Commit 2832d1ad authored by Patrick McCormick's avatar Patrick McCormick
Browse files

Merge remote-tracking branch 'stelleg-gpu10.x/gpu/10.x' into gpu-abi

parents 3a8c9cd8 29b6491b

clang/lib/Driver/Tapir.cpp

+1 −0
Original line number Diff line number Diff line
@@ -49,6 +49,7 @@ TapirTargetID clang::parseTapirTarget(const ArgList &Args) { 
      // TODO: Deprecate non-opencilk support...

      .Case("cilk", TapirTargetID::Cilk)

      .Case("cilkr", TapirTargetID::CilkR)

      .Case("gpu", TapirTargetID::GPU)

      .Default(TapirTargetID::Last_TapirTargetID);



  return TapirTarget;

clang/lib/Driver/ToolChain.cpp

+4 −0
Original line number Diff line number Diff line
@@ -1212,6 +1212,10 @@ void ToolChain::AddTapirRuntimeLibArgs(const ArgList &Args, 
      getDriver().Diag(diag::warn_drv_opencl_target_disabled);

    break;



  case TapirTargetID::GPU:

    CmdArgs.push_back("-lllvm-gpu");

    break; 



  case TapirTargetID::Serial:

  case TapirTargetID::None:

    break;

clang/lib/Frontend/CompilerInvocation.cpp

+2 −0
Original line number Diff line number Diff line
@@ -3707,6 +3707,8 @@ bool CompilerInvocation::CreateFromArgs(CompilerInvocation &Res, 
      LangOpts.Tapir = TapirTargetID::Realm;

    else if (Name == "opencl")

      LangOpts.Tapir = TapirTargetID::OpenCL;

    else if (Name == "gpu")

      LangOpts.Tapir = TapirTargetID::GPU;

    else if (Name == "serial")

      LangOpts.Tapir = TapirTargetID::Serial;

    else

llvm/include/llvm/Transforms/Tapir/GPUABI.h

0 → 100644
+107 −0
Original line number Diff line number Diff line 
//===- GPUABI.h - Interface to the Kitsune GPU back end ------*- C++ -*--===//

//

//                     The LLVM Compiler Infrastructure

//

// This file is distributed under the University of Illinois Open Source

// License. See LICENSE.TXT for details.

//

//===----------------------------------------------------------------------===//

//

// This file implements the Kitsune GPU ABI to convert Tapir instructions to

// calls into the Kitsune runtime system for NVIDIA GPU code.

//

//===----------------------------------------------------------------------===//

#ifndef GPU_ABI_H_

#define GPU_ABI_H_



#include "llvm/Transforms/Tapir/LoweringUtils.h"

#include "llvm/Transforms/Tapir/TapirLoopInfo.h"



namespace llvm {



class DataLayout;

class TargetMachine;

class LLVMLoop; 



class GPUABI : public TapirTarget {

  LLVMLoop *LOP = nullptr;

public:

  GPUABI(Module &M) : TapirTarget(M) {}

  ~GPUABI() {}

  Value *lowerGrainsizeCall(CallInst *GrainsizeCall) override final;

  void lowerSync(SyncInst &SI) override final;



  void addHelperAttributes(Function &F) override final {}

  void preProcessFunction(Function &F, TaskInfo &TI,

                          bool OutliningTapirLoops) override final;

  void postProcessFunction(Function &F, bool OutliningTapirLoops)

    override final;

  void postProcessHelper(Function &F) override final;

  void preProcessOutlinedTask(Function &F, Instruction *DetachPt,

                              Instruction *TaskFrameCreate,

                              bool IsSpawner) override final;

  void postProcessOutlinedTask(Function &F, Instruction *DetachPt,

                                       Instruction *TaskFrameCreate,

                                       bool IsSpawner) override final;

  void preProcessRootSpawner(Function &F) override final;

  void postProcessRootSpawner(Function &F) override final;



  void processSubTaskCall(TaskOutlineInfo &TOI, DominatorTree &DT)

    override final;



  LoopOutlineProcessor *getLoopOutlineProcessor(const TapirLoopInfo *TL)

    const override final;

};



class LLVMLoop : public LoopOutlineProcessor {

  friend class GPUABI; 



private:

  static unsigned NextKernelID;

  unsigned MyKernelID;

  Module LLVMM;

  TargetMachine *LLVMTargetMachine;

  GlobalVariable *LLVMGlobal;



  FunctionCallee GetThreadIdx = nullptr;

  FunctionCallee GPUInit = nullptr;

  FunctionCallee GPULaunchKernel = nullptr;

  FunctionCallee GPUWaitKernel = nullptr;



  SmallVector<Value *, 5> OrderedInputs; 

public:

  LLVMLoop(Module &M);



  void setupLoopOutlineArgs(

      Function &F, ValueSet &HelperArgs, SmallVectorImpl<Value *> &HelperInputs,

      ValueSet &InputSet, const SmallVectorImpl<Value *> &LCArgs,

      const SmallVectorImpl<Value *> &LCInputs,

      const ValueSet &TLInputsFixed)

    override final;

  unsigned getIVArgIndex(const Function &F, const ValueSet &Args) const

    override final;

  unsigned getLimitArgIndex(const Function &F, const ValueSet &Args) const

    override final;

  void postProcessOutline(TapirLoopInfo &TL, TaskOutlineInfo &Out,

                          ValueToValueMapTy &VMap) override final;

  void processOutlinedLoopCall(TapirLoopInfo &TL, TaskOutlineInfo &TOI,

                               DominatorTree &DT) override final;

};

}



#endif

/*

#include "llvm/Transforms/Tapir/LoopSpawningTI.h"

#include "llvm/Transforms/Tapir/LoweringUtils.h"

#include "llvm/ADT/DenseMap.h"



using namespace llvm; 



class GPU : public LoopOutlineProcessor {

public:

  GPU(Module &M) : LoopOutlineProcessor(M) {}

  void postProcessOutline(TapirLoopInfo &TL, TaskOutlineInfo &Out,

                          ValueToValueMapTy &VMap) override final;  

  GlobalVariable* LLVMKernel; 

};

*/

llvm/include/llvm/Transforms/Tapir/TapirTargetIDs.h

+1 −0
Original line number Diff line number Diff line
@@ -26,6 +26,7 @@ enum class TapirTargetID { 
  OpenCilk, // Lower to OpenCilk ABI

  OpenCL,   // Lower to OpenCL ABI

  OpenMP,   // Lower to OpenMP

  GPU,   // Lower to OpenCL

  Qthreads, // Lower to Qthreads

  Realm,    // Lower to Realm

  Last_TapirTargetID