-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathvec_unit.h
1212 lines (887 loc) · 34.6 KB
/
vec_unit.h
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
/*
* Copyright (C) 2009-2012 Simon A. Berger
*
* This file is part of papara.
*
* papara is free software: you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation, either version 3 of the License, or
* (at your option) any later version.
*
* papara is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with papara. If not, see <http://www.gnu.org/licenses/>.
*/
#ifndef __vec_unit_h
#define __vec_unit_h
#include <iostream>
#include <stdexcept>
#include <cassert>
#include <cstdio>
#include <stdint.h>
// the convenient x86intrin.h is not available on ancient gcc/msvc, so pull in the best manually.
// sse3 is the absolute baseline. sssssssse3 and sse4.666 are nice to have (abs and 32bit min)
// and AVX is a completely different beast...
#ifdef __MMX__
#include <mmintrin.h>
#endif
#ifdef __SSE__
#include <xmmintrin.h>
#endif
#ifdef __SSE2__
#include <emmintrin.h>
#endif
#ifdef __SSE3__
#include <pmmintrin.h>
#endif
#ifdef __SSE4A__
#include <ammintrin.h>
#endif
#if defined (__SSE4_2__) || defined (__SSE4_1__)
#include <smmintrin.h>
#endif
#ifdef __SSSE3__
#include <tmmintrin.h>
#endif
#ifdef __SSE4A__
#include <ammintrin.h>
#endif
#if defined (__SSE4_2__) || defined (__SSE4_1__)
#include <smmintrin.h>
#endif
#ifdef __AVX__
#include <immintrin.h>
#endif
#ifndef _MSC_VER
#include <x86intrin.h>
#endif
#ifdef __AVX__
#define HAVE_AVX
//#include <immintrin.h>
#endif
#ifdef min
#error min defined as macro. this is evil. Please #define NOMINMAX before including any windows headers.
#endif
#ifdef max
#error max defined as macro. this is evil. Please #define NOMINMAX before including any windows headers.
#endif
const size_t required_alignment = 16;
template<class T, size_t W>
struct vector_unit {
};
// vector unit specialization: SSE 8x16bit integer
template<>
struct vector_unit<short, 8> {
const static bool do_checks = false;
typedef __m128i vec_t;
typedef short T;
const static T POS_MAX_VALUE = 0x7fff;
const static T LARGE_VALUE = 32000;
const static T SMALL_VALUE = -32000;
const static T BIAS = 0;
const static size_t W = 8;
static inline vec_t setzero() {
return set1(0);
}
static inline vec_t set1( T val ) {
return _mm_set1_epi16( val );
}
static inline void store( const vec_t &v, T *addr ) {
if( do_checks && addr == 0 ) {
throw std::runtime_error( "store: addr == 0" );
}
//std::cout << "store to: " << addr << "\n";
_mm_store_si128( (vec_t*)addr, v );
}
static inline const vec_t load( const T* addr ) {
return _mm_load_si128( (vec_t*)addr );
}
static inline const vec_t bit_and( const vec_t &a, const vec_t &b ) {
return _mm_and_si128( a, b );
}
static inline const vec_t bit_or( const vec_t &a, const vec_t &b ) {
return _mm_or_si128( a, b );
}
static inline const vec_t bit_andnot( const vec_t &a, const vec_t &b ) {
return _mm_andnot_si128( a, b );
}
static inline const vec_t bit_invert( const vec_t &a ) {
return _mm_xor_si128( a, set1(T(0xffff)) );
}
// static inline const vec_t bit_invert( const vec_t &a ) {
// //return _mm_andnot_pd(a, set1(0xffff));
// }
static inline const vec_t add( const vec_t &a, const vec_t &b ) {
return _mm_add_epi16( a, b );
}
static inline const vec_t adds( const vec_t &a, const vec_t &b ) {
return _mm_adds_epi16( a, b );
}
static inline const vec_t sub( const vec_t &a, const vec_t &b ) {
return _mm_sub_epi16( a, b );
}
static inline const vec_t cmp_zero( const vec_t &a ) {
return _mm_cmpeq_epi16( a, setzero() );
}
static inline const vec_t cmp_eq( const vec_t &a, const vec_t &b ) {
return _mm_cmpeq_epi16( a, b );
}
static inline const vec_t cmp_lt( const vec_t &a, const vec_t &b ) {
return _mm_cmplt_epi16( a, b );
}
static inline const vec_t min( const vec_t &a, const vec_t &b ) {
return _mm_min_epi16( a, b );
}
static inline const vec_t max( const vec_t &a, const vec_t &b ) {
return _mm_max_epi16( a, b );
}
static inline const vec_t abs_diff( const vec_t &a, const vec_t &b ) {
// i don't really like this function, as ideally this would just be abs(sub(a,b)),
// but there doesn't seem to be a fast way to implement abs on pre SSSSSSE3.
// The max(sub(a,b),sub(b,a)) work-around seems to be the next-best thing in this special case.
#ifdef __SSSE3__
return _mm_abs_epi16(sub(a,b));
#else
// FIXME: is there a faster method for boring old CPUs?
return max( sub(a,b), sub(b,a) );
// #error missing SSSSSSSSSE3
#endif
}
static inline void assert_alignment( T * p ) {
assert( size_t(p) % required_alignment == 0 );
}
};
#ifndef __AVX__
// vector unit specialization: future AVX 16x16bit integer
template<>
struct vector_unit<short, 16> {
const static bool do_checks = false;
struct vec_t {
__m128i l;
__m128i h;
vec_t( const __m128i &ll, const __m128i &hh ) : l(ll), h(hh) {}
};
typedef short T;
const static T POS_MAX_VALUE = 0x7fff;
const static T LARGE_VALUE = 32000;
const static T SMALL_VALUE = -32000;
const static T BIAS = 0;
const static size_t W = 16;
static inline vec_t setzero() {
return set1(0);
}
static inline vec_t set1( T val ) {
return vec_t(_mm_set1_epi16( val ), _mm_set1_epi16( val ));
}
static inline void store( const vec_t &v, T *addr ) {
if( do_checks && addr == 0 ) {
throw std::runtime_error( "store: addr == 0" );
}
//std::cout << "store to: " << addr << "\n";
_mm_store_si128( (__m128i*)addr, v.l );
_mm_store_si128( (__m128i*)(addr + 8), v.h );
}
static inline const vec_t load( const T* addr ) {
return vec_t( _mm_load_si128( (__m128i*)addr ), _mm_load_si128( (__m128i*)(addr + 8 )) );
}
static inline const vec_t bit_and( const vec_t &a, const vec_t &b ) {
return vec_t(_mm_and_si128( a.l, b.l ), _mm_and_si128( a.h, b.h ));
}
static inline const vec_t bit_or( const vec_t &a, const vec_t &b ) {
return vec_t(_mm_or_si128( a.l, b.l ), _mm_or_si128( a.h, b.h ));
}
static inline const vec_t bit_andnot( const vec_t &a, const vec_t &b ) {
return vec_t( _mm_andnot_si128( a.l, b.l ), _mm_andnot_si128( a.h, b.h ) );
}
static inline const vec_t bit_invert( const vec_t &a ) {
return vec_t( _mm_xor_si128( a.l, _mm_set1_epi16(T(0xffff)) ), _mm_xor_si128( a.h, _mm_set1_epi16(T(0xffff)) ) );
}
// static inline const vec_t bit_invert( const vec_t &a ) {
// //return _mm_andnot_pd(a, set1(0xffff));
// }
static inline const vec_t add( const vec_t &a, const vec_t &b ) {
return vec_t(_mm_add_epi16( a.l, b.l ), _mm_add_epi16( a.h, b.h ));
}
static inline const vec_t adds( const vec_t &a, const vec_t &b ) {
return vec_t(_mm_adds_epi16( a.l, b.l ), _mm_adds_epi16( a.h, b.h ));
}
static inline const vec_t sub( const vec_t &a, const vec_t &b ) {
return vec_t(_mm_sub_epi16( a.l, b.l ),_mm_sub_epi16( a.h, b.h ));
}
static inline const vec_t cmp_zero( const vec_t &a ) {
return vec_t(_mm_cmpeq_epi16( a.l, _mm_setzero_si128() ), _mm_cmpeq_epi16( a.h, _mm_setzero_si128() ) );
}
static inline const vec_t cmp_eq( const vec_t &a, const vec_t &b ) {
return vec_t(_mm_cmpeq_epi16( a.l, b.l ), _mm_cmpeq_epi16( a.h, b.h ) );
}
static inline const vec_t cmp_lt( const vec_t &a, const vec_t &b ) {
return vec_t(_mm_cmplt_epi16( a.l, b.l ), _mm_cmplt_epi16( a.h, b.h ) );
}
static inline const vec_t min( const vec_t &a, const vec_t &b ) {
return vec_t(_mm_min_epi16( a.l, b.l ),_mm_min_epi16( a.h, b.h ));
}
static inline const vec_t max( const vec_t &a, const vec_t &b ) {
return vec_t(_mm_max_epi16( a.l, b.l ), _mm_max_epi16( a.h, b.h ));
}
static inline const vec_t abs_diff( const vec_t &a, const vec_t &b ) {
// i don't really like this function, as ideally this would just be abs(sub(a,b)),
// but there doesn't seem to be a fast way to implement abs on pre SSSSSSE3.
// The max(sub(a,b),sub(b,a)) work-around seems to be the next-best thing in this special case.
#ifdef __SSSE3__
const vec_t &tmp = sub(a,b);
return vec_t(_mm_abs_epi16(tmp.l), _mm_abs_epi16(tmp.h));
#else
// FIXME: is there a faster method for boring old CPUs?
return max( sub(a,b), sub(b,a) );
// #error missing SSSSSSSSSE3
#endif
}
static inline void assert_alignment( T * p ) {
assert( size_t(p) % required_alignment == 0 );
}
};
#endif
template<>
struct vector_unit<int, 4> {
const static bool do_checks = false;
typedef __m128i vec_t;
typedef int T;
const static T POS_MAX_VALUE = 0x7fffffff;
const static T LARGE_VALUE = 2100000000;
const static T SMALL_VALUE = -2100000000;
const static T BIAS = 0;
const static size_t W = 4;
static inline vec_t setzero() {
return set1(0);
}
static inline vec_t set1( T val ) {
return _mm_set1_epi32( val );
}
static inline void store( const vec_t &v, T *addr ) {
if( do_checks && addr == 0 ) {
throw std::runtime_error( "store: addr == 0" );
}
//std::cout << "store to: " << addr << "\n";
_mm_store_si128( (vec_t*)addr, v );
}
static inline const vec_t load( const T* addr ) {
return _mm_load_si128( (vec_t*)addr );
}
static inline const vec_t bit_and( const vec_t &a, const vec_t &b ) {
return _mm_and_si128( a, b );
}
static inline const vec_t bit_andnot( const vec_t &a, const vec_t &b ) {
return _mm_andnot_si128( a, b );
}
// static inline const vec_t bit_invert( const vec_t &a ) {
// //return _mm_andnot_pd(a, set1(0xffff));
// }
static inline const vec_t add( const vec_t &a, const vec_t &b ) {
return _mm_add_epi32( a, b );
}
static inline const vec_t adds( const vec_t &a, const vec_t &b ) {
// there is no saturating add for 32bit int. Just hope that nothing bad will happen.
// Treat saturation as kind of 'best effort hint'
return _mm_add_epi32( a, b );
}
static inline const vec_t sub( const vec_t &a, const vec_t &b ) {
return _mm_sub_epi32( a, b );
}
static inline const vec_t cmp_zero( const vec_t &a ) {
return _mm_cmpeq_epi32( a, setzero() );
}
static inline const vec_t cmp_eq( const vec_t &a, const vec_t &b ) {
return _mm_cmpeq_epi32( a, b );
}
static inline const vec_t cmp_lt( const vec_t &a, const vec_t &b ) {
return _mm_cmplt_epi32( a, b );
}
static inline const vec_t min( const vec_t &a, const vec_t &b ) {
// sse 4.1, no shit! what were they smoking...
#ifdef __SSE4_1__
return _mm_min_epi32( a, b );
#else
const vec_t ma = _mm_cmplt_epi32( a, b );
return _mm_or_si128( _mm_and_si128( ma, a ), _mm_andnot_si128( ma, b ) );
#endif
}
static inline const vec_t max( const vec_t &a, const vec_t &b ) {
#ifdef __SSE4_1__
return _mm_max_epi32( a, b );
#else
const vec_t ma = _mm_cmpgt_epi32( a, b );
return _mm_or_si128( _mm_and_si128( ma, a ), _mm_andnot_si128( ma, b ) );
#endif
}
static inline const vec_t abs_diff( const vec_t &a, const vec_t &b ) {
// i don't really like this function, as ideally this would just be abs(sub(a,b)),
// but there doesn't seem to be a fast way to implement abs on pre SSSSSSE3.
// The max(sub(a,b),sub(b,a)) work-around seems to be the next-best thing in this special case.
#ifdef __SSSE3__
return _mm_abs_epi32(sub(a,b));
#else
// FIXME: is there a faster method for boring old CPUs?
return max( sub(a,b), sub(b,a) );
// #error missing SSSSSSSSSE3
#endif
}
static inline void assert_alignment( T * p ) {
assert( size_t(p) % required_alignment == 0 );
}
};
#ifndef WIN32
#ifdef __clang__
#pragma clang diagnostic push
#pragma clang diagnostic ignored "-Wgnu"
#endif
template<>
struct vector_unit<float, 4> {
const static bool do_checks = false;
typedef __m128 vec_t;
typedef float T;
const static int SIGN_MASK_INT = 0x7FFFFFFF;
// const static T LARGE_VALUE = 1e8;
// const static T SMALL_VALUE = -1e8;
// const static T BIAS = 0;
const static T LARGE_VALUE;
const static T SMALL_VALUE;
const static T BIAS;
const static size_t W = 4;
static inline vec_t cast_from_int( const __m128i &iv ) {
return _mm_castsi128_ps( iv );
}
static inline vec_t setzero() {
return set1(0);
}
static inline vec_t set1( T val ) {
return _mm_set1_ps( val );
}
static inline void store( const vec_t &v, T *addr ) {
if( do_checks && addr == 0 ) {
throw std::runtime_error( "store: addr == 0" );
}
//std::cout << "store to: " << addr << "\n";
_mm_store_ps( (T*)addr, v );
}
static inline const vec_t load( const T* addr ) {
return _mm_load_ps( (T*)addr );
}
#if 1
static inline const vec_t bit_and( const vec_t &a, const vec_t &b ) {
return _mm_and_ps( a, b );
}
static inline const vec_t bit_andnot( const vec_t &a, const vec_t &b ) {
return _mm_andnot_ps( a, b );
}
#endif
// static inline const vec_t bit_invert( const vec_t &a ) {
// //return _mm_andnot_pd(a, set1(0xffff));
// }
static inline const vec_t add( const vec_t &a, const vec_t &b ) {
return _mm_add_ps( a, b );
}
static inline const vec_t mul( const vec_t &a, const vec_t &b ) {
return _mm_mul_ps( a, b );
}
static inline const vec_t adds( const vec_t &a, const vec_t &b ) {
// float add is always saturating, kind of!?
return _mm_add_ps( a, b );
}
static inline const vec_t sub( const vec_t &a, const vec_t &b ) {
return _mm_sub_ps( a, b );
}
static inline const vec_t cmp_zero( const vec_t &a ) {
return _mm_cmpeq_ps( a, setzero() );
}
static inline const vec_t cmp_eq( const vec_t &a, const vec_t &b ) {
// TODO: think about what this really means. Maybe use something epsilon-based as a default for float?
return _mm_cmpeq_ps( a, b );
}
static inline const vec_t cmp_lt( const vec_t &a, const vec_t &b ) {
return _mm_cmplt_ps( a, b );
}
static inline const vec_t min( const vec_t &a, const vec_t &b ) {
return _mm_min_ps( a, b );
}
static inline const vec_t max( const vec_t &a, const vec_t &b ) {
return _mm_max_ps( a, b );
}
static inline const vec_t abs_diff( const vec_t &a, const vec_t &b ) {
// i don't really like this function, as ideally this would just be abs(sub(a,b)),
// but there doesn't seem to be a fast way to implement abs on pre SSSSSSE3.
// The max(sub(a,b),sub(b,a)) work-around seems to be the next-best thing in this special case.
//unsigned int SIGN_MASK[4] = {0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF};
//unsigned int SIGN_MASK = 0x7FFFFFFF;
const float *SIGN_MASK_PTR = (float*)&SIGN_MASK_INT;
static float SIGN_MASK = *SIGN_MASK_PTR;
return bit_and(sub(a,b), set1(SIGN_MASK) ); // TODO: could this case any alignment problems?
//return bit_and(sub(a,b), set1(*((float*)&SIGN_MASK_INT) )); // TODO: could this case any alignment problems?
//return bit_and(sub(a,b), load((float*)SIGN_MASK ));
}
};
const vector_unit<float,4>::T vector_unit<float,4>::LARGE_VALUE = 1e8;
const vector_unit<float,4>::T vector_unit<float,4>::SMALL_VALUE = -1e8;
const vector_unit<float,4>::T vector_unit<float,4>::BIAS = 0;
template<>
struct vector_unit<double, 2> {
const static bool do_checks = false;
typedef __m128d vec_t;
typedef double T;
// const static uint64_t SIGN_MASK_U64 = 0x7FFFFFFFFFFFFFFF;
const static T LARGE_VALUE;// = 1e8;
const static T SMALL_VALUE;// = -1e8;
const static T BIAS;// = 0;
const static size_t W = 2;
static inline vec_t cast_from_int( const __m128i &iv ) {
return _mm_castsi128_pd( iv );
}
static inline vec_t setzero() {
return set1(0);
}
static inline vec_t set1( T val ) {
return _mm_set1_pd( val );
}
static inline void store( const vec_t &v, T *addr ) {
if( do_checks && addr == 0 ) {
throw std::runtime_error( "store: addr == 0" );
}
//std::cout << "store to: " << addr << "\n";
_mm_store_pd( (T*)addr, v );
}
static inline const vec_t load( const T* addr ) {
return _mm_load_pd( (T*)addr );
}
#if 1
static inline const vec_t bit_and( const vec_t &a, const vec_t &b ) {
return _mm_and_pd( a, b );
}
static inline const vec_t bit_or( const vec_t &a, const vec_t &b ) {
return _mm_or_pd( a, b );
}
static inline const vec_t bit_andnot( const vec_t &a, const vec_t &b ) {
return _mm_andnot_pd( a, b );
}
#endif
// static inline const vec_t bit_invert( const vec_t &a ) {
// //return _mm_andnot_pd(a, set1(0xffff));
// }
static inline const vec_t add( const vec_t &a, const vec_t &b ) {
return _mm_add_pd( a, b );
}
static inline const vec_t mul( const vec_t &a, const vec_t &b ) {
return _mm_mul_pd( a, b );
}
static inline const vec_t adds( const vec_t &a, const vec_t &b ) {
// float add is always saturating, kind of!?
return _mm_add_pd( a, b );
}
static inline const vec_t sub( const vec_t &a, const vec_t &b ) {
return _mm_sub_pd( a, b );
}
static inline const vec_t cmp_zero( const vec_t &a ) {
return _mm_cmpeq_pd( a, setzero() );
}
static inline const vec_t cmp_eq( const vec_t &a, const vec_t &b ) {
// TODO: think about what this really means. Maybe use something epsilon-based as a default for float?
return _mm_cmpeq_pd( a, b );
}
static inline const vec_t cmp_lt( const vec_t &a, const vec_t &b ) {
return _mm_cmplt_pd( a, b );
}
static inline const vec_t min( const vec_t &a, const vec_t &b ) {
return _mm_min_pd( a, b );
}
static inline const vec_t max( const vec_t &a, const vec_t &b ) {
return _mm_max_pd( a, b );
}
// deactivated for now, beacuse of the missing ULL suffix in pre c++11
#if 0
static inline const vec_t abs_diff( const vec_t &a, const vec_t &b ) {
// i don't really like this function, as ideally this would just be abs(sub(a,b)),
// but there doesn't seem to be a fast way to implement abs on pre SSSSSSE3.
// The max(sub(a,b),sub(b,a)) work-around seems to be the next-best thing in this special case.
//unsigned int SIGN_MASK[4] = {0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF};
//unsigned int SIGN_MASK = 0x7FFFFFFF;
const uint64_t SIGN_MASK_U64x = 0x7fffffffffffffffULL;
const double *SIGN_MASK_PTR = (double*)&SIGN_MASK_U64x;
double SIGN_MASK = *SIGN_MASK_PTR;
return bit_and(sub(a,b), set1(SIGN_MASK) ); // TODO: could this cause any alignment problems?
//return bit_and(sub(a,b), set1(*((float*)&SIGN_MASK_INT) )); // TODO: could this case any alignment problems?
//return bit_and(sub(a,b), load((float*)SIGN_MASK ));
}
#endif
static inline const vec_t abs( const vec_t &a ) {
// const uint64_t SIGN_MASK_U64x = 0x7fffffffffffffff;
//
// const static double *SIGN_MASK_PTR = (double*)&SIGN_MASK_U64x;
// static double SIGN_MASK = *SIGN_MASK_PTR;
//
// return bit_and( a, set1(SIGN_MASK));
//
//
return max( a, sub( setzero(), a ));
}
static inline void println( const vec_t & v ) {
double tmp[2];
_mm_storeu_pd( tmp, v );
printf( "%f %f\n", tmp[0], tmp[1] );
}
};
const double vector_unit<double,2>::LARGE_VALUE = 1e8;
const double vector_unit<double,2>::SMALL_VALUE = -1e8;
const double vector_unit<double,2>::BIAS = 0;
#ifdef __clang__
#pragma clang diagnostic pop
#endif
#endif
#ifdef HAVE_AVX
// AVX is the most pointless thing in the world, as far as integers are concerned.
// waiting for AVX5.7
// AVX 16x16bit vector unit
template<>
struct vector_unit<short, 16> {
const static bool do_checks = false;
typedef __m256i vec_t;
typedef short T;
const static T SMALL_VALUE = -32000;
const static T BIAS = 0;
const static size_t W = 8;
static inline vec_t setzero() {
return _mm256_setzero_si256();
}
static inline vec_t set1( T val ) {
return _mm256_set1_epi16( val );
}
static inline void store( const vec_t &v, T *addr ) {
if( do_checks && addr == 0 ) {
throw std::runtime_error( "store: addr == 0" );
}
//std::cout << "store to: " << addr << "\n";
_mm256_store_si256( (vec_t*)addr, v );
}
static inline const vec_t load( T* addr ) {
return _mm256_load_si256( (vec_t*)addr );
}
static inline const vec_t bit_and( const vec_t &a, const vec_t &b ) {
__m128i lowa = _mm256_castsi256_si128(a);
const __m128i lowb = _mm256_castsi256_si128(b);
lowa = _mm_and_si128( lowa, lowb );
__m128i higha = _mm256_extractf128_si256(a,1);
const __m128i highb = _mm256_extractf128_si256(b,1);
higha = _mm_and_si128( higha, highb );
return _mm256_insertf128_si256( _mm256_castsi128_si256(lowa), higha, 1 );
}
static inline const vec_t bit_andnot( const vec_t &a, const vec_t &b ) {
__m128i lowa = _mm256_castsi256_si128(a);
const __m128i lowb = _mm256_castsi256_si128(b);
lowa = _mm_andnot_si128( lowa, lowb );
__m128i higha = _mm256_extractf128_si256(a,1);
const __m128i highb = _mm256_extractf128_si256(b,1);
higha = _mm_andnot_si128( higha, highb );
return _mm256_insertf128_si256( _mm256_castsi128_si256(lowa), higha, 1 );
}
// static inline const vec_t bit_invert( const vec_t &a ) {
// //return _mm_andnot_pd(a, set1(0xffff));
// }
static inline const vec_t add( const vec_t &a, const vec_t &b ) {
const __m128i lowa = _mm256_extractf128_si256(a,0);
const __m128i lowb = _mm256_extractf128_si256(b,0);
//lowa = _mm_add_epi16( lowa, lowb );
const __m128i higha = _mm256_extractf128_si256(a,1);
const __m128i highb = _mm256_extractf128_si256(b,1);
//higha = _mm_add_epi16( higha, highb );
return _mm256_insertf128_si256( _mm256_castsi128_si256(_mm_add_epi16( lowa, lowb )), _mm_add_epi16( higha, highb ), 1 );
}
static inline const vec_t sub( const vec_t &a, const vec_t &b ) {
__m128i lowa = _mm256_castsi256_si128(a);
const __m128i lowb = _mm256_castsi256_si128(b);
lowa = _mm_sub_epi16( lowa, lowb );
__m128i higha = _mm256_extractf128_si256(a,1);
const __m128i highb = _mm256_extractf128_si256(b,1);
higha = _mm_sub_epi16( higha, highb );
return _mm256_insertf128_si256( _mm256_castsi128_si256(lowa), higha, 1 );
}
static inline const vec_t cmp_zero( const vec_t &a ) {
return cmp_eq( a, setzero() );
}
static inline const vec_t cmp_eq( const vec_t &a, const vec_t &b ) {
__m128i lowa = _mm256_castsi256_si128(a);
const __m128i lowb = _mm256_castsi256_si128(b);
lowa = _mm_cmpeq_epi16( lowa, lowb );
__m128i higha = _mm256_extractf128_si256(a,1);
const __m128i highb = _mm256_extractf128_si256(b,1);
higha = _mm_cmpeq_epi16( higha, highb );
return _mm256_insertf128_si256( _mm256_castsi128_si256(lowa), higha, 1 );
}
static inline const vec_t cmp_lt( const vec_t &a, const vec_t &b ) {
__m128i lowa = _mm256_castsi256_si128(a);
const __m128i lowb = _mm256_castsi256_si128(b);
lowa = _mm_cmplt_epi16( lowa, lowb );
__m128i higha = _mm256_extractf128_si256(a,1);
const __m128i highb = _mm256_extractf128_si256(b,1);
higha = _mm_cmplt_epi16( higha, highb );
return _mm256_insertf128_si256( _mm256_castsi128_si256(lowa), higha, 1 );
}
static inline const vec_t min( const vec_t &a, const vec_t &b ) {
__m128i lowa = _mm256_castsi256_si128(a);
const __m128i lowb = _mm256_castsi256_si128(b);
lowa = _mm_min_epi16( lowa, lowb );
__m128i higha = _mm256_extractf128_si256(a,1);
const __m128i highb = _mm256_extractf128_si256(b,1);
higha = _mm_min_epi16( higha, highb );
return _mm256_insertf128_si256( _mm256_castsi128_si256(lowa), higha, 1 );
}
static inline const vec_t max( const vec_t &a, const vec_t &b ) {
__m128i lowa = _mm256_castsi256_si128(a);
const __m128i lowb = _mm256_castsi256_si128(b);
lowa = _mm_max_epi16( lowa, lowb );
__m128i higha = _mm256_extractf128_si256(a,1);
const __m128i highb = _mm256_extractf128_si256(b,1);
higha = _mm_max_epi16( higha, highb );
return _mm256_insertf128_si256( _mm256_castsi128_si256(lowa), higha, 1 );
}
};
template<>
struct vector_unit<double, 4> {
const static bool do_checks = false;
typedef __m256d vec_t;
typedef double T;
// const static uint64_t SIGN_MASK_U64 = 0x7FFFFFFFFFFFFFFF;
const static T LARGE_VALUE = 1e8;
const static T SMALL_VALUE = -1e8;
const static T BIAS = 0;
const static size_t W = 2;
static inline vec_t setzero() {
return _mm256_setzero_pd();
}
static inline vec_t set1( T val ) {
return _mm256_set1_pd( val );
}
static inline void store( const vec_t &v, T *addr ) {
if( do_checks && addr == 0 ) {
throw std::runtime_error( "store: addr == 0" );
}
//std::cout << "store to: " << addr << "\n";
_mm256_store_pd( (T*)addr, v );
}
static inline const vec_t load( const T* addr ) {
return _mm256_load_pd( (T*)addr );
}
#if 1
static inline const vec_t bit_and( const vec_t &a, const vec_t &b ) {
return _mm256_and_pd( a, b );
}
static inline const vec_t bit_or( const vec_t &a, const vec_t &b ) {
return _mm256_or_pd( a, b );
}
static inline const vec_t bit_andnot( const vec_t &a, const vec_t &b ) {
return _mm256_andnot_pd( a, b );
}
#endif
// static inline const vec_t bit_invert( const vec_t &a ) {
// //return _mm_andnot_pd(a, set1(0xffff));
// }
static inline const vec_t add( const vec_t &a, const vec_t &b ) {
return _mm256_add_pd( a, b );
}
static inline const vec_t mul( const vec_t &a, const vec_t &b ) {
return _mm256_mul_pd( a, b );
}
static inline const vec_t adds( const vec_t &a, const vec_t &b ) {
// float add is always saturating, kind of!?
return _mm256_add_pd( a, b );
}
static inline const vec_t sub( const vec_t &a, const vec_t &b ) {
return _mm256_sub_pd( a, b );
}
static inline const vec_t cmp_zero( const vec_t &a ) {
return _mm256_cmp_pd( a, setzero(), _CMP_EQ_OQ );
}
static inline const vec_t cmp_eq( const vec_t &a, const vec_t &b ) {
// TODO: think about what this really means. Maybe use something epsilon-based as a default for float?
return _mm256_cmp_pd( a, b, _CMP_EQ_OQ );
}
static inline const vec_t cmp_lt( const vec_t &a, const vec_t &b ) {
return _mm256_cmp_pd( a, b, _CMP_LT_OQ );
}
static inline const vec_t min( const vec_t &a, const vec_t &b ) {
return _mm256_min_pd( a, b );
}
static inline const vec_t max( const vec_t &a, const vec_t &b ) {
return _mm256_max_pd( a, b );
}
static inline const vec_t abs_diff( const vec_t &a, const vec_t &b ) {
// i don't really like this function, as ideally this would just be abs(sub(a,b)),
// but there doesn't seem to be a fast way to implement abs on pre SSSSSSE3.
// The max(sub(a,b),sub(b,a)) work-around seems to be the next-best thing in this special case.
//unsigned int SIGN_MASK[4] = {0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF};
//unsigned int SIGN_MASK = 0x7FFFFFFF;
const uint64_t SIGN_MASK_U64x = 0x7fffffffffffffff;
const double *SIGN_MASK_PTR = (double*)&SIGN_MASK_U64x;
double SIGN_MASK = *SIGN_MASK_PTR;