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 } 40 llvm_unreachable("invalid enum"); 41 } 42 43 CudaVersion CudaStringToVersion(const llvm::Twine &S) { 44 return llvm::StringSwitch<CudaVersion>(S.str()) 45 .Case("7.0", CudaVersion::CUDA_70) 46 .Case("7.5", CudaVersion::CUDA_75) 47 .Case("8.0", CudaVersion::CUDA_80) 48 .Case("9.0", CudaVersion::CUDA_90) 49 .Case("9.1", CudaVersion::CUDA_91) 50 .Case("9.2", CudaVersion::CUDA_92) 51 .Case("10.0", CudaVersion::CUDA_100) 52 .Case("10.1", CudaVersion::CUDA_101) 53 .Case("10.2", CudaVersion::CUDA_102) 54 .Case("11.0", CudaVersion::CUDA_110) 55 .Case("11.1", CudaVersion::CUDA_111) 56 .Case("11.2", CudaVersion::CUDA_112) 57 .Default(CudaVersion::UNKNOWN); 58 } 59 60 struct CudaArchToStringMap { 61 CudaArch arch; 62 const char *arch_name; 63 const char *virtual_arch_name; 64 }; 65 66 #define SM2(sm, ca) \ 67 { CudaArch::SM_##sm, "sm_" #sm, ca } 68 #define SM(sm) SM2(sm, "compute_" #sm) 69 #define GFX(gpu) \ 70 { CudaArch::GFX##gpu, "gfx" #gpu, "compute_amdgcn" } 71 CudaArchToStringMap arch_names[] = { 72 // clang-format off 73 {CudaArch::UNUSED, "", ""}, 74 SM2(20, "compute_20"), SM2(21, "compute_20"), // Fermi 75 SM(30), SM(32), SM(35), SM(37), // Kepler 76 SM(50), SM(52), SM(53), // Maxwell 77 SM(60), SM(61), SM(62), // Pascal 78 SM(70), SM(72), // Volta 79 SM(75), // Turing 80 SM(80), SM(86), // Ampere 81 GFX(600), // gfx600 82 GFX(601), // gfx601 83 GFX(602), // gfx602 84 GFX(700), // gfx700 85 GFX(701), // gfx701 86 GFX(702), // gfx702 87 GFX(703), // gfx703 88 GFX(704), // gfx704 89 GFX(705), // gfx705 90 GFX(801), // gfx801 91 GFX(802), // gfx802 92 GFX(803), // gfx803 93 GFX(805), // gfx805 94 GFX(810), // gfx810 95 GFX(900), // gfx900 96 GFX(902), // gfx902 97 GFX(904), // gfx903 98 GFX(906), // gfx906 99 GFX(908), // gfx908 100 GFX(909), // gfx909 101 GFX(90a), // gfx90a 102 GFX(90c), // gfx90c 103 GFX(1010), // gfx1010 104 GFX(1011), // gfx1011 105 GFX(1012), // gfx1012 106 GFX(1030), // gfx1030 107 GFX(1031), // gfx1031 108 GFX(1032), // gfx1032 109 GFX(1033), // gfx1033 110 // clang-format on 111 }; 112 #undef SM 113 #undef SM2 114 #undef GFX 115 116 const char *CudaArchToString(CudaArch A) { 117 auto result = std::find_if( 118 std::begin(arch_names), std::end(arch_names), 119 [A](const CudaArchToStringMap &map) { return A == map.arch; }); 120 if (result == std::end(arch_names)) 121 return "unknown"; 122 return result->arch_name; 123 } 124 125 const char *CudaArchToVirtualArchString(CudaArch A) { 126 auto result = std::find_if( 127 std::begin(arch_names), std::end(arch_names), 128 [A](const CudaArchToStringMap &map) { return A == map.arch; }); 129 if (result == std::end(arch_names)) 130 return "unknown"; 131 return result->virtual_arch_name; 132 } 133 134 CudaArch StringToCudaArch(llvm::StringRef S) { 135 auto result = std::find_if( 136 std::begin(arch_names), std::end(arch_names), 137 [S](const CudaArchToStringMap &map) { return S == map.arch_name; }); 138 if (result == std::end(arch_names)) 139 return CudaArch::UNKNOWN; 140 return result->arch; 141 } 142 143 CudaVersion MinVersionForCudaArch(CudaArch A) { 144 if (A == CudaArch::UNKNOWN) 145 return CudaVersion::UNKNOWN; 146 147 // AMD GPUs do not depend on CUDA versions. 148 if (IsAMDGpuArch(A)) 149 return CudaVersion::CUDA_70; 150 151 switch (A) { 152 case CudaArch::SM_20: 153 case CudaArch::SM_21: 154 case CudaArch::SM_30: 155 case CudaArch::SM_32: 156 case CudaArch::SM_35: 157 case CudaArch::SM_37: 158 case CudaArch::SM_50: 159 case CudaArch::SM_52: 160 case CudaArch::SM_53: 161 return CudaVersion::CUDA_70; 162 case CudaArch::SM_60: 163 case CudaArch::SM_61: 164 case CudaArch::SM_62: 165 return CudaVersion::CUDA_80; 166 case CudaArch::SM_70: 167 return CudaVersion::CUDA_90; 168 case CudaArch::SM_72: 169 return CudaVersion::CUDA_91; 170 case CudaArch::SM_75: 171 return CudaVersion::CUDA_100; 172 case CudaArch::SM_80: 173 return CudaVersion::CUDA_110; 174 case CudaArch::SM_86: 175 return CudaVersion::CUDA_111; 176 default: 177 llvm_unreachable("invalid enum"); 178 } 179 } 180 181 CudaVersion MaxVersionForCudaArch(CudaArch A) { 182 // AMD GPUs do not depend on CUDA versions. 183 if (IsAMDGpuArch(A)) 184 return CudaVersion::LATEST; 185 186 switch (A) { 187 case CudaArch::UNKNOWN: 188 return CudaVersion::UNKNOWN; 189 case CudaArch::SM_20: 190 case CudaArch::SM_21: 191 return CudaVersion::CUDA_80; 192 default: 193 return CudaVersion::LATEST; 194 } 195 } 196 197 CudaVersion ToCudaVersion(llvm::VersionTuple Version) { 198 int IVer = 199 Version.getMajor() * 10 + Version.getMinor().getValueOr(0); 200 switch(IVer) { 201 case 70: 202 return CudaVersion::CUDA_70; 203 case 75: 204 return CudaVersion::CUDA_75; 205 case 80: 206 return CudaVersion::CUDA_80; 207 case 90: 208 return CudaVersion::CUDA_90; 209 case 91: 210 return CudaVersion::CUDA_91; 211 case 92: 212 return CudaVersion::CUDA_92; 213 case 100: 214 return CudaVersion::CUDA_100; 215 case 101: 216 return CudaVersion::CUDA_101; 217 case 102: 218 return CudaVersion::CUDA_102; 219 case 110: 220 return CudaVersion::CUDA_110; 221 case 111: 222 return CudaVersion::CUDA_111; 223 case 112: 224 return CudaVersion::CUDA_112; 225 default: 226 return CudaVersion::UNKNOWN; 227 } 228 } 229 230 bool CudaFeatureEnabled(llvm::VersionTuple Version, CudaFeature Feature) { 231 return CudaFeatureEnabled(ToCudaVersion(Version), Feature); 232 } 233 234 bool CudaFeatureEnabled(CudaVersion Version, CudaFeature Feature) { 235 switch (Feature) { 236 case CudaFeature::CUDA_USES_NEW_LAUNCH: 237 return Version >= CudaVersion::CUDA_92; 238 case CudaFeature::CUDA_USES_FATBIN_REGISTER_END: 239 return Version >= CudaVersion::CUDA_101; 240 } 241 llvm_unreachable("Unknown CUDA feature."); 242 } 243 } // namespace clang 244