libs/lcms2/src/cmsopt.c

   1 //---------------------------------------------------------------------------------
   2 //
   3 //  Little Color Management System
   4 //  Copyright (c) 1998-2023 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     cmsUInt32Number nInputs;
  53     cmsUInt32Number 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     cmsUInt32Number nCurves;      // Number of curves
  98     cmsUInt32Number 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 after 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(CMSREGISTER const cmsUInt16Number Input[],
 292                  CMSREGISTER cmsUInt16Number Output[],
 293                  CMSREGISTER const struct _cms_interp_struc* p)
 294 {
 295     Output[0] = Input[0];
 296
 297     cmsUNUSED_PARAMETER(p);
 298 }
 299
 300 static
 301 void PrelinEval16(CMSREGISTER const cmsUInt16Number Input[],
 302                   CMSREGISTER cmsUInt16Number Output[],
 303                   CMSREGISTER const void* D)
 304 {
 305     Prelin16Data* p16 = (Prelin16Data*) D;
 306     cmsUInt16Number  StageABC[MAX_INPUT_DIMENSIONS];
 307     cmsUInt16Number  StageDEF[cmsMAXCHANNELS];
 308     cmsUInt32Number 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                                cmsUInt32Number nInputs, cmsToneCurve** In,
 354                                cmsUInt32Number nOutputs, cmsToneCurve** Out )
 355 {
 356     cmsUInt32Number 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     if (p16->EvalCurveOut16 == NULL)
 383     {
 384         _cmsFree(ContextID, p16);
 385         return NULL;
 386     }
 387
 388     p16 -> ParamsCurveOut16 = (cmsInterpParams**) _cmsCalloc(ContextID, nOutputs, sizeof(cmsInterpParams* ));
 389     if (p16->ParamsCurveOut16 == NULL)
 390     {
 391
 392         _cmsFree(ContextID, p16->EvalCurveOut16);
 393         _cmsFree(ContextID, p16);
 394         return NULL;
 395     }
 396
 397     for (i=0; i < nOutputs; i++) {
 398
 399         if (Out == NULL) {
 400             p16 ->ParamsCurveOut16[i] = NULL;
 401             p16 -> EvalCurveOut16[i] = Eval16nop1D;
 402         }
 403         else {
 404
 405             p16 ->ParamsCurveOut16[i] = Out[i] ->InterpParams;
 406             p16 -> EvalCurveOut16[i] = p16 ->ParamsCurveOut16[i]->Interpolation.Lerp16;
 407         }
 408     }
 409
 410     return p16;
 411 }
 412
 413
 414
 415 // Resampling ---------------------------------------------------------------------------------
 416
 417 #define PRELINEARIZATION_POINTS 4096
 418
 419 // Sampler implemented by another LUT. This is a clean way to precalculate the devicelink 3D CLUT for
 420 // almost any transform. We use floating point precision and then convert from floating point to 16 bits.
 421 static
 422 cmsInt32Number XFormSampler16(CMSREGISTER const cmsUInt16Number In[],
 423                               CMSREGISTER cmsUInt16Number Out[],
 424                               CMSREGISTER void* Cargo)
 425 {
 426     cmsPipeline* Lut = (cmsPipeline*) Cargo;
 427     cmsFloat32Number InFloat[cmsMAXCHANNELS], OutFloat[cmsMAXCHANNELS];
 428     cmsUInt32Number i;
 429
 430     _cmsAssert(Lut -> InputChannels < cmsMAXCHANNELS);
 431     _cmsAssert(Lut -> OutputChannels < cmsMAXCHANNELS);
 432
 433     // From 16 bit to floating point
 434     for (i=0; i < Lut ->InputChannels; i++)
 435         InFloat[i] = (cmsFloat32Number) (In[i] / 65535.0);
 436
 437     // Evaluate in floating point
 438     cmsPipelineEvalFloat(InFloat, OutFloat, Lut);
 439
 440     // Back to 16 bits representation
 441     for (i=0; i < Lut ->OutputChannels; i++)
 442         Out[i] = _cmsQuickSaturateWord(OutFloat[i] * 65535.0);
 443
 444     // Always succeed
 445     return TRUE;
 446 }
 447
 448 // Try to see if the curves of a given MPE are linear
 449 static
 450 cmsBool AllCurvesAreLinear(cmsStage* mpe)
 451 {
 452     cmsToneCurve** Curves;
 453     cmsUInt32Number i, n;
 454
 455     Curves = _cmsStageGetPtrToCurveSet(mpe);
 456     if (Curves == NULL) return FALSE;
 457
 458     n = cmsStageOutputChannels(mpe);
 459
 460     for (i=0; i < n; i++) {
 461         if (!cmsIsToneCurveLinear(Curves[i])) return FALSE;
 462     }
 463
 464     return TRUE;
 465 }
 466
 467 // This function replaces a specific node placed in "At" by the "Value" numbers. Its purpose
 468 // is to fix scum dot on broken profiles/transforms. Works on 1, 3 and 4 channels
 469 static
 470 cmsBool  PatchLUT(cmsStage* CLUT, cmsUInt16Number At[], cmsUInt16Number Value[],
 471                   cmsUInt32Number nChannelsOut, cmsUInt32Number nChannelsIn)
 472 {
 473     _cmsStageCLutData* Grid = (_cmsStageCLutData*) CLUT ->Data;
 474     cmsInterpParams* p16  = Grid ->Params;
 475     cmsFloat64Number px, py, pz, pw;
 476     int        x0, y0, z0, w0;
 477     int        i, index;
 478
 479     if (CLUT -> Type != cmsSigCLutElemType) {
 480         cmsSignalError(CLUT->ContextID, cmsERROR_INTERNAL, "(internal) Attempt to PatchLUT on non-lut stage");
 481         return FALSE;
 482     }
 483
 484     if (nChannelsIn == 4) {
 485
 486         px = ((cmsFloat64Number) At[0] * (p16->Domain[0])) / 65535.0;
 487         py = ((cmsFloat64Number) At[1] * (p16->Domain[1])) / 65535.0;
 488         pz = ((cmsFloat64Number) At[2] * (p16->Domain[2])) / 65535.0;
 489         pw = ((cmsFloat64Number) At[3] * (p16->Domain[3])) / 65535.0;
 490
 491         x0 = (int) floor(px);
 492         y0 = (int) floor(py);
 493         z0 = (int) floor(pz);
 494         w0 = (int) floor(pw);
 495
 496         if (((px - x0) != 0) ||
 497             ((py - y0) != 0) ||
 498             ((pz - z0) != 0) ||
 499             ((pw - w0) != 0)) return FALSE; // Not on exact node
 500
 501         index = (int) p16 -> opta[3] * x0 +
 502                 (int) p16 -> opta[2] * y0 +
 503                 (int) p16 -> opta[1] * z0 +
 504                 (int) p16 -> opta[0] * w0;
 505     }
 506     else
 507         if (nChannelsIn == 3) {
 508
 509             px = ((cmsFloat64Number) At[0] * (p16->Domain[0])) / 65535.0;
 510             py = ((cmsFloat64Number) At[1] * (p16->Domain[1])) / 65535.0;
 511             pz = ((cmsFloat64Number) At[2] * (p16->Domain[2])) / 65535.0;
 512
 513             x0 = (int) floor(px);
 514             y0 = (int) floor(py);
 515             z0 = (int) floor(pz);
 516
 517             if (((px - x0) != 0) ||
 518                 ((py - y0) != 0) ||
 519                 ((pz - z0) != 0)) return FALSE;  // Not on exact node
 520
 521             index = (int) p16 -> opta[2] * x0 +
 522                     (int) p16 -> opta[1] * y0 +
 523                     (int) p16 -> opta[0] * z0;
 524         }
 525         else
 526             if (nChannelsIn == 1) {
 527
 528                 px = ((cmsFloat64Number) At[0] * (p16->Domain[0])) / 65535.0;
 529
 530                 x0 = (int) floor(px);
 531
 532                 if (((px - x0) != 0)) return FALSE; // Not on exact node
 533
 534                 index = (int) p16 -> opta[0] * x0;
 535             }
 536             else {
 537                 cmsSignalError(CLUT->ContextID, cmsERROR_INTERNAL, "(internal) %d Channels are not supported on PatchLUT", nChannelsIn);
 538                 return FALSE;
 539             }
 540
 541     for (i = 0; i < (int) nChannelsOut; i++)
 542         Grid->Tab.T[index + i] = Value[i];
 543
 544     return TRUE;
 545 }
 546
 547 // Auxiliary, to see if two values are equal or very different
 548 static
 549 cmsBool WhitesAreEqual(cmsUInt32Number n, cmsUInt16Number White1[], cmsUInt16Number White2[] )
 550 {
 551     cmsUInt32Number i;
 552
 553     for (i=0; i < n; i++) {
 554
 555         if (abs(White1[i] - White2[i]) > 0xf000) return TRUE;  // Values are so extremely different that the fixup should be avoided
 556         if (White1[i] != White2[i]) return FALSE;
 557     }
 558     return TRUE;
 559 }
 560
 561
 562 // Locate the node for the white point and fix it to pure white in order to avoid scum dot.
 563 static
 564 cmsBool FixWhiteMisalignment(cmsPipeline* Lut, cmsColorSpaceSignature EntryColorSpace, cmsColorSpaceSignature ExitColorSpace)
 565 {
 566     cmsUInt16Number *WhitePointIn, *WhitePointOut;
 567     cmsUInt16Number  WhiteIn[cmsMAXCHANNELS], WhiteOut[cmsMAXCHANNELS], ObtainedOut[cmsMAXCHANNELS];
 568     cmsUInt32Number i, nOuts, nIns;
 569     cmsStage *PreLin = NULL, *CLUT = NULL, *PostLin = NULL;
 570
 571     if (!_cmsEndPointsBySpace(EntryColorSpace,
 572         &WhitePointIn, NULL, &nIns)) return FALSE;
 573
 574     if (!_cmsEndPointsBySpace(ExitColorSpace,
 575         &WhitePointOut, NULL, &nOuts)) return FALSE;
 576
 577     // It needs to be fixed?
 578     if (Lut ->InputChannels != nIns) return FALSE;
 579     if (Lut ->OutputChannels != nOuts) return FALSE;
 580
 581     cmsPipelineEval16(WhitePointIn, ObtainedOut, Lut);
 582
 583     if (WhitesAreEqual(nOuts, WhitePointOut, ObtainedOut)) return TRUE; // whites already match
 584
 585     // Check if the LUT comes as Prelin, CLUT or Postlin. We allow all combinations
 586     if (!cmsPipelineCheckAndRetreiveStages(Lut, 3, cmsSigCurveSetElemType, cmsSigCLutElemType, cmsSigCurveSetElemType, &PreLin, &CLUT, &PostLin))
 587         if (!cmsPipelineCheckAndRetreiveStages(Lut, 2, cmsSigCurveSetElemType, cmsSigCLutElemType, &PreLin, &CLUT))
 588             if (!cmsPipelineCheckAndRetreiveStages(Lut, 2, cmsSigCLutElemType, cmsSigCurveSetElemType, &CLUT, &PostLin))
 589                 if (!cmsPipelineCheckAndRetreiveStages(Lut, 1, cmsSigCLutElemType, &CLUT))
 590                     return FALSE;
 591
 592     // We need to interpolate white points of both, pre and post curves
 593     if (PreLin) {
 594
 595         cmsToneCurve** Curves = _cmsStageGetPtrToCurveSet(PreLin);
 596
 597         for (i=0; i < nIns; i++) {
 598             WhiteIn[i] = cmsEvalToneCurve16(Curves[i], WhitePointIn[i]);
 599         }
 600     }
 601     else {
 602         for (i=0; i < nIns; i++)
 603             WhiteIn[i] = WhitePointIn[i];
 604     }
 605
 606     // If any post-linearization, we need to find how is represented white before the curve, do
 607     // a reverse interpolation in this case.
 608     if (PostLin) {
 609
 610         cmsToneCurve** Curves = _cmsStageGetPtrToCurveSet(PostLin);
 611
 612         for (i=0; i < nOuts; i++) {
 613
 614             cmsToneCurve* InversePostLin = cmsReverseToneCurve(Curves[i]);
 615             if (InversePostLin == NULL) {
 616                 WhiteOut[i] = WhitePointOut[i];
 617
 618             } else {
 619
 620                 WhiteOut[i] = cmsEvalToneCurve16(InversePostLin, WhitePointOut[i]);
 621                 cmsFreeToneCurve(InversePostLin);
 622             }
 623         }
 624     }
 625     else {
 626         for (i=0; i < nOuts; i++)
 627             WhiteOut[i] = WhitePointOut[i];
 628     }
 629
 630     // Ok, proceed with patching. May fail and we don't care if it fails
 631     PatchLUT(CLUT, WhiteIn, WhiteOut, nOuts, nIns);
 632
 633     return TRUE;
 634 }
 635
 636 // -----------------------------------------------------------------------------------------------------------------------------------------------
 637 // This function creates simple LUT from complex ones. The generated LUT has an optional set of
 638 // prelinearization curves, a CLUT of nGridPoints and optional postlinearization tables.
 639 // These curves have to exist in the original LUT in order to be used in the simplified output.
 640 // Caller may also use the flags to allow this feature.
 641 // LUTS with all curves will be simplified to a single curve. Parametric curves are lost.
 642 // This function should be used on 16-bits LUTS only, as floating point losses precision when simplified
 643 // -----------------------------------------------------------------------------------------------------------------------------------------------
 644
 645 static
 646 cmsBool OptimizeByResampling(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags)
 647 {
 648     cmsPipeline* Src = NULL;
 649     cmsPipeline* Dest = NULL;
 650     cmsStage* CLUT;
 651     cmsStage *KeepPreLin = NULL, *KeepPostLin = NULL;
 652     cmsUInt32Number nGridPoints;
 653     cmsColorSpaceSignature ColorSpace, OutputColorSpace;
 654     cmsStage *NewPreLin = NULL;
 655     cmsStage *NewPostLin = NULL;
 656     _cmsStageCLutData* DataCLUT;
 657     cmsToneCurve** DataSetIn;
 658     cmsToneCurve** DataSetOut;
 659     Prelin16Data* p16;
 660
 661     // This is a lossy optimization! does not apply in floating-point cases
 662     if (_cmsFormatterIsFloat(*InputFormat) || _cmsFormatterIsFloat(*OutputFormat)) return FALSE;
 663
 664     ColorSpace       = _cmsICCcolorSpace((int) T_COLORSPACE(*InputFormat));
 665     OutputColorSpace = _cmsICCcolorSpace((int) T_COLORSPACE(*OutputFormat));
 666
 667     // Color space must be specified
 668     if (ColorSpace == (cmsColorSpaceSignature)0 ||
 669         OutputColorSpace == (cmsColorSpaceSignature)0) return FALSE;
 670
 671     nGridPoints = _cmsReasonableGridpointsByColorspace(ColorSpace, *dwFlags);
 672
 673     // For empty LUTs, 2 points are enough
 674     if (cmsPipelineStageCount(*Lut) == 0)
 675         nGridPoints = 2;
 676
 677     Src = *Lut;
 678
 679     // Allocate an empty LUT
 680     Dest =  cmsPipelineAlloc(Src ->ContextID, Src ->InputChannels, Src ->OutputChannels);
 681     if (!Dest) return FALSE;
 682
 683     // Prelinearization tables are kept unless indicated by flags
 684     if (*dwFlags & cmsFLAGS_CLUT_PRE_LINEARIZATION) {
 685
 686         // Get a pointer to the prelinearization element
 687         cmsStage* PreLin = cmsPipelineGetPtrToFirstStage(Src);
 688
 689         // Check if suitable
 690         if (PreLin && PreLin ->Type == cmsSigCurveSetElemType) {
 691
 692             // Maybe this is a linear tram, so we can avoid the whole stuff
 693             if (!AllCurvesAreLinear(PreLin)) {
 694
 695                 // All seems ok, proceed.
 696                 NewPreLin = cmsStageDup(PreLin);
 697                 if(!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, NewPreLin))
 698                     goto Error;
 699
 700                 // Remove prelinearization. Since we have duplicated the curve
 701                 // in destination LUT, the sampling should be applied after this stage.
 702                 cmsPipelineUnlinkStage(Src, cmsAT_BEGIN, &KeepPreLin);
 703             }
 704         }
 705     }
 706
 707     // Allocate the CLUT
 708     CLUT = cmsStageAllocCLut16bit(Src ->ContextID, nGridPoints, Src ->InputChannels, Src->OutputChannels, NULL);
 709     if (CLUT == NULL) goto Error;
 710
 711     // Add the CLUT to the destination LUT
 712     if (!cmsPipelineInsertStage(Dest, cmsAT_END, CLUT)) {
 713         goto Error;
 714     }
 715
 716     // Postlinearization tables are kept unless indicated by flags
 717     if (*dwFlags & cmsFLAGS_CLUT_POST_LINEARIZATION) {
 718
 719         // Get a pointer to the postlinearization if present
 720         cmsStage* PostLin = cmsPipelineGetPtrToLastStage(Src);
 721
 722         // Check if suitable
 723         if (PostLin && cmsStageType(PostLin) == cmsSigCurveSetElemType) {
 724
 725             // Maybe this is a linear tram, so we can avoid the whole stuff
 726             if (!AllCurvesAreLinear(PostLin)) {
 727
 728                 // All seems ok, proceed.
 729                 NewPostLin = cmsStageDup(PostLin);
 730                 if (!cmsPipelineInsertStage(Dest, cmsAT_END, NewPostLin))
 731                     goto Error;
 732
 733                 // In destination LUT, the sampling should be applied after this stage.
 734                 cmsPipelineUnlinkStage(Src, cmsAT_END, &KeepPostLin);
 735             }
 736         }
 737     }
 738
 739     // Now its time to do the sampling. We have to ignore pre/post linearization
 740     // The source LUT without pre/post curves is passed as parameter.
 741     if (!cmsStageSampleCLut16bit(CLUT, XFormSampler16, (void*) Src, 0)) {
 742 Error:
 743         // Ops, something went wrong, Restore stages
 744         if (KeepPreLin != NULL) {
 745             if (!cmsPipelineInsertStage(Src, cmsAT_BEGIN, KeepPreLin)) {
 746                 _cmsAssert(0); // This never happens
 747             }
 748         }
 749         if (KeepPostLin != NULL) {
 750             if (!cmsPipelineInsertStage(Src, cmsAT_END,   KeepPostLin)) {
 751                 _cmsAssert(0); // This never happens
 752             }
 753         }
 754         cmsPipelineFree(Dest);
 755         return FALSE;
 756     }
 757
 758     // Done.
 759
 760     if (KeepPreLin != NULL) cmsStageFree(KeepPreLin);
 761     if (KeepPostLin != NULL) cmsStageFree(KeepPostLin);
 762     cmsPipelineFree(Src);
 763
 764     DataCLUT = (_cmsStageCLutData*) CLUT ->Data;
 765
 766     if (NewPreLin == NULL) DataSetIn = NULL;
 767     else DataSetIn = ((_cmsStageToneCurvesData*) NewPreLin ->Data) ->TheCurves;
 768
 769     if (NewPostLin == NULL) DataSetOut = NULL;
 770     else  DataSetOut = ((_cmsStageToneCurvesData*) NewPostLin ->Data) ->TheCurves;
 771
 772
 773     if (DataSetIn == NULL && DataSetOut == NULL) {
 774
 775         _cmsPipelineSetOptimizationParameters(Dest, (_cmsPipelineEval16Fn) DataCLUT->Params->Interpolation.Lerp16, DataCLUT->Params, NULL, NULL);
 776     }
 777     else {
 778
 779         p16 = PrelinOpt16alloc(Dest ->ContextID,
 780             DataCLUT ->Params,
 781             Dest ->InputChannels,
 782             DataSetIn,
 783             Dest ->OutputChannels,
 784             DataSetOut);
 785
 786         _cmsPipelineSetOptimizationParameters(Dest, PrelinEval16, (void*) p16, PrelinOpt16free, Prelin16dup);
 787     }
 788
 789
 790     // Don't fix white on absolute colorimetric
 791     if (Intent == INTENT_ABSOLUTE_COLORIMETRIC)
 792         *dwFlags |= cmsFLAGS_NOWHITEONWHITEFIXUP;
 793
 794     if (!(*dwFlags & cmsFLAGS_NOWHITEONWHITEFIXUP)) {
 795
 796         FixWhiteMisalignment(Dest, ColorSpace, OutputColorSpace);
 797     }
 798
 799     *Lut = Dest;
 800     return TRUE;
 801
 802     cmsUNUSED_PARAMETER(Intent);
 803 }
 804
 805
 806 // -----------------------------------------------------------------------------------------------------------------------------------------------
 807 // Fixes the gamma balancing of transform. This is described in my paper "Prelinearization Stages on
 808 // Color-Management Application-Specific Integrated Circuits (ASICs)" presented at NIP24. It only works
 809 // for RGB transforms. See the paper for more details
 810 // -----------------------------------------------------------------------------------------------------------------------------------------------
 811
 812
 813 // Normalize endpoints by slope limiting max and min. This assures endpoints as well.
 814 // Descending curves are handled as well.
 815 static
 816 void SlopeLimiting(cmsToneCurve* g)
 817 {
 818     int BeginVal, EndVal;
 819     int AtBegin = (int) floor((cmsFloat64Number) g ->nEntries * 0.02 + 0.5);   // Cutoff at 2%
 820     int AtEnd   = (int) g ->nEntries - AtBegin - 1;                                  // And 98%
 821     cmsFloat64Number Val, Slope, beta;
 822     int i;
 823
 824     if (cmsIsToneCurveDescending(g)) {
 825         BeginVal = 0xffff; EndVal = 0;
 826     }
 827     else {
 828         BeginVal = 0; EndVal = 0xffff;
 829     }
 830
 831     // Compute slope and offset for begin of curve
 832     Val   = g ->Table16[AtBegin];
 833     Slope = (Val - BeginVal) / AtBegin;
 834     beta  = Val - Slope * AtBegin;
 835
 836     for (i=0; i < AtBegin; i++)
 837         g ->Table16[i] = _cmsQuickSaturateWord(i * Slope + beta);
 838
 839     // Compute slope and offset for the end
 840     Val   = g ->Table16[AtEnd];
 841     Slope = (EndVal - Val) / AtBegin;   // AtBegin holds the X interval, which is same in both cases
 842     beta  = Val - Slope * AtEnd;
 843
 844     for (i = AtEnd; i < (int) g ->nEntries; i++)
 845         g ->Table16[i] = _cmsQuickSaturateWord(i * Slope + beta);
 846 }
 847
 848
 849 // Precomputes tables for 8-bit on input devicelink.
 850 static
 851 Prelin8Data* PrelinOpt8alloc(cmsContext ContextID, const cmsInterpParams* p, cmsToneCurve* G[3])
 852 {
 853     int i;
 854     cmsUInt16Number Input[3];
 855     cmsS15Fixed16Number v1, v2, v3;
 856     Prelin8Data* p8;
 857
 858     p8 = (Prelin8Data*)_cmsMallocZero(ContextID, sizeof(Prelin8Data));
 859     if (p8 == NULL) return NULL;
 860
 861     // Since this only works for 8 bit input, values comes always as x * 257,
 862     // we can safely take msb byte (x << 8 + x)
 863
 864     for (i=0; i < 256; i++) {
 865
 866         if (G != NULL) {
 867
 868             // Get 16-bit representation
 869             Input[0] = cmsEvalToneCurve16(G[0], FROM_8_TO_16(i));
 870             Input[1] = cmsEvalToneCurve16(G[1], FROM_8_TO_16(i));
 871             Input[2] = cmsEvalToneCurve16(G[2], FROM_8_TO_16(i));
 872         }
 873         else {
 874             Input[0] = FROM_8_TO_16(i);
 875             Input[1] = FROM_8_TO_16(i);
 876             Input[2] = FROM_8_TO_16(i);
 877         }
 878
 879
 880         // Move to 0..1.0 in fixed domain
 881         v1 = _cmsToFixedDomain((int) (Input[0] * p -> Domain[0]));
 882         v2 = _cmsToFixedDomain((int) (Input[1] * p -> Domain[1]));
 883         v3 = _cmsToFixedDomain((int) (Input[2] * p -> Domain[2]));
 884
 885         // Store the precalculated table of nodes
 886         p8 ->X0[i] = (p->opta[2] * FIXED_TO_INT(v1));
 887         p8 ->Y0[i] = (p->opta[1] * FIXED_TO_INT(v2));
 888         p8 ->Z0[i] = (p->opta[0] * FIXED_TO_INT(v3));
 889
 890         // Store the precalculated table of offsets
 891         p8 ->rx[i] = (cmsUInt16Number) FIXED_REST_TO_INT(v1);
 892         p8 ->ry[i] = (cmsUInt16Number) FIXED_REST_TO_INT(v2);
 893         p8 ->rz[i] = (cmsUInt16Number) FIXED_REST_TO_INT(v3);
 894     }
 895
 896     p8 ->ContextID = ContextID;
 897     p8 ->p = p;
 898
 899     return p8;
 900 }
 901
 902 static
 903 void Prelin8free(cmsContext ContextID, void* ptr)
 904 {
 905     _cmsFree(ContextID, ptr);
 906 }
 907
 908 static
 909 void* Prelin8dup(cmsContext ContextID, const void* ptr)
 910 {
 911     return _cmsDupMem(ContextID, ptr, sizeof(Prelin8Data));
 912 }
 913
 914
 915
 916 // A optimized interpolation for 8-bit input.
 917 #define DENS(i,j,k) (LutTable[(i)+(j)+(k)+OutChan])
 918 static CMS_NO_SANITIZE
 919 void PrelinEval8(CMSREGISTER const cmsUInt16Number Input[],
 920                  CMSREGISTER cmsUInt16Number Output[],
 921                  CMSREGISTER const void* D)
 922 {
 923
 924     cmsUInt8Number         r, g, b;
 925     cmsS15Fixed16Number    rx, ry, rz;
 926     cmsS15Fixed16Number    c0, c1, c2, c3, Rest;
 927     int                    OutChan;
 928     CMSREGISTER cmsS15Fixed16Number X0, X1, Y0, Y1, Z0, Z1;
 929     Prelin8Data* p8 = (Prelin8Data*) D;
 930     CMSREGISTER const cmsInterpParams* p = p8 ->p;
 931     int                    TotalOut = (int) p -> nOutputs;
 932     const cmsUInt16Number* LutTable = (const cmsUInt16Number*) p->Table;
 933
 934     r = (cmsUInt8Number) (Input[0] >> 8);
 935     g = (cmsUInt8Number) (Input[1] >> 8);
 936     b = (cmsUInt8Number) (Input[2] >> 8);
 937
 938     X0 = (cmsS15Fixed16Number) p8->X0[r];
 939     Y0 = (cmsS15Fixed16Number) p8->Y0[g];
 940     Z0 = (cmsS15Fixed16Number) p8->Z0[b];
 941
 942     rx = p8 ->rx[r];
 943     ry = p8 ->ry[g];
 944     rz = p8 ->rz[b];
 945
 946     X1 = X0 + (cmsS15Fixed16Number)((rx == 0) ? 0 :  p ->opta[2]);
 947     Y1 = Y0 + (cmsS15Fixed16Number)((ry == 0) ? 0 :  p ->opta[1]);
 948     Z1 = Z0 + (cmsS15Fixed16Number)((rz == 0) ? 0 :  p ->opta[0]);
 949
 950
 951     // These are the 6 Tetrahedral
 952     for (OutChan=0; OutChan < TotalOut; OutChan++) {
 953
 954         c0 = DENS(X0, Y0, Z0);
 955
 956         if (rx >= ry && ry >= rz)
 957         {
 958             c1 = DENS(X1, Y0, Z0) - c0;
 959             c2 = DENS(X1, Y1, Z0) - DENS(X1, Y0, Z0);
 960             c3 = DENS(X1, Y1, Z1) - DENS(X1, Y1, Z0);
 961         }
 962         else
 963             if (rx >= rz && rz >= ry)
 964             {
 965                 c1 = DENS(X1, Y0, Z0) - c0;
 966                 c2 = DENS(X1, Y1, Z1) - DENS(X1, Y0, Z1);
 967                 c3 = DENS(X1, Y0, Z1) - DENS(X1, Y0, Z0);
 968             }
 969             else
 970                 if (rz >= rx && rx >= ry)
 971                 {
 972                     c1 = DENS(X1, Y0, Z1) - DENS(X0, Y0, Z1);
 973                     c2 = DENS(X1, Y1, Z1) - DENS(X1, Y0, Z1);
 974                     c3 = DENS(X0, Y0, Z1) - c0;
 975                 }
 976                 else
 977                     if (ry >= rx && rx >= rz)
 978                     {
 979                         c1 = DENS(X1, Y1, Z0) - DENS(X0, Y1, Z0);
 980                         c2 = DENS(X0, Y1, Z0) - c0;
 981                         c3 = DENS(X1, Y1, Z1) - DENS(X1, Y1, Z0);
 982                     }
 983                     else
 984                         if (ry >= rz && rz >= rx)
 985                         {
 986                             c1 = DENS(X1, Y1, Z1) - DENS(X0, Y1, Z1);
 987                             c2 = DENS(X0, Y1, Z0) - c0;
 988                             c3 = DENS(X0, Y1, Z1) - DENS(X0, Y1, Z0);
 989                         }
 990                         else
 991                             if (rz >= ry && ry >= rx)
 992                             {
 993                                 c1 = DENS(X1, Y1, Z1) - DENS(X0, Y1, Z1);
 994                                 c2 = DENS(X0, Y1, Z1) - DENS(X0, Y0, Z1);
 995                                 c3 = DENS(X0, Y0, Z1) - c0;
 996                             }
 997                             else  {
 998                                 c1 = c2 = c3 = 0;
 999                             }
1000
1001         Rest = c1 * rx + c2 * ry + c3 * rz + 0x8001;
1002         Output[OutChan] = (cmsUInt16Number) (c0 + ((Rest + (Rest >> 16)) >> 16));
1003
1004     }
1005 }
1006
1007 #undef DENS
1008
1009
1010 // Curves that contain wide empty areas are not optimizeable
1011 static
1012 cmsBool IsDegenerated(const cmsToneCurve* g)
1013 {
1014     cmsUInt32Number i, Zeros = 0, Poles = 0;
1015     cmsUInt32Number nEntries = g ->nEntries;
1016
1017     for (i=0; i < nEntries; i++) {
1018
1019         if (g ->Table16[i] == 0x0000) Zeros++;
1020         if (g ->Table16[i] == 0xffff) Poles++;
1021     }
1022
1023     if (Zeros == 1 && Poles == 1) return FALSE;  // For linear tables
1024     if (Zeros > (nEntries / 20)) return TRUE;  // Degenerated, many zeros
1025     if (Poles > (nEntries / 20)) return TRUE;  // Degenerated, many poles
1026
1027     return FALSE;
1028 }
1029
1030 // --------------------------------------------------------------------------------------------------------------
1031 // We need xput over here
1032
1033 static
1034 cmsBool OptimizeByComputingLinearization(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags)
1035 {
1036     cmsPipeline* OriginalLut;
1037     cmsUInt32Number nGridPoints;
1038     cmsToneCurve *Trans[cmsMAXCHANNELS], *TransReverse[cmsMAXCHANNELS];
1039     cmsUInt32Number t, i;
1040     cmsFloat32Number v, In[cmsMAXCHANNELS], Out[cmsMAXCHANNELS];
1041     cmsBool lIsSuitable, lIsLinear;
1042     cmsPipeline* OptimizedLUT = NULL, *LutPlusCurves = NULL;
1043     cmsStage* OptimizedCLUTmpe;
1044     cmsColorSpaceSignature ColorSpace, OutputColorSpace;
1045     cmsStage* OptimizedPrelinMpe;
1046     cmsToneCurve** OptimizedPrelinCurves;
1047     _cmsStageCLutData* OptimizedPrelinCLUT;
1048
1049
1050     // This is a lossy optimization! does not apply in floating-point cases
1051     if (_cmsFormatterIsFloat(*InputFormat) || _cmsFormatterIsFloat(*OutputFormat)) return FALSE;
1052
1053     // Only on chunky RGB
1054     if (T_COLORSPACE(*InputFormat)  != PT_RGB) return FALSE;
1055     if (T_PLANAR(*InputFormat)) return FALSE;
1056
1057     if (T_COLORSPACE(*OutputFormat) != PT_RGB) return FALSE;
1058     if (T_PLANAR(*OutputFormat)) return FALSE;
1059
1060     // On 16 bits, user has to specify the feature
1061     if (!_cmsFormatterIs8bit(*InputFormat)) {
1062         if (!(*dwFlags & cmsFLAGS_CLUT_PRE_LINEARIZATION)) return FALSE;
1063     }
1064
1065     OriginalLut = *Lut;
1066
1067     ColorSpace       = _cmsICCcolorSpace((int) T_COLORSPACE(*InputFormat));
1068     OutputColorSpace = _cmsICCcolorSpace((int) T_COLORSPACE(*OutputFormat));
1069
1070     // Color space must be specified
1071     if (ColorSpace == (cmsColorSpaceSignature)0 ||
1072         OutputColorSpace == (cmsColorSpaceSignature)0) return FALSE;
1073
1074     nGridPoints      = _cmsReasonableGridpointsByColorspace(ColorSpace, *dwFlags);
1075
1076     // Empty gamma containers
1077     memset(Trans, 0, sizeof(Trans));
1078     memset(TransReverse, 0, sizeof(TransReverse));
1079
1080     // If the last stage of the original lut are curves, and those curves are
1081     // degenerated, it is likely the transform is squeezing and clipping
1082     // the output from previous CLUT. We cannot optimize this case
1083     {
1084         cmsStage* last = cmsPipelineGetPtrToLastStage(OriginalLut);
1085
1086         if (last == NULL) goto Error;
1087         if (cmsStageType(last) == cmsSigCurveSetElemType) {
1088
1089             _cmsStageToneCurvesData* Data = (_cmsStageToneCurvesData*)cmsStageData(last);
1090             for (i = 0; i < Data->nCurves; i++) {
1091                 if (IsDegenerated(Data->TheCurves[i]))
1092                     goto Error;
1093             }
1094         }
1095     }
1096
1097     for (t = 0; t < OriginalLut ->InputChannels; t++) {
1098         Trans[t] = cmsBuildTabulatedToneCurve16(OriginalLut ->ContextID, PRELINEARIZATION_POINTS, NULL);
1099         if (Trans[t] == NULL) goto Error;
1100     }
1101
1102     // Populate the curves
1103     for (i=0; i < PRELINEARIZATION_POINTS; i++) {
1104
1105         v = (cmsFloat32Number) ((cmsFloat64Number) i / (PRELINEARIZATION_POINTS - 1));
1106
1107         // Feed input with a gray ramp
1108         for (t=0; t < OriginalLut ->InputChannels; t++)
1109             In[t] = v;
1110
1111         // Evaluate the gray value
1112         cmsPipelineEvalFloat(In, Out, OriginalLut);
1113
1114         // Store result in curve
1115         for (t=0; t < OriginalLut ->InputChannels; t++)
1116             Trans[t] ->Table16[i] = _cmsQuickSaturateWord(Out[t] * 65535.0);
1117     }
1118
1119     // Slope-limit the obtained curves
1120     for (t = 0; t < OriginalLut ->InputChannels; t++)
1121         SlopeLimiting(Trans[t]);
1122
1123     // Check for validity
1124     lIsSuitable = TRUE;
1125     lIsLinear   = TRUE;
1126     for (t=0; (lIsSuitable && (t < OriginalLut ->InputChannels)); t++) {
1127
1128         // Exclude if already linear
1129         if (!cmsIsToneCurveLinear(Trans[t]))
1130             lIsLinear = FALSE;
1131
1132         // Exclude if non-monotonic
1133         if (!cmsIsToneCurveMonotonic(Trans[t]))
1134             lIsSuitable = FALSE;
1135
1136         if (IsDegenerated(Trans[t]))
1137             lIsSuitable = FALSE;
1138     }
1139
1140     // If it is not suitable, just quit
1141     if (!lIsSuitable) goto Error;
1142
1143     // Invert curves if possible
1144     for (t = 0; t < OriginalLut ->InputChannels; t++) {
1145         TransReverse[t] = cmsReverseToneCurveEx(PRELINEARIZATION_POINTS, Trans[t]);
1146         if (TransReverse[t] == NULL) goto Error;
1147     }
1148
1149     // Now inset the reversed curves at the begin of transform
1150     LutPlusCurves = cmsPipelineDup(OriginalLut);
1151     if (LutPlusCurves == NULL) goto Error;
1152
1153     if (!cmsPipelineInsertStage(LutPlusCurves, cmsAT_BEGIN, cmsStageAllocToneCurves(OriginalLut ->ContextID, OriginalLut ->InputChannels, TransReverse)))
1154         goto Error;
1155
1156     // Create the result LUT
1157     OptimizedLUT = cmsPipelineAlloc(OriginalLut ->ContextID, OriginalLut ->InputChannels, OriginalLut ->OutputChannels);
1158     if (OptimizedLUT == NULL) goto Error;
1159
1160     OptimizedPrelinMpe = cmsStageAllocToneCurves(OriginalLut ->ContextID, OriginalLut ->InputChannels, Trans);
1161
1162     // Create and insert the curves at the beginning
1163     if (!cmsPipelineInsertStage(OptimizedLUT, cmsAT_BEGIN, OptimizedPrelinMpe))
1164         goto Error;
1165
1166     // Allocate the CLUT for result
1167     OptimizedCLUTmpe = cmsStageAllocCLut16bit(OriginalLut ->ContextID, nGridPoints, OriginalLut ->InputChannels, OriginalLut ->OutputChannels, NULL);
1168
1169     // Add the CLUT to the destination LUT
1170     if (!cmsPipelineInsertStage(OptimizedLUT, cmsAT_END, OptimizedCLUTmpe))
1171         goto Error;
1172
1173     // Resample the LUT
1174     if (!cmsStageSampleCLut16bit(OptimizedCLUTmpe, XFormSampler16, (void*) LutPlusCurves, 0)) goto Error;
1175
1176     // Free resources
1177     for (t = 0; t < OriginalLut ->InputChannels; t++) {
1178
1179         if (Trans[t]) cmsFreeToneCurve(Trans[t]);
1180         if (TransReverse[t]) cmsFreeToneCurve(TransReverse[t]);
1181     }
1182
1183     cmsPipelineFree(LutPlusCurves);
1184
1185
1186     OptimizedPrelinCurves = _cmsStageGetPtrToCurveSet(OptimizedPrelinMpe);
1187     OptimizedPrelinCLUT   = (_cmsStageCLutData*) OptimizedCLUTmpe ->Data;
1188
1189     // Set the evaluator if 8-bit
1190     if (_cmsFormatterIs8bit(*InputFormat)) {
1191
1192         Prelin8Data* p8 = PrelinOpt8alloc(OptimizedLUT ->ContextID,
1193                                                 OptimizedPrelinCLUT ->Params,
1194                                                 OptimizedPrelinCurves);
1195         if (p8 == NULL) return FALSE;
1196
1197         _cmsPipelineSetOptimizationParameters(OptimizedLUT, PrelinEval8, (void*) p8, Prelin8free, Prelin8dup);
1198
1199     }
1200     else
1201     {
1202         Prelin16Data* p16 = PrelinOpt16alloc(OptimizedLUT ->ContextID,
1203             OptimizedPrelinCLUT ->Params,
1204             3, OptimizedPrelinCurves, 3, NULL);
1205         if (p16 == NULL) return FALSE;
1206
1207         _cmsPipelineSetOptimizationParameters(OptimizedLUT, PrelinEval16, (void*) p16, PrelinOpt16free, Prelin16dup);
1208
1209     }
1210
1211     // Don't fix white on absolute colorimetric
1212     if (Intent == INTENT_ABSOLUTE_COLORIMETRIC)
1213         *dwFlags |= cmsFLAGS_NOWHITEONWHITEFIXUP;
1214
1215     if (!(*dwFlags & cmsFLAGS_NOWHITEONWHITEFIXUP)) {
1216
1217         if (!FixWhiteMisalignment(OptimizedLUT, ColorSpace, OutputColorSpace)) {
1218
1219             return FALSE;
1220         }
1221     }
1222
1223     // And return the obtained LUT
1224
1225     cmsPipelineFree(OriginalLut);
1226     *Lut = OptimizedLUT;
1227     return TRUE;
1228
1229 Error:
1230
1231     for (t = 0; t < OriginalLut ->InputChannels; t++) {
1232
1233         if (Trans[t]) cmsFreeToneCurve(Trans[t]);
1234         if (TransReverse[t]) cmsFreeToneCurve(TransReverse[t]);
1235     }
1236
1237     if (LutPlusCurves != NULL) cmsPipelineFree(LutPlusCurves);
1238     if (OptimizedLUT != NULL) cmsPipelineFree(OptimizedLUT);
1239
1240     return FALSE;
1241
1242     cmsUNUSED_PARAMETER(Intent);
1243     cmsUNUSED_PARAMETER(lIsLinear);
1244 }
1245
1246
1247 // Curves optimizer ------------------------------------------------------------------------------------------------------------------
1248
1249 static
1250 void CurvesFree(cmsContext ContextID, void* ptr)
1251 {
1252      Curves16Data* Data = (Curves16Data*) ptr;
1253      cmsUInt32Number i;
1254
1255      for (i=0; i < Data -> nCurves; i++) {
1256
1257          _cmsFree(ContextID, Data ->Curves[i]);
1258      }
1259
1260      _cmsFree(ContextID, Data ->Curves);
1261      _cmsFree(ContextID, ptr);
1262 }
1263
1264 static
1265 void* CurvesDup(cmsContext ContextID, const void* ptr)
1266 {
1267     Curves16Data* Data = (Curves16Data*)_cmsDupMem(ContextID, ptr, sizeof(Curves16Data));
1268     cmsUInt32Number i;
1269
1270     if (Data == NULL) return NULL;
1271
1272     Data->Curves = (cmsUInt16Number**) _cmsDupMem(ContextID, Data->Curves, Data->nCurves * sizeof(cmsUInt16Number*));
1273
1274     for (i=0; i < Data -> nCurves; i++) {
1275         Data->Curves[i] = (cmsUInt16Number*) _cmsDupMem(ContextID, Data->Curves[i], Data->nElements * sizeof(cmsUInt16Number));
1276     }
1277
1278     return (void*) Data;
1279 }
1280
1281 // Precomputes tables for 8-bit on input devicelink.
1282 static
1283 Curves16Data* CurvesAlloc(cmsContext ContextID, cmsUInt32Number nCurves, cmsUInt32Number nElements, cmsToneCurve** G)
1284 {
1285     cmsUInt32Number i, j;
1286     Curves16Data* c16;
1287
1288     c16 = (Curves16Data*)_cmsMallocZero(ContextID, sizeof(Curves16Data));
1289     if (c16 == NULL) return NULL;
1290
1291     c16 ->nCurves = nCurves;
1292     c16 ->nElements = nElements;
1293
1294     c16->Curves = (cmsUInt16Number**) _cmsCalloc(ContextID, nCurves, sizeof(cmsUInt16Number*));
1295     if (c16->Curves == NULL) {
1296         _cmsFree(ContextID, c16);
1297         return NULL;
1298     }
1299
1300     for (i=0; i < nCurves; i++) {
1301
1302         c16->Curves[i] = (cmsUInt16Number*) _cmsCalloc(ContextID, nElements, sizeof(cmsUInt16Number));
1303
1304         if (c16->Curves[i] == NULL) {
1305
1306             for (j=0; j < i; j++) {
1307                 _cmsFree(ContextID, c16->Curves[j]);
1308             }
1309             _cmsFree(ContextID, c16->Curves);
1310             _cmsFree(ContextID, c16);
1311             return NULL;
1312         }
1313
1314         if (nElements == 256U) {
1315
1316             for (j=0; j < nElements; j++) {
1317
1318                 c16 ->Curves[i][j] = cmsEvalToneCurve16(G[i], FROM_8_TO_16(j));
1319             }
1320         }
1321         else {
1322
1323             for (j=0; j < nElements; j++) {
1324                 c16 ->Curves[i][j] = cmsEvalToneCurve16(G[i], (cmsUInt16Number) j);
1325             }
1326         }
1327     }
1328
1329     return c16;
1330 }
1331
1332 static
1333 void FastEvaluateCurves8(CMSREGISTER const cmsUInt16Number In[],
1334                          CMSREGISTER cmsUInt16Number Out[],
1335                          CMSREGISTER const void* D)
1336 {
1337     Curves16Data* Data = (Curves16Data*) D;
1338     int x;
1339     cmsUInt32Number i;
1340
1341     for (i=0; i < Data ->nCurves; i++) {
1342
1343          x = (In[i] >> 8);
1344          Out[i] = Data -> Curves[i][x];
1345     }
1346 }
1347
1348
1349 static
1350 void FastEvaluateCurves16(CMSREGISTER const cmsUInt16Number In[],
1351                           CMSREGISTER cmsUInt16Number Out[],
1352                           CMSREGISTER const void* D)
1353 {
1354     Curves16Data* Data = (Curves16Data*) D;
1355     cmsUInt32Number i;
1356
1357     for (i=0; i < Data ->nCurves; i++) {
1358          Out[i] = Data -> Curves[i][In[i]];
1359     }
1360 }
1361
1362
1363 static
1364 void FastIdentity16(CMSREGISTER const cmsUInt16Number In[],
1365                     CMSREGISTER cmsUInt16Number Out[],
1366                     CMSREGISTER const void* D)
1367 {
1368     cmsPipeline* Lut = (cmsPipeline*) D;
1369     cmsUInt32Number i;
1370
1371     for (i=0; i < Lut ->InputChannels; i++) {
1372          Out[i] = In[i];
1373     }
1374 }
1375
1376
1377 // If the target LUT holds only curves, the optimization procedure is to join all those
1378 // curves together. That only works on curves and does not work on matrices.
1379 static
1380 cmsBool OptimizeByJoiningCurves(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags)
1381 {
1382     cmsToneCurve** GammaTables = NULL;
1383     cmsFloat32Number InFloat[cmsMAXCHANNELS], OutFloat[cmsMAXCHANNELS];
1384     cmsUInt32Number i, j;
1385     cmsPipeline* Src = *Lut;
1386     cmsPipeline* Dest = NULL;
1387     cmsStage* mpe;
1388     cmsStage* ObtainedCurves = NULL;
1389
1390
1391     // This is a lossy optimization! does not apply in floating-point cases
1392     if (_cmsFormatterIsFloat(*InputFormat) || _cmsFormatterIsFloat(*OutputFormat)) return FALSE;
1393
1394     //  Only curves in this LUT?
1395     for (mpe = cmsPipelineGetPtrToFirstStage(Src);
1396          mpe != NULL;
1397          mpe = cmsStageNext(mpe)) {
1398             if (cmsStageType(mpe) != cmsSigCurveSetElemType) return FALSE;
1399     }
1400
1401     // Allocate an empty LUT
1402     Dest =  cmsPipelineAlloc(Src ->ContextID, Src ->InputChannels, Src ->OutputChannels);
1403     if (Dest == NULL) return FALSE;
1404
1405     // Create target curves
1406     GammaTables = (cmsToneCurve**) _cmsCalloc(Src ->ContextID, Src ->InputChannels, sizeof(cmsToneCurve*));
1407     if (GammaTables == NULL) goto Error;
1408
1409     for (i=0; i < Src ->InputChannels; i++) {
1410         GammaTables[i] = cmsBuildTabulatedToneCurve16(Src ->ContextID, PRELINEARIZATION_POINTS, NULL);
1411         if (GammaTables[i] == NULL) goto Error;
1412     }
1413
1414     // Compute 16 bit result by using floating point
1415     for (i=0; i < PRELINEARIZATION_POINTS; i++) {
1416
1417         for (j=0; j < Src ->InputChannels; j++)
1418             InFloat[j] = (cmsFloat32Number) ((cmsFloat64Number) i / (PRELINEARIZATION_POINTS - 1));
1419
1420         cmsPipelineEvalFloat(InFloat, OutFloat, Src);
1421
1422         for (j=0; j < Src ->InputChannels; j++)
1423             GammaTables[j] -> Table16[i] = _cmsQuickSaturateWord(OutFloat[j] * 65535.0);
1424     }
1425
1426     ObtainedCurves = cmsStageAllocToneCurves(Src ->ContextID, Src ->InputChannels, GammaTables);
1427     if (ObtainedCurves == NULL) goto Error;
1428
1429     for (i=0; i < Src ->InputChannels; i++) {
1430         cmsFreeToneCurve(GammaTables[i]);
1431         GammaTables[i] = NULL;
1432     }
1433
1434     if (GammaTables != NULL) {
1435         _cmsFree(Src->ContextID, GammaTables);
1436         GammaTables = NULL;
1437     }
1438
1439     // Maybe the curves are linear at the end
1440     if (!AllCurvesAreLinear(ObtainedCurves)) {
1441        _cmsStageToneCurvesData* Data;
1442
1443         if (!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, ObtainedCurves))
1444             goto Error;
1445         Data = (_cmsStageToneCurvesData*) cmsStageData(ObtainedCurves);
1446         ObtainedCurves = NULL;
1447
1448         // If the curves are to be applied in 8 bits, we can save memory
1449         if (_cmsFormatterIs8bit(*InputFormat)) {
1450              Curves16Data* c16 = CurvesAlloc(Dest ->ContextID, Data ->nCurves, 256, Data ->TheCurves);
1451
1452              if (c16 == NULL) goto Error;
1453              *dwFlags |= cmsFLAGS_NOCACHE;
1454             _cmsPipelineSetOptimizationParameters(Dest, FastEvaluateCurves8, c16, CurvesFree, CurvesDup);
1455
1456         }
1457         else {
1458              Curves16Data* c16 = CurvesAlloc(Dest ->ContextID, Data ->nCurves, 65536, Data ->TheCurves);
1459
1460              if (c16 == NULL) goto Error;
1461              *dwFlags |= cmsFLAGS_NOCACHE;
1462             _cmsPipelineSetOptimizationParameters(Dest, FastEvaluateCurves16, c16, CurvesFree, CurvesDup);
1463         }
1464     }
1465     else {
1466
1467         // LUT optimizes to nothing. Set the identity LUT
1468         cmsStageFree(ObtainedCurves);
1469         ObtainedCurves = NULL;
1470
1471         if (!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, cmsStageAllocIdentity(Dest ->ContextID, Src ->InputChannels)))
1472             goto Error;
1473
1474         *dwFlags |= cmsFLAGS_NOCACHE;
1475         _cmsPipelineSetOptimizationParameters(Dest, FastIdentity16, (void*) Dest, NULL, NULL);
1476     }
1477
1478     // We are done.
1479     cmsPipelineFree(Src);
1480     *Lut = Dest;
1481     return TRUE;
1482
1483 Error:
1484
1485     if (ObtainedCurves != NULL) cmsStageFree(ObtainedCurves);
1486     if (GammaTables != NULL) {
1487         for (i=0; i < Src ->InputChannels; i++) {
1488             if (GammaTables[i] != NULL) cmsFreeToneCurve(GammaTables[i]);
1489         }
1490
1491         _cmsFree(Src ->ContextID, GammaTables);
1492     }
1493
1494     if (Dest != NULL) cmsPipelineFree(Dest);
1495     return FALSE;
1496
1497     cmsUNUSED_PARAMETER(Intent);
1498     cmsUNUSED_PARAMETER(InputFormat);
1499     cmsUNUSED_PARAMETER(OutputFormat);
1500     cmsUNUSED_PARAMETER(dwFlags);
1501 }
1502
1503 // -------------------------------------------------------------------------------------------------------------------------------------
1504 // LUT is Shaper - Matrix - Matrix - Shaper, which is very frequent when combining two matrix-shaper profiles
1505
1506
1507 static
1508 void  FreeMatShaper(cmsContext ContextID, void* Data)
1509 {
1510     if (Data != NULL) _cmsFree(ContextID, Data);
1511 }
1512
1513 static
1514 void* DupMatShaper(cmsContext ContextID, const void* Data)
1515 {
1516     return _cmsDupMem(ContextID, Data, sizeof(MatShaper8Data));
1517 }
1518
1519
1520 // A fast matrix-shaper evaluator for 8 bits. This is a bit tricky since I'm using 1.14 signed fixed point
1521 // to accomplish some performance. Actually it takes 256x3 16 bits tables and 16385 x 3 tables of 8 bits,
1522 // in total about 50K, and the performance boost is huge!
1523 static CMS_NO_SANITIZE
1524 void MatShaperEval16(CMSREGISTER const cmsUInt16Number In[],
1525                      CMSREGISTER cmsUInt16Number Out[],
1526                      CMSREGISTER const void* D)
1527 {
1528     MatShaper8Data* p = (MatShaper8Data*) D;
1529     cmsS1Fixed14Number l1, l2, l3, r, g, b;
1530     cmsUInt32Number ri, gi, bi;
1531
1532     // In this case (and only in this case!) we can use this simplification since
1533     // In[] is assured to come from a 8 bit number. (a << 8 | a)
1534     ri = In[0] & 0xFFU;
1535     gi = In[1] & 0xFFU;
1536     bi = In[2] & 0xFFU;
1537
1538     // Across first shaper, which also converts to 1.14 fixed point
1539     r = p->Shaper1R[ri];
1540     g = p->Shaper1G[gi];
1541     b = p->Shaper1B[bi];
1542
1543     // Evaluate the matrix in 1.14 fixed point
1544     l1 =  (p->Mat[0][0] * r + p->Mat[0][1] * g + p->Mat[0][2] * b + p->Off[0] + 0x2000) >> 14;
1545     l2 =  (p->Mat[1][0] * r + p->Mat[1][1] * g + p->Mat[1][2] * b + p->Off[1] + 0x2000) >> 14;
1546     l3 =  (p->Mat[2][0] * r + p->Mat[2][1] * g + p->Mat[2][2] * b + p->Off[2] + 0x2000) >> 14;
1547
1548     // Now we have to clip to 0..1.0 range
1549     ri = (l1 < 0) ? 0 : ((l1 > 16384) ? 16384U : (cmsUInt32Number) l1);
1550     gi = (l2 < 0) ? 0 : ((l2 > 16384) ? 16384U : (cmsUInt32Number) l2);
1551     bi = (l3 < 0) ? 0 : ((l3 > 16384) ? 16384U : (cmsUInt32Number) l3);
1552
1553     // And across second shaper,
1554     Out[0] = p->Shaper2R[ri];
1555     Out[1] = p->Shaper2G[gi];
1556     Out[2] = p->Shaper2B[bi];
1557
1558 }
1559
1560 // This table converts from 8 bits to 1.14 after applying the curve
1561 static
1562 void FillFirstShaper(cmsS1Fixed14Number* Table, cmsToneCurve* Curve)
1563 {
1564     int i;
1565     cmsFloat32Number R, y;
1566
1567     for (i=0; i < 256; i++) {
1568
1569         R   = (cmsFloat32Number) (i / 255.0);
1570         y   = cmsEvalToneCurveFloat(Curve, R);
1571
1572         if (y < 131072.0)
1573             Table[i] = DOUBLE_TO_1FIXED14(y);
1574         else
1575             Table[i] = 0x7fffffff;
1576     }
1577 }
1578
1579 // This table converts form 1.14 (being 0x4000 the last entry) to 8 bits after applying the curve
1580 static
1581 void FillSecondShaper(cmsUInt16Number* Table, cmsToneCurve* Curve, cmsBool Is8BitsOutput)
1582 {
1583     int i;
1584     cmsFloat32Number R, Val;
1585
1586     for (i=0; i < 16385; i++) {
1587
1588         R   = (cmsFloat32Number) (i / 16384.0);
1589         Val = cmsEvalToneCurveFloat(Curve, R);    // Val comes 0..1.0
1590
1591         if (Val < 0)
1592             Val = 0;
1593
1594         if (Val > 1.0)
1595             Val = 1.0;
1596
1597         if (Is8BitsOutput) {
1598
1599             // If 8 bits output, we can optimize further by computing the / 257 part.
1600             // first we compute the resulting byte and then we store the byte times
1601             // 257. This quantization allows to round very quick by doing a >> 8, but
1602             // since the low byte is always equal to msb, we can do a & 0xff and this works!
1603             cmsUInt16Number w = _cmsQuickSaturateWord(Val * 65535.0);
1604             cmsUInt8Number  b = FROM_16_TO_8(w);
1605
1606             Table[i] = FROM_8_TO_16(b);
1607         }
1608         else Table[i]  = _cmsQuickSaturateWord(Val * 65535.0);
1609     }
1610 }
1611
1612 // Compute the matrix-shaper structure
1613 static
1614 cmsBool SetMatShaper(cmsPipeline* Dest, cmsToneCurve* Curve1[3], cmsMAT3* Mat, cmsVEC3* Off, cmsToneCurve* Curve2[3], cmsUInt32Number* OutputFormat)
1615 {
1616     MatShaper8Data* p;
1617     int i, j;
1618     cmsBool Is8Bits = _cmsFormatterIs8bit(*OutputFormat);
1619
1620     // Allocate a big chuck of memory to store precomputed tables
1621     p = (MatShaper8Data*) _cmsMalloc(Dest ->ContextID, sizeof(MatShaper8Data));
1622     if (p == NULL) return FALSE;
1623
1624     p -> ContextID = Dest -> ContextID;
1625
1626     // Precompute tables
1627     FillFirstShaper(p ->Shaper1R, Curve1[0]);
1628     FillFirstShaper(p ->Shaper1G, Curve1[1]);
1629     FillFirstShaper(p ->Shaper1B, Curve1[2]);
1630
1631     FillSecondShaper(p ->Shaper2R, Curve2[0], Is8Bits);
1632     FillSecondShaper(p ->Shaper2G, Curve2[1], Is8Bits);
1633     FillSecondShaper(p ->Shaper2B, Curve2[2], Is8Bits);
1634
1635     // Convert matrix to nFixed14. Note that those values may take more than 16 bits
1636     for (i=0; i < 3; i++) {
1637         for (j=0; j < 3; j++) {
1638             p ->Mat[i][j] = DOUBLE_TO_1FIXED14(Mat->v[i].n[j]);
1639         }
1640     }
1641
1642     for (i=0; i < 3; i++) {
1643
1644         if (Off == NULL) {
1645             p ->Off[i] = 0;
1646         }
1647         else {
1648             p ->Off[i] = DOUBLE_TO_1FIXED14(Off->n[i]);
1649         }
1650     }
1651
1652     // Mark as optimized for faster formatter
1653     if (Is8Bits)
1654         *OutputFormat |= OPTIMIZED_SH(1);
1655
1656     // Fill function pointers
1657     _cmsPipelineSetOptimizationParameters(Dest, MatShaperEval16, (void*) p, FreeMatShaper, DupMatShaper);
1658     return TRUE;
1659 }
1660
1661 //  8 bits on input allows matrix-shaper boot up to 25 Mpixels per second on RGB. That's fast!
1662 static
1663 cmsBool OptimizeMatrixShaper(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags)
1664 {
1665        cmsStage* Curve1, *Curve2;
1666        cmsStage* Matrix1, *Matrix2;
1667        cmsMAT3 res;
1668        cmsBool IdentityMat;
1669        cmsPipeline* Dest, *Src;
1670        cmsFloat64Number* Offset;
1671
1672        // Only works on RGB to RGB
1673        if (T_CHANNELS(*InputFormat) != 3 || T_CHANNELS(*OutputFormat) != 3) return FALSE;
1674
1675        // Only works on 8 bit input
1676        if (!_cmsFormatterIs8bit(*InputFormat)) return FALSE;
1677
1678        // Seems suitable, proceed
1679        Src = *Lut;
1680
1681        // Check for:
1682        //
1683        //    shaper-matrix-matrix-shaper
1684        //    shaper-matrix-shaper
1685        //
1686        // Both of those constructs are possible (first because abs. colorimetric).
1687        // additionally, In the first case, the input matrix offset should be zero.
1688
1689        IdentityMat = FALSE;
1690        if (cmsPipelineCheckAndRetreiveStages(Src, 4,
1691               cmsSigCurveSetElemType, cmsSigMatrixElemType, cmsSigMatrixElemType, cmsSigCurveSetElemType,
1692               &Curve1, &Matrix1, &Matrix2, &Curve2)) {
1693
1694               // Get both matrices
1695               _cmsStageMatrixData* Data1 = (_cmsStageMatrixData*)cmsStageData(Matrix1);
1696               _cmsStageMatrixData* Data2 = (_cmsStageMatrixData*)cmsStageData(Matrix2);
1697
1698               // Only RGB to RGB
1699               if (Matrix1->InputChannels != 3 || Matrix1->OutputChannels != 3 ||
1700                   Matrix2->InputChannels != 3 || Matrix2->OutputChannels != 3) return FALSE;
1701
1702               // Input offset should be zero
1703               if (Data1->Offset != NULL) return FALSE;
1704
1705               // Multiply both matrices to get the result
1706               _cmsMAT3per(&res, (cmsMAT3*)Data2->Double, (cmsMAT3*)Data1->Double);
1707
1708               // Only 2nd matrix has offset, or it is zero
1709               Offset = Data2->Offset;
1710
1711               // Now the result is in res + Data2 -> Offset. Maybe is a plain identity?
1712               if (_cmsMAT3isIdentity(&res) && Offset == NULL) {
1713
1714                      // We can get rid of full matrix
1715                      IdentityMat = TRUE;
1716               }
1717
1718        }
1719        else {
1720
1721               if (cmsPipelineCheckAndRetreiveStages(Src, 3,
1722                      cmsSigCurveSetElemType, cmsSigMatrixElemType, cmsSigCurveSetElemType,
1723                      &Curve1, &Matrix1, &Curve2)) {
1724
1725                      _cmsStageMatrixData* Data = (_cmsStageMatrixData*)cmsStageData(Matrix1);
1726
1727                      // Copy the matrix to our result
1728                      memcpy(&res, Data->Double, sizeof(res));
1729
1730                      // Preserve the Odffset (may be NULL as a zero offset)
1731                      Offset = Data->Offset;
1732
1733                      if (_cmsMAT3isIdentity(&res) && Offset == NULL) {
1734
1735                             // We can get rid of full matrix
1736                             IdentityMat = TRUE;
1737                      }
1738               }
1739               else
1740                      return FALSE; // Not optimizeable this time
1741
1742        }
1743
1744       // Allocate an empty LUT
1745     Dest =  cmsPipelineAlloc(Src ->ContextID, Src ->InputChannels, Src ->OutputChannels);
1746     if (!Dest) return FALSE;
1747
1748     // Assamble the new LUT
1749     if (!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, cmsStageDup(Curve1)))
1750         goto Error;
1751
1752     if (!IdentityMat) {
1753
1754            if (!cmsPipelineInsertStage(Dest, cmsAT_END, cmsStageAllocMatrix(Dest->ContextID, 3, 3, (const cmsFloat64Number*)&res, Offset)))
1755                   goto Error;
1756     }
1757
1758     if (!cmsPipelineInsertStage(Dest, cmsAT_END, cmsStageDup(Curve2)))
1759         goto Error;
1760
1761     // If identity on matrix, we can further optimize the curves, so call the join curves routine
1762     if (IdentityMat) {
1763
1764         OptimizeByJoiningCurves(&Dest, Intent, InputFormat, OutputFormat, dwFlags);
1765     }
1766     else {
1767         _cmsStageToneCurvesData* mpeC1 = (_cmsStageToneCurvesData*) cmsStageData(Curve1);
1768         _cmsStageToneCurvesData* mpeC2 = (_cmsStageToneCurvesData*) cmsStageData(Curve2);
1769
1770         // In this particular optimization, cache does not help as it takes more time to deal with
1771         // the cache that with the pixel handling
1772         *dwFlags |= cmsFLAGS_NOCACHE;
1773
1774         // Setup the optimizarion routines
1775         SetMatShaper(Dest, mpeC1 ->TheCurves, &res, (cmsVEC3*) Offset, mpeC2->TheCurves, OutputFormat);
1776     }
1777
1778     cmsPipelineFree(Src);
1779     *Lut = Dest;
1780     return TRUE;
1781 Error:
1782     // Leave Src unchanged
1783     cmsPipelineFree(Dest);
1784     return FALSE;
1785 }
1786
1787
1788 // -------------------------------------------------------------------------------------------------------------------------------------
1789 // Optimization plug-ins
1790
1791 // List of optimizations
1792 typedef struct _cmsOptimizationCollection_st {
1793
1794     _cmsOPToptimizeFn  OptimizePtr;
1795
1796     struct _cmsOptimizationCollection_st *Next;
1797
1798 } _cmsOptimizationCollection;
1799
1800
1801 // The built-in list. We currently implement 4 types of optimizations. Joining of curves, matrix-shaper, linearization and resampling
1802 static _cmsOptimizationCollection DefaultOptimization[] = {
1803
1804     { OptimizeByJoiningCurves,            &DefaultOptimization[1] },
1805     { OptimizeMatrixShaper,               &DefaultOptimization[2] },
1806     { OptimizeByComputingLinearization,   &DefaultOptimization[3] },
1807     { OptimizeByResampling,               NULL }
1808 };
1809
1810 // The linked list head
1811 _cmsOptimizationPluginChunkType _cmsOptimizationPluginChunk = { NULL };
1812
1813
1814 // Duplicates the zone of memory used by the plug-in in the new context
1815 static
1816 void DupPluginOptimizationList(struct _cmsContext_struct* ctx,
1817                                const struct _cmsContext_struct* src)
1818 {
1819    _cmsOptimizationPluginChunkType newHead = { NULL };
1820    _cmsOptimizationCollection*  entry;
1821    _cmsOptimizationCollection*  Anterior = NULL;
1822    _cmsOptimizationPluginChunkType* head = (_cmsOptimizationPluginChunkType*) src->chunks[OptimizationPlugin];
1823
1824     _cmsAssert(ctx != NULL);
1825     _cmsAssert(head != NULL);
1826
1827     // Walk the list copying all nodes
1828    for (entry = head->OptimizationCollection;
1829         entry != NULL;
1830         entry = entry ->Next) {
1831
1832             _cmsOptimizationCollection *newEntry = ( _cmsOptimizationCollection *) _cmsSubAllocDup(ctx ->MemPool, entry, sizeof(_cmsOptimizationCollection));
1833
1834             if (newEntry == NULL)
1835                 return;
1836
1837             // We want to keep the linked list order, so this is a little bit tricky
1838             newEntry -> Next = NULL;
1839             if (Anterior)
1840                 Anterior -> Next = newEntry;
1841
1842             Anterior = newEntry;
1843
1844             if (newHead.OptimizationCollection == NULL)
1845                 newHead.OptimizationCollection = newEntry;
1846     }
1847
1848   ctx ->chunks[OptimizationPlugin] = _cmsSubAllocDup(ctx->MemPool, &newHead, sizeof(_cmsOptimizationPluginChunkType));
1849 }
1850
1851 void  _cmsAllocOptimizationPluginChunk(struct _cmsContext_struct* ctx,
1852                                          const struct _cmsContext_struct* src)
1853 {
1854   if (src != NULL) {
1855
1856         // Copy all linked list
1857        DupPluginOptimizationList(ctx, src);
1858     }
1859     else {
1860         static _cmsOptimizationPluginChunkType OptimizationPluginChunkType = { NULL };
1861         ctx ->chunks[OptimizationPlugin] = _cmsSubAllocDup(ctx ->MemPool, &OptimizationPluginChunkType, sizeof(_cmsOptimizationPluginChunkType));
1862     }
1863 }
1864
1865
1866 // Register new ways to optimize
1867 cmsBool  _cmsRegisterOptimizationPlugin(cmsContext ContextID, cmsPluginBase* Data)
1868 {
1869     cmsPluginOptimization* Plugin = (cmsPluginOptimization*) Data;
1870     _cmsOptimizationPluginChunkType* ctx = ( _cmsOptimizationPluginChunkType*) _cmsContextGetClientChunk(ContextID, OptimizationPlugin);
1871     _cmsOptimizationCollection* fl;
1872
1873     if (Data == NULL) {
1874
1875         ctx->OptimizationCollection = NULL;
1876         return TRUE;
1877     }
1878
1879     // Optimizer callback is required
1880     if (Plugin ->OptimizePtr == NULL) return FALSE;
1881
1882     fl = (_cmsOptimizationCollection*) _cmsPluginMalloc(ContextID, sizeof(_cmsOptimizationCollection));
1883     if (fl == NULL) return FALSE;
1884
1885     // Copy the parameters
1886     fl ->OptimizePtr = Plugin ->OptimizePtr;
1887
1888     // Keep linked list
1889     fl ->Next = ctx->OptimizationCollection;
1890
1891     // Set the head
1892     ctx ->OptimizationCollection = fl;
1893
1894     // All is ok
1895     return TRUE;
1896 }
1897
1898 // The entry point for LUT optimization
1899 cmsBool CMSEXPORT _cmsOptimizePipeline(cmsContext ContextID,
1900                              cmsPipeline**    PtrLut,
1901                              cmsUInt32Number  Intent,
1902                              cmsUInt32Number* InputFormat,
1903                              cmsUInt32Number* OutputFormat,
1904                              cmsUInt32Number* dwFlags)
1905 {
1906     _cmsOptimizationPluginChunkType* ctx = ( _cmsOptimizationPluginChunkType*) _cmsContextGetClientChunk(ContextID, OptimizationPlugin);
1907     _cmsOptimizationCollection* Opts;
1908     cmsBool AnySuccess = FALSE;
1909     cmsStage* mpe;
1910
1911     // A CLUT is being asked, so force this specific optimization
1912     if (*dwFlags & cmsFLAGS_FORCE_CLUT) {
1913
1914         PreOptimize(*PtrLut);
1915         return OptimizeByResampling(PtrLut, Intent, InputFormat, OutputFormat, dwFlags);
1916     }
1917
1918     // Anything to optimize?
1919     if ((*PtrLut) ->Elements == NULL) {
1920         _cmsPipelineSetOptimizationParameters(*PtrLut, FastIdentity16, (void*) *PtrLut, NULL, NULL);
1921         return TRUE;
1922     }
1923
1924     // Named color pipelines cannot be optimized
1925     for (mpe = cmsPipelineGetPtrToFirstStage(*PtrLut);
1926         mpe != NULL;
1927         mpe = cmsStageNext(mpe)) {
1928         if (cmsStageType(mpe) == cmsSigNamedColorElemType) return FALSE;
1929     }
1930
1931     // Try to get rid of identities and trivial conversions.
1932     AnySuccess = PreOptimize(*PtrLut);
1933
1934     // After removal do we end with an identity?
1935     if ((*PtrLut) ->Elements == NULL) {
1936         _cmsPipelineSetOptimizationParameters(*PtrLut, FastIdentity16, (void*) *PtrLut, NULL, NULL);
1937         return TRUE;
1938     }
1939
1940     // Do not optimize, keep all precision
1941     if (*dwFlags & cmsFLAGS_NOOPTIMIZE)
1942         return FALSE;
1943
1944     // Try plug-in optimizations
1945     for (Opts = ctx->OptimizationCollection;
1946          Opts != NULL;
1947          Opts = Opts ->Next) {
1948
1949             // If one schema succeeded, we are done
1950             if (Opts ->OptimizePtr(PtrLut, Intent, InputFormat, OutputFormat, dwFlags)) {
1951
1952                 return TRUE;    // Optimized!
1953             }
1954     }
1955
1956    // Try built-in optimizations
1957     for (Opts = DefaultOptimization;
1958          Opts != NULL;
1959          Opts = Opts ->Next) {
1960
1961             if (Opts ->OptimizePtr(PtrLut, Intent, InputFormat, OutputFormat, dwFlags)) {
1962
1963                 return TRUE;
1964             }
1965     }
1966
1967     // Only simple optimizations succeeded
1968     return AnySuccess;
1969 }
1970
1971
1972