1 /* { dg-do run { target avx512fp16 } } */
2 /* { dg-options "-O2 -mavx512fp16 -mavx512vl" } */
3
4 static void do_test (void);
5
6 #define DO_TEST do_test
7 #define AVX512FP16
8
9 #include <immintrin.h>
10 #include "avx512-check.h"
11
12 __m256h a1 = { -39.3f16, -180.9f16, 13.4f16, 35.4f16, -41.1f16, -14.4f16, 24.5f16, 53.54f16,
13 238.4f16, -134.8f16, 24.5f16, 35.6f16, -346.7f16, -43.4f16, -535.3f16, 324.7f16 };
14 __m256h a2 = { 82.5f16, 21.4f16, 24.4f16, 53.4f16, 23.5f16, -24.4f16, -34.5f16, -32.5f16,
15 23.6f16, -13.4f16, 24.5f16, 35.5f16, -34.4f16, -24.5f16, -34.5f16, 13.5f16 };
16
17 __m128h b1 = { 1.25f16, 2.25f16, -0.25f16, 4.0f16, -2.0f16, 4.0f16, -3.0f16, 2.0f16 };
18 __m128h b2 = { -0.5f16, -1.0f16, 1.0f16, -1.0f16, 1.0f16, 1.0f16, 2.0f16, 4.0f16 };
19 __m128h b3 = { 1.25f16, 2.25f16, -4.25f16, 4.0f16, -2.4f16, 4.0f16, -3.0f, 2.0f16 };
20 __m128h b4 = { -4.5f16, 7.6f16, 0.7f16, -8.2f16, 2.1f16, 2.4f16, -2.0f16, 1.4f16 };
21
22 __attribute__((noinline, noclone)) _Float16
23 test_reduce_256_add_ph (__m256h a)
24 {
25 return _mm256_reduce_add_ph (a);
26 }
27
28 __attribute__((noinline, noclone)) _Float16
29 test_reduce_256_mul_ph (__m256h a)
30 {
31 return _mm256_reduce_mul_ph (a);
32 }
33
34 __attribute__((noinline, noclone)) _Float16
35 test_reduce_256_max_ph (__m256h a)
36 {
37 return _mm256_reduce_max_ph (a);
38 }
39
40 __attribute__((noinline, noclone)) _Float16
41 test_reduce_256_min_ph (__m256h a)
42 {
43 return _mm256_reduce_min_ph (a);
44 }
45
46 __attribute__((noinline, noclone)) _Float16
47 test_reduce_add_ph (__m128h b)
48 {
49 return _mm_reduce_add_ph (b);
50 }
51
52 __attribute__((noinline, noclone)) _Float16
53 test_reduce_mul_ph (__m128h b)
54 {
55 return _mm_reduce_mul_ph (b);
56 }
57
58 __attribute__((noinline, noclone)) _Float16
59 test_reduce_max_ph (__m128h b)
60 {
61 return _mm_reduce_max_ph (b);
62 }
63
64 __attribute__((noinline, noclone)) _Float16
65 test_reduce_min_ph (__m128h b)
66 {
67 return _mm_reduce_min_ph (b);
68 }
69
70 #define SIZE 16
71 #define REF_ADDMUL(op, a) \
72 __m128h __a1 = _mm_setzero_ph (); \
73 for (int i = 0; i < 8; i++) { \
74 __a1[i] = (_Float16) a[i] op (_Float16) a[i + 8]; \
75 } \
76 _Float16 __c0 = __a1[0] op __a1[4]; \
77 _Float16 __c1 = __a1[1] op __a1[5]; \
78 _Float16 __c2 = __a1[2] op __a1[6]; \
79 _Float16 __c3 = __a1[3] op __a1[7]; \
80 _Float16 __d0 = __c0 op __c2; \
81 _Float16 __d1 = __c1 op __c3; \
82 _Float16 __e0 = __d0 op __d1; \
83 r3 = __e0
84
85 #define TESTOP(opname, op, a) \
86 do { \
87 _Float16 r1 = _mm256_reduce_##opname##_ph (a); \
88 _Float16 r2 = test_reduce_256_##opname##_ph (a); \
89 _Float16 r3 = a[0]; \
90 if (r1 != r2) { \
91 __builtin_abort (); \
92 } \
93 REF_ADDMUL (op, a); \
94 if (r1 != r3) { \
95 __builtin_abort (); \
96 } \
97 } while (0)
98
99 #define TEST_ADDMUL_PH(a) \
100 do { \
101 TESTOP (add, +, a); \
102 TESTOP (mul, *, a); \
103 } while (0)
104
105 static void
106 test_256_addmul_ph (void)
107 {
108 TEST_ADDMUL_PH (a1);
109 TEST_ADDMUL_PH (a2);
110 }
111
112 #undef TESTOP
113 #define TESTOP(opname, op, a) \
114 do { \
115 _Float16 r1 = _mm256_reduce_##opname##_ph (a); \
116 _Float16 r2 = test_reduce_256_##opname##_ph (a); \
117 _Float16 r3 = a[0]; \
118 if (r1 != r2) { \
119 __builtin_abort (); \
120 } \
121 for (int i = 1; i < SIZE; i++) \
122 r3 = r3 op a[i]; \
123 if (r1 != r3) { \
124 __builtin_abort (); \
125 } \
126 } while (0)
127
128 #define TEST_MINMAX_PH(a) \
129 do { \
130 TESTOP (min, < a[i] ? r3 :, a); \
131 TESTOP (max, > a[i] ? r3 :, a); \
132 } while (0)
133
134 static void
135 test_256_minmax_ph (void)
136 {
137 TEST_MINMAX_PH (a1);
138 TEST_MINMAX_PH (a2);
139 }
140
141 static void
142 test_256_ph (void)
143 {
144 test_256_addmul_ph ();
145 test_256_minmax_ph ();
146 }
147
148 #undef SIZE
149 #define SIZE 8
150
151 #undef REF_ADDMUL
152 #define REF_ADDMUL(op, a) \
153 _Float16 __c0 = a[0] op a[4]; \
154 _Float16 __c1 = a[1] op a[5]; \
155 _Float16 __c2 = a[2] op a[6]; \
156 _Float16 __c3 = a[3] op a[7]; \
157 _Float16 __d0 = __c0 op __c2; \
158 _Float16 __d1 = __c1 op __c3; \
159 _Float16 __e0 = __d0 op __d1; \
160 r3 = __e0
161
162 #undef TESTOP
163 #define TESTOP(opname, op, a) \
164 do { \
165 _Float16 r1 = _mm_reduce_##opname##_ph (a); \
166 _Float16 r2 = test_reduce_##opname##_ph (a); \
167 _Float16 r3 = a[0]; \
168 if (r1 != r2) { \
169 __builtin_abort (); \
170 } \
171 REF_ADDMUL (op, a); \
172 if (r1 != r3) { \
173 __builtin_abort (); \
174 } \
175 } while (0)
176
177 #undef TEST_ADDMUL_PH
178 #define TEST_ADDMUL_PH(a) \
179 do { \
180 TESTOP (add, +, a); \
181 TESTOP (mul, *, a); \
182 } while (0)
183
184 static void
185 test_128_addmul_ph (void)
186 {
187 TEST_ADDMUL_PH (b1);
188 TEST_ADDMUL_PH (b2);
189 TEST_ADDMUL_PH (b3);
190 TEST_ADDMUL_PH (b4);
191 }
192
193 #undef TESTOP
194 #define TESTOP(opname, op, b) \
195 do { \
196 _Float16 r1 = _mm_reduce_##opname##_ph (b); \
197 _Float16 r2 = test_reduce_##opname##_ph (b); \
198 _Float16 r3 = b[0]; \
199 if (r1 != r2) { \
200 __builtin_abort (); \
201 } \
202 for (int i = 1; i < SIZE; i++) \
203 r3 = r3 op b[i]; \
204 if (r1 != r3) { \
205 __builtin_abort (); \
206 } \
207 } while (0)
208
209 #undef TEST_MINMAX_PH
210 #define TEST_MINMAX_PH(b) \
211 do { \
212 TESTOP (min, < b[i] ? r3 :, b); \
213 TESTOP (max, > b[i] ? r3 :, b); \
214 } while (0)
215
216 static void
217 test_128_minmax_ph (void)
218 {
219 TEST_MINMAX_PH (b1);
220 TEST_MINMAX_PH (b2);
221 TEST_MINMAX_PH (b3);
222 TEST_MINMAX_PH (b4);
223 }
224
225 static void
226 test_128_ph (void)
227 {
228 test_128_addmul_ph ();
229 test_128_minmax_ph ();
230 }
231
232 static void
233 do_test (void)
234 {
235 test_256_ph ();
236 test_128_ph ();
237 }
238
239
240 #undef SIZE
241 #undef REF_ADDMUL
242 #undef TESTOP
243 #undef TEST_ADDMUL_PH
244 #undef TEST_MINMAX_PH