diff --git a/llvm/lib/Transforms/DFG2LLVM_WrapperAPI/DFG2LLVM_WrapperAPI.cpp b/llvm/lib/Transforms/DFG2LLVM_WrapperAPI/DFG2LLVM_WrapperAPI.cpp
index ac5fa450c0956fc056caf732418cda9e52e3655c..ecec258dfe6ef45377bdf2890a5bcd1a31fedf6e 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 eb72b3dd7425dd59ad2212741b78d5954d35e64c..541efe4e1dae7ec0b62fd041396cb34cd6f9e519 100644
--- a/llvm/lib/Transforms/FuseHPVMTensorNodes/FuseHPVMTensorNodes.cpp
+++ b/llvm/lib/Transforms/FuseHPVMTensorNodes/FuseHPVMTensorNodes.cpp
@@ -92,8 +92,7 @@ static IntrinsicInst *isValidHPVMTensorNode(DFNode *N) {
if(dyn_cast<IntrinsicInst>(&*I)){
II = dyn_cast<IntrinsicInst>(&*I);
if ((II->getCalledFunction()->getName()).startswith("llvm.visc.tensor")){
-
- errs()<<"**** WATCH *** " << *II << "\n\n\n";
+ errs()<<"** Tensor Intrinsic = " << *II << "\n";
}
}