?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Check for SYS/GL during library init. Reason is that
[AROS.git] / workbench / libs / lcms2 / src / cmsopt.c
blob6d8c9e7aef5001aa7db16d98102738970cdeddb0
1
2 //---------------------------------------------------------------------------------
3 //
4 // Little Color Management System
5 // Copyright (c) 1998-2011 Marti Maria Saguer
6 //
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:
13 //
14 // The above copyright notice and this permission notice shall be included in
15 // all copies or substantial portions of the Software.
16 //
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.
24 //
25 //---------------------------------------------------------------------------------
26 //
27
28 #include "lcms2_internal.h"
29
30
31 //----------------------------------------------------------------------------------
32
33 // Optimization for 8 bits, Shaper-CLUT (3 inputs only)
34 typedef struct {
35
36 cmsContext ContextID;
37
38 const cmsInterpParams* p; // Tetrahedrical interpolation parameters. This is a not-owned pointer.
39
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
42
43
44 } Prelin8Data;
45
46
47 // Generic optimization for 16 bits Shaper-CLUT-Shaper (any inputs)
48 typedef struct {
49
50 cmsContext ContextID;
51
52 // Number of channels
53 int nInputs;
54 int nOutputs;
55
56 _cmsInterpFn16 EvalCurveIn16[MAX_INPUT_DIMENSIONS]; // The maximum number of input channels is known in advance
57 cmsInterpParams* ParamsCurveIn16[MAX_INPUT_DIMENSIONS];
58
59 _cmsInterpFn16 EvalCLUT; // The evaluator for 3D grid
60 const cmsInterpParams* CLUTparams; // (not-owned pointer)
61
62
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)
65
66
67 } Prelin16Data;
68
69
70 // Optimization for matrix-shaper in 8 bits. Numbers are operated in n.14 signed, tables are stored in 1.14 fixed
71
72 typedef cmsInt32Number cmsS1Fixed14Number; // Note that this may hold more than 16 bits!
73
74 #define DOUBLE_TO_1FIXED14(x) ((cmsS1Fixed14Number) floor((x) * 16384.0 + 0.5))
75
76 typedef struct {
77
78 cmsContext ContextID;
79
80 cmsS1Fixed14Number Shaper1R[256]; // from 0..255 to 1.14 (0.0...1.0)
81 cmsS1Fixed14Number Shaper1G[256];
82 cmsS1Fixed14Number Shaper1B[256];
83
84 cmsS1Fixed14Number Mat[3][3]; // n.14 to n.14 (needs a saturation after that)
85 cmsS1Fixed14Number Off[3];
86
87 cmsUInt16Number Shaper2R[16385]; // 1.14 to 0..255
88 cmsUInt16Number Shaper2G[16385];
89 cmsUInt16Number Shaper2B[16385];
90
91 } MatShaper8Data;
92
93 // Curves, optimization is shared between 8 and 16 bits
94 typedef struct {
95
96 cmsContext ContextID;
97
98 int nCurves; // Number of curves
99 int nElements; // Elements in curves
100 cmsUInt16Number** Curves; // Points to a dynamically allocated array
101
102 } Curves16Data;
103
104
105 // Simple optimizations ----------------------------------------------------------------------------------------------------------
106
107
108 // Remove an element in linked chain
109 static
110 void _RemoveElement(cmsStage** head)
111 {
112 cmsStage* mpe = *head;
113 cmsStage* next = mpe ->Next;
114 *head = next;
115 cmsStageFree(mpe);
116 }
117
118 // Remove all identities in chain. Note that pt actually is a double pointer to the element that holds the pointer.
119 static
120 cmsBool _Remove1Op(cmsPipeline* Lut, cmsStageSignature UnaryOp)
121 {
122 cmsStage** pt = &Lut ->Elements;
123 cmsBool AnyOpt = FALSE;
124
125 while (*pt != NULL) {
126
127 if ((*pt) ->Implements == UnaryOp) {
128 _RemoveElement(pt);
129 AnyOpt = TRUE;
130 }
131 else
132 pt = &((*pt) -> Next);
133 }
134
135 return AnyOpt;
136 }
137
138 // Same, but only if two adjacent elements are found
139 static
140 cmsBool _Remove2Op(cmsPipeline* Lut, cmsStageSignature Op1, cmsStageSignature Op2)
141 {
142 cmsStage** pt1;
143 cmsStage** pt2;
144 cmsBool AnyOpt = FALSE;
145
146 pt1 = &Lut ->Elements;
147 if (*pt1 == NULL) return AnyOpt;
148
149 while (*pt1 != NULL) {
150
151 pt2 = &((*pt1) -> Next);
152 if (*pt2 == NULL) return AnyOpt;
153
154 if ((*pt1) ->Implements == Op1 && (*pt2) ->Implements == Op2) {
155 _RemoveElement(pt2);
156 _RemoveElement(pt1);
157 AnyOpt = TRUE;
158 }
159 else
160 pt1 = &((*pt1) -> Next);
161 }
162
163 return AnyOpt;
164 }
165
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.
168 static
169 cmsBool PreOptimize(cmsPipeline* Lut)
170 {
171 cmsBool AnyOpt = FALSE, Opt;
172
173 do {
174
175 Opt = FALSE;
176
177 // Remove all identities
178 Opt |= _Remove1Op(Lut, cmsSigIdentityElemType);
179
180 // Remove XYZ2Lab followed by Lab2XYZ
181 Opt |= _Remove2Op(Lut, cmsSigXYZ2LabElemType, cmsSigLab2XYZElemType);
182
183 // Remove Lab2XYZ followed by XYZ2Lab
184 Opt |= _Remove2Op(Lut, cmsSigLab2XYZElemType, cmsSigXYZ2LabElemType);
185
186 // Remove V4 to V2 followed by V2 to V4
187 Opt |= _Remove2Op(Lut, cmsSigLabV4toV2, cmsSigLabV2toV4);
188
189 // Remove V2 to V4 followed by V4 to V2
190 Opt |= _Remove2Op(Lut, cmsSigLabV2toV4, cmsSigLabV4toV2);
191
192 // Remove float pcs Lab conversions
193 Opt |= _Remove2Op(Lut, cmsSigLab2FloatPCS, cmsSigFloatPCS2Lab);
194
195 // Remove float pcs Lab conversions
196 Opt |= _Remove2Op(Lut, cmsSigXYZ2FloatPCS, cmsSigFloatPCS2XYZ);
197
198 if (Opt) AnyOpt = TRUE;
199
200 } while (Opt);
201
202 return AnyOpt;
203 }
204
205 static
206 void Eval16nop1D(register const cmsUInt16Number Input[],
207 register cmsUInt16Number Output[],
208 register const struct _cms_interp_struc* p)
209 {
210 Output[0] = Input[0];
211
212 cmsUNUSED_PARAMETER(p);
213 }
214
215 static
216 void PrelinEval16(register const cmsUInt16Number Input[],
217 register cmsUInt16Number Output[],
218 register const void* D)
219 {
220 Prelin16Data* p16 = (Prelin16Data*) D;
221 cmsUInt16Number StageABC[MAX_INPUT_DIMENSIONS];
222 cmsUInt16Number StageDEF[cmsMAXCHANNELS];
223 int i;
224
225 for (i=0; i < p16 ->nInputs; i++) {
226
227 p16 ->EvalCurveIn16[i](&Input[i], &StageABC[i], p16 ->ParamsCurveIn16[i]);
228 }
229
230 p16 ->EvalCLUT(StageABC, StageDEF, p16 ->CLUTparams);
231
232 for (i=0; i < p16 ->nOutputs; i++) {
233
234 p16 ->EvalCurveOut16[i](&StageDEF[i], &Output[i], p16 ->ParamsCurveOut16[i]);
235 }
236 }
237
238
239 static
240 void PrelinOpt16free(cmsContext ContextID, void* ptr)
241 {
242 Prelin16Data* p16 = (Prelin16Data*) ptr;
243
244 _cmsFree(ContextID, p16 ->EvalCurveOut16);
245 _cmsFree(ContextID, p16 ->ParamsCurveOut16);
246
247 _cmsFree(ContextID, p16);
248 }
249
250 static
251 void* Prelin16dup(cmsContext ContextID, const void* ptr)
252 {
253 Prelin16Data* p16 = (Prelin16Data*) ptr;
254 Prelin16Data* Duped = _cmsDupMem(ContextID, p16, sizeof(Prelin16Data));
255
256 if (Duped == NULL) return NULL;
257
258 Duped ->EvalCurveOut16 = _cmsDupMem(ContextID, p16 ->EvalCurveOut16, p16 ->nOutputs * sizeof(_cmsInterpFn16));
259 Duped ->ParamsCurveOut16 = _cmsDupMem(ContextID, p16 ->ParamsCurveOut16, p16 ->nOutputs * sizeof(cmsInterpParams* ));
260
261 return Duped;
262 }
263
264
265 static
266 Prelin16Data* PrelinOpt16alloc(cmsContext ContextID,
267 const cmsInterpParams* ColorMap,
268 int nInputs, cmsToneCurve** In,
269 int nOutputs, cmsToneCurve** Out )
270 {
271 int i;
272 Prelin16Data* p16 = _cmsMallocZero(ContextID, sizeof(Prelin16Data));
273 if (p16 == NULL) return NULL;
274
275 p16 ->nInputs = nInputs;
276 p16 -> nOutputs = nOutputs;
277
278
279 for (i=0; i < nInputs; i++) {
280
281 if (In == NULL) {
282 p16 -> ParamsCurveIn16[i] = NULL;
283 p16 -> EvalCurveIn16[i] = Eval16nop1D;
284
285 }
286 else {
287 p16 -> ParamsCurveIn16[i] = In[i] ->InterpParams;
288 p16 -> EvalCurveIn16[i] = p16 ->ParamsCurveIn16[i]->Interpolation.Lerp16;
289 }
290 }
291
292 p16 ->CLUTparams = ColorMap;
293 p16 ->EvalCLUT = ColorMap ->Interpolation.Lerp16;
294
295
296 p16 -> EvalCurveOut16 = (_cmsInterpFn16*) _cmsCalloc(ContextID, nOutputs, sizeof(_cmsInterpFn16));
297 p16 -> ParamsCurveOut16 = (cmsInterpParams**) _cmsCalloc(ContextID, nOutputs, sizeof(cmsInterpParams* ));
298
299 for (i=0; i < nOutputs; i++) {
300
301 if (Out == NULL) {
302 p16 ->ParamsCurveOut16[i] = NULL;
303 p16 -> EvalCurveOut16[i] = Eval16nop1D;
304 }
305 else {
306
307 p16 ->ParamsCurveOut16[i] = Out[i] ->InterpParams;
308 p16 -> EvalCurveOut16[i] = p16 ->ParamsCurveOut16[i]->Interpolation.Lerp16;
309 }
310 }
311
312 return p16;
313 }
314
315
316
317 // Resampling ---------------------------------------------------------------------------------
318
319 #define PRELINEARIZATION_POINTS 4096
320
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.
323 static
324 int XFormSampler16(register const cmsUInt16Number In[], register cmsUInt16Number Out[], register void* Cargo)
325 {
326 cmsPipeline* Lut = (cmsPipeline*) Cargo;
327 cmsFloat32Number InFloat[cmsMAXCHANNELS], OutFloat[cmsMAXCHANNELS];
328 cmsUInt32Number i;
329
330 _cmsAssert(Lut -> InputChannels < cmsMAXCHANNELS);
331 _cmsAssert(Lut -> OutputChannels < cmsMAXCHANNELS);
332
333 // From 16 bit to floating point
334 for (i=0; i < Lut ->InputChannels; i++)
335 InFloat[i] = (cmsFloat32Number) (In[i] / 65535.0);
336
337 // Evaluate in floating point
338 cmsPipelineEvalFloat(InFloat, OutFloat, Lut);
339
340 // Back to 16 bits representation
341 for (i=0; i < Lut ->OutputChannels; i++)
342 Out[i] = _cmsQuickSaturateWord(OutFloat[i] * 65535.0);
343
344 // Always succeed
345 return TRUE;
346 }
347
348 // Try to see if the curves of a given MPE are linear
349 static
350 cmsBool AllCurvesAreLinear(cmsStage* mpe)
351 {
352 cmsToneCurve** Curves;
353 cmsUInt32Number i, n;
354
355 Curves = _cmsStageGetPtrToCurveSet(mpe);
356 if (Curves == NULL) return FALSE;
357
358 n = cmsStageOutputChannels(mpe);
359
360 for (i=0; i < n; i++) {
361 if (!cmsIsToneCurveLinear(Curves[i])) return FALSE;
362 }
363
364 return TRUE;
365 }
366
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
369 static
370 cmsBool PatchLUT(cmsStage* CLUT, cmsUInt16Number At[], cmsUInt16Number Value[],
371 int nChannelsOut, int nChannelsIn)
372 {
373 _cmsStageCLutData* Grid = (_cmsStageCLutData*) CLUT ->Data;
374 cmsInterpParams* p16 = Grid ->Params;
375 cmsFloat64Number px, py, pz, pw;
376 int x0, y0, z0, w0;
377 int i, index;
378
379 if (CLUT -> Type != cmsSigCLutElemType) {
380 cmsSignalError(CLUT->ContextID, cmsERROR_INTERNAL, "(internal) Attempt to PatchLUT on non-lut stage");
381 return FALSE;
382 }
383
384 if (nChannelsIn == 4) {
385
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;
390
391 x0 = (int) floor(px);
392 y0 = (int) floor(py);
393 z0 = (int) floor(pz);
394 w0 = (int) floor(pw);
395
396 if (((px - x0) != 0) ||
397 ((py - y0) != 0) ||
398 ((pz - z0) != 0) ||
399 ((pw - w0) != 0)) return FALSE; // Not on exact node
400
401 index = p16 -> opta[3] * x0 +
402 p16 -> opta[2] * y0 +
403 p16 -> opta[1] * z0 +
404 p16 -> opta[0] * w0;
405 }
406 else
407 if (nChannelsIn == 3) {
408
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;
412
413 x0 = (int) floor(px);
414 y0 = (int) floor(py);
415 z0 = (int) floor(pz);
416
417 if (((px - x0) != 0) ||
418 ((py - y0) != 0) ||
419 ((pz - z0) != 0)) return FALSE; // Not on exact node
420
421 index = p16 -> opta[2] * x0 +
422 p16 -> opta[1] * y0 +
423 p16 -> opta[0] * z0;
424 }
425 else
426 if (nChannelsIn == 1) {
427
428 px = ((cmsFloat64Number) At[0] * (p16->Domain[0])) / 65535.0;
429
430 x0 = (int) floor(px);
431
432 if (((px - x0) != 0)) return FALSE; // Not on exact node
433
434 index = p16 -> opta[0] * x0;
435 }
436 else {
437 cmsSignalError(CLUT->ContextID, cmsERROR_INTERNAL, "(internal) %d Channels are not supported on PatchLUT", nChannelsIn);
438 return FALSE;
439 }
440
441 for (i=0; i < nChannelsOut; i++)
442 Grid -> Tab.T[index + i] = Value[i];
443
444 return TRUE;
445 }
446
447 // Auxiliar, to see if two values are equal or very different
448 static
449 cmsBool WhitesAreEqual(int n, cmsUInt16Number White1[], cmsUInt16Number White2[] )
450 {
451 int i;
452
453 for (i=0; i < n; i++) {
454
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;
457 }
458 return TRUE;
459 }
460
461
462 // Locate the node for the white point and fix it to pure white in order to avoid scum dot.
463 static
464 cmsBool FixWhiteMisalignment(cmsPipeline* Lut, cmsColorSpaceSignature EntryColorSpace, cmsColorSpaceSignature ExitColorSpace)
465 {
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;
470
471 if (!_cmsEndPointsBySpace(EntryColorSpace,
472 &WhitePointIn, NULL, &nIns)) return FALSE;
473
474 if (!_cmsEndPointsBySpace(ExitColorSpace,
475 &WhitePointOut, NULL, &nOuts)) return FALSE;
476
477 // It needs to be fixed?
478 if (Lut ->InputChannels != nIns) return FALSE;
479 if (Lut ->OutputChannels != nOuts) return FALSE;
480
481 cmsPipelineEval16(WhitePointIn, ObtainedOut, Lut);
482
483 if (WhitesAreEqual(nOuts, WhitePointOut, ObtainedOut)) return TRUE; // whites already match
484
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))
490 return FALSE;
491
492 // We need to interpolate white points of both, pre and post curves
493 if (PreLin) {
494
495 cmsToneCurve** Curves = _cmsStageGetPtrToCurveSet(PreLin);
496
497 for (i=0; i < nIns; i++) {
498 WhiteIn[i] = cmsEvalToneCurve16(Curves[i], WhitePointIn[i]);
499 }
500 }
501 else {
502 for (i=0; i < nIns; i++)
503 WhiteIn[i] = WhitePointIn[i];
504 }
505
506 // If any post-linearization, we need to find how is represented white before the curve, do
507 // a reverse interpolation in this case.
508 if (PostLin) {
509
510 cmsToneCurve** Curves = _cmsStageGetPtrToCurveSet(PostLin);
511
512 for (i=0; i < nOuts; i++) {
513
514 cmsToneCurve* InversePostLin = cmsReverseToneCurve(Curves[i]);
515 WhiteOut[i] = cmsEvalToneCurve16(InversePostLin, WhitePointOut[i]);
516 cmsFreeToneCurve(InversePostLin);
517 }
518 }
519 else {
520 for (i=0; i < nOuts; i++)
521 WhiteOut[i] = WhitePointOut[i];
522 }
523
524 // Ok, proceed with patching. May fail and we don't care if it fails
525 PatchLUT(CLUT, WhiteIn, WhiteOut, nOuts, nIns);
526
527 return TRUE;
528 }
529
530 // -----------------------------------------------------------------------------------------------------------------------------------------------
531 // This function creates simple LUT from complex ones. The generated LUT has an optional set of
532 // prelinearization curves, a CLUT of nGridPoints and optional postlinearization tables.
533 // These curves have to exist in the original LUT in order to be used in the simplified output.
534 // Caller may also use the flags to allow this feature.
535 // LUTS with all curves will be simplified to a single curve. Parametric curves are lost.
536 // This function should be used on 16-bits LUTS only, as floating point losses precision when simplified
537 // -----------------------------------------------------------------------------------------------------------------------------------------------
538
539 static
540 cmsBool OptimizeByResampling(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags)
541 {
542 cmsPipeline* Src = NULL;
543 cmsPipeline* Dest = NULL;
544 cmsStage* mpe;
545 cmsStage* CLUT;
546 cmsStage *KeepPreLin = NULL, *KeepPostLin = NULL;
547 int nGridPoints;
548 cmsColorSpaceSignature ColorSpace, OutputColorSpace;
549 cmsStage *NewPreLin = NULL;
550 cmsStage *NewPostLin = NULL;
551 _cmsStageCLutData* DataCLUT;
552 cmsToneCurve** DataSetIn;
553 cmsToneCurve** DataSetOut;
554 Prelin16Data* p16;
555
556 // This is a loosy optimization! does not apply in floating-point cases
557 if (_cmsFormatterIsFloat(*InputFormat) || _cmsFormatterIsFloat(*OutputFormat)) return FALSE;
558
559 ColorSpace = _cmsICCcolorSpace(T_COLORSPACE(*InputFormat));
560 OutputColorSpace = _cmsICCcolorSpace(T_COLORSPACE(*OutputFormat));
561 nGridPoints = _cmsReasonableGridpointsByColorspace(ColorSpace, *dwFlags);
562
563 // For empty LUTs, 2 points are enough
564 if (cmsPipelineStageCount(*Lut) == 0)
565 nGridPoints = 2;
566
567 Src = *Lut;
568
569 // Named color pipelines cannot be optimized either
570 for (mpe = cmsPipelineGetPtrToFirstStage(Src);
571 mpe != NULL;
572 mpe = cmsStageNext(mpe)) {
573 if (cmsStageType(mpe) == cmsSigNamedColorElemType) return FALSE;
574 }
575
576 // Allocate an empty LUT
577 Dest = cmsPipelineAlloc(Src ->ContextID, Src ->InputChannels, Src ->OutputChannels);
578 if (!Dest) return FALSE;
579
580 // Prelinearization tables are kept unless indicated by flags
581 if (*dwFlags & cmsFLAGS_CLUT_PRE_LINEARIZATION) {
582
583 // Get a pointer to the prelinearization element
584 cmsStage* PreLin = cmsPipelineGetPtrToFirstStage(Src);
585
586 // Check if suitable
587 if (PreLin ->Type == cmsSigCurveSetElemType) {
588
589 // Maybe this is a linear tram, so we can avoid the whole stuff
590 if (!AllCurvesAreLinear(PreLin)) {
591
592 // All seems ok, proceed.
593 NewPreLin = cmsStageDup(PreLin);
594 if(!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, NewPreLin))
595 goto Error;
596
597 // Remove prelinearization. Since we have duplicated the curve
598 // in destination LUT, the sampling shoud be applied after this stage.
599 cmsPipelineUnlinkStage(Src, cmsAT_BEGIN, &KeepPreLin);
600 }
601 }
602 }
603
604 // Allocate the CLUT
605 CLUT = cmsStageAllocCLut16bit(Src ->ContextID, nGridPoints, Src ->InputChannels, Src->OutputChannels, NULL);
606 if (CLUT == NULL) return FALSE;
607
608 // Add the CLUT to the destination LUT
609 if (!cmsPipelineInsertStage(Dest, cmsAT_END, CLUT)) {
610 goto Error;
611 }
612
613 // Postlinearization tables are kept unless indicated by flags
614 if (*dwFlags & cmsFLAGS_CLUT_POST_LINEARIZATION) {
615
616 // Get a pointer to the postlinearization if present
617 cmsStage* PostLin = cmsPipelineGetPtrToLastStage(Src);
618
619 // Check if suitable
620 if (cmsStageType(PostLin) == cmsSigCurveSetElemType) {
621
622 // Maybe this is a linear tram, so we can avoid the whole stuff
623 if (!AllCurvesAreLinear(PostLin)) {
624
625 // All seems ok, proceed.
626 NewPostLin = cmsStageDup(PostLin);
627 if (!cmsPipelineInsertStage(Dest, cmsAT_END, NewPostLin))
628 goto Error;
629
630 // In destination LUT, the sampling shoud be applied after this stage.
631 cmsPipelineUnlinkStage(Src, cmsAT_END, &KeepPostLin);
632 }
633 }
634 }
635
636 // Now its time to do the sampling. We have to ignore pre/post linearization
637 // The source LUT whithout pre/post curves is passed as parameter.
638 if (!cmsStageSampleCLut16bit(CLUT, XFormSampler16, (void*) Src, 0)) {
639 Error:
640 // Ops, something went wrong, Restore stages
641 if (KeepPreLin != NULL) {
642 if (!cmsPipelineInsertStage(Src, cmsAT_BEGIN, KeepPreLin)) {
643 _cmsAssert(0); // This never happens
644 }
645 }
646 if (KeepPostLin != NULL) {
647 if (!cmsPipelineInsertStage(Src, cmsAT_END, KeepPostLin)) {
648 _cmsAssert(0); // This never happens
649 }
650 }
651 cmsPipelineFree(Dest);
652 return FALSE;
653 }
654
655 // Done.
656
657 if (KeepPreLin != NULL) cmsStageFree(KeepPreLin);
658 if (KeepPostLin != NULL) cmsStageFree(KeepPostLin);
659 cmsPipelineFree(Src);
660
661 DataCLUT = (_cmsStageCLutData*) CLUT ->Data;
662
663 if (NewPreLin == NULL) DataSetIn = NULL;
664 else DataSetIn = ((_cmsStageToneCurvesData*) NewPreLin ->Data) ->TheCurves;
665
666 if (NewPostLin == NULL) DataSetOut = NULL;
667 else DataSetOut = ((_cmsStageToneCurvesData*) NewPostLin ->Data) ->TheCurves;
668
669
670 if (DataSetIn == NULL && DataSetOut == NULL) {
671
672 _cmsPipelineSetOptimizationParameters(Dest, (_cmsOPTeval16Fn) DataCLUT->Params->Interpolation.Lerp16, DataCLUT->Params, NULL, NULL);
673 }
674 else {
675
676 p16 = PrelinOpt16alloc(Dest ->ContextID,
677 DataCLUT ->Params,
678 Dest ->InputChannels,
679 DataSetIn,
680 Dest ->OutputChannels,
681 DataSetOut);
682
683 _cmsPipelineSetOptimizationParameters(Dest, PrelinEval16, (void*) p16, PrelinOpt16free, Prelin16dup);
684 }
685
686
687 // Don't fix white on absolute colorimetric
688 if (Intent == INTENT_ABSOLUTE_COLORIMETRIC)
689 *dwFlags |= cmsFLAGS_NOWHITEONWHITEFIXUP;
690
691 if (!(*dwFlags & cmsFLAGS_NOWHITEONWHITEFIXUP)) {
692
693 FixWhiteMisalignment(Dest, ColorSpace, OutputColorSpace);
694 }
695
696 *Lut = Dest;
697 return TRUE;
698
699 cmsUNUSED_PARAMETER(Intent);
700 }
701
702
703 // -----------------------------------------------------------------------------------------------------------------------------------------------
704 // Fixes the gamma balancing of transform. This is described in my paper "Prelinearization Stages on
705 // Color-Management Application-Specific Integrated Circuits (ASICs)" presented at NIP24. It only works
706 // for RGB transforms. See the paper for more details
707 // -----------------------------------------------------------------------------------------------------------------------------------------------
708
709
710 // Normalize endpoints by slope limiting max and min. This assures endpoints as well.
711 // Descending curves are handled as well.
712 static
713 void SlopeLimiting(cmsToneCurve* g)
714 {
715 int BeginVal, EndVal;
716 int AtBegin = (int) floor((cmsFloat64Number) g ->nEntries * 0.02 + 0.5); // Cutoff at 2%
717 int AtEnd = g ->nEntries - AtBegin - 1; // And 98%
718 cmsFloat64Number Val, Slope, beta;
719 int i;
720
721 if (cmsIsToneCurveDescending(g)) {
722 BeginVal = 0xffff; EndVal = 0;
723 }
724 else {
725 BeginVal = 0; EndVal = 0xffff;
726 }
727
728 // Compute slope and offset for begin of curve
729 Val = g ->Table16[AtBegin];
730 Slope = (Val - BeginVal) / AtBegin;
731 beta = Val - Slope * AtBegin;
732
733 for (i=0; i < AtBegin; i++)
734 g ->Table16[i] = _cmsQuickSaturateWord(i * Slope + beta);
735
736 // Compute slope and offset for the end
737 Val = g ->Table16[AtEnd];
738 Slope = (EndVal - Val) / AtBegin; // AtBegin holds the X interval, which is same in both cases
739 beta = Val - Slope * AtEnd;
740
741 for (i = AtEnd; i < (int) g ->nEntries; i++)
742 g ->Table16[i] = _cmsQuickSaturateWord(i * Slope + beta);
743 }
744
745
746 // Precomputes tables for 8-bit on input devicelink.
747 static
748 Prelin8Data* PrelinOpt8alloc(cmsContext ContextID, const cmsInterpParams* p, cmsToneCurve* G[3])
749 {
750 int i;
751 cmsUInt16Number Input[3];
752 cmsS15Fixed16Number v1, v2, v3;
753 Prelin8Data* p8;
754
755 p8 = _cmsMallocZero(ContextID, sizeof(Prelin8Data));
756 if (p8 == NULL) return NULL;
757
758 // Since this only works for 8 bit input, values comes always as x * 257,
759 // we can safely take msb byte (x << 8 + x)
760
761 for (i=0; i < 256; i++) {
762
763 if (G != NULL) {
764
765 // Get 16-bit representation
766 Input[0] = cmsEvalToneCurve16(G[0], FROM_8_TO_16(i));
767 Input[1] = cmsEvalToneCurve16(G[1], FROM_8_TO_16(i));
768 Input[2] = cmsEvalToneCurve16(G[2], FROM_8_TO_16(i));
769 }
770 else {
771 Input[0] = FROM_8_TO_16(i);
772 Input[1] = FROM_8_TO_16(i);
773 Input[2] = FROM_8_TO_16(i);
774 }
775
776
777 // Move to 0..1.0 in fixed domain
778 v1 = _cmsToFixedDomain(Input[0] * p -> Domain[0]);
779 v2 = _cmsToFixedDomain(Input[1] * p -> Domain[1]);
780 v3 = _cmsToFixedDomain(Input[2] * p -> Domain[2]);
781
782 // Store the precalculated table of nodes
783 p8 ->X0[i] = (p->opta[2] * FIXED_TO_INT(v1));
784 p8 ->Y0[i] = (p->opta[1] * FIXED_TO_INT(v2));
785 p8 ->Z0[i] = (p->opta[0] * FIXED_TO_INT(v3));
786
787 // Store the precalculated table of offsets
788 p8 ->rx[i] = (cmsUInt16Number) FIXED_REST_TO_INT(v1);
789 p8 ->ry[i] = (cmsUInt16Number) FIXED_REST_TO_INT(v2);
790 p8 ->rz[i] = (cmsUInt16Number) FIXED_REST_TO_INT(v3);
791 }
792
793 p8 ->ContextID = ContextID;
794 p8 ->p = p;
795
796 return p8;
797 }
798
799 static
800 void Prelin8free(cmsContext ContextID, void* ptr)
801 {
802 _cmsFree(ContextID, ptr);
803 }
804
805 static
806 void* Prelin8dup(cmsContext ContextID, const void* ptr)
807 {
808 return _cmsDupMem(ContextID, ptr, sizeof(Prelin8Data));
809 }
810
811
812
813 // A optimized interpolation for 8-bit input.
814 #define DENS(i,j,k) (LutTable[(i)+(j)+(k)+OutChan])
815 static
816 void PrelinEval8(register const cmsUInt16Number Input[],
817 register cmsUInt16Number Output[],
818 register const void* D)
819 {
820
821 cmsUInt8Number r, g, b;
822 cmsS15Fixed16Number rx, ry, rz;
823 cmsS15Fixed16Number c0, c1, c2, c3, Rest;
824 int OutChan;
825 register cmsS15Fixed16Number X0, X1, Y0, Y1, Z0, Z1;
826 Prelin8Data* p8 = (Prelin8Data*) D;
827 register const cmsInterpParams* p = p8 ->p;
828 int TotalOut = p -> nOutputs;
829 const cmsUInt16Number* LutTable = p -> Table;
830
831 r = Input[0] >> 8;
832 g = Input[1] >> 8;
833 b = Input[2] >> 8;
834
835 X0 = X1 = p8->X0[r];
836 Y0 = Y1 = p8->Y0[g];
837 Z0 = Z1 = p8->Z0[b];
838
839 rx = p8 ->rx[r];
840 ry = p8 ->ry[g];
841 rz = p8 ->rz[b];
842
843 X1 = X0 + ((rx == 0) ? 0 : p ->opta[2]);
844 Y1 = Y0 + ((ry == 0) ? 0 : p ->opta[1]);
845 Z1 = Z0 + ((rz == 0) ? 0 : p ->opta[0]);
846
847
848 // These are the 6 Tetrahedral
849 for (OutChan=0; OutChan < TotalOut; OutChan++) {
850
851 c0 = DENS(X0, Y0, Z0);
852
853 if (rx >= ry && ry >= rz)
854 {
855 c1 = DENS(X1, Y0, Z0) - c0;
856 c2 = DENS(X1, Y1, Z0) - DENS(X1, Y0, Z0);
857 c3 = DENS(X1, Y1, Z1) - DENS(X1, Y1, Z0);
858 }
859 else
860 if (rx >= rz && rz >= ry)
861 {
862 c1 = DENS(X1, Y0, Z0) - c0;
863 c2 = DENS(X1, Y1, Z1) - DENS(X1, Y0, Z1);
864 c3 = DENS(X1, Y0, Z1) - DENS(X1, Y0, Z0);
865 }
866 else
867 if (rz >= rx && rx >= ry)
868 {
869 c1 = DENS(X1, Y0, Z1) - DENS(X0, Y0, Z1);
870 c2 = DENS(X1, Y1, Z1) - DENS(X1, Y0, Z1);
871 c3 = DENS(X0, Y0, Z1) - c0;
872 }
873 else
874 if (ry >= rx && rx >= rz)
875 {
876 c1 = DENS(X1, Y1, Z0) - DENS(X0, Y1, Z0);
877 c2 = DENS(X0, Y1, Z0) - c0;
878 c3 = DENS(X1, Y1, Z1) - DENS(X1, Y1, Z0);
879 }
880 else
881 if (ry >= rz && rz >= rx)
882 {
883 c1 = DENS(X1, Y1, Z1) - DENS(X0, Y1, Z1);
884 c2 = DENS(X0, Y1, Z0) - c0;
885 c3 = DENS(X0, Y1, Z1) - DENS(X0, Y1, Z0);
886 }
887 else
888 if (rz >= ry && ry >= rx)
889 {
890 c1 = DENS(X1, Y1, Z1) - DENS(X0, Y1, Z1);
891 c2 = DENS(X0, Y1, Z1) - DENS(X0, Y0, Z1);
892 c3 = DENS(X0, Y0, Z1) - c0;
893 }
894 else {
895 c1 = c2 = c3 = 0;
896 }
897
898
899 Rest = c1 * rx + c2 * ry + c3 * rz + 0x8001;
900 Output[OutChan] = (cmsUInt16Number)c0 + ((Rest + (Rest>>16))>>16);
901
902 }
903 }
904
905 #undef DENS
906
907
908 // Curves that contain wide empty areas are not optimizeable
909 static
910 cmsBool IsDegenerated(const cmsToneCurve* g)
911 {
912 int i, Zeros = 0, Poles = 0;
913 int nEntries = g ->nEntries;
914
915 for (i=0; i < nEntries; i++) {
916
917 if (g ->Table16[i] == 0x0000) Zeros++;
918 if (g ->Table16[i] == 0xffff) Poles++;
919 }
920
921 if (Zeros == 1 && Poles == 1) return FALSE; // For linear tables
922 if (Zeros > (nEntries / 4)) return TRUE; // Degenerated, mostly zeros
923 if (Poles > (nEntries / 4)) return TRUE; // Degenerated, mostly poles
924
925 return FALSE;
926 }
927
928 // --------------------------------------------------------------------------------------------------------------
929 // We need xput over here
930
931 static
932 cmsBool OptimizeByComputingLinearization(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags)
933 {
934 cmsPipeline* OriginalLut;
935 int nGridPoints;
936 cmsToneCurve *Trans[cmsMAXCHANNELS], *TransReverse[cmsMAXCHANNELS];
937 cmsUInt32Number t, i;
938 cmsFloat32Number v, In[cmsMAXCHANNELS], Out[cmsMAXCHANNELS];
939 cmsBool lIsSuitable, lIsLinear;
940 cmsPipeline* OptimizedLUT = NULL, *LutPlusCurves = NULL;
941 cmsStage* OptimizedCLUTmpe;
942 cmsColorSpaceSignature ColorSpace, OutputColorSpace;
943 cmsStage* OptimizedPrelinMpe;
944 cmsStage* mpe;
945 cmsToneCurve** OptimizedPrelinCurves;
946 _cmsStageCLutData* OptimizedPrelinCLUT;
947
948
949 // This is a loosy optimization! does not apply in floating-point cases
950 if (_cmsFormatterIsFloat(*InputFormat) || _cmsFormatterIsFloat(*OutputFormat)) return FALSE;
951
952 // Only on RGB
953 if (T_COLORSPACE(*InputFormat) != PT_RGB) return FALSE;
954 if (T_COLORSPACE(*OutputFormat) != PT_RGB) return FALSE;
955
956
957 // On 16 bits, user has to specify the feature
958 if (!_cmsFormatterIs8bit(*InputFormat)) {
959 if (!(*dwFlags & cmsFLAGS_CLUT_PRE_LINEARIZATION)) return FALSE;
960 }
961
962 OriginalLut = *Lut;
963
964 // Named color pipelines cannot be optimized either
965 for (mpe = cmsPipelineGetPtrToFirstStage(OriginalLut);
966 mpe != NULL;
967 mpe = cmsStageNext(mpe)) {
968 if (cmsStageType(mpe) == cmsSigNamedColorElemType) return FALSE;
969 }
970
971 ColorSpace = _cmsICCcolorSpace(T_COLORSPACE(*InputFormat));
972 OutputColorSpace = _cmsICCcolorSpace(T_COLORSPACE(*OutputFormat));
973 nGridPoints = _cmsReasonableGridpointsByColorspace(ColorSpace, *dwFlags);
974
975 // Empty gamma containers
976 memset(Trans, 0, sizeof(Trans));
977 memset(TransReverse, 0, sizeof(TransReverse));
978
979 for (t = 0; t < OriginalLut ->InputChannels; t++) {
980 Trans[t] = cmsBuildTabulatedToneCurve16(OriginalLut ->ContextID, PRELINEARIZATION_POINTS, NULL);
981 if (Trans[t] == NULL) goto Error;
982 }
983
984 // Populate the curves
985 for (i=0; i < PRELINEARIZATION_POINTS; i++) {
986
987 v = (cmsFloat32Number) ((cmsFloat64Number) i / (PRELINEARIZATION_POINTS - 1));
988
989 // Feed input with a gray ramp
990 for (t=0; t < OriginalLut ->InputChannels; t++)
991 In[t] = v;
992
993 // Evaluate the gray value
994 cmsPipelineEvalFloat(In, Out, OriginalLut);
995
996 // Store result in curve
997 for (t=0; t < OriginalLut ->InputChannels; t++)
998 Trans[t] ->Table16[i] = _cmsQuickSaturateWord(Out[t] * 65535.0);
999 }
1000
1001 // Slope-limit the obtained curves
1002 for (t = 0; t < OriginalLut ->InputChannels; t++)
1003 SlopeLimiting(Trans[t]);
1004
1005 // Check for validity
1006 lIsSuitable = TRUE;
1007 lIsLinear = TRUE;
1008 for (t=0; (lIsSuitable && (t < OriginalLut ->InputChannels)); t++) {
1009
1010 // Exclude if already linear
1011 if (!cmsIsToneCurveLinear(Trans[t]))
1012 lIsLinear = FALSE;
1013
1014 // Exclude if non-monotonic
1015 if (!cmsIsToneCurveMonotonic(Trans[t]))
1016 lIsSuitable = FALSE;
1017
1018 if (IsDegenerated(Trans[t]))
1019 lIsSuitable = FALSE;
1020 }
1021
1022 // If it is not suitable, just quit
1023 if (!lIsSuitable) goto Error;
1024
1025 // Invert curves if possible
1026 for (t = 0; t < OriginalLut ->InputChannels; t++) {
1027 TransReverse[t] = cmsReverseToneCurveEx(PRELINEARIZATION_POINTS, Trans[t]);
1028 if (TransReverse[t] == NULL) goto Error;
1029 }
1030
1031 // Now inset the reversed curves at the begin of transform
1032 LutPlusCurves = cmsPipelineDup(OriginalLut);
1033 if (LutPlusCurves == NULL) goto Error;
1034
1035 if (!cmsPipelineInsertStage(LutPlusCurves, cmsAT_BEGIN, cmsStageAllocToneCurves(OriginalLut ->ContextID, OriginalLut ->InputChannels, TransReverse)))
1036 goto Error;
1037
1038 // Create the result LUT
1039 OptimizedLUT = cmsPipelineAlloc(OriginalLut ->ContextID, OriginalLut ->InputChannels, OriginalLut ->OutputChannels);
1040 if (OptimizedLUT == NULL) goto Error;
1041
1042 OptimizedPrelinMpe = cmsStageAllocToneCurves(OriginalLut ->ContextID, OriginalLut ->InputChannels, Trans);
1043
1044 // Create and insert the curves at the beginning
1045 if (!cmsPipelineInsertStage(OptimizedLUT, cmsAT_BEGIN, OptimizedPrelinMpe))
1046 goto Error;
1047
1048 // Allocate the CLUT for result
1049 OptimizedCLUTmpe = cmsStageAllocCLut16bit(OriginalLut ->ContextID, nGridPoints, OriginalLut ->InputChannels, OriginalLut ->OutputChannels, NULL);
1050
1051 // Add the CLUT to the destination LUT
1052 if (!cmsPipelineInsertStage(OptimizedLUT, cmsAT_END, OptimizedCLUTmpe))
1053 goto Error;
1054
1055 // Resample the LUT
1056 if (!cmsStageSampleCLut16bit(OptimizedCLUTmpe, XFormSampler16, (void*) LutPlusCurves, 0)) goto Error;
1057
1058 // Free resources
1059 for (t = 0; t < OriginalLut ->InputChannels; t++) {
1060
1061 if (Trans[t]) cmsFreeToneCurve(Trans[t]);
1062 if (TransReverse[t]) cmsFreeToneCurve(TransReverse[t]);
1063 }
1064
1065 cmsPipelineFree(LutPlusCurves);
1066
1067
1068 OptimizedPrelinCurves = _cmsStageGetPtrToCurveSet(OptimizedPrelinMpe);
1069 OptimizedPrelinCLUT = (_cmsStageCLutData*) OptimizedCLUTmpe ->Data;
1070
1071 // Set the evaluator if 8-bit
1072 if (_cmsFormatterIs8bit(*InputFormat)) {
1073
1074 Prelin8Data* p8 = PrelinOpt8alloc(OptimizedLUT ->ContextID,
1075 OptimizedPrelinCLUT ->Params,
1076 OptimizedPrelinCurves);
1077 if (p8 == NULL) return FALSE;
1078
1079 _cmsPipelineSetOptimizationParameters(OptimizedLUT, PrelinEval8, (void*) p8, Prelin8free, Prelin8dup);
1080
1081 }
1082 else
1083 {
1084 Prelin16Data* p16 = PrelinOpt16alloc(OptimizedLUT ->ContextID,
1085 OptimizedPrelinCLUT ->Params,
1086 3, OptimizedPrelinCurves, 3, NULL);
1087 if (p16 == NULL) return FALSE;
1088
1089 _cmsPipelineSetOptimizationParameters(OptimizedLUT, PrelinEval16, (void*) p16, PrelinOpt16free, Prelin16dup);
1090
1091 }
1092
1093 // Don't fix white on absolute colorimetric
1094 if (Intent == INTENT_ABSOLUTE_COLORIMETRIC)
1095 *dwFlags |= cmsFLAGS_NOWHITEONWHITEFIXUP;
1096
1097 if (!(*dwFlags & cmsFLAGS_NOWHITEONWHITEFIXUP)) {
1098
1099 if (!FixWhiteMisalignment(OptimizedLUT, ColorSpace, OutputColorSpace)) {
1100
1101 return FALSE;
1102 }
1103 }
1104
1105 // And return the obtained LUT
1106
1107 cmsPipelineFree(OriginalLut);
1108 *Lut = OptimizedLUT;
1109 return TRUE;
1110
1111 Error:
1112
1113 for (t = 0; t < OriginalLut ->InputChannels; t++) {
1114
1115 if (Trans[t]) cmsFreeToneCurve(Trans[t]);
1116 if (TransReverse[t]) cmsFreeToneCurve(TransReverse[t]);
1117 }
1118
1119 if (LutPlusCurves != NULL) cmsPipelineFree(LutPlusCurves);
1120 if (OptimizedLUT != NULL) cmsPipelineFree(OptimizedLUT);
1121
1122 return FALSE;
1123
1124 cmsUNUSED_PARAMETER(Intent);
1125 }
1126
1127
1128 // Curves optimizer ------------------------------------------------------------------------------------------------------------------
1129
1130 static
1131 void CurvesFree(cmsContext ContextID, void* ptr)
1132 {
1133 Curves16Data* Data = (Curves16Data*) ptr;
1134 int i;
1135
1136 for (i=0; i < Data -> nCurves; i++) {
1137
1138 _cmsFree(ContextID, Data ->Curves[i]);
1139 }
1140
1141 _cmsFree(ContextID, Data ->Curves);
1142 _cmsFree(ContextID, ptr);
1143 }
1144
1145 static
1146 void* CurvesDup(cmsContext ContextID, const void* ptr)
1147 {
1148 Curves16Data* Data = _cmsDupMem(ContextID, ptr, sizeof(Curves16Data));
1149 int i;
1150
1151 if (Data == NULL) return NULL;
1152
1153 Data ->Curves = _cmsDupMem(ContextID, Data ->Curves, Data ->nCurves * sizeof(cmsUInt16Number*));
1154
1155 for (i=0; i < Data -> nCurves; i++) {
1156 Data ->Curves[i] = _cmsDupMem(ContextID, Data ->Curves[i], Data -> nElements * sizeof(cmsUInt16Number));
1157 }
1158
1159 return (void*) Data;
1160 }
1161
1162 // Precomputes tables for 8-bit on input devicelink.
1163 static
1164 Curves16Data* CurvesAlloc(cmsContext ContextID, int nCurves, int nElements, cmsToneCurve** G)
1165 {
1166 int i, j;
1167 Curves16Data* c16;
1168
1169 c16 = _cmsMallocZero(ContextID, sizeof(Curves16Data));
1170 if (c16 == NULL) return NULL;
1171
1172 c16 ->nCurves = nCurves;
1173 c16 ->nElements = nElements;
1174
1175 c16 ->Curves = _cmsCalloc(ContextID, nCurves, sizeof(cmsUInt16Number*));
1176 if (c16 ->Curves == NULL) return NULL;
1177
1178 for (i=0; i < nCurves; i++) {
1179
1180 c16->Curves[i] = _cmsCalloc(ContextID, nElements, sizeof(cmsUInt16Number));
1181
1182 if (c16->Curves[i] == NULL) {
1183
1184 for (j=0; j < i; j++) {
1185 _cmsFree(ContextID, c16->Curves[j]);
1186 }
1187 _cmsFree(ContextID, c16->Curves);
1188 _cmsFree(ContextID, c16);
1189 return NULL;
1190 }
1191
1192 if (nElements == 256) {
1193
1194 for (j=0; j < nElements; j++) {
1195
1196 c16 ->Curves[i][j] = cmsEvalToneCurve16(G[i], FROM_8_TO_16(j));
1197 }
1198 }
1199 else {
1200
1201 for (j=0; j < nElements; j++) {
1202 c16 ->Curves[i][j] = cmsEvalToneCurve16(G[i], (cmsUInt16Number) j);
1203 }
1204 }
1205 }
1206
1207 return c16;
1208 }
1209
1210 static
1211 void FastEvaluateCurves8(register const cmsUInt16Number In[],
1212 register cmsUInt16Number Out[],
1213 register const void* D)
1214 {
1215 Curves16Data* Data = (Curves16Data*) D;
1216 cmsUInt8Number x;
1217 int i;
1218
1219 for (i=0; i < Data ->nCurves; i++) {
1220
1221 x = (In[i] >> 8);
1222 Out[i] = Data -> Curves[i][x];
1223 }
1224 }
1225
1226
1227 static
1228 void FastEvaluateCurves16(register const cmsUInt16Number In[],
1229 register cmsUInt16Number Out[],
1230 register const void* D)
1231 {
1232 Curves16Data* Data = (Curves16Data*) D;
1233 int i;
1234
1235 for (i=0; i < Data ->nCurves; i++) {
1236 Out[i] = Data -> Curves[i][In[i]];
1237 }
1238 }
1239
1240
1241 static
1242 void FastIdentity16(register const cmsUInt16Number In[],
1243 register cmsUInt16Number Out[],
1244 register const void* D)
1245 {
1246 cmsPipeline* Lut = (cmsPipeline*) D;
1247 cmsUInt32Number i;
1248
1249 for (i=0; i < Lut ->InputChannels; i++) {
1250 Out[i] = In[i];
1251 }
1252 }
1253
1254
1255 // If the target LUT holds only curves, the optimization procedure is to join all those
1256 // curves together. That only works on curves and does not work on matrices.
1257 static
1258 cmsBool OptimizeByJoiningCurves(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags)
1259 {
1260 cmsToneCurve** GammaTables = NULL;
1261 cmsFloat32Number InFloat[cmsMAXCHANNELS], OutFloat[cmsMAXCHANNELS];
1262 cmsUInt32Number i, j;
1263 cmsPipeline* Src = *Lut;
1264 cmsPipeline* Dest = NULL;
1265 cmsStage* mpe;
1266 cmsStage* ObtainedCurves = NULL;
1267
1268
1269 // This is a loosy optimization! does not apply in floating-point cases
1270 if (_cmsFormatterIsFloat(*InputFormat) || _cmsFormatterIsFloat(*OutputFormat)) return FALSE;
1271
1272 // Only curves in this LUT?
1273 for (mpe = cmsPipelineGetPtrToFirstStage(Src);
1274 mpe != NULL;
1275 mpe = cmsStageNext(mpe)) {
1276 if (cmsStageType(mpe) != cmsSigCurveSetElemType) return FALSE;
1277 }
1278
1279 // Allocate an empty LUT
1280 Dest = cmsPipelineAlloc(Src ->ContextID, Src ->InputChannels, Src ->OutputChannels);
1281 if (Dest == NULL) return FALSE;
1282
1283 // Create target curves
1284 GammaTables = (cmsToneCurve**) _cmsCalloc(Src ->ContextID, Src ->InputChannels, sizeof(cmsToneCurve*));
1285 if (GammaTables == NULL) goto Error;
1286
1287 for (i=0; i < Src ->InputChannels; i++) {
1288 GammaTables[i] = cmsBuildTabulatedToneCurve16(Src ->ContextID, PRELINEARIZATION_POINTS, NULL);
1289 if (GammaTables[i] == NULL) goto Error;
1290 }
1291
1292 // Compute 16 bit result by using floating point
1293 for (i=0; i < PRELINEARIZATION_POINTS; i++) {
1294
1295 for (j=0; j < Src ->InputChannels; j++)
1296 InFloat[j] = (cmsFloat32Number) ((cmsFloat64Number) i / (PRELINEARIZATION_POINTS - 1));
1297
1298 cmsPipelineEvalFloat(InFloat, OutFloat, Src);
1299
1300 for (j=0; j < Src ->InputChannels; j++)
1301 GammaTables[j] -> Table16[i] = _cmsQuickSaturateWord(OutFloat[j] * 65535.0);
1302 }
1303
1304 ObtainedCurves = cmsStageAllocToneCurves(Src ->ContextID, Src ->InputChannels, GammaTables);
1305 if (ObtainedCurves == NULL) goto Error;
1306
1307 for (i=0; i < Src ->InputChannels; i++) {
1308 cmsFreeToneCurve(GammaTables[i]);
1309 GammaTables[i] = NULL;
1310 }
1311
1312 if (GammaTables != NULL) _cmsFree(Src ->ContextID, GammaTables);
1313
1314 // Maybe the curves are linear at the end
1315 if (!AllCurvesAreLinear(ObtainedCurves)) {
1316
1317 if (!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, ObtainedCurves))
1318 goto Error;
1319
1320 // If the curves are to be applied in 8 bits, we can save memory
1321 if (_cmsFormatterIs8bit(*InputFormat)) {
1322
1323 _cmsStageToneCurvesData* Data = (_cmsStageToneCurvesData*) ObtainedCurves ->Data;
1324 Curves16Data* c16 = CurvesAlloc(Dest ->ContextID, Data ->nCurves, 256, Data ->TheCurves);
1325
1326 if (c16 == NULL) goto Error;
1327 *dwFlags |= cmsFLAGS_NOCACHE;
1328 _cmsPipelineSetOptimizationParameters(Dest, FastEvaluateCurves8, c16, CurvesFree, CurvesDup);
1329
1330 }
1331 else {
1332
1333 _cmsStageToneCurvesData* Data = (_cmsStageToneCurvesData*) cmsStageData(ObtainedCurves);
1334 Curves16Data* c16 = CurvesAlloc(Dest ->ContextID, Data ->nCurves, 65536, Data ->TheCurves);
1335
1336 if (c16 == NULL) goto Error;
1337 *dwFlags |= cmsFLAGS_NOCACHE;
1338 _cmsPipelineSetOptimizationParameters(Dest, FastEvaluateCurves16, c16, CurvesFree, CurvesDup);
1339 }
1340 }
1341 else {
1342
1343 // LUT optimizes to nothing. Set the identity LUT
1344 cmsStageFree(ObtainedCurves);
1345
1346 if (!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, cmsStageAllocIdentity(Dest ->ContextID, Src ->InputChannels)))
1347 goto Error;
1348
1349 *dwFlags |= cmsFLAGS_NOCACHE;
1350 _cmsPipelineSetOptimizationParameters(Dest, FastIdentity16, (void*) Dest, NULL, NULL);
1351 }
1352
1353 // We are done.
1354 cmsPipelineFree(Src);
1355 *Lut = Dest;
1356 return TRUE;
1357
1358 Error:
1359
1360 if (ObtainedCurves != NULL) cmsStageFree(ObtainedCurves);
1361 if (GammaTables != NULL) {
1362 for (i=0; i < Src ->InputChannels; i++) {
1363 if (GammaTables[i] != NULL) cmsFreeToneCurve(GammaTables[i]);
1364 }
1365
1366 _cmsFree(Src ->ContextID, GammaTables);
1367 }
1368
1369 if (Dest != NULL) cmsPipelineFree(Dest);
1370 return FALSE;
1371
1372 cmsUNUSED_PARAMETER(Intent);
1373 cmsUNUSED_PARAMETER(InputFormat);
1374 cmsUNUSED_PARAMETER(OutputFormat);
1375 cmsUNUSED_PARAMETER(dwFlags);
1376 }
1377
1378 // -------------------------------------------------------------------------------------------------------------------------------------
1379 // LUT is Shaper - Matrix - Matrix - Shaper, which is very frequent when combining two matrix-shaper profiles
1380
1381
1382 static
1383 void FreeMatShaper(cmsContext ContextID, void* Data)
1384 {
1385 if (Data != NULL) _cmsFree(ContextID, Data);
1386 }
1387
1388 static
1389 void* DupMatShaper(cmsContext ContextID, const void* Data)
1390 {
1391 return _cmsDupMem(ContextID, Data, sizeof(MatShaper8Data));
1392 }
1393
1394
1395 // A fast matrix-shaper evaluator for 8 bits. This is a bit ticky since I'm using 1.14 signed fixed point
1396 // to accomplish some performance. Actually it takes 256x3 16 bits tables and 16385 x 3 tables of 8 bits,
1397 // in total about 50K, and the performance boost is huge!
1398 static
1399 void MatShaperEval16(register const cmsUInt16Number In[],
1400 register cmsUInt16Number Out[],
1401 register const void* D)
1402 {
1403 MatShaper8Data* p = (MatShaper8Data*) D;
1404 cmsS1Fixed14Number l1, l2, l3, r, g, b;
1405 cmsUInt32Number ri, gi, bi;
1406
1407 // In this case (and only in this case!) we can use this simplification since
1408 // In[] is assured to come from a 8 bit number. (a << 8 | a)
1409 ri = In[0] & 0xFF;
1410 gi = In[1] & 0xFF;
1411 bi = In[2] & 0xFF;
1412
1413 // Across first shaper, which also converts to 1.14 fixed point
1414 r = p->Shaper1R[ri];
1415 g = p->Shaper1G[gi];
1416 b = p->Shaper1B[bi];
1417
1418 // Evaluate the matrix in 1.14 fixed point
1419 l1 = (p->Mat[0][0] * r + p->Mat[0][1] * g + p->Mat[0][2] * b + p->Off[0] + 0x2000) >> 14;
1420 l2 = (p->Mat[1][0] * r + p->Mat[1][1] * g + p->Mat[1][2] * b + p->Off[1] + 0x2000) >> 14;
1421 l3 = (p->Mat[2][0] * r + p->Mat[2][1] * g + p->Mat[2][2] * b + p->Off[2] + 0x2000) >> 14;
1422
1423 // Now we have to clip to 0..1.0 range
1424 ri = (l1 < 0) ? 0 : ((l1 > 16384) ? 16384 : l1);
1425 gi = (l2 < 0) ? 0 : ((l2 > 16384) ? 16384 : l2);
1426 bi = (l3 < 0) ? 0 : ((l3 > 16384) ? 16384 : l3);
1427
1428 // And across second shaper,
1429 Out[0] = p->Shaper2R[ri];
1430 Out[1] = p->Shaper2G[gi];
1431 Out[2] = p->Shaper2B[bi];
1432
1433 }
1434
1435 // This table converts from 8 bits to 1.14 after applying the curve
1436 static
1437 void FillFirstShaper(cmsS1Fixed14Number* Table, cmsToneCurve* Curve)
1438 {
1439 int i;
1440 cmsFloat32Number R, y;
1441
1442 for (i=0; i < 256; i++) {
1443
1444 R = (cmsFloat32Number) (i / 255.0);
1445 y = cmsEvalToneCurveFloat(Curve, R);
1446
1447 Table[i] = DOUBLE_TO_1FIXED14(y);
1448 }
1449 }
1450
1451 // This table converts form 1.14 (being 0x4000 the last entry) to 8 bits after applying the curve
1452 static
1453 void FillSecondShaper(cmsUInt16Number* Table, cmsToneCurve* Curve, cmsBool Is8BitsOutput)
1454 {
1455 int i;
1456 cmsFloat32Number R, Val;
1457
1458 for (i=0; i < 16385; i++) {
1459
1460 R = (cmsFloat32Number) (i / 16384.0);
1461 Val = cmsEvalToneCurveFloat(Curve, R); // Val comes 0..1.0
1462
1463 if (Is8BitsOutput) {
1464
1465 // If 8 bits output, we can optimize further by computing the / 257 part.
1466 // first we compute the resulting byte and then we store the byte times
1467 // 257. This quantization allows to round very quick by doing a >> 8, but
1468 // since the low byte is always equal to msb, we can do a & 0xff and this works!
1469 cmsUInt16Number w = _cmsQuickSaturateWord(Val * 65535.0);
1470 cmsUInt8Number b = FROM_16_TO_8(w);
1471
1472 Table[i] = FROM_8_TO_16(b);
1473 }
1474 else Table[i] = _cmsQuickSaturateWord(Val * 65535.0);
1475 }
1476 }
1477
1478 // Compute the matrix-shaper structure
1479 static
1480 cmsBool SetMatShaper(cmsPipeline* Dest, cmsToneCurve* Curve1[3], cmsMAT3* Mat, cmsVEC3* Off, cmsToneCurve* Curve2[3], cmsUInt32Number* OutputFormat)
1481 {
1482 MatShaper8Data* p;
1483 int i, j;
1484 cmsBool Is8Bits = _cmsFormatterIs8bit(*OutputFormat);
1485
1486 // Allocate a big chuck of memory to store precomputed tables
1487 p = (MatShaper8Data*) _cmsMalloc(Dest ->ContextID, sizeof(MatShaper8Data));
1488 if (p == NULL) return FALSE;
1489
1490 p -> ContextID = Dest -> ContextID;
1491
1492 // Precompute tables
1493 FillFirstShaper(p ->Shaper1R, Curve1[0]);
1494 FillFirstShaper(p ->Shaper1G, Curve1[1]);
1495 FillFirstShaper(p ->Shaper1B, Curve1[2]);
1496
1497 FillSecondShaper(p ->Shaper2R, Curve2[0], Is8Bits);
1498 FillSecondShaper(p ->Shaper2G, Curve2[1], Is8Bits);
1499 FillSecondShaper(p ->Shaper2B, Curve2[2], Is8Bits);
1500
1501 // Convert matrix to nFixed14. Note that those values may take more than 16 bits as
1502 for (i=0; i < 3; i++) {
1503 for (j=0; j < 3; j++) {
1504 p ->Mat[i][j] = DOUBLE_TO_1FIXED14(Mat->v[i].n[j]);
1505 }
1506 }
1507
1508 for (i=0; i < 3; i++) {
1509
1510 if (Off == NULL) {
1511 p ->Off[i] = 0;
1512 }
1513 else {
1514 p ->Off[i] = DOUBLE_TO_1FIXED14(Off->n[i]);
1515 }
1516 }
1517
1518 // Mark as optimized for faster formatter
1519 if (Is8Bits)
1520 *OutputFormat |= OPTIMIZED_SH(1);
1521
1522 // Fill function pointers
1523 _cmsPipelineSetOptimizationParameters(Dest, MatShaperEval16, (void*) p, FreeMatShaper, DupMatShaper);
1524 return TRUE;
1525 }
1526
1527 // 8 bits on input allows matrix-shaper boot up to 25 Mpixels per second on RGB. That's fast!
1528 // TODO: Allow a third matrix for abs. colorimetric
1529 static
1530 cmsBool OptimizeMatrixShaper(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags)
1531 {
1532 cmsStage* Curve1, *Curve2;
1533 cmsStage* Matrix1, *Matrix2;
1534 _cmsStageMatrixData* Data1;
1535 _cmsStageMatrixData* Data2;
1536 cmsMAT3 res;
1537 cmsBool IdentityMat;
1538 cmsPipeline* Dest, *Src;
1539
1540 // Only works on RGB to RGB
1541 if (T_CHANNELS(*InputFormat) != 3 || T_CHANNELS(*OutputFormat) != 3) return FALSE;
1542
1543 // Only works on 8 bit input
1544 if (!_cmsFormatterIs8bit(*InputFormat)) return FALSE;
1545
1546 // Seems suitable, proceed
1547 Src = *Lut;
1548
1549 // Check for shaper-matrix-matrix-shaper structure, that is what this optimizer stands for
1550 if (!cmsPipelineCheckAndRetreiveStages(Src, 4,
1551 cmsSigCurveSetElemType, cmsSigMatrixElemType, cmsSigMatrixElemType, cmsSigCurveSetElemType,
1552 &Curve1, &Matrix1, &Matrix2, &Curve2)) return FALSE;
1553
1554 // Get both matrices
1555 Data1 = (_cmsStageMatrixData*) cmsStageData(Matrix1);
1556 Data2 = (_cmsStageMatrixData*) cmsStageData(Matrix2);
1557
1558 // Input offset should be zero
1559 if (Data1 ->Offset != NULL) return FALSE;
1560
1561 // Multiply both matrices to get the result
1562 _cmsMAT3per(&res, (cmsMAT3*) Data2 ->Double, (cmsMAT3*) Data1 ->Double);
1563
1564 // Now the result is in res + Data2 -> Offset. Maybe is a plain identity?
1565 IdentityMat = FALSE;
1566 if (_cmsMAT3isIdentity(&res) && Data2 ->Offset == NULL) {
1567
1568 // We can get rid of full matrix
1569 IdentityMat = TRUE;
1570 }
1571
1572 // Allocate an empty LUT
1573 Dest = cmsPipelineAlloc(Src ->ContextID, Src ->InputChannels, Src ->OutputChannels);
1574 if (!Dest) return FALSE;
1575
1576 // Assamble the new LUT
1577 if (!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, cmsStageDup(Curve1)))
1578 goto Error;
1579
1580 if (!IdentityMat)
1581 if (!cmsPipelineInsertStage(Dest, cmsAT_END, cmsStageAllocMatrix(Dest ->ContextID, 3, 3, (const cmsFloat64Number*) &res, Data2 ->Offset)))
1582 goto Error;
1583 if (!cmsPipelineInsertStage(Dest, cmsAT_END, cmsStageDup(Curve2)))
1584 goto Error;
1585
1586 // If identity on matrix, we can further optimize the curves, so call the join curves routine
1587 if (IdentityMat) {
1588
1589 OptimizeByJoiningCurves(&Dest, Intent, InputFormat, OutputFormat, dwFlags);
1590 }
1591 else {
1592 _cmsStageToneCurvesData* mpeC1 = (_cmsStageToneCurvesData*) cmsStageData(Curve1);
1593 _cmsStageToneCurvesData* mpeC2 = (_cmsStageToneCurvesData*) cmsStageData(Curve2);
1594
1595 // In this particular optimization, caché does not help as it takes more time to deal with
1596 // the caché that with the pixel handling
1597 *dwFlags |= cmsFLAGS_NOCACHE;
1598
1599 // Setup the optimizarion routines
1600 SetMatShaper(Dest, mpeC1 ->TheCurves, &res, (cmsVEC3*) Data2 ->Offset, mpeC2->TheCurves, OutputFormat);
1601 }
1602
1603 cmsPipelineFree(Src);
1604 *Lut = Dest;
1605 return TRUE;
1606 Error:
1607 // Leave Src unchanged
1608 cmsPipelineFree(Dest);
1609 return FALSE;
1610 }
1611
1612
1613 // -------------------------------------------------------------------------------------------------------------------------------------
1614 // Optimization plug-ins
1615
1616 // List of optimizations
1617 typedef struct _cmsOptimizationCollection_st {
1618
1619 _cmsOPToptimizeFn OptimizePtr;
1620
1621 struct _cmsOptimizationCollection_st *Next;
1622
1623 } _cmsOptimizationCollection;
1624
1625
1626 // The built-in list. We currently implement 4 types of optimizations. Joining of curves, matrix-shaper, linearization and resampling
1627 static _cmsOptimizationCollection DefaultOptimization[] = {
1628
1629 { OptimizeByJoiningCurves, &DefaultOptimization[1] },
1630 { OptimizeMatrixShaper, &DefaultOptimization[2] },
1631 { OptimizeByComputingLinearization, &DefaultOptimization[3] },
1632 { OptimizeByResampling, NULL }
1633 };
1634
1635 // The linked list head
1636 static _cmsOptimizationCollection* OptimizationCollection = DefaultOptimization;
1637
1638 // Register new ways to optimize
1639 cmsBool _cmsRegisterOptimizationPlugin(cmsContext id, cmsPluginBase* Data)
1640 {
1641 cmsPluginOptimization* Plugin = (cmsPluginOptimization*) Data;
1642 _cmsOptimizationCollection* fl;
1643
1644 if (Data == NULL) {
1645
1646 OptimizationCollection = DefaultOptimization;
1647 return TRUE;
1648 }
1649
1650 // Optimizer callback is required
1651 if (Plugin ->OptimizePtr == NULL) return FALSE;
1652
1653 fl = (_cmsOptimizationCollection*) _cmsPluginMalloc(id, sizeof(_cmsOptimizationCollection));
1654 if (fl == NULL) return FALSE;
1655
1656 // Copy the parameters
1657 fl ->OptimizePtr = Plugin ->OptimizePtr;
1658
1659 // Keep linked list
1660 fl ->Next = OptimizationCollection;
1661 OptimizationCollection = fl;
1662
1663 // All is ok
1664 return TRUE;
1665 }
1666
1667 // The entry point for LUT optimization
1668 cmsBool _cmsOptimizePipeline(cmsPipeline** PtrLut,
1669 int Intent,
1670 cmsUInt32Number* InputFormat,
1671 cmsUInt32Number* OutputFormat,
1672 cmsUInt32Number* dwFlags)
1673 {
1674 _cmsOptimizationCollection* Opts;
1675 cmsBool AnySuccess = FALSE;
1676
1677 // A CLUT is being asked, so force this specific optimization
1678 if (*dwFlags & cmsFLAGS_FORCE_CLUT) {
1679
1680 PreOptimize(*PtrLut);
1681 return OptimizeByResampling(PtrLut, Intent, InputFormat, OutputFormat, dwFlags);
1682 }
1683
1684 // Anything to optimize?
1685 if ((*PtrLut) ->Elements == NULL) {
1686 _cmsPipelineSetOptimizationParameters(*PtrLut, FastIdentity16, (void*) *PtrLut, NULL, NULL);
1687 return TRUE;
1688 }
1689
1690 // Try to get rid of identities and trivial conversions.
1691 AnySuccess = PreOptimize(*PtrLut);
1692
1693 // After removal do we end with an identity?
1694 if ((*PtrLut) ->Elements == NULL) {
1695 _cmsPipelineSetOptimizationParameters(*PtrLut, FastIdentity16, (void*) *PtrLut, NULL, NULL);
1696 return TRUE;
1697 }
1698
1699 // Do not optimize, keep all precision
1700 if (*dwFlags & cmsFLAGS_NOOPTIMIZE)
1701 return FALSE;
1702
1703 // Try built-in optimizations and plug-in
1704 for (Opts = OptimizationCollection;
1705 Opts != NULL;
1706 Opts = Opts ->Next) {
1707
1708 // If one schema succeeded, we are done
1709 if (Opts ->OptimizePtr(PtrLut, Intent, InputFormat, OutputFormat, dwFlags)) {
1710
1711 return TRUE; // Optimized!
1712 }
1713 }
1714
1715 // Only simple optimizations succeeded
1716 return AnySuccess;
1717 }
1718
1719
1720
Mirror of AROS' svn.