From 91d350a2e573c2f62825ab3baef84a3fbc14e549 Mon Sep 17 00:00:00 2001
From: Akash Kothari <akashk4@tyler.cs.illinois.edu>
Date: Tue, 8 Dec 2020 12:17:37 -0600
Subject: [PATCH] Add support for node ID intrinsic
---
llvm/include/llvm/IR/IntrinsicsVISC.td | 3 +
llvm/lib/Transforms/ClearDFG/ClearDFG.cpp | 23 +-
.../DFG2LLVM_CUDNN/DFG2LLVM_CUDNN.cpp | 46 +++-
.../DFG2LLVM_WrapperAPI.cpp | 198 ++++++++----------
.../FuseHPVMTensorNodes.cpp | 76 +++++--
llvm/lib/Transforms/GenVISC/GenVISC.cpp | 98 +--------
6 files changed, 217 insertions(+), 227 deletions(-)
diff --git a/llvm/include/llvm/IR/IntrinsicsVISC.td b/llvm/include/llvm/IR/IntrinsicsVISC.td
index ab22372d80..404903648f 100644
--- a/llvm/include/llvm/IR/IntrinsicsVISC.td
+++ b/llvm/include/llvm/IR/IntrinsicsVISC.td
@@ -325,4 +325,7 @@ let TargetPrefix = "visc" in {
llvm_i32_ty,
llvm_i32_ty], []>;
+ def int_visc_node_id : Intrinsic<[llvm_ptr_ty], [llvm_i32_ty], []>;
+
+
}
diff --git a/llvm/lib/Transforms/ClearDFG/ClearDFG.cpp b/llvm/lib/Transforms/ClearDFG/ClearDFG.cpp
index 37f74325a7..84f9bec04f 100644
--- a/llvm/lib/Transforms/ClearDFG/ClearDFG.cpp
+++ b/llvm/lib/Transforms/ClearDFG/ClearDFG.cpp
@@ -59,6 +59,7 @@ private:
//Functions
void deleteNode(DFNode* N);
+
public:
// Constructor
TreeTraversal(Module &_M, BuildDFG &_DFG) : M(_M), DFG(_DFG) { }
@@ -88,6 +89,7 @@ public:
};
bool ClearDFG::runOnModule(Module &M) {
+
errs() << "\nCLEARDFG PASS\n";
// Get the BuildDFG Analysis Results:
// - Dataflow graph
@@ -118,6 +120,19 @@ bool ClearDFG::runOnModule(Module &M) {
VC->replaceAllUsesWith(UndefValue::get(VC->getType()));
VC->eraseFromParent();
+
+ Function* VN = M.getFunction("llvm.visc.node.id");
+ if (VN != NULL){ // Delete visc.node.id intrinsic calls if they exist
+ for(Value::user_iterator ui = VN->user_begin(), ue = VN->user_end(); ui != ue; ui++) {
+ Instruction* I = dyn_cast<Instruction>(*ui);
+ I->eraseFromParent();
+ }
+
+ VN->replaceAllUsesWith(UndefValue::get(VN->getType()));
+ VN->eraseFromParent();
+ }
+
+
// Visitor for Code Generation Graph Traversal
TreeTraversal *Visitor = new TreeTraversal(M, DFG);
@@ -125,7 +140,9 @@ bool ClearDFG::runOnModule(Module &M) {
for (auto rootNode: Roots) {
Visitor->visit(rootNode);
}
+
delete Visitor;
+
return true;
}
@@ -149,7 +166,7 @@ void TreeTraversal::deleteNode(DFNode* N) {
char ClearDFG::ID = 0;
static RegisterPass<ClearDFG> X("clearDFG",
- "Delete all DFG functions for which code has been generated",
- false /* does not modify the CFG */,
- true /* transformation, not just analysis */);
+ "Delete all DFG functions for which code has been generated",
+ false /* does not modify the CFG */,
+ true /* transformation, not just analysis */);
diff --git a/llvm/lib/Transforms/DFG2LLVM_CUDNN/DFG2LLVM_CUDNN.cpp b/llvm/lib/Transforms/DFG2LLVM_CUDNN/DFG2LLVM_CUDNN.cpp
index abc4e9ef89..f18325588c 100644
--- a/llvm/lib/Transforms/DFG2LLVM_CUDNN/DFG2LLVM_CUDNN.cpp
+++ b/llvm/lib/Transforms/DFG2LLVM_CUDNN/DFG2LLVM_CUDNN.cpp
@@ -145,7 +145,7 @@ void CGT_CUDNN::initRuntimeAPI() {
runtimeModule = parseIRFile(runtimeAPI.str(), Err, M.getContext());
if(runtimeModule == nullptr)
DEBUG(errs() << Err.getMessage());
- else
+ else
DEBUG(errs() << "Successfully loaded hpvm-tensor-rt API module\n");
// Get or insert Global declarations for
@@ -158,7 +158,15 @@ void CGT_CUDNN::initRuntimeAPI() {
// Find visc.init and visc.cleanup calls, and add placeholder methods
// for initialization and cleanup of the hpvm tensor runtime
-
+ /*
+ LLVMContext &C = M.getContext();
+ auto *FuncType = FunctionType::get(Type::getVoidTy(C), ArrayRef<Type *>({Type::getInt32Ty(C)}), false);
+ llvm_hpvm_initTensorRt = M.getOrInsertFunction(StringRef("llvm_hpvm_initTensorRt"), FuncType);
+ FuncType = FunctionType::get(Type::getVoidTy(C), ArrayRef<Type *>({}), false);
+ llvm_hpvm_cleanupTensorRt = M.getOrInsertFunction(StringRef("llvm_hpvm_cleanupTensorRt"), FuncType);
+ FuncType = FunctionType::get(Type::getVoidTy(C), ArrayRef<Type *>({Type::getInt8PtrTy(C), Type::getInt32Ty(C)}), false);
+ hpvm_request_tensor = M.getOrInsertFunction(StringRef("hpvm_request_tensor"), FuncType);
+*/
Function* VI = M.getFunction("llvm.visc.init");
assert(VI->getNumUses() == 1 && "__visc__init should only be used once\n");
InitCall = cast<Instruction>(*VI->user_begin());
@@ -166,6 +174,7 @@ void CGT_CUDNN::initRuntimeAPI() {
ArrayRef<Value*>(ConstantInt::get(Type::getInt32Ty(M.getContext()), 0)),
"", InitCall);
+
Function* VC = M.getFunction("llvm.visc.cleanup");
assert(VC->getNumUses() == 1 && "__visc__clear should only be used once\n");
CleanupCall = cast<Instruction>(*VC->user_begin());
@@ -264,9 +273,13 @@ void CGT_CUDNN::codeGen(DFLeafNode* N) {
if (BuildDFG::isViscIntrinsic(I)) {
IntrinsicInst* II = dyn_cast<IntrinsicInst>(I);
- assert((II->getCalledFunction()->getName()).startswith("llvm.visc.tensor")
- && "Only HPVM tensor intrinsics allowed in ApproxHPVM leaf nodes\n");
+ //assert((II->getCalledFunction()->getName()).startswith("llvm.visc.tensor")
+ // && "Only HPVM tensor intrinsics allowed in ApproxHPVM leaf nodes\n");
+ //if (!(II->getCalledFunction()->getName()).startswith("llvm.visc.tensor")){
+ //continue; // skip non-tensor ops
+ //}
+
/********************* Handle VISC Tensor intrinsics ********************/
switch (II->getIntrinsicID()) {
@@ -427,7 +440,8 @@ void CGT_CUDNN::codeGen(DFLeafNode* N) {
{ /* llvm.visc.tensor.relu */
DEBUG(errs() << F_cudnn->getName() << "\t: Handling tensor_pool_max\n");
- // Argument list - tensorPooling(input, poolFunction, window_height, window_width, vertical_pad, horizontal_pad,
+ // Argument list - tensorPooling(input, poolFunction, window_height,
+ // window_width, vertical_pad, horizontal_pad,
// vertical_stride, horizontal_stride);
std::vector<Value*> Args;
Args.push_back(II->getOperand(0));
@@ -540,6 +554,28 @@ void CGT_CUDNN::codeGen(DFLeafNode* N) {
IItoRemove.push_back(II);
}
break;
+
+ case Intrinsic::visc_node_id:
+ { /* llvm.visc.node.id */
+ DEBUG(errs() << F_cudnn->getName() << "\t: Handling Node ID Intrinsic \n");
+ // Get uint32 argument
+ Value *Op = II->getOperand(0);
+
+ // Argument list for the runtime call
+ std::vector<Value*> Args;
+ Args.push_back(II->getOperand(0));
+
+ // Create hpvm-tensor-rt function call
+ Constant* tensor_set_node_id;
+ DECLARE(tensor_set_node_id);
+ CallInst::Create(tensor_set_node_id, Args, "", II);
+
+ // Mark to remove at the end
+ IItoRemove.push_back(II);
+ }
+ break;
+
+
default:
llvm_unreachable("Unknown VISC Intrinsic!");
break;
diff --git a/llvm/lib/Transforms/DFG2LLVM_WrapperAPI/DFG2LLVM_WrapperAPI.cpp b/llvm/lib/Transforms/DFG2LLVM_WrapperAPI/DFG2LLVM_WrapperAPI.cpp
index ac5fa450c0..ecec258dfe 100644
--- a/llvm/lib/Transforms/DFG2LLVM_WrapperAPI/DFG2LLVM_WrapperAPI.cpp
+++ b/llvm/lib/Transforms/DFG2LLVM_WrapperAPI/DFG2LLVM_WrapperAPI.cpp
@@ -73,6 +73,7 @@ private:
std::vector<Value*> Args;
std::vector<IntrinsicInst*> IIs;
+ std::vector<IntrinsicInst*> IIs_remove; // Intrinsics to remove
AbstractState *current;
public:
@@ -88,6 +89,10 @@ public:
return M;
}
+ Module *getRtModule() {
+ return RtM;
+ }
+
void addArgument(Value *Arg) {
Args.push_back(Arg);
}
@@ -96,6 +101,10 @@ public:
IIs.push_back(II);
}
+ void addIntrinsicToRemove(IntrinsicInst *II) {
+ IIs_remove.push_back(II);
+ }
+
IntrinsicInst *getIntrinsicInstAt(unsigned idx) {
return IIs[idx];
}
@@ -267,6 +276,7 @@ public:
void transition(CodeGenStateMachine *Mch, IntrinsicInst *II) override;
};
+
void InitialState::transition(CodeGenStateMachine *Mch, IntrinsicInst *II) {
if (II) { // Not end of instruction stream
switch (II->getIntrinsicID()) {
@@ -288,6 +298,31 @@ void InitialState::transition(CodeGenStateMachine *Mch, IntrinsicInst *II) {
Mch->setCurrent(new FullyConnectedLayer_1());
}
break;
+
+ case Intrinsic::visc_node_id:
+ {
+
+ DEBUG(errs() << "\t: Handling __visc_node_id \n");
+ // Get uint32 node ID
+ Value *Op = II->getOperand(0);
+
+ std::vector<Value*> Args;
+ Args.push_back(Op);
+
+ Module *M = Mch->getModule();
+ Module *RtM = Mch->getRtModule();
+
+ Constant* visc_node_id_call =
+ M->getOrInsertFunction(StringRef("tensor_set_node_id"),
+ RtM->getFunction(StringRef("tensor_set_node_id"))->getFunctionType());
+
+ CallInst::Create(visc_node_id_call, Args, "", II);
+
+ Mch->addIntrinsicToRemove(II);
+ Mch->setCurrent(new InitialState());
+ }
+ break;
+
default: // Other HPVM intrinsic
{
Mch->addIntrinsicInst(II);
@@ -438,14 +473,15 @@ void ConvolutionLayer_1::transition(CodeGenStateMachine *Mch,
Mch->addArgument(ConvII->getOperand(3)); // 2nd numeric arg of conv
Mch->addArgument(ConvII->getOperand(4)); // 3rd numeric arg of conv
Mch->addArgument(ConvII->getOperand(5)); // 4th numeric arg of conv
-// Mch->addArgument(ConstantInt::get(
-// Type::getInt32Ty(Mch->getModule()->getContext()), 0));
-// Mch->addArgument(ConstantInt::get(
-// Type::getInt32Ty(Mch->getModule()->getContext()), 0));
-// Mch->addArgument(ConstantInt::get(
-// Type::getInt32Ty(Mch->getModule()->getContext()), 0));
-// Mch->addArgument(ConstantInt::get(
-// Type::getInt32Ty(Mch->getModule()->getContext()), 0));
+
+ // Mch->addArgument(ConstantInt::get(
+ // Type::getInt32Ty(Mch->getModule()->getContext()), 0));
+ // Mch->addArgument(ConstantInt::get(
+ // Type::getInt32Ty(Mch->getModule()->getContext()), 0));
+ // Mch->addArgument(ConstantInt::get(
+ // Type::getInt32Ty(Mch->getModule()->getContext()), 0));
+ // Mch->addArgument(ConstantInt::get(
+ // Type::getInt32Ty(Mch->getModule()->getContext()), 0));
// No pooling
// 0 for unused pool arguments:
@@ -470,9 +506,9 @@ void ConvolutionLayer_2::transition(CodeGenStateMachine *Mch,
switch (II->getIntrinsicID()) {
case Intrinsic::visc_tensor_tanh:
{
- // Type of activation : TanH
-// Mch->addArgument(ConstantInt::get(
-// Type::getInt32Ty(Mch->getModule()->getContext()), 0));
+ // Type of activation : TanH
+ // Mch->addArgument(ConstantInt::get(
+ // Type::getInt32Ty(Mch->getModule()->getContext()), 0));
Mch->addIntrinsicInst(II);
Mch->setCurrent(new ConvolutionLayer_3());
@@ -480,9 +516,9 @@ void ConvolutionLayer_2::transition(CodeGenStateMachine *Mch,
break;
case Intrinsic::visc_tensor_relu:
{
- // Type of activation : ReLU
-// Mch->addArgument(ConstantInt::get(
-// Type::getInt32Ty(Mch->getModule()->getContext()), 1));
+ // Type of activation : ReLU
+ // Mch->addArgument(ConstantInt::get(
+ // Type::getInt32Ty(Mch->getModule()->getContext()), 1));
Mch->addIntrinsicInst(II);
Mch->setCurrent(new ConvolutionLayer_3());
@@ -490,9 +526,9 @@ void ConvolutionLayer_2::transition(CodeGenStateMachine *Mch,
break;
case Intrinsic::visc_tensor_clipped_relu:
{
- // Type of activation : Clipped ReLU
-// Mch->addArgument(ConstantInt::get(
-// Type::getInt32Ty(Mch->getModule()->getContext()), 2));
+ // Type of activation : Clipped ReLU
+ // Mch->addArgument(ConstantInt::get(
+ // Type::getInt32Ty(Mch->getModule()->getContext()), 2));
Mch->addIntrinsicInst(II);
Mch->setCurrent(new ConvolutionLayer_3());
@@ -613,6 +649,7 @@ void ConvolutionLayer_3::transition(CodeGenStateMachine *Mch,
// pool min FIXME: 2: supported?
Mch->addArgument(ConstantInt::get(
Type::getInt32Ty(Mch->getModule()->getContext()), 2));
+
// pool_size_v, pool_size_h, pool pad_v,
// pool_pad_h, pool_stride_v, pool_stride_h
for (int i = 1; i < 7; i++) {
@@ -737,10 +774,11 @@ void CodeGenStateMachine::codeGen(DFNode *N, Function *F, const StringRef &strRe
assert( ( (current->getStateID() == AbstractState::ID::FULLY_CONNECTED_LAYER) ||
(current->getStateID() == AbstractState::ID::CONVOLUTION_LAYER) ||
(current->getStateID() == AbstractState::ID::SINGLE_TENSOR_OPERATION) ) &&
- "Unsupported instruction sequence for the Wrapper API.\n" );
+ "Unsupported instruction sequence for the Wrapper API.\n" );
if ((current->getStateID() == AbstractState::ID::FULLY_CONNECTED_LAYER) ||
(current->getStateID() == AbstractState::ID::CONVOLUTION_LAYER)) {
+
// Layer Operation.
DEBUG(errs() << "Layer Instruction Sequence. Validating ...\n");
// We have a valid instruction sequence.
@@ -765,6 +803,7 @@ void CodeGenStateMachine::codeGen(DFNode *N, Function *F, const StringRef &strRe
Constant* wrapper_ConvLayer2 =
M->getOrInsertFunction(StringRef("wrapper_ConvLayer2"),
RtM->getFunction(StringRef("wrapper_ConvLayer2"))->getFunctionType());
+
DEBUG(errs() << *wrapper_ConvLayer2);
// FIXME: get last (float) arguments from clipped relu intrinsic. For now, 0
@@ -839,12 +878,14 @@ void CodeGenStateMachine::codeGen(DFNode *N, Function *F, const StringRef &strRe
CI->insertBefore(IIlast);
IIlast->replaceAllUsesWith(CI);
- } else { // SINGLE_TENSOR_OPERATION
+ }
+ else { // SINGLE_TENSOR_OPERATION
assert((IIs.size() == 1) &&
"Unexpected size of intrinsics vector in code gen state machine.\n");
assert(Args.empty() && "Unexpected arguments found in coge gen state machine.\n");
IntrinsicInst *TensorII = IIs[0];
-errs() << "TensorII: " << *TensorII << "\n";
+
+ errs() << "TensorII: " << *TensorII << "\n";
switch (TensorII->getIntrinsicID()) {
case Intrinsic::visc_tensor_group_convolution:
@@ -893,8 +934,9 @@ errs() << "TensorII: " << *TensorII << "\n";
case Intrinsic::visc_tensor_batchnorm:
{ /* llvm.hpvm.tensor.batchnorm */
+
// Tensor batchnorm is not in place.
- // FIXME: Add Check for InPlace Analysis
+ // FIXME: Add Check for InPlace Analysis
DEBUG(errs() << F->getName() << "\t: Handling tensor batch normalization \n");
// Argument list for the runtime call
@@ -933,18 +975,18 @@ errs() << "TensorII: " << *TensorII << "\n";
case Intrinsic::visc_tensor_add:
{ /* llvm.hpvm.tensor.add */
- DEBUG(errs() << F->getName() << "\t: Handling tensor add\n");
- // Tensor add(a,b) is in place for argument a.
-// Value *Op = TensorII->getOperand(0);
+ DEBUG(errs() << F->getName() << "\t: Handling tensorAdd\n");
+ // Tensor add(a,b) is in place for argument a.
+ // Value *Op = TensorII->getOperand(0);
// Test the intrinsic operand for in place operation.
-// bool inplace = isValidOperandForInPlaceOperation(Op, F, N, IPP);
+ // bool inplace = isValidOperandForInPlaceOperation(Op, F, N, IPP);
// Code generation will not continue if this is false, because the target
// may provide an in place operation(safe choice)
// FIXME: remove this comment - must check for in-place
-// assert(inplace &&
-// "Operand not valid for in place operation. Code gen aborted.\n");
+ // assert(inplace &&
+ // "Operand not valid for in place operation. Code gen aborted.\n");
// Argument list for the runtime call
@@ -1047,13 +1089,11 @@ errs() << "TensorII: " << *TensorII << "\n";
Value *Op = TensorII->getOperand(0);
// Test the intrinsic operand for in place operation.
- bool inplace = isValidOperandForInPlaceOperation(Op, F, N, IPP);
+ //-- bool inplace = isValidOperandForInPlaceOperation(Op, F, N, IPP);
// Code generation will not continue if this is false, because the target
// may provide an in place operation(safe choice)
- assert(inplace &&
- "Operand not valid for in place operation. Code gen aborted.\n");
-
- // Argument list for the runtime call
+ //-- assert(inplace &&
+ //-- "Operand not valid for in place operation. Code gen aborted.\n");
// Create string for node name, as first argument for wrapper API call
Constant *ConstArray = ConstantDataArray::getString(M->getContext(),
@@ -1110,15 +1150,6 @@ errs() << "TensorII: " << *TensorII << "\n";
// Tensor softmax(a) is in place for argument a.
Value *Op = TensorII->getOperand(0);
- // Test the intrinsic operand for in place operation.
- bool inplace = isValidOperandForInPlaceOperation(Op, F, N, IPP);
- // Code generation will not continue if this is false, because the target
- // may provide an in place operation(safe choice)
- assert(inplace &&
- "Operand not valid for in place operation. Code gen aborted.\n");
-
- // Argument list for the runtime call
-
// Create string for node name, as first argument for wrapper API call
Constant *ConstArray = ConstantDataArray::getString(M->getContext(),
strRef, true);
@@ -1146,62 +1177,8 @@ errs() << "TensorII: " << *TensorII << "\n";
TensorII->replaceAllUsesWith(TensorII->getOperand(0));
}
break;
-/*
- case Intrinsic::visc_image_fft_transform:
- { // llvm.hpvm.image.fft.transform - Or another image intrinsic
- // All will be treated as not in place
- DEBUG(errs() << F->getName() << "\t: Handling fft transform \n");
-
- // Create argument list for the runtime call - stored in Args
- // All interfaces will have a string as first argument, which will be
- // used to identify the dataflow node at runtime
- // Create string for node name, as first argument for wrapper API call
- Constant *ConstArray = ConstantDataArray::getString(M->getContext(),
- strRef, true);
- GlobalVariable *GV = new GlobalVariable(*M,ConstArray->getType(),
- true, GlobalValue::ExternalLinkage, ConstArray, "");
- // Create GEP expression to access it
- Constant* Int_0 = ConstantInt::get(Type::getInt32Ty(M->getContext()), 0);
- Constant* GEPIndices[] = { Int_0, Int_0 };
- Constant* GEPConst =
- ConstantExpr::getGetElementPtr(GV->getType()->getPointerElementType(),
- GV, GEPIndices);
-
- Args.push_back(GEPConst);
-
- // Here, use you will access the appropriate arruments of the intrinsic
- // and push_back, in order to create the argument list of runtime call
- Args.push_back(TensorII->getOperand(0));
- Args.push_back(TensorII->getOperand(1));
- Args.push_back(TensorII->getOperand(2));
- Args.push_back(TensorII->getOperand(3));
- Args.push_back(TensorII->getOperand(4));
- Args.push_back(TensorII->getOperand(5));
-
- Constant *conv_mode = ConstantInt::get(Type::getInt32Ty(M->getContext()), 1);
- Args.push_back(conv_mode);
-
- Args.push_back(TensorII->getOperand(7));
-
- // Done with argument list.
-
- // Create wrapper API runtime function call
- // Appropriately set the name of the function of the runtime that you
- // want to call
- // Note: the Constant * is what we need to pass to the callInst.
- // This name does not have to match, but does so for similarity.
- Constant* wrapper_tensorGroupConvolution;
- M->getOrInsertFunction(StringRef("wrapper_tensorGroupConvolution"),
- RtM->getFunction(StringRef("wrapper_tensorGroupConvolution"))->getFunctionType());
- CallInst* CI = CallInst::Create(wrapper_tensorGroupConvolution,
- Args, "", TensorII);
- // We can replace the call to hpvm.tensor.xxx with the runtime call
- TensorII->replaceAllUsesWith(CI);
- }
- break;
-
-*/
+
default:
llvm_unreachable("Unknown VISC Intrinsic!");
break;
@@ -1219,6 +1196,13 @@ errs() << "TensorII: " << *TensorII << "\n";
(*ri)->eraseFromParent();
}
+
+ for (std::vector<IntrinsicInst *>::reverse_iterator ri = IIs_remove.rbegin(),
+ re = IIs_remove.rend(); ri != re; ++ri) {
+ DEBUG(errs() << "Erasing: " << **ri << "\n");
+ (*ri)->eraseFromParent();
+ }
+
}
// DFG2LLVM_WrapperAPI - The first implementation.
@@ -1226,6 +1210,8 @@ errs() << "TensorII: " << *TensorII << "\n";
struct DFG2LLVM_WrapperAPI : public DFG2LLVM {
static char ID; // Pass identification, replacement for typeid
DFG2LLVM_WrapperAPI() : DFG2LLVM(ID) {}
+
+
private:
public:
@@ -1380,20 +1366,13 @@ void CGT_WrapperAPI::codeGen(DFLeafNode* N) {
return;
}
-// For wrapper API, we generate code for every leaf node.
-// No need to check for hints from frontend
-// // Generate code only if it has the right hint
-// if (!checkPreferredTarget(N, visc::PROMISE_TARGET)) {
-// errs() << "Skipping node: "<< N->getFuncPointer()->getName() << "\n";
-// return;
-// }
// Increment the node ID, for current node.
++nodeID;
// Get the function associated with the dataflow node
Function *F = N->getFuncPointer();
-errs() << "Node Function: " << *F << "\n";
+ errs() << "Node Function: " << *F << "\n";
// Look up if we have visited this function before. If we have, then just
// get the cloned function pointer from DFNode. Otherwise, create the cloned
// function and add it to the DFNode GenFunc.
@@ -1405,6 +1384,8 @@ errs() << "Node Function: " << *F << "\n";
// Clone the function
ValueToValueMapTy VMap;
std::string FName(F->getName().data());//Twine FName = F->getName();
+
+
F_wrapper_api = CloneFunction(F, VMap);
F_wrapper_api->setName(FName+"_wrapper_api");
F_wrapper_api->removeFromParent();
@@ -1461,15 +1442,12 @@ errs() << "Node Function: " << *F << "\n";
//CGM.codeGen(N, F_wrapper_api, N->getFuncPointer()->getName(), *IPP);
CGM.codeGen(N, F_wrapper_api, StringRef(std::to_string(nodeID)), *IPP);
-//errs() << "-----------------------------------\n";
-//errs() << *F_wrapper_api << "\n";
-
return;
}
bool DFG2LLVM_WrapperAPI::runOnModule(Module &M) {
- errs() << "\nDFG2LLVM_WrapperAPI PASS\n";
+ errs() << "\nDFG2LLVM_WrapperAPI PASS\n";
// Get the BuildDFG Analysis Results:
// - Dataflow graph
BuildDFG &DFG = getAnalysis<BuildDFG>();
@@ -1477,9 +1455,8 @@ bool DFG2LLVM_WrapperAPI::runOnModule(Module &M) {
// Get the In Place Analysis Results
InPlaceDFGAnalysis::InPlaceDFGParameter IPP =
(getAnalysis<InPlaceDFGAnalysisWrapper>()).getIPP();
- // Print results
-// printInPlaceDFGParameter(IPP);
+
std::vector<DFInternalNode*> Roots = DFG.getRoots();
// Visitor for Code Generation Graph Traversal
@@ -1496,6 +1473,7 @@ bool DFG2LLVM_WrapperAPI::runOnModule(Module &M) {
//TODO: Edit module epilogue to remove the VISC intrinsic declarations
delete CGTVisitor;
+
return true;
}
diff --git a/llvm/lib/Transforms/FuseHPVMTensorNodes/FuseHPVMTensorNodes.cpp b/llvm/lib/Transforms/FuseHPVMTensorNodes/FuseHPVMTensorNodes.cpp
index d9a3c588b5..541efe4e1d 100644
--- a/llvm/lib/Transforms/FuseHPVMTensorNodes/FuseHPVMTensorNodes.cpp
+++ b/llvm/lib/Transforms/FuseHPVMTensorNodes/FuseHPVMTensorNodes.cpp
@@ -78,17 +78,35 @@ static bool isIncomingEdgeArgument(unsigned argno,
return false;
}
+
// Check that this is a valid HPVM Tensor Node (starts with an HPVM intrinsic)
// Return the node intrinsic function
static IntrinsicInst *isValidHPVMTensorNode(DFNode *N) {
+
Function *F = N->getFuncPointer();
- IntrinsicInst *II = dyn_cast<IntrinsicInst>(&*(inst_begin(F)));
- assert(II &&
- "HPVM tensor intrinsic expected as first instruction of HPVM tensor node\n");
- assert((II->getCalledFunction()->getName()).startswith("llvm.visc.tensor") &&
- "Only HPVM tensor intrinsics allowed in ApproxHPVM leaf nodes\n");
+ //IntrinsicInst *II = dyn_cast<IntrinsicInst>(&*(inst_begin(F)));
+
+ IntrinsicInst *II;
+ for (auto I = inst_begin(F), E = inst_end(F); I != E; I++){
+
+ if(dyn_cast<IntrinsicInst>(&*I)){
+ II = dyn_cast<IntrinsicInst>(&*I);
+ if ((II->getCalledFunction()->getName()).startswith("llvm.visc.tensor")){
+ errs()<<"** Tensor Intrinsic = " << *II << "\n";
+ }
+
+ }
+ }
+
+ //assert(II &&
+ // "HPVM tensor intrinsic expected as first instruction of HPVM tensor node\n");
+
+ //assert((II->getCalledFunction()->getName()).startswith("llvm.visc.tensor") &&
+ // "Only HPVM tensor intrinsics allowed in ApproxHPVM leaf nodes\n");
+
return II;
}
+
// Returns the next node in a node sequence, or NULL if it does not exist.
// We consider two nodes a sequence if SrcN has a single successor, DstN,
@@ -340,9 +358,9 @@ Function* FuseHPVMTensorNodes::createEmptyDFNodeFunction(IntrinsicInst* II1,
the body of the fused function instead *
* OutVs: This maps the output struct field index to the stored value */
void FuseHPVMTensorNodes::inlineFirstNodeFunction(Module &M, Function *F1,
- Function *Ffused,
- ValueMap<Value*, Value*> &VMap,
- std::vector<Value*> &OutVs) {
+ Function *Ffused,
+ ValueMap<Value*, Value*> &VMap,
+ std::vector<Value*> &OutVs) {
ReturnInst *RI = cast<ReturnInst>(Ffused->getEntryBlock().getTerminator());
@@ -356,8 +374,9 @@ void FuseHPVMTensorNodes::inlineFirstNodeFunction(Module &M, Function *F1,
}
IntrinsicInst* II = dyn_cast<IntrinsicInst>(I);
- assert((II->getCalledFunction()->getName()).startswith("llvm.visc.tensor")
- && "Only HPVM tensor intrinsics allowed in ApproxHPVM leaf nodes\n");
+ assert ( ((II->getCalledFunction()->getName()).startswith("llvm.visc.tensor")
+ || (II->getCalledFunction()->getName()).startswith("llvm.visc.node.id") )
+ && "Only HPVM tensor intrinsics allowed in ApproxHPVM leaf nodes\n");
std::vector<Value*> Args;
for(unsigned i = 0; i < II->getNumArgOperands(); i++) {
@@ -370,6 +389,7 @@ void FuseHPVMTensorNodes::inlineFirstNodeFunction(Module &M, Function *F1,
Args.push_back(VMap[V]);
}
}
+
Function *F = Intrinsic::getDeclaration(&M, II->getIntrinsicID());
CallInst* CI =
CallInst::Create(F, Args,
@@ -409,9 +429,14 @@ void FuseHPVMTensorNodes::inlineSecondNodeFunction(Module &M, Function *F2,
Instruction *I = &(*f2_i);
if ((BuildDFG::isViscIntrinsic(I))) {
IntrinsicInst* II = dyn_cast<IntrinsicInst>(I);
- assert((II->getCalledFunction()->getName()).startswith("llvm.visc.tensor")
+ assert( ((II->getCalledFunction()->getName()).startswith("llvm.visc.tensor")
+ || (II->getCalledFunction()->getName()).startswith("llvm.visc.node.id"))
&& "Only HPVM tensor intrinsics allowed in ApproxHPVM leaf nodes\n");
+ if ( (II->getCalledFunction()->getName()).startswith("llvm.visc.node.id")) {
+ continue; // Skip adding visc.node.id calls in nodes other than first node
+ }
+
std::vector<Value*> Args;
for(unsigned i = 0; i < II->getNumArgOperands(); i++) {
Value *V = II->getArgOperand(i);
@@ -506,10 +531,11 @@ Function* FuseHPVMTensorNodes::createLeafDFNodeFunction(IntrinsicInst* II1,
++fused_arg_it;
}
-// for(const auto& v: FusedValueMap) {
-// errs() << "key = " << *(v.first) << "\t";
-// errs() << "value = " << *(v.second) << "\n";
-// }
+
+ // for(const auto& v: FusedValueMap) {
+ // errs() << "key = " << *(v.first) << "\t";
+ // errs() << "value = " << *(v.second) << "\n";
+ // }
// Invoke function that inlines F1 into Ffused, using and updating mappings
inlineFirstNodeFunction(M, F1, Ffused, FusedValueMap, OutValues);
@@ -670,6 +696,7 @@ void FuseHPVMTensorNodes::updateParentNodeFunction(IntrinsicInst* II1,
DEBUG(errs() << "Erasing: " << **ib << "\n");
(*ib)->eraseFromParent();
}
+
II2->replaceAllUsesWith(IInew);
II2->eraseFromParent();
@@ -792,6 +819,7 @@ void FindFusionTargetsTraversal::codeGen(DFInternalNode *N) {
return;
}
+
void FindFusionTargetsTraversal::codeGen(DFLeafNode *N) {
DEBUG(errs() << "Inside leaf node: "
<< N->getFuncPointer()->getName() << "\n");
@@ -802,9 +830,9 @@ void FindFusionTargetsTraversal::codeGen(DFLeafNode *N) {
return;
}
-// if(N->getTargetHint() != visc::PROMISE_TARGET) {
+
if(!preferredTargetIncludes(N, visc::PROMISE_TARGET)) {
- // Only fuse if we plan to target PROMISE
+ // Only fuse if we plan to target PROMISE/Layers API
// The CUDNN backend would be able to generate calls for the fused node,
// but not the other way around
DEBUG(errs() << "No PROMISE hint. Skipping node: "
@@ -820,6 +848,14 @@ void FindFusionTargetsTraversal::codeGen(DFLeafNode *N) {
std::vector<IntrinsicInst*> CurrentNodeSequence;
switch(II->getIntrinsicID()) {
+
+ /*case Intrinsic::visc_node_id:
+ { // Found beginning of pattern conv-bias-activation-pooling.
+
+ }
+ break;
+ */
+
case Intrinsic::visc_tensor_convolution:
{ // Found beginning of pattern conv-bias-activation-pooling.
// Look for the rest
@@ -931,9 +967,9 @@ void FindFusionTargetsTraversal::codeGen(DFLeafNode *N) {
}
bool FuseHPVMTensorNodesWrapper::runOnModule(Module &M) {
- errs() << "\nFUSE HPVM TENSOR NODES PASS\n";
-// Get the BuildDFG Analysis Results:
+ errs() << "\nFUSE HPVM TENSOR NODES PASS\n";
+ // Get the BuildDFG Analysis Results:
// - Dataflow graph
BuildDFG &DFG = getAnalysis<BuildDFG>();
@@ -952,7 +988,7 @@ bool FuseHPVMTensorNodesWrapper::runOnModule(Module &M) {
FuseHPVMTensorNodes::FusionTargets &FTs = FTTVisitor->getFusionTargets();
FuseHPVMTensorNodes Fuse;
-// Fuse.printFusionTargets(FTs);
+ // Fuse.printFusionTargets(FTs);
Fuse.run(M, FTs);
diff --git a/llvm/lib/Transforms/GenVISC/GenVISC.cpp b/llvm/lib/Transforms/GenVISC/GenVISC.cpp
index faab312087..a4d9f2c2a4 100644
--- a/llvm/lib/Transforms/GenVISC/GenVISC.cpp
+++ b/llvm/lib/Transforms/GenVISC/GenVISC.cpp
@@ -27,21 +27,14 @@
#include "llvm/SupportVISC/VISCUtils.h"
-#define TIMER(X) do { if (VISCTimer) { X; } } while (0)
-
using namespace llvm;
using namespace viscUtils;
-// VISC Command line option to use timer or not
-static cl::opt<bool>
-VISCTimer("visc-timers-gen", cl::desc("Enable GenVISC timer"));
-
namespace genvisc {
// Helper Functions
-static inline ConstantInt* getTimerID(Module&, enum visc_TimerID);
static Function* transformReturnTypeToStruct(Function* F);
static Type* getReturnTypeFromReturnInst(Function* F);
@@ -178,6 +171,9 @@ IS_VISC_CALL(tensor_tanh)
IS_VISC_CALL(tensor_sigmoid)
IS_VISC_CALL(tensor_softmax)
+IS_VISC_CALL(node_id)
+
+
// Return the constant integer represented by value V
static unsigned getNumericValue(Value* V) {
assert(isa<ConstantInt>(V)
@@ -816,53 +812,17 @@ bool GenVISC::runOnModule(Module &M) {
// Load Runtime API Module
SMDiagnostic Err;
- char* LLVM_SRC_ROOT = getenv("LLVM_SRC_ROOT");
- assert(LLVM_SRC_ROOT != NULL &&
- "Define LLVM_SRC_ROOT environment variable!");
-
- Twine llvmSrcRoot = LLVM_SRC_ROOT;
- Twine runtimeAPI = llvmSrcRoot+"/../build/projects/visc-rt/visc-rt.ll";
- errs() << llvmSrcRoot << "\n";
-
- std::unique_ptr<Module> runtimeModule = parseIRFile(runtimeAPI.str(), Err, M.getContext());
- // Module* runtimeModule =
- // ParseIRFile("/home/psrivas2/current-src/projects/visc-rt/visc-rt.ll",
- // Err, M.getContext());
-
- if(runtimeModule == NULL)
- DEBUG(errs() << Err.getMessage());
- else
- DEBUG(errs() << "Successfully loaded visc-rt API module\n");
-
- 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 context in main
DEBUG(errs() << "Locate __visc__init()\n");
Function* VI = M.getFunction("__visc__init");
assert(VI->getNumUses() == 1 && "__visc__init should only be used once");
Instruction* I = cast<Instruction>(*VI->user_begin());
- DEBUG(errs() << "Initialize Timer Set\n");
- initializeTimerSet(I);
- switchToTimer(visc_TimerID_NONE, I);
-
// Insert print instruction at visc exit
DEBUG(errs() << "Locate __visc__cleanup()\n");
Function* VC = M.getFunction("__visc__cleanup");
assert(VC->getNumUses() == 1 && "__visc__cleanup should only be used once");
I = cast<Instruction>(*VC->user_begin());
- printTimerSet(I);
-
DEBUG(errs() << "-------- Searching for launch sites ----------\n");
@@ -1308,6 +1268,12 @@ bool GenVISC::runOnModule(Module &M) {
if (isVISCCall_tensor_softmax(I)) {
ReplaceCallWithIntrinsic(I, Intrinsic::visc_tensor_softmax, &toBeErased);
}
+
+ // New Intrinsic to set Node ID
+ if (isVISCCall_node_id(I)) {
+ ReplaceCallWithIntrinsic(I, Intrinsic::visc_node_id, &toBeErased);
+ }
+
}
// Erase the __visc__node calls
@@ -1530,7 +1496,6 @@ void GenVISC::genHost(CallInst* CI, Function* KernelF, unsigned levels, unsigned
StructType* ArgStructTy = StructType::create(ArgList, "struct.arg", true);
DEBUG(errs() << *ArgStructTy << "\n");
- switchToTimer(visc_TimerID_ARG_PACK, CI);
// Insert alloca inst for this argument struct type
AllocaInst* AI = new AllocaInst(ArgStructTy, "in.addr", CI);
@@ -1544,8 +1509,6 @@ void GenVISC::genHost(CallInst* CI, Function* KernelF, unsigned levels, unsigned
"args",
CI);
- switchToTimer(visc_TimerID_NONE, CI);
-
// Bitcast Root function to i8*
Constant* Root_i8ptr = ConstantExpr::getPointerCast(Root, Type::getInt8PtrTy(Ctx));
// Replace CI with launch call to a Root function
@@ -1575,49 +1538,6 @@ void GenVISC::genHost(CallInst* CI, Function* KernelF, unsigned levels, unsigned
// Get result (optional)
}
-void GenVISC::initializeTimerSet(Instruction* InsertBefore) {
- Value* TimerSetAddr;
- StoreInst* SI;
- TIMER(TimerSet = new GlobalVariable(*M,
- Type::getInt8PtrTy(M->getContext()),
- false,
- GlobalValue::CommonLinkage,
- Constant::getNullValue(Type::getInt8PtrTy(M->getContext())),
- "viscTimerSet_GenVISC"));
- 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 GenVISC::switchToTimer(enum visc_TimerID timer, Instruction* InsertBefore) {
- Value* switchArgs[] = {TimerSet, getTimerID(*M, timer)};
- TIMER(CallInst::Create(llvm_visc_switchToTimer,
- ArrayRef<Value*>(switchArgs, 2),
- "",
- InsertBefore));
-}
-
-void GenVISC::printTimerSet(Instruction* InsertBefore) {
- Value* TimerName;
- TIMER(TimerName = getStringPointer("GenVISC_Timer", InsertBefore));
- Value* printArgs[] = {TimerSet, TimerName};
- TIMER(CallInst::Create(llvm_visc_printTimerSet,
- ArrayRef<Value*>(printArgs, 2),
- "",
- InsertBefore));
-}
-
-static inline ConstantInt* getTimerID(Module& M, enum visc_TimerID timer) {
- return ConstantInt::get(Type::getInt32Ty(M.getContext()), timer);
-}
-
static Function* transformReturnTypeToStruct(Function* F) {
// Currently only works for void return types
DEBUG(errs() << "Transforming return type of function to Struct: " << F->getName() << "\n");
--
GitLab