1 /* { dg-do run } */
2 /* { dg-options "-O2 -mavx5124vnniw" } */
3 /* { dg-require-effective-target avx5124vnniw } */
4
5 #define DEFAULT_VALUE 0x7ffffffe
6
7 #define AVX5124VNNIW
8 #include "avx512f-helper.h"
9
10 #define SIZE (AVX512F_LEN / 32)
11
12 #include "avx512f-mask-type.h"
13
14 void
CALC(short * src1,short * src2,short * src3,short * src4,int * prev_dst,short * mult,int * dst)15 CALC (short *src1, short* src2, short *src3,
16 short *src4, int* prev_dst, short *mult, int *dst)
17 {
18 int i;
19
20 for (i = 0; i < SIZE; i++)
21 {
22 int p1dword, p2dword;
23 long long int tmp;
24 dst[i] = prev_dst[i];
25 p1dword = (int)(src1[2*i ]) * (int)(mult[0]);
26 p2dword = (int)(src1[2*i+1]) * (int)(mult[1]);
27 tmp = (long long)dst[i] + p1dword + p2dword;
28 if (tmp > 0x7fffffff)
29 dst[i] = 0x7fffffff;
30 else
31 dst[i] += p1dword + p2dword;
32
33 p1dword = (int)(src2[2*i ]) * (int)(mult[2]);
34 p2dword = (int)(src2[2*i+1]) * (int)(mult[3]);
35 tmp = (long long)dst[i] + p1dword + p2dword;
36 if (tmp > 0x7fffffff)
37 dst[i] = 0x7fffffff;
38 else
39 dst[i] += p1dword + p2dword;
40
41 p1dword = (int)(src3[2*i ]) * (int)(mult[4]);
42 p2dword = (int)(src3[2*i+1]) * (int)(mult[5]);
43 tmp = (long long)dst[i] + p1dword + p2dword;
44 if (tmp > 0x7fffffff)
45 dst[i] = 0x7fffffff;
46 else
47 dst[i] += p1dword + p2dword;
48
49 p1dword = (int)(src4[2*i ]) * (int)(mult[6]);
50 p2dword = (int)(src4[2*i+1]) * (int)(mult[7]);
51 tmp = (long long)dst[i] + p1dword + p2dword;
52 if (tmp > 0x7fffffff)
53 dst[i] = 0x7fffffff;
54 else
55 dst[i] += p1dword + p2dword;
56 }
57 }
58
59 void
TEST(void)60 TEST (void)
61 {
62 int i;
63 UNION_TYPE (AVX512F_LEN, i_w) src1, src2, src3, src4;
64 UNION_TYPE (AVX512F_LEN, i_d) src5, dst, res1, res2, res3;
65 UNION_TYPE (128, i_w) mult;
66 MASK_TYPE mask = MASK_VALUE;
67 int res_ref[SIZE];
68
69 for (i = 0; i < SIZE * 2; i++)
70 {
71 src1.a[i] = 2 + 7 * i % 291;
72 src2.a[i] = 3 + 11 * (i % 377) * i;
73 src3.a[i] = src1.a[i] * src1.a[i];
74 src4.a[i] = src2.a[i] * src2.a[i];
75 }
76 for (i = 0; i < 8; i++)
77 mult.a[i] = 3 + i * 2;
78
79 for (i = 0; i < SIZE; i++)
80 src5.a[i] = DEFAULT_VALUE;
81
82 CALC (src1.a, src2.a, src3.a, src4.a, src5.a, mult.a, res_ref);
83
84 res1.x = INTRINSIC (_4dpwssds_epi32) ( src5.x, src1.x, src2.x, src3.x, src4.x, &mult.x);
85 res2.x = INTRINSIC (_mask_4dpwssds_epi32) (src5.x, mask, src1.x, src2.x, src3.x, src4.x, &mult.x);
86 res3.x = INTRINSIC (_maskz_4dpwssds_epi32) (mask, src5.x, src1.x, src2.x, src3.x, src4.x, &mult.x);
87
88 if (UNION_CHECK (AVX512F_LEN, i_d) (res1, res_ref))
89 abort ();
90
91 MASK_MERGE (i_d) (res_ref, mask, SIZE);
92 if (UNION_CHECK (AVX512F_LEN, i_d) (res2, res_ref))
93 abort ();
94
95 MASK_ZERO (i_d) (res_ref, mask, SIZE);
96 if (UNION_CHECK (AVX512F_LEN, i_d) (res3, res_ref))
97 abort ();
98 }
99