-
Notifications
You must be signed in to change notification settings - Fork 166
/
xbrz.cpp
1293 lines (1060 loc) · 41.2 KB
/
xbrz.cpp
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
// ****************************************************************************
// * This file is part of the HqMAME project. It is distributed under *
// * GNU General Public License: http://www.gnu.org/licenses/gpl.html *
// * Copyright (C) Zenju (zenju AT gmx DOT de) - All Rights Reserved *
// * *
// * Additionally and as a special exception, the author gives permission *
// * to link the code of this program with the MAME library (or with modified *
// * versions of MAME that use the same license as MAME), and distribute *
// * linked combinations including the two. You must obey the GNU General *
// * Public License in all respects for all of the code used other than MAME. *
// * If you modify this file, you may extend this exception to your version *
// * of the file, but you are not obligated to do so. If you do not wish to *
// * do so, delete this exception statement from your version. *
// ****************************************************************************
#include "xbrz.h"
#include <cmath>
#include <cassert>
#include <algorithm>
#include <limits>
#include <vector>
namespace
{
template <uint32_t N> inline
unsigned char getByte(uint32_t val) { return static_cast<unsigned char>((val >> (8 * N)) & 0xff); }
// adjusted for RGBA
// - Durante
inline unsigned char getRed (uint32_t val) { return getByte<0>(val); }
inline unsigned char getGreen(uint32_t val) { return getByte<1>(val); }
inline unsigned char getBlue (uint32_t val) { return getByte<2>(val); }
inline unsigned char getAlpha(uint32_t val) { return getByte<3>(val); }
template <class T> inline
T abs(T value)
{
static_assert(std::numeric_limits<T>::is_signed, "abs() requires signed types");
return value < 0 ? -value : value;
}
const uint32_t redMask = 0x00ff0000;
const uint32_t greenMask = 0x0000ff00;
const uint32_t blueMask = 0x000000ff;
template <unsigned int N, unsigned int M> inline
void alphaBlend(uint32_t& dst, uint32_t col) //blend color over destination with opacity N / M
{
static_assert(N < 256, "possible overflow of (col & redMask) * N");
static_assert(M < 256, "possible overflow of (col & redMask ) * N + (dst & redMask ) * (M - N)");
static_assert(0 < N && N < M, "");
//dst = (redMask & ((col & redMask ) * N + (dst & redMask ) * (M - N)) / M) | //this works because 8 upper bits are free
// (greenMask & ((col & greenMask) * N + (dst & greenMask) * (M - N)) / M) |
// (blueMask & ((col & blueMask ) * N + (dst & blueMask ) * (M - N)) / M);
// the upper 8 bits are not free in our case, so we need to do this differently
// could probably be MUCH faster
// - Durante
uint8_t a = (((col ) >> 24) * N + ((dst ) >> 24) * (M - N) ) / M;
uint8_t r = (((col & redMask) >> 16) * N + ((dst & redMask) >> 16) * (M - N) ) / M;
uint8_t g = (((col & greenMask) >> 8) * N + ((dst & greenMask) >> 8) * (M - N) ) / M;
uint8_t b = (((col & blueMask) ) * N + ((dst & blueMask) ) * (M - N) ) / M;
dst = (a << 24) | (r << 16) | (g << 8) | (b << 0);
}
uint32_t* byteAdvance( uint32_t* ptr, int bytes) { return reinterpret_cast< uint32_t*>(reinterpret_cast< char*>(ptr) + bytes); }
const uint32_t* byteAdvance(const uint32_t* ptr, int bytes) { return reinterpret_cast<const uint32_t*>(reinterpret_cast<const char*>(ptr) + bytes); }
//fill block with the given color
inline
void fillBlock(uint32_t* trg, int pitch, uint32_t col, int blockWidth, int blockHeight)
{
//for (int y = 0; y < blockHeight; ++y, trg = byteAdvance(trg, pitch))
// std::fill(trg, trg + blockWidth, col);
for (int y = 0; y < blockHeight; ++y, trg = byteAdvance(trg, pitch))
for (int x = 0; x < blockWidth; ++x)
trg[x] = col;
}
inline
void fillBlock(uint32_t* trg, int pitch, uint32_t col, int n) { fillBlock(trg, pitch, col, n, n); }
#ifdef _MSC_VER
#define FORCE_INLINE __forceinline
#elif defined __GNUC__
#define FORCE_INLINE __attribute__((always_inline)) inline
#else
#define FORCE_INLINE inline
#endif
enum RotationDegree //clock-wise
{
ROT_0,
ROT_90,
ROT_180,
ROT_270
};
//calculate input matrix coordinates after rotation at compile time
template <RotationDegree rotDeg, size_t I, size_t J, size_t N>
struct MatrixRotation;
template <size_t I, size_t J, size_t N>
struct MatrixRotation<ROT_0, I, J, N>
{
static const size_t I_old = I;
static const size_t J_old = J;
};
template <RotationDegree rotDeg, size_t I, size_t J, size_t N> //(i, j) = (row, col) indices, N = size of (square) matrix
struct MatrixRotation
{
static const size_t I_old = N - 1 - MatrixRotation<static_cast<RotationDegree>(rotDeg - 1), I, J, N>::J_old; //old coordinates before rotation!
static const size_t J_old = MatrixRotation<static_cast<RotationDegree>(rotDeg - 1), I, J, N>::I_old; //
};
template <size_t N, RotationDegree rotDeg>
class OutputMatrix
{
public:
OutputMatrix(uint32_t* out, int outWidth) : //access matrix area, top-left at position "out" for image with given width
out_(out),
outWidth_(outWidth) {}
template <size_t I, size_t J>
uint32_t& ref() const
{
static const size_t I_old = MatrixRotation<rotDeg, I, J, N>::I_old;
static const size_t J_old = MatrixRotation<rotDeg, I, J, N>::J_old;
return *(out_ + J_old + I_old * outWidth_);
}
private:
uint32_t* out_;
const int outWidth_;
};
template <class T> inline
T square(T value) { return value * value; }
/*
inline
void rgbtoLuv(uint32_t c, double& L, double& u, double& v)
{
//http://www.easyrgb.com/index.php?X=MATH&H=02#text2
double r = getRed (c) / 255.0;
double g = getGreen(c) / 255.0;
double b = getBlue (c) / 255.0;
if ( r > 0.04045 )
r = std::pow(( ( r + 0.055 ) / 1.055 ) , 2.4);
else
r /= 12.92;
if ( g > 0.04045 )
g = std::pow(( ( g + 0.055 ) / 1.055 ) , 2.4);
else
g /= 12.92;
if ( b > 0.04045 )
b = std::pow(( ( b + 0.055 ) / 1.055 ) , 2.4);
else
b /= 12.92;
r *= 100;
g *= 100;
b *= 100;
double x = 0.4124564 * r + 0.3575761 * g + 0.1804375 * b;
double y = 0.2126729 * r + 0.7151522 * g + 0.0721750 * b;
double z = 0.0193339 * r + 0.1191920 * g + 0.9503041 * b;
//---------------------
double var_U = 4 * x / ( x + 15 * y + 3 * z );
double var_V = 9 * y / ( x + 15 * y + 3 * z );
double var_Y = y / 100;
if ( var_Y > 0.008856 ) var_Y = std::pow(var_Y , 1.0/3 );
else var_Y = 7.787 * var_Y + 16.0 / 116;
const double ref_X = 95.047; //Observer= 2?, Illuminant= D65
const double ref_Y = 100.000;
const double ref_Z = 108.883;
const double ref_U = ( 4 * ref_X ) / ( ref_X + ( 15 * ref_Y ) + ( 3 * ref_Z ) );
const double ref_V = ( 9 * ref_Y ) / ( ref_X + ( 15 * ref_Y ) + ( 3 * ref_Z ) );
L = ( 116 * var_Y ) - 16;
u = 13 * L * ( var_U - ref_U );
v = 13 * L * ( var_V - ref_V );
}
*/
inline
void rgbtoLab(uint32_t c, unsigned char& L, signed char& A, signed char& B)
{
//code: http://www.easyrgb.com/index.php?X=MATH
//test: http://www.workwithcolor.com/color-converter-01.htm
//------RGB to XYZ------
double r = getRed (c) / 255.0;
double g = getGreen(c) / 255.0;
double b = getBlue (c) / 255.0;
r = r > 0.04045 ? std::pow(( r + 0.055 ) / 1.055, 2.4) : r / 12.92;
r = g > 0.04045 ? std::pow(( g + 0.055 ) / 1.055, 2.4) : g / 12.92;
r = b > 0.04045 ? std::pow(( b + 0.055 ) / 1.055, 2.4) : b / 12.92;
r *= 100;
g *= 100;
b *= 100;
double x = 0.4124564 * r + 0.3575761 * g + 0.1804375 * b;
double y = 0.2126729 * r + 0.7151522 * g + 0.0721750 * b;
double z = 0.0193339 * r + 0.1191920 * g + 0.9503041 * b;
//------XYZ to Lab------
const double refX = 95.047; //
const double refY = 100.000; //Observer= 2?, Illuminant= D65
const double refZ = 108.883; //
double var_X = x / refX;
double var_Y = y / refY;
double var_Z = z / refZ;
var_X = var_X > 0.008856 ? std::pow(var_X, 1.0 / 3) : 7.787 * var_X + 4.0 / 29;
var_Y = var_Y > 0.008856 ? std::pow(var_Y, 1.0 / 3) : 7.787 * var_Y + 4.0 / 29;
var_Z = var_Z > 0.008856 ? std::pow(var_Z, 1.0 / 3) : 7.787 * var_Z + 4.0 / 29;
L = static_cast<unsigned char>(116 * var_Y - 16);
A = static_cast< signed char>(500 * (var_X - var_Y));
B = static_cast< signed char>(200 * (var_Y - var_Z));
};
inline
double distLAB(uint32_t pix1, uint32_t pix2)
{
unsigned char L1 = 0; //[0, 100]
signed char a1 = 0; //[-128, 127]
signed char b1 = 0; //[-128, 127]
rgbtoLab(pix1, L1, a1, b1);
unsigned char L2 = 0;
signed char a2 = 0;
signed char b2 = 0;
rgbtoLab(pix2, L2, a2, b2);
//-----------------------------
//http://www.easyrgb.com/index.php?X=DELT
//Delta E/CIE76
return std::sqrt(square(1.0 * L1 - L2) +
square(1.0 * a1 - a2) +
square(1.0 * b1 - b2));
}
/*
inline
void rgbtoHsl(uint32_t c, double& h, double& s, double& l)
{
//http://www.easyrgb.com/index.php?X=MATH&H=18#text18
const int r = getRed (c);
const int g = getGreen(c);
const int b = getBlue (c);
const int varMin = numeric::min(r, g, b);
const int varMax = numeric::max(r, g, b);
const int delMax = varMax - varMin;
l = (varMax + varMin) / 2.0 / 255.0;
if (delMax == 0) //gray, no chroma...
{
h = 0;
s = 0;
}
else
{
s = l < 0.5 ?
delMax / (1.0 * varMax + varMin) :
delMax / (2.0 * 255 - varMax - varMin);
double delR = ((varMax - r) / 6.0 + delMax / 2.0) / delMax;
double delG = ((varMax - g) / 6.0 + delMax / 2.0) / delMax;
double delB = ((varMax - b) / 6.0 + delMax / 2.0) / delMax;
if (r == varMax)
h = delB - delG;
else if (g == varMax)
h = 1 / 3.0 + delR - delB;
else if (b == varMax)
h = 2 / 3.0 + delG - delR;
if (h < 0)
h += 1;
if (h > 1)
h -= 1;
}
}
inline
double distHSL(uint32_t pix1, uint32_t pix2, double lightningWeight)
{
double h1 = 0;
double s1 = 0;
double l1 = 0;
rgbtoHsl(pix1, h1, s1, l1);
double h2 = 0;
double s2 = 0;
double l2 = 0;
rgbtoHsl(pix2, h2, s2, l2);
//HSL is in cylindric coordinatates where L represents height, S radius, H angle,
//however we interpret the cylinder as a bi-conic solid with top/bottom radius 0, middle radius 1
assert(0 <= h1 && h1 <= 1);
assert(0 <= h2 && h2 <= 1);
double r1 = l1 < 0.5 ?
l1 * 2 :
2 - l1 * 2;
double x1 = r1 * s1 * std::cos(h1 * 2 * numeric::pi);
double y1 = r1 * s1 * std::sin(h1 * 2 * numeric::pi);
double z1 = l1;
double r2 = l2 < 0.5 ?
l2 * 2 :
2 - l2 * 2;
double x2 = r2 * s2 * std::cos(h2 * 2 * numeric::pi);
double y2 = r2 * s2 * std::sin(h2 * 2 * numeric::pi);
double z2 = l2;
return 255 * std::sqrt(square(x1 - x2) + square(y1 - y2) + square(lightningWeight * (z1 - z2)));
}
*/
inline
double distRGB(uint32_t pix1, uint32_t pix2)
{
const double r_diff = static_cast<int>(getRed (pix1)) - getRed (pix2);
const double g_diff = static_cast<int>(getGreen(pix1)) - getGreen(pix2);
const double b_diff = static_cast<int>(getBlue (pix1)) - getBlue (pix2);
//euklidean RGB distance
return std::sqrt(square(r_diff) + square(g_diff) + square(b_diff));
}
inline
double distNonLinearRGB(uint32_t pix1, uint32_t pix2)
{
//non-linear rgb: http://www.compuphase.com/cmetric.htm
const double r_diff = static_cast<int>(getRed (pix1)) - getRed (pix2);
const double g_diff = static_cast<int>(getGreen(pix1)) - getGreen(pix2);
const double b_diff = static_cast<int>(getBlue (pix1)) - getBlue (pix2);
const double r_avg = (static_cast<double>(getRed(pix1)) + getRed(pix2)) / 2;
return std::sqrt((2 + r_avg / 255) * square(r_diff) + 4 * square(g_diff) + (2 + (255 - r_avg) / 255) * square(b_diff));
}
inline
double distYCbCr(uint32_t pix1, uint32_t pix2, double lumaWeight)
{
//http://en.wikipedia.org/wiki/YCbCr#ITU-R_BT.601_conversion
//YCbCr conversion is a matrix multiplication => take advantage of linearity by subtracting first!
const int r_diff = static_cast<int>(getRed (pix1)) - getRed (pix2); //we may delay division by 255 to after matrix multiplication
const int g_diff = static_cast<int>(getGreen(pix1)) - getGreen(pix2); //
const int b_diff = static_cast<int>(getBlue (pix1)) - getBlue (pix2); //substraction for int is noticeable faster than for double!
//const double k_b = 0.0722; //ITU-R BT.709 conversion
//const double k_r = 0.2126; //
const double k_b = 0.0593; //ITU-R BT.2020 conversion
const double k_r = 0.2627; //
const double k_g = 1 - k_b - k_r;
const double scale_b = 0.5 / (1 - k_b);
const double scale_r = 0.5 / (1 - k_r);
const double y = k_r * r_diff + k_g * g_diff + k_b * b_diff; //[!], analog YCbCr!
const double c_b = scale_b * (b_diff - y);
const double c_r = scale_r * (r_diff - y);
//we skip division by 255 to have similar range like other distance functions
return std::sqrt(square(lumaWeight * y) + square(c_b) + square(c_r));
}
struct DistYCbCrBuffer //30% perf boost compared to distYCbCr()!
{
public:
DistYCbCrBuffer() : buffer(256 * 256 * 256)
{
for (uint32_t i = 0; i < 256 * 256 * 256; ++i) //startup time: 114 ms on Intel Core i5 (four cores)
{
const int r_diff = getByte<2>(i) * 2 - 255;
const int g_diff = getByte<1>(i) * 2 - 255;
const int b_diff = getByte<0>(i) * 2 - 255;
const double k_b = 0.0593; //ITU-R BT.2020 conversion
const double k_r = 0.2627; //
const double k_g = 1 - k_b - k_r;
const double scale_b = 0.5 / (1 - k_b);
const double scale_r = 0.5 / (1 - k_r);
const double y = k_r * r_diff + k_g * g_diff + k_b * b_diff; //[!], analog YCbCr!
const double c_b = scale_b * (b_diff - y);
const double c_r = scale_r * (r_diff - y);
buffer[i] = static_cast<float>(std::sqrt(square(y) + square(c_b) + square(c_r)));
}
}
double dist(uint32_t pix1, uint32_t pix2) const
{
//if (pix1 == pix2) -> 8% perf degradation!
// return 0;
//if (pix1 > pix2)
// std::swap(pix1, pix2); -> 30% perf degradation!!!
const int r_diff = static_cast<int>(getRed (pix1)) - getRed (pix2);
const int g_diff = static_cast<int>(getGreen(pix1)) - getGreen(pix2);
const int b_diff = static_cast<int>(getBlue (pix1)) - getBlue (pix2);
return buffer[(((r_diff + 255) / 2) << 16) | //slightly reduce precision (division by 2) to squeeze value into single byte
(((g_diff + 255) / 2) << 8) |
(( b_diff + 255) / 2)];
}
private:
std::vector<float> buffer; //consumes 64 MB memory; using double is 2% faster, but takes 128 MB
};
DistYCbCrBuffer *distYCbCrBuffer = nullptr;
inline
double distYUV(uint32_t pix1, uint32_t pix2, double luminanceWeight)
{
//perf: it's not worthwhile to buffer the YUV-conversion, the direct code is faster by ~ 6%
//since RGB -> YUV conversion is essentially a matrix multiplication, we can calculate the RGB diff before the conversion (distributive property)
const double r_diff = static_cast<int>(getRed (pix1)) - getRed (pix2);
const double g_diff = static_cast<int>(getGreen(pix1)) - getGreen(pix2);
const double b_diff = static_cast<int>(getBlue (pix1)) - getBlue (pix2);
//http://en.wikipedia.org/wiki/YUV#Conversion_to.2Ffrom_RGB
const double w_b = 0.114;
const double w_r = 0.299;
const double w_g = 1 - w_r - w_b;
const double u_max = 0.436;
const double v_max = 0.615;
const double scale_u = u_max / (1 - w_b);
const double scale_v = v_max / (1 - w_r);
double y = w_r * r_diff + w_g * g_diff + w_b * b_diff;//value range: 255 * [-1, 1]
double u = scale_u * (b_diff - y); //value range: 255 * 2 * u_max * [-1, 1]
double v = scale_v * (r_diff - y); //value range: 255 * 2 * v_max * [-1, 1]
#ifdef _DEBUG
const double eps = 0.5;
assert(abs(y) <= 255 + eps);
assert(abs(u) <= 255 * 2 * u_max + eps);
assert(abs(v) <= 255 * 2 * v_max + eps);
#endif
return std::sqrt(square(luminanceWeight * y) + square(u) + square(v));
}
enum BlendType
{
BLEND_NONE = 0,
BLEND_NORMAL, //a normal indication to blend
BLEND_DOMINANT, //a strong indication to blend
//attention: BlendType must fit into the value range of 2 bit!!!
};
struct BlendResult
{
BlendType
/**/blend_f, blend_g,
/**/blend_j, blend_k;
};
struct Kernel_4x4 //kernel for preprocessing step
{
uint32_t
/**/a, b, c, d,
/**/e, f, g, h,
/**/i, j, k, l,
/**/m, n, o, p;
};
/*
input kernel area naming convention:
-----------------
| A | B | C | D |
----|---|---|---|
| E | F | G | H | //evaluate the four corners between F, G, J, K
----|---|---|---| //input pixel is at position F
| I | J | K | L |
----|---|---|---|
| M | N | O | P |
-----------------
*/
template <class ColorDistance>
FORCE_INLINE //detect blend direction
BlendResult preProcessCorners(const Kernel_4x4& ker, const xbrz::ScalerCfg& cfg) //result: F, G, J, K corners of "GradientType"
{
BlendResult result = {};
if ((ker.f == ker.g &&
ker.j == ker.k) ||
(ker.f == ker.j &&
ker.g == ker.k))
return result;
auto dist = [&](uint32_t pix1, uint32_t pix2) { return ColorDistance::dist(pix1, pix2, cfg.luminanceWeight_); };
const int weight = 4;
double jg = dist(ker.i, ker.f) + dist(ker.f, ker.c) + dist(ker.n, ker.k) + dist(ker.k, ker.h) + weight * dist(ker.j, ker.g);
double fk = dist(ker.e, ker.j) + dist(ker.j, ker.o) + dist(ker.b, ker.g) + dist(ker.g, ker.l) + weight * dist(ker.f, ker.k);
if (jg < fk) //test sample: 70% of values max(jg, fk) / min(jg, fk) are between 1.1 and 3.7 with median being 1.8
{
const bool dominantGradient = cfg.dominantDirectionThreshold * jg < fk;
if (ker.f != ker.g && ker.f != ker.j)
result.blend_f = dominantGradient ? BLEND_DOMINANT : BLEND_NORMAL;
if (ker.k != ker.j && ker.k != ker.g)
result.blend_k = dominantGradient ? BLEND_DOMINANT : BLEND_NORMAL;
}
else if (fk < jg)
{
const bool dominantGradient = cfg.dominantDirectionThreshold * fk < jg;
if (ker.j != ker.f && ker.j != ker.k)
result.blend_j = dominantGradient ? BLEND_DOMINANT : BLEND_NORMAL;
if (ker.g != ker.f && ker.g != ker.k)
result.blend_g = dominantGradient ? BLEND_DOMINANT : BLEND_NORMAL;
}
return result;
}
struct Kernel_3x3
{
uint32_t
/**/a, b, c,
/**/d, e, f,
/**/g, h, i;
};
#define DEF_GETTER(x) template <RotationDegree rotDeg> uint32_t inline get_##x(const Kernel_3x3& ker) { return ker.x; }
//we cannot and NEED NOT write "ker.##x" since ## concatenates preprocessor tokens but "." is not a token
DEF_GETTER(a) DEF_GETTER(b) DEF_GETTER(c)
DEF_GETTER(d) DEF_GETTER(e) DEF_GETTER(f)
DEF_GETTER(g) DEF_GETTER(h) DEF_GETTER(i)
#undef DEF_GETTER
#define DEF_GETTER(x, y) template <> inline uint32_t get_##x<ROT_90>(const Kernel_3x3& ker) { return ker.y; }
DEF_GETTER(a, g) DEF_GETTER(b, d) DEF_GETTER(c, a)
DEF_GETTER(d, h) DEF_GETTER(e, e) DEF_GETTER(f, b)
DEF_GETTER(g, i) DEF_GETTER(h, f) DEF_GETTER(i, c)
#undef DEF_GETTER
#define DEF_GETTER(x, y) template <> inline uint32_t get_##x<ROT_180>(const Kernel_3x3& ker) { return ker.y; }
DEF_GETTER(a, i) DEF_GETTER(b, h) DEF_GETTER(c, g)
DEF_GETTER(d, f) DEF_GETTER(e, e) DEF_GETTER(f, d)
DEF_GETTER(g, c) DEF_GETTER(h, b) DEF_GETTER(i, a)
#undef DEF_GETTER
#define DEF_GETTER(x, y) template <> inline uint32_t get_##x<ROT_270>(const Kernel_3x3& ker) { return ker.y; }
DEF_GETTER(a, c) DEF_GETTER(b, f) DEF_GETTER(c, i)
DEF_GETTER(d, b) DEF_GETTER(e, e) DEF_GETTER(f, h)
DEF_GETTER(g, a) DEF_GETTER(h, d) DEF_GETTER(i, g)
#undef DEF_GETTER
//compress four blend types into a single byte
inline BlendType getTopL (unsigned char b) { return static_cast<BlendType>(0x3 & b); }
inline BlendType getTopR (unsigned char b) { return static_cast<BlendType>(0x3 & (b >> 2)); }
inline BlendType getBottomR(unsigned char b) { return static_cast<BlendType>(0x3 & (b >> 4)); }
inline BlendType getBottomL(unsigned char b) { return static_cast<BlendType>(0x3 & (b >> 6)); }
inline void setTopL (unsigned char& b, BlendType bt) { b |= bt; } //buffer is assumed to be initialized before preprocessing!
inline void setTopR (unsigned char& b, BlendType bt) { b |= (bt << 2); }
inline void setBottomR(unsigned char& b, BlendType bt) { b |= (bt << 4); }
inline void setBottomL(unsigned char& b, BlendType bt) { b |= (bt << 6); }
inline bool blendingNeeded(unsigned char b) { return b != 0; }
template <RotationDegree rotDeg> inline
unsigned char rotateBlendInfo(unsigned char b) { return b; }
template <> inline unsigned char rotateBlendInfo<ROT_90 >(unsigned char b) { return ((b << 2) | (b >> 6)) & 0xff; }
template <> inline unsigned char rotateBlendInfo<ROT_180>(unsigned char b) { return ((b << 4) | (b >> 4)) & 0xff; }
template <> inline unsigned char rotateBlendInfo<ROT_270>(unsigned char b) { return ((b << 6) | (b >> 2)) & 0xff; }
#ifdef _DEBUG
int debugPixelX = -1;
int debugPixelY = 84;
bool breakIntoDebugger = false;
#endif
/*
input kernel area naming convention:
-------------
| A | B | C |
----|---|---|
| D | E | F | //input pixel is at position E
----|---|---|
| G | H | I |
-------------
*/
template <class Scaler, class ColorDistance, RotationDegree rotDeg>
FORCE_INLINE //perf: quite worth it!
void scalePixel(const Kernel_3x3& ker,
uint32_t* target, int trgWidth,
unsigned char blendInfo, //result of preprocessing all four corners of pixel "e"
const xbrz::ScalerCfg& cfg)
{
#define a get_a<rotDeg>(ker)
#define b get_b<rotDeg>(ker)
#define c get_c<rotDeg>(ker)
#define d get_d<rotDeg>(ker)
#define e get_e<rotDeg>(ker)
#define f get_f<rotDeg>(ker)
#define g get_g<rotDeg>(ker)
#define h get_h<rotDeg>(ker)
#define i get_i<rotDeg>(ker)
const unsigned char blend = rotateBlendInfo<rotDeg>(blendInfo);
if (getBottomR(blend) >= BLEND_NORMAL)
{
auto eq = [&](uint32_t pix1, uint32_t pix2) { return ColorDistance::dist(pix1, pix2, cfg.luminanceWeight_) < cfg.equalColorTolerance_; };
auto dist = [&](uint32_t pix1, uint32_t pix2) { return ColorDistance::dist(pix1, pix2, cfg.luminanceWeight_); };
const bool doLineBlend = [&]() -> bool
{
if (getBottomR(blend) >= BLEND_DOMINANT)
return true;
//make sure there is no second blending in an adjacent rotation for this pixel: handles insular pixels, mario eyes
if (getTopR(blend) != BLEND_NONE && !eq(e, g)) //but support double-blending for 90? corners
return false;
if (getBottomL(blend) != BLEND_NONE && !eq(e, c))
return false;
//no full blending for L-shapes; blend corner only (handles "mario mushroom eyes")
if (!eq(e, i) && eq(g, h) && eq(h , i) && eq(i, f) && eq(f, c))
return false;
return true;
}();
const uint32_t px = dist(e, f) <= dist(e, h) ? f : h; //choose most similar color
OutputMatrix<Scaler::scale, rotDeg> out(target, trgWidth);
if (doLineBlend)
{
const double fg = dist(f, g); //test sample: 70% of values max(fg, hc) / min(fg, hc) are between 1.1 and 3.7 with median being 1.9
const double hc = dist(h, c); //
const bool haveShallowLine = cfg.steepDirectionThreshold * fg <= hc && e != g && d != g;
const bool haveSteepLine = cfg.steepDirectionThreshold * hc <= fg && e != c && b != c;
if (haveShallowLine)
{
if (haveSteepLine)
Scaler::blendLineSteepAndShallow(px, out);
else
Scaler::blendLineShallow(px, out);
}
else
{
if (haveSteepLine)
Scaler::blendLineSteep(px, out);
else
Scaler::blendLineDiagonal(px,out);
}
}
else
Scaler::blendCorner(px, out);
}
#undef a
#undef b
#undef c
#undef d
#undef e
#undef f
#undef g
#undef h
#undef i
}
template <class Scaler, class ColorDistance> //scaler policy: see "Scaler2x" reference implementation
void scaleImage(const uint32_t* src, uint32_t* trg, int srcWidth, int srcHeight, const xbrz::ScalerCfg& cfg, int yFirst, int yLast)
{
yFirst = std::max(yFirst, 0);
yLast = std::min(yLast, srcHeight);
if (yFirst >= yLast || srcWidth <= 0)
return;
const int trgWidth = srcWidth * Scaler::scale;
//"use" space at the end of the image as temporary buffer for "on the fly preprocessing": we even could use larger area of
//"sizeof(uint32_t) * srcWidth * (yLast - yFirst)" bytes without risk of accidental overwriting before accessing
const int bufferSize = srcWidth;
unsigned char* preProcBuffer = reinterpret_cast<unsigned char*>(trg + yLast * Scaler::scale * trgWidth) - bufferSize;
std::fill(preProcBuffer, preProcBuffer + bufferSize, 0);
static_assert(BLEND_NONE == 0, "");
//initialize preprocessing buffer for first row: detect upper left and right corner blending
//this cannot be optimized for adjacent processing stripes; we must not allow for a memory race condition!
if (yFirst > 0)
{
const int y = yFirst - 1;
const uint32_t* s_m1 = src + srcWidth * std::max(y - 1, 0);
const uint32_t* s_0 = src + srcWidth * y; //center line
const uint32_t* s_p1 = src + srcWidth * std::min(y + 1, srcHeight - 1);
const uint32_t* s_p2 = src + srcWidth * std::min(y + 2, srcHeight - 1);
for (int x = 0; x < srcWidth; ++x)
{
const int x_m1 = std::max(x - 1, 0);
const int x_p1 = std::min(x + 1, srcWidth - 1);
const int x_p2 = std::min(x + 2, srcWidth - 1);
Kernel_4x4 ker = {}; //perf: initialization is negligible
ker.a = s_m1[x_m1]; //read sequentially from memory as far as possible
ker.b = s_m1[x];
ker.c = s_m1[x_p1];
ker.d = s_m1[x_p2];
ker.e = s_0[x_m1];
ker.f = s_0[x];
ker.g = s_0[x_p1];
ker.h = s_0[x_p2];
ker.i = s_p1[x_m1];
ker.j = s_p1[x];
ker.k = s_p1[x_p1];
ker.l = s_p1[x_p2];
ker.m = s_p2[x_m1];
ker.n = s_p2[x];
ker.o = s_p2[x_p1];
ker.p = s_p2[x_p2];
const BlendResult res = preProcessCorners<ColorDistance>(ker, cfg);
/*
preprocessing blend result:
---------
| F | G | //evalute corner between F, G, J, K
----|---| //input pixel is at position F
| J | K |
---------
*/
setTopR(preProcBuffer[x], res.blend_j);
if (x + 1 < bufferSize)
setTopL(preProcBuffer[x + 1], res.blend_k);
}
}
//------------------------------------------------------------------------------------
for (int y = yFirst; y < yLast; ++y)
{
uint32_t* out = trg + Scaler::scale * y * trgWidth; //consider MT "striped" access
const uint32_t* s_m1 = src + srcWidth * std::max(y - 1, 0);
const uint32_t* s_0 = src + srcWidth * y; //center line
const uint32_t* s_p1 = src + srcWidth * std::min(y + 1, srcHeight - 1);
const uint32_t* s_p2 = src + srcWidth * std::min(y + 2, srcHeight - 1);
unsigned char blend_xy1 = 0; //corner blending for current (x, y + 1) position
for (int x = 0; x < srcWidth; ++x, out += Scaler::scale)
{
#ifdef _DEBUG
breakIntoDebugger = debugPixelX == x && debugPixelY == y;
#endif
//all those bounds checks have only insignificant impact on performance!
const int x_m1 = std::max(x - 1, 0); //perf: prefer array indexing to additional pointers!
const int x_p1 = std::min(x + 1, srcWidth - 1);
const int x_p2 = std::min(x + 2, srcWidth - 1);
Kernel_4x4 ker4 = {}; //perf: initialization is negligible
ker4.a = s_m1[x_m1]; //read sequentially from memory as far as possible
ker4.b = s_m1[x];
ker4.c = s_m1[x_p1];
ker4.d = s_m1[x_p2];
ker4.e = s_0[x_m1];
ker4.f = s_0[x];
ker4.g = s_0[x_p1];
ker4.h = s_0[x_p2];
ker4.i = s_p1[x_m1];
ker4.j = s_p1[x];
ker4.k = s_p1[x_p1];
ker4.l = s_p1[x_p2];
ker4.m = s_p2[x_m1];
ker4.n = s_p2[x];
ker4.o = s_p2[x_p1];
ker4.p = s_p2[x_p2];
//evaluate the four corners on bottom-right of current pixel
unsigned char blend_xy = 0; //for current (x, y) position
{
const BlendResult res = preProcessCorners<ColorDistance>(ker4, cfg);
/*
preprocessing blend result:
---------
| F | G | //evalute corner between F, G, J, K
----|---| //current input pixel is at position F
| J | K |
---------
*/
blend_xy = preProcBuffer[x];
setBottomR(blend_xy, res.blend_f); //all four corners of (x, y) have been determined at this point due to processing sequence!
setTopR(blend_xy1, res.blend_j); //set 2nd known corner for (x, y + 1)
preProcBuffer[x] = blend_xy1; //store on current buffer position for use on next row
blend_xy1 = 0;
setTopL(blend_xy1, res.blend_k); //set 1st known corner for (x + 1, y + 1) and buffer for use on next column
if (x + 1 < bufferSize) //set 3rd known corner for (x + 1, y)
setBottomL(preProcBuffer[x + 1], res.blend_g);
}
//fill block of size scale * scale with the given color
fillBlock(out, trgWidth * sizeof(uint32_t), ker4.f, Scaler::scale); //place *after* preprocessing step, to not overwrite the results while processing the the last pixel!
//blend four corners of current pixel
if (blendingNeeded(blend_xy)) //good 20% perf-improvement
{
Kernel_3x3 ker3 = {}; //perf: initialization is negligible
ker3.a = ker4.a;
ker3.b = ker4.b;
ker3.c = ker4.c;
ker3.d = ker4.e;
ker3.e = ker4.f;
ker3.f = ker4.g;
ker3.g = ker4.i;
ker3.h = ker4.j;
ker3.i = ker4.k;
scalePixel<Scaler, ColorDistance, ROT_0 >(ker3, out, trgWidth, blend_xy, cfg);
scalePixel<Scaler, ColorDistance, ROT_90 >(ker3, out, trgWidth, blend_xy, cfg);
scalePixel<Scaler, ColorDistance, ROT_180>(ker3, out, trgWidth, blend_xy, cfg);
scalePixel<Scaler, ColorDistance, ROT_270>(ker3, out, trgWidth, blend_xy, cfg);
}
}
}
}
//------------------------------------------------------------------------------------
struct Scaler2x
{
static const int scale = 2;
template <class OutputMatrix>
static void blendLineShallow(uint32_t col, OutputMatrix& out)
{
alphaBlend<1, 4>(out.template ref<scale - 1, 0>(), col);
alphaBlend<3, 4>(out.template ref<scale - 1, 1>(), col);
}
template <class OutputMatrix>
static void blendLineSteep(uint32_t col, OutputMatrix& out)
{
alphaBlend<1, 4>(out.template ref<0, scale - 1>(), col);
alphaBlend<3, 4>(out.template ref<1, scale - 1>(), col);
}
template <class OutputMatrix>
static void blendLineSteepAndShallow(uint32_t col, OutputMatrix& out)
{
alphaBlend<1, 4>(out.template ref<1, 0>(), col);
alphaBlend<1, 4>(out.template ref<0, 1>(), col);
alphaBlend<5, 6>(out.template ref<1, 1>(), col); //[!] fixes 7/8 used in xBR
}
template <class OutputMatrix>
static void blendLineDiagonal(uint32_t col, OutputMatrix& out)
{
alphaBlend<1, 2>(out.template ref<1, 1>(), col);
}
template <class OutputMatrix>
static void blendCorner(uint32_t col, OutputMatrix& out)
{
//model a round corner
alphaBlend<21, 100>(out.template ref<1, 1>(), col); //exact: 1 - pi/4 = 0.2146018366
}
};
struct Scaler3x
{
static const int scale = 3;
template <class OutputMatrix>
static void blendLineShallow(uint32_t col, OutputMatrix& out)
{
alphaBlend<1, 4>(out.template ref<scale - 1, 0>(), col);
alphaBlend<1, 4>(out.template ref<scale - 2, 2>(), col);
alphaBlend<3, 4>(out.template ref<scale - 1, 1>(), col);
out.template ref<scale - 1, 2>() = col;
}
template <class OutputMatrix>
static void blendLineSteep(uint32_t col, OutputMatrix& out)
{
alphaBlend<1, 4>(out.template ref<0, scale - 1>(), col);
alphaBlend<1, 4>(out.template ref<2, scale - 2>(), col);
alphaBlend<3, 4>(out.template ref<1, scale - 1>(), col);
out.template ref<2, scale - 1>() = col;
}
template <class OutputMatrix>
static void blendLineSteepAndShallow(uint32_t col, OutputMatrix& out)
{
alphaBlend<1, 4>(out.template ref<2, 0>(), col);
alphaBlend<1, 4>(out.template ref<0, 2>(), col);
alphaBlend<3, 4>(out.template ref<2, 1>(), col);
alphaBlend<3, 4>(out.template ref<1, 2>(), col);
out.template ref<2, 2>() = col;
}
template <class OutputMatrix>
static void blendLineDiagonal(uint32_t col, OutputMatrix& out)
{
alphaBlend<1, 8>(out.template ref<1, 2>(), col);
alphaBlend<1, 8>(out.template ref<2, 1>(), col);
alphaBlend<7, 8>(out.template ref<2, 2>(), col);
}
template <class OutputMatrix>
static void blendCorner(uint32_t col, OutputMatrix& out)
{
//model a round corner
alphaBlend<45, 100>(out.template ref<2, 2>(), col); //exact: 0.4545939598
//alphaBlend<14, 1000>(out.template ref<2, 1>(), col); //0.01413008627 -> negligible
//alphaBlend<14, 1000>(out.template ref<1, 2>(), col); //0.01413008627
}
};
struct Scaler4x
{
static const int scale = 4;
template <class OutputMatrix>
static void blendLineShallow(uint32_t col, OutputMatrix& out)
{
alphaBlend<1, 4>(out.template ref<scale - 1, 0>(), col);
alphaBlend<1, 4>(out.template ref<scale - 2, 2>(), col);
alphaBlend<3, 4>(out.template ref<scale - 1, 1>(), col);
alphaBlend<3, 4>(out.template ref<scale - 2, 3>(), col);
out.template ref<scale - 1, 2>() = col;
out.template ref<scale - 1, 3>() = col;
}
template <class OutputMatrix>
static void blendLineSteep(uint32_t col, OutputMatrix& out)
{
alphaBlend<1, 4>(out.template ref<0, scale - 1>(), col);
alphaBlend<1, 4>(out.template ref<2, scale - 2>(), col);
alphaBlend<3, 4>(out.template ref<1, scale - 1>(), col);