From e7c3ae27f2b8c9f20828f04dfd2125176911a9f6 Mon Sep 17 00:00:00 2001
From: Maria Kotsifakou <kotsifa2@illinois.edu>
Date: Tue, 25 Nov 2014 23:02:10 +0000
Subject: [PATCH] Computing the local and global sizes for the two-level
hierarchy case in dfg2llvm-nvptx.
---
.../DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp | 40 ++++++++++---------
1 file changed, 21 insertions(+), 19 deletions(-)
diff --git a/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp b/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
index 6140b97c42..93059ccde2 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, unsigned _gridDim = 0, std::vector<Value*>
- _globalWGSize = std::vector<Value*>(),
- std::vector<unsigned>* _inArgMap = NULL, unsigned _blockDim = 0,
+ Kernel(Function* _KF, std::vector<unsigned>* _inArgMap = NULL,
+ unsigned _gridDim = 0, std::vector<Value*>
+ _globalWGSize = std::vector<Value*>(), unsigned _blockDim = 0,
std::vector<Value*> _localWGSize = std::vector<Value*>())
- : KernelFunction(_KF),
- gridDim(_gridDim), globalWGSize(_globalWGSize), inArgMap(_inArgMap),
+ : KernelFunction(_KF), inArgMap(_inArgMap),
+ gridDim(_gridDim), globalWGSize(_globalWGSize),
blockDim(_blockDim), localWGSize(_localWGSize) {
assert(gridDim == globalWGSize.size()
@@ -69,11 +69,11 @@ public:
}
Function* KernelFunction;
+ std::vector<unsigned>* inArgMap;
unsigned gridDim;
unsigned blockDim;
std::vector<Value*> globalWGSize;
std::vector<Value*> localWGSize;
- std::vector<unsigned>* inArgMap;
};
// Helper function declarations
@@ -716,8 +716,8 @@ void CodeGenTraversal::codeGen(DFLeafNode* N) {
if (!pLevel || !pReplFactor) {
KernelLaunchNode = PNode;
- kernel = new Kernel(NULL, N->getNumOfDim(), N->getDimLimits(), N->getInArgMap());
- // TODO: Find a better way of choosing parameters
+ kernel = new Kernel(NULL, N->getInArgMap(), N->getNumOfDim(), N->getDimLimits());
+ // TODO: Find a good way of choosing parameters - is this required?
//kernel->gridDim = N->getNumOfDim();
//kernel->blockDim = N->getNumOfDim();
//kernel->globalWGSize = N->getDimLimits();
@@ -736,17 +736,19 @@ void CodeGenTraversal::codeGen(DFLeafNode* N) {
errs() << "*************** Entering else part **************\n";
KernelLaunchNode = PNode->getParent();
assert((PNode->getNumOfDim() == N->getNumOfDim()) && "Dimension number must match");
- /*
- KernelLaunchNode = PNode->getParent();
- kernel->gridDim = PNode->getNumOfDim();
- kernel->blockDim = N->getNumOfDim();
- assert((kernel->gridDim == kernel->blockDim) && "Dimension number must match");
- std::vector<Value*> numOfBlocks = PNode->getDimLimits();
- kernel->localWGSize = N->getDimLimits();
- for (unsigned i = 0; i < kernel->gridDim; i++) {
- //BinaryOperator* MulInst = BinaryOperator::CreateMul(kernel->localWGSize[i],numOfBlocks[i]);
- //kernel->globalWGSize.push_back(MulInst);
- }*/
+ // Contains the instructions generating the kernel configuration parameters
+ std::vector<Value*> globalWGSizeInsts;
+ for (unsigned i = 0; i < PNode->getNumOfDim(); i++) {
+ BinaryOperator* MulInst = BinaryOperator::CreateMul(PNode->getDimLimits()[i],N->getDimLimits()[i]);
+ globalWGSizeInsts.push_back(MulInst);
+ }
+ kernel = new Kernel(NULL, // kernel function
+ N->getInArgMap(), // kenel argument mapping
+ PNode->getNumOfDim(), // gridDim
+ globalWGSizeInsts, // grid size
+ N->getNumOfDim(), // blockDim
+ N->getDimLimits()); // block size
+
}
std::vector<IntrinsicInst *> IItoRemove;
--
GitLab