1 #include <arm_neon.h>
2 #include "arm-neon-ref.h"
3 #include "compute-ref-data.h"
4
5 /* Expected results for vrev16. */
6 VECT_VAR_DECL(expected_vrev16,int,8,8) [] = { 0xf1, 0xf0, 0xf3, 0xf2,
7 0xf5, 0xf4, 0xf7, 0xf6 };
8 VECT_VAR_DECL(expected_vrev16,uint,8,8) [] = { 0xf1, 0xf0, 0xf3, 0xf2,
9 0xf5, 0xf4, 0xf7, 0xf6 };
10 VECT_VAR_DECL(expected_vrev16,poly,8,8) [] = { 0xf1, 0xf0, 0xf3, 0xf2,
11 0xf5, 0xf4, 0xf7, 0xf6 };
12 VECT_VAR_DECL(expected_vrev16,int,8,16) [] = { 0xf1, 0xf0, 0xf3, 0xf2,
13 0xf5, 0xf4, 0xf7, 0xf6,
14 0xf9, 0xf8, 0xfb, 0xfa,
15 0xfd, 0xfc, 0xff, 0xfe };
16 VECT_VAR_DECL(expected_vrev16,uint,8,16) [] = { 0xf1, 0xf0, 0xf3, 0xf2,
17 0xf5, 0xf4, 0xf7, 0xf6,
18 0xf9, 0xf8, 0xfb, 0xfa,
19 0xfd, 0xfc, 0xff, 0xfe };
20 VECT_VAR_DECL(expected_vrev16,poly,8,16) [] = { 0xf1, 0xf0, 0xf3, 0xf2,
21 0xf5, 0xf4, 0xf7, 0xf6,
22 0xf9, 0xf8, 0xfb, 0xfa,
23 0xfd, 0xfc, 0xff, 0xfe };
24
25 /* Expected results for vrev32. */
26 VECT_VAR_DECL(expected_vrev32,int,8,8) [] = { 0xf3, 0xf2, 0xf1, 0xf0,
27 0xf7, 0xf6, 0xf5, 0xf4 };
28 VECT_VAR_DECL(expected_vrev32,int,16,4) [] = { 0xfff1, 0xfff0, 0xfff3, 0xfff2 };
29 VECT_VAR_DECL(expected_vrev32,uint,8,8) [] = { 0xf3, 0xf2, 0xf1, 0xf0,
30 0xf7, 0xf6, 0xf5, 0xf4 };
31 VECT_VAR_DECL(expected_vrev32,uint,16,4) [] = { 0xfff1, 0xfff0, 0xfff3, 0xfff2 };
32 VECT_VAR_DECL(expected_vrev32,poly,8,8) [] = { 0xf3, 0xf2, 0xf1, 0xf0,
33 0xf7, 0xf6, 0xf5, 0xf4 };
34 VECT_VAR_DECL(expected_vrev32,poly,16,4) [] = { 0xfff1, 0xfff0, 0xfff3, 0xfff2 };
35 VECT_VAR_DECL(expected_vrev32,int,8,16) [] = { 0xf3, 0xf2, 0xf1, 0xf0,
36 0xf7, 0xf6, 0xf5, 0xf4,
37 0xfb, 0xfa, 0xf9, 0xf8,
38 0xff, 0xfe, 0xfd, 0xfc };
39 VECT_VAR_DECL(expected_vrev32,int,16,8) [] = { 0xfff1, 0xfff0, 0xfff3, 0xfff2,
40 0xfff5, 0xfff4, 0xfff7, 0xfff6 };
41 VECT_VAR_DECL(expected_vrev32,uint,8,16) [] = { 0xf3, 0xf2, 0xf1, 0xf0,
42 0xf7, 0xf6, 0xf5, 0xf4,
43 0xfb, 0xfa, 0xf9, 0xf8,
44 0xff, 0xfe, 0xfd, 0xfc };
45 VECT_VAR_DECL(expected_vrev32,uint,16,8) [] = { 0xfff1, 0xfff0, 0xfff3, 0xfff2,
46 0xfff5, 0xfff4, 0xfff7, 0xfff6 };
47 VECT_VAR_DECL(expected_vrev32,poly,8,16) [] = { 0xf3, 0xf2, 0xf1, 0xf0,
48 0xf7, 0xf6, 0xf5, 0xf4,
49 0xfb, 0xfa, 0xf9, 0xf8,
50 0xff, 0xfe, 0xfd, 0xfc };
51 VECT_VAR_DECL(expected_vrev32,poly,16,8) [] = { 0xfff1, 0xfff0, 0xfff3, 0xfff2,
52 0xfff5, 0xfff4, 0xfff7, 0xfff6 };
53
54 /* Expected results for vrev64. */
55 VECT_VAR_DECL(expected_vrev64,int,8,8) [] = { 0xf7, 0xf6, 0xf5, 0xf4,
56 0xf3, 0xf2, 0xf1, 0xf0 };
57 VECT_VAR_DECL(expected_vrev64,int,16,4) [] = { 0xfff3, 0xfff2, 0xfff1, 0xfff0 };
58 VECT_VAR_DECL(expected_vrev64,int,32,2) [] = { 0xfffffff1, 0xfffffff0 };
59 VECT_VAR_DECL(expected_vrev64,uint,8,8) [] = { 0xf7, 0xf6, 0xf5, 0xf4, 0xf3,
60 0xf2, 0xf1, 0xf0 };
61 VECT_VAR_DECL(expected_vrev64,uint,16,4) [] = { 0xfff3, 0xfff2, 0xfff1, 0xfff0 };
62 VECT_VAR_DECL(expected_vrev64,uint,32,2) [] = { 0xfffffff1, 0xfffffff0 };
63 VECT_VAR_DECL(expected_vrev64,poly,8,8) [] = { 0xf7, 0xf6, 0xf5, 0xf4,
64 0xf3, 0xf2, 0xf1, 0xf0 };
65 VECT_VAR_DECL(expected_vrev64,poly,16,4) [] = { 0xfff3, 0xfff2, 0xfff1, 0xfff0 };
66 #if defined (FP16_SUPPORTED)
67 VECT_VAR_DECL (expected_vrev64, hfloat, 16, 4) [] = { 0xca80, 0xcb00,
68 0xcb80, 0xcc00 };
69 #endif
70 VECT_VAR_DECL(expected_vrev64,hfloat,32,2) [] = { 0xc1700000, 0xc1800000 };
71 VECT_VAR_DECL(expected_vrev64,int,8,16) [] = { 0xf7, 0xf6, 0xf5, 0xf4,
72 0xf3, 0xf2, 0xf1, 0xf0,
73 0xff, 0xfe, 0xfd, 0xfc,
74 0xfb, 0xfa, 0xf9, 0xf8 };
75 VECT_VAR_DECL(expected_vrev64,int,16,8) [] = { 0xfff3, 0xfff2, 0xfff1, 0xfff0,
76 0xfff7, 0xfff6, 0xfff5, 0xfff4 };
77 VECT_VAR_DECL(expected_vrev64,int,32,4) [] = { 0xfffffff1, 0xfffffff0,
78 0xfffffff3, 0xfffffff2 };
79 VECT_VAR_DECL(expected_vrev64,uint,8,16) [] = { 0xf7, 0xf6, 0xf5, 0xf4,
80 0xf3, 0xf2, 0xf1, 0xf0,
81 0xff, 0xfe, 0xfd, 0xfc,
82 0xfb, 0xfa, 0xf9, 0xf8 };
83 VECT_VAR_DECL(expected_vrev64,uint,16,8) [] = { 0xfff3, 0xfff2, 0xfff1, 0xfff0,
84 0xfff7, 0xfff6, 0xfff5, 0xfff4 };
85 VECT_VAR_DECL(expected_vrev64,uint,32,4) [] = { 0xfffffff1, 0xfffffff0,
86 0xfffffff3, 0xfffffff2 };
87 VECT_VAR_DECL(expected_vrev64,poly,8,16) [] = { 0xf7, 0xf6, 0xf5, 0xf4,
88 0xf3, 0xf2, 0xf1, 0xf0,
89 0xff, 0xfe, 0xfd, 0xfc,
90 0xfb, 0xfa, 0xf9, 0xf8 };
91 VECT_VAR_DECL(expected_vrev64,poly,16,8) [] = { 0xfff3, 0xfff2, 0xfff1, 0xfff0,
92 0xfff7, 0xfff6, 0xfff5, 0xfff4 };
93 #if defined (FP16_SUPPORTED)
94 VECT_VAR_DECL (expected_vrev64, hfloat, 16, 8) [] = { 0xca80, 0xcb00,
95 0xcb80, 0xcc00,
96 0xc880, 0xc900,
97 0xc980, 0xca00 };
98 #endif
99 VECT_VAR_DECL(expected_vrev64,hfloat,32,4) [] = { 0xc1700000, 0xc1800000,
100 0xc1500000, 0xc1600000 };
101
102 void exec_vrev (void)
103 {
104 /* Basic test: y=vrev(x), then store the result. */
105 #define TEST_VREV(Q, T1, T2, W, N, W2) \
106 VECT_VAR(vector_res, T1, W, N) = \
107 vrev##W2##Q##_##T2##W(VECT_VAR(vector, T1, W, N)); \
108 vst1##Q##_##T2##W(VECT_VAR(result, T1, W, N), VECT_VAR(vector_res, T1, W, N))
109
110 DECL_VARIABLE_ALL_VARIANTS(vector);
111 DECL_VARIABLE_ALL_VARIANTS(vector_res);
112
113 clean_results ();
114
115 /* Initialize input "vector" from "buffer". */
116 TEST_MACRO_ALL_VARIANTS_2_5(VLOAD, vector, buffer);
117 #if defined (FP16_SUPPORTED)
118 VLOAD (vector, buffer, , float, f, 16, 4);
119 VLOAD (vector, buffer, q, float, f, 16, 8);
120 #endif
121 VLOAD(vector, buffer, , float, f, 32, 2);
122 VLOAD(vector, buffer, q, float, f, 32, 4);
123
124 /* Check vrev in each of the existing combinations. */
125 #define TEST_MSG "VREV16"
126 TEST_VREV(, int, s, 8, 8, 16);
127 TEST_VREV(, uint, u, 8, 8, 16);
128 TEST_VREV(, poly, p, 8, 8, 16);
129 TEST_VREV(q, int, s, 8, 16, 16);
130 TEST_VREV(q, uint, u, 8, 16, 16);
131 TEST_VREV(q, poly, p, 8, 16, 16);
132
133 CHECK(TEST_MSG, int, 8, 8, PRIx8, expected_vrev16, "");
134 CHECK(TEST_MSG, uint, 8, 8, PRIx8, expected_vrev16, "");
135 CHECK_POLY(TEST_MSG, poly, 8, 8, PRIx8, expected_vrev16, "");
136 CHECK(TEST_MSG, int, 8, 16, PRIx8, expected_vrev16, "");
137 CHECK(TEST_MSG, uint, 8, 16, PRIx8, expected_vrev16, "");
138 CHECK_POLY(TEST_MSG, poly, 8, 16, PRIx8, expected_vrev16, "");
139
140 #undef TEST_MSG
141 #define TEST_MSG "VREV32"
142 TEST_VREV(, int, s, 8, 8, 32);
143 TEST_VREV(, int, s, 16, 4, 32);
144 TEST_VREV(, uint, u, 8, 8, 32);
145 TEST_VREV(, uint, u, 16, 4, 32);
146 TEST_VREV(, poly, p, 8, 8, 32);
147 TEST_VREV(, poly, p, 16, 4, 32);
148 TEST_VREV(q, int, s, 8, 16, 32);
149 TEST_VREV(q, int, s, 16, 8, 32);
150 TEST_VREV(q, uint, u, 8, 16, 32);
151 TEST_VREV(q, uint, u, 16, 8, 32);
152 TEST_VREV(q, poly, p, 8, 16, 32);
153 TEST_VREV(q, poly, p, 16, 8, 32);
154
155 CHECK(TEST_MSG, int, 8, 8, PRIx8, expected_vrev32, "");
156 CHECK(TEST_MSG, int, 16, 4, PRIx16, expected_vrev32, "");
157 CHECK(TEST_MSG, uint, 8, 8, PRIx8, expected_vrev32, "");
158 CHECK(TEST_MSG, uint, 16, 4, PRIx16, expected_vrev32, "");
159 CHECK_POLY(TEST_MSG, poly, 8, 8, PRIx8, expected_vrev32, "");
160 CHECK_POLY(TEST_MSG, poly, 16, 4, PRIx16, expected_vrev32, "");
161 CHECK(TEST_MSG, int, 8, 16, PRIx8, expected_vrev32, "");
162 CHECK(TEST_MSG, int, 16, 8, PRIx16, expected_vrev32, "");
163 CHECK(TEST_MSG, uint, 8, 16, PRIx8, expected_vrev32, "");
164 CHECK(TEST_MSG, uint, 16, 8, PRIx16, expected_vrev32, "");
165 CHECK_POLY(TEST_MSG, poly, 8, 16, PRIx8, expected_vrev32, "");
166 CHECK_POLY(TEST_MSG, poly, 16, 8, PRIx16, expected_vrev32, "");
167
168 #undef TEST_MSG
169 #define TEST_MSG "VREV64"
170 TEST_VREV(, int, s, 8, 8, 64);
171 TEST_VREV(, int, s, 16, 4, 64);
172 TEST_VREV(, int, s, 32, 2, 64);
173 TEST_VREV(, uint, u, 8, 8, 64);
174 TEST_VREV(, uint, u, 16, 4, 64);
175 TEST_VREV(, uint, u, 32, 2, 64);
176 TEST_VREV(, poly, p, 8, 8, 64);
177 TEST_VREV(, poly, p, 16, 4, 64);
178 TEST_VREV(q, int, s, 8, 16, 64);
179 TEST_VREV(q, int, s, 16, 8, 64);
180 TEST_VREV(q, int, s, 32, 4, 64);
181 TEST_VREV(q, uint, u, 8, 16, 64);
182 TEST_VREV(q, uint, u, 16, 8, 64);
183 TEST_VREV(q, uint, u, 32, 4, 64);
184 TEST_VREV(q, poly, p, 8, 16, 64);
185 TEST_VREV(q, poly, p, 16, 8, 64);
186
187 CHECK(TEST_MSG, int, 8, 8, PRIx8, expected_vrev64, "");
188 CHECK(TEST_MSG, int, 16, 4, PRIx16, expected_vrev64, "");
189 CHECK(TEST_MSG, int, 32, 2, PRIx32, expected_vrev64, "");
190 CHECK(TEST_MSG, uint, 8, 8, PRIx8, expected_vrev64, "");
191 CHECK(TEST_MSG, uint, 16, 4, PRIx16, expected_vrev64, "");
192 CHECK(TEST_MSG, uint, 32, 2, PRIx32, expected_vrev64, "");
193 CHECK_POLY(TEST_MSG, poly, 8, 8, PRIx8, expected_vrev64, "");
194 CHECK_POLY(TEST_MSG, poly, 16, 4, PRIx16, expected_vrev64, "");
195 CHECK(TEST_MSG, int, 8, 16, PRIx8, expected_vrev64, "");
196 CHECK(TEST_MSG, int, 16, 8, PRIx16, expected_vrev64, "");
197 CHECK(TEST_MSG, int, 32, 4, PRIx32, expected_vrev64, "");
198 CHECK(TEST_MSG, uint, 8, 16, PRIx8, expected_vrev64, "");
199 CHECK(TEST_MSG, uint, 16, 8, PRIx16, expected_vrev64, "");
200 CHECK(TEST_MSG, uint, 32, 4, PRIx32, expected_vrev64, "");
201 CHECK_POLY(TEST_MSG, poly, 8, 16, PRIx8, expected_vrev64, "");
202 CHECK_POLY(TEST_MSG, poly, 16, 8, PRIx16, expected_vrev64, "");
203
204 #if defined (FP16_SUPPORTED)
205 TEST_VREV (, float, f, 16, 4, 64);
206 TEST_VREV (q, float, f, 16, 8, 64);
207 CHECK_FP(TEST_MSG, float, 16, 4, PRIx32, expected_vrev64, "");
208 CHECK_FP(TEST_MSG, float, 16, 8, PRIx32, expected_vrev64, "");
209 #endif
210 TEST_VREV(, float, f, 32, 2, 64);
211 TEST_VREV(q, float, f, 32, 4, 64);
212 CHECK_FP(TEST_MSG, float, 32, 2, PRIx32, expected_vrev64, "");
213 CHECK_FP(TEST_MSG, float, 32, 4, PRIx32, expected_vrev64, "");
214 }
215
216 int main (void)
217 {
218 exec_vrev ();
219 return 0;
220 }