1/* 2 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 3 * 4 * This code is free software; you can redistribute it and/or modify it 5 * under the terms of the GNU General Public License version 2 only, as 6 * published by the Free Software Foundation. Oracle designates this 7 * particular file as subject to the "Classpath" exception as provided 8 * by Oracle in the LICENSE file that accompanied this code. 9 * 10 * This code is distributed in the hope that it will be useful, but WITHOUT 11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 13 * version 2 for more details (a copy is included in the LICENSE file that 14 * accompanied this code). 15 * 16 * You should have received a copy of the GNU General Public License version 17 * 2 along with this work; if not, write to the Free Software Foundation, 18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 19 * 20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 21 * or visit www.oracle.com if you need additional information or have any 22 * questions. 23 */ 24 25// This file is available under and governed by the GNU General Public 26// License version 2 only, as published by the Free Software Foundation. 27// However, the following notice accompanied the original version of this 28// file: 29// 30//--------------------------------------------------------------------------------- 31// 32// Little Color Management System 33// Copyright (c) 1998-2016 Marti Maria Saguer 34// 35// Permission is hereby granted, free of charge, to any person obtaining 36// a copy of this software and associated documentation files (the "Software"), 37// to deal in the Software without restriction, including without limitation 38// the rights to use, copy, modify, merge, publish, distribute, sublicense, 39// and/or sell copies of the Software, and to permit persons to whom the Software 40// is furnished to do so, subject to the following conditions: 41// 42// The above copyright notice and this permission notice shall be included in 43// all copies or substantial portions of the Software. 44// 45// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 46// EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO 47// THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 48// NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE 49// LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION 50// OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION 51// WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. 52// 53//--------------------------------------------------------------------------------- 54// 55 56#include "lcms2_internal.h" 57 58 59//---------------------------------------------------------------------------------- 60 61// Optimization for 8 bits, Shaper-CLUT (3 inputs only) 62typedef struct { 63 64 cmsContext ContextID; 65 66 const cmsInterpParams* p; // Tetrahedrical interpolation parameters. This is a not-owned pointer. 67 68 cmsUInt16Number rx[256], ry[256], rz[256]; 69 cmsUInt32Number X0[256], Y0[256], Z0[256]; // Precomputed nodes and offsets for 8-bit input data 70 71 72} Prelin8Data; 73 74 75// Generic optimization for 16 bits Shaper-CLUT-Shaper (any inputs) 76typedef struct { 77 78 cmsContext ContextID; 79 80 // Number of channels 81 int nInputs; 82 int nOutputs; 83 84 _cmsInterpFn16 EvalCurveIn16[MAX_INPUT_DIMENSIONS]; // The maximum number of input channels is known in advance 85 cmsInterpParams* ParamsCurveIn16[MAX_INPUT_DIMENSIONS]; 86 87 _cmsInterpFn16 EvalCLUT; // The evaluator for 3D grid 88 const cmsInterpParams* CLUTparams; // (not-owned pointer) 89 90 91 _cmsInterpFn16* EvalCurveOut16; // Points to an array of curve evaluators in 16 bits (not-owned pointer) 92 cmsInterpParams** ParamsCurveOut16; // Points to an array of references to interpolation params (not-owned pointer) 93 94 95} Prelin16Data; 96 97 98// Optimization for matrix-shaper in 8 bits. Numbers are operated in n.14 signed, tables are stored in 1.14 fixed 99 100typedef cmsInt32Number cmsS1Fixed14Number; // Note that this may hold more than 16 bits! 101 102#define DOUBLE_TO_1FIXED14(x) ((cmsS1Fixed14Number) floor((x) * 16384.0 + 0.5)) 103 104typedef struct { 105 106 cmsContext ContextID; 107 108 cmsS1Fixed14Number Shaper1R[256]; // from 0..255 to 1.14 (0.0...1.0) 109 cmsS1Fixed14Number Shaper1G[256]; 110 cmsS1Fixed14Number Shaper1B[256]; 111 112 cmsS1Fixed14Number Mat[3][3]; // n.14 to n.14 (needs a saturation after that) 113 cmsS1Fixed14Number Off[3]; 114 115 cmsUInt16Number Shaper2R[16385]; // 1.14 to 0..255 116 cmsUInt16Number Shaper2G[16385]; 117 cmsUInt16Number Shaper2B[16385]; 118 119} MatShaper8Data; 120 121// Curves, optimization is shared between 8 and 16 bits 122typedef struct { 123 124 cmsContext ContextID; 125 126 int nCurves; // Number of curves 127 int nElements; // Elements in curves 128 cmsUInt16Number** Curves; // Points to a dynamically allocated array 129 130} Curves16Data; 131 132 133// Simple optimizations ---------------------------------------------------------------------------------------------------------- 134 135 136// Remove an element in linked chain 137static 138void _RemoveElement(cmsStage** head) 139{ 140 cmsStage* mpe = *head; 141 cmsStage* next = mpe ->Next; 142 *head = next; 143 cmsStageFree(mpe); 144} 145 146// Remove all identities in chain. Note that pt actually is a double pointer to the element that holds the pointer. 147static 148cmsBool _Remove1Op(cmsPipeline* Lut, cmsStageSignature UnaryOp) 149{ 150 cmsStage** pt = &Lut ->Elements; 151 cmsBool AnyOpt = FALSE; 152 153 while (*pt != NULL) { 154 155 if ((*pt) ->Implements == UnaryOp) { 156 _RemoveElement(pt); 157 AnyOpt = TRUE; 158 } 159 else 160 pt = &((*pt) -> Next); 161 } 162 163 return AnyOpt; 164} 165 166// Same, but only if two adjacent elements are found 167static 168cmsBool _Remove2Op(cmsPipeline* Lut, cmsStageSignature Op1, cmsStageSignature Op2) 169{ 170 cmsStage** pt1; 171 cmsStage** pt2; 172 cmsBool AnyOpt = FALSE; 173 174 pt1 = &Lut ->Elements; 175 if (*pt1 == NULL) return AnyOpt; 176 177 while (*pt1 != NULL) { 178 179 pt2 = &((*pt1) -> Next); 180 if (*pt2 == NULL) return AnyOpt; 181 182 if ((*pt1) ->Implements == Op1 && (*pt2) ->Implements == Op2) { 183 _RemoveElement(pt2); 184 _RemoveElement(pt1); 185 AnyOpt = TRUE; 186 } 187 else 188 pt1 = &((*pt1) -> Next); 189 } 190 191 return AnyOpt; 192} 193 194 195static 196cmsBool CloseEnoughFloat(cmsFloat64Number a, cmsFloat64Number b) 197{ 198 return fabs(b - a) < 0.00001f; 199} 200 201static 202cmsBool isFloatMatrixIdentity(const cmsMAT3* a) 203{ 204 cmsMAT3 Identity; 205 int i, j; 206 207 _cmsMAT3identity(&Identity); 208 209 for (i = 0; i < 3; i++) 210 for (j = 0; j < 3; j++) 211 if (!CloseEnoughFloat(a->v[i].n[j], Identity.v[i].n[j])) return FALSE; 212 213 return TRUE; 214} 215// if two adjacent matrices are found, multiply them. 216static 217cmsBool _MultiplyMatrix(cmsPipeline* Lut) 218{ 219 cmsStage** pt1; 220 cmsStage** pt2; 221 cmsStage* chain; 222 cmsBool AnyOpt = FALSE; 223 224 pt1 = &Lut->Elements; 225 if (*pt1 == NULL) return AnyOpt; 226 227 while (*pt1 != NULL) { 228 229 pt2 = &((*pt1)->Next); 230 if (*pt2 == NULL) return AnyOpt; 231 232 if ((*pt1)->Implements == cmsSigMatrixElemType && (*pt2)->Implements == cmsSigMatrixElemType) { 233 234 // Get both matrices 235 _cmsStageMatrixData* m1 = (_cmsStageMatrixData*) cmsStageData(*pt1); 236 _cmsStageMatrixData* m2 = (_cmsStageMatrixData*) cmsStageData(*pt2); 237 cmsMAT3 res; 238 239 // Input offset and output offset should be zero to use this optimization 240 if (m1->Offset != NULL || m2 ->Offset != NULL || 241 cmsStageInputChannels(*pt1) != 3 || cmsStageOutputChannels(*pt1) != 3 || 242 cmsStageInputChannels(*pt2) != 3 || cmsStageOutputChannels(*pt2) != 3) 243 return FALSE; 244 245 // Multiply both matrices to get the result 246 _cmsMAT3per(&res, (cmsMAT3*)m2->Double, (cmsMAT3*)m1->Double); 247 248 // Get the next in chain afer the matrices 249 chain = (*pt2)->Next; 250 251 // Remove both matrices 252 _RemoveElement(pt2); 253 _RemoveElement(pt1); 254 255 // Now what if the result is a plain identity? 256 if (!isFloatMatrixIdentity(&res)) { 257 258 // We can not get rid of full matrix 259 cmsStage* Multmat = cmsStageAllocMatrix(Lut->ContextID, 3, 3, (const cmsFloat64Number*) &res, NULL); 260 if (Multmat == NULL) return FALSE; // Should never happen 261 262 // Recover the chain 263 Multmat->Next = chain; 264 *pt1 = Multmat; 265 } 266 267 AnyOpt = TRUE; 268 } 269 else 270 pt1 = &((*pt1)->Next); 271 } 272 273 return AnyOpt; 274} 275 276 277// Preoptimize just gets rif of no-ops coming paired. Conversion from v2 to v4 followed 278// by a v4 to v2 and vice-versa. The elements are then discarded. 279static 280cmsBool PreOptimize(cmsPipeline* Lut) 281{ 282 cmsBool AnyOpt = FALSE, Opt; 283 284 do { 285 286 Opt = FALSE; 287 288 // Remove all identities 289 Opt |= _Remove1Op(Lut, cmsSigIdentityElemType); 290 291 // Remove XYZ2Lab followed by Lab2XYZ 292 Opt |= _Remove2Op(Lut, cmsSigXYZ2LabElemType, cmsSigLab2XYZElemType); 293 294 // Remove Lab2XYZ followed by XYZ2Lab 295 Opt |= _Remove2Op(Lut, cmsSigLab2XYZElemType, cmsSigXYZ2LabElemType); 296 297 // Remove V4 to V2 followed by V2 to V4 298 Opt |= _Remove2Op(Lut, cmsSigLabV4toV2, cmsSigLabV2toV4); 299 300 // Remove V2 to V4 followed by V4 to V2 301 Opt |= _Remove2Op(Lut, cmsSigLabV2toV4, cmsSigLabV4toV2); 302 303 // Remove float pcs Lab conversions 304 Opt |= _Remove2Op(Lut, cmsSigLab2FloatPCS, cmsSigFloatPCS2Lab); 305 306 // Remove float pcs Lab conversions 307 Opt |= _Remove2Op(Lut, cmsSigXYZ2FloatPCS, cmsSigFloatPCS2XYZ); 308 309 // Simplify matrix. 310 Opt |= _MultiplyMatrix(Lut); 311 312 if (Opt) AnyOpt = TRUE; 313 314 } while (Opt); 315 316 return AnyOpt; 317} 318 319static 320void Eval16nop1D(register const cmsUInt16Number Input[], 321 register cmsUInt16Number Output[], 322 register const struct _cms_interp_struc* p) 323{ 324 Output[0] = Input[0]; 325 326 cmsUNUSED_PARAMETER(p); 327} 328 329static 330void PrelinEval16(register const cmsUInt16Number Input[], 331 register cmsUInt16Number Output[], 332 register const void* D) 333{ 334 Prelin16Data* p16 = (Prelin16Data*) D; 335 cmsUInt16Number StageABC[MAX_INPUT_DIMENSIONS]; 336 cmsUInt16Number StageDEF[cmsMAXCHANNELS]; 337 int i; 338 339 for (i=0; i < p16 ->nInputs; i++) { 340 341 p16 ->EvalCurveIn16[i](&Input[i], &StageABC[i], p16 ->ParamsCurveIn16[i]); 342 } 343 344 p16 ->EvalCLUT(StageABC, StageDEF, p16 ->CLUTparams); 345 346 for (i=0; i < p16 ->nOutputs; i++) { 347 348 p16 ->EvalCurveOut16[i](&StageDEF[i], &Output[i], p16 ->ParamsCurveOut16[i]); 349 } 350} 351 352 353static 354void PrelinOpt16free(cmsContext ContextID, void* ptr) 355{ 356 Prelin16Data* p16 = (Prelin16Data*) ptr; 357 358 _cmsFree(ContextID, p16 ->EvalCurveOut16); 359 _cmsFree(ContextID, p16 ->ParamsCurveOut16); 360 361 _cmsFree(ContextID, p16); 362} 363 364static 365void* Prelin16dup(cmsContext ContextID, const void* ptr) 366{ 367 Prelin16Data* p16 = (Prelin16Data*) ptr; 368 Prelin16Data* Duped = (Prelin16Data*) _cmsDupMem(ContextID, p16, sizeof(Prelin16Data)); 369 370 if (Duped == NULL) return NULL; 371 372 Duped->EvalCurveOut16 = (_cmsInterpFn16*) _cmsDupMem(ContextID, p16->EvalCurveOut16, p16->nOutputs * sizeof(_cmsInterpFn16)); 373 Duped->ParamsCurveOut16 = (cmsInterpParams**)_cmsDupMem(ContextID, p16->ParamsCurveOut16, p16->nOutputs * sizeof(cmsInterpParams*)); 374 375 return Duped; 376} 377 378 379static 380Prelin16Data* PrelinOpt16alloc(cmsContext ContextID, 381 const cmsInterpParams* ColorMap, 382 int nInputs, cmsToneCurve** In, 383 int nOutputs, cmsToneCurve** Out ) 384{ 385 int i; 386 Prelin16Data* p16 = (Prelin16Data*)_cmsMallocZero(ContextID, sizeof(Prelin16Data)); 387 if (p16 == NULL) return NULL; 388 389 p16 ->nInputs = nInputs; 390 p16 -> nOutputs = nOutputs; 391 392 393 for (i=0; i < nInputs; i++) { 394 395 if (In == NULL) { 396 p16 -> ParamsCurveIn16[i] = NULL; 397 p16 -> EvalCurveIn16[i] = Eval16nop1D; 398 399 } 400 else { 401 p16 -> ParamsCurveIn16[i] = In[i] ->InterpParams; 402 p16 -> EvalCurveIn16[i] = p16 ->ParamsCurveIn16[i]->Interpolation.Lerp16; 403 } 404 } 405 406 p16 ->CLUTparams = ColorMap; 407 p16 ->EvalCLUT = ColorMap ->Interpolation.Lerp16; 408 409 410 p16 -> EvalCurveOut16 = (_cmsInterpFn16*) _cmsCalloc(ContextID, nOutputs, sizeof(_cmsInterpFn16)); 411 p16 -> ParamsCurveOut16 = (cmsInterpParams**) _cmsCalloc(ContextID, nOutputs, sizeof(cmsInterpParams* )); 412 413 for (i=0; i < nOutputs; i++) { 414 415 if (Out == NULL) { 416 p16 ->ParamsCurveOut16[i] = NULL; 417 p16 -> EvalCurveOut16[i] = Eval16nop1D; 418 } 419 else { 420 421 p16 ->ParamsCurveOut16[i] = Out[i] ->InterpParams; 422 p16 -> EvalCurveOut16[i] = p16 ->ParamsCurveOut16[i]->Interpolation.Lerp16; 423 } 424 } 425 426 return p16; 427} 428 429 430 431// Resampling --------------------------------------------------------------------------------- 432 433#define PRELINEARIZATION_POINTS 4096 434 435// Sampler implemented by another LUT. This is a clean way to precalculate the devicelink 3D CLUT for 436// almost any transform. We use floating point precision and then convert from floating point to 16 bits. 437static 438int XFormSampler16(register const cmsUInt16Number In[], register cmsUInt16Number Out[], register void* Cargo) 439{ 440 cmsPipeline* Lut = (cmsPipeline*) Cargo; 441 cmsFloat32Number InFloat[cmsMAXCHANNELS], OutFloat[cmsMAXCHANNELS]; 442 cmsUInt32Number i; 443 444 _cmsAssert(Lut -> InputChannels < cmsMAXCHANNELS); 445 _cmsAssert(Lut -> OutputChannels < cmsMAXCHANNELS); 446 447 // From 16 bit to floating point 448 for (i=0; i < Lut ->InputChannels; i++) 449 InFloat[i] = (cmsFloat32Number) (In[i] / 65535.0); 450 451 // Evaluate in floating point 452 cmsPipelineEvalFloat(InFloat, OutFloat, Lut); 453 454 // Back to 16 bits representation 455 for (i=0; i < Lut ->OutputChannels; i++) 456 Out[i] = _cmsQuickSaturateWord(OutFloat[i] * 65535.0); 457 458 // Always succeed 459 return TRUE; 460} 461 462// Try to see if the curves of a given MPE are linear 463static 464cmsBool AllCurvesAreLinear(cmsStage* mpe) 465{ 466 cmsToneCurve** Curves; 467 cmsUInt32Number i, n; 468 469 Curves = _cmsStageGetPtrToCurveSet(mpe); 470 if (Curves == NULL) return FALSE; 471 472 n = cmsStageOutputChannels(mpe); 473 474 for (i=0; i < n; i++) { 475 if (!cmsIsToneCurveLinear(Curves[i])) return FALSE; 476 } 477 478 return TRUE; 479} 480 481// This function replaces a specific node placed in "At" by the "Value" numbers. Its purpose 482// is to fix scum dot on broken profiles/transforms. Works on 1, 3 and 4 channels 483static 484cmsBool PatchLUT(cmsStage* CLUT, cmsUInt16Number At[], cmsUInt16Number Value[], 485 int nChannelsOut, int nChannelsIn) 486{ 487 _cmsStageCLutData* Grid = (_cmsStageCLutData*) CLUT ->Data; 488 cmsInterpParams* p16 = Grid ->Params; 489 cmsFloat64Number px, py, pz, pw; 490 int x0, y0, z0, w0; 491 int i, index; 492 493 if (CLUT -> Type != cmsSigCLutElemType) { 494 cmsSignalError(CLUT->ContextID, cmsERROR_INTERNAL, "(internal) Attempt to PatchLUT on non-lut stage"); 495 return FALSE; 496 } 497 498 if (nChannelsIn == 4) { 499 500 px = ((cmsFloat64Number) At[0] * (p16->Domain[0])) / 65535.0; 501 py = ((cmsFloat64Number) At[1] * (p16->Domain[1])) / 65535.0; 502 pz = ((cmsFloat64Number) At[2] * (p16->Domain[2])) / 65535.0; 503 pw = ((cmsFloat64Number) At[3] * (p16->Domain[3])) / 65535.0; 504 505 x0 = (int) floor(px); 506 y0 = (int) floor(py); 507 z0 = (int) floor(pz); 508 w0 = (int) floor(pw); 509 510 if (((px - x0) != 0) || 511 ((py - y0) != 0) || 512 ((pz - z0) != 0) || 513 ((pw - w0) != 0)) return FALSE; // Not on exact node 514 515 index = p16 -> opta[3] * x0 + 516 p16 -> opta[2] * y0 + 517 p16 -> opta[1] * z0 + 518 p16 -> opta[0] * w0; 519 } 520 else 521 if (nChannelsIn == 3) { 522 523 px = ((cmsFloat64Number) At[0] * (p16->Domain[0])) / 65535.0; 524 py = ((cmsFloat64Number) At[1] * (p16->Domain[1])) / 65535.0; 525 pz = ((cmsFloat64Number) At[2] * (p16->Domain[2])) / 65535.0; 526 527 x0 = (int) floor(px); 528 y0 = (int) floor(py); 529 z0 = (int) floor(pz); 530 531 if (((px - x0) != 0) || 532 ((py - y0) != 0) || 533 ((pz - z0) != 0)) return FALSE; // Not on exact node 534 535 index = p16 -> opta[2] * x0 + 536 p16 -> opta[1] * y0 + 537 p16 -> opta[0] * z0; 538 } 539 else 540 if (nChannelsIn == 1) { 541 542 px = ((cmsFloat64Number) At[0] * (p16->Domain[0])) / 65535.0; 543 544 x0 = (int) floor(px); 545 546 if (((px - x0) != 0)) return FALSE; // Not on exact node 547 548 index = p16 -> opta[0] * x0; 549 } 550 else { 551 cmsSignalError(CLUT->ContextID, cmsERROR_INTERNAL, "(internal) %d Channels are not supported on PatchLUT", nChannelsIn); 552 return FALSE; 553 } 554 555 for (i=0; i < nChannelsOut; i++) 556 Grid -> Tab.T[index + i] = Value[i]; 557 558 return TRUE; 559} 560 561// Auxiliary, to see if two values are equal or very different 562static 563cmsBool WhitesAreEqual(int n, cmsUInt16Number White1[], cmsUInt16Number White2[] ) 564{ 565 int i; 566 567 for (i=0; i < n; i++) { 568 569 if (abs(White1[i] - White2[i]) > 0xf000) return TRUE; // Values are so extremely different that the fixup should be avoided 570 if (White1[i] != White2[i]) return FALSE; 571 } 572 return TRUE; 573} 574 575 576// Locate the node for the white point and fix it to pure white in order to avoid scum dot. 577static 578cmsBool FixWhiteMisalignment(cmsPipeline* Lut, cmsColorSpaceSignature EntryColorSpace, cmsColorSpaceSignature ExitColorSpace) 579{ 580 cmsUInt16Number *WhitePointIn, *WhitePointOut; 581 cmsUInt16Number WhiteIn[cmsMAXCHANNELS], WhiteOut[cmsMAXCHANNELS], ObtainedOut[cmsMAXCHANNELS]; 582 cmsUInt32Number i, nOuts, nIns; 583 cmsStage *PreLin = NULL, *CLUT = NULL, *PostLin = NULL; 584 585 if (!_cmsEndPointsBySpace(EntryColorSpace, 586 &WhitePointIn, NULL, &nIns)) return FALSE; 587 588 if (!_cmsEndPointsBySpace(ExitColorSpace, 589 &WhitePointOut, NULL, &nOuts)) return FALSE; 590 591 // It needs to be fixed? 592 if (Lut ->InputChannels != nIns) return FALSE; 593 if (Lut ->OutputChannels != nOuts) return FALSE; 594 595 cmsPipelineEval16(WhitePointIn, ObtainedOut, Lut); 596 597 if (WhitesAreEqual(nOuts, WhitePointOut, ObtainedOut)) return TRUE; // whites already match 598 599 // Check if the LUT comes as Prelin, CLUT or Postlin. We allow all combinations 600 if (!cmsPipelineCheckAndRetreiveStages(Lut, 3, cmsSigCurveSetElemType, cmsSigCLutElemType, cmsSigCurveSetElemType, &PreLin, &CLUT, &PostLin)) 601 if (!cmsPipelineCheckAndRetreiveStages(Lut, 2, cmsSigCurveSetElemType, cmsSigCLutElemType, &PreLin, &CLUT)) 602 if (!cmsPipelineCheckAndRetreiveStages(Lut, 2, cmsSigCLutElemType, cmsSigCurveSetElemType, &CLUT, &PostLin)) 603 if (!cmsPipelineCheckAndRetreiveStages(Lut, 1, cmsSigCLutElemType, &CLUT)) 604 return FALSE; 605 606 // We need to interpolate white points of both, pre and post curves 607 if (PreLin) { 608 609 cmsToneCurve** Curves = _cmsStageGetPtrToCurveSet(PreLin); 610 611 for (i=0; i < nIns; i++) { 612 WhiteIn[i] = cmsEvalToneCurve16(Curves[i], WhitePointIn[i]); 613 } 614 } 615 else { 616 for (i=0; i < nIns; i++) 617 WhiteIn[i] = WhitePointIn[i]; 618 } 619 620 // If any post-linearization, we need to find how is represented white before the curve, do 621 // a reverse interpolation in this case. 622 if (PostLin) { 623 624 cmsToneCurve** Curves = _cmsStageGetPtrToCurveSet(PostLin); 625 626 for (i=0; i < nOuts; i++) { 627 628 cmsToneCurve* InversePostLin = cmsReverseToneCurve(Curves[i]); 629 if (InversePostLin == NULL) { 630 WhiteOut[i] = WhitePointOut[i]; 631 632 } else { 633 634 WhiteOut[i] = cmsEvalToneCurve16(InversePostLin, WhitePointOut[i]); 635 cmsFreeToneCurve(InversePostLin); 636 } 637 } 638 } 639 else { 640 for (i=0; i < nOuts; i++) 641 WhiteOut[i] = WhitePointOut[i]; 642 } 643 644 // Ok, proceed with patching. May fail and we don't care if it fails 645 PatchLUT(CLUT, WhiteIn, WhiteOut, nOuts, nIns); 646 647 return TRUE; 648} 649 650// ----------------------------------------------------------------------------------------------------------------------------------------------- 651// This function creates simple LUT from complex ones. The generated LUT has an optional set of 652// prelinearization curves, a CLUT of nGridPoints and optional postlinearization tables. 653// These curves have to exist in the original LUT in order to be used in the simplified output. 654// Caller may also use the flags to allow this feature. 655// LUTS with all curves will be simplified to a single curve. Parametric curves are lost. 656// This function should be used on 16-bits LUTS only, as floating point losses precision when simplified 657// ----------------------------------------------------------------------------------------------------------------------------------------------- 658 659static 660cmsBool OptimizeByResampling(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags) 661{ 662 cmsPipeline* Src = NULL; 663 cmsPipeline* Dest = NULL; 664 cmsStage* mpe; 665 cmsStage* CLUT; 666 cmsStage *KeepPreLin = NULL, *KeepPostLin = NULL; 667 int nGridPoints; 668 cmsColorSpaceSignature ColorSpace, OutputColorSpace; 669 cmsStage *NewPreLin = NULL; 670 cmsStage *NewPostLin = NULL; 671 _cmsStageCLutData* DataCLUT; 672 cmsToneCurve** DataSetIn; 673 cmsToneCurve** DataSetOut; 674 Prelin16Data* p16; 675 676 // This is a loosy optimization! does not apply in floating-point cases 677 if (_cmsFormatterIsFloat(*InputFormat) || _cmsFormatterIsFloat(*OutputFormat)) return FALSE; 678 679 ColorSpace = _cmsICCcolorSpace(T_COLORSPACE(*InputFormat)); 680 OutputColorSpace = _cmsICCcolorSpace(T_COLORSPACE(*OutputFormat)); 681 nGridPoints = _cmsReasonableGridpointsByColorspace(ColorSpace, *dwFlags); 682 683 // For empty LUTs, 2 points are enough 684 if (cmsPipelineStageCount(*Lut) == 0) 685 nGridPoints = 2; 686 687 Src = *Lut; 688 689 // Named color pipelines cannot be optimized either 690 for (mpe = cmsPipelineGetPtrToFirstStage(Src); 691 mpe != NULL; 692 mpe = cmsStageNext(mpe)) { 693 if (cmsStageType(mpe) == cmsSigNamedColorElemType) return FALSE; 694 } 695 696 // Allocate an empty LUT 697 Dest = cmsPipelineAlloc(Src ->ContextID, Src ->InputChannels, Src ->OutputChannels); 698 if (!Dest) return FALSE; 699 700 // Prelinearization tables are kept unless indicated by flags 701 if (*dwFlags & cmsFLAGS_CLUT_PRE_LINEARIZATION) { 702 703 // Get a pointer to the prelinearization element 704 cmsStage* PreLin = cmsPipelineGetPtrToFirstStage(Src); 705 706 // Check if suitable 707 if (PreLin && PreLin ->Type == cmsSigCurveSetElemType) { 708 709 // Maybe this is a linear tram, so we can avoid the whole stuff 710 if (!AllCurvesAreLinear(PreLin)) { 711 712 // All seems ok, proceed. 713 NewPreLin = cmsStageDup(PreLin); 714 if(!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, NewPreLin)) 715 goto Error; 716 717 // Remove prelinearization. Since we have duplicated the curve 718 // in destination LUT, the sampling shoud be applied after this stage. 719 cmsPipelineUnlinkStage(Src, cmsAT_BEGIN, &KeepPreLin); 720 } 721 } 722 } 723 724 // Allocate the CLUT 725 CLUT = cmsStageAllocCLut16bit(Src ->ContextID, nGridPoints, Src ->InputChannels, Src->OutputChannels, NULL); 726 if (CLUT == NULL) return FALSE; 727 728 // Add the CLUT to the destination LUT 729 if (!cmsPipelineInsertStage(Dest, cmsAT_END, CLUT)) { 730 goto Error; 731 } 732 733 // Postlinearization tables are kept unless indicated by flags 734 if (*dwFlags & cmsFLAGS_CLUT_POST_LINEARIZATION) { 735 736 // Get a pointer to the postlinearization if present 737 cmsStage* PostLin = cmsPipelineGetPtrToLastStage(Src); 738 739 // Check if suitable 740 if (PostLin && cmsStageType(PostLin) == cmsSigCurveSetElemType) { 741 742 // Maybe this is a linear tram, so we can avoid the whole stuff 743 if (!AllCurvesAreLinear(PostLin)) { 744 745 // All seems ok, proceed. 746 NewPostLin = cmsStageDup(PostLin); 747 if (!cmsPipelineInsertStage(Dest, cmsAT_END, NewPostLin)) 748 goto Error; 749 750 // In destination LUT, the sampling shoud be applied after this stage. 751 cmsPipelineUnlinkStage(Src, cmsAT_END, &KeepPostLin); 752 } 753 } 754 } 755 756 // Now its time to do the sampling. We have to ignore pre/post linearization 757 // The source LUT whithout pre/post curves is passed as parameter. 758 if (!cmsStageSampleCLut16bit(CLUT, XFormSampler16, (void*) Src, 0)) { 759Error: 760 // Ops, something went wrong, Restore stages 761 if (KeepPreLin != NULL) { 762 if (!cmsPipelineInsertStage(Src, cmsAT_BEGIN, KeepPreLin)) { 763 _cmsAssert(0); // This never happens 764 } 765 } 766 if (KeepPostLin != NULL) { 767 if (!cmsPipelineInsertStage(Src, cmsAT_END, KeepPostLin)) { 768 _cmsAssert(0); // This never happens 769 } 770 } 771 cmsPipelineFree(Dest); 772 return FALSE; 773 } 774 775 // Done. 776 777 if (KeepPreLin != NULL) cmsStageFree(KeepPreLin); 778 if (KeepPostLin != NULL) cmsStageFree(KeepPostLin); 779 cmsPipelineFree(Src); 780 781 DataCLUT = (_cmsStageCLutData*) CLUT ->Data; 782 783 if (NewPreLin == NULL) DataSetIn = NULL; 784 else DataSetIn = ((_cmsStageToneCurvesData*) NewPreLin ->Data) ->TheCurves; 785 786 if (NewPostLin == NULL) DataSetOut = NULL; 787 else DataSetOut = ((_cmsStageToneCurvesData*) NewPostLin ->Data) ->TheCurves; 788 789 790 if (DataSetIn == NULL && DataSetOut == NULL) { 791 792 _cmsPipelineSetOptimizationParameters(Dest, (_cmsOPTeval16Fn) DataCLUT->Params->Interpolation.Lerp16, DataCLUT->Params, NULL, NULL); 793 } 794 else { 795 796 p16 = PrelinOpt16alloc(Dest ->ContextID, 797 DataCLUT ->Params, 798 Dest ->InputChannels, 799 DataSetIn, 800 Dest ->OutputChannels, 801 DataSetOut); 802 803 _cmsPipelineSetOptimizationParameters(Dest, PrelinEval16, (void*) p16, PrelinOpt16free, Prelin16dup); 804 } 805 806 807 // Don't fix white on absolute colorimetric 808 if (Intent == INTENT_ABSOLUTE_COLORIMETRIC) 809 *dwFlags |= cmsFLAGS_NOWHITEONWHITEFIXUP; 810 811 if (!(*dwFlags & cmsFLAGS_NOWHITEONWHITEFIXUP)) { 812 813 FixWhiteMisalignment(Dest, ColorSpace, OutputColorSpace); 814 } 815 816 *Lut = Dest; 817 return TRUE; 818 819 cmsUNUSED_PARAMETER(Intent); 820} 821 822 823// ----------------------------------------------------------------------------------------------------------------------------------------------- 824// Fixes the gamma balancing of transform. This is described in my paper "Prelinearization Stages on 825// Color-Management Application-Specific Integrated Circuits (ASICs)" presented at NIP24. It only works 826// for RGB transforms. See the paper for more details 827// ----------------------------------------------------------------------------------------------------------------------------------------------- 828 829 830// Normalize endpoints by slope limiting max and min. This assures endpoints as well. 831// Descending curves are handled as well. 832static 833void SlopeLimiting(cmsToneCurve* g) 834{ 835 int BeginVal, EndVal; 836 int AtBegin = (int) floor((cmsFloat64Number) g ->nEntries * 0.02 + 0.5); // Cutoff at 2% 837 int AtEnd = g ->nEntries - AtBegin - 1; // And 98% 838 cmsFloat64Number Val, Slope, beta; 839 int i; 840 841 if (cmsIsToneCurveDescending(g)) { 842 BeginVal = 0xffff; EndVal = 0; 843 } 844 else { 845 BeginVal = 0; EndVal = 0xffff; 846 } 847 848 // Compute slope and offset for begin of curve 849 Val = g ->Table16[AtBegin]; 850 Slope = (Val - BeginVal) / AtBegin; 851 beta = Val - Slope * AtBegin; 852 853 for (i=0; i < AtBegin; i++) 854 g ->Table16[i] = _cmsQuickSaturateWord(i * Slope + beta); 855 856 // Compute slope and offset for the end 857 Val = g ->Table16[AtEnd]; 858 Slope = (EndVal - Val) / AtBegin; // AtBegin holds the X interval, which is same in both cases 859 beta = Val - Slope * AtEnd; 860 861 for (i = AtEnd; i < (int) g ->nEntries; i++) 862 g ->Table16[i] = _cmsQuickSaturateWord(i * Slope + beta); 863} 864 865 866// Precomputes tables for 8-bit on input devicelink. 867static 868Prelin8Data* PrelinOpt8alloc(cmsContext ContextID, const cmsInterpParams* p, cmsToneCurve* G[3]) 869{ 870 int i; 871 cmsUInt16Number Input[3]; 872 cmsS15Fixed16Number v1, v2, v3; 873 Prelin8Data* p8; 874 875 p8 = (Prelin8Data*)_cmsMallocZero(ContextID, sizeof(Prelin8Data)); 876 if (p8 == NULL) return NULL; 877 878 // Since this only works for 8 bit input, values comes always as x * 257, 879 // we can safely take msb byte (x << 8 + x) 880 881 for (i=0; i < 256; i++) { 882 883 if (G != NULL) { 884 885 // Get 16-bit representation 886 Input[0] = cmsEvalToneCurve16(G[0], FROM_8_TO_16(i)); 887 Input[1] = cmsEvalToneCurve16(G[1], FROM_8_TO_16(i)); 888 Input[2] = cmsEvalToneCurve16(G[2], FROM_8_TO_16(i)); 889 } 890 else { 891 Input[0] = FROM_8_TO_16(i); 892 Input[1] = FROM_8_TO_16(i); 893 Input[2] = FROM_8_TO_16(i); 894 } 895 896 897 // Move to 0..1.0 in fixed domain 898 v1 = _cmsToFixedDomain(Input[0] * p -> Domain[0]); 899 v2 = _cmsToFixedDomain(Input[1] * p -> Domain[1]); 900 v3 = _cmsToFixedDomain(Input[2] * p -> Domain[2]); 901 902 // Store the precalculated table of nodes 903 p8 ->X0[i] = (p->opta[2] * FIXED_TO_INT(v1)); 904 p8 ->Y0[i] = (p->opta[1] * FIXED_TO_INT(v2)); 905 p8 ->Z0[i] = (p->opta[0] * FIXED_TO_INT(v3)); 906 907 // Store the precalculated table of offsets 908 p8 ->rx[i] = (cmsUInt16Number) FIXED_REST_TO_INT(v1); 909 p8 ->ry[i] = (cmsUInt16Number) FIXED_REST_TO_INT(v2); 910 p8 ->rz[i] = (cmsUInt16Number) FIXED_REST_TO_INT(v3); 911 } 912 913 p8 ->ContextID = ContextID; 914 p8 ->p = p; 915 916 return p8; 917} 918 919static 920void Prelin8free(cmsContext ContextID, void* ptr) 921{ 922 _cmsFree(ContextID, ptr); 923} 924 925static 926void* Prelin8dup(cmsContext ContextID, const void* ptr) 927{ 928 return _cmsDupMem(ContextID, ptr, sizeof(Prelin8Data)); 929} 930 931 932 933// A optimized interpolation for 8-bit input. 934#define DENS(i,j,k) (LutTable[(i)+(j)+(k)+OutChan]) 935static 936void PrelinEval8(register const cmsUInt16Number Input[], 937 register cmsUInt16Number Output[], 938 register const void* D) 939{ 940 941 cmsUInt8Number r, g, b; 942 cmsS15Fixed16Number rx, ry, rz; 943 cmsS15Fixed16Number c0, c1, c2, c3, Rest; 944 int OutChan; 945 register cmsS15Fixed16Number X0, X1, Y0, Y1, Z0, Z1; 946 Prelin8Data* p8 = (Prelin8Data*) D; 947 register const cmsInterpParams* p = p8 ->p; 948 int TotalOut = p -> nOutputs; 949 const cmsUInt16Number* LutTable = (const cmsUInt16Number*) p->Table; 950 951 r = Input[0] >> 8; 952 g = Input[1] >> 8; 953 b = Input[2] >> 8; 954 955 X0 = X1 = p8->X0[r]; 956 Y0 = Y1 = p8->Y0[g]; 957 Z0 = Z1 = p8->Z0[b]; 958 959 rx = p8 ->rx[r]; 960 ry = p8 ->ry[g]; 961 rz = p8 ->rz[b]; 962 963 X1 = X0 + ((rx == 0) ? 0 : p ->opta[2]); 964 Y1 = Y0 + ((ry == 0) ? 0 : p ->opta[1]); 965 Z1 = Z0 + ((rz == 0) ? 0 : p ->opta[0]); 966 967 968 // These are the 6 Tetrahedral 969 for (OutChan=0; OutChan < TotalOut; OutChan++) { 970 971 c0 = DENS(X0, Y0, Z0); 972 973 if (rx >= ry && ry >= rz) 974 { 975 c1 = DENS(X1, Y0, Z0) - c0; 976 c2 = DENS(X1, Y1, Z0) - DENS(X1, Y0, Z0); 977 c3 = DENS(X1, Y1, Z1) - DENS(X1, Y1, Z0); 978 } 979 else 980 if (rx >= rz && rz >= ry) 981 { 982 c1 = DENS(X1, Y0, Z0) - c0; 983 c2 = DENS(X1, Y1, Z1) - DENS(X1, Y0, Z1); 984 c3 = DENS(X1, Y0, Z1) - DENS(X1, Y0, Z0); 985 } 986 else 987 if (rz >= rx && rx >= ry) 988 { 989 c1 = DENS(X1, Y0, Z1) - DENS(X0, Y0, Z1); 990 c2 = DENS(X1, Y1, Z1) - DENS(X1, Y0, Z1); 991 c3 = DENS(X0, Y0, Z1) - c0; 992 } 993 else 994 if (ry >= rx && rx >= rz) 995 { 996 c1 = DENS(X1, Y1, Z0) - DENS(X0, Y1, Z0); 997 c2 = DENS(X0, Y1, Z0) - c0; 998 c3 = DENS(X1, Y1, Z1) - DENS(X1, Y1, Z0); 999 } 1000 else 1001 if (ry >= rz && rz >= rx) 1002 { 1003 c1 = DENS(X1, Y1, Z1) - DENS(X0, Y1, Z1); 1004 c2 = DENS(X0, Y1, Z0) - c0; 1005 c3 = DENS(X0, Y1, Z1) - DENS(X0, Y1, Z0); 1006 } 1007 else 1008 if (rz >= ry && ry >= rx) 1009 { 1010 c1 = DENS(X1, Y1, Z1) - DENS(X0, Y1, Z1); 1011 c2 = DENS(X0, Y1, Z1) - DENS(X0, Y0, Z1); 1012 c3 = DENS(X0, Y0, Z1) - c0; 1013 } 1014 else { 1015 c1 = c2 = c3 = 0; 1016 } 1017 1018 1019 Rest = c1 * rx + c2 * ry + c3 * rz + 0x8001; 1020 Output[OutChan] = (cmsUInt16Number)c0 + ((Rest + (Rest>>16))>>16); 1021 1022 } 1023} 1024 1025#undef DENS 1026 1027 1028// Curves that contain wide empty areas are not optimizeable 1029static 1030cmsBool IsDegenerated(const cmsToneCurve* g) 1031{ 1032 int i, Zeros = 0, Poles = 0; 1033 int nEntries = g ->nEntries; 1034 1035 for (i=0; i < nEntries; i++) { 1036 1037 if (g ->Table16[i] == 0x0000) Zeros++; 1038 if (g ->Table16[i] == 0xffff) Poles++; 1039 } 1040 1041 if (Zeros == 1 && Poles == 1) return FALSE; // For linear tables 1042 if (Zeros > (nEntries / 20)) return TRUE; // Degenerated, many zeros 1043 if (Poles > (nEntries / 20)) return TRUE; // Degenerated, many poles 1044 1045 return FALSE; 1046} 1047 1048// -------------------------------------------------------------------------------------------------------------- 1049// We need xput over here 1050 1051static 1052cmsBool OptimizeByComputingLinearization(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags) 1053{ 1054 cmsPipeline* OriginalLut; 1055 int nGridPoints; 1056 cmsToneCurve *Trans[cmsMAXCHANNELS], *TransReverse[cmsMAXCHANNELS]; 1057 cmsUInt32Number t, i; 1058 cmsFloat32Number v, In[cmsMAXCHANNELS], Out[cmsMAXCHANNELS]; 1059 cmsBool lIsSuitable, lIsLinear; 1060 cmsPipeline* OptimizedLUT = NULL, *LutPlusCurves = NULL; 1061 cmsStage* OptimizedCLUTmpe; 1062 cmsColorSpaceSignature ColorSpace, OutputColorSpace; 1063 cmsStage* OptimizedPrelinMpe; 1064 cmsStage* mpe; 1065 cmsToneCurve** OptimizedPrelinCurves; 1066 _cmsStageCLutData* OptimizedPrelinCLUT; 1067 1068 1069 // This is a loosy optimization! does not apply in floating-point cases 1070 if (_cmsFormatterIsFloat(*InputFormat) || _cmsFormatterIsFloat(*OutputFormat)) return FALSE; 1071 1072 // Only on chunky RGB 1073 if (T_COLORSPACE(*InputFormat) != PT_RGB) return FALSE; 1074 if (T_PLANAR(*InputFormat)) return FALSE; 1075 1076 if (T_COLORSPACE(*OutputFormat) != PT_RGB) return FALSE; 1077 if (T_PLANAR(*OutputFormat)) return FALSE; 1078 1079 // On 16 bits, user has to specify the feature 1080 if (!_cmsFormatterIs8bit(*InputFormat)) { 1081 if (!(*dwFlags & cmsFLAGS_CLUT_PRE_LINEARIZATION)) return FALSE; 1082 } 1083 1084 OriginalLut = *Lut; 1085 1086 // Named color pipelines cannot be optimized either 1087 for (mpe = cmsPipelineGetPtrToFirstStage(OriginalLut); 1088 mpe != NULL; 1089 mpe = cmsStageNext(mpe)) { 1090 if (cmsStageType(mpe) == cmsSigNamedColorElemType) return FALSE; 1091 } 1092 1093 ColorSpace = _cmsICCcolorSpace(T_COLORSPACE(*InputFormat)); 1094 OutputColorSpace = _cmsICCcolorSpace(T_COLORSPACE(*OutputFormat)); 1095 nGridPoints = _cmsReasonableGridpointsByColorspace(ColorSpace, *dwFlags); 1096 1097 // Empty gamma containers 1098 memset(Trans, 0, sizeof(Trans)); 1099 memset(TransReverse, 0, sizeof(TransReverse)); 1100 1101 // If the last stage of the original lut are curves, and those curves are 1102 // degenerated, it is likely the transform is squeezing and clipping 1103 // the output from previous CLUT. We cannot optimize this case 1104 { 1105 cmsStage* last = cmsPipelineGetPtrToLastStage(OriginalLut); 1106 1107 if (cmsStageType(last) == cmsSigCurveSetElemType) { 1108 1109 _cmsStageToneCurvesData* Data = (_cmsStageToneCurvesData*)cmsStageData(last); 1110 for (i = 0; i < Data->nCurves; i++) { 1111 if (IsDegenerated(Data->TheCurves[i])) 1112 goto Error; 1113 } 1114 } 1115 } 1116 1117 for (t = 0; t < OriginalLut ->InputChannels; t++) { 1118 Trans[t] = cmsBuildTabulatedToneCurve16(OriginalLut ->ContextID, PRELINEARIZATION_POINTS, NULL); 1119 if (Trans[t] == NULL) goto Error; 1120 } 1121 1122 // Populate the curves 1123 for (i=0; i < PRELINEARIZATION_POINTS; i++) { 1124 1125 v = (cmsFloat32Number) ((cmsFloat64Number) i / (PRELINEARIZATION_POINTS - 1)); 1126 1127 // Feed input with a gray ramp 1128 for (t=0; t < OriginalLut ->InputChannels; t++) 1129 In[t] = v; 1130 1131 // Evaluate the gray value 1132 cmsPipelineEvalFloat(In, Out, OriginalLut); 1133 1134 // Store result in curve 1135 for (t=0; t < OriginalLut ->InputChannels; t++) 1136 Trans[t] ->Table16[i] = _cmsQuickSaturateWord(Out[t] * 65535.0); 1137 } 1138 1139 // Slope-limit the obtained curves 1140 for (t = 0; t < OriginalLut ->InputChannels; t++) 1141 SlopeLimiting(Trans[t]); 1142 1143 // Check for validity 1144 lIsSuitable = TRUE; 1145 lIsLinear = TRUE; 1146 for (t=0; (lIsSuitable && (t < OriginalLut ->InputChannels)); t++) { 1147 1148 // Exclude if already linear 1149 if (!cmsIsToneCurveLinear(Trans[t])) 1150 lIsLinear = FALSE; 1151 1152 // Exclude if non-monotonic 1153 if (!cmsIsToneCurveMonotonic(Trans[t])) 1154 lIsSuitable = FALSE; 1155 1156 if (IsDegenerated(Trans[t])) 1157 lIsSuitable = FALSE; 1158 } 1159 1160 // If it is not suitable, just quit 1161 if (!lIsSuitable) goto Error; 1162 1163 // Invert curves if possible 1164 for (t = 0; t < OriginalLut ->InputChannels; t++) { 1165 TransReverse[t] = cmsReverseToneCurveEx(PRELINEARIZATION_POINTS, Trans[t]); 1166 if (TransReverse[t] == NULL) goto Error; 1167 } 1168 1169 // Now inset the reversed curves at the begin of transform 1170 LutPlusCurves = cmsPipelineDup(OriginalLut); 1171 if (LutPlusCurves == NULL) goto Error; 1172 1173 if (!cmsPipelineInsertStage(LutPlusCurves, cmsAT_BEGIN, cmsStageAllocToneCurves(OriginalLut ->ContextID, OriginalLut ->InputChannels, TransReverse))) 1174 goto Error; 1175 1176 // Create the result LUT 1177 OptimizedLUT = cmsPipelineAlloc(OriginalLut ->ContextID, OriginalLut ->InputChannels, OriginalLut ->OutputChannels); 1178 if (OptimizedLUT == NULL) goto Error; 1179 1180 OptimizedPrelinMpe = cmsStageAllocToneCurves(OriginalLut ->ContextID, OriginalLut ->InputChannels, Trans); 1181 1182 // Create and insert the curves at the beginning 1183 if (!cmsPipelineInsertStage(OptimizedLUT, cmsAT_BEGIN, OptimizedPrelinMpe)) 1184 goto Error; 1185 1186 // Allocate the CLUT for result 1187 OptimizedCLUTmpe = cmsStageAllocCLut16bit(OriginalLut ->ContextID, nGridPoints, OriginalLut ->InputChannels, OriginalLut ->OutputChannels, NULL); 1188 1189 // Add the CLUT to the destination LUT 1190 if (!cmsPipelineInsertStage(OptimizedLUT, cmsAT_END, OptimizedCLUTmpe)) 1191 goto Error; 1192 1193 // Resample the LUT 1194 if (!cmsStageSampleCLut16bit(OptimizedCLUTmpe, XFormSampler16, (void*) LutPlusCurves, 0)) goto Error; 1195 1196 // Free resources 1197 for (t = 0; t < OriginalLut ->InputChannels; t++) { 1198 1199 if (Trans[t]) cmsFreeToneCurve(Trans[t]); 1200 if (TransReverse[t]) cmsFreeToneCurve(TransReverse[t]); 1201 } 1202 1203 cmsPipelineFree(LutPlusCurves); 1204 1205 1206 OptimizedPrelinCurves = _cmsStageGetPtrToCurveSet(OptimizedPrelinMpe); 1207 OptimizedPrelinCLUT = (_cmsStageCLutData*) OptimizedCLUTmpe ->Data; 1208 1209 // Set the evaluator if 8-bit 1210 if (_cmsFormatterIs8bit(*InputFormat)) { 1211 1212 Prelin8Data* p8 = PrelinOpt8alloc(OptimizedLUT ->ContextID, 1213 OptimizedPrelinCLUT ->Params, 1214 OptimizedPrelinCurves); 1215 if (p8 == NULL) return FALSE; 1216 1217 _cmsPipelineSetOptimizationParameters(OptimizedLUT, PrelinEval8, (void*) p8, Prelin8free, Prelin8dup); 1218 1219 } 1220 else 1221 { 1222 Prelin16Data* p16 = PrelinOpt16alloc(OptimizedLUT ->ContextID, 1223 OptimizedPrelinCLUT ->Params, 1224 3, OptimizedPrelinCurves, 3, NULL); 1225 if (p16 == NULL) return FALSE; 1226 1227 _cmsPipelineSetOptimizationParameters(OptimizedLUT, PrelinEval16, (void*) p16, PrelinOpt16free, Prelin16dup); 1228 1229 } 1230 1231 // Don't fix white on absolute colorimetric 1232 if (Intent == INTENT_ABSOLUTE_COLORIMETRIC) 1233 *dwFlags |= cmsFLAGS_NOWHITEONWHITEFIXUP; 1234 1235 if (!(*dwFlags & cmsFLAGS_NOWHITEONWHITEFIXUP)) { 1236 1237 if (!FixWhiteMisalignment(OptimizedLUT, ColorSpace, OutputColorSpace)) { 1238 1239 return FALSE; 1240 } 1241 } 1242 1243 // And return the obtained LUT 1244 1245 cmsPipelineFree(OriginalLut); 1246 *Lut = OptimizedLUT; 1247 return TRUE; 1248 1249Error: 1250 1251 for (t = 0; t < OriginalLut ->InputChannels; t++) { 1252 1253 if (Trans[t]) cmsFreeToneCurve(Trans[t]); 1254 if (TransReverse[t]) cmsFreeToneCurve(TransReverse[t]); 1255 } 1256 1257 if (LutPlusCurves != NULL) cmsPipelineFree(LutPlusCurves); 1258 if (OptimizedLUT != NULL) cmsPipelineFree(OptimizedLUT); 1259 1260 return FALSE; 1261 1262 cmsUNUSED_PARAMETER(Intent); 1263} 1264 1265 1266// Curves optimizer ------------------------------------------------------------------------------------------------------------------ 1267 1268static 1269void CurvesFree(cmsContext ContextID, void* ptr) 1270{ 1271 Curves16Data* Data = (Curves16Data*) ptr; 1272 int i; 1273 1274 for (i=0; i < Data -> nCurves; i++) { 1275 1276 _cmsFree(ContextID, Data ->Curves[i]); 1277 } 1278 1279 _cmsFree(ContextID, Data ->Curves); 1280 _cmsFree(ContextID, ptr); 1281} 1282 1283static 1284void* CurvesDup(cmsContext ContextID, const void* ptr) 1285{ 1286 Curves16Data* Data = (Curves16Data*)_cmsDupMem(ContextID, ptr, sizeof(Curves16Data)); 1287 int i; 1288 1289 if (Data == NULL) return NULL; 1290 1291 Data->Curves = (cmsUInt16Number**) _cmsDupMem(ContextID, Data->Curves, Data->nCurves * sizeof(cmsUInt16Number*)); 1292 1293 for (i=0; i < Data -> nCurves; i++) { 1294 Data->Curves[i] = (cmsUInt16Number*) _cmsDupMem(ContextID, Data->Curves[i], Data->nElements * sizeof(cmsUInt16Number)); 1295 } 1296 1297 return (void*) Data; 1298} 1299 1300// Precomputes tables for 8-bit on input devicelink. 1301static 1302Curves16Data* CurvesAlloc(cmsContext ContextID, int nCurves, int nElements, cmsToneCurve** G) 1303{ 1304 int i, j; 1305 Curves16Data* c16; 1306 1307 c16 = (Curves16Data*)_cmsMallocZero(ContextID, sizeof(Curves16Data)); 1308 if (c16 == NULL) return NULL; 1309 1310 c16 ->nCurves = nCurves; 1311 c16 ->nElements = nElements; 1312 1313 c16->Curves = (cmsUInt16Number**) _cmsCalloc(ContextID, nCurves, sizeof(cmsUInt16Number*)); 1314 if (c16 ->Curves == NULL) return NULL; 1315 1316 for (i=0; i < nCurves; i++) { 1317 1318 c16->Curves[i] = (cmsUInt16Number*) _cmsCalloc(ContextID, nElements, sizeof(cmsUInt16Number)); 1319 1320 if (c16->Curves[i] == NULL) { 1321 1322 for (j=0; j < i; j++) { 1323 _cmsFree(ContextID, c16->Curves[j]); 1324 } 1325 _cmsFree(ContextID, c16->Curves); 1326 _cmsFree(ContextID, c16); 1327 return NULL; 1328 } 1329 1330 if (nElements == 256) { 1331 1332 for (j=0; j < nElements; j++) { 1333 1334 c16 ->Curves[i][j] = cmsEvalToneCurve16(G[i], FROM_8_TO_16(j)); 1335 } 1336 } 1337 else { 1338 1339 for (j=0; j < nElements; j++) { 1340 c16 ->Curves[i][j] = cmsEvalToneCurve16(G[i], (cmsUInt16Number) j); 1341 } 1342 } 1343 } 1344 1345 return c16; 1346} 1347 1348static 1349void FastEvaluateCurves8(register const cmsUInt16Number In[], 1350 register cmsUInt16Number Out[], 1351 register const void* D) 1352{ 1353 Curves16Data* Data = (Curves16Data*) D; 1354 cmsUInt8Number x; 1355 int i; 1356 1357 for (i=0; i < Data ->nCurves; i++) { 1358 1359 x = (In[i] >> 8); 1360 Out[i] = Data -> Curves[i][x]; 1361 } 1362} 1363 1364 1365static 1366void FastEvaluateCurves16(register const cmsUInt16Number In[], 1367 register cmsUInt16Number Out[], 1368 register const void* D) 1369{ 1370 Curves16Data* Data = (Curves16Data*) D; 1371 int i; 1372 1373 for (i=0; i < Data ->nCurves; i++) { 1374 Out[i] = Data -> Curves[i][In[i]]; 1375 } 1376} 1377 1378 1379static 1380void FastIdentity16(register const cmsUInt16Number In[], 1381 register cmsUInt16Number Out[], 1382 register const void* D) 1383{ 1384 cmsPipeline* Lut = (cmsPipeline*) D; 1385 cmsUInt32Number i; 1386 1387 for (i=0; i < Lut ->InputChannels; i++) { 1388 Out[i] = In[i]; 1389 } 1390} 1391 1392 1393// If the target LUT holds only curves, the optimization procedure is to join all those 1394// curves together. That only works on curves and does not work on matrices. 1395static 1396cmsBool OptimizeByJoiningCurves(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags) 1397{ 1398 cmsToneCurve** GammaTables = NULL; 1399 cmsFloat32Number InFloat[cmsMAXCHANNELS], OutFloat[cmsMAXCHANNELS]; 1400 cmsUInt32Number i, j; 1401 cmsPipeline* Src = *Lut; 1402 cmsPipeline* Dest = NULL; 1403 cmsStage* mpe; 1404 cmsStage* ObtainedCurves = NULL; 1405 1406 1407 // This is a loosy optimization! does not apply in floating-point cases 1408 if (_cmsFormatterIsFloat(*InputFormat) || _cmsFormatterIsFloat(*OutputFormat)) return FALSE; 1409 1410 // Only curves in this LUT? 1411 for (mpe = cmsPipelineGetPtrToFirstStage(Src); 1412 mpe != NULL; 1413 mpe = cmsStageNext(mpe)) { 1414 if (cmsStageType(mpe) != cmsSigCurveSetElemType) return FALSE; 1415 } 1416 1417 // Allocate an empty LUT 1418 Dest = cmsPipelineAlloc(Src ->ContextID, Src ->InputChannels, Src ->OutputChannels); 1419 if (Dest == NULL) return FALSE; 1420 1421 // Create target curves 1422 GammaTables = (cmsToneCurve**) _cmsCalloc(Src ->ContextID, Src ->InputChannels, sizeof(cmsToneCurve*)); 1423 if (GammaTables == NULL) goto Error; 1424 1425 for (i=0; i < Src ->InputChannels; i++) { 1426 GammaTables[i] = cmsBuildTabulatedToneCurve16(Src ->ContextID, PRELINEARIZATION_POINTS, NULL); 1427 if (GammaTables[i] == NULL) goto Error; 1428 } 1429 1430 // Compute 16 bit result by using floating point 1431 for (i=0; i < PRELINEARIZATION_POINTS; i++) { 1432 1433 for (j=0; j < Src ->InputChannels; j++) 1434 InFloat[j] = (cmsFloat32Number) ((cmsFloat64Number) i / (PRELINEARIZATION_POINTS - 1)); 1435 1436 cmsPipelineEvalFloat(InFloat, OutFloat, Src); 1437 1438 for (j=0; j < Src ->InputChannels; j++) 1439 GammaTables[j] -> Table16[i] = _cmsQuickSaturateWord(OutFloat[j] * 65535.0); 1440 } 1441 1442 ObtainedCurves = cmsStageAllocToneCurves(Src ->ContextID, Src ->InputChannels, GammaTables); 1443 if (ObtainedCurves == NULL) goto Error; 1444 1445 for (i=0; i < Src ->InputChannels; i++) { 1446 cmsFreeToneCurve(GammaTables[i]); 1447 GammaTables[i] = NULL; 1448 } 1449 1450 if (GammaTables != NULL) { 1451 _cmsFree(Src->ContextID, GammaTables); 1452 GammaTables = NULL; 1453 } 1454 1455 // Maybe the curves are linear at the end 1456 if (!AllCurvesAreLinear(ObtainedCurves)) { 1457 1458 if (!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, ObtainedCurves)) 1459 goto Error; 1460 1461 // If the curves are to be applied in 8 bits, we can save memory 1462 if (_cmsFormatterIs8bit(*InputFormat)) { 1463 1464 _cmsStageToneCurvesData* Data = (_cmsStageToneCurvesData*) ObtainedCurves ->Data; 1465 Curves16Data* c16 = CurvesAlloc(Dest ->ContextID, Data ->nCurves, 256, Data ->TheCurves); 1466 1467 if (c16 == NULL) goto Error; 1468 *dwFlags |= cmsFLAGS_NOCACHE; 1469 _cmsPipelineSetOptimizationParameters(Dest, FastEvaluateCurves8, c16, CurvesFree, CurvesDup); 1470 1471 } 1472 else { 1473 1474 _cmsStageToneCurvesData* Data = (_cmsStageToneCurvesData*) cmsStageData(ObtainedCurves); 1475 Curves16Data* c16 = CurvesAlloc(Dest ->ContextID, Data ->nCurves, 65536, Data ->TheCurves); 1476 1477 if (c16 == NULL) goto Error; 1478 *dwFlags |= cmsFLAGS_NOCACHE; 1479 _cmsPipelineSetOptimizationParameters(Dest, FastEvaluateCurves16, c16, CurvesFree, CurvesDup); 1480 } 1481 } 1482 else { 1483 1484 // LUT optimizes to nothing. Set the identity LUT 1485 cmsStageFree(ObtainedCurves); 1486 1487 if (!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, cmsStageAllocIdentity(Dest ->ContextID, Src ->InputChannels))) 1488 goto Error; 1489 1490 *dwFlags |= cmsFLAGS_NOCACHE; 1491 _cmsPipelineSetOptimizationParameters(Dest, FastIdentity16, (void*) Dest, NULL, NULL); 1492 } 1493 1494 // We are done. 1495 cmsPipelineFree(Src); 1496 *Lut = Dest; 1497 return TRUE; 1498 1499Error: 1500 1501 if (ObtainedCurves != NULL) cmsStageFree(ObtainedCurves); 1502 if (GammaTables != NULL) { 1503 for (i=0; i < Src ->InputChannels; i++) { 1504 if (GammaTables[i] != NULL) cmsFreeToneCurve(GammaTables[i]); 1505 } 1506 1507 _cmsFree(Src ->ContextID, GammaTables); 1508 } 1509 1510 if (Dest != NULL) cmsPipelineFree(Dest); 1511 return FALSE; 1512 1513 cmsUNUSED_PARAMETER(Intent); 1514 cmsUNUSED_PARAMETER(InputFormat); 1515 cmsUNUSED_PARAMETER(OutputFormat); 1516 cmsUNUSED_PARAMETER(dwFlags); 1517} 1518 1519// ------------------------------------------------------------------------------------------------------------------------------------- 1520// LUT is Shaper - Matrix - Matrix - Shaper, which is very frequent when combining two matrix-shaper profiles 1521 1522 1523static 1524void FreeMatShaper(cmsContext ContextID, void* Data) 1525{ 1526 if (Data != NULL) _cmsFree(ContextID, Data); 1527} 1528 1529static 1530void* DupMatShaper(cmsContext ContextID, const void* Data) 1531{ 1532 return _cmsDupMem(ContextID, Data, sizeof(MatShaper8Data)); 1533} 1534 1535 1536// A fast matrix-shaper evaluator for 8 bits. This is a bit ticky since I'm using 1.14 signed fixed point 1537// to accomplish some performance. Actually it takes 256x3 16 bits tables and 16385 x 3 tables of 8 bits, 1538// in total about 50K, and the performance boost is huge! 1539static 1540void MatShaperEval16(register const cmsUInt16Number In[], 1541 register cmsUInt16Number Out[], 1542 register const void* D) 1543{ 1544 MatShaper8Data* p = (MatShaper8Data*) D; 1545 cmsS1Fixed14Number l1, l2, l3, r, g, b; 1546 cmsUInt32Number ri, gi, bi; 1547 1548 // In this case (and only in this case!) we can use this simplification since 1549 // In[] is assured to come from a 8 bit number. (a << 8 | a) 1550 ri = In[0] & 0xFF; 1551 gi = In[1] & 0xFF; 1552 bi = In[2] & 0xFF; 1553 1554 // Across first shaper, which also converts to 1.14 fixed point 1555 r = p->Shaper1R[ri]; 1556 g = p->Shaper1G[gi]; 1557 b = p->Shaper1B[bi]; 1558 1559 // Evaluate the matrix in 1.14 fixed point 1560 l1 = (p->Mat[0][0] * r + p->Mat[0][1] * g + p->Mat[0][2] * b + p->Off[0] + 0x2000) >> 14; 1561 l2 = (p->Mat[1][0] * r + p->Mat[1][1] * g + p->Mat[1][2] * b + p->Off[1] + 0x2000) >> 14; 1562 l3 = (p->Mat[2][0] * r + p->Mat[2][1] * g + p->Mat[2][2] * b + p->Off[2] + 0x2000) >> 14; 1563 1564 // Now we have to clip to 0..1.0 range 1565 ri = (l1 < 0) ? 0 : ((l1 > 16384) ? 16384 : l1); 1566 gi = (l2 < 0) ? 0 : ((l2 > 16384) ? 16384 : l2); 1567 bi = (l3 < 0) ? 0 : ((l3 > 16384) ? 16384 : l3); 1568 1569 // And across second shaper, 1570 Out[0] = p->Shaper2R[ri]; 1571 Out[1] = p->Shaper2G[gi]; 1572 Out[2] = p->Shaper2B[bi]; 1573 1574} 1575 1576// This table converts from 8 bits to 1.14 after applying the curve 1577static 1578void FillFirstShaper(cmsS1Fixed14Number* Table, cmsToneCurve* Curve) 1579{ 1580 int i; 1581 cmsFloat32Number R, y; 1582 1583 for (i=0; i < 256; i++) { 1584 1585 R = (cmsFloat32Number) (i / 255.0); 1586 y = cmsEvalToneCurveFloat(Curve, R); 1587 1588 Table[i] = DOUBLE_TO_1FIXED14(y); 1589 } 1590} 1591 1592// This table converts form 1.14 (being 0x4000 the last entry) to 8 bits after applying the curve 1593static 1594void FillSecondShaper(cmsUInt16Number* Table, cmsToneCurve* Curve, cmsBool Is8BitsOutput) 1595{ 1596 int i; 1597 cmsFloat32Number R, Val; 1598 1599 for (i=0; i < 16385; i++) { 1600 1601 R = (cmsFloat32Number) (i / 16384.0); 1602 Val = cmsEvalToneCurveFloat(Curve, R); // Val comes 0..1.0 1603 1604 if (Is8BitsOutput) { 1605 1606 // If 8 bits output, we can optimize further by computing the / 257 part. 1607 // first we compute the resulting byte and then we store the byte times 1608 // 257. This quantization allows to round very quick by doing a >> 8, but 1609 // since the low byte is always equal to msb, we can do a & 0xff and this works! 1610 cmsUInt16Number w = _cmsQuickSaturateWord(Val * 65535.0); 1611 cmsUInt8Number b = FROM_16_TO_8(w); 1612 1613 Table[i] = FROM_8_TO_16(b); 1614 } 1615 else Table[i] = _cmsQuickSaturateWord(Val * 65535.0); 1616 } 1617} 1618 1619// Compute the matrix-shaper structure 1620static 1621cmsBool SetMatShaper(cmsPipeline* Dest, cmsToneCurve* Curve1[3], cmsMAT3* Mat, cmsVEC3* Off, cmsToneCurve* Curve2[3], cmsUInt32Number* OutputFormat) 1622{ 1623 MatShaper8Data* p; 1624 int i, j; 1625 cmsBool Is8Bits = _cmsFormatterIs8bit(*OutputFormat); 1626 1627 // Allocate a big chuck of memory to store precomputed tables 1628 p = (MatShaper8Data*) _cmsMalloc(Dest ->ContextID, sizeof(MatShaper8Data)); 1629 if (p == NULL) return FALSE; 1630 1631 p -> ContextID = Dest -> ContextID; 1632 1633 // Precompute tables 1634 FillFirstShaper(p ->Shaper1R, Curve1[0]); 1635 FillFirstShaper(p ->Shaper1G, Curve1[1]); 1636 FillFirstShaper(p ->Shaper1B, Curve1[2]); 1637 1638 FillSecondShaper(p ->Shaper2R, Curve2[0], Is8Bits); 1639 FillSecondShaper(p ->Shaper2G, Curve2[1], Is8Bits); 1640 FillSecondShaper(p ->Shaper2B, Curve2[2], Is8Bits); 1641 1642 // Convert matrix to nFixed14. Note that those values may take more than 16 bits as 1643 for (i=0; i < 3; i++) { 1644 for (j=0; j < 3; j++) { 1645 p ->Mat[i][j] = DOUBLE_TO_1FIXED14(Mat->v[i].n[j]); 1646 } 1647 } 1648 1649 for (i=0; i < 3; i++) { 1650 1651 if (Off == NULL) { 1652 p ->Off[i] = 0; 1653 } 1654 else { 1655 p ->Off[i] = DOUBLE_TO_1FIXED14(Off->n[i]); 1656 } 1657 } 1658 1659 // Mark as optimized for faster formatter 1660 if (Is8Bits) 1661 *OutputFormat |= OPTIMIZED_SH(1); 1662 1663 // Fill function pointers 1664 _cmsPipelineSetOptimizationParameters(Dest, MatShaperEval16, (void*) p, FreeMatShaper, DupMatShaper); 1665 return TRUE; 1666} 1667 1668// 8 bits on input allows matrix-shaper boot up to 25 Mpixels per second on RGB. That's fast! 1669static 1670cmsBool OptimizeMatrixShaper(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags) 1671{ 1672 cmsStage* Curve1, *Curve2; 1673 cmsStage* Matrix1, *Matrix2; 1674 cmsMAT3 res; 1675 cmsBool IdentityMat; 1676 cmsPipeline* Dest, *Src; 1677 cmsFloat64Number* Offset; 1678 1679 // Only works on RGB to RGB 1680 if (T_CHANNELS(*InputFormat) != 3 || T_CHANNELS(*OutputFormat) != 3) return FALSE; 1681 1682 // Only works on 8 bit input 1683 if (!_cmsFormatterIs8bit(*InputFormat)) return FALSE; 1684 1685 // Seems suitable, proceed 1686 Src = *Lut; 1687 1688 // Check for: 1689 // 1690 // shaper-matrix-matrix-shaper 1691 // shaper-matrix-shaper 1692 // 1693 // Both of those constructs are possible (first because abs. colorimetric). 1694 // additionally, In the first case, the input matrix offset should be zero. 1695 1696 IdentityMat = FALSE; 1697 if (cmsPipelineCheckAndRetreiveStages(Src, 4, 1698 cmsSigCurveSetElemType, cmsSigMatrixElemType, cmsSigMatrixElemType, cmsSigCurveSetElemType, 1699 &Curve1, &Matrix1, &Matrix2, &Curve2)) { 1700 1701 // Get both matrices 1702 _cmsStageMatrixData* Data1 = (_cmsStageMatrixData*)cmsStageData(Matrix1); 1703 _cmsStageMatrixData* Data2 = (_cmsStageMatrixData*)cmsStageData(Matrix2); 1704 1705 // Input offset should be zero 1706 if (Data1->Offset != NULL) return FALSE; 1707 1708 // Multiply both matrices to get the result 1709 _cmsMAT3per(&res, (cmsMAT3*)Data2->Double, (cmsMAT3*)Data1->Double); 1710 1711 // Only 2nd matrix has offset, or it is zero 1712 Offset = Data2->Offset; 1713 1714 // Now the result is in res + Data2 -> Offset. Maybe is a plain identity? 1715 if (_cmsMAT3isIdentity(&res) && Offset == NULL) { 1716 1717 // We can get rid of full matrix 1718 IdentityMat = TRUE; 1719 } 1720 1721 } 1722 else { 1723 1724 if (cmsPipelineCheckAndRetreiveStages(Src, 3, 1725 cmsSigCurveSetElemType, cmsSigMatrixElemType, cmsSigCurveSetElemType, 1726 &Curve1, &Matrix1, &Curve2)) { 1727 1728 _cmsStageMatrixData* Data = (_cmsStageMatrixData*)cmsStageData(Matrix1); 1729 1730 // Copy the matrix to our result 1731 memcpy(&res, Data->Double, sizeof(res)); 1732 1733 // Preserve the Odffset (may be NULL as a zero offset) 1734 Offset = Data->Offset; 1735 1736 if (_cmsMAT3isIdentity(&res) && Offset == NULL) { 1737 1738 // We can get rid of full matrix 1739 IdentityMat = TRUE; 1740 } 1741 } 1742 else 1743 return FALSE; // Not optimizeable this time 1744 1745 } 1746 1747 // Allocate an empty LUT 1748 Dest = cmsPipelineAlloc(Src ->ContextID, Src ->InputChannels, Src ->OutputChannels); 1749 if (!Dest) return FALSE; 1750 1751 // Assamble the new LUT 1752 if (!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, cmsStageDup(Curve1))) 1753 goto Error; 1754 1755 if (!IdentityMat) { 1756 1757 if (!cmsPipelineInsertStage(Dest, cmsAT_END, cmsStageAllocMatrix(Dest->ContextID, 3, 3, (const cmsFloat64Number*)&res, Offset))) 1758 goto Error; 1759 } 1760 1761 if (!cmsPipelineInsertStage(Dest, cmsAT_END, cmsStageDup(Curve2))) 1762 goto Error; 1763 1764 // If identity on matrix, we can further optimize the curves, so call the join curves routine 1765 if (IdentityMat) { 1766 1767 OptimizeByJoiningCurves(&Dest, Intent, InputFormat, OutputFormat, dwFlags); 1768 } 1769 else { 1770 _cmsStageToneCurvesData* mpeC1 = (_cmsStageToneCurvesData*) cmsStageData(Curve1); 1771 _cmsStageToneCurvesData* mpeC2 = (_cmsStageToneCurvesData*) cmsStageData(Curve2); 1772 1773 // In this particular optimization, cach� does not help as it takes more time to deal with 1774 // the cach� that with the pixel handling 1775 *dwFlags |= cmsFLAGS_NOCACHE; 1776 1777 // Setup the optimizarion routines 1778 SetMatShaper(Dest, mpeC1 ->TheCurves, &res, (cmsVEC3*) Offset, mpeC2->TheCurves, OutputFormat); 1779 } 1780 1781 cmsPipelineFree(Src); 1782 *Lut = Dest; 1783 return TRUE; 1784Error: 1785 // Leave Src unchanged 1786 cmsPipelineFree(Dest); 1787 return FALSE; 1788} 1789 1790 1791// ------------------------------------------------------------------------------------------------------------------------------------- 1792// Optimization plug-ins 1793 1794// List of optimizations 1795typedef struct _cmsOptimizationCollection_st { 1796 1797 _cmsOPToptimizeFn OptimizePtr; 1798 1799 struct _cmsOptimizationCollection_st *Next; 1800 1801} _cmsOptimizationCollection; 1802 1803 1804// The built-in list. We currently implement 4 types of optimizations. Joining of curves, matrix-shaper, linearization and resampling 1805static _cmsOptimizationCollection DefaultOptimization[] = { 1806 1807 { OptimizeByJoiningCurves, &DefaultOptimization[1] }, 1808 { OptimizeMatrixShaper, &DefaultOptimization[2] }, 1809 { OptimizeByComputingLinearization, &DefaultOptimization[3] }, 1810 { OptimizeByResampling, NULL } 1811}; 1812 1813// The linked list head 1814_cmsOptimizationPluginChunkType _cmsOptimizationPluginChunk = { NULL }; 1815 1816 1817// Duplicates the zone of memory used by the plug-in in the new context 1818static 1819void DupPluginOptimizationList(struct _cmsContext_struct* ctx, 1820 const struct _cmsContext_struct* src) 1821{ 1822 _cmsOptimizationPluginChunkType newHead = { NULL }; 1823 _cmsOptimizationCollection* entry; 1824 _cmsOptimizationCollection* Anterior = NULL; 1825 _cmsOptimizationPluginChunkType* head = (_cmsOptimizationPluginChunkType*) src->chunks[OptimizationPlugin]; 1826 1827 _cmsAssert(ctx != NULL); 1828 _cmsAssert(head != NULL); 1829 1830 // Walk the list copying all nodes 1831 for (entry = head->OptimizationCollection; 1832 entry != NULL; 1833 entry = entry ->Next) { 1834 1835 _cmsOptimizationCollection *newEntry = ( _cmsOptimizationCollection *) _cmsSubAllocDup(ctx ->MemPool, entry, sizeof(_cmsOptimizationCollection)); 1836 1837 if (newEntry == NULL) 1838 return; 1839 1840 // We want to keep the linked list order, so this is a little bit tricky 1841 newEntry -> Next = NULL; 1842 if (Anterior) 1843 Anterior -> Next = newEntry; 1844 1845 Anterior = newEntry; 1846 1847 if (newHead.OptimizationCollection == NULL) 1848 newHead.OptimizationCollection = newEntry; 1849 } 1850 1851 ctx ->chunks[OptimizationPlugin] = _cmsSubAllocDup(ctx->MemPool, &newHead, sizeof(_cmsOptimizationPluginChunkType)); 1852} 1853 1854void _cmsAllocOptimizationPluginChunk(struct _cmsContext_struct* ctx, 1855 const struct _cmsContext_struct* src) 1856{ 1857 if (src != NULL) { 1858 1859 // Copy all linked list 1860 DupPluginOptimizationList(ctx, src); 1861 } 1862 else { 1863 static _cmsOptimizationPluginChunkType OptimizationPluginChunkType = { NULL }; 1864 ctx ->chunks[OptimizationPlugin] = _cmsSubAllocDup(ctx ->MemPool, &OptimizationPluginChunkType, sizeof(_cmsOptimizationPluginChunkType)); 1865 } 1866} 1867 1868 1869// Register new ways to optimize 1870cmsBool _cmsRegisterOptimizationPlugin(cmsContext ContextID, cmsPluginBase* Data) 1871{ 1872 cmsPluginOptimization* Plugin = (cmsPluginOptimization*) Data; 1873 _cmsOptimizationPluginChunkType* ctx = ( _cmsOptimizationPluginChunkType*) _cmsContextGetClientChunk(ContextID, OptimizationPlugin); 1874 _cmsOptimizationCollection* fl; 1875 1876 if (Data == NULL) { 1877 1878 ctx->OptimizationCollection = NULL; 1879 return TRUE; 1880 } 1881 1882 // Optimizer callback is required 1883 if (Plugin ->OptimizePtr == NULL) return FALSE; 1884 1885 fl = (_cmsOptimizationCollection*) _cmsPluginMalloc(ContextID, sizeof(_cmsOptimizationCollection)); 1886 if (fl == NULL) return FALSE; 1887 1888 // Copy the parameters 1889 fl ->OptimizePtr = Plugin ->OptimizePtr; 1890 1891 // Keep linked list 1892 fl ->Next = ctx->OptimizationCollection; 1893 1894 // Set the head 1895 ctx ->OptimizationCollection = fl; 1896 1897 // All is ok 1898 return TRUE; 1899} 1900 1901// The entry point for LUT optimization 1902cmsBool _cmsOptimizePipeline(cmsContext ContextID, 1903 cmsPipeline** PtrLut, 1904 int Intent, 1905 cmsUInt32Number* InputFormat, 1906 cmsUInt32Number* OutputFormat, 1907 cmsUInt32Number* dwFlags) 1908{ 1909 _cmsOptimizationPluginChunkType* ctx = ( _cmsOptimizationPluginChunkType*) _cmsContextGetClientChunk(ContextID, OptimizationPlugin); 1910 _cmsOptimizationCollection* Opts; 1911 cmsBool AnySuccess = FALSE; 1912 1913 // A CLUT is being asked, so force this specific optimization 1914 if (*dwFlags & cmsFLAGS_FORCE_CLUT) { 1915 1916 PreOptimize(*PtrLut); 1917 return OptimizeByResampling(PtrLut, Intent, InputFormat, OutputFormat, dwFlags); 1918 } 1919 1920 // Anything to optimize? 1921 if ((*PtrLut) ->Elements == NULL) { 1922 _cmsPipelineSetOptimizationParameters(*PtrLut, FastIdentity16, (void*) *PtrLut, NULL, NULL); 1923 return TRUE; 1924 } 1925 1926 // Try to get rid of identities and trivial conversions. 1927 AnySuccess = PreOptimize(*PtrLut); 1928 1929 // After removal do we end with an identity? 1930 if ((*PtrLut) ->Elements == NULL) { 1931 _cmsPipelineSetOptimizationParameters(*PtrLut, FastIdentity16, (void*) *PtrLut, NULL, NULL); 1932 return TRUE; 1933 } 1934 1935 // Do not optimize, keep all precision 1936 if (*dwFlags & cmsFLAGS_NOOPTIMIZE) 1937 return FALSE; 1938 1939 // Try plug-in optimizations 1940 for (Opts = ctx->OptimizationCollection; 1941 Opts != NULL; 1942 Opts = Opts ->Next) { 1943 1944 // If one schema succeeded, we are done 1945 if (Opts ->OptimizePtr(PtrLut, Intent, InputFormat, OutputFormat, dwFlags)) { 1946 1947 return TRUE; // Optimized! 1948 } 1949 } 1950 1951 // Try built-in optimizations 1952 for (Opts = DefaultOptimization; 1953 Opts != NULL; 1954 Opts = Opts ->Next) { 1955 1956 if (Opts ->OptimizePtr(PtrLut, Intent, InputFormat, OutputFormat, dwFlags)) { 1957 1958 return TRUE; 1959 } 1960 } 1961 1962 // Only simple optimizations succeeded 1963 return AnySuccess; 1964} 1965 1966 1967 1968