1 | /* |
2 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
3 | * |
4 | * This code is free software; you can redistribute it and/or modify it |
5 | * under the terms of the GNU General Public License version 2 only, as |
6 | * published by the Free Software Foundation. Oracle designates this |
7 | * particular file as subject to the "Classpath" exception as provided |
8 | * by Oracle in the LICENSE file that accompanied this code. |
9 | * |
10 | * This code is distributed in the hope that it will be useful, but WITHOUT |
11 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
12 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
13 | * version 2 for more details (a copy is included in the LICENSE file that |
14 | * accompanied this code). |
15 | * |
16 | * You should have received a copy of the GNU General Public License version |
17 | * 2 along with this work; if not, write to the Free Software Foundation, |
18 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
19 | * |
20 | * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
21 | * or visit www.oracle.com if you need additional information or have any |
22 | * questions. |
23 | */ |
24 | |
25 | // This file is available under and governed by the GNU General Public |
26 | // License version 2 only, as published by the Free Software Foundation. |
27 | // However, the following notice accompanied the original version of this |
28 | // file: |
29 | // |
30 | //--------------------------------------------------------------------------------- |
31 | // |
32 | // Little Color Management System |
33 | // Copyright (c) 1998-2017 Marti Maria Saguer |
34 | // |
35 | // Permission is hereby granted, free of charge, to any person obtaining |
36 | // a copy of this software and associated documentation files (the "Software"), |
37 | // to deal in the Software without restriction, including without limitation |
38 | // the rights to use, copy, modify, merge, publish, distribute, sublicense, |
39 | // and/or sell copies of the Software, and to permit persons to whom the Software |
40 | // is furnished to do so, subject to the following conditions: |
41 | // |
42 | // The above copyright notice and this permission notice shall be included in |
43 | // all copies or substantial portions of the Software. |
44 | // |
45 | // THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, |
46 | // EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO |
47 | // THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND |
48 | // NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE |
49 | // LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION |
50 | // OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION |
51 | // WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. |
52 | // |
53 | //--------------------------------------------------------------------------------- |
54 | // |
55 | |
56 | #include "lcms2_internal.h" |
57 | |
58 | |
59 | //---------------------------------------------------------------------------------- |
60 | |
61 | // Optimization for 8 bits, Shaper-CLUT (3 inputs only) |
62 | typedef struct { |
63 | |
64 | cmsContext ContextID; |
65 | |
66 | const cmsInterpParams* p; // Tetrahedrical interpolation parameters. This is a not-owned pointer. |
67 | |
68 | cmsUInt16Number rx[256], ry[256], rz[256]; |
69 | cmsUInt32Number X0[256], Y0[256], Z0[256]; // Precomputed nodes and offsets for 8-bit input data |
70 | |
71 | |
72 | } Prelin8Data; |
73 | |
74 | |
75 | // Generic optimization for 16 bits Shaper-CLUT-Shaper (any inputs) |
76 | typedef struct { |
77 | |
78 | cmsContext ContextID; |
79 | |
80 | // Number of channels |
81 | cmsUInt32Number nInputs; |
82 | cmsUInt32Number nOutputs; |
83 | |
84 | _cmsInterpFn16 EvalCurveIn16[MAX_INPUT_DIMENSIONS]; // The maximum number of input channels is known in advance |
85 | cmsInterpParams* ParamsCurveIn16[MAX_INPUT_DIMENSIONS]; |
86 | |
87 | _cmsInterpFn16 EvalCLUT; // The evaluator for 3D grid |
88 | const cmsInterpParams* CLUTparams; // (not-owned pointer) |
89 | |
90 | |
91 | _cmsInterpFn16* EvalCurveOut16; // Points to an array of curve evaluators in 16 bits (not-owned pointer) |
92 | cmsInterpParams** ParamsCurveOut16; // Points to an array of references to interpolation params (not-owned pointer) |
93 | |
94 | |
95 | } Prelin16Data; |
96 | |
97 | |
98 | // Optimization for matrix-shaper in 8 bits. Numbers are operated in n.14 signed, tables are stored in 1.14 fixed |
99 | |
100 | typedef cmsInt32Number cmsS1Fixed14Number; // Note that this may hold more than 16 bits! |
101 | |
102 | #define DOUBLE_TO_1FIXED14(x) ((cmsS1Fixed14Number) floor((x) * 16384.0 + 0.5)) |
103 | |
104 | typedef struct { |
105 | |
106 | cmsContext ContextID; |
107 | |
108 | cmsS1Fixed14Number Shaper1R[256]; // from 0..255 to 1.14 (0.0...1.0) |
109 | cmsS1Fixed14Number Shaper1G[256]; |
110 | cmsS1Fixed14Number Shaper1B[256]; |
111 | |
112 | cmsS1Fixed14Number Mat[3][3]; // n.14 to n.14 (needs a saturation after that) |
113 | cmsS1Fixed14Number Off[3]; |
114 | |
115 | cmsUInt16Number Shaper2R[16385]; // 1.14 to 0..255 |
116 | cmsUInt16Number Shaper2G[16385]; |
117 | cmsUInt16Number Shaper2B[16385]; |
118 | |
119 | } MatShaper8Data; |
120 | |
121 | // Curves, optimization is shared between 8 and 16 bits |
122 | typedef struct { |
123 | |
124 | cmsContext ContextID; |
125 | |
126 | cmsUInt32Number nCurves; // Number of curves |
127 | cmsUInt32Number nElements; // Elements in curves |
128 | cmsUInt16Number** Curves; // Points to a dynamically allocated array |
129 | |
130 | } Curves16Data; |
131 | |
132 | |
133 | // Simple optimizations ---------------------------------------------------------------------------------------------------------- |
134 | |
135 | |
136 | // Remove an element in linked chain |
137 | static |
138 | void _RemoveElement(cmsStage** head) |
139 | { |
140 | cmsStage* mpe = *head; |
141 | cmsStage* next = mpe ->Next; |
142 | *head = next; |
143 | cmsStageFree(mpe); |
144 | } |
145 | |
146 | // Remove all identities in chain. Note that pt actually is a double pointer to the element that holds the pointer. |
147 | static |
148 | cmsBool _Remove1Op(cmsPipeline* Lut, cmsStageSignature UnaryOp) |
149 | { |
150 | cmsStage** pt = &Lut ->Elements; |
151 | cmsBool AnyOpt = FALSE; |
152 | |
153 | while (*pt != NULL) { |
154 | |
155 | if ((*pt) ->Implements == UnaryOp) { |
156 | _RemoveElement(pt); |
157 | AnyOpt = TRUE; |
158 | } |
159 | else |
160 | pt = &((*pt) -> Next); |
161 | } |
162 | |
163 | return AnyOpt; |
164 | } |
165 | |
166 | // Same, but only if two adjacent elements are found |
167 | static |
168 | cmsBool _Remove2Op(cmsPipeline* Lut, cmsStageSignature Op1, cmsStageSignature Op2) |
169 | { |
170 | cmsStage** pt1; |
171 | cmsStage** pt2; |
172 | cmsBool AnyOpt = FALSE; |
173 | |
174 | pt1 = &Lut ->Elements; |
175 | if (*pt1 == NULL) return AnyOpt; |
176 | |
177 | while (*pt1 != NULL) { |
178 | |
179 | pt2 = &((*pt1) -> Next); |
180 | if (*pt2 == NULL) return AnyOpt; |
181 | |
182 | if ((*pt1) ->Implements == Op1 && (*pt2) ->Implements == Op2) { |
183 | _RemoveElement(pt2); |
184 | _RemoveElement(pt1); |
185 | AnyOpt = TRUE; |
186 | } |
187 | else |
188 | pt1 = &((*pt1) -> Next); |
189 | } |
190 | |
191 | return AnyOpt; |
192 | } |
193 | |
194 | |
195 | static |
196 | cmsBool CloseEnoughFloat(cmsFloat64Number a, cmsFloat64Number b) |
197 | { |
198 | return fabs(b - a) < 0.00001f; |
199 | } |
200 | |
201 | static |
202 | cmsBool isFloatMatrixIdentity(const cmsMAT3* a) |
203 | { |
204 | cmsMAT3 Identity; |
205 | int i, j; |
206 | |
207 | _cmsMAT3identity(&Identity); |
208 | |
209 | for (i = 0; i < 3; i++) |
210 | for (j = 0; j < 3; j++) |
211 | if (!CloseEnoughFloat(a->v[i].n[j], Identity.v[i].n[j])) return FALSE; |
212 | |
213 | return TRUE; |
214 | } |
215 | // if two adjacent matrices are found, multiply them. |
216 | static |
217 | cmsBool _MultiplyMatrix(cmsPipeline* Lut) |
218 | { |
219 | cmsStage** pt1; |
220 | cmsStage** pt2; |
221 | cmsStage* chain; |
222 | cmsBool AnyOpt = FALSE; |
223 | |
224 | pt1 = &Lut->Elements; |
225 | if (*pt1 == NULL) return AnyOpt; |
226 | |
227 | while (*pt1 != NULL) { |
228 | |
229 | pt2 = &((*pt1)->Next); |
230 | if (*pt2 == NULL) return AnyOpt; |
231 | |
232 | if ((*pt1)->Implements == cmsSigMatrixElemType && (*pt2)->Implements == cmsSigMatrixElemType) { |
233 | |
234 | // Get both matrices |
235 | _cmsStageMatrixData* m1 = (_cmsStageMatrixData*) cmsStageData(*pt1); |
236 | _cmsStageMatrixData* m2 = (_cmsStageMatrixData*) cmsStageData(*pt2); |
237 | cmsMAT3 res; |
238 | |
239 | // Input offset and output offset should be zero to use this optimization |
240 | if (m1->Offset != NULL || m2 ->Offset != NULL || |
241 | cmsStageInputChannels(*pt1) != 3 || cmsStageOutputChannels(*pt1) != 3 || |
242 | cmsStageInputChannels(*pt2) != 3 || cmsStageOutputChannels(*pt2) != 3) |
243 | return FALSE; |
244 | |
245 | // Multiply both matrices to get the result |
246 | _cmsMAT3per(&res, (cmsMAT3*)m2->Double, (cmsMAT3*)m1->Double); |
247 | |
248 | // Get the next in chain after the matrices |
249 | chain = (*pt2)->Next; |
250 | |
251 | // Remove both matrices |
252 | _RemoveElement(pt2); |
253 | _RemoveElement(pt1); |
254 | |
255 | // Now what if the result is a plain identity? |
256 | if (!isFloatMatrixIdentity(&res)) { |
257 | |
258 | // We can not get rid of full matrix |
259 | cmsStage* Multmat = cmsStageAllocMatrix(Lut->ContextID, 3, 3, (const cmsFloat64Number*) &res, NULL); |
260 | if (Multmat == NULL) return FALSE; // Should never happen |
261 | |
262 | // Recover the chain |
263 | Multmat->Next = chain; |
264 | *pt1 = Multmat; |
265 | } |
266 | |
267 | AnyOpt = TRUE; |
268 | } |
269 | else |
270 | pt1 = &((*pt1)->Next); |
271 | } |
272 | |
273 | return AnyOpt; |
274 | } |
275 | |
276 | |
277 | // Preoptimize just gets rif of no-ops coming paired. Conversion from v2 to v4 followed |
278 | // by a v4 to v2 and vice-versa. The elements are then discarded. |
279 | static |
280 | cmsBool PreOptimize(cmsPipeline* Lut) |
281 | { |
282 | cmsBool AnyOpt = FALSE, Opt; |
283 | |
284 | do { |
285 | |
286 | Opt = FALSE; |
287 | |
288 | // Remove all identities |
289 | Opt |= _Remove1Op(Lut, cmsSigIdentityElemType); |
290 | |
291 | // Remove XYZ2Lab followed by Lab2XYZ |
292 | Opt |= _Remove2Op(Lut, cmsSigXYZ2LabElemType, cmsSigLab2XYZElemType); |
293 | |
294 | // Remove Lab2XYZ followed by XYZ2Lab |
295 | Opt |= _Remove2Op(Lut, cmsSigLab2XYZElemType, cmsSigXYZ2LabElemType); |
296 | |
297 | // Remove V4 to V2 followed by V2 to V4 |
298 | Opt |= _Remove2Op(Lut, cmsSigLabV4toV2, cmsSigLabV2toV4); |
299 | |
300 | // Remove V2 to V4 followed by V4 to V2 |
301 | Opt |= _Remove2Op(Lut, cmsSigLabV2toV4, cmsSigLabV4toV2); |
302 | |
303 | // Remove float pcs Lab conversions |
304 | Opt |= _Remove2Op(Lut, cmsSigLab2FloatPCS, cmsSigFloatPCS2Lab); |
305 | |
306 | // Remove float pcs Lab conversions |
307 | Opt |= _Remove2Op(Lut, cmsSigXYZ2FloatPCS, cmsSigFloatPCS2XYZ); |
308 | |
309 | // Simplify matrix. |
310 | Opt |= _MultiplyMatrix(Lut); |
311 | |
312 | if (Opt) AnyOpt = TRUE; |
313 | |
314 | } while (Opt); |
315 | |
316 | return AnyOpt; |
317 | } |
318 | |
319 | static |
320 | void Eval16nop1D(register const cmsUInt16Number Input[], |
321 | register cmsUInt16Number Output[], |
322 | register const struct _cms_interp_struc* p) |
323 | { |
324 | Output[0] = Input[0]; |
325 | |
326 | cmsUNUSED_PARAMETER(p); |
327 | } |
328 | |
329 | static |
330 | void PrelinEval16(register const cmsUInt16Number Input[], |
331 | register cmsUInt16Number Output[], |
332 | register const void* D) |
333 | { |
334 | Prelin16Data* p16 = (Prelin16Data*) D; |
335 | cmsUInt16Number StageABC[MAX_INPUT_DIMENSIONS]; |
336 | cmsUInt16Number StageDEF[cmsMAXCHANNELS]; |
337 | cmsUInt32Number i; |
338 | |
339 | for (i=0; i < p16 ->nInputs; i++) { |
340 | |
341 | p16 ->EvalCurveIn16[i](&Input[i], &StageABC[i], p16 ->ParamsCurveIn16[i]); |
342 | } |
343 | |
344 | p16 ->EvalCLUT(StageABC, StageDEF, p16 ->CLUTparams); |
345 | |
346 | for (i=0; i < p16 ->nOutputs; i++) { |
347 | |
348 | p16 ->EvalCurveOut16[i](&StageDEF[i], &Output[i], p16 ->ParamsCurveOut16[i]); |
349 | } |
350 | } |
351 | |
352 | |
353 | static |
354 | void PrelinOpt16free(cmsContext ContextID, void* ptr) |
355 | { |
356 | Prelin16Data* p16 = (Prelin16Data*) ptr; |
357 | |
358 | _cmsFree(ContextID, p16 ->EvalCurveOut16); |
359 | _cmsFree(ContextID, p16 ->ParamsCurveOut16); |
360 | |
361 | _cmsFree(ContextID, p16); |
362 | } |
363 | |
364 | static |
365 | void* Prelin16dup(cmsContext ContextID, const void* ptr) |
366 | { |
367 | Prelin16Data* p16 = (Prelin16Data*) ptr; |
368 | Prelin16Data* Duped = (Prelin16Data*) _cmsDupMem(ContextID, p16, sizeof(Prelin16Data)); |
369 | |
370 | if (Duped == NULL) return NULL; |
371 | |
372 | Duped->EvalCurveOut16 = (_cmsInterpFn16*) _cmsDupMem(ContextID, p16->EvalCurveOut16, p16->nOutputs * sizeof(_cmsInterpFn16)); |
373 | Duped->ParamsCurveOut16 = (cmsInterpParams**)_cmsDupMem(ContextID, p16->ParamsCurveOut16, p16->nOutputs * sizeof(cmsInterpParams*)); |
374 | |
375 | return Duped; |
376 | } |
377 | |
378 | |
379 | static |
380 | Prelin16Data* PrelinOpt16alloc(cmsContext ContextID, |
381 | const cmsInterpParams* ColorMap, |
382 | cmsUInt32Number nInputs, cmsToneCurve** In, |
383 | cmsUInt32Number nOutputs, cmsToneCurve** Out ) |
384 | { |
385 | cmsUInt32Number i; |
386 | Prelin16Data* p16 = (Prelin16Data*)_cmsMallocZero(ContextID, sizeof(Prelin16Data)); |
387 | if (p16 == NULL) return NULL; |
388 | |
389 | p16 ->nInputs = nInputs; |
390 | p16 ->nOutputs = nOutputs; |
391 | |
392 | |
393 | for (i=0; i < nInputs; i++) { |
394 | |
395 | if (In == NULL) { |
396 | p16 -> ParamsCurveIn16[i] = NULL; |
397 | p16 -> EvalCurveIn16[i] = Eval16nop1D; |
398 | |
399 | } |
400 | else { |
401 | p16 -> ParamsCurveIn16[i] = In[i] ->InterpParams; |
402 | p16 -> EvalCurveIn16[i] = p16 ->ParamsCurveIn16[i]->Interpolation.Lerp16; |
403 | } |
404 | } |
405 | |
406 | p16 ->CLUTparams = ColorMap; |
407 | p16 ->EvalCLUT = ColorMap ->Interpolation.Lerp16; |
408 | |
409 | |
410 | p16 -> EvalCurveOut16 = (_cmsInterpFn16*) _cmsCalloc(ContextID, nOutputs, sizeof(_cmsInterpFn16)); |
411 | p16 -> ParamsCurveOut16 = (cmsInterpParams**) _cmsCalloc(ContextID, nOutputs, sizeof(cmsInterpParams* )); |
412 | |
413 | for (i=0; i < nOutputs; i++) { |
414 | |
415 | if (Out == NULL) { |
416 | p16 ->ParamsCurveOut16[i] = NULL; |
417 | p16 -> EvalCurveOut16[i] = Eval16nop1D; |
418 | } |
419 | else { |
420 | |
421 | p16 ->ParamsCurveOut16[i] = Out[i] ->InterpParams; |
422 | p16 -> EvalCurveOut16[i] = p16 ->ParamsCurveOut16[i]->Interpolation.Lerp16; |
423 | } |
424 | } |
425 | |
426 | return p16; |
427 | } |
428 | |
429 | |
430 | |
431 | // Resampling --------------------------------------------------------------------------------- |
432 | |
433 | #define PRELINEARIZATION_POINTS 4096 |
434 | |
435 | // Sampler implemented by another LUT. This is a clean way to precalculate the devicelink 3D CLUT for |
436 | // almost any transform. We use floating point precision and then convert from floating point to 16 bits. |
437 | static |
438 | cmsInt32Number XFormSampler16(register const cmsUInt16Number In[], register cmsUInt16Number Out[], register void* Cargo) |
439 | { |
440 | cmsPipeline* Lut = (cmsPipeline*) Cargo; |
441 | cmsFloat32Number InFloat[cmsMAXCHANNELS], OutFloat[cmsMAXCHANNELS]; |
442 | cmsUInt32Number i; |
443 | |
444 | _cmsAssert(Lut -> InputChannels < cmsMAXCHANNELS); |
445 | _cmsAssert(Lut -> OutputChannels < cmsMAXCHANNELS); |
446 | |
447 | // From 16 bit to floating point |
448 | for (i=0; i < Lut ->InputChannels; i++) |
449 | InFloat[i] = (cmsFloat32Number) (In[i] / 65535.0); |
450 | |
451 | // Evaluate in floating point |
452 | cmsPipelineEvalFloat(InFloat, OutFloat, Lut); |
453 | |
454 | // Back to 16 bits representation |
455 | for (i=0; i < Lut ->OutputChannels; i++) |
456 | Out[i] = _cmsQuickSaturateWord(OutFloat[i] * 65535.0); |
457 | |
458 | // Always succeed |
459 | return TRUE; |
460 | } |
461 | |
462 | // Try to see if the curves of a given MPE are linear |
463 | static |
464 | cmsBool AllCurvesAreLinear(cmsStage* mpe) |
465 | { |
466 | cmsToneCurve** Curves; |
467 | cmsUInt32Number i, n; |
468 | |
469 | Curves = _cmsStageGetPtrToCurveSet(mpe); |
470 | if (Curves == NULL) return FALSE; |
471 | |
472 | n = cmsStageOutputChannels(mpe); |
473 | |
474 | for (i=0; i < n; i++) { |
475 | if (!cmsIsToneCurveLinear(Curves[i])) return FALSE; |
476 | } |
477 | |
478 | return TRUE; |
479 | } |
480 | |
481 | // This function replaces a specific node placed in "At" by the "Value" numbers. Its purpose |
482 | // is to fix scum dot on broken profiles/transforms. Works on 1, 3 and 4 channels |
483 | static |
484 | cmsBool PatchLUT(cmsStage* CLUT, cmsUInt16Number At[], cmsUInt16Number Value[], |
485 | cmsUInt32Number nChannelsOut, cmsUInt32Number nChannelsIn) |
486 | { |
487 | _cmsStageCLutData* Grid = (_cmsStageCLutData*) CLUT ->Data; |
488 | cmsInterpParams* p16 = Grid ->Params; |
489 | cmsFloat64Number px, py, pz, pw; |
490 | int x0, y0, z0, w0; |
491 | int i, index; |
492 | |
493 | if (CLUT -> Type != cmsSigCLutElemType) { |
494 | cmsSignalError(CLUT->ContextID, cmsERROR_INTERNAL, "(internal) Attempt to PatchLUT on non-lut stage" ); |
495 | return FALSE; |
496 | } |
497 | |
498 | if (nChannelsIn == 4) { |
499 | |
500 | px = ((cmsFloat64Number) At[0] * (p16->Domain[0])) / 65535.0; |
501 | py = ((cmsFloat64Number) At[1] * (p16->Domain[1])) / 65535.0; |
502 | pz = ((cmsFloat64Number) At[2] * (p16->Domain[2])) / 65535.0; |
503 | pw = ((cmsFloat64Number) At[3] * (p16->Domain[3])) / 65535.0; |
504 | |
505 | x0 = (int) floor(px); |
506 | y0 = (int) floor(py); |
507 | z0 = (int) floor(pz); |
508 | w0 = (int) floor(pw); |
509 | |
510 | if (((px - x0) != 0) || |
511 | ((py - y0) != 0) || |
512 | ((pz - z0) != 0) || |
513 | ((pw - w0) != 0)) return FALSE; // Not on exact node |
514 | |
515 | index = (int) p16 -> opta[3] * x0 + |
516 | (int) p16 -> opta[2] * y0 + |
517 | (int) p16 -> opta[1] * z0 + |
518 | (int) p16 -> opta[0] * w0; |
519 | } |
520 | else |
521 | if (nChannelsIn == 3) { |
522 | |
523 | px = ((cmsFloat64Number) At[0] * (p16->Domain[0])) / 65535.0; |
524 | py = ((cmsFloat64Number) At[1] * (p16->Domain[1])) / 65535.0; |
525 | pz = ((cmsFloat64Number) At[2] * (p16->Domain[2])) / 65535.0; |
526 | |
527 | x0 = (int) floor(px); |
528 | y0 = (int) floor(py); |
529 | z0 = (int) floor(pz); |
530 | |
531 | if (((px - x0) != 0) || |
532 | ((py - y0) != 0) || |
533 | ((pz - z0) != 0)) return FALSE; // Not on exact node |
534 | |
535 | index = (int) p16 -> opta[2] * x0 + |
536 | (int) p16 -> opta[1] * y0 + |
537 | (int) p16 -> opta[0] * z0; |
538 | } |
539 | else |
540 | if (nChannelsIn == 1) { |
541 | |
542 | px = ((cmsFloat64Number) At[0] * (p16->Domain[0])) / 65535.0; |
543 | |
544 | x0 = (int) floor(px); |
545 | |
546 | if (((px - x0) != 0)) return FALSE; // Not on exact node |
547 | |
548 | index = (int) p16 -> opta[0] * x0; |
549 | } |
550 | else { |
551 | cmsSignalError(CLUT->ContextID, cmsERROR_INTERNAL, "(internal) %d Channels are not supported on PatchLUT" , nChannelsIn); |
552 | return FALSE; |
553 | } |
554 | |
555 | for (i = 0; i < (int) nChannelsOut; i++) |
556 | Grid->Tab.T[index + i] = Value[i]; |
557 | |
558 | return TRUE; |
559 | } |
560 | |
561 | // Auxiliary, to see if two values are equal or very different |
562 | static |
563 | cmsBool WhitesAreEqual(cmsUInt32Number n, cmsUInt16Number White1[], cmsUInt16Number White2[] ) |
564 | { |
565 | cmsUInt32Number i; |
566 | |
567 | for (i=0; i < n; i++) { |
568 | |
569 | if (abs(White1[i] - White2[i]) > 0xf000) return TRUE; // Values are so extremely different that the fixup should be avoided |
570 | if (White1[i] != White2[i]) return FALSE; |
571 | } |
572 | return TRUE; |
573 | } |
574 | |
575 | |
576 | // Locate the node for the white point and fix it to pure white in order to avoid scum dot. |
577 | static |
578 | cmsBool FixWhiteMisalignment(cmsPipeline* Lut, cmsColorSpaceSignature EntryColorSpace, cmsColorSpaceSignature ExitColorSpace) |
579 | { |
580 | cmsUInt16Number *WhitePointIn, *WhitePointOut; |
581 | cmsUInt16Number WhiteIn[cmsMAXCHANNELS], WhiteOut[cmsMAXCHANNELS], ObtainedOut[cmsMAXCHANNELS]; |
582 | cmsUInt32Number i, nOuts, nIns; |
583 | cmsStage *PreLin = NULL, *CLUT = NULL, *PostLin = NULL; |
584 | |
585 | if (!_cmsEndPointsBySpace(EntryColorSpace, |
586 | &WhitePointIn, NULL, &nIns)) return FALSE; |
587 | |
588 | if (!_cmsEndPointsBySpace(ExitColorSpace, |
589 | &WhitePointOut, NULL, &nOuts)) return FALSE; |
590 | |
591 | // It needs to be fixed? |
592 | if (Lut ->InputChannels != nIns) return FALSE; |
593 | if (Lut ->OutputChannels != nOuts) return FALSE; |
594 | |
595 | cmsPipelineEval16(WhitePointIn, ObtainedOut, Lut); |
596 | |
597 | if (WhitesAreEqual(nOuts, WhitePointOut, ObtainedOut)) return TRUE; // whites already match |
598 | |
599 | // Check if the LUT comes as Prelin, CLUT or Postlin. We allow all combinations |
600 | if (!cmsPipelineCheckAndRetreiveStages(Lut, 3, cmsSigCurveSetElemType, cmsSigCLutElemType, cmsSigCurveSetElemType, &PreLin, &CLUT, &PostLin)) |
601 | if (!cmsPipelineCheckAndRetreiveStages(Lut, 2, cmsSigCurveSetElemType, cmsSigCLutElemType, &PreLin, &CLUT)) |
602 | if (!cmsPipelineCheckAndRetreiveStages(Lut, 2, cmsSigCLutElemType, cmsSigCurveSetElemType, &CLUT, &PostLin)) |
603 | if (!cmsPipelineCheckAndRetreiveStages(Lut, 1, cmsSigCLutElemType, &CLUT)) |
604 | return FALSE; |
605 | |
606 | // We need to interpolate white points of both, pre and post curves |
607 | if (PreLin) { |
608 | |
609 | cmsToneCurve** Curves = _cmsStageGetPtrToCurveSet(PreLin); |
610 | |
611 | for (i=0; i < nIns; i++) { |
612 | WhiteIn[i] = cmsEvalToneCurve16(Curves[i], WhitePointIn[i]); |
613 | } |
614 | } |
615 | else { |
616 | for (i=0; i < nIns; i++) |
617 | WhiteIn[i] = WhitePointIn[i]; |
618 | } |
619 | |
620 | // If any post-linearization, we need to find how is represented white before the curve, do |
621 | // a reverse interpolation in this case. |
622 | if (PostLin) { |
623 | |
624 | cmsToneCurve** Curves = _cmsStageGetPtrToCurveSet(PostLin); |
625 | |
626 | for (i=0; i < nOuts; i++) { |
627 | |
628 | cmsToneCurve* InversePostLin = cmsReverseToneCurve(Curves[i]); |
629 | if (InversePostLin == NULL) { |
630 | WhiteOut[i] = WhitePointOut[i]; |
631 | |
632 | } else { |
633 | |
634 | WhiteOut[i] = cmsEvalToneCurve16(InversePostLin, WhitePointOut[i]); |
635 | cmsFreeToneCurve(InversePostLin); |
636 | } |
637 | } |
638 | } |
639 | else { |
640 | for (i=0; i < nOuts; i++) |
641 | WhiteOut[i] = WhitePointOut[i]; |
642 | } |
643 | |
644 | // Ok, proceed with patching. May fail and we don't care if it fails |
645 | PatchLUT(CLUT, WhiteIn, WhiteOut, nOuts, nIns); |
646 | |
647 | return TRUE; |
648 | } |
649 | |
650 | // ----------------------------------------------------------------------------------------------------------------------------------------------- |
651 | // This function creates simple LUT from complex ones. The generated LUT has an optional set of |
652 | // prelinearization curves, a CLUT of nGridPoints and optional postlinearization tables. |
653 | // These curves have to exist in the original LUT in order to be used in the simplified output. |
654 | // Caller may also use the flags to allow this feature. |
655 | // LUTS with all curves will be simplified to a single curve. Parametric curves are lost. |
656 | // This function should be used on 16-bits LUTS only, as floating point losses precision when simplified |
657 | // ----------------------------------------------------------------------------------------------------------------------------------------------- |
658 | |
659 | static |
660 | cmsBool OptimizeByResampling(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags) |
661 | { |
662 | cmsPipeline* Src = NULL; |
663 | cmsPipeline* Dest = NULL; |
664 | cmsStage* mpe; |
665 | cmsStage* CLUT; |
666 | cmsStage *KeepPreLin = NULL, *KeepPostLin = NULL; |
667 | cmsUInt32Number nGridPoints; |
668 | cmsColorSpaceSignature ColorSpace, OutputColorSpace; |
669 | cmsStage *NewPreLin = NULL; |
670 | cmsStage *NewPostLin = NULL; |
671 | _cmsStageCLutData* DataCLUT; |
672 | cmsToneCurve** DataSetIn; |
673 | cmsToneCurve** DataSetOut; |
674 | Prelin16Data* p16; |
675 | |
676 | // This is a loosy optimization! does not apply in floating-point cases |
677 | if (_cmsFormatterIsFloat(*InputFormat) || _cmsFormatterIsFloat(*OutputFormat)) return FALSE; |
678 | |
679 | ColorSpace = _cmsICCcolorSpace((int) T_COLORSPACE(*InputFormat)); |
680 | OutputColorSpace = _cmsICCcolorSpace((int) T_COLORSPACE(*OutputFormat)); |
681 | |
682 | // Color space must be specified |
683 | if (ColorSpace == (cmsColorSpaceSignature)0 || |
684 | OutputColorSpace == (cmsColorSpaceSignature)0) return FALSE; |
685 | |
686 | nGridPoints = _cmsReasonableGridpointsByColorspace(ColorSpace, *dwFlags); |
687 | |
688 | // For empty LUTs, 2 points are enough |
689 | if (cmsPipelineStageCount(*Lut) == 0) |
690 | nGridPoints = 2; |
691 | |
692 | Src = *Lut; |
693 | |
694 | // Named color pipelines cannot be optimized either |
695 | for (mpe = cmsPipelineGetPtrToFirstStage(Src); |
696 | mpe != NULL; |
697 | mpe = cmsStageNext(mpe)) { |
698 | if (cmsStageType(mpe) == cmsSigNamedColorElemType) return FALSE; |
699 | } |
700 | |
701 | // Allocate an empty LUT |
702 | Dest = cmsPipelineAlloc(Src ->ContextID, Src ->InputChannels, Src ->OutputChannels); |
703 | if (!Dest) return FALSE; |
704 | |
705 | // Prelinearization tables are kept unless indicated by flags |
706 | if (*dwFlags & cmsFLAGS_CLUT_PRE_LINEARIZATION) { |
707 | |
708 | // Get a pointer to the prelinearization element |
709 | cmsStage* PreLin = cmsPipelineGetPtrToFirstStage(Src); |
710 | |
711 | // Check if suitable |
712 | if (PreLin && PreLin ->Type == cmsSigCurveSetElemType) { |
713 | |
714 | // Maybe this is a linear tram, so we can avoid the whole stuff |
715 | if (!AllCurvesAreLinear(PreLin)) { |
716 | |
717 | // All seems ok, proceed. |
718 | NewPreLin = cmsStageDup(PreLin); |
719 | if(!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, NewPreLin)) |
720 | goto Error; |
721 | |
722 | // Remove prelinearization. Since we have duplicated the curve |
723 | // in destination LUT, the sampling should be applied after this stage. |
724 | cmsPipelineUnlinkStage(Src, cmsAT_BEGIN, &KeepPreLin); |
725 | } |
726 | } |
727 | } |
728 | |
729 | // Allocate the CLUT |
730 | CLUT = cmsStageAllocCLut16bit(Src ->ContextID, nGridPoints, Src ->InputChannels, Src->OutputChannels, NULL); |
731 | if (CLUT == NULL) goto Error; |
732 | |
733 | // Add the CLUT to the destination LUT |
734 | if (!cmsPipelineInsertStage(Dest, cmsAT_END, CLUT)) { |
735 | goto Error; |
736 | } |
737 | |
738 | // Postlinearization tables are kept unless indicated by flags |
739 | if (*dwFlags & cmsFLAGS_CLUT_POST_LINEARIZATION) { |
740 | |
741 | // Get a pointer to the postlinearization if present |
742 | cmsStage* PostLin = cmsPipelineGetPtrToLastStage(Src); |
743 | |
744 | // Check if suitable |
745 | if (PostLin && cmsStageType(PostLin) == cmsSigCurveSetElemType) { |
746 | |
747 | // Maybe this is a linear tram, so we can avoid the whole stuff |
748 | if (!AllCurvesAreLinear(PostLin)) { |
749 | |
750 | // All seems ok, proceed. |
751 | NewPostLin = cmsStageDup(PostLin); |
752 | if (!cmsPipelineInsertStage(Dest, cmsAT_END, NewPostLin)) |
753 | goto Error; |
754 | |
755 | // In destination LUT, the sampling should be applied after this stage. |
756 | cmsPipelineUnlinkStage(Src, cmsAT_END, &KeepPostLin); |
757 | } |
758 | } |
759 | } |
760 | |
761 | // Now its time to do the sampling. We have to ignore pre/post linearization |
762 | // The source LUT without pre/post curves is passed as parameter. |
763 | if (!cmsStageSampleCLut16bit(CLUT, XFormSampler16, (void*) Src, 0)) { |
764 | Error: |
765 | // Ops, something went wrong, Restore stages |
766 | if (KeepPreLin != NULL) { |
767 | if (!cmsPipelineInsertStage(Src, cmsAT_BEGIN, KeepPreLin)) { |
768 | _cmsAssert(0); // This never happens |
769 | } |
770 | } |
771 | if (KeepPostLin != NULL) { |
772 | if (!cmsPipelineInsertStage(Src, cmsAT_END, KeepPostLin)) { |
773 | _cmsAssert(0); // This never happens |
774 | } |
775 | } |
776 | cmsPipelineFree(Dest); |
777 | return FALSE; |
778 | } |
779 | |
780 | // Done. |
781 | |
782 | if (KeepPreLin != NULL) cmsStageFree(KeepPreLin); |
783 | if (KeepPostLin != NULL) cmsStageFree(KeepPostLin); |
784 | cmsPipelineFree(Src); |
785 | |
786 | DataCLUT = (_cmsStageCLutData*) CLUT ->Data; |
787 | |
788 | if (NewPreLin == NULL) DataSetIn = NULL; |
789 | else DataSetIn = ((_cmsStageToneCurvesData*) NewPreLin ->Data) ->TheCurves; |
790 | |
791 | if (NewPostLin == NULL) DataSetOut = NULL; |
792 | else DataSetOut = ((_cmsStageToneCurvesData*) NewPostLin ->Data) ->TheCurves; |
793 | |
794 | |
795 | if (DataSetIn == NULL && DataSetOut == NULL) { |
796 | |
797 | _cmsPipelineSetOptimizationParameters(Dest, (_cmsOPTeval16Fn) DataCLUT->Params->Interpolation.Lerp16, DataCLUT->Params, NULL, NULL); |
798 | } |
799 | else { |
800 | |
801 | p16 = PrelinOpt16alloc(Dest ->ContextID, |
802 | DataCLUT ->Params, |
803 | Dest ->InputChannels, |
804 | DataSetIn, |
805 | Dest ->OutputChannels, |
806 | DataSetOut); |
807 | |
808 | _cmsPipelineSetOptimizationParameters(Dest, PrelinEval16, (void*) p16, PrelinOpt16free, Prelin16dup); |
809 | } |
810 | |
811 | |
812 | // Don't fix white on absolute colorimetric |
813 | if (Intent == INTENT_ABSOLUTE_COLORIMETRIC) |
814 | *dwFlags |= cmsFLAGS_NOWHITEONWHITEFIXUP; |
815 | |
816 | if (!(*dwFlags & cmsFLAGS_NOWHITEONWHITEFIXUP)) { |
817 | |
818 | FixWhiteMisalignment(Dest, ColorSpace, OutputColorSpace); |
819 | } |
820 | |
821 | *Lut = Dest; |
822 | return TRUE; |
823 | |
824 | cmsUNUSED_PARAMETER(Intent); |
825 | } |
826 | |
827 | |
828 | // ----------------------------------------------------------------------------------------------------------------------------------------------- |
829 | // Fixes the gamma balancing of transform. This is described in my paper "Prelinearization Stages on |
830 | // Color-Management Application-Specific Integrated Circuits (ASICs)" presented at NIP24. It only works |
831 | // for RGB transforms. See the paper for more details |
832 | // ----------------------------------------------------------------------------------------------------------------------------------------------- |
833 | |
834 | |
835 | // Normalize endpoints by slope limiting max and min. This assures endpoints as well. |
836 | // Descending curves are handled as well. |
837 | static |
838 | void SlopeLimiting(cmsToneCurve* g) |
839 | { |
840 | int BeginVal, EndVal; |
841 | int AtBegin = (int) floor((cmsFloat64Number) g ->nEntries * 0.02 + 0.5); // Cutoff at 2% |
842 | int AtEnd = (int) g ->nEntries - AtBegin - 1; // And 98% |
843 | cmsFloat64Number Val, Slope, beta; |
844 | int i; |
845 | |
846 | if (cmsIsToneCurveDescending(g)) { |
847 | BeginVal = 0xffff; EndVal = 0; |
848 | } |
849 | else { |
850 | BeginVal = 0; EndVal = 0xffff; |
851 | } |
852 | |
853 | // Compute slope and offset for begin of curve |
854 | Val = g ->Table16[AtBegin]; |
855 | Slope = (Val - BeginVal) / AtBegin; |
856 | beta = Val - Slope * AtBegin; |
857 | |
858 | for (i=0; i < AtBegin; i++) |
859 | g ->Table16[i] = _cmsQuickSaturateWord(i * Slope + beta); |
860 | |
861 | // Compute slope and offset for the end |
862 | Val = g ->Table16[AtEnd]; |
863 | Slope = (EndVal - Val) / AtBegin; // AtBegin holds the X interval, which is same in both cases |
864 | beta = Val - Slope * AtEnd; |
865 | |
866 | for (i = AtEnd; i < (int) g ->nEntries; i++) |
867 | g ->Table16[i] = _cmsQuickSaturateWord(i * Slope + beta); |
868 | } |
869 | |
870 | |
871 | // Precomputes tables for 8-bit on input devicelink. |
872 | static |
873 | Prelin8Data* PrelinOpt8alloc(cmsContext ContextID, const cmsInterpParams* p, cmsToneCurve* G[3]) |
874 | { |
875 | int i; |
876 | cmsUInt16Number Input[3]; |
877 | cmsS15Fixed16Number v1, v2, v3; |
878 | Prelin8Data* p8; |
879 | |
880 | p8 = (Prelin8Data*)_cmsMallocZero(ContextID, sizeof(Prelin8Data)); |
881 | if (p8 == NULL) return NULL; |
882 | |
883 | // Since this only works for 8 bit input, values comes always as x * 257, |
884 | // we can safely take msb byte (x << 8 + x) |
885 | |
886 | for (i=0; i < 256; i++) { |
887 | |
888 | if (G != NULL) { |
889 | |
890 | // Get 16-bit representation |
891 | Input[0] = cmsEvalToneCurve16(G[0], FROM_8_TO_16(i)); |
892 | Input[1] = cmsEvalToneCurve16(G[1], FROM_8_TO_16(i)); |
893 | Input[2] = cmsEvalToneCurve16(G[2], FROM_8_TO_16(i)); |
894 | } |
895 | else { |
896 | Input[0] = FROM_8_TO_16(i); |
897 | Input[1] = FROM_8_TO_16(i); |
898 | Input[2] = FROM_8_TO_16(i); |
899 | } |
900 | |
901 | |
902 | // Move to 0..1.0 in fixed domain |
903 | v1 = _cmsToFixedDomain((int) (Input[0] * p -> Domain[0])); |
904 | v2 = _cmsToFixedDomain((int) (Input[1] * p -> Domain[1])); |
905 | v3 = _cmsToFixedDomain((int) (Input[2] * p -> Domain[2])); |
906 | |
907 | // Store the precalculated table of nodes |
908 | p8 ->X0[i] = (p->opta[2] * FIXED_TO_INT(v1)); |
909 | p8 ->Y0[i] = (p->opta[1] * FIXED_TO_INT(v2)); |
910 | p8 ->Z0[i] = (p->opta[0] * FIXED_TO_INT(v3)); |
911 | |
912 | // Store the precalculated table of offsets |
913 | p8 ->rx[i] = (cmsUInt16Number) FIXED_REST_TO_INT(v1); |
914 | p8 ->ry[i] = (cmsUInt16Number) FIXED_REST_TO_INT(v2); |
915 | p8 ->rz[i] = (cmsUInt16Number) FIXED_REST_TO_INT(v3); |
916 | } |
917 | |
918 | p8 ->ContextID = ContextID; |
919 | p8 ->p = p; |
920 | |
921 | return p8; |
922 | } |
923 | |
924 | static |
925 | void Prelin8free(cmsContext ContextID, void* ptr) |
926 | { |
927 | _cmsFree(ContextID, ptr); |
928 | } |
929 | |
930 | static |
931 | void* Prelin8dup(cmsContext ContextID, const void* ptr) |
932 | { |
933 | return _cmsDupMem(ContextID, ptr, sizeof(Prelin8Data)); |
934 | } |
935 | |
936 | |
937 | |
938 | // A optimized interpolation for 8-bit input. |
939 | #define DENS(i,j,k) (LutTable[(i)+(j)+(k)+OutChan]) |
940 | static |
941 | void PrelinEval8(register const cmsUInt16Number Input[], |
942 | register cmsUInt16Number Output[], |
943 | register const void* D) |
944 | { |
945 | |
946 | cmsUInt8Number r, g, b; |
947 | cmsS15Fixed16Number rx, ry, rz; |
948 | cmsS15Fixed16Number c0, c1, c2, c3, Rest; |
949 | int OutChan; |
950 | register cmsS15Fixed16Number X0, X1, Y0, Y1, Z0, Z1; |
951 | Prelin8Data* p8 = (Prelin8Data*) D; |
952 | register const cmsInterpParams* p = p8 ->p; |
953 | int TotalOut = (int) p -> nOutputs; |
954 | const cmsUInt16Number* LutTable = (const cmsUInt16Number*) p->Table; |
955 | |
956 | r = (cmsUInt8Number) (Input[0] >> 8); |
957 | g = (cmsUInt8Number) (Input[1] >> 8); |
958 | b = (cmsUInt8Number) (Input[2] >> 8); |
959 | |
960 | X0 = X1 = (cmsS15Fixed16Number) p8->X0[r]; |
961 | Y0 = Y1 = (cmsS15Fixed16Number) p8->Y0[g]; |
962 | Z0 = Z1 = (cmsS15Fixed16Number) p8->Z0[b]; |
963 | |
964 | rx = p8 ->rx[r]; |
965 | ry = p8 ->ry[g]; |
966 | rz = p8 ->rz[b]; |
967 | |
968 | X1 = X0 + (cmsS15Fixed16Number)((rx == 0) ? 0 : p ->opta[2]); |
969 | Y1 = Y0 + (cmsS15Fixed16Number)((ry == 0) ? 0 : p ->opta[1]); |
970 | Z1 = Z0 + (cmsS15Fixed16Number)((rz == 0) ? 0 : p ->opta[0]); |
971 | |
972 | |
973 | // These are the 6 Tetrahedral |
974 | for (OutChan=0; OutChan < TotalOut; OutChan++) { |
975 | |
976 | c0 = DENS(X0, Y0, Z0); |
977 | |
978 | if (rx >= ry && ry >= rz) |
979 | { |
980 | c1 = DENS(X1, Y0, Z0) - c0; |
981 | c2 = DENS(X1, Y1, Z0) - DENS(X1, Y0, Z0); |
982 | c3 = DENS(X1, Y1, Z1) - DENS(X1, Y1, Z0); |
983 | } |
984 | else |
985 | if (rx >= rz && rz >= ry) |
986 | { |
987 | c1 = DENS(X1, Y0, Z0) - c0; |
988 | c2 = DENS(X1, Y1, Z1) - DENS(X1, Y0, Z1); |
989 | c3 = DENS(X1, Y0, Z1) - DENS(X1, Y0, Z0); |
990 | } |
991 | else |
992 | if (rz >= rx && rx >= ry) |
993 | { |
994 | c1 = DENS(X1, Y0, Z1) - DENS(X0, Y0, Z1); |
995 | c2 = DENS(X1, Y1, Z1) - DENS(X1, Y0, Z1); |
996 | c3 = DENS(X0, Y0, Z1) - c0; |
997 | } |
998 | else |
999 | if (ry >= rx && rx >= rz) |
1000 | { |
1001 | c1 = DENS(X1, Y1, Z0) - DENS(X0, Y1, Z0); |
1002 | c2 = DENS(X0, Y1, Z0) - c0; |
1003 | c3 = DENS(X1, Y1, Z1) - DENS(X1, Y1, Z0); |
1004 | } |
1005 | else |
1006 | if (ry >= rz && rz >= rx) |
1007 | { |
1008 | c1 = DENS(X1, Y1, Z1) - DENS(X0, Y1, Z1); |
1009 | c2 = DENS(X0, Y1, Z0) - c0; |
1010 | c3 = DENS(X0, Y1, Z1) - DENS(X0, Y1, Z0); |
1011 | } |
1012 | else |
1013 | if (rz >= ry && ry >= rx) |
1014 | { |
1015 | c1 = DENS(X1, Y1, Z1) - DENS(X0, Y1, Z1); |
1016 | c2 = DENS(X0, Y1, Z1) - DENS(X0, Y0, Z1); |
1017 | c3 = DENS(X0, Y0, Z1) - c0; |
1018 | } |
1019 | else { |
1020 | c1 = c2 = c3 = 0; |
1021 | } |
1022 | |
1023 | Rest = c1 * rx + c2 * ry + c3 * rz + 0x8001; |
1024 | Output[OutChan] = (cmsUInt16Number) (c0 + ((Rest + (Rest >> 16)) >> 16)); |
1025 | |
1026 | } |
1027 | } |
1028 | |
1029 | #undef DENS |
1030 | |
1031 | |
1032 | // Curves that contain wide empty areas are not optimizeable |
1033 | static |
1034 | cmsBool IsDegenerated(const cmsToneCurve* g) |
1035 | { |
1036 | cmsUInt32Number i, Zeros = 0, Poles = 0; |
1037 | cmsUInt32Number nEntries = g ->nEntries; |
1038 | |
1039 | for (i=0; i < nEntries; i++) { |
1040 | |
1041 | if (g ->Table16[i] == 0x0000) Zeros++; |
1042 | if (g ->Table16[i] == 0xffff) Poles++; |
1043 | } |
1044 | |
1045 | if (Zeros == 1 && Poles == 1) return FALSE; // For linear tables |
1046 | if (Zeros > (nEntries / 20)) return TRUE; // Degenerated, many zeros |
1047 | if (Poles > (nEntries / 20)) return TRUE; // Degenerated, many poles |
1048 | |
1049 | return FALSE; |
1050 | } |
1051 | |
1052 | // -------------------------------------------------------------------------------------------------------------- |
1053 | // We need xput over here |
1054 | |
1055 | static |
1056 | cmsBool OptimizeByComputingLinearization(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags) |
1057 | { |
1058 | cmsPipeline* OriginalLut; |
1059 | cmsUInt32Number nGridPoints; |
1060 | cmsToneCurve *Trans[cmsMAXCHANNELS], *TransReverse[cmsMAXCHANNELS]; |
1061 | cmsUInt32Number t, i; |
1062 | cmsFloat32Number v, In[cmsMAXCHANNELS], Out[cmsMAXCHANNELS]; |
1063 | cmsBool lIsSuitable, lIsLinear; |
1064 | cmsPipeline* OptimizedLUT = NULL, *LutPlusCurves = NULL; |
1065 | cmsStage* OptimizedCLUTmpe; |
1066 | cmsColorSpaceSignature ColorSpace, OutputColorSpace; |
1067 | cmsStage* OptimizedPrelinMpe; |
1068 | cmsStage* mpe; |
1069 | cmsToneCurve** OptimizedPrelinCurves; |
1070 | _cmsStageCLutData* OptimizedPrelinCLUT; |
1071 | |
1072 | |
1073 | // This is a loosy optimization! does not apply in floating-point cases |
1074 | if (_cmsFormatterIsFloat(*InputFormat) || _cmsFormatterIsFloat(*OutputFormat)) return FALSE; |
1075 | |
1076 | // Only on chunky RGB |
1077 | if (T_COLORSPACE(*InputFormat) != PT_RGB) return FALSE; |
1078 | if (T_PLANAR(*InputFormat)) return FALSE; |
1079 | |
1080 | if (T_COLORSPACE(*OutputFormat) != PT_RGB) return FALSE; |
1081 | if (T_PLANAR(*OutputFormat)) return FALSE; |
1082 | |
1083 | // On 16 bits, user has to specify the feature |
1084 | if (!_cmsFormatterIs8bit(*InputFormat)) { |
1085 | if (!(*dwFlags & cmsFLAGS_CLUT_PRE_LINEARIZATION)) return FALSE; |
1086 | } |
1087 | |
1088 | OriginalLut = *Lut; |
1089 | |
1090 | // Named color pipelines cannot be optimized either |
1091 | for (mpe = cmsPipelineGetPtrToFirstStage(OriginalLut); |
1092 | mpe != NULL; |
1093 | mpe = cmsStageNext(mpe)) { |
1094 | if (cmsStageType(mpe) == cmsSigNamedColorElemType) return FALSE; |
1095 | } |
1096 | |
1097 | ColorSpace = _cmsICCcolorSpace((int) T_COLORSPACE(*InputFormat)); |
1098 | OutputColorSpace = _cmsICCcolorSpace((int) T_COLORSPACE(*OutputFormat)); |
1099 | |
1100 | // Color space must be specified |
1101 | if (ColorSpace == (cmsColorSpaceSignature)0 || |
1102 | OutputColorSpace == (cmsColorSpaceSignature)0) return FALSE; |
1103 | |
1104 | nGridPoints = _cmsReasonableGridpointsByColorspace(ColorSpace, *dwFlags); |
1105 | |
1106 | // Empty gamma containers |
1107 | memset(Trans, 0, sizeof(Trans)); |
1108 | memset(TransReverse, 0, sizeof(TransReverse)); |
1109 | |
1110 | // If the last stage of the original lut are curves, and those curves are |
1111 | // degenerated, it is likely the transform is squeezing and clipping |
1112 | // the output from previous CLUT. We cannot optimize this case |
1113 | { |
1114 | cmsStage* last = cmsPipelineGetPtrToLastStage(OriginalLut); |
1115 | |
1116 | if (cmsStageType(last) == cmsSigCurveSetElemType) { |
1117 | |
1118 | _cmsStageToneCurvesData* Data = (_cmsStageToneCurvesData*)cmsStageData(last); |
1119 | for (i = 0; i < Data->nCurves; i++) { |
1120 | if (IsDegenerated(Data->TheCurves[i])) |
1121 | goto Error; |
1122 | } |
1123 | } |
1124 | } |
1125 | |
1126 | for (t = 0; t < OriginalLut ->InputChannels; t++) { |
1127 | Trans[t] = cmsBuildTabulatedToneCurve16(OriginalLut ->ContextID, PRELINEARIZATION_POINTS, NULL); |
1128 | if (Trans[t] == NULL) goto Error; |
1129 | } |
1130 | |
1131 | // Populate the curves |
1132 | for (i=0; i < PRELINEARIZATION_POINTS; i++) { |
1133 | |
1134 | v = (cmsFloat32Number) ((cmsFloat64Number) i / (PRELINEARIZATION_POINTS - 1)); |
1135 | |
1136 | // Feed input with a gray ramp |
1137 | for (t=0; t < OriginalLut ->InputChannels; t++) |
1138 | In[t] = v; |
1139 | |
1140 | // Evaluate the gray value |
1141 | cmsPipelineEvalFloat(In, Out, OriginalLut); |
1142 | |
1143 | // Store result in curve |
1144 | for (t=0; t < OriginalLut ->InputChannels; t++) |
1145 | Trans[t] ->Table16[i] = _cmsQuickSaturateWord(Out[t] * 65535.0); |
1146 | } |
1147 | |
1148 | // Slope-limit the obtained curves |
1149 | for (t = 0; t < OriginalLut ->InputChannels; t++) |
1150 | SlopeLimiting(Trans[t]); |
1151 | |
1152 | // Check for validity |
1153 | lIsSuitable = TRUE; |
1154 | lIsLinear = TRUE; |
1155 | for (t=0; (lIsSuitable && (t < OriginalLut ->InputChannels)); t++) { |
1156 | |
1157 | // Exclude if already linear |
1158 | if (!cmsIsToneCurveLinear(Trans[t])) |
1159 | lIsLinear = FALSE; |
1160 | |
1161 | // Exclude if non-monotonic |
1162 | if (!cmsIsToneCurveMonotonic(Trans[t])) |
1163 | lIsSuitable = FALSE; |
1164 | |
1165 | if (IsDegenerated(Trans[t])) |
1166 | lIsSuitable = FALSE; |
1167 | } |
1168 | |
1169 | // If it is not suitable, just quit |
1170 | if (!lIsSuitable) goto Error; |
1171 | |
1172 | // Invert curves if possible |
1173 | for (t = 0; t < OriginalLut ->InputChannels; t++) { |
1174 | TransReverse[t] = cmsReverseToneCurveEx(PRELINEARIZATION_POINTS, Trans[t]); |
1175 | if (TransReverse[t] == NULL) goto Error; |
1176 | } |
1177 | |
1178 | // Now inset the reversed curves at the begin of transform |
1179 | LutPlusCurves = cmsPipelineDup(OriginalLut); |
1180 | if (LutPlusCurves == NULL) goto Error; |
1181 | |
1182 | if (!cmsPipelineInsertStage(LutPlusCurves, cmsAT_BEGIN, cmsStageAllocToneCurves(OriginalLut ->ContextID, OriginalLut ->InputChannels, TransReverse))) |
1183 | goto Error; |
1184 | |
1185 | // Create the result LUT |
1186 | OptimizedLUT = cmsPipelineAlloc(OriginalLut ->ContextID, OriginalLut ->InputChannels, OriginalLut ->OutputChannels); |
1187 | if (OptimizedLUT == NULL) goto Error; |
1188 | |
1189 | OptimizedPrelinMpe = cmsStageAllocToneCurves(OriginalLut ->ContextID, OriginalLut ->InputChannels, Trans); |
1190 | |
1191 | // Create and insert the curves at the beginning |
1192 | if (!cmsPipelineInsertStage(OptimizedLUT, cmsAT_BEGIN, OptimizedPrelinMpe)) |
1193 | goto Error; |
1194 | |
1195 | // Allocate the CLUT for result |
1196 | OptimizedCLUTmpe = cmsStageAllocCLut16bit(OriginalLut ->ContextID, nGridPoints, OriginalLut ->InputChannels, OriginalLut ->OutputChannels, NULL); |
1197 | |
1198 | // Add the CLUT to the destination LUT |
1199 | if (!cmsPipelineInsertStage(OptimizedLUT, cmsAT_END, OptimizedCLUTmpe)) |
1200 | goto Error; |
1201 | |
1202 | // Resample the LUT |
1203 | if (!cmsStageSampleCLut16bit(OptimizedCLUTmpe, XFormSampler16, (void*) LutPlusCurves, 0)) goto Error; |
1204 | |
1205 | // Free resources |
1206 | for (t = 0; t < OriginalLut ->InputChannels; t++) { |
1207 | |
1208 | if (Trans[t]) cmsFreeToneCurve(Trans[t]); |
1209 | if (TransReverse[t]) cmsFreeToneCurve(TransReverse[t]); |
1210 | } |
1211 | |
1212 | cmsPipelineFree(LutPlusCurves); |
1213 | |
1214 | |
1215 | OptimizedPrelinCurves = _cmsStageGetPtrToCurveSet(OptimizedPrelinMpe); |
1216 | OptimizedPrelinCLUT = (_cmsStageCLutData*) OptimizedCLUTmpe ->Data; |
1217 | |
1218 | // Set the evaluator if 8-bit |
1219 | if (_cmsFormatterIs8bit(*InputFormat)) { |
1220 | |
1221 | Prelin8Data* p8 = PrelinOpt8alloc(OptimizedLUT ->ContextID, |
1222 | OptimizedPrelinCLUT ->Params, |
1223 | OptimizedPrelinCurves); |
1224 | if (p8 == NULL) { |
1225 | cmsPipelineFree(OptimizedLUT); |
1226 | return FALSE; |
1227 | } |
1228 | |
1229 | _cmsPipelineSetOptimizationParameters(OptimizedLUT, PrelinEval8, (void*) p8, Prelin8free, Prelin8dup); |
1230 | |
1231 | } |
1232 | else |
1233 | { |
1234 | Prelin16Data* p16 = PrelinOpt16alloc(OptimizedLUT ->ContextID, |
1235 | OptimizedPrelinCLUT ->Params, |
1236 | 3, OptimizedPrelinCurves, 3, NULL); |
1237 | if (p16 == NULL) { |
1238 | cmsPipelineFree(OptimizedLUT); |
1239 | return FALSE; |
1240 | } |
1241 | |
1242 | _cmsPipelineSetOptimizationParameters(OptimizedLUT, PrelinEval16, (void*) p16, PrelinOpt16free, Prelin16dup); |
1243 | |
1244 | } |
1245 | |
1246 | // Don't fix white on absolute colorimetric |
1247 | if (Intent == INTENT_ABSOLUTE_COLORIMETRIC) |
1248 | *dwFlags |= cmsFLAGS_NOWHITEONWHITEFIXUP; |
1249 | |
1250 | if (!(*dwFlags & cmsFLAGS_NOWHITEONWHITEFIXUP)) { |
1251 | |
1252 | if (!FixWhiteMisalignment(OptimizedLUT, ColorSpace, OutputColorSpace)) { |
1253 | |
1254 | return FALSE; |
1255 | } |
1256 | } |
1257 | |
1258 | // And return the obtained LUT |
1259 | |
1260 | cmsPipelineFree(OriginalLut); |
1261 | *Lut = OptimizedLUT; |
1262 | return TRUE; |
1263 | |
1264 | Error: |
1265 | |
1266 | for (t = 0; t < OriginalLut ->InputChannels; t++) { |
1267 | |
1268 | if (Trans[t]) cmsFreeToneCurve(Trans[t]); |
1269 | if (TransReverse[t]) cmsFreeToneCurve(TransReverse[t]); |
1270 | } |
1271 | |
1272 | if (LutPlusCurves != NULL) cmsPipelineFree(LutPlusCurves); |
1273 | if (OptimizedLUT != NULL) cmsPipelineFree(OptimizedLUT); |
1274 | |
1275 | return FALSE; |
1276 | |
1277 | cmsUNUSED_PARAMETER(Intent); |
1278 | cmsUNUSED_PARAMETER(lIsLinear); |
1279 | } |
1280 | |
1281 | |
1282 | // Curves optimizer ------------------------------------------------------------------------------------------------------------------ |
1283 | |
1284 | static |
1285 | void CurvesFree(cmsContext ContextID, void* ptr) |
1286 | { |
1287 | Curves16Data* Data = (Curves16Data*) ptr; |
1288 | cmsUInt32Number i; |
1289 | |
1290 | for (i=0; i < Data -> nCurves; i++) { |
1291 | |
1292 | _cmsFree(ContextID, Data ->Curves[i]); |
1293 | } |
1294 | |
1295 | _cmsFree(ContextID, Data ->Curves); |
1296 | _cmsFree(ContextID, ptr); |
1297 | } |
1298 | |
1299 | static |
1300 | void* CurvesDup(cmsContext ContextID, const void* ptr) |
1301 | { |
1302 | Curves16Data* Data = (Curves16Data*)_cmsDupMem(ContextID, ptr, sizeof(Curves16Data)); |
1303 | cmsUInt32Number i; |
1304 | |
1305 | if (Data == NULL) return NULL; |
1306 | |
1307 | Data->Curves = (cmsUInt16Number**) _cmsDupMem(ContextID, Data->Curves, Data->nCurves * sizeof(cmsUInt16Number*)); |
1308 | |
1309 | for (i=0; i < Data -> nCurves; i++) { |
1310 | Data->Curves[i] = (cmsUInt16Number*) _cmsDupMem(ContextID, Data->Curves[i], Data->nElements * sizeof(cmsUInt16Number)); |
1311 | } |
1312 | |
1313 | return (void*) Data; |
1314 | } |
1315 | |
1316 | // Precomputes tables for 8-bit on input devicelink. |
1317 | static |
1318 | Curves16Data* CurvesAlloc(cmsContext ContextID, cmsUInt32Number nCurves, cmsUInt32Number nElements, cmsToneCurve** G) |
1319 | { |
1320 | cmsUInt32Number i, j; |
1321 | Curves16Data* c16; |
1322 | |
1323 | c16 = (Curves16Data*)_cmsMallocZero(ContextID, sizeof(Curves16Data)); |
1324 | if (c16 == NULL) return NULL; |
1325 | |
1326 | c16 ->nCurves = nCurves; |
1327 | c16 ->nElements = nElements; |
1328 | |
1329 | c16->Curves = (cmsUInt16Number**) _cmsCalloc(ContextID, nCurves, sizeof(cmsUInt16Number*)); |
1330 | if (c16->Curves == NULL) { |
1331 | _cmsFree(ContextID, c16); |
1332 | return NULL; |
1333 | } |
1334 | |
1335 | for (i=0; i < nCurves; i++) { |
1336 | |
1337 | c16->Curves[i] = (cmsUInt16Number*) _cmsCalloc(ContextID, nElements, sizeof(cmsUInt16Number)); |
1338 | |
1339 | if (c16->Curves[i] == NULL) { |
1340 | |
1341 | for (j=0; j < i; j++) { |
1342 | _cmsFree(ContextID, c16->Curves[j]); |
1343 | } |
1344 | _cmsFree(ContextID, c16->Curves); |
1345 | _cmsFree(ContextID, c16); |
1346 | return NULL; |
1347 | } |
1348 | |
1349 | if (nElements == 256U) { |
1350 | |
1351 | for (j=0; j < nElements; j++) { |
1352 | |
1353 | c16 ->Curves[i][j] = cmsEvalToneCurve16(G[i], FROM_8_TO_16(j)); |
1354 | } |
1355 | } |
1356 | else { |
1357 | |
1358 | for (j=0; j < nElements; j++) { |
1359 | c16 ->Curves[i][j] = cmsEvalToneCurve16(G[i], (cmsUInt16Number) j); |
1360 | } |
1361 | } |
1362 | } |
1363 | |
1364 | return c16; |
1365 | } |
1366 | |
1367 | static |
1368 | void FastEvaluateCurves8(register const cmsUInt16Number In[], |
1369 | register cmsUInt16Number Out[], |
1370 | register const void* D) |
1371 | { |
1372 | Curves16Data* Data = (Curves16Data*) D; |
1373 | int x; |
1374 | cmsUInt32Number i; |
1375 | |
1376 | for (i=0; i < Data ->nCurves; i++) { |
1377 | |
1378 | x = (In[i] >> 8); |
1379 | Out[i] = Data -> Curves[i][x]; |
1380 | } |
1381 | } |
1382 | |
1383 | |
1384 | static |
1385 | void FastEvaluateCurves16(register const cmsUInt16Number In[], |
1386 | register cmsUInt16Number Out[], |
1387 | register const void* D) |
1388 | { |
1389 | Curves16Data* Data = (Curves16Data*) D; |
1390 | cmsUInt32Number i; |
1391 | |
1392 | for (i=0; i < Data ->nCurves; i++) { |
1393 | Out[i] = Data -> Curves[i][In[i]]; |
1394 | } |
1395 | } |
1396 | |
1397 | |
1398 | static |
1399 | void FastIdentity16(register const cmsUInt16Number In[], |
1400 | register cmsUInt16Number Out[], |
1401 | register const void* D) |
1402 | { |
1403 | cmsPipeline* Lut = (cmsPipeline*) D; |
1404 | cmsUInt32Number i; |
1405 | |
1406 | for (i=0; i < Lut ->InputChannels; i++) { |
1407 | Out[i] = In[i]; |
1408 | } |
1409 | } |
1410 | |
1411 | |
1412 | // If the target LUT holds only curves, the optimization procedure is to join all those |
1413 | // curves together. That only works on curves and does not work on matrices. |
1414 | static |
1415 | cmsBool OptimizeByJoiningCurves(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags) |
1416 | { |
1417 | cmsToneCurve** GammaTables = NULL; |
1418 | cmsFloat32Number InFloat[cmsMAXCHANNELS], OutFloat[cmsMAXCHANNELS]; |
1419 | cmsUInt32Number i, j; |
1420 | cmsPipeline* Src = *Lut; |
1421 | cmsPipeline* Dest = NULL; |
1422 | cmsStage* mpe; |
1423 | cmsStage* ObtainedCurves = NULL; |
1424 | |
1425 | |
1426 | // This is a loosy optimization! does not apply in floating-point cases |
1427 | if (_cmsFormatterIsFloat(*InputFormat) || _cmsFormatterIsFloat(*OutputFormat)) return FALSE; |
1428 | |
1429 | // Only curves in this LUT? |
1430 | for (mpe = cmsPipelineGetPtrToFirstStage(Src); |
1431 | mpe != NULL; |
1432 | mpe = cmsStageNext(mpe)) { |
1433 | if (cmsStageType(mpe) != cmsSigCurveSetElemType) return FALSE; |
1434 | } |
1435 | |
1436 | // Allocate an empty LUT |
1437 | Dest = cmsPipelineAlloc(Src ->ContextID, Src ->InputChannels, Src ->OutputChannels); |
1438 | if (Dest == NULL) return FALSE; |
1439 | |
1440 | // Create target curves |
1441 | GammaTables = (cmsToneCurve**) _cmsCalloc(Src ->ContextID, Src ->InputChannels, sizeof(cmsToneCurve*)); |
1442 | if (GammaTables == NULL) goto Error; |
1443 | |
1444 | for (i=0; i < Src ->InputChannels; i++) { |
1445 | GammaTables[i] = cmsBuildTabulatedToneCurve16(Src ->ContextID, PRELINEARIZATION_POINTS, NULL); |
1446 | if (GammaTables[i] == NULL) goto Error; |
1447 | } |
1448 | |
1449 | // Compute 16 bit result by using floating point |
1450 | for (i=0; i < PRELINEARIZATION_POINTS; i++) { |
1451 | |
1452 | for (j=0; j < Src ->InputChannels; j++) |
1453 | InFloat[j] = (cmsFloat32Number) ((cmsFloat64Number) i / (PRELINEARIZATION_POINTS - 1)); |
1454 | |
1455 | cmsPipelineEvalFloat(InFloat, OutFloat, Src); |
1456 | |
1457 | for (j=0; j < Src ->InputChannels; j++) |
1458 | GammaTables[j] -> Table16[i] = _cmsQuickSaturateWord(OutFloat[j] * 65535.0); |
1459 | } |
1460 | |
1461 | ObtainedCurves = cmsStageAllocToneCurves(Src ->ContextID, Src ->InputChannels, GammaTables); |
1462 | if (ObtainedCurves == NULL) goto Error; |
1463 | |
1464 | for (i=0; i < Src ->InputChannels; i++) { |
1465 | cmsFreeToneCurve(GammaTables[i]); |
1466 | GammaTables[i] = NULL; |
1467 | } |
1468 | |
1469 | if (GammaTables != NULL) { |
1470 | _cmsFree(Src->ContextID, GammaTables); |
1471 | GammaTables = NULL; |
1472 | } |
1473 | |
1474 | // Maybe the curves are linear at the end |
1475 | if (!AllCurvesAreLinear(ObtainedCurves)) { |
1476 | |
1477 | if (!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, ObtainedCurves)) |
1478 | goto Error; |
1479 | |
1480 | // If the curves are to be applied in 8 bits, we can save memory |
1481 | if (_cmsFormatterIs8bit(*InputFormat)) { |
1482 | |
1483 | _cmsStageToneCurvesData* Data = (_cmsStageToneCurvesData*) ObtainedCurves ->Data; |
1484 | Curves16Data* c16 = CurvesAlloc(Dest ->ContextID, Data ->nCurves, 256, Data ->TheCurves); |
1485 | |
1486 | if (c16 == NULL) goto Error; |
1487 | *dwFlags |= cmsFLAGS_NOCACHE; |
1488 | _cmsPipelineSetOptimizationParameters(Dest, FastEvaluateCurves8, c16, CurvesFree, CurvesDup); |
1489 | |
1490 | } |
1491 | else { |
1492 | |
1493 | _cmsStageToneCurvesData* Data = (_cmsStageToneCurvesData*) cmsStageData(ObtainedCurves); |
1494 | Curves16Data* c16 = CurvesAlloc(Dest ->ContextID, Data ->nCurves, 65536, Data ->TheCurves); |
1495 | |
1496 | if (c16 == NULL) goto Error; |
1497 | *dwFlags |= cmsFLAGS_NOCACHE; |
1498 | _cmsPipelineSetOptimizationParameters(Dest, FastEvaluateCurves16, c16, CurvesFree, CurvesDup); |
1499 | } |
1500 | } |
1501 | else { |
1502 | |
1503 | // LUT optimizes to nothing. Set the identity LUT |
1504 | cmsStageFree(ObtainedCurves); |
1505 | ObtainedCurves = NULL; |
1506 | |
1507 | if (!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, cmsStageAllocIdentity(Dest ->ContextID, Src ->InputChannels))) |
1508 | goto Error; |
1509 | |
1510 | *dwFlags |= cmsFLAGS_NOCACHE; |
1511 | _cmsPipelineSetOptimizationParameters(Dest, FastIdentity16, (void*) Dest, NULL, NULL); |
1512 | } |
1513 | |
1514 | // We are done. |
1515 | cmsPipelineFree(Src); |
1516 | *Lut = Dest; |
1517 | return TRUE; |
1518 | |
1519 | Error: |
1520 | |
1521 | if (ObtainedCurves != NULL) cmsStageFree(ObtainedCurves); |
1522 | if (GammaTables != NULL) { |
1523 | for (i=0; i < Src ->InputChannels; i++) { |
1524 | if (GammaTables[i] != NULL) cmsFreeToneCurve(GammaTables[i]); |
1525 | } |
1526 | |
1527 | _cmsFree(Src ->ContextID, GammaTables); |
1528 | } |
1529 | |
1530 | if (Dest != NULL) cmsPipelineFree(Dest); |
1531 | return FALSE; |
1532 | |
1533 | cmsUNUSED_PARAMETER(Intent); |
1534 | cmsUNUSED_PARAMETER(InputFormat); |
1535 | cmsUNUSED_PARAMETER(OutputFormat); |
1536 | cmsUNUSED_PARAMETER(dwFlags); |
1537 | } |
1538 | |
1539 | // ------------------------------------------------------------------------------------------------------------------------------------- |
1540 | // LUT is Shaper - Matrix - Matrix - Shaper, which is very frequent when combining two matrix-shaper profiles |
1541 | |
1542 | |
1543 | static |
1544 | void FreeMatShaper(cmsContext ContextID, void* Data) |
1545 | { |
1546 | if (Data != NULL) _cmsFree(ContextID, Data); |
1547 | } |
1548 | |
1549 | static |
1550 | void* DupMatShaper(cmsContext ContextID, const void* Data) |
1551 | { |
1552 | return _cmsDupMem(ContextID, Data, sizeof(MatShaper8Data)); |
1553 | } |
1554 | |
1555 | |
1556 | // A fast matrix-shaper evaluator for 8 bits. This is a bit ticky since I'm using 1.14 signed fixed point |
1557 | // to accomplish some performance. Actually it takes 256x3 16 bits tables and 16385 x 3 tables of 8 bits, |
1558 | // in total about 50K, and the performance boost is huge! |
1559 | static |
1560 | void MatShaperEval16(register const cmsUInt16Number In[], |
1561 | register cmsUInt16Number Out[], |
1562 | register const void* D) |
1563 | { |
1564 | MatShaper8Data* p = (MatShaper8Data*) D; |
1565 | cmsS1Fixed14Number l1, l2, l3, r, g, b; |
1566 | cmsUInt32Number ri, gi, bi; |
1567 | |
1568 | // In this case (and only in this case!) we can use this simplification since |
1569 | // In[] is assured to come from a 8 bit number. (a << 8 | a) |
1570 | ri = In[0] & 0xFFU; |
1571 | gi = In[1] & 0xFFU; |
1572 | bi = In[2] & 0xFFU; |
1573 | |
1574 | // Across first shaper, which also converts to 1.14 fixed point |
1575 | r = p->Shaper1R[ri]; |
1576 | g = p->Shaper1G[gi]; |
1577 | b = p->Shaper1B[bi]; |
1578 | |
1579 | // Evaluate the matrix in 1.14 fixed point |
1580 | l1 = (p->Mat[0][0] * r + p->Mat[0][1] * g + p->Mat[0][2] * b + p->Off[0] + 0x2000) >> 14; |
1581 | l2 = (p->Mat[1][0] * r + p->Mat[1][1] * g + p->Mat[1][2] * b + p->Off[1] + 0x2000) >> 14; |
1582 | l3 = (p->Mat[2][0] * r + p->Mat[2][1] * g + p->Mat[2][2] * b + p->Off[2] + 0x2000) >> 14; |
1583 | |
1584 | // Now we have to clip to 0..1.0 range |
1585 | ri = (l1 < 0) ? 0 : ((l1 > 16384) ? 16384U : (cmsUInt32Number) l1); |
1586 | gi = (l2 < 0) ? 0 : ((l2 > 16384) ? 16384U : (cmsUInt32Number) l2); |
1587 | bi = (l3 < 0) ? 0 : ((l3 > 16384) ? 16384U : (cmsUInt32Number) l3); |
1588 | |
1589 | // And across second shaper, |
1590 | Out[0] = p->Shaper2R[ri]; |
1591 | Out[1] = p->Shaper2G[gi]; |
1592 | Out[2] = p->Shaper2B[bi]; |
1593 | |
1594 | } |
1595 | |
1596 | // This table converts from 8 bits to 1.14 after applying the curve |
1597 | static |
1598 | void FillFirstShaper(cmsS1Fixed14Number* Table, cmsToneCurve* Curve) |
1599 | { |
1600 | int i; |
1601 | cmsFloat32Number R, y; |
1602 | |
1603 | for (i=0; i < 256; i++) { |
1604 | |
1605 | R = (cmsFloat32Number) (i / 255.0); |
1606 | y = cmsEvalToneCurveFloat(Curve, R); |
1607 | |
1608 | if (y < 131072.0) |
1609 | Table[i] = DOUBLE_TO_1FIXED14(y); |
1610 | else |
1611 | Table[i] = 0x7fffffff; |
1612 | } |
1613 | } |
1614 | |
1615 | // This table converts form 1.14 (being 0x4000 the last entry) to 8 bits after applying the curve |
1616 | static |
1617 | void FillSecondShaper(cmsUInt16Number* Table, cmsToneCurve* Curve, cmsBool Is8BitsOutput) |
1618 | { |
1619 | int i; |
1620 | cmsFloat32Number R, Val; |
1621 | |
1622 | for (i=0; i < 16385; i++) { |
1623 | |
1624 | R = (cmsFloat32Number) (i / 16384.0); |
1625 | Val = cmsEvalToneCurveFloat(Curve, R); // Val comes 0..1.0 |
1626 | |
1627 | if (Val < 0) |
1628 | Val = 0; |
1629 | |
1630 | if (Val > 1.0) |
1631 | Val = 1.0; |
1632 | |
1633 | if (Is8BitsOutput) { |
1634 | |
1635 | // If 8 bits output, we can optimize further by computing the / 257 part. |
1636 | // first we compute the resulting byte and then we store the byte times |
1637 | // 257. This quantization allows to round very quick by doing a >> 8, but |
1638 | // since the low byte is always equal to msb, we can do a & 0xff and this works! |
1639 | cmsUInt16Number w = _cmsQuickSaturateWord(Val * 65535.0); |
1640 | cmsUInt8Number b = FROM_16_TO_8(w); |
1641 | |
1642 | Table[i] = FROM_8_TO_16(b); |
1643 | } |
1644 | else Table[i] = _cmsQuickSaturateWord(Val * 65535.0); |
1645 | } |
1646 | } |
1647 | |
1648 | // Compute the matrix-shaper structure |
1649 | static |
1650 | cmsBool SetMatShaper(cmsPipeline* Dest, cmsToneCurve* Curve1[3], cmsMAT3* Mat, cmsVEC3* Off, cmsToneCurve* Curve2[3], cmsUInt32Number* OutputFormat) |
1651 | { |
1652 | MatShaper8Data* p; |
1653 | int i, j; |
1654 | cmsBool Is8Bits = _cmsFormatterIs8bit(*OutputFormat); |
1655 | |
1656 | // Allocate a big chuck of memory to store precomputed tables |
1657 | p = (MatShaper8Data*) _cmsMalloc(Dest ->ContextID, sizeof(MatShaper8Data)); |
1658 | if (p == NULL) return FALSE; |
1659 | |
1660 | p -> ContextID = Dest -> ContextID; |
1661 | |
1662 | // Precompute tables |
1663 | FillFirstShaper(p ->Shaper1R, Curve1[0]); |
1664 | FillFirstShaper(p ->Shaper1G, Curve1[1]); |
1665 | FillFirstShaper(p ->Shaper1B, Curve1[2]); |
1666 | |
1667 | FillSecondShaper(p ->Shaper2R, Curve2[0], Is8Bits); |
1668 | FillSecondShaper(p ->Shaper2G, Curve2[1], Is8Bits); |
1669 | FillSecondShaper(p ->Shaper2B, Curve2[2], Is8Bits); |
1670 | |
1671 | // Convert matrix to nFixed14. Note that those values may take more than 16 bits |
1672 | for (i=0; i < 3; i++) { |
1673 | for (j=0; j < 3; j++) { |
1674 | p ->Mat[i][j] = DOUBLE_TO_1FIXED14(Mat->v[i].n[j]); |
1675 | } |
1676 | } |
1677 | |
1678 | for (i=0; i < 3; i++) { |
1679 | |
1680 | if (Off == NULL) { |
1681 | p ->Off[i] = 0; |
1682 | } |
1683 | else { |
1684 | p ->Off[i] = DOUBLE_TO_1FIXED14(Off->n[i]); |
1685 | } |
1686 | } |
1687 | |
1688 | // Mark as optimized for faster formatter |
1689 | if (Is8Bits) |
1690 | *OutputFormat |= OPTIMIZED_SH(1); |
1691 | |
1692 | // Fill function pointers |
1693 | _cmsPipelineSetOptimizationParameters(Dest, MatShaperEval16, (void*) p, FreeMatShaper, DupMatShaper); |
1694 | return TRUE; |
1695 | } |
1696 | |
1697 | // 8 bits on input allows matrix-shaper boot up to 25 Mpixels per second on RGB. That's fast! |
1698 | static |
1699 | cmsBool OptimizeMatrixShaper(cmsPipeline** Lut, cmsUInt32Number Intent, cmsUInt32Number* InputFormat, cmsUInt32Number* OutputFormat, cmsUInt32Number* dwFlags) |
1700 | { |
1701 | cmsStage* Curve1, *Curve2; |
1702 | cmsStage* Matrix1, *Matrix2; |
1703 | cmsMAT3 res; |
1704 | cmsBool IdentityMat; |
1705 | cmsPipeline* Dest, *Src; |
1706 | cmsFloat64Number* Offset; |
1707 | |
1708 | // Only works on RGB to RGB |
1709 | if (T_CHANNELS(*InputFormat) != 3 || T_CHANNELS(*OutputFormat) != 3) return FALSE; |
1710 | |
1711 | // Only works on 8 bit input |
1712 | if (!_cmsFormatterIs8bit(*InputFormat)) return FALSE; |
1713 | |
1714 | // Seems suitable, proceed |
1715 | Src = *Lut; |
1716 | |
1717 | // Check for: |
1718 | // |
1719 | // shaper-matrix-matrix-shaper |
1720 | // shaper-matrix-shaper |
1721 | // |
1722 | // Both of those constructs are possible (first because abs. colorimetric). |
1723 | // additionally, In the first case, the input matrix offset should be zero. |
1724 | |
1725 | IdentityMat = FALSE; |
1726 | if (cmsPipelineCheckAndRetreiveStages(Src, 4, |
1727 | cmsSigCurveSetElemType, cmsSigMatrixElemType, cmsSigMatrixElemType, cmsSigCurveSetElemType, |
1728 | &Curve1, &Matrix1, &Matrix2, &Curve2)) { |
1729 | |
1730 | // Get both matrices |
1731 | _cmsStageMatrixData* Data1 = (_cmsStageMatrixData*)cmsStageData(Matrix1); |
1732 | _cmsStageMatrixData* Data2 = (_cmsStageMatrixData*)cmsStageData(Matrix2); |
1733 | |
1734 | // Input offset should be zero |
1735 | if (Data1->Offset != NULL) return FALSE; |
1736 | |
1737 | // Multiply both matrices to get the result |
1738 | _cmsMAT3per(&res, (cmsMAT3*)Data2->Double, (cmsMAT3*)Data1->Double); |
1739 | |
1740 | // Only 2nd matrix has offset, or it is zero |
1741 | Offset = Data2->Offset; |
1742 | |
1743 | // Now the result is in res + Data2 -> Offset. Maybe is a plain identity? |
1744 | if (_cmsMAT3isIdentity(&res) && Offset == NULL) { |
1745 | |
1746 | // We can get rid of full matrix |
1747 | IdentityMat = TRUE; |
1748 | } |
1749 | |
1750 | } |
1751 | else { |
1752 | |
1753 | if (cmsPipelineCheckAndRetreiveStages(Src, 3, |
1754 | cmsSigCurveSetElemType, cmsSigMatrixElemType, cmsSigCurveSetElemType, |
1755 | &Curve1, &Matrix1, &Curve2)) { |
1756 | |
1757 | _cmsStageMatrixData* Data = (_cmsStageMatrixData*)cmsStageData(Matrix1); |
1758 | |
1759 | // Copy the matrix to our result |
1760 | memcpy(&res, Data->Double, sizeof(res)); |
1761 | |
1762 | // Preserve the Odffset (may be NULL as a zero offset) |
1763 | Offset = Data->Offset; |
1764 | |
1765 | if (_cmsMAT3isIdentity(&res) && Offset == NULL) { |
1766 | |
1767 | // We can get rid of full matrix |
1768 | IdentityMat = TRUE; |
1769 | } |
1770 | } |
1771 | else |
1772 | return FALSE; // Not optimizeable this time |
1773 | |
1774 | } |
1775 | |
1776 | // Allocate an empty LUT |
1777 | Dest = cmsPipelineAlloc(Src ->ContextID, Src ->InputChannels, Src ->OutputChannels); |
1778 | if (!Dest) return FALSE; |
1779 | |
1780 | // Assamble the new LUT |
1781 | if (!cmsPipelineInsertStage(Dest, cmsAT_BEGIN, cmsStageDup(Curve1))) |
1782 | goto Error; |
1783 | |
1784 | if (!IdentityMat) { |
1785 | |
1786 | if (!cmsPipelineInsertStage(Dest, cmsAT_END, cmsStageAllocMatrix(Dest->ContextID, 3, 3, (const cmsFloat64Number*)&res, Offset))) |
1787 | goto Error; |
1788 | } |
1789 | |
1790 | if (!cmsPipelineInsertStage(Dest, cmsAT_END, cmsStageDup(Curve2))) |
1791 | goto Error; |
1792 | |
1793 | // If identity on matrix, we can further optimize the curves, so call the join curves routine |
1794 | if (IdentityMat) { |
1795 | |
1796 | OptimizeByJoiningCurves(&Dest, Intent, InputFormat, OutputFormat, dwFlags); |
1797 | } |
1798 | else { |
1799 | _cmsStageToneCurvesData* mpeC1 = (_cmsStageToneCurvesData*) cmsStageData(Curve1); |
1800 | _cmsStageToneCurvesData* mpeC2 = (_cmsStageToneCurvesData*) cmsStageData(Curve2); |
1801 | |
1802 | // In this particular optimization, caché does not help as it takes more time to deal with |
1803 | // the caché that with the pixel handling |
1804 | *dwFlags |= cmsFLAGS_NOCACHE; |
1805 | |
1806 | // Setup the optimizarion routines |
1807 | SetMatShaper(Dest, mpeC1 ->TheCurves, &res, (cmsVEC3*) Offset, mpeC2->TheCurves, OutputFormat); |
1808 | } |
1809 | |
1810 | cmsPipelineFree(Src); |
1811 | *Lut = Dest; |
1812 | return TRUE; |
1813 | Error: |
1814 | // Leave Src unchanged |
1815 | cmsPipelineFree(Dest); |
1816 | return FALSE; |
1817 | } |
1818 | |
1819 | |
1820 | // ------------------------------------------------------------------------------------------------------------------------------------- |
1821 | // Optimization plug-ins |
1822 | |
1823 | // List of optimizations |
1824 | typedef struct _cmsOptimizationCollection_st { |
1825 | |
1826 | _cmsOPToptimizeFn OptimizePtr; |
1827 | |
1828 | struct _cmsOptimizationCollection_st *Next; |
1829 | |
1830 | } _cmsOptimizationCollection; |
1831 | |
1832 | |
1833 | // The built-in list. We currently implement 4 types of optimizations. Joining of curves, matrix-shaper, linearization and resampling |
1834 | static _cmsOptimizationCollection DefaultOptimization[] = { |
1835 | |
1836 | { OptimizeByJoiningCurves, &DefaultOptimization[1] }, |
1837 | { OptimizeMatrixShaper, &DefaultOptimization[2] }, |
1838 | { OptimizeByComputingLinearization, &DefaultOptimization[3] }, |
1839 | { OptimizeByResampling, NULL } |
1840 | }; |
1841 | |
1842 | // The linked list head |
1843 | _cmsOptimizationPluginChunkType _cmsOptimizationPluginChunk = { NULL }; |
1844 | |
1845 | |
1846 | // Duplicates the zone of memory used by the plug-in in the new context |
1847 | static |
1848 | void DupPluginOptimizationList(struct _cmsContext_struct* ctx, |
1849 | const struct _cmsContext_struct* src) |
1850 | { |
1851 | _cmsOptimizationPluginChunkType newHead = { NULL }; |
1852 | _cmsOptimizationCollection* entry; |
1853 | _cmsOptimizationCollection* Anterior = NULL; |
1854 | _cmsOptimizationPluginChunkType* head = (_cmsOptimizationPluginChunkType*) src->chunks[OptimizationPlugin]; |
1855 | |
1856 | _cmsAssert(ctx != NULL); |
1857 | _cmsAssert(head != NULL); |
1858 | |
1859 | // Walk the list copying all nodes |
1860 | for (entry = head->OptimizationCollection; |
1861 | entry != NULL; |
1862 | entry = entry ->Next) { |
1863 | |
1864 | _cmsOptimizationCollection *newEntry = ( _cmsOptimizationCollection *) _cmsSubAllocDup(ctx ->MemPool, entry, sizeof(_cmsOptimizationCollection)); |
1865 | |
1866 | if (newEntry == NULL) |
1867 | return; |
1868 | |
1869 | // We want to keep the linked list order, so this is a little bit tricky |
1870 | newEntry -> Next = NULL; |
1871 | if (Anterior) |
1872 | Anterior -> Next = newEntry; |
1873 | |
1874 | Anterior = newEntry; |
1875 | |
1876 | if (newHead.OptimizationCollection == NULL) |
1877 | newHead.OptimizationCollection = newEntry; |
1878 | } |
1879 | |
1880 | ctx ->chunks[OptimizationPlugin] = _cmsSubAllocDup(ctx->MemPool, &newHead, sizeof(_cmsOptimizationPluginChunkType)); |
1881 | } |
1882 | |
1883 | void _cmsAllocOptimizationPluginChunk(struct _cmsContext_struct* ctx, |
1884 | const struct _cmsContext_struct* src) |
1885 | { |
1886 | if (src != NULL) { |
1887 | |
1888 | // Copy all linked list |
1889 | DupPluginOptimizationList(ctx, src); |
1890 | } |
1891 | else { |
1892 | static _cmsOptimizationPluginChunkType OptimizationPluginChunkType = { NULL }; |
1893 | ctx ->chunks[OptimizationPlugin] = _cmsSubAllocDup(ctx ->MemPool, &OptimizationPluginChunkType, sizeof(_cmsOptimizationPluginChunkType)); |
1894 | } |
1895 | } |
1896 | |
1897 | |
1898 | // Register new ways to optimize |
1899 | cmsBool _cmsRegisterOptimizationPlugin(cmsContext ContextID, cmsPluginBase* Data) |
1900 | { |
1901 | cmsPluginOptimization* Plugin = (cmsPluginOptimization*) Data; |
1902 | _cmsOptimizationPluginChunkType* ctx = ( _cmsOptimizationPluginChunkType*) _cmsContextGetClientChunk(ContextID, OptimizationPlugin); |
1903 | _cmsOptimizationCollection* fl; |
1904 | |
1905 | if (Data == NULL) { |
1906 | |
1907 | ctx->OptimizationCollection = NULL; |
1908 | return TRUE; |
1909 | } |
1910 | |
1911 | // Optimizer callback is required |
1912 | if (Plugin ->OptimizePtr == NULL) return FALSE; |
1913 | |
1914 | fl = (_cmsOptimizationCollection*) _cmsPluginMalloc(ContextID, sizeof(_cmsOptimizationCollection)); |
1915 | if (fl == NULL) return FALSE; |
1916 | |
1917 | // Copy the parameters |
1918 | fl ->OptimizePtr = Plugin ->OptimizePtr; |
1919 | |
1920 | // Keep linked list |
1921 | fl ->Next = ctx->OptimizationCollection; |
1922 | |
1923 | // Set the head |
1924 | ctx ->OptimizationCollection = fl; |
1925 | |
1926 | // All is ok |
1927 | return TRUE; |
1928 | } |
1929 | |
1930 | // The entry point for LUT optimization |
1931 | cmsBool _cmsOptimizePipeline(cmsContext ContextID, |
1932 | cmsPipeline** PtrLut, |
1933 | cmsUInt32Number Intent, |
1934 | cmsUInt32Number* InputFormat, |
1935 | cmsUInt32Number* OutputFormat, |
1936 | cmsUInt32Number* dwFlags) |
1937 | { |
1938 | _cmsOptimizationPluginChunkType* ctx = ( _cmsOptimizationPluginChunkType*) _cmsContextGetClientChunk(ContextID, OptimizationPlugin); |
1939 | _cmsOptimizationCollection* Opts; |
1940 | cmsBool AnySuccess = FALSE; |
1941 | |
1942 | // A CLUT is being asked, so force this specific optimization |
1943 | if (*dwFlags & cmsFLAGS_FORCE_CLUT) { |
1944 | |
1945 | PreOptimize(*PtrLut); |
1946 | return OptimizeByResampling(PtrLut, Intent, InputFormat, OutputFormat, dwFlags); |
1947 | } |
1948 | |
1949 | // Anything to optimize? |
1950 | if ((*PtrLut) ->Elements == NULL) { |
1951 | _cmsPipelineSetOptimizationParameters(*PtrLut, FastIdentity16, (void*) *PtrLut, NULL, NULL); |
1952 | return TRUE; |
1953 | } |
1954 | |
1955 | // Try to get rid of identities and trivial conversions. |
1956 | AnySuccess = PreOptimize(*PtrLut); |
1957 | |
1958 | // After removal do we end with an identity? |
1959 | if ((*PtrLut) ->Elements == NULL) { |
1960 | _cmsPipelineSetOptimizationParameters(*PtrLut, FastIdentity16, (void*) *PtrLut, NULL, NULL); |
1961 | return TRUE; |
1962 | } |
1963 | |
1964 | // Do not optimize, keep all precision |
1965 | if (*dwFlags & cmsFLAGS_NOOPTIMIZE) |
1966 | return FALSE; |
1967 | |
1968 | // Try plug-in optimizations |
1969 | for (Opts = ctx->OptimizationCollection; |
1970 | Opts != NULL; |
1971 | Opts = Opts ->Next) { |
1972 | |
1973 | // If one schema succeeded, we are done |
1974 | if (Opts ->OptimizePtr(PtrLut, Intent, InputFormat, OutputFormat, dwFlags)) { |
1975 | |
1976 | return TRUE; // Optimized! |
1977 | } |
1978 | } |
1979 | |
1980 | // Try built-in optimizations |
1981 | for (Opts = DefaultOptimization; |
1982 | Opts != NULL; |
1983 | Opts = Opts ->Next) { |
1984 | |
1985 | if (Opts ->OptimizePtr(PtrLut, Intent, InputFormat, OutputFormat, dwFlags)) { |
1986 | |
1987 | return TRUE; |
1988 | } |
1989 | } |
1990 | |
1991 | // Only simple optimizations succeeded |
1992 | return AnySuccess; |
1993 | } |
1994 | |
1995 | |
1996 | |
1997 | |