1 /* { dg-do run } */
2 /* { dg-options "-O2 -mavx512f" } */
3 /* { dg-require-effective-target avx512f } */
4
5 #define AVX512F
6
7 #include "avx512f-helper.h"
8
9 #define SIZE (AVX512F_LEN / 32)
10 #include "avx512f-mask-type.h"
11 #include "math.h"
12
13 static void
CALC(float * dst,float * src1,int * ind,float * src2)14 CALC (float *dst, float *src1, int *ind, float *src2)
15 {
16 int i;
17
18 for (i = 0; i < SIZE; i++)
19 {
20 unsigned long long offset = ind[i] & (SIZE - 1);
21 unsigned long long cond = ind[i] & SIZE;
22
23 dst[i] = cond ? src2[offset] : src1[offset];
24 }
25 }
26
27 void
TEST(void)28 TEST (void)
29 {
30 int i, k;
31 UNION_TYPE (AVX512F_LEN,) s1, s2, res;
32 UNION_TYPE (AVX512F_LEN, i_d) ind;
33 float res_ref[SIZE];
34
35 union
36 {
37 float f;
38 int i;
39 } ind_copy[SIZE];
40
41 MASK_TYPE mask = MASK_VALUE;
42
43 for (i = 0; i < SIZE; i++)
44 {
45 /* Some of the integer indexes may be interpreted as floating point
46 values in mask-merge mode, that's why we use IND_COPY. */
47 ind.a[i] = ind_copy[i].i = 17 * (i << 1);
48 s1.a[i] = 42.5 * i + 1;
49 s2.a[i] = 22.5 * i;
50
51 res.a[i] = DEFAULT_VALUE;
52 }
53
54 CALC (res_ref, s1.a, ind.a, s2.a);
55
56 res.x = INTRINSIC (_mask2_permutex2var_ps) (s1.x, ind.x, mask, s2.x);
57
58 /* Standard MASK_MERGE cannot be used since VPERMI2PS in mask-merge mode
59 merges vectors of two different types (_m512 and __m512i). */
60 for (k = 0; k < SIZE; k++)
61 res_ref[k] = (mask & (1LL << k)) ? res_ref[k] : ind_copy[k].f;
62
63 if (UNION_CHECK (AVX512F_LEN,) (res, res_ref))
64 abort ();
65 }
66