xref: /freebsd/contrib/llvm-project/clang/lib/Basic/Cuda.cpp (revision ebacd8013fe5f7fdf9f6a5b286f6680dd2891036)
1 #include "clang/Basic/Cuda.h"
2 
3 #include "llvm/ADT/StringRef.h"
4 #include "llvm/ADT/StringSwitch.h"
5 #include "llvm/ADT/Twine.h"
6 #include "llvm/Support/ErrorHandling.h"
7 #include "llvm/Support/VersionTuple.h"
8 
9 namespace clang {
10 
11 const char *CudaVersionToString(CudaVersion V) {
12   switch (V) {
13   case CudaVersion::UNKNOWN:
14     return "unknown";
15   case CudaVersion::CUDA_70:
16     return "7.0";
17   case CudaVersion::CUDA_75:
18     return "7.5";
19   case CudaVersion::CUDA_80:
20     return "8.0";
21   case CudaVersion::CUDA_90:
22     return "9.0";
23   case CudaVersion::CUDA_91:
24     return "9.1";
25   case CudaVersion::CUDA_92:
26     return "9.2";
27   case CudaVersion::CUDA_100:
28     return "10.0";
29   case CudaVersion::CUDA_101:
30     return "10.1";
31   case CudaVersion::CUDA_102:
32     return "10.2";
33   case CudaVersion::CUDA_110:
34     return "11.0";
35   case CudaVersion::CUDA_111:
36     return "11.1";
37   case CudaVersion::CUDA_112:
38     return "11.2";
39   case CudaVersion::CUDA_113:
40     return "11.3";
41   case CudaVersion::CUDA_114:
42     return "11.4";
43   case CudaVersion::CUDA_115:
44     return "11.5";
45   case CudaVersion::NEW:
46     return "";
47   }
48   llvm_unreachable("invalid enum");
49 }
50 
51 CudaVersion CudaStringToVersion(const llvm::Twine &S) {
52   return llvm::StringSwitch<CudaVersion>(S.str())
53       .Case("7.0", CudaVersion::CUDA_70)
54       .Case("7.5", CudaVersion::CUDA_75)
55       .Case("8.0", CudaVersion::CUDA_80)
56       .Case("9.0", CudaVersion::CUDA_90)
57       .Case("9.1", CudaVersion::CUDA_91)
58       .Case("9.2", CudaVersion::CUDA_92)
59       .Case("10.0", CudaVersion::CUDA_100)
60       .Case("10.1", CudaVersion::CUDA_101)
61       .Case("10.2", CudaVersion::CUDA_102)
62       .Case("11.0", CudaVersion::CUDA_110)
63       .Case("11.1", CudaVersion::CUDA_111)
64       .Case("11.2", CudaVersion::CUDA_112)
65       .Case("11.3", CudaVersion::CUDA_113)
66       .Case("11.4", CudaVersion::CUDA_114)
67       .Case("11.5", CudaVersion::CUDA_115)
68       .Default(CudaVersion::UNKNOWN);
69 }
70 
71 namespace {
72 struct CudaArchToStringMap {
73   CudaArch arch;
74   const char *arch_name;
75   const char *virtual_arch_name;
76 };
77 } // namespace
78 
79 #define SM2(sm, ca)                                                            \
80   { CudaArch::SM_##sm, "sm_" #sm, ca }
81 #define SM(sm) SM2(sm, "compute_" #sm)
82 #define GFX(gpu)                                                               \
83   { CudaArch::GFX##gpu, "gfx" #gpu, "compute_amdgcn" }
84 static const CudaArchToStringMap arch_names[] = {
85     // clang-format off
86     {CudaArch::UNUSED, "", ""},
87     SM2(20, "compute_20"), SM2(21, "compute_20"), // Fermi
88     SM(30), SM(32), SM(35), SM(37),  // Kepler
89     SM(50), SM(52), SM(53),          // Maxwell
90     SM(60), SM(61), SM(62),          // Pascal
91     SM(70), SM(72),                  // Volta
92     SM(75),                          // Turing
93     SM(80), SM(86),                  // Ampere
94     GFX(600),  // gfx600
95     GFX(601),  // gfx601
96     GFX(602),  // gfx602
97     GFX(700),  // gfx700
98     GFX(701),  // gfx701
99     GFX(702),  // gfx702
100     GFX(703),  // gfx703
101     GFX(704),  // gfx704
102     GFX(705),  // gfx705
103     GFX(801),  // gfx801
104     GFX(802),  // gfx802
105     GFX(803),  // gfx803
106     GFX(805),  // gfx805
107     GFX(810),  // gfx810
108     GFX(900),  // gfx900
109     GFX(902),  // gfx902
110     GFX(904),  // gfx903
111     GFX(906),  // gfx906
112     GFX(908),  // gfx908
113     GFX(909),  // gfx909
114     GFX(90a),  // gfx90a
115     GFX(90c),  // gfx90c
116     GFX(940),  // gfx940
117     GFX(1010), // gfx1010
118     GFX(1011), // gfx1011
119     GFX(1012), // gfx1012
120     GFX(1013), // gfx1013
121     GFX(1030), // gfx1030
122     GFX(1031), // gfx1031
123     GFX(1032), // gfx1032
124     GFX(1033), // gfx1033
125     GFX(1034), // gfx1034
126     GFX(1035), // gfx1035
127     GFX(1036), // gfx1036
128     GFX(1100), // gfx1100
129     GFX(1101), // gfx1101
130     GFX(1102), // gfx1102
131     GFX(1103), // gfx1103
132     {CudaArch::Generic, "generic", ""},
133     // clang-format on
134 };
135 #undef SM
136 #undef SM2
137 #undef GFX
138 
139 const char *CudaArchToString(CudaArch A) {
140   auto result = std::find_if(
141       std::begin(arch_names), std::end(arch_names),
142       [A](const CudaArchToStringMap &map) { return A == map.arch; });
143   if (result == std::end(arch_names))
144     return "unknown";
145   return result->arch_name;
146 }
147 
148 const char *CudaArchToVirtualArchString(CudaArch A) {
149   auto result = std::find_if(
150       std::begin(arch_names), std::end(arch_names),
151       [A](const CudaArchToStringMap &map) { return A == map.arch; });
152   if (result == std::end(arch_names))
153     return "unknown";
154   return result->virtual_arch_name;
155 }
156 
157 CudaArch StringToCudaArch(llvm::StringRef S) {
158   auto result = std::find_if(
159       std::begin(arch_names), std::end(arch_names),
160       [S](const CudaArchToStringMap &map) { return S == map.arch_name; });
161   if (result == std::end(arch_names))
162     return CudaArch::UNKNOWN;
163   return result->arch;
164 }
165 
166 CudaVersion MinVersionForCudaArch(CudaArch A) {
167   if (A == CudaArch::UNKNOWN)
168     return CudaVersion::UNKNOWN;
169 
170   // AMD GPUs do not depend on CUDA versions.
171   if (IsAMDGpuArch(A))
172     return CudaVersion::CUDA_70;
173 
174   switch (A) {
175   case CudaArch::SM_20:
176   case CudaArch::SM_21:
177   case CudaArch::SM_30:
178   case CudaArch::SM_32:
179   case CudaArch::SM_35:
180   case CudaArch::SM_37:
181   case CudaArch::SM_50:
182   case CudaArch::SM_52:
183   case CudaArch::SM_53:
184     return CudaVersion::CUDA_70;
185   case CudaArch::SM_60:
186   case CudaArch::SM_61:
187   case CudaArch::SM_62:
188     return CudaVersion::CUDA_80;
189   case CudaArch::SM_70:
190     return CudaVersion::CUDA_90;
191   case CudaArch::SM_72:
192     return CudaVersion::CUDA_91;
193   case CudaArch::SM_75:
194     return CudaVersion::CUDA_100;
195   case CudaArch::SM_80:
196     return CudaVersion::CUDA_110;
197   case CudaArch::SM_86:
198     return CudaVersion::CUDA_111;
199   default:
200     llvm_unreachable("invalid enum");
201   }
202 }
203 
204 CudaVersion MaxVersionForCudaArch(CudaArch A) {
205   // AMD GPUs do not depend on CUDA versions.
206   if (IsAMDGpuArch(A))
207     return CudaVersion::NEW;
208 
209   switch (A) {
210   case CudaArch::UNKNOWN:
211     return CudaVersion::UNKNOWN;
212   case CudaArch::SM_20:
213   case CudaArch::SM_21:
214     return CudaVersion::CUDA_80;
215   case CudaArch::SM_30:
216     return CudaVersion::CUDA_110;
217   default:
218     return CudaVersion::NEW;
219   }
220 }
221 
222 CudaVersion ToCudaVersion(llvm::VersionTuple Version) {
223   int IVer = Version.getMajor() * 10 + Version.getMinor().value_or(0);
224   switch(IVer) {
225   case 70:
226     return CudaVersion::CUDA_70;
227   case 75:
228     return CudaVersion::CUDA_75;
229   case 80:
230     return CudaVersion::CUDA_80;
231   case 90:
232     return CudaVersion::CUDA_90;
233   case 91:
234     return CudaVersion::CUDA_91;
235   case 92:
236     return CudaVersion::CUDA_92;
237   case 100:
238     return CudaVersion::CUDA_100;
239   case 101:
240     return CudaVersion::CUDA_101;
241   case 102:
242     return CudaVersion::CUDA_102;
243   case 110:
244     return CudaVersion::CUDA_110;
245   case 111:
246     return CudaVersion::CUDA_111;
247   case 112:
248     return CudaVersion::CUDA_112;
249   case 113:
250     return CudaVersion::CUDA_113;
251   case 114:
252     return CudaVersion::CUDA_114;
253   case 115:
254     return CudaVersion::CUDA_115;
255   default:
256     return CudaVersion::UNKNOWN;
257   }
258 }
259 
260 bool CudaFeatureEnabled(llvm::VersionTuple  Version, CudaFeature Feature) {
261   return CudaFeatureEnabled(ToCudaVersion(Version), Feature);
262 }
263 
264 bool CudaFeatureEnabled(CudaVersion Version, CudaFeature Feature) {
265   switch (Feature) {
266   case CudaFeature::CUDA_USES_NEW_LAUNCH:
267     return Version >= CudaVersion::CUDA_92;
268   case CudaFeature::CUDA_USES_FATBIN_REGISTER_END:
269     return Version >= CudaVersion::CUDA_101;
270   }
271   llvm_unreachable("Unknown CUDA feature.");
272 }
273 } // namespace clang
274