-
Notifications
You must be signed in to change notification settings - Fork 549
/
PacketMath.h
executable file
·1283 lines (1140 loc) · 49.4 KB
/
PacketMath.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
// This file is part of Eigen, a lightweight C++ template library
// for linear algebra.
//
// Copyright (C) 2008-2009 Gael Guennebaud <[email protected]>
//
// This Source Code Form is subject to the terms of the Mozilla
// Public License v. 2.0. If a copy of the MPL was not distributed
// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
#ifndef EIGEN_PACKET_MATH_SSE_H
#define EIGEN_PACKET_MATH_SSE_H
namespace Eigen {
namespace internal {
#ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8
#endif
#if !defined(EIGEN_VECTORIZE_AVX) && !defined(EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS)
// 32 bits => 8 registers
// 64 bits => 16 registers
#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS (2*sizeof(void*))
#endif
#ifdef EIGEN_VECTORIZE_FMA
#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD
#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD 1
#endif
#endif
#if ((defined EIGEN_VECTORIZE_AVX) && (EIGEN_COMP_GNUC_STRICT || EIGEN_COMP_MINGW) && (__GXX_ABI_VERSION < 1004)) || EIGEN_OS_QNX
// With GCC's default ABI version, a __m128 or __m256 are the same types and therefore we cannot
// have overloads for both types without linking error.
// One solution is to increase ABI version using -fabi-version=4 (or greater).
// Otherwise, we workaround this inconvenience by wrapping 128bit types into the following helper
// structure:
template<typename T>
struct eigen_packet_wrapper
{
EIGEN_ALWAYS_INLINE operator T&() { return m_val; }
EIGEN_ALWAYS_INLINE operator const T&() const { return m_val; }
EIGEN_ALWAYS_INLINE eigen_packet_wrapper() {}
EIGEN_ALWAYS_INLINE eigen_packet_wrapper(const T &v) : m_val(v) {}
EIGEN_ALWAYS_INLINE eigen_packet_wrapper& operator=(const T &v) {
m_val = v;
return *this;
}
T m_val;
};
typedef eigen_packet_wrapper<__m128> Packet4f;
typedef eigen_packet_wrapper<__m128i> Packet4i;
typedef eigen_packet_wrapper<__m128d> Packet2d;
#else
typedef __m128 Packet4f;
typedef __m128i Packet4i;
typedef __m128d Packet2d;
#endif
template<> struct is_arithmetic<__m128> { enum { value = true }; };
template<> struct is_arithmetic<__m128i> { enum { value = true }; };
template<> struct is_arithmetic<__m128d> { enum { value = true }; };
#define EIGEN_SSE_SHUFFLE_MASK(p,q,r,s) ((s)<<6|(r)<<4|(q)<<2|(p))
#define vec4f_swizzle1(v,p,q,r,s) \
(_mm_castsi128_ps(_mm_shuffle_epi32( _mm_castps_si128(v), EIGEN_SSE_SHUFFLE_MASK(p,q,r,s))))
#define vec4i_swizzle1(v,p,q,r,s) \
(_mm_shuffle_epi32( v, EIGEN_SSE_SHUFFLE_MASK(p,q,r,s)))
#define vec2d_swizzle1(v,p,q) \
(_mm_castsi128_pd(_mm_shuffle_epi32( _mm_castpd_si128(v), EIGEN_SSE_SHUFFLE_MASK(2*p,2*p+1,2*q,2*q+1))))
#define vec4f_swizzle2(a,b,p,q,r,s) \
(_mm_shuffle_ps( (a), (b), EIGEN_SSE_SHUFFLE_MASK(p,q,r,s)))
#define vec4i_swizzle2(a,b,p,q,r,s) \
(_mm_castps_si128( (_mm_shuffle_ps( _mm_castsi128_ps(a), _mm_castsi128_ps(b), EIGEN_SSE_SHUFFLE_MASK(p,q,r,s)))))
#define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \
const Packet4f p4f_##NAME = pset1<Packet4f>(X)
#define _EIGEN_DECLARE_CONST_Packet2d(NAME,X) \
const Packet2d p2d_##NAME = pset1<Packet2d>(X)
#define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \
const Packet4f p4f_##NAME = pset1frombits<Packet4f>(X)
#define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
const Packet4i p4i_##NAME = pset1<Packet4i>(X)
// Use the packet_traits defined in AVX/PacketMath.h instead if we're going
// to leverage AVX instructions.
#ifndef EIGEN_VECTORIZE_AVX
template <>
struct packet_traits<float> : default_packet_traits {
typedef Packet4f type;
typedef Packet4f half;
enum {
Vectorizable = 1,
AlignedOnScalar = 1,
size = 4,
HasHalfPacket = 0,
HasDiv = 1,
HasSin = EIGEN_FAST_MATH,
HasCos = EIGEN_FAST_MATH,
HasLog = 1,
HasLog1p = 1,
HasExpm1 = 1,
HasNdtri = 1,
HasExp = 1,
HasBessel = 1,
HasSqrt = 1,
HasRsqrt = 1,
HasTanh = EIGEN_FAST_MATH,
HasErf = EIGEN_FAST_MATH,
HasBlend = 1,
HasFloor = 1
#ifdef EIGEN_VECTORIZE_SSE4_1
,
HasRound = 1,
HasCeil = 1
#endif
};
};
template <>
struct packet_traits<double> : default_packet_traits {
typedef Packet2d type;
typedef Packet2d half;
enum {
Vectorizable = 1,
AlignedOnScalar = 1,
size=2,
HasHalfPacket = 0,
HasDiv = 1,
HasExp = 1,
HasSqrt = 1,
HasRsqrt = 1,
HasBlend = 1
#ifdef EIGEN_VECTORIZE_SSE4_1
,
HasRound = 1,
HasFloor = 1,
HasCeil = 1
#endif
};
};
#endif
template<> struct packet_traits<int> : default_packet_traits
{
typedef Packet4i type;
typedef Packet4i half;
enum {
Vectorizable = 1,
AlignedOnScalar = 1,
size=4,
HasBlend = 1
};
};
template<> struct unpacket_traits<Packet4f> {
typedef float type;
typedef Packet4f half;
typedef Packet4i integer_packet;
enum {size=4, alignment=Aligned16, vectorizable=true, masked_load_available=false, masked_store_available=false};
};
template<> struct unpacket_traits<Packet2d> {
typedef double type;
typedef Packet2d half;
enum {size=2, alignment=Aligned16, vectorizable=true, masked_load_available=false, masked_store_available=false};
};
template<> struct unpacket_traits<Packet4i> {
typedef int type;
typedef Packet4i half;
enum {size=4, alignment=Aligned16, vectorizable=false, masked_load_available=false, masked_store_available=false};
};
#ifndef EIGEN_VECTORIZE_AVX
template<> struct scalar_div_cost<float,true> { enum { value = 7 }; };
template<> struct scalar_div_cost<double,true> { enum { value = 8 }; };
#endif
#if EIGEN_COMP_MSVC==1500
// Workaround MSVC 9 internal compiler error.
// TODO: It has been detected with win64 builds (amd64), so let's check whether it also happens in 32bits+SSE mode
// TODO: let's check whether there does not exist a better fix, like adding a pset0() function. (it crashed on pset1(0)).
template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) { return _mm_set_ps(from,from,from,from); }
template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) { return _mm_set_pd(from,from); }
template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int& from) { return _mm_set_epi32(from,from,from,from); }
#else
template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float& from) { return _mm_set_ps1(from); }
template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) { return _mm_set1_pd(from); }
template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int& from) { return _mm_set1_epi32(from); }
#endif
template<> EIGEN_STRONG_INLINE Packet4f pset1frombits<Packet4f>(unsigned int from) { return _mm_castsi128_ps(pset1<Packet4i>(from)); }
template<> EIGEN_STRONG_INLINE Packet4f pzero(const Packet4f& /*a*/) { return _mm_setzero_ps(); }
template<> EIGEN_STRONG_INLINE Packet2d pzero(const Packet2d& /*a*/) { return _mm_setzero_pd(); }
template<> EIGEN_STRONG_INLINE Packet4i pzero(const Packet4i& /*a*/) { return _mm_setzero_si128(); }
// GCC generates a shufps instruction for _mm_set1_ps/_mm_load1_ps instead of the more efficient pshufd instruction.
// However, using inrinsics for pset1 makes gcc to generate crappy code in some cases (see bug 203)
// Using inline assembly is also not an option because then gcc fails to reorder properly the instructions.
// Therefore, we introduced the pload1 functions to be used in product kernels for which bug 203 does not apply.
// Also note that with AVX, we want it to generate a vbroadcastss.
#if EIGEN_COMP_GNUC_STRICT && (!defined __AVX__)
template<> EIGEN_STRONG_INLINE Packet4f pload1<Packet4f>(const float *from) {
return vec4f_swizzle1(_mm_load_ss(from),0,0,0,0);
}
#endif
template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a) { return _mm_add_ps(pset1<Packet4f>(a), _mm_set_ps(3,2,1,0)); }
template<> EIGEN_STRONG_INLINE Packet2d plset<Packet2d>(const double& a) { return _mm_add_pd(pset1<Packet2d>(a),_mm_set_pd(1,0)); }
template<> EIGEN_STRONG_INLINE Packet4i plset<Packet4i>(const int& a) { return _mm_add_epi32(pset1<Packet4i>(a),_mm_set_epi32(3,2,1,0)); }
template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_add_ps(a,b); }
template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_add_pd(a,b); }
template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_add_epi32(a,b); }
template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_sub_ps(a,b); }
template<> EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_sub_pd(a,b); }
template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_sub_epi32(a,b); }
template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a)
{
const Packet4f mask = _mm_castsi128_ps(_mm_setr_epi32(0x80000000,0x80000000,0x80000000,0x80000000));
return _mm_xor_ps(a,mask);
}
template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a)
{
const Packet2d mask = _mm_castsi128_pd(_mm_setr_epi32(0x0,0x80000000,0x0,0x80000000));
return _mm_xor_pd(a,mask);
}
template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a)
{
return psub(Packet4i(_mm_setr_epi32(0,0,0,0)), a);
}
template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; }
template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; }
template<> EIGEN_STRONG_INLINE Packet4i pconj(const Packet4i& a) { return a; }
template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_mul_ps(a,b); }
template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_mul_pd(a,b); }
template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b)
{
#ifdef EIGEN_VECTORIZE_SSE4_1
return _mm_mullo_epi32(a,b);
#else
// this version is slightly faster than 4 scalar products
return vec4i_swizzle1(
vec4i_swizzle2(
_mm_mul_epu32(a,b),
_mm_mul_epu32(vec4i_swizzle1(a,1,0,3,2),
vec4i_swizzle1(b,1,0,3,2)),
0,2,0,2),
0,2,1,3);
#endif
}
template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_div_ps(a,b); }
template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_div_pd(a,b); }
// for some weird raisons, it has to be overloaded for packet of integers
template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return padd(pmul(a,b), c); }
#ifdef EIGEN_VECTORIZE_FMA
template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return _mm_fmadd_ps(a,b,c); }
template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return _mm_fmadd_pd(a,b,c); }
#endif
#ifdef EIGEN_VECTORIZE_SSE4_1
template<> EIGEN_DEVICE_FUNC inline Packet4f pselect(const Packet4f& mask, const Packet4f& a, const Packet4f& b) { return _mm_blendv_ps(b,a,mask); }
template<> EIGEN_DEVICE_FUNC inline Packet2d pselect(const Packet2d& mask, const Packet2d& a, const Packet2d& b) { return _mm_blendv_pd(b,a,mask); }
#endif
template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) {
#if EIGEN_COMP_GNUC && EIGEN_COMP_GNUC < 63
// There appears to be a bug in GCC, by which the optimizer may
// flip the argument order in calls to _mm_min_ps, so we have to
// resort to inline ASM here. This is supposed to be fixed in gcc6.3,
// see also: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72867
#ifdef EIGEN_VECTORIZE_AVX
Packet4f res;
asm("vminps %[a], %[b], %[res]" : [res] "=x" (res) : [a] "x" (a), [b] "x" (b));
#else
Packet4f res = b;
asm("minps %[a], %[res]" : [res] "+x" (res) : [a] "x" (a));
#endif
return res;
#else
// Arguments are reversed to match NaN propagation behavior of std::min.
return _mm_min_ps(b, a);
#endif
}
template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) {
#if EIGEN_COMP_GNUC && EIGEN_COMP_GNUC < 63
// There appears to be a bug in GCC, by which the optimizer may
// flip the argument order in calls to _mm_min_pd, so we have to
// resort to inline ASM here. This is supposed to be fixed in gcc6.3,
// see also: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72867
#ifdef EIGEN_VECTORIZE_AVX
Packet2d res;
asm("vminpd %[a], %[b], %[res]" : [res] "=x" (res) : [a] "x" (a), [b] "x" (b));
#else
Packet2d res = b;
asm("minpd %[a], %[res]" : [res] "+x" (res) : [a] "x" (a));
#endif
return res;
#else
// Arguments are reversed to match NaN propagation behavior of std::min.
return _mm_min_pd(b, a);
#endif
}
template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b)
{
#ifdef EIGEN_VECTORIZE_SSE4_1
return _mm_min_epi32(a,b);
#else
// after some bench, this version *is* faster than a scalar implementation
Packet4i mask = _mm_cmplt_epi32(a,b);
return _mm_or_si128(_mm_and_si128(mask,a),_mm_andnot_si128(mask,b));
#endif
}
template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) {
#if EIGEN_COMP_GNUC && EIGEN_COMP_GNUC < 63
// There appears to be a bug in GCC, by which the optimizer may
// flip the argument order in calls to _mm_max_ps, so we have to
// resort to inline ASM here. This is supposed to be fixed in gcc6.3,
// see also: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72867
#ifdef EIGEN_VECTORIZE_AVX
Packet4f res;
asm("vmaxps %[a], %[b], %[res]" : [res] "=x" (res) : [a] "x" (a), [b] "x" (b));
#else
Packet4f res = b;
asm("maxps %[a], %[res]" : [res] "+x" (res) : [a] "x" (a));
#endif
return res;
#else
// Arguments are reversed to match NaN propagation behavior of std::max.
return _mm_max_ps(b, a);
#endif
}
template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) {
#if EIGEN_COMP_GNUC && EIGEN_COMP_GNUC < 63
// There appears to be a bug in GCC, by which the optimizer may
// flip the argument order in calls to _mm_max_pd, so we have to
// resort to inline ASM here. This is supposed to be fixed in gcc6.3,
// see also: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72867
#ifdef EIGEN_VECTORIZE_AVX
Packet2d res;
asm("vmaxpd %[a], %[b], %[res]" : [res] "=x" (res) : [a] "x" (a), [b] "x" (b));
#else
Packet2d res = b;
asm("maxpd %[a], %[res]" : [res] "+x" (res) : [a] "x" (a));
#endif
return res;
#else
// Arguments are reversed to match NaN propagation behavior of std::max.
return _mm_max_pd(b, a);
#endif
}
template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b)
{
#ifdef EIGEN_VECTORIZE_SSE4_1
return _mm_max_epi32(a,b);
#else
// after some bench, this version *is* faster than a scalar implementation
Packet4i mask = _mm_cmpgt_epi32(a,b);
return _mm_or_si128(_mm_and_si128(mask,a),_mm_andnot_si128(mask,b));
#endif
}
template<> EIGEN_STRONG_INLINE Packet4f pcmp_le(const Packet4f& a, const Packet4f& b) { return _mm_cmple_ps(a,b); }
template<> EIGEN_STRONG_INLINE Packet4f pcmp_lt(const Packet4f& a, const Packet4f& b) { return _mm_cmplt_ps(a,b); }
template<> EIGEN_STRONG_INLINE Packet4f pcmp_eq(const Packet4f& a, const Packet4f& b) { return _mm_cmpeq_ps(a,b); }
template<> EIGEN_STRONG_INLINE Packet4i pcmp_eq(const Packet4i& a, const Packet4i& b) { return _mm_cmpeq_epi32(a,b); }
template<> EIGEN_STRONG_INLINE Packet2d pcmp_eq(const Packet2d& a, const Packet2d& b) { return _mm_cmpeq_pd(a,b); }
template<> EIGEN_STRONG_INLINE Packet4f pcmp_lt_or_nan(const Packet4f& a, const Packet4f& b) { return _mm_cmpnge_ps(a,b); }
template<> EIGEN_STRONG_INLINE Packet4i ptrue<Packet4i>(const Packet4i& a) { return _mm_cmpeq_epi32(a, a); }
template<> EIGEN_STRONG_INLINE Packet4f
ptrue<Packet4f>(const Packet4f& a) {
Packet4i b = _mm_castps_si128(a);
return _mm_castsi128_ps(_mm_cmpeq_epi32(b, b));
}
template<> EIGEN_STRONG_INLINE Packet2d
ptrue<Packet2d>(const Packet2d& a) {
Packet4i b = _mm_castpd_si128(a);
return _mm_castsi128_pd(_mm_cmpeq_epi32(b, b));
}
template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_and_ps(a,b); }
template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_and_pd(a,b); }
template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_and_si128(a,b); }
template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_or_ps(a,b); }
template<> EIGEN_STRONG_INLINE Packet2d por<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_or_pd(a,b); }
template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_or_si128(a,b); }
template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_xor_ps(a,b); }
template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_xor_pd(a,b); }
template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_xor_si128(a,b); }
template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_andnot_ps(b,a); }
template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_andnot_pd(b,a); }
template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_andnot_si128(b,a); }
template<int N> EIGEN_STRONG_INLINE Packet4i pshiftright(Packet4i a) { return _mm_srli_epi32(a,N); }
template<int N> EIGEN_STRONG_INLINE Packet4i pshiftleft(Packet4i a) { return _mm_slli_epi32(a,N); }
#ifdef EIGEN_VECTORIZE_SSE4_1
template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a) { return _mm_round_ps(a, 0); }
template<> EIGEN_STRONG_INLINE Packet2d pround<Packet2d>(const Packet2d& a) { return _mm_round_pd(a, 0); }
template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(const Packet4f& a) { return _mm_ceil_ps(a); }
template<> EIGEN_STRONG_INLINE Packet2d pceil<Packet2d>(const Packet2d& a) { return _mm_ceil_pd(a); }
template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a) { return _mm_floor_ps(a); }
template<> EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(const Packet2d& a) { return _mm_floor_pd(a); }
#else
template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a)
{
const Packet4f cst_1 = pset1<Packet4f>(1.0f);
Packet4i emm0 = _mm_cvttps_epi32(a);
Packet4f tmp = _mm_cvtepi32_ps(emm0);
/* if greater, substract 1 */
Packet4f mask = _mm_cmpgt_ps(tmp, a);
mask = pand(mask, cst_1);
return psub(tmp, mask);
}
// WARNING: this pfloor implementation makes sense for small inputs only,
// It is currently only used by pexp and not exposed through HasFloor.
template<> EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(const Packet2d& a)
{
const Packet2d cst_1 = pset1<Packet2d>(1.0);
Packet4i emm0 = _mm_cvttpd_epi32(a);
Packet2d tmp = _mm_cvtepi32_pd(emm0);
/* if greater, substract 1 */
Packet2d mask = _mm_cmpgt_pd(tmp, a);
mask = pand(mask, cst_1);
return psub(tmp, mask);
}
#endif
template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_ps(from); }
template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(const double* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_pd(from); }
template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int* from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_si128(reinterpret_cast<const __m128i*>(from)); }
#if EIGEN_COMP_MSVC
template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from) {
EIGEN_DEBUG_UNALIGNED_LOAD
#if (EIGEN_COMP_MSVC==1600)
// NOTE Some version of MSVC10 generates bad code when using _mm_loadu_ps
// (i.e., it does not generate an unaligned load!!
__m128 res = _mm_loadl_pi(_mm_set1_ps(0.0f), (const __m64*)(from));
res = _mm_loadh_pi(res, (const __m64*)(from+2));
return res;
#else
return _mm_loadu_ps(from);
#endif
}
#else
// NOTE: with the code below, MSVC's compiler crashes!
template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from)
{
EIGEN_DEBUG_UNALIGNED_LOAD
return _mm_loadu_ps(from);
}
#endif
template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from)
{
EIGEN_DEBUG_UNALIGNED_LOAD
return _mm_loadu_pd(from);
}
template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from)
{
EIGEN_DEBUG_UNALIGNED_LOAD
return _mm_loadu_si128(reinterpret_cast<const __m128i*>(from));
}
template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from)
{
return vec4f_swizzle1(_mm_castpd_ps(_mm_load_sd(reinterpret_cast<const double*>(from))), 0, 0, 1, 1);
}
template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double* from)
{ return pset1<Packet2d>(from[0]); }
template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int* from)
{
Packet4i tmp;
tmp = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(from));
return vec4i_swizzle1(tmp, 0, 0, 1, 1);
}
template<> EIGEN_STRONG_INLINE void pstore<float>(float* to, const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_ps(to, from); }
template<> EIGEN_STRONG_INLINE void pstore<double>(double* to, const Packet2d& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_pd(to, from); }
template<> EIGEN_STRONG_INLINE void pstore<int>(int* to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_si128(reinterpret_cast<__m128i*>(to), from); }
template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet2d& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_pd(to, from); }
template<> EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_ps(to, from); }
template<> EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_si128(reinterpret_cast<__m128i*>(to), from); }
template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride)
{
return _mm_set_ps(from[3*stride], from[2*stride], from[1*stride], from[0*stride]);
}
template<> EIGEN_DEVICE_FUNC inline Packet2d pgather<double, Packet2d>(const double* from, Index stride)
{
return _mm_set_pd(from[1*stride], from[0*stride]);
}
template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int, Packet4i>(const int* from, Index stride)
{
return _mm_set_epi32(from[3*stride], from[2*stride], from[1*stride], from[0*stride]);
}
template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride)
{
to[stride*0] = _mm_cvtss_f32(from);
to[stride*1] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 1));
to[stride*2] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 2));
to[stride*3] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 3));
}
template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to, const Packet2d& from, Index stride)
{
to[stride*0] = _mm_cvtsd_f64(from);
to[stride*1] = _mm_cvtsd_f64(_mm_shuffle_pd(from, from, 1));
}
template<> EIGEN_DEVICE_FUNC inline void pscatter<int, Packet4i>(int* to, const Packet4i& from, Index stride)
{
to[stride*0] = _mm_cvtsi128_si32(from);
to[stride*1] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 1));
to[stride*2] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 2));
to[stride*3] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 3));
}
// some compilers might be tempted to perform multiple moves instead of using a vector path.
template<> EIGEN_STRONG_INLINE void pstore1<Packet4f>(float* to, const float& a)
{
Packet4f pa = _mm_set_ss(a);
pstore(to, Packet4f(vec4f_swizzle1(pa,0,0,0,0)));
}
// some compilers might be tempted to perform multiple moves instead of using a vector path.
template<> EIGEN_STRONG_INLINE void pstore1<Packet2d>(double* to, const double& a)
{
Packet2d pa = _mm_set_sd(a);
pstore(to, Packet2d(vec2d_swizzle1(pa,0,0)));
}
#if EIGEN_COMP_PGI && EIGEN_COMP_PGI < 1900
typedef const void * SsePrefetchPtrType;
#else
typedef const char * SsePrefetchPtrType;
#endif
#ifndef EIGEN_VECTORIZE_AVX
template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); }
template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); }
template<> EIGEN_STRONG_INLINE void prefetch<int>(const int* addr) { _mm_prefetch((SsePrefetchPtrType)(addr), _MM_HINT_T0); }
#endif
#if EIGEN_COMP_MSVC_STRICT && EIGEN_OS_WIN64
// The temporary variable fixes an internal compilation error in vs <= 2008 and a wrong-result bug in vs 2010
// Direct of the struct members fixed bug #62.
template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { return a.m128_f32[0]; }
template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { return a.m128d_f64[0]; }
template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { int x = _mm_cvtsi128_si32(a); return x; }
#elif EIGEN_COMP_MSVC_STRICT
// The temporary variable fixes an internal compilation error in vs <= 2008 and a wrong-result bug in vs 2010
template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { float x = _mm_cvtss_f32(a); return x; }
template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { double x = _mm_cvtsd_f64(a); return x; }
template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { int x = _mm_cvtsi128_si32(a); return x; }
#else
template<> EIGEN_STRONG_INLINE float pfirst<Packet4f>(const Packet4f& a) { return _mm_cvtss_f32(a); }
template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { return _mm_cvtsd_f64(a); }
template<> EIGEN_STRONG_INLINE int pfirst<Packet4i>(const Packet4i& a) { return _mm_cvtsi128_si32(a); }
#endif
template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a)
{ return _mm_shuffle_ps(a,a,0x1B); }
template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a)
{ return _mm_shuffle_pd(a,a,0x1); }
template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a)
{ return _mm_shuffle_epi32(a,0x1B); }
template<> EIGEN_STRONG_INLINE Packet4f pabs(const Packet4f& a)
{
const Packet4f mask = _mm_castsi128_ps(_mm_setr_epi32(0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF));
return _mm_and_ps(a,mask);
}
template<> EIGEN_STRONG_INLINE Packet2d pabs(const Packet2d& a)
{
const Packet2d mask = _mm_castsi128_pd(_mm_setr_epi32(0xFFFFFFFF,0x7FFFFFFF,0xFFFFFFFF,0x7FFFFFFF));
return _mm_and_pd(a,mask);
}
template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a)
{
#ifdef EIGEN_VECTORIZE_SSSE3
return _mm_abs_epi32(a);
#else
Packet4i aux = _mm_srai_epi32(a,31);
return _mm_sub_epi32(_mm_xor_si128(a,aux),aux);
#endif
}
template<> EIGEN_STRONG_INLINE Packet4f pfrexp<Packet4f>(const Packet4f& a, Packet4f& exponent) {
return pfrexp_float(a,exponent);
}
template<> EIGEN_STRONG_INLINE Packet4f pldexp<Packet4f>(const Packet4f& a, const Packet4f& exponent) {
return pldexp_float(a,exponent);
}
template<> EIGEN_STRONG_INLINE Packet2d pldexp<Packet2d>(const Packet2d& a, const Packet2d& exponent) {
const Packet4i cst_1023_0 = _mm_setr_epi32(1023, 1023, 0, 0);
Packet4i emm0 = _mm_cvttpd_epi32(exponent);
emm0 = padd(emm0, cst_1023_0);
emm0 = _mm_slli_epi32(emm0, 20);
emm0 = _mm_shuffle_epi32(emm0, _MM_SHUFFLE(1,2,0,3));
return pmul(a, Packet2d(_mm_castsi128_pd(emm0)));
}
// with AVX, the default implementations based on pload1 are faster
#ifndef __AVX__
template<> EIGEN_STRONG_INLINE void
pbroadcast4<Packet4f>(const float *a,
Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3)
{
a3 = pload<Packet4f>(a);
a0 = vec4f_swizzle1(a3, 0,0,0,0);
a1 = vec4f_swizzle1(a3, 1,1,1,1);
a2 = vec4f_swizzle1(a3, 2,2,2,2);
a3 = vec4f_swizzle1(a3, 3,3,3,3);
}
template<> EIGEN_STRONG_INLINE void
pbroadcast4<Packet2d>(const double *a,
Packet2d& a0, Packet2d& a1, Packet2d& a2, Packet2d& a3)
{
#ifdef EIGEN_VECTORIZE_SSE3
a0 = _mm_loaddup_pd(a+0);
a1 = _mm_loaddup_pd(a+1);
a2 = _mm_loaddup_pd(a+2);
a3 = _mm_loaddup_pd(a+3);
#else
a1 = pload<Packet2d>(a);
a0 = vec2d_swizzle1(a1, 0,0);
a1 = vec2d_swizzle1(a1, 1,1);
a3 = pload<Packet2d>(a+2);
a2 = vec2d_swizzle1(a3, 0,0);
a3 = vec2d_swizzle1(a3, 1,1);
#endif
}
#endif
EIGEN_STRONG_INLINE void punpackp(Packet4f* vecs)
{
vecs[1] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0x55));
vecs[2] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0xAA));
vecs[3] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0xFF));
vecs[0] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0x00));
}
#ifdef EIGEN_VECTORIZE_SSE3
template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
{
return _mm_hadd_ps(_mm_hadd_ps(vecs[0], vecs[1]),_mm_hadd_ps(vecs[2], vecs[3]));
}
template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
{
return _mm_hadd_pd(vecs[0], vecs[1]);
}
#else
template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
{
Packet4f tmp0, tmp1, tmp2;
tmp0 = _mm_unpacklo_ps(vecs[0], vecs[1]);
tmp1 = _mm_unpackhi_ps(vecs[0], vecs[1]);
tmp2 = _mm_unpackhi_ps(vecs[2], vecs[3]);
tmp0 = _mm_add_ps(tmp0, tmp1);
tmp1 = _mm_unpacklo_ps(vecs[2], vecs[3]);
tmp1 = _mm_add_ps(tmp1, tmp2);
tmp2 = _mm_movehl_ps(tmp1, tmp0);
tmp0 = _mm_movelh_ps(tmp0, tmp1);
return _mm_add_ps(tmp0, tmp2);
}
template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
{
return _mm_add_pd(_mm_unpacklo_pd(vecs[0], vecs[1]), _mm_unpackhi_pd(vecs[0], vecs[1]));
}
#endif // SSE3
template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
{
// Disable SSE3 _mm_hadd_pd that is extremely slow on all existing Intel's architectures
// (from Nehalem to Haswell)
// #ifdef EIGEN_VECTORIZE_SSE3
// Packet4f tmp = _mm_add_ps(a, vec4f_swizzle1(a,2,3,2,3));
// return pfirst<Packet4f>(_mm_hadd_ps(tmp, tmp));
// #else
Packet4f tmp = _mm_add_ps(a, _mm_movehl_ps(a,a));
return pfirst<Packet4f>(_mm_add_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
// #endif
}
template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a)
{
// Disable SSE3 _mm_hadd_pd that is extremely slow on all existing Intel's architectures
// (from Nehalem to Haswell)
// #ifdef EIGEN_VECTORIZE_SSE3
// return pfirst<Packet2d>(_mm_hadd_pd(a, a));
// #else
return pfirst<Packet2d>(_mm_add_sd(a, _mm_unpackhi_pd(a,a)));
// #endif
}
#ifdef EIGEN_VECTORIZE_SSSE3
template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
{
return _mm_hadd_epi32(_mm_hadd_epi32(vecs[0], vecs[1]),_mm_hadd_epi32(vecs[2], vecs[3]));
}
template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
{
Packet4i tmp0 = _mm_hadd_epi32(a,a);
return pfirst<Packet4i>(_mm_hadd_epi32(tmp0,tmp0));
}
#else
template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
{
Packet4i tmp = _mm_add_epi32(a, _mm_unpackhi_epi64(a,a));
return pfirst(tmp) + pfirst<Packet4i>(_mm_shuffle_epi32(tmp, 1));
}
template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
{
Packet4i tmp0, tmp1, tmp2;
tmp0 = _mm_unpacklo_epi32(vecs[0], vecs[1]);
tmp1 = _mm_unpackhi_epi32(vecs[0], vecs[1]);
tmp2 = _mm_unpackhi_epi32(vecs[2], vecs[3]);
tmp0 = _mm_add_epi32(tmp0, tmp1);
tmp1 = _mm_unpacklo_epi32(vecs[2], vecs[3]);
tmp1 = _mm_add_epi32(tmp1, tmp2);
tmp2 = _mm_unpacklo_epi64(tmp0, tmp1);
tmp0 = _mm_unpackhi_epi64(tmp0, tmp1);
return _mm_add_epi32(tmp0, tmp2);
}
#endif
// Other reduction functions:
// mul
template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a)
{
Packet4f tmp = _mm_mul_ps(a, _mm_movehl_ps(a,a));
return pfirst<Packet4f>(_mm_mul_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
}
template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a)
{
return pfirst<Packet2d>(_mm_mul_sd(a, _mm_unpackhi_pd(a,a)));
}
template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a)
{
// after some experiments, it is seems this is the fastest way to implement it
// for GCC (eg., reusing pmul is very slow !)
// TODO try to call _mm_mul_epu32 directly
EIGEN_ALIGN16 int aux[4];
pstore(aux, a);
return (aux[0] * aux[1]) * (aux[2] * aux[3]);
}
// min
template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a)
{
Packet4f tmp = _mm_min_ps(a, _mm_movehl_ps(a,a));
return pfirst<Packet4f>(_mm_min_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
}
template<> EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a)
{
return pfirst<Packet2d>(_mm_min_sd(a, _mm_unpackhi_pd(a,a)));
}
template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a)
{
#ifdef EIGEN_VECTORIZE_SSE4_1
Packet4i tmp = _mm_min_epi32(a, _mm_shuffle_epi32(a, _MM_SHUFFLE(0,0,3,2)));
return pfirst<Packet4i>(_mm_min_epi32(tmp,_mm_shuffle_epi32(tmp, 1)));
#else
// after some experiments, it is seems this is the fastest way to implement it
// for GCC (eg., it does not like using std::min after the pstore !!)
EIGEN_ALIGN16 int aux[4];
pstore(aux, a);
int aux0 = aux[0]<aux[1] ? aux[0] : aux[1];
int aux2 = aux[2]<aux[3] ? aux[2] : aux[3];
return aux0<aux2 ? aux0 : aux2;
#endif // EIGEN_VECTORIZE_SSE4_1
}
// max
template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a)
{
Packet4f tmp = _mm_max_ps(a, _mm_movehl_ps(a,a));
return pfirst<Packet4f>(_mm_max_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
}
template<> EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a)
{
return pfirst<Packet2d>(_mm_max_sd(a, _mm_unpackhi_pd(a,a)));
}
template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a)
{
#ifdef EIGEN_VECTORIZE_SSE4_1
Packet4i tmp = _mm_max_epi32(a, _mm_shuffle_epi32(a, _MM_SHUFFLE(0,0,3,2)));
return pfirst<Packet4i>(_mm_max_epi32(tmp,_mm_shuffle_epi32(tmp, 1)));
#else
// after some experiments, it is seems this is the fastest way to implement it
// for GCC (eg., it does not like using std::min after the pstore !!)
EIGEN_ALIGN16 int aux[4];
pstore(aux, a);
int aux0 = aux[0]>aux[1] ? aux[0] : aux[1];
int aux2 = aux[2]>aux[3] ? aux[2] : aux[3];
return aux0>aux2 ? aux0 : aux2;
#endif // EIGEN_VECTORIZE_SSE4_1
}
// not needed yet
// template<> EIGEN_STRONG_INLINE bool predux_all(const Packet4f& x)
// {
// return _mm_movemask_ps(x) == 0xF;
// }
template<> EIGEN_STRONG_INLINE bool predux_any(const Packet4f& x)
{
return _mm_movemask_ps(x) != 0x0;
}
#if EIGEN_COMP_GNUC
// template <> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c)
// {
// Packet4f res = b;
// asm("mulps %[a], %[b] \n\taddps %[c], %[b]" : [b] "+x" (res) : [a] "x" (a), [c] "x" (c));
// return res;
// }
// EIGEN_STRONG_INLINE Packet4i _mm_alignr_epi8(const Packet4i& a, const Packet4i& b, const int i)
// {
// Packet4i res = a;
// asm("palignr %[i], %[a], %[b] " : [b] "+x" (res) : [a] "x" (a), [i] "i" (i));
// return res;
// }
#endif
#ifdef EIGEN_VECTORIZE_SSSE3
// SSSE3 versions
template<int Offset>
struct palign_impl<Offset,Packet4f>
{
static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second)
{
if (Offset!=0)
first = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(second), _mm_castps_si128(first), Offset*4));
}
};
template<int Offset>
struct palign_impl<Offset,Packet4i>
{
static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second)
{
if (Offset!=0)
first = _mm_alignr_epi8(second,first, Offset*4);
}
};
template<int Offset>
struct palign_impl<Offset,Packet2d>
{
static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second)
{
if (Offset==1)
first = _mm_castsi128_pd(_mm_alignr_epi8(_mm_castpd_si128(second), _mm_castpd_si128(first), 8));
}
};
#else
// SSE2 versions
template<int Offset>
struct palign_impl<Offset,Packet4f>
{
static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second)
{
if (Offset==1)
{
first = _mm_move_ss(first,second);
first = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(first),0x39));
}
else if (Offset==2)
{
first = _mm_movehl_ps(first,first);
first = _mm_movelh_ps(first,second);
}
else if (Offset==3)
{
first = _mm_move_ss(first,second);
first = _mm_shuffle_ps(first,second,0x93);
}
}
};
template<int Offset>
struct palign_impl<Offset,Packet4i>
{
static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second)
{
if (Offset==1)
{
first = _mm_castps_si128(_mm_move_ss(_mm_castsi128_ps(first),_mm_castsi128_ps(second)));
first = _mm_shuffle_epi32(first,0x39);
}
else if (Offset==2)
{
first = _mm_castps_si128(_mm_movehl_ps(_mm_castsi128_ps(first),_mm_castsi128_ps(first)));
first = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(first),_mm_castsi128_ps(second)));
}
else if (Offset==3)
{
first = _mm_castps_si128(_mm_move_ss(_mm_castsi128_ps(first),_mm_castsi128_ps(second)));
first = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(first),_mm_castsi128_ps(second),0x93));
}
}
};
template<int Offset>
struct palign_impl<Offset,Packet2d>
{
static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second)
{
if (Offset==1)
{
first = _mm_castps_pd(_mm_movehl_ps(_mm_castpd_ps(first),_mm_castpd_ps(first)));
first = _mm_castps_pd(_mm_movelh_ps(_mm_castpd_ps(first),_mm_castpd_ps(second)));
}
}
};
#endif
EIGEN_DEVICE_FUNC inline void
ptranspose(PacketBlock<Packet4f,4>& kernel) {
_MM_TRANSPOSE4_PS(kernel.packet[0], kernel.packet[1], kernel.packet[2], kernel.packet[3]);
}
EIGEN_DEVICE_FUNC inline void
ptranspose(PacketBlock<Packet2d,2>& kernel) {
__m128d tmp = _mm_unpackhi_pd(kernel.packet[0], kernel.packet[1]);
kernel.packet[0] = _mm_unpacklo_pd(kernel.packet[0], kernel.packet[1]);
kernel.packet[1] = tmp;
}
EIGEN_DEVICE_FUNC inline void
ptranspose(PacketBlock<Packet4i,4>& kernel) {
__m128i T0 = _mm_unpacklo_epi32(kernel.packet[0], kernel.packet[1]);
__m128i T1 = _mm_unpacklo_epi32(kernel.packet[2], kernel.packet[3]);
__m128i T2 = _mm_unpackhi_epi32(kernel.packet[0], kernel.packet[1]);
__m128i T3 = _mm_unpackhi_epi32(kernel.packet[2], kernel.packet[3]);
kernel.packet[0] = _mm_unpacklo_epi64(T0, T1);
kernel.packet[1] = _mm_unpackhi_epi64(T0, T1);
kernel.packet[2] = _mm_unpacklo_epi64(T2, T3);
kernel.packet[3] = _mm_unpackhi_epi64(T2, T3);
}
template<> EIGEN_STRONG_INLINE Packet4i pblend(const Selector<4>& ifPacket, const Packet4i& thenPacket, const Packet4i& elsePacket) {
const __m128i zero = _mm_setzero_si128();
const __m128i select = _mm_set_epi32(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]);
__m128i false_mask = _mm_cmpeq_epi32(select, zero);
#ifdef EIGEN_VECTORIZE_SSE4_1
return _mm_blendv_epi8(thenPacket, elsePacket, false_mask);
#else
return _mm_or_si128(_mm_andnot_si128(false_mask, thenPacket), _mm_and_si128(false_mask, elsePacket));
#endif
}
template<> EIGEN_STRONG_INLINE Packet4f pblend(const Selector<4>& ifPacket, const Packet4f& thenPacket, const Packet4f& elsePacket) {
const __m128 zero = _mm_setzero_ps();
const __m128 select = _mm_set_ps(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]);
__m128 false_mask = _mm_cmpeq_ps(select, zero);
#ifdef EIGEN_VECTORIZE_SSE4_1
return _mm_blendv_ps(thenPacket, elsePacket, false_mask);
#else
return _mm_or_ps(_mm_andnot_ps(false_mask, thenPacket), _mm_and_ps(false_mask, elsePacket));
#endif