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) 1998-2023 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)
  62 typedef 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)
  76 typedef struct {
  77 
  78     cmsContext ContextID;
  79 
  80     // Number of channels
  81     cmsUInt32Number nInputs;
  82     cmsUInt32Number 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 
 100 typedef 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 
 104 typedef 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
 122 typedef struct {
 123 
 124     cmsContext ContextID;
 125 
 126     cmsUInt32Number nCurves;      // Number of curves
 127     cmsUInt32Number 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
 137 static
 138 void _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.
 147 static
 148 cmsBool _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
 167 static
 168 cmsBool _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 
 195 static
 196 cmsBool CloseEnoughFloat(cmsFloat64Number a, cmsFloat64Number b)
 197 {
 198        return fabs(b - a) < 0.00001f;
 199 }
 200 
 201 static
 202 cmsBool  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.
 216 static
 217 cmsBool _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 after 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.
 279 static
 280 cmsBool 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 
 319 static
 320 void Eval16nop1D(CMSREGISTER const cmsUInt16Number Input[],
 321                  CMSREGISTER cmsUInt16Number Output[],
 322                  CMSREGISTER const struct _cms_interp_struc* p)
 323 {
 324     Output[0] = Input[0];
 325 
 326     cmsUNUSED_PARAMETER(p);
 327 }
 328 
 329 static
 330 void PrelinEval16(CMSREGISTER const cmsUInt16Number Input[],
 331                   CMSREGISTER cmsUInt16Number Output[],
 332                   CMSREGISTER const void* D)
 333 {
 334     Prelin16Data* p16 = (Prelin16Data*) D;
 335     cmsUInt16Number  StageABC[MAX_INPUT_DIMENSIONS];
 336     cmsUInt16Number  StageDEF[cmsMAXCHANNELS];
 337     cmsUInt32Number 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 
 353 static
 354 void 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 
 364 static
 365 void* 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 
 379 static
 380 Prelin16Data* PrelinOpt16alloc(cmsContext ContextID,
 381                                const cmsInterpParams* ColorMap,
 382                                cmsUInt32Number nInputs, cmsToneCurve** In,
 383                                cmsUInt32Number nOutputs, cmsToneCurve** Out )
 384 {
 385     cmsUInt32Number 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     if (p16->EvalCurveOut16 == NULL)
 412     {
 413         _cmsFree(ContextID, p16);
 414         return NULL;
 415     }
 416 
 417     p16 -> ParamsCurveOut16 = (cmsInterpParams**) _cmsCalloc(ContextID, nOutputs, sizeof(cmsInterpParams* ));
 418     if (p16->ParamsCurveOut16 == NULL)
 419     {
 420 
 421         _cmsFree(ContextID, p16->EvalCurveOut16);
 422         _cmsFree(ContextID, p16);
 423         return NULL;
 424     }
 425 
 426     for (i=0; i < nOutputs; i++) {
 427 
 428         if (Out == NULL) {
 429             p16 ->ParamsCurveOut16[i] = NULL;
 430             p16 -> EvalCurveOut16[i] = Eval16nop1D;
 431         }
 432         else {
 433 
 434             p16 ->ParamsCurveOut16[i] = Out[i] ->InterpParams;
 435             p16 -> EvalCurveOut16[i] = p16 ->ParamsCurveOut16[i]->Interpolation.Lerp16;
 436         }
 437     }
 438 
 439     return p16;
 440 }
 441 
 442 
 443 
 444 // Resampling ---------------------------------------------------------------------------------
 445 
 446 #define PRELINEARIZATION_POINTS 4096
 447 
 448 // Sampler implemented by another LUT. This is a clean way to precalculate the devicelink 3D CLUT for
 449 // almost any transform. We use floating point precision and then convert from floating point to 16 bits.
 450 static
 451 cmsInt32Number XFormSampler16(CMSREGISTER const cmsUInt16Number In[],
 452                               CMSREGISTER cmsUInt16Number Out[],
 453                               CMSREGISTER void* Cargo)
 454 {
 455     cmsPipeline* Lut = (cmsPipeline*) Cargo;
 456     cmsFloat32Number InFloat[cmsMAXCHANNELS], OutFloat[cmsMAXCHANNELS];
 457     cmsUInt32Number i;
 458 
 459     _cmsAssert(Lut -> InputChannels < cmsMAXCHANNELS);
 460     _cmsAssert(Lut -> OutputChannels < cmsMAXCHANNELS);
 461 
 462     // From 16 bit to floating point
 463     for (i=0; i < Lut ->InputChannels; i++)
 464         InFloat[i] = (cmsFloat32Number) (In[i] / 65535.0);
 465 
 466     // Evaluate in floating point
 467     cmsPipelineEvalFloat(InFloat, OutFloat, Lut);
 468 
 469     // Back to 16 bits representation
 470     for (i=0; i < Lut ->OutputChannels; i++)
 471         Out[i] = _cmsQuickSaturateWord(OutFloat[i] * 65535.0);
 472 
 473     // Always succeed
 474     return TRUE;
 475 }
 476 
 477 // Try to see if the curves of a given MPE are linear
 478 static
 479 cmsBool AllCurvesAreLinear(cmsStage* mpe)
 480 {
 481     cmsToneCurve** Curves;
 482     cmsUInt32Number i, n;
 483 
 484     Curves = _cmsStageGetPtrToCurveSet(mpe);
 485     if (Curves == NULL) return FALSE;
 486 
 487     n = cmsStageOutputChannels(mpe);
 488 
 489     for (i=0; i < n; i++) {
 490         if (!cmsIsToneCurveLinear(Curves[i])) return FALSE;
 491     }
 492 
 493     return TRUE;
 494 }
 495 
 496 // This function replaces a specific node placed in "At" by the "Value" numbers. Its purpose
 497 // is to fix scum dot on broken profiles/transforms. Works on 1, 3 and 4 channels
 498 static
 499 cmsBool  PatchLUT(cmsStage* CLUT, cmsUInt16Number At[], cmsUInt16Number Value[],
 500                   cmsUInt32Number nChannelsOut, cmsUInt32Number nChannelsIn)
 501 {
 502     _cmsStageCLutData* Grid = (_cmsStageCLutData*) CLUT ->Data;
 503     cmsInterpParams* p16  = Grid ->Params;
 504     cmsFloat64Number px, py, pz, pw;
 505     int        x0, y0, z0, w0;
 506     int        i, index;
 507 
 508     if (CLUT -> Type != cmsSigCLutElemType) {
 509         cmsSignalError(CLUT->ContextID, cmsERROR_INTERNAL, "(internal) Attempt to PatchLUT on non-lut stage");
 510         return FALSE;
 511     }
 512 
 513     if (nChannelsIn == 4) {
 514 
 515         px = ((cmsFloat64Number) At[0] * (p16->Domain[0])) / 65535.0;
 516         py = ((cmsFloat64Number) At[1] * (p16->Domain[1])) / 65535.0;
 517         pz = ((cmsFloat64Number) At[2] * (p16->Domain[2])) / 65535.0;
 518         pw = ((cmsFloat64Number) At[3] * (p16->Domain[3])) / 65535.0;
 519 
 520         x0 = (int) floor(px);
 521         y0 = (int) floor(py);
 522         z0 = (int) floor(pz);
 523         w0 = (int) floor(pw);
 524 
 525         if (((px - x0) != 0) ||
 526             ((py - y0) != 0) ||
 527             ((pz - z0) != 0) ||
 528             ((pw - w0) != 0)) return FALSE; // Not on exact node
 529 
 530         index = (int) p16 -> opta[3] * x0 +
 531                 (int) p16 -> opta[2] * y0 +
 532                 (int) p16 -> opta[1] * z0 +
 533                 (int) p16 -> opta[0] * w0;
 534     }
 535     else
 536         if (nChannelsIn == 3) {
 537 
 538             px = ((cmsFloat64Number) At[0] * (p16->Domain[0])) / 65535.0;
 539             py = ((cmsFloat64Number) At[1] * (p16->Domain[1])) / 65535.0;
 540             pz = ((cmsFloat64Number) At[2] * (p16->Domain[2])) / 65535.0;
 541 
 542             x0 = (int) floor(px);
 543             y0 = (int) floor(py);
 544             z0 = (int) floor(pz);
 545 
 546             if (((px - x0) != 0) ||
 547                 ((py - y0) != 0) ||
 548                 ((pz - z0) != 0)) return FALSE;  // Not on exact node
 549 
 550             index = (int) p16 -> opta[2] * x0 +
 551                     (int) p16 -> opta[1] * y0 +
 552                     (int) p16 -> opta[0] * z0;
 553         }
 554         else
 555             if (nChannelsIn == 1) {
 556 
 557                 px = ((cmsFloat64Number) At[0] * (p16->Domain[0])) / 65535.0;
 558 
 559                 x0 = (int) floor(px);
 560 
 561                 if (((px - x0) != 0)) return FALSE; // Not on exact node
 562 
 563                 index = (int) p16 -> opta[0] * x0;
 564             }
 565             else {
 566                 cmsSignalError(CLUT->ContextID, cmsERROR_INTERNAL, "(internal) %d Channels are not supported on PatchLUT", nChannelsIn);
 567                 return FALSE;
 568             }
 569 
 570     for (i = 0; i < (int) nChannelsOut; i++)
 571         Grid->Tab.T[index + i] = Value[i];
 572 
 573     return TRUE;
 574 }
 575 
 576 // Auxiliary, to see if two values are equal or very different
 577 static
 578 cmsBool WhitesAreEqual(cmsUInt32Number n, cmsUInt16Number White1[], cmsUInt16Number White2[] )
 579 {
 580     cmsUInt32Number i;
 581 
 582     for (i=0; i < n; i++) {
 583 
 584         if (abs(White1[i] - White2[i]) > 0xf000) return TRUE;  // Values are so extremely different that the fixup should be avoided
 585         if (White1[i] != White2[i]) return FALSE;
 586     }
 587     return TRUE;
 588 }
 589 
 590 
 591 // Locate the node for the white point and fix it to pure white in order to avoid scum dot.
 592 static
 593 cmsBool FixWhiteMisalignment(cmsPipeline* Lut, cmsColorSpaceSignature EntryColorSpace, cmsColorSpaceSignature ExitColorSpace)
 594 {
 595     cmsUInt16Number *WhitePointIn, *WhitePointOut;
 596     cmsUInt16Number  WhiteIn[cmsMAXCHANNELS], WhiteOut[cmsMAXCHANNELS], ObtainedOut[cmsMAXCHANNELS];
 597     cmsUInt32Number i, nOuts, nIns;
 598     cmsStage *PreLin = NULL, *CLUT = NULL, *PostLin = NULL;
 599 
 600     if (!_cmsEndPointsBySpace(EntryColorSpace,
 601         &WhitePointIn, NULL, &nIns)) return FALSE;
 602 
 603     if (!_cmsEndPointsBySpace(ExitColorSpace,
 604         &WhitePointOut, NULL, &nOuts)) return FALSE;
 605 
 606     // It needs to be fixed?
 607     if (Lut ->InputChannels != nIns) return FALSE;
 608     if (Lut ->OutputChannels != nOuts) return FALSE;
 609 
 610     cmsPipelineEval16(WhitePointIn, ObtainedOut, Lut);
 611 
 612     if (WhitesAreEqual(nOuts, WhitePointOut, ObtainedOut)) return TRUE; // whites already match
 613 
 614     // Check if the LUT comes as Prelin, CLUT or Postlin. We allow all combinations
 615     if (!cmsPipelineCheckAndRetreiveStages(Lut, 3, cmsSigCurveSetElemType, cmsSigCLutElemType, cmsSigCurveSetElemType, &PreLin, &CLUT, &PostLin))
 616         if (!cmsPipelineCheckAndRetreiveStages(Lut, 2, cmsSigCurveSetElemType, cmsSigCLutElemType, &PreLin, &CLUT))
 617             if (!cmsPipelineCheckAndRetreiveStages(Lut, 2, cmsSigCLutElemType, cmsSigCurveSetElemType, &CLUT, &PostLin))
 618                 if (!cmsPipelineCheckAndRetreiveStages(Lut, 1, cmsSigCLutElemType, &CLUT))
 619                     return FALSE;
 620 
 621     // We need to interpolate white points of both, pre and post curves
 622     if (PreLin) {
 623 
 624         cmsToneCurve** Curves = _cmsStageGetPtrToCurveSet(PreLin);
 625 
 626         for (i=0; i < nIns; i++) {
 627             WhiteIn[i] = cmsEvalToneCurve16(Curves[i], WhitePointIn[i]);
 628         }
 629     }
 630     else {
 631         for (i=0; i < nIns; i++)
 632             WhiteIn[i] = WhitePointIn[i];
 633     }
 634 
 635     // If any post-linearization, we need to find how is represented white before the curve, do
 636     // a reverse interpolation in this case.
 637     if (PostLin) {
 638 
 639         cmsToneCurve** Curves = _cmsStageGetPtrToCurveSet(PostLin);
 640 
 641         for (i=0; i < nOuts; i++) {
 642 
 643             cmsToneCurve* InversePostLin = cmsReverseToneCurve(Curves[i]);
 644             if (InversePostLin == NULL) {
 645                 WhiteOut[i] = WhitePointOut[i];
 646 
 647             } else {
 648 
 649                 WhiteOut[i] = cmsEvalToneCurve16(InversePostLin, WhitePointOut[i]);
 650                 cmsFreeToneCurve(InversePostLin);
 651             }
 652         }
 653     }
 654     else {
 655         for (i=0; i < nOuts; i++)
 656             WhiteOut[i] = WhitePointOut[i];
 657     }
 658 
 659     // Ok, proceed with patching. May fail and we don't care if it fails
 660     PatchLUT(CLUT, WhiteIn, WhiteOut, nOuts, nIns);
 661 
 662     return TRUE;
 663 }
 664 
 665 // -----------------------------------------------------------------------------------------------------------------------------------------------
 666 // This function creates simple LUT from complex ones. The generated LUT has an optional set of
 667 // prelinearization curves, a CLUT of nGridPoints and optional postlinearization tables.
 668 // These curves have to exist in the original LUT in order to be used in the simplified output.
 669 // Caller may also use the flags to allow this feature.
 670 // LUTS with all curves will be simplified to a single curve. Parametric curves are lost.
 671 // This function should be used on 16-bits LUTS only, as floating point losses precision when simplified
 672 // -----------------------------------------------------------------------------------------------------------------------------------------------
 673 
 674 static
 675 cmsBool OptimizeByResampling(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags)
 676 {
 677     cmsPipeline* Src = NULL;
 678     cmsPipeline* Dest = NULL;
 679     cmsStage* CLUT;
 680     cmsStage *KeepPreLin = NULL, *KeepPostLin = NULL;
 681     cmsUInt32Number nGridPoints;
 682     cmsColorSpaceSignature ColorSpace, OutputColorSpace;
 683     cmsStage *NewPreLin = NULL;
 684     cmsStage *NewPostLin = NULL;
 685     _cmsStageCLutData* DataCLUT;
 686     cmsToneCurve** DataSetIn;
 687     cmsToneCurve** DataSetOut;
 688     Prelin16Data* p16;
 689 
 690     // This is a lossy optimization! does not apply in floating-point cases
 691     if (_cmsFormatterIsFloat(*InputFormat) || _cmsFormatterIsFloat(*OutputFormat)) return FALSE;
 692 
 693     ColorSpace       = _cmsICCcolorSpace((int) T_COLORSPACE(*InputFormat));
 694     OutputColorSpace = _cmsICCcolorSpace((int) T_COLORSPACE(*OutputFormat));
 695 
 696     // Color space must be specified
 697     if (ColorSpace == (cmsColorSpaceSignature)0 ||
 698         OutputColorSpace == (cmsColorSpaceSignature)0) return FALSE;
 699 
 700     nGridPoints = _cmsReasonableGridpointsByColorspace(ColorSpace, *dwFlags);
 701 
 702     // For empty LUTs, 2 points are enough
 703     if (cmsPipelineStageCount(*Lut) == 0)
 704         nGridPoints = 2;
 705 
 706     Src = *Lut;
 707 
 708     // Allocate an empty LUT
 709     Dest =  cmsPipelineAlloc(Src ->ContextID, Src ->InputChannels, Src ->OutputChannels);
 710     if (!Dest) return FALSE;
 711 
 712     // Prelinearization tables are kept unless indicated by flags
 713     if (*dwFlags & cmsFLAGS_CLUT_PRE_LINEARIZATION) {
 714 
 715         // Get a pointer to the prelinearization element
 716         cmsStage* PreLin = cmsPipelineGetPtrToFirstStage(Src);
 717 
 718         // Check if suitable
 719         if (PreLin && PreLin ->Type == cmsSigCurveSetElemType) {
 720 
 721             // Maybe this is a linear tram, so we can avoid the whole stuff
 722             if (!AllCurvesAreLinear(PreLin)) {
 723 
 724                 // All seems ok, proceed.
 725                 NewPreLin = cmsStageDup(PreLin);
 726                 if(!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, NewPreLin))
 727                     goto Error;
 728 
 729                 // Remove prelinearization. Since we have duplicated the curve
 730                 // in destination LUT, the sampling should be applied after this stage.
 731                 cmsPipelineUnlinkStage(Src, cmsAT_BEGIN, &KeepPreLin);
 732             }
 733         }
 734     }
 735 
 736     // Allocate the CLUT
 737     CLUT = cmsStageAllocCLut16bit(Src ->ContextID, nGridPoints, Src ->InputChannels, Src->OutputChannels, NULL);
 738     if (CLUT == NULL) goto Error;
 739 
 740     // Add the CLUT to the destination LUT
 741     if (!cmsPipelineInsertStage(Dest, cmsAT_END, CLUT)) {
 742         goto Error;
 743     }
 744 
 745     // Postlinearization tables are kept unless indicated by flags
 746     if (*dwFlags & cmsFLAGS_CLUT_POST_LINEARIZATION) {
 747 
 748         // Get a pointer to the postlinearization if present
 749         cmsStage* PostLin = cmsPipelineGetPtrToLastStage(Src);
 750 
 751         // Check if suitable
 752         if (PostLin && cmsStageType(PostLin) == cmsSigCurveSetElemType) {
 753 
 754             // Maybe this is a linear tram, so we can avoid the whole stuff
 755             if (!AllCurvesAreLinear(PostLin)) {
 756 
 757                 // All seems ok, proceed.
 758                 NewPostLin = cmsStageDup(PostLin);
 759                 if (!cmsPipelineInsertStage(Dest, cmsAT_END, NewPostLin))
 760                     goto Error;
 761 
 762                 // In destination LUT, the sampling should be applied after this stage.
 763                 cmsPipelineUnlinkStage(Src, cmsAT_END, &KeepPostLin);
 764             }
 765         }
 766     }
 767 
 768     // Now its time to do the sampling. We have to ignore pre/post linearization
 769     // The source LUT without pre/post curves is passed as parameter.
 770     if (!cmsStageSampleCLut16bit(CLUT, XFormSampler16, (void*) Src, 0)) {
 771 Error:
 772         // Ops, something went wrong, Restore stages
 773         if (KeepPreLin != NULL) {
 774             if (!cmsPipelineInsertStage(Src, cmsAT_BEGIN, KeepPreLin)) {
 775                 _cmsAssert(0); // This never happens
 776             }
 777         }
 778         if (KeepPostLin != NULL) {
 779             if (!cmsPipelineInsertStage(Src, cmsAT_END,   KeepPostLin)) {
 780                 _cmsAssert(0); // This never happens
 781             }
 782         }
 783         cmsPipelineFree(Dest);
 784         return FALSE;
 785     }
 786 
 787     // Done.
 788 
 789     if (KeepPreLin != NULL) cmsStageFree(KeepPreLin);
 790     if (KeepPostLin != NULL) cmsStageFree(KeepPostLin);
 791     cmsPipelineFree(Src);
 792 
 793     DataCLUT = (_cmsStageCLutData*) CLUT ->Data;
 794 
 795     if (NewPreLin == NULL) DataSetIn = NULL;
 796     else DataSetIn = ((_cmsStageToneCurvesData*) NewPreLin ->Data) ->TheCurves;
 797 
 798     if (NewPostLin == NULL) DataSetOut = NULL;
 799     else  DataSetOut = ((_cmsStageToneCurvesData*) NewPostLin ->Data) ->TheCurves;
 800 
 801 
 802     if (DataSetIn == NULL && DataSetOut == NULL) {
 803 
 804         _cmsPipelineSetOptimizationParameters(Dest, (_cmsPipelineEval16Fn) DataCLUT->Params->Interpolation.Lerp16, DataCLUT->Params, NULL, NULL);
 805     }
 806     else {
 807 
 808         p16 = PrelinOpt16alloc(Dest ->ContextID,
 809             DataCLUT ->Params,
 810             Dest ->InputChannels,
 811             DataSetIn,
 812             Dest ->OutputChannels,
 813             DataSetOut);
 814 
 815         _cmsPipelineSetOptimizationParameters(Dest, PrelinEval16, (void*) p16, PrelinOpt16free, Prelin16dup);
 816     }
 817 
 818 
 819     // Don't fix white on absolute colorimetric
 820     if (Intent == INTENT_ABSOLUTE_COLORIMETRIC)
 821         *dwFlags |= cmsFLAGS_NOWHITEONWHITEFIXUP;
 822 
 823     if (!(*dwFlags & cmsFLAGS_NOWHITEONWHITEFIXUP)) {
 824 
 825         FixWhiteMisalignment(Dest, ColorSpace, OutputColorSpace);
 826     }
 827 
 828     *Lut = Dest;
 829     return TRUE;
 830 
 831     cmsUNUSED_PARAMETER(Intent);
 832 }
 833 
 834 
 835 // -----------------------------------------------------------------------------------------------------------------------------------------------
 836 // Fixes the gamma balancing of transform. This is described in my paper "Prelinearization Stages on
 837 // Color-Management Application-Specific Integrated Circuits (ASICs)" presented at NIP24. It only works
 838 // for RGB transforms. See the paper for more details
 839 // -----------------------------------------------------------------------------------------------------------------------------------------------
 840 
 841 
 842 // Normalize endpoints by slope limiting max and min. This assures endpoints as well.
 843 // Descending curves are handled as well.
 844 static
 845 void SlopeLimiting(cmsToneCurve* g)
 846 {
 847     int BeginVal, EndVal;
 848     int AtBegin = (int) floor((cmsFloat64Number) g ->nEntries * 0.02 + 0.5);   // Cutoff at 2%
 849     int AtEnd   = (int) g ->nEntries - AtBegin - 1;                                  // And 98%
 850     cmsFloat64Number Val, Slope, beta;
 851     int i;
 852 
 853     if (cmsIsToneCurveDescending(g)) {
 854         BeginVal = 0xffff; EndVal = 0;
 855     }
 856     else {
 857         BeginVal = 0; EndVal = 0xffff;
 858     }
 859 
 860     // Compute slope and offset for begin of curve
 861     Val   = g ->Table16[AtBegin];
 862     Slope = (Val - BeginVal) / AtBegin;
 863     beta  = Val - Slope * AtBegin;
 864 
 865     for (i=0; i < AtBegin; i++)
 866         g ->Table16[i] = _cmsQuickSaturateWord(i * Slope + beta);
 867 
 868     // Compute slope and offset for the end
 869     Val   = g ->Table16[AtEnd];
 870     Slope = (EndVal - Val) / AtBegin;   // AtBegin holds the X interval, which is same in both cases
 871     beta  = Val - Slope * AtEnd;
 872 
 873     for (i = AtEnd; i < (int) g ->nEntries; i++)
 874         g ->Table16[i] = _cmsQuickSaturateWord(i * Slope + beta);
 875 }
 876 
 877 
 878 // Precomputes tables for 8-bit on input devicelink.
 879 static
 880 Prelin8Data* PrelinOpt8alloc(cmsContext ContextID, const cmsInterpParams* p, cmsToneCurve* G[3])
 881 {
 882     int i;
 883     cmsUInt16Number Input[3];
 884     cmsS15Fixed16Number v1, v2, v3;
 885     Prelin8Data* p8;
 886 
 887     p8 = (Prelin8Data*)_cmsMallocZero(ContextID, sizeof(Prelin8Data));
 888     if (p8 == NULL) return NULL;
 889 
 890     // Since this only works for 8 bit input, values comes always as x * 257,
 891     // we can safely take msb byte (x << 8 + x)
 892 
 893     for (i=0; i < 256; i++) {
 894 
 895         if (G != NULL) {
 896 
 897             // Get 16-bit representation
 898             Input[0] = cmsEvalToneCurve16(G[0], FROM_8_TO_16(i));
 899             Input[1] = cmsEvalToneCurve16(G[1], FROM_8_TO_16(i));
 900             Input[2] = cmsEvalToneCurve16(G[2], FROM_8_TO_16(i));
 901         }
 902         else {
 903             Input[0] = FROM_8_TO_16(i);
 904             Input[1] = FROM_8_TO_16(i);
 905             Input[2] = FROM_8_TO_16(i);
 906         }
 907 
 908 
 909         // Move to 0..1.0 in fixed domain
 910         v1 = _cmsToFixedDomain((int) (Input[0] * p -> Domain[0]));
 911         v2 = _cmsToFixedDomain((int) (Input[1] * p -> Domain[1]));
 912         v3 = _cmsToFixedDomain((int) (Input[2] * p -> Domain[2]));
 913 
 914         // Store the precalculated table of nodes
 915         p8 ->X0[i] = (p->opta[2] * FIXED_TO_INT(v1));
 916         p8 ->Y0[i] = (p->opta[1] * FIXED_TO_INT(v2));
 917         p8 ->Z0[i] = (p->opta[0] * FIXED_TO_INT(v3));
 918 
 919         // Store the precalculated table of offsets
 920         p8 ->rx[i] = (cmsUInt16Number) FIXED_REST_TO_INT(v1);
 921         p8 ->ry[i] = (cmsUInt16Number) FIXED_REST_TO_INT(v2);
 922         p8 ->rz[i] = (cmsUInt16Number) FIXED_REST_TO_INT(v3);
 923     }
 924 
 925     p8 ->ContextID = ContextID;
 926     p8 ->p = p;
 927 
 928     return p8;
 929 }
 930 
 931 static
 932 void Prelin8free(cmsContext ContextID, void* ptr)
 933 {
 934     _cmsFree(ContextID, ptr);
 935 }
 936 
 937 static
 938 void* Prelin8dup(cmsContext ContextID, const void* ptr)
 939 {
 940     return _cmsDupMem(ContextID, ptr, sizeof(Prelin8Data));
 941 }
 942 
 943 
 944 
 945 // A optimized interpolation for 8-bit input.
 946 #define DENS(i,j,k) (LutTable[(i)+(j)+(k)+OutChan])
 947 static CMS_NO_SANITIZE
 948 void PrelinEval8(CMSREGISTER const cmsUInt16Number Input[],
 949                  CMSREGISTER cmsUInt16Number Output[],
 950                  CMSREGISTER const void* D)
 951 {
 952 
 953     cmsUInt8Number         r, g, b;
 954     cmsS15Fixed16Number    rx, ry, rz;
 955     cmsS15Fixed16Number    c0, c1, c2, c3, Rest;
 956     int                    OutChan;
 957     CMSREGISTER cmsS15Fixed16Number X0, X1, Y0, Y1, Z0, Z1;
 958     Prelin8Data* p8 = (Prelin8Data*) D;
 959     CMSREGISTER const cmsInterpParams* p = p8 ->p;
 960     int                    TotalOut = (int) p -> nOutputs;
 961     const cmsUInt16Number* LutTable = (const cmsUInt16Number*) p->Table;
 962 
 963     r = (cmsUInt8Number) (Input[0] >> 8);
 964     g = (cmsUInt8Number) (Input[1] >> 8);
 965     b = (cmsUInt8Number) (Input[2] >> 8);
 966 
 967     X0 = (cmsS15Fixed16Number) p8->X0[r];
 968     Y0 = (cmsS15Fixed16Number) p8->Y0[g];
 969     Z0 = (cmsS15Fixed16Number) p8->Z0[b];
 970 
 971     rx = p8 ->rx[r];
 972     ry = p8 ->ry[g];
 973     rz = p8 ->rz[b];
 974 
 975     X1 = X0 + (cmsS15Fixed16Number)((rx == 0) ? 0 :  p ->opta[2]);
 976     Y1 = Y0 + (cmsS15Fixed16Number)((ry == 0) ? 0 :  p ->opta[1]);
 977     Z1 = Z0 + (cmsS15Fixed16Number)((rz == 0) ? 0 :  p ->opta[0]);
 978 
 979 
 980     // These are the 6 Tetrahedral
 981     for (OutChan=0; OutChan < TotalOut; OutChan++) {
 982 
 983         c0 = DENS(X0, Y0, Z0);
 984 
 985         if (rx >= ry && ry >= rz)
 986         {
 987             c1 = DENS(X1, Y0, Z0) - c0;
 988             c2 = DENS(X1, Y1, Z0) - DENS(X1, Y0, Z0);
 989             c3 = DENS(X1, Y1, Z1) - DENS(X1, Y1, Z0);
 990         }
 991         else
 992             if (rx >= rz && rz >= ry)
 993             {
 994                 c1 = DENS(X1, Y0, Z0) - c0;
 995                 c2 = DENS(X1, Y1, Z1) - DENS(X1, Y0, Z1);
 996                 c3 = DENS(X1, Y0, Z1) - DENS(X1, Y0, Z0);
 997             }
 998             else
 999                 if (rz >= rx && rx >= ry)
1000                 {
1001                     c1 = DENS(X1, Y0, Z1) - DENS(X0, Y0, Z1);
1002                     c2 = DENS(X1, Y1, Z1) - DENS(X1, Y0, Z1);
1003                     c3 = DENS(X0, Y0, Z1) - c0;
1004                 }
1005                 else
1006                     if (ry >= rx && rx >= rz)
1007                     {
1008                         c1 = DENS(X1, Y1, Z0) - DENS(X0, Y1, Z0);
1009                         c2 = DENS(X0, Y1, Z0) - c0;
1010                         c3 = DENS(X1, Y1, Z1) - DENS(X1, Y1, Z0);
1011                     }
1012                     else
1013                         if (ry >= rz && rz >= rx)
1014                         {
1015                             c1 = DENS(X1, Y1, Z1) - DENS(X0, Y1, Z1);
1016                             c2 = DENS(X0, Y1, Z0) - c0;
1017                             c3 = DENS(X0, Y1, Z1) - DENS(X0, Y1, Z0);
1018                         }
1019                         else
1020                             if (rz >= ry && ry >= rx)
1021                             {
1022                                 c1 = DENS(X1, Y1, Z1) - DENS(X0, Y1, Z1);
1023                                 c2 = DENS(X0, Y1, Z1) - DENS(X0, Y0, Z1);
1024                                 c3 = DENS(X0, Y0, Z1) - c0;
1025                             }
1026                             else  {
1027                                 c1 = c2 = c3 = 0;
1028                             }
1029 
1030         Rest = c1 * rx + c2 * ry + c3 * rz + 0x8001;
1031         Output[OutChan] = (cmsUInt16Number) (c0 + ((Rest + (Rest >> 16)) >> 16));
1032 
1033     }
1034 }
1035 
1036 #undef DENS
1037 
1038 
1039 // Curves that contain wide empty areas are not optimizeable
1040 static
1041 cmsBool IsDegenerated(const cmsToneCurve* g)
1042 {
1043     cmsUInt32Number i, Zeros = 0, Poles = 0;
1044     cmsUInt32Number nEntries = g ->nEntries;
1045 
1046     for (i=0; i < nEntries; i++) {
1047 
1048         if (g ->Table16[i] == 0x0000) Zeros++;
1049         if (g ->Table16[i] == 0xffff) Poles++;
1050     }
1051 
1052     if (Zeros == 1 && Poles == 1) return FALSE;  // For linear tables
1053     if (Zeros > (nEntries / 20)) return TRUE;  // Degenerated, many zeros
1054     if (Poles > (nEntries / 20)) return TRUE;  // Degenerated, many poles
1055 
1056     return FALSE;
1057 }
1058 
1059 // --------------------------------------------------------------------------------------------------------------
1060 // We need xput over here
1061 
1062 static
1063 cmsBool OptimizeByComputingLinearization(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags)
1064 {
1065     cmsPipeline* OriginalLut;
1066     cmsUInt32Number nGridPoints;
1067     cmsToneCurve *Trans[cmsMAXCHANNELS], *TransReverse[cmsMAXCHANNELS];
1068     cmsUInt32Number t, i;
1069     cmsFloat32Number v, In[cmsMAXCHANNELS], Out[cmsMAXCHANNELS];
1070     cmsBool lIsSuitable, lIsLinear;
1071     cmsPipeline* OptimizedLUT = NULL, *LutPlusCurves = NULL;
1072     cmsStage* OptimizedCLUTmpe;
1073     cmsColorSpaceSignature ColorSpace, OutputColorSpace;
1074     cmsStage* OptimizedPrelinMpe;
1075     cmsToneCurve** OptimizedPrelinCurves;
1076     _cmsStageCLutData* OptimizedPrelinCLUT;
1077 
1078 
1079     // This is a lossy optimization! does not apply in floating-point cases
1080     if (_cmsFormatterIsFloat(*InputFormat) || _cmsFormatterIsFloat(*OutputFormat)) return FALSE;
1081 
1082     // Only on chunky RGB
1083     if (T_COLORSPACE(*InputFormat)  != PT_RGB) return FALSE;
1084     if (T_PLANAR(*InputFormat)) return FALSE;
1085 
1086     if (T_COLORSPACE(*OutputFormat) != PT_RGB) return FALSE;
1087     if (T_PLANAR(*OutputFormat)) return FALSE;
1088 
1089     // On 16 bits, user has to specify the feature
1090     if (!_cmsFormatterIs8bit(*InputFormat)) {
1091         if (!(*dwFlags & cmsFLAGS_CLUT_PRE_LINEARIZATION)) return FALSE;
1092     }
1093 
1094     OriginalLut = *Lut;
1095 
1096     ColorSpace       = _cmsICCcolorSpace((int) T_COLORSPACE(*InputFormat));
1097     OutputColorSpace = _cmsICCcolorSpace((int) T_COLORSPACE(*OutputFormat));
1098 
1099     // Color space must be specified
1100     if (ColorSpace == (cmsColorSpaceSignature)0 ||
1101         OutputColorSpace == (cmsColorSpaceSignature)0) return FALSE;
1102 
1103     nGridPoints      = _cmsReasonableGridpointsByColorspace(ColorSpace, *dwFlags);
1104 
1105     // Empty gamma containers
1106     memset(Trans, 0, sizeof(Trans));
1107     memset(TransReverse, 0, sizeof(TransReverse));
1108 
1109     // If the last stage of the original lut are curves, and those curves are
1110     // degenerated, it is likely the transform is squeezing and clipping
1111     // the output from previous CLUT. We cannot optimize this case
1112     {
1113         cmsStage* last = cmsPipelineGetPtrToLastStage(OriginalLut);
1114 
1115         if (last == NULL) goto Error;
1116         if (cmsStageType(last) == cmsSigCurveSetElemType) {
1117 
1118             _cmsStageToneCurvesData* Data = (_cmsStageToneCurvesData*)cmsStageData(last);
1119             for (i = 0; i < Data->nCurves; i++) {
1120                 if (IsDegenerated(Data->TheCurves[i]))
1121                     goto Error;
1122             }
1123         }
1124     }
1125 
1126     for (t = 0; t < OriginalLut ->InputChannels; t++) {
1127         Trans[t] = cmsBuildTabulatedToneCurve16(OriginalLut ->ContextID, PRELINEARIZATION_POINTS, NULL);
1128         if (Trans[t] == NULL) goto Error;
1129     }
1130 
1131     // Populate the curves
1132     for (i=0; i < PRELINEARIZATION_POINTS; i++) {
1133 
1134         v = (cmsFloat32Number) ((cmsFloat64Number) i / (PRELINEARIZATION_POINTS - 1));
1135 
1136         // Feed input with a gray ramp
1137         for (t=0; t < OriginalLut ->InputChannels; t++)
1138             In[t] = v;
1139 
1140         // Evaluate the gray value
1141         cmsPipelineEvalFloat(In, Out, OriginalLut);
1142 
1143         // Store result in curve
1144         for (t=0; t < OriginalLut ->InputChannels; t++)
1145             Trans[t] ->Table16[i] = _cmsQuickSaturateWord(Out[t] * 65535.0);
1146     }
1147 
1148     // Slope-limit the obtained curves
1149     for (t = 0; t < OriginalLut ->InputChannels; t++)
1150         SlopeLimiting(Trans[t]);
1151 
1152     // Check for validity
1153     lIsSuitable = TRUE;
1154     lIsLinear   = TRUE;
1155     for (t=0; (lIsSuitable && (t < OriginalLut ->InputChannels)); t++) {
1156 
1157         // Exclude if already linear
1158         if (!cmsIsToneCurveLinear(Trans[t]))
1159             lIsLinear = FALSE;
1160 
1161         // Exclude if non-monotonic
1162         if (!cmsIsToneCurveMonotonic(Trans[t]))
1163             lIsSuitable = FALSE;
1164 
1165         if (IsDegenerated(Trans[t]))
1166             lIsSuitable = FALSE;
1167     }
1168 
1169     // If it is not suitable, just quit
1170     if (!lIsSuitable) goto Error;
1171 
1172     // Invert curves if possible
1173     for (t = 0; t < OriginalLut ->InputChannels; t++) {
1174         TransReverse[t] = cmsReverseToneCurveEx(PRELINEARIZATION_POINTS, Trans[t]);
1175         if (TransReverse[t] == NULL) goto Error;
1176     }
1177 
1178     // Now inset the reversed curves at the begin of transform
1179     LutPlusCurves = cmsPipelineDup(OriginalLut);
1180     if (LutPlusCurves == NULL) goto Error;
1181 
1182     if (!cmsPipelineInsertStage(LutPlusCurves, cmsAT_BEGIN, cmsStageAllocToneCurves(OriginalLut ->ContextID, OriginalLut ->InputChannels, TransReverse)))
1183         goto Error;
1184 
1185     // Create the result LUT
1186     OptimizedLUT = cmsPipelineAlloc(OriginalLut ->ContextID, OriginalLut ->InputChannels, OriginalLut ->OutputChannels);
1187     if (OptimizedLUT == NULL) goto Error;
1188 
1189     OptimizedPrelinMpe = cmsStageAllocToneCurves(OriginalLut ->ContextID, OriginalLut ->InputChannels, Trans);
1190 
1191     // Create and insert the curves at the beginning
1192     if (!cmsPipelineInsertStage(OptimizedLUT, cmsAT_BEGIN, OptimizedPrelinMpe))
1193         goto Error;
1194 
1195     // Allocate the CLUT for result
1196     OptimizedCLUTmpe = cmsStageAllocCLut16bit(OriginalLut ->ContextID, nGridPoints, OriginalLut ->InputChannels, OriginalLut ->OutputChannels, NULL);
1197 
1198     // Add the CLUT to the destination LUT
1199     if (!cmsPipelineInsertStage(OptimizedLUT, cmsAT_END, OptimizedCLUTmpe))
1200         goto Error;
1201 
1202     // Resample the LUT
1203     if (!cmsStageSampleCLut16bit(OptimizedCLUTmpe, XFormSampler16, (void*) LutPlusCurves, 0)) goto Error;
1204 
1205     // Free resources
1206     for (t = 0; t < OriginalLut ->InputChannels; t++) {
1207 
1208         if (Trans[t]) cmsFreeToneCurve(Trans[t]);
1209         if (TransReverse[t]) cmsFreeToneCurve(TransReverse[t]);
1210     }
1211 
1212     cmsPipelineFree(LutPlusCurves);
1213 
1214 
1215     OptimizedPrelinCurves = _cmsStageGetPtrToCurveSet(OptimizedPrelinMpe);
1216     OptimizedPrelinCLUT   = (_cmsStageCLutData*) OptimizedCLUTmpe ->Data;
1217 
1218     // Set the evaluator if 8-bit
1219     if (_cmsFormatterIs8bit(*InputFormat)) {
1220 
1221         Prelin8Data* p8 = PrelinOpt8alloc(OptimizedLUT ->ContextID,
1222                                                 OptimizedPrelinCLUT ->Params,
1223                                                 OptimizedPrelinCurves);
1224         if (p8 == NULL) return FALSE;
1225 
1226         _cmsPipelineSetOptimizationParameters(OptimizedLUT, PrelinEval8, (void*) p8, Prelin8free, Prelin8dup);
1227 
1228     }
1229     else
1230     {
1231         Prelin16Data* p16 = PrelinOpt16alloc(OptimizedLUT ->ContextID,
1232             OptimizedPrelinCLUT ->Params,
1233             3, OptimizedPrelinCurves, 3, NULL);
1234         if (p16 == NULL) return FALSE;
1235 
1236         _cmsPipelineSetOptimizationParameters(OptimizedLUT, PrelinEval16, (void*) p16, PrelinOpt16free, Prelin16dup);
1237 
1238     }
1239 
1240     // Don't fix white on absolute colorimetric
1241     if (Intent == INTENT_ABSOLUTE_COLORIMETRIC)
1242         *dwFlags |= cmsFLAGS_NOWHITEONWHITEFIXUP;
1243 
1244     if (!(*dwFlags & cmsFLAGS_NOWHITEONWHITEFIXUP)) {
1245 
1246         if (!FixWhiteMisalignment(OptimizedLUT, ColorSpace, OutputColorSpace)) {
1247 
1248             return FALSE;
1249         }
1250     }
1251 
1252     // And return the obtained LUT
1253 
1254     cmsPipelineFree(OriginalLut);
1255     *Lut = OptimizedLUT;
1256     return TRUE;
1257 
1258 Error:
1259 
1260     for (t = 0; t < OriginalLut ->InputChannels; t++) {
1261 
1262         if (Trans[t]) cmsFreeToneCurve(Trans[t]);
1263         if (TransReverse[t]) cmsFreeToneCurve(TransReverse[t]);
1264     }
1265 
1266     if (LutPlusCurves != NULL) cmsPipelineFree(LutPlusCurves);
1267     if (OptimizedLUT != NULL) cmsPipelineFree(OptimizedLUT);
1268 
1269     return FALSE;
1270 
1271     cmsUNUSED_PARAMETER(Intent);
1272     cmsUNUSED_PARAMETER(lIsLinear);
1273 }
1274 
1275 
1276 // Curves optimizer ------------------------------------------------------------------------------------------------------------------
1277 
1278 static
1279 void CurvesFree(cmsContext ContextID, void* ptr)
1280 {
1281      Curves16Data* Data = (Curves16Data*) ptr;
1282      cmsUInt32Number i;
1283 
1284      for (i=0; i < Data -> nCurves; i++) {
1285 
1286          _cmsFree(ContextID, Data ->Curves[i]);
1287      }
1288 
1289      _cmsFree(ContextID, Data ->Curves);
1290      _cmsFree(ContextID, ptr);
1291 }
1292 
1293 static
1294 void* CurvesDup(cmsContext ContextID, const void* ptr)
1295 {
1296     Curves16Data* Data = (Curves16Data*)_cmsDupMem(ContextID, ptr, sizeof(Curves16Data));
1297     cmsUInt32Number i;
1298 
1299     if (Data == NULL) return NULL;
1300 
1301     Data->Curves = (cmsUInt16Number**) _cmsDupMem(ContextID, Data->Curves, Data->nCurves * sizeof(cmsUInt16Number*));
1302 
1303     for (i=0; i < Data -> nCurves; i++) {
1304         Data->Curves[i] = (cmsUInt16Number*) _cmsDupMem(ContextID, Data->Curves[i], Data->nElements * sizeof(cmsUInt16Number));
1305     }
1306 
1307     return (void*) Data;
1308 }
1309 
1310 // Precomputes tables for 8-bit on input devicelink.
1311 static
1312 Curves16Data* CurvesAlloc(cmsContext ContextID, cmsUInt32Number nCurves, cmsUInt32Number nElements, cmsToneCurve** G)
1313 {
1314     cmsUInt32Number i, j;
1315     Curves16Data* c16;
1316 
1317     c16 = (Curves16Data*)_cmsMallocZero(ContextID, sizeof(Curves16Data));
1318     if (c16 == NULL) return NULL;
1319 
1320     c16 ->nCurves = nCurves;
1321     c16 ->nElements = nElements;
1322 
1323     c16->Curves = (cmsUInt16Number**) _cmsCalloc(ContextID, nCurves, sizeof(cmsUInt16Number*));
1324     if (c16->Curves == NULL) {
1325         _cmsFree(ContextID, c16);
1326         return NULL;
1327     }
1328 
1329     for (i=0; i < nCurves; i++) {
1330 
1331         c16->Curves[i] = (cmsUInt16Number*) _cmsCalloc(ContextID, nElements, sizeof(cmsUInt16Number));
1332 
1333         if (c16->Curves[i] == NULL) {
1334 
1335             for (j=0; j < i; j++) {
1336                 _cmsFree(ContextID, c16->Curves[j]);
1337             }
1338             _cmsFree(ContextID, c16->Curves);
1339             _cmsFree(ContextID, c16);
1340             return NULL;
1341         }
1342 
1343         if (nElements == 256U) {
1344 
1345             for (j=0; j < nElements; j++) {
1346 
1347                 c16 ->Curves[i][j] = cmsEvalToneCurve16(G[i], FROM_8_TO_16(j));
1348             }
1349         }
1350         else {
1351 
1352             for (j=0; j < nElements; j++) {
1353                 c16 ->Curves[i][j] = cmsEvalToneCurve16(G[i], (cmsUInt16Number) j);
1354             }
1355         }
1356     }
1357 
1358     return c16;
1359 }
1360 
1361 static
1362 void FastEvaluateCurves8(CMSREGISTER const cmsUInt16Number In[],
1363                          CMSREGISTER cmsUInt16Number Out[],
1364                          CMSREGISTER const void* D)
1365 {
1366     Curves16Data* Data = (Curves16Data*) D;
1367     int x;
1368     cmsUInt32Number i;
1369 
1370     for (i=0; i < Data ->nCurves; i++) {
1371 
1372          x = (In[i] >> 8);
1373          Out[i] = Data -> Curves[i][x];
1374     }
1375 }
1376 
1377 
1378 static
1379 void FastEvaluateCurves16(CMSREGISTER const cmsUInt16Number In[],
1380                           CMSREGISTER cmsUInt16Number Out[],
1381                           CMSREGISTER const void* D)
1382 {
1383     Curves16Data* Data = (Curves16Data*) D;
1384     cmsUInt32Number i;
1385 
1386     for (i=0; i < Data ->nCurves; i++) {
1387          Out[i] = Data -> Curves[i][In[i]];
1388     }
1389 }
1390 
1391 
1392 static
1393 void FastIdentity16(CMSREGISTER const cmsUInt16Number In[],
1394                     CMSREGISTER cmsUInt16Number Out[],
1395                     CMSREGISTER const void* D)
1396 {
1397     cmsPipeline* Lut = (cmsPipeline*) D;
1398     cmsUInt32Number i;
1399 
1400     for (i=0; i < Lut ->InputChannels; i++) {
1401          Out[i] = In[i];
1402     }
1403 }
1404 
1405 
1406 // If the target LUT holds only curves, the optimization procedure is to join all those
1407 // curves together. That only works on curves and does not work on matrices.
1408 static
1409 cmsBool OptimizeByJoiningCurves(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags)
1410 {
1411     cmsToneCurve** GammaTables = NULL;
1412     cmsFloat32Number InFloat[cmsMAXCHANNELS], OutFloat[cmsMAXCHANNELS];
1413     cmsUInt32Number i, j;
1414     cmsPipeline* Src = *Lut;
1415     cmsPipeline* Dest = NULL;
1416     cmsStage* mpe;
1417     cmsStage* ObtainedCurves = NULL;
1418 
1419 
1420     // This is a lossy optimization! does not apply in floating-point cases
1421     if (_cmsFormatterIsFloat(*InputFormat) || _cmsFormatterIsFloat(*OutputFormat)) return FALSE;
1422 
1423     //  Only curves in this LUT?
1424     for (mpe = cmsPipelineGetPtrToFirstStage(Src);
1425          mpe != NULL;
1426          mpe = cmsStageNext(mpe)) {
1427             if (cmsStageType(mpe) != cmsSigCurveSetElemType) return FALSE;
1428     }
1429 
1430     // Allocate an empty LUT
1431     Dest =  cmsPipelineAlloc(Src ->ContextID, Src ->InputChannels, Src ->OutputChannels);
1432     if (Dest == NULL) return FALSE;
1433 
1434     // Create target curves
1435     GammaTables = (cmsToneCurve**) _cmsCalloc(Src ->ContextID, Src ->InputChannels, sizeof(cmsToneCurve*));
1436     if (GammaTables == NULL) goto Error;
1437 
1438     for (i=0; i < Src ->InputChannels; i++) {
1439         GammaTables[i] = cmsBuildTabulatedToneCurve16(Src ->ContextID, PRELINEARIZATION_POINTS, NULL);
1440         if (GammaTables[i] == NULL) goto Error;
1441     }
1442 
1443     // Compute 16 bit result by using floating point
1444     for (i=0; i < PRELINEARIZATION_POINTS; i++) {
1445 
1446         for (j=0; j < Src ->InputChannels; j++)
1447             InFloat[j] = (cmsFloat32Number) ((cmsFloat64Number) i / (PRELINEARIZATION_POINTS - 1));
1448 
1449         cmsPipelineEvalFloat(InFloat, OutFloat, Src);
1450 
1451         for (j=0; j < Src ->InputChannels; j++)
1452             GammaTables[j] -> Table16[i] = _cmsQuickSaturateWord(OutFloat[j] * 65535.0);
1453     }
1454 
1455     ObtainedCurves = cmsStageAllocToneCurves(Src ->ContextID, Src ->InputChannels, GammaTables);
1456     if (ObtainedCurves == NULL) goto Error;
1457 
1458     for (i=0; i < Src ->InputChannels; i++) {
1459         cmsFreeToneCurve(GammaTables[i]);
1460         GammaTables[i] = NULL;
1461     }
1462 
1463     if (GammaTables != NULL) {
1464         _cmsFree(Src->ContextID, GammaTables);
1465         GammaTables = NULL;
1466     }
1467 
1468     // Maybe the curves are linear at the end
1469     if (!AllCurvesAreLinear(ObtainedCurves)) {
1470        _cmsStageToneCurvesData* Data;
1471 
1472         if (!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, ObtainedCurves))
1473             goto Error;
1474         Data = (_cmsStageToneCurvesData*) cmsStageData(ObtainedCurves);
1475         ObtainedCurves = NULL;
1476 
1477         // If the curves are to be applied in 8 bits, we can save memory
1478         if (_cmsFormatterIs8bit(*InputFormat)) {
1479              Curves16Data* c16 = CurvesAlloc(Dest ->ContextID, Data ->nCurves, 256, Data ->TheCurves);
1480 
1481              if (c16 == NULL) goto Error;
1482              *dwFlags |= cmsFLAGS_NOCACHE;
1483             _cmsPipelineSetOptimizationParameters(Dest, FastEvaluateCurves8, c16, CurvesFree, CurvesDup);
1484 
1485         }
1486         else {
1487              Curves16Data* c16 = CurvesAlloc(Dest ->ContextID, Data ->nCurves, 65536, Data ->TheCurves);
1488 
1489              if (c16 == NULL) goto Error;
1490              *dwFlags |= cmsFLAGS_NOCACHE;
1491             _cmsPipelineSetOptimizationParameters(Dest, FastEvaluateCurves16, c16, CurvesFree, CurvesDup);
1492         }
1493     }
1494     else {
1495 
1496         // LUT optimizes to nothing. Set the identity LUT
1497         cmsStageFree(ObtainedCurves);
1498         ObtainedCurves = NULL;
1499 
1500         if (!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, cmsStageAllocIdentity(Dest ->ContextID, Src ->InputChannels)))
1501             goto Error;
1502 
1503         *dwFlags |= cmsFLAGS_NOCACHE;
1504         _cmsPipelineSetOptimizationParameters(Dest, FastIdentity16, (void*) Dest, NULL, NULL);
1505     }
1506 
1507     // We are done.
1508     cmsPipelineFree(Src);
1509     *Lut = Dest;
1510     return TRUE;
1511 
1512 Error:
1513 
1514     if (ObtainedCurves != NULL) cmsStageFree(ObtainedCurves);
1515     if (GammaTables != NULL) {
1516         for (i=0; i < Src ->InputChannels; i++) {
1517             if (GammaTables[i] != NULL) cmsFreeToneCurve(GammaTables[i]);
1518         }
1519 
1520         _cmsFree(Src ->ContextID, GammaTables);
1521     }
1522 
1523     if (Dest != NULL) cmsPipelineFree(Dest);
1524     return FALSE;
1525 
1526     cmsUNUSED_PARAMETER(Intent);
1527     cmsUNUSED_PARAMETER(InputFormat);
1528     cmsUNUSED_PARAMETER(OutputFormat);
1529     cmsUNUSED_PARAMETER(dwFlags);
1530 }
1531 
1532 // -------------------------------------------------------------------------------------------------------------------------------------
1533 // LUT is Shaper - Matrix - Matrix - Shaper, which is very frequent when combining two matrix-shaper profiles
1534 
1535 
1536 static
1537 void  FreeMatShaper(cmsContext ContextID, void* Data)
1538 {
1539     if (Data != NULL) _cmsFree(ContextID, Data);
1540 }
1541 
1542 static
1543 void* DupMatShaper(cmsContext ContextID, const void* Data)
1544 {
1545     return _cmsDupMem(ContextID, Data, sizeof(MatShaper8Data));
1546 }
1547 
1548 
1549 // A fast matrix-shaper evaluator for 8 bits. This is a bit tricky since I'm using 1.14 signed fixed point
1550 // to accomplish some performance. Actually it takes 256x3 16 bits tables and 16385 x 3 tables of 8 bits,
1551 // in total about 50K, and the performance boost is huge!
1552 static CMS_NO_SANITIZE
1553 void MatShaperEval16(CMSREGISTER const cmsUInt16Number In[],
1554                      CMSREGISTER cmsUInt16Number Out[],
1555                      CMSREGISTER const void* D)
1556 {
1557     MatShaper8Data* p = (MatShaper8Data*) D;
1558     cmsS1Fixed14Number l1, l2, l3, r, g, b;
1559     cmsUInt32Number ri, gi, bi;
1560 
1561     // In this case (and only in this case!) we can use this simplification since
1562     // In[] is assured to come from a 8 bit number. (a << 8 | a)
1563     ri = In[0] & 0xFFU;
1564     gi = In[1] & 0xFFU;
1565     bi = In[2] & 0xFFU;
1566 
1567     // Across first shaper, which also converts to 1.14 fixed point
1568     r = p->Shaper1R[ri];
1569     g = p->Shaper1G[gi];
1570     b = p->Shaper1B[bi];
1571 
1572     // Evaluate the matrix in 1.14 fixed point
1573     l1 =  (p->Mat[0][0] * r + p->Mat[0][1] * g + p->Mat[0][2] * b + p->Off[0] + 0x2000) >> 14;
1574     l2 =  (p->Mat[1][0] * r + p->Mat[1][1] * g + p->Mat[1][2] * b + p->Off[1] + 0x2000) >> 14;
1575     l3 =  (p->Mat[2][0] * r + p->Mat[2][1] * g + p->Mat[2][2] * b + p->Off[2] + 0x2000) >> 14;
1576 
1577     // Now we have to clip to 0..1.0 range
1578     ri = (l1 < 0) ? 0 : ((l1 > 16384) ? 16384U : (cmsUInt32Number) l1);
1579     gi = (l2 < 0) ? 0 : ((l2 > 16384) ? 16384U : (cmsUInt32Number) l2);
1580     bi = (l3 < 0) ? 0 : ((l3 > 16384) ? 16384U : (cmsUInt32Number) l3);
1581 
1582     // And across second shaper,
1583     Out[0] = p->Shaper2R[ri];
1584     Out[1] = p->Shaper2G[gi];
1585     Out[2] = p->Shaper2B[bi];
1586 
1587 }
1588 
1589 // This table converts from 8 bits to 1.14 after applying the curve
1590 static
1591 void FillFirstShaper(cmsS1Fixed14Number* Table, cmsToneCurve* Curve)
1592 {
1593     int i;
1594     cmsFloat32Number R, y;
1595 
1596     for (i=0; i < 256; i++) {
1597 
1598         R   = (cmsFloat32Number) (i / 255.0);
1599         y   = cmsEvalToneCurveFloat(Curve, R);
1600 
1601         if (y < 131072.0)
1602             Table[i] = DOUBLE_TO_1FIXED14(y);
1603         else
1604             Table[i] = 0x7fffffff;
1605     }
1606 }
1607 
1608 // This table converts form 1.14 (being 0x4000 the last entry) to 8 bits after applying the curve
1609 static
1610 void FillSecondShaper(cmsUInt16Number* Table, cmsToneCurve* Curve, cmsBool Is8BitsOutput)
1611 {
1612     int i;
1613     cmsFloat32Number R, Val;
1614 
1615     for (i=0; i < 16385; i++) {
1616 
1617         R   = (cmsFloat32Number) (i / 16384.0);
1618         Val = cmsEvalToneCurveFloat(Curve, R);    // Val comes 0..1.0
1619 
1620         if (Val < 0)
1621             Val = 0;
1622 
1623         if (Val > 1.0)
1624             Val = 1.0;
1625 
1626         if (Is8BitsOutput) {
1627 
1628             // If 8 bits output, we can optimize further by computing the / 257 part.
1629             // first we compute the resulting byte and then we store the byte times
1630             // 257. This quantization allows to round very quick by doing a >> 8, but
1631             // since the low byte is always equal to msb, we can do a & 0xff and this works!
1632             cmsUInt16Number w = _cmsQuickSaturateWord(Val * 65535.0);
1633             cmsUInt8Number  b = FROM_16_TO_8(w);
1634 
1635             Table[i] = FROM_8_TO_16(b);
1636         }
1637         else Table[i]  = _cmsQuickSaturateWord(Val * 65535.0);
1638     }
1639 }
1640 
1641 // Compute the matrix-shaper structure
1642 static
1643 cmsBool SetMatShaper(cmsPipeline* Dest, cmsToneCurve* Curve1[3], cmsMAT3* Mat, cmsVEC3* Off, cmsToneCurve* Curve2[3], cmsUInt32Number* OutputFormat)
1644 {
1645     MatShaper8Data* p;
1646     int i, j;
1647     cmsBool Is8Bits = _cmsFormatterIs8bit(*OutputFormat);
1648 
1649     // Allocate a big chuck of memory to store precomputed tables
1650     p = (MatShaper8Data*) _cmsMalloc(Dest ->ContextID, sizeof(MatShaper8Data));
1651     if (p == NULL) return FALSE;
1652 
1653     p -> ContextID = Dest -> ContextID;
1654 
1655     // Precompute tables
1656     FillFirstShaper(p ->Shaper1R, Curve1[0]);
1657     FillFirstShaper(p ->Shaper1G, Curve1[1]);
1658     FillFirstShaper(p ->Shaper1B, Curve1[2]);
1659 
1660     FillSecondShaper(p ->Shaper2R, Curve2[0], Is8Bits);
1661     FillSecondShaper(p ->Shaper2G, Curve2[1], Is8Bits);
1662     FillSecondShaper(p ->Shaper2B, Curve2[2], Is8Bits);
1663 
1664     // Convert matrix to nFixed14. Note that those values may take more than 16 bits
1665     for (i=0; i < 3; i++) {
1666         for (j=0; j < 3; j++) {
1667             p ->Mat[i][j] = DOUBLE_TO_1FIXED14(Mat->v[i].n[j]);
1668         }
1669     }
1670 
1671     for (i=0; i < 3; i++) {
1672 
1673         if (Off == NULL) {
1674             p ->Off[i] = 0;
1675         }
1676         else {
1677             p ->Off[i] = DOUBLE_TO_1FIXED14(Off->n[i]);
1678         }
1679     }
1680 
1681     // Mark as optimized for faster formatter
1682     if (Is8Bits)
1683         *OutputFormat |= OPTIMIZED_SH(1);
1684 
1685     // Fill function pointers
1686     _cmsPipelineSetOptimizationParameters(Dest, MatShaperEval16, (void*) p, FreeMatShaper, DupMatShaper);
1687     return TRUE;
1688 }
1689 
1690 //  8 bits on input allows matrix-shaper boot up to 25 Mpixels per second on RGB. That's fast!
1691 static
1692 cmsBool OptimizeMatrixShaper(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags)
1693 {
1694        cmsStage* Curve1, *Curve2;
1695        cmsStage* Matrix1, *Matrix2;
1696        cmsMAT3 res;
1697        cmsBool IdentityMat;
1698        cmsPipeline* Dest, *Src;
1699        cmsFloat64Number* Offset;
1700 
1701        // Only works on RGB to RGB
1702        if (T_CHANNELS(*InputFormat) != 3 || T_CHANNELS(*OutputFormat) != 3) return FALSE;
1703 
1704        // Only works on 8 bit input
1705        if (!_cmsFormatterIs8bit(*InputFormat)) return FALSE;
1706 
1707        // Seems suitable, proceed
1708        Src = *Lut;
1709 
1710        // Check for:
1711        //
1712        //    shaper-matrix-matrix-shaper
1713        //    shaper-matrix-shaper
1714        //
1715        // Both of those constructs are possible (first because abs. colorimetric).
1716        // additionally, In the first case, the input matrix offset should be zero.
1717 
1718        IdentityMat = FALSE;
1719        if (cmsPipelineCheckAndRetreiveStages(Src, 4,
1720               cmsSigCurveSetElemType, cmsSigMatrixElemType, cmsSigMatrixElemType, cmsSigCurveSetElemType,
1721               &Curve1, &Matrix1, &Matrix2, &Curve2)) {
1722 
1723               // Get both matrices
1724               _cmsStageMatrixData* Data1 = (_cmsStageMatrixData*)cmsStageData(Matrix1);
1725               _cmsStageMatrixData* Data2 = (_cmsStageMatrixData*)cmsStageData(Matrix2);
1726 
1727               // Only RGB to RGB
1728               if (Matrix1->InputChannels != 3 || Matrix1->OutputChannels != 3 ||
1729                   Matrix2->InputChannels != 3 || Matrix2->OutputChannels != 3) return FALSE;
1730 
1731               // Input offset should be zero
1732               if (Data1->Offset != NULL) return FALSE;
1733 
1734               // Multiply both matrices to get the result
1735               _cmsMAT3per(&res, (cmsMAT3*)Data2->Double, (cmsMAT3*)Data1->Double);
1736 
1737               // Only 2nd matrix has offset, or it is zero
1738               Offset = Data2->Offset;
1739 
1740               // Now the result is in res + Data2 -> Offset. Maybe is a plain identity?
1741               if (_cmsMAT3isIdentity(&res) && Offset == NULL) {
1742 
1743                      // We can get rid of full matrix
1744                      IdentityMat = TRUE;
1745               }
1746 
1747        }
1748        else {
1749 
1750               if (cmsPipelineCheckAndRetreiveStages(Src, 3,
1751                      cmsSigCurveSetElemType, cmsSigMatrixElemType, cmsSigCurveSetElemType,
1752                      &Curve1, &Matrix1, &Curve2)) {
1753 
1754                      _cmsStageMatrixData* Data = (_cmsStageMatrixData*)cmsStageData(Matrix1);
1755 
1756                      // Copy the matrix to our result
1757                      memcpy(&res, Data->Double, sizeof(res));
1758 
1759                      // Preserve the Odffset (may be NULL as a zero offset)
1760                      Offset = Data->Offset;
1761 
1762                      if (_cmsMAT3isIdentity(&res) && Offset == NULL) {
1763 
1764                             // We can get rid of full matrix
1765                             IdentityMat = TRUE;
1766                      }
1767               }
1768               else
1769                      return FALSE; // Not optimizeable this time
1770 
1771        }
1772 
1773       // Allocate an empty LUT
1774     Dest =  cmsPipelineAlloc(Src ->ContextID, Src ->InputChannels, Src ->OutputChannels);
1775     if (!Dest) return FALSE;
1776 
1777     // Assamble the new LUT
1778     if (!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, cmsStageDup(Curve1)))
1779         goto Error;
1780 
1781     if (!IdentityMat) {
1782 
1783            if (!cmsPipelineInsertStage(Dest, cmsAT_END, cmsStageAllocMatrix(Dest->ContextID, 3, 3, (const cmsFloat64Number*)&res, Offset)))
1784                   goto Error;
1785     }
1786 
1787     if (!cmsPipelineInsertStage(Dest, cmsAT_END, cmsStageDup(Curve2)))
1788         goto Error;
1789 
1790     // If identity on matrix, we can further optimize the curves, so call the join curves routine
1791     if (IdentityMat) {
1792 
1793         OptimizeByJoiningCurves(&Dest, Intent, InputFormat, OutputFormat, dwFlags);
1794     }
1795     else {
1796         _cmsStageToneCurvesData* mpeC1 = (_cmsStageToneCurvesData*) cmsStageData(Curve1);
1797         _cmsStageToneCurvesData* mpeC2 = (_cmsStageToneCurvesData*) cmsStageData(Curve2);
1798 
1799         // In this particular optimization, cache does not help as it takes more time to deal with
1800         // the cache that with the pixel handling
1801         *dwFlags |= cmsFLAGS_NOCACHE;
1802 
1803         // Setup the optimizarion routines
1804         SetMatShaper(Dest, mpeC1 ->TheCurves, &res, (cmsVEC3*) Offset, mpeC2->TheCurves, OutputFormat);
1805     }
1806 
1807     cmsPipelineFree(Src);
1808     *Lut = Dest;
1809     return TRUE;
1810 Error:
1811     // Leave Src unchanged
1812     cmsPipelineFree(Dest);
1813     return FALSE;
1814 }
1815 
1816 
1817 // -------------------------------------------------------------------------------------------------------------------------------------
1818 // Optimization plug-ins
1819 
1820 // List of optimizations
1821 typedef struct _cmsOptimizationCollection_st {
1822 
1823     _cmsOPToptimizeFn  OptimizePtr;
1824 
1825     struct _cmsOptimizationCollection_st *Next;
1826 
1827 } _cmsOptimizationCollection;
1828 
1829 
1830 // The built-in list. We currently implement 4 types of optimizations. Joining of curves, matrix-shaper, linearization and resampling
1831 static _cmsOptimizationCollection DefaultOptimization[] = {
1832 
1833     { OptimizeByJoiningCurves,            &DefaultOptimization[1] },
1834     { OptimizeMatrixShaper,               &DefaultOptimization[2] },
1835     { OptimizeByComputingLinearization,   &DefaultOptimization[3] },
1836     { OptimizeByResampling,               NULL }
1837 };
1838 
1839 // The linked list head
1840 _cmsOptimizationPluginChunkType _cmsOptimizationPluginChunk = { NULL };
1841 
1842 
1843 // Duplicates the zone of memory used by the plug-in in the new context
1844 static
1845 void DupPluginOptimizationList(struct _cmsContext_struct* ctx,
1846                                const struct _cmsContext_struct* src)
1847 {
1848    _cmsOptimizationPluginChunkType newHead = { NULL };
1849    _cmsOptimizationCollection*  entry;
1850    _cmsOptimizationCollection*  Anterior = NULL;
1851    _cmsOptimizationPluginChunkType* head = (_cmsOptimizationPluginChunkType*) src->chunks[OptimizationPlugin];
1852 
1853     _cmsAssert(ctx != NULL);
1854     _cmsAssert(head != NULL);
1855 
1856     // Walk the list copying all nodes
1857    for (entry = head->OptimizationCollection;
1858         entry != NULL;
1859         entry = entry ->Next) {
1860 
1861             _cmsOptimizationCollection *newEntry = ( _cmsOptimizationCollection *) _cmsSubAllocDup(ctx ->MemPool, entry, sizeof(_cmsOptimizationCollection));
1862 
1863             if (newEntry == NULL)
1864                 return;
1865 
1866             // We want to keep the linked list order, so this is a little bit tricky
1867             newEntry -> Next = NULL;
1868             if (Anterior)
1869                 Anterior -> Next = newEntry;
1870 
1871             Anterior = newEntry;
1872 
1873             if (newHead.OptimizationCollection == NULL)
1874                 newHead.OptimizationCollection = newEntry;
1875     }
1876 
1877   ctx ->chunks[OptimizationPlugin] = _cmsSubAllocDup(ctx->MemPool, &newHead, sizeof(_cmsOptimizationPluginChunkType));
1878 }
1879 
1880 void  _cmsAllocOptimizationPluginChunk(struct _cmsContext_struct* ctx,
1881                                          const struct _cmsContext_struct* src)
1882 {
1883   if (src != NULL) {
1884 
1885         // Copy all linked list
1886        DupPluginOptimizationList(ctx, src);
1887     }
1888     else {
1889         static _cmsOptimizationPluginChunkType OptimizationPluginChunkType = { NULL };
1890         ctx ->chunks[OptimizationPlugin] = _cmsSubAllocDup(ctx ->MemPool, &OptimizationPluginChunkType, sizeof(_cmsOptimizationPluginChunkType));
1891     }
1892 }
1893 
1894 
1895 // Register new ways to optimize
1896 cmsBool  _cmsRegisterOptimizationPlugin(cmsContext ContextID, cmsPluginBase* Data)
1897 {
1898     cmsPluginOptimization* Plugin = (cmsPluginOptimization*) Data;
1899     _cmsOptimizationPluginChunkType* ctx = ( _cmsOptimizationPluginChunkType*) _cmsContextGetClientChunk(ContextID, OptimizationPlugin);
1900     _cmsOptimizationCollection* fl;
1901 
1902     if (Data == NULL) {
1903 
1904         ctx->OptimizationCollection = NULL;
1905         return TRUE;
1906     }
1907 
1908     // Optimizer callback is required
1909     if (Plugin ->OptimizePtr == NULL) return FALSE;
1910 
1911     fl = (_cmsOptimizationCollection*) _cmsPluginMalloc(ContextID, sizeof(_cmsOptimizationCollection));
1912     if (fl == NULL) return FALSE;
1913 
1914     // Copy the parameters
1915     fl ->OptimizePtr = Plugin ->OptimizePtr;
1916 
1917     // Keep linked list
1918     fl ->Next = ctx->OptimizationCollection;
1919 
1920     // Set the head
1921     ctx ->OptimizationCollection = fl;
1922 
1923     // All is ok
1924     return TRUE;
1925 }
1926 
1927 // The entry point for LUT optimization
1928 cmsBool CMSEXPORT _cmsOptimizePipeline(cmsContext ContextID,
1929                              cmsPipeline**    PtrLut,
1930                              cmsUInt32Number  Intent,
1931                              cmsUInt32Number* InputFormat,
1932                              cmsUInt32Number* OutputFormat,
1933                              cmsUInt32Number* dwFlags)
1934 {
1935     _cmsOptimizationPluginChunkType* ctx = ( _cmsOptimizationPluginChunkType*) _cmsContextGetClientChunk(ContextID, OptimizationPlugin);
1936     _cmsOptimizationCollection* Opts;
1937     cmsBool AnySuccess = FALSE;
1938     cmsStage* mpe;
1939 
1940     // A CLUT is being asked, so force this specific optimization
1941     if (*dwFlags & cmsFLAGS_FORCE_CLUT) {
1942 
1943         PreOptimize(*PtrLut);
1944         return OptimizeByResampling(PtrLut, Intent, InputFormat, OutputFormat, dwFlags);
1945     }
1946 
1947     // Anything to optimize?
1948     if ((*PtrLut) ->Elements == NULL) {
1949         _cmsPipelineSetOptimizationParameters(*PtrLut, FastIdentity16, (void*) *PtrLut, NULL, NULL);
1950         return TRUE;
1951     }
1952 
1953     // Named color pipelines cannot be optimized
1954     for (mpe = cmsPipelineGetPtrToFirstStage(*PtrLut);
1955         mpe != NULL;
1956         mpe = cmsStageNext(mpe)) {
1957         if (cmsStageType(mpe) == cmsSigNamedColorElemType) return FALSE;
1958     }
1959 
1960     // Try to get rid of identities and trivial conversions.
1961     AnySuccess = PreOptimize(*PtrLut);
1962 
1963     // After removal do we end with an identity?
1964     if ((*PtrLut) ->Elements == NULL) {
1965         _cmsPipelineSetOptimizationParameters(*PtrLut, FastIdentity16, (void*) *PtrLut, NULL, NULL);
1966         return TRUE;
1967     }
1968 
1969     // Do not optimize, keep all precision
1970     if (*dwFlags & cmsFLAGS_NOOPTIMIZE)
1971         return FALSE;
1972 
1973     // Try plug-in optimizations
1974     for (Opts = ctx->OptimizationCollection;
1975          Opts != NULL;
1976          Opts = Opts ->Next) {
1977 
1978             // If one schema succeeded, we are done
1979             if (Opts ->OptimizePtr(PtrLut, Intent, InputFormat, OutputFormat, dwFlags)) {
1980 
1981                 return TRUE;    // Optimized!
1982             }
1983     }
1984 
1985    // Try built-in optimizations
1986     for (Opts = DefaultOptimization;
1987          Opts != NULL;
1988          Opts = Opts ->Next) {
1989 
1990             if (Opts ->OptimizePtr(PtrLut, Intent, InputFormat, OutputFormat, dwFlags)) {
1991 
1992                 return TRUE;
1993             }
1994     }
1995 
1996     // Only simple optimizations succeeded
1997     return AnySuccess;
1998 }
1999 
2000 
2001