16Data*)_cmsMallocZero(ContextID, sizeof(Prelin16Data)); 358 if (p16 == NULL) return NULL; 359 360 p16 ->nInputs = nInputs; 361 p16 -> nOutputs = nOutputs; 362 363 364 for (i=0; i < nInputs; i++) { 365 366 if (In == NULL) { 367 p16 -> ParamsCurveIn16[i] = NULL; 368 p16 -> EvalCurveIn16[i] = Eval16nop1D; 369 370 } 371 else { 372 p16 -> ParamsCurveIn16[i] = In[i] ->InterpParams; 373 p16 -> EvalCurveIn16[i] = p16 ->ParamsCurveIn16[i]->Interpolation.Lerp16; 374 } 375 } 376 377 p16 ->CLUTparams = ColorMap; 378 p16 ->EvalCLUT = ColorMap ->Interpolation.Lerp16; 379 380 381 p16 -> EvalCurveOut16 = (_cmsInterpFn16*) _cmsCalloc(ContextID, nOutputs, sizeof(_cmsInterpFn16)); 382 p16 -> ParamsCurveOut16 = (cmsInterpParams**) _cmsCalloc(ContextID, nOutputs, sizeof(cmsInterpParams* )); 383 384 for (i=0; i < nOutputs; i++) { 385 386 if (Out == NULL) { 387 p16 ->ParamsCurveOut16[i] = NULL; 388 p16 -> EvalCurveOut16[i] = Eval16nop1D; 389 } 390 else { 391 392 p16 ->ParamsCurveOut16[i] = Out[i] ->InterpParams; 393 p16 -> EvalCurveOut16[i] = p16 ->ParamsCurveOut16[i]->Interpolation.Lerp16; 394 } 395 } 396 397 return p16; 398 } 399 400 401 402 // Resampling --------------------------------------------------------------------------------- 403 404 #define PRELINEARIZATION_POINTS 4096 405 406 // Sampler implemented by another LUT. This is a clean way to precalculate the devicelink 3D CLUT for 407 // almost any transform. We use floating point precision and then convert from floating point to 16 bits. 408 static 409 int XFormSampler16(register const cmsUInt16Number In[], register cmsUInt16Number Out[], register void* Cargo) 410 { 411 cmsPipeline* Lut = (cmsPipeline*) Cargo; 412 cmsFloat32Number InFloat[cmsMAXCHANNELS], OutFloat[cmsMAXCHANNELS]; 413 cmsUInt32Number i; 414 415 _cmsAssert(Lut -> InputChannels < cmsMAXCHANNELS); 416 _cmsAssert(Lut -> OutputChannels < cmsMAXCHANNELS); 417 418 // From 16 bit to floating point 419 for (i=0; i < Lut ->InputChannels; i++) 420 InFloat[i] = (cmsFloat32Number) (In[i] / 65535.0); 421 422 // Evaluate in floating point 423 cmsPipelineEvalFloat(InFloat, OutFloat, Lut); 424 425 // Back to 16 bits representation 426 for (i=0; i < Lut ->OutputChannels; i++) 427 Out[i] = _cmsQuickSaturateWord(OutFloat[i] * 65535.0); 428 429 // Always succeed 430 return TRUE; 431 } 432 433 // Try to see if the curves of a given MPE are linear 434 static 435 cmsBool AllCurvesAreLinear(cmsStage* mpe) 436 { 437 cmsToneCurve** Curves; 438 cmsUInt32Number i, n; 439 440 Curves = _cmsStageGetPtrToCurveSet(mpe); 441 if (Curves == NULL) return FALSE; 442 443 n = cmsStageOutputChannels(mpe); 444 445 for (i=0; i < n; i++) { 446 if (!cmsIsToneCurveLinear(Curves[i])) return FALSE; 447 } 448 449 return TRUE; 450 } 451 452 // This function replaces a specific node placed in "At" by the "Value" numbers. Its purpose 453 // is to fix scum dot on broken profiles/transforms. Works on 1, 3 and 4 channels 454 static 455 cmsBool PatchLUT(cmsStage* CLUT, cmsUInt16Number At[], cmsUInt16Number Value[], 456 int nChannelsOut, int nChannelsIn) 457 { 458 _cmsStageCLutData* Grid = (_cmsStageCLutData*) CLUT ->Data; 459 cmsInterpParams* p16 = Grid ->Params; 460 cmsFloat64Number px, py, pz, pw; 461 int x0, y0, z0, w0; 462 int i, index; 463 464 if (CLUT -> Type != cmsSigCLutElemType) { 465 cmsSignalError(CLUT->ContextID, cmsERROR_INTERNAL, "(internal) Attempt to PatchLUT on non-lut stage"); 466 return FALSE; 467 } 468 469 if (nChannelsIn == 4) { 470 471 px = ((cmsFloat64Number) At[0] * (p16->Domain[0])) / 65535.0; 472 py = ((cmsFloat64Number) At[1] * (p16->Domain[1])) / 65535.0; 473 pz = ((cmsFloat64Number) At[2] * (p16->Domain[2])) / 65535.0; 474 pw = ((cmsFloat64Number) At[3] * (p16->Domain[3])) / 65535.0; 475 476 x0 = (int) floor(px); 477 y0 = (int) floor(py); 478 z0 = (int) floor(pz); 479 w0 = (int) floor(pw); 480 481 if (((px - x0) != 0) || 482 ((py - y0) != 0) || 483 ((pz - z0) != 0) || 484 ((pw - w0) != 0)) return FALSE; // Not on exact node 485 486 index = p16 -> opta[3] * x0 + 487 p16 -> opta[2] * y0 + 488 p16 -> opta[1] * z0 + 489 p16 -> opta[0] * w0; 490 } 491 else 492 if (nChannelsIn == 3) { 493 494 px = ((cmsFloat64Number) At[0] * (p16->Domain[0])) / 65535.0; 495 py = ((cmsFloat64Number) At[1] * (p16->Domain[1])) / 65535.0; 496 pz = ((cmsFloat64Number) At[2] * (p16->Domain[2])) / 65535.0; 497 498 x0 = (int) floor(px); 499 y0 = (int) floor(py); 500 z0 = (int) floor(pz); 501 502 if (((px - x0) != 0) || 503 ((py - y0) != 0) || 504 ((pz - z0) != 0)) return FALSE; // Not on exact node 505 506 index = p16 -> opta[2] * x0 + 507 p16 -> opta[1] * y0 + 508 p16 -> opta[0] * z0; 509 } 510 else 511 if (nChannelsIn == 1) { 512 513 px = ((cmsFloat64Number) At[0] * (p16->Domain[0])) / 65535.0; 514 515 x0 = (int) floor(px); 516 517 if (((px - x0) != 0)) return FALSE; // Not on exact node 518 519 index = p16 -> opta[0] * x0; 520 } 521 else { 522 cmsSignalError(CLUT->ContextID, cmsERROR_INTERNAL, "(internal) %d Channels are not supported on PatchLUT", nChannelsIn); 523 return FALSE; 524 } 525 526 for (i=0; i < nChannelsOut; i++) 527 Grid -> Tab.T[index + i] = Value[i]; 528 529 return TRUE; 530 } 531 532 // Auxiliary, to see if two values are equal or very different 533 static 534 cmsBool WhitesAreEqual(int n, cmsUInt16Number White1[], cmsUInt16Number White2[] ) 535 { 536 int i; 537 538 for (i=0; i < n; i++) { 539 540 if (abs(White1[i] - White2[i]) > 0xf000) return TRUE; // Values are so extremely different that the fixup should be avoided 541 if (White1[i] != White2[i]) return FALSE; 542 } 543 return TRUE; 544 } 545 546 547 // Locate the node for the white point and fix it to pure white in order to avoid scum dot. 548 static 549 cmsBool FixWhiteMisalignment(cmsPipeline* Lut, cmsColorSpaceSignature EntryColorSpace, cmsColorSpaceSignature ExitColorSpace) 550 { 551 cmsUInt16Number *WhitePointIn, *WhitePointOut; 552 cmsUInt16Number WhiteIn[cmsMAXCHANNELS], WhiteOut[cmsMAXCHANNELS], ObtainedOut[cmsMAXCHANNELS]; 553 cmsUInt32Number i, nOuts, nIns; 554 cmsStage *PreLin = NULL, *CLUT = NULL, *PostLin = NULL; 555 556 if (!_cmsEndPointsBySpace(EntryColorSpace, 557 &WhitePointIn, NULL, &nIns)) return FALSE; 558 559 if (!_cmsEndPointsBySpace(ExitColorSpace, 560 &WhitePointOut, NULL, &nOuts)) return FALSE; 561 562 // It needs to be fixed? 563 if (Lut ->InputChannels != nIns) return FALSE; 564 if (Lut ->OutputChannels != nOuts) return FALSE; 565 566 cmsPipelineEval16(WhitePointIn, ObtainedOut, Lut); 567 568 if (WhitesAreEqual(nOuts, WhitePointOut, ObtainedOut)) return TRUE; // whites already match 569 570 // Check if the LUT comes as Prelin, CLUT or Postlin. We allow all combinations 571 if (!cmsPipelineCheckAndRetreiveStages(Lut, 3, cmsSigCurveSetElemType, cmsSigCLutElemType, cmsSigCurveSetElemType, &PreLin, &CLUT, &PostLin)) 572 if (!cmsPipelineCheckAndRetreiveStages(Lut, 2, cmsSigCurveSetElemType, cmsSigCLutElemType, &PreLin, &CLUT)) 573 if (!cmsPipelineCheckAndRetreiveStages(Lut, 2, cmsSigCLutElemType, cmsSigCurveSetElemType, &CLUT, &PostLin)) 574 if (!cmsPipelineCheckAndRetreiveStages(Lut, 1, cmsSigCLutElemType, &CLUT)) 575 return FALSE; 576 577 // We need to interpolate white points of both, pre and post curves 578 if (PreLin) { 579 580 cmsToneCurve** Curves = _cmsStageGetPtrToCurveSet(PreLin); 581 582 for (i=0; i < nIns; i++) { 583 WhiteIn[i] = cmsEvalToneCurve16(Curves[i], WhitePointIn[i]); 584 } 585 } 586 else { 587 for (i=0; i < nIns; i++) 588 WhiteIn[i] = WhitePointIn[i]; 589 } 590 591 // If any post-linearization, we need to find how is represented white before the curve, do 592 // a reverse interpolation in this case. 593 if (PostLin) { 594 595 cmsToneCurve** Curves = _cmsStageGetPtrToCurveSet(PostLin); 596 597 for (i=0; i < nOuts; i++) { 598 599 cmsToneCurve* InversePostLin = cmsReverseToneCurve(Curves[i]); 600 if (InversePostLin == NULL) { 601 WhiteOut[i] = WhitePointOut[i]; 602 603 } else { 604 605 WhiteOut[i] = cmsEvalToneCurve16(InversePostLin, WhitePointOut[i]); 606 cmsFreeToneCurve(InversePostLin); 607 } 608 } 609 } 610 else { 611 for (i=0; i < nOuts; i++) 612 WhiteOut[i] = WhitePointOut[i]; 613 } 614 615 // Ok, proceed with patching. May fail and we don't care if it fails 616 PatchLUT(CLUT, WhiteIn, WhiteOut, nOuts, nIns); 617 618 return TRUE; 619 } 620 621 // ----------------------------------------------------------------------------------------------------------------------------------------------- 622 // This function creates simple LUT from complex ones. The generated LUT has an optional set of 623 // prelinearization curves, a CLUT of nGridPoints and optional postlinearization tables. 624 // These curves have to exist in the original LUT in order to be used in the simplified output. 625 // Caller may also use the flags to allow this feature. 626 // LUTS with all curves will be simplified to a single curve. Parametric curves are lost. 627 // This function should be used on 16-bits LUTS only, as floating point losses precision when simplified 628 // ----------------------------------------------------------------------------------------------------------------------------------------------- 629 630 static 631 cmsBool OptimizeByResampling(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags) 632 { 633 cmsPipeline* Src = NULL; 634 cmsPipeline* Dest = NULL; 635 cmsStage* mpe; 636 cmsStage* CLUT; 637 cmsStage *KeepPreLin = NULL, *KeepPostLin = NULL; 638 int nGridPoints; 639 cmsColorSpaceSignature ColorSpace, OutputColorSpace; 640 cmsStage *NewPreLin = NULL; 641 cmsStage *NewPostLin = NULL; 642 _cmsStageCLutData* DataCLUT; 643 cmsToneCurve** DataSetIn; 644 cmsToneCurve** DataSetOut; 645 Prelin16Data* p16; 646 647 // This is a loosy optimization! does not apply in floating-point cases 648 if (_cmsFormatterIsFloat(*InputFormat) || _cmsFormatterIsFloat(*OutputFormat)) return FALSE; 649 650 ColorSpace = _cmsICCcolorSpace(T_COLORSPACE(*InputFormat)); 651 OutputColorSpace = _cmsICCcolorSpace(T_COLORSPACE(*OutputFormat)); 652 nGridPoints = _cmsReasonableGridpointsByColorspace(ColorSpace, *dwFlags); 653 654 // For empty LUTs, 2 points are enough 655 if (cmsPipelineStageCount(*Lut) == 0) 656 nGridPoints = 2; 657 658 Src = *Lut; 659 660 // Named color pipelines cannot be optimized either 661 for (mpe = cmsPipelineGetPtrToFirstStage(Src); 662 mpe != NULL; 663 mpe = cmsStageNext(mpe)) { 664 if (cmsStageType(mpe) == cmsSigNamedColorElemType) return FALSE; 665 } 666 667 // Allocate an empty LUT 668 Dest = cmsPipelineAlloc(Src ->ContextID, Src ->InputChannels, Src ->OutputChannels); 669 if (!Dest) return FALSE; 670 671 // Prelinearization tables are kept unless indicated by flags 672 if (*dwFlags & cmsFLAGS_CLUT_PRE_LINEARIZATION) { 673 674 // Get a pointer to the prelinearization element 675 cmsStage* PreLin = cmsPipelineGetPtrToFirstStage(Src); 676 677 // Check if suitable 678 if (PreLin && PreLin ->Type == cmsSigCurveSetElemType) { 679 680 // Maybe this is a linear tram, so we can avoid the whole stuff 681 if (!AllCurvesAreLinear(PreLin)) { 682 683 // All seems ok, proceed. 684 NewPreLin = cmsStageDup(PreLin); 685 if(!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, NewPreLin)) 686 goto Error; 687 688 // Remove prelinearization. Since we have duplicated the curve 689 // in destination LUT, the sampling shoud be applied after this stage. 690 cmsPipelineUnlinkStage(Src, cmsAT_BEGIN, &KeepPreLin); 691 } 692 } 693 } 694 695 // Allocate the CLUT 696 CLUT = cmsStageAllocCLut16bit(Src ->ContextID, nGridPoints, Src ->InputChannels, Src->OutputChannels, NULL); 697 if (CLUT == NULL) goto Error; 698 699 // Add the CLUT to the destination LUT 700 if (!cmsPipelineInsertStage(Dest, cmsAT_END, CLUT)) { 701 goto Error; 702 } 703 704 // Postlinearization tables are kept unless indicated by flags 705 if (*dwFlags & cmsFLAGS_CLUT_POST_LINEARIZATION) { 706 707 // Get a pointer to the postlinearization if present 708 cmsStage* PostLin = cmsPipelineGetPtrToLastStage(Src); 709 710 // Check if suitable 711 if (PostLin && cmsStageType(PostLin) == cmsSigCurveSetElemType) { 712 713 // Maybe this is a linear tram, so we can avoid the whole stuff 714 if (!AllCurvesAreLinear(PostLin)) { 715 716 // All seems ok, proceed. 717 NewPostLin = cmsStageDup(PostLin); 718 if (!cmsPipelineInsertStage(Dest, cmsAT_END, NewPostLin)) 719 goto Error; 720 721 // In destination LUT, the sampling shoud be applied after this stage. 722 cmsPipelineUnlinkStage(Src, cmsAT_END, &KeepPostLin); 723 } 724 } 725 } 726 727 // Now its time to do the sampling. We have to ignore pre/post linearization 728 // The source LUT whithout pre/post curves is passed as parameter. 729 if (!cmsStageSampleCLut16bit(CLUT, XFormSampler16, (void*) Src, 0)) { 730 Error: 731 // Ops, something went wrong, Restore stages 732 if (KeepPreLin != NULL) { 733 if (!cmsPipelineInsertStage(Src, cmsAT_BEGIN, KeepPreLin)) { 734 _cmsAssert(0); // This never happens 735 } 736 } 737 if (KeepPostLin != NULL) { 738 if (!cmsPipelineInsertStage(Src, cmsAT_END, KeepPostLin)) { 739 _cmsAssert(0); // This never happens 740 } 741 } 742 cmsPipelineFree(Dest); 743 return FALSE; 744 } 745 746 // Done. 747 748 if (KeepPreLin != NULL) cmsStageFree(KeepPreLin); 749 if (KeepPostLin != NULL) cmsStageFree(KeepPostLin); 750 cmsPipelineFree(Src); 751 752 DataCLUT = (_cmsStageCLutData*) CLUT ->Data; 753 754 if (NewPreLin == NULL) DataSetIn = NULL; 755 else DataSetIn = ((_cmsStageToneCurvesData*) NewPreLin ->Data) ->TheCurves; 756 757 if (NewPostLin == NULL) DataSetOut = NULL; 758 else DataSetOut = ((_cmsStageToneCurvesData*) NewPostLin ->Data) ->TheCurves; 759 760 761 if (DataSetIn == NULL && DataSetOut == NULL) { 762 763 _cmsPipelineSetOptimizationParameters(Dest, (_cmsOPTeval16Fn) DataCLUT->Params->Interpolation.Lerp16, DataCLUT->Params, NULL, NULL); 764 } 765 else { 766 767 p16 = PrelinOpt16alloc(Dest ->ContextID, 768 DataCLUT ->Params, 769 Dest ->InputChannels, 770 DataSetIn, 771 Dest ->OutputChannels, 772 DataSetOut); 773 774 _cmsPipelineSetOptimizationParameters(Dest, PrelinEval16, (void*) p16, PrelinOpt16free, Prelin16dup); 775 } 776 777 778 // Don't fix white on absolute colorimetric 779 if (Intent == INTENT_ABSOLUTE_COLORIMETRIC) 780 *dwFlags |= cmsFLAGS_NOWHITEONWHITEFIXUP; 781 782 if (!(*dwFlags & cmsFLAGS_NOWHITEONWHITEFIXUP)) { 783 784 FixWhiteMisalignment(Dest, ColorSpace, OutputColorSpace); 785 } 786 787 *Lut = Dest; 788 return TRUE; 789 790 cmsUNUSED_PARAMETER(Intent); 791 } 792 793 794 // ----------------------------------------------------------------------------------------------------------------------------------------------- 795 // Fixes the gamma balancing of transform. This is described in my paper "Prelinearization Stages on 796 // Color-Management Application-Specific Integrated Circuits (ASICs)" presented at NIP24. It only works 797 // for RGB transforms. See the paper for more details 798 // ----------------------------------------------------------------------------------------------------------------------------------------------- 799 800 801 // Normalize endpoints by slope limiting max and min. This assures endpoints as well. 802 // Descending curves are handled as well. 803 static 804 void SlopeLimiting(cmsToneCurve* g) 805 { 806 int BeginVal, EndVal; 807 int AtBegin = (int) floor((cmsFloat64Number) g ->nEntries * 0.02 + 0.5); // Cutoff at 2% 808 int AtEnd = g ->nEntries - AtBegin - 1; // And 98% 809 cmsFloat64Number Val, Slope, beta; 810 int i; 811 812 if (cmsIsToneCurveDescending(g)) { 813 BeginVal = 0xffff; EndVal = 0; 814 } 815 else { 816 BeginVal = 0; EndVal = 0xffff; 817 } 818 819 // Compute slope and offset for begin of curve 820 Val = g ->Table16[AtBegin]; 821 Slope = (Val - BeginVal) / AtBegin; 822 beta = Val - Slope * AtBegin; 823 824 for (i=0; i < AtBegin; i++) 825 g ->Table16[i] = _cmsQuickSaturateWord(i * Slope + beta); 826 827 // Compute slope and offset for the end 828 Val = g ->Table16[AtEnd]; 829 Slope = (EndVal - Val) / AtBegin; // AtBegin holds the X interval, which is same in both cases 830 beta = Val - Slope * AtEnd; 831 832 for (i = AtEnd; i < (int) g ->nEntries; i++) 833 g ->Table16[i] = _cmsQuickSaturateWord(i * Slope + beta); 834 } 835 836 837 // Precomputes tables for 8-bit on input devicelink. 838 static 839 Prelin8Data* PrelinOpt8alloc(cmsContext ContextID, const cmsInterpParams* p, cmsToneCurve* G[3]) 840 { 841 int i; 842 cmsUInt16Number Input[3]; 843 cmsS15Fixed16Number v1, v2, v3; 844 Prelin8Data* p8; 845 846 p8 = (Prelin8Data*)_cmsMallocZero(ContextID, sizeof(Prelin8Data)); 847 if (p8 == NULL) return NULL; 848 849 // Since this only works for 8 bit input, values comes always as x * 257, 850 // we can safely take msb byte (x << 8 + x) 851 852 for (i=0; i < 256; i++) { 853 854 if (G != NULL) { 855 856 // Get 16-bit representation 857 Input[0] = cmsEvalToneCurve16(G[0], FROM_8_TO_16(i)); 858 Input[1] = cmsEvalToneCurve16(G[1], FROM_8_TO_16(i)); 859 Input[2] = cmsEvalToneCurve16(G[2], FROM_8_TO_16(i)); 860 } 861 else { 862 Input[0] = FROM_8_TO_16(i); 863 Input[1] = FROM_8_TO_16(i); 864 Input[2] = FROM_8_TO_16(i); 865 } 866 867 868 // Move to 0..1.0 in fixed domain 869 v1 = _cmsToFixedDomain(Input[0] * p -> Domain[0]); 870 v2 = _cmsToFixedDomain(Input[1] * p -> Domain[1]); 871 v3 = _cmsToFixedDomain(Input[2] * p -> Domain[2]); 872 873 // Store the precalculated table of nodes 874 p8 ->X0[i] = (p->opta[2] * FIXED_TO_INT(v1)); 875 p8 ->Y0[i] = (p->opta[1] * FIXED_TO_INT(v2)); 876 p8 ->Z0[i] = (p->opta[0] * FIXED_TO_INT(v3)); 877 878 // Store the precalculated table of offsets 879 p8 ->rx[i] = (cmsUInt16Number) FIXED_REST_TO_INT(v1); 880 p8 ->ry[i] = (cmsUInt16Number) FIXED_REST_TO_INT(v2); 881 p8 ->rz[i] = (cmsUInt16Number) FIXED_REST_TO_INT(v3); 882 } 883 884 p8 ->ContextID = ContextID; 885 p8 ->p = p; 886 887 return p8; 888 } 889 890 static 891 void Prelin8free(cmsContext ContextID, void* ptr) 892 { 893 _cmsFree(ContextID, ptr); 894 } 895 896 static 897 void* Prelin8dup(cmsContext ContextID, const void* ptr) 898 { 899 return _cmsDupMem(ContextID, ptr, sizeof(Prelin8Data)); 900 } 901 902 903 904 // A optimized interpolation for 8-bit input. 905 #define DENS(i,j,k) (LutTable[(i)+(j)+(k)+OutChan]) 906 static 907 void PrelinEval8(register const cmsUInt16Number Input[], 908 register cmsUInt16Number Output[], 909 register const void* D) 910 { 911 912 cmsUInt8Number r, g, b; 913 cmsS15Fixed16Number rx, ry, rz; 914 cmsS15Fixed16Number c0, c1, c2, c3, Rest; 915 int OutChan; 916 register cmsS15Fixed16Number X0, X1, Y0, Y1, Z0, Z1; 917 Prelin8Data* p8 = (Prelin8Data*) D; 918 register const cmsInterpParams* p = p8 ->p; 919 int TotalOut = p -> nOutputs; 920 const cmsUInt16Number* LutTable = (const cmsUInt16Number*) p->Table; 921 922 r = Input[0] >> 8; 923 g = Input[1] >> 8; 924 b = Input[2] >> 8; 925 926 X0 = X1 = p8->X0[r]; 927 Y0 = Y1 = p8->Y0[g]; 928 Z0 = Z1 = p8->Z0[b]; 929 930 rx = p8 ->rx[r]; 931 ry = p8 ->ry[g]; 932 rz = p8 ->rz[b]; 933 934 X1 = X0 + ((rx == 0) ? 0 : p ->opta[2]); 935 Y1 = Y0 + ((ry == 0) ? 0 : p ->opta[1]); 936 Z1 = Z0 + ((rz == 0) ? 0 : p ->opta[0]); 937 938 939 // These are the 6 Tetrahedral 940 for (OutChan=0; OutChan < TotalOut; OutChan++) { 941 942 c0 = DENS(X0, Y0, Z0); 943 944 if (rx >= ry && ry >= rz) 945 { 946 c1 = DENS(X1, Y0, Z0) - c0; 947 c2 = DENS(X1, Y1, Z0) - DENS(X1, Y0, Z0); 948 c3 = DENS(X1, Y1, Z1) - DENS(X1, Y1, Z0); 949 } 950 else 951 if (rx >= rz && rz >= ry) 952 { 953 c1 = DENS(X1, Y0, Z0) - c0; 954 c2 = DENS(X1, Y1, Z1) - DENS(X1, Y0, Z1); 955 c3 = DENS(X1, Y0, Z1) - DENS(X1, Y0, Z0); 956 } 957 else 958 if (rz >= rx && rx >= ry) 959 { 960 c1 = DENS(X1, Y0, Z1) - DENS(X0, Y0, Z1); 961 c2 = DENS(X1, Y1, Z1) - DENS(X1, Y0, Z1); 962 c3 = DENS(X0, Y0, Z1) - c0; 963 } 964 else 965 if (ry >= rx && rx >= rz) 966 { 967 c1 = DENS(X1, Y1, Z0) - DENS(X0, Y1, Z0); 968 c2 = DENS(X0, Y1, Z0) - c0; 969 c3 = DENS(X1, Y1, Z1) - DENS(X1, Y1, Z0); 970 } 971 else 972 if (ry >= rz && rz >= rx) 973 { 974 c1 = DENS(X1, Y1, Z1) - DENS(X0, Y1, Z1); 975 c2 = DENS(X0, Y1, Z0) - c0; 976 c3 = DENS(X0, Y1, Z1) - DENS(X0, Y1, Z0); 977 } 978 else 979 if (rz >= ry && ry >= rx) 980 { 981 c1 = DENS(X1, Y1, Z1) - DENS(X0, Y1, Z1); 982 c2 = DENS(X0, Y1, Z1) - DENS(X0, Y0, Z1); 983 c3 = DENS(X0, Y0, Z1) - c0; 984 } 985 else { 986 c1 = c2 = c3 = 0; 987 } 988 989 990 Rest = c1 * rx + c2 * ry + c3 * rz + 0x8001; 991 Output[OutChan] = (cmsUInt16Number)c0 + ((Rest + (Rest>>16))>>16); 992 993 } 994 } 995 996 #undef DENS 997 998 999 // Curves that contain wide empty areas are not optimizeable 1000 static 1001 cmsBool IsDegenerated(const cmsToneCurve* g) 1002 { 1003 int i, Zeros = 0, Poles = 0; 1004 int nEntries = g ->nEntries; 1005 1006 for (i=0; i < nEntries; i++) { 1007 1008 if (g ->Table16[i] == 0x0000) Zeros++; 1009 if (g ->Table16[i] == 0xffff) Poles++; 1010 } 1011 1012 if (Zeros == 1 && Poles == 1) return FALSE; // For linear tables 1013 if (Zeros > (nEntries / 20)) return TRUE; // Degenerated, many zeros 1014 if (Poles > (nEntries / 20)) return TRUE; // Degenerated, many poles 1015 1016 return FALSE; 1017 } 1018 1019 // -------------------------------------------------------------------------------------------------------------- 1020 // We need xput over here 1021 1022 static 1023 cmsBool OptimizeByComputingLinearization(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags) 1024 { 1025 cmsPipeline* OriginalLut; 1026 int nGridPoints; 1027 cmsToneCurve *Trans[cmsMAXCHANNELS], *TransReverse[cmsMAXCHANNELS]; 1028 cmsUInt32Number t, i; 1029 cmsFloat32Number v, In[cmsMAXCHANNELS], Out[cmsMAXCHANNELS]; 1030 cmsBool lIsSuitable, lIsLinear; 1031 cmsPipeline* OptimizedLUT = NULL, *LutPlusCurves = NULL; 1032 cmsStage* OptimizedCLUTmpe; 1033 cmsColorSpaceSignature ColorSpace, OutputColorSpace; 1034 cmsStage* OptimizedPrelinMpe; 1035 cmsStage* mpe; 1036 cmsToneCurve** OptimizedPrelinCurves; 1037 _cmsStageCLutData* OptimizedPrelinCLUT; 1038 1039 1040 // This is a loosy optimization! does not apply in floating-point cases 1041 if (_cmsFormatterIsFloat(*InputFormat) || _cmsFormatterIsFloat(*OutputFormat)) return FALSE; 1042 1043 // Only on chunky RGB 1044 if (T_COLORSPACE(*InputFormat) != PT_RGB) return FALSE; 1045 if (T_PLANAR(*InputFormat)) return FALSE; 1046 1047 if (T_COLORSPACE(*OutputFormat) != PT_RGB) return FALSE; 1048 if (T_PLANAR(*OutputFormat)) return FALSE; 1049 1050 // On 16 bits, user has to specify the feature 1051 if (!_cmsFormatterIs8bit(*InputFormat)) { 1052 if (!(*dwFlags & cmsFLAGS_CLUT_PRE_LINEARIZATION)) return FALSE; 1053 } 1054 1055 OriginalLut = *Lut; 1056 1057 // Named color pipelines cannot be optimized either 1058 for (mpe = cmsPipelineGetPtrToFirstStage(OriginalLut); 1059 mpe != NULL; 1060 mpe = cmsStageNext(mpe)) { 1061 if (cmsStageType(mpe) == cmsSigNamedColorElemType) return FALSE; 1062 } 1063 1064 ColorSpace = _cmsICCcolorSpace(T_COLORSPACE(*InputFormat)); 1065 OutputColorSpace = _cmsICCcolorSpace(T_COLORSPACE(*OutputFormat)); 1066 nGridPoints = _cmsReasonableGridpointsByColorspace(ColorSpace, *dwFlags); 1067 1068 // Empty gamma containers 1069 memset(Trans, 0, sizeof(Trans)); 1070 memset(TransReverse, 0, sizeof(TransReverse)); 1071 1072 // If the last stage of the original lut are curves, and those curves are 1073 // degenerated, it is likely the transform is squeezing and clipping 1074 // the output from previous CLUT. We cannot optimize this case 1075 { 1076 cmsStage* last = cmsPipelineGetPtrToLastStage(OriginalLut); 1077 1078 if (cmsStageType(last) == cmsSigCurveSetElemType) { 1079 1080 _cmsStageToneCurvesData* Data = (_cmsStageToneCurvesData*)cmsStageData(last); 1081 for (i = 0; i < Data->nCurves; i++) { 1082 if (IsDegenerated(Data->TheCurves[i])) 1083 goto Error; 1084 } 1085 } 1086 } 1087 1088 for (t = 0; t < OriginalLut ->InputChannels; t++) { 1089 Trans[t] = cmsBuildTabulatedToneCurve16(OriginalLut ->ContextID, PRELINEARIZATION_POINTS, NULL); 1090 if (Trans[t] == NULL) goto Error; 1091 } 1092 1093 // Populate the curves 1094 for (i=0; i < PRELINEARIZATION_POINTS; i++) { 1095 1096 v = (cmsFloat32Number) ((cmsFloat64Number) i / (PRELINEARIZATION_POINTS - 1)); 1097 1098 // Feed input with a gray ramp 1099 for (t=0; t < OriginalLut ->InputChannels; t++) 1100 In[t] = v; 1101 1102 // Evaluate the gray value 1103 cmsPipelineEvalFloat(In, Out, OriginalLut); 1104 1105 // Store result in curve 1106 for (t=0; t < OriginalLut ->InputChannels; t++) 1107 Trans[t] ->Table16[i] = _cmsQuickSaturateWord(Out[t] * 65535.0); 1108 } 1109 1110 // Slope-limit the obtained curves 1111 for (t = 0; t < OriginalLut ->InputChannels; t++) 1112 SlopeLimiting(Trans[t]); 1113 1114 // Check for validity 1115 lIsSuitable = TRUE; 1116 lIsLinear = TRUE; 1117 for (t=0; (lIsSuitable && (t < OriginalLut ->InputChannels)); t++) { 1118 1119 // Exclude if already linear 1120 if (!cmsIsToneCurveLinear(Trans[t])) 1121 lIsLinear = FALSE; 1122 1123 // Exclude if non-monotonic 1124 if (!cmsIsToneCurveMonotonic(Trans[t])) 1125 lIsSuitable = FALSE; 1126 1127 if (IsDegenerated(Trans[t])) 1128 lIsSuitable = FALSE; 1129 } 1130 1131 // If it is not suitable, just quit 1132 if (!lIsSuitable) goto Error; 1133 1134 // Invert curves if possible 1135 for (t = 0; t < OriginalLut ->InputChannels; t++) { 1136 TransReverse[t] = cmsReverseToneCurveEx(PRELINEARIZATION_POINTS, Trans[t]); 1137 if (TransReverse[t] == NULL) goto Error; 1138 } 1139 1140 // Now inset the reversed curves at the begin of transform 1141 LutPlusCurves = cmsPipelineDup(OriginalLut); 1142 if (LutPlusCurves == NULL) goto Error; 1143 1144 if (!cmsPipelineInsertStage(LutPlusCurves, cmsAT_BEGIN, cmsStageAllocToneCurves(OriginalLut ->ContextID, OriginalLut ->InputChannels, TransReverse))) 1145 goto Error; 1146 1147 // Create the result LUT 1148 OptimizedLUT = cmsPipelineAlloc(OriginalLut ->ContextID, OriginalLut ->InputChannels, OriginalLut ->OutputChannels); 1149 if (OptimizedLUT == NULL) goto Error; 1150 1151 OptimizedPrelinMpe = cmsStageAllocToneCurves(OriginalLut ->ContextID, OriginalLut ->InputChannels, Trans); 1152 1153 // Create and insert the curves at the beginning 1154 if (!cmsPipelineInsertStage(OptimizedLUT, cmsAT_BEGIN, OptimizedPrelinMpe)) 1155 goto Error; 1156 1157 // Allocate the CLUT for result 1158 OptimizedCLUTmpe = cmsStageAllocCLut16bit(OriginalLut ->ContextID, nGridPoints, OriginalLut ->InputChannels, OriginalLut ->OutputChannels, NULL); 1159 1160 // Add the CLUT to the destination LUT 1161 if (!cmsPipelineInsertStage(OptimizedLUT, cmsAT_END, OptimizedCLUTmpe)) 1162 goto Error; 1163 1164 // Resample the LUT 1165 if (!cmsStageSampleCLut16bit(OptimizedCLUTmpe, XFormSampler16, (void*) LutPlusCurves, 0)) goto Error; 1166 1167 // Free resources 1168 for (t = 0; t < OriginalLut ->InputChannels; t++) { 1169 1170 if (Trans[t]) cmsFreeToneCurve(Trans[t]); 1171 if (TransReverse[t]) cmsFreeToneCurve(TransReverse[t]); 1172 } 1173 1174 cmsPipelineFree(LutPlusCurves); 1175 1176 1177 OptimizedPrelinCurves = _cmsStageGetPtrToCurveSet(OptimizedPrelinMpe); 1178 OptimizedPrelinCLUT = (_cmsStageCLutData*) OptimizedCLUTmpe ->Data; 1179 1180 // Set the evaluator if 8-bit 1181 if (_cmsFormatterIs8bit(*InputFormat)) { 1182 1183 Prelin8Data* p8 = PrelinOpt8alloc(OptimizedLUT ->ContextID, 1184 OptimizedPrelinCLUT ->Params, 1185 OptimizedPrelinCurves); 1186 if (p8 == NULL) return FALSE; 1187 1188 _cmsPipelineSetOptimizationParameters(OptimizedLUT, PrelinEval8, (void*) p8, Prelin8free, Prelin8dup); 1189 1190 } 1191 else 1192 { 1193 Prelin16Data* p16 = PrelinOpt16alloc(OptimizedLUT ->ContextID, 1194 OptimizedPrelinCLUT ->Params, 1195 3, OptimizedPrelinCurves, 3, NULL); 1196 if (p16 == NULL) return FALSE; 1197 1198 _cmsPipelineSetOptimizationParameters(OptimizedLUT, PrelinEval16, (void*) p16, PrelinOpt16free, Prelin16dup); 1199 1200 } 1201 1202 // Don't fix white on absolute colorimetric 1203 if (Intent == INTENT_ABSOLUTE_COLORIMETRIC) 1204 *dwFlags |= cmsFLAGS_NOWHITEONWHITEFIXUP; 1205 1206 if (!(*dwFlags & cmsFLAGS_NOWHITEONWHITEFIXUP)) { 1207 1208 if (!FixWhiteMisalignment(OptimizedLUT, ColorSpace, OutputColorSpace)) { 1209 1210 return FALSE; 1211 } 1212 } 1213 1214 // And return the obtained LUT 1215 1216 cmsPipelineFree(OriginalLut); 1217 *Lut = OptimizedLUT; 1218 return TRUE; 1219 1220 Error: 1221 1222 for (t = 0; t < OriginalLut ->InputChannels; t++) { 1223 1224 if (Trans[t]) cmsFreeToneCurve(Trans[t]); 1225 if (TransReverse[t]) cmsFreeToneCurve(TransReverse[t]); 1226 } 1227 1228 if (LutPlusCurves != NULL) cmsPipelineFree(LutPlusCurves); 1229 if (OptimizedLUT != NULL) cmsPipelineFree(OptimizedLUT); 1230 1231 return FALSE; 1232 1233 cmsUNUSED_PARAMETER(Intent); 1234 } 1235 1236 1237 // Curves optimizer ------------------------------------------------------------------------------------------------------------------ 1238 1239 static 1240 void CurvesFree(cmsContext ContextID, void* ptr) 1241 { 1242 Curves16Data* Data = (Curves16Data*) ptr; 1243 int i; 1244 1245 for (i=0; i < Data -> nCurves; i++) { 1246 1247 _cmsFree(ContextID, Data ->Curves[i]); 1248 } 1249 1250 _cmsFree(ContextID, Data ->Curves); 1251 _cmsFree(ContextID, ptr); 1252 } 1253 1254 static 1255 void* CurvesDup(cmsContext ContextID, const void* ptr) 1256 { 1257 Curves16Data* Data = (Curves16Data*)_cmsDupMem(ContextID, ptr, sizeof(Curves16Data)); 1258 int i; 1259 1260 if (Data == NULL) return NULL; 1261 1262 Data->Curves = (cmsUInt16Number**) _cmsDupMem(ContextID, Data->Curves, Data->nCurves * sizeof(cmsUInt16Number*)); 1263 1264 for (i=0; i < Data -> nCurves; i++) { 1265 Data->Curves[i] = (cmsUInt16Number*) _cmsDupMem(ContextID, Data->Curves[i], Data->nElements * sizeof(cmsUInt16Number)); 1266 } 1267 1268 return (void*) Data; 1269 } 1270 1271 // Precomputes tables for 8-bit on input devicelink. 1272 static 1273 Curves16Data* CurvesAlloc(cmsContext ContextID, int nCurves, int nElements, cmsToneCurve** G) 1274 { 1275 int i, j; 1276 Curves16Data* c16; 1277 1278 c16 = (Curves16Data*)_cmsMallocZero(ContextID, sizeof(Curves16Data)); 1279 if (c16 == NULL) return NULL; 1280 1281 c16 ->nCurves = nCurves; 1282 c16 ->nElements = nElements; 1283 1284 c16->Curves = (cmsUInt16Number**) _cmsCalloc(ContextID, nCurves, sizeof(cmsUInt16Number*)); 1285 if (c16 ->Curves == NULL) return NULL; 1286 1287 for (i=0; i < nCurves; i++) { 1288 1289 c16->Curves[i] = (cmsUInt16Number*) _cmsCalloc(ContextID, nElements, sizeof(cmsUInt16Number)); 1290 1291 if (c16->Curves[i] == NULL) { 1292 1293 for (j=0; j < i; j++) { 1294 _cmsFree(ContextID, c16->Curves[j]); 1295 } 1296 _cmsFree(ContextID, c16->Curves); 1297 _cmsFree(ContextID, c16); 1298 return NULL; 1299 } 1300 1301 if (nElements == 256) { 1302 1303 for (j=0; j < nElements; j++) { 1304 1305 c16 ->Curves[i][j] = cmsEvalToneCurve16(G[i], FROM_8_TO_16(j)); 1306 } 1307 } 1308 else { 1309 1310 for (j=0; j < nElements; j++) { 1311 c16 ->Curves[i][j] = cmsEvalToneCurve16(G[i], (cmsUInt16Number) j); 1312 } 1313 } 1314 } 1315 1316 return c16; 1317 } 1318 1319 static 1320 void FastEvaluateCurves8(register const cmsUInt16Number In[], 1321 register cmsUInt16Number Out[], 1322 register const void* D) 1323 { 1324 Curves16Data* Data = (Curves16Data*) D; 1325 cmsUInt8Number x; 1326 int i; 1327 1328 for (i=0; i < Data ->nCurves; i++) { 1329 1330 x = (In[i] >> 8); 1331 Out[i] = Data -> Curves[i][x]; 1332 } 1333 } 1334 1335 1336 static 1337 void FastEvaluateCurves16(register const cmsUInt16Number In[], 1338 register cmsUInt16Number Out[], 1339 register const void* D) 1340 { 1341 Curves16Data* Data = (Curves16Data*) D; 1342 int i; 1343 1344 for (i=0; i < Data ->nCurves; i++) { 1345 Out[i] = Data -> Curves[i][In[i]]; 1346 } 1347 } 1348 1349 1350 static 1351 void FastIdentity16(register const cmsUInt16Number In[], 1352 register cmsUInt16Number Out[], 1353 register const void* D) 1354 { 1355 cmsPipeline* Lut = (cmsPipeline*) D; 1356 cmsUInt32Number i; 1357 1358 for (i=0; i < Lut ->InputChannels; i++) { 1359 Out[i] = In[i]; 1360 } 1361 } 1362 1363 1364 // If the target LUT holds only curves, the optimization procedure is to join all those 1365 // curves together. That only works on curves and does not work on matrices. 1366 static 1367 cmsBool OptimizeByJoiningCurves(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags) 1368 { 1369 cmsToneCurve** GammaTables = NULL; 1370 cmsFloat32Number InFloat[cmsMAXCHANNELS], OutFloat[cmsMAXCHANNELS]; 1371 cmsUInt32Number i, j; 1372 cmsPipeline* Src = *Lut; 1373 cmsPipeline* Dest = NULL; 1374 cmsStage* mpe; 1375 cmsStage* ObtainedCurves = NULL; 1376 1377 1378 // This is a loosy optimization! does not apply in floating-point cases 1379 if (_cmsFormatterIsFloat(*InputFormat) || _cmsFormatterIsFloat(*OutputFormat)) return FALSE; 1380 1381 // Only curves in this LUT? 1382 for (mpe = cmsPipelineGetPtrToFirstStage(Src); 1383 mpe != NULL; 1384 mpe = cmsStageNext(mpe)) { 1385 if (cmsStageType(mpe) != cmsSigCurveSetElemType) return FALSE; 1386 } 1387 1388 // Allocate an empty LUT 1389 Dest = cmsPipelineAlloc(Src ->ContextID, Src ->InputChannels, Src ->OutputChannels); 1390 if (Dest == NULL) return FALSE; 1391 1392 // Create target curves 1393 GammaTables = (cmsToneCurve**) _cmsCalloc(Src ->ContextID, Src ->InputChannels, sizeof(cmsToneCurve*)); 1394 if (GammaTables == NULL) goto Error; 1395 1396 for (i=0; i < Src ->InputChannels; i++) { 1397 GammaTables[i] = cmsBuildTabulatedToneCurve16(Src ->ContextID, PRELINEARIZATION_POINTS, NULL); 1398 if (GammaTables[i] == NULL) goto Error; 1399 } 1400 1401 // Compute 16 bit result by using floating point 1402 for (i=0; i < PRELINEARIZATION_POINTS; i++) { 1403 1404 for (j=0; j < Src ->InputChannels; j++) 1405 InFloat[j] = (cmsFloat32Number) ((cmsFloat64Number) i / (PRELINEARIZATION_POINTS - 1)); 1406 1407 cmsPipelineEvalFloat(InFloat, OutFloat, Src); 1408 1409 for (j=0; j < Src ->InputChannels; j++) 1410 GammaTables[j] -> Table16[i] = _cmsQuickSaturateWord(OutFloat[j] * 65535.0); 1411 } 1412 1413 ObtainedCurves = cmsStageAllocToneCurves(Src ->ContextID, Src ->InputChannels, GammaTables); 1414 if (ObtainedCurves == NULL) goto Error; 1415 1416 for (i=0; i < Src ->InputChannels; i++) { 1417 cmsFreeToneCurve(GammaTables[i]); 1418 GammaTables[i] = NULL; 1419 } 1420 1421 if (GammaTables != NULL) { 1422 _cmsFree(Src->ContextID, GammaTables); 1423 GammaTables = NULL; 1424 } 1425 1426 // Maybe the curves are linear at the end 1427 if (!AllCurvesAreLinear(ObtainedCurves)) { 1428 1429 if (!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, ObtainedCurves)) 1430 goto Error; 1431 1432 // If the curves are to be applied in 8 bits, we can save memory 1433 if (_cmsFormatterIs8bit(*InputFormat)) { 1434 1435 _cmsStageToneCurvesData* Data = (_cmsStageToneCurvesData*) ObtainedCurves ->Data; 1436 Curves16Data* c16 = CurvesAlloc(Dest ->ContextID, Data ->nCurves, 256, Data ->TheCurves); 1437 1438 if (c16 == NULL) goto Error; 1439 *dwFlags |= cmsFLAGS_NOCACHE; 1440 _cmsPipelineSetOptimizationParameters(Dest, FastEvaluateCurves8, c16, CurvesFree, CurvesDup); 1441 1442 } 1443 else { 1444 1445 _cmsStageToneCurvesData* Data = (_cmsStageToneCurvesData*) cmsStageData(ObtainedCurves); 1446 Curves16Data* c16 = CurvesAlloc(Dest ->ContextID, Data ->nCurves, 65536, Data ->TheCurves); 1447 1448 if (c16 == NULL) goto Error; 1449 *dwFlags |= cmsFLAGS_NOCACHE; 1450 _cmsPipelineSetOptimizationParameters(Dest, FastEvaluateCurves16, c16, CurvesFree, CurvesDup); 1451 } 1452 } 1453 else { 1454 1455 // LUT optimizes to nothing. Set the identity LUT 1456 cmsStageFree(ObtainedCurves); 1457 1458 if (!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, cmsStageAllocIdentity(Dest ->ContextID, Src ->InputChannels))) 1459 goto Error; 1460 1461 *dwFlags |= cmsFLAGS_NOCACHE; 1462 _cmsPipelineSetOptimizationParameters(Dest, FastIdentity16, (void*) Dest, NULL, NULL); 1463 } 1464 1465 // We are done. 1466 cmsPipelineFree(Src); 1467 *Lut = Dest; 1468 return TRUE; 1469 1470 Error: 1471 1472 if (ObtainedCurves != NULL) cmsStageFree(ObtainedCurves); 1473 if (GammaTables != NULL) { 1474 for (i=0; i < Src ->InputChannels; i++) { 1475 if (GammaTables[i] != NULL) cmsFreeToneCurve(GammaTables[i]); 1476 } 1477 1478 _cmsFree(Src ->ContextID, GammaTables); 1479 } 1480 1481 if (Dest != NULL) cmsPipelineFree(Dest); 1482 return FALSE; 1483 1484 cmsUNUSED_PARAMETER(Intent); 1485 cmsUNUSED_PARAMETER(InputFormat); 1486 cmsUNUSED_PARAMETER(OutputFormat); 1487 cmsUNUSED_PARAMETER(dwFlags); 1488 } 1489 1490 // ------------------------------------------------------------------------------------------------------------------------------------- 1491 // LUT is Shaper - Matrix - Matrix - Shaper, which is very frequent when combining two matrix-shaper profiles 1492 1493 1494 static 1495 void FreeMatShaper(cmsContext ContextID, void* Data) 1496 { 1497 if (Data != NULL) _cmsFree(ContextID, Data); 1498 } 1499 1500 static 1501 void* DupMatShaper(cmsContext ContextID, const void* Data) 1502 { 1503 return _cmsDupMem(ContextID, Data, sizeof(MatShaper8Data)); 1504 } 1505 1506 1507 // A fast matrix-shaper evaluator for 8 bits. This is a bit ticky since I'm using 1.14 signed fixed point 1508 // to accomplish some performance. Actually it takes 256x3 16 bits tables and 16385 x 3 tables of 8 bits, 1509 // in total about 50K, and the performance boost is huge! 1510 static 1511 void MatShaperEval16(register const cmsUInt16Number In[], 1512 register cmsUInt16Number Out[], 1513 register const void* D) 1514 { 1515 MatShaper8Data* p = (MatShaper8Data*) D; 1516 cmsS1Fixed14Number l1, l2, l3, r, g, b; 1517 cmsUInt32Number ri, gi, bi; 1518 1519 // In this case (and only in this case!) we can use this simplification since 1520 // In[] is assured to come from a 8 bit number. (a << 8 | a) 1521 ri = In[0] & 0xFF; 1522 gi = In[1] & 0xFF; 1523 bi = In[2] & 0xFF; 1524 1525 // Across first shaper, which also converts to 1.14 fixed point 1526 r = p->Shaper1R[ri]; 1527 g = p->Shaper1G[gi]; 1528 b = p->Shaper1B[bi]; 1529 1530 // Evaluate the matrix in 1.14 fixed point 1531 l1 = (p->Mat[0][0] * r + p->Mat[0][1] * g + p->Mat[0][2] * b + p->Off[0] + 0x2000) >> 14; 1532 l2 = (p->Mat[1][0] * r + p->Mat[1][1] * g + p->Mat[1][2] * b + p->Off[1] + 0x2000) >> 14; 1533 l3 = (p->Mat[2][0] * r + p->Mat[2][1] * g + p->Mat[2][2] * b + p->Off[2] + 0x2000) >> 14; 1534 1535 // Now we have to clip to 0..1.0 range 1536 ri = (l1 < 0) ? 0 : ((l1 > 16384) ? 16384 : l1); 1537 gi = (l2 < 0) ? 0 : ((l2 > 16384) ? 16384 : l2); 1538 bi = (l3 < 0) ? 0 : ((l3 > 16384) ? 16384 : l3); 1539 1540 // And across second shaper, 1541 Out[0] = p->Shaper2R[ri]; 1542 Out[1] = p->Shaper2G[gi]; 1543 Out[2] = p->Shaper2B[bi]; 1544 1545 } 1546 1547 // This table converts from 8 bits to 1.14 after applying the curve 1548 static 1549 cmsBool FillFirstShaper(cmsS1Fixed14Number* Table, cmsToneCurve* Curve) 1550 { 1551 int i; 1552 cmsFloat32Number R, y; 1553 1554 for (i=0; i < 256; i++) { 1555 1556 R = (cmsFloat32Number) (i / 255.0); 1557 y = cmsEvalToneCurveFloat(Curve, R); 1558 if (isinf(y)) 1559 return FALSE; 1560 1561 Table[i] = DOUBLE_TO_1FIXED14(y); 1562 } 1563 return TRUE; 1564 } 1565 1566 // This table converts form 1.14 (being 0x4000 the last entry) to 8 bits after applying the curve 1567 static 1568 cmsBool FillSecondShaper(cmsUInt16Number* Table, cmsToneCurve* Curve, cmsBool Is8BitsOutput) 1569 { 1570 int i; 1571 cmsFloat32Number R, Val; 1572 1573 for (i=0; i < 16385; i++) { 1574 1575 R = (cmsFloat32Number) (i / 16384.0); 1576 Val = cmsEvalToneCurveFloat(Curve, R); // Val comes 0..1.0 1577 if (isinf(Val)) 1578 return FALSE; 1579 1580 if (Is8BitsOutput) { 1581 1582 // If 8 bits output, we can optimize further by computing the / 257 part. 1583 // first we compute the resulting byte and then we store the byte times 1584 // 257. This quantization allows to round very quick by doing a >> 8, but 1585 // since the low byte is always equal to msb, we can do a & 0xff and this works! 1586 cmsUInt16Number w = _cmsQuickSaturateWord(Val * 65535.0); 1587 cmsUInt8Number b = FROM_16_TO_8(w); 1588 1589 Table[i] = FROM_8_TO_16(b); 1590 } 1591 else Table[i] = _cmsQuickSaturateWord(Val * 65535.0); 1592 } 1593 return TRUE; 1594 } 1595 1596 // Compute the matrix-shaper structure 1597 static 1598 cmsBool SetMatShaper(cmsPipeline* Dest, cmsToneCurve* Curve1[3], cmsMAT3* Mat, cmsVEC3* Off, cmsToneCurve* Curve2[3], cmsUInt32Number* OutputFormat) 1599 { 1600 MatShaper8Data* p; 1601 int i, j; 1602 cmsBool Is8Bits = _cmsFormatterIs8bit(*OutputFormat); 1603 1604 // Allocate a big chuck of memory to store precomputed tables 1605 p = (MatShaper8Data*) _cmsMalloc(Dest ->ContextID, sizeof(MatShaper8Data)); 1606 if (p == NULL) return FALSE; 1607 1608 p -> ContextID = Dest -> ContextID; 1609 1610 // Precompute tables 1611 if (!FillFirstShaper(p ->Shaper1R, Curve1[0])) 1612 goto Error; 1613 if (!FillFirstShaper(p ->Shaper1G, Curve1[1])) 1614 goto Error; 1615 if (!FillFirstShaper(p ->Shaper1B, Curve1[2])) 1616 goto Error; 1617 1618 if (!FillSecondShaper(p ->Shaper2R, Curve2[0], Is8Bits)) 1619 goto Error; 1620 if (!FillSecondShaper(p ->Shaper2G, Curve2[1], Is8Bits)) 1621 goto Error; 1622 if (!FillSecondShaper(p ->Shaper2B, Curve2[2], Is8Bits)) 1623 goto Error; 1624 1625 // Convert matrix to nFixed14. Note that those values may take more than 16 bits as 1626 for (i=0; i < 3; i++) { 1627 for (j=0; j < 3; j++) { 1628 p ->Mat[i][j] = DOUBLE_TO_1FIXED14(Mat->v[i].n[j]); 1629 } 1630 } 1631 1632 for (i=0; i < 3; i++) { 1633 1634 if (Off == NULL) { 1635 p ->Off[i] = 0; 1636 } 1637 else { 1638 p ->Off[i] = DOUBLE_TO_1FIXED14(Off->n[i]); 1639 } 1640 } 1641 1642 // Mark as optimized for faster formatter 1643 if (Is8Bits) 1644 *OutputFormat |= OPTIMIZED_SH(1); 1645 1646 // Fill function pointers 1647 _cmsPipelineSetOptimizationParameters(Dest, MatShaperEval16, (void*) p, FreeMatShaper, DupMatShaper); 1648 return TRUE; 1649 Error: 1650 _cmsFree(Dest->ContextID, p); 1651 return FALSE; 1652 } 1653 1654 // 8 bits on input allows matrix-shaper boot up to 25 Mpixels per second on RGB. That's fast! 1655 static 1656 cmsBool OptimizeMatrixShaper(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags) 1657 { 1658 cmsStage* Curve1, *Curve2; 1659 cmsStage* Matrix1, *Matrix2; 1660 cmsMAT3 res; 1661 cmsBool IdentityMat; 1662 cmsPipeline* Dest, *Src; 1663 cmsFloat64Number* Offset; 1664 1665 // Only works on RGB to RGB 1666 if (T_CHANNELS(*InputFormat) != 3 || T_CHANNELS(*OutputFormat) != 3) return FALSE; 1667 1668 // Only works on 8 bit input 1669 if (!_cmsFormatterIs8bit(*InputFormat)) return FALSE; 1670 1671 // Seems suitable, proceed 1672 Src = *Lut; 1673 1674 // Check for: 1675 // 1676 // shaper-matrix-matrix-shaper 1677 // shaper-matrix-shaper 1678 // 1679 // Both of those constructs are possible (first because abs. colorimetric). 1680 // additionally, In the first case, the input matrix offset should be zero. 1681 1682 IdentityMat = FALSE; 1683 if (cmsPipelineCheckAndRetreiveStages(Src, 4, 1684 cmsSigCurveSetElemType, cmsSigMatrixElemType, cmsSigMatrixElemType, cmsSigCurveSetElemType, 1685 &Curve1, &Matrix1, &Matrix2, &Curve2)) { 1686 1687 // Get both matrices 1688 _cmsStageMatrixData* Data1 = (_cmsStageMatrixData*)cmsStageData(Matrix1); 1689 _cmsStageMatrixData* Data2 = (_cmsStageMatrixData*)cmsStageData(Matrix2); 1690 1691 // Input offset should be zero 1692 if (Data1->Offset != NULL) return FALSE; 1693 1694 // Multiply both matrices to get the result 1695 _cmsMAT3per(&res, (cmsMAT3*)Data2->Double, (cmsMAT3*)Data1->Double); 1696 1697 // Only 2nd matrix has offset, or it is zero 1698 Offset = Data2->Offset; 1699 1700 // Now the result is in res + Data2 -> Offset. Maybe is a plain identity? 1701 if (_cmsMAT3isIdentity(&res) && Offset == NULL) { 1702 1703 // We can get rid of full matrix 1704 IdentityMat = TRUE; 1705 } 1706 1707 } 1708 else { 1709 1710 if (cmsPipelineCheckAndRetreiveStages(Src, 3, 1711 cmsSigCurveSetElemType, cmsSigMatrixElemType, cmsSigCurveSetElemType, 1712 &Curve1, &Matrix1, &Curve2)) { 1713 1714 _cmsStageMatrixData* Data = (_cmsStageMatrixData*)cmsStageData(Matrix1); 1715 1716 // Copy the matrix to our result 1717 memcpy(&res, Data->Double, sizeof(res)); 1718 1719 // Preserve the Odffset (may be NULL as a zero offset) 1720 Offset = Data->Offset; 1721 1722 if (_cmsMAT3isIdentity(&res) && Offset == NULL) { 1723 1724 // We can get rid of full matrix 1725 IdentityMat = TRUE; 1726 } 1727 } 1728 else 1729 return FALSE; // Not optimizeable this time 1730 1731 } 1732 1733 // Allocate an empty LUT 1734 Dest = cmsPipelineAlloc(Src ->ContextID, Src ->InputChannels, Src ->OutputChannels); 1735 if (!Dest) return FALSE; 1736 1737 // Assamble the new LUT 1738 if (!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, cmsStageDup(Curve1))) 1739 goto Error; 1740 1741 if (!IdentityMat) { 1742 1743 if (!cmsPipelineInsertStage(Dest, cmsAT_END, cmsStageAllocMatrix(Dest->ContextID, 3, 3, (const cmsFloat64Number*)&res, Offset))) 1744 goto Error; 1745 } 1746 1747 if (!cmsPipelineInsertStage(Dest, cmsAT_END, cmsStageDup(Curve2))) 1748 goto Error; 1749 1750 // If identity on matrix, we can further optimize the curves, so call the join curves routine 1751 if (IdentityMat) { 1752 1753 OptimizeByJoiningCurves(&Dest, Intent, InputFormat, OutputFormat, dwFlags); 1754 } 1755 else { 1756 _cmsStageToneCurvesData* mpeC1 = (_cmsStageToneCurvesData*) cmsStageData(Curve1); 1757 _cmsStageToneCurvesData* mpeC2 = (_cmsStageToneCurvesData*) cmsStageData(Curve2); 1758 1759 // In this particular optimization, cache does not help as it takes more time to deal with 1760 // the cache that with the pixel handling 1761 *dwFlags |= cmsFLAGS_NOCACHE; 1762 1763 // Setup the optimizarion routines 1764 if (!SetMatShaper(Dest, mpeC1 ->TheCurves, &res, (cmsVEC3*) Offset, mpeC2->TheCurves, OutputFormat)) 1765 goto Error; 1766 } 1767 1768 cmsPipelineFree(Src); 1769 *Lut = Dest; 1770 return TRUE; 1771 Error: 1772 // Leave Src unchanged 1773 cmsPipelineFree(Dest); 1774 return FALSE; 1775 } 1776 1777 1778 // ------------------------------------------------------------------------------------------------------------------------------------- 1779 // Optimization plug-ins 1780 1781 // List of optimizations 1782 typedef struct _cmsOptimizationCollection_st { 1783 1784 _cmsOPToptimizeFn OptimizePtr; 1785 1786 struct _cmsOptimizationCollection_st *Next; 1787 1788 } _cmsOptimizationCollection; 1789 1790 1791 // The built-in list. We currently implement 4 types of optimizations. Joining of curves, matrix-shaper, linearization and resampling 1792 static _cmsOptimizationCollection DefaultOptimization[] = { 1793 1794 { OptimizeByJoiningCurves, &DefaultOptimization[1] }, 1795 { OptimizeMatrixShaper, &DefaultOptimization[2] }, 1796 { OptimizeByComputingLinearization, &DefaultOptimization[3] }, 1797 { OptimizeByResampling, NULL } 1798 }; 1799 1800 // The linked list head 1801 _cmsOptimizationPluginChunkType _cmsOptimizationPluginChunk = { NULL }; 1802 1803 1804 // Duplicates the zone of memory used by the plug-in in the new context 1805 static 1806 void DupPluginOptimizationList(struct _cmsContext_struct* ctx, 1807 const struct _cmsContext_struct* src) 1808 { 1809 _cmsOptimizationPluginChunkType newHead = { NULL }; 1810 _cmsOptimizationCollection* entry; 1811 _cmsOptimizationCollection* Anterior = NULL; 1812 _cmsOptimizationPluginChunkType* head = (_cmsOptimizationPluginChunkType*) src->chunks[OptimizationPlugin]; 1813 1814 _cmsAssert(ctx != NULL); 1815 _cmsAssert(head != NULL); 1816 1817 // Walk the list copying all nodes 1818 for (entry = head->OptimizationCollection; 1819 entry != NULL; 1820 entry = entry ->Next) { 1821 1822 _cmsOptimizationCollection *newEntry = ( _cmsOptimizationCollection *) _cmsSubAllocDup(ctx ->MemPool, entry, sizeof(_cmsOptimizationCollection)); 1823 1824 if (newEntry == NULL) 1825 return; 1826 1827 // We want to keep the linked list order, so this is a little bit tricky 1828 newEntry -> Next = NULL; 1829 if (Anterior) 1830 Anterior -> Next = newEntry; 1831 1832 Anterior = newEntry; 1833 1834 if (newHead.OptimizationCollection == NULL) 1835 newHead.OptimizationCollection = newEntry; 1836 } 1837 1838 ctx ->chunks[OptimizationPlugin] = _cmsSubAllocDup(ctx->MemPool, &newHead, sizeof(_cmsOptimizationPluginChunkType)); 1839 } 1840 1841 void _cmsAllocOptimizationPluginChunk(struct _cmsContext_struct* ctx, 1842 const struct _cmsContext_struct* src) 1843 { 1844 if (src != NULL) { 1845 1846 // Copy all linked list 1847 DupPluginOptimizationList(ctx, src); 1848 } 1849 else { 1850 static _cmsOptimizationPluginChunkType OptimizationPluginChunkType = { NULL }; 1851 ctx ->chunks[OptimizationPlugin] = _cmsSubAllocDup(ctx ->MemPool, &OptimizationPluginChunkType, sizeof(_cmsOptimizationPluginChunkType)); 1852 } 1853 } 1854 1855 1856 // Register new ways to optimize 1857 cmsBool _cmsRegisterOptimizationPlugin(cmsContext ContextID, cmsPluginBase* Data) 1858 { 1859 cmsPluginOptimization* Plugin = (cmsPluginOptimization*) Data; 1860 _cmsOptimizationPluginChunkType* ctx = ( _cmsOptimizationPluginChunkType*) _cmsContextGetClientChunk(ContextID, OptimizationPlugin); 1861 _cmsOptimizationCollection* fl; 1862 1863 if (Data == NULL) { 1864 1865 ctx->OptimizationCollection = NULL; 1866 return TRUE; 1867 } 1868 1869 // Optimizer callback is required 1870 if (Plugin ->OptimizePtr == NULL) return FALSE; 1871 1872 fl = (_cmsOptimizationCollection*) _cmsPluginMalloc(ContextID, sizeof(_cmsOptimizationCollection)); 1873 if (fl == NULL) return FALSE; 1874 1875 // Copy the parameters 1876 fl ->OptimizePtr = Plugin ->OptimizePtr; 1877 1878 // Keep linked list 1879 fl ->Next = ctx->OptimizationCollection; 1880 1881 // Set the head 1882 ctx ->OptimizationCollection = fl; 1883 1884 // All is ok 1885 return TRUE; 1886 } 1887 1888 // The entry point for LUT optimization 1889 cmsBool _cmsOptimizePipeline(cmsContext ContextID, 1890 cmsPipeline** PtrLut, 1891 int Intent, 1892 cmsUInt32Number* InputFormat, 1893 cmsUInt32Number* OutputFormat, 1894 cmsUInt32Number* dwFlags) 1895 { 1896 _cmsOptimizationPluginChunkType* ctx = ( _cmsOptimizationPluginChunkType*) _cmsContextGetClientChunk(ContextID, OptimizationPlugin); 1897 _cmsOptimizationCollection* Opts; 1898 cmsBool AnySuccess = FALSE; 1899 1900 // A CLUT is being asked, so force this specific optimization 1901 if (*dwFlags & cmsFLAGS_FORCE_CLUT) { 1902 1903 PreOptimize(*PtrLut); 1904 return OptimizeByResampling(PtrLut, Intent, InputFormat, OutputFormat, dwFlags); 1905 } 1906 1907 // Anything to optimize? 1908 if ((*PtrLut) ->Elements == NULL) { 1909 _cmsPipelineSetOptimizationParameters(*PtrLut, FastIdentity16, (void*) *PtrLut, NULL, NULL); 1910 return TRUE; 1911 } 1912 1913 // Try to get rid of identities and trivial conversions. 1914 AnySuccess = PreOptimize(*PtrLut); 1915 1916 // After removal do we end with an identity? 1917 if ((*PtrLut) ->Elements == NULL) { 1918 _cmsPipelineSetOptimizationParameters(*PtrLut, FastIdentity16, (void*) *PtrLut, NULL, NULL); 1919 return TRUE; 1920 } 1921 1922 // Do not optimize, keep all precision 1923 if (*dwFlags & cmsFLAGS_NOOPTIMIZE) 1924 return FALSE; 1925 1926 // Try plug-in optimizations 1927 for (Opts = ctx->OptimizationCollection; 1928 Opts != NULL; 1929 Opts = Opts ->Next) { 1930 1931 // If one schema succeeded, we are done 1932 if (Opts ->OptimizePtr(PtrLut, Intent, InputFormat, OutputFormat, dwFlags)) { 1933 1934 return TRUE; // Optimized! 1935 } 1936 } 1937 1938 // Try built-in optimizations 1939 for (Opts = DefaultOptimization; 1940 Opts != NULL; 1941 Opts = Opts ->Next) { 1942 1943 if (Opts ->OptimizePtr(PtrLut, Intent, InputFormat, OutputFormat, dwFlags)) { 1944 1945 return TRUE; 1946 } 1947 } 1948 1949 // Only simple optimizations succeeded 1950 return AnySuccess; 1951 } 1952 1953 1954 1955