From b9461eeadc955b5f3b2461199eb572cd8dd38e7a Mon Sep 17 00:00:00 2001
From: Prakalp Srivastava <psrivas2@illinois.edu>
Date: Thu, 6 Nov 2014 04:36:20 +0000
Subject: [PATCH] (1) Modified test case to use i64 for pointer data size and
not i32. Using i32 is incompatible with visc-rt (2) Set GenFunc pointer to
NULL. Completely removed FMap from both backend passes. Now generated func
info is contained in DFNode itself. (3) Added Makefile to visc-rt for easy
compilation to generate visc-rt.ll
---
llvm/include/llvm/IR/DFGraph.h | 2 +-
.../DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp | 25 ++++----
.../Transforms/DFG2LLVM_X86/DFG2LLVM_X86.cpp | 58 ++++++++-----------
llvm/projects/visc-rt/Makefile | 11 ++++
llvm/projects/visc-rt/visc-rt.cpp | 4 +-
.../VISC/MatrixMultiplication/visc_gemm.ll | 31 +++++-----
6 files changed, 65 insertions(+), 66 deletions(-)
create mode 100644 llvm/projects/visc-rt/Makefile
diff --git a/llvm/include/llvm/IR/DFGraph.h b/llvm/include/llvm/IR/DFGraph.h
index 6093fe9297..2ab5f91c83 100644
--- a/llvm/include/llvm/IR/DFGraph.h
+++ b/llvm/include/llvm/IR/DFGraph.h
@@ -167,7 +167,7 @@ private:
// Important things that make up a Dataflow Node
IntrinsicInst* II; ///< Associated IntrinsicInst/Value
Function* FuncPointer; ///< Associated Function
- Function* GenFunc; ///< Associated Function generated by backend
+ Function* GenFunc = NULL; ///< Associated Function generated by backend
DFInternalNode* Parent; ///< Pointer to parent dataflow Node
unsigned NumOfDim; ///< Number of dimensions
std::vector<Value*> DimLimits; ///< Number of instances in each dimension
diff --git a/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp b/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
index d08a3ae170..8d37637fdc 100644
--- a/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
+++ b/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
@@ -76,7 +76,6 @@ namespace {
// extra index and dimension arguments. This map also serves to find out if
// we already have an index and dim extended function copy or not (i.e.,
// "Have we visited this function before?")
- ValueMap<Function*, Function*> FMap;
DenseMap<DFNode*, Value*> OutputMap;
// VISC Runtime API
@@ -288,7 +287,7 @@ namespace {
assert(OutputMap.count(SrcDF)
&& "Source node call not found. Dependency violation!");
- // Find CallInst associated with the Source DFNode using FMap
+ // Find CallInst associated with the Source DFNode using OutputMap
Value* CI = OutputMap[SrcDF];
// Extract element at source position from this call instruction
@@ -323,7 +322,7 @@ namespace {
void CodeGenTraversal::insertRuntimeCalls(DFInternalNode* N, const Twine& FileName, const Twine& KernelName) {
// Check if clone already exists. If it does, it means we have visited this
// function before and nothing else needs to be done for this leaf node.
- assert(N->getGenFunc() != NULL && "Code already generated for this node");
+ assert(N->getGenFunc() == NULL && "Code already generated for this node");
DEBUG(errs() << "Generating kernel call code\n");
@@ -355,7 +354,7 @@ namespace {
ReturnInst* RI = ReturnInst::Create(M.getContext(),
UndefValue::get(F_X86->getReturnType()), BB);
- //Add old func: new func pair to the FMap
+ //Add the generated function info to DFNode
N->setGenFunc(F_X86, DFNode::X86);
// FIXME: Adding Index and Dim arguments are probably not required except
@@ -645,13 +644,11 @@ namespace {
Function *F = N->getFuncPointer();
// Look up if we have visited this function before. If we have, then just
- // get the cloned function pointer from FMap. Otherwise, create the cloned
- // function and add it to the FMap.
- Function *F_nvptx;
- if(FMap.count(F)) {
- F_nvptx = FMap[F];
- }
- else {
+ // get the cloned function pointer from DFNode. Otherwise, create the cloned
+ // function and add it to the DFNode GenFunc.
+ Function *F_nvptx = N->getGenFunc();
+ if(F_nvptx == NULL) {
+ errs() << "GenFunc is NULL\n";
// Clone the function
ValueToValueMapTy VMap;
F_nvptx = CloneFunction(F, VMap, true);
@@ -662,8 +659,10 @@ namespace {
DEBUG(errs() << *F_nvptx->getType());
DEBUG(errs() << *F_nvptx);
- //Add old func: new func pair to the FMap
- FMap[F] = F_nvptx;
+ //Add generated function info to DFNode
+ N->setGenFunc(F_nvptx, DFNode::PTX);
+ } {
+ errs() << "GenFunc is not NULL\n" << *F_nvptx;
}
transformFunctionToVoid(F_nvptx);
diff --git a/llvm/lib/Transforms/DFG2LLVM_X86/DFG2LLVM_X86.cpp b/llvm/lib/Transforms/DFG2LLVM_X86/DFG2LLVM_X86.cpp
index 8b61d42af9..2659e9e444 100644
--- a/llvm/lib/Transforms/DFG2LLVM_X86/DFG2LLVM_X86.cpp
+++ b/llvm/lib/Transforms/DFG2LLVM_X86/DFG2LLVM_X86.cpp
@@ -56,11 +56,7 @@ private:
Module &M;
BuildDFG &DFG;
- // 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
- // we already have an index and dim extended function copy or not (i.e.,
- // "Have we visited this function before?")
- ValueMap<Function*, Function*> FMap;
+ // Map from Node to Instruction containing output of a node
DenseMap<DFNode*, Value*> OutputMap;
// VISC Runtime API
@@ -82,7 +78,6 @@ private:
Instruction* InsertBefore);
void invokeChild_PTX(DFNode* C, Function* F_X86, ValueToValueMapTy &VMap,
Instruction* InsertBefore);
- void codeGenLaunch(DFInternalNode* Root);
void codeGen(DFInternalNode* N);
void codeGen(DFLeafNode* N);
public:
@@ -91,6 +86,8 @@ public:
initRuntimeAPI();
}
+ void codeGenLaunch(DFInternalNode* Root);
+
virtual void visit(DFInternalNode* N) {
// Follows a bottom-up approach for code generation.
// First generate code for all the child nodes
@@ -112,6 +109,7 @@ public:
DEBUG(errs() << "DONE: Generating Code for Node (L) - " << N->getFuncPointer()->getName() << "\n");
}
+
};
bool DFG2LLVM_X86::runOnModule(Module &M) {
@@ -130,6 +128,14 @@ bool DFG2LLVM_X86::runOnModule(Module &M) {
// Initiate code generation for root DFNode
CGTVisitor->visit(Root);
+
+
+ // Go ahead and replace the launch intrinsic with pthread call, otherwise return now.
+ // TODO: Later on, we might like to do this in a separate pass, which would
+ // allow us the flexibility to switch between complete static code generation
+ // for DFG or having a customized runtime+scheduler
+ CGTVisitor->codeGenLaunch(Root);
+
delete CGTVisitor;
return true;
}
@@ -323,7 +329,7 @@ void CodeGenTraversal::codeGenLaunch(DFInternalNode* Root) {
DEBUG(errs() << "Created Empty Launch Function\n");
// Find the X86 function generated for Root and
- Function* RootF_X86 = FMap[Root->getFuncPointer()];
+ Function* RootF_X86 = Root->getGenFunc();
// Generate a call to RootF_X86 with null parameters for now
std::vector<Value*>Args;
for(unsigned i=0; i< RootF_X86->getFunctionType()->getNumParams(); i++) {
@@ -416,7 +422,7 @@ Value* CodeGenTraversal::getInValueAt(DFNode* Child, unsigned i, Function* Paren
assert(OutputMap.count(SrcDF)
&& "Source node call not found. Dependency violation!");
- // Find CallInst associated with the Source DFNode using FMap
+ // Find CallInst associated with the Source DFNode using OutputMap
Value* CI = OutputMap[SrcDF];
// Extract element at source position from this call instruction
@@ -483,13 +489,12 @@ void CodeGenTraversal::invokeChild_X86(DFNode* C, Function* F_X86,
}
void CodeGenTraversal::codeGen(DFInternalNode* N) {
- Function* F = N->getFuncPointer();
-
// Check if clone already exists. If it does, it means we have visited this
// function before and nothing else needs to be done for this leaf node.
- if(FMap.count(F))
+ if(N->getGenFunc() != NULL)
return;
+ Function* F = N->getFuncPointer();
// Create of clone of F with no instructions. Only the type is the same as F
// without the extra arguments.
Function* F_X86;
@@ -515,10 +520,7 @@ void CodeGenTraversal::codeGen(DFInternalNode* N) {
ReturnInst* RI = ReturnInst::Create(F_X86->getContext(),
UndefValue::get(F_X86->getReturnType()), BB);
- //Add old func: new func pair to the FMap
- // FIXME: We do not require the FMap probably. Only one of setGenFunc or FMap
- // is required
- FMap[F] = F_X86;
+ //Add generated function info to DFNode
N->setGenFunc(F_X86, DFNode::X86);
// Add Index and Dim arguments except for the root node
@@ -591,16 +593,6 @@ void CodeGenTraversal::codeGen(DFInternalNode* N) {
ReturnInst* newRI = ReturnInst::Create(F_X86->getContext(), retVal);
ReplaceInstWithInst(RI, newRI);
- // If it is a root node, we can go ahead and replace the launch intrinsic with
- // pthead call, otherwise return now.
- // TODO: Later on, we might like to do this in a separate pass, which would
- // allow us the flexibility to switch between complete static code generation
- // for DFG or having a customized runtime+scheduler
- if(!N->isRoot())
- return;
-
- codeGenLaunch(N);
-
}
// Code generation for leaf nodes
@@ -611,6 +603,11 @@ void CodeGenTraversal::codeGen(DFLeafNode* N) {
return;
}
+ // Check if clone already exists. If it does, it means we have visited this
+ // function before and nothing else needs to be done for this leaf node.
+ if(N->getGenFunc() != NULL)
+ return;
+
std::vector<IntrinsicInst *> IItoRemove;
std::vector<std::pair<IntrinsicInst *, Value *> > IItoReplace;
BuildDFG::HandleToDFNode Leaf_HandleToDFNodeMap;
@@ -618,21 +615,14 @@ void CodeGenTraversal::codeGen(DFLeafNode* N) {
// Get the function associated woth the dataflow node
Function *F = N->getFuncPointer();
- // Check if clone already exists. If it does, it means we have visited this
- // function before and nothing else needs to be done for this leaf node.
- if(FMap.count(F))
- return;
-
// Clone the function, if we are seeing this function for the first time.
Function *F_X86;
ValueToValueMapTy VMap;
F_X86 = CloneFunction(F, VMap, true);
// Insert the cloned function into the module
M.getFunctionList().push_back(F_X86);
- // Add old func: new func pair to the FMap
- // FIXME: We do not require the FMap probably. Only one of setGenFunc or FMap
- // is required
- FMap[F] = F_X86;
+
+ // Add generated function info to DFNode
N->setGenFunc(F_X86, DFNode::X86);
// Add the new argument to the argument list
diff --git a/llvm/projects/visc-rt/Makefile b/llvm/projects/visc-rt/Makefile
new file mode 100644
index 0000000000..68e3e5f79c
--- /dev/null
+++ b/llvm/projects/visc-rt/Makefile
@@ -0,0 +1,11 @@
+LLVM_INSTALL:=/home/psrivas2/Hetero/VISC/Code/trunk/llvm-install
+TARGET:=visc-rt
+LLVM_CC:=$(LLVM_INSTALL)/bin/clang
+
+all: $(TARGET:%=%.ll)
+
+$(TARGET:%=%.ll):%.ll:%.cpp
+ $(LLVM_CC) -O3 -S -emit-llvm -I /usr/local/cuda/include $< -o $@
+
+clean :
+ rm -f $(TARGET).ll
diff --git a/llvm/projects/visc-rt/visc-rt.cpp b/llvm/projects/visc-rt/visc-rt.cpp
index 0a5eab8a87..6c032289e7 100644
--- a/llvm/projects/visc-rt/visc-rt.cpp
+++ b/llvm/projects/visc-rt/visc-rt.cpp
@@ -95,7 +95,7 @@ void* llvm_visc_ptx_input_ptr(void* graphID, void* input, int arg_index, size_t
CL_MEM_COPY_HOST_PTR, size, input, &errcode);
checkErr(errcode, CL_SUCCESS, "Failure to allocate memory on device");
errcode |= clSetKernelArg(Context->clKernel, arg_index, sizeof(cl_mem), (void*)&d_input);
- checkErr(errcode, CL_SUCCESS, "Failure to set constant input argument");
+ checkErr(errcode, CL_SUCCESS, "Failure to set pointer input argument");
return d_input;
}
@@ -107,7 +107,7 @@ void* llvm_visc_ptx_output_ptr(void* graphID, int arg_index, size_t size) {
size, NULL, &errcode);
checkErr(errcode, CL_SUCCESS, "Failure to allocate memory on device");
errcode |= clSetKernelArg(Context->clKernel, arg_index, sizeof(cl_mem), (void*)&d_output);
- checkErr(errcode, CL_SUCCESS, "Failure to set constant input argument");
+ checkErr(errcode, CL_SUCCESS, "Failure to set pointer output argument");
return d_output;
}
diff --git a/llvm/test/VISC/MatrixMultiplication/visc_gemm.ll b/llvm/test/VISC/MatrixMultiplication/visc_gemm.ll
index 9b19a8c9a1..2f4a6f8277 100644
--- a/llvm/test/VISC/MatrixMultiplication/visc_gemm.ll
+++ b/llvm/test/VISC/MatrixMultiplication/visc_gemm.ll
@@ -131,8 +131,8 @@ declare i32 @printf(i8* nocapture, ...) #1
; --------------- VISC Intrinsics ---------------
; Return Type of VISC Compute Matrix Mul
-%rtype = type {float*, i32}
-%struct.arg = type <{ float*, i32, float*, i32, float*, i32, i32, i32, i32, %rtype }>
+%rtype = type {float*, i64}
+%struct.arg = type <{ float*, i64, float*, i64, float*, i64, i32, i32, i32, %rtype }>
; Function Attrs: nounwind
declare i8* @llvm.visc.launch(i8*, i8*) #0
@@ -178,12 +178,12 @@ declare void @llvm.visc.bind.output(i8*, i32, i32)
; ----------------- VISC intrinsics end ------------------
; Function Attrs: nounwind uwtable
-define %rtype @matrixMul(float* nocapture %A, i32 %bytes_A, float* nocapture %B, i32 %bytes_B, float* %C, i32 %bytes_C, i32 %k, i32 %n, i32 %m) #0 {
+define %rtype @matrixMul(float* nocapture %A, i64 %bytes_A, float* nocapture %B, i64 %bytes_B, float* %C, i64 %bytes_C, i32 %k, i32 %n, i32 %m) #0 {
entry:
;%puts = tail call i32 @puts(i8* getelementptr inbounds ([17 x i8]* @str, i64 0, i64 0))
; ------------------------- VISC changes ------------------
- ; Replace get_global_id calls with calls to getNode followed but getNumNodeInstances.x
+ ; Replace get_global_id calls with calls to getNode followed by getNumNodeInstances.x
; Replaced statement --
; -- %call1 = tail call i32 (i32, ...)* bitcast (i32 (...)* @get_global_id to i32 (i32, ...)*)(i32 0) #5
; -- %call2 = tail call i32 (i32, ...)* bitcast (i32 (...)* @get_global_id to i32 (i32, ...)*)(i32 1) #5
@@ -232,13 +232,13 @@ for.end: ; preds = %for.body, %entry
;%puts42 = tail call i32 @puts(i8* getelementptr inbounds ([20 x i8]* @str11, i64 0, i64 0))
;%puts43 = tail call i32 @puts(i8* getelementptr inbounds ([17 x i8]* @str12, i64 0, i64 0))
%.fca.0.insert = insertvalue %rtype undef, float* %C, 0
- %.fca.1.insert = insertvalue %rtype %.fca.0.insert, i32 %bytes_C, 1
+ %.fca.1.insert = insertvalue %rtype %.fca.0.insert, i64 %bytes_C, 1
ret %rtype %.fca.1.insert
}
; ----------------- VISC SGEMM root node ----------------
-define %rtype @MatrixMulRoot(float* %h_A, i32 %bytes_A, float* %h_B, i32 %bytes_B, float* %h_C, i32 %bytes_C, i32 %WA, i32 %WB, i32 %HA) {
- %kernel = call i8* @llvm.visc.createNode2D(i8* bitcast (%rtype (float*, i32, float*, i32, float*, i32, i32, i32, i32)* @matrixMul to i8*), i32 %WB, i32 %HA)
+define %rtype @MatrixMulRoot(float* %h_A, i64 %bytes_A, float* %h_B, i64 %bytes_B, float* %h_C, i64 %bytes_C, i32 %WA, i32 %WB, i32 %HA) {
+ %kernel = call i8* @llvm.visc.createNode2D(i8* bitcast (%rtype (float*, i64, float*, i64, float*, i64, i32, i32, i32)* @matrixMul to i8*), i32 %WB, i32 %HA)
; Bind Inputs
call void @llvm.visc.bind.input(i8* %kernel, i32 0, i32 0); h_A
call void @llvm.visc.bind.input(i8* %kernel, i32 1, i32 1); bytes_A
@@ -256,7 +256,7 @@ define %rtype @MatrixMulRoot(float* %h_A, i32 %bytes_A, float* %h_B, i32 %bytes_
}
; Function Attrs: noinline nounwind uwtable
-;define %rtype @computeMatrixMul(float* nocapture %h_A, i32 %bytes_A, float* nocapture %h_B, i32 %bytes_B, float* %h_C, i32 %bytes_C, i32 %k, i32 %n, i32 %m) #3 {
+;define %rtype @computeMatrixMul(float* nocapture %h_A, i64 %bytes_A, float* nocapture %h_B, i64 %bytes_B, float* %h_C, i64 %bytes_C, i32 %k, i32 %n, i32 %m) #3 {
;entry:
; %cmp18 = icmp eq i32 %m, 0
; %cmp215 = icmp eq i32 %n, 0
@@ -272,7 +272,7 @@ define %rtype @MatrixMulRoot(float* %h_A, i32 %bytes_A, float* %h_B, i32 %bytes_
;
;for.body3.us: ; preds = %for.body3.us, %for.body3.lr.ph.us
; %j.016.us = phi i32 [ 0, %for.body3.lr.ph.us ], [ %inc.us, %for.body3.us ]
-; %call.us = tail call %rtype @matrixMul(float* %h_A, i32 undef, float* %h_B, i32 undef, float* %h_C, i32 %bytes_C, i32 %k, i32 %n, i32 undef, i32 undef, i32 undef)
+; %call.us = tail call %rtype @matrixMul(float* %h_A, i64 undef, float* %h_B, i64 undef, float* %h_C, i64 %bytes_C, i32 %k, i32 %n, i32 undef, i32 undef, i32 undef)
; %inc.us = add i32 %j.016.us, 1
; %exitcond = icmp eq i32 %inc.us, %n
; br i1 %exitcond, label %for.inc4.us, label %for.body3.us
@@ -347,19 +347,18 @@ randomInit.exit41: ; preds = %for.body.i40
%in.addr.HA = getelementptr %struct.arg* %in.addr, i32 0, i32 8
store float* %0, float** %in.addr.h_A
- store i32 4194304, i32* %in.addr.bytes_A
+ store i64 4194304, i64* %in.addr.bytes_A
store float* %1, float** %in.addr.h_B
- store i32 4194304, i32* %in.addr.bytes_B
+ store i64 4194304, i64* %in.addr.bytes_B
store float* %2, float** %in.addr.h_C
- store i32 4194304, i32* %in.addr.bytes_C
+ store i64 4194304, i64* %in.addr.bytes_C
store i32 1024, i32* %in.addr.WA
store i32 1024, i32* %in.addr.WB
store i32 1024, i32* %in.addr.HA
; Change type to i8* and VISC Launch call
%args = bitcast %struct.arg* %in.addr to i8*
- %graphID = call i8* @llvm.visc.launch(i8* bitcast (%rtype (float*, i32, float*, i32, float*, i32, i32, i32, i32)* @MatrixMulRoot to i8*), i8* %args)
- ;tail call void @computeMatrixMul(float* %0, i32 4194304, float* %1, i32 4194304, float* %2, i32 4194304)
+ %graphID = call i8* @llvm.visc.launch(i8* bitcast (%rtype (float*, i64, float*, i64, float*, i64, i32, i32, i32)* @MatrixMulRoot to i8*), i8* %args)
; Wait for result
call void @llvm.visc.wait(i8* %graphID)
@@ -367,8 +366,8 @@ randomInit.exit41: ; preds = %for.body.i40
; Get the result
%out.addr = getelementptr %struct.arg* %in.addr, i32 0, i32 9
%out = load %rtype* %out.addr
- ; -------------------------------- Completed VISC Launch Call --------------------------------
-
+ ; -------------------------------- Completed VISC Launch Call --------------------------------
+
%3 = extractvalue %rtype %out, 0
%call14 = tail call i32 @checkResults(float* %0, float* %1, float* %3)
%tobool = icmp eq i32 %call14, 0
--
GitLab