diff --git a/hpvm/lib/Transforms/CMakeLists.txt b/hpvm/lib/Transforms/CMakeLists.txt
index bb044cd756883449dc775fd7742ac575aa5815a1..b18cd4551ba33e0c315a416164b45e6282098aeb 100644
--- a/hpvm/lib/Transforms/CMakeLists.txt
+++ b/hpvm/lib/Transforms/CMakeLists.txt
@@ -5,9 +5,6 @@ add_subdirectory(DFG2LLVM_CPU)
add_subdirectory(GenHPVM)
add_subdirectory(LocalMem)
add_subdirectory(DFG2LLVM_WrapperAPI)
-add_subdirectory(ReplaceIntrinsics)
add_subdirectory(DFG2LLVM_CUDNN)
-add_subdirectory(ExtractHPVMLeafNodes)
add_subdirectory(FuseHPVMTensorNodes)
-add_subdirectory(InlineTensorCalls)
add_subdirectory(InPlaceDFG)
diff --git a/hpvm/lib/Transforms/ExtractHPVMLeafNodes/CMakeLists.txt b/hpvm/lib/Transforms/ExtractHPVMLeafNodes/CMakeLists.txt
deleted file mode 100644
index bb943f9100e628c87c865dfbdce80fc094ebb23e..0000000000000000000000000000000000000000
--- a/hpvm/lib/Transforms/ExtractHPVMLeafNodes/CMakeLists.txt
+++ /dev/null
@@ -1,14 +0,0 @@
-if(WIN32 OR CYGWIN)
- set(LLVM_LINK_COMPONENTS Core Support)
-endif()
-
-add_llvm_library( ExtractHPVMLeafNodes
- MODULE
- ExtractHPVMLeafNodes.cpp
-
- DEPENDS
- intrinsics_gen
- PLUGIN_TOOL
- opt
- )
-
diff --git a/hpvm/lib/Transforms/ExtractHPVMLeafNodes/ExtractHPVMLeafNodes.cpp b/hpvm/lib/Transforms/ExtractHPVMLeafNodes/ExtractHPVMLeafNodes.cpp
deleted file mode 100644
index 031503adeddd6c070ca06f3012fa0c2e5362f92c..0000000000000000000000000000000000000000
--- a/hpvm/lib/Transforms/ExtractHPVMLeafNodes/ExtractHPVMLeafNodes.cpp
+++ /dev/null
@@ -1,248 +0,0 @@
-//===------------------- ExtractHPVMLeafNodeGenFunctions.cpp -----------------===//
-//
-// The LLVM Compiler Infrastructure
-//
-// This file is distributed under the University of Illinois Open Source
-//
-// License. See LICENSE.TXT for details.
-//
-//===----------------------------------------------------------------------===//
-
-#define DEBUG_TYPE "ExtractHPVMLeafNodes"
-
-#include "llvm/Support/SourceMgr.h"
-#include "llvm/Pass.h"
-
-#include "SupportHPVM/DFGTreeTraversal.h"
-#include "ExtractHPVMLeafNodes/ExtractHPVMLeafNodes.h"
-
-#include "llvm/Transforms/Utils/BasicBlockUtils.h"
-#include "llvm/IR/IRPrintingPasses.h"
-#include "llvm/IR/LegacyPassManager.h"
-#include "llvm/Support/ToolOutputFile.h"
-#include "llvm/Support/FileSystem.h"
-
-using namespace llvm;
-using namespace builddfg;
-using namespace extracthpvmleaf;
-using namespace dfg2llvm;
-
-namespace {
-
-class PrintLeafNodes : public DFGTreeTraversal {
- public:
- virtual void process(DFInternalNode* N) override;
- virtual void process(DFLeafNode* N) override;
-
- // Constructor
- PrintLeafNodes(Module &_M, BuildDFG &_DFG) : DFGTreeTraversal(_M, _DFG) {}
-
-};
-
-}
-
-void PrintLeafNodes::process(DFInternalNode* N) {
- DEBUG(errs() << "Analysing Node: " << N->getFuncPointer()->getName() << "\n");
- return; // nothing to do
-}
-
-void PrintLeafNodes::process(DFLeafNode* N) {
- DEBUG(errs() << "Analysing Node: " << N->getFuncPointer()->getName() << "\n");
- if((N->isDummyNode())) {
- DEBUG(errs() << "Skipping Dummy Node: " << N->getFuncPointer()->getName() << "\n");
- return;
- }
-
- // Find function generated for node
- Function *F = N->getGenFuncForTarget(hpvm::CPU_TARGET);
- assert(F != NULL
- && "This pass is invoked after code generation for x86 is completed.\nFound leaf node for which code generation has not happened!\n");
- assert(N->hasCPUGenFuncForTarget(hpvm::CPU_TARGET) &&
- "The generated function from x86 pass is not an x86 function\n");
-
- std::string module_name = std::string("./build/") + std::string(F->getName().str().c_str()) + std::string("_module.ll");
- Twine tw(module_name);
- // Create a new module for the node function
- //Twine tw = Twine(F->getName()).concat(Twine("_module.ll"));
- Module *m = new Module(tw.str(), F->getParent()->getContext());
- // Create a new function for F. It will be written to a new module.
- ValueToValueMapTy VMap;
- Function *ClonedF = CloneFunction(F, VMap);
- // Remove it from current module
- ClonedF->removeFromParent();
- // Insert it to the newly created module for it
- m->getFunctionList().push_back(ClonedF);
-
- std::vector<Instruction*> ItoRemove;
-
- for (inst_iterator i = inst_begin(ClonedF), e = inst_end(ClonedF); i != e; ++i) {
- Instruction *I = &(*i);
- errs() << *I << "\n";
-
- if (CallInst *CI = dyn_cast<CallInst>(I)) {
- errs() << "Found call instruction\n";
-
- Function *CalledF = CI->getCalledFunction();
- StringRef CallName = CalledF->getName();
- errs() << "CallName: " << CallName << "\n";
-
-// if (CallName.startswith("llvm_hpvm")) { //TODO
- if ((CallName.startswith("llvm_hpvm")) || (CallName.startswith("tensor"))) { //TODO
-// errs() << "This is an HPVM runtime call. Include its declaration.\n";
- errs() << "This is an HPVM runtime call or tensor. Include its declaration.\n";
-
- FunctionType *CalledFType = CalledF->getFunctionType();
-
- std::vector<Value*> Fargs;
- for (unsigned argno = 0; argno < CI->getNumArgOperands(); argno++) {
- Fargs.push_back(CI->getArgOperand(argno));
- }
- Function *FDecl = dyn_cast<Function>((m->getOrInsertFunction(CallName, CalledFType)).getCallee());
- CallInst *NewCI = CallInst::Create(CalledFType, FDecl, Fargs, CallName, CI);
- errs() << "NewCI: " << *NewCI << "\n";
- CI->replaceAllUsesWith(NewCI);
- ItoRemove.push_back(CI);
- }
- }
- }
-
- for (unsigned i = 0; i < ItoRemove.size() ; i++) {
- ItoRemove[i]->eraseFromParent();
- }
-
- ItoRemove.clear();
-
- // Print new module
- legacy::PassManager Passes;
-
- errs() << "Writing to File --- " << tw.str() << "\n";
- std::error_code EC;
- ToolOutputFile Out(tw.str(), EC, sys::fs::F_None);
- if (EC) {
- errs() << EC.message() << '\n';
- }
-
- Passes.add(createPrintModulePass(Out.os()));
- Passes.run(*m);
- // Declare success.
- Out.keep();
-
- // Any call that is to F, needs to call the new external function
- // Edit initial module to do so
- // This is the name with which the function is called now
- StringRef FName = ClonedF->getName();
- FunctionType *FType = F->getFunctionType();
-
- // This is a node function, so it is only called through the dataflow graph
- assert(F->hasOneUse() && "F is an HPVM node function\n");
-
-/*
- errs() << "F uses: " << F->getNumUses() << "\n" ;
- for(Value::user_iterator ui = F->user_begin(),
- ue = F->user_end(); ui!=ue; ++ui) {
- errs() << "use : "<< **ui << "\n";
- }
-*/
-
- // Get the parent node's generated x86 function
- DFInternalNode *ParentNode = N->getParent();
- Function *PGenF = ParentNode->getGenFuncForTarget(hpvm::CPU_TARGET);
- assert(PGenF != NULL
- && "This pass is invoked after code generation for x86 is completed.\nFound node for which code generation has not happened!\n");
- assert(ParentNode->hasCPUGenFuncForTarget(hpvm::CPU_TARGET) &&
- "The generated function from x86 pass is not an x86 function\n");
-
- for (inst_iterator i = inst_begin(PGenF), e = inst_end(PGenF); i != e; ++i) {
- Instruction *I = &(*i);
- errs() << *I << "\n";
-
- if (CallInst *CI = dyn_cast<CallInst>(I)) {
- errs() << "Found call instruction\n";
-
- StringRef CallName = CI->getCalledFunction()->getName();
- errs() << "CallName: " << CallName << "\n";
- errs() << "F->getName(): " << F->getName() << "\n";
-
- if (CallName == F->getName()) {
- // Found the call to the leaf node function we moved to the other module.
- // Replace the call
- std::vector<Value*> Fargs;
- for (unsigned argno = 0; argno < CI->getNumArgOperands(); argno++) {
- Fargs.push_back(CI->getArgOperand(argno));
- }
- Function *FDecl = dyn_cast<Function>(M.getOrInsertFunction(FName, FType).getCallee());
- CallInst *NewCI = CallInst::Create(FType, FDecl, Fargs, FName, CI);
- errs() << "NewCI: " << *NewCI << "\n";
- CI->replaceAllUsesWith(NewCI);
- ItoRemove.push_back(CI);
- }
- }
- }
-
- for (unsigned i = 0; i < ItoRemove.size() ; i++) {
- ItoRemove[i]->eraseFromParent();
- }
-
- // Clean up
- ClonedF->eraseFromParent();
- delete m;
-
- F->replaceAllUsesWith(UndefValue::get(F->getType()));
- F->eraseFromParent();
-
- return;
-}
-
-void ExtractHPVMLeafNodeFunctions::run(Module &M, BuildDFG &DFG) {
-
- errs() << "\nEXTRACT HPVM LEAF NODE FUNCTIONS PASS\n";
-
- std::vector<DFInternalNode*> Roots = DFG.getRoots();
-
- // Visitor for Graph Traversal
- PrintLeafNodes *LeafVisitor = new PrintLeafNodes(M, DFG);
-
- // Iterate over all the DFGs
- // Analyse the edges for parameters that are valid to be used in place
- for (auto rootNode: Roots) {
- LeafVisitor->visit(rootNode);
- }
-
- delete LeafVisitor;
- return;
-}
-
-namespace {
-struct ExtractHPVMLeafNodeGenFunctionsWrapper : public ModulePass {
- static char ID;
- ExtractHPVMLeafNodeGenFunctionsWrapper() : ModulePass(ID) {}
-
- bool runOnModule(Module &) override;
-
- void getAnalysisUsage(AnalysisUsage &AU) const override;
-};
-} // end anonymous namespace
-
-void ExtractHPVMLeafNodeGenFunctionsWrapper::getAnalysisUsage(AnalysisUsage &AU) const {
- AU.addRequired<BuildDFG>();
- AU.addPreserved<BuildDFG>();
-}
-
-bool ExtractHPVMLeafNodeGenFunctionsWrapper::runOnModule(Module &M) {
- // Get the BuildDFG Analysis Results:
- // - Dataflow graph
- BuildDFG &DFG = getAnalysis<BuildDFG>();
-
- ExtractHPVMLeafNodeFunctions ELNF;
- ELNF.run(M, DFG);
-
- return false;
-}
-
-char ExtractHPVMLeafNodeGenFunctionsWrapper::ID = 0;
-static RegisterPass<ExtractHPVMLeafNodeGenFunctionsWrapper> X(
- "hpvm-extract-leaf-gen",
- "Pass to extract leaf nodes to modules in HPVM",
- false /* does not modify the CFG */,
-true /* transformation, not just analysis */);
-
diff --git a/hpvm/lib/Transforms/ExtractHPVMLeafNodes/ExtractHPVMLeafNodes.exports b/hpvm/lib/Transforms/ExtractHPVMLeafNodes/ExtractHPVMLeafNodes.exports
deleted file mode 100644
index 139597f9cb07c5d48bed18984ec4747f4b4f3438..0000000000000000000000000000000000000000
--- a/hpvm/lib/Transforms/ExtractHPVMLeafNodes/ExtractHPVMLeafNodes.exports
+++ /dev/null
@@ -1,2 +0,0 @@
-
-
diff --git a/hpvm/lib/Transforms/ExtractHPVMLeafNodes/LLVMBuild.txt b/hpvm/lib/Transforms/ExtractHPVMLeafNodes/LLVMBuild.txt
deleted file mode 100644
index 73ac540f06e86e9e7f0201b993d2c1e11270158e..0000000000000000000000000000000000000000
--- a/hpvm/lib/Transforms/ExtractHPVMLeafNodes/LLVMBuild.txt
+++ /dev/null
@@ -1,21 +0,0 @@
-;===- ./lib/Transforms/DFG2LLVM_WrapperAPI/LLVMBuild.txt -------*- Conf -*--===;
-;
-; The LLVM Compiler Infrastructure
-;
-; This file is distributed under the University of Illinois Open Source
-; License. See LICENSE.TXT for details.
-;
-;===------------------------------------------------------------------------===;
-;
-; This is an LLVMBuild description file for the components in this subdirectory.
-;
-; For more information on the LLVMBuild system, please see:
-;
-; http://llvm.org/docs/LLVMBuild.html
-;
-;===------------------------------------------------------------------------===;
-
-[component_0]
-type = Library
-name = ExtractHPVMLeafNodes
-parent = Transforms
diff --git a/hpvm/lib/Transforms/InlineTensorCalls/CMakeLists.txt b/hpvm/lib/Transforms/InlineTensorCalls/CMakeLists.txt
deleted file mode 100644
index 29dd2c8431362b28a1d5683eadb2c9eb867696ff..0000000000000000000000000000000000000000
--- a/hpvm/lib/Transforms/InlineTensorCalls/CMakeLists.txt
+++ /dev/null
@@ -1,14 +0,0 @@
-if(WIN32 OR CYGWIN)
- set(LLVM_LINK_COMPONENTS Core Support)
-endif()
-
-add_llvm_library( InlineTensorCalls
- MODULE
- InlineTensorCalls.cpp
-
- DEPENDS
- intrinsics_gen
- PLUGIN_TOOL
- opt
- )
-
diff --git a/hpvm/lib/Transforms/InlineTensorCalls/InlineTensorCalls.cpp b/hpvm/lib/Transforms/InlineTensorCalls/InlineTensorCalls.cpp
deleted file mode 100644
index d31434341cf65939768d0acb7a0051d453909971..0000000000000000000000000000000000000000
--- a/hpvm/lib/Transforms/InlineTensorCalls/InlineTensorCalls.cpp
+++ /dev/null
@@ -1,77 +0,0 @@
-//=== InlineApproxHPVMCalls.cpp ===//
-//
-// The LLVM Compiler Infrastructure
-//
-// This file is distributed under the University of Illinois Open Source
-// License. See LICENSE.TXT for details.
-//
-//===----------------------------------------------------------------------===//
-#define ENABLE_ASSERTS
-
-#define DEBUG_TYPE "INLINE_APPROXHPVM_CALLS"
-#include "llvm/IR/Function.h"
-#include "llvm/IR/Module.h"
-#include "llvm/Pass.h"
-
-#include "llvm/IR/InstIterator.h"
-
-#include "llvm/Support/raw_ostream.h"
-#include "llvm/Analysis/InlineCost.h"
-
-#include "llvm/Transforms/Utils/BasicBlockUtils.h"
-#include "llvm/Transforms/Utils/Cloning.h"
-#include "llvm/IRReader/IRReader.h"
-#include "llvm/Linker/Linker.h"
-#include "llvm/Support/SourceMgr.h"
-#include "llvm/IR/CallSite.h"
-#include "llvm/ADT/SetVector.h"
-#include <sstream>
-
-using namespace llvm;
-
-
-namespace {
-
- struct InlineApproxHPVMCalls : public ModulePass {
- static char ID; // Pass identification, replacement for typeid
- InlineApproxHPVMCalls() : ModulePass(ID) {}
-
- bool runOnModule(Module &M) override {
-
- InlineFunctionInfo IFI;
- SmallSetVector<CallSite, 16> Calls;
- bool Changed = false;
- SmallVector<Function *, 16> InlinedFunctions;
- for (Function &F : M){
- if (!F.isDeclaration() && F.getName().startswith("tensor") ) {
- //errs()<<"Function = "<<*&F<<"\n";
- Calls.clear();
-
- for (User *U : F.users())
- if (auto CS = CallSite(U))
- if (CS.getCalledFunction() == &F)
- Calls.insert(CS);
-
- for (CallSite CS : Calls)
- // FIXME: We really shouldn't be able to fail to inline at this point!
- // We should do something to log or check the inline failures here.
- Changed |= InlineFunction(CS, IFI);
-
- }
- }
-
- return true;
- }
-
- };
-
-
-} // End of namespace
-
-char InlineApproxHPVMCalls::ID = 0;
-static RegisterPass<InlineApproxHPVMCalls> X("inline-tensor-calls",
- "Inline ApproxHPVM tensor library function calls (CPU version)",
- true /* modifies the CFG */,
- true /* transformation, *
- * not just analysis */);
-
diff --git a/hpvm/lib/Transforms/InlineTensorCalls/InlineTensorCalls.exports b/hpvm/lib/Transforms/InlineTensorCalls/InlineTensorCalls.exports
deleted file mode 100644
index 139597f9cb07c5d48bed18984ec4747f4b4f3438..0000000000000000000000000000000000000000
--- a/hpvm/lib/Transforms/InlineTensorCalls/InlineTensorCalls.exports
+++ /dev/null
@@ -1,2 +0,0 @@
-
-
diff --git a/hpvm/lib/Transforms/InlineTensorCalls/LLVMBuild.txt b/hpvm/lib/Transforms/InlineTensorCalls/LLVMBuild.txt
deleted file mode 100644
index c160516a6477d367893495e39f5fd4d00366f6f0..0000000000000000000000000000000000000000
--- a/hpvm/lib/Transforms/InlineTensorCalls/LLVMBuild.txt
+++ /dev/null
@@ -1,21 +0,0 @@
-;===- ./lib/Transforms/DFG2LLVM_WrapperAPI/LLVMBuild.txt -------*- Conf -*--===;
-;
-; The LLVM Compiler Infrastructure
-;
-; This file is distributed under the University of Illinois Open Source
-; License. See LICENSE.TXT for details.
-;
-;===------------------------------------------------------------------------===;
-;
-; This is an LLVMBuild description file for the components in this subdirectory.
-;
-; For more information on the LLVMBuild system, please see:
-;
-; http://llvm.org/docs/LLVMBuild.html
-;
-;===------------------------------------------------------------------------===;
-
-[component_0]
-type = Library
-name = InlineTensorCalls
-parent = Transforms
diff --git a/hpvm/lib/Transforms/ReplaceIntrinsics/CMakeLists.txt b/hpvm/lib/Transforms/ReplaceIntrinsics/CMakeLists.txt
deleted file mode 100644
index 460aabcc27b51a2d94dabee3e9c4c60d14803ea9..0000000000000000000000000000000000000000
--- a/hpvm/lib/Transforms/ReplaceIntrinsics/CMakeLists.txt
+++ /dev/null
@@ -1,14 +0,0 @@
-if(WIN32 OR CYGWIN)
- set(LLVM_LINK_COMPONENTS Core Support)
-endif()
-
-add_llvm_library( ReplaceIntrinsics
- MODULE
- ReplaceIntrinsics.cpp
-
- DEPENDS
- intrinsics_gen
- PLUGIN_TOOL
- opt
- )
-
diff --git a/hpvm/lib/Transforms/ReplaceIntrinsics/LLVMBuild.txt b/hpvm/lib/Transforms/ReplaceIntrinsics/LLVMBuild.txt
deleted file mode 100644
index 95739b3d4d1c3a68cc5014dc85fb26d3b1fc6ac5..0000000000000000000000000000000000000000
--- a/hpvm/lib/Transforms/ReplaceIntrinsics/LLVMBuild.txt
+++ /dev/null
@@ -1,21 +0,0 @@
-;===- ./lib/Transforms/DFG2LLVM_WrapperAPI/LLVMBuild.txt -------*- Conf -*--===;
-;
-; The LLVM Compiler Infrastructure
-;
-; This file is distributed under the University of Illinois Open Source
-; License. See LICENSE.TXT for details.
-;
-;===------------------------------------------------------------------------===;
-;
-; This is an LLVMBuild description file for the components in this subdirectory.
-;
-; For more information on the LLVMBuild system, please see:
-;
-; http://llvm.org/docs/LLVMBuild.html
-;
-;===------------------------------------------------------------------------===;
-
-[component_0]
-type = Library
-name = ReplaceIntrinsics
-parent = Transforms
diff --git a/hpvm/lib/Transforms/ReplaceIntrinsics/ReplaceIntrinsics.cpp b/hpvm/lib/Transforms/ReplaceIntrinsics/ReplaceIntrinsics.cpp
deleted file mode 100644
index 45ad0ece23568a41fbf532b92918a582ebbae505..0000000000000000000000000000000000000000
--- a/hpvm/lib/Transforms/ReplaceIntrinsics/ReplaceIntrinsics.cpp
+++ /dev/null
@@ -1,495 +0,0 @@
-//=== ReplaceApproxHPVMIntrinsicsWithFCalls.cpp ===//
-//
-// The LLVM Compiler Infrastructure
-//
-// This file is distributed under the University of Illinois Open Source
-// License. See LICENSE.TXT for details.
-//
-//===----------------------------------------------------------------------===//
-#define ENABLE_ASSERTS
-
-#define DEBUG_TYPE "REPLACE_APPROXHPVM_INTRINSICS_WITH_FCALLS"
-
-#include "llvm/IR/DataLayout.h"
-#include "llvm/IR/IRBuilder.h"
-#include "llvm/IR/Module.h"
-#include "llvm/Pass.h"
-#include "llvm/IR/InstIterator.h"
-#include "llvm/Transforms/Utils/ValueMapper.h"
-#include "llvm/Transforms/Utils/BasicBlockUtils.h"
-#include "llvm/Transforms/Utils/Cloning.h"
-#include "llvm/IRReader/IRReader.h"
-#include "llvm/Linker/Linker.h"
-#include "llvm/Support/SourceMgr.h"
-#include "llvm/Support/FileSystem.h"
-#include "llvm/IR/Attributes.h"
-#include "llvm-c/Core.h"
-
-#include "SupportHPVM/DFG2LLVM.h"
-#include "InPlaceDFG/InPlaceDFGAnalysis.h"
-
-#include <sstream>
-
-using namespace llvm;
-using namespace builddfg;
-using namespace dfg2llvm;
-
-// TODO: We still need in place analysis, if calls have the same interface
-using namespace inplacedfg;
-
-namespace {
-// Helper class declarations
-
-// Replace ApproxHPVM intrinsics with LLVM function calls.
-// aiming to go through the CPU backend code generation.
-
-struct DFG2LLVM_ReplaceApproxHPVMIntrinsicsWithFCalls : public DFG2LLVM {
- static char ID; // Pass identification, replacement for typeid
- DFG2LLVM_ReplaceApproxHPVMIntrinsicsWithFCalls() : DFG2LLVM(ID) {}
-
-private:
-public:
- void getAnalysisUsage(AnalysisUsage &AU) const {
- AU.addRequired<BuildDFG>();
- AU.addRequired<InPlaceDFGAnalysisWrapper>();
- AU.addPreserved<BuildDFG>();
- AU.addPreserved<InPlaceDFGAnalysisWrapper>();
- }
-
- bool runOnModule(Module &M);
-};
-
-// Visitor for Code generation traversal (tree traversal for now)
-class CGT_ReplaceApproxHPVMIntrinsicsWithFCalls : public CodeGenTraversal {
-
-private:
- // Member variables
- InPlaceDFGAnalysis::InPlaceDFGParameter *IPP;
-
- // VISC Runtime API and Tensor runtime API
-
- /* TODO: I believe that TensorRt is not needed, since we will have llvm
- implementations linked in, so init and cleanup calls can be removed and
- relevant code also, but I leave in in for now until verified. */
- FunctionCallee llvm_hpvm_initTensorRt;
- FunctionCallee llvm_hpvm_cleanupTensorRt;
- // Constant* hpvm_request_tensor; DONE: request tensor will not be used
-
- // Functions
- bool isValidOperandForInPlaceOperation(Value *Op, Function *Fgen, DFNode *N);
-
- // Virtual Functions
- void init();
- void initRuntimeAPI();
- void codeGen(DFInternalNode *N);
- void codeGen(DFLeafNode *N);
-
-public:
- // Constructor
- CGT_ReplaceApproxHPVMIntrinsicsWithFCalls(
- Module &_M, BuildDFG &_DFG, InPlaceDFGAnalysis::InPlaceDFGParameter &_IPP)
- : CodeGenTraversal(_M, _DFG), IPP(&_IPP) {
- initRuntimeAPI();
- }
-};
-
-bool CGT_ReplaceApproxHPVMIntrinsicsWithFCalls::
- isValidOperandForInPlaceOperation(Value *Op, Function *Fgen, DFNode *N) {
- // We only expect the if branch to be taken
- if (Argument *Arg = dyn_cast<Argument>(Op)) {
- DEBUG(errs() << *Arg << "\t: argument, candidate for in place\n");
- assert((Arg->getParent() == Fgen) &&
- "Extra Parameter in body of Function\n");
- // Candidae parameter is a function argument
- // In this case, consult the result of in place analysis
- // Find position in arg list
- unsigned pos = Arg->getArgNo();
- // If this parameter cannot be used for in place operation
- // code gen cannot continue
- if (IPP->at(N)[pos]) {
- DEBUG(errs() << *Arg << "\t: argument, suitable for in place\n");
- return true;
- } else {
- DEBUG(errs() << *Arg << "\t: argument, not suitable for in place\n");
- return false;
- }
- } else {
- // If it is not an argument, then it needs to be the result of
- // another intrinsic. These are new objects that are allocated,
- // and consumed by next intrinsic. Alternatively, the intrinsic
- // could have been replaced by a call to an LLVM function.
- // We do not expect a merge pass to have run before the replacement pass,
- // therefore we do not expect to go in the else branch.
- DEBUG(errs() << *Op << "\t: Test for result of intrinsic operation\n");
- if (dyn_cast<IntrinsicInst>(Op)) {
- DEBUG(errs() << *Arg << "\t: local, suitable for in place\n");
- return true;
- } else if (CallInst *CI = dyn_cast<CallInst>(Op)) {
- if ((CI->getCalledFunction()->getName()).startswith("tensor"))
- return true;
- else
- return false;
- } else {
- DEBUG(errs() << *Arg << "\t: local, not suitable for in place\n");
- return false;
- }
- }
-}
-
-void CGT_ReplaceApproxHPVMIntrinsicsWithFCalls::init() {}
-
-// Initialize the VISC runtime API. This makes it easier to insert these calls
-void CGT_ReplaceApproxHPVMIntrinsicsWithFCalls::initRuntimeAPI() {
-
- // Load Runtime API Module
- SMDiagnostic Err;
- runtimeModule = parseIRFile(TENSOR_RT_LL, Err, M.getContext());
- if (runtimeModule == nullptr)
- DEBUG(errs() << Err.getMessage());
- else
- DEBUG(errs() << "Successfully loaded hpvm-tensor-rt API module\n");
-
- // Get or insert Global declarations for
- // - initialization
- // - cleanup
- // - request a tensor
- DECLARE(llvm_hpvm_initTensorRt);
- DECLARE(llvm_hpvm_cleanupTensorRt);
- // DECLARE(hpvm_request_tensor);
-
- // Find hpvm.init and visc.cleanup calls, and add placeholder methods
- // for initialization and cleanup of the hpvm tensor runtime
-
- Function *VI = M.getFunction("llvm.hpvm.init");
- assert(VI->getNumUses() == 1 && "__hpvm__init should only be used once\n");
- InitCall = cast<Instruction>(*VI->user_begin());
- CallInst::Create(
- llvm_hpvm_initTensorRt,
- ArrayRef<Value *>(ConstantInt::get(Type::getInt32Ty(M.getContext()), 0)),
- "", InitCall);
-
- Function *VC = M.getFunction("llvm.hpvm.cleanup");
- assert(VC->getNumUses() == 1 && "__hpvm__clear should only be used once\n");
- CleanupCall = cast<Instruction>(*VC->user_begin());
- CallInst::Create(llvm_hpvm_cleanupTensorRt, ArrayRef<Value *>(), "",
- CleanupCall);
-}
-
-void CGT_ReplaceApproxHPVMIntrinsicsWithFCalls::codeGen(DFInternalNode *N) {
- errs() << "Inside node: " << N->getFuncPointer()->getName() << "\n";
- errs() << "Skipping internal node\n";
-}
-
-void CGT_ReplaceApproxHPVMIntrinsicsWithFCalls::codeGen(DFLeafNode *N) {
-
- // Skip if it is a dummy node
- if (N->isDummyNode()) {
- DEBUG(errs() << "Skipping dummy node\n");
- return;
- }
-
- // Abort if it is an allocation node
- if (N->isAllocationNode()) {
- assert(false && "Allocation Node not expected in ApproxHPVM");
- return;
- }
-
- // Search for intrinsic only if it has the right hint
- if (!checkPreferredTarget(N, hpvm::CPU_TARGET)) {
- errs() << "Skipping node: " << N->getFuncPointer()->getName() << "\n";
- return;
- }
-
- // Get the function associated with the dataflow node
- Function *F = N->getFuncPointer();
- errs() << "function name = " << F->getName() << "\n";
-
- std::vector<IntrinsicInst *> IItoRemove;
-
- for (inst_iterator i = inst_begin(F), e = inst_end(F); i != e; ++i) {
- Instruction *I = &(*i);
- if (BuildDFG::isHPVMIntrinsic(I)) {
- IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
- assert(
- (II->getCalledFunction()->getName()).startswith("llvm.hpvm.tensor") &&
- "Only HPVM tensor intrinsics allowed in ApproxHPVM leaf nodes\n");
- /********************* Handle VISC Tensor intrinsics ********************/
- // We replace them with calls to functions with implementations at the
- // LLVM level
- switch (II->getIntrinsicID()) {
-
- case Intrinsic::hpvm_tensor_convolution: { /* llvm.hpvm.tensor.convolution
- */
- DEBUG(errs() << F->getName() << "\t: Handling tensor convolution \n");
-
- // Argument list for the runtime call
- std::vector<Value *> Args;
- Args.push_back(II->getOperand(0));
- Args.push_back(II->getOperand(1));
- Args.push_back(II->getOperand(2));
- Args.push_back(II->getOperand(3));
- Args.push_back(II->getOperand(4));
- Args.push_back(II->getOperand(5));
-
- Constant *conv_mode =
- ConstantInt::get(Type::getInt32Ty(M.getContext()), 1);
- Constant *conv_precision =
- ConstantInt::get(Type::getInt32Ty(M.getContext()), 0);
-
- Args.push_back(conv_mode);
- Args.push_back(conv_precision);
-
- // Create function call
- FunctionCallee tensorConvolutionCPU;
- DECLARE(tensorConvolutionCPU);
-
- CallInst *CI = CallInst::Create(tensorConvolutionCPU, Args, "", II);
- // We can replace the call to hpvm.tensor.mul with the LLVM call
- II->replaceAllUsesWith(CI);
-
- // Mark to remove at the end
- IItoRemove.push_back(II);
- } break;
-
- case Intrinsic::hpvm_tensor_mul: { /* llvm.hpvm.tensor.mul */
- DEBUG(errs() << F->getName() << "\t: Handling tensor mul\n");
-
- // Argument list for the runtime call
- std::vector<Value *> Args;
- Args.push_back(II->getOperand(0));
- Args.push_back(II->getOperand(1));
-
- // Create function call
- FunctionCallee tensorGemmCPU;
- DECLARE(tensorGemmCPU);
-
- CallInst *CI = CallInst::Create(tensorGemmCPU, Args, "", II);
- // We can replace the call to hpvm.tensor.mul with the LLVM call
- II->replaceAllUsesWith(CI);
-
- // Mark to remove at the end
- IItoRemove.push_back(II);
- } break;
-
- case Intrinsic::hpvm_tensor_add: { /* llvm.hpvm.tensor.add */
- DEBUG(errs() << F->getName() << "\t: Handling tensor add\n");
- // Tensor add(a,b) is in place for argument a.
- Value *Op = II->getOperand(0);
-
- // Test the intrinsic operand for in place operation.
- bool inplace = isValidOperandForInPlaceOperation(Op, F, N);
- // Code generation cannot continue if this is false, because the target
- // only provides an in place operation
-
- // FIXME: remove this comment - must check for in-place
- // assert(inplace &&
- // "Operand not valid for in place operation. Code gen
- // aborted.\n");
-
- // Argument list for the runtime call
- std::vector<Value *> Args;
- Args.push_back(II->getOperand(0));
- Args.push_back(II->getOperand(1));
-
- // Create function call
- FunctionCallee tensorAddCPU;
- DECLARE(tensorAddCPU);
- CallInst::Create(tensorAddCPU, Args, "", II);
- // We can replace the call to hpvm.tensor.add with the 1st argument
- // that, due to in place operation, now contains the result
- II->replaceAllUsesWith(II->getOperand(0));
-
- // Mark to remove at the end
- IItoRemove.push_back(II);
- } break;
-
- case Intrinsic::hpvm_tensor_pool_max:
- case Intrinsic::hpvm_tensor_pool_mean: { /* llvm.hpvm.tensor.relu */
- DEBUG(errs() << F->getName() << "\t: Handling tensor_pool_max\n");
- // Tensor relu(a) is in place for argument a.
- Value *Op = II->getOperand(0);
-
- // Test the intrinsic operand for in place operation.
- bool inplace = isValidOperandForInPlaceOperation(Op, F, N);
- // Code generation cannot continue if this is false, because the target
- // only provides an in place operation
- assert(inplace &&
- "Operand not valid for in place operation. Code gen aborted.\n");
-
- // Argument list - tensorPooling(input, poolFunction, window_height,
- // window_width, vertical_pad, horizontal_pad,
- // vertical_stride, horizontal_stride);
- std::vector<Value *> Args;
- Args.push_back(II->getOperand(0));
-
- int pool_type = 0;
- if (II->getIntrinsicID() == Intrinsic::hpvm_tensor_pool_max) {
- pool_type = 0;
- }
- if (II->getIntrinsicID() == Intrinsic::hpvm_tensor_pool_mean) {
- pool_type = 1;
- }
-
- Constant *constPoolType =
- ConstantInt::get(Type::getInt32Ty(M.getContext()), pool_type);
- Args.push_back(constPoolType); // ID for max pool. Min/Avg have
- // different IDs (non-zero)
- Args.push_back(II->getOperand(1));
- Args.push_back(II->getOperand(2));
- Args.push_back(II->getOperand(3));
- Args.push_back(II->getOperand(4));
- Args.push_back(II->getOperand(5));
- Args.push_back(II->getOperand(6));
-
- // Create function call
- FunctionCallee tensorPoolingCPU;
- DECLARE(tensorPoolingCPU);
- CallInst *CI = CallInst::Create(tensorPoolingCPU, Args, "", II);
-
- // Replacing intrinsic result uses with the result of the LLVM call
- II->replaceAllUsesWith(CI);
-
- // Mark to remove at the end
- IItoRemove.push_back(II);
- } break;
-
- case Intrinsic::hpvm_tensor_relu:
- case Intrinsic::hpvm_tensor_clipped_relu:
- case Intrinsic::hpvm_tensor_tanh: { /* llvm.hpvm.tensor.relu */
- DEBUG(errs() << F->getName()
- << "\t: Handling tensor activation functions \n");
- // Tensor relu(a) is in place for argument a.
- Value *Op = II->getOperand(0);
-
- // Test the intrinsic operand for in place operation.
- bool inplace = isValidOperandForInPlaceOperation(Op, F, N);
- // Code generation cannot continue if this is false, because the target
- // only provides an in place operation
- assert(inplace &&
- "Operand not valid for in place operation. Code gen aborted.\n");
-
- // Argument list for the runtime call
- std::vector<Value *> Args;
- Args.push_back(II->getOperand(0));
-
- if (II->getIntrinsicID() == Intrinsic::hpvm_tensor_relu) {
- // Create function call
- FunctionCallee tensorReluCPU;
- DECLARE(tensorReluCPU);
- CallInst::Create(tensorReluCPU, Args, "", II);
- } else if (II->getIntrinsicID() ==
- Intrinsic::hpvm_tensor_clipped_relu) {
- // Create function call
- //-- FunctionCallee tensorClippedRelu;
- FunctionCallee tensorRelu2CPU;
- DECLARE(tensorRelu2CPU);
- CallInst::Create(tensorRelu2CPU, Args, "", II);
- } else if (II->getIntrinsicID() == Intrinsic::hpvm_tensor_tanh) {
- // Create function call
- FunctionCallee tensorTanhCPU;
- errs() << "tensorTanh Call = \n\n";
- DECLARE(tensorTanhCPU);
- // errs()<<"tensorTanh Call = "<<*tensorTanh<<"\l";
- CallInst::Create(tensorTanhCPU, Args, "", II);
- }
-
- // We can replace the call to hpvm.tensor.relu with the 1st argument
- // that, due to in place operation, now contains the result
- II->replaceAllUsesWith(II->getOperand(0));
-
- // Mark to remove at the end
- IItoRemove.push_back(II);
- } break;
-
- case Intrinsic::hpvm_tensor_softmax: { /* llvm.hpvm.tensor.softmax */
- DEBUG(errs() << F->getName() << "\t: Handling tensor softmax\n");
- // Tensor relu(a) is in place for argument a.
- Value *Op = II->getOperand(0);
-
- // Test the intrinsic operand for in place operation.
- bool inplace = isValidOperandForInPlaceOperation(Op, F, N);
- // Code generation cannot continue if this is false, because the target
- // only provides an in place operation
- assert(inplace &&
- "Operand not valid for in place operation. Code gen aborted.\n");
-
- // Argument list for the runtime call
- std::vector<Value *> Args;
- Args.push_back(II->getOperand(0));
-
- // Create function call
- FunctionCallee tensorSoftmaxCPU;
- DECLARE(tensorSoftmaxCPU);
- CallInst::Create(tensorSoftmaxCPU, Args, "", II);
- // We can replace the call to hpvm.tensor.softmax with the 1st argument
- // that, due to in place operation, now contains the result
- II->replaceAllUsesWith(II->getOperand(0));
-
- // Mark to remove at the end
- IItoRemove.push_back(II);
- } break;
-
- default:
- llvm_unreachable("Unknown VISC Intrinsic!");
- break;
- }
- }
- }
-
- // We need to do this explicitly: DCE pass may not remove them.
- // Traverse the vector backwards, otherwise definitions are deleted while
- // their subsequent uses are still around.
- for (std::vector<IntrinsicInst *>::reverse_iterator ri = IItoRemove.rbegin(),
- re = IItoRemove.rend();
- ri != re; ++ri) {
- DEBUG(errs() << "Erasing: " << **ri << "\n");
- errs() << "Erasing: " << **ri << "\n";
- (*ri)->eraseFromParent();
- }
-
- return;
-}
-
-bool DFG2LLVM_ReplaceApproxHPVMIntrinsicsWithFCalls::runOnModule(Module &M) {
- errs() << "\nDFG2LLVM_ReplaceApproxHPVMIntrinsicsWithFCalls PASS\n";
-
- // Get the BuildDFG Analysis Results:
- // - Dataflow graph
- BuildDFG &DFG = getAnalysis<BuildDFG>();
-
- // Get the In Place Analysis Results
- InPlaceDFGAnalysis::InPlaceDFGParameter IPP =
- (getAnalysis<InPlaceDFGAnalysisWrapper>()).getIPP();
- // Print results
- printInPlaceDFGParameter(IPP);
-
- std::vector<DFInternalNode *> Roots = DFG.getRoots();
-
- // Visitor for Code Generation Graph Traversal
- CGT_ReplaceApproxHPVMIntrinsicsWithFCalls *CGTVisitor =
- new CGT_ReplaceApproxHPVMIntrinsicsWithFCalls(M, DFG, IPP);
-
- // Iterate over all the DFGs and produce code for each one of them
- for (auto rootNode : Roots) {
- // Initiate code generation for root DFNode
- CGTVisitor->visit(rootNode);
- }
-
- // TODO: Edit module epilogue to remove the VISC intrinsic declarations
- delete CGTVisitor;
-
- return true;
-}
-
-/******************************************************************************
- * Helper functions *
- ******************************************************************************/
-
-} // End of namespace
-
-char DFG2LLVM_ReplaceApproxHPVMIntrinsicsWithFCalls::ID = 0;
-static RegisterPass<DFG2LLVM_ReplaceApproxHPVMIntrinsicsWithFCalls> X("replace-intrinsics",
- "Replace ApproxHPVM intrinsics with LLVM calls",
- false /* does not modify the CFG */,
- true /* transformation, *
- * not just analysis */);
diff --git a/hpvm/lib/Transforms/ReplaceIntrinsics/ReplaceIntrinsics.exports b/hpvm/lib/Transforms/ReplaceIntrinsics/ReplaceIntrinsics.exports
deleted file mode 100644
index 139597f9cb07c5d48bed18984ec4747f4b4f3438..0000000000000000000000000000000000000000
--- a/hpvm/lib/Transforms/ReplaceIntrinsics/ReplaceIntrinsics.exports
+++ /dev/null
@@ -1,2 +0,0 @@
-
-