From 1601e64f7a029546671ff2a7273cdc4b73ebb118 Mon Sep 17 00:00:00 2001
From: Maria Kotsifakou <kotsifa2@illinois.edu>
Date: Mon, 24 Nov 2014 03:35:42 +0000
Subject: [PATCH] Added kernel argument mapping information for dfg2llvm_nvptx.
---
llvm/include/llvm/IR/DFGraph.h | 12 +++++++++
.../DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp | 26 +++++++++++++------
llvm/lib/Transforms/Makefile | 2 +-
3 files changed, 31 insertions(+), 9 deletions(-)
diff --git a/llvm/include/llvm/IR/DFGraph.h b/llvm/include/llvm/IR/DFGraph.h
index 2ab5f91c83..06e8e499b2 100644
--- a/llvm/include/llvm/IR/DFGraph.h
+++ b/llvm/include/llvm/IR/DFGraph.h
@@ -335,6 +335,7 @@ public:
bool isEntryNode();
bool isExitNode();
DFEdge* getInDFEdgeAt(unsigned inPort);
+ std::vector<unsigned>* getInArgMap();
int getAncestorHops(DFNode* N);
virtual void applyDFNodeVisitor(DFNodeVisitor &V) = 0;
@@ -565,6 +566,17 @@ DFEdge* DFNode::getInDFEdgeAt(unsigned inPort) {
}
return NULL;
}
+
+std::vector<unsigned>* DFNode::getInArgMap() {
+ std::vector<unsigned>* map = new std::vector<unsigned>(InDFEdges.size());
+ for (unsigned i = 0; i < InDFEdges.size(); i++) {
+ DFEdge* E = getInDFEdgeAt(i);
+ unsigned pos = E->getSourcePosition();
+ map[i] = pos;
+ }
+ return map;
+}
+
int DFNode::getAncestorHops(DFNode* N) {
DFNode* temp = this->getParent();
int hops = 1;
diff --git a/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp b/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
index c3e49e3904..6140b97c42 100644
--- a/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
+++ b/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
@@ -55,11 +55,12 @@ public:
class Kernel {
public:
Kernel(Function* _KF, unsigned _gridDim = 0, std::vector<Value*>
- _globalWGSize = std::vector<Value*>(), unsigned _blockDim = 0,
+ _globalWGSize = std::vector<Value*>(),
+ std::vector<unsigned>* _inArgMap = NULL, unsigned _blockDim = 0,
std::vector<Value*> _localWGSize = std::vector<Value*>())
: KernelFunction(_KF),
- gridDim(_gridDim), globalWGSize(_globalWGSize), blockDim(_blockDim),
- localWGSize(_localWGSize) {
+ gridDim(_gridDim), globalWGSize(_globalWGSize), inArgMap(_inArgMap),
+ blockDim(_blockDim), localWGSize(_localWGSize) {
assert(gridDim == globalWGSize.size()
&& "gridDim should be same as the size of vector globalWGSize");
@@ -72,9 +73,9 @@ public:
unsigned blockDim;
std::vector<Value*> globalWGSize;
std::vector<Value*> localWGSize;
+ std::vector<unsigned>* inArgMap;
};
-
// Helper function declarations
static void getExecuteNodeParams(Value* &, Value* &, Value* &, Kernel*,
ValueToValueMapTy&, Instruction*);
@@ -685,8 +686,16 @@ void CodeGenTraversal::codeGen(DFInternalNode* N) {
} else {
DEBUG(errs() << "Found intermediate node. Getting size parameters.\n");
- //TODO : Check that the arguments order of root to intermediate matches
- // the intermediate to leaf.
+ // Keep track of the arguments order.
+
+ std::vector<unsigned>* map1 = N->getInArgMap();
+ std::vector<unsigned>* map2 = kernel->getInArgMap();
+
+ // The limit is the size of map2, because this is the number of kernel arguments
+ for (unsigned i = 0; i < map2.size(); i++) {
+ map2[i] = map1[map2[i]];
+ }
+ free(map1);
}
@@ -707,7 +716,7 @@ void CodeGenTraversal::codeGen(DFLeafNode* N) {
if (!pLevel || !pReplFactor) {
KernelLaunchNode = PNode;
- kernel = new Kernel(NULL, N->getNumOfDim(), N->getDimLimits());
+ kernel = new Kernel(NULL, N->getNumOfDim(), N->getDimLimits(), N->getInArgMap());
// TODO: Find a better way of choosing parameters
//kernel->gridDim = N->getNumOfDim();
//kernel->blockDim = N->getNumOfDim();
@@ -725,11 +734,12 @@ void CodeGenTraversal::codeGen(DFLeafNode* N) {
}
else {
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();
- // TODO: Handle different number of dimensions
assert((kernel->gridDim == kernel->blockDim) && "Dimension number must match");
std::vector<Value*> numOfBlocks = PNode->getDimLimits();
kernel->localWGSize = N->getDimLimits();
diff --git a/llvm/lib/Transforms/Makefile b/llvm/lib/Transforms/Makefile
index 575ed37a67..69748f6f0d 100644
--- a/llvm/lib/Transforms/Makefile
+++ b/llvm/lib/Transforms/Makefile
@@ -9,7 +9,7 @@
LEVEL = ../..
PARALLEL_DIRS = Utils Instrumentation Scalar InstCombine IPO Vectorize Hello \
- ObjCARC BuildDFG DFG2LLVM_NVPTX DFG2LLVM_X86
+ ObjCARC BuildDFG DFG2LLVM_NVPTX DFG2LLVM_X86 ClearDFG
include $(LEVEL)/Makefile.config
--
GitLab