1
2
3
4
5
6
7 package cpu
8
9 import "runtime"
10
11 const cacheLineSize = 64
12
13 func initOptions() {
14 options = []option{
15 {Name: "adx", Feature: &X86.HasADX},
16 {Name: "aes", Feature: &X86.HasAES},
17 {Name: "avx", Feature: &X86.HasAVX},
18 {Name: "avx2", Feature: &X86.HasAVX2},
19 {Name: "avx512", Feature: &X86.HasAVX512},
20 {Name: "avx512f", Feature: &X86.HasAVX512F},
21 {Name: "avx512cd", Feature: &X86.HasAVX512CD},
22 {Name: "avx512er", Feature: &X86.HasAVX512ER},
23 {Name: "avx512pf", Feature: &X86.HasAVX512PF},
24 {Name: "avx512vl", Feature: &X86.HasAVX512VL},
25 {Name: "avx512bw", Feature: &X86.HasAVX512BW},
26 {Name: "avx512dq", Feature: &X86.HasAVX512DQ},
27 {Name: "avx512ifma", Feature: &X86.HasAVX512IFMA},
28 {Name: "avx512vbmi", Feature: &X86.HasAVX512VBMI},
29 {Name: "avx512vnniw", Feature: &X86.HasAVX5124VNNIW},
30 {Name: "avx5124fmaps", Feature: &X86.HasAVX5124FMAPS},
31 {Name: "avx512vpopcntdq", Feature: &X86.HasAVX512VPOPCNTDQ},
32 {Name: "avx512vpclmulqdq", Feature: &X86.HasAVX512VPCLMULQDQ},
33 {Name: "avx512vnni", Feature: &X86.HasAVX512VNNI},
34 {Name: "avx512gfni", Feature: &X86.HasAVX512GFNI},
35 {Name: "avx512vaes", Feature: &X86.HasAVX512VAES},
36 {Name: "avx512vbmi2", Feature: &X86.HasAVX512VBMI2},
37 {Name: "avx512bitalg", Feature: &X86.HasAVX512BITALG},
38 {Name: "avx512bf16", Feature: &X86.HasAVX512BF16},
39 {Name: "amxtile", Feature: &X86.HasAMXTile},
40 {Name: "amxint8", Feature: &X86.HasAMXInt8},
41 {Name: "amxbf16", Feature: &X86.HasAMXBF16},
42 {Name: "bmi1", Feature: &X86.HasBMI1},
43 {Name: "bmi2", Feature: &X86.HasBMI2},
44 {Name: "cx16", Feature: &X86.HasCX16},
45 {Name: "erms", Feature: &X86.HasERMS},
46 {Name: "fma", Feature: &X86.HasFMA},
47 {Name: "osxsave", Feature: &X86.HasOSXSAVE},
48 {Name: "pclmulqdq", Feature: &X86.HasPCLMULQDQ},
49 {Name: "popcnt", Feature: &X86.HasPOPCNT},
50 {Name: "rdrand", Feature: &X86.HasRDRAND},
51 {Name: "rdseed", Feature: &X86.HasRDSEED},
52 {Name: "sse3", Feature: &X86.HasSSE3},
53 {Name: "sse41", Feature: &X86.HasSSE41},
54 {Name: "sse42", Feature: &X86.HasSSE42},
55 {Name: "ssse3", Feature: &X86.HasSSSE3},
56 {Name: "avxifma", Feature: &X86.HasAVXIFMA},
57 {Name: "avxvnni", Feature: &X86.HasAVXVNNI},
58 {Name: "avxvnniint8", Feature: &X86.HasAVXVNNIInt8},
59
60
61 {Name: "sse2", Feature: &X86.HasSSE2, Required: runtime.GOARCH == "amd64"},
62 }
63 }
64
65 func archInit() {
66
67
68 const (
69
70 cpuid_AVXVNNI = 1 << 4
71
72
73 cpuid_SSE3 = 1 << 0
74 cpuid_PCLMULQDQ = 1 << 1
75 cpuid_AVX512VBMI = 1 << 1
76 cpuid_AVX512VBMI2 = 1 << 6
77 cpuid_SSSE3 = 1 << 9
78 cpuid_AVX512GFNI = 1 << 8
79 cpuid_AVX512VAES = 1 << 9
80 cpuid_AVX512VNNI = 1 << 11
81 cpuid_AVX512BITALG = 1 << 12
82 cpuid_FMA = 1 << 12
83 cpuid_AVX512VPOPCNTDQ = 1 << 14
84 cpuid_SSE41 = 1 << 19
85 cpuid_SSE42 = 1 << 20
86 cpuid_POPCNT = 1 << 23
87 cpuid_AES = 1 << 25
88 cpuid_OSXSAVE = 1 << 27
89 cpuid_AVX = 1 << 28
90
91
92 cpuid_BMI1 = 1 << 3
93 cpuid_AVX2 = 1 << 5
94 cpuid_BMI2 = 1 << 8
95 cpuid_ERMS = 1 << 9
96 cpuid_AVX512F = 1 << 16
97 cpuid_AVX512DQ = 1 << 17
98 cpuid_ADX = 1 << 19
99 cpuid_AVX512CD = 1 << 28
100 cpuid_SHA = 1 << 29
101 cpuid_AVX512BW = 1 << 30
102 cpuid_AVX512VL = 1 << 31
103
104
105 cpuid_AVX512_VBMI = 1 << 1
106 cpuid_AVX512_VBMI2 = 1 << 6
107 cpuid_GFNI = 1 << 8
108 cpuid_AVX512VPCLMULQDQ = 1 << 10
109 cpuid_AVX512_BITALG = 1 << 12
110
111
112 cpuid_FSRM = 1 << 4
113
114 cpuid_RDTSCP = 1 << 27
115 )
116
117 const (
118
119 cpuid_SSE2 = 1 << 26
120
121 cpuid_CX16 = 1 << 13
122 cpuid_RDRAND = 1 << 30
123
124 cpuid_RDSEED = 1 << 18
125 cpuid_AVX512IFMA = 1 << 21
126 cpuid_AVX512PF = 1 << 26
127 cpuid_AVX512ER = 1 << 27
128
129 cpuid_AVX5124VNNIW = 1 << 2
130 cpuid_AVX5124FMAPS = 1 << 3
131 cpuid_AMXBF16 = 1 << 22
132 cpuid_AMXTile = 1 << 24
133 cpuid_AMXInt8 = 1 << 25
134
135 cpuid_AVX512BF16 = 1 << 5
136 cpuid_AVXIFMA = 1 << 23
137
138 cpuid_AVXVNNIInt8 = 1 << 4
139 )
140
141 Initialized = true
142
143 maxID, _, _, _ := cpuid(0, 0)
144
145 if maxID < 1 {
146 return
147 }
148
149 _, _, ecx1, edx1 := cpuid(1, 0)
150 X86.HasSSE2 = isSet(edx1, cpuid_SSE2)
151
152 X86.HasSSE3 = isSet(ecx1, cpuid_SSE3)
153 X86.HasPCLMULQDQ = isSet(ecx1, cpuid_PCLMULQDQ)
154 X86.HasSSSE3 = isSet(ecx1, cpuid_SSSE3)
155 X86.HasFMA = isSet(ecx1, cpuid_FMA)
156 X86.HasCX16 = isSet(ecx1, cpuid_CX16)
157 X86.HasSSE41 = isSet(ecx1, cpuid_SSE41)
158 X86.HasSSE42 = isSet(ecx1, cpuid_SSE42)
159 X86.HasPOPCNT = isSet(ecx1, cpuid_POPCNT)
160 X86.HasAES = isSet(ecx1, cpuid_AES)
161 X86.HasOSXSAVE = isSet(ecx1, cpuid_OSXSAVE)
162 X86.HasRDRAND = isSet(ecx1, cpuid_RDRAND)
163
164 var osSupportsAVX, osSupportsAVX512 bool
165
166 if X86.HasOSXSAVE {
167 eax, _ := xgetbv()
168
169 osSupportsAVX = isSet(eax, 1<<1) && isSet(eax, 1<<2)
170
171 if runtime.GOOS == "darwin" {
172
173 osSupportsAVX512 = osSupportsAVX && darwinSupportsAVX512()
174 } else {
175
176 osSupportsAVX512 = osSupportsAVX && isSet(eax, 1<<5) && isSet(eax, 1<<6) && isSet(eax, 1<<7)
177 }
178 }
179
180 X86.HasAVX = isSet(ecx1, cpuid_AVX) && osSupportsAVX
181
182 if maxID < 7 {
183 return
184 }
185
186 eax7, ebx7, ecx7, edx7 := cpuid(7, 0)
187 X86.HasBMI1 = isSet(ebx7, cpuid_BMI1)
188 X86.HasAVX2 = isSet(ebx7, cpuid_AVX2) && osSupportsAVX
189 X86.HasBMI2 = isSet(ebx7, cpuid_BMI2)
190 X86.HasERMS = isSet(ebx7, cpuid_ERMS)
191 X86.HasRDSEED = isSet(ebx7, cpuid_RDSEED)
192 X86.HasADX = isSet(ebx7, cpuid_ADX)
193
194 X86.HasAVX512 = isSet(ebx7, cpuid_AVX512F) && osSupportsAVX512
195 if X86.HasAVX512 {
196 X86.HasAVX512F = true
197 X86.HasAVX512CD = isSet(ebx7, cpuid_AVX512CD)
198 X86.HasAVX512ER = isSet(ebx7, cpuid_AVX512ER)
199 X86.HasAVX512PF = isSet(ebx7, cpuid_AVX512PF)
200 X86.HasAVX512VL = isSet(ebx7, cpuid_AVX512VL)
201 X86.HasAVX512BW = isSet(ebx7, cpuid_AVX512BW)
202 X86.HasAVX512DQ = isSet(ebx7, cpuid_AVX512DQ)
203 X86.HasAVX512IFMA = isSet(ebx7, cpuid_AVX512IFMA)
204 X86.HasAVX512VBMI = isSet(ecx7, cpuid_AVX512_VBMI)
205 X86.HasAVX5124VNNIW = isSet(edx7, cpuid_AVX5124VNNIW)
206 X86.HasAVX5124FMAPS = isSet(edx7, cpuid_AVX5124FMAPS)
207 X86.HasAVX512VPOPCNTDQ = isSet(ecx7, cpuid_AVX512VPOPCNTDQ)
208 X86.HasAVX512VPCLMULQDQ = isSet(ecx7, cpuid_AVX512VPCLMULQDQ)
209 X86.HasAVX512VNNI = isSet(ecx7, cpuid_AVX512VNNI)
210 X86.HasAVX512GFNI = isSet(ecx7, cpuid_AVX512GFNI)
211 X86.HasAVX512VAES = isSet(ecx7, cpuid_AVX512VAES)
212 X86.HasAVX512VBMI2 = isSet(ecx7, cpuid_AVX512VBMI2)
213 X86.HasAVX512BITALG = isSet(ecx7, cpuid_AVX512BITALG)
214 }
215
216 X86.HasAMXTile = isSet(edx7, cpuid_AMXTile)
217 X86.HasAMXInt8 = isSet(edx7, cpuid_AMXInt8)
218 X86.HasAMXBF16 = isSet(edx7, cpuid_AMXBF16)
219
220
221 if eax7 >= 1 {
222 eax71, _, _, edx71 := cpuid(7, 1)
223 if X86.HasAVX512 {
224 X86.HasAVX512BF16 = isSet(eax71, cpuid_AVX512BF16)
225 }
226 if X86.HasAVX {
227 X86.HasAVXIFMA = isSet(eax71, cpuid_AVXIFMA)
228 X86.HasAVXVNNI = isSet(eax71, cpuid_AVXVNNI)
229 X86.HasAVXVNNIInt8 = isSet(edx71, cpuid_AVXVNNIInt8)
230 }
231 }
232 }
233
234 func isSet(hwc uint32, value uint32) bool {
235 return hwc&value != 0
236 }
237
View as plain text