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 {CudaArch::Generic, "generic", ""}, 128 // clang-format on 129 }; 130 #undef SM 131 #undef SM2 132 #undef GFX 133 134 const char *CudaArchToString(CudaArch A) { 135 auto result = std::find_if( 136 std::begin(arch_names), std::end(arch_names), 137 [A](const CudaArchToStringMap &map) { return A == map.arch; }); 138 if (result == std::end(arch_names)) 139 return "unknown"; 140 return result->arch_name; 141 } 142 143 const char *CudaArchToVirtualArchString(CudaArch A) { 144 auto result = std::find_if( 145 std::begin(arch_names), std::end(arch_names), 146 [A](const CudaArchToStringMap &map) { return A == map.arch; }); 147 if (result == std::end(arch_names)) 148 return "unknown"; 149 return result->virtual_arch_name; 150 } 151 152 CudaArch StringToCudaArch(llvm::StringRef S) { 153 auto result = std::find_if( 154 std::begin(arch_names), std::end(arch_names), 155 [S](const CudaArchToStringMap &map) { return S == map.arch_name; }); 156 if (result == std::end(arch_names)) 157 return CudaArch::UNKNOWN; 158 return result->arch; 159 } 160 161 CudaVersion MinVersionForCudaArch(CudaArch A) { 162 if (A == CudaArch::UNKNOWN) 163 return CudaVersion::UNKNOWN; 164 165 // AMD GPUs do not depend on CUDA versions. 166 if (IsAMDGpuArch(A)) 167 return CudaVersion::CUDA_70; 168 169 switch (A) { 170 case CudaArch::SM_20: 171 case CudaArch::SM_21: 172 case CudaArch::SM_30: 173 case CudaArch::SM_32: 174 case CudaArch::SM_35: 175 case CudaArch::SM_37: 176 case CudaArch::SM_50: 177 case CudaArch::SM_52: 178 case CudaArch::SM_53: 179 return CudaVersion::CUDA_70; 180 case CudaArch::SM_60: 181 case CudaArch::SM_61: 182 case CudaArch::SM_62: 183 return CudaVersion::CUDA_80; 184 case CudaArch::SM_70: 185 return CudaVersion::CUDA_90; 186 case CudaArch::SM_72: 187 return CudaVersion::CUDA_91; 188 case CudaArch::SM_75: 189 return CudaVersion::CUDA_100; 190 case CudaArch::SM_80: 191 return CudaVersion::CUDA_110; 192 case CudaArch::SM_86: 193 return CudaVersion::CUDA_111; 194 default: 195 llvm_unreachable("invalid enum"); 196 } 197 } 198 199 CudaVersion MaxVersionForCudaArch(CudaArch A) { 200 // AMD GPUs do not depend on CUDA versions. 201 if (IsAMDGpuArch(A)) 202 return CudaVersion::NEW; 203 204 switch (A) { 205 case CudaArch::UNKNOWN: 206 return CudaVersion::UNKNOWN; 207 case CudaArch::SM_20: 208 case CudaArch::SM_21: 209 return CudaVersion::CUDA_80; 210 case CudaArch::SM_30: 211 return CudaVersion::CUDA_110; 212 default: 213 return CudaVersion::NEW; 214 } 215 } 216 217 CudaVersion ToCudaVersion(llvm::VersionTuple Version) { 218 int IVer = 219 Version.getMajor() * 10 + Version.getMinor().getValueOr(0); 220 switch(IVer) { 221 case 70: 222 return CudaVersion::CUDA_70; 223 case 75: 224 return CudaVersion::CUDA_75; 225 case 80: 226 return CudaVersion::CUDA_80; 227 case 90: 228 return CudaVersion::CUDA_90; 229 case 91: 230 return CudaVersion::CUDA_91; 231 case 92: 232 return CudaVersion::CUDA_92; 233 case 100: 234 return CudaVersion::CUDA_100; 235 case 101: 236 return CudaVersion::CUDA_101; 237 case 102: 238 return CudaVersion::CUDA_102; 239 case 110: 240 return CudaVersion::CUDA_110; 241 case 111: 242 return CudaVersion::CUDA_111; 243 case 112: 244 return CudaVersion::CUDA_112; 245 case 113: 246 return CudaVersion::CUDA_113; 247 case 114: 248 return CudaVersion::CUDA_114; 249 case 115: 250 return CudaVersion::CUDA_115; 251 default: 252 return CudaVersion::UNKNOWN; 253 } 254 } 255 256 bool CudaFeatureEnabled(llvm::VersionTuple Version, CudaFeature Feature) { 257 return CudaFeatureEnabled(ToCudaVersion(Version), Feature); 258 } 259 260 bool CudaFeatureEnabled(CudaVersion Version, CudaFeature Feature) { 261 switch (Feature) { 262 case CudaFeature::CUDA_USES_NEW_LAUNCH: 263 return Version >= CudaVersion::CUDA_92; 264 case CudaFeature::CUDA_USES_FATBIN_REGISTER_END: 265 return Version >= CudaVersion::CUDA_101; 266 } 267 llvm_unreachable("Unknown CUDA feature."); 268 } 269 } // namespace clang 270