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_VAES = 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_VPCLMULQDQ = 1 << 10
61 cpuid_AVX512VPCLMULQDQ = 1 << 10
62 cpuid_AVX512_BITALG = 1 << 12
63
64
65 cpuid_FSRM = 1 << 4
66
67 cpuid_RDTSCP = 1 << 27
68 )
69
70 var maxExtendedFunctionInformation uint32
71
72 func doinit() {
73 options = []option{
74 {Name: "adx", Feature: &X86.HasADX},
75 {Name: "aes", Feature: &X86.HasAES},
76 {Name: "erms", Feature: &X86.HasERMS},
77 {Name: "fsrm", Feature: &X86.HasFSRM},
78 {Name: "pclmulqdq", Feature: &X86.HasPCLMULQDQ},
79 {Name: "rdtscp", Feature: &X86.HasRDTSCP},
80 {Name: "sha", Feature: &X86.HasSHA},
81 {Name: "vpclmulqdq", Feature: &X86.HasAVX512VPCLMULQDQ},
82 }
83 level := getGOAMD64level()
84 if level < 2 {
85
86
87 options = append(options,
88 option{Name: "popcnt", Feature: &X86.HasPOPCNT},
89 option{Name: "sse3", Feature: &X86.HasSSE3},
90 option{Name: "sse41", Feature: &X86.HasSSE41},
91 option{Name: "sse42", Feature: &X86.HasSSE42},
92 option{Name: "ssse3", Feature: &X86.HasSSSE3})
93 }
94 if level < 3 {
95
96
97 options = append(options,
98 option{Name: "avx", Feature: &X86.HasAVX},
99 option{Name: "avx2", Feature: &X86.HasAVX2},
100 option{Name: "bmi1", Feature: &X86.HasBMI1},
101 option{Name: "bmi2", Feature: &X86.HasBMI2},
102 option{Name: "fma", Feature: &X86.HasFMA})
103 }
104 if level < 4 {
105
106
107 options = append(options,
108 option{Name: "avx512f", Feature: &X86.HasAVX512F},
109 option{Name: "avx512cd", Feature: &X86.HasAVX512CD},
110 option{Name: "avx512bw", Feature: &X86.HasAVX512BW},
111 option{Name: "avx512dq", Feature: &X86.HasAVX512DQ},
112 option{Name: "avx512vl", Feature: &X86.HasAVX512VL},
113 )
114 }
115
116 maxID, _, _, _ := cpuid(0, 0)
117
118 if maxID < 1 {
119 osInit()
120 return
121 }
122
123 maxExtendedFunctionInformation, _, _, _ = cpuid(0x80000000, 0)
124
125 _, _, ecx1, _ := cpuid(1, 0)
126
127 X86.HasSSE3 = isSet(ecx1, cpuid_SSE3)
128 X86.HasPCLMULQDQ = isSet(ecx1, cpuid_PCLMULQDQ)
129 X86.HasSSSE3 = isSet(ecx1, cpuid_SSSE3)
130 X86.HasSSE41 = isSet(ecx1, cpuid_SSE41)
131 X86.HasSSE42 = isSet(ecx1, cpuid_SSE42)
132 X86.HasPOPCNT = isSet(ecx1, cpuid_POPCNT)
133 X86.HasAES = isSet(ecx1, cpuid_AES)
134
135
136
137
138 X86.HasOSXSAVE = isSet(ecx1, cpuid_OSXSAVE)
139
140 osSupportsAVX := false
141 osSupportsAVX512 := false
142
143 if X86.HasOSXSAVE {
144 eax, _ := xgetbv()
145
146 osSupportsAVX = isSet(eax, 1<<1) && isSet(eax, 1<<2)
147
148
149
150
151
152 osSupportsAVX512 = osSupportsAVX && isSet(eax, 1<<5) && isSet(eax, 1<<6) && isSet(eax, 1<<7)
153 }
154
155 X86.HasAVX = isSet(ecx1, cpuid_AVX) && osSupportsAVX
156
157
158
159
160
161
162
163 X86.HasFMA = isSet(ecx1, cpuid_FMA) && X86.HasAVX && X86.HasOSXSAVE
164
165 if maxID < 7 {
166 osInit()
167 return
168 }
169
170 eax7, ebx7, ecx7, edx7 := cpuid(7, 0)
171 X86.HasBMI1 = isSet(ebx7, cpuid_BMI1)
172 X86.HasAVX2 = isSet(ebx7, cpuid_AVX2) && osSupportsAVX
173 X86.HasBMI2 = isSet(ebx7, cpuid_BMI2)
174 X86.HasERMS = isSet(ebx7, cpuid_ERMS)
175 X86.HasADX = isSet(ebx7, cpuid_ADX)
176 X86.HasSHA = isSet(ebx7, cpuid_SHA)
177 X86.HasVAES = isSet(ecx7, cpuid_VAES) && X86.HasAVX
178 X86.HasVPCLMULQDQ = isSet(ecx7, cpuid_VPCLMULQDQ)
179
180 X86.HasAVX512F = isSet(ebx7, cpuid_AVX512F) && osSupportsAVX512
181 if X86.HasAVX512F {
182 X86.HasAVX512CD = isSet(ebx7, cpuid_AVX512CD)
183 X86.HasAVX512BW = isSet(ebx7, cpuid_AVX512BW)
184 X86.HasAVX512DQ = isSet(ebx7, cpuid_AVX512DQ)
185 X86.HasAVX512VL = isSet(ebx7, cpuid_AVX512VL)
186 X86.HasAVX512GFNI = isSet(ecx7, cpuid_AVX512GFNI)
187 X86.HasAVX512BITALG = isSet(ecx7, cpuid_AVX512BITALG)
188 X86.HasAVX512VPOPCNTDQ = isSet(ecx7, cpuid_AVX512VPOPCNTDQ)
189 X86.HasAVX512VBMI = isSet(ecx7, cpuid_AVX512VBMI)
190 X86.HasAVX512VBMI2 = isSet(ecx7, cpuid_AVX512VBMI2)
191 X86.HasAVX512VAES = isSet(ecx7, cpuid_VAES) && X86.HasAES && isSet(ebx7, cpuid_AVX512VL)
192 X86.HasAVX512VNNI = isSet(ecx7, cpuid_AVX512VNNI)
193 X86.HasAVX512VPCLMULQDQ = isSet(ecx7, cpuid_AVX512VPCLMULQDQ)
194 X86.HasAVX512VBMI = isSet(ecx7, cpuid_AVX512_VBMI)
195 X86.HasAVX512VBMI2 = isSet(ecx7, cpuid_AVX512_VBMI2)
196 X86.HasGFNI = isSet(ecx7, cpuid_GFNI)
197 X86.HasAVX512BITALG = isSet(ecx7, cpuid_AVX512_BITALG)
198 }
199
200 X86.HasFSRM = isSet(edx7, cpuid_FSRM)
201
202 var maxExtendedInformation uint32
203 maxExtendedInformation, _, _, _ = cpuid(0x80000000, 0)
204
205 if maxExtendedInformation < 0x80000001 {
206 osInit()
207 return
208 }
209
210 _, _, _, edxExt1 := cpuid(0x80000001, 0)
211 X86.HasRDTSCP = isSet(edxExt1, cpuid_RDTSCP)
212
213 doDerived = func() {
214
215
216
217
218
219
220
221 X86.HasAVX512 = X86.HasAVX512F && X86.HasAVX512CD && X86.HasAVX512BW && X86.HasAVX512DQ && X86.HasAVX512VL
222 }
223
224 if eax7 >= 1 {
225 eax71, _, _, _ := cpuid(7, 1)
226 if X86.HasAVX {
227 X86.HasAVXVNNI = isSet(eax71, cpuid_AVXVNNI)
228 }
229 }
230
231 osInit()
232 }
233
234 func isSet(hwc uint32, value uint32) bool {
235 return hwc&value != 0
236 }
237
238
239
240
241 func Name() string {
242 if maxExtendedFunctionInformation < 0x80000004 {
243 return ""
244 }
245
246 data := make([]byte, 0, 3*4*4)
247
248 var eax, ebx, ecx, edx uint32
249 eax, ebx, ecx, edx = cpuid(0x80000002, 0)
250 data = appendBytes(data, eax, ebx, ecx, edx)
251 eax, ebx, ecx, edx = cpuid(0x80000003, 0)
252 data = appendBytes(data, eax, ebx, ecx, edx)
253 eax, ebx, ecx, edx = cpuid(0x80000004, 0)
254 data = appendBytes(data, eax, ebx, ecx, edx)
255
256
257 for len(data) > 0 && data[0] == ' ' {
258 data = data[1:]
259 }
260
261
262 for i, c := range data {
263 if c == '\x00' {
264 data = data[:i]
265 break
266 }
267 }
268
269 return string(data)
270 }
271
272 func appendBytes(b []byte, args ...uint32) []byte {
273 for _, arg := range args {
274 b = append(b,
275 byte((arg >> 0)),
276 byte((arg >> 8)),
277 byte((arg >> 16)),
278 byte((arg >> 24)))
279 }
280 return b
281 }
282
View as plain text