1 /**************************************************************************
2 *
3 * Copyright 2008 Dennis Smit
4 * All Rights Reserved.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * on the rights to use, copy, modify, merge, publish, distribute, sub
10 * license, and/or sell copies of the Software, and to permit persons to whom
11 * the Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
15 * Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
20 * AUTHORS, COPYRIGHT HOLDERS, AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
21 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
22 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
23 * USE OR OTHER DEALINGS IN THE SOFTWARE.
24 *
25 ***************************************************************************/
26
27 /**
28 * @file
29 * CPU feature detection.
30 *
31 * @author Dennis Smit
32 * @author Based on the work of Eric Anholt <anholt@FreeBSD.org>
33 */
34
35 #ifndef _UTIL_CPU_DETECT_H
36 #define _UTIL_CPU_DETECT_H
37
38 #include <stdbool.h>
39
40 #include "pipe/p_config.h"
41 #include "util/u_atomic.h"
42 #include "util/u_thread.h"
43
44
45 #ifdef __cplusplus
46 extern "C" {
47 #endif
48
49 enum cpu_family {
50 CPU_UNKNOWN,
51
52 CPU_AMD_ZEN1_ZEN2,
53 CPU_AMD_ZEN_HYGON,
54 CPU_AMD_ZEN3,
55 CPU_AMD_ZEN_NEXT,
56 CPU_AMD_LAST,
57
58 CPU_S390X,
59 };
60
61 typedef uint32_t util_affinity_mask[UTIL_MAX_CPUS / 32];
62
63 struct util_cpu_caps_t {
64 /**
65 * Initialized to 0 and set to non-zero with an atomic after the entire
66 * struct has been initialized.
67 */
68 uint32_t detect_done;
69
70 /**
71 * Number of CPUs available to the process.
72 *
73 * This will be less than or equal to \c max_cpus. This is the number of
74 * CPUs that are online and available to the process.
75 */
76 int16_t nr_cpus;
77
78 /**
79 * Maximum number of CPUs that can be online in the system.
80 *
81 * This will be greater than or equal to \c nr_cpus. This is the number of
82 * CPUs installed in the system. \c nr_cpus will be less if some CPUs are
83 * offline.
84 */
85 int16_t max_cpus;
86
87 enum cpu_family family;
88
89 /* Feature flags */
90 int x86_cpu_type;
91 unsigned cacheline;
92
93 unsigned has_intel:1;
94 unsigned has_tsc:1;
95 unsigned has_mmx:1;
96 unsigned has_mmx2:1;
97 unsigned has_sse:1;
98 unsigned has_sse2:1;
99 unsigned has_sse3:1;
100 unsigned has_ssse3:1;
101 unsigned has_sse4_1:1;
102 unsigned has_sse4_2:1;
103 unsigned has_popcnt:1;
104 unsigned has_avx:1;
105 unsigned has_avx2:1;
106 unsigned has_f16c:1;
107 unsigned has_fma:1;
108 unsigned has_3dnow:1;
109 unsigned has_3dnow_ext:1;
110 unsigned has_xop:1;
111 unsigned has_altivec:1;
112 unsigned has_vsx:1;
113 unsigned has_daz:1;
114 unsigned has_neon:1;
115 unsigned has_msa:1;
116
117 unsigned has_avx512f:1;
118 unsigned has_avx512dq:1;
119 unsigned has_avx512ifma:1;
120 unsigned has_avx512pf:1;
121 unsigned has_avx512er:1;
122 unsigned has_avx512cd:1;
123 unsigned has_avx512bw:1;
124 unsigned has_avx512vl:1;
125 unsigned has_avx512vbmi:1;
126
127 unsigned num_L3_caches;
128 unsigned num_cpu_mask_bits;
129
130 uint16_t cpu_to_L3[UTIL_MAX_CPUS];
131 /* Affinity masks for each L3 cache. */
132 util_affinity_mask *L3_affinity_mask;
133 };
134
135 #define U_CPU_INVALID_L3 0xffff
136
137 void util_cpu_detect(void);
138
139 static inline ATTRIBUTE_CONST const struct util_cpu_caps_t *
util_get_cpu_caps(void)140 util_get_cpu_caps(void)
141 {
142 extern struct util_cpu_caps_t util_cpu_caps;
143
144 /* On most CPU architectures, an atomic read is simply a regular memory
145 * load instruction with some extra compiler magic to prevent code
146 * re-ordering around it. The perf impact of doing this check should be
147 * negligible in most cases.
148 *
149 * Also, even though it looks like a bit of a lie, we've declared this
150 * function with ATTRIBUTE_CONST. The GCC docs say:
151 *
152 * "Calls to functions whose return value is not affected by changes to
153 * the observable state of the program and that have no observable
154 * effects on such state other than to return a value may lend
155 * themselves to optimizations such as common subexpression elimination.
156 * Declaring such functions with the const attribute allows GCC to avoid
157 * emitting some calls in repeated invocations of the function with the
158 * same argument values."
159 *
160 * The word "observable" is important here. With the exception of a
161 * llvmpipe debug flag behind an environment variable and a few unit tests,
162 * all of which emulate worse CPUs, this function neither affects nor is
163 * affected by any "observable" state. It has its own internal state for
164 * sure, but that state is such that it appears to return exactly the same
165 * value with the same internal data every time.
166 */
167 if (unlikely(!p_atomic_read(&util_cpu_caps.detect_done)))
168 util_cpu_detect();
169
170 return &util_cpu_caps;
171 }
172
173 #ifdef __cplusplus
174 }
175 #endif
176
177
178 #endif /* _UTIL_CPU_DETECT_H */
179