diff --git a/hpvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp b/hpvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
index 97e0f3411b7d74e97b1b6aca3d9f7d09f0650d72..642202bf055adcf534b7f1b44b3ab71858347ecb 100644
--- a/hpvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
+++ b/hpvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
@@ -953,8 +953,13 @@ void CGT_NVPTX::codeGen(DFLeafNode* N) {
// (2) Parent does not have multiple instances
errs() << "pLevel = " << pLevel << "\n";
errs() << "pReplFactor = " << pReplFactor << "\n";
- if (!pLevel || !pReplFactor) {
+ assert((pLevel > 0) && "Root not allowed to be chosen as Kernel Node.");
+
+ // Only these options are supported
+ enum XLevelHierarchy{ONE_LEVEL, TWO_LEVEL} SelectedHierarchy;
+ if(pLevel == 1 || !pReplFactor) {
errs() << "*************** Kernel Gen: 1-Level Hierarchy **************\n";
+ SelectedHierarchy = ONE_LEVEL;
KernelLaunchNode = PNode;
kernel = new Kernel(NULL,
N,
@@ -967,6 +972,8 @@ void CGT_NVPTX::codeGen(DFLeafNode* N) {
else {
// Converting a 2-level DFG to opencl kernel
errs() << "*************** Kernel Gen: 2-Level Hierarchy **************\n";
+ assert((pLevel >= 2) && "Selected node not nested deep enough to be Kernel Node.");
+ SelectedHierarchy = TWO_LEVEL;
KernelLaunchNode = PNode->getParent();
assert((PNode->getNumOfDim() == N->getNumOfDim()) && "Dimension number must match");
// Contains the instructions generating the kernel configuration parameters
@@ -982,7 +989,7 @@ void CGT_NVPTX::codeGen(DFLeafNode* N) {
}
- std::vector<IntrinsicInst *> IItoRemove;
+ std::vector<Instruction *> IItoRemove;
BuildDFG::HandleToDFNode Leaf_HandleToDFNodeMap;
// Get the function associated with the dataflow node
@@ -1127,6 +1134,33 @@ void CGT_NVPTX::codeGen(DFLeafNode* N) {
F_nvptx = changeArgAddrspace(F_nvptx, SharedMemArgs, SHARED_ADDRSPACE);
F_nvptx = changeArgAddrspace(F_nvptx, GlobalMemArgs, GLOBAL_ADDRSPACE);
+// Function to replace call instructions to functions in the kernel
+ std::map<Function *, Function *> OrgToClonedFuncMap;
+ std::vector<Function *> FuncToBeRemoved;
+ auto CloneAndReplaceCall = [&] (CallInst *CI, Function *OrgFunc) {
+ Function* NewFunc;
+ // Check if the called function has already been cloned before.
+ auto It = OrgToClonedFuncMap.find(OrgFunc);
+ if(It == OrgToClonedFuncMap.end()) {
+ ValueToValueMapTy VMap;
+ NewFunc = CloneFunction(OrgFunc, VMap);
+ OrgToClonedFuncMap[OrgFunc] = NewFunc;
+ FuncToBeRemoved.push_back(NewFunc);
+ } else {
+ NewFunc = (*It).second;
+ }
+ // Replace the calls to this function
+ std::vector<Value*> args;
+ for(unsigned i=0; i < CI->getNumArgOperands(); i++) {
+ args.push_back(CI->getArgOperand(i));
+ }
+ CallInst* Inst = CallInst::Create(NewFunc, args,
+ OrgFunc->getReturnType()->isVoidTy()? "" : CI->getName(), CI);
+ CI->replaceAllUsesWith(Inst);
+ IItoRemove.push_back(CI);
+ return NewFunc;
+ };
+
// Go through all the instructions
for (inst_iterator i = inst_begin(F_nvptx), e = inst_end(F_nvptx); i != e; ++i) {
@@ -1212,16 +1246,12 @@ void CGT_NVPTX::codeGen(DFLeafNode* N) {
// The following is to find which function to call
Function * OpenCLFunction;
- int parentLevel = N->getParent()->getLevel();
- int parentReplFactor = N->getParent()->getNumOfDim();
- DEBUG(errs() << "Parent Level = " << parentLevel << "\n");
- DEBUG(errs() << "Parent Repl factor = " << parentReplFactor << "\n");
FunctionType* FT =
FunctionType::get(Type::getInt64Ty(KernelM->getContext()),
Type::getInt32Ty(KernelM->getContext()),
false);
- if ((!parentLevel || !parentReplFactor) && ArgDFNode == N) {
+ if (SelectedHierarchy == ONE_LEVEL && ArgDFNode == N) {
// We only have one level in the hierarchy or the parent node is not
// replicated. This indicates that the parent node is the kernel
// launch, so we need to specify a global id.
@@ -1296,14 +1326,12 @@ void CGT_NVPTX::codeGen(DFLeafNode* N) {
// The following is to find which function to call
Function * OpenCLFunction;
- int parentLevel = ParentDFNode->getLevel();
- int parentReplFactor = ParentDFNode->getNumOfDim();
FunctionType* FT =
FunctionType::get(Type::getInt64Ty(KernelM->getContext()),
Type::getInt32Ty(KernelM->getContext()),
false);
- if ((N == ArgDFNode) && (!parentLevel || !parentReplFactor)) {
+ if (N == ArgDFNode && SelectedHierarchy == ONE_LEVEL) {
// We only have one level in the hierarchy or the parent node is not
// replicated. This indicates that the parent node is the kernel
// launch, so the instances are global_size (gridDim x blockDim)
@@ -1412,9 +1440,9 @@ void CGT_NVPTX::codeGen(DFLeafNode* N) {
&& "Only handling sin(float) and cos(float)!");
std::string name;
if(II->getIntrinsicID() == Intrinsic::sin)
- name = "_Z3sinf";
+ name = "sin";
else
- name = "_Z3cosf";
+ name = "cos";
FunctionType* SinCosFT = FunctionType::get(II->getType(),
Type::getFloatTy(KernelM->getContext()),
@@ -1465,11 +1493,16 @@ void CGT_NVPTX::codeGen(DFLeafNode* N) {
}
else {
- // Clone the function
- ValueToValueMapTy VMap;
- Function* newCalleeF = CloneFunction(calleeF, VMap);
- newCalleeF->removeFromParent(); //TODO: MARIA check
- KernelM->getFunctionList().push_back(newCalleeF);
+ // Check if the called function has already been cloned before.
+ Function *NewFunc = CloneAndReplaceCall(CI, calleeF);
+ // Iterate over the new function to see if it calls any other functions
+ // in the module.
+ for(inst_iterator i = inst_begin(NewFunc), e = inst_end(NewFunc); i != e; ++i) {
+ if(auto *Call = dyn_cast<CallInst>(&*i)) {
+ Function *CalledFunc = cast<Function>(Call->getCalledValue()->stripPointerCasts());
+ CloneAndReplaceCall(Call, CalledFunc);
+ }
+ }
}
//TODO: how to handle address space qualifiers in load/store
}
@@ -1480,10 +1513,15 @@ void CGT_NVPTX::codeGen(DFLeafNode* N) {
// have assumed theworst memory behaviour for these function calls
// 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");
- (*ri)->eraseFromParent();
+ for (auto *I : IItoRemove) {
+ DEBUG(errs() << "Erasing: " << *I << "\n");
+ I->eraseFromParent();
+ }
+
+ // Removed the cloned functions from the parent module into the new module
+ for(auto *F : FuncToBeRemoved) {
+ F->removeFromParent(); //TODO: MARIA check
+ KernelM->getFunctionList().push_back(F);
}
addCLMetadata(F_nvptx);
@@ -1936,7 +1974,7 @@ static Value* genWorkGroupPtr(Module &M, std::vector<Value*> WGSize, ValueToValu
// Get generated PTX binary name
static std::string getPTXFilename(const Module& M) {
std::string moduleID = M.getModuleIdentifier();
- moduleID.append(".nvptx.s");
+ moduleID.append(".kernels.cl");
return moduleID;
}