From 07032ca7715e479974d7da4b07d717ca17689ab1 Mon Sep 17 00:00:00 2001
From: Prakalp Srivastava <prakalps@gmail.com>
Date: Sun, 3 May 2015 14:14:17 -0500
Subject: [PATCH] (1) Added timer support for DFG2LLVM_X86 pass. New timers for
calcualting pthread creation overhead added
---
llvm/include/llvm/SupportVISC/VISCTimer.h | 7 ++
.../Transforms/DFG2LLVM_X86/DFG2LLVM_X86.cpp | 68 +++++++++++++++++--
llvm/projects/visc-rt/visc-rt.cpp | 5 +-
llvm/test/VISC/parboil/common/mk/visc.mk | 2 +-
4 files changed, 73 insertions(+), 9 deletions(-)
diff --git a/llvm/include/llvm/SupportVISC/VISCTimer.h b/llvm/include/llvm/SupportVISC/VISCTimer.h
index e1fb2762f3..9151049914 100644
--- a/llvm/include/llvm/SupportVISC/VISCTimer.h
+++ b/llvm/include/llvm/SupportVISC/VISCTimer.h
@@ -86,6 +86,7 @@ enum visc_TimerID {
visc_TimerID_OVERLAP, /* Time double-counted in asynchronous and
* host activity: automatically filled in,
* not intended for direct usage */
+ // GPU FUNCTION
visc_TimerID_INIT_CTX,
visc_TimerID_CLEAR_CTX,
visc_TimerID_COPY_SCALAR,
@@ -96,6 +97,12 @@ enum visc_TimerID {
visc_TimerID_MEM_TRACK,
visc_TimerID_MEM_UNTRACK,
visc_TimerID_MISC,
+ // LAUNCH FUNCTION
+ visc_TimerID_PTHREAD_CREATE,
+ visc_TimerID_ARG_UNPACK,
+ visc_TimerID_COMPUTATION,
+ visc_TimerID_OUTPUT_PACK,
+ visc_TimerID_OUTPUT_UNPACK,
visc_TimerID_LAST /* Number of timer IDs */
};
diff --git a/llvm/lib/Transforms/DFG2LLVM_X86/DFG2LLVM_X86.cpp b/llvm/lib/Transforms/DFG2LLVM_X86/DFG2LLVM_X86.cpp
index 88c9a1ee3d..bd6aa65aa1 100644
--- a/llvm/lib/Transforms/DFG2LLVM_X86/DFG2LLVM_X86.cpp
+++ b/llvm/lib/Transforms/DFG2LLVM_X86/DFG2LLVM_X86.cpp
@@ -78,6 +78,8 @@ private:
Constant* llvm_visc_switchToTimer;
Constant* llvm_visc_printTimerSet;
+ GlobalVariable* TimerSet;
+
//Functions
void initRuntimeAPI();
std::vector<IntrinsicInst*>* getWaitList(Value* LI);
@@ -185,6 +187,36 @@ void CodeGenTraversal::initRuntimeAPI() {
runtimeModule->getFunction("llvm_visc_x86_wait")->getFunctionType());
DEBUG(errs() << *llvm_visc_x86_wait);
+ llvm_visc_initializeTimerSet = M.getOrInsertFunction("llvm_visc_initializeTimerSet",
+ runtimeModule->getFunction("llvm_visc_initializeTimerSet")->getFunctionType());
+ DEBUG(errs() << *llvm_visc_initializeTimerSet);
+
+ llvm_visc_switchToTimer = M.getOrInsertFunction("llvm_visc_switchToTimer",
+ runtimeModule->getFunction("llvm_visc_switchToTimer")->getFunctionType());
+ DEBUG(errs() << *llvm_visc_switchToTimer);
+
+ llvm_visc_printTimerSet = M.getOrInsertFunction("llvm_visc_printTimerSet",
+ runtimeModule->getFunction("llvm_visc_printTimerSet")->getFunctionType());
+ DEBUG(errs() << *llvm_visc_printTimerSet);
+
+ // Insert init timer in main
+ for (Module::iterator mi = M.begin(), me = M.end(); mi != me; ++mi) {
+ Function* f = &*mi;
+ errs() << "Function: " << f->getName() << "\n";
+ if(f->getName().equals("main")) {
+ DEBUG(errs() << "Initializing context" << "\n");
+ for(inst_iterator i = inst_begin(f), e = inst_end(f); i != e; ++i) {
+ //if(Label)
+ }
+ errs() << "First instruction: " << *inst_begin(f) << "\n";
+ Instruction* I = &*inst_begin(f);
+ initializeTimerSet(I);
+ switchToTimer(visc_TimerID_NONE, I);
+ }
+ }
+
+
+
}
/* Returns vector of all wait instructions
@@ -313,12 +345,13 @@ Value* CodeGenTraversal::addLoop(Instruction* I, Value* limit, const Twine& inde
}
void CodeGenTraversal::codeGenLaunch(DFInternalNode* Root) {
- // TODO: Place an assert to check if the constant passed bu launch intrinsic
+ // TODO: Place an assert to check if the constant passed by launch intrinsic
// as the number of arguments to DFG is same as the number of arguments of the
// root of DFG
DEBUG(errs() << "Generating Launch Function\n");
// Get Launch Instruction
IntrinsicInst* LI = Root->getInstruction();
+ switchToTimer(visc_TimerID_PTHREAD_CREATE, LI);
DEBUG(errs() << "Generating Launch Function\n");
/* Now we have all the necessary global declarations necessary to generate the
@@ -351,6 +384,7 @@ void CodeGenTraversal::codeGenLaunch(DFInternalNode* Root) {
ReturnInst* RI = ReturnInst::Create(AppFunc->getContext(),
Constant::getNullValue(AppFunc->getReturnType()),
BB);
+ switchToTimer(visc_TimerID_ARG_UNPACK, RI);
DEBUG(errs() << "Created Empty Launch Function\n");
// Find the X86 function generated for Root and
@@ -391,6 +425,9 @@ void CodeGenTraversal::codeGenLaunch(DFInternalNode* Root) {
argNum++;
data = GEP;
}
+ // Add timers around Call to RootF_X86 function
+ switchToTimer(visc_TimerID_COMPUTATION, CI);
+ switchToTimer(visc_TimerID_OUTPUT_PACK, RI);
// Code for returning the output
CastInst* OutputAddrCast = CastInst::CreatePointerCast(data,
@@ -398,6 +435,7 @@ void CodeGenTraversal::codeGenLaunch(DFInternalNode* Root) {
CI->getName()+".addr",
RI);
new StoreInst(CI, OutputAddrCast, RI);
+ switchToTimer(visc_TimerID_NONE, RI);
DEBUG(errs() << "Application specific function:\n");
DEBUG(errs() << *AppFunc << "\n");
@@ -420,6 +458,7 @@ void CodeGenTraversal::codeGenLaunch(DFInternalNode* Root) {
ArrayRef<Value*>(LaunchInst),
"");
ReplaceInstWithInst(waitI, waitI_X86);
+ printTimerSet(waitI_X86->getNextNode());
DEBUG(errs() << *waitI_X86 << "\n");
}
@@ -832,22 +871,37 @@ void CodeGenTraversal::codeGen(DFLeafNode* N) {
}
void CodeGenTraversal::initializeTimerSet(Instruction* InsertBefore) {
- TIMER(CallInst::Create(llvm_visc_initializeTimerSet,
- None,
- "",
- InsertBefore));
+ Value* TimerSetAddr;
+ StoreInst* SI;
+ TIMER(TimerSet = new GlobalVariable(M,
+ Type::getInt8PtrTy(M.getContext()),
+ false,
+ GlobalValue::CommonLinkage,
+ Constant::getNullValue(Type::getInt8PtrTy(M.getContext())),
+ "viscTimerSet_X86"));
+ DEBUG(errs() << "Inserting GV: " << *TimerSet->getType() << *TimerSet << "\n");
+ DEBUG(errs() << "Inserting call to: " << *llvm_visc_initializeTimerSet << "\n");
+
+ TIMER(TimerSetAddr = CallInst::Create(llvm_visc_initializeTimerSet,
+ None,
+ "",
+ InsertBefore));
+ DEBUG(errs() << "TimerSetAddress = " << *TimerSetAddr << "\n");
+ TIMER(SI = new StoreInst(TimerSetAddr, TimerSet, InsertBefore));
+ DEBUG(errs() << "Store Timer Address in Global variable: " << *SI << "\n");
}
void CodeGenTraversal::switchToTimer(enum visc_TimerID timer, Instruction* InsertBefore) {
+ Value* switchArgs[] = {TimerSet, getTimerID(M, timer)};
TIMER(CallInst::Create(llvm_visc_switchToTimer,
- ArrayRef<Value*>(getTimerID(M, timer)),
+ ArrayRef<Value*>(switchArgs, 2),
"",
InsertBefore));
}
void CodeGenTraversal::printTimerSet(Instruction* InsertBefore) {
TIMER(CallInst::Create(llvm_visc_printTimerSet,
- None,
+ ArrayRef<Value*>(TimerSet),
"",
InsertBefore));
}
diff --git a/llvm/projects/visc-rt/visc-rt.cpp b/llvm/projects/visc-rt/visc-rt.cpp
index de348979bf..27ab2b0238 100644
--- a/llvm/projects/visc-rt/visc-rt.cpp
+++ b/llvm/projects/visc-rt/visc-rt.cpp
@@ -804,7 +804,8 @@ visc_PrintTimerSet(struct visc_TimerSet *timers)
const char *categories[] = {
"IO", "Kernel", "Copy", "Driver", "Copy Async", "Compute", "Overlap",
"Init_Ctx", "Clear_Ctx", "Copy_Scalar", "Copy_Ptr", "Mem_Free",
- "Read_Output", "Setup", "Mem_Track", "Mem_Untrack", "Misc"
+ "Read_Output", "Setup", "Mem_Track", "Mem_Untrack", "Misc",
+ "Pthread_Create", "Arg_Unpack", "Computation", "Output_Pack", "Output_Unpack"
};
@@ -1184,7 +1185,9 @@ void llvm_visc_switchToTimer(void** timerSet, enum visc_TimerID timer) {
visc_SwitchToTimer((visc_TimerSet*)(*timerSet), timer);
}
void llvm_visc_printTimerSet(void** timerSet) {
+ cout<< "Entered Print timer set\n";
visc_PrintTimerSet((visc_TimerSet*) (*timerSet));
+ cout<< "Exiting Print timer set\n";
}
void* llvm_visc_initializeTimerSet() {
diff --git a/llvm/test/VISC/parboil/common/mk/visc.mk b/llvm/test/VISC/parboil/common/mk/visc.mk
index a6c54b943e..68532ef623 100644
--- a/llvm/test/VISC/parboil/common/mk/visc.mk
+++ b/llvm/test/VISC/parboil/common/mk/visc.mk
@@ -23,7 +23,7 @@ VISC_OPTFLAGS = -load LLVMBuildDFG.so -load LLVMDFG2LLVM_NVPTX.so -load LLVMDFG2
HOST_LINKFLAGS =
ifeq ($(NO_TIMER),)
- VISC_OPTFLAGS+=-visc-timers
+ VISC_OPTFLAGS+=-visc-timers-x86 -visc-timers-ptx
endif
--
GitLab