diff --git a/llvm/include/llvm/IR/DFGraph.h b/llvm/include/llvm/IR/DFGraph.h
index b6471a4cc1420458be65467df814780ea28196e4..a1ac2cb45e216acb643cab0cd29b0b5e26d2d5e3 100644
--- a/llvm/include/llvm/IR/DFGraph.h
+++ b/llvm/include/llvm/IR/DFGraph.h
@@ -335,7 +335,9 @@ public:
bool isEntryNode();
bool isExitNode();
DFEdge* getInDFEdgeAt(unsigned inPort);
+ DFEdge* getOutDFEdgeAt(unsigned outPort);
std::vector<unsigned> getInArgMap();
+ std::vector<unsigned> getOutArgMap();
int getAncestorHops(DFNode* N);
virtual void applyDFNodeVisitor(DFNodeVisitor &V) = 0;
@@ -567,6 +569,19 @@ DFEdge* DFNode::getInDFEdgeAt(unsigned inPort) {
return NULL;
}
+DFEdge* DFNode::getOutDFEdgeAt(unsigned outPort) {
+
+ // Cannot perform check for the number of outputs here,
+ // it depends on the node's return type
+
+ for(outdfedge_iterator i = outdfedge_begin(), e = outdfedge_end(); i != e; ++i) {
+ DFEdge* E = *i;
+ if(outPort == E->getSourcePosition())
+ return E;
+ }
+ return NULL;
+}
+
std::vector<unsigned> DFNode::getInArgMap() {
std::vector<unsigned> map(InDFEdges.size());
for (unsigned i = 0; i < InDFEdges.size(); i++) {
@@ -577,6 +592,16 @@ std::vector<unsigned> DFNode::getInArgMap() {
return map;
}
+std::vector<unsigned> DFNode::getOutArgMap() {
+ std::vector<unsigned> map(OutDFEdges.size());
+ for (unsigned i = 0; i < OutDFEdges.size(); i++) {
+ DFEdge* E = getOutDFEdgeAt(i);
+ unsigned pos = E->getDestPosition();
+ map[pos] = i;
+ }
+ 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 47b7e18856406a3f1853550587fd662a267a3e47..e8c027686aba7fa19dfb0d3344f208e5ba2302b0 100644
--- a/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
+++ b/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
@@ -54,13 +54,14 @@ public:
// calls
class Kernel {
public:
- Kernel(Function* _KF, DFLeafNode* _KLeafNode,std::vector<unsigned> _inArgMap =
+ Kernel(Function* _KF, DFLeafNode* _KLeafNode, std::vector<unsigned> _inArgMap =
+ std::vector<unsigned>(), std::vector<unsigned> _outArgMap =
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), KernelLeafNode(_KLeafNode), inArgMap(_inArgMap),
- gridDim(_gridDim), globalWGSize(_globalWGSize),
+ outArgMap(_outArgMap), gridDim(_gridDim), globalWGSize(_globalWGSize),
blockDim(_blockDim), localWGSize(_localWGSize) {
assert(gridDim == globalWGSize.size()
@@ -72,6 +73,7 @@ public:
Function* KernelFunction;
DFLeafNode* KernelLeafNode;
std::vector<unsigned> inArgMap;
+ std::vector<unsigned> outArgMap;
unsigned gridDim;
unsigned blockDim;
std::vector<Value*> globalWGSize;
@@ -85,6 +87,13 @@ public:
inArgMap = map;
}
+ std::vector<unsigned> getOutArgMap() {
+ return outArgMap;
+ }
+ void setOutArgMap(std::vector<unsigned> map) {
+ outArgMap = map;
+ }
+
void setLocalWGSize(std::vector<Value*> V) {
localWGSize = V;
}
@@ -648,6 +657,8 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, Kernel* K, const Tw
// Get OutputType of this node
StructType* OutTy = N->getOutputType();
Value *retVal = UndefValue::get(F_X86->getReturnType());
+ // Find the kernel's output arg map, to use instead of the bindings
+ std::vector<unsigned> outArgMap = kernel->getOutArgMap();
// Find all the input edges to exit node
for (unsigned i=0; i < OutTy->getNumElements(); i++) {
DEBUG(errs() << "Output Edge " << i << "\n");
@@ -682,7 +693,10 @@ void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, Kernel* K, const Tw
// Extract element at source position from this call instruction
std::vector<unsigned> IndexList;
- IndexList.push_back(E->getSourcePosition());
+ // i is the destination of DFEdge E
+ // Use the mapping instead of the bindings
+// IndexList.push_back(E->getSourcePosition());
+ IndexList.push_back(outArgMap[i]);
DEBUG(errs() << "Going to generate ExtarctVal inst from "<< *CI <<"\n");
ExtractValueInst* EI = ExtractValueInst::Create(CI, IndexList,
"",RI);
@@ -722,14 +736,32 @@ void CodeGenTraversal::codeGen(DFInternalNode* N) {
} else {
DEBUG(errs() << "Found intermediate node. Getting size parameters.\n");
// Keep track of the arguments order.
- std::vector<unsigned> map1 = N->getInArgMap();
- std::vector<unsigned> map2 = kernel->getInArgMap();
+ std::vector<unsigned> inmap1 = N->getInArgMap();
+ std::vector<unsigned> inmap2 = 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]];
+ // TODO: Verify when we have incoming edges from more than one nodes
+ // The limit is the size of inmap2, because this is the number of kernel arguments
+ for (unsigned i = 0; i < inmap2.size(); i++) {
+ inmap2[i] = inmap1[inmap2[i]];
+ }
+ kernel->setInArgMap(inmap2);
+
+ // Keep track of the output arguments order.
+ std::vector<unsigned> outmap1 = N->getOutArgMap();
+ std::vector<unsigned> outmap2 = kernel->getOutArgMap();
+
+ // TODO: Change when we have incoming edges to the dummy exit node from more
+ // than one nodes. In this case, the number of bindings is the same, but
+ // their destination position, thus the index in outmap1, is not
+ // 0 ... outmap2.size()-1
+ // The limit is the size of outmap2, because this is the number of kernel
+ // output arguments for which the mapping matters
+ // For now, it reasonable to assume that all the kernel arguments are returned,
+ // maybe plys some others from other nodes, thus outmap2.size() <= outmap1.size()
+ for (unsigned i = 0; i < outmap2.size(); i++) {
+ outmap1[i] = outmap2[outmap1[i]];
}
- kernel->setInArgMap(map2);
+ kernel->setOutArgMap(outmap1);
// Track the source of local dimlimits for the kernel
// Dimension limit can either be a constant or an argument of parent
@@ -779,7 +811,12 @@ void CodeGenTraversal::codeGen(DFLeafNode* N) {
// (2) Parent does not have multiple instances
if (!pLevel || !pReplFactor) {
KernelLaunchNode = PNode;
- kernel = new Kernel(NULL, N, N->getInArgMap(), N->getNumOfDim(), N->getDimLimits());
+ kernel = new Kernel(NULL,
+ N,
+ N->getInArgMap(),
+ N->getOutArgMap(),
+ N->getNumOfDim(),
+ N->getDimLimits());
}
else {
// Converting a 2-level DFG to opencl kernel
@@ -790,6 +827,7 @@ void CodeGenTraversal::codeGen(DFLeafNode* N) {
kernel = new Kernel(NULL, // kernel function
N, // kernel leaf node
N->getInArgMap(), // kenel argument mapping
+ N->getOutArgMap(), // kernel output mapping from the leaf to the interemediate node
PNode->getNumOfDim(), // gridDim
PNode->getDimLimits(),// grid size
N->getNumOfDim(), // blockDim