2 /* { dg-options "-O2 -mavx512f" } */
3 /* { dg-require-effective-target avx512f } */
5 #include "avx512f-check.h"
7 #include "avx512f-helper.h"
9 #define SIZE (128 / 32)
10 #include "avx512f-mask-type.h"
13 calc_max (float *r
, float *s1
, float *s2
)
15 r
[0] = s1
[0] > s2
[0] ? s1
[0] : s2
[0];
17 for (i
= 1; i
< SIZE
; i
++)
27 union128 res1
, res2
, res3
, res4
, res5
, src1
, src2
;
28 MASK_TYPE mask
= MASK_VALUE
;
32 for (i
= 0; i
< SIZE
; i
++)
34 src1
.a
[i
] = 1.5 + 34.67 * i
* sign
;
35 src2
.a
[i
] = -22.17 * i
* sign
+ 1.0;
36 res1
.a
[i
] = DEFAULT_VALUE
;
37 res4
.a
[i
] = DEFAULT_VALUE
;
41 res1
.x
= _mm_mask_max_ss (res1
.x
, mask
, src1
.x
, src2
.x
);
42 res2
.x
= _mm_maskz_max_ss (mask
, src1
.x
, src2
.x
);
43 res3
.x
= _mm_max_round_ss (src1
.x
, src2
.x
, _MM_FROUND_NO_EXC
);
44 res4
.x
= _mm_mask_max_round_ss (res4
.x
, mask
, src1
.x
, src2
.x
, _MM_FROUND_NO_EXC
);
45 res5
.x
= _mm_maskz_max_round_ss (mask
, src1
.x
, src2
.x
, _MM_FROUND_NO_EXC
);
47 calc_max (res_ref
, src1
.a
, src2
.a
);
49 MASK_MERGE () (res_ref
, mask
, 1);
50 if (check_union128 (res1
, res_ref
))
53 MASK_ZERO () (res_ref
, mask
, 1);
54 if (check_union128 (res2
, res_ref
))
57 calc_max (res_ref
, src1
.a
, src2
.a
);
59 if (check_union128 (res3
, res_ref
))
62 MASK_MERGE () (res_ref
, mask
, 1);
63 if (check_union128 (res4
, res_ref
))
66 MASK_ZERO () (res_ref
, mask
, 1);
67 if (check_union128 (res5
, res_ref
))