| 1 |
//===--- Cuda.h - Cuda ToolChain Implementations ----------------*- C++ -*-===// |
| 2 |
// |
| 3 |
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 |
// See https://llvm.org/LICENSE.txt for license information. |
| 5 |
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 |
// |
| 7 |
//===----------------------------------------------------------------------===// |
| 8 |
|
| 9 |
#ifndef LLVM_CLANG_LIB_DRIVER_TOOLCHAINS_CUDA_H |
| 10 |
#define LLVM_CLANG_LIB_DRIVER_TOOLCHAINS_CUDA_H |
| 11 |
|
| 12 |
#include "clang/Basic/Cuda.h" |
| 13 |
#include "clang/Driver/Action.h" |
| 14 |
#include "clang/Driver/Multilib.h" |
| 15 |
#include "clang/Driver/Tool.h" |
| 16 |
#include "clang/Driver/ToolChain.h" |
| 17 |
#include "llvm/Support/Compiler.h" |
| 18 |
#include "llvm/Support/VersionTuple.h" |
| 19 |
#include <bitset> |
| 20 |
#include <set> |
| 21 |
#include <vector> |
| 22 |
|
| 23 |
namespace clang { |
| 24 |
namespace driver { |
| 25 |
|
| 26 |
/// A class to find a viable CUDA installation |
| 27 |
class CudaInstallationDetector { |
| 28 |
private: |
| 29 |
const Driver &D; |
| 30 |
bool IsValid = false; |
| 31 |
CudaVersion Version = CudaVersion::UNKNOWN; |
| 32 |
std::string InstallPath; |
| 33 |
std::string BinPath; |
| 34 |
std::string LibDevicePath; |
| 35 |
std::string IncludePath; |
| 36 |
llvm::StringMap<std::string> LibDeviceMap; |
| 37 |
|
| 38 |
// CUDA architectures for which we have raised an error in |
| 39 |
// CheckCudaVersionSupportsArch. |
| 40 |
mutable std::bitset<(int)CudaArch::LAST> ArchsWithBadVersion; |
| 41 |
|
| 42 |
public: |
| 43 |
CudaInstallationDetector(const Driver &D, const llvm::Triple &HostTriple, |
| 44 |
const llvm::opt::ArgList &Args); |
| 45 |
|
| 46 |
void AddCudaIncludeArgs(const llvm::opt::ArgList &DriverArgs, |
| 47 |
llvm::opt::ArgStringList &CC1Args) const; |
| 48 |
|
| 49 |
/// Emit an error if Version does not support the given Arch. |
| 50 |
/// |
| 51 |
/// If either Version or Arch is unknown, does not emit an error. Emits at |
| 52 |
/// most one error per Arch. |
| 53 |
void CheckCudaVersionSupportsArch(CudaArch Arch) const; |
| 54 |
|
| 55 |
/// Check whether we detected a valid Cuda install. |
| 56 |
bool isValid() const { return IsValid; } |
| 57 |
/// Print information about the detected CUDA installation. |
| 58 |
void print(raw_ostream &OS) const; |
| 59 |
|
| 60 |
/// Get the detected Cuda install's version. |
| 61 |
CudaVersion version() const { |
| 62 |
return Version == CudaVersion::NEW ? CudaVersion::PARTIALLY_SUPPORTED |
| 63 |
: Version; |
| 64 |
} |
| 65 |
/// Get the detected Cuda installation path. |
| 66 |
StringRef getInstallPath() const { return InstallPath; } |
| 67 |
/// Get the detected path to Cuda's bin directory. |
| 68 |
StringRef getBinPath() const { return BinPath; } |
| 69 |
/// Get the detected Cuda Include path. |
| 70 |
StringRef getIncludePath() const { return IncludePath; } |
| 71 |
/// Get the detected Cuda device library path. |
| 72 |
StringRef getLibDevicePath() const { return LibDevicePath; } |
| 73 |
/// Get libdevice file for given architecture |
| 74 |
std::string getLibDeviceFile(StringRef Gpu) const { |
| 75 |
return LibDeviceMap.lookup(Gpu); |
| 76 |
} |
| 77 |
void WarnIfUnsupportedVersion(); |
| 78 |
}; |
| 79 |
|
| 80 |
namespace tools { |
| 81 |
namespace NVPTX { |
| 82 |
|
| 83 |
// Run ptxas, the NVPTX assembler. |
| 84 |
class LLVM_LIBRARY_VISIBILITY Assembler : public Tool { |
| 85 |
public: |
| 86 |
Assembler(const ToolChain &TC) : Tool("NVPTX::Assembler", "ptxas", TC) {} |
| 87 |
|
| 88 |
bool hasIntegratedCPP() const override { return false; } |
| 89 |
|
| 90 |
void ConstructJob(Compilation &C, const JobAction &JA, |
| 91 |
const InputInfo &Output, const InputInfoList &Inputs, |
| 92 |
const llvm::opt::ArgList &TCArgs, |
| 93 |
const char *LinkingOutput) const override; |
| 94 |
}; |
| 95 |
|
| 96 |
// Runs fatbinary, which combines GPU object files ("cubin" files) and/or PTX |
| 97 |
// assembly into a single output file. |
| 98 |
class LLVM_LIBRARY_VISIBILITY FatBinary : public Tool { |
| 99 |
public: |
| 100 |
FatBinary(const ToolChain &TC) : Tool("NVPTX::Linker", "fatbinary", TC) {} |
| 101 |
|
| 102 |
bool hasIntegratedCPP() const override { return false; } |
| 103 |
|
| 104 |
void ConstructJob(Compilation &C, const JobAction &JA, |
| 105 |
const InputInfo &Output, const InputInfoList &Inputs, |
| 106 |
const llvm::opt::ArgList &TCArgs, |
| 107 |
const char *LinkingOutput) const override; |
| 108 |
}; |
| 109 |
|
| 110 |
// Runs nvlink, which links GPU object files ("cubin" files) into a single file. |
| 111 |
class LLVM_LIBRARY_VISIBILITY Linker : public Tool { |
| 112 |
public: |
| 113 |
Linker(const ToolChain &TC) : Tool("NVPTX::Linker", "fatbinary", TC) {} |
| 114 |
|
| 115 |
bool hasIntegratedCPP() const override { return false; } |
| 116 |
|
| 117 |
void ConstructJob(Compilation &C, const JobAction &JA, |
| 118 |
const InputInfo &Output, const InputInfoList &Inputs, |
| 119 |
const llvm::opt::ArgList &TCArgs, |
| 120 |
const char *LinkingOutput) const override; |
| 121 |
}; |
| 122 |
|
| 123 |
void getNVPTXTargetFeatures(const Driver &D, const llvm::Triple &Triple, |
| 124 |
const llvm::opt::ArgList &Args, |
| 125 |
std::vector<StringRef> &Features); |
| 126 |
|
| 127 |
} // end namespace NVPTX |
| 128 |
} // end namespace tools |
| 129 |
|
| 130 |
namespace toolchains { |
| 131 |
|
| 132 |
class LLVM_LIBRARY_VISIBILITY NVPTXToolChain : public ToolChain { |
| 133 |
public: |
| 134 |
NVPTXToolChain(const Driver &D, const llvm::Triple &Triple, |
| 135 |
const llvm::Triple &HostTriple, |
| 136 |
const llvm::opt::ArgList &Args); |
| 137 |
|
| 138 |
NVPTXToolChain(const Driver &D, const llvm::Triple &Triple, |
| 139 |
const llvm::opt::ArgList &Args); |
| 140 |
|
| 141 |
llvm::opt::DerivedArgList * |
| 142 |
TranslateArgs(const llvm::opt::DerivedArgList &Args, StringRef BoundArch, |
| 143 |
Action::OffloadKind DeviceOffloadKind) const override; |
| 144 |
|
| 145 |
// Never try to use the integrated assembler with CUDA; always fork out to |
| 146 |
// ptxas. |
| 147 |
bool useIntegratedAs() const override { return false; } |
| 148 |
bool isCrossCompiling() const override { return true; } |
| 149 |
bool isPICDefault() const override { return false; } |
| 150 |
bool isPIEDefault(const llvm::opt::ArgList &Args) const override { |
| 151 |
return false; |
| 152 |
} |
| 153 |
bool isPICDefaultForced() const override { return false; } |
| 154 |
bool SupportsProfiling() const override { return false; } |
| 155 |
|
| 156 |
bool IsMathErrnoDefault() const override { return false; } |
| 157 |
|
| 158 |
bool supportsDebugInfoOption(const llvm::opt::Arg *A) const override; |
| 159 |
void adjustDebugInfoKind(codegenoptions::DebugInfoKind &DebugInfoKind, |
| 160 |
const llvm::opt::ArgList &Args) const override; |
| 161 |
|
| 162 |
// NVPTX supports only DWARF2. |
| 163 |
unsigned GetDefaultDwarfVersion() const override { return 2; } |
| 164 |
unsigned getMaxDwarfVersion() const override { return 2; } |
| 165 |
|
| 166 |
CudaInstallationDetector CudaInstallation; |
| 167 |
|
| 168 |
protected: |
| 169 |
Tool *buildAssembler() const override; // ptxas. |
| 170 |
Tool *buildLinker() const override; // nvlink. |
| 171 |
}; |
| 172 |
|
| 173 |
class LLVM_LIBRARY_VISIBILITY CudaToolChain : public NVPTXToolChain { |
| 174 |
public: |
| 175 |
CudaToolChain(const Driver &D, const llvm::Triple &Triple, |
| 176 |
const ToolChain &HostTC, const llvm::opt::ArgList &Args); |
| 177 |
|
| 178 |
const llvm::Triple *getAuxTriple() const override { |
| 179 |
return &HostTC.getTriple(); |
| 180 |
} |
| 181 |
|
| 182 |
std::string getInputFilename(const InputInfo &Input) const override; |
| 183 |
|
| 184 |
llvm::opt::DerivedArgList * |
| 185 |
TranslateArgs(const llvm::opt::DerivedArgList &Args, StringRef BoundArch, |
| 186 |
Action::OffloadKind DeviceOffloadKind) const override; |
| 187 |
void |
| 188 |
addClangTargetOptions(const llvm::opt::ArgList &DriverArgs, |
| 189 |
llvm::opt::ArgStringList &CC1Args, |
| 190 |
Action::OffloadKind DeviceOffloadKind) const override; |
| 191 |
|
| 192 |
llvm::DenormalMode getDefaultDenormalModeForType( |
| 193 |
const llvm::opt::ArgList &DriverArgs, const JobAction &JA, |
| 194 |
const llvm::fltSemantics *FPType = nullptr) const override; |
| 195 |
|
| 196 |
void AddCudaIncludeArgs(const llvm::opt::ArgList &DriverArgs, |
| 197 |
llvm::opt::ArgStringList &CC1Args) const override; |
| 198 |
|
| 199 |
void addClangWarningOptions(llvm::opt::ArgStringList &CC1Args) const override; |
| 200 |
CXXStdlibType GetCXXStdlibType(const llvm::opt::ArgList &Args) const override; |
| 201 |
void |
| 202 |
AddClangSystemIncludeArgs(const llvm::opt::ArgList &DriverArgs, |
| 203 |
llvm::opt::ArgStringList &CC1Args) const override; |
| 204 |
void AddClangCXXStdlibIncludeArgs( |
| 205 |
const llvm::opt::ArgList &Args, |
| 206 |
llvm::opt::ArgStringList &CC1Args) const override; |
| 207 |
void AddIAMCUIncludeArgs(const llvm::opt::ArgList &DriverArgs, |
| 208 |
llvm::opt::ArgStringList &CC1Args) const override; |
| 209 |
|
| 210 |
SanitizerMask getSupportedSanitizers() const override; |
| 211 |
|
| 212 |
VersionTuple |
| 213 |
computeMSVCVersion(const Driver *D, |
| 214 |
const llvm::opt::ArgList &Args) const override; |
| 215 |
|
| 216 |
const ToolChain &HostTC; |
| 217 |
|
| 218 |
/// Uses nvptx-arch tool to get arch of the system GPU. Will return error |
| 219 |
/// if unable to find one. |
| 220 |
virtual Expected<SmallVector<std::string>> |
| 221 |
getSystemGPUArchs(const llvm::opt::ArgList &Args) const override; |
| 222 |
|
| 223 |
protected: |
| 224 |
Tool *buildAssembler() const override; // ptxas |
| 225 |
Tool *buildLinker() const override; // fatbinary (ok, not really a linker) |
| 226 |
}; |
| 227 |
|
| 228 |
} // end namespace toolchains |
| 229 |
} // end namespace driver |
| 230 |
} // end namespace clang |
| 231 |
|
| 232 |
#endif // LLVM_CLANG_LIB_DRIVER_TOOLCHAINS_CUDA_H |