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 
 216 // if two adjacent matrices are found, multiply them.
 217 static
 218 cmsBool _MultiplyMatrix(cmsPipeline* Lut)
 219 {
 220        cmsStage** pt1;
 221        cmsStage** pt2;
 222        cmsStage*  chain;
 223        cmsBool AnyOpt = FALSE;
 224 
 225        pt1 = &Lut->Elements;
 226        if (*pt1 == NULL) return AnyOpt;
 227 
 228        while (*pt1 != NULL) {
 229 
 230               pt2 = &((*pt1)->Next);
 231               if (*pt2 == NULL) return AnyOpt;
 232 
 233               if ((*pt1)->Implements == cmsSigMatrixElemType && (*pt2)->Implements == cmsSigMatrixElemType) {
 234 
 235                      // Get both matrices
 236                      _cmsStageMatrixData* m1 = (_cmsStageMatrixData*) cmsStageData(*pt1);
 237                      _cmsStageMatrixData* m2 = (_cmsStageMatrixData*) cmsStageData(*pt2);
 238                      cmsMAT3 res;
 239 
 240                      // Input offset and output offset should be zero to use this optimization
 241                      if (m1->Offset != NULL || m2 ->Offset != NULL ||
 242                             cmsStageInputChannels(*pt1) != 3 || cmsStageOutputChannels(*pt1) != 3 ||
 243                             cmsStageInputChannels(*pt2) != 3 || cmsStageOutputChannels(*pt2) != 3)
 244                             return FALSE;
 245 
 246                      // Multiply both matrices to get the result
 247                      _cmsMAT3per(&res, (cmsMAT3*)m2->Double, (cmsMAT3*)m1->Double);
 248 
 249                      // Get the next in chain after the matrices
 250                      chain = (*pt2)->Next;
 251 
 252                      // Remove both matrices
 253                      _RemoveElement(pt2);
 254                      _RemoveElement(pt1);
 255 
 256                      // Now what if the result is a plain identity?
 257                      if (!isFloatMatrixIdentity(&res)) {
 258 
 259                             // We can not get rid of full matrix
 260                             cmsStage* Multmat = cmsStageAllocMatrix(Lut->ContextID, 3, 3, (const cmsFloat64Number*) &res, NULL);
 261                             if (Multmat == NULL) return FALSE;  // Should never happen
 262 
 263                             // Recover the chain
 264                             Multmat->Next = chain;
 265                             *pt1 = Multmat;
 266                      }
 267 
 268                      AnyOpt = TRUE;
 269               }
 270               else
 271                      pt1 = &((*pt1)->Next);
 272        }
 273 
 274        return AnyOpt;
 275 }
 276 
 277 
 278 // Preoptimize just gets rif of no-ops coming paired. Conversion from v2 to v4 followed
 279 // by a v4 to v2 and vice-versa. The elements are then discarded.
 280 static
 281 cmsBool PreOptimize(cmsPipeline* Lut)
 282 {
 283     cmsBool AnyOpt = FALSE, Opt;
 284 
 285     do {
 286 
 287         Opt = FALSE;
 288 
 289         // Remove all identities
 290         Opt |= _Remove1Op(Lut, cmsSigIdentityElemType);
 291 
 292         // Remove XYZ2Lab followed by Lab2XYZ
 293         Opt |= _Remove2Op(Lut, cmsSigXYZ2LabElemType, cmsSigLab2XYZElemType);
 294 
 295         // Remove Lab2XYZ followed by XYZ2Lab
 296         Opt |= _Remove2Op(Lut, cmsSigLab2XYZElemType, cmsSigXYZ2LabElemType);
 297 
 298         // Remove V4 to V2 followed by V2 to V4
 299         Opt |= _Remove2Op(Lut, cmsSigLabV4toV2, cmsSigLabV2toV4);
 300 
 301         // Remove V2 to V4 followed by V4 to V2
 302         Opt |= _Remove2Op(Lut, cmsSigLabV2toV4, cmsSigLabV4toV2);
 303 
 304         // Remove float pcs Lab conversions
 305         Opt |= _Remove2Op(Lut, cmsSigLab2FloatPCS, cmsSigFloatPCS2Lab);
 306 
 307         // Remove float pcs Lab conversions
 308         Opt |= _Remove2Op(Lut, cmsSigXYZ2FloatPCS, cmsSigFloatPCS2XYZ);
 309 
 310         // Simplify matrix.
 311         Opt |= _MultiplyMatrix(Lut);
 312 
 313         if (Opt) AnyOpt = TRUE;
 314 
 315     } while (Opt);
 316 
 317     return AnyOpt;
 318 }
 319 
 320 static
 321 void Eval16nop1D(CMSREGISTER const cmsUInt16Number Input[],
 322                  CMSREGISTER cmsUInt16Number Output[],
 323                  CMSREGISTER const struct _cms_interp_struc* p)
 324 {
 325     Output[0] = Input[0];
 326 
 327     cmsUNUSED_PARAMETER(p);
 328 }
 329 
 330 static
 331 void PrelinEval16(CMSREGISTER const cmsUInt16Number Input[],
 332                   CMSREGISTER cmsUInt16Number Output[],
 333                   CMSREGISTER const void* D)
 334 {
 335     Prelin16Data* p16 = (Prelin16Data*) D;
 336     cmsUInt16Number  StageABC[MAX_INPUT_DIMENSIONS];
 337     cmsUInt16Number  StageDEF[cmsMAXCHANNELS];
 338     cmsUInt32Number i;
 339 
 340     for (i=0; i < p16 ->nInputs; i++) {
 341 
 342         p16 ->EvalCurveIn16[i](&Input[i], &StageABC[i], p16 ->ParamsCurveIn16[i]);
 343     }
 344 
 345     p16 ->EvalCLUT(StageABC, StageDEF, p16 ->CLUTparams);
 346 
 347     for (i=0; i < p16 ->nOutputs; i++) {
 348 
 349         p16 ->EvalCurveOut16[i](&StageDEF[i], &Output[i], p16 ->ParamsCurveOut16[i]);
 350     }
 351 }
 352 
 353 
 354 static
 355 void PrelinOpt16free(cmsContext ContextID, void* ptr)
 356 {
 357     Prelin16Data* p16 = (Prelin16Data*) ptr;
 358 
 359     _cmsFree(ContextID, p16 ->EvalCurveOut16);
 360     _cmsFree(ContextID, p16 ->ParamsCurveOut16);
 361 
 362     _cmsFree(ContextID, p16);
 363 }
 364 
 365 static
 366 void* Prelin16dup(cmsContext ContextID, const void* ptr)
 367 {
 368     Prelin16Data* p16 = (Prelin16Data*) ptr;
 369     Prelin16Data* Duped = (Prelin16Data*) _cmsDupMem(ContextID, p16, sizeof(Prelin16Data));
 370 
 371     if (Duped == NULL) return NULL;
 372 
 373     Duped->EvalCurveOut16 = (_cmsInterpFn16*) _cmsDupMem(ContextID, p16->EvalCurveOut16, p16->nOutputs * sizeof(_cmsInterpFn16));
 374     Duped->ParamsCurveOut16 = (cmsInterpParams**)_cmsDupMem(ContextID, p16->ParamsCurveOut16, p16->nOutputs * sizeof(cmsInterpParams*));
 375 
 376     return Duped;
 377 }
 378 
 379 
 380 static
 381 Prelin16Data* PrelinOpt16alloc(cmsContext ContextID,
 382                                const cmsInterpParams* ColorMap,
 383                                cmsUInt32Number nInputs, cmsToneCurve** In,
 384                                cmsUInt32Number nOutputs, cmsToneCurve** Out )
 385 {
 386     cmsUInt32Number i;
 387     Prelin16Data* p16 = (Prelin16Data*)_cmsMallocZero(ContextID, sizeof(Prelin16Data));
 388     if (p16 == NULL) return NULL;
 389 
 390     p16 ->nInputs = nInputs;
 391     p16 ->nOutputs = nOutputs;
 392 
 393 
 394     for (i=0; i < nInputs; i++) {
 395 
 396         if (In == NULL) {
 397             p16 -> ParamsCurveIn16[i] = NULL;
 398             p16 -> EvalCurveIn16[i] = Eval16nop1D;
 399 
 400         }
 401         else {
 402             p16 -> ParamsCurveIn16[i] = In[i] ->InterpParams;
 403             p16 -> EvalCurveIn16[i] = p16 ->ParamsCurveIn16[i]->Interpolation.Lerp16;
 404         }
 405     }
 406 
 407     p16 ->CLUTparams = ColorMap;
 408     p16 ->EvalCLUT   = ColorMap ->Interpolation.Lerp16;
 409 
 410 
 411     p16 -> EvalCurveOut16 = (_cmsInterpFn16*) _cmsCalloc(ContextID, nOutputs, sizeof(_cmsInterpFn16));
 412     if (p16->EvalCurveOut16 == NULL)
 413     {
 414         _cmsFree(ContextID, p16);
 415         return NULL;
 416     }
 417 
 418     p16 -> ParamsCurveOut16 = (cmsInterpParams**) _cmsCalloc(ContextID, nOutputs, sizeof(cmsInterpParams* ));
 419     if (p16->ParamsCurveOut16 == NULL)
 420     {
 421 
 422         _cmsFree(ContextID, p16->EvalCurveOut16);
 423         _cmsFree(ContextID, p16);
 424         return NULL;
 425     }
 426 
 427     for (i=0; i < nOutputs; i++) {
 428 
 429         if (Out == NULL) {
 430             p16 ->ParamsCurveOut16[i] = NULL;
 431             p16 -> EvalCurveOut16[i] = Eval16nop1D;
 432         }
 433         else {
 434 
 435             p16 ->ParamsCurveOut16[i] = Out[i] ->InterpParams;
 436             p16 -> EvalCurveOut16[i] = p16 ->ParamsCurveOut16[i]->Interpolation.Lerp16;
 437         }
 438     }
 439 
 440     return p16;
 441 }
 442 
 443 
 444 
 445 // Resampling ---------------------------------------------------------------------------------
 446 
 447 #define PRELINEARIZATION_POINTS 4096
 448 
 449 // Sampler implemented by another LUT. This is a clean way to precalculate the devicelink 3D CLUT for
 450 // almost any transform. We use floating point precision and then convert from floating point to 16 bits.
 451 static
 452 cmsInt32Number XFormSampler16(CMSREGISTER const cmsUInt16Number In[],
 453                               CMSREGISTER cmsUInt16Number Out[],
 454                               CMSREGISTER void* Cargo)
 455 {
 456     cmsPipeline* Lut = (cmsPipeline*) Cargo;
 457     cmsFloat32Number InFloat[cmsMAXCHANNELS], OutFloat[cmsMAXCHANNELS];
 458     cmsUInt32Number i;
 459 
 460     _cmsAssert(Lut -> InputChannels < cmsMAXCHANNELS);
 461     _cmsAssert(Lut -> OutputChannels < cmsMAXCHANNELS);
 462 
 463     // From 16 bit to floating point
 464     for (i=0; i < Lut ->InputChannels; i++)
 465         InFloat[i] = (cmsFloat32Number) (In[i] / 65535.0);
 466 
 467     // Evaluate in floating point
 468     cmsPipelineEvalFloat(InFloat, OutFloat, Lut);
 469 
 470     // Back to 16 bits representation
 471     for (i=0; i < Lut ->OutputChannels; i++)
 472         Out[i] = _cmsQuickSaturateWord(OutFloat[i] * 65535.0);
 473 
 474     // Always succeed
 475     return TRUE;
 476 }
 477 
 478 // Try to see if the curves of a given MPE are linear
 479 static
 480 cmsBool AllCurvesAreLinear(cmsStage* mpe)
 481 {
 482     cmsToneCurve** Curves;
 483     cmsUInt32Number i, n;
 484 
 485     Curves = _cmsStageGetPtrToCurveSet(mpe);
 486     if (Curves == NULL) return FALSE;
 487 
 488     n = cmsStageOutputChannels(mpe);
 489 
 490     for (i=0; i < n; i++) {
 491         if (!cmsIsToneCurveLinear(Curves[i])) return FALSE;
 492     }
 493 
 494     return TRUE;
 495 }
 496 
 497 // This function replaces a specific node placed in "At" by the "Value" numbers. Its purpose
 498 // is to fix scum dot on broken profiles/transforms. Works on 1, 3 and 4 channels
 499 static
 500 cmsBool  PatchLUT(cmsStage* CLUT, cmsUInt16Number At[], cmsUInt16Number Value[],
 501                   cmsUInt32Number nChannelsOut, cmsUInt32Number nChannelsIn)
 502 {
 503     _cmsStageCLutData* Grid = (_cmsStageCLutData*) CLUT ->Data;
 504     cmsInterpParams* p16  = Grid ->Params;
 505     cmsFloat64Number px, py, pz, pw;
 506     int        x0, y0, z0, w0;
 507     int        i, index;
 508 
 509     if (CLUT -> Type != cmsSigCLutElemType) {
 510         cmsSignalError(CLUT->ContextID, cmsERROR_INTERNAL, "(internal) Attempt to PatchLUT on non-lut stage");
 511         return FALSE;
 512     }
 513 
 514     if (nChannelsIn == 4) {
 515 
 516         px = ((cmsFloat64Number) At[0] * (p16->Domain[0])) / 65535.0;
 517         py = ((cmsFloat64Number) At[1] * (p16->Domain[1])) / 65535.0;
 518         pz = ((cmsFloat64Number) At[2] * (p16->Domain[2])) / 65535.0;
 519         pw = ((cmsFloat64Number) At[3] * (p16->Domain[3])) / 65535.0;
 520 
 521         x0 = (int) floor(px);
 522         y0 = (int) floor(py);
 523         z0 = (int) floor(pz);
 524         w0 = (int) floor(pw);
 525 
 526         if (((px - x0) != 0) ||
 527             ((py - y0) != 0) ||
 528             ((pz - z0) != 0) ||
 529             ((pw - w0) != 0)) return FALSE; // Not on exact node
 530 
 531         index = (int) p16 -> opta[3] * x0 +
 532                 (int) p16 -> opta[2] * y0 +
 533                 (int) p16 -> opta[1] * z0 +
 534                 (int) p16 -> opta[0] * w0;
 535     }
 536     else
 537         if (nChannelsIn == 3) {
 538 
 539             px = ((cmsFloat64Number) At[0] * (p16->Domain[0])) / 65535.0;
 540             py = ((cmsFloat64Number) At[1] * (p16->Domain[1])) / 65535.0;
 541             pz = ((cmsFloat64Number) At[2] * (p16->Domain[2])) / 65535.0;
 542 
 543             x0 = (int) floor(px);
 544             y0 = (int) floor(py);
 545             z0 = (int) floor(pz);
 546 
 547             if (((px - x0) != 0) ||
 548                 ((py - y0) != 0) ||
 549                 ((pz - z0) != 0)) return FALSE;  // Not on exact node
 550 
 551             index = (int) p16 -> opta[2] * x0 +
 552                     (int) p16 -> opta[1] * y0 +
 553                     (int) p16 -> opta[0] * z0;
 554         }
 555         else
 556             if (nChannelsIn == 1) {
 557 
 558                 px = ((cmsFloat64Number) At[0] * (p16->Domain[0])) / 65535.0;
 559 
 560                 x0 = (int) floor(px);
 561 
 562                 if (((px - x0) != 0)) return FALSE; // Not on exact node
 563 
 564                 index = (int) p16 -> opta[0] * x0;
 565             }
 566             else {
 567                 cmsSignalError(CLUT->ContextID, cmsERROR_INTERNAL, "(internal) %d Channels are not supported on PatchLUT", nChannelsIn);
 568                 return FALSE;
 569             }
 570 
 571     for (i = 0; i < (int) nChannelsOut; i++)
 572         Grid->Tab.T[index + i] = Value[i];
 573 
 574     return TRUE;
 575 }
 576 
 577 // Auxiliary, to see if two values are equal or very different
 578 static
 579 cmsBool WhitesAreEqual(cmsUInt32Number n, cmsUInt16Number White1[], cmsUInt16Number White2[] )
 580 {
 581     cmsUInt32Number i;
 582 
 583     for (i=0; i < n; i++) {
 584 
 585         if (abs(White1[i] - White2[i]) > 0xf000) return TRUE;  // Values are so extremely different that the fixup should be avoided
 586         if (White1[i] != White2[i]) return FALSE;
 587     }
 588     return TRUE;
 589 }
 590 
 591 
 592 // Locate the node for the white point and fix it to pure white in order to avoid scum dot.
 593 static
 594 cmsBool FixWhiteMisalignment(cmsPipeline* Lut, cmsColorSpaceSignature EntryColorSpace, cmsColorSpaceSignature ExitColorSpace)
 595 {
 596     cmsUInt16Number *WhitePointIn, *WhitePointOut;
 597     cmsUInt16Number  WhiteIn[cmsMAXCHANNELS], WhiteOut[cmsMAXCHANNELS], ObtainedOut[cmsMAXCHANNELS];
 598     cmsUInt32Number i, nOuts, nIns;
 599     cmsStage *PreLin = NULL, *CLUT = NULL, *PostLin = NULL;
 600 
 601     if (!_cmsEndPointsBySpace(EntryColorSpace,
 602         &WhitePointIn, NULL, &nIns)) return FALSE;
 603 
 604     if (!_cmsEndPointsBySpace(ExitColorSpace,
 605         &WhitePointOut, NULL, &nOuts)) return FALSE;
 606 
 607     // It needs to be fixed?
 608     if (Lut ->InputChannels != nIns) return FALSE;
 609     if (Lut ->OutputChannels != nOuts) return FALSE;
 610 
 611     cmsPipelineEval16(WhitePointIn, ObtainedOut, Lut);
 612 
 613     if (WhitesAreEqual(nOuts, WhitePointOut, ObtainedOut)) return TRUE; // whites already match
 614 
 615     // Check if the LUT comes as Prelin, CLUT or Postlin. We allow all combinations
 616     if (!cmsPipelineCheckAndRetreiveStages(Lut, 3, cmsSigCurveSetElemType, cmsSigCLutElemType, cmsSigCurveSetElemType, &PreLin, &CLUT, &PostLin))
 617         if (!cmsPipelineCheckAndRetreiveStages(Lut, 2, cmsSigCurveSetElemType, cmsSigCLutElemType, &PreLin, &CLUT))
 618             if (!cmsPipelineCheckAndRetreiveStages(Lut, 2, cmsSigCLutElemType, cmsSigCurveSetElemType, &CLUT, &PostLin))
 619                 if (!cmsPipelineCheckAndRetreiveStages(Lut, 1, cmsSigCLutElemType, &CLUT))
 620                     return FALSE;
 621 
 622     // We need to interpolate white points of both, pre and post curves
 623     if (PreLin) {
 624 
 625         cmsToneCurve** Curves = _cmsStageGetPtrToCurveSet(PreLin);
 626 
 627         for (i=0; i < nIns; i++) {
 628             WhiteIn[i] = cmsEvalToneCurve16(Curves[i], WhitePointIn[i]);
 629         }
 630     }
 631     else {
 632         for (i=0; i < nIns; i++)
 633             WhiteIn[i] = WhitePointIn[i];
 634     }
 635 
 636     // If any post-linearization, we need to find how is represented white before the curve, do
 637     // a reverse interpolation in this case.
 638     if (PostLin) {
 639 
 640         cmsToneCurve** Curves = _cmsStageGetPtrToCurveSet(PostLin);
 641 
 642         for (i=0; i < nOuts; i++) {
 643 
 644             cmsToneCurve* InversePostLin = cmsReverseToneCurve(Curves[i]);
 645             if (InversePostLin == NULL) {
 646                 WhiteOut[i] = WhitePointOut[i];
 647 
 648             } else {
 649 
 650                 WhiteOut[i] = cmsEvalToneCurve16(InversePostLin, WhitePointOut[i]);
 651                 cmsFreeToneCurve(InversePostLin);
 652             }
 653         }
 654     }
 655     else {
 656         for (i=0; i < nOuts; i++)
 657             WhiteOut[i] = WhitePointOut[i];
 658     }
 659 
 660     // Ok, proceed with patching. May fail and we don't care if it fails
 661     PatchLUT(CLUT, WhiteIn, WhiteOut, nOuts, nIns);
 662 
 663     return TRUE;
 664 }
 665 
 666 // -----------------------------------------------------------------------------------------------------------------------------------------------
 667 // This function creates simple LUT from complex ones. The generated LUT has an optional set of
 668 // prelinearization curves, a CLUT of nGridPoints and optional postlinearization tables.
 669 // These curves have to exist in the original LUT in order to be used in the simplified output.
 670 // Caller may also use the flags to allow this feature.
 671 // LUTS with all curves will be simplified to a single curve. Parametric curves are lost.
 672 // This function should be used on 16-bits LUTS only, as floating point losses precision when simplified
 673 // -----------------------------------------------------------------------------------------------------------------------------------------------
 674 
 675 static
 676 cmsBool OptimizeByResampling(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags)
 677 {
 678     cmsPipeline* Src = NULL;
 679     cmsPipeline* Dest = NULL;
 680     cmsStage* CLUT;
 681     cmsStage *KeepPreLin = NULL, *KeepPostLin = NULL;
 682     cmsUInt32Number nGridPoints;
 683     cmsColorSpaceSignature ColorSpace, OutputColorSpace;
 684     cmsStage *NewPreLin = NULL;
 685     cmsStage *NewPostLin = NULL;
 686     _cmsStageCLutData* DataCLUT;
 687     cmsToneCurve** DataSetIn;
 688     cmsToneCurve** DataSetOut;
 689     Prelin16Data* p16;
 690 
 691     // This is a lossy optimization! does not apply in floating-point cases
 692     if (_cmsFormatterIsFloat(*InputFormat) || _cmsFormatterIsFloat(*OutputFormat)) return FALSE;
 693 
 694     ColorSpace       = _cmsICCcolorSpace((int) T_COLORSPACE(*InputFormat));
 695     OutputColorSpace = _cmsICCcolorSpace((int) T_COLORSPACE(*OutputFormat));
 696 
 697     // Color space must be specified
 698     if (ColorSpace == (cmsColorSpaceSignature)0 ||
 699         OutputColorSpace == (cmsColorSpaceSignature)0) return FALSE;
 700 
 701     nGridPoints = _cmsReasonableGridpointsByColorspace(ColorSpace, *dwFlags);
 702 
 703     // For empty LUTs, 2 points are enough
 704     if (cmsPipelineStageCount(*Lut) == 0)
 705         nGridPoints = 2;
 706 
 707     Src = *Lut;
 708 
 709     // Allocate an empty LUT
 710     Dest =  cmsPipelineAlloc(Src ->ContextID, Src ->InputChannels, Src ->OutputChannels);
 711     if (!Dest) return FALSE;
 712 
 713     // Prelinearization tables are kept unless indicated by flags
 714     if (*dwFlags & cmsFLAGS_CLUT_PRE_LINEARIZATION) {
 715 
 716         // Get a pointer to the prelinearization element
 717         cmsStage* PreLin = cmsPipelineGetPtrToFirstStage(Src);
 718 
 719         // Check if suitable
 720         if (PreLin && PreLin ->Type == cmsSigCurveSetElemType) {
 721 
 722             // Maybe this is a linear tram, so we can avoid the whole stuff
 723             if (!AllCurvesAreLinear(PreLin)) {
 724 
 725                 // All seems ok, proceed.
 726                 NewPreLin = cmsStageDup(PreLin);
 727                 if(!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, NewPreLin))
 728                     goto Error;
 729 
 730                 // Remove prelinearization. Since we have duplicated the curve
 731                 // in destination LUT, the sampling should be applied after this stage.
 732                 cmsPipelineUnlinkStage(Src, cmsAT_BEGIN, &KeepPreLin);
 733             }
 734         }
 735     }
 736 
 737     // Allocate the CLUT
 738     CLUT = cmsStageAllocCLut16bit(Src ->ContextID, nGridPoints, Src ->InputChannels, Src->OutputChannels, NULL);
 739     if (CLUT == NULL) goto Error;
 740 
 741     // Add the CLUT to the destination LUT
 742     if (!cmsPipelineInsertStage(Dest, cmsAT_END, CLUT)) {
 743         goto Error;
 744     }
 745 
 746     // Postlinearization tables are kept unless indicated by flags
 747     if (*dwFlags & cmsFLAGS_CLUT_POST_LINEARIZATION) {
 748 
 749         // Get a pointer to the postlinearization if present
 750         cmsStage* PostLin = cmsPipelineGetPtrToLastStage(Src);
 751 
 752         // Check if suitable
 753         if (PostLin && cmsStageType(PostLin) == cmsSigCurveSetElemType) {
 754 
 755             // Maybe this is a linear tram, so we can avoid the whole stuff
 756             if (!AllCurvesAreLinear(PostLin)) {
 757 
 758                 // All seems ok, proceed.
 759                 NewPostLin = cmsStageDup(PostLin);
 760                 if (!cmsPipelineInsertStage(Dest, cmsAT_END, NewPostLin))
 761                     goto Error;
 762 
 763                 // In destination LUT, the sampling should be applied after this stage.
 764                 cmsPipelineUnlinkStage(Src, cmsAT_END, &KeepPostLin);
 765             }
 766         }
 767     }
 768 
 769     // Now its time to do the sampling. We have to ignore pre/post linearization
 770     // The source LUT without pre/post curves is passed as parameter.
 771     if (!cmsStageSampleCLut16bit(CLUT, XFormSampler16, (void*) Src, 0)) {
 772 Error:
 773         // Ops, something went wrong, Restore stages
 774         if (KeepPreLin != NULL) {
 775             if (!cmsPipelineInsertStage(Src, cmsAT_BEGIN, KeepPreLin)) {
 776                 _cmsAssert(0); // This never happens
 777             }
 778         }
 779         if (KeepPostLin != NULL) {
 780             if (!cmsPipelineInsertStage(Src, cmsAT_END,   KeepPostLin)) {
 781                 _cmsAssert(0); // This never happens
 782             }
 783         }
 784         cmsPipelineFree(Dest);
 785         return FALSE;
 786     }
 787 
 788     // Done.
 789 
 790     if (KeepPreLin != NULL) cmsStageFree(KeepPreLin);
 791     if (KeepPostLin != NULL) cmsStageFree(KeepPostLin);
 792     cmsPipelineFree(Src);
 793 
 794     DataCLUT = (_cmsStageCLutData*) CLUT ->Data;
 795 
 796     if (NewPreLin == NULL) DataSetIn = NULL;
 797     else DataSetIn = ((_cmsStageToneCurvesData*) NewPreLin ->Data) ->TheCurves;
 798 
 799     if (NewPostLin == NULL) DataSetOut = NULL;
 800     else  DataSetOut = ((_cmsStageToneCurvesData*) NewPostLin ->Data) ->TheCurves;
 801 
 802 
 803     if (DataSetIn == NULL && DataSetOut == NULL) {
 804 
 805         _cmsPipelineSetOptimizationParameters(Dest, (_cmsPipelineEval16Fn) DataCLUT->Params->Interpolation.Lerp16, DataCLUT->Params, NULL, NULL);
 806     }
 807     else {
 808 
 809         p16 = PrelinOpt16alloc(Dest ->ContextID,
 810             DataCLUT ->Params,
 811             Dest ->InputChannels,
 812             DataSetIn,
 813             Dest ->OutputChannels,
 814             DataSetOut);
 815 
 816         _cmsPipelineSetOptimizationParameters(Dest, PrelinEval16, (void*) p16, PrelinOpt16free, Prelin16dup);
 817     }
 818 
 819 
 820     // Don't fix white on absolute colorimetric
 821     if (Intent == INTENT_ABSOLUTE_COLORIMETRIC)
 822         *dwFlags |= cmsFLAGS_NOWHITEONWHITEFIXUP;
 823 
 824     if (!(*dwFlags & cmsFLAGS_NOWHITEONWHITEFIXUP)) {
 825 
 826         FixWhiteMisalignment(Dest, ColorSpace, OutputColorSpace);
 827     }
 828 
 829     *Lut = Dest;
 830     return TRUE;
 831 
 832     cmsUNUSED_PARAMETER(Intent);
 833 }
 834 
 835 
 836 // -----------------------------------------------------------------------------------------------------------------------------------------------
 837 // Fixes the gamma balancing of transform. This is described in my paper "Prelinearization Stages on
 838 // Color-Management Application-Specific Integrated Circuits (ASICs)" presented at NIP24. It only works
 839 // for RGB transforms. See the paper for more details
 840 // -----------------------------------------------------------------------------------------------------------------------------------------------
 841 
 842 
 843 // Normalize endpoints by slope limiting max and min. This assures endpoints as well.
 844 // Descending curves are handled as well.
 845 static
 846 void SlopeLimiting(cmsToneCurve* g)
 847 {
 848     int BeginVal, EndVal;
 849     int AtBegin = (int) floor((cmsFloat64Number) g ->nEntries * 0.02 + 0.5);   // Cutoff at 2%
 850     int AtEnd   = (int) g ->nEntries - AtBegin - 1;                                  // And 98%
 851     cmsFloat64Number Val, Slope, beta;
 852     int i;
 853 
 854     if (cmsIsToneCurveDescending(g)) {
 855         BeginVal = 0xffff; EndVal = 0;
 856     }
 857     else {
 858         BeginVal = 0; EndVal = 0xffff;
 859     }
 860 
 861     // Compute slope and offset for begin of curve
 862     Val   = g ->Table16[AtBegin];
 863     Slope = (Val - BeginVal) / AtBegin;
 864     beta  = Val - Slope * AtBegin;
 865 
 866     for (i=0; i < AtBegin; i++)
 867         g ->Table16[i] = _cmsQuickSaturateWord(i * Slope + beta);
 868 
 869     // Compute slope and offset for the end
 870     Val   = g ->Table16[AtEnd];
 871     Slope = (EndVal - Val) / AtBegin;   // AtBegin holds the X interval, which is same in both cases
 872     beta  = Val - Slope * AtEnd;
 873 
 874     for (i = AtEnd; i < (int) g ->nEntries; i++)
 875         g ->Table16[i] = _cmsQuickSaturateWord(i * Slope + beta);
 876 }
 877 
 878 
 879 // Precomputes tables for 8-bit on input devicelink.
 880 static
 881 Prelin8Data* PrelinOpt8alloc(cmsContext ContextID, const cmsInterpParams* p, cmsToneCurve* G[3])
 882 {
 883     int i;
 884     cmsUInt16Number Input[3];
 885     cmsS15Fixed16Number v1, v2, v3;
 886     Prelin8Data* p8;
 887 
 888     p8 = (Prelin8Data*)_cmsMallocZero(ContextID, sizeof(Prelin8Data));
 889     if (p8 == NULL) return NULL;
 890 
 891     // Since this only works for 8 bit input, values comes always as x * 257,
 892     // we can safely take msb byte (x << 8 + x)
 893 
 894     for (i=0; i < 256; i++) {
 895 
 896         if (G != NULL) {
 897 
 898             // Get 16-bit representation
 899             Input[0] = cmsEvalToneCurve16(G[0], FROM_8_TO_16(i));
 900             Input[1] = cmsEvalToneCurve16(G[1], FROM_8_TO_16(i));
 901             Input[2] = cmsEvalToneCurve16(G[2], FROM_8_TO_16(i));
 902         }
 903         else {
 904             Input[0] = FROM_8_TO_16(i);
 905             Input[1] = FROM_8_TO_16(i);
 906             Input[2] = FROM_8_TO_16(i);
 907         }
 908 
 909 
 910         // Move to 0..1.0 in fixed domain
 911         v1 = _cmsToFixedDomain((int) (Input[0] * p -> Domain[0]));
 912         v2 = _cmsToFixedDomain((int) (Input[1] * p -> Domain[1]));
 913         v3 = _cmsToFixedDomain((int) (Input[2] * p -> Domain[2]));
 914 
 915         // Store the precalculated table of nodes
 916         p8 ->X0[i] = (p->opta[2] * FIXED_TO_INT(v1));
 917         p8 ->Y0[i] = (p->opta[1] * FIXED_TO_INT(v2));
 918         p8 ->Z0[i] = (p->opta[0] * FIXED_TO_INT(v3));
 919 
 920         // Store the precalculated table of offsets
 921         p8 ->rx[i] = (cmsUInt16Number) FIXED_REST_TO_INT(v1);
 922         p8 ->ry[i] = (cmsUInt16Number) FIXED_REST_TO_INT(v2);
 923         p8 ->rz[i] = (cmsUInt16Number) FIXED_REST_TO_INT(v3);
 924     }
 925 
 926     p8 ->ContextID = ContextID;
 927     p8 ->p = p;
 928 
 929     return p8;
 930 }
 931 
 932 static
 933 void Prelin8free(cmsContext ContextID, void* ptr)
 934 {
 935     _cmsFree(ContextID, ptr);
 936 }
 937 
 938 static
 939 void* Prelin8dup(cmsContext ContextID, const void* ptr)
 940 {
 941     return _cmsDupMem(ContextID, ptr, sizeof(Prelin8Data));
 942 }
 943 
 944 
 945 
 946 // A optimized interpolation for 8-bit input.
 947 #define DENS(i,j,k) (LutTable[(i)+(j)+(k)+OutChan])
 948 static CMS_NO_SANITIZE
 949 void PrelinEval8(CMSREGISTER const cmsUInt16Number Input[],
 950                  CMSREGISTER cmsUInt16Number Output[],
 951                  CMSREGISTER const void* D)
 952 {
 953 
 954     cmsUInt8Number         r, g, b;
 955     cmsS15Fixed16Number    rx, ry, rz;
 956     cmsS15Fixed16Number    c0, c1, c2, c3, Rest;
 957     int                    OutChan;
 958     CMSREGISTER cmsS15Fixed16Number X0, X1, Y0, Y1, Z0, Z1;
 959     Prelin8Data* p8 = (Prelin8Data*) D;
 960     CMSREGISTER const cmsInterpParams* p = p8 ->p;
 961     int                    TotalOut = (int) p -> nOutputs;
 962     const cmsUInt16Number* LutTable = (const cmsUInt16Number*) p->Table;
 963 
 964     r = (cmsUInt8Number) (Input[0] >> 8);
 965     g = (cmsUInt8Number) (Input[1] >> 8);
 966     b = (cmsUInt8Number) (Input[2] >> 8);
 967 
 968     X0 = (cmsS15Fixed16Number) p8->X0[r];
 969     Y0 = (cmsS15Fixed16Number) p8->Y0[g];
 970     Z0 = (cmsS15Fixed16Number) p8->Z0[b];
 971 
 972     rx = p8 ->rx[r];
 973     ry = p8 ->ry[g];
 974     rz = p8 ->rz[b];
 975 
 976     X1 = X0 + (cmsS15Fixed16Number)((rx == 0) ? 0 :  p ->opta[2]);
 977     Y1 = Y0 + (cmsS15Fixed16Number)((ry == 0) ? 0 :  p ->opta[1]);
 978     Z1 = Z0 + (cmsS15Fixed16Number)((rz == 0) ? 0 :  p ->opta[0]);
 979 
 980 
 981     // These are the 6 Tetrahedral
 982     for (OutChan=0; OutChan < TotalOut; OutChan++) {
 983 
 984         c0 = DENS(X0, Y0, Z0);
 985 
 986         if (rx >= ry && ry >= rz)
 987         {
 988             c1 = DENS(X1, Y0, Z0) - c0;
 989             c2 = DENS(X1, Y1, Z0) - DENS(X1, Y0, Z0);
 990             c3 = DENS(X1, Y1, Z1) - DENS(X1, Y1, Z0);
 991         }
 992         else
 993             if (rx >= rz && rz >= ry)
 994             {
 995                 c1 = DENS(X1, Y0, Z0) - c0;
 996                 c2 = DENS(X1, Y1, Z1) - DENS(X1, Y0, Z1);
 997                 c3 = DENS(X1, Y0, Z1) - DENS(X1, Y0, Z0);
 998             }
 999             else
1000                 if (rz >= rx && rx >= ry)
1001                 {
1002                     c1 = DENS(X1, Y0, Z1) - DENS(X0, Y0, Z1);
1003                     c2 = DENS(X1, Y1, Z1) - DENS(X1, Y0, Z1);
1004                     c3 = DENS(X0, Y0, Z1) - c0;
1005                 }
1006                 else
1007                     if (ry >= rx && rx >= rz)
1008                     {
1009                         c1 = DENS(X1, Y1, Z0) - DENS(X0, Y1, Z0);
1010                         c2 = DENS(X0, Y1, Z0) - c0;
1011                         c3 = DENS(X1, Y1, Z1) - DENS(X1, Y1, Z0);
1012                     }
1013                     else
1014                         if (ry >= rz && rz >= rx)
1015                         {
1016                             c1 = DENS(X1, Y1, Z1) - DENS(X0, Y1, Z1);
1017                             c2 = DENS(X0, Y1, Z0) - c0;
1018                             c3 = DENS(X0, Y1, Z1) - DENS(X0, Y1, Z0);
1019                         }
1020                         else
1021                             if (rz >= ry && ry >= rx)
1022                             {
1023                                 c1 = DENS(X1, Y1, Z1) - DENS(X0, Y1, Z1);
1024                                 c2 = DENS(X0, Y1, Z1) - DENS(X0, Y0, Z1);
1025                                 c3 = DENS(X0, Y0, Z1) - c0;
1026                             }
1027                             else  {
1028                                 c1 = c2 = c3 = 0;
1029                             }
1030 
1031         Rest = c1 * rx + c2 * ry + c3 * rz + 0x8001;
1032         Output[OutChan] = (cmsUInt16Number) (c0 + ((Rest + (Rest >> 16)) >> 16));
1033 
1034     }
1035 }
1036 
1037 #undef DENS
1038 
1039 
1040 // Curves that contain wide empty areas are not optimizeable
1041 static
1042 cmsBool IsDegenerated(const cmsToneCurve* g)
1043 {
1044     cmsUInt32Number i, Zeros = 0, Poles = 0;
1045     cmsUInt32Number nEntries = g ->nEntries;
1046 
1047     for (i=0; i < nEntries; i++) {
1048 
1049         if (g ->Table16[i] == 0x0000) Zeros++;
1050         if (g ->Table16[i] == 0xffff) Poles++;
1051     }
1052 
1053     if (Zeros == 1 && Poles == 1) return FALSE;  // For linear tables
1054     if (Zeros > (nEntries / 20)) return TRUE;  // Degenerated, many zeros
1055     if (Poles > (nEntries / 20)) return TRUE;  // Degenerated, many poles
1056 
1057     return FALSE;
1058 }
1059 
1060 // --------------------------------------------------------------------------------------------------------------
1061 // We need xput over here
1062 
1063 static
1064 cmsBool OptimizeByComputingLinearization(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags)
1065 {
1066     cmsPipeline* OriginalLut;
1067     cmsUInt32Number nGridPoints;
1068     cmsToneCurve *Trans[cmsMAXCHANNELS], *TransReverse[cmsMAXCHANNELS];
1069     cmsUInt32Number t, i;
1070     cmsFloat32Number v, In[cmsMAXCHANNELS], Out[cmsMAXCHANNELS];
1071     cmsBool lIsSuitable, lIsLinear;
1072     cmsPipeline* OptimizedLUT = NULL, *LutPlusCurves = NULL;
1073     cmsStage* OptimizedCLUTmpe;
1074     cmsColorSpaceSignature ColorSpace, OutputColorSpace;
1075     cmsStage* OptimizedPrelinMpe;
1076     cmsToneCurve** OptimizedPrelinCurves;
1077     _cmsStageCLutData* OptimizedPrelinCLUT;
1078 
1079 
1080     // This is a lossy optimization! does not apply in floating-point cases
1081     if (_cmsFormatterIsFloat(*InputFormat) || _cmsFormatterIsFloat(*OutputFormat)) return FALSE;
1082 
1083     // Only on chunky RGB
1084     if (T_COLORSPACE(*InputFormat)  != PT_RGB) return FALSE;
1085     if (T_PLANAR(*InputFormat)) return FALSE;
1086 
1087     if (T_COLORSPACE(*OutputFormat) != PT_RGB) return FALSE;
1088     if (T_PLANAR(*OutputFormat)) return FALSE;
1089 
1090     // On 16 bits, user has to specify the feature
1091     if (!_cmsFormatterIs8bit(*InputFormat)) {
1092         if (!(*dwFlags & cmsFLAGS_CLUT_PRE_LINEARIZATION)) return FALSE;
1093     }
1094 
1095     OriginalLut = *Lut;
1096 
1097     ColorSpace       = _cmsICCcolorSpace((int) T_COLORSPACE(*InputFormat));
1098     OutputColorSpace = _cmsICCcolorSpace((int) T_COLORSPACE(*OutputFormat));
1099 
1100     // Color space must be specified
1101     if (ColorSpace == (cmsColorSpaceSignature)0 ||
1102         OutputColorSpace == (cmsColorSpaceSignature)0) return FALSE;
1103 
1104     nGridPoints      = _cmsReasonableGridpointsByColorspace(ColorSpace, *dwFlags);
1105 
1106     // Empty gamma containers
1107     memset(Trans, 0, sizeof(Trans));
1108     memset(TransReverse, 0, sizeof(TransReverse));
1109 
1110     // If the last stage of the original lut are curves, and those curves are
1111     // degenerated, it is likely the transform is squeezing and clipping
1112     // the output from previous CLUT. We cannot optimize this case
1113     {
1114         cmsStage* last = cmsPipelineGetPtrToLastStage(OriginalLut);
1115 
1116         if (last == NULL) goto Error;
1117         if (cmsStageType(last) == cmsSigCurveSetElemType) {
1118 
1119             _cmsStageToneCurvesData* Data = (_cmsStageToneCurvesData*)cmsStageData(last);
1120             for (i = 0; i < Data->nCurves; i++) {
1121                 if (IsDegenerated(Data->TheCurves[i]))
1122                     goto Error;
1123             }
1124         }
1125     }
1126 
1127     for (t = 0; t < OriginalLut ->InputChannels; t++) {
1128         Trans[t] = cmsBuildTabulatedToneCurve16(OriginalLut ->ContextID, PRELINEARIZATION_POINTS, NULL);
1129         if (Trans[t] == NULL) goto Error;
1130     }
1131 
1132     // Populate the curves
1133     for (i=0; i < PRELINEARIZATION_POINTS; i++) {
1134 
1135         v = (cmsFloat32Number) ((cmsFloat64Number) i / (PRELINEARIZATION_POINTS - 1));
1136 
1137         // Feed input with a gray ramp
1138         for (t=0; t < OriginalLut ->InputChannels; t++)
1139             In[t] = v;
1140 
1141         // Evaluate the gray value
1142         cmsPipelineEvalFloat(In, Out, OriginalLut);
1143 
1144         // Store result in curve
1145         for (t=0; t < OriginalLut ->InputChannels; t++)
1146         {
1147             if (Trans[t]->Table16 != NULL)
1148                 Trans[t] ->Table16[i] = _cmsQuickSaturateWord(Out[t] * 65535.0);
1149         }
1150     }
1151 
1152     // Slope-limit the obtained curves
1153     for (t = 0; t < OriginalLut ->InputChannels; t++)
1154         SlopeLimiting(Trans[t]);
1155 
1156     // Check for validity. lIsLinear is here for debug purposes
1157     lIsSuitable = TRUE;
1158     lIsLinear   = TRUE;
1159     for (t=0; (lIsSuitable && (t < OriginalLut ->InputChannels)); t++) {
1160 
1161         // Exclude if already linear
1162         if (!cmsIsToneCurveLinear(Trans[t]))
1163             lIsLinear = FALSE;
1164 
1165         // Exclude if non-monotonic
1166         if (!cmsIsToneCurveMonotonic(Trans[t]))
1167             lIsSuitable = FALSE;
1168 
1169         if (IsDegenerated(Trans[t]))
1170             lIsSuitable = FALSE;
1171     }
1172 
1173     // If it is not suitable, just quit
1174     if (!lIsSuitable) goto Error;
1175 
1176     // Invert curves if possible
1177     for (t = 0; t < OriginalLut ->InputChannels; t++) {
1178         TransReverse[t] = cmsReverseToneCurveEx(PRELINEARIZATION_POINTS, Trans[t]);
1179         if (TransReverse[t] == NULL) goto Error;
1180     }
1181 
1182     // Now inset the reversed curves at the begin of transform
1183     LutPlusCurves = cmsPipelineDup(OriginalLut);
1184     if (LutPlusCurves == NULL) goto Error;
1185 
1186     if (!cmsPipelineInsertStage(LutPlusCurves, cmsAT_BEGIN, cmsStageAllocToneCurves(OriginalLut ->ContextID, OriginalLut ->InputChannels, TransReverse)))
1187         goto Error;
1188 
1189     // Create the result LUT
1190     OptimizedLUT = cmsPipelineAlloc(OriginalLut ->ContextID, OriginalLut ->InputChannels, OriginalLut ->OutputChannels);
1191     if (OptimizedLUT == NULL) goto Error;
1192 
1193     OptimizedPrelinMpe = cmsStageAllocToneCurves(OriginalLut ->ContextID, OriginalLut ->InputChannels, Trans);
1194 
1195     // Create and insert the curves at the beginning
1196     if (!cmsPipelineInsertStage(OptimizedLUT, cmsAT_BEGIN, OptimizedPrelinMpe))
1197         goto Error;
1198 
1199     // Allocate the CLUT for result
1200     OptimizedCLUTmpe = cmsStageAllocCLut16bit(OriginalLut ->ContextID, nGridPoints, OriginalLut ->InputChannels, OriginalLut ->OutputChannels, NULL);
1201 
1202     // Add the CLUT to the destination LUT
1203     if (!cmsPipelineInsertStage(OptimizedLUT, cmsAT_END, OptimizedCLUTmpe))
1204         goto Error;
1205 
1206     // Resample the LUT
1207     if (!cmsStageSampleCLut16bit(OptimizedCLUTmpe, XFormSampler16, (void*) LutPlusCurves, 0)) goto Error;
1208 
1209     // Free resources
1210     for (t = 0; t < OriginalLut ->InputChannels; t++) {
1211 
1212         if (Trans[t]) cmsFreeToneCurve(Trans[t]);
1213         if (TransReverse[t]) cmsFreeToneCurve(TransReverse[t]);
1214     }
1215 
1216     cmsPipelineFree(LutPlusCurves);
1217 
1218 
1219     OptimizedPrelinCurves = _cmsStageGetPtrToCurveSet(OptimizedPrelinMpe);
1220     OptimizedPrelinCLUT   = (_cmsStageCLutData*) OptimizedCLUTmpe ->Data;
1221 
1222     // Set the evaluator if 8-bit
1223     if (_cmsFormatterIs8bit(*InputFormat)) {
1224 
1225         Prelin8Data* p8 = PrelinOpt8alloc(OptimizedLUT ->ContextID,
1226                                                 OptimizedPrelinCLUT ->Params,
1227                                                 OptimizedPrelinCurves);
1228         if (p8 == NULL) return FALSE;
1229 
1230         _cmsPipelineSetOptimizationParameters(OptimizedLUT, PrelinEval8, (void*) p8, Prelin8free, Prelin8dup);
1231 
1232     }
1233     else
1234     {
1235         Prelin16Data* p16 = PrelinOpt16alloc(OptimizedLUT ->ContextID,
1236             OptimizedPrelinCLUT ->Params,
1237             3, OptimizedPrelinCurves, 3, NULL);
1238         if (p16 == NULL) return FALSE;
1239 
1240         _cmsPipelineSetOptimizationParameters(OptimizedLUT, PrelinEval16, (void*) p16, PrelinOpt16free, Prelin16dup);
1241 
1242     }
1243 
1244     // Don't fix white on absolute colorimetric
1245     if (Intent == INTENT_ABSOLUTE_COLORIMETRIC)
1246         *dwFlags |= cmsFLAGS_NOWHITEONWHITEFIXUP;
1247 
1248     if (!(*dwFlags & cmsFLAGS_NOWHITEONWHITEFIXUP)) {
1249 
1250         if (!FixWhiteMisalignment(OptimizedLUT, ColorSpace, OutputColorSpace)) {
1251 
1252             return FALSE;
1253         }
1254     }
1255 
1256     // And return the obtained LUT
1257 
1258     cmsPipelineFree(OriginalLut);
1259     *Lut = OptimizedLUT;
1260     return TRUE;
1261 
1262 Error:
1263 
1264     for (t = 0; t < OriginalLut ->InputChannels; t++) {
1265 
1266         if (Trans[t]) cmsFreeToneCurve(Trans[t]);
1267         if (TransReverse[t]) cmsFreeToneCurve(TransReverse[t]);
1268     }
1269 
1270     if (LutPlusCurves != NULL) cmsPipelineFree(LutPlusCurves);
1271     if (OptimizedLUT != NULL) cmsPipelineFree(OptimizedLUT);
1272 
1273     return FALSE;
1274 
1275     cmsUNUSED_PARAMETER(Intent);
1276     cmsUNUSED_PARAMETER(lIsLinear);
1277 }
1278 
1279 
1280 // Curves optimizer ------------------------------------------------------------------------------------------------------------------
1281 
1282 static
1283 void CurvesFree(cmsContext ContextID, void* ptr)
1284 {
1285      Curves16Data* Data = (Curves16Data*) ptr;
1286      cmsUInt32Number i;
1287 
1288      for (i=0; i < Data -> nCurves; i++) {
1289 
1290          _cmsFree(ContextID, Data ->Curves[i]);
1291      }
1292 
1293      _cmsFree(ContextID, Data ->Curves);
1294      _cmsFree(ContextID, ptr);
1295 }
1296 
1297 static
1298 void* CurvesDup(cmsContext ContextID, const void* ptr)
1299 {
1300     Curves16Data* Data = (Curves16Data*)_cmsDupMem(ContextID, ptr, sizeof(Curves16Data));
1301     cmsUInt32Number i;
1302 
1303     if (Data == NULL) return NULL;
1304 
1305     Data->Curves = (cmsUInt16Number**) _cmsDupMem(ContextID, Data->Curves, Data->nCurves * sizeof(cmsUInt16Number*));
1306 
1307     for (i=0; i < Data -> nCurves; i++) {
1308         Data->Curves[i] = (cmsUInt16Number*) _cmsDupMem(ContextID, Data->Curves[i], Data->nElements * sizeof(cmsUInt16Number));
1309     }
1310 
1311     return (void*) Data;
1312 }
1313 
1314 // Precomputes tables for 8-bit on input devicelink.
1315 static
1316 Curves16Data* CurvesAlloc(cmsContext ContextID, cmsUInt32Number nCurves, cmsUInt32Number nElements, cmsToneCurve** G)
1317 {
1318     cmsUInt32Number i, j;
1319     Curves16Data* c16;
1320 
1321     c16 = (Curves16Data*)_cmsMallocZero(ContextID, sizeof(Curves16Data));
1322     if (c16 == NULL) return NULL;
1323 
1324     c16 ->nCurves = nCurves;
1325     c16 ->nElements = nElements;
1326 
1327     c16->Curves = (cmsUInt16Number**) _cmsCalloc(ContextID, nCurves, sizeof(cmsUInt16Number*));
1328     if (c16->Curves == NULL) {
1329         _cmsFree(ContextID, c16);
1330         return NULL;
1331     }
1332 
1333     for (i=0; i < nCurves; i++) {
1334 
1335         c16->Curves[i] = (cmsUInt16Number*) _cmsCalloc(ContextID, nElements, sizeof(cmsUInt16Number));
1336 
1337         if (c16->Curves[i] == NULL) {
1338 
1339             for (j=0; j < i; j++) {
1340                 _cmsFree(ContextID, c16->Curves[j]);
1341             }
1342             _cmsFree(ContextID, c16->Curves);
1343             _cmsFree(ContextID, c16);
1344             return NULL;
1345         }
1346 
1347         if (nElements == 256U) {
1348 
1349             for (j=0; j < nElements; j++) {
1350 
1351                 c16 ->Curves[i][j] = cmsEvalToneCurve16(G[i], FROM_8_TO_16(j));
1352             }
1353         }
1354         else {
1355 
1356             for (j=0; j < nElements; j++) {
1357                 c16 ->Curves[i][j] = cmsEvalToneCurve16(G[i], (cmsUInt16Number) j);
1358             }
1359         }
1360     }
1361 
1362     return c16;
1363 }
1364 
1365 static
1366 void FastEvaluateCurves8(CMSREGISTER const cmsUInt16Number In[],
1367                          CMSREGISTER cmsUInt16Number Out[],
1368                          CMSREGISTER const void* D)
1369 {
1370     Curves16Data* Data = (Curves16Data*) D;
1371     int x;
1372     cmsUInt32Number i;
1373 
1374     for (i=0; i < Data ->nCurves; i++) {
1375 
1376          x = (In[i] >> 8);
1377          Out[i] = Data -> Curves[i][x];
1378     }
1379 }
1380 
1381 
1382 static
1383 void FastEvaluateCurves16(CMSREGISTER const cmsUInt16Number In[],
1384                           CMSREGISTER cmsUInt16Number Out[],
1385                           CMSREGISTER const void* D)
1386 {
1387     Curves16Data* Data = (Curves16Data*) D;
1388     cmsUInt32Number i;
1389 
1390     for (i=0; i < Data ->nCurves; i++) {
1391          Out[i] = Data -> Curves[i][In[i]];
1392     }
1393 }
1394 
1395 
1396 static
1397 void FastIdentity16(CMSREGISTER const cmsUInt16Number In[],
1398                     CMSREGISTER cmsUInt16Number Out[],
1399                     CMSREGISTER const void* D)
1400 {
1401     cmsPipeline* Lut = (cmsPipeline*) D;
1402     cmsUInt32Number i;
1403 
1404     for (i=0; i < Lut ->InputChannels; i++) {
1405          Out[i] = In[i];
1406     }
1407 }
1408 
1409 
1410 // If the target LUT holds only curves, the optimization procedure is to join all those
1411 // curves together. That only works on curves and does not work on matrices.
1412 static
1413 cmsBool OptimizeByJoiningCurves(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags)
1414 {
1415     cmsToneCurve** GammaTables = NULL;
1416     cmsFloat32Number InFloat[cmsMAXCHANNELS], OutFloat[cmsMAXCHANNELS];
1417     cmsUInt32Number i, j;
1418     cmsPipeline* Src = *Lut;
1419     cmsPipeline* Dest = NULL;
1420     cmsStage* mpe;
1421     cmsStage* ObtainedCurves = NULL;
1422 
1423 
1424     // This is a lossy optimization! does not apply in floating-point cases
1425     if (_cmsFormatterIsFloat(*InputFormat) || _cmsFormatterIsFloat(*OutputFormat)) return FALSE;
1426 
1427     //  Only curves in this LUT?
1428     for (mpe = cmsPipelineGetPtrToFirstStage(Src);
1429          mpe != NULL;
1430          mpe = cmsStageNext(mpe)) {
1431             if (cmsStageType(mpe) != cmsSigCurveSetElemType) return FALSE;
1432     }
1433 
1434     // Allocate an empty LUT
1435     Dest =  cmsPipelineAlloc(Src ->ContextID, Src ->InputChannels, Src ->OutputChannels);
1436     if (Dest == NULL) return FALSE;
1437 
1438     // Create target curves
1439     GammaTables = (cmsToneCurve**) _cmsCalloc(Src ->ContextID, Src ->InputChannels, sizeof(cmsToneCurve*));
1440     if (GammaTables == NULL) goto Error;
1441 
1442     for (i=0; i < Src ->InputChannels; i++) {
1443         GammaTables[i] = cmsBuildTabulatedToneCurve16(Src ->ContextID, PRELINEARIZATION_POINTS, NULL);
1444         if (GammaTables[i] == NULL) goto Error;
1445     }
1446 
1447     // Compute 16 bit result by using floating point
1448     for (i=0; i < PRELINEARIZATION_POINTS; i++) {
1449 
1450         for (j=0; j < Src ->InputChannels; j++)
1451             InFloat[j] = (cmsFloat32Number) ((cmsFloat64Number) i / (PRELINEARIZATION_POINTS - 1));
1452 
1453         cmsPipelineEvalFloat(InFloat, OutFloat, Src);
1454 
1455         for (j=0; j < Src ->InputChannels; j++)
1456             GammaTables[j] -> Table16[i] = _cmsQuickSaturateWord(OutFloat[j] * 65535.0);
1457     }
1458 
1459     ObtainedCurves = cmsStageAllocToneCurves(Src ->ContextID, Src ->InputChannels, GammaTables);
1460     if (ObtainedCurves == NULL) goto Error;
1461 
1462     for (i=0; i < Src ->InputChannels; i++) {
1463         cmsFreeToneCurve(GammaTables[i]);
1464         GammaTables[i] = NULL;
1465     }
1466 
1467     if (GammaTables != NULL) {
1468         _cmsFree(Src->ContextID, GammaTables);
1469         GammaTables = NULL;
1470     }
1471 
1472     // Maybe the curves are linear at the end
1473     if (!AllCurvesAreLinear(ObtainedCurves)) {
1474        _cmsStageToneCurvesData* Data;
1475 
1476         if (!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, ObtainedCurves))
1477             goto Error;
1478         Data = (_cmsStageToneCurvesData*) cmsStageData(ObtainedCurves);
1479         ObtainedCurves = NULL;
1480 
1481         // If the curves are to be applied in 8 bits, we can save memory
1482         if (_cmsFormatterIs8bit(*InputFormat)) {
1483              Curves16Data* c16 = CurvesAlloc(Dest ->ContextID, Data ->nCurves, 256, Data ->TheCurves);
1484 
1485              if (c16 == NULL) goto Error;
1486              *dwFlags |= cmsFLAGS_NOCACHE;
1487             _cmsPipelineSetOptimizationParameters(Dest, FastEvaluateCurves8, c16, CurvesFree, CurvesDup);
1488 
1489         }
1490         else {
1491              Curves16Data* c16 = CurvesAlloc(Dest ->ContextID, Data ->nCurves, 65536, Data ->TheCurves);
1492 
1493              if (c16 == NULL) goto Error;
1494              *dwFlags |= cmsFLAGS_NOCACHE;
1495             _cmsPipelineSetOptimizationParameters(Dest, FastEvaluateCurves16, c16, CurvesFree, CurvesDup);
1496         }
1497     }
1498     else {
1499 
1500         // LUT optimizes to nothing. Set the identity LUT
1501         cmsStageFree(ObtainedCurves);
1502         ObtainedCurves = NULL;
1503 
1504         if (!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, cmsStageAllocIdentity(Dest ->ContextID, Src ->InputChannels)))
1505             goto Error;
1506 
1507         *dwFlags |= cmsFLAGS_NOCACHE;
1508         _cmsPipelineSetOptimizationParameters(Dest, FastIdentity16, (void*) Dest, NULL, NULL);
1509     }
1510 
1511     // We are done.
1512     cmsPipelineFree(Src);
1513     *Lut = Dest;
1514     return TRUE;
1515 
1516 Error:
1517 
1518     if (ObtainedCurves != NULL) cmsStageFree(ObtainedCurves);
1519     if (GammaTables != NULL) {
1520         for (i=0; i < Src ->InputChannels; i++) {
1521             if (GammaTables[i] != NULL) cmsFreeToneCurve(GammaTables[i]);
1522         }
1523 
1524         _cmsFree(Src ->ContextID, GammaTables);
1525     }
1526 
1527     if (Dest != NULL) cmsPipelineFree(Dest);
1528     return FALSE;
1529 
1530     cmsUNUSED_PARAMETER(Intent);
1531     cmsUNUSED_PARAMETER(InputFormat);
1532     cmsUNUSED_PARAMETER(OutputFormat);
1533     cmsUNUSED_PARAMETER(dwFlags);
1534 }
1535 
1536 // -------------------------------------------------------------------------------------------------------------------------------------
1537 // LUT is Shaper - Matrix - Matrix - Shaper, which is very frequent when combining two matrix-shaper profiles
1538 
1539 
1540 static
1541 void  FreeMatShaper(cmsContext ContextID, void* Data)
1542 {
1543     if (Data != NULL) _cmsFree(ContextID, Data);
1544 }
1545 
1546 static
1547 void* DupMatShaper(cmsContext ContextID, const void* Data)
1548 {
1549     return _cmsDupMem(ContextID, Data, sizeof(MatShaper8Data));
1550 }
1551 
1552 
1553 // A fast matrix-shaper evaluator for 8 bits. This is a bit tricky since I'm using 1.14 signed fixed point
1554 // to accomplish some performance. Actually it takes 256x3 16 bits tables and 16385 x 3 tables of 8 bits,
1555 // in total about 50K, and the performance boost is huge!
1556 static CMS_NO_SANITIZE
1557 void MatShaperEval16(CMSREGISTER const cmsUInt16Number In[],
1558                      CMSREGISTER cmsUInt16Number Out[],
1559                      CMSREGISTER const void* D)
1560 {
1561     MatShaper8Data* p = (MatShaper8Data*) D;
1562     cmsS1Fixed14Number l1, l2, l3, r, g, b;
1563     cmsUInt32Number ri, gi, bi;
1564 
1565     // In this case (and only in this case!) we can use this simplification since
1566     // In[] is assured to come from a 8 bit number. (a << 8 | a)
1567     ri = In[0] & 0xFFU;
1568     gi = In[1] & 0xFFU;
1569     bi = In[2] & 0xFFU;
1570 
1571     // Across first shaper, which also converts to 1.14 fixed point
1572     r = p->Shaper1R[ri];
1573     g = p->Shaper1G[gi];
1574     b = p->Shaper1B[bi];
1575 
1576     // Evaluate the matrix in 1.14 fixed point
1577     l1 =  (p->Mat[0][0] * r + p->Mat[0][1] * g + p->Mat[0][2] * b + p->Off[0] + 0x2000) >> 14;
1578     l2 =  (p->Mat[1][0] * r + p->Mat[1][1] * g + p->Mat[1][2] * b + p->Off[1] + 0x2000) >> 14;
1579     l3 =  (p->Mat[2][0] * r + p->Mat[2][1] * g + p->Mat[2][2] * b + p->Off[2] + 0x2000) >> 14;
1580 
1581     // Now we have to clip to 0..1.0 range
1582     ri = (l1 < 0) ? 0 : ((l1 > 16384) ? 16384U : (cmsUInt32Number) l1);
1583     gi = (l2 < 0) ? 0 : ((l2 > 16384) ? 16384U : (cmsUInt32Number) l2);
1584     bi = (l3 < 0) ? 0 : ((l3 > 16384) ? 16384U : (cmsUInt32Number) l3);
1585 
1586     // And across second shaper,
1587     Out[0] = p->Shaper2R[ri];
1588     Out[1] = p->Shaper2G[gi];
1589     Out[2] = p->Shaper2B[bi];
1590 
1591 }
1592 
1593 // This table converts from 8 bits to 1.14 after applying the curve
1594 static
1595 void FillFirstShaper(cmsS1Fixed14Number* Table, cmsToneCurve* Curve)
1596 {
1597     int i;
1598     cmsFloat32Number R, y;
1599 
1600     for (i=0; i < 256; i++) {
1601 
1602         R   = (cmsFloat32Number) (i / 255.0);
1603         y   = cmsEvalToneCurveFloat(Curve, R);
1604 
1605         if (y < 131072.0)
1606             Table[i] = DOUBLE_TO_1FIXED14(y);
1607         else
1608             Table[i] = 0x7fffffff;
1609     }
1610 }
1611 
1612 // This table converts form 1.14 (being 0x4000 the last entry) to 8 bits after applying the curve
1613 static
1614 void FillSecondShaper(cmsUInt16Number* Table, cmsToneCurve* Curve, cmsBool Is8BitsOutput)
1615 {
1616     int i;
1617     cmsFloat32Number R, Val;
1618 
1619     for (i=0; i < 16385; i++) {
1620 
1621         R   = (cmsFloat32Number) (i / 16384.0);
1622         Val = cmsEvalToneCurveFloat(Curve, R);    // Val comes 0..1.0
1623 
1624         if (Val < 0)
1625             Val = 0;
1626 
1627         if (Val > 1.0)
1628             Val = 1.0;
1629 
1630         if (Is8BitsOutput) {
1631 
1632             // If 8 bits output, we can optimize further by computing the / 257 part.
1633             // first we compute the resulting byte and then we store the byte times
1634             // 257. This quantization allows to round very quick by doing a >> 8, but
1635             // since the low byte is always equal to msb, we can do a & 0xff and this works!
1636             cmsUInt16Number w = _cmsQuickSaturateWord(Val * 65535.0);
1637             cmsUInt8Number  b = FROM_16_TO_8(w);
1638 
1639             Table[i] = FROM_8_TO_16(b);
1640         }
1641         else Table[i]  = _cmsQuickSaturateWord(Val * 65535.0);
1642     }
1643 }
1644 
1645 // Compute the matrix-shaper structure
1646 static
1647 cmsBool SetMatShaper(cmsPipeline* Dest, cmsToneCurve* Curve1[3], cmsMAT3* Mat, cmsVEC3* Off, cmsToneCurve* Curve2[3], cmsUInt32Number* OutputFormat)
1648 {
1649     MatShaper8Data* p;
1650     int i, j;
1651     cmsBool Is8Bits = _cmsFormatterIs8bit(*OutputFormat);
1652 
1653     // Allocate a big chuck of memory to store precomputed tables
1654     p = (MatShaper8Data*) _cmsMalloc(Dest ->ContextID, sizeof(MatShaper8Data));
1655     if (p == NULL) return FALSE;
1656 
1657     p -> ContextID = Dest -> ContextID;
1658 
1659     // Precompute tables
1660     FillFirstShaper(p ->Shaper1R, Curve1[0]);
1661     FillFirstShaper(p ->Shaper1G, Curve1[1]);
1662     FillFirstShaper(p ->Shaper1B, Curve1[2]);
1663 
1664     FillSecondShaper(p ->Shaper2R, Curve2[0], Is8Bits);
1665     FillSecondShaper(p ->Shaper2G, Curve2[1], Is8Bits);
1666     FillSecondShaper(p ->Shaper2B, Curve2[2], Is8Bits);
1667 
1668     // Convert matrix to nFixed14. Note that those values may take more than 16 bits
1669     for (i=0; i < 3; i++) {
1670         for (j=0; j < 3; j++) {
1671             p ->Mat[i][j] = DOUBLE_TO_1FIXED14(Mat->v[i].n[j]);
1672         }
1673     }
1674 
1675     for (i=0; i < 3; i++) {
1676 
1677         if (Off == NULL) {
1678             p ->Off[i] = 0;
1679         }
1680         else {
1681             p ->Off[i] = DOUBLE_TO_1FIXED14(Off->n[i]);
1682         }
1683     }
1684 
1685     // Mark as optimized for faster formatter
1686     if (Is8Bits)
1687         *OutputFormat |= OPTIMIZED_SH(1);
1688 
1689     // Fill function pointers
1690     _cmsPipelineSetOptimizationParameters(Dest, MatShaperEval16, (void*) p, FreeMatShaper, DupMatShaper);
1691     return TRUE;
1692 }
1693 
1694 //  8 bits on input allows matrix-shaper boot up to 25 Mpixels per second on RGB. That's fast!
1695 static
1696 cmsBool OptimizeMatrixShaper(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags)
1697 {
1698        cmsStage* Curve1, *Curve2;
1699        cmsStage* Matrix1, *Matrix2;
1700        cmsMAT3 res;
1701        cmsBool IdentityMat;
1702        cmsPipeline* Dest, *Src;
1703        cmsFloat64Number* Offset;
1704 
1705        // Only works on RGB to RGB
1706        if (T_CHANNELS(*InputFormat) != 3 || T_CHANNELS(*OutputFormat) != 3) return FALSE;
1707 
1708        // Only works on 8 bit input
1709        if (!_cmsFormatterIs8bit(*InputFormat)) return FALSE;
1710 
1711        // Seems suitable, proceed
1712        Src = *Lut;
1713 
1714        // Check for:
1715        //
1716        //    shaper-matrix-matrix-shaper
1717        //    shaper-matrix-shaper
1718        //
1719        // Both of those constructs are possible (first because abs. colorimetric).
1720        // additionally, In the first case, the input matrix offset should be zero.
1721 
1722        IdentityMat = FALSE;
1723        if (cmsPipelineCheckAndRetreiveStages(Src, 4,
1724               cmsSigCurveSetElemType, cmsSigMatrixElemType, cmsSigMatrixElemType, cmsSigCurveSetElemType,
1725               &Curve1, &Matrix1, &Matrix2, &Curve2)) {
1726 
1727               // Get both matrices
1728               _cmsStageMatrixData* Data1 = (_cmsStageMatrixData*)cmsStageData(Matrix1);
1729               _cmsStageMatrixData* Data2 = (_cmsStageMatrixData*)cmsStageData(Matrix2);
1730 
1731               // Only RGB to RGB
1732               if (Matrix1->InputChannels != 3 || Matrix1->OutputChannels != 3 ||
1733                   Matrix2->InputChannels != 3 || Matrix2->OutputChannels != 3) return FALSE;
1734 
1735               // Input offset should be zero
1736               if (Data1->Offset != NULL) return FALSE;
1737 
1738               // Multiply both matrices to get the result
1739               _cmsMAT3per(&res, (cmsMAT3*)Data2->Double, (cmsMAT3*)Data1->Double);
1740 
1741               // Only 2nd matrix has offset, or it is zero
1742               Offset = Data2->Offset;
1743 
1744               // Now the result is in res + Data2 -> Offset. Maybe is a plain identity?
1745               if (_cmsMAT3isIdentity(&res) && Offset == NULL) {
1746 
1747                      // We can get rid of full matrix
1748                      IdentityMat = TRUE;
1749               }
1750 
1751        }
1752        else {
1753 
1754               if (cmsPipelineCheckAndRetreiveStages(Src, 3,
1755                      cmsSigCurveSetElemType, cmsSigMatrixElemType, cmsSigCurveSetElemType,
1756                      &Curve1, &Matrix1, &Curve2)) {
1757 
1758                      _cmsStageMatrixData* Data = (_cmsStageMatrixData*)cmsStageData(Matrix1);
1759 
1760                      if (Matrix1->InputChannels != 3 || Matrix1->OutputChannels != 3) return FALSE;
1761 
1762                      // Copy the matrix to our result
1763                      memcpy(&res, Data->Double, sizeof(res));
1764 
1765                      // Preserve the Odffset (may be NULL as a zero offset)
1766                      Offset = Data->Offset;
1767 
1768                      if (_cmsMAT3isIdentity(&res) && Offset == NULL) {
1769 
1770                             // We can get rid of full matrix
1771                             IdentityMat = TRUE;
1772                      }
1773               }
1774               else
1775                      return FALSE; // Not optimizeable this time
1776 
1777        }
1778 
1779       // Allocate an empty LUT
1780     Dest =  cmsPipelineAlloc(Src ->ContextID, Src ->InputChannels, Src ->OutputChannels);
1781     if (!Dest) return FALSE;
1782 
1783     // Assamble the new LUT
1784     if (!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, cmsStageDup(Curve1)))
1785         goto Error;
1786 
1787     if (!IdentityMat) {
1788 
1789            if (!cmsPipelineInsertStage(Dest, cmsAT_END, cmsStageAllocMatrix(Dest->ContextID, 3, 3, (const cmsFloat64Number*)&res, Offset)))
1790                   goto Error;
1791     }
1792 
1793     if (!cmsPipelineInsertStage(Dest, cmsAT_END, cmsStageDup(Curve2)))
1794         goto Error;
1795 
1796     // If identity on matrix, we can further optimize the curves, so call the join curves routine
1797     if (IdentityMat) {
1798 
1799         OptimizeByJoiningCurves(&Dest, Intent, InputFormat, OutputFormat, dwFlags);
1800     }
1801     else {
1802         _cmsStageToneCurvesData* mpeC1 = (_cmsStageToneCurvesData*) cmsStageData(Curve1);
1803         _cmsStageToneCurvesData* mpeC2 = (_cmsStageToneCurvesData*) cmsStageData(Curve2);
1804 
1805         // In this particular optimization, cache does not help as it takes more time to deal with
1806         // the cache than with the pixel handling
1807         *dwFlags |= cmsFLAGS_NOCACHE;
1808 
1809         // Setup the optimizarion routines
1810         SetMatShaper(Dest, mpeC1 ->TheCurves, &res, (cmsVEC3*) Offset, mpeC2->TheCurves, OutputFormat);
1811     }
1812 
1813     cmsPipelineFree(Src);
1814     *Lut = Dest;
1815     return TRUE;
1816 Error:
1817     // Leave Src unchanged
1818     cmsPipelineFree(Dest);
1819     return FALSE;
1820 }
1821 
1822 
1823 // -------------------------------------------------------------------------------------------------------------------------------------
1824 // Optimization plug-ins
1825 
1826 // List of optimizations
1827 typedef struct _cmsOptimizationCollection_st {
1828 
1829     _cmsOPToptimizeFn  OptimizePtr;
1830 
1831     struct _cmsOptimizationCollection_st *Next;
1832 
1833 } _cmsOptimizationCollection;
1834 
1835 
1836 // The built-in list. We currently implement 4 types of optimizations. Joining of curves, matrix-shaper, linearization and resampling
1837 static _cmsOptimizationCollection DefaultOptimization[] = {
1838 
1839     { OptimizeByJoiningCurves,            &DefaultOptimization[1] },
1840     { OptimizeMatrixShaper,               &DefaultOptimization[2] },
1841     { OptimizeByComputingLinearization,   &DefaultOptimization[3] },
1842     { OptimizeByResampling,               NULL }
1843 };
1844 
1845 // The linked list head
1846 _cmsOptimizationPluginChunkType _cmsOptimizationPluginChunk = { NULL };
1847 
1848 
1849 // Duplicates the zone of memory used by the plug-in in the new context
1850 static
1851 void DupPluginOptimizationList(struct _cmsContext_struct* ctx,
1852                                const struct _cmsContext_struct* src)
1853 {
1854    _cmsOptimizationPluginChunkType newHead = { NULL };
1855    _cmsOptimizationCollection*  entry;
1856    _cmsOptimizationCollection*  Anterior = NULL;
1857    _cmsOptimizationPluginChunkType* head = (_cmsOptimizationPluginChunkType*) src->chunks[OptimizationPlugin];
1858 
1859     _cmsAssert(ctx != NULL);
1860     _cmsAssert(head != NULL);
1861 
1862     // Walk the list copying all nodes
1863    for (entry = head->OptimizationCollection;
1864         entry != NULL;
1865         entry = entry ->Next) {
1866 
1867             _cmsOptimizationCollection *newEntry = ( _cmsOptimizationCollection *) _cmsSubAllocDup(ctx ->MemPool, entry, sizeof(_cmsOptimizationCollection));
1868 
1869             if (newEntry == NULL)
1870                 return;
1871 
1872             // We want to keep the linked list order, so this is a little bit tricky
1873             newEntry -> Next = NULL;
1874             if (Anterior)
1875                 Anterior -> Next = newEntry;
1876 
1877             Anterior = newEntry;
1878 
1879             if (newHead.OptimizationCollection == NULL)
1880                 newHead.OptimizationCollection = newEntry;
1881     }
1882 
1883   ctx ->chunks[OptimizationPlugin] = _cmsSubAllocDup(ctx->MemPool, &newHead, sizeof(_cmsOptimizationPluginChunkType));
1884 }
1885 
1886 void  _cmsAllocOptimizationPluginChunk(struct _cmsContext_struct* ctx,
1887                                          const struct _cmsContext_struct* src)
1888 {
1889   if (src != NULL) {
1890 
1891         // Copy all linked list
1892        DupPluginOptimizationList(ctx, src);
1893     }
1894     else {
1895         static _cmsOptimizationPluginChunkType OptimizationPluginChunkType = { NULL };
1896         ctx ->chunks[OptimizationPlugin] = _cmsSubAllocDup(ctx ->MemPool, &OptimizationPluginChunkType, sizeof(_cmsOptimizationPluginChunkType));
1897     }
1898 }
1899 
1900 
1901 // Register new ways to optimize
1902 cmsBool  _cmsRegisterOptimizationPlugin(cmsContext ContextID, cmsPluginBase* Data)
1903 {
1904     cmsPluginOptimization* Plugin = (cmsPluginOptimization*) Data;
1905     _cmsOptimizationPluginChunkType* ctx = ( _cmsOptimizationPluginChunkType*) _cmsContextGetClientChunk(ContextID, OptimizationPlugin);
1906     _cmsOptimizationCollection* fl;
1907 
1908     if (Data == NULL) {
1909 
1910         ctx->OptimizationCollection = NULL;
1911         return TRUE;
1912     }
1913 
1914     // Optimizer callback is required
1915     if (Plugin ->OptimizePtr == NULL) return FALSE;
1916 
1917     fl = (_cmsOptimizationCollection*) _cmsPluginMalloc(ContextID, sizeof(_cmsOptimizationCollection));
1918     if (fl == NULL) return FALSE;
1919 
1920     // Copy the parameters
1921     fl ->OptimizePtr = Plugin ->OptimizePtr;
1922 
1923     // Keep linked list
1924     fl ->Next = ctx->OptimizationCollection;
1925 
1926     // Set the head
1927     ctx ->OptimizationCollection = fl;
1928 
1929     // All is ok
1930     return TRUE;
1931 }
1932 
1933 // The entry point for LUT optimization
1934 cmsBool CMSEXPORT _cmsOptimizePipeline(cmsContext ContextID,
1935                              cmsPipeline**    PtrLut,
1936                              cmsUInt32Number  Intent,
1937                              cmsUInt32Number* InputFormat,
1938                              cmsUInt32Number* OutputFormat,
1939                              cmsUInt32Number* dwFlags)
1940 {
1941     _cmsOptimizationPluginChunkType* ctx = ( _cmsOptimizationPluginChunkType*) _cmsContextGetClientChunk(ContextID, OptimizationPlugin);
1942     _cmsOptimizationCollection* Opts;
1943     cmsBool AnySuccess = FALSE;
1944     cmsStage* mpe;
1945 
1946     // A CLUT is being asked, so force this specific optimization
1947     if (*dwFlags & cmsFLAGS_FORCE_CLUT) {
1948 
1949         PreOptimize(*PtrLut);
1950         return OptimizeByResampling(PtrLut, Intent, InputFormat, OutputFormat, dwFlags);
1951     }
1952 
1953     // Anything to optimize?
1954     if ((*PtrLut) ->Elements == NULL) {
1955         _cmsPipelineSetOptimizationParameters(*PtrLut, FastIdentity16, (void*) *PtrLut, NULL, NULL);
1956         return TRUE;
1957     }
1958 
1959     // Named color pipelines cannot be optimized
1960     for (mpe = cmsPipelineGetPtrToFirstStage(*PtrLut);
1961         mpe != NULL;
1962         mpe = cmsStageNext(mpe)) {
1963             if (cmsStageType(mpe) == cmsSigNamedColorElemType) return FALSE;
1964     }
1965 
1966     // Try to get rid of identities and trivial conversions.
1967     AnySuccess = PreOptimize(*PtrLut);
1968 
1969     // After removal do we end with an identity?
1970     if ((*PtrLut) ->Elements == NULL) {
1971         _cmsPipelineSetOptimizationParameters(*PtrLut, FastIdentity16, (void*) *PtrLut, NULL, NULL);
1972         return TRUE;
1973     }
1974 
1975     // Do not optimize, keep all precision
1976     if (*dwFlags & cmsFLAGS_NOOPTIMIZE)
1977         return FALSE;
1978 
1979     // Try plug-in optimizations
1980     for (Opts = ctx->OptimizationCollection;
1981          Opts != NULL;
1982          Opts = Opts ->Next) {
1983 
1984             // If one schema succeeded, we are done
1985             if (Opts ->OptimizePtr(PtrLut, Intent, InputFormat, OutputFormat, dwFlags)) {
1986 
1987                 return TRUE;    // Optimized!
1988             }
1989     }
1990 
1991    // Try built-in optimizations
1992     for (Opts = DefaultOptimization;
1993          Opts != NULL;
1994          Opts = Opts ->Next) {
1995 
1996             if (Opts ->OptimizePtr(PtrLut, Intent, InputFormat, OutputFormat, dwFlags)) {
1997 
1998                 return TRUE;
1999             }
2000     }
2001 
2002     // Only simple optimizations succeeded
2003     return AnySuccess;
2004 }
2005 
2006 
2007