diff options
Diffstat (limited to 'third_party/lcms/src/cmsopt.c')
-rw-r--r-- | third_party/lcms/src/cmsopt.c | 273 |
1 files changed, 208 insertions, 65 deletions
diff --git a/third_party/lcms/src/cmsopt.c b/third_party/lcms/src/cmsopt.c index 76de01554c..abe26b93af 100644 --- a/third_party/lcms/src/cmsopt.c +++ b/third_party/lcms/src/cmsopt.c @@ -1,7 +1,7 @@ //--------------------------------------------------------------------------------- // // Little Color Management System -// Copyright (c) 1998-2011 Marti Maria Saguer +// Copyright (c) 1998-2016 Marti Maria Saguer // // Permission is hereby granted, free of charge, to any person obtaining // a copy of this software and associated documentation files (the "Software"), @@ -162,6 +162,89 @@ cmsBool _Remove2Op(cmsPipeline* Lut, cmsStageSignature Op1, cmsStageSignature Op return AnyOpt; } + +static +cmsBool CloseEnoughFloat(cmsFloat64Number a, cmsFloat64Number b) +{ + return fabs(b - a) < 0.00001f; +} + +static +cmsBool isFloatMatrixIdentity(const cmsMAT3* a) +{ + cmsMAT3 Identity; + int i, j; + + _cmsMAT3identity(&Identity); + + for (i = 0; i < 3; i++) + for (j = 0; j < 3; j++) + if (!CloseEnoughFloat(a->v[i].n[j], Identity.v[i].n[j])) return FALSE; + + return TRUE; +} +// if two adjacent matrices are found, multiply them. +static +cmsBool _MultiplyMatrix(cmsPipeline* Lut) +{ + cmsStage** pt1; + cmsStage** pt2; + cmsStage* chain; + cmsBool AnyOpt = FALSE; + + pt1 = &Lut->Elements; + if (*pt1 == NULL) return AnyOpt; + + while (*pt1 != NULL) { + + pt2 = &((*pt1)->Next); + if (*pt2 == NULL) return AnyOpt; + + if ((*pt1)->Implements == cmsSigMatrixElemType && (*pt2)->Implements == cmsSigMatrixElemType) { + + // Get both matrices + _cmsStageMatrixData* m1 = (_cmsStageMatrixData*) cmsStageData(*pt1); + _cmsStageMatrixData* m2 = (_cmsStageMatrixData*) cmsStageData(*pt2); + cmsMAT3 res; + + // Input offset and output offset should be zero to use this optimization + if (m1->Offset != NULL || m2 ->Offset != NULL || + cmsStageInputChannels(*pt1) != 3 || cmsStageOutputChannels(*pt1) != 3 || + cmsStageInputChannels(*pt2) != 3 || cmsStageOutputChannels(*pt2) != 3) + return FALSE; + + // Multiply both matrices to get the result + _cmsMAT3per(&res, (cmsMAT3*)m2->Double, (cmsMAT3*)m1->Double); + + // Get the next in chain afer the matrices + chain = (*pt2)->Next; + + // Remove both matrices + _RemoveElement(pt2); + _RemoveElement(pt1); + + // Now what if the result is a plain identity? + if (!isFloatMatrixIdentity(&res)) { + + // We can not get rid of full matrix + cmsStage* Multmat = cmsStageAllocMatrix(Lut->ContextID, 3, 3, (const cmsFloat64Number*) &res, NULL); + if (Multmat == NULL) return FALSE; // Should never happen + + // Recover the chain + Multmat->Next = chain; + *pt1 = Multmat; + } + + AnyOpt = TRUE; + } + else + pt1 = &((*pt1)->Next); + } + + return AnyOpt; +} + + // Preoptimize just gets rif of no-ops coming paired. Conversion from v2 to v4 followed // by a v4 to v2 and vice-versa. The elements are then discarded. static @@ -194,6 +277,9 @@ cmsBool PreOptimize(cmsPipeline* Lut) // Remove float pcs Lab conversions Opt |= _Remove2Op(Lut, cmsSigXYZ2FloatPCS, cmsSigFloatPCS2XYZ); + // Simplify matrix. + Opt |= _MultiplyMatrix(Lut); + if (Opt) AnyOpt = TRUE; } while (Opt); @@ -250,12 +336,12 @@ static void* Prelin16dup(cmsContext ContextID, const void* ptr) { Prelin16Data* p16 = (Prelin16Data*) ptr; - Prelin16Data* Duped = _cmsDupMem(ContextID, p16, sizeof(Prelin16Data)); + Prelin16Data* Duped = (Prelin16Data*) _cmsDupMem(ContextID, p16, sizeof(Prelin16Data)); if (Duped == NULL) return NULL; - Duped ->EvalCurveOut16 = (_cmsInterpFn16*)_cmsDupMem(ContextID, p16 ->EvalCurveOut16, p16 ->nOutputs * sizeof(_cmsInterpFn16)); - Duped ->ParamsCurveOut16 = (cmsInterpParams**)_cmsDupMem(ContextID, p16 ->ParamsCurveOut16, p16 ->nOutputs * sizeof(cmsInterpParams* )); + Duped->EvalCurveOut16 = (_cmsInterpFn16*) _cmsDupMem(ContextID, p16->EvalCurveOut16, p16->nOutputs * sizeof(_cmsInterpFn16)); + Duped->ParamsCurveOut16 = (cmsInterpParams**)_cmsDupMem(ContextID, p16->ParamsCurveOut16, p16->nOutputs * sizeof(cmsInterpParams*)); return Duped; } @@ -268,7 +354,7 @@ Prelin16Data* PrelinOpt16alloc(cmsContext ContextID, int nOutputs, cmsToneCurve** Out ) { int i; - Prelin16Data* p16 = _cmsMallocZero(ContextID, sizeof(Prelin16Data)); + Prelin16Data* p16 = (Prelin16Data*)_cmsMallocZero(ContextID, sizeof(Prelin16Data)); if (p16 == NULL) return NULL; p16 ->nInputs = nInputs; @@ -380,10 +466,6 @@ cmsBool PatchLUT(cmsStage* CLUT, cmsUInt16Number At[], cmsUInt16Number Value[], return FALSE; } - if (nChannelsIn != 1 && nChannelsIn != 3 && nChannelsIn != 4) { - cmsSignalError(CLUT->ContextID, cmsERROR_INTERNAL, "(internal) %d Channels are not supported on PatchLUT", nChannelsIn); - return FALSE; - } if (nChannelsIn == 4) { px = ((cmsFloat64Number) At[0] * (p16->Domain[0])) / 65535.0; @@ -447,7 +529,7 @@ cmsBool PatchLUT(cmsStage* CLUT, cmsUInt16Number At[], cmsUInt16Number Value[], return TRUE; } -// Auxiliar, to see if two values are equal or very different +// Auxiliary, to see if two values are equal or very different static cmsBool WhitesAreEqual(int n, cmsUInt16Number White1[], cmsUInt16Number White2[] ) { @@ -455,7 +537,7 @@ cmsBool WhitesAreEqual(int n, cmsUInt16Number White1[], cmsUInt16Number White2[] for (i=0; i < n; i++) { - if (abs(White1[i] - White2[i]) > 0xf000) return TRUE; // Values are so extremly different that the fixup should be avoided + if (abs(White1[i] - White2[i]) > 0xf000) return TRUE; // Values are so extremely different that the fixup should be avoided if (White1[i] != White2[i]) return FALSE; } return TRUE; @@ -593,7 +675,7 @@ cmsBool OptimizeByResampling(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt3 cmsStage* PreLin = cmsPipelineGetPtrToFirstStage(Src); // Check if suitable - if (PreLin ->Type == cmsSigCurveSetElemType) { + if (PreLin && PreLin ->Type == cmsSigCurveSetElemType) { // Maybe this is a linear tram, so we can avoid the whole stuff if (!AllCurvesAreLinear(PreLin)) { @@ -626,7 +708,7 @@ cmsBool OptimizeByResampling(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt3 cmsStage* PostLin = cmsPipelineGetPtrToLastStage(Src); // Check if suitable - if (cmsStageType(PostLin) == cmsSigCurveSetElemType) { + if (PostLin && cmsStageType(PostLin) == cmsSigCurveSetElemType) { // Maybe this is a linear tram, so we can avoid the whole stuff if (!AllCurvesAreLinear(PostLin)) { @@ -835,7 +917,7 @@ void PrelinEval8(register const cmsUInt16Number Input[], Prelin8Data* p8 = (Prelin8Data*) D; register const cmsInterpParams* p = p8 ->p; int TotalOut = p -> nOutputs; - const cmsUInt16Number* LutTable = (const cmsUInt16Number*)p -> Table; + const cmsUInt16Number* LutTable = (const cmsUInt16Number*) p->Table; r = Input[0] >> 8; g = Input[1] >> 8; @@ -928,8 +1010,8 @@ cmsBool IsDegenerated(const cmsToneCurve* g) } if (Zeros == 1 && Poles == 1) return FALSE; // For linear tables - if (Zeros > (nEntries / 4)) return TRUE; // Degenerated, mostly zeros - if (Poles > (nEntries / 4)) return TRUE; // Degenerated, mostly poles + if (Zeros > (nEntries / 20)) return TRUE; // Degenerated, many zeros + if (Poles > (nEntries / 20)) return TRUE; // Degenerated, many poles return FALSE; } @@ -951,17 +1033,19 @@ cmsBool OptimizeByComputingLinearization(cmsPipeline** Lut, cmsUInt32Number Inte cmsColorSpaceSignature ColorSpace, OutputColorSpace; cmsStage* OptimizedPrelinMpe; cmsStage* mpe; - cmsToneCurve** OptimizedPrelinCurves; - _cmsStageCLutData* OptimizedPrelinCLUT; + cmsToneCurve** OptimizedPrelinCurves; + _cmsStageCLutData* OptimizedPrelinCLUT; // This is a loosy optimization! does not apply in floating-point cases if (_cmsFormatterIsFloat(*InputFormat) || _cmsFormatterIsFloat(*OutputFormat)) return FALSE; - // Only on RGB + // Only on chunky RGB if (T_COLORSPACE(*InputFormat) != PT_RGB) return FALSE; - if (T_COLORSPACE(*OutputFormat) != PT_RGB) return FALSE; + if (T_PLANAR(*InputFormat)) return FALSE; + if (T_COLORSPACE(*OutputFormat) != PT_RGB) return FALSE; + if (T_PLANAR(*OutputFormat)) return FALSE; // On 16 bits, user has to specify the feature if (!_cmsFormatterIs8bit(*InputFormat)) { @@ -985,6 +1069,22 @@ cmsBool OptimizeByComputingLinearization(cmsPipeline** Lut, cmsUInt32Number Inte memset(Trans, 0, sizeof(Trans)); memset(TransReverse, 0, sizeof(TransReverse)); + // If the last stage of the original lut are curves, and those curves are + // degenerated, it is likely the transform is squeezing and clipping + // the output from previous CLUT. We cannot optimize this case + { + cmsStage* last = cmsPipelineGetPtrToLastStage(OriginalLut); + + if (cmsStageType(last) == cmsSigCurveSetElemType) { + + _cmsStageToneCurvesData* Data = (_cmsStageToneCurvesData*)cmsStageData(last); + for (i = 0; i < Data->nCurves; i++) { + if (IsDegenerated(Data->TheCurves[i])) + goto Error; + } + } + } + for (t = 0; t < OriginalLut ->InputChannels; t++) { Trans[t] = cmsBuildTabulatedToneCurve16(OriginalLut ->ContextID, PRELINEARIZATION_POINTS, NULL); if (Trans[t] == NULL) goto Error; @@ -1159,10 +1259,10 @@ void* CurvesDup(cmsContext ContextID, const void* ptr) if (Data == NULL) return NULL; - Data ->Curves = (cmsUInt16Number**)_cmsDupMem(ContextID, Data ->Curves, Data ->nCurves * sizeof(cmsUInt16Number*)); + Data->Curves = (cmsUInt16Number**) _cmsDupMem(ContextID, Data->Curves, Data->nCurves * sizeof(cmsUInt16Number*)); for (i=0; i < Data -> nCurves; i++) { - Data ->Curves[i] = (cmsUInt16Number*)_cmsDupMem(ContextID, Data ->Curves[i], Data -> nElements * sizeof(cmsUInt16Number)); + Data->Curves[i] = (cmsUInt16Number*) _cmsDupMem(ContextID, Data->Curves[i], Data->nElements * sizeof(cmsUInt16Number)); } return (void*) Data; @@ -1181,12 +1281,12 @@ Curves16Data* CurvesAlloc(cmsContext ContextID, int nCurves, int nElements, cmsT c16 ->nCurves = nCurves; c16 ->nElements = nElements; - c16 ->Curves = (cmsUInt16Number**)_cmsCalloc(ContextID, nCurves, sizeof(cmsUInt16Number*)); + c16->Curves = (cmsUInt16Number**) _cmsCalloc(ContextID, nCurves, sizeof(cmsUInt16Number*)); if (c16 ->Curves == NULL) return NULL; for (i=0; i < nCurves; i++) { - c16->Curves[i] = (cmsUInt16Number*)_cmsCalloc(ContextID, nElements, sizeof(cmsUInt16Number)); + c16->Curves[i] = (cmsUInt16Number*) _cmsCalloc(ContextID, nElements, sizeof(cmsUInt16Number)); if (c16->Curves[i] == NULL) { @@ -1318,7 +1418,10 @@ cmsBool OptimizeByJoiningCurves(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUI GammaTables[i] = NULL; } - if (GammaTables != NULL) _cmsFree(Src ->ContextID, GammaTables); + if (GammaTables != NULL) { + _cmsFree(Src->ContextID, GammaTables); + GammaTables = NULL; + } // Maybe the curves are linear at the end if (!AllCurvesAreLinear(ObtainedCurves)) { @@ -1543,55 +1646,89 @@ cmsBool SetMatShaper(cmsPipeline* Dest, cmsToneCurve* Curve1[3], cmsMAT3* Mat, c // Fill function pointers _cmsPipelineSetOptimizationParameters(Dest, MatShaperEval16, (void*) p, FreeMatShaper, DupMatShaper); return TRUE; -Error: - _cmsFree(Dest->ContextID, p); - return FALSE; + Error: + _cmsFree(Dest->ContextID, p); + return FALSE; } // 8 bits on input allows matrix-shaper boot up to 25 Mpixels per second on RGB. That's fast! -// TODO: Allow a third matrix for abs. colorimetric static cmsBool OptimizeMatrixShaper(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags) { - cmsStage* Curve1, *Curve2; - cmsStage* Matrix1, *Matrix2; - _cmsStageMatrixData* Data1; - _cmsStageMatrixData* Data2; - cmsMAT3 res; - cmsBool IdentityMat; - cmsPipeline* Dest, *Src; + cmsStage* Curve1, *Curve2; + cmsStage* Matrix1, *Matrix2; + cmsMAT3 res; + cmsBool IdentityMat; + cmsPipeline* Dest, *Src; + cmsFloat64Number* Offset; - // Only works on RGB to RGB - if (T_CHANNELS(*InputFormat) != 3 || T_CHANNELS(*OutputFormat) != 3) return FALSE; + // Only works on RGB to RGB + if (T_CHANNELS(*InputFormat) != 3 || T_CHANNELS(*OutputFormat) != 3) return FALSE; - // Only works on 8 bit input - if (!_cmsFormatterIs8bit(*InputFormat)) return FALSE; + // Only works on 8 bit input + if (!_cmsFormatterIs8bit(*InputFormat)) return FALSE; - // Seems suitable, proceed - Src = *Lut; + // Seems suitable, proceed + Src = *Lut; - // Check for shaper-matrix-matrix-shaper structure, that is what this optimizer stands for - if (!cmsPipelineCheckAndRetreiveStages(Src, 4, - cmsSigCurveSetElemType, cmsSigMatrixElemType, cmsSigMatrixElemType, cmsSigCurveSetElemType, - &Curve1, &Matrix1, &Matrix2, &Curve2)) return FALSE; + // Check for: + // + // shaper-matrix-matrix-shaper + // shaper-matrix-shaper + // + // Both of those constructs are possible (first because abs. colorimetric). + // additionally, In the first case, the input matrix offset should be zero. - // Get both matrices - Data1 = (_cmsStageMatrixData*) cmsStageData(Matrix1); - Data2 = (_cmsStageMatrixData*) cmsStageData(Matrix2); + IdentityMat = FALSE; + if (cmsPipelineCheckAndRetreiveStages(Src, 4, + cmsSigCurveSetElemType, cmsSigMatrixElemType, cmsSigMatrixElemType, cmsSigCurveSetElemType, + &Curve1, &Matrix1, &Matrix2, &Curve2)) { - // Input offset should be zero - if (Data1 ->Offset != NULL) return FALSE; + // Get both matrices + _cmsStageMatrixData* Data1 = (_cmsStageMatrixData*)cmsStageData(Matrix1); + _cmsStageMatrixData* Data2 = (_cmsStageMatrixData*)cmsStageData(Matrix2); - // Multiply both matrices to get the result - _cmsMAT3per(&res, (cmsMAT3*) Data2 ->Double, (cmsMAT3*) Data1 ->Double); + // Input offset should be zero + if (Data1->Offset != NULL) return FALSE; - // Now the result is in res + Data2 -> Offset. Maybe is a plain identity? - IdentityMat = FALSE; - if (_cmsMAT3isIdentity(&res) && Data2 ->Offset == NULL) { + // Multiply both matrices to get the result + _cmsMAT3per(&res, (cmsMAT3*)Data2->Double, (cmsMAT3*)Data1->Double); - // We can get rid of full matrix - IdentityMat = TRUE; - } + // Only 2nd matrix has offset, or it is zero + Offset = Data2->Offset; + + // Now the result is in res + Data2 -> Offset. Maybe is a plain identity? + if (_cmsMAT3isIdentity(&res) && Offset == NULL) { + + // We can get rid of full matrix + IdentityMat = TRUE; + } + + } + else { + + if (cmsPipelineCheckAndRetreiveStages(Src, 3, + cmsSigCurveSetElemType, cmsSigMatrixElemType, cmsSigCurveSetElemType, + &Curve1, &Matrix1, &Curve2)) { + + _cmsStageMatrixData* Data = (_cmsStageMatrixData*)cmsStageData(Matrix1); + + // Copy the matrix to our result + memcpy(&res, Data->Double, sizeof(res)); + + // Preserve the Odffset (may be NULL as a zero offset) + Offset = Data->Offset; + + if (_cmsMAT3isIdentity(&res) && Offset == NULL) { + + // We can get rid of full matrix + IdentityMat = TRUE; + } + } + else + return FALSE; // Not optimizeable this time + + } // Allocate an empty LUT Dest = cmsPipelineAlloc(Src ->ContextID, Src ->InputChannels, Src ->OutputChannels); @@ -1601,9 +1738,12 @@ cmsBool OptimizeMatrixShaper(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt3 if (!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, cmsStageDup(Curve1))) goto Error; - if (!IdentityMat) - if (!cmsPipelineInsertStage(Dest, cmsAT_END, cmsStageAllocMatrix(Dest ->ContextID, 3, 3, (const cmsFloat64Number*) &res, Data2 ->Offset))) - goto Error; + if (!IdentityMat) { + + if (!cmsPipelineInsertStage(Dest, cmsAT_END, cmsStageAllocMatrix(Dest->ContextID, 3, 3, (const cmsFloat64Number*)&res, Offset))) + goto Error; + } + if (!cmsPipelineInsertStage(Dest, cmsAT_END, cmsStageDup(Curve2))) goto Error; @@ -1616,12 +1756,12 @@ cmsBool OptimizeMatrixShaper(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt3 _cmsStageToneCurvesData* mpeC1 = (_cmsStageToneCurvesData*) cmsStageData(Curve1); _cmsStageToneCurvesData* mpeC2 = (_cmsStageToneCurvesData*) cmsStageData(Curve2); - // In this particular optimization, cach?does not help as it takes more time to deal with - // the cach?that with the pixel handling + // In this particular optimization, cache does not help as it takes more time to deal with + // the cache that with the pixel handling *dwFlags |= cmsFLAGS_NOCACHE; // Setup the optimizarion routines - if (!SetMatShaper(Dest, mpeC1 ->TheCurves, &res, (cmsVEC3*) Data2 ->Offset, mpeC2->TheCurves, OutputFormat)) + if (!SetMatShaper(Dest, mpeC1 ->TheCurves, &res, (cmsVEC3*) Offset, mpeC2->TheCurves, OutputFormat)) goto Error; } @@ -1809,3 +1949,6 @@ cmsBool _cmsOptimizePipeline(cmsContext ContextID, // Only simple optimizations succeeded return AnySuccess; } + + + |