diff --git a/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp b/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
index c402848cbd1328ba5a0e790586e4030ac9ca6f2c..7df93bd099e837335656eb6f5f30aa34e4cb0f41 100644
--- a/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
+++ b/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
@@ -55,7 +55,7 @@ public:
class Kernel {
public:
Kernel(Function* _KF, std::vector<unsigned> _inArgMap =
- std::vector<unsigned>(), unsigned _gridDim = 0, std::vector<Value*>
+ std::vector<unsigned>(), unsigned _gridDim = 0, std::vector<Value*>
_globalWGSize = std::vector<Value*>(),
unsigned _blockDim = 0,
std::vector<Value*> _localWGSize = std::vector<Value*>())
@@ -79,11 +79,17 @@ public:
std::vector<unsigned> getInArgMap() {
return inArgMap;
}
+
+ bool hasLocalWG() {
+ return blockDim == 0;
+ }
};
// Helper function declarations
static void getExecuteNodeParams(Value* &, Value* &, Value* &, Kernel*,
ValueToValueMapTy&, Instruction*);
+static Value* genWorkGroupPtr(std::vector<Value*>&, ValueToValueMapTy&,
+ Instruction*, const Twine& WGName = "WGSize");
static bool hasAttribute(Function*, unsigned, Attribute::AttrKind);
static std::string getPTXFilename(const Module&);
static std::string getFilenameFromModule(const Module& M);
@@ -723,6 +729,10 @@ void CodeGenTraversal::codeGen(DFLeafNode* N) {
int pLevel = PNode->getLevel();
int pReplFactor = PNode->getNumOfDim();
+ // Choose parent node as kernel launch if:
+ // (1) Parent is the top level node i.e., Root of DFG
+ // OR
+ // (2) Parent does not have multiple instances
if (!pLevel || !pReplFactor) {
KernelLaunchNode = PNode;
kernel = new Kernel(NULL, N->getInArgMap(), N->getNumOfDim(), N->getDimLimits());
@@ -742,7 +752,8 @@ void CodeGenTraversal::codeGen(DFLeafNode* N) {
//}
}
else {
- errs() << "*************** Entering else part **************\n";
+ // Converting a 2-level DFG to opencl kernel
+ errs() << "*************** Kernel Gen: 2-Level Hierarchy **************\n";
KernelLaunchNode = PNode->getParent();
assert((PNode->getNumOfDim() == N->getNumOfDim()) && "Dimension number must match");
// Contains the instructions generating the kernel configuration parameters
@@ -1272,53 +1283,65 @@ void CodeGenTraversal::transformFunctionToVoid(Function* F) {
static void getExecuteNodeParams(Value* &workDim, Value* &LocalWGPtr, Value*
&GlobalWGPtr, Kernel* kernel, ValueToValueMapTy& VMap, Instruction* IB) {
- // Get int64_t and or ease of use
- Type* Int64Ty = Type::getInt64Ty(getGlobalContext());
-
// Assign number of dimenstions a constant value
workDim = ConstantInt::get(Type::getInt32Ty(getGlobalContext()), kernel->gridDim);
- // For now, local work group size if null
- LocalWGPtr = Constant::getNullValue(Type::getInt64PtrTy(getGlobalContext()));
+ // If local work group size if null
+ if(!kernel->hasLocalWG()) {
+ LocalWGPtr = Constant::getNullValue(Type::getInt64PtrTy(getGlobalContext()));
+ }
+ else {
+ LocalWGPtr = genWorkGroupPtr(kernel->localWGSize, VMap, IB, "LocalWGSize");
+ }
- // Global Work Group type is [#dim x i64]
- Type* GlobalWGTy = ArrayType::get(Int64Ty, kernel->gridDim);
+ GlobalWGPtr = genWorkGroupPtr(kernel->globalWGSize, VMap, IB, "GlobalWGSize");
+ DEBUG(errs() << "Pointer to global work group: " << *GlobalWGPtr << "\n");
+}
+
+static Value* genWorkGroupPtr(std::vector<Value*>& WGSize, ValueToValueMapTy& VMap, Instruction* IB, const Twine& WGName) {
+ Value* WGPtr;
+ // Get int64_t and or ease of use
+ Type* Int64Ty = Type::getInt64Ty(getGlobalContext());
+
+ // Work Group type is [#dim x i64]
+ Type* WGTy = ArrayType::get(Int64Ty, WGSize.size());
// Allocate space of Global work group data on stack and get pointer to
// first element.
- AllocaInst* GlobalWG = new AllocaInst(GlobalWGTy, "GlobalWGSize", IB);
- GlobalWGPtr = BitCastInst::CreatePointerCast(GlobalWG, Int64Ty->getPointerTo(), GlobalWG->getName()+".0", IB);
- Value* nextDim = GlobalWGPtr;
- DEBUG(errs() << *GlobalWGPtr << "\n");
+ AllocaInst* WG = new AllocaInst(WGTy, WGName, IB);
+ WGPtr = BitCastInst::CreatePointerCast(WG, Int64Ty->getPointerTo(), WG->getName()+".0", IB);
+ Value* nextDim = WGPtr;
+ DEBUG(errs() << *WGPtr << "\n");
// Iterate over the number of dimensions and store the global work group
// size in that dimension
- for(unsigned i=0; i < kernel->gridDim; i++) {
- assert(kernel->globalWGSize[i]->getType()->isIntegerTy() && "Dimension not an integer type!");
- if(kernel->globalWGSize[i]->getType() != Int64Ty) {
+ for(unsigned i=0; i < WGSize.size(); i++) {
+ assert(WGSize[i]->getType()->isIntegerTy() && "Dimension not an integer type!");
+ if(WGSize[i]->getType() != Int64Ty) {
// If number of dimensions are mentioned in any other integer format,
// generate code to extend it to i64. We need to use the mapped value in
// the new generated function, hence the use of VMap
- kernel->globalWGSize[i] = BitCastInst::CreateIntegerCast(VMap[kernel->globalWGSize[i]], Int64Ty, true, "", IB);
- StoreInst* SI = new StoreInst(kernel->globalWGSize[i], nextDim, IB);
+ // FIXME: Why are we changing the kernel WGSize vector here?
+ WGSize[i] = BitCastInst::CreateIntegerCast(VMap[WGSize[i]], Int64Ty, true, "", IB);
+ StoreInst* SI = new StoreInst(WGSize[i], nextDim, IB);
DEBUG(errs() << "\tZero extended work group size: " << *SI << "\n");
} else {
// Store the value representing work group size in ith dimension on
// stack
- StoreInst* SI = new StoreInst(VMap[kernel->globalWGSize[i]], nextDim, IB);
+ StoreInst* SI = new StoreInst(VMap[WGSize[i]], nextDim, IB);
DEBUG(errs() << "\t Work group size: " << *SI << "\n");
}
- if(i+1 < kernel->gridDim) {
+ if(i+1 < WGSize.size()) {
// Move to next dimension
GetElementPtrInst* GEP = GetElementPtrInst::Create(nextDim,
ArrayRef<Value*>(ConstantInt::get(Int64Ty, 1)),
- GlobalWG->getName()+"."+Twine(i+1),
+ WG->getName()+"."+Twine(i+1),
IB);
DEBUG(errs() << "\tPointer to next dimension on stack: " << *GEP << "\n");
nextDim = GEP;
}
}
+ return WGPtr;
- DEBUG(errs() << "Pointer to global work group: " << *GlobalWGPtr << "\n");
}
// Find if argument has the given attribute