2 //---------------------------------------------------------------------------------
4 // Little Color Management System
5 // Copyright (c) 1998-2011 Marti Maria Saguer
7 // Permission is hereby granted, free of charge, to any person obtaining
8 // a copy of this software and associated documentation files (the "Software"),
9 // to deal in the Software without restriction, including without limitation
10 // the rights to use, copy, modify, merge, publish, distribute, sublicense,
11 // and/or sell copies of the Software, and to permit persons to whom the Software
12 // is furnished to do so, subject to the following conditions:
14 // The above copyright notice and this permission notice shall be included in
15 // all copies or substantial portions of the Software.
17 // THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
18 // EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO
19 // THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
20 // NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE
21 // LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
22 // OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
23 // WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
25 //---------------------------------------------------------------------------------
28 #include "lcms2_internal.h"
31 //----------------------------------------------------------------------------------
33 // Optimization for 8 bits, Shaper-CLUT (3 inputs only)
38 const cmsInterpParams* p; // Tetrahedrical interpolation parameters. This is a not-owned pointer.
40 cmsUInt16Number rx[256], ry[256], rz[256];
41 cmsUInt32Number X0[256], Y0[256], Z0[256]; // Precomputed nodes and offsets for 8-bit input data
47 // Generic optimization for 16 bits Shaper-CLUT-Shaper (any inputs)
56 _cmsInterpFn16 EvalCurveIn16[MAX_INPUT_DIMENSIONS]; // The maximum number of input channels is known in advance
57 cmsInterpParams* ParamsCurveIn16[MAX_INPUT_DIMENSIONS];
59 _cmsInterpFn16 EvalCLUT; // The evaluator for 3D grid
60 const cmsInterpParams* CLUTparams; // (not-owned pointer)
63 _cmsInterpFn16* EvalCurveOut16; // Points to an array of curve evaluators in 16 bits (not-owned pointer)
64 cmsInterpParams** ParamsCurveOut16; // Points to an array of references to interpolation params (not-owned pointer)
70 // Optimization for matrix-shaper in 8 bits. Numbers are operated in n.14 signed, tables are stored in 1.14 fixed
72 typedef cmsInt32Number cmsS1Fixed14Number; // Note that this may hold more than 16 bits!
74 #define DOUBLE_TO_1FIXED14(x) ((cmsS1Fixed14Number) floor((x) * 16384.0 + 0.5))
80 cmsS1Fixed14Number Shaper1R[256]; // from 0..255 to 1.14 (0.0...1.0)
81 cmsS1Fixed14Number Shaper1G[256];
82 cmsS1Fixed14Number Shaper1B[256];
84 cmsS1Fixed14Number Mat[3][3]; // n.14 to n.14 (needs a saturation after that)
85 cmsS1Fixed14Number Off[3];
87 cmsUInt16Number Shaper2R[16385]; // 1.14 to 0..255
88 cmsUInt16Number Shaper2G[16385];
89 cmsUInt16Number Shaper2B[16385];
93 // Curves, optimization is shared between 8 and 16 bits
98 int nCurves; // Number of curves
99 int nElements; // Elements in curves
100 cmsUInt16Number** Curves; // Points to a dynamically allocated array
105 // Simple optimizations ----------------------------------------------------------------------------------------------------------
108 // Remove an element in linked chain
110 void _RemoveElement(cmsStage** head)
112 cmsStage* mpe = *head;
113 cmsStage* next = mpe ->Next;
118 // Remove all identities in chain. Note that pt actually is a double pointer to the element that holds the pointer.
120 cmsBool _Remove1Op(cmsPipeline* Lut, cmsStageSignature UnaryOp)
122 cmsStage** pt = &Lut ->Elements;
123 cmsBool AnyOpt = FALSE;
125 while (*pt != NULL) {
127 if ((*pt) ->Implements == UnaryOp) {
132 pt = &((*pt) -> Next);
138 // Same, but only if two adjacent elements are found
140 cmsBool _Remove2Op(cmsPipeline* Lut, cmsStageSignature Op1, cmsStageSignature Op2)
144 cmsBool AnyOpt = FALSE;
146 pt1 = &Lut ->Elements;
147 if (*pt1 == NULL) return AnyOpt;
149 while (*pt1 != NULL) {
151 pt2 = &((*pt1) -> Next);
152 if (*pt2 == NULL) return AnyOpt;
154 if ((*pt1) ->Implements == Op1 && (*pt2) ->Implements == Op2) {
160 pt1 = &((*pt1) -> Next);
166 // Preoptimize just gets rif of no-ops coming paired. Conversion from v2 to v4 followed
167 // by a v4 to v2 and vice-versa. The elements are then discarded.
169 cmsBool PreOptimize(cmsPipeline* Lut)
171 cmsBool AnyOpt = FALSE, Opt;
177 // Remove all identities
178 Opt |= _Remove1Op(Lut, cmsSigIdentityElemType);
180 // Remove XYZ2Lab followed by Lab2XYZ
181 Opt |= _Remove2Op(Lut, cmsSigXYZ2LabElemType, cmsSigLab2XYZElemType);
183 // Remove Lab2XYZ followed by XYZ2Lab
184 Opt |= _Remove2Op(Lut, cmsSigLab2XYZElemType, cmsSigXYZ2LabElemType);
186 // Remove V4 to V2 followed by V2 to V4
187 Opt |= _Remove2Op(Lut, cmsSigLabV4toV2, cmsSigLabV2toV4);
189 // Remove V2 to V4 followed by V4 to V2
190 Opt |= _Remove2Op(Lut, cmsSigLabV2toV4, cmsSigLabV4toV2);
192 // Remove float pcs Lab conversions
193 Opt |= _Remove2Op(Lut, cmsSigLab2FloatPCS, cmsSigFloatPCS2Lab);
195 // Remove float pcs Lab conversions
196 Opt |= _Remove2Op(Lut, cmsSigXYZ2FloatPCS, cmsSigFloatPCS2XYZ);
198 if (Opt) AnyOpt = TRUE;
206 void Eval16nop1D(register const cmsUInt16Number Input[],
207 register cmsUInt16Number Output[],
208 register const struct _cms_interp_struc* p)
210 Output[0] = Input[0];
212 cmsUNUSED_PARAMETER(p);
216 void PrelinEval16(register const cmsUInt16Number Input[],
217 register cmsUInt16Number Output[],
218 register const void* D)
220 Prelin16Data* p16 = (Prelin16Data*) D;
221 cmsUInt16Number StageABC[MAX_INPUT_DIMENSIONS];
222 cmsUInt16Number StageDEF[cmsMAXCHANNELS];
225 for (i=0; i < p16 ->nInputs; i++) {
227 p16 ->EvalCurveIn16[i](&Input[i], &StageABC[i], p16 ->ParamsCurveIn16[i]);
230 p16 ->EvalCLUT(StageABC, StageDEF, p16 ->CLUTparams);
232 for (i=0; i < p16 ->nOutputs; i++) {
234 p16 ->EvalCurveOut16[i](&StageDEF[i], &Output[i], p16 ->ParamsCurveOut16[i]);
240 void PrelinOpt16free(cmsContext ContextID, void* ptr)
242 Prelin16Data* p16 = (Prelin16Data*) ptr;
244 _cmsFree(ContextID, p16 ->EvalCurveOut16);
245 _cmsFree(ContextID, p16 ->ParamsCurveOut16);
247 _cmsFree(ContextID, p16);
251 void* Prelin16dup(cmsContext ContextID, const void* ptr)
253 Prelin16Data* p16 = (Prelin16Data*) ptr;
254 Prelin16Data* Duped = _cmsDupMem(ContextID, p16, sizeof(Prelin16Data));
256 if (Duped == NULL) return NULL;
258 Duped ->EvalCurveOut16 = _cmsDupMem(ContextID, p16 ->EvalCurveOut16, p16 ->nOutputs * sizeof(_cmsInterpFn16));
259 Duped ->ParamsCurveOut16 = _cmsDupMem(ContextID, p16 ->ParamsCurveOut16, p16 ->nOutputs * sizeof(cmsInterpParams* ));
266 Prelin16Data* PrelinOpt16alloc(cmsContext ContextID,
267 const cmsInterpParams* ColorMap,
268 int nInputs, cmsToneCurve** In,
269 int nOutputs, cmsToneCurve** Out )
272 Prelin16Data* p16 = _cmsMallocZero(ContextID, sizeof(Prelin16Data));
273 if (p16 == NULL) return NULL;
275 p16 ->nInputs = nInputs;
276 p16 -> nOutputs = nOutputs;
279 for (i=0; i < nInputs; i++) {
282 p16 -> ParamsCurveIn16[i] = NULL;
283 p16 -> EvalCurveIn16[i] = Eval16nop1D;
287 p16 -> ParamsCurveIn16[i] = In[i] ->InterpParams;
288 p16 -> EvalCurveIn16[i] = p16 ->ParamsCurveIn16[i]->Interpolation.Lerp16;
292 p16 ->CLUTparams = ColorMap;
293 p16 ->EvalCLUT = ColorMap ->Interpolation.Lerp16;
296 p16 -> EvalCurveOut16 = (_cmsInterpFn16*) _cmsCalloc(ContextID, nOutputs, sizeof(_cmsInterpFn16));
297 p16 -> ParamsCurveOut16 = (cmsInterpParams**) _cmsCalloc(ContextID, nOutputs, sizeof(cmsInterpParams* ));
299 for (i=0; i < nOutputs; i++) {
302 p16 ->ParamsCurveOut16[i] = NULL;
303 p16 -> EvalCurveOut16[i] = Eval16nop1D;
307 p16 ->ParamsCurveOut16[i] = Out[i] ->InterpParams;
308 p16 -> EvalCurveOut16[i] = p16 ->ParamsCurveOut16[i]->Interpolation.Lerp16;
317 // Resampling ---------------------------------------------------------------------------------
319 #define PRELINEARIZATION_POINTS 4096
321 // Sampler implemented by another LUT. This is a clean way to precalculate the devicelink 3D CLUT for
322 // almost any transform. We use floating point precision and then convert from floating point to 16 bits.
324 int XFormSampler16(register const cmsUInt16Number In[], register cmsUInt16Number Out[], register void* Cargo)
326 cmsPipeline* Lut = (cmsPipeline*) Cargo;
327 cmsFloat32Number InFloat[cmsMAXCHANNELS], OutFloat[cmsMAXCHANNELS];
330 _cmsAssert(Lut -> InputChannels < cmsMAXCHANNELS);
331 _cmsAssert(Lut -> OutputChannels < cmsMAXCHANNELS);
333 // From 16 bit to floating point
334 for (i=0; i < Lut ->InputChannels; i++)
335 InFloat[i] = (cmsFloat32Number) (In[i] / 65535.0);
337 // Evaluate in floating point
338 cmsPipelineEvalFloat(InFloat, OutFloat, Lut);
340 // Back to 16 bits representation
341 for (i=0; i < Lut ->OutputChannels; i++)
342 Out[i] = _cmsQuickSaturateWord(OutFloat[i] * 65535.0);
348 // Try to see if the curves of a given MPE are linear
350 cmsBool AllCurvesAreLinear(cmsStage* mpe)
352 cmsToneCurve** Curves;
353 cmsUInt32Number i, n;
355 Curves = _cmsStageGetPtrToCurveSet(mpe);
356 if (Curves == NULL) return FALSE;
358 n = cmsStageOutputChannels(mpe);
360 for (i=0; i < n; i++) {
361 if (!cmsIsToneCurveLinear(Curves[i])) return FALSE;
367 // This function replaces a specific node placed in "At" by the "Value" numbers. Its purpose
368 // is to fix scum dot on broken profiles/transforms. Works on 1, 3 and 4 channels
370 cmsBool PatchLUT(cmsStage* CLUT, cmsUInt16Number At[], cmsUInt16Number Value[],
371 int nChannelsOut, int nChannelsIn)
373 _cmsStageCLutData* Grid = (_cmsStageCLutData*) CLUT ->Data;
374 cmsInterpParams* p16 = Grid ->Params;
375 cmsFloat64Number px, py, pz, pw;
379 if (CLUT -> Type != cmsSigCLutElemType) {
380 cmsSignalError(CLUT->ContextID, cmsERROR_INTERNAL, "(internal) Attempt to PatchLUT on non-lut stage");
384 if (nChannelsIn == 4) {
386 px = ((cmsFloat64Number) At[0] * (p16->Domain[0])) / 65535.0;
387 py = ((cmsFloat64Number) At[1] * (p16->Domain[1])) / 65535.0;
388 pz = ((cmsFloat64Number) At[2] * (p16->Domain[2])) / 65535.0;
389 pw = ((cmsFloat64Number) At[3] * (p16->Domain[3])) / 65535.0;
391 x0 = (int) floor(px);
392 y0 = (int) floor(py);
393 z0 = (int) floor(pz);
394 w0 = (int) floor(pw);
396 if (((px - x0) != 0) ||
399 ((pw - w0) != 0)) return FALSE; // Not on exact node
401 index = p16 -> opta[3] * x0 +
402 p16 -> opta[2] * y0 +
403 p16 -> opta[1] * z0 +
407 if (nChannelsIn == 3) {
409 px = ((cmsFloat64Number) At[0] * (p16->Domain[0])) / 65535.0;
410 py = ((cmsFloat64Number) At[1] * (p16->Domain[1])) / 65535.0;
411 pz = ((cmsFloat64Number) At[2] * (p16->Domain[2])) / 65535.0;
413 x0 = (int) floor(px);
414 y0 = (int) floor(py);
415 z0 = (int) floor(pz);
417 if (((px - x0) != 0) ||
419 ((pz - z0) != 0)) return FALSE; // Not on exact node
421 index = p16 -> opta[2] * x0 +
422 p16 -> opta[1] * y0 +
426 if (nChannelsIn == 1) {
428 px = ((cmsFloat64Number) At[0] * (p16->Domain[0])) / 65535.0;
430 x0 = (int) floor(px);
432 if (((px - x0) != 0)) return FALSE; // Not on exact node
434 index = p16 -> opta[0] * x0;
437 cmsSignalError(CLUT->ContextID, cmsERROR_INTERNAL, "(internal) %d Channels are not supported on PatchLUT", nChannelsIn);
441 for (i=0; i < nChannelsOut; i++)
442 Grid -> Tab.T[index + i] = Value[i];
447 // Auxiliar, to see if two values are equal or very different
449 cmsBool WhitesAreEqual(int n, cmsUInt16Number White1[], cmsUInt16Number White2[] )
453 for (i=0; i < n; i++) {
455 if (abs(White1[i] - White2[i]) > 0xf000) return TRUE; // Values are so extremly different that the fixup should be avoided
456 if (White1[i] != White2[i]) return FALSE;
462 // Locate the node for the white point and fix it to pure white in order to avoid scum dot.
464 cmsBool FixWhiteMisalignment(cmsPipeline* Lut, cmsColorSpaceSignature EntryColorSpace, cmsColorSpaceSignature ExitColorSpace)
466 cmsUInt16Number *WhitePointIn, *WhitePointOut;
467 cmsUInt16Number WhiteIn[cmsMAXCHANNELS], WhiteOut[cmsMAXCHANNELS], ObtainedOut[cmsMAXCHANNELS];
468 cmsUInt32Number i, nOuts, nIns;
469 cmsStage *PreLin = NULL, *CLUT = NULL, *PostLin = NULL;
471 if (!_cmsEndPointsBySpace(EntryColorSpace,
472 &WhitePointIn, NULL, &nIns)) return FALSE;
474 if (!_cmsEndPointsBySpace(ExitColorSpace,
475 &WhitePointOut, NULL, &nOuts)) return FALSE;
477 // It needs to be fixed?
478 if (Lut ->InputChannels != nIns) return FALSE;
479 if (Lut ->OutputChannels != nOuts) return FALSE;
481 cmsPipelineEval16(WhitePointIn, ObtainedOut, Lut);
483 if (WhitesAreEqual(nOuts, WhitePointOut, ObtainedOut)) return TRUE; // whites already match
485 // Check if the LUT comes as Prelin, CLUT or Postlin. We allow all combinations
486 if (!cmsPipelineCheckAndRetreiveStages(Lut, 3, cmsSigCurveSetElemType, cmsSigCLutElemType, cmsSigCurveSetElemType, &PreLin, &CLUT, &PostLin))
487 if (!cmsPipelineCheckAndRetreiveStages(Lut, 2, cmsSigCurveSetElemType, cmsSigCLutElemType, &PreLin, &CLUT))
488 if (!cmsPipelineCheckAndRetreiveStages(Lut, 2, cmsSigCLutElemType, cmsSigCurveSetElemType, &CLUT, &PostLin))
489 if (!cmsPipelineCheckAndRetreiveStages(Lut, 1, cmsSigCLutElemType, &CLUT))
492 // We need to interpolate white points of both, pre and post curves
495 cmsToneCurve** Curves = _cmsStageGetPtrToCurveSet(PreLin);
497 for (i=0; i < nIns; i++) {
498 WhiteIn[i] = cmsEvalToneCurve16(Curves[i], WhitePointIn[i]);
502 for (i=0; i < nIns; i++)
503 WhiteIn[i] = WhitePointIn[i];
506 // If any post-linearization, we need to find how is represented white before the curve, do
507 // a reverse interpolation in this case.
510 cmsToneCurve** Curves = _cmsStageGetPtrToCurveSet(PostLin);
512 for (i=0; i < nOuts; i++) {
514 cmsToneCurve* InversePostLin = cmsReverseToneCurve(Curves[i]);
515 if (InversePostLin == NULL) {
516 WhiteOut[i] = WhitePointOut[i];
520 WhiteOut[i] = cmsEvalToneCurve16(InversePostLin, WhitePointOut[i]);
521 cmsFreeToneCurve(InversePostLin);
526 for (i=0; i < nOuts; i++)
527 WhiteOut[i] = WhitePointOut[i];
530 // Ok, proceed with patching. May fail and we don't care if it fails
531 PatchLUT(CLUT, WhiteIn, WhiteOut, nOuts, nIns);
536 // -----------------------------------------------------------------------------------------------------------------------------------------------
537 // This function creates simple LUT from complex ones. The generated LUT has an optional set of
538 // prelinearization curves, a CLUT of nGridPoints and optional postlinearization tables.
539 // These curves have to exist in the original LUT in order to be used in the simplified output.
540 // Caller may also use the flags to allow this feature.
541 // LUTS with all curves will be simplified to a single curve. Parametric curves are lost.
542 // This function should be used on 16-bits LUTS only, as floating point losses precision when simplified
543 // -----------------------------------------------------------------------------------------------------------------------------------------------
546 cmsBool OptimizeByResampling(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags)
548 cmsPipeline* Src = NULL;
549 cmsPipeline* Dest = NULL;
552 cmsStage *KeepPreLin = NULL, *KeepPostLin = NULL;
554 cmsColorSpaceSignature ColorSpace, OutputColorSpace;
555 cmsStage *NewPreLin = NULL;
556 cmsStage *NewPostLin = NULL;
557 _cmsStageCLutData* DataCLUT;
558 cmsToneCurve** DataSetIn;
559 cmsToneCurve** DataSetOut;
562 // This is a loosy optimization! does not apply in floating-point cases
563 if (_cmsFormatterIsFloat(*InputFormat) || _cmsFormatterIsFloat(*OutputFormat)) return FALSE;
565 ColorSpace = _cmsICCcolorSpace(T_COLORSPACE(*InputFormat));
566 OutputColorSpace = _cmsICCcolorSpace(T_COLORSPACE(*OutputFormat));
567 nGridPoints = _cmsReasonableGridpointsByColorspace(ColorSpace, *dwFlags);
569 // For empty LUTs, 2 points are enough
570 if (cmsPipelineStageCount(*Lut) == 0)
575 // Named color pipelines cannot be optimized either
576 for (mpe = cmsPipelineGetPtrToFirstStage(Src);
578 mpe = cmsStageNext(mpe)) {
579 if (cmsStageType(mpe) == cmsSigNamedColorElemType) return FALSE;
582 // Allocate an empty LUT
583 Dest = cmsPipelineAlloc(Src ->ContextID, Src ->InputChannels, Src ->OutputChannels);
584 if (!Dest) return FALSE;
586 // Prelinearization tables are kept unless indicated by flags
587 if (*dwFlags & cmsFLAGS_CLUT_PRE_LINEARIZATION) {
589 // Get a pointer to the prelinearization element
590 cmsStage* PreLin = cmsPipelineGetPtrToFirstStage(Src);
593 if (PreLin ->Type == cmsSigCurveSetElemType) {
595 // Maybe this is a linear tram, so we can avoid the whole stuff
596 if (!AllCurvesAreLinear(PreLin)) {
598 // All seems ok, proceed.
599 NewPreLin = cmsStageDup(PreLin);
600 if(!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, NewPreLin))
603 // Remove prelinearization. Since we have duplicated the curve
604 // in destination LUT, the sampling shoud be applied after this stage.
605 cmsPipelineUnlinkStage(Src, cmsAT_BEGIN, &KeepPreLin);
611 CLUT = cmsStageAllocCLut16bit(Src ->ContextID, nGridPoints, Src ->InputChannels, Src->OutputChannels, NULL);
612 if (CLUT == NULL) return FALSE;
614 // Add the CLUT to the destination LUT
615 if (!cmsPipelineInsertStage(Dest, cmsAT_END, CLUT)) {
619 // Postlinearization tables are kept unless indicated by flags
620 if (*dwFlags & cmsFLAGS_CLUT_POST_LINEARIZATION) {
622 // Get a pointer to the postlinearization if present
623 cmsStage* PostLin = cmsPipelineGetPtrToLastStage(Src);
626 if (cmsStageType(PostLin) == cmsSigCurveSetElemType) {
628 // Maybe this is a linear tram, so we can avoid the whole stuff
629 if (!AllCurvesAreLinear(PostLin)) {
631 // All seems ok, proceed.
632 NewPostLin = cmsStageDup(PostLin);
633 if (!cmsPipelineInsertStage(Dest, cmsAT_END, NewPostLin))
636 // In destination LUT, the sampling shoud be applied after this stage.
637 cmsPipelineUnlinkStage(Src, cmsAT_END, &KeepPostLin);
642 // Now its time to do the sampling. We have to ignore pre/post linearization
643 // The source LUT whithout pre/post curves is passed as parameter.
644 if (!cmsStageSampleCLut16bit(CLUT, XFormSampler16, (void*) Src, 0)) {
646 // Ops, something went wrong, Restore stages
647 if (KeepPreLin != NULL) {
648 if (!cmsPipelineInsertStage(Src, cmsAT_BEGIN, KeepPreLin)) {
649 _cmsAssert(0); // This never happens
652 if (KeepPostLin != NULL) {
653 if (!cmsPipelineInsertStage(Src, cmsAT_END, KeepPostLin)) {
654 _cmsAssert(0); // This never happens
657 cmsPipelineFree(Dest);
663 if (KeepPreLin != NULL) cmsStageFree(KeepPreLin);
664 if (KeepPostLin != NULL) cmsStageFree(KeepPostLin);
665 cmsPipelineFree(Src);
667 DataCLUT = (_cmsStageCLutData*) CLUT ->Data;
669 if (NewPreLin == NULL) DataSetIn = NULL;
670 else DataSetIn = ((_cmsStageToneCurvesData*) NewPreLin ->Data) ->TheCurves;
672 if (NewPostLin == NULL) DataSetOut = NULL;
673 else DataSetOut = ((_cmsStageToneCurvesData*) NewPostLin ->Data) ->TheCurves;
676 if (DataSetIn == NULL && DataSetOut == NULL) {
678 _cmsPipelineSetOptimizationParameters(Dest, (_cmsOPTeval16Fn) DataCLUT->Params->Interpolation.Lerp16, DataCLUT->Params, NULL, NULL);
682 p16 = PrelinOpt16alloc(Dest ->ContextID,
684 Dest ->InputChannels,
686 Dest ->OutputChannels,
689 _cmsPipelineSetOptimizationParameters(Dest, PrelinEval16, (void*) p16, PrelinOpt16free, Prelin16dup);
693 // Don't fix white on absolute colorimetric
694 if (Intent == INTENT_ABSOLUTE_COLORIMETRIC)
695 *dwFlags |= cmsFLAGS_NOWHITEONWHITEFIXUP;
697 if (!(*dwFlags & cmsFLAGS_NOWHITEONWHITEFIXUP)) {
699 FixWhiteMisalignment(Dest, ColorSpace, OutputColorSpace);
705 cmsUNUSED_PARAMETER(Intent);
709 // -----------------------------------------------------------------------------------------------------------------------------------------------
710 // Fixes the gamma balancing of transform. This is described in my paper "Prelinearization Stages on
711 // Color-Management Application-Specific Integrated Circuits (ASICs)" presented at NIP24. It only works
712 // for RGB transforms. See the paper for more details
713 // -----------------------------------------------------------------------------------------------------------------------------------------------
716 // Normalize endpoints by slope limiting max and min. This assures endpoints as well.
717 // Descending curves are handled as well.
719 void SlopeLimiting(cmsToneCurve* g)
721 int BeginVal, EndVal;
722 int AtBegin = (int) floor((cmsFloat64Number) g ->nEntries * 0.02 + 0.5); // Cutoff at 2%
723 int AtEnd = g ->nEntries - AtBegin - 1; // And 98%
724 cmsFloat64Number Val, Slope, beta;
727 if (cmsIsToneCurveDescending(g)) {
728 BeginVal = 0xffff; EndVal = 0;
731 BeginVal = 0; EndVal = 0xffff;
734 // Compute slope and offset for begin of curve
735 Val = g ->Table16[AtBegin];
736 Slope = (Val - BeginVal) / AtBegin;
737 beta = Val - Slope * AtBegin;
739 for (i=0; i < AtBegin; i++)
740 g ->Table16[i] = _cmsQuickSaturateWord(i * Slope + beta);
742 // Compute slope and offset for the end
743 Val = g ->Table16[AtEnd];
744 Slope = (EndVal - Val) / AtBegin; // AtBegin holds the X interval, which is same in both cases
745 beta = Val - Slope * AtEnd;
747 for (i = AtEnd; i < (int) g ->nEntries; i++)
748 g ->Table16[i] = _cmsQuickSaturateWord(i * Slope + beta);
752 // Precomputes tables for 8-bit on input devicelink.
754 Prelin8Data* PrelinOpt8alloc(cmsContext ContextID, const cmsInterpParams* p, cmsToneCurve* G[3])
757 cmsUInt16Number Input[3];
758 cmsS15Fixed16Number v1, v2, v3;
761 p8 = _cmsMallocZero(ContextID, sizeof(Prelin8Data));
762 if (p8 == NULL) return NULL;
764 // Since this only works for 8 bit input, values comes always as x * 257,
765 // we can safely take msb byte (x << 8 + x)
767 for (i=0; i < 256; i++) {
771 // Get 16-bit representation
772 Input[0] = cmsEvalToneCurve16(G[0], FROM_8_TO_16(i));
773 Input[1] = cmsEvalToneCurve16(G[1], FROM_8_TO_16(i));
774 Input[2] = cmsEvalToneCurve16(G[2], FROM_8_TO_16(i));
777 Input[0] = FROM_8_TO_16(i);
778 Input[1] = FROM_8_TO_16(i);
779 Input[2] = FROM_8_TO_16(i);
783 // Move to 0..1.0 in fixed domain
784 v1 = _cmsToFixedDomain(Input[0] * p -> Domain[0]);
785 v2 = _cmsToFixedDomain(Input[1] * p -> Domain[1]);
786 v3 = _cmsToFixedDomain(Input[2] * p -> Domain[2]);
788 // Store the precalculated table of nodes
789 p8 ->X0[i] = (p->opta[2] * FIXED_TO_INT(v1));
790 p8 ->Y0[i] = (p->opta[1] * FIXED_TO_INT(v2));
791 p8 ->Z0[i] = (p->opta[0] * FIXED_TO_INT(v3));
793 // Store the precalculated table of offsets
794 p8 ->rx[i] = (cmsUInt16Number) FIXED_REST_TO_INT(v1);
795 p8 ->ry[i] = (cmsUInt16Number) FIXED_REST_TO_INT(v2);
796 p8 ->rz[i] = (cmsUInt16Number) FIXED_REST_TO_INT(v3);
799 p8 ->ContextID = ContextID;
806 void Prelin8free(cmsContext ContextID, void* ptr)
808 _cmsFree(ContextID, ptr);
812 void* Prelin8dup(cmsContext ContextID, const void* ptr)
814 return _cmsDupMem(ContextID, ptr, sizeof(Prelin8Data));
819 // A optimized interpolation for 8-bit input.
820 #define DENS(i,j,k) (LutTable[(i)+(j)+(k)+OutChan])
822 void PrelinEval8(register const cmsUInt16Number Input[],
823 register cmsUInt16Number Output[],
824 register const void* D)
827 cmsUInt8Number r, g, b;
828 cmsS15Fixed16Number rx, ry, rz;
829 cmsS15Fixed16Number c0, c1, c2, c3, Rest;
831 register cmsS15Fixed16Number X0, X1, Y0, Y1, Z0, Z1;
832 Prelin8Data* p8 = (Prelin8Data*) D;
833 register const cmsInterpParams* p = p8 ->p;
834 int TotalOut = p -> nOutputs;
835 const cmsUInt16Number* LutTable = p -> Table;
849 X1 = X0 + ((rx == 0) ? 0 : p ->opta[2]);
850 Y1 = Y0 + ((ry == 0) ? 0 : p ->opta[1]);
851 Z1 = Z0 + ((rz == 0) ? 0 : p ->opta[0]);
854 // These are the 6 Tetrahedral
855 for (OutChan=0; OutChan < TotalOut; OutChan++) {
857 c0 = DENS(X0, Y0, Z0);
859 if (rx >= ry && ry >= rz)
861 c1 = DENS(X1, Y0, Z0) - c0;
862 c2 = DENS(X1, Y1, Z0) - DENS(X1, Y0, Z0);
863 c3 = DENS(X1, Y1, Z1) - DENS(X1, Y1, Z0);
866 if (rx >= rz && rz >= ry)
868 c1 = DENS(X1, Y0, Z0) - c0;
869 c2 = DENS(X1, Y1, Z1) - DENS(X1, Y0, Z1);
870 c3 = DENS(X1, Y0, Z1) - DENS(X1, Y0, Z0);
873 if (rz >= rx && rx >= ry)
875 c1 = DENS(X1, Y0, Z1) - DENS(X0, Y0, Z1);
876 c2 = DENS(X1, Y1, Z1) - DENS(X1, Y0, Z1);
877 c3 = DENS(X0, Y0, Z1) - c0;
880 if (ry >= rx && rx >= rz)
882 c1 = DENS(X1, Y1, Z0) - DENS(X0, Y1, Z0);
883 c2 = DENS(X0, Y1, Z0) - c0;
884 c3 = DENS(X1, Y1, Z1) - DENS(X1, Y1, Z0);
887 if (ry >= rz && rz >= rx)
889 c1 = DENS(X1, Y1, Z1) - DENS(X0, Y1, Z1);
890 c2 = DENS(X0, Y1, Z0) - c0;
891 c3 = DENS(X0, Y1, Z1) - DENS(X0, Y1, Z0);
894 if (rz >= ry && ry >= rx)
896 c1 = DENS(X1, Y1, Z1) - DENS(X0, Y1, Z1);
897 c2 = DENS(X0, Y1, Z1) - DENS(X0, Y0, Z1);
898 c3 = DENS(X0, Y0, Z1) - c0;
905 Rest = c1 * rx + c2 * ry + c3 * rz + 0x8001;
906 Output[OutChan] = (cmsUInt16Number)c0 + ((Rest + (Rest>>16))>>16);
914 // Curves that contain wide empty areas are not optimizeable
916 cmsBool IsDegenerated(const cmsToneCurve* g)
918 int i, Zeros = 0, Poles = 0;
919 int nEntries = g ->nEntries;
921 for (i=0; i < nEntries; i++) {
923 if (g ->Table16[i] == 0x0000) Zeros++;
924 if (g ->Table16[i] == 0xffff) Poles++;
927 if (Zeros == 1 && Poles == 1) return FALSE; // For linear tables
928 if (Zeros > (nEntries / 4)) return TRUE; // Degenerated, mostly zeros
929 if (Poles > (nEntries / 4)) return TRUE; // Degenerated, mostly poles
934 // --------------------------------------------------------------------------------------------------------------
935 // We need xput over here
938 cmsBool OptimizeByComputingLinearization(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags)
940 cmsPipeline* OriginalLut;
942 cmsToneCurve *Trans[cmsMAXCHANNELS], *TransReverse[cmsMAXCHANNELS];
943 cmsUInt32Number t, i;
944 cmsFloat32Number v, In[cmsMAXCHANNELS], Out[cmsMAXCHANNELS];
945 cmsBool lIsSuitable, lIsLinear;
946 cmsPipeline* OptimizedLUT = NULL, *LutPlusCurves = NULL;
947 cmsStage* OptimizedCLUTmpe;
948 cmsColorSpaceSignature ColorSpace, OutputColorSpace;
949 cmsStage* OptimizedPrelinMpe;
951 cmsToneCurve** OptimizedPrelinCurves;
952 _cmsStageCLutData* OptimizedPrelinCLUT;
955 // This is a loosy optimization! does not apply in floating-point cases
956 if (_cmsFormatterIsFloat(*InputFormat) || _cmsFormatterIsFloat(*OutputFormat)) return FALSE;
959 if (T_COLORSPACE(*InputFormat) != PT_RGB) return FALSE;
960 if (T_COLORSPACE(*OutputFormat) != PT_RGB) return FALSE;
963 // On 16 bits, user has to specify the feature
964 if (!_cmsFormatterIs8bit(*InputFormat)) {
965 if (!(*dwFlags & cmsFLAGS_CLUT_PRE_LINEARIZATION)) return FALSE;
970 // Named color pipelines cannot be optimized either
971 for (mpe = cmsPipelineGetPtrToFirstStage(OriginalLut);
973 mpe = cmsStageNext(mpe)) {
974 if (cmsStageType(mpe) == cmsSigNamedColorElemType) return FALSE;
977 ColorSpace = _cmsICCcolorSpace(T_COLORSPACE(*InputFormat));
978 OutputColorSpace = _cmsICCcolorSpace(T_COLORSPACE(*OutputFormat));
979 nGridPoints = _cmsReasonableGridpointsByColorspace(ColorSpace, *dwFlags);
981 // Empty gamma containers
982 memset(Trans, 0, sizeof(Trans));
983 memset(TransReverse, 0, sizeof(TransReverse));
985 for (t = 0; t < OriginalLut ->InputChannels; t++) {
986 Trans[t] = cmsBuildTabulatedToneCurve16(OriginalLut ->ContextID, PRELINEARIZATION_POINTS, NULL);
987 if (Trans[t] == NULL) goto Error;
990 // Populate the curves
991 for (i=0; i < PRELINEARIZATION_POINTS; i++) {
993 v = (cmsFloat32Number) ((cmsFloat64Number) i / (PRELINEARIZATION_POINTS - 1));
995 // Feed input with a gray ramp
996 for (t=0; t < OriginalLut ->InputChannels; t++)
999 // Evaluate the gray value
1000 cmsPipelineEvalFloat(In, Out, OriginalLut);
1002 // Store result in curve
1003 for (t=0; t < OriginalLut ->InputChannels; t++)
1004 Trans[t] ->Table16[i] = _cmsQuickSaturateWord(Out[t] * 65535.0);
1007 // Slope-limit the obtained curves
1008 for (t = 0; t < OriginalLut ->InputChannels; t++)
1009 SlopeLimiting(Trans[t]);
1011 // Check for validity
1014 for (t=0; (lIsSuitable && (t < OriginalLut ->InputChannels)); t++) {
1016 // Exclude if already linear
1017 if (!cmsIsToneCurveLinear(Trans[t]))
1020 // Exclude if non-monotonic
1021 if (!cmsIsToneCurveMonotonic(Trans[t]))
1022 lIsSuitable = FALSE;
1024 if (IsDegenerated(Trans[t]))
1025 lIsSuitable = FALSE;
1028 // If it is not suitable, just quit
1029 if (!lIsSuitable) goto Error;
1031 // Invert curves if possible
1032 for (t = 0; t < OriginalLut ->InputChannels; t++) {
1033 TransReverse[t] = cmsReverseToneCurveEx(PRELINEARIZATION_POINTS, Trans[t]);
1034 if (TransReverse[t] == NULL) goto Error;
1037 // Now inset the reversed curves at the begin of transform
1038 LutPlusCurves = cmsPipelineDup(OriginalLut);
1039 if (LutPlusCurves == NULL) goto Error;
1041 if (!cmsPipelineInsertStage(LutPlusCurves, cmsAT_BEGIN, cmsStageAllocToneCurves(OriginalLut ->ContextID, OriginalLut ->InputChannels, TransReverse)))
1044 // Create the result LUT
1045 OptimizedLUT = cmsPipelineAlloc(OriginalLut ->ContextID, OriginalLut ->InputChannels, OriginalLut ->OutputChannels);
1046 if (OptimizedLUT == NULL) goto Error;
1048 OptimizedPrelinMpe = cmsStageAllocToneCurves(OriginalLut ->ContextID, OriginalLut ->InputChannels, Trans);
1050 // Create and insert the curves at the beginning
1051 if (!cmsPipelineInsertStage(OptimizedLUT, cmsAT_BEGIN, OptimizedPrelinMpe))
1054 // Allocate the CLUT for result
1055 OptimizedCLUTmpe = cmsStageAllocCLut16bit(OriginalLut ->ContextID, nGridPoints, OriginalLut ->InputChannels, OriginalLut ->OutputChannels, NULL);
1057 // Add the CLUT to the destination LUT
1058 if (!cmsPipelineInsertStage(OptimizedLUT, cmsAT_END, OptimizedCLUTmpe))
1062 if (!cmsStageSampleCLut16bit(OptimizedCLUTmpe, XFormSampler16, (void*) LutPlusCurves, 0)) goto Error;
1065 for (t = 0; t < OriginalLut ->InputChannels; t++) {
1067 if (Trans[t]) cmsFreeToneCurve(Trans[t]);
1068 if (TransReverse[t]) cmsFreeToneCurve(TransReverse[t]);
1071 cmsPipelineFree(LutPlusCurves);
1074 OptimizedPrelinCurves = _cmsStageGetPtrToCurveSet(OptimizedPrelinMpe);
1075 OptimizedPrelinCLUT = (_cmsStageCLutData*) OptimizedCLUTmpe ->Data;
1077 // Set the evaluator if 8-bit
1078 if (_cmsFormatterIs8bit(*InputFormat)) {
1080 Prelin8Data* p8 = PrelinOpt8alloc(OptimizedLUT ->ContextID,
1081 OptimizedPrelinCLUT ->Params,
1082 OptimizedPrelinCurves);
1083 if (p8 == NULL) return FALSE;
1085 _cmsPipelineSetOptimizationParameters(OptimizedLUT, PrelinEval8, (void*) p8, Prelin8free, Prelin8dup);
1090 Prelin16Data* p16 = PrelinOpt16alloc(OptimizedLUT ->ContextID,
1091 OptimizedPrelinCLUT ->Params,
1092 3, OptimizedPrelinCurves, 3, NULL);
1093 if (p16 == NULL) return FALSE;
1095 _cmsPipelineSetOptimizationParameters(OptimizedLUT, PrelinEval16, (void*) p16, PrelinOpt16free, Prelin16dup);
1099 // Don't fix white on absolute colorimetric
1100 if (Intent == INTENT_ABSOLUTE_COLORIMETRIC)
1101 *dwFlags |= cmsFLAGS_NOWHITEONWHITEFIXUP;
1103 if (!(*dwFlags & cmsFLAGS_NOWHITEONWHITEFIXUP)) {
1105 if (!FixWhiteMisalignment(OptimizedLUT, ColorSpace, OutputColorSpace)) {
1111 // And return the obtained LUT
1113 cmsPipelineFree(OriginalLut);
1114 *Lut = OptimizedLUT;
1119 for (t = 0; t < OriginalLut ->InputChannels; t++) {
1121 if (Trans[t]) cmsFreeToneCurve(Trans[t]);
1122 if (TransReverse[t]) cmsFreeToneCurve(TransReverse[t]);
1125 if (LutPlusCurves != NULL) cmsPipelineFree(LutPlusCurves);
1126 if (OptimizedLUT != NULL) cmsPipelineFree(OptimizedLUT);
1130 cmsUNUSED_PARAMETER(Intent);
1134 // Curves optimizer ------------------------------------------------------------------------------------------------------------------
1137 void CurvesFree(cmsContext ContextID, void* ptr)
1139 Curves16Data* Data = (Curves16Data*) ptr;
1142 for (i=0; i < Data -> nCurves; i++) {
1144 _cmsFree(ContextID, Data ->Curves[i]);
1147 _cmsFree(ContextID, Data ->Curves);
1148 _cmsFree(ContextID, ptr);
1152 void* CurvesDup(cmsContext ContextID, const void* ptr)
1154 Curves16Data* Data = _cmsDupMem(ContextID, ptr, sizeof(Curves16Data));
1157 if (Data == NULL) return NULL;
1159 Data ->Curves = _cmsDupMem(ContextID, Data ->Curves, Data ->nCurves * sizeof(cmsUInt16Number*));
1161 for (i=0; i < Data -> nCurves; i++) {
1162 Data ->Curves[i] = _cmsDupMem(ContextID, Data ->Curves[i], Data -> nElements * sizeof(cmsUInt16Number));
1165 return (void*) Data;
1168 // Precomputes tables for 8-bit on input devicelink.
1170 Curves16Data* CurvesAlloc(cmsContext ContextID, int nCurves, int nElements, cmsToneCurve** G)
1175 c16 = _cmsMallocZero(ContextID, sizeof(Curves16Data));
1176 if (c16 == NULL) return NULL;
1178 c16 ->nCurves = nCurves;
1179 c16 ->nElements = nElements;
1181 c16 ->Curves = _cmsCalloc(ContextID, nCurves, sizeof(cmsUInt16Number*));
1182 if (c16 ->Curves == NULL) return NULL;
1184 for (i=0; i < nCurves; i++) {
1186 c16->Curves[i] = _cmsCalloc(ContextID, nElements, sizeof(cmsUInt16Number));
1188 if (c16->Curves[i] == NULL) {
1190 for (j=0; j < i; j++) {
1191 _cmsFree(ContextID, c16->Curves[j]);
1193 _cmsFree(ContextID, c16->Curves);
1194 _cmsFree(ContextID, c16);
1198 if (nElements == 256) {
1200 for (j=0; j < nElements; j++) {
1202 c16 ->Curves[i][j] = cmsEvalToneCurve16(G[i], FROM_8_TO_16(j));
1207 for (j=0; j < nElements; j++) {
1208 c16 ->Curves[i][j] = cmsEvalToneCurve16(G[i], (cmsUInt16Number) j);
1217 void FastEvaluateCurves8(register const cmsUInt16Number In[],
1218 register cmsUInt16Number Out[],
1219 register const void* D)
1221 Curves16Data* Data = (Curves16Data*) D;
1225 for (i=0; i < Data ->nCurves; i++) {
1228 Out[i] = Data -> Curves[i][x];
1234 void FastEvaluateCurves16(register const cmsUInt16Number In[],
1235 register cmsUInt16Number Out[],
1236 register const void* D)
1238 Curves16Data* Data = (Curves16Data*) D;
1241 for (i=0; i < Data ->nCurves; i++) {
1242 Out[i] = Data -> Curves[i][In[i]];
1248 void FastIdentity16(register const cmsUInt16Number In[],
1249 register cmsUInt16Number Out[],
1250 register const void* D)
1252 cmsPipeline* Lut = (cmsPipeline*) D;
1255 for (i=0; i < Lut ->InputChannels; i++) {
1261 // If the target LUT holds only curves, the optimization procedure is to join all those
1262 // curves together. That only works on curves and does not work on matrices.
1264 cmsBool OptimizeByJoiningCurves(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags)
1266 cmsToneCurve** GammaTables = NULL;
1267 cmsFloat32Number InFloat[cmsMAXCHANNELS], OutFloat[cmsMAXCHANNELS];
1268 cmsUInt32Number i, j;
1269 cmsPipeline* Src = *Lut;
1270 cmsPipeline* Dest = NULL;
1272 cmsStage* ObtainedCurves = NULL;
1275 // This is a loosy optimization! does not apply in floating-point cases
1276 if (_cmsFormatterIsFloat(*InputFormat) || _cmsFormatterIsFloat(*OutputFormat)) return FALSE;
1278 // Only curves in this LUT?
1279 for (mpe = cmsPipelineGetPtrToFirstStage(Src);
1281 mpe = cmsStageNext(mpe)) {
1282 if (cmsStageType(mpe) != cmsSigCurveSetElemType) return FALSE;
1285 // Allocate an empty LUT
1286 Dest = cmsPipelineAlloc(Src ->ContextID, Src ->InputChannels, Src ->OutputChannels);
1287 if (Dest == NULL) return FALSE;
1289 // Create target curves
1290 GammaTables = (cmsToneCurve**) _cmsCalloc(Src ->ContextID, Src ->InputChannels, sizeof(cmsToneCurve*));
1291 if (GammaTables == NULL) goto Error;
1293 for (i=0; i < Src ->InputChannels; i++) {
1294 GammaTables[i] = cmsBuildTabulatedToneCurve16(Src ->ContextID, PRELINEARIZATION_POINTS, NULL);
1295 if (GammaTables[i] == NULL) goto Error;
1298 // Compute 16 bit result by using floating point
1299 for (i=0; i < PRELINEARIZATION_POINTS; i++) {
1301 for (j=0; j < Src ->InputChannels; j++)
1302 InFloat[j] = (cmsFloat32Number) ((cmsFloat64Number) i / (PRELINEARIZATION_POINTS - 1));
1304 cmsPipelineEvalFloat(InFloat, OutFloat, Src);
1306 for (j=0; j < Src ->InputChannels; j++)
1307 GammaTables[j] -> Table16[i] = _cmsQuickSaturateWord(OutFloat[j] * 65535.0);
1310 ObtainedCurves = cmsStageAllocToneCurves(Src ->ContextID, Src ->InputChannels, GammaTables);
1311 if (ObtainedCurves == NULL) goto Error;
1313 for (i=0; i < Src ->InputChannels; i++) {
1314 cmsFreeToneCurve(GammaTables[i]);
1315 GammaTables[i] = NULL;
1318 if (GammaTables != NULL) _cmsFree(Src ->ContextID, GammaTables);
1320 // Maybe the curves are linear at the end
1321 if (!AllCurvesAreLinear(ObtainedCurves)) {
1323 if (!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, ObtainedCurves))
1326 // If the curves are to be applied in 8 bits, we can save memory
1327 if (_cmsFormatterIs8bit(*InputFormat)) {
1329 _cmsStageToneCurvesData* Data = (_cmsStageToneCurvesData*) ObtainedCurves ->Data;
1330 Curves16Data* c16 = CurvesAlloc(Dest ->ContextID, Data ->nCurves, 256, Data ->TheCurves);
1332 if (c16 == NULL) goto Error;
1333 *dwFlags |= cmsFLAGS_NOCACHE;
1334 _cmsPipelineSetOptimizationParameters(Dest, FastEvaluateCurves8, c16, CurvesFree, CurvesDup);
1339 _cmsStageToneCurvesData* Data = (_cmsStageToneCurvesData*) cmsStageData(ObtainedCurves);
1340 Curves16Data* c16 = CurvesAlloc(Dest ->ContextID, Data ->nCurves, 65536, Data ->TheCurves);
1342 if (c16 == NULL) goto Error;
1343 *dwFlags |= cmsFLAGS_NOCACHE;
1344 _cmsPipelineSetOptimizationParameters(Dest, FastEvaluateCurves16, c16, CurvesFree, CurvesDup);
1349 // LUT optimizes to nothing. Set the identity LUT
1350 cmsStageFree(ObtainedCurves);
1352 if (!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, cmsStageAllocIdentity(Dest ->ContextID, Src ->InputChannels)))
1355 *dwFlags |= cmsFLAGS_NOCACHE;
1356 _cmsPipelineSetOptimizationParameters(Dest, FastIdentity16, (void*) Dest, NULL, NULL);
1360 cmsPipelineFree(Src);
1366 if (ObtainedCurves != NULL) cmsStageFree(ObtainedCurves);
1367 if (GammaTables != NULL) {
1368 for (i=0; i < Src ->InputChannels; i++) {
1369 if (GammaTables[i] != NULL) cmsFreeToneCurve(GammaTables[i]);
1372 _cmsFree(Src ->ContextID, GammaTables);
1375 if (Dest != NULL) cmsPipelineFree(Dest);
1378 cmsUNUSED_PARAMETER(Intent);
1379 cmsUNUSED_PARAMETER(InputFormat);
1380 cmsUNUSED_PARAMETER(OutputFormat);
1381 cmsUNUSED_PARAMETER(dwFlags);
1384 // -------------------------------------------------------------------------------------------------------------------------------------
1385 // LUT is Shaper - Matrix - Matrix - Shaper, which is very frequent when combining two matrix-shaper profiles
1389 void FreeMatShaper(cmsContext ContextID, void* Data)
1391 if (Data != NULL) _cmsFree(ContextID, Data);
1395 void* DupMatShaper(cmsContext ContextID, const void* Data)
1397 return _cmsDupMem(ContextID, Data, sizeof(MatShaper8Data));
1401 // A fast matrix-shaper evaluator for 8 bits. This is a bit ticky since I'm using 1.14 signed fixed point
1402 // to accomplish some performance. Actually it takes 256x3 16 bits tables and 16385 x 3 tables of 8 bits,
1403 // in total about 50K, and the performance boost is huge!
1405 void MatShaperEval16(register const cmsUInt16Number In[],
1406 register cmsUInt16Number Out[],
1407 register const void* D)
1409 MatShaper8Data* p = (MatShaper8Data*) D;
1410 cmsS1Fixed14Number l1, l2, l3, r, g, b;
1411 cmsUInt32Number ri, gi, bi;
1413 // In this case (and only in this case!) we can use this simplification since
1414 // In[] is assured to come from a 8 bit number. (a << 8 | a)
1419 // Across first shaper, which also converts to 1.14 fixed point
1420 r = p->Shaper1R[ri];
1421 g = p->Shaper1G[gi];
1422 b = p->Shaper1B[bi];
1424 // Evaluate the matrix in 1.14 fixed point
1425 l1 = (p->Mat[0][0] * r + p->Mat[0][1] * g + p->Mat[0][2] * b + p->Off[0] + 0x2000) >> 14;
1426 l2 = (p->Mat[1][0] * r + p->Mat[1][1] * g + p->Mat[1][2] * b + p->Off[1] + 0x2000) >> 14;
1427 l3 = (p->Mat[2][0] * r + p->Mat[2][1] * g + p->Mat[2][2] * b + p->Off[2] + 0x2000) >> 14;
1429 // Now we have to clip to 0..1.0 range
1430 ri = (l1 < 0) ? 0 : ((l1 > 16384) ? 16384 : l1);
1431 gi = (l2 < 0) ? 0 : ((l2 > 16384) ? 16384 : l2);
1432 bi = (l3 < 0) ? 0 : ((l3 > 16384) ? 16384 : l3);
1434 // And across second shaper,
1435 Out[0] = p->Shaper2R[ri];
1436 Out[1] = p->Shaper2G[gi];
1437 Out[2] = p->Shaper2B[bi];
1441 // This table converts from 8 bits to 1.14 after applying the curve
1443 void FillFirstShaper(cmsS1Fixed14Number* Table, cmsToneCurve* Curve)
1446 cmsFloat32Number R, y;
1448 for (i=0; i < 256; i++) {
1450 R = (cmsFloat32Number) (i / 255.0);
1451 y = cmsEvalToneCurveFloat(Curve, R);
1453 Table[i] = DOUBLE_TO_1FIXED14(y);
1457 // This table converts form 1.14 (being 0x4000 the last entry) to 8 bits after applying the curve
1459 void FillSecondShaper(cmsUInt16Number* Table, cmsToneCurve* Curve, cmsBool Is8BitsOutput)
1462 cmsFloat32Number R, Val;
1464 for (i=0; i < 16385; i++) {
1466 R = (cmsFloat32Number) (i / 16384.0);
1467 Val = cmsEvalToneCurveFloat(Curve, R); // Val comes 0..1.0
1469 if (Is8BitsOutput) {
1471 // If 8 bits output, we can optimize further by computing the / 257 part.
1472 // first we compute the resulting byte and then we store the byte times
1473 // 257. This quantization allows to round very quick by doing a >> 8, but
1474 // since the low byte is always equal to msb, we can do a & 0xff and this works!
1475 cmsUInt16Number w = _cmsQuickSaturateWord(Val * 65535.0);
1476 cmsUInt8Number b = FROM_16_TO_8(w);
1478 Table[i] = FROM_8_TO_16(b);
1480 else Table[i] = _cmsQuickSaturateWord(Val * 65535.0);
1484 // Compute the matrix-shaper structure
1486 cmsBool SetMatShaper(cmsPipeline* Dest, cmsToneCurve* Curve1[3], cmsMAT3* Mat, cmsVEC3* Off, cmsToneCurve* Curve2[3], cmsUInt32Number* OutputFormat)
1490 cmsBool Is8Bits = _cmsFormatterIs8bit(*OutputFormat);
1492 // Allocate a big chuck of memory to store precomputed tables
1493 p = (MatShaper8Data*) _cmsMalloc(Dest ->ContextID, sizeof(MatShaper8Data));
1494 if (p == NULL) return FALSE;
1496 p -> ContextID = Dest -> ContextID;
1498 // Precompute tables
1499 FillFirstShaper(p ->Shaper1R, Curve1[0]);
1500 FillFirstShaper(p ->Shaper1G, Curve1[1]);
1501 FillFirstShaper(p ->Shaper1B, Curve1[2]);
1503 FillSecondShaper(p ->Shaper2R, Curve2[0], Is8Bits);
1504 FillSecondShaper(p ->Shaper2G, Curve2[1], Is8Bits);
1505 FillSecondShaper(p ->Shaper2B, Curve2[2], Is8Bits);
1507 // Convert matrix to nFixed14. Note that those values may take more than 16 bits as
1508 for (i=0; i < 3; i++) {
1509 for (j=0; j < 3; j++) {
1510 p ->Mat[i][j] = DOUBLE_TO_1FIXED14(Mat->v[i].n[j]);
1514 for (i=0; i < 3; i++) {
1520 p ->Off[i] = DOUBLE_TO_1FIXED14(Off->n[i]);
1524 // Mark as optimized for faster formatter
1526 *OutputFormat |= OPTIMIZED_SH(1);
1528 // Fill function pointers
1529 _cmsPipelineSetOptimizationParameters(Dest, MatShaperEval16, (void*) p, FreeMatShaper, DupMatShaper);
1533 // 8 bits on input allows matrix-shaper boot up to 25 Mpixels per second on RGB. That's fast!
1534 // TODO: Allow a third matrix for abs. colorimetric
1536 cmsBool OptimizeMatrixShaper(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags)
1538 cmsStage* Curve1, *Curve2;
1539 cmsStage* Matrix1, *Matrix2;
1540 _cmsStageMatrixData* Data1;
1541 _cmsStageMatrixData* Data2;
1543 cmsBool IdentityMat;
1544 cmsPipeline* Dest, *Src;
1546 // Only works on RGB to RGB
1547 if (T_CHANNELS(*InputFormat) != 3 || T_CHANNELS(*OutputFormat) != 3) return FALSE;
1549 // Only works on 8 bit input
1550 if (!_cmsFormatterIs8bit(*InputFormat)) return FALSE;
1552 // Seems suitable, proceed
1555 // Check for shaper-matrix-matrix-shaper structure, that is what this optimizer stands for
1556 if (!cmsPipelineCheckAndRetreiveStages(Src, 4,
1557 cmsSigCurveSetElemType, cmsSigMatrixElemType, cmsSigMatrixElemType, cmsSigCurveSetElemType,
1558 &Curve1, &Matrix1, &Matrix2, &Curve2)) return FALSE;
1560 // Get both matrices
1561 Data1 = (_cmsStageMatrixData*) cmsStageData(Matrix1);
1562 Data2 = (_cmsStageMatrixData*) cmsStageData(Matrix2);
1564 // Input offset should be zero
1565 if (Data1 ->Offset != NULL) return FALSE;
1567 // Multiply both matrices to get the result
1568 _cmsMAT3per(&res, (cmsMAT3*) Data2 ->Double, (cmsMAT3*) Data1 ->Double);
1570 // Now the result is in res + Data2 -> Offset. Maybe is a plain identity?
1571 IdentityMat = FALSE;
1572 if (_cmsMAT3isIdentity(&res) && Data2 ->Offset == NULL) {
1574 // We can get rid of full matrix
1578 // Allocate an empty LUT
1579 Dest = cmsPipelineAlloc(Src ->ContextID, Src ->InputChannels, Src ->OutputChannels);
1580 if (!Dest) return FALSE;
1582 // Assamble the new LUT
1583 if (!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, cmsStageDup(Curve1)))
1587 if (!cmsPipelineInsertStage(Dest, cmsAT_END, cmsStageAllocMatrix(Dest ->ContextID, 3, 3, (const cmsFloat64Number*) &res, Data2 ->Offset)))
1589 if (!cmsPipelineInsertStage(Dest, cmsAT_END, cmsStageDup(Curve2)))
1592 // If identity on matrix, we can further optimize the curves, so call the join curves routine
1595 OptimizeByJoiningCurves(&Dest, Intent, InputFormat, OutputFormat, dwFlags);
1598 _cmsStageToneCurvesData* mpeC1 = (_cmsStageToneCurvesData*) cmsStageData(Curve1);
1599 _cmsStageToneCurvesData* mpeC2 = (_cmsStageToneCurvesData*) cmsStageData(Curve2);
1601 // In this particular optimization, cach� does not help as it takes more time to deal with
1602 // the cach� that with the pixel handling
1603 *dwFlags |= cmsFLAGS_NOCACHE;
1605 // Setup the optimizarion routines
1606 SetMatShaper(Dest, mpeC1 ->TheCurves, &res, (cmsVEC3*) Data2 ->Offset, mpeC2->TheCurves, OutputFormat);
1609 cmsPipelineFree(Src);
1613 // Leave Src unchanged
1614 cmsPipelineFree(Dest);
1619 // -------------------------------------------------------------------------------------------------------------------------------------
1620 // Optimization plug-ins
1622 // List of optimizations
1623 typedef struct _cmsOptimizationCollection_st {
1625 _cmsOPToptimizeFn OptimizePtr;
1627 struct _cmsOptimizationCollection_st *Next;
1629 } _cmsOptimizationCollection;
1632 // The built-in list. We currently implement 4 types of optimizations. Joining of curves, matrix-shaper, linearization and resampling
1633 static _cmsOptimizationCollection DefaultOptimization[] = {
1635 { OptimizeByJoiningCurves, &DefaultOptimization[1] },
1636 { OptimizeMatrixShaper, &DefaultOptimization[2] },
1637 { OptimizeByComputingLinearization, &DefaultOptimization[3] },
1638 { OptimizeByResampling, NULL }
1641 // The linked list head
1642 _cmsOptimizationPluginChunkType _cmsOptimizationPluginChunk = { NULL };
1645 // Duplicates the zone of memory used by the plug-in in the new context
1647 void DupPluginOptimizationList(struct _cmsContext_struct* ctx,
1648 const struct _cmsContext_struct* src)
1650 _cmsOptimizationPluginChunkType newHead = { NULL };
1651 _cmsOptimizationCollection* entry;
1652 _cmsOptimizationCollection* Anterior = NULL;
1653 _cmsOptimizationPluginChunkType* head = (_cmsOptimizationPluginChunkType*) src->chunks[OptimizationPlugin];
1655 _cmsAssert(ctx != NULL);
1656 _cmsAssert(head != NULL);
1658 // Walk the list copying all nodes
1659 for (entry = head->OptimizationCollection;
1661 entry = entry ->Next) {
1663 _cmsOptimizationCollection *newEntry = ( _cmsOptimizationCollection *) _cmsSubAllocDup(ctx ->MemPool, entry, sizeof(_cmsOptimizationCollection));
1665 if (newEntry == NULL)
1668 // We want to keep the linked list order, so this is a little bit tricky
1669 newEntry -> Next = NULL;
1671 Anterior -> Next = newEntry;
1673 Anterior = newEntry;
1675 if (newHead.OptimizationCollection == NULL)
1676 newHead.OptimizationCollection = newEntry;
1679 ctx ->chunks[OptimizationPlugin] = _cmsSubAllocDup(ctx->MemPool, &newHead, sizeof(_cmsOptimizationPluginChunkType));
1682 void _cmsAllocOptimizationPluginChunk(struct _cmsContext_struct* ctx,
1683 const struct _cmsContext_struct* src)
1687 // Copy all linked list
1688 DupPluginOptimizationList(ctx, src);
1691 static _cmsOptimizationPluginChunkType OptimizationPluginChunkType = { NULL };
1692 ctx ->chunks[OptimizationPlugin] = _cmsSubAllocDup(ctx ->MemPool, &OptimizationPluginChunkType, sizeof(_cmsOptimizationPluginChunkType));
1697 // Register new ways to optimize
1698 cmsBool _cmsRegisterOptimizationPlugin(cmsContext ContextID, cmsPluginBase* Data)
1700 cmsPluginOptimization* Plugin = (cmsPluginOptimization*) Data;
1701 _cmsOptimizationPluginChunkType* ctx = ( _cmsOptimizationPluginChunkType*) _cmsContextGetClientChunk(ContextID, OptimizationPlugin);
1702 _cmsOptimizationCollection* fl;
1706 ctx->OptimizationCollection = NULL;
1710 // Optimizer callback is required
1711 if (Plugin ->OptimizePtr == NULL) return FALSE;
1713 fl = (_cmsOptimizationCollection*) _cmsPluginMalloc(ContextID, sizeof(_cmsOptimizationCollection));
1714 if (fl == NULL) return FALSE;
1716 // Copy the parameters
1717 fl ->OptimizePtr = Plugin ->OptimizePtr;
1720 fl ->Next = ctx->OptimizationCollection;
1723 ctx ->OptimizationCollection = fl;
1729 // The entry point for LUT optimization
1730 cmsBool _cmsOptimizePipeline(cmsContext ContextID,
1731 cmsPipeline** PtrLut,
1733 cmsUInt32Number* InputFormat,
1734 cmsUInt32Number* OutputFormat,
1735 cmsUInt32Number* dwFlags)
1737 _cmsOptimizationPluginChunkType* ctx = ( _cmsOptimizationPluginChunkType*) _cmsContextGetClientChunk(ContextID, OptimizationPlugin);
1738 _cmsOptimizationCollection* Opts;
1739 cmsBool AnySuccess = FALSE;
1741 // A CLUT is being asked, so force this specific optimization
1742 if (*dwFlags & cmsFLAGS_FORCE_CLUT) {
1744 PreOptimize(*PtrLut);
1745 return OptimizeByResampling(PtrLut, Intent, InputFormat, OutputFormat, dwFlags);
1748 // Anything to optimize?
1749 if ((*PtrLut) ->Elements == NULL) {
1750 _cmsPipelineSetOptimizationParameters(*PtrLut, FastIdentity16, (void*) *PtrLut, NULL, NULL);
1754 // Try to get rid of identities and trivial conversions.
1755 AnySuccess = PreOptimize(*PtrLut);
1757 // After removal do we end with an identity?
1758 if ((*PtrLut) ->Elements == NULL) {
1759 _cmsPipelineSetOptimizationParameters(*PtrLut, FastIdentity16, (void*) *PtrLut, NULL, NULL);
1763 // Do not optimize, keep all precision
1764 if (*dwFlags & cmsFLAGS_NOOPTIMIZE)
1767 // Try plug-in optimizations
1768 for (Opts = ctx->OptimizationCollection;
1770 Opts = Opts ->Next) {
1772 // If one schema succeeded, we are done
1773 if (Opts ->OptimizePtr(PtrLut, Intent, InputFormat, OutputFormat, dwFlags)) {
1775 return TRUE; // Optimized!
1779 // Try built-in optimizations
1780 for (Opts = DefaultOptimization;
1782 Opts = Opts ->Next) {
1784 if (Opts ->OptimizePtr(PtrLut, Intent, InputFormat, OutputFormat, dwFlags)) {
1790 // Only simple optimizations succeeded