Debian Codegraph @

gcc/testsuite/gcc.target/i386/avx10-minmax-helper.h

raw
1 #ifndef AVX10MINMAX_HELPERFUNC_INCLUDED
2 #define AVX10MINMAX_HELPERFUNC_INCLUDED
3
4 #include <math.h>
5 #include <limits.h>
6 #include <stdbool.h>
7 #include <float.h>
8 #include "avx512f-helper.h"
9 #define SNAN_float __builtin_nansf ("")
10 #define SNAN_flag_float 0x7fa00000
11 #define QNAN_float __builtin_nanf ("")
12 #define QNAN_flag_float 0x7fc00000
13 #define SNAN_double ((double)__builtin_nans (""))
14 #define SNAN_flag_double 0x7ff4000000000000
15 #define QNAN_double ((double)__builtin_nan (""))
16 #define QNAN_flag_double 0x7ff8000000000000
17 #define SNAN__Float16 ((_Float16)__builtin_nansf16 (""))
18 #define SNAN_flag__Float16 0x7d00
19 #define QNAN__Float16 ((_Float16)__builtin_nanf16 (""))
20 #define QNAN_flag__Float16 0x7e00
21 #define SNAN___bf16 ((__bf16)__builtin_nansf16b (""))
22 #define SNAN_flag___bf16 0x7fa0
23 #define QNAN___bf16 ((__bf16)__builtin_nanf (""))
24 #define QNAN_flag___bf16 0x7fc0
25 #define ISNAN(x) (x != x)
26 #define ABS_float(x) fabsf (x)
27 #define ABS_double(x) fabs (x)
28 #define ABS__Float16(x) __builtin_fabsf16 (x)
29 #define ABS___bf16(x) __builtin_fabsf (x)
30
31 #define Union_Data(typef, typei) \
32 typedef union \
33 { \
34 typef f; \
35 typei i; \
36 } union_##typef;
37
38 Union_Data(float, int)
39 Union_Data(double, long long)
40 Union_Data(__bf16, short)
41 Union_Data(_Float16, short)
42
43 #define IS_SNAN(union_x, type) ((union_x.i & SNAN_flag_##type) == union_snan.i)
44
45 #define IS_QNAN(union_x, type) ((union_x.i & QNAN_flag_##type) == union_qnan.i)
46
47 #define CHECK_EXP_MINMAX(UNION_TYPE, VALUE_TYPE, INT_TYPE) \
48 static int \
49 __attribute__((noinline, unused)) \
50 check_minmax_##UNION_TYPE (UNION_TYPE u, const VALUE_TYPE *v) \
51 { \
52 int i; \
53 int err = 0; \
54 for (i = 0; i < ARRAY_SIZE (u.a); i++) \
55 { \
56 union_##VALUE_TYPE union_x, union_y; \
57 union_x.f = u.a[i]; \
58 union_y.f = v[i]; \
59 if (union_x.i != union_y.i) \
60 { \
61 err++; \
62 PRINTF ("%i: " "%f" " != " "%f" "\n", \
63 i, v[i], u.a[i]); \
64 } \
65 } \
66 return err; \
67 }
68
69 #if defined (AVX10_512BIT)
70 CHECK_EXP_MINMAX (union512, float, int)
71 CHECK_EXP_MINMAX (union512d, double, long int)
72 CHECK_EXP_MINMAX (union512bf16_bf, __bf16, short int)
73 CHECK_EXP_MINMAX (union512h, _Float16, short int)
74 #endif
75 CHECK_EXP_MINMAX (union256, float, int)
76 CHECK_EXP_MINMAX (union256d, double, long int)
77 CHECK_EXP_MINMAX (union128, float, int)
78 CHECK_EXP_MINMAX (union128d, double, long int)
79 CHECK_EXP_MINMAX (union256bf16_bf, __bf16, short int)
80 CHECK_EXP_MINMAX (union128bf16_bf, __bf16, short int)
81 CHECK_EXP_MINMAX (union256h, _Float16, short int)
82 CHECK_EXP_MINMAX (union128h, _Float16, short int)
83
84 #define UNION_CHECK_MINMAX(SIZE, NAME) EVAL(check_minmax_union, SIZE, NAME)
85
86 #define CMP(res, x, y, type, value, op1, np, op2, zero, num, mag) \
87 { \
88 union_##type union_a, union_b; \
89 union_a.f = x; \
90 union_b.f = y; \
91 union_##type union_snan, union_qnan; \
92 union_snan.f = SNAN_##type; \
93 union_qnan.f = QNAN_##type; \
94 bool flag = false; \
95 if(num) \
96 { \
97 if(ISNAN(x) && ISNAN(y)) \
98 { \
99 if(IS_SNAN(union_a,type) || (IS_QNAN(union_a,type) && IS_QNAN(union_b,type))) \
100 { \
101 union_a.i |= value; \
102 res = union_a.f; \
103 flag = true; \
104 } \
105 else \
106 { \
107 union_b.i |= value; \
108 res = union_b.f; \
109 flag = true; \
110 } \
111 } \
112 else if(ISNAN(x)) \
113 { \
114 res = y; \
115 flag = true; \
116 } \
117 else if(ISNAN(y)) \
118 { \
119 res = x; \
120 flag = true; \
121 } \
122 } \
123 else \
124 { \
125 if(IS_SNAN(union_a,type) || (IS_QNAN(union_a,type) && !IS_SNAN(union_b,type))) \
126 { \
127 union_a.i |= value; \
128 res = union_a.f; \
129 flag = true; \
130 } \
131 else if(ISNAN(y)) \
132 { \
133 union_b.i |= value; \
134 res = union_b.f; \
135 flag = true; \
136 } \
137 } \
138 if(!flag) \
139 { \
140 if(!mag) \
141 { \
142 if((x == zero && y == - zero) || (x == - zero && y == zero)) \
143 res = np zero; \
144 else if(x op1 y) \
145 res = x; \
146 else \
147 res = y; \
148 } \
149 else \
150 { \
151 if(ABS_##type(x) op2 ABS_##type(y)) \
152 res = x; \
153 else if(ABS_##type(y) op2 ABS_##type(x)) \
154 res = y; \
155 else \
156 { \
157 if((x == zero && y == - zero) || (x == - zero && y == zero)) \
158 res = np zero; \
159 else if(x op1 y) \
160 res = x; \
161 else \
162 res = y; \
163 } \
164 } \
165 } \
166 }
167
168 #define MINMAX(type, value, zero) \
169 type \
170 minmax_##type (type * a, type * b, int imm) \
171 { \
172 int op_select = imm & 0x03; \
173 int sign_control = (imm & 0x0C) >> 2; \
174 int nan_prop_select = (imm & 0x10) >> 4; \
175 type tmp; \
176 if(nan_prop_select == 0) \
177 if(op_select == 0) \
178 CMP(tmp, *a, *b, type, value, <=, -, <, zero, false, false) \
179 else if(op_select == 1) \
180 CMP(tmp, *a, *b, type, value, >=, +, >, zero, false, false) \
181 else if(op_select == 2) \
182 CMP(tmp, *a, *b, type, value, <=, -, <, zero, false, true) \
183 else \
184 CMP(tmp, *a, *b, type, value, >=, +, >, zero, false, true) \
185 else \
186 if(op_select == 0) \
187 CMP(tmp, *a, *b, type, value, <=, -, <, zero, true, false) \
188 else if(op_select == 1) \
189 CMP(tmp, *a, *b, type, value, >=, +, >, zero, true, false) \
190 else if(op_select == 2) \
191 CMP(tmp, *a, *b, type, value, <=, -, <, zero, true, true) \
192 else \
193 CMP(tmp, *a, *b, type, value, >=, +, >, zero, true, true) \
194 if(!ISNAN(tmp)) \
195 if(sign_control == 0 && !ISNAN(*a)) \
196 if((tmp < 0 && *a > 0) || (tmp > 0 && *a < 0)) \
197 tmp = -tmp; \
198 else if(sign_control == 2) \
199 if(tmp < 0) tmp = -tmp; \
200 else if(sign_control == 3) \
201 if(tmp > 0) tmp = -tmp; \
202 return tmp; \
203 }
204
205
206 MINMAX(double, 0x7ff8000000000000, 0.0)
207 MINMAX(float, 0x7fc00000, 0.0f)
208 MINMAX(_Float16, 0x7e00, 0.0f16)
209 MINMAX(__bf16, 0x7fc0, 0.0bf16)
210
211 #define UNIT_TEST(R, InsnSuffix, MaskType, type) \
212 sign = -1; \
213 for (i = 0; i < SIZE; i++) \
214 { \
215 src1.a[i] = i % 2 ? SNAN_##type : 1.5 + 34.67 * i * sign; \
216 src2.a[i] = i % 3 ? QNAN_##type : -22.17 * i * sign; \
217 sign = sign * -1; \
218 } \
219 for (i = 0; i < SIZE; i++) \
220 res2.a[i] = DEFAULT_VALUE; \
221 res1.x = INTRINSIC(_minmax_##InsnSuffix) (src1.x, src2.x, R); \
222 res2.x = INTRINSIC(_mask_minmax_##InsnSuffix) (res2.x, mask, src1.x, src2.x, R); \
223 res3.x = INTRINSIC(_maskz_minmax_##InsnSuffix) (mask, src1.x, src2.x, R); \
224 CALC (res_ref, src1.a, src2.a, R); \
225 if (UNION_CHECK_MINMAX (AVX512F_LEN, MaskType) (res1, res_ref)) \
226 abort(); \
227 MASK_MERGE (MaskType) (res_ref, mask, SIZE); \
228 if (UNION_CHECK_MINMAX (AVX512F_LEN, MaskType) (res2, res_ref)) \
229 abort(); \
230 MASK_ZERO (MaskType) (res_ref, mask, SIZE); \
231 if (UNION_CHECK_MINMAX (AVX512F_LEN, MaskType) (res3, res_ref)) \
232 abort();
233
234 #define SCALAR_UNIT_TEST(R, InsnSuffix, MaskType, type) \
235 sign = -1; \
236 for (i = 0; i < SIZE; i++) \
237 { \
238 src1.a[i] = i % 2 ? SNAN_##type : 1.5 + 34.67 * i * sign; \
239 src2.a[i] = i % 3 ? QNAN_##type : -22.17 * i * sign; \
240 sign = sign * -1; \
241 } \
242 for (i = 0; i < SIZE; i++) \
243 res2.a[i] = DEFAULT_VALUE; \
244 res1.x = _mm_minmax_##InsnSuffix (src1.x, src2.x, R); \
245 res2.x = _mm_mask_minmax_##InsnSuffix (res2.x, mask, src1.x, src2.x, R); \
246 res3.x = _mm_maskz_minmax_##InsnSuffix (mask, src1.x, src2.x, R); \
247 CALC (res_ref, src1.a, src2.a, R); \
248 if (UNION_CHECK_MINMAX (128, MaskType) (res1, res_ref)) \
249 abort(); \
250 MASK_MERGE (MaskType) (res_ref, mask, 1); \
251 if (UNION_CHECK_MINMAX (128, MaskType) (res2, res_ref)) \
252 abort(); \
253 MASK_ZERO (MaskType) (res_ref, mask, 1); \
254 if (UNION_CHECK_MINMAX (128, MaskType) (res3, res_ref)) \
255 abort();
256
257 #endif