diff --git a/llvm/lib/Transforms/ClearDFG/ClearDFG.cpp b/llvm/lib/Transforms/ClearDFG/ClearDFG.cpp
index 05869fa0dfca5b742d828f330e2f17fd21b28aa6..4b392b2c966d802664598c3f2d09331910d6e555 100644
--- a/llvm/lib/Transforms/ClearDFG/ClearDFG.cpp
+++ b/llvm/lib/Transforms/ClearDFG/ClearDFG.cpp
@@ -65,12 +65,12 @@ public:
virtual void visit(DFInternalNode* N) {
// Follows a bottom-up approach for code generation.
// First generate code for all the child nodes
- DEBUG(errs() << "Erasing Node (I) - " << N->getFuncPointer()->getName() << "\n");
for(DFGraph::children_iterator i = N->getChildGraph()->begin(),
e = N->getChildGraph()->end(); i != e; ++i) {
DFNode* child = *i;
child->applyDFNodeVisitor(*this);
}
+ DEBUG(errs() << "Erasing Node (I) - " << N->getFuncPointer()->getName() << "\n");
// Generate code for this internal node now. This way all the cloned
// functions for children exist.
deleteNode(N);
@@ -80,7 +80,7 @@ public:
virtual void visit(DFLeafNode* N) {
DEBUG(errs() << "Erasing Node (L) - " << N->getFuncPointer()->getName() << "\n");
deleteNode(N);
- //errs() << "DONE: Generating Code for Node (L) - " << N->getFuncPointer()->getName() << "\n";
+ errs() << "DONE" << "\n";
}
};
diff --git a/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp b/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
index 190b9af288bbd77f11e720fd979dc6eb0832bb82..869ca5d5927ffb59db62f1056f77ea357daa77d0 100644
--- a/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
+++ b/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
@@ -54,12 +54,12 @@ public:
// calls
class Kernel {
public:
- Kernel(Function* _KF, std::vector<unsigned> _inArgMap =
+ Kernel(Function* _KF, DFLeafNode* _KLeafNode,std::vector<unsigned> _inArgMap =
std::vector<unsigned>(), unsigned _gridDim = 0, std::vector<Value*>
_globalWGSize = std::vector<Value*>(),
unsigned _blockDim = 0,
std::vector<Value*> _localWGSize = std::vector<Value*>())
- : KernelFunction(_KF), inArgMap(_inArgMap),
+ : KernelFunction(_KF), KernelLeafNode(_KLeafNode), inArgMap(_inArgMap),
gridDim(_gridDim), globalWGSize(_globalWGSize),
blockDim(_blockDim), localWGSize(_localWGSize) {
@@ -70,6 +70,7 @@ public:
}
Function* KernelFunction;
+ DFLeafNode* KernelLeafNode;
std::vector<unsigned> inArgMap;
unsigned gridDim;
unsigned blockDim;
@@ -169,7 +170,7 @@ private:
Argument* getArgumentAt(Function* F, unsigned offset);
Value* getInValueAt(DFNode* Child, unsigned i, Function* ParentF_X86,
Instruction* InsertBefore);
- void insertRuntimeCalls(DFInternalNode* N, const Twine& FileName);
+ void insertRuntimeCalls(DFInternalNode* N, Kernel* K, const Twine& FileName);
void codeGen(DFInternalNode* N);
void codeGen(DFLeafNode* N);
@@ -318,16 +319,16 @@ void CodeGenTraversal::addIdxDimArgs(Function* F) {
/* Traverse the function F argument list to get argument at offset*/
Argument* CodeGenTraversal::getArgumentAt(Function* F, unsigned offset) {
+ DEBUG(errs() << "Finding argument " << offset << ":\n");
assert((F->getFunctionType()->getNumParams() > offset && offset >= 0)
&& "Invalid offset to access arguments!");
-
Argument* arg;
Function::arg_iterator i = F->arg_begin(), e = F->arg_end();
for(; offset != 0 && i!=e; i++) {
offset--;
}
arg = i;
- DEBUG(errs() << *arg <<"\n");
+ DEBUG(errs() << "\t" << *arg <<"\n");
return arg;
}
@@ -388,7 +389,7 @@ Value* CodeGenTraversal::getStringPointer(const Twine& S, Instruction* IB, const
// used to generate a function to associate with this leaf node. The function
// is responsible for all the memory allocation/transfer and invoking the
// kernel call on the device
-void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileName) {
+void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, Kernel* K, const Twine& FileName) {
// Check if clone already exists. If it does, it means we have visited this
// function before.
assert(N->getGenFunc() == NULL && "Code already generated for this node");
@@ -398,7 +399,7 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileNa
Value* False = ConstantInt::get(Type::getInt1Ty(M.getContext()), 0);
// If kernel struct has not been initialized with kernel function, then fail
- assert(kernel != NULL && "No kernel found!!");
+ assert(K != NULL && "No kernel found!!");
DEBUG(errs() << "Generating kernel call code\n");
@@ -441,6 +442,8 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileNa
if(!N->isRoot())
addIdxDimArgs(F_X86);
+ /* TODO: Use this code to verufy if this is a good pattern for PTX kernel
+
// Sort children in topological order before code generation for kernel call
N->getChildGraph()->sortChildren();
@@ -464,6 +467,8 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileNa
assert(C->isDummyNode() == false && "Internal Node only contains dummy nodes!");
Function* CF = C->getFuncPointer();
+ */
+ Function* KF = K->KernelLeafNode->getFuncPointer();
// Initialize context
DEBUG(errs() << "Initializing context" << "\n");
CallInst::Create(llvm_visc_ptx_initContext, None, "", RI);
@@ -472,26 +477,26 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileNa
// Initialize command queue
Value* fileStr = getStringPointer(FileName, RI, "Filename");
DEBUG(errs() << "Kernel Filename constant: " << *fileStr << "\n");
- DEBUG(errs() << "Generating code for kernel - " << kernel->KernelFunction->getName()<< "\n");
- Value* kernelStr = getStringPointer(kernel->KernelFunction->getName(), RI,"KernelName");
+ DEBUG(errs() << "Generating code for kernel - " << K->KernelFunction->getName()<< "\n");
+ Value* kernelStr = getStringPointer(K->KernelFunction->getName(), RI,"KernelName");
Value* LaunchInstArgs[] = {fileStr, kernelStr};
DEBUG(errs() << "Inserting launch call" << "\n");
CallInst* GraphID = CallInst::Create(llvm_visc_ptx_launch,
ArrayRef<Value*>(LaunchInstArgs, 2),
- "graph"+CF->getName(),
+ "graph"+KF->getName(),
RI);
DEBUG(errs() << *GraphID << "\n");
// Iterate over the required input edges of the node and use the visc-rt API
// to set inputs
DEBUG(errs() << "Iterate over input edges of node and insert visc api\n");
std::vector<OutputPtr> OutputPointers;
- for(unsigned i=0; i<CF->getFunctionType()->getNumParams(); i++) {
+ for(unsigned i=0; i<KF->getFunctionType()->getNumParams(); i++) {
// The kernel object gives us the mapping of arguments from kernel launch
// node function (F_X86) to kernel (kernel->KF)
- Value* inputVal = getArgumentAt(F_X86, kernel->getInArgMap()[i]);
+ Value* inputVal = getArgumentAt(F_X86, K->getInArgMap()[i]);
DEBUG(errs() << "\tArgument "<< i<< " = " << *inputVal << "\n");
// input value has been obtained.
@@ -502,11 +507,11 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileNa
if(inputVal->getType()->isPointerTy()) {
// Pointer Input
// CheckAttribute
- Value* isOutput = (hasAttribute(CF, i, Attribute::Out))? True : False;
- Value* isInput = ((hasAttribute(CF, i, Attribute::Out))
- && !(hasAttribute(CF, i, Attribute::In)))? False : True;
+ Value* isOutput = (hasAttribute(KF, i, Attribute::Out))? True : False;
+ Value* isInput = ((hasAttribute(KF, i, Attribute::Out))
+ && !(hasAttribute(KF, i, Attribute::In)))? False : True;
- Argument* A = getArgumentAt(CF, i);
+ Argument* A = getArgumentAt(KF, i);
if(isOutput == True) {
DEBUG(errs() << *A << " is an OUTPUT argument\n");
}
@@ -519,7 +524,7 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileNa
Type::getInt8PtrTy(M.getContext()),
inputVal->getName()+".i8ptr",
RI);
- Value* inputSize = getArgumentAt(F_X86, kernel->getInArgMap()[i+1]);
+ Value* inputSize = getArgumentAt(F_X86, K->getInArgMap()[i+1]);
assert(inputSize->getType() == Type::getInt64Ty(M.getContext())
&& "Pointer type input must always be followed by size (integer type)");
Value* setInputArgs[] = {GraphID,
@@ -560,11 +565,11 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileNa
DEBUG(errs() << "Setup output edges of node and insert visc api\n");
// Set output if struct is not an empty struct
- StructType* OutputTy = C->getOutputType();
+ StructType* OutputTy = K->KernelLeafNode->getOutputType();
Value *outputSize, *d_Output;
if(!OutputTy->isEmptyTy()) {
// Not an empty struct
- unsigned outputIndex = CF->getFunctionType()->getNumParams();
+ unsigned outputIndex = KF->getFunctionType()->getNumParams();
outputSize = ConstantExpr::getSizeOf(OutputTy);
Value* setOutputArgs[] = {GraphID,
Constant::getNullValue(Type::getInt8PtrTy(M.getContext())),
@@ -576,7 +581,7 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileNa
d_Output = CallInst::Create(llvm_visc_ptx_argument_ptr,
ArrayRef<Value*>(setOutputArgs, 6),
- "d_output."+CF->getName(),
+ "d_output."+KF->getName(),
RI);
}
@@ -586,7 +591,7 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileNa
// pass it as an argument to ExecNode
Value *workDim, *LocalWGPtr, *GlobalWGPtr;
- getExecuteNodeParams(workDim, LocalWGPtr, GlobalWGPtr, kernel, VMap, RI);
+ getExecuteNodeParams(workDim, LocalWGPtr, GlobalWGPtr, K, VMap, RI);
Value* ExecNodeArgs[] = {GraphID,
workDim,
LocalWGPtr,
@@ -594,7 +599,7 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileNa
};
CallInst* Event = CallInst::Create(llvm_visc_ptx_executeNode,
ArrayRef<Value*>(ExecNodeArgs, 4),
- "event."+CF->getName(),
+ "event."+KF->getName(),
RI);
DEBUG(errs() << "Execute Node Call: " << *Event << "\n");
@@ -613,13 +618,13 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileNa
};
CallInst* h_Output = CallInst::Create(llvm_visc_ptx_getOutput,
ArrayRef<Value*>(GetOutputArgs, 4),
- "h_output."+CF->getName()+".addr",
+ "h_output."+KF->getName()+".addr",
RI);
// Read each device pointer listed in output struct
// Load the output struct
- CastInst* BI = BitCastInst::CreatePointerCast(h_Output, CF->getReturnType()->getPointerTo(), "output.ptr", RI);
- Value* KernelOutput = new LoadInst(BI, "output."+CF->getName(), RI);
- OutputMap[C] = KernelOutput;
+ CastInst* BI = BitCastInst::CreatePointerCast(h_Output, KF->getReturnType()->getPointerTo(), "output.ptr", RI);
+ Value* KernelOutput = new LoadInst(BI, "output."+KF->getName(), RI);
+ OutputMap[K->KernelLeafNode] = KernelOutput;
}
// Read all the pointer arguments which had side effects i.e., had out
@@ -639,7 +644,7 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileNa
DEBUG(errs() << "*** Generating epilogue code for the function****\n");
// Generate code for output bindings
// Get Exit node
- C = N->getChildGraph()->getExit();
+ DFNode* C = N->getChildGraph()->getExit();
// Get OutputType of this node
StructType* OutTy = N->getOutputType();
Value *retVal = UndefValue::get(F_X86->getReturnType());
@@ -665,6 +670,10 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileNa
else {
// edge is from a internal node
// Check - code should already be generated for this source dfnode
+ // FIXME: Since the 2-level kernel code gen has aspecific structure, we
+ // can assume the SrcDF is same as Kernel Leaf node.
+ // Use outArgMap to get correct mapping
+ SrcDF = K->KernelLeafNode;
assert(OutputMap.count(SrcDF)
&& "Source node call not found. Dependency violation!");
@@ -707,7 +716,7 @@ void CodeGenTraversal::codeGen(DFInternalNode* N) {
// Now the remaining nodes to be visited should be ignored
KernelLaunchNode = NULL;
errs() << "Insert Runtime calls\n";
- insertRuntimeCalls(N, getPTXFilename(M));
+ insertRuntimeCalls(N, kernel, getPTXFilename(M));
writeKernelsModule();
} else {
@@ -770,7 +779,7 @@ void CodeGenTraversal::codeGen(DFLeafNode* N) {
// (2) Parent does not have multiple instances
if (!pLevel || !pReplFactor) {
KernelLaunchNode = PNode;
- kernel = new Kernel(NULL, N->getInArgMap(), N->getNumOfDim(), N->getDimLimits());
+ kernel = new Kernel(NULL, N, N->getInArgMap(), N->getNumOfDim(), N->getDimLimits());
}
else {
// Converting a 2-level DFG to opencl kernel
@@ -779,6 +788,7 @@ void CodeGenTraversal::codeGen(DFLeafNode* N) {
assert((PNode->getNumOfDim() == N->getNumOfDim()) && "Dimension number must match");
// Contains the instructions generating the kernel configuration parameters
kernel = new Kernel(NULL, // kernel function
+ N, // kernel leaf node
N->getInArgMap(), // kenel argument mapping
PNode->getNumOfDim(), // gridDim
PNode->getDimLimits(),// grid size
@@ -1316,9 +1326,17 @@ static void getExecuteNodeParams(Value* &workDim, Value* &LocalWGPtr, Value*
LocalWGPtr = Constant::getNullValue(Type::getInt64PtrTy(getGlobalContext()));
}
else {
+ for(unsigned i = 0; i < kernel->localWGSize.size(); i++) {
+ if(isa<Argument>(kernel->localWGSize[i]))
+ kernel->localWGSize[i] = VMap[kernel->localWGSize[i]];
+ }
LocalWGPtr = genWorkGroupPtr(kernel->localWGSize, VMap, IB, "LocalWGSize");
}
+ for(unsigned i = 0; i < kernel->globalWGSize.size(); i++) {
+ if(isa<Argument>(kernel->globalWGSize[i]))
+ kernel->globalWGSize[i] = VMap[kernel->globalWGSize[i]];
+ }
// For OpenCL, global work group size is the total bumber of instances in each
// dimension. So, multiply local and global dim limits.
std::vector<Value*> globalWGSizeInsts;
@@ -1350,16 +1368,7 @@ static Value* genWorkGroupPtr(std::vector<Value*> WGSize, ValueToValueMapTy& VMa
// size in that dimension
for(unsigned i=0; i < WGSize.size(); i++) {
assert(WGSize[i]->getType()->isIntegerTy() && "Dimension not an integer type!");
-
- // If WGSize[i] is not a constant or a instruction, use mapped value in the new function
- Value* WGSizeMapped;
- if(isa<Argument>(WGSize[i]))
- WGSizeMapped = VMap[WGSize[i]];
- else {
- WGSizeMapped = WGSize[i];
- errs() << "Mapping value is not required: ";
- errs() << *WGSize[i] << "\n";
- }
+
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
@@ -1367,8 +1376,7 @@ static Value* genWorkGroupPtr(std::vector<Value*> WGSize, ValueToValueMapTy& VMa
// FIXME: Why are we changing the kernel WGSize vector here?
errs() << "Not i64. Zero extend required.\n";
errs() << *WGSize[i] << "\n";
- errs() << *WGSizeMapped << "\n";
- CastInst* CI = BitCastInst::CreateIntegerCast(WGSizeMapped, Int64Ty, true, "", IB);
+ CastInst* CI = BitCastInst::CreateIntegerCast(WGSize[i], Int64Ty, true, "", IB);
errs() << "Bitcast done.\n";
StoreInst* SI = new StoreInst(CI, nextDim, IB);
errs() << "Zero extend done.\n";
@@ -1376,7 +1384,7 @@ static Value* genWorkGroupPtr(std::vector<Value*> WGSize, ValueToValueMapTy& VMa
} else {
// Store the value representing work group size in ith dimension on
// stack
- StoreInst* SI = new StoreInst(WGSizeMapped, nextDim, IB);
+ StoreInst* SI = new StoreInst(WGSize[i], nextDim, IB);
DEBUG(errs() << "\t Work group size: " << *SI << "\n");
}