numo-narray-alt 0.10.5 → 0.11.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- checksums.yaml +4 -4
- data/README.md +9 -0
- data/ext/numo/narray/extconf.rb +2 -0
- data/ext/numo/narray/numo/narray.h +3 -3
- data/ext/numo/narray/src/mh/math/sqrt.h +372 -0
- data/ext/numo/narray/src/mh/mulsum.h +1360 -74
- data/ext/numo/narray/src/mh/op/add.h +20 -0
- data/ext/numo/narray/src/mh/op/binary_func.h +542 -0
- data/ext/numo/narray/src/mh/op/div.h +20 -0
- data/ext/numo/narray/src/mh/op/mul.h +20 -0
- data/ext/numo/narray/src/mh/op/sub.h +20 -0
- data/ext/numo/narray/src/mh/sort.h +4 -4
- data/ext/numo/narray/src/t_bit.c +0 -5
- data/ext/numo/narray/src/t_dcomplex.c +0 -5
- data/ext/numo/narray/src/t_dfloat.c +35 -3
- data/ext/numo/narray/src/t_int16.c +0 -5
- data/ext/numo/narray/src/t_int32.c +0 -5
- data/ext/numo/narray/src/t_int64.c +0 -5
- data/ext/numo/narray/src/t_int8.c +0 -5
- data/ext/numo/narray/src/t_robject.c +0 -5
- data/ext/numo/narray/src/t_scomplex.c +0 -5
- data/ext/numo/narray/src/t_sfloat.c +35 -3
- data/ext/numo/narray/src/t_uint16.c +0 -5
- data/ext/numo/narray/src/t_uint32.c +0 -5
- data/ext/numo/narray/src/t_uint64.c +0 -5
- data/ext/numo/narray/src/t_uint8.c +0 -5
- metadata +2 -2
|
@@ -75,4 +75,24 @@
|
|
|
75
75
|
DEF_BINARY_SELF_FUNC(add, dfloat, numo_cDFloat) \
|
|
76
76
|
DEF_BINARY_FUNC(add, '+', dfloat, numo_cDFloat)
|
|
77
77
|
|
|
78
|
+
#define DEF_NARRAY_SFLT_ADD_AVX_METHOD_FUNC() \
|
|
79
|
+
DEF_BINARY_SFLT_AVX_ITER_FUNC(add, _mm256_add_ps) \
|
|
80
|
+
DEF_BINARY_SELF_FUNC(add, sfloat, numo_cSFloat) \
|
|
81
|
+
DEF_BINARY_FUNC(add, '+', sfloat, numo_cSFloat)
|
|
82
|
+
|
|
83
|
+
#define DEF_NARRAY_DFLT_ADD_AVX_METHOD_FUNC() \
|
|
84
|
+
DEF_BINARY_DFLT_AVX_ITER_FUNC(add, _mm256_add_pd) \
|
|
85
|
+
DEF_BINARY_SELF_FUNC(add, dfloat, numo_cDFloat) \
|
|
86
|
+
DEF_BINARY_FUNC(add, '+', dfloat, numo_cDFloat)
|
|
87
|
+
|
|
88
|
+
#define DEF_NARRAY_SFLT_ADD_NEON_METHOD_FUNC() \
|
|
89
|
+
DEF_BINARY_SFLT_NEON_ITER_FUNC(add, vaddq_f32) \
|
|
90
|
+
DEF_BINARY_SELF_FUNC(add, sfloat, numo_cSFloat) \
|
|
91
|
+
DEF_BINARY_FUNC(add, '+', sfloat, numo_cSFloat)
|
|
92
|
+
|
|
93
|
+
#define DEF_NARRAY_DFLT_ADD_NEON_METHOD_FUNC() \
|
|
94
|
+
DEF_BINARY_DFLT_NEON_ITER_FUNC(add, vaddq_f64) \
|
|
95
|
+
DEF_BINARY_SELF_FUNC(add, dfloat, numo_cDFloat) \
|
|
96
|
+
DEF_BINARY_FUNC(add, '+', dfloat, numo_cDFloat)
|
|
97
|
+
|
|
78
98
|
#endif /* NUMO_NARRAY_MH_OP_ADD_H */
|
|
@@ -420,4 +420,546 @@
|
|
|
420
420
|
} \
|
|
421
421
|
}
|
|
422
422
|
|
|
423
|
+
#define DEF_BINARY_SFLT_AVX_ITER_FUNC(fOpFunc, fSimdOp) \
|
|
424
|
+
static void iter_sfloat_##fOpFunc(na_loop_t* const lp) { \
|
|
425
|
+
size_t i = 0; \
|
|
426
|
+
ITER_BINARY_INIT_VARS() \
|
|
427
|
+
\
|
|
428
|
+
size_t cnt; \
|
|
429
|
+
size_t cnt_simd_loop = -1; \
|
|
430
|
+
__m256 a; \
|
|
431
|
+
__m256 b; \
|
|
432
|
+
size_t num_pack; \
|
|
433
|
+
num_pack = AVX_ALIGNMENT_SIZE / sizeof(sfloat); \
|
|
434
|
+
\
|
|
435
|
+
if (is_aligned(p1, sizeof(sfloat)) && is_aligned(p2, sizeof(sfloat)) && \
|
|
436
|
+
is_aligned(p3, sizeof(sfloat))) { \
|
|
437
|
+
if (s1 == sizeof(sfloat) && s2 == sizeof(sfloat) && s3 == sizeof(sfloat)) { \
|
|
438
|
+
if ((n >= num_pack) && \
|
|
439
|
+
is_same_aligned3( \
|
|
440
|
+
&((sfloat*)p1)[i], &((sfloat*)p2)[i], &((sfloat*)p3)[i], AVX_ALIGNMENT_SIZE \
|
|
441
|
+
)) { \
|
|
442
|
+
cnt = get_count_of_elements_not_aligned_to_simd_size( \
|
|
443
|
+
&((sfloat*)p1)[i], AVX_ALIGNMENT_SIZE, sizeof(sfloat) \
|
|
444
|
+
); \
|
|
445
|
+
if (p1 == p3) { \
|
|
446
|
+
for (i = 0; i < cnt; i++) { \
|
|
447
|
+
((sfloat*)p1)[i] = m_##fOpFunc(((sfloat*)p1)[i], ((sfloat*)p2)[i]); \
|
|
448
|
+
} \
|
|
449
|
+
} else { \
|
|
450
|
+
for (i = 0; i < cnt; i++) { \
|
|
451
|
+
((sfloat*)p3)[i] = m_##fOpFunc(((sfloat*)p1)[i], ((sfloat*)p2)[i]); \
|
|
452
|
+
} \
|
|
453
|
+
} \
|
|
454
|
+
cnt_simd_loop = (n - i) % num_pack; \
|
|
455
|
+
if (p1 == p3) { \
|
|
456
|
+
for (; i < n - cnt_simd_loop; i += num_pack) { \
|
|
457
|
+
a = _mm256_load_ps(&((sfloat*)p1)[i]); \
|
|
458
|
+
b = _mm256_load_ps(&((sfloat*)p2)[i]); \
|
|
459
|
+
a = fSimdOp(a, b); \
|
|
460
|
+
_mm256_store_ps(&((sfloat*)p1)[i], a); \
|
|
461
|
+
} \
|
|
462
|
+
} else { \
|
|
463
|
+
for (; i < n - cnt_simd_loop; i += num_pack) { \
|
|
464
|
+
a = _mm256_load_ps(&((sfloat*)p1)[i]); \
|
|
465
|
+
b = _mm256_load_ps(&((sfloat*)p2)[i]); \
|
|
466
|
+
a = fSimdOp(a, b); \
|
|
467
|
+
_mm256_stream_ps(&((sfloat*)p3)[i], a); \
|
|
468
|
+
} \
|
|
469
|
+
} \
|
|
470
|
+
} \
|
|
471
|
+
if (cnt_simd_loop != 0) { \
|
|
472
|
+
if (p1 == p3) { \
|
|
473
|
+
for (; i < n; i++) { \
|
|
474
|
+
((sfloat*)p1)[i] = m_##fOpFunc(((sfloat*)p1)[i], ((sfloat*)p2)[i]); \
|
|
475
|
+
} \
|
|
476
|
+
} else { \
|
|
477
|
+
for (; i < n; i++) { \
|
|
478
|
+
((sfloat*)p3)[i] = m_##fOpFunc(((sfloat*)p1)[i], ((sfloat*)p2)[i]); \
|
|
479
|
+
} \
|
|
480
|
+
} \
|
|
481
|
+
} \
|
|
482
|
+
return; \
|
|
483
|
+
} \
|
|
484
|
+
if (is_aligned_step(s1, sizeof(sfloat)) && is_aligned_step(s2, sizeof(sfloat)) && \
|
|
485
|
+
is_aligned_step(s3, sizeof(sfloat))) { \
|
|
486
|
+
if (s2 == 0) { \
|
|
487
|
+
if (s1 == sizeof(sfloat) && s3 == sizeof(sfloat)) { \
|
|
488
|
+
b = _mm256_broadcast_ss(&((sfloat*)p2)[0]); \
|
|
489
|
+
if ((n >= num_pack) && \
|
|
490
|
+
is_same_aligned2(&((sfloat*)p1)[i], &((sfloat*)p3)[i], AVX_ALIGNMENT_SIZE)) { \
|
|
491
|
+
cnt = get_count_of_elements_not_aligned_to_simd_size( \
|
|
492
|
+
&((sfloat*)p1)[i], AVX_ALIGNMENT_SIZE, sizeof(sfloat) \
|
|
493
|
+
); \
|
|
494
|
+
if (p1 == p3) { \
|
|
495
|
+
for (i = 0; i < cnt; i++) { \
|
|
496
|
+
((sfloat*)p1)[i] = m_##fOpFunc(((sfloat*)p1)[i], *(sfloat*)p2); \
|
|
497
|
+
} \
|
|
498
|
+
} else { \
|
|
499
|
+
for (i = 0; i < cnt; i++) { \
|
|
500
|
+
((sfloat*)p3)[i] = m_##fOpFunc(((sfloat*)p1)[i], *(sfloat*)p2); \
|
|
501
|
+
} \
|
|
502
|
+
} \
|
|
503
|
+
cnt_simd_loop = (n - i) % num_pack; \
|
|
504
|
+
if (p1 == p3) { \
|
|
505
|
+
for (; i < n - cnt_simd_loop; i += num_pack) { \
|
|
506
|
+
a = _mm256_load_ps(&((sfloat*)p1)[i]); \
|
|
507
|
+
a = fSimdOp(a, b); \
|
|
508
|
+
_mm256_store_ps(&((sfloat*)p1)[i], a); \
|
|
509
|
+
} \
|
|
510
|
+
} else { \
|
|
511
|
+
for (; i < n - cnt_simd_loop; i += num_pack) { \
|
|
512
|
+
a = _mm256_load_ps(&((sfloat*)p1)[i]); \
|
|
513
|
+
a = fSimdOp(a, b); \
|
|
514
|
+
_mm256_stream_ps(&((sfloat*)p3)[i], a); \
|
|
515
|
+
} \
|
|
516
|
+
} \
|
|
517
|
+
} \
|
|
518
|
+
if (cnt_simd_loop != 0) { \
|
|
519
|
+
if (p1 == p3) { \
|
|
520
|
+
for (; i < n; i++) { \
|
|
521
|
+
((sfloat*)p1)[i] = m_##fOpFunc(((sfloat*)p1)[i], *(sfloat*)p2); \
|
|
522
|
+
} \
|
|
523
|
+
} else { \
|
|
524
|
+
for (; i < n; i++) { \
|
|
525
|
+
((sfloat*)p3)[i] = m_##fOpFunc(((sfloat*)p1)[i], *(sfloat*)p2); \
|
|
526
|
+
} \
|
|
527
|
+
} \
|
|
528
|
+
} \
|
|
529
|
+
} else { \
|
|
530
|
+
for (i = 0; i < n; i++) { \
|
|
531
|
+
*(sfloat*)p3 = m_##fOpFunc(*(sfloat*)p1, *(sfloat*)p2); \
|
|
532
|
+
p1 += s1; \
|
|
533
|
+
p3 += s3; \
|
|
534
|
+
} \
|
|
535
|
+
} \
|
|
536
|
+
} else { \
|
|
537
|
+
if (p1 == p3) { \
|
|
538
|
+
for (i = 0; i < n; i++) { \
|
|
539
|
+
*(sfloat*)p1 = m_##fOpFunc(*(sfloat*)p1, *(sfloat*)p2); \
|
|
540
|
+
p1 += s1; \
|
|
541
|
+
p2 += s2; \
|
|
542
|
+
} \
|
|
543
|
+
} else { \
|
|
544
|
+
for (i = 0; i < n; i++) { \
|
|
545
|
+
*(sfloat*)p3 = m_##fOpFunc(*(sfloat*)p1, *(sfloat*)p2); \
|
|
546
|
+
p1 += s1; \
|
|
547
|
+
p2 += s2; \
|
|
548
|
+
p3 += s3; \
|
|
549
|
+
} \
|
|
550
|
+
} \
|
|
551
|
+
} \
|
|
552
|
+
return; \
|
|
553
|
+
} \
|
|
554
|
+
} \
|
|
555
|
+
\
|
|
556
|
+
for (i = 0; i < n; i++) { \
|
|
557
|
+
sfloat x; \
|
|
558
|
+
sfloat y; \
|
|
559
|
+
sfloat z; \
|
|
560
|
+
GET_DATA_STRIDE(p1, s1, sfloat, x); \
|
|
561
|
+
GET_DATA_STRIDE(p2, s2, sfloat, y); \
|
|
562
|
+
z = m_##fOpFunc(x, y); \
|
|
563
|
+
SET_DATA_STRIDE(p3, s3, sfloat, z); \
|
|
564
|
+
} \
|
|
565
|
+
}
|
|
566
|
+
|
|
567
|
+
#define DEF_BINARY_DFLT_AVX_ITER_FUNC(fOpFunc, fSimdOp) \
|
|
568
|
+
static void iter_dfloat_##fOpFunc(na_loop_t* const lp) { \
|
|
569
|
+
size_t i = 0; \
|
|
570
|
+
ITER_BINARY_INIT_VARS() \
|
|
571
|
+
\
|
|
572
|
+
size_t cnt; \
|
|
573
|
+
size_t cnt_simd_loop = -1; \
|
|
574
|
+
__m256d a; \
|
|
575
|
+
__m256d b; \
|
|
576
|
+
size_t num_pack; \
|
|
577
|
+
num_pack = AVX_ALIGNMENT_SIZE / sizeof(dfloat); \
|
|
578
|
+
\
|
|
579
|
+
if (is_aligned(p1, sizeof(dfloat)) && is_aligned(p2, sizeof(dfloat)) && \
|
|
580
|
+
is_aligned(p3, sizeof(dfloat))) { \
|
|
581
|
+
if (s1 == sizeof(dfloat) && s2 == sizeof(dfloat) && s3 == sizeof(dfloat)) { \
|
|
582
|
+
if ((n >= num_pack) && \
|
|
583
|
+
is_same_aligned3( \
|
|
584
|
+
&((dfloat*)p1)[i], &((dfloat*)p2)[i], &((dfloat*)p3)[i], AVX_ALIGNMENT_SIZE \
|
|
585
|
+
)) { \
|
|
586
|
+
cnt = get_count_of_elements_not_aligned_to_simd_size( \
|
|
587
|
+
&((dfloat*)p1)[i], AVX_ALIGNMENT_SIZE, sizeof(dfloat) \
|
|
588
|
+
); \
|
|
589
|
+
if (p1 == p3) { \
|
|
590
|
+
for (i = 0; i < cnt; i++) { \
|
|
591
|
+
((dfloat*)p1)[i] = m_##fOpFunc(((dfloat*)p1)[i], ((dfloat*)p2)[i]); \
|
|
592
|
+
} \
|
|
593
|
+
} else { \
|
|
594
|
+
for (i = 0; i < cnt; i++) { \
|
|
595
|
+
((dfloat*)p3)[i] = m_##fOpFunc(((dfloat*)p1)[i], ((dfloat*)p2)[i]); \
|
|
596
|
+
} \
|
|
597
|
+
} \
|
|
598
|
+
cnt_simd_loop = (n - i) % num_pack; \
|
|
599
|
+
if (p1 == p3) { \
|
|
600
|
+
for (; i < n - cnt_simd_loop; i += num_pack) { \
|
|
601
|
+
a = _mm256_load_pd(&((dfloat*)p1)[i]); \
|
|
602
|
+
b = _mm256_load_pd(&((dfloat*)p2)[i]); \
|
|
603
|
+
a = fSimdOp(a, b); \
|
|
604
|
+
_mm256_store_pd(&((dfloat*)p1)[i], a); \
|
|
605
|
+
} \
|
|
606
|
+
} else { \
|
|
607
|
+
for (; i < n - cnt_simd_loop; i += num_pack) { \
|
|
608
|
+
a = _mm256_load_pd(&((dfloat*)p1)[i]); \
|
|
609
|
+
b = _mm256_load_pd(&((dfloat*)p2)[i]); \
|
|
610
|
+
a = fSimdOp(a, b); \
|
|
611
|
+
_mm256_stream_pd(&((dfloat*)p3)[i], a); \
|
|
612
|
+
} \
|
|
613
|
+
} \
|
|
614
|
+
} \
|
|
615
|
+
if (cnt_simd_loop != 0) { \
|
|
616
|
+
if (p1 == p3) { \
|
|
617
|
+
for (; i < n; i++) { \
|
|
618
|
+
((dfloat*)p1)[i] = m_##fOpFunc(((dfloat*)p1)[i], ((dfloat*)p2)[i]); \
|
|
619
|
+
} \
|
|
620
|
+
} else { \
|
|
621
|
+
for (; i < n; i++) { \
|
|
622
|
+
((dfloat*)p3)[i] = m_##fOpFunc(((dfloat*)p1)[i], ((dfloat*)p2)[i]); \
|
|
623
|
+
} \
|
|
624
|
+
} \
|
|
625
|
+
} \
|
|
626
|
+
return; \
|
|
627
|
+
} \
|
|
628
|
+
if (is_aligned_step(s1, sizeof(dfloat)) && is_aligned_step(s2, sizeof(dfloat)) && \
|
|
629
|
+
is_aligned_step(s3, sizeof(dfloat))) { \
|
|
630
|
+
if (s2 == 0) { \
|
|
631
|
+
if (s1 == sizeof(dfloat) && s3 == sizeof(dfloat)) { \
|
|
632
|
+
b = _mm256_broadcast_sd(&((dfloat*)p2)[0]); \
|
|
633
|
+
if ((n >= num_pack) && \
|
|
634
|
+
is_same_aligned2(&((dfloat*)p1)[i], &((dfloat*)p3)[i], AVX_ALIGNMENT_SIZE)) { \
|
|
635
|
+
cnt = get_count_of_elements_not_aligned_to_simd_size( \
|
|
636
|
+
&((dfloat*)p1)[i], AVX_ALIGNMENT_SIZE, sizeof(dfloat) \
|
|
637
|
+
); \
|
|
638
|
+
if (p1 == p3) { \
|
|
639
|
+
for (; i < cnt; i++) { \
|
|
640
|
+
((dfloat*)p1)[i] = m_##fOpFunc(((dfloat*)p1)[i], *(dfloat*)p2); \
|
|
641
|
+
} \
|
|
642
|
+
} else { \
|
|
643
|
+
for (; i < cnt; i++) { \
|
|
644
|
+
((dfloat*)p3)[i] = m_##fOpFunc(((dfloat*)p1)[i], *(dfloat*)p2); \
|
|
645
|
+
} \
|
|
646
|
+
} \
|
|
647
|
+
cnt_simd_loop = (n - i) % num_pack; \
|
|
648
|
+
if (p1 == p3) { \
|
|
649
|
+
for (; i < n - cnt_simd_loop; i += num_pack) { \
|
|
650
|
+
a = _mm256_load_pd(&((dfloat*)p1)[i]); \
|
|
651
|
+
a = fSimdOp(a, b); \
|
|
652
|
+
_mm256_store_pd(&((dfloat*)p1)[i], a); \
|
|
653
|
+
} \
|
|
654
|
+
} else { \
|
|
655
|
+
for (; i < n - cnt_simd_loop; i += num_pack) { \
|
|
656
|
+
a = _mm256_load_pd(&((dfloat*)p1)[i]); \
|
|
657
|
+
a = fSimdOp(a, b); \
|
|
658
|
+
_mm256_stream_pd(&((dfloat*)p3)[i], a); \
|
|
659
|
+
} \
|
|
660
|
+
} \
|
|
661
|
+
} \
|
|
662
|
+
if (cnt_simd_loop != 0) { \
|
|
663
|
+
if (p1 == p3) { \
|
|
664
|
+
for (; i < n; i++) { \
|
|
665
|
+
((dfloat*)p1)[i] = m_##fOpFunc(((dfloat*)p1)[i], *(dfloat*)p2); \
|
|
666
|
+
} \
|
|
667
|
+
} else { \
|
|
668
|
+
for (; i < n; i++) { \
|
|
669
|
+
((dfloat*)p3)[i] = m_##fOpFunc(((dfloat*)p1)[i], *(dfloat*)p2); \
|
|
670
|
+
} \
|
|
671
|
+
} \
|
|
672
|
+
} \
|
|
673
|
+
} else { \
|
|
674
|
+
for (i = 0; i < n; i++) { \
|
|
675
|
+
*(dfloat*)p3 = m_##fOpFunc(*(dfloat*)p1, *(dfloat*)p2); \
|
|
676
|
+
p1 += s1; \
|
|
677
|
+
p3 += s3; \
|
|
678
|
+
} \
|
|
679
|
+
} \
|
|
680
|
+
} else { \
|
|
681
|
+
if (p1 == p3) { \
|
|
682
|
+
for (i = 0; i < n; i++) { \
|
|
683
|
+
*(dfloat*)p1 = m_##fOpFunc(*(dfloat*)p1, *(dfloat*)p2); \
|
|
684
|
+
p1 += s1; \
|
|
685
|
+
p2 += s2; \
|
|
686
|
+
} \
|
|
687
|
+
} else { \
|
|
688
|
+
for (i = 0; i < n; i++) { \
|
|
689
|
+
*(dfloat*)p3 = m_##fOpFunc(*(dfloat*)p1, *(dfloat*)p2); \
|
|
690
|
+
p1 += s1; \
|
|
691
|
+
p2 += s2; \
|
|
692
|
+
p3 += s3; \
|
|
693
|
+
} \
|
|
694
|
+
} \
|
|
695
|
+
} \
|
|
696
|
+
return; \
|
|
697
|
+
} \
|
|
698
|
+
} \
|
|
699
|
+
\
|
|
700
|
+
for (i = 0; i < n; i++) { \
|
|
701
|
+
dfloat x; \
|
|
702
|
+
dfloat y; \
|
|
703
|
+
dfloat z; \
|
|
704
|
+
GET_DATA_STRIDE(p1, s1, dfloat, x); \
|
|
705
|
+
GET_DATA_STRIDE(p2, s2, dfloat, y); \
|
|
706
|
+
z = m_##fOpFunc(x, y); \
|
|
707
|
+
SET_DATA_STRIDE(p3, s3, dfloat, z); \
|
|
708
|
+
} \
|
|
709
|
+
}
|
|
710
|
+
|
|
711
|
+
#define DEF_BINARY_SFLT_NEON_ITER_FUNC(fOpFunc, fSimdOp) \
|
|
712
|
+
static void iter_sfloat_##fOpFunc(na_loop_t* const lp) { \
|
|
713
|
+
size_t i = 0; \
|
|
714
|
+
ITER_BINARY_INIT_VARS() \
|
|
715
|
+
\
|
|
716
|
+
size_t cnt; \
|
|
717
|
+
size_t cnt_simd_loop = -1; \
|
|
718
|
+
float32x4_t a; \
|
|
719
|
+
float32x4_t b; \
|
|
720
|
+
size_t num_pack; \
|
|
721
|
+
num_pack = NEON_ALIGNMENT_SIZE / sizeof(sfloat); \
|
|
722
|
+
\
|
|
723
|
+
if (is_aligned(p1, sizeof(sfloat)) && is_aligned(p2, sizeof(sfloat)) && \
|
|
724
|
+
is_aligned(p3, sizeof(sfloat))) { \
|
|
725
|
+
if (s1 == sizeof(sfloat) && s2 == sizeof(sfloat) && s3 == sizeof(sfloat)) { \
|
|
726
|
+
if ((n >= num_pack) && \
|
|
727
|
+
is_same_aligned3( \
|
|
728
|
+
&((sfloat*)p1)[i], &((sfloat*)p2)[i], &((sfloat*)p3)[i], NEON_ALIGNMENT_SIZE \
|
|
729
|
+
)) { \
|
|
730
|
+
cnt = get_count_of_elements_not_aligned_to_simd_size( \
|
|
731
|
+
&((sfloat*)p1)[i], NEON_ALIGNMENT_SIZE, sizeof(sfloat) \
|
|
732
|
+
); \
|
|
733
|
+
if (p1 == p3) { \
|
|
734
|
+
for (i = 0; i < cnt; i++) { \
|
|
735
|
+
((sfloat*)p1)[i] = m_##fOpFunc(((sfloat*)p1)[i], ((sfloat*)p2)[i]); \
|
|
736
|
+
} \
|
|
737
|
+
} else { \
|
|
738
|
+
for (i = 0; i < cnt; i++) { \
|
|
739
|
+
((sfloat*)p3)[i] = m_##fOpFunc(((sfloat*)p1)[i], ((sfloat*)p2)[i]); \
|
|
740
|
+
} \
|
|
741
|
+
} \
|
|
742
|
+
cnt_simd_loop = (n - i) % num_pack; \
|
|
743
|
+
for (; i < n - cnt_simd_loop; i += num_pack) { \
|
|
744
|
+
a = vld1q_f32(&((sfloat*)p1)[i]); \
|
|
745
|
+
b = vld1q_f32(&((sfloat*)p2)[i]); \
|
|
746
|
+
a = fSimdOp(a, b); \
|
|
747
|
+
vst1q_f32(&((sfloat*)p3)[i], a); \
|
|
748
|
+
} \
|
|
749
|
+
} \
|
|
750
|
+
if (cnt_simd_loop != 0) { \
|
|
751
|
+
if (p1 == p3) { \
|
|
752
|
+
for (; i < n; i++) { \
|
|
753
|
+
((sfloat*)p1)[i] = m_##fOpFunc(((sfloat*)p1)[i], ((sfloat*)p2)[i]); \
|
|
754
|
+
} \
|
|
755
|
+
} else { \
|
|
756
|
+
for (; i < n; i++) { \
|
|
757
|
+
((sfloat*)p3)[i] = m_##fOpFunc(((sfloat*)p1)[i], ((sfloat*)p2)[i]); \
|
|
758
|
+
} \
|
|
759
|
+
} \
|
|
760
|
+
} \
|
|
761
|
+
return; \
|
|
762
|
+
} \
|
|
763
|
+
if (is_aligned_step(s1, sizeof(sfloat)) && is_aligned_step(s2, sizeof(sfloat)) && \
|
|
764
|
+
is_aligned_step(s3, sizeof(sfloat))) { \
|
|
765
|
+
if (s2 == 0) { \
|
|
766
|
+
if (s1 == sizeof(sfloat) && s3 == sizeof(sfloat)) { \
|
|
767
|
+
b = vld1q_dup_f32(&((sfloat*)p2)[0]); \
|
|
768
|
+
if ((n >= num_pack) && \
|
|
769
|
+
is_same_aligned2(&((sfloat*)p1)[i], &((sfloat*)p3)[i], NEON_ALIGNMENT_SIZE)) { \
|
|
770
|
+
cnt = get_count_of_elements_not_aligned_to_simd_size( \
|
|
771
|
+
&((sfloat*)p1)[i], NEON_ALIGNMENT_SIZE, sizeof(sfloat) \
|
|
772
|
+
); \
|
|
773
|
+
if (p1 == p3) { \
|
|
774
|
+
for (i = 0; i < cnt; i++) { \
|
|
775
|
+
((sfloat*)p1)[i] = m_##fOpFunc(((sfloat*)p1)[i], *(sfloat*)p2); \
|
|
776
|
+
} \
|
|
777
|
+
} else { \
|
|
778
|
+
for (i = 0; i < cnt; i++) { \
|
|
779
|
+
((sfloat*)p3)[i] = m_##fOpFunc(((sfloat*)p1)[i], *(sfloat*)p2); \
|
|
780
|
+
} \
|
|
781
|
+
} \
|
|
782
|
+
cnt_simd_loop = (n - i) % num_pack; \
|
|
783
|
+
for (; i < n - cnt_simd_loop; i += num_pack) { \
|
|
784
|
+
a = vld1q_f32(&((sfloat*)p1)[i]); \
|
|
785
|
+
a = fSimdOp(a, b); \
|
|
786
|
+
vst1q_f32(&((sfloat*)p3)[i], a); \
|
|
787
|
+
} \
|
|
788
|
+
} \
|
|
789
|
+
if (cnt_simd_loop != 0) { \
|
|
790
|
+
if (p1 == p3) { \
|
|
791
|
+
for (; i < n; i++) { \
|
|
792
|
+
((sfloat*)p1)[i] = m_##fOpFunc(((sfloat*)p1)[i], *(sfloat*)p2); \
|
|
793
|
+
} \
|
|
794
|
+
} else { \
|
|
795
|
+
for (; i < n; i++) { \
|
|
796
|
+
((sfloat*)p3)[i] = m_##fOpFunc(((sfloat*)p1)[i], *(sfloat*)p2); \
|
|
797
|
+
} \
|
|
798
|
+
} \
|
|
799
|
+
} \
|
|
800
|
+
} else { \
|
|
801
|
+
for (i = 0; i < n; i++) { \
|
|
802
|
+
*(sfloat*)p3 = m_##fOpFunc(*(sfloat*)p1, *(sfloat*)p2); \
|
|
803
|
+
p1 += s1; \
|
|
804
|
+
p3 += s3; \
|
|
805
|
+
} \
|
|
806
|
+
} \
|
|
807
|
+
} else { \
|
|
808
|
+
if (p1 == p3) { \
|
|
809
|
+
for (i = 0; i < n; i++) { \
|
|
810
|
+
*(sfloat*)p1 = m_##fOpFunc(*(sfloat*)p1, *(sfloat*)p2); \
|
|
811
|
+
p1 += s1; \
|
|
812
|
+
p2 += s2; \
|
|
813
|
+
} \
|
|
814
|
+
} else { \
|
|
815
|
+
for (i = 0; i < n; i++) { \
|
|
816
|
+
*(sfloat*)p3 = m_##fOpFunc(*(sfloat*)p1, *(sfloat*)p2); \
|
|
817
|
+
p1 += s1; \
|
|
818
|
+
p2 += s2; \
|
|
819
|
+
p3 += s3; \
|
|
820
|
+
} \
|
|
821
|
+
} \
|
|
822
|
+
} \
|
|
823
|
+
return; \
|
|
824
|
+
} \
|
|
825
|
+
} \
|
|
826
|
+
\
|
|
827
|
+
for (i = 0; i < n; i++) { \
|
|
828
|
+
sfloat x; \
|
|
829
|
+
sfloat y; \
|
|
830
|
+
sfloat z; \
|
|
831
|
+
GET_DATA_STRIDE(p1, s1, sfloat, x); \
|
|
832
|
+
GET_DATA_STRIDE(p2, s2, sfloat, y); \
|
|
833
|
+
z = m_##fOpFunc(x, y); \
|
|
834
|
+
SET_DATA_STRIDE(p3, s3, sfloat, z); \
|
|
835
|
+
} \
|
|
836
|
+
}
|
|
837
|
+
|
|
838
|
+
#define DEF_BINARY_DFLT_NEON_ITER_FUNC(fOpFunc, fSimdOp) \
|
|
839
|
+
static void iter_dfloat_##fOpFunc(na_loop_t* const lp) { \
|
|
840
|
+
size_t i = 0; \
|
|
841
|
+
ITER_BINARY_INIT_VARS() \
|
|
842
|
+
\
|
|
843
|
+
size_t cnt; \
|
|
844
|
+
size_t cnt_simd_loop = -1; \
|
|
845
|
+
float64x2_t a; \
|
|
846
|
+
float64x2_t b; \
|
|
847
|
+
size_t num_pack; \
|
|
848
|
+
num_pack = NEON_ALIGNMENT_SIZE / sizeof(dfloat); \
|
|
849
|
+
\
|
|
850
|
+
if (is_aligned(p1, sizeof(dfloat)) && is_aligned(p2, sizeof(dfloat)) && \
|
|
851
|
+
is_aligned(p3, sizeof(dfloat))) { \
|
|
852
|
+
if (s1 == sizeof(dfloat) && s2 == sizeof(dfloat) && s3 == sizeof(dfloat)) { \
|
|
853
|
+
if ((n >= num_pack) && \
|
|
854
|
+
is_same_aligned3( \
|
|
855
|
+
&((dfloat*)p1)[i], &((dfloat*)p2)[i], &((dfloat*)p3)[i], NEON_ALIGNMENT_SIZE \
|
|
856
|
+
)) { \
|
|
857
|
+
cnt = get_count_of_elements_not_aligned_to_simd_size( \
|
|
858
|
+
&((dfloat*)p1)[i], NEON_ALIGNMENT_SIZE, sizeof(dfloat) \
|
|
859
|
+
); \
|
|
860
|
+
if (p1 == p3) { \
|
|
861
|
+
for (i = 0; i < cnt; i++) { \
|
|
862
|
+
((dfloat*)p1)[i] = m_##fOpFunc(((dfloat*)p1)[i], ((dfloat*)p2)[i]); \
|
|
863
|
+
} \
|
|
864
|
+
} else { \
|
|
865
|
+
for (i = 0; i < cnt; i++) { \
|
|
866
|
+
((dfloat*)p3)[i] = m_##fOpFunc(((dfloat*)p1)[i], ((dfloat*)p2)[i]); \
|
|
867
|
+
} \
|
|
868
|
+
} \
|
|
869
|
+
cnt_simd_loop = (n - i) % num_pack; \
|
|
870
|
+
for (; i < n - cnt_simd_loop; i += num_pack) { \
|
|
871
|
+
a = vld1q_f64(&((dfloat*)p1)[i]); \
|
|
872
|
+
b = vld1q_f64(&((dfloat*)p2)[i]); \
|
|
873
|
+
a = fSimdOp(a, b); \
|
|
874
|
+
vst1q_f64(&((dfloat*)p3)[i], a); \
|
|
875
|
+
} \
|
|
876
|
+
} \
|
|
877
|
+
if (cnt_simd_loop != 0) { \
|
|
878
|
+
if (p1 == p3) { \
|
|
879
|
+
for (; i < n; i++) { \
|
|
880
|
+
((dfloat*)p1)[i] = m_##fOpFunc(((dfloat*)p1)[i], ((dfloat*)p2)[i]); \
|
|
881
|
+
} \
|
|
882
|
+
} else { \
|
|
883
|
+
for (; i < n; i++) { \
|
|
884
|
+
((dfloat*)p3)[i] = m_##fOpFunc(((dfloat*)p1)[i], ((dfloat*)p2)[i]); \
|
|
885
|
+
} \
|
|
886
|
+
} \
|
|
887
|
+
} \
|
|
888
|
+
return; \
|
|
889
|
+
} \
|
|
890
|
+
if (is_aligned_step(s1, sizeof(dfloat)) && is_aligned_step(s2, sizeof(dfloat)) && \
|
|
891
|
+
is_aligned_step(s3, sizeof(dfloat))) { \
|
|
892
|
+
if (s2 == 0) { \
|
|
893
|
+
if (s1 == sizeof(dfloat) && s3 == sizeof(dfloat)) { \
|
|
894
|
+
b = vld1q_dup_f64(&((dfloat*)p2)[0]); \
|
|
895
|
+
if ((n >= num_pack) && \
|
|
896
|
+
is_same_aligned2(&((dfloat*)p1)[i], &((dfloat*)p3)[i], NEON_ALIGNMENT_SIZE)) { \
|
|
897
|
+
cnt = get_count_of_elements_not_aligned_to_simd_size( \
|
|
898
|
+
&((dfloat*)p1)[i], NEON_ALIGNMENT_SIZE, sizeof(dfloat) \
|
|
899
|
+
); \
|
|
900
|
+
if (p1 == p3) { \
|
|
901
|
+
for (; i < cnt; i++) { \
|
|
902
|
+
((dfloat*)p1)[i] = m_##fOpFunc(((dfloat*)p1)[i], *(dfloat*)p2); \
|
|
903
|
+
} \
|
|
904
|
+
} else { \
|
|
905
|
+
for (; i < cnt; i++) { \
|
|
906
|
+
((dfloat*)p3)[i] = m_##fOpFunc(((dfloat*)p1)[i], *(dfloat*)p2); \
|
|
907
|
+
} \
|
|
908
|
+
} \
|
|
909
|
+
cnt_simd_loop = (n - i) % num_pack; \
|
|
910
|
+
for (; i < n - cnt_simd_loop; i += num_pack) { \
|
|
911
|
+
a = vld1q_f64(&((dfloat*)p1)[i]); \
|
|
912
|
+
a = fSimdOp(a, b); \
|
|
913
|
+
vst1q_f64(&((dfloat*)p3)[i], a); \
|
|
914
|
+
} \
|
|
915
|
+
} \
|
|
916
|
+
if (cnt_simd_loop != 0) { \
|
|
917
|
+
if (p1 == p3) { \
|
|
918
|
+
for (; i < n; i++) { \
|
|
919
|
+
((dfloat*)p1)[i] = m_##fOpFunc(((dfloat*)p1)[i], *(dfloat*)p2); \
|
|
920
|
+
} \
|
|
921
|
+
} else { \
|
|
922
|
+
for (; i < n; i++) { \
|
|
923
|
+
((dfloat*)p3)[i] = m_##fOpFunc(((dfloat*)p1)[i], *(dfloat*)p2); \
|
|
924
|
+
} \
|
|
925
|
+
} \
|
|
926
|
+
} \
|
|
927
|
+
} else { \
|
|
928
|
+
for (i = 0; i < n; i++) { \
|
|
929
|
+
*(dfloat*)p3 = m_##fOpFunc(*(dfloat*)p1, *(dfloat*)p2); \
|
|
930
|
+
p1 += s1; \
|
|
931
|
+
p3 += s3; \
|
|
932
|
+
} \
|
|
933
|
+
} \
|
|
934
|
+
} else { \
|
|
935
|
+
if (p1 == p3) { \
|
|
936
|
+
for (i = 0; i < n; i++) { \
|
|
937
|
+
*(dfloat*)p1 = m_##fOpFunc(*(dfloat*)p1, *(dfloat*)p2); \
|
|
938
|
+
p1 += s1; \
|
|
939
|
+
p2 += s2; \
|
|
940
|
+
} \
|
|
941
|
+
} else { \
|
|
942
|
+
for (i = 0; i < n; i++) { \
|
|
943
|
+
*(dfloat*)p3 = m_##fOpFunc(*(dfloat*)p1, *(dfloat*)p2); \
|
|
944
|
+
p1 += s1; \
|
|
945
|
+
p2 += s2; \
|
|
946
|
+
p3 += s3; \
|
|
947
|
+
} \
|
|
948
|
+
} \
|
|
949
|
+
} \
|
|
950
|
+
return; \
|
|
951
|
+
} \
|
|
952
|
+
} \
|
|
953
|
+
\
|
|
954
|
+
for (i = 0; i < n; i++) { \
|
|
955
|
+
dfloat x; \
|
|
956
|
+
dfloat y; \
|
|
957
|
+
dfloat z; \
|
|
958
|
+
GET_DATA_STRIDE(p1, s1, dfloat, x); \
|
|
959
|
+
GET_DATA_STRIDE(p2, s2, dfloat, y); \
|
|
960
|
+
z = m_##fOpFunc(x, y); \
|
|
961
|
+
SET_DATA_STRIDE(p3, s3, dfloat, z); \
|
|
962
|
+
} \
|
|
963
|
+
}
|
|
964
|
+
|
|
423
965
|
#endif /* NUMO_NARRAY_MH_OP_BINARY_FUNC_H */
|
|
@@ -115,4 +115,24 @@
|
|
|
115
115
|
DEF_BINARY_SELF_FUNC(div, dfloat, numo_cDFloat) \
|
|
116
116
|
DEF_BINARY_FUNC(div, '/', dfloat, numo_cDFloat)
|
|
117
117
|
|
|
118
|
+
#define DEF_NARRAY_SFLT_DIV_AVX_METHOD_FUNC() \
|
|
119
|
+
DEF_BINARY_SFLT_AVX_ITER_FUNC(div, _mm256_div_ps) \
|
|
120
|
+
DEF_BINARY_SELF_FUNC(div, sfloat, numo_cSFloat) \
|
|
121
|
+
DEF_BINARY_FUNC(div, '/', sfloat, numo_cSFloat)
|
|
122
|
+
|
|
123
|
+
#define DEF_NARRAY_DFLT_DIV_AVX_METHOD_FUNC() \
|
|
124
|
+
DEF_BINARY_DFLT_AVX_ITER_FUNC(div, _mm256_div_pd) \
|
|
125
|
+
DEF_BINARY_SELF_FUNC(div, dfloat, numo_cDFloat) \
|
|
126
|
+
DEF_BINARY_FUNC(div, '/', dfloat, numo_cDFloat)
|
|
127
|
+
|
|
128
|
+
#define DEF_NARRAY_SFLT_DIV_NEON_METHOD_FUNC() \
|
|
129
|
+
DEF_BINARY_SFLT_NEON_ITER_FUNC(div, vdivq_f32) \
|
|
130
|
+
DEF_BINARY_SELF_FUNC(div, sfloat, numo_cSFloat) \
|
|
131
|
+
DEF_BINARY_FUNC(div, '/', sfloat, numo_cSFloat)
|
|
132
|
+
|
|
133
|
+
#define DEF_NARRAY_DFLT_DIV_NEON_METHOD_FUNC() \
|
|
134
|
+
DEF_BINARY_DFLT_NEON_ITER_FUNC(div, vdivq_f64) \
|
|
135
|
+
DEF_BINARY_SELF_FUNC(div, dfloat, numo_cDFloat) \
|
|
136
|
+
DEF_BINARY_FUNC(div, '/', dfloat, numo_cDFloat)
|
|
137
|
+
|
|
118
138
|
#endif /* NUMO_NARRAY_MH_OP_DIV_H */
|