diff --git a/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp b/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
index d7cdfa2e40cc09daf0987900d4619ac923d6e1dc..425cabd0b485db37c9ea1481e05905f43aef8523 100644
--- a/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
+++ b/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
@@ -443,8 +443,8 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileNa
DEBUG(errs() << "Initializing commandQ" << "\n");
// Initialize command queue
Value* fileStr = getStringPointer(FileName, RI, "Filename");
- errs() << *fileStr << "\n";
- errs() << "Generating code for kernel - " << kernel->KernelFunction->getName()<< "\n";
+ 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");
Value* LaunchInstArgs[] = {fileStr, kernelStr};
@@ -458,7 +458,6 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileNa
// 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++) {
@@ -469,6 +468,7 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileNa
// type on target machine, but for pointers, the size of data would be the
// next integer argument
if(inputVal->getType()->isPointerTy()) {
+ // Pointer Input
// CheckAttribute
Value* isOutput = (hasAttribute(CF, i, Attribute::Out))? True : False;
Value* isInput = ((hasAttribute(CF, i, Attribute::Out))
@@ -482,11 +482,11 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileNa
errs() << *A << " is an INPUT argument\n";
}
+
Value* inputValI8Ptr = CastInst::CreatePointerCast(inputVal,
Type::getInt8PtrTy(M.getContext()),
inputVal->getName()+".i8ptr",
RI);
- // Pointer Input
Value* inputSize = getInValueAt(C, i+1, F_X86, RI);
assert(inputSize->getType() == Type::getInt64Ty(M.getContext())
&& "Pointer type input must always be followed by size (integer type)");
@@ -503,7 +503,8 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileNa
// memory to read device memory later
if(isOutput == True) OutputPointers.push_back(OutputPtr(inputValI8Ptr, d_ptr, inputSize));
}
- else { // Scalar Input
+ else {
+ // Scalar Input
// Store the scalar value on stack and then pass the pointer to its
// location
AllocaInst* inputValPtr = new AllocaInst(inputVal->getType(), inputVal->getName()+".ptr", RI);
@@ -525,15 +526,6 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileNa
}
DEBUG(errs() << "Setup output edges of node and insert visc api\n");
- // Setup output
- // FIXME: Note - There is a tricky question. In X86 we do not need to care
- // about pointer inputs which modify data in memory implicitly (without
- // showing it as output). There is no extra cost needed to handle such inputs
- // For PTX, we need to read back such data from device memory to host memory.
- // The cost is huge and hence we need to differentiate between readonly
- // pointer inputs vs read/write pointer inputs. Currently supporting only a
- // simple model in which all input edges are readonly and output is
- // writeonly.
// Set output
StructType* OutputTy = C->getOutputType();
@@ -554,8 +546,6 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileNa
// Enqueue kernel
// Need work dim, localworksize, globalworksize
- // FIXME: Talk to DFG2LLVM_PTX pass to figure out the workdim, loacal work
- // size and global work size
// Allocate size_t[numDims] space on stack. Store the work group sizes and
// pass it as an argument to ExecNode
@@ -571,11 +561,13 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileNa
"event."+CF->getName(),
RI);
DEBUG(errs() << "Execute Node Call: " << *Event << "\n");
+
// Wait for Kernel to Finish
CallInst::Create(llvm_visc_ptx_wait,
ArrayRef<Value*>(GraphID),
"",
RI);
+
// Read Output Struct
Value* GetOutputArgs[] = {GraphID,
Constant::getNullValue(Type::getInt8PtrTy(M.getContext())),
@@ -589,7 +581,8 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileNa
// 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, "", RI);
+ Value* KernelOutput = new LoadInst(BI, "output."+CF->getName(), RI);
+ OutputMap[C] = KernelOutput;
// Read all the pointer arguments which had side effects i.e., had out
// attribute
@@ -602,37 +595,6 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileNa
ArrayRef<Value*>(GetOutputArgs, 4),
"", RI);
}
- /*for(unsigned i=0; i < OutputTy->getNumElements(); i++) {
- Type* elemTy = OutputTy->getElementType(i);
- if(elemTy->isPointerTy()) {
- // Pointer type
- assert(OutputTy->getElementType(i+1) == Type::getInt64Ty(M.getContext())
- && "Every Pointer type must be followed by an integer");
- ExtractValueInst* d_ptr = ExtractValueInst::Create(KernelOutput, ArrayRef<unsigned>(i), "", RI);
- // Change d_ptr to i8*
- CastInst* d_ptr_i8 = BitCastInst::CreatePointerCast(d_ptr, Type::getInt8PtrTy(M.getContext()), "", RI);
- ExtractValueInst* len = ExtractValueInst::Create(KernelOutput, ArrayRef<unsigned>(i+1), "", RI);
- // GetOutputPtr call
- Value* GetOutputArgs[] = {GraphID,
- d_ptr_i8,
- len};
- CallInst* h_ptr_i8 = CallInst::Create(llvm_visc_ptx_getOutput,
- ArrayRef<Value*>(GetOutputArgs, 3),
- "",
- RI);
- // Change h_ptr to correct type
- CastInst* h_ptr = CastInst::CreatePointerCast(h_ptr_i8,
- cast<StructType>(KernelOutput->getType())->getElementType(i),
- "",
- RI);
- KernelOutput = InsertValueInst::Create(KernelOutput, h_ptr, ArrayRef<unsigned>(i), "", RI);
-
- }
- }*/
-
- // Prepare output
- KernelOutput->setName("output."+CF->getName());
- OutputMap[C] = KernelOutput;
DEBUG(errs() << "*** Generating epilogue code for the function****\n");
// Generate code for output bindings
@@ -681,6 +643,7 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileNa
IdxList.push_back(i);
retVal = InsertValueInst::Create(retVal, inputVal, IdxList, "", RI);
}
+
DEBUG(errs() << "Extracted all\n");
retVal->setName("output");
ReturnInst* newRI = ReturnInst::Create(F_X86->getContext(), retVal);