From c16f8fe6d317cc30f30f4662c3f6ac42d33c53b0 Mon Sep 17 00:00:00 2001
From: Maria Kotsifakou <kotsifa2@illinois.edu>
Date: Tue, 11 Nov 2014 17:29:16 +0000
Subject: [PATCH] dfrg2llvm_nvptx pass provides information about the kernel
name, local and global sizes.
---
.../DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp | 41 +++++++++++++++++--
1 file changed, 37 insertions(+), 4 deletions(-)
diff --git a/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp b/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
index fda7a18f1c..4fc3934246 100644
--- a/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
+++ b/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
@@ -72,6 +72,12 @@ namespace {
Module &KernelM;
BuildDFG &DFG;
DFNode * KernelLaunchNode;
+ struct { Function * KF;
+ unsigned gridDim;
+ unsigned blockDim;
+ std::vector<Value*> localWGSize;
+ std::vector<Value*> globalWGSize;
+ } kernel;
// Map from Old function associated with DFNode to new cloned function with
// extra index and dimension arguments. This map also serves to find out if
@@ -614,8 +620,10 @@ namespace {
insertRuntimeCalls(N, getKernelsModuleName(M), "matrixMul");
} else {
- DEBUG(errs() << "Found intermediate node. Generating device code.\n");
- //TODO
+ DEBUG(errs() << "Found intermediate node. Getting size parameters.\n");
+ //TODO : Check that the arguments order of root to intermediate matches
+ // the intermediate to leaf.
+
}
}
@@ -633,10 +641,34 @@ namespace {
int pLevel = PNode->getLevel();
int pReplFactor = PNode->getNumOfDim();
- if (!pLevel || !pReplFactor)
+ if (!pLevel || !pReplFactor) {
KernelLaunchNode = PNode;
- else
+ // TODO: Find a better way of choosing parameters
+ kernel.gridDim = N->getNumOfDim();
+ kernel.blockDim = N->getNumOfDim();
+ kernel.globalWGSize = N->getDimLimits();
+ IntegerType* IntTy = Type::getInt32Ty(KernelM.getContext());
+ // TODO: How to choose the div factor;
+ ConstantInt* divFactor = ConstantInt::getSigned(IntTy, (int64_t) 16);
+ std::vector<Value*> tmp(kernel.gridDim, divFactor);
+ for (unsigned i = 0; i < kernel.gridDim; i++) {
+ BinaryOperator* SDivInst = BinaryOperator::CreateSDiv(kernel.globalWGSize[i],tmp[i]);
+ kernel.localWGSize.push_back(SDivInst);
+ }
+ }
+ else {
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();
+ for (unsigned i = 0; i < kernel.gridDim; i++) {
+ BinaryOperator* MulInst = BinaryOperator::CreateMul(kernel.localWGSize[i],numOfBlocks[i]);
+ kernel.globalWGSize.push_back(MulInst);
+ }
+ }
std::vector<IntrinsicInst *> IItoRemove;
BuildDFG::HandleToDFNode Leaf_HandleToDFNodeMap;
@@ -890,6 +922,7 @@ namespace {
(*ri)->eraseFromParent();
addCLMetadata(F_nvptx);
+ kernel.KF = F_nvptx;
DEBUG(errs() << KernelM);
return;
--
GitLab