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(1010), // gfx1010 117 GFX(1011), // gfx1011 118 GFX(1012), // gfx1012 119 GFX(1013), // gfx1013 120 GFX(1030), // gfx1030 121 GFX(1031), // gfx1031 122 GFX(1032), // gfx1032 123 GFX(1033), // gfx1033 124 GFX(1034), // gfx1034 125 GFX(1035), // gfx1035 126 // clang-format on 127 }; 128 #undef SM 129 #undef SM2 130 #undef GFX 131 132 const char *CudaArchToString(CudaArch A) { 133 auto result = std::find_if( 134 std::begin(arch_names), std::end(arch_names), 135 [A](const CudaArchToStringMap &map) { return A == map.arch; }); 136 if (result == std::end(arch_names)) 137 return "unknown"; 138 return result->arch_name; 139 } 140 141 const char *CudaArchToVirtualArchString(CudaArch A) { 142 auto result = std::find_if( 143 std::begin(arch_names), std::end(arch_names), 144 [A](const CudaArchToStringMap &map) { return A == map.arch; }); 145 if (result == std::end(arch_names)) 146 return "unknown"; 147 return result->virtual_arch_name; 148 } 149 150 CudaArch StringToCudaArch(llvm::StringRef S) { 151 auto result = std::find_if( 152 std::begin(arch_names), std::end(arch_names), 153 [S](const CudaArchToStringMap &map) { return S == map.arch_name; }); 154 if (result == std::end(arch_names)) 155 return CudaArch::UNKNOWN; 156 return result->arch; 157 } 158 159 CudaVersion MinVersionForCudaArch(CudaArch A) { 160 if (A == CudaArch::UNKNOWN) 161 return CudaVersion::UNKNOWN; 162 163 // AMD GPUs do not depend on CUDA versions. 164 if (IsAMDGpuArch(A)) 165 return CudaVersion::CUDA_70; 166 167 switch (A) { 168 case CudaArch::SM_20: 169 case CudaArch::SM_21: 170 case CudaArch::SM_30: 171 case CudaArch::SM_32: 172 case CudaArch::SM_35: 173 case CudaArch::SM_37: 174 case CudaArch::SM_50: 175 case CudaArch::SM_52: 176 case CudaArch::SM_53: 177 return CudaVersion::CUDA_70; 178 case CudaArch::SM_60: 179 case CudaArch::SM_61: 180 case CudaArch::SM_62: 181 return CudaVersion::CUDA_80; 182 case CudaArch::SM_70: 183 return CudaVersion::CUDA_90; 184 case CudaArch::SM_72: 185 return CudaVersion::CUDA_91; 186 case CudaArch::SM_75: 187 return CudaVersion::CUDA_100; 188 case CudaArch::SM_80: 189 return CudaVersion::CUDA_110; 190 case CudaArch::SM_86: 191 return CudaVersion::CUDA_111; 192 default: 193 llvm_unreachable("invalid enum"); 194 } 195 } 196 197 CudaVersion MaxVersionForCudaArch(CudaArch A) { 198 // AMD GPUs do not depend on CUDA versions. 199 if (IsAMDGpuArch(A)) 200 return CudaVersion::NEW; 201 202 switch (A) { 203 case CudaArch::UNKNOWN: 204 return CudaVersion::UNKNOWN; 205 case CudaArch::SM_20: 206 case CudaArch::SM_21: 207 return CudaVersion::CUDA_80; 208 case CudaArch::SM_30: 209 return CudaVersion::CUDA_110; 210 default: 211 return CudaVersion::NEW; 212 } 213 } 214 215 CudaVersion ToCudaVersion(llvm::VersionTuple Version) { 216 int IVer = 217 Version.getMajor() * 10 + Version.getMinor().getValueOr(0); 218 switch(IVer) { 219 case 70: 220 return CudaVersion::CUDA_70; 221 case 75: 222 return CudaVersion::CUDA_75; 223 case 80: 224 return CudaVersion::CUDA_80; 225 case 90: 226 return CudaVersion::CUDA_90; 227 case 91: 228 return CudaVersion::CUDA_91; 229 case 92: 230 return CudaVersion::CUDA_92; 231 case 100: 232 return CudaVersion::CUDA_100; 233 case 101: 234 return CudaVersion::CUDA_101; 235 case 102: 236 return CudaVersion::CUDA_102; 237 case 110: 238 return CudaVersion::CUDA_110; 239 case 111: 240 return CudaVersion::CUDA_111; 241 case 112: 242 return CudaVersion::CUDA_112; 243 case 113: 244 return CudaVersion::CUDA_113; 245 case 114: 246 return CudaVersion::CUDA_114; 247 case 115: 248 return CudaVersion::CUDA_115; 249 default: 250 return CudaVersion::UNKNOWN; 251 } 252 } 253 254 bool CudaFeatureEnabled(llvm::VersionTuple Version, CudaFeature Feature) { 255 return CudaFeatureEnabled(ToCudaVersion(Version), Feature); 256 } 257 258 bool CudaFeatureEnabled(CudaVersion Version, CudaFeature Feature) { 259 switch (Feature) { 260 case CudaFeature::CUDA_USES_NEW_LAUNCH: 261 return Version >= CudaVersion::CUDA_92; 262 case CudaFeature::CUDA_USES_FATBIN_REGISTER_END: 263 return Version >= CudaVersion::CUDA_101; 264 } 265 llvm_unreachable("Unknown CUDA feature."); 266 } 267 } // namespace clang 268