-
-
Notifications
You must be signed in to change notification settings - Fork 41
/
ProcessRGB.cpp
4210 lines (3510 loc) · 165 KB
/
ProcessRGB.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
#include <array>
#include <string.h>
#include <limits>
#ifdef __ARM_NEON
# include <arm_neon.h>
#endif
#include "Dither.hpp"
#include "ForceInline.hpp"
#include "Math.hpp"
#include "ProcessCommon.hpp"
#include "ProcessRGB.hpp"
#include "Tables.hpp"
#include "Vector.hpp"
#if defined __SSE4_1__ || defined __AVX2__ || defined _MSC_VER
# ifdef _MSC_VER
# include <intrin.h>
# include <Windows.h>
# define _bswap(x) _byteswap_ulong(x)
# define _bswap64(x) _byteswap_uint64(x)
# else
# include <x86intrin.h>
# endif
#endif
#ifndef _bswap
# define _bswap(x) __builtin_bswap32(x)
# define _bswap64(x) __builtin_bswap64(x)
#endif
static const uint32_t MaxError = 1065369600; // ((38+76+14) * 255)^2
// common T-/H-mode table
static uint8_t tableTH[8] = { 3, 6, 11, 16, 23, 32, 41, 64 };
// thresholds for the early compression-mode decision scheme
// default: 0.03, 0.09, and 0.38
float ecmd_threshold[3] = { 0.03f, 0.09f, 0.38f };
static const uint8_t ModeUndecided = 0;
static const uint8_t ModePlanar = 0x1;
static const uint8_t ModeTH = 0x2;
const unsigned int R = 2;
const unsigned int G = 1;
const unsigned int B = 0;
struct Luma
{
#ifdef __AVX2__
float max, min;
uint8_t minIdx = 255, maxIdx = 255;
__m128i luma8;
#elif defined __ARM_NEON && defined __aarch64__
float max, min;
uint8_t minIdx = 255, maxIdx = 255;
uint8x16_t luma8;
#else
uint8_t max = 0, min = 255, maxIdx = 0, minIdx = 0;
uint8_t val[16];
#endif
};
#ifdef __AVX2__
struct Plane
{
uint64_t plane;
uint64_t error;
__m256i sum4;
};
#endif
#if defined __AVX2__ || (defined __ARM_NEON && defined __aarch64__)
struct Channels
{
#ifdef __AVX2__
__m128i r8, g8, b8;
#elif defined __ARM_NEON && defined __aarch64__
uint8x16x2_t r, g, b;
#endif
};
#endif
namespace
{
static etcpak_force_inline uint8_t clamp( uint8_t min, int16_t val, uint8_t max )
{
return val < min ? min : ( val > max ? max : val );
}
static etcpak_force_inline uint8_t clampMin( uint8_t min, int16_t val )
{
return val < min ? min : val;
}
static etcpak_force_inline uint8_t clampMax( int16_t val, uint8_t max )
{
return val > max ? max : val;
}
// slightly faster than std::sort
static void insertionSort( uint8_t* arr1, uint8_t* arr2 )
{
for( uint8_t i = 1; i < 16; ++i )
{
uint8_t value = arr1[i];
uint8_t hole = i;
for( ; hole > 0 && value < arr1[hole - 1]; --hole )
{
arr1[hole] = arr1[hole - 1];
arr2[hole] = arr2[hole - 1];
}
arr1[hole] = value;
arr2[hole] = i;
}
}
//converts indices from |a0|a1|e0|e1|i0|i1|m0|m1|b0|b1|f0|f1|j0|j1|n0|n1|c0|c1|g0|g1|k0|k1|o0|o1|d0|d1|h0|h1|l0|l1|p0|p1| previously used by T- and H-modes
// into |p0|o0|n0|m0|l0|k0|j0|i0|h0|g0|f0|e0|d0|c0|b0|a0|p1|o1|n1|m1|l1|k1|j1|i1|h1|g1|f1|e1|d1|c1|b1|a1| which should be used for all modes.
// NO WARRANTY --- SEE STATEMENT IN TOP OF FILE (C) Ericsson AB 2005-2013. All Rights Reserved.
static etcpak_force_inline int indexConversion( int pixelIndices )
{
int correctIndices = 0;
int LSB[4][4];
int MSB[4][4];
int shift = 0;
for( int y = 3; y >= 0; y-- )
{
for( int x = 3; x >= 0; x-- )
{
LSB[x][y] = ( pixelIndices >> shift ) & 1;
shift++;
MSB[x][y] = ( pixelIndices >> shift ) & 1;
shift++;
}
}
shift = 0;
for( int x = 0; x < 4; x++ )
{
for( int y = 0; y < 4; y++ )
{
correctIndices |= ( LSB[x][y] << shift );
correctIndices |= ( MSB[x][y] << ( 16 + shift ) );
shift++;
}
}
return correctIndices;
}
// Swapping two RGB-colors
// NO WARRANTY --- SEE STATEMENT IN TOP OF FILE (C) Ericsson AB 2005-2013. All Rights Reserved.
static etcpak_force_inline void swapColors( uint8_t( colors )[2][3] )
{
uint8_t temp = colors[0][R];
colors[0][R] = colors[1][R];
colors[1][R] = temp;
temp = colors[0][G];
colors[0][G] = colors[1][G];
colors[1][G] = temp;
temp = colors[0][B];
colors[0][B] = colors[1][B];
colors[1][B] = temp;
}
// calculates quantized colors for T or H modes
void compressColor( uint8_t( currColor )[2][3], uint8_t( quantColor )[2][3], bool t_mode )
{
if( t_mode )
{
quantColor[0][R] = clampMax( 15 * ( currColor[0][R] + 8 ) / 255, 15 );
quantColor[0][G] = clampMax( 15 * ( currColor[0][G] + 8 ) / 255, 15 );
quantColor[0][B] = clampMax( 15 * ( currColor[0][B] + 8 ) / 255, 15 );
}
else // clamped to [1,14] to get a wider range
{
quantColor[0][R] = clamp( 1, 15 * ( currColor[0][R] + 8 ) / 255, 14 );
quantColor[0][G] = clamp( 1, 15 * ( currColor[0][G] + 8 ) / 255, 14 );
quantColor[0][B] = clamp( 1, 15 * ( currColor[0][B] + 8 ) / 255, 14 );
}
// clamped to [1,14] to get a wider range
quantColor[1][R] = clamp( 1, 15 * ( currColor[1][R] + 8 ) / 255, 14 );
quantColor[1][G] = clamp( 1, 15 * ( currColor[1][G] + 8 ) / 255, 14 );
quantColor[1][B] = clamp( 1, 15 * ( currColor[1][B] + 8 ) / 255, 14 );
}
// three decoding functions come from ETCPACK v2.74 and are slightly changed.
static etcpak_force_inline void decompressColor( uint8_t( colorsRGB444 )[2][3], uint8_t( colors )[2][3] )
{
// The color should be retrieved as:
//
// c = round(255/(r_bits^2-1))*comp_color
//
// This is similar to bit replication
//
// Note -- this code only work for bit replication from 4 bits and up --- 3 bits needs
// two copy operations.
colors[0][R] = ( colorsRGB444[0][R] << 4 ) | colorsRGB444[0][R];
colors[0][G] = ( colorsRGB444[0][G] << 4 ) | colorsRGB444[0][G];
colors[0][B] = ( colorsRGB444[0][B] << 4 ) | colorsRGB444[0][B];
colors[1][R] = ( colorsRGB444[1][R] << 4 ) | colorsRGB444[1][R];
colors[1][G] = ( colorsRGB444[1][G] << 4 ) | colorsRGB444[1][G];
colors[1][B] = ( colorsRGB444[1][B] << 4 ) | colorsRGB444[1][B];
}
// calculates the paint colors from the block colors
// using a distance d and one of the H- or T-patterns.
static void calculatePaintColors59T( uint8_t d, uint8_t( colors )[2][3], uint8_t( pColors )[4][3] )
{
//////////////////////////////////////////////
//
// C3 C1 C4----C1---C2
// | | |
// | | |
// |-------| |
// | | |
// | | |
// C4 C2 C3
//
//////////////////////////////////////////////
// C4
pColors[3][R] = clampMin( 0, colors[1][R] - tableTH[d] );
pColors[3][G] = clampMin( 0, colors[1][G] - tableTH[d] );
pColors[3][B] = clampMin( 0, colors[1][B] - tableTH[d] );
// C3
pColors[0][R] = colors[0][R];
pColors[0][G] = colors[0][G];
pColors[0][B] = colors[0][B];
// C2
pColors[1][R] = clampMax( colors[1][R] + tableTH[d], 255 );
pColors[1][G] = clampMax( colors[1][G] + tableTH[d], 255 );
pColors[1][B] = clampMax( colors[1][B] + tableTH[d], 255 );
// C1
pColors[2][R] = colors[1][R];
pColors[2][G] = colors[1][G];
pColors[2][B] = colors[1][B];
}
static void calculatePaintColors58H( uint8_t d, uint8_t( colors )[2][3], uint8_t( pColors )[4][3] )
{
pColors[3][R] = clampMin( 0, colors[1][R] - tableTH[d] );
pColors[3][G] = clampMin( 0, colors[1][G] - tableTH[d] );
pColors[3][B] = clampMin( 0, colors[1][B] - tableTH[d] );
// C1
pColors[0][R] = clampMax( colors[0][R] + tableTH[d], 255 );
pColors[0][G] = clampMax( colors[0][G] + tableTH[d], 255 );
pColors[0][B] = clampMax( colors[0][B] + tableTH[d], 255 );
// C2
pColors[1][R] = clampMin( 0, colors[0][R] - tableTH[d] );
pColors[1][G] = clampMin( 0, colors[0][G] - tableTH[d] );
pColors[1][B] = clampMin( 0, colors[0][B] - tableTH[d] );
// C3
pColors[2][R] = clampMax( colors[1][R] + tableTH[d], 255 );
pColors[2][G] = clampMax( colors[1][G] + tableTH[d], 255 );
pColors[2][B] = clampMax( colors[1][B] + tableTH[d], 255 );
}
#if defined _MSC_VER && !defined __clang__
static etcpak_force_inline unsigned long _bit_scan_forward( unsigned long mask )
{
unsigned long ret;
_BitScanForward( &ret, mask );
return ret;
}
#endif
typedef std::array<uint16_t, 4> v4i;
#ifdef __AVX2__
static etcpak_force_inline __m256i Sum4_AVX2( const uint8_t* data) noexcept
{
__m128i d0 = _mm_loadu_si128(((__m128i*)data) + 0);
__m128i d1 = _mm_loadu_si128(((__m128i*)data) + 1);
__m128i d2 = _mm_loadu_si128(((__m128i*)data) + 2);
__m128i d3 = _mm_loadu_si128(((__m128i*)data) + 3);
__m128i dm0 = _mm_and_si128(d0, _mm_set1_epi32(0x00FFFFFF));
__m128i dm1 = _mm_and_si128(d1, _mm_set1_epi32(0x00FFFFFF));
__m128i dm2 = _mm_and_si128(d2, _mm_set1_epi32(0x00FFFFFF));
__m128i dm3 = _mm_and_si128(d3, _mm_set1_epi32(0x00FFFFFF));
__m256i t0 = _mm256_cvtepu8_epi16(dm0);
__m256i t1 = _mm256_cvtepu8_epi16(dm1);
__m256i t2 = _mm256_cvtepu8_epi16(dm2);
__m256i t3 = _mm256_cvtepu8_epi16(dm3);
__m256i sum0 = _mm256_add_epi16(t0, t1);
__m256i sum1 = _mm256_add_epi16(t2, t3);
__m256i s0 = _mm256_permute2x128_si256(sum0, sum1, (0) | (3 << 4)); // 0, 0, 3, 3
__m256i s1 = _mm256_permute2x128_si256(sum0, sum1, (1) | (2 << 4)); // 1, 1, 2, 2
__m256i s2 = _mm256_permute4x64_epi64(s0, _MM_SHUFFLE(1, 3, 0, 2));
__m256i s3 = _mm256_permute4x64_epi64(s0, _MM_SHUFFLE(0, 2, 1, 3));
__m256i s4 = _mm256_permute4x64_epi64(s1, _MM_SHUFFLE(3, 1, 0, 2));
__m256i s5 = _mm256_permute4x64_epi64(s1, _MM_SHUFFLE(2, 0, 1, 3));
__m256i sum5 = _mm256_add_epi16(s2, s3); // 3, 0, 3, 0
__m256i sum6 = _mm256_add_epi16(s4, s5); // 2, 1, 1, 2
return _mm256_add_epi16(sum5, sum6); // 3+2, 0+1, 3+1, 3+2
}
static etcpak_force_inline __m256i Average_AVX2( const __m256i data) noexcept
{
__m256i a = _mm256_add_epi16(data, _mm256_set1_epi16(4));
return _mm256_srli_epi16(a, 3);
}
static etcpak_force_inline __m128i CalcErrorBlock_AVX2( const __m256i data, const v4i a[8]) noexcept
{
//
__m256i a0 = _mm256_load_si256((__m256i*)a[0].data());
__m256i a1 = _mm256_load_si256((__m256i*)a[4].data());
// err = 8 * ( sq( average[0] ) + sq( average[1] ) + sq( average[2] ) );
__m256i a4 = _mm256_madd_epi16(a0, a0);
__m256i a5 = _mm256_madd_epi16(a1, a1);
__m256i a6 = _mm256_hadd_epi32(a4, a5);
__m256i a7 = _mm256_slli_epi32(a6, 3);
__m256i a8 = _mm256_add_epi32(a7, _mm256_set1_epi32(0x3FFFFFFF)); // Big value to prevent negative values, but small enough to prevent overflow
// average is not swapped
// err -= block[0] * 2 * average[0];
// err -= block[1] * 2 * average[1];
// err -= block[2] * 2 * average[2];
__m256i a2 = _mm256_slli_epi16(a0, 1);
__m256i a3 = _mm256_slli_epi16(a1, 1);
__m256i b0 = _mm256_madd_epi16(a2, data);
__m256i b1 = _mm256_madd_epi16(a3, data);
__m256i b2 = _mm256_hadd_epi32(b0, b1);
__m256i b3 = _mm256_sub_epi32(a8, b2);
__m256i b4 = _mm256_hadd_epi32(b3, b3);
__m256i b5 = _mm256_permutevar8x32_epi32(b4, _mm256_set_epi32(0, 0, 0, 0, 5, 1, 4, 0));
return _mm256_castsi256_si128(b5);
}
static etcpak_force_inline void ProcessAverages_AVX2(const __m256i d, v4i a[8] ) noexcept
{
__m256i t = _mm256_add_epi16(_mm256_mullo_epi16(d, _mm256_set1_epi16(31)), _mm256_set1_epi16(128));
__m256i c = _mm256_srli_epi16(_mm256_add_epi16(t, _mm256_srli_epi16(t, 8)), 8);
__m256i c1 = _mm256_shuffle_epi32(c, _MM_SHUFFLE(3, 2, 3, 2));
__m256i diff = _mm256_sub_epi16(c, c1);
diff = _mm256_max_epi16(diff, _mm256_set1_epi16(-4));
diff = _mm256_min_epi16(diff, _mm256_set1_epi16(3));
__m256i co = _mm256_add_epi16(c1, diff);
c = _mm256_blend_epi16(co, c, 0xF0);
__m256i a0 = _mm256_or_si256(_mm256_slli_epi16(c, 3), _mm256_srli_epi16(c, 2));
_mm256_store_si256((__m256i*)a[4].data(), a0);
__m256i t0 = _mm256_add_epi16(_mm256_mullo_epi16(d, _mm256_set1_epi16(15)), _mm256_set1_epi16(128));
__m256i t1 = _mm256_srli_epi16(_mm256_add_epi16(t0, _mm256_srli_epi16(t0, 8)), 8);
__m256i t2 = _mm256_or_si256(t1, _mm256_slli_epi16(t1, 4));
_mm256_store_si256((__m256i*)a[0].data(), t2);
}
static etcpak_force_inline uint64_t EncodeAverages_AVX2( const v4i a[8], size_t idx ) noexcept
{
uint64_t d = ( idx << 24 );
size_t base = idx << 1;
__m128i a0 = _mm_load_si128((const __m128i*)a[base].data());
__m128i r0, r1;
if( ( idx & 0x2 ) == 0 )
{
r0 = _mm_srli_epi16(a0, 4);
__m128i a1 = _mm_unpackhi_epi64(r0, r0);
r1 = _mm_slli_epi16(a1, 4);
}
else
{
__m128i a1 = _mm_and_si128(a0, _mm_set1_epi16(-8));
r0 = _mm_unpackhi_epi64(a1, a1);
__m128i a2 = _mm_sub_epi16(a1, r0);
__m128i a3 = _mm_srai_epi16(a2, 3);
r1 = _mm_and_si128(a3, _mm_set1_epi16(0x07));
}
__m128i r2 = _mm_or_si128(r0, r1);
// do missing swap for average values
__m128i r3 = _mm_shufflelo_epi16(r2, _MM_SHUFFLE(3, 0, 1, 2));
__m128i r4 = _mm_packus_epi16(r3, _mm_setzero_si128());
d |= _mm_cvtsi128_si32(r4);
return d;
}
static etcpak_force_inline uint64_t CheckSolid_AVX2( const uint8_t* src ) noexcept
{
__m256i d0 = _mm256_loadu_si256(((__m256i*)src) + 0);
__m256i d1 = _mm256_loadu_si256(((__m256i*)src) + 1);
__m256i c = _mm256_broadcastd_epi32(_mm256_castsi256_si128(d0));
__m256i c0 = _mm256_cmpeq_epi8(d0, c);
__m256i c1 = _mm256_cmpeq_epi8(d1, c);
__m256i m = _mm256_and_si256(c0, c1);
if (!_mm256_testc_si256(m, _mm256_set1_epi32(-1)))
{
return 0;
}
return 0x02000000 |
( (unsigned int)( src[0] & 0xF8 ) << 16 ) |
( (unsigned int)( src[1] & 0xF8 ) << 8 ) |
( (unsigned int)( src[2] & 0xF8 ) );
}
static etcpak_force_inline __m128i PrepareAverages_AVX2( v4i a[8], const uint8_t* src) noexcept
{
__m256i sum4 = Sum4_AVX2( src );
ProcessAverages_AVX2(Average_AVX2( sum4 ), a );
return CalcErrorBlock_AVX2( sum4, a);
}
static etcpak_force_inline __m128i PrepareAverages_AVX2( v4i a[8], const __m256i sum4) noexcept
{
ProcessAverages_AVX2(Average_AVX2( sum4 ), a );
return CalcErrorBlock_AVX2( sum4, a);
}
static etcpak_force_inline void FindBestFit_4x2_AVX2( uint32_t terr[2][8], uint32_t tsel[8], v4i a[8], const uint32_t offset, const uint8_t* data) noexcept
{
__m256i sel0 = _mm256_setzero_si256();
__m256i sel1 = _mm256_setzero_si256();
for (unsigned int j = 0; j < 2; ++j)
{
unsigned int bid = offset + 1 - j;
__m256i squareErrorSum = _mm256_setzero_si256();
__m128i a0 = _mm_loadl_epi64((const __m128i*)a[bid].data());
__m256i a1 = _mm256_broadcastq_epi64(a0);
// Processing one full row each iteration
for (size_t i = 0; i < 8; i += 4)
{
__m128i rgb = _mm_loadu_si128((const __m128i*)(data + i * 4));
__m256i rgb16 = _mm256_cvtepu8_epi16(rgb);
__m256i d = _mm256_sub_epi16(a1, rgb16);
// The scaling values are divided by two and rounded, to allow the differences to be in the range of signed int16
// This produces slightly different results, but is significant faster
__m256i pixel0 = _mm256_madd_epi16(d, _mm256_set_epi16(0, 38, 76, 14, 0, 38, 76, 14, 0, 38, 76, 14, 0, 38, 76, 14));
__m256i pixel1 = _mm256_packs_epi32(pixel0, pixel0);
__m256i pixel2 = _mm256_hadd_epi16(pixel1, pixel1);
__m128i pixel3 = _mm256_castsi256_si128(pixel2);
__m128i pix0 = _mm_broadcastw_epi16(pixel3);
__m128i pix1 = _mm_broadcastw_epi16(_mm_srli_epi32(pixel3, 16));
__m256i pixel = _mm256_insertf128_si256(_mm256_castsi128_si256(pix0), pix1, 1);
// Processing first two pixels of the row
{
__m256i pix = _mm256_abs_epi16(pixel);
// Taking the absolute value is way faster. The values are only used to sort, so the result will be the same.
// Since the selector table is symmetrical, we need to calculate the difference only for half of the entries.
__m256i error0 = _mm256_abs_epi16(_mm256_sub_epi16(pix, _mm256_broadcastsi128_si256(g_table128_SIMD[0])));
__m256i error1 = _mm256_abs_epi16(_mm256_sub_epi16(pix, _mm256_broadcastsi128_si256(g_table128_SIMD[1])));
__m256i minIndex0 = _mm256_and_si256(_mm256_cmpgt_epi16(error0, error1), _mm256_set1_epi16(1));
__m256i minError = _mm256_min_epi16(error0, error1);
// Exploiting symmetry of the selector table and use the sign bit
// This produces slightly different results, but is significant faster
__m256i minIndex1 = _mm256_srli_epi16(pixel, 15);
// Interleaving values so madd instruction can be used
__m256i minErrorLo = _mm256_permute4x64_epi64(minError, _MM_SHUFFLE(1, 1, 0, 0));
__m256i minErrorHi = _mm256_permute4x64_epi64(minError, _MM_SHUFFLE(3, 3, 2, 2));
__m256i minError2 = _mm256_unpacklo_epi16(minErrorLo, minErrorHi);
// Squaring the minimum error to produce correct values when adding
__m256i squareError = _mm256_madd_epi16(minError2, minError2);
squareErrorSum = _mm256_add_epi32(squareErrorSum, squareError);
// Packing selector bits
__m256i minIndexLo2 = _mm256_sll_epi16(minIndex0, _mm_cvtsi64_si128(i + j * 8));
__m256i minIndexHi2 = _mm256_sll_epi16(minIndex1, _mm_cvtsi64_si128(i + j * 8));
sel0 = _mm256_or_si256(sel0, minIndexLo2);
sel1 = _mm256_or_si256(sel1, minIndexHi2);
}
pixel3 = _mm256_extracti128_si256(pixel2, 1);
pix0 = _mm_broadcastw_epi16(pixel3);
pix1 = _mm_broadcastw_epi16(_mm_srli_epi32(pixel3, 16));
pixel = _mm256_insertf128_si256(_mm256_castsi128_si256(pix0), pix1, 1);
// Processing second two pixels of the row
{
__m256i pix = _mm256_abs_epi16(pixel);
// Taking the absolute value is way faster. The values are only used to sort, so the result will be the same.
// Since the selector table is symmetrical, we need to calculate the difference only for half of the entries.
__m256i error0 = _mm256_abs_epi16(_mm256_sub_epi16(pix, _mm256_broadcastsi128_si256(g_table128_SIMD[0])));
__m256i error1 = _mm256_abs_epi16(_mm256_sub_epi16(pix, _mm256_broadcastsi128_si256(g_table128_SIMD[1])));
__m256i minIndex0 = _mm256_and_si256(_mm256_cmpgt_epi16(error0, error1), _mm256_set1_epi16(1));
__m256i minError = _mm256_min_epi16(error0, error1);
// Exploiting symmetry of the selector table and use the sign bit
__m256i minIndex1 = _mm256_srli_epi16(pixel, 15);
// Interleaving values so madd instruction can be used
__m256i minErrorLo = _mm256_permute4x64_epi64(minError, _MM_SHUFFLE(1, 1, 0, 0));
__m256i minErrorHi = _mm256_permute4x64_epi64(minError, _MM_SHUFFLE(3, 3, 2, 2));
__m256i minError2 = _mm256_unpacklo_epi16(minErrorLo, minErrorHi);
// Squaring the minimum error to produce correct values when adding
__m256i squareError = _mm256_madd_epi16(minError2, minError2);
squareErrorSum = _mm256_add_epi32(squareErrorSum, squareError);
// Packing selector bits
__m256i minIndexLo2 = _mm256_sll_epi16(minIndex0, _mm_cvtsi64_si128(i + j * 8));
__m256i minIndexHi2 = _mm256_sll_epi16(minIndex1, _mm_cvtsi64_si128(i + j * 8));
__m256i minIndexLo3 = _mm256_slli_epi16(minIndexLo2, 2);
__m256i minIndexHi3 = _mm256_slli_epi16(minIndexHi2, 2);
sel0 = _mm256_or_si256(sel0, minIndexLo3);
sel1 = _mm256_or_si256(sel1, minIndexHi3);
}
}
data += 8 * 4;
_mm256_store_si256((__m256i*)terr[1 - j], squareErrorSum);
}
// Interleave selector bits
__m256i minIndexLo0 = _mm256_unpacklo_epi16(sel0, sel1);
__m256i minIndexHi0 = _mm256_unpackhi_epi16(sel0, sel1);
__m256i minIndexLo1 = _mm256_permute2x128_si256(minIndexLo0, minIndexHi0, (0) | (2 << 4));
__m256i minIndexHi1 = _mm256_permute2x128_si256(minIndexLo0, minIndexHi0, (1) | (3 << 4));
__m256i minIndexHi2 = _mm256_slli_epi32(minIndexHi1, 1);
__m256i sel = _mm256_or_si256(minIndexLo1, minIndexHi2);
_mm256_store_si256((__m256i*)tsel, sel);
}
static etcpak_force_inline void FindBestFit_2x4_AVX2( uint32_t terr[2][8], uint32_t tsel[8], v4i a[8], const uint32_t offset, const uint8_t* data) noexcept
{
__m256i sel0 = _mm256_setzero_si256();
__m256i sel1 = _mm256_setzero_si256();
__m256i squareErrorSum0 = _mm256_setzero_si256();
__m256i squareErrorSum1 = _mm256_setzero_si256();
__m128i a0 = _mm_loadl_epi64((const __m128i*)a[offset + 1].data());
__m128i a1 = _mm_loadl_epi64((const __m128i*)a[offset + 0].data());
__m128i a2 = _mm_broadcastq_epi64(a0);
__m128i a3 = _mm_broadcastq_epi64(a1);
__m256i a4 = _mm256_insertf128_si256(_mm256_castsi128_si256(a2), a3, 1);
// Processing one full row each iteration
for (size_t i = 0; i < 16; i += 4)
{
__m128i rgb = _mm_loadu_si128((const __m128i*)(data + i * 4));
__m256i rgb16 = _mm256_cvtepu8_epi16(rgb);
__m256i d = _mm256_sub_epi16(a4, rgb16);
// The scaling values are divided by two and rounded, to allow the differences to be in the range of signed int16
// This produces slightly different results, but is significant faster
__m256i pixel0 = _mm256_madd_epi16(d, _mm256_set_epi16(0, 38, 76, 14, 0, 38, 76, 14, 0, 38, 76, 14, 0, 38, 76, 14));
__m256i pixel1 = _mm256_packs_epi32(pixel0, pixel0);
__m256i pixel2 = _mm256_hadd_epi16(pixel1, pixel1);
__m128i pixel3 = _mm256_castsi256_si128(pixel2);
__m128i pix0 = _mm_broadcastw_epi16(pixel3);
__m128i pix1 = _mm_broadcastw_epi16(_mm_srli_epi32(pixel3, 16));
__m256i pixel = _mm256_insertf128_si256(_mm256_castsi128_si256(pix0), pix1, 1);
// Processing first two pixels of the row
{
__m256i pix = _mm256_abs_epi16(pixel);
// Taking the absolute value is way faster. The values are only used to sort, so the result will be the same.
// Since the selector table is symmetrical, we need to calculate the difference only for half of the entries.
__m256i error0 = _mm256_abs_epi16(_mm256_sub_epi16(pix, _mm256_broadcastsi128_si256(g_table128_SIMD[0])));
__m256i error1 = _mm256_abs_epi16(_mm256_sub_epi16(pix, _mm256_broadcastsi128_si256(g_table128_SIMD[1])));
__m256i minIndex0 = _mm256_and_si256(_mm256_cmpgt_epi16(error0, error1), _mm256_set1_epi16(1));
__m256i minError = _mm256_min_epi16(error0, error1);
// Exploiting symmetry of the selector table and use the sign bit
__m256i minIndex1 = _mm256_srli_epi16(pixel, 15);
// Interleaving values so madd instruction can be used
__m256i minErrorLo = _mm256_permute4x64_epi64(minError, _MM_SHUFFLE(1, 1, 0, 0));
__m256i minErrorHi = _mm256_permute4x64_epi64(minError, _MM_SHUFFLE(3, 3, 2, 2));
__m256i minError2 = _mm256_unpacklo_epi16(minErrorLo, minErrorHi);
// Squaring the minimum error to produce correct values when adding
__m256i squareError = _mm256_madd_epi16(minError2, minError2);
squareErrorSum0 = _mm256_add_epi32(squareErrorSum0, squareError);
// Packing selector bits
__m256i minIndexLo2 = _mm256_sll_epi16(minIndex0, _mm_cvtsi64_si128(i));
__m256i minIndexHi2 = _mm256_sll_epi16(minIndex1, _mm_cvtsi64_si128(i));
sel0 = _mm256_or_si256(sel0, minIndexLo2);
sel1 = _mm256_or_si256(sel1, minIndexHi2);
}
pixel3 = _mm256_extracti128_si256(pixel2, 1);
pix0 = _mm_broadcastw_epi16(pixel3);
pix1 = _mm_broadcastw_epi16(_mm_srli_epi32(pixel3, 16));
pixel = _mm256_insertf128_si256(_mm256_castsi128_si256(pix0), pix1, 1);
// Processing second two pixels of the row
{
__m256i pix = _mm256_abs_epi16(pixel);
// Taking the absolute value is way faster. The values are only used to sort, so the result will be the same.
// Since the selector table is symmetrical, we need to calculate the difference only for half of the entries.
__m256i error0 = _mm256_abs_epi16(_mm256_sub_epi16(pix, _mm256_broadcastsi128_si256(g_table128_SIMD[0])));
__m256i error1 = _mm256_abs_epi16(_mm256_sub_epi16(pix, _mm256_broadcastsi128_si256(g_table128_SIMD[1])));
__m256i minIndex0 = _mm256_and_si256(_mm256_cmpgt_epi16(error0, error1), _mm256_set1_epi16(1));
__m256i minError = _mm256_min_epi16(error0, error1);
// Exploiting symmetry of the selector table and use the sign bit
__m256i minIndex1 = _mm256_srli_epi16(pixel, 15);
// Interleaving values so madd instruction can be used
__m256i minErrorLo = _mm256_permute4x64_epi64(minError, _MM_SHUFFLE(1, 1, 0, 0));
__m256i minErrorHi = _mm256_permute4x64_epi64(minError, _MM_SHUFFLE(3, 3, 2, 2));
__m256i minError2 = _mm256_unpacklo_epi16(minErrorLo, minErrorHi);
// Squaring the minimum error to produce correct values when adding
__m256i squareError = _mm256_madd_epi16(minError2, minError2);
squareErrorSum1 = _mm256_add_epi32(squareErrorSum1, squareError);
// Packing selector bits
__m256i minIndexLo2 = _mm256_sll_epi16(minIndex0, _mm_cvtsi64_si128(i));
__m256i minIndexHi2 = _mm256_sll_epi16(minIndex1, _mm_cvtsi64_si128(i));
__m256i minIndexLo3 = _mm256_slli_epi16(minIndexLo2, 2);
__m256i minIndexHi3 = _mm256_slli_epi16(minIndexHi2, 2);
sel0 = _mm256_or_si256(sel0, minIndexLo3);
sel1 = _mm256_or_si256(sel1, minIndexHi3);
}
}
_mm256_store_si256((__m256i*)terr[1], squareErrorSum0);
_mm256_store_si256((__m256i*)terr[0], squareErrorSum1);
// Interleave selector bits
__m256i minIndexLo0 = _mm256_unpacklo_epi16(sel0, sel1);
__m256i minIndexHi0 = _mm256_unpackhi_epi16(sel0, sel1);
__m256i minIndexLo1 = _mm256_permute2x128_si256(minIndexLo0, minIndexHi0, (0) | (2 << 4));
__m256i minIndexHi1 = _mm256_permute2x128_si256(minIndexLo0, minIndexHi0, (1) | (3 << 4));
__m256i minIndexHi2 = _mm256_slli_epi32(minIndexHi1, 1);
__m256i sel = _mm256_or_si256(minIndexLo1, minIndexHi2);
_mm256_store_si256((__m256i*)tsel, sel);
}
static etcpak_force_inline uint64_t EncodeSelectors_AVX2( uint64_t d, const uint32_t terr[2][8], const uint32_t tsel[8], const bool rotate) noexcept
{
size_t tidx[2];
// Get index of minimum error (terr[0] and terr[1])
__m256i err0 = _mm256_load_si256((const __m256i*)terr[0]);
__m256i err1 = _mm256_load_si256((const __m256i*)terr[1]);
__m256i errLo = _mm256_permute2x128_si256(err0, err1, (0) | (2 << 4));
__m256i errHi = _mm256_permute2x128_si256(err0, err1, (1) | (3 << 4));
__m256i errMin0 = _mm256_min_epu32(errLo, errHi);
__m256i errMin1 = _mm256_shuffle_epi32(errMin0, _MM_SHUFFLE(2, 3, 0, 1));
__m256i errMin2 = _mm256_min_epu32(errMin0, errMin1);
__m256i errMin3 = _mm256_shuffle_epi32(errMin2, _MM_SHUFFLE(1, 0, 3, 2));
__m256i errMin4 = _mm256_min_epu32(errMin3, errMin2);
__m256i errMin5 = _mm256_permute2x128_si256(errMin4, errMin4, (0) | (0 << 4));
__m256i errMin6 = _mm256_permute2x128_si256(errMin4, errMin4, (1) | (1 << 4));
__m256i errMask0 = _mm256_cmpeq_epi32(errMin5, err0);
__m256i errMask1 = _mm256_cmpeq_epi32(errMin6, err1);
uint32_t mask0 = _mm256_movemask_epi8(errMask0);
uint32_t mask1 = _mm256_movemask_epi8(errMask1);
tidx[0] = _bit_scan_forward(mask0) >> 2;
tidx[1] = _bit_scan_forward(mask1) >> 2;
d |= tidx[0] << 26;
d |= tidx[1] << 29;
unsigned int t0 = tsel[tidx[0]];
unsigned int t1 = tsel[tidx[1]];
if (!rotate)
{
t0 &= 0xFF00FF00;
t1 &= 0x00FF00FF;
}
else
{
t0 &= 0xCCCCCCCC;
t1 &= 0x33333333;
}
// Flip selectors from sign bit
unsigned int t2 = (t0 | t1) ^ 0xFFFF0000;
return d | static_cast<uint64_t>(_bswap(t2)) << 32;
}
static etcpak_force_inline __m128i r6g7b6_AVX2(__m128 cof, __m128 chf, __m128 cvf) noexcept
{
__m128i co = _mm_cvttps_epi32(cof);
__m128i ch = _mm_cvttps_epi32(chf);
__m128i cv = _mm_cvttps_epi32(cvf);
__m128i coh = _mm_packus_epi32(co, ch);
__m128i cv0 = _mm_packus_epi32(cv, _mm_setzero_si128());
__m256i cohv0 = _mm256_inserti128_si256(_mm256_castsi128_si256(coh), cv0, 1);
__m256i cohv1 = _mm256_min_epu16(cohv0, _mm256_set1_epi16(1023));
__m256i cohv2 = _mm256_sub_epi16(cohv1, _mm256_set1_epi16(15));
__m256i cohv3 = _mm256_srai_epi16(cohv2, 1);
__m256i cohvrb0 = _mm256_add_epi16(cohv3, _mm256_set1_epi16(11));
__m256i cohvrb1 = _mm256_add_epi16(cohv3, _mm256_set1_epi16(4));
__m256i cohvg0 = _mm256_add_epi16(cohv3, _mm256_set1_epi16(9));
__m256i cohvg1 = _mm256_add_epi16(cohv3, _mm256_set1_epi16(6));
__m256i cohvrb2 = _mm256_srai_epi16(cohvrb0, 7);
__m256i cohvrb3 = _mm256_srai_epi16(cohvrb1, 7);
__m256i cohvg2 = _mm256_srai_epi16(cohvg0, 8);
__m256i cohvg3 = _mm256_srai_epi16(cohvg1, 8);
__m256i cohvrb4 = _mm256_sub_epi16(cohvrb0, cohvrb2);
__m256i cohvrb5 = _mm256_sub_epi16(cohvrb4, cohvrb3);
__m256i cohvg4 = _mm256_sub_epi16(cohvg0, cohvg2);
__m256i cohvg5 = _mm256_sub_epi16(cohvg4, cohvg3);
__m256i cohvrb6 = _mm256_srai_epi16(cohvrb5, 3);
__m256i cohvg6 = _mm256_srai_epi16(cohvg5, 2);
__m256i cohv4 = _mm256_blend_epi16(cohvg6, cohvrb6, 0x55);
__m128i cohv5 = _mm_packus_epi16(_mm256_castsi256_si128(cohv4), _mm256_extracti128_si256(cohv4, 1));
return _mm_shuffle_epi8(cohv5, _mm_setr_epi8(6, 5, 4, -1, 2, 1, 0, -1, 10, 9, 8, -1, -1, -1, -1, -1));
}
static etcpak_force_inline Plane Planar_AVX2( const Channels& ch, uint8_t& mode, bool useHeuristics )
{
__m128i t0 = _mm_sad_epu8( ch.r8, _mm_setzero_si128() );
__m128i t1 = _mm_sad_epu8( ch.g8, _mm_setzero_si128() );
__m128i t2 = _mm_sad_epu8( ch.b8, _mm_setzero_si128() );
__m128i r8s = _mm_shuffle_epi8( ch.r8, _mm_set_epi8( 0xF, 0xE, 0xB, 0xA, 0x7, 0x6, 0x3, 0x2, 0xD, 0xC, 0x9, 0x8, 0x5, 0x4, 0x1, 0x0 ) );
__m128i g8s = _mm_shuffle_epi8( ch.g8, _mm_set_epi8( 0xF, 0xE, 0xB, 0xA, 0x7, 0x6, 0x3, 0x2, 0xD, 0xC, 0x9, 0x8, 0x5, 0x4, 0x1, 0x0 ) );
__m128i b8s = _mm_shuffle_epi8( ch.b8, _mm_set_epi8( 0xF, 0xE, 0xB, 0xA, 0x7, 0x6, 0x3, 0x2, 0xD, 0xC, 0x9, 0x8, 0x5, 0x4, 0x1, 0x0 ) );
__m128i s0 = _mm_sad_epu8( r8s, _mm_setzero_si128() );
__m128i s1 = _mm_sad_epu8( g8s, _mm_setzero_si128() );
__m128i s2 = _mm_sad_epu8( b8s, _mm_setzero_si128() );
__m256i sr0 = _mm256_insertf128_si256( _mm256_castsi128_si256( t0 ), s0, 1 );
__m256i sg0 = _mm256_insertf128_si256( _mm256_castsi128_si256( t1 ), s1, 1 );
__m256i sb0 = _mm256_insertf128_si256( _mm256_castsi128_si256( t2 ), s2, 1 );
__m256i sr1 = _mm256_slli_epi64( sr0, 32 );
__m256i sg1 = _mm256_slli_epi64( sg0, 16 );
__m256i srb = _mm256_or_si256( sr1, sb0 );
__m256i srgb = _mm256_or_si256( srb, sg1 );
if( mode != ModePlanar && useHeuristics )
{
Plane plane;
plane.sum4 = _mm256_permute4x64_epi64( srgb, _MM_SHUFFLE( 2, 3, 0, 1 ) );
return plane;
}
__m128i t3 = _mm_castps_si128( _mm_shuffle_ps( _mm_castsi128_ps( t0 ), _mm_castsi128_ps( t1 ), _MM_SHUFFLE( 2, 0, 2, 0 ) ) );
__m128i t4 = _mm_shuffle_epi32( t2, _MM_SHUFFLE( 3, 1, 2, 0 ) );
__m128i t5 = _mm_hadd_epi32( t3, t4 );
__m128i t6 = _mm_shuffle_epi32( t5, _MM_SHUFFLE( 1, 1, 1, 1 ) );
__m128i t7 = _mm_shuffle_epi32( t5, _MM_SHUFFLE( 2, 2, 2, 2 ) );
__m256i sr = _mm256_broadcastw_epi16( t5 );
__m256i sg = _mm256_broadcastw_epi16( t6 );
__m256i sb = _mm256_broadcastw_epi16( t7 );
__m256i r08 = _mm256_cvtepu8_epi16( ch.r8 );
__m256i g08 = _mm256_cvtepu8_epi16( ch.g8 );
__m256i b08 = _mm256_cvtepu8_epi16( ch.b8 );
__m256i r16 = _mm256_slli_epi16( r08, 4 );
__m256i g16 = _mm256_slli_epi16( g08, 4 );
__m256i b16 = _mm256_slli_epi16( b08, 4 );
__m256i difR0 = _mm256_sub_epi16( r16, sr );
__m256i difG0 = _mm256_sub_epi16( g16, sg );
__m256i difB0 = _mm256_sub_epi16( b16, sb );
__m256i difRyz = _mm256_madd_epi16( difR0, _mm256_set_epi16( 255, 85, -85, -255, 255, 85, -85, -255, 255, 85, -85, -255, 255, 85, -85, -255 ) );
__m256i difGyz = _mm256_madd_epi16( difG0, _mm256_set_epi16( 255, 85, -85, -255, 255, 85, -85, -255, 255, 85, -85, -255, 255, 85, -85, -255 ) );
__m256i difByz = _mm256_madd_epi16( difB0, _mm256_set_epi16( 255, 85, -85, -255, 255, 85, -85, -255, 255, 85, -85, -255, 255, 85, -85, -255 ) );
__m256i difRxz = _mm256_madd_epi16( difR0, _mm256_set_epi16( 255, 255, 255, 255, 85, 85, 85, 85, -85, -85, -85, -85, -255, -255, -255, -255 ) );
__m256i difGxz = _mm256_madd_epi16( difG0, _mm256_set_epi16( 255, 255, 255, 255, 85, 85, 85, 85, -85, -85, -85, -85, -255, -255, -255, -255 ) );
__m256i difBxz = _mm256_madd_epi16( difB0, _mm256_set_epi16( 255, 255, 255, 255, 85, 85, 85, 85, -85, -85, -85, -85, -255, -255, -255, -255 ) );
__m256i difRGyz = _mm256_hadd_epi32( difRyz, difGyz );
__m256i difByzxz = _mm256_hadd_epi32( difByz, difBxz );
__m256i difRGxz = _mm256_hadd_epi32( difRxz, difGxz );
__m128i sumRGyz = _mm_add_epi32( _mm256_castsi256_si128( difRGyz ), _mm256_extracti128_si256( difRGyz, 1 ) );
__m128i sumByzxz = _mm_add_epi32( _mm256_castsi256_si128( difByzxz ), _mm256_extracti128_si256( difByzxz, 1 ) );
__m128i sumRGxz = _mm_add_epi32( _mm256_castsi256_si128( difRGxz ), _mm256_extracti128_si256( difRGxz, 1 ) );
__m128i sumRGByz = _mm_hadd_epi32( sumRGyz, sumByzxz );
__m128i sumRGByzxz = _mm_hadd_epi32( sumRGxz, sumByzxz );
__m128i sumRGBxz = _mm_shuffle_epi32( sumRGByzxz, _MM_SHUFFLE( 2, 3, 1, 0 ) );
__m128 sumRGByzf = _mm_cvtepi32_ps( sumRGByz );
__m128 sumRGBxzf = _mm_cvtepi32_ps( sumRGBxz );
const float value = ( 255 * 255 * 8.0f + 85 * 85 * 8.0f ) * 16.0f;
__m128 scale = _mm_set1_ps( -4.0f / value );
__m128 af = _mm_mul_ps( sumRGBxzf, scale );
__m128 bf = _mm_mul_ps( sumRGByzf, scale );
__m128 df = _mm_mul_ps( _mm_cvtepi32_ps( t5 ), _mm_set1_ps( 4.0f / 16.0f ) );
// calculating the three colors RGBO, RGBH, and RGBV. RGB = df - af * x - bf * y;
__m128 cof0 = _mm_fnmadd_ps( af, _mm_set1_ps( -255.0f ), _mm_fnmadd_ps( bf, _mm_set1_ps( -255.0f ), df ) );
__m128 chf0 = _mm_fnmadd_ps( af, _mm_set1_ps( 425.0f ), _mm_fnmadd_ps( bf, _mm_set1_ps( -255.0f ), df ) );
__m128 cvf0 = _mm_fnmadd_ps( af, _mm_set1_ps( -255.0f ), _mm_fnmadd_ps( bf, _mm_set1_ps( 425.0f ), df ) );
// convert to r6g7b6
__m128i cohv = r6g7b6_AVX2( cof0, chf0, cvf0 );
uint64_t rgbho = _mm_extract_epi64( cohv, 0 );
uint32_t rgbv0 = _mm_extract_epi32( cohv, 2 );
// Error calculation
uint64_t error = 0;
if( !useHeuristics )
{
auto ro0 = ( rgbho >> 48 ) & 0x3F;
auto go0 = ( rgbho >> 40 ) & 0x7F;
auto bo0 = ( rgbho >> 32 ) & 0x3F;
auto ro1 = ( ro0 >> 4 ) | ( ro0 << 2 );
auto go1 = ( go0 >> 6 ) | ( go0 << 1 );
auto bo1 = ( bo0 >> 4 ) | ( bo0 << 2 );
auto ro2 = ( ro1 << 2 ) + 2;
auto go2 = ( go1 << 2 ) + 2;
auto bo2 = ( bo1 << 2 ) + 2;
__m256i ro3 = _mm256_set1_epi16( ro2 );
__m256i go3 = _mm256_set1_epi16( go2 );
__m256i bo3 = _mm256_set1_epi16( bo2 );
auto rh0 = ( rgbho >> 16 ) & 0x3F;
auto gh0 = ( rgbho >> 8 ) & 0x7F;
auto bh0 = ( rgbho >> 0 ) & 0x3F;
auto rh1 = ( rh0 >> 4 ) | ( rh0 << 2 );
auto gh1 = ( gh0 >> 6 ) | ( gh0 << 1 );
auto bh1 = ( bh0 >> 4 ) | ( bh0 << 2 );
auto rh2 = rh1 - ro1;
auto gh2 = gh1 - go1;
auto bh2 = bh1 - bo1;
__m256i rh3 = _mm256_set1_epi16( rh2 );
__m256i gh3 = _mm256_set1_epi16( gh2 );
__m256i bh3 = _mm256_set1_epi16( bh2 );
auto rv0 = ( rgbv0 >> 16 ) & 0x3F;
auto gv0 = ( rgbv0 >> 8 ) & 0x7F;
auto bv0 = ( rgbv0 >> 0 ) & 0x3F;
auto rv1 = ( rv0 >> 4 ) | ( rv0 << 2 );
auto gv1 = ( gv0 >> 6 ) | ( gv0 << 1 );
auto bv1 = ( bv0 >> 4 ) | ( bv0 << 2 );
auto rv2 = rv1 - ro1;
auto gv2 = gv1 - go1;
auto bv2 = bv1 - bo1;
__m256i rv3 = _mm256_set1_epi16( rv2 );
__m256i gv3 = _mm256_set1_epi16( gv2 );
__m256i bv3 = _mm256_set1_epi16( bv2 );
__m256i x = _mm256_set_epi16( 3, 3, 3, 3, 2, 2, 2, 2, 1, 1, 1, 1, 0, 0, 0, 0 );
__m256i rh4 = _mm256_mullo_epi16( rh3, x );
__m256i gh4 = _mm256_mullo_epi16( gh3, x );
__m256i bh4 = _mm256_mullo_epi16( bh3, x );
__m256i y = _mm256_set_epi16( 3, 2, 1, 0, 3, 2, 1, 0, 3, 2, 1, 0, 3, 2, 1, 0 );
__m256i rv4 = _mm256_mullo_epi16( rv3, y );
__m256i gv4 = _mm256_mullo_epi16( gv3, y );
__m256i bv4 = _mm256_mullo_epi16( bv3, y );
__m256i rxy = _mm256_add_epi16( rh4, rv4 );
__m256i gxy = _mm256_add_epi16( gh4, gv4 );
__m256i bxy = _mm256_add_epi16( bh4, bv4 );
__m256i rp0 = _mm256_add_epi16( rxy, ro3 );
__m256i gp0 = _mm256_add_epi16( gxy, go3 );
__m256i bp0 = _mm256_add_epi16( bxy, bo3 );
__m256i rp1 = _mm256_srai_epi16( rp0, 2 );
__m256i gp1 = _mm256_srai_epi16( gp0, 2 );
__m256i bp1 = _mm256_srai_epi16( bp0, 2 );
__m256i rp2 = _mm256_max_epi16( _mm256_min_epi16( rp1, _mm256_set1_epi16( 255 ) ), _mm256_setzero_si256() );
__m256i gp2 = _mm256_max_epi16( _mm256_min_epi16( gp1, _mm256_set1_epi16( 255 ) ), _mm256_setzero_si256() );
__m256i bp2 = _mm256_max_epi16( _mm256_min_epi16( bp1, _mm256_set1_epi16( 255 ) ), _mm256_setzero_si256() );
__m256i rdif = _mm256_sub_epi16( r08, rp2 );
__m256i gdif = _mm256_sub_epi16( g08, gp2 );
__m256i bdif = _mm256_sub_epi16( b08, bp2 );
__m256i rerr = _mm256_mullo_epi16( rdif, _mm256_set1_epi16( 38 ) );
__m256i gerr = _mm256_mullo_epi16( gdif, _mm256_set1_epi16( 76 ) );
__m256i berr = _mm256_mullo_epi16( bdif, _mm256_set1_epi16( 14 ) );
__m256i sum0 = _mm256_add_epi16( rerr, gerr );
__m256i sum1 = _mm256_add_epi16( sum0, berr );
__m256i sum2 = _mm256_madd_epi16( sum1, sum1 );
__m128i sum3 = _mm_add_epi32( _mm256_castsi256_si128( sum2 ), _mm256_extracti128_si256( sum2, 1 ) );
uint32_t err0 = _mm_extract_epi32( sum3, 0 );
uint32_t err1 = _mm_extract_epi32( sum3, 1 );
uint32_t err2 = _mm_extract_epi32( sum3, 2 );
uint32_t err3 = _mm_extract_epi32( sum3, 3 );
error = err0 + err1 + err2 + err3;
}
/**/
uint32_t rgbv = ( rgbv0 & 0x3F ) | ( ( rgbv0 >> 2 ) & 0x1FC0 ) | ( ( rgbv0 >> 3 ) & 0x7E000 );
uint64_t rgbho0_ = ( rgbho & 0x3F0000003F ) | ( ( rgbho >> 2 ) & 0x1FC000001FC0 ) | ( ( rgbho >> 3 ) & 0x7E0000007E000 );
uint64_t rgbho0 = ( rgbho0_ & 0x7FFFF ) | ( ( rgbho0_ >> 13 ) & 0x3FFFF80000 );
uint32_t hi = rgbv | ((rgbho0 & 0x1FFF) << 19);
rgbho0 >>= 13;
uint32_t lo = ( rgbho0 & 0x1 ) | ( ( rgbho0 & 0x1FE ) << 1 ) | ( ( rgbho0 & 0x600 ) << 2 ) | ( ( rgbho0 & 0x3F800 ) << 5 ) | ( ( rgbho0 & 0x1FC0000 ) << 6 );