1 #include <arm_neon.h>
2 #include "arm-neon-ref.h"
3 #include "compute-ref-data.h"
4
5 /* Expected results. */
6 VECT_VAR_DECL(expected,int,16,4) [] = { 0xffc0, 0xffc4, 0xffc8, 0xffcc };
7 VECT_VAR_DECL(expected,int,32,2) [] = { 0xfffffde0, 0xfffffe02 };
8 VECT_VAR_DECL(expected,uint,16,4) [] = { 0xbbc0, 0xc004, 0xc448, 0xc88c };
9 VECT_VAR_DECL(expected,uint,32,2) [] = { 0xfffface0, 0xffffb212 };
10 #if defined (__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
11 VECT_VAR_DECL(expected, hfloat, 16, 4) [] = { 0xddb3, 0xdd58, 0xdcfd, 0xdca1 };
12 #endif
13 VECT_VAR_DECL(expected,hfloat,32,2) [] = { 0xc3b66666, 0xc3ab0000 };
14 VECT_VAR_DECL(expected,int,16,8) [] = { 0xffc0, 0xffc4, 0xffc8, 0xffcc,
15 0xffd0, 0xffd4, 0xffd8, 0xffdc };
16 VECT_VAR_DECL(expected,int,32,4) [] = { 0xfffffde0, 0xfffffe02,
17 0xfffffe24, 0xfffffe46 };
18 VECT_VAR_DECL(expected,uint,16,8) [] = { 0xbbc0, 0xc004, 0xc448, 0xc88c,
19 0xccd0, 0xd114, 0xd558, 0xd99c };
20 VECT_VAR_DECL(expected,uint,32,4) [] = { 0xfffface0, 0xffffb212,
21 0xffffb744, 0xffffbc76 };
22 #if defined (__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
23 VECT_VAR_DECL(expected, hfloat, 16, 8) [] = { 0xddb3, 0xdd58, 0xdcfd, 0xdca1,
24 0xdc46, 0xdbd6, 0xdb20, 0xda69 };
25 #endif
26 VECT_VAR_DECL(expected,hfloat,32,4) [] = { 0xc3b66666, 0xc3ab0000,
27 0xc39f9999, 0xc3943333 };
28
29 #define TEST_MSG "VMUL_LANE"
30 void exec_vmul_lane (void)
31 {
32 #define DECL_VMUL(VAR) \
33 DECL_VARIABLE(VAR, int, 16, 4); \
34 DECL_VARIABLE(VAR, int, 32, 2); \
35 DECL_VARIABLE(VAR, uint, 16, 4); \
36 DECL_VARIABLE(VAR, uint, 32, 2); \
37 DECL_VARIABLE(VAR, float, 32, 2); \
38 DECL_VARIABLE(VAR, int, 16, 8); \
39 DECL_VARIABLE(VAR, int, 32, 4); \
40 DECL_VARIABLE(VAR, uint, 16, 8); \
41 DECL_VARIABLE(VAR, uint, 32, 4); \
42 DECL_VARIABLE(VAR, float, 32, 4)
43
44 /* vector_res = vmul_lane(vector,vector2,lane), then store the result. */
45 #define TEST_VMUL_LANE(Q, T1, T2, W, N, N2, L) \
46 VECT_VAR(vector_res, T1, W, N) = \
47 vmul##Q##_lane_##T2##W(VECT_VAR(vector, T1, W, N), \
48 VECT_VAR(vector2, T1, W, N2), \
49 L); \
50 vst1##Q##_##T2##W(VECT_VAR(result, T1, W, N), \
51 VECT_VAR(vector_res, T1, W, N))
52
53 DECL_VMUL(vector);
54 DECL_VMUL(vector_res);
55 #if defined (__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
56 DECL_VARIABLE(vector, float, 16, 4);
57 DECL_VARIABLE(vector, float, 16, 8);
58 DECL_VARIABLE(vector_res, float, 16, 4);
59 DECL_VARIABLE(vector_res, float, 16, 8);
60 #endif
61
62 DECL_VARIABLE(vector2, int, 16, 4);
63 DECL_VARIABLE(vector2, int, 32, 2);
64 DECL_VARIABLE(vector2, uint, 16, 4);
65 DECL_VARIABLE(vector2, uint, 32, 2);
66 #if defined (__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
67 DECL_VARIABLE(vector2, float, 16, 4);
68 #endif
69 DECL_VARIABLE(vector2, float, 32, 2);
70
71 clean_results ();
72
73 /* Initialize vector from pre-initialized values. */
74 VLOAD(vector, buffer, , int, s, 16, 4);
75 VLOAD(vector, buffer, , int, s, 32, 2);
76 VLOAD(vector, buffer, , uint, u, 16, 4);
77 VLOAD(vector, buffer, , uint, u, 32, 2);
78 #if defined (__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
79 VLOAD(vector, buffer, , float, f, 16, 4);
80 #endif
81 VLOAD(vector, buffer, , float, f, 32, 2);
82 VLOAD(vector, buffer, q, int, s, 16, 8);
83 VLOAD(vector, buffer, q, int, s, 32, 4);
84 VLOAD(vector, buffer, q, uint, u, 16, 8);
85 VLOAD(vector, buffer, q, uint, u, 32, 4);
86 #if defined (__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
87 VLOAD(vector, buffer, q, float, f, 16, 8);
88 #endif
89 VLOAD(vector, buffer, q, float, f, 32, 4);
90
91 /* Initialize vector2. */
92 VDUP(vector2, , int, s, 16, 4, 0x4);
93 VDUP(vector2, , int, s, 32, 2, 0x22);
94 VDUP(vector2, , uint, u, 16, 4, 0x444);
95 VDUP(vector2, , uint, u, 32, 2, 0x532);
96 #if defined (__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
97 VDUP(vector2, , float, f, 16, 4, 22.8f);
98 #endif
99 VDUP(vector2, , float, f, 32, 2, 22.8f);
100
101 /* Choose lane arbitrarily. */
102 TEST_VMUL_LANE(, int, s, 16, 4, 4, 2);
103 TEST_VMUL_LANE(, int, s, 32, 2, 2, 1);
104 TEST_VMUL_LANE(, uint, u, 16, 4, 4, 2);
105 TEST_VMUL_LANE(, uint, u, 32, 2, 2, 1);
106 #if defined (__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
107 TEST_VMUL_LANE(, float, f, 16, 4, 4, 1);
108 #endif
109 TEST_VMUL_LANE(, float, f, 32, 2, 2, 1);
110 TEST_VMUL_LANE(q, int, s, 16, 8, 4, 2);
111 TEST_VMUL_LANE(q, int, s, 32, 4, 2, 0);
112 TEST_VMUL_LANE(q, uint, u, 16, 8, 4, 2);
113 TEST_VMUL_LANE(q, uint, u, 32, 4, 2, 1);
114 #if defined (__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
115 TEST_VMUL_LANE(q, float, f, 16, 8, 4, 0);
116 #endif
117 TEST_VMUL_LANE(q, float, f, 32, 4, 2, 0);
118
119 CHECK(TEST_MSG, int, 16, 4, PRIx16, expected, "");
120 CHECK(TEST_MSG, int, 32, 2, PRIx32, expected, "");
121 CHECK(TEST_MSG, uint, 16, 4, PRIx16, expected, "");
122 CHECK(TEST_MSG, uint, 32, 2, PRIx32, expected, "");
123 #if defined (__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
124 CHECK_FP(TEST_MSG, float, 16, 4, PRIx16, expected, "");
125 #endif
126 CHECK_FP(TEST_MSG, float, 32, 2, PRIx32, expected, "");
127 CHECK(TEST_MSG, int, 16, 8, PRIx16, expected, "");
128 CHECK(TEST_MSG, int, 32, 4, PRIx32, expected, "");
129 CHECK(TEST_MSG, uint, 16, 8, PRIx16, expected, "");
130 CHECK(TEST_MSG, uint, 32, 4, PRIx32, expected, "");
131 #if defined (__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
132 CHECK_FP(TEST_MSG, float, 16, 8, PRIx16, expected, "");
133 #endif
134 CHECK_FP(TEST_MSG, float, 32, 4, PRIx32, expected, "");
135 }
136
137 int main (void)
138 {
139 exec_vmul_lane ();
140 return 0;
141 }