1 //===--- AMDGPU.cpp - AMDGPU 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 #include "AMDGPU.h" 10 #include "CommonArgs.h" 11 #include "InputInfo.h" 12 #include "clang/Driver/Compilation.h" 13 #include "clang/Driver/DriverDiagnostic.h" 14 #include "llvm/Option/ArgList.h" 15 #include "llvm/Support/Path.h" 16 #include "llvm/Support/VirtualFileSystem.h" 17 18 using namespace clang::driver; 19 using namespace clang::driver::tools; 20 using namespace clang::driver::toolchains; 21 using namespace clang; 22 using namespace llvm::opt; 23 24 RocmInstallationDetector::RocmInstallationDetector( 25 const Driver &D, const llvm::Triple &HostTriple, 26 const llvm::opt::ArgList &Args) 27 : D(D) { 28 struct Candidate { 29 std::string Path; 30 bool StrictChecking; 31 32 Candidate(std::string Path, bool StrictChecking = false) 33 : Path(Path), StrictChecking(StrictChecking) {} 34 }; 35 36 SmallVector<Candidate, 4> Candidates; 37 38 if (Args.hasArg(clang::driver::options::OPT_rocm_path_EQ)) { 39 Candidates.emplace_back( 40 Args.getLastArgValue(clang::driver::options::OPT_rocm_path_EQ).str()); 41 } else { 42 // Try to find relative to the compiler binary. 43 const char *InstallDir = D.getInstalledDir(); 44 45 // Check both a normal Unix prefix position of the clang binary, as well as 46 // the Windows-esque layout the ROCm packages use with the host architecture 47 // subdirectory of bin. 48 49 StringRef ParentDir = llvm::sys::path::parent_path(InstallDir); 50 if (ParentDir == HostTriple.getArchName()) 51 ParentDir = llvm::sys::path::parent_path(ParentDir); 52 53 if (ParentDir == "bin") { 54 Candidates.emplace_back(llvm::sys::path::parent_path(ParentDir).str(), 55 /*StrictChecking=*/true); 56 } 57 58 Candidates.emplace_back(D.SysRoot + "/opt/rocm"); 59 } 60 61 bool NoBuiltinLibs = Args.hasArg(options::OPT_nogpulib); 62 63 for (const auto &Candidate : Candidates) { 64 InstallPath = Candidate.Path; 65 if (InstallPath.empty() || !D.getVFS().exists(InstallPath)) 66 continue; 67 68 // FIXME: The install path situation is a real mess. 69 70 // For a cmake install, these are placed directly in 71 // ${INSTALL_PREFIX}/lib 72 73 // In the separate OpenCL builds, the bitcode libraries are placed in 74 // ${OPENCL_ROOT}/lib/x86_64/bitcode/* 75 76 // For the rocm installed packages, these are placed at 77 // /opt/rocm/opencl/lib/x86_64/bitcode 78 79 // An additional copy is installed, in scattered locations between 80 // /opt/rocm/hcc/rocdl/oclc 81 // /opt/rocm/hcc/rocdl/ockl 82 // /opt/rocm/hcc/rocdl/lib 83 // 84 // Yet another complete set is installed to 85 // /opt/rocm/hcc/rocdl/lib 86 87 // For now just recognize the opencl package layout. 88 89 // BinPath = InstallPath + "/bin"; 90 llvm::sys::path::append(IncludePath, InstallPath, "include"); 91 llvm::sys::path::append(LibDevicePath, InstallPath, "lib"); 92 93 auto &FS = D.getVFS(); 94 95 // We don't need the include path for OpenCL, since clang already ships with 96 // the default header. 97 98 bool CheckLibDevice = (!NoBuiltinLibs || Candidate.StrictChecking); 99 if (CheckLibDevice && !FS.exists(LibDevicePath)) 100 continue; 101 102 const StringRef Suffix(".amdgcn.bc"); 103 104 std::error_code EC; 105 for (llvm::sys::fs::directory_iterator LI(LibDevicePath, EC), LE; 106 !EC && LI != LE; LI = LI.increment(EC)) { 107 StringRef FilePath = LI->path(); 108 StringRef FileName = llvm::sys::path::filename(FilePath); 109 if (!FileName.endswith(Suffix)) 110 continue; 111 112 StringRef BaseName = FileName.drop_back(Suffix.size()); 113 114 if (BaseName == "ocml") { 115 OCML = FilePath; 116 } else if (BaseName == "ockl") { 117 OCKL = FilePath; 118 } else if (BaseName == "opencl") { 119 OpenCL = FilePath; 120 } else if (BaseName == "hip") { 121 HIP = FilePath; 122 } else if (BaseName == "oclc_finite_only_off") { 123 FiniteOnly.Off = FilePath; 124 } else if (BaseName == "oclc_finite_only_on") { 125 FiniteOnly.On = FilePath; 126 } else if (BaseName == "oclc_daz_opt_on") { 127 DenormalsAreZero.On = FilePath; 128 } else if (BaseName == "oclc_daz_opt_off") { 129 DenormalsAreZero.Off = FilePath; 130 } else if (BaseName == "oclc_correctly_rounded_sqrt_on") { 131 CorrectlyRoundedSqrt.On = FilePath; 132 } else if (BaseName == "oclc_correctly_rounded_sqrt_off") { 133 CorrectlyRoundedSqrt.Off = FilePath; 134 } else if (BaseName == "oclc_unsafe_math_on") { 135 UnsafeMath.On = FilePath; 136 } else if (BaseName == "oclc_unsafe_math_off") { 137 UnsafeMath.Off = FilePath; 138 } else if (BaseName == "oclc_wavefrontsize64_on") { 139 WavefrontSize64.On = FilePath; 140 } else if (BaseName == "oclc_wavefrontsize64_off") { 141 WavefrontSize64.Off = FilePath; 142 } else { 143 // Process all bitcode filenames that look like 144 // ocl_isa_version_XXX.amdgcn.bc 145 const StringRef DeviceLibPrefix = "oclc_isa_version_"; 146 if (!BaseName.startswith(DeviceLibPrefix)) 147 continue; 148 149 StringRef IsaVersionNumber = 150 BaseName.drop_front(DeviceLibPrefix.size()); 151 152 llvm::Twine GfxName = Twine("gfx") + IsaVersionNumber; 153 SmallString<8> Tmp; 154 LibDeviceMap.insert( 155 std::make_pair(GfxName.toStringRef(Tmp), FilePath.str())); 156 } 157 } 158 159 if (!NoBuiltinLibs) { 160 // Check that the required non-target libraries are all available. 161 if (!allGenericLibsValid()) 162 continue; 163 164 // Check that we have found at least one libdevice that we can link in if 165 // -nobuiltinlib hasn't been specified. 166 if (LibDeviceMap.empty()) 167 continue; 168 } 169 170 IsValid = true; 171 break; 172 } 173 } 174 175 void RocmInstallationDetector::print(raw_ostream &OS) const { 176 if (isValid()) 177 OS << "Found ROCm installation: " << InstallPath << '\n'; 178 } 179 180 void amdgpu::Linker::ConstructJob(Compilation &C, const JobAction &JA, 181 const InputInfo &Output, 182 const InputInfoList &Inputs, 183 const ArgList &Args, 184 const char *LinkingOutput) const { 185 186 std::string Linker = getToolChain().GetProgramPath(getShortName()); 187 ArgStringList CmdArgs; 188 AddLinkerInputs(getToolChain(), Inputs, Args, CmdArgs, JA); 189 CmdArgs.push_back("-shared"); 190 CmdArgs.push_back("-o"); 191 CmdArgs.push_back(Output.getFilename()); 192 C.addCommand(std::make_unique<Command>(JA, *this, Args.MakeArgString(Linker), 193 CmdArgs, Inputs)); 194 } 195 196 void amdgpu::getAMDGPUTargetFeatures(const Driver &D, 197 const llvm::opt::ArgList &Args, 198 std::vector<StringRef> &Features) { 199 if (const Arg *dAbi = Args.getLastArg(options::OPT_mamdgpu_debugger_abi)) 200 D.Diag(diag::err_drv_clang_unsupported) << dAbi->getAsString(Args); 201 202 if (Args.getLastArg(options::OPT_mwavefrontsize64)) { 203 Features.push_back("-wavefrontsize16"); 204 Features.push_back("-wavefrontsize32"); 205 Features.push_back("+wavefrontsize64"); 206 } 207 if (Args.getLastArg(options::OPT_mno_wavefrontsize64)) { 208 Features.push_back("-wavefrontsize16"); 209 Features.push_back("+wavefrontsize32"); 210 Features.push_back("-wavefrontsize64"); 211 } 212 213 handleTargetFeaturesGroup( 214 Args, Features, options::OPT_m_amdgpu_Features_Group); 215 } 216 217 /// AMDGPU Toolchain 218 AMDGPUToolChain::AMDGPUToolChain(const Driver &D, const llvm::Triple &Triple, 219 const ArgList &Args) 220 : Generic_ELF(D, Triple, Args), 221 OptionsDefault({{options::OPT_O, "3"}, 222 {options::OPT_cl_std_EQ, "CL1.2"}}) {} 223 224 Tool *AMDGPUToolChain::buildLinker() const { 225 return new tools::amdgpu::Linker(*this); 226 } 227 228 DerivedArgList * 229 AMDGPUToolChain::TranslateArgs(const DerivedArgList &Args, StringRef BoundArch, 230 Action::OffloadKind DeviceOffloadKind) const { 231 232 DerivedArgList *DAL = 233 Generic_ELF::TranslateArgs(Args, BoundArch, DeviceOffloadKind); 234 235 // Do nothing if not OpenCL (-x cl) 236 if (!Args.getLastArgValue(options::OPT_x).equals("cl")) 237 return DAL; 238 239 if (!DAL) 240 DAL = new DerivedArgList(Args.getBaseArgs()); 241 for (auto *A : Args) 242 DAL->append(A); 243 244 const OptTable &Opts = getDriver().getOpts(); 245 246 // Phase 1 (.cl -> .bc) 247 if (Args.hasArg(options::OPT_c) && Args.hasArg(options::OPT_emit_llvm)) { 248 DAL->AddFlagArg(nullptr, Opts.getOption(getTriple().isArch64Bit() 249 ? options::OPT_m64 250 : options::OPT_m32)); 251 252 // Have to check OPT_O4, OPT_O0 & OPT_Ofast separately 253 // as they defined that way in Options.td 254 if (!Args.hasArg(options::OPT_O, options::OPT_O0, options::OPT_O4, 255 options::OPT_Ofast)) 256 DAL->AddJoinedArg(nullptr, Opts.getOption(options::OPT_O), 257 getOptionDefault(options::OPT_O)); 258 } 259 260 return DAL; 261 } 262 263 bool AMDGPUToolChain::getDefaultDenormsAreZeroForTarget( 264 llvm::AMDGPU::GPUKind Kind) { 265 266 // Assume nothing without a specific target. 267 if (Kind == llvm::AMDGPU::GK_NONE) 268 return false; 269 270 const unsigned ArchAttr = llvm::AMDGPU::getArchAttrAMDGCN(Kind); 271 272 // Default to enabling f32 denormals by default on subtargets where fma is 273 // fast with denormals 274 const bool BothDenormAndFMAFast = 275 (ArchAttr & llvm::AMDGPU::FEATURE_FAST_FMA_F32) && 276 (ArchAttr & llvm::AMDGPU::FEATURE_FAST_DENORMAL_F32); 277 return !BothDenormAndFMAFast; 278 } 279 280 llvm::DenormalMode AMDGPUToolChain::getDefaultDenormalModeForType( 281 const llvm::opt::ArgList &DriverArgs, const JobAction &JA, 282 const llvm::fltSemantics *FPType) const { 283 // Denormals should always be enabled for f16 and f64. 284 if (!FPType || FPType != &llvm::APFloat::IEEEsingle()) 285 return llvm::DenormalMode::getIEEE(); 286 287 if (JA.getOffloadingDeviceKind() == Action::OFK_HIP || 288 JA.getOffloadingDeviceKind() == Action::OFK_Cuda) { 289 auto Kind = llvm::AMDGPU::parseArchAMDGCN(JA.getOffloadingArch()); 290 if (FPType && FPType == &llvm::APFloat::IEEEsingle() && 291 DriverArgs.hasFlag(options::OPT_fcuda_flush_denormals_to_zero, 292 options::OPT_fno_cuda_flush_denormals_to_zero, 293 getDefaultDenormsAreZeroForTarget(Kind))) 294 return llvm::DenormalMode::getPreserveSign(); 295 296 return llvm::DenormalMode::getIEEE(); 297 } 298 299 const StringRef GpuArch = DriverArgs.getLastArgValue(options::OPT_mcpu_EQ); 300 auto Kind = llvm::AMDGPU::parseArchAMDGCN(GpuArch); 301 302 // TODO: There are way too many flags that change this. Do we need to check 303 // them all? 304 bool DAZ = DriverArgs.hasArg(options::OPT_cl_denorms_are_zero) || 305 getDefaultDenormsAreZeroForTarget(Kind); 306 307 // Outputs are flushed to zero (FTZ), preserving sign. Denormal inputs are 308 // also implicit treated as zero (DAZ). 309 return DAZ ? llvm::DenormalMode::getPreserveSign() : 310 llvm::DenormalMode::getIEEE(); 311 } 312 313 /// ROCM Toolchain 314 ROCMToolChain::ROCMToolChain(const Driver &D, const llvm::Triple &Triple, 315 const ArgList &Args) 316 : AMDGPUToolChain(D, Triple, Args), 317 RocmInstallation(D, Triple, Args) { } 318 319 void AMDGPUToolChain::addClangTargetOptions( 320 const llvm::opt::ArgList &DriverArgs, 321 llvm::opt::ArgStringList &CC1Args, 322 Action::OffloadKind DeviceOffloadingKind) const { 323 // Default to "hidden" visibility, as object level linking will not be 324 // supported for the foreseeable future. 325 if (!DriverArgs.hasArg(options::OPT_fvisibility_EQ, 326 options::OPT_fvisibility_ms_compat)) { 327 CC1Args.push_back("-fvisibility"); 328 CC1Args.push_back("hidden"); 329 CC1Args.push_back("-fapply-global-visibility-to-externs"); 330 } 331 } 332 333 void ROCMToolChain::addClangTargetOptions( 334 const llvm::opt::ArgList &DriverArgs, llvm::opt::ArgStringList &CC1Args, 335 Action::OffloadKind DeviceOffloadingKind) const { 336 AMDGPUToolChain::addClangTargetOptions(DriverArgs, CC1Args, 337 DeviceOffloadingKind); 338 339 if (DriverArgs.hasArg(options::OPT_nogpulib)) 340 return; 341 342 if (!RocmInstallation.isValid()) { 343 getDriver().Diag(diag::err_drv_no_rocm_installation); 344 return; 345 } 346 347 // Get the device name and canonicalize it 348 const StringRef GpuArch = DriverArgs.getLastArgValue(options::OPT_mcpu_EQ); 349 auto Kind = llvm::AMDGPU::parseArchAMDGCN(GpuArch); 350 const StringRef CanonArch = llvm::AMDGPU::getArchNameAMDGCN(Kind); 351 std::string LibDeviceFile = RocmInstallation.getLibDeviceFile(CanonArch); 352 if (LibDeviceFile.empty()) { 353 getDriver().Diag(diag::err_drv_no_rocm_device_lib) << GpuArch; 354 return; 355 } 356 357 const unsigned ArchAttr = llvm::AMDGPU::getArchAttrAMDGCN(Kind); 358 static bool HasWave32 = (ArchAttr & llvm::AMDGPU::FEATURE_WAVE32); 359 360 bool Wave64 = !HasWave32 || DriverArgs.hasFlag( 361 options::OPT_mwavefrontsize64, options::OPT_mno_wavefrontsize64, false); 362 363 // TODO: There are way too many flags that change this. Do we need to check 364 // them all? 365 bool DAZ = DriverArgs.hasArg(options::OPT_cl_denorms_are_zero) || 366 getDefaultDenormsAreZeroForTarget(Kind); 367 bool FiniteOnly = DriverArgs.hasArg(options::OPT_cl_finite_math_only); 368 369 bool UnsafeMathOpt = 370 DriverArgs.hasArg(options::OPT_cl_unsafe_math_optimizations); 371 bool FastRelaxedMath = DriverArgs.hasArg(options::OPT_cl_fast_relaxed_math); 372 bool CorrectSqrt = 373 DriverArgs.hasArg(options::OPT_cl_fp32_correctly_rounded_divide_sqrt); 374 375 // Add the OpenCL specific bitcode library. 376 CC1Args.push_back("-mlink-builtin-bitcode"); 377 CC1Args.push_back(DriverArgs.MakeArgString(RocmInstallation.getOpenCLPath())); 378 379 // Add the generic set of libraries. 380 RocmInstallation.addCommonBitcodeLibCC1Args( 381 DriverArgs, CC1Args, LibDeviceFile, Wave64, DAZ, FiniteOnly, 382 UnsafeMathOpt, FastRelaxedMath, CorrectSqrt); 383 } 384 385 void RocmInstallationDetector::addCommonBitcodeLibCC1Args( 386 const llvm::opt::ArgList &DriverArgs, llvm::opt::ArgStringList &CC1Args, 387 StringRef LibDeviceFile, bool Wave64, bool DAZ, bool FiniteOnly, 388 bool UnsafeMathOpt, bool FastRelaxedMath, bool CorrectSqrt) const { 389 static const char LinkBitcodeFlag[] = "-mlink-builtin-bitcode"; 390 391 CC1Args.push_back(LinkBitcodeFlag); 392 CC1Args.push_back(DriverArgs.MakeArgString(getOCMLPath())); 393 394 CC1Args.push_back(LinkBitcodeFlag); 395 CC1Args.push_back(DriverArgs.MakeArgString(getOCKLPath())); 396 397 CC1Args.push_back(LinkBitcodeFlag); 398 CC1Args.push_back(DriverArgs.MakeArgString(getDenormalsAreZeroPath(DAZ))); 399 400 CC1Args.push_back(LinkBitcodeFlag); 401 CC1Args.push_back(DriverArgs.MakeArgString( 402 getUnsafeMathPath(UnsafeMathOpt || FastRelaxedMath))); 403 404 CC1Args.push_back(LinkBitcodeFlag); 405 CC1Args.push_back(DriverArgs.MakeArgString( 406 getFiniteOnlyPath(FiniteOnly || FastRelaxedMath))); 407 408 CC1Args.push_back(LinkBitcodeFlag); 409 CC1Args.push_back( 410 DriverArgs.MakeArgString(getCorrectlyRoundedSqrtPath(CorrectSqrt))); 411 412 CC1Args.push_back(LinkBitcodeFlag); 413 CC1Args.push_back(DriverArgs.MakeArgString(getWavefrontSize64Path(Wave64))); 414 415 CC1Args.push_back(LinkBitcodeFlag); 416 CC1Args.push_back(DriverArgs.MakeArgString(LibDeviceFile)); 417 } 418