CoreLLVMHIP.h

Go to the documentation of this file.

#ifndef PROTEUS_CORE_LLVM_HIP_H
#define PROTEUS_CORE_LLVM_HIP_H
 
#include "proteus/Error.h"
#include "proteus/TimeTracing.h"
#include "proteus/impl/CoreLLVM.h"
#include "proteus/impl/Debug.h"
#include "proteus/impl/Logger.h"
#include "proteus/impl/Utils.h"
#include "proteus/impl/UtilsHIP.h"
 
#include <llvm/Bitcode/BitcodeWriter.h>
#include <llvm/CodeGen/MachineModuleInfo.h>
#include <llvm/IR/DiagnosticPrinter.h>
#include <llvm/IR/Function.h>
#include <llvm/IR/LegacyPassManager.h>
#include <llvm/IR/Module.h>
#include <llvm/IR/Verifier.h>
#include <llvm/LTO/LTO.h>
#include <llvm/MC/MCSubtargetInfo.h>
#include <llvm/Support/CodeGen.h>
#include <llvm/Support/FileSystem.h>
#include <llvm/Support/MemoryBuffer.h>
#include <llvm/Support/Path.h>
#include <llvm/Support/Signals.h>
#include <llvm/Support/TargetSelect.h>
#include <llvm/Support/WithColor.h>
#include <llvm/Target/TargetMachine.h>
 
#include <optional>
 
#if LLVM_VERSION_MAJOR >= 18
#include <lld/Common/Driver.h>
LLD_HAS_DRIVER(elf)
#endif
 
namespace proteus {
 
using namespace llvm;
 
namespace detail {
 
inline const SmallVector<StringRef> &gridDimXFnName() {
  static SmallVector<StringRef> Names = {
      "_ZNK17__HIP_CoordinatesI13__HIP_GridDimE3__XcvjEv",
      "llvm.amdgcn.num.workgroups.x", "_ZL20__hip_get_grid_dim_xv"};
  return Names;
}
 
inline const SmallVector<StringRef> &gridDimYFnName() {
  static SmallVector<StringRef> Names = {
      "_ZNK17__HIP_CoordinatesI13__HIP_GridDimE3__YcvjEv",
      "llvm.amdgcn.num.workgroups.y", "_ZL20__hip_get_grid_dim_yv"};
  return Names;
}
 
inline const SmallVector<StringRef> &gridDimZFnName() {
  static SmallVector<StringRef> Names = {
      "_ZNK17__HIP_CoordinatesI13__HIP_GridDimE3__ZcvjEv",
      "llvm.amdgcn.num.workgroups.z", "_ZL20__hip_get_grid_dim_zv"};
  return Names;
}
 
inline const SmallVector<StringRef> &blockDimXFnName() {
  static SmallVector<StringRef> Names = {
      "_ZNK17__HIP_CoordinatesI14__HIP_BlockDimE3__XcvjEv",
      "llvm.amdgcn.workgroup.size.x", "_ZL21__hip_get_block_dim_xv"};
  return Names;
}
 
inline const SmallVector<StringRef> &blockDimYFnName() {
  static SmallVector<StringRef> Names = {
      "_ZNK17__HIP_CoordinatesI14__HIP_BlockDimE3__YcvjEv",
      "llvm.amdgcn.workgroup.size.y", "_ZL21__hip_get_block_dim_yv"};
  return Names;
}
 
inline const SmallVector<StringRef> &blockDimZFnName() {
  static SmallVector<StringRef> Names = {
      "_ZNK17__HIP_CoordinatesI14__HIP_BlockDimE3__ZcvjEv",
      "llvm.amdgcn.workgroup.size.z", "_ZL21__hip_get_block_dim_zv"};
  return Names;
}
 
inline const SmallVector<StringRef> &blockIdxXFnName() {
  static SmallVector<StringRef> Names = {
      "_ZNK17__HIP_CoordinatesI14__HIP_BlockIdxE3__XcvjEv",
      "llvm.amdgcn.workgroup.id.x"};
  return Names;
};
 
inline const SmallVector<StringRef> &blockIdxYFnName() {
  static SmallVector<StringRef> Names = {
      "_ZNK17__HIP_CoordinatesI14__HIP_BlockIdxE3__YcvjEv",
      "llvm.amdgcn.workgroup.id.y"};
  return Names;
};
 
inline const SmallVector<StringRef> &blockIdxZFnName() {
  static SmallVector<StringRef> Names = {
      "_ZNK17__HIP_CoordinatesI14__HIP_BlockIdxE3__ZcvjEv",
      "llvm.amdgcn.workgroup.id.z"};
  return Names;
}
 
inline const SmallVector<StringRef> &threadIdxXFnName() {
  static SmallVector<StringRef> Names = {
      "_ZNK17__HIP_CoordinatesI15__HIP_ThreadIdxE3__XcvjEv",
      "llvm.amdgcn.workitem.id.x"};
  return Names;
};
 
inline const SmallVector<StringRef> &threadIdxYFnName() {
  static SmallVector<StringRef> Names = {
      "_ZNK17__HIP_CoordinatesI15__HIP_ThreadIdxE3__YcvjEv",
      "llvm.amdgcn.workitem.id.y"};
  return Names;
};
 
inline const SmallVector<StringRef> &threadIdxZFnName() {
  static SmallVector<StringRef> Names = {
      "_ZNK17__HIP_CoordinatesI15__HIP_ThreadIdxE3__ZcvjEv",
      "llvm.amdgcn.workitem.id.z"};
  return Names;
};
 
inline Expected<sys::fs::TempFile> createTempFile(StringRef Prefix,
                                                  StringRef Suffix) {
  SmallString<128> TmpDir;
  sys::path::system_temp_directory(true, TmpDir);
 
  SmallString<64> FileName;
  FileName.append(Prefix);
  FileName.append(Suffix.empty() ? "-%%%%%%%" : "-%%%%%%%.");
  FileName.append(Suffix);
  sys::path::append(TmpDir, FileName);
  return sys::fs::TempFile::create(TmpDir);
}
 
#if LLVM_VERSION_MAJOR >= 18
inline SmallVector<std::unique_ptr<sys::fs::TempFile>>
codegenSerial(Module &M, StringRef DeviceArch,
              [[maybe_unused]] char OptLevel = '3', int CodegenOptLevel = 3) {
  TIMESCOPE("proteus::codegenSerial");
  SmallVector<std::unique_ptr<sys::fs::TempFile>> ObjectFiles;
 
  auto ExpectedTM =
      proteus::detail::createTargetMachine(M, DeviceArch, CodegenOptLevel);
  if (!ExpectedTM)
    reportFatalError(toString(ExpectedTM.takeError()));
 
  std::unique_ptr<TargetMachine> TM = std::move(*ExpectedTM);
  TargetLibraryInfoImpl TLII(Triple(M.getTargetTriple()));
  M.setDataLayout(TM->createDataLayout());
 
  legacy::PassManager PM;
  PM.add(new TargetLibraryInfoWrapperPass(TLII));
  MachineModuleInfoWrapperPass *MMIWP =
#if LLVM_VERSION_MAJOR >= 20
      new MachineModuleInfoWrapperPass(TM.get());
#else
      new MachineModuleInfoWrapperPass(
          reinterpret_cast<LLVMTargetMachine *>(TM.get()));
#endif
 
  SmallVector<char, 4096> ObjectCode;
  raw_svector_ostream OS(ObjectCode);
  auto ExpectedF = createTempFile("object", "o");
  if (auto E = ExpectedF.takeError())
    reportFatalError("Error creating object tmp file " +
                     toString(std::move(E)));
  auto ObjectFile = std::move(*ExpectedF);
  auto FileStream = std::make_unique<CachedFileStream>(
      std::make_unique<llvm::raw_fd_ostream>(ObjectFile.FD, false));
  TM->addPassesToEmitFile(PM, *FileStream->OS, nullptr,
                          CodeGenFileType::ObjectFile,
                          /* DisableVerify */ true, MMIWP);
 
  std::unique_ptr<sys::fs::TempFile> ObjectFilePtr =
      std::make_unique<sys::fs::TempFile>(std::move(ObjectFile));
  ObjectFiles.emplace_back(std::move(ObjectFilePtr));
 
  PM.run(M);
  if (Error E = FileStream->commit())
    reportFatalError("Error committing object tmp file stream " +
                     toString(std::move(E)));
 
  return ObjectFiles;
}
 
inline SmallVector<std::unique_ptr<sys::fs::TempFile>>
codegenParallel(Module &M, StringRef DeviceArch,
                const OptimizationPipelineConfig &OptConfig =
                    OptimizationPipelineConfig(std::nullopt, '3', 3)) {
  TIMESCOPE("proteus::codegenParallel");
  // Use regular LTO with parallelism enabled to parallelize codegen.
  std::atomic<bool> LTOError = false;
 
  auto DiagnosticHandler = [&](const DiagnosticInfo &DI) {
    std::string ErrStorage;
    raw_string_ostream OS(ErrStorage);
    DiagnosticPrinterRawOStream DP(OS);
    DI.print(DP);
 
    switch (DI.getSeverity()) {
    case DS_Error:
      WithColor::error(errs(), "[proteus codegen]") << ErrStorage << "\n";
      LTOError = true;
      break;
    case DS_Warning:
      WithColor::warning(errs(), "[proteus codegen]") << ErrStorage << "\n";
      break;
    case DS_Note:
      WithColor::note(errs(), "[proteus codegen]") << ErrStorage << "\n";
      break;
    case DS_Remark:
      WithColor::remark(errs()) << ErrStorage << "\n";
      break;
    }
  };
 
  // Create TargetMachine and extract options/features.
  auto ExpectedTM = proteus::detail::createTargetMachine(
      M, DeviceArch, OptConfig.CodegenOptLevel);
  if (!ExpectedTM)
    reportFatalError(toString(ExpectedTM.takeError()));
  std::unique_ptr<TargetMachine> TM = std::move(*ExpectedTM);
 
  lto::Config Conf;
  Conf.CPU = DeviceArch;
 
  // Propagate attributes from TargetMachine to LTO Config.
  std::string FeatureStr = TM->getMCSubtargetInfo()->getFeatureString().str();
  if (!FeatureStr.empty()) {
    SmallVector<StringRef> Features;
    StringRef(FeatureStr).split(Features, ',');
    for (auto &F : Features)
      Conf.MAttrs.push_back(F.str());
  } else {
    Conf.MAttrs = {};
  }
 
  // FIX: Propagate TargetOptions (e.g. UnsafeFPMath, etc.)
  Conf.Options = TM->Options;
 
  Conf.DisableVerify = true;
  Conf.TimeTraceEnabled = false;
  Conf.DebugPassManager = false;
  Conf.VerifyEach = false;
  Conf.DiagHandler = DiagnosticHandler;
  Conf.OptLevel = OptConfig.OptLevel;
  // Parallel codegen lets LTO own optimization, so custom textual pipelines
  // must be forwarded to the LTO configuration instead of run beforehand.
  if (OptConfig.PassPipeline)
    Conf.OptPipeline = OptConfig.PassPipeline.value();
  Conf.CGOptLevel = static_cast<CodeGenOptLevel>(OptConfig.CodegenOptLevel);
 
  unsigned ParallelCodeGenParallelismLevel =
      std::max(1u, std::thread::hardware_concurrency());
  lto::LTO L(std::move(Conf), {}, ParallelCodeGenParallelismLevel);
 
  // Ensure module has the correct DataLayout prior to emitting bitcode.
  M.setDataLayout(TM->createDataLayout());
 
  SmallString<0> BitcodeBuf;
  raw_svector_ostream BitcodeOS(BitcodeBuf);
  WriteBitcodeToFile(M, BitcodeOS);
 
  // TODO: Module identifier can be empty because you always have on module to
  // link. However, in the general case, with multiple modules, each one must
  // have a unique identifier for LTO to work correctly.
  auto IF = cantFail(lto::InputFile::create(
      MemoryBufferRef{BitcodeBuf, M.getModuleIdentifier()}));
 
  std::set<std::string> PrevailingSymbols;
  auto BuildResolutions = [&]() {
    // Save the input file and the buffer associated with its memory.
    const auto Symbols = IF->symbols();
    SmallVector<lto::SymbolResolution, 16> Resolutions(Symbols.size());
    size_t SymbolIdx = 0;
    for (auto &Sym : Symbols) {
      lto::SymbolResolution &Res = Resolutions[SymbolIdx];
      SymbolIdx++;
 
      // All defined symbols are prevailing.
      Res.Prevailing = !Sym.isUndefined() &&
                       PrevailingSymbols.insert(Sym.getName().str()).second;
 
      Res.VisibleToRegularObj =
          Res.Prevailing &&
          Sym.getVisibility() != GlobalValue::HiddenVisibility &&
          !Sym.canBeOmittedFromSymbolTable();
 
      Res.ExportDynamic =
          Sym.getVisibility() != GlobalValue::HiddenVisibility &&
          (!Sym.canBeOmittedFromSymbolTable());
 
      Res.FinalDefinitionInLinkageUnit =
          Sym.getVisibility() != GlobalValue::DefaultVisibility &&
          (!Sym.isUndefined() && !Sym.isCommon());
 
      // Device linking does not support linker redefined symbols (e.g.
      // --wrap).
      Res.LinkerRedefined = false;
 
      if (Config::get().ProteusDebugOutput) {
        auto PrintSymbol = [](const lto::InputFile::Symbol &Sym,
                              lto::SymbolResolution &Res) {
          auto &OutStream = Logger::logs("proteus");
          OutStream << "Vis: ";
          switch (Sym.getVisibility()) {
          case GlobalValue::HiddenVisibility:
            OutStream << 'H';
            break;
          case GlobalValue::ProtectedVisibility:
            OutStream << 'P';
            break;
          case GlobalValue::DefaultVisibility:
            OutStream << 'D';
            break;
          }
 
          OutStream << " Sym: ";
          auto PrintBool = [&](char C, bool B) { OutStream << (B ? C : '-'); };
          PrintBool('U', Sym.isUndefined());
          PrintBool('C', Sym.isCommon());
          PrintBool('W', Sym.isWeak());
          PrintBool('I', Sym.isIndirect());
          PrintBool('O', Sym.canBeOmittedFromSymbolTable());
          PrintBool('T', Sym.isTLS());
          PrintBool('X', Sym.isExecutable());
          OutStream << ' ' << Sym.getName();
          OutStream << "| P " << Res.Prevailing;
          OutStream << " V " << Res.VisibleToRegularObj;
          OutStream << " E " << Res.ExportDynamic;
          OutStream << " F " << Res.FinalDefinitionInLinkageUnit;
          OutStream << "\n";
        };
 
        PrintSymbol(Sym, Res);
      }
    }
 
    // Add the bitcode file with its resolved symbols to the LTO job.
    cantFail(L.add(std::move(IF), Resolutions));
  };
 
  BuildResolutions();
 
  // Run the LTO job to compile the bitcode.
  size_t MaxTasks = L.getMaxTasks();
  SmallVector<std::unique_ptr<sys::fs::TempFile>> ObjectFiles{MaxTasks};
 
  auto AddStream =
      [&](size_t Task,
          const Twine & /*ModuleName*/) -> std::unique_ptr<CachedFileStream> {
    std::string TaskStr = Task ? "." + std::to_string(Task) : "";
    auto ExpectedF = createTempFile("lto-shard" + TaskStr, "o");
    if (auto E = ExpectedF.takeError())
      reportFatalError("Error creating tmp file " + toString(std::move(E)));
    ObjectFiles[Task] =
        std::make_unique<sys::fs::TempFile>(std::move(*ExpectedF));
    auto Ret = std::make_unique<CachedFileStream>(
        std::make_unique<llvm::raw_fd_ostream>(ObjectFiles[Task]->FD, false));
    if (!Ret)
      reportFatalError("Error creating CachedFileStream");
    return Ret;
  };
 
  if (Error E = L.run(AddStream))
    reportFatalError("Error: " + toString(std::move(E)));
 
  if (LTOError)
    reportFatalError(toString(
        createStringError(inconvertibleErrorCode(),
                          "Errors encountered inside the LTO pipeline.")));
 
  return ObjectFiles;
}
#endif
 
inline std::unique_ptr<MemoryBuffer> codegenRTC(Module &M,
                                                StringRef DeviceArch) {
  TIMESCOPE("proteus::codegenRTC");
  char *BinOut;
  size_t BinSize;
 
  SmallString<4096> ModuleBuf;
  raw_svector_ostream ModuleBufOS(ModuleBuf);
  WriteBitcodeToFile(M, ModuleBufOS);
 
  hiprtcLinkState HipLinkStatePtr;
 
  // NOTE: This code is an example of passing custom, AMD-specific
  // options to the compiler/linker.
  // NOTE: Unrolling can have a dramatic (time-consuming) effect on JIT
  // compilation time and on the resulting optimization, better or worse
  // depending on code specifics.
  std::string MArchOpt = ("-march=" + DeviceArch).str();
 
  // NOTE: We used to pass these options as well. "-mllvm",
  // "-unroll-threshold=1000",
  //  We removed them cause we saw on bezier they cause slowdowns
  const char *OptArgs[] = {MArchOpt.c_str()};
  std::vector<hiprtcJIT_option> JITOptions = {
      HIPRTC_JIT_IR_TO_ISA_OPT_EXT, HIPRTC_JIT_IR_TO_ISA_OPT_COUNT_EXT};
  size_t OptArgsSize = 1;
  const void *JITOptionsValues[] = {(void *)OptArgs, (void *)(OptArgsSize)};
  proteusHiprtcErrCheck(hiprtcLinkCreate(JITOptions.size(), JITOptions.data(),
                                         (void **)JITOptionsValues,
                                         &HipLinkStatePtr));
  // NOTE: the following version of te code does not set options.
  // proteusHiprtcErrCheck(hiprtcLinkCreate(0, nullptr, nullptr,
  // &hip_link_state_ptr));
 
  proteusHiprtcErrCheck(hiprtcLinkAddData(
      HipLinkStatePtr, HIPRTC_JIT_INPUT_LLVM_BITCODE, (void *)ModuleBuf.data(),
      ModuleBuf.size(), "", 0, nullptr, nullptr));
  proteusHiprtcErrCheck(
      hiprtcLinkComplete(HipLinkStatePtr, (void **)&BinOut, &BinSize));
 
  return MemoryBuffer::getMemBuffer(StringRef{BinOut, BinSize});
}
 
} // namespace detail
 
inline void setLaunchBoundsForKernel(Function &F, int MaxNumWorkGroups,
                                     int MinBlocksPerSM = 0) {
  // TODO: fix calculation of launch bounds.
  // TODO: find maximum (hardcoded 1024) from device info.
  // TODO: Setting as 1, BlockSize to replicate launch bounds settings
  F.addFnAttr("amdgpu-flat-work-group-size",
              "1," + std::to_string(std::min(1024, MaxNumWorkGroups)));
  // F->addFnAttr("amdgpu-waves-per-eu", std::to_string(WavesPerEU));
  if (MinBlocksPerSM != 0) {
    // NOTE: We are missing a heuristic to define the `WavesPerEU`, as such we
    // still need to study it. I restrict the waves by setting min equal to max
    // and disallowing any heuristics that HIP will use internally.
    // For more information please check:
    // https://clang.llvm.org/docs/AttributeReference.html#amdgpu-waves-per-eu
    F.addFnAttr("amdgpu-waves-per-eu", std::to_string(MinBlocksPerSM) + "," +
                                           std::to_string(MinBlocksPerSM));
  }
 
  PROTEUS_DBG(Logger::logs("proteus")
              << " => Set Workgroup size " << MaxNumWorkGroups
              << " WavesPerEU (unused) " << MinBlocksPerSM << "\n");
}
 
inline std::unique_ptr<MemoryBuffer>
codegenObject(Module &M, StringRef DeviceArch,
              [[maybe_unused]] SmallPtrSetImpl<void *> &GlobalLinkedBinaries,
              CodegenOption CGOption = CodegenOption::RTC,
              const OptimizationPipelineConfig &OptConfig =
                  OptimizationPipelineConfig(std::nullopt, '3', 3)) {
  TIMESCOPE("proteus::codegenObjectHIP");
  assert(GlobalLinkedBinaries.empty() &&
         "Expected empty linked binaries for HIP");
  Timer T(Config::get().ProteusEnableTimers);
  SmallVector<std::unique_ptr<sys::fs::TempFile>> ObjectFiles;
  switch (CGOption) {
  case CodegenOption::RTC: {
    auto Ret = detail::codegenRTC(M, DeviceArch);
    PROTEUS_TIMER_OUTPUT(Logger::outs("proteus")
                         << "Codegen RTC " << T.elapsed() << " ms\n");
    return Ret;
  }
#if LLVM_VERSION_MAJOR >= 18
  case CodegenOption::Serial:
    ObjectFiles = detail::codegenSerial(M, DeviceArch);
    break;
  case CodegenOption::Parallel:
    ObjectFiles = detail::codegenParallel(M, DeviceArch, OptConfig);
    break;
#endif
  default:
    reportFatalError("Unknown Codegen Option");
  }
 
  if (ObjectFiles.empty())
    reportFatalError("Expected non-empty vector of object files");
 
#if LLVM_VERSION_MAJOR >= 18
  auto ExpectedF = detail::createTempFile("proteus-jit", "o");
  if (auto E = ExpectedF.takeError())
    reportFatalError("Error creating shared object file " +
                     toString(std::move(E)));
 
  auto SharedObject = std::move(*ExpectedF);
 
  std::vector<const char *> Args{"ld.lld", "--no-undefined", "-shared", "-o",
                                 SharedObject.TmpName.c_str()};
  for (auto &File : ObjectFiles) {
    if (!File)
      continue;
    Args.push_back(File->TmpName.c_str());
  }
 
  if (Config::get().ProteusDebugOutput) {
    for (auto &Arg : Args) {
      Logger::logs("proteus") << Arg << " ";
    }
    Logger::logs("proteus") << "\n";
  }
 
  PROTEUS_TIMER_OUTPUT(Logger::outs("proteus")
                       << "Codegen object " << toString(CGOption) << "["
                       << ObjectFiles.size() << "] " << T.elapsed() << " ms\n");
 
  T.reset();
  // The LLD linker interface is not thread-safe, so we use a mutex.
  static std::mutex Mutex;
  {
    std::lock_guard LockGuard{Mutex};
    lld::Result S = lld::lldMain(Args, llvm::outs(), llvm::errs(),
                                 {{lld::Gnu, &lld::elf::link}});
    if (S.retCode)
      reportFatalError("Error: lld failed");
  }
 
  ErrorOr<std::unique_ptr<MemoryBuffer>> Buffer =
      MemoryBuffer::getFileAsStream(SharedObject.TmpName);
  if (!Buffer)
    reportFatalError("Error reading file: " + Buffer.getError().message());
 
  // Remove temporary files.
  for (auto &File : ObjectFiles) {
    if (!File)
      continue;
    if (auto E = File->discard())
      reportFatalError("Error removing object tmp file " +
                       toString(std::move(E)));
  }
  if (auto E = SharedObject.discard())
    reportFatalError("Error removing shared object tmp file " +
                     toString(std::move(E)));
 
  PROTEUS_TIMER_OUTPUT(Logger::outs("proteus")
                       << "Codegen linking " << T.elapsed() << " ms\n");
 
  return std::move(*Buffer);
#else
  reportFatalError("Expected LLVM18 for non-RTC codegen");
#endif
}
 
} // namespace proteus
 
#endif