Changed the order of code generation to postorder (leaves first, parent after children).

Added a (simple, to be reconsidered) way of finding a kernel.

llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp

@@ -66,14 +66,16 @@ namespace {
     CodeGenTraversal(Module &_M, BuildDFG &_DFG) : M(_M), DFG(_DFG) { }
 
     virtual void visit(DFInternalNode* N) {
-      DEBUG(errs() << "Generating Code for Node (I) - " << N->getFuncPointer()->getName() << "\n");
-      codeGen(N);
-      DEBUG(errs() << "DONE" << "\n");
       for(DFGraph::children_iterator i = N->getChildGraph()->begin(),
           e = N->getChildGraph()->end(); i != e; ++i) {
         DFNode* child = *i;
         child->applyDFNodeVisitor(*this);
       }
+
+      DEBUG(errs() << "Generating Code for Node (I) - " << N->getFuncPointer()->getName() << "\n");
+      codeGen(N);
+      DEBUG(errs() << "DONE" << "\n");
+
     }
 
     virtual void visit(DFLeafNode* N) {
@@ -210,11 +212,12 @@ namespace {
             // The following is to find which function to call
             Function * OpenCLFunction;
             int parentLevel = ParentDFNode->getLevel();
+            int parentReplFactor = ParentDFNode->getNumOfDim();
 
-            if (!parentLevel) {
-            // We only have one level in the hierarchy, but still need to
-            // specify a global id
-              assert((dim != 2) && "Invalid dimension argument");
+            if (!parentLevel || !parentReplFactor) {
+            // We only have one level in the hierarchy or the parent node is not
+            // replicated. This indicates that the parent node is the kernel
+            // launch, so we need to specify a global id
 
               FunctionType* FT =
                 FunctionType::get(Type::getInt32Ty(getGlobalContext()),
@@ -284,10 +287,12 @@ namespace {
             // The following is to find which function to call
             Function * OpenCLFunction;
             int parentLevel = ParentDFNode->getLevel();
+            int parentReplFactor = ParentDFNode->getNumOfDim();
 
-            if (!parentLevel) {
-            // We only have one level in the hierarchy, thus the node
-            // instances are global_size (gridDim x blockDim)
+            if (!parentLevel || !parentReplFactor) {
+            // We only have one level in the hierarchy or the parent node is not
+            // replicated. This indicates that the parent node is the kernel
+            // launch, so the instances are global_size (gridDim x blockDim)
               FunctionType* FT =
                 FunctionType::get(Type::getInt32Ty(getGlobalContext()),
                                   std::vector<Type*>(1, Type::getInt32Ty(getGlobalContext())),