-
Notifications
You must be signed in to change notification settings - Fork 21
/
simd_utils.h
3636 lines (3238 loc) · 96.6 KB
/
simd_utils.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
/*
* Project : SIMD_Utils
* Version : 0.2.6
* Author : JishinMaster
* Licence : BSD-2
*/
#pragma once
#ifdef __cplusplus
extern "C" {
#endif
#define MAJOR_VERSION 0
#define MINOR_VERSION 2
#define SUB_VERSION 5
#ifdef OMP
#include <omp.h>
#endif
#include <math.h>
#include <fenv.h>
#include <stdint.h>
#include <stdio.h>
#include "simd_utils_constants.h"
#include "mysincosf.h"
/* if the user insures that all of their pointers are aligned,
* they can use ALWAYS_ALIGNED to hope for some minor speedup on small vectors
*/
static inline int isAligned(uintptr_t ptr, size_t alignment)
{
#ifndef ALWAYS_ALIGNED
#ifndef ARM // ARM manages disalignment in hardware
if (((uintptr_t) (ptr) % alignment) == 0)
return 1;
return 0;
#else
return 1;
#endif
#else
return 1;
#endif
}
static inline int areAligned2(uintptr_t ptr1, uintptr_t ptr2, size_t alignment)
{
#ifndef ALWAYS_ALIGNED
#ifndef ARM // ARM manages disalignment in hardware
if (((uintptr_t) (ptr1) % alignment) == 0)
if (((uintptr_t) (ptr2) % alignment) == 0)
return 1;
return 0;
#else
return 1;
#endif
#else
return 1;
#endif
}
static inline int areAligned3(uintptr_t ptr1, uintptr_t ptr2, uintptr_t ptr3, size_t alignment)
{
#ifndef ALWAYS_ALIGNED
#ifndef ARM // ARM manages disalignment in hardware
if (((uintptr_t) (ptr1) % alignment) == 0)
if (((uintptr_t) (ptr2) % alignment) == 0)
if (((uintptr_t) (ptr3) % alignment) == 0)
return 1;
return 0;
#else
return 1;
#endif
#else
return 1;
#endif
}
static inline void simd_utils_get_version(void)
{
printf("Simd Utils Version : %d.%d.%d\n", MAJOR_VERSION, MINOR_VERSION, SUB_VERSION);
}
// Produce value of bit n. n must be less than 32.
#define Bit(n) ((uint32_t) 1 << (n))
// Create a mask of n bits in the low bits. n must be less than 32.
#define Mask(n) (Bit(n) - 1)
/* Convert an IEEE-754 16-bit binary floating-point encoding to an IEEE-754
32-bit binary floating-point encoding.
This code has not been tested.
*/
//From https://stackoverflow.com/questions/71120169/bit-shifting-a-half-float-into-a-float/71125539#71125539
static inline uint32_t Float16ToFloat32(uint16_t x)
{
/* Separate the sign encoding (1 bit starting at bit 15), the exponent
encoding (5 bits starting at bit 10), and the primary significand
(fraction) encoding (10 bits starting at bit 0).
*/
uint32_t s = x >> 15;
uint32_t e = x >> 10 & Mask( 5);
uint32_t f = x & Mask(10);
// Left-adjust the significand field.
f <<= 23 - 10;
// Switch to handle subnormal numbers, normal numbers, and infinities/NaNs.
switch (e)
{
// Exponent code is subnormal.
case 0:
// Zero does need any changes, but subnormals need normalization.
if (f != 0)
{
/* Set the 32-bit exponent code corresponding to the 16-bit
subnormal exponent.
*/
e = 1 + (127 - 15);
/* Normalize the significand by shifting until its leading
bit moves out of the field. (This code could benefit from
a find-first-set instruction or possibly using a conversion
from integer to floating-point to do the normalization.)
*/
while (f < Bit(23))
{
f <<= 1;
e -= 1;
}
// Remove the leading bit.
f &= Mask(23);
}
break;
// Exponent code is normal.
default:
e += 127 - 15; // Adjust from 16-bit bias to 32-bit bias.
break;
// Exponent code indicates infinity or NaN.
case 31:
e = 255; // Set 32-bit exponent code for infinity or NaN.
break;
}
// Assemble and return the 32-bit encoding.
return s << 31 | e << 23 | f;
}
static inline uint32_t float_as_uint32 (float a)
{
return *(uint32_t *)&a;
}
static inline float uint32_as_float (uint32_t a)
{
return *(float *)&a;
}
//From https://stackoverflow.com/questions/76799117/how-to-convert-a-float-to-a-half-type-and-the-other-way-around-in-c
static inline uint16_t float2half_rn (float a)
{
uint32_t ia = float_as_uint32 (a);
uint16_t ir;
ir = (ia >> 16) & 0x8000;
if ((ia & 0x7f800000) == 0x7f800000) {
if ((ia & 0x7fffffff) == 0x7f800000) {
ir |= 0x7c00; /* infinity */
} else {
ir |= 0x7e00 | ((ia >> (24 - 11)) & 0x1ff); /* NaN, quietened */
}
} else if ((ia & 0x7f800000) >= 0x33000000) {
int shift = (int)((ia >> 23) & 0xff) - 127;
if (shift > 15) {
ir |= 0x7c00; /* infinity */
} else {
ia = (ia & 0x007fffff) | 0x00800000; /* extract mantissa */
if (shift < -14) { /* denormal */
ir |= ia >> (-1 - shift);
ia = ia << (32 - (-1 - shift));
} else { /* normal */
ir |= ia >> (24 - 11);
ia = ia << (32 - (24 - 11));
ir = ir + ((14 + shift) << 10);
}
/* IEEE-754 round to nearest of even */
if ((ia > 0x80000000) || ((ia == 0x80000000) && (ir & 1))) {
ir++;
}
}
}
return ir;
}
static inline void fp32tofp16_C (float* src, uint16_t* dst, size_t len)
{
for (int i = 0; i < len; i ++) {
dst[i] = float2half_rn(src[i]);
}
}
static inline void fp16tofp32_C (uint16_t* src, float* dst, size_t len)
{
for (int i = 0; i < len; i ++) {
dst[i] = uint32_as_float(Float16ToFloat32(src[i]));
}
}
#ifdef SSE
#ifdef NO_SSE3
#define NO_SSE4
static inline __m128 _mm_movehdup_ps(__m128 __X)
{
return _mm_shuffle_ps(__X, __X, 0xF5);
}
static inline __m128 _mm_moveldup_ps(__m128 __X)
{
return _mm_shuffle_ps(__X, __X, 0xA0);
}
#endif
#ifdef NO_SSE4
static inline __m128i _mm_cmpeq_epi64(__m128i __X, __m128i __Y)
{
int64_t *ptr_x = (int64_t *) &__X;
int64_t *ptr_y = (int64_t *) &__Y;
__m128i ret;
int64_t *ptr_ret = (int64_t *) &ret;
ptr_ret[0] = (ptr_x[0] == ptr_y[0]) ? 0xFFFFFFFFFFFFFFFF : 0;
ptr_ret[1] = (ptr_x[1] == ptr_y[1]) ? 0xFFFFFFFFFFFFFFFF : 0;
return ret;
}
static inline __m128d _mm_blendv_pd(__m128d __X, __m128d __Y, __m128d __M)
{
__m128d b_tmp = _mm_and_pd(__Y, __M);
__m128d a_tmp = _mm_and_pd(__X, _mm_cmpeq_pd(__M, *(__m128d *) _pd_zero));
return _mm_or_pd(a_tmp, b_tmp);
}
static inline __m128 _mm_blendv_ps(__m128 __X, __m128 __Y, __m128 __M)
{
__m128 b_tmp = _mm_and_ps(__Y, __M);
__m128 a_tmp = _mm_and_ps(__X, _mm_cmpeq_ps(__M, *(__m128 *) _ps_zero));
return _mm_or_ps(a_tmp, b_tmp);
}
static inline __m128i _mm_stream_load_si128(__m128i *__X)
{
return _mm_load_si128(__X);
}
static inline __m128 _mm_round_ps(__m128 X, int mode)
{
__m128 ret;
__m128i reti;
unsigned int old_mode = _MM_GET_ROUNDING_MODE();
switch (mode) {
case _MM_FROUND_TRUNC:
case _MM_ROUND_TOWARD_ZERO:
case ROUNDTOZERO:
_MM_SET_ROUNDING_MODE(_MM_ROUND_TOWARD_ZERO);
break;
case ROUNDTOCEIL:
case _MM_ROUND_UP:
_MM_SET_ROUNDING_MODE(_MM_ROUND_UP);
break;
case ROUNDTOFLOOR:
case _MM_ROUND_DOWN:
_MM_SET_ROUNDING_MODE(_MM_ROUND_DOWN);
break;
default:
//_MM_SET_ROUNDING_MODE(_MM_ROUND_NEAREST);
break;
}
reti = _mm_cvtps_epi32(X);
ret = _mm_cvtepi32_ps(reti);
_MM_SET_ROUNDING_MODE(old_mode);
return ret;
}
/* not accurate but might do the trick for most cases
where the full range is not needed */
static inline __m128i _mm_packus_epi32(__m128i a, __m128i b)
{
return _mm_packs_epi32(a, b);
}
// https://gist.github.com/cxd4/8137986
#define SWAP(d3, d2, d1, d0) ((d3 << 6) | (d2 << 4) | (d1 << 2) | (d0 << 0))
static __m128i _mm_mullo_epi32(__m128i a, __m128i b)
{
__m128i prod_m; /* alternating FFFFFFFF00000000FFFFFFFF00000000 */
__m128i prod_n; /* alternating 00000000FFFFFFFF00000000FFFFFFFF */
prod_n = _mm_mul_epu32(a, b);
a = _mm_shuffle_epi32(a, SWAP(2, 3, 0, 1)); /* old SWAP(3,2,1,0) */
b = _mm_shuffle_epi32(b, SWAP(2, 3, 0, 1)); /* old SWAP(3,2,1,0) */
prod_m = _mm_mul_epu32(a, b);
/*
* prod_m = { a[0] * b[0], a[2] * b[2] }
* prod_n = { a[1] * b[1], a[3] * b[3] }
*/
a = _mm_unpacklo_epi32(prod_n, prod_m);
a = _mm_slli_si128(a, 64 / 8);
a = _mm_srli_si128(a, 64 / 8);
b = _mm_unpackhi_epi32(prod_n, prod_m);
b = _mm_slli_si128(b, 64 / 8);
b = _mm_or_si128(b, a); /* Ans = (hi << 64) | (lo & 0x00000000FFFFFFFF) */
return (b);
}
// Or : https://stackoverflow.com/questions/17264399/fastest-way-to-multiply-two-vectors-of-32bit-integers-in-c-with-sse
// Vec4i operator * (Vec4i const & a, Vec4i const & b) {
/*
__m128i a13 = _mm_shuffle_epi32(a, 0xF5); // (-,a3,-,a1)
__m128i b13 = _mm_shuffle_epi32(b, 0xF5); // (-,b3,-,b1)
__m128i prod02 = _mm_mul_epu32(a, b); // (-,a2*b2,-,a0*b0)
__m128i prod13 = _mm_mul_epu32(a13, b13); // (-,a3*b3,-,a1*b1)
__m128i prod01 = _mm_unpacklo_epi32(prod02,prod13); // (-,-,a1*b1,a0*b0)
__m128i prod23 = _mm_unpackhi_epi32(prod02,prod13); // (-,-,a3*b3,a2*b2)
__m128i prod = _mm_unpacklo_epi64(prod01,prod23); // (ab3,ab2,ab1,ab0)
*/
#endif
#ifndef ARM
#include "sse_mathfun.h"
#else /* ARM */
#include "neon_mathfun.h"
#endif /* ARM */
//_mm_movehl_ps vs shuffle(1,0,3,2)?
static inline v2sd _mm_cvtps_pd_high(v4sf in){
#if defined(__aarch64__)
return vcvt_high_f64_f32(in);
#else
return _mm_cvtps_pd(_mm_movehl_ps(in, in));
#endif
}
static inline v4sf _mm_cvtpd2_ps(v2sd low, v2sd high){
#if defined(__aarch64__)
return vcombine_f32(vcvt_f32_f64(low), vcvt_f32_f64(high));
#else
return _mm_movelh_ps(_mm_cvtpd_ps(low), _mm_cvtpd_ps(high));
#endif
}
static inline v4sfx2 _mm_load2_ps(float const *mem_addr)
{
#ifdef ARM
return vld2q_f32(mem_addr);
#else
v4sf tmp1 = _mm_load_ps(mem_addr);
v4sf tmp2 = _mm_load_ps(mem_addr + SSE_LEN_FLOAT);
v4sfx2 ret;
ret.val[0] = _mm_shuffle_ps(tmp1, tmp2, _MM_SHUFFLE(2, 0, 2, 0));
ret.val[1] = _mm_shuffle_ps(tmp1, tmp2, _MM_SHUFFLE(3, 1, 3, 1));
return ret;
#endif
}
static inline v4sfx2 _mm_load2u_ps(float const *mem_addr)
{
#ifdef ARM
return vld2q_f32(mem_addr);
#else
v4sf tmp1 = _mm_loadu_ps(mem_addr);
v4sf tmp2 = _mm_loadu_ps(mem_addr + SSE_LEN_FLOAT);
v4sfx2 ret;
ret.val[0] = _mm_shuffle_ps(tmp1, tmp2, _MM_SHUFFLE(2, 0, 2, 0));
ret.val[1] = _mm_shuffle_ps(tmp1, tmp2, _MM_SHUFFLE(3, 1, 3, 1));
return ret;
#endif
}
static inline void _mm_store2_ps(float *mem_addr, v4sfx2 a)
{
#ifdef ARM
vst2q_f32(mem_addr, a);
#else
v4sf tmp1 = _mm_unpacklo_ps(a.val[0], a.val[1]);
v4sf tmp2 = _mm_unpackhi_ps(a.val[0], a.val[1]);
_mm_store_ps(mem_addr, tmp1);
_mm_store_ps(mem_addr + SSE_LEN_FLOAT, tmp2);
#endif
}
static inline void _mm_store2u_ps(float *mem_addr, v4sfx2 a)
{
#ifdef ARM
vst2q_f32(mem_addr, a);
#else
v4sf tmp1 = _mm_unpacklo_ps(a.val[0], a.val[1]);
v4sf tmp2 = _mm_unpackhi_ps(a.val[0], a.val[1]);
_mm_storeu_ps(mem_addr, tmp1);
_mm_storeu_ps(mem_addr + SSE_LEN_FLOAT, tmp2);
#endif
}
static inline v2sdx2 _mm_load2_pd(double const *mem_addr)
{
#if defined(__aarch64__)
return vld2q_f64(mem_addr);
#else
v2sd tmp1 = _mm_load_pd(mem_addr);
v2sd tmp2 = _mm_load_pd(mem_addr + SSE_LEN_DOUBLE);
v2sdx2 ret;
ret.val[0] = _mm_shuffle_pd(tmp1, tmp2, _MM_SHUFFLE2(0, 0));
ret.val[1] = _mm_shuffle_pd(tmp1, tmp2, _MM_SHUFFLE2(1, 1));
return ret;
#endif
}
static inline v2sdx2 _mm_load2u_pd(double const *mem_addr)
{
#if defined(__aarch64__)
return vld2q_f64(mem_addr);
#else
v2sd tmp1 = _mm_loadu_pd(mem_addr);
v2sd tmp2 = _mm_loadu_pd(mem_addr + SSE_LEN_DOUBLE);
v2sdx2 ret;
ret.val[0] = _mm_shuffle_pd(tmp1, tmp2, _MM_SHUFFLE2(0, 0));
ret.val[1] = _mm_shuffle_pd(tmp1, tmp2, _MM_SHUFFLE2(1, 1));
return ret;
#endif
}
static inline void _mm_store2_pd(double *mem_addr, v2sdx2 a)
{
#if defined(__aarch64__)
vst2q_f64(mem_addr, a);
#else
v2sd tmp1 = _mm_unpacklo_pd(a.val[0], a.val[1]);
v2sd tmp2 = _mm_unpackhi_pd(a.val[0], a.val[1]);
_mm_store_pd(mem_addr, tmp1);
_mm_store_pd(mem_addr + SSE_LEN_DOUBLE, tmp2);
#endif
}
static inline void _mm_store2u_pd(double *mem_addr, v2sdx2 a)
{
#if defined(__aarch64__)
vst2q_f64(mem_addr, a);
#else
v2sd tmp1 = _mm_unpacklo_pd(a.val[0], a.val[1]);
v2sd tmp2 = _mm_unpackhi_pd(a.val[0], a.val[1]);
_mm_storeu_pd(mem_addr, tmp1);
_mm_storeu_pd(mem_addr + SSE_LEN_DOUBLE, tmp2);
#endif
}
static inline __m128 _mm_fmadd_ps_custom(__m128 a, __m128 b, __m128 c)
{
// Haswell comes with avx2 and fma
// ARM has vmla instead of fma in 32bits
#if defined(ARM) || defined(FMA)
return _mm_fmadd_ps(a, b, c);
#else
return _mm_add_ps(_mm_mul_ps(a, b), c);
#endif
}
//B is a scalar, some optimizations for ARM NEON
static inline __m128 _mm_fmadd1_ps_custom(__m128 a, __m128 b, __m128 c)
{
#if defined(ARM)
return _mm_fmadd1_ps(a, b, c);
#else
return _mm_fmadd_ps_custom(a,b,c);
#endif
}
#ifndef ARM
//B is a scalar, some optimizations for ARM NEON
static inline __m128 _mm_mul1_ps(__m128 a, __m128 b)
{
return _mm_mul_ps(a, b);
}
#endif
static inline __m128 _mm_fmaddsub_ps_custom(__m128 a, __m128 b, __m128 c)
{
#ifndef FMA // Haswell comes with avx2 and fma
return _mm_addsub_ps(_mm_mul_ps(a, b), c);
#else /* FMA */
return _mm_fmaddsub_ps(a, b, c);
#endif /* FMA */
}
static inline __m128 _mm_fmsubadd_ps_custom(__m128 a, __m128 b, __m128 c)
{
#if !defined(FMA) || defined(ARM)
v4sf d = _mm_mul_ps(*(v4sf *) _ps_conj_mask, c);
return _mm_addsub_ps(_mm_mul_ps(a, b), d);
#else /* FMA */
return _mm_fmsubadd_ps(a, b, c);
#endif /* FMA */
}
static inline __m128 _mm_fnmadd_ps_custom(__m128 a, __m128 b, __m128 c)
{
// Haswell comes with avx2 and fma
// ARM has vmla instead of fma in 32bits
#if defined(ARM) || defined(FMA)
return _mm_fnmadd_ps(a, b, c);
#else
return _mm_sub_ps(c, _mm_mul_ps(a, b));
#endif
}
//B is a scalar, some optimizations for ARM NEON
static inline __m128 _mm_fnmadd1_ps_custom(__m128 a, __m128 b, __m128 c)
{
#if defined(ARM)
return _mm_fnmadd1_ps(a, b, c);
#else
return _mm_fnmadd_ps_custom(a,b,c);
#endif
}
static inline __m128d _mm_fmadd_pd_custom(__m128d a, __m128d b, __m128d c)
{
#ifndef FMA // Haswell comes with avx2 and fma
return _mm_add_pd(_mm_mul_pd(a, b), c);
#else /* FMA */
return _mm_fmadd_pd(a, b, c);
#endif /* FMA */
}
static inline __m128d _mm_fnmadd_pd_custom(__m128d a, __m128d b, __m128d c)
{
#ifndef FMA // Haswell comes with avx2 and fma
return _mm_sub_pd(c, _mm_mul_pd(a, b));
#else /* FMA */
return _mm_fnmadd_pd(a, b, c);
#endif /* FMA */
}
//B is a scalar, some optimizations for ARM NEON
static inline __m128d _mm_fmadd1_pd_custom(__m128d a, __m128d b, __m128d c)
{
#if defined(ARM)
return _mm_fmadd1_pd(a, b, c);
#else
return _mm_fmadd_pd_custom(a,b,c);
#endif
}
//B is a scalar, some optimizations for ARM NEON
static inline __m128d _mm_fnmadd1_pd_custom(__m128d a, __m128d b, __m128d c)
{
#if defined(ARM)
return _mm_fnmadd1_pd(a, b, c);
#else
return _mm_fnmadd_pd_custom(a,b,c);
#endif
}
#ifndef ARM
//B is a scalar, some optimizations for ARM NEON
static inline __m128d _mm_mul1_pd(__m128d a, __m128d b)
{
return _mm_mul_pd(a, b);
}
#endif
// Work in progress
// in SSE, missing _mm_cvtepi64_pd, _mm_cvttpd_epi64
// See : https://stackoverflow.com/questions/41144668/how-to-efficiently-perform-double-int64-conversions-with-sse-avx
static inline v2sd _mm_cvtepi64_pd_custom(v2sid x)
{
#if 0
//Signed
x = _mm_add_epi64(x, _mm_castpd_si128(*(v2sd *) _pd_epi64_mask));
return _mm_sub_pd(_mm_castsi128_pd(x), *(v2sd *) _pd_epi64_mask);
#else
// unsigned
x = _mm_or_si128(x, _mm_castpd_si128(*(v2sd *) _pd_PDEPI64U));
return _mm_sub_pd(_mm_castsi128_pd(x), *(v2sd *) _pd_PDEPI64U);
#endif
}
static inline v2sd _mm_cvtepi64_pd_signed_custom(v2sid x)
{
// Signed
x = _mm_add_epi64(x, _mm_castpd_si128(*(v2sd *) _pd_epi64_mask));
return _mm_sub_pd(_mm_castsi128_pd(x), *(v2sd *) _pd_epi64_mask);
}
static inline v2sid _mm_cvtpd_epi64_custom(v2sd x)
{
#if defined(__aarch64__)
return vcvtq_s64_f64(x);
#else //ARM
// Signed
#if 1
x = _mm_add_pd(x, *(v2sd *) _pd_epi64_mask);
return _mm_sub_epi64(
_mm_castpd_si128(x),
_mm_castpd_si128(*(v2sd *) _pd_epi64_mask));
#else
// Unsigned
x = _mm_add_pd(x, *(v2sd *) _pd_PDEPI64U); //_mm_set1_pd(0x0010000000000000));
return _mm_xor_si128(
_mm_castpd_si128(x),
_mm_castpd_si128(*(v2sd *) _pd_PDEPI64U));
#endif
#endif // ARM
}
#include "simd_utils_sse_double.h"
#include "simd_utils_sse_float.h"
#include "simd_utils_sse_int32.h"
#endif /* SSE */
#ifdef AVX
#ifndef __clang__
#ifndef __INTEL_COMPILER
#ifndef __cplusplus // TODO : it seems to be defined with G++ 9.2 and not GCC 9.2
static inline __m256 _mm256_set_m128(__m128 H, __m128 L) // not present on every GCC version
{
return _mm256_insertf128_ps(_mm256_castps128_ps256(L), H, 1);
}
#endif
#endif
#endif /* __clang__ */
static inline __m256 _mm256_fmadd_ps_custom(__m256 a, __m256 b, __m256 c)
{
#ifndef FMA // Haswell comes with avx2 and fma
return _mm256_add_ps(_mm256_mul_ps(a, b), c);
#else /* FMA */
return _mm256_fmadd_ps(a, b, c);
#endif /* FMA */
}
static inline __m256 _mm256_fmaddsub_ps_custom(__m256 a, __m256 b, __m256 c)
{
#ifndef FMA // Haswell comes with avx2 and fma
return _mm256_addsub_ps(_mm256_mul_ps(a, b), c);
#else /* FMA */
return _mm256_fmaddsub_ps(a, b, c);
#endif /* FMA */
}
static inline __m256 _mm256_fnmadd_ps_custom(__m256 a, __m256 b, __m256 c)
{
#ifndef FMA // Haswell comes with avx2 and fma
return _mm256_sub_ps(c, _mm256_mul_ps(a, b));
#else /* FMA */
return _mm256_fnmadd_ps(a, b, c);
#endif /* FMA */
}
static inline __m256d _mm256_fmadd_pd_custom(__m256d a, __m256d b, __m256d c)
{
#ifndef FMA // Haswell comes with avx2 and fma
return _mm256_add_pd(_mm256_mul_pd(a, b), c);
#else /* FMA */
return _mm256_fmadd_pd(a, b, c);
#endif /* FMA */
}
static inline __m256d _mm256_fnmadd_pd_custom(__m256d a, __m256d b, __m256d c)
{
#ifndef FMA // Haswell comes with avx2 and fma
return _mm256_sub_pd(c, _mm256_mul_pd(a, b));
#else /* FMA */
return _mm256_fnmadd_pd(a, b, c);
#endif /* FMA */
}
// https://stackoverflow.com/questions/41144668/how-to-efficiently-perform-double-int64-conversions-with-sse-avx
// Only works for inputs in the range: [-2^51, 2^51]
static inline __m256i _mm256_cvtpd_epi64_custom(__m256d x)
{
x = _mm256_add_pd(x, *(v4sd *) _pd256_epi64_mask);
return _mm256_sub_epi64(
_mm256_castpd_si256(x),
_mm256_castpd_si256(*(v4sd *) _pd256_epi64_mask));
}
/*
static inline v4sid _mm256_cvtpd_epi64_custom(v4sd x)
{
x = _mm256_add_pd(x, *(v4sd *) _pd256_PDEPI64U);
return _mm256_xor_si256(
_mm256_castpd_si256(x),
_mm256_castpd_si256(*(v4sd *) _pd256_PDEPI64U));
}
*/
static inline v4sd _mm256_cvtepi64_pd_custom(v4sid x)
{
x = _mm256_or_si256(x, _mm256_castpd_si256(*(v4sd *) _pd256_PDEPI64U));
return _mm256_sub_pd(_mm256_castsi256_pd(x), *(v4sd *) _pd256_PDEPI64U);
}
static inline v4sd _mm256_cvtepi64_pd_signed_custom(v4sid x)
{
x = _mm256_add_epi64(x, _mm256_castpd_si256(*(v4sd *) _pd256_epi64_mask));
return _mm256_sub_pd(_mm256_castsi256_pd(x), *(v4sd *) _pd256_epi64_mask);
}
#include "avx_mathfun.h"
static inline v8sfx2 _mm256_load2_ps(float const *mem_addr)
{
v4sfx2 src_1 = _mm_load2_ps(mem_addr);
v4sfx2 src_2 = _mm_load2_ps(mem_addr + 2 * SSE_LEN_FLOAT);
v8sfx2 ret;
ret.val[0] = _mm256_set_m128(src_2.val[0], src_1.val[0]);
ret.val[1] = _mm256_set_m128(src_2.val[1], src_1.val[1]);
return ret;
}
static inline v8sfx2 _mm256_load2u_ps(float const *mem_addr)
{
v4sfx2 src_1 = _mm_load2u_ps(mem_addr);
v4sfx2 src_2 = _mm_load2u_ps(mem_addr + 2 * SSE_LEN_FLOAT);
v8sfx2 ret;
ret.val[0] = _mm256_set_m128(src_2.val[0], src_1.val[0]);
ret.val[1] = _mm256_set_m128(src_2.val[1], src_1.val[1]);
return ret;
}
static inline void _mm256_store2_ps(float *mem_addr, v8sfx2 a)
{
v8sf cplx0 = _mm256_unpacklo_ps(a.val[0], a.val[1]);
v8sf cplx1 = _mm256_unpackhi_ps(a.val[0], a.val[1]);
v8sf perm0 = _mm256_permute2f128_ps(cplx0, cplx1, 0x20); // permute mask [cplx1(127:0],cplx0[127:0])
v8sf perm1 = _mm256_permute2f128_ps(cplx0, cplx1, 0x31); // permute mask [cplx1(255:128],cplx0[255:128])
_mm256_store_ps(mem_addr, perm0);
_mm256_store_ps(mem_addr + AVX_LEN_FLOAT, perm1);
}
static inline void _mm256_store2u_ps(float *mem_addr, v8sfx2 a)
{
v8sf cplx0 = _mm256_unpacklo_ps(a.val[0], a.val[1]);
v8sf cplx1 = _mm256_unpackhi_ps(a.val[0], a.val[1]);
v8sf perm0 = _mm256_permute2f128_ps(cplx0, cplx1, 0x20); // permute mask [cplx1(127:0],cplx0[127:0])
v8sf perm1 = _mm256_permute2f128_ps(cplx0, cplx1, 0x31); // permute mask [cplx1(255:128],cplx0[255:128])
_mm256_storeu_ps(mem_addr, perm0);
_mm256_storeu_ps(mem_addr + AVX_LEN_FLOAT, perm1);
}
static inline v4sdx2 _mm256_load2_pd(double const *mem_addr)
{
v2sdx2 src_1 = _mm_load2_pd(mem_addr);
v2sdx2 src_2 = _mm_load2_pd(mem_addr + 2 * SSE_LEN_DOUBLE);
v4sdx2 ret;
ret.val[0] = _mm256_set_m128d(src_2.val[0], src_1.val[0]);
ret.val[1] = _mm256_set_m128d(src_2.val[1], src_1.val[1]);
return ret;
}
static inline v4sdx2 _mm256_load2u_pd(double const *mem_addr)
{
v2sdx2 src_1 = _mm_load2u_pd(mem_addr);
v2sdx2 src_2 = _mm_load2u_pd(mem_addr + 2 * SSE_LEN_DOUBLE);
v4sdx2 ret;
ret.val[0] = _mm256_set_m128d(src_2.val[0], src_1.val[0]);
ret.val[1] = _mm256_set_m128d(src_2.val[1], src_1.val[1]);
return ret;
}
static inline void _mm256_store2_pd(double *mem_addr, v4sdx2 a)
{
v4sd cplx0 = _mm256_unpacklo_pd(a.val[0], a.val[1]);
v4sd cplx1 = _mm256_unpackhi_pd(a.val[0], a.val[1]);
v4sd perm0 = _mm256_permute2f128_pd(cplx0, cplx1, 0x20); // permute mask [cplx1(127:0],cplx0[127:0])
v4sd perm1 = _mm256_permute2f128_pd(cplx0, cplx1, 0x31); // permute mask [cplx1(255:128],cplx0[255:128])
_mm256_store_pd(mem_addr, perm0);
_mm256_store_pd(mem_addr + AVX_LEN_DOUBLE, perm1);
}
static inline void _mm256_store2u_pd(double *mem_addr, v4sdx2 a)
{
v4sd cplx0 = _mm256_unpacklo_pd(a.val[0], a.val[1]);
v4sd cplx1 = _mm256_unpackhi_pd(a.val[0], a.val[1]);
v4sd perm0 = _mm256_permute2f128_pd(cplx0, cplx1, 0x20); // permute mask [cplx1(127:0],cplx0[127:0])
v4sd perm1 = _mm256_permute2f128_pd(cplx0, cplx1, 0x31); // permute mask [cplx1(255:128],cplx0[255:128])
_mm256_storeu_pd(mem_addr, perm0);
_mm256_storeu_pd(mem_addr + AVX_LEN_DOUBLE, perm1);
}
#include "simd_utils_avx_double.h"
#include "simd_utils_avx_float.h"
#include "simd_utils_avx_int32.h"
#endif /* AVX */
#ifdef AVX512
static const float _ps512_conj_mask[16] __attribute__((aligned(64))) = {1.0f, -1.0f, 1.0f, -1.0f, 1.0f, -1.0f, 1.0f, -1.0f,
1.0f, -1.0f, 1.0f, -1.0f, 1.0f, -1.0f, 1.0f, -1.0f};
static inline __m512 _mm512_fmadd_ps_custom(__m512 a, __m512 b, __m512 c)
{
return _mm512_fmadd_ps(a, b, c);
}
static inline __m512 _mm512_fmaddsub_ps_custom(__m512 a, __m512 b, __m512 c)
{
return _mm512_fmaddsub_ps(a, b, c);
}
static inline __m512 _mm512_fnmadd_ps_custom(__m512 a, __m512 b, __m512 c)
{
return _mm512_fnmadd_ps(a, b, c);
}
static inline __m512d _mm512_fmadd_pd_custom(__m512d a, __m512d b, __m512d c)
{
return _mm512_fmadd_pd(a, b, c);
}
static inline __m512d _mm512_fnmadd_pd_custom(__m512d a, __m512d b, __m512d c)
{
return _mm512_fnmadd_pd(a, b, c);
}
#include "avx512_mathfun.h"
static inline v16sfx2 _mm512_load2_ps(float const *mem_addr)
{
v16sf vec1 = _mm512_load_ps(mem_addr); // load 0 1 2 3 4 5 6 7
v16sf vec2 = _mm512_load_ps(mem_addr + AVX512_LEN_FLOAT); // load 8 9 10 11 12 13 14 15
v16sfx2 ret;
ret.val[0] = _mm512_permutex2var_ps(vec2, *(v16si *) _pi32_512_idx_re, vec1);
ret.val[1] = _mm512_permutex2var_ps(vec2, *(v16si *) _pi32_512_idx_im, vec1);
return ret;
}
static inline v16sfx2 _mm512_load2u_ps(float const *mem_addr)
{
v16sf vec1 = _mm512_loadu_ps(mem_addr); // load 0 1 2 3 4 5 6 7
v16sf vec2 = _mm512_loadu_ps(mem_addr + AVX512_LEN_FLOAT); // load 8 9 10 11 12 13 14 15
v16sfx2 ret;
ret.val[0] = _mm512_permutex2var_ps(vec2, *(v16si *) _pi32_512_idx_re, vec1);
ret.val[1] = _mm512_permutex2var_ps(vec2, *(v16si *) _pi32_512_idx_im, vec1);
return ret;
}
static inline void _mm512_store2_ps(float *mem_addr, v16sfx2 a)
{
v16sf tmp1 = _mm512_permutex2var_ps(a.val[1], *(v16si *) _pi32_512_idx_cplx_lo, a.val[0]);
v16sf tmp2 = _mm512_permutex2var_ps(a.val[1], *(v16si *) _pi32_512_idx_cplx_hi, a.val[0]);
_mm512_store_ps(mem_addr, tmp1);
_mm512_store_ps(mem_addr + AVX512_LEN_FLOAT, tmp2);
}
static inline void _mm512_store2u_ps(float *mem_addr, v16sfx2 a)
{
v16sf tmp1 = _mm512_permutex2var_ps(a.val[1], *(v16si *) _pi32_512_idx_cplx_lo, a.val[0]);
v16sf tmp2 = _mm512_permutex2var_ps(a.val[1], *(v16si *) _pi32_512_idx_cplx_hi, a.val[0]);
_mm512_storeu_ps(mem_addr, tmp1);
_mm512_storeu_ps(mem_addr + AVX512_LEN_FLOAT, tmp2);
}
static inline v8sdx2 _mm512_load2_pd(double const *mem_addr)
{
v8sd vec1 = _mm512_load_pd(mem_addr); // load 0 1 2 3 4 5 6 7
v8sd vec2 = _mm512_load_pd(mem_addr + AVX512_LEN_DOUBLE); // load 8 9 10 11 12 13 14 15
v8sdx2 ret;
ret.val[0] = _mm512_permutex2var_pd(vec2, *(v8sid *) _pi64_512_idx_re, vec1);
ret.val[1] = _mm512_permutex2var_pd(vec2, *(v8sid *) _pi64_512_idx_im, vec1);
return ret;
}
static inline v8sdx2 _mm512_load2u_pd(double const *mem_addr)
{
v8sd vec1 = _mm512_loadu_pd(mem_addr); // load 0 1 2 3 4 5 6 7
v8sd vec2 = _mm512_loadu_pd(mem_addr + AVX512_LEN_DOUBLE); // load 8 9 10 11 12 13 14 15
v8sdx2 ret;
ret.val[0] = _mm512_permutex2var_pd(vec2, *(v8sid *) _pi64_512_idx_re, vec1);
ret.val[1] = _mm512_permutex2var_pd(vec2, *(v8sid *) _pi64_512_idx_im, vec1);
return ret;
}
static inline void _mm512_store2_pd(double *mem_addr, v8sdx2 a)
{
v8sd tmp1 = _mm512_permutex2var_pd(a.val[1], *(v8sid *) _pi64_512_idx_cplx_lo, a.val[0]);
v8sd tmp2 = _mm512_permutex2var_pd(a.val[1], *(v8sid *) _pi64_512_idx_cplx_hi, a.val[0]);
_mm512_store_pd(mem_addr, tmp1);
_mm512_store_pd(mem_addr + AVX512_LEN_DOUBLE, tmp2);
}
static inline void _mm512_store2u_pd(double *mem_addr, v8sdx2 a)
{
v8sd tmp1 = _mm512_permutex2var_pd(a.val[1], *(v8sid *) _pi64_512_idx_cplx_lo, a.val[0]);
v8sd tmp2 = _mm512_permutex2var_pd(a.val[1], *(v8sid *) _pi64_512_idx_cplx_hi, a.val[0]);
_mm512_storeu_pd(mem_addr, tmp1);
_mm512_storeu_pd(mem_addr + AVX512_LEN_DOUBLE, tmp2);
}
#include "simd_utils_avx512_double.h"
#include "simd_utils_avx512_float.h"
#include "simd_utils_avx512_int32.h"
#endif /* AVX512 */
#ifdef ICC
#include "simd_utils_svml.h"
#endif
#ifdef AMDLIBM
#include "simd_utils_amdlibm.h"
#endif
#ifdef RISCV /* RISCV */
#ifndef __linux__
/* Get current value of CLOCK and store it in TP. */
int clock_gettime(clockid_t clock_id, struct timespec *tp)
{
struct timeval tv;
int retval = gettimeofday(&tv, NULL);
if (retval == 0)
/* Convert into `timespec'. */
TIMEVAL_TO_TIMESPEC(&tv, tp);
return retval;
}
#endif
static inline uint32_t _MM_GET_ROUNDING_MODE()
{
uint32_t reg;
asm volatile("frrm %0"
: "=r"(reg));
return reg;
}
static inline void _MM_SET_ROUNDING_MODE(uint32_t mode)
{
uint32_t reg;
switch (mode) {
case _MM_ROUND_NEAREST: //IEEE 754 round to nearest even
asm volatile("fsrmi %0,0"
: "=r"(reg));
break;
case _MM_ROUND_TOWARD_ZERO: // trunc
asm volatile("fsrmi %0,1"
: "=r"(reg));
break;
case _MM_ROUND_DOWN:
asm volatile("fsrmi %0,2"
: "=r"(reg));
break;
case _MM_ROUND_UP:
asm volatile("fsrmi %0,3"
: "=r"(reg));
break;
case _MM_ROUND_AWAY: //C round away from zero
asm volatile("fsrmi %0,4"
: "=r"(reg));
break;
default:
printf("_MM_SET_ROUNDING_MODE wrong mod requested %d\n", mode);
}
}
#include "simd_utils_riscv_double.h"
#include "simd_utils_riscv_float.h"
#include "simd_utils_riscv_int.h"
#endif /* RISCV */
#ifdef ALTIVEC
#include <altivec.h>
// Compare and perm operations => perm unit
// On e6500, VPERM operations take 2 cycles. VFPU operations take 6 cycles.
// Complex FPU operations take 7 cycles (and block the unit for 2 cycles)