| 1 |
#include "clang/Basic/Cuda.h" |
| 2 |
|
| 3 |
#include "llvm/ADT/StringRef.h" |
| 4 |
#include "llvm/ADT/Twine.h" |
| 5 |
#include "llvm/Support/ErrorHandling.h" |
| 6 |
#include "llvm/Support/VersionTuple.h" |
| 7 |
|
| 8 |
namespace clang { |
| 9 |
|
| 10 |
struct CudaVersionMapEntry { |
| 11 |
const char *Name; |
| 12 |
CudaVersion Version; |
| 13 |
llvm::VersionTuple TVersion; |
| 14 |
}; |
| 15 |
#define CUDA_ENTRY(major, minor) \ |
| 16 |
{ \ |
| 17 |
#major "." #minor, CudaVersion::CUDA_##major##minor, \ |
| 18 |
llvm::VersionTuple(major, minor) \ |
| 19 |
} |
| 20 |
|
| 21 |
static const CudaVersionMapEntry CudaNameVersionMap[] = { |
| 22 |
CUDA_ENTRY(7, 0), |
| 23 |
CUDA_ENTRY(7, 5), |
| 24 |
CUDA_ENTRY(8, 0), |
| 25 |
CUDA_ENTRY(9, 0), |
| 26 |
CUDA_ENTRY(9, 1), |
| 27 |
CUDA_ENTRY(9, 2), |
| 28 |
CUDA_ENTRY(10, 0), |
| 29 |
CUDA_ENTRY(10, 1), |
| 30 |
CUDA_ENTRY(10, 2), |
| 31 |
CUDA_ENTRY(11, 0), |
| 32 |
CUDA_ENTRY(11, 1), |
| 33 |
CUDA_ENTRY(11, 2), |
| 34 |
CUDA_ENTRY(11, 3), |
| 35 |
CUDA_ENTRY(11, 4), |
| 36 |
CUDA_ENTRY(11, 5), |
| 37 |
CUDA_ENTRY(11, 6), |
| 38 |
CUDA_ENTRY(11, 7), |
| 39 |
CUDA_ENTRY(11, 8), |
| 40 |
{"", CudaVersion::NEW, llvm::VersionTuple(std::numeric_limits<int>::max())}, |
| 41 |
{"unknown", CudaVersion::UNKNOWN, {}} // End of list tombstone. |
| 42 |
}; |
| 43 |
#undef CUDA_ENTRY |
| 44 |
|
| 45 |
const char *CudaVersionToString(CudaVersion V) { |
| 46 |
for (auto *I = CudaNameVersionMap; I->Version != CudaVersion::UNKNOWN; ++I) |
| 47 |
if (I->Version == V) |
| 48 |
return I->Name; |
| 49 |
|
| 50 |
return CudaVersionToString(CudaVersion::UNKNOWN); |
| 51 |
} |
| 52 |
|
| 53 |
CudaVersion CudaStringToVersion(const llvm::Twine &S) { |
| 54 |
std::string VS = S.str(); |
| 55 |
for (auto *I = CudaNameVersionMap; I->Version != CudaVersion::UNKNOWN; ++I) |
| 56 |
if (I->Name == VS) |
| 57 |
return I->Version; |
| 58 |
return CudaVersion::UNKNOWN; |
| 59 |
} |
| 60 |
|
| 61 |
CudaVersion ToCudaVersion(llvm::VersionTuple Version) { |
| 62 |
for (auto *I = CudaNameVersionMap; I->Version != CudaVersion::UNKNOWN; ++I) |
| 63 |
if (I->TVersion == Version) |
| 64 |
return I->Version; |
| 65 |
return CudaVersion::UNKNOWN; |
| 66 |
} |
| 67 |
|
| 68 |
namespace { |
| 69 |
struct CudaArchToStringMap { |
| 70 |
CudaArch arch; |
| 71 |
const char *arch_name; |
| 72 |
const char *virtual_arch_name; |
| 73 |
}; |
| 74 |
} // namespace |
| 75 |
|
| 76 |
#define SM2(sm, ca) \ |
| 77 |
{ CudaArch::SM_##sm, "sm_" #sm, ca } |
| 78 |
#define SM(sm) SM2(sm, "compute_" #sm) |
| 79 |
#define GFX(gpu) \ |
| 80 |
{ CudaArch::GFX##gpu, "gfx" #gpu, "compute_amdgcn" } |
| 81 |
static const CudaArchToStringMap arch_names[] = { |
| 82 |
// clang-format off |
| 83 |
{CudaArch::UNUSED, "", ""}, |
| 84 |
SM2(20, "compute_20"), SM2(21, "compute_20"), // Fermi |
| 85 |
SM(30), SM(32), SM(35), SM(37), // Kepler |
| 86 |
SM(50), SM(52), SM(53), // Maxwell |
| 87 |
SM(60), SM(61), SM(62), // Pascal |
| 88 |
SM(70), SM(72), // Volta |
| 89 |
SM(75), // Turing |
| 90 |
SM(80), SM(86), // Ampere |
| 91 |
SM(87), // Jetson/Drive AGX Orin |
| 92 |
SM(89), // Ada Lovelace |
| 93 |
SM(90), // Hopper |
| 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 |
case CudaArch::SM_87: |
| 200 |
return CudaVersion::CUDA_114; |
| 201 |
case CudaArch::SM_89: |
| 202 |
case CudaArch::SM_90: |
| 203 |
return CudaVersion::CUDA_118; |
| 204 |
default: |
| 205 |
llvm_unreachable("invalid enum"); |
| 206 |
} |
| 207 |
} |
| 208 |
|
| 209 |
CudaVersion MaxVersionForCudaArch(CudaArch A) { |
| 210 |
// AMD GPUs do not depend on CUDA versions. |
| 211 |
if (IsAMDGpuArch(A)) |
| 212 |
return CudaVersion::NEW; |
| 213 |
|
| 214 |
switch (A) { |
| 215 |
case CudaArch::UNKNOWN: |
| 216 |
return CudaVersion::UNKNOWN; |
| 217 |
case CudaArch::SM_20: |
| 218 |
case CudaArch::SM_21: |
| 219 |
return CudaVersion::CUDA_80; |
| 220 |
case CudaArch::SM_30: |
| 221 |
return CudaVersion::CUDA_110; |
| 222 |
default: |
| 223 |
return CudaVersion::NEW; |
| 224 |
} |
| 225 |
} |
| 226 |
|
| 227 |
bool CudaFeatureEnabled(llvm::VersionTuple Version, CudaFeature Feature) { |
| 228 |
return CudaFeatureEnabled(ToCudaVersion(Version), Feature); |
| 229 |
} |
| 230 |
|
| 231 |
bool CudaFeatureEnabled(CudaVersion Version, CudaFeature Feature) { |
| 232 |
switch (Feature) { |
| 233 |
case CudaFeature::CUDA_USES_NEW_LAUNCH: |
| 234 |
return Version >= CudaVersion::CUDA_92; |
| 235 |
case CudaFeature::CUDA_USES_FATBIN_REGISTER_END: |
| 236 |
return Version >= CudaVersion::CUDA_101; |
| 237 |
} |
| 238 |
llvm_unreachable("Unknown CUDA feature."); |
| 239 |
} |
| 240 |
} // namespace clang |