1
2
3
4
5
6
7 package cpu
8
9 const CacheLinePadSize = 64
10
11
12 func cpuid(eaxArg, ecxArg uint32) (eax, ebx, ecx, edx uint32)
13
14
15 func xgetbv() (eax, edx uint32)
16
17
18 func getGOAMD64level() int32
19
20 const (
21
22 cpuid_AVXVNNI = 1 << 4
23
24
25 cpuid_SSE3 = 1 << 0
26 cpuid_PCLMULQDQ = 1 << 1
27 cpuid_AVX512VBMI = 1 << 1
28 cpuid_AVX512VBMI2 = 1 << 6
29 cpuid_SSSE3 = 1 << 9
30 cpuid_AVX512GFNI = 1 << 8
31 cpuid_AVX512VAES = 1 << 9
32 cpuid_AVX512VNNI = 1 << 11
33 cpuid_AVX512BITALG = 1 << 12
34 cpuid_FMA = 1 << 12
35 cpuid_AVX512VPOPCNTDQ = 1 << 14
36 cpuid_SSE41 = 1 << 19
37 cpuid_SSE42 = 1 << 20
38 cpuid_POPCNT = 1 << 23
39 cpuid_AES = 1 << 25
40 cpuid_OSXSAVE = 1 << 27
41 cpuid_AVX = 1 << 28
42
43
44 cpuid_BMI1 = 1 << 3
45 cpuid_AVX2 = 1 << 5
46 cpuid_BMI2 = 1 << 8
47 cpuid_ERMS = 1 << 9
48 cpuid_AVX512F = 1 << 16
49 cpuid_AVX512DQ = 1 << 17
50 cpuid_ADX = 1 << 19
51 cpuid_AVX512CD = 1 << 28
52 cpuid_SHA = 1 << 29
53 cpuid_AVX512BW = 1 << 30
54 cpuid_AVX512VL = 1 << 31
55
56
57 cpuid_AVX512_VBMI = 1 << 1
58 cpuid_AVX512_VBMI2 = 1 << 6
59 cpuid_GFNI = 1 << 8
60 cpuid_AVX512VPCLMULQDQ = 1 << 10
61 cpuid_AVX512_BITALG = 1 << 12
62
63
64 cpuid_FSRM = 1 << 4
65
66 cpuid_RDTSCP = 1 << 27
67 )
68
69 var maxExtendedFunctionInformation uint32
70
71 func doinit() {
72 options = []option{
73 {Name: "adx", Feature: &X86.HasADX},
74 {Name: "aes", Feature: &X86.HasAES},
75 {Name: "erms", Feature: &X86.HasERMS},
76 {Name: "fsrm", Feature: &X86.HasFSRM},
77 {Name: "pclmulqdq", Feature: &X86.HasPCLMULQDQ},
78 {Name: "rdtscp", Feature: &X86.HasRDTSCP},
79 {Name: "sha", Feature: &X86.HasSHA},
80 {Name: "vpclmulqdq", Feature: &X86.HasAVX512VPCLMULQDQ},
81 }
82 level := getGOAMD64level()
83 if level < 2 {
84
85
86 options = append(options,
87 option{Name: "popcnt", Feature: &X86.HasPOPCNT},
88 option{Name: "sse3", Feature: &X86.HasSSE3},
89 option{Name: "sse41", Feature: &X86.HasSSE41},
90 option{Name: "sse42", Feature: &X86.HasSSE42},
91 option{Name: "ssse3", Feature: &X86.HasSSSE3})
92 }
93 if level < 3 {
94
95
96 options = append(options,
97 option{Name: "avx", Feature: &X86.HasAVX},
98 option{Name: "avx2", Feature: &X86.HasAVX2},
99 option{Name: "bmi1", Feature: &X86.HasBMI1},
100 option{Name: "bmi2", Feature: &X86.HasBMI2},
101 option{Name: "fma", Feature: &X86.HasFMA})
102 }
103 if level < 4 {
104
105
106 options = append(options,
107 option{Name: "avx512f", Feature: &X86.HasAVX512F},
108 option{Name: "avx512cd", Feature: &X86.HasAVX512CD},
109 option{Name: "avx512bw", Feature: &X86.HasAVX512BW},
110 option{Name: "avx512dq", Feature: &X86.HasAVX512DQ},
111 option{Name: "avx512vl", Feature: &X86.HasAVX512VL},
112 )
113 }
114
115 maxID, _, _, _ := cpuid(0, 0)
116
117 if maxID < 1 {
118 osInit()
119 return
120 }
121
122 maxExtendedFunctionInformation, _, _, _ = cpuid(0x80000000, 0)
123
124 _, _, ecx1, _ := cpuid(1, 0)
125
126 X86.HasSSE3 = isSet(ecx1, cpuid_SSE3)
127 X86.HasPCLMULQDQ = isSet(ecx1, cpuid_PCLMULQDQ)
128 X86.HasSSSE3 = isSet(ecx1, cpuid_SSSE3)
129 X86.HasSSE41 = isSet(ecx1, cpuid_SSE41)
130 X86.HasSSE42 = isSet(ecx1, cpuid_SSE42)
131 X86.HasPOPCNT = isSet(ecx1, cpuid_POPCNT)
132 X86.HasAES = isSet(ecx1, cpuid_AES)
133
134
135
136
137 X86.HasOSXSAVE = isSet(ecx1, cpuid_OSXSAVE)
138
139
140
141
142
143 X86.HasFMA = isSet(ecx1, cpuid_FMA) && X86.HasOSXSAVE
144
145 osSupportsAVX := false
146 osSupportsAVX512 := false
147
148 if X86.HasOSXSAVE {
149 eax, _ := xgetbv()
150
151 osSupportsAVX = isSet(eax, 1<<1) && isSet(eax, 1<<2)
152
153
154
155
156
157 osSupportsAVX512 = osSupportsAVX && isSet(eax, 1<<5) && isSet(eax, 1<<6) && isSet(eax, 1<<7)
158 }
159
160 X86.HasAVX = isSet(ecx1, cpuid_AVX) && osSupportsAVX
161
162 if maxID < 7 {
163 osInit()
164 return
165 }
166
167 eax7, ebx7, ecx7, edx7 := cpuid(7, 0)
168 X86.HasBMI1 = isSet(ebx7, cpuid_BMI1)
169 X86.HasAVX2 = isSet(ebx7, cpuid_AVX2) && osSupportsAVX
170 X86.HasBMI2 = isSet(ebx7, cpuid_BMI2)
171 X86.HasERMS = isSet(ebx7, cpuid_ERMS)
172 X86.HasADX = isSet(ebx7, cpuid_ADX)
173 X86.HasSHA = isSet(ebx7, cpuid_SHA)
174
175 X86.HasAVX512F = isSet(ebx7, cpuid_AVX512F) && osSupportsAVX512
176 if X86.HasAVX512F {
177 X86.HasAVX512CD = isSet(ebx7, cpuid_AVX512CD)
178 X86.HasAVX512BW = isSet(ebx7, cpuid_AVX512BW)
179 X86.HasAVX512DQ = isSet(ebx7, cpuid_AVX512DQ)
180 X86.HasAVX512VL = isSet(ebx7, cpuid_AVX512VL)
181 X86.HasAVX512GFNI = isSet(ecx7, cpuid_AVX512GFNI)
182 X86.HasAVX512BITALG = isSet(ecx7, cpuid_AVX512BITALG)
183 X86.HasAVX512VPOPCNTDQ = isSet(ecx7, cpuid_AVX512VPOPCNTDQ)
184 X86.HasAVX512VBMI = isSet(ecx7, cpuid_AVX512VBMI)
185 X86.HasAVX512VBMI2 = isSet(ecx7, cpuid_AVX512VBMI2)
186 X86.HasAVX512VAES = isSet(ecx7, cpuid_AVX512VAES)
187 X86.HasAVX512VNNI = isSet(ecx7, cpuid_AVX512VNNI)
188 X86.HasAVX512VPCLMULQDQ = isSet(ecx7, cpuid_AVX512VPCLMULQDQ)
189 X86.HasAVX512VBMI = isSet(ecx7, cpuid_AVX512_VBMI)
190 X86.HasAVX512VBMI2 = isSet(ecx7, cpuid_AVX512_VBMI2)
191 X86.HasGFNI = isSet(ecx7, cpuid_GFNI)
192 X86.HasAVX512BITALG = isSet(ecx7, cpuid_AVX512_BITALG)
193 }
194
195 X86.HasFSRM = isSet(edx7, cpuid_FSRM)
196
197 var maxExtendedInformation uint32
198 maxExtendedInformation, _, _, _ = cpuid(0x80000000, 0)
199
200 if maxExtendedInformation < 0x80000001 {
201 osInit()
202 return
203 }
204
205 _, _, _, edxExt1 := cpuid(0x80000001, 0)
206 X86.HasRDTSCP = isSet(edxExt1, cpuid_RDTSCP)
207
208 doDerived = func() {
209
210
211
212
213
214
215
216 X86.HasAVX512 = X86.HasAVX512F && X86.HasAVX512CD && X86.HasAVX512BW && X86.HasAVX512DQ && X86.HasAVX512VL
217 }
218
219 if eax7 >= 1 {
220 eax71, _, _, _ := cpuid(7, 1)
221 if X86.HasAVX {
222 X86.HasAVXVNNI = isSet(4, eax71)
223 }
224 }
225
226 osInit()
227 }
228
229 func isSet(hwc uint32, value uint32) bool {
230 return hwc&value != 0
231 }
232
233
234
235
236 func Name() string {
237 if maxExtendedFunctionInformation < 0x80000004 {
238 return ""
239 }
240
241 data := make([]byte, 0, 3*4*4)
242
243 var eax, ebx, ecx, edx uint32
244 eax, ebx, ecx, edx = cpuid(0x80000002, 0)
245 data = appendBytes(data, eax, ebx, ecx, edx)
246 eax, ebx, ecx, edx = cpuid(0x80000003, 0)
247 data = appendBytes(data, eax, ebx, ecx, edx)
248 eax, ebx, ecx, edx = cpuid(0x80000004, 0)
249 data = appendBytes(data, eax, ebx, ecx, edx)
250
251
252 for len(data) > 0 && data[0] == ' ' {
253 data = data[1:]
254 }
255
256
257 for i, c := range data {
258 if c == '\x00' {
259 data = data[:i]
260 break
261 }
262 }
263
264 return string(data)
265 }
266
267 func appendBytes(b []byte, args ...uint32) []byte {
268 for _, arg := range args {
269 b = append(b,
270 byte((arg >> 0)),
271 byte((arg >> 8)),
272 byte((arg >> 16)),
273 byte((arg >> 24)))
274 }
275 return b
276 }
277
View as plain text