From 49ae0d441b03cdc502b14e3b1e887bdef117289e Mon Sep 17 00:00:00 2001
From: Akash Kothari <akashk4@tyler.cs.illinois.edu>
Date: Mon, 21 Dec 2020 12:12:45 -0600
Subject: [PATCH] Add LLVM-9-ported InPlaceDFG analysis pass
---
hpvm/lib/Transforms/BuildDFG/BuildDFG.cpp | 35 +++-
hpvm/lib/Transforms/CMakeLists.txt | 1 +
.../DFG2LLVM_WrapperAPI/CMakeLists.txt | 4 +-
.../DFG2LLVM_WrapperAPI.cpp | 163 +++++++++---------
.../DFG2LLVM_WrapperAPI.exports | 2 +
hpvm/lib/Transforms/GenHPVM/GenHPVM.cpp | 97 ++++++++---
.../InPlaceDFG/InPlaceDFGAnalysis.cpp | 7 +-
7 files changed, 195 insertions(+), 114 deletions(-)
diff --git a/hpvm/lib/Transforms/BuildDFG/BuildDFG.cpp b/hpvm/lib/Transforms/BuildDFG/BuildDFG.cpp
index 3177f86005..a8c807704a 100644
--- a/hpvm/lib/Transforms/BuildDFG/BuildDFG.cpp
+++ b/hpvm/lib/Transforms/BuildDFG/BuildDFG.cpp
@@ -53,6 +53,8 @@ bool BuildDFG::runOnModule(Module &M) {
// Intrinsic Instruction has been initialized from this point on.
Function *F = cast<Function>(II->getOperand(0)->stripPointerCasts());
Root = DFInternalNode::Create(II, F, hpvmUtils::getPreferredTarget(F));
+ errs() << "INTRINSIC: " << II << "\n";
+ errs() << "ROOT NODE" << Root << "\n";
Roots.push_back(Root);
BuildGraph(Root, F);
@@ -174,6 +176,9 @@ bool BuildDFG::isTypeCongruent(Type *L, Type *R) {
// Handles all the createNodeXX hpvm intrinsics.
void BuildDFG::handleCreateNode(DFInternalNode *N, IntrinsicInst *II) {
+ errs() << "************ HANDLE CREATE NODE *********\n";
+ II->print(errs());
+ errs() << "\n";
bool isInternalNode = false;
Function *F = cast<Function>((II->getOperand(0))->stripPointerCasts());
@@ -206,6 +211,7 @@ void BuildDFG::handleCreateNode(DFInternalNode *N, IntrinsicInst *II) {
// dataflow graph
DFInternalNode *childDFNode = DFInternalNode::Create(
II, F, hpvmUtils::getPreferredTarget(F), N, numOfDim, dimLimits);
+ errs() << "INTERNAL NODE: " << childDFNode << "\n";
N->addChildToDFGraph(childDFNode);
HandleToDFNodeMap[II] = childDFNode;
BuildGraph(childDFNode, F);
@@ -213,18 +219,26 @@ void BuildDFG::handleCreateNode(DFInternalNode *N, IntrinsicInst *II) {
// Create Leaf DFnode and add it to the map.
DFLeafNode *childDFNode = DFLeafNode::Create(
II, F, hpvmUtils::getPreferredTarget(F), N, numOfDim, dimLimits);
+ errs() << "LEAF NODE: " << childDFNode << "\n";
N->addChildToDFGraph(childDFNode);
HandleToDFNodeMap[II] = childDFNode;
}
}
void BuildDFG::handleCreateEdge(DFInternalNode *N, IntrinsicInst *II) {
+ errs() << "************ HANDLE CREATE EDGE *********\n";
+ II->print(errs());
+ errs() << "\n";
// The DFNode structures must be in the map before the edge is processed
HandleToDFNode::iterator DFI = HandleToDFNodeMap.find(II->getOperand(0));
assert(DFI != HandleToDFNodeMap.end());
DFI = HandleToDFNodeMap.find(II->getOperand(1));
assert(DFI != HandleToDFNodeMap.end());
-
+
+ errs() << "NODE TO MAP OPERAND 0: " << II->getOperand(0) << "\n";
+ errs() << "NODE TO MAP OPERAND 1: " << II->getOperand(1) << "\n";
+ errs() << "SRC NODE: " << HandleToDFNodeMap[II->getOperand(0)] << "\n";
+ errs() << "DEST NODE: " << HandleToDFNodeMap[II->getOperand(1)] << "\n";
DFNode *SrcDF = HandleToDFNodeMap[II->getOperand(0)];
DFNode *DestDF = HandleToDFNodeMap[II->getOperand(1)];
@@ -258,16 +272,23 @@ void BuildDFG::handleCreateEdge(DFInternalNode *N, IntrinsicInst *II) {
DestPosition, DestTy, isStreaming);
HandleToDFEdgeMap[II] = newDFEdge;
+ errs() << "NEW EDGE: " << newDFEdge << "\n";
// Add Edge to the dataflow graph associated with the parent node
N->addEdgeToDFGraph(newDFEdge);
}
void BuildDFG::handleBindInput(DFInternalNode *N, IntrinsicInst *II) {
+ errs() << "************ HANDLE BIND INPUT *********\n";
+ II->print(errs());
+ errs() << "\n";
// The DFNode structures must be in the map before the edge is processed
HandleToDFNode::iterator DFI = HandleToDFNodeMap.find(II->getOperand(0));
assert(DFI != HandleToDFNodeMap.end());
-
+
+ errs() << "NODE TP MAP: " << II->getOperand(0) << "\n";
+ errs() << "SRC NODE: " << N->getChildGraph()->getEntry() << "\n";
+ errs() << "DEST NODE: " << HandleToDFNodeMap[II->getOperand(0)] << "\n";
DFNode *SrcDF = N->getChildGraph()->getEntry();
DFNode *DestDF = HandleToDFNodeMap[II->getOperand(0)];
@@ -298,16 +319,23 @@ void BuildDFG::handleBindInput(DFInternalNode *N, IntrinsicInst *II) {
DestPosition, DestTy, isStreaming);
HandleToDFEdgeMap[II] = newDFEdge;
+ errs() << "NEW EDGE: " << newDFEdge << "\n";
// Add Edge to the dataflow graph associated with the parent node
N->addEdgeToDFGraph(newDFEdge);
}
void BuildDFG::handleBindOutput(DFInternalNode *N, IntrinsicInst *II) {
+ errs() << "************ HANDLE BIND OUTPUT *********\n";
+ II->print(errs());
+ errs() << "\n";
// The DFNode structures must be in the map before the edge is processed
HandleToDFNode::iterator DFI = HandleToDFNodeMap.find(II->getOperand(0));
assert(DFI != HandleToDFNodeMap.end());
-
+
+ errs() << "NODE TP MAP: " << II->getOperand(0) << "\n";
+ errs() << "SRC NODE: " << HandleToDFNodeMap[II->getOperand(0)] << "\n";
+ errs() << "DEST NODE: " << N->getChildGraph()->getExit() << "\n";
DFNode *SrcDF = HandleToDFNodeMap[II->getOperand(0)];
DFNode *DestDF = N->getChildGraph()->getExit();
@@ -338,6 +366,7 @@ void BuildDFG::handleBindOutput(DFInternalNode *N, IntrinsicInst *II) {
DestPosition, DestTy, isStreaming);
HandleToDFEdgeMap[II] = newDFEdge;
+ errs() << "NEW EDGE: " << newDFEdge << "\n";
// Add Edge to the dataflow graph associated with the parent node
N->addEdgeToDFGraph(newDFEdge);
diff --git a/hpvm/lib/Transforms/CMakeLists.txt b/hpvm/lib/Transforms/CMakeLists.txt
index 5421808ecd..af8c711d33 100644
--- a/hpvm/lib/Transforms/CMakeLists.txt
+++ b/hpvm/lib/Transforms/CMakeLists.txt
@@ -4,6 +4,7 @@ add_subdirectory(DFG2LLVM_OpenCL)
add_subdirectory(DFG2LLVM_CPU)
add_subdirectory(GenHPVM)
add_subdirectory(LocalMem)
+add_subdirectory(HPVM2NGRAPH)
add_subdirectory(DFG2LLVM_WrapperAPI)
add_subdirectory(ReplaceIntrinsics)
add_subdirectory(DFG2LLVM_CUDNN)
diff --git a/hpvm/lib/Transforms/DFG2LLVM_WrapperAPI/CMakeLists.txt b/hpvm/lib/Transforms/DFG2LLVM_WrapperAPI/CMakeLists.txt
index 22c219d0a1..d643e595da 100644
--- a/hpvm/lib/Transforms/DFG2LLVM_WrapperAPI/CMakeLists.txt
+++ b/hpvm/lib/Transforms/DFG2LLVM_WrapperAPI/CMakeLists.txt
@@ -2,7 +2,8 @@ if(WIN32 OR CYGWIN)
set(LLVM_LINK_COMPONENTS Core Support)
endif()
-add_llvm_loadable_module( LLVMDFG2LLVM_WrapperAPI
+add_llvm_library( LLVMDFG2LLVM_WrapperAPI
+ MODULE
DFG2LLVM_WrapperAPI.cpp
DEPENDS
@@ -10,3 +11,4 @@ add_llvm_loadable_module( LLVMDFG2LLVM_WrapperAPI
PLUGIN_TOOL
opt
)
+
diff --git a/hpvm/lib/Transforms/DFG2LLVM_WrapperAPI/DFG2LLVM_WrapperAPI.cpp b/hpvm/lib/Transforms/DFG2LLVM_WrapperAPI/DFG2LLVM_WrapperAPI.cpp
index ecec258dfe..e8a3f59ffa 100644
--- a/hpvm/lib/Transforms/DFG2LLVM_WrapperAPI/DFG2LLVM_WrapperAPI.cpp
+++ b/hpvm/lib/Transforms/DFG2LLVM_WrapperAPI/DFG2LLVM_WrapperAPI.cpp
@@ -22,10 +22,11 @@
#include "llvm/Support/SourceMgr.h"
#include "llvm/Support/FileSystem.h"
#include "llvm/IR/Attributes.h"
+#include "llvm/IR/DerivedTypes.h"
#include "llvm-c/Core.h"
-#include "llvm/SupportVISC/VISCTimer.h"
-#include "llvm/SupportVISC/DFG2LLVM.h"
-#include "llvm/InPlaceDFG/InPlaceDFGAnalysis.h"
+
+#include "SupportHPVM/DFG2LLVM.h"
+#include "InPlaceDFG/InPlaceDFGAnalysis.h"
#include <sstream>
#include <fstream>
@@ -280,7 +281,7 @@ public:
void InitialState::transition(CodeGenStateMachine *Mch, IntrinsicInst *II) {
if (II) { // Not end of instruction stream
switch (II->getIntrinsicID()) {
- case Intrinsic::visc_tensor_convolution:
+ case Intrinsic::hpvm_tensor_convolution:
{
Mch->addIntrinsicInst(II);
Mch->addArgument(II->getOperand(0)); // conv input
@@ -289,7 +290,7 @@ void InitialState::transition(CodeGenStateMachine *Mch, IntrinsicInst *II) {
Mch->setCurrent(new ConvolutionLayer_1());
}
break;
- case Intrinsic::visc_tensor_mul:
+ case Intrinsic::hpvm_tensor_mul:
{
Mch->addIntrinsicInst(II);
Mch->addArgument(II->getOperand(0)); // 1st gemm input
@@ -299,10 +300,10 @@ void InitialState::transition(CodeGenStateMachine *Mch, IntrinsicInst *II) {
}
break;
- case Intrinsic::visc_node_id:
+ case Intrinsic::hpvm_node_id:
{
- DEBUG(errs() << "\t: Handling __visc_node_id \n");
+ DEBUG(errs() << "\t: Handling __hpvm_node_id \n");
// Get uint32 node ID
Value *Op = II->getOperand(0);
@@ -312,11 +313,11 @@ void InitialState::transition(CodeGenStateMachine *Mch, IntrinsicInst *II) {
Module *M = Mch->getModule();
Module *RtM = Mch->getRtModule();
- Constant* visc_node_id_call =
+ FunctionCallee hpvm_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);
+ CallInst::Create(hpvm_node_id_call, Args, "", II);
Mch->addIntrinsicToRemove(II);
Mch->setCurrent(new InitialState());
@@ -346,7 +347,7 @@ void FullyConnectedLayer_1::transition(CodeGenStateMachine *Mch,
IntrinsicInst *II) {
if (II) { // Not end of instruction stream
switch (II->getIntrinsicID()) {
- case Intrinsic::visc_tensor_add:
+ case Intrinsic::hpvm_tensor_add:
{
IntrinsicInst *MulII = Mch->getIntrinsicInstAt(0);
assert((MulII == II->getOperand(0)) &&
@@ -372,7 +373,7 @@ void FullyConnectedLayer_2::transition(CodeGenStateMachine *Mch,
IntrinsicInst *II) {
if (II) { // Not end of instruction stream
switch (II->getIntrinsicID()) {
- case Intrinsic::visc_tensor_tanh:
+ case Intrinsic::hpvm_tensor_tanh:
{
// Type of activation : TanH
Mch->addArgument(ConstantInt::get(
@@ -383,7 +384,7 @@ void FullyConnectedLayer_2::transition(CodeGenStateMachine *Mch,
Mch->setCurrent(new FullyConnectedLayer_3());
}
break;
- case Intrinsic::visc_tensor_relu:
+ case Intrinsic::hpvm_tensor_relu:
{
// Type of activation : ReLU
Mch->addArgument(ConstantInt::get(
@@ -394,7 +395,7 @@ void FullyConnectedLayer_2::transition(CodeGenStateMachine *Mch,
Mch->setCurrent(new FullyConnectedLayer_3());
}
break;
- case Intrinsic::visc_tensor_clipped_relu:
+ case Intrinsic::hpvm_tensor_clipped_relu:
{
// Type of activation : Clipped ReLU
Mch->addArgument(ConstantInt::get(
@@ -441,7 +442,7 @@ void ConvolutionLayer_1::transition(CodeGenStateMachine *Mch,
IntrinsicInst *II) {
if (II) { // Not end of instruction stream
switch (II->getIntrinsicID()) {
- case Intrinsic::visc_tensor_add:
+ case Intrinsic::hpvm_tensor_add:
{
IntrinsicInst *ConvII = Mch->getIntrinsicInstAt(0);
assert((ConvII == II->getOperand(0)) &&
@@ -504,7 +505,7 @@ void ConvolutionLayer_2::transition(CodeGenStateMachine *Mch,
IntrinsicInst *II) {
if (II) { // Not end of instruction stream
switch (II->getIntrinsicID()) {
- case Intrinsic::visc_tensor_tanh:
+ case Intrinsic::hpvm_tensor_tanh:
{
// Type of activation : TanH
// Mch->addArgument(ConstantInt::get(
@@ -514,7 +515,7 @@ void ConvolutionLayer_2::transition(CodeGenStateMachine *Mch,
Mch->setCurrent(new ConvolutionLayer_3());
}
break;
- case Intrinsic::visc_tensor_relu:
+ case Intrinsic::hpvm_tensor_relu:
{
// Type of activation : ReLU
// Mch->addArgument(ConstantInt::get(
@@ -524,7 +525,7 @@ void ConvolutionLayer_2::transition(CodeGenStateMachine *Mch,
Mch->setCurrent(new ConvolutionLayer_3());
}
break;
- case Intrinsic::visc_tensor_clipped_relu:
+ case Intrinsic::hpvm_tensor_clipped_relu:
{
// Type of activation : Clipped ReLU
// Mch->addArgument(ConstantInt::get(
@@ -534,7 +535,7 @@ void ConvolutionLayer_2::transition(CodeGenStateMachine *Mch,
Mch->setCurrent(new ConvolutionLayer_3());
}
break;
- case Intrinsic::visc_tensor_pool_max:
+ case Intrinsic::hpvm_tensor_pool_max:
{
// pool max
Mch->addArgument(ConstantInt::get(
@@ -552,7 +553,7 @@ void ConvolutionLayer_2::transition(CodeGenStateMachine *Mch,
Mch->setCurrent(new ConvolutionLayer_4());
}
break;
- case Intrinsic::visc_tensor_pool_min:
+ case Intrinsic::hpvm_tensor_pool_min:
{
// pool min FIXME: 2: supported?
Mch->addArgument(ConstantInt::get(
@@ -570,7 +571,7 @@ void ConvolutionLayer_2::transition(CodeGenStateMachine *Mch,
Mch->setCurrent(new ConvolutionLayer_4());
}
break;
- case Intrinsic::visc_tensor_pool_mean:
+ case Intrinsic::hpvm_tensor_pool_mean:
{
// pool mean
Mch->addArgument(ConstantInt::get(
@@ -614,7 +615,7 @@ void ConvolutionLayer_3::transition(CodeGenStateMachine *Mch,
IntrinsicInst *II) {
if (II) { // Not end of instruction stream
switch (II->getIntrinsicID()) {
- case Intrinsic::visc_tensor_pool_max:
+ case Intrinsic::hpvm_tensor_pool_max:
{
// pool max
Mch->addArgument(ConstantInt::get(
@@ -630,13 +631,13 @@ void ConvolutionLayer_3::transition(CodeGenStateMachine *Mch,
IntrinsicInst *ActII = Mch->getIntrinsicInstAt(2);
// Due to previous switch, we know it is a TanH, ReLU, or Clipped ReLU
Intrinsic::ID ActIID = ActII->getIntrinsicID();
- if (ActIID == Intrinsic::visc_tensor_tanh) {
+ if (ActIID == Intrinsic::hpvm_tensor_tanh) {
Mch->addArgument(ConstantInt::get(
Type::getInt32Ty(Mch->getModule()->getContext()), 0));
- } else if (ActIID == Intrinsic::visc_tensor_relu) {
+ } else if (ActIID == Intrinsic::hpvm_tensor_relu) {
Mch->addArgument(ConstantInt::get(
Type::getInt32Ty(Mch->getModule()->getContext()), 1));
- } else { //ActIID == Intrinsic::visc_tensor_clipped_relu
+ } else { //ActIID == Intrinsic::hpvm_tensor_clipped_relu
Mch->addArgument(ConstantInt::get(
Type::getInt32Ty(Mch->getModule()->getContext()), 2));
}
@@ -644,7 +645,7 @@ void ConvolutionLayer_3::transition(CodeGenStateMachine *Mch,
Mch->setCurrent(new ConvolutionLayer_4());
}
break;
- case Intrinsic::visc_tensor_pool_min:
+ case Intrinsic::hpvm_tensor_pool_min:
{
// pool min FIXME: 2: supported?
Mch->addArgument(ConstantInt::get(
@@ -661,13 +662,13 @@ void ConvolutionLayer_3::transition(CodeGenStateMachine *Mch,
IntrinsicInst *ActII = Mch->getIntrinsicInstAt(2);
// Due to previous switch, we know it is a TanH, ReLU, or Clipped ReLU
Intrinsic::ID ActIID = ActII->getIntrinsicID();
- if (ActIID == Intrinsic::visc_tensor_tanh) {
+ if (ActIID == Intrinsic::hpvm_tensor_tanh) {
Mch->addArgument(ConstantInt::get(
Type::getInt32Ty(Mch->getModule()->getContext()), 0));
- } else if (ActIID == Intrinsic::visc_tensor_relu) {
+ } else if (ActIID == Intrinsic::hpvm_tensor_relu) {
Mch->addArgument(ConstantInt::get(
Type::getInt32Ty(Mch->getModule()->getContext()), 1));
- } else { //ActIID == Intrinsic::visc_tensor_clipped_relu
+ } else { //ActIID == Intrinsic::hpvm_tensor_clipped_relu
Mch->addArgument(ConstantInt::get(
Type::getInt32Ty(Mch->getModule()->getContext()), 2));
}
@@ -675,7 +676,7 @@ void ConvolutionLayer_3::transition(CodeGenStateMachine *Mch,
Mch->setCurrent(new ConvolutionLayer_4());
}
break;
- case Intrinsic::visc_tensor_pool_mean:
+ case Intrinsic::hpvm_tensor_pool_mean:
{
// pool max
Mch->addArgument(ConstantInt::get(
@@ -691,13 +692,13 @@ void ConvolutionLayer_3::transition(CodeGenStateMachine *Mch,
IntrinsicInst *ActII = Mch->getIntrinsicInstAt(2);
// Due to previous switch, we know it is a TanH, ReLU, or Clipped ReLU
Intrinsic::ID ActIID = ActII->getIntrinsicID();
- if (ActIID == Intrinsic::visc_tensor_tanh) {
+ if (ActIID == Intrinsic::hpvm_tensor_tanh) {
Mch->addArgument(ConstantInt::get(
Type::getInt32Ty(Mch->getModule()->getContext()), 0));
- } else if (ActIID == Intrinsic::visc_tensor_relu) {
+ } else if (ActIID == Intrinsic::hpvm_tensor_relu) {
Mch->addArgument(ConstantInt::get(
Type::getInt32Ty(Mch->getModule()->getContext()), 1));
- } else { //ActIID == Intrinsic::visc_tensor_clipped_relu
+ } else { //ActIID == Intrinsic::hpvm_tensor_clipped_relu
Mch->addArgument(ConstantInt::get(
Type::getInt32Ty(Mch->getModule()->getContext()), 2));
}
@@ -723,13 +724,13 @@ void ConvolutionLayer_3::transition(CodeGenStateMachine *Mch,
IntrinsicInst *ActII = Mch->getIntrinsicInstAt(2);
// Due to previous switch, we know it is a TanH, ReLU, or Clipped ReLU
Intrinsic::ID ActIID = ActII->getIntrinsicID();
- if (ActIID == Intrinsic::visc_tensor_tanh) {
+ if (ActIID == Intrinsic::hpvm_tensor_tanh) {
Mch->addArgument(ConstantInt::get(
Type::getInt32Ty(Mch->getModule()->getContext()), 0));
- } else if (ActIID == Intrinsic::visc_tensor_relu) {
+ } else if (ActIID == Intrinsic::hpvm_tensor_relu) {
Mch->addArgument(ConstantInt::get(
Type::getInt32Ty(Mch->getModule()->getContext()), 1));
- } else { //ActIID == Intrinsic::visc_tensor_clipped_relu
+ } else { //ActIID == Intrinsic::hpvm_tensor_clipped_relu
Mch->addArgument(ConstantInt::get(
Type::getInt32Ty(Mch->getModule()->getContext()), 2));
}
@@ -800,11 +801,10 @@ void CodeGenStateMachine::codeGen(DFNode *N, Function *F, const StringRef &strRe
switch (current->getStateID()) {
case AbstractState::ID::CONVOLUTION_LAYER:
{
- Constant* wrapper_ConvLayer2 =
+ FunctionCallee 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
Args.push_back(ConstantFP::get(Type::getFloatTy(M->getContext()), (double) 0));
@@ -835,10 +835,9 @@ void CodeGenStateMachine::codeGen(DFNode *N, Function *F, const StringRef &strRe
break;
case AbstractState::ID::FULLY_CONNECTED_LAYER:
{
- Constant* wrapper_FCLayer =
+ FunctionCallee wrapper_FCLayer =
M->getOrInsertFunction(StringRef("wrapper_FCLayer"),
RtM->getFunction(StringRef("wrapper_FCLayer"))->getFunctionType());
- DEBUG(errs() << *wrapper_FCLayer);
// FIXME: get last (float) arguments from clipped relu intrinsic. For now, 0
Args.push_back(ConstantFP::get(Type::getFloatTy(M->getContext()), (double) 0));
@@ -888,7 +887,7 @@ void CodeGenStateMachine::codeGen(DFNode *N, Function *F, const StringRef &strRe
errs() << "TensorII: " << *TensorII << "\n";
switch (TensorII->getIntrinsicID()) {
- case Intrinsic::visc_tensor_group_convolution:
+ case Intrinsic::hpvm_tensor_group_convolution:
{ /* llvm.hpvm.tensor.group.conv */
// Tensor group conv is not in place.
DEBUG(errs() << F->getName() << "\t: Handling tensor group convolution \n");
@@ -922,7 +921,7 @@ void CodeGenStateMachine::codeGen(DFNode *N, Function *F, const StringRef &strRe
Args.push_back(TensorII->getOperand(7));
// Create wrapper API runtime function call
- Constant* wrapper_tensorGroupConvolution =
+ FunctionCallee wrapper_tensorGroupConvolution =
M->getOrInsertFunction(StringRef("wrapper_tensorGroupConvolution"),
RtM->getFunction(StringRef("wrapper_tensorGroupConvolution"))->getFunctionType());
CallInst* CI = CallInst::Create(wrapper_tensorGroupConvolution,
@@ -932,7 +931,7 @@ void CodeGenStateMachine::codeGen(DFNode *N, Function *F, const StringRef &strRe
}
break;
- case Intrinsic::visc_tensor_batchnorm:
+ case Intrinsic::hpvm_tensor_batchnorm:
{ /* llvm.hpvm.tensor.batchnorm */
// Tensor batchnorm is not in place.
@@ -963,7 +962,7 @@ void CodeGenStateMachine::codeGen(DFNode *N, Function *F, const StringRef &strRe
Args.push_back(TensorII->getOperand(5));
// Create wrapper API runtime function call
- Constant* wrapper_tensorBatchNorm =
+ FunctionCallee wrapper_tensorBatchNorm =
M->getOrInsertFunction(StringRef("wrapper_tensorBatchNorm"),
RtM->getFunction(StringRef("wrapper_tensorBatchNorm"))->getFunctionType());
CallInst* CI = CallInst::Create(wrapper_tensorBatchNorm,
@@ -973,7 +972,7 @@ void CodeGenStateMachine::codeGen(DFNode *N, Function *F, const StringRef &strRe
}
break;
- case Intrinsic::visc_tensor_add:
+ case Intrinsic::hpvm_tensor_add:
{ /* llvm.hpvm.tensor.add */
DEBUG(errs() << F->getName() << "\t: Handling tensorAdd\n");
@@ -1009,7 +1008,7 @@ void CodeGenStateMachine::codeGen(DFNode *N, Function *F, const StringRef &strRe
Args.push_back(TensorII->getOperand(1));
// Create wrapper API runtime function call
- Constant* wrapper_tensorAdd =
+ FunctionCallee wrapper_tensorAdd =
M->getOrInsertFunction(StringRef("wrapper_tensorAdd"),
RtM->getFunction(StringRef("wrapper_tensorAdd"))->getFunctionType());
CallInst::Create(wrapper_tensorAdd, Args, "", TensorII);
@@ -1019,9 +1018,9 @@ void CodeGenStateMachine::codeGen(DFNode *N, Function *F, const StringRef &strRe
}
break;
- case Intrinsic::visc_tensor_pool_max:
- case Intrinsic::visc_tensor_pool_mean:
- case Intrinsic::visc_tensor_pool_min:
+ case Intrinsic::hpvm_tensor_pool_max:
+ case Intrinsic::hpvm_tensor_pool_mean:
+ case Intrinsic::hpvm_tensor_pool_min:
{
DEBUG(errs() << F->getName() << "\t: Handling tensor pooling functions\n");
@@ -1046,13 +1045,13 @@ void CodeGenStateMachine::codeGen(DFNode *N, Function *F, const StringRef &strRe
Args.push_back(TensorII->getOperand(0));
int pool_type = 0;
- if (TensorII->getIntrinsicID() == Intrinsic::visc_tensor_pool_max) {
+ if (TensorII->getIntrinsicID() == Intrinsic::hpvm_tensor_pool_max) {
pool_type = 0;
}
- if (TensorII->getIntrinsicID() == Intrinsic::visc_tensor_pool_mean) {
+ if (TensorII->getIntrinsicID() == Intrinsic::hpvm_tensor_pool_mean) {
pool_type = 1;
}
- if (TensorII->getIntrinsicID() == Intrinsic::visc_tensor_pool_min) {
+ if (TensorII->getIntrinsicID() == Intrinsic::hpvm_tensor_pool_min) {
pool_type = 2;
}
@@ -1068,10 +1067,9 @@ void CodeGenStateMachine::codeGen(DFNode *N, Function *F, const StringRef &strRe
Args.push_back(TensorII->getOperand(6));
// Create wrapper API runtime function call
- Constant* wrapper_tensorPooling =
+ FunctionCallee wrapper_tensorPooling =
M->getOrInsertFunction(StringRef("wrapper_tensorPooling"),
RtM->getFunction(StringRef("wrapper_tensorPooling"))->getFunctionType());
- DEBUG(errs() << *wrapper_tensorPooling);
CallInst* CI = CallInst::Create(wrapper_tensorPooling, Args, "", TensorII);
// Replacing intrinsic result uses with the result of the tensor runtime operation
@@ -1079,9 +1077,9 @@ void CodeGenStateMachine::codeGen(DFNode *N, Function *F, const StringRef &strRe
}
break;
- case Intrinsic::visc_tensor_relu:
- case Intrinsic::visc_tensor_clipped_relu:
- case Intrinsic::visc_tensor_tanh:
+ case Intrinsic::hpvm_tensor_relu:
+ case Intrinsic::hpvm_tensor_clipped_relu:
+ case Intrinsic::hpvm_tensor_tanh:
{
DEBUG(errs() << F->getName() << "\t: Handling tensor activation functions\n");
@@ -1111,28 +1109,25 @@ void CodeGenStateMachine::codeGen(DFNode *N, Function *F, const StringRef &strRe
Args.push_back(TensorII->getOperand(0));
- if (TensorII->getIntrinsicID() == Intrinsic::visc_tensor_relu) {
+ if (TensorII->getIntrinsicID() == Intrinsic::hpvm_tensor_relu) {
// Create wrapper API runtime function call
- Constant* wrapper_tensorRelu =
+ FunctionCallee wrapper_tensorRelu =
M->getOrInsertFunction(StringRef("wrapper_tensorRelu"),
RtM->getFunction(StringRef("wrapper_tensorRelu"))->getFunctionType());
- DEBUG(errs() << *wrapper_tensorRelu);
CallInst::Create(wrapper_tensorRelu, Args, "", TensorII);
}
- else if (TensorII->getIntrinsicID() == Intrinsic::visc_tensor_clipped_relu) {
+ else if (TensorII->getIntrinsicID() == Intrinsic::hpvm_tensor_clipped_relu) {
// Create wrapper API runtime function call
- Constant* wrapper_tensorClippedRelu =
+ FunctionCallee wrapper_tensorClippedRelu =
M->getOrInsertFunction(StringRef("wrapper_tensorClippedRelu"),
RtM->getFunction(StringRef("wrapper_tensorClippedRelu"))->getFunctionType());
- DEBUG(errs() << *wrapper_tensorClippedRelu);
CallInst::Create(wrapper_tensorClippedRelu, Args, "", TensorII);
}
- else if (TensorII->getIntrinsicID() == Intrinsic::visc_tensor_tanh) {
+ else if (TensorII->getIntrinsicID() == Intrinsic::hpvm_tensor_tanh) {
// Create wrapper API runtime function call
- Constant* wrapper_tensorTanh =
+ FunctionCallee wrapper_tensorTanh =
M->getOrInsertFunction(StringRef("wrapper_tensorTanh"),
RtM->getFunction(StringRef("wrapper_tensorTanh"))->getFunctionType());
- DEBUG(errs() << *wrapper_tensorTanh);
CallInst::Create(wrapper_tensorTanh, Args, "", TensorII);
}
@@ -1143,8 +1138,8 @@ void CodeGenStateMachine::codeGen(DFNode *N, Function *F, const StringRef &strRe
}
break;
- case Intrinsic::visc_tensor_softmax:
- { /* llvm.visc.tensor.softmax */
+ case Intrinsic::hpvm_tensor_softmax:
+ { /* llvm.hpvm.tensor.softmax */
DEBUG(errs() << F->getName() << "\t: Handling tensor softmax\n");
// Tensor softmax(a) is in place for argument a.
@@ -1167,10 +1162,9 @@ void CodeGenStateMachine::codeGen(DFNode *N, Function *F, const StringRef &strRe
Args.push_back(TensorII->getOperand(0));
// Create wrapper API runtime function call
- Constant* wrapper_tensorSoftmax =
+ FunctionCallee wrapper_tensorSoftmax =
M->getOrInsertFunction(StringRef("wrapper_tensorSoftmax"),
RtM->getFunction(StringRef("wrapper_tensorSoftmax"))->getFunctionType());
- DEBUG(errs() << *wrapper_tensorSoftmax);
CallInst::Create(wrapper_tensorSoftmax, Args, "", TensorII);
// We can replace the call to hpvm.tensor.softmax with the 1st argument
// that, due to in place operation, now contains the result
@@ -1238,13 +1232,13 @@ private:
InPlaceDFGAnalysis::InPlaceDFGParameter *IPP;
- // VISC Runtime API and Tensor runtime API
- Constant* llvm_hpvm_initApproxhpvmRt;
- Constant* llvm_hpvm_cleanupApproxhpvmRt;
- Constant* hpvm_request_tensor;
+ // HPVM Runtime API and Tensor runtime API
+ FunctionCallee llvm_hpvm_initApproxhpvmRt;
+ FunctionCallee llvm_hpvm_cleanupApproxhpvmRt;
+ FunctionCallee hpvm_request_tensor;
- Constant* llvm_hpvm_initializeRuntimeController;
- Constant* llvm_hpvm_clearRuntimeController;
+ FunctionCallee llvm_hpvm_initializeRuntimeController;
+ FunctionCallee llvm_hpvm_clearRuntimeController;
// Functions
@@ -1304,11 +1298,11 @@ void CGT_WrapperAPI::initRuntimeAPI() {
DECLARE(llvm_hpvm_initializeRuntimeController);
DECLARE(llvm_hpvm_clearRuntimeController);
- // Find visc.init and visc.cleanup calls, and add placeholder methods
+ // Find hpvm.init and visc.cleanup calls, and add placeholder methods
// for initialization and cleanup of the hpvm tensor runtime
- Function* VI = M.getFunction("llvm.visc.init");
- assert(VI->getNumUses() == 1 && "__visc__init should only be used once\n");
+ Function* VI = M.getFunction("llvm.hpvm.init");
+ assert(VI->getNumUses() == 1 && "__hpvm__init should only be used once\n");
InitCall = cast<Instruction>(*VI->user_begin());
CallInst::Create(llvm_hpvm_initApproxhpvmRt,
ArrayRef<Value*>(ConstantInt::get(Type::getInt32Ty(M.getContext()), 0)),
@@ -1339,8 +1333,8 @@ void CGT_WrapperAPI::initRuntimeAPI() {
ArrayRef<Value*> RTCInitArgs = {ConfsGEPConst, QRangesGEPConst};
CallInst::Create(llvm_hpvm_initializeRuntimeController, RTCInitArgs, "", InitCall);
- Function* VC = M.getFunction("llvm.visc.cleanup");
- assert(VC->getNumUses() == 1 && "__visc__clear should only be used once\n");
+ Function* VC = M.getFunction("llvm.hpvm.cleanup");
+ assert(VC->getNumUses() == 1 && "__hpvm__clear should only be used once\n");
CleanupCall = cast<Instruction>(*VC->user_begin());
CallInst::Create(llvm_hpvm_cleanupApproxhpvmRt, ArrayRef<Value*>(), "", CleanupCall);
CallInst::Create(llvm_hpvm_clearRuntimeController, ArrayRef<Value*>(), "", CleanupCall);
@@ -1376,7 +1370,7 @@ void CGT_WrapperAPI::codeGen(DFLeafNode* 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.
- Function *F_wrapper_api = N->getGenFuncForTarget(visc::PROMISE_TARGET);
+ Function *F_wrapper_api = N->getGenFuncForTarget(hpvm::PROMISE_TARGET);
assert((F_wrapper_api == NULL) &&
"Error: Visiting a node for which code already generated");
@@ -1391,7 +1385,7 @@ void CGT_WrapperAPI::codeGen(DFLeafNode* N) {
F_wrapper_api->removeFromParent();
M.getFunctionList().push_back(F_wrapper_api);
- N->addGenFunc(F_wrapper_api, visc::PROMISE_TARGET, true);
+ N->addGenFunc(F_wrapper_api, hpvm::PROMISE_TARGET, true);
/* Removing HPVM in/out/inout function attributes */
for(Function::arg_iterator ai = F_wrapper_api->arg_begin(), ae = F_wrapper_api->arg_end();
@@ -1407,9 +1401,9 @@ void CGT_WrapperAPI::codeGen(DFLeafNode* N) {
// Adding nounwind to generated function : FIXME: needed?
DEBUG(errs() << "Adding nounwind to generated function\n");
- F_wrapper_api->addAttribute(AttributeSet::FunctionIndex, Attribute::NoUnwind);
+ F_wrapper_api->addAttribute(AttributeList::FunctionIndex, Attribute::NoUnwind);
- // Add llvm_visc_requestTensor calls for every pointer argument of the function
+ // Add llvm_hpvm_requestTensor calls for every pointer argument of the function
// (they are all expected to be tensors), at the beginning of the function.
// This is the first instruction of the function, insert them before this
Instruction* FI = &*(F_wrapper_api->getEntryBlock().begin());
@@ -1530,3 +1524,4 @@ static RegisterPass<DFG2LLVM_WrapperAPI> X("dfg2llvm-wrapperapi",
true /* transformation, *
* not just analysis */);
+
diff --git a/hpvm/lib/Transforms/DFG2LLVM_WrapperAPI/DFG2LLVM_WrapperAPI.exports b/hpvm/lib/Transforms/DFG2LLVM_WrapperAPI/DFG2LLVM_WrapperAPI.exports
index e69de29bb2..139597f9cb 100644
--- a/hpvm/lib/Transforms/DFG2LLVM_WrapperAPI/DFG2LLVM_WrapperAPI.exports
+++ b/hpvm/lib/Transforms/DFG2LLVM_WrapperAPI/DFG2LLVM_WrapperAPI.exports
@@ -0,0 +1,2 @@
+
+
diff --git a/hpvm/lib/Transforms/GenHPVM/GenHPVM.cpp b/hpvm/lib/Transforms/GenHPVM/GenHPVM.cpp
index 6c3dcd75f6..2d62ffda0e 100644
--- a/hpvm/lib/Transforms/GenHPVM/GenHPVM.cpp
+++ b/hpvm/lib/Transforms/GenHPVM/GenHPVM.cpp
@@ -169,6 +169,21 @@ IS_HPVM_CALL(requestMemory)
IS_HPVM_CALL(attributes)
IS_HPVM_CALL(hint)
+// Tensor Operators
+IS_HPVM_CALL(tensor_mul)
+IS_HPVM_CALL(tensor_convolution)
+IS_HPVM_CALL(tensor_group_convolution)
+IS_HPVM_CALL(tensor_batchnorm)
+IS_HPVM_CALL(tensor_add)
+IS_HPVM_CALL(tensor_pool_max)
+IS_HPVM_CALL(tensor_pool_min)
+IS_HPVM_CALL(tensor_pool_mean)
+IS_HPVM_CALL(tensor_relu)
+IS_HPVM_CALL(tensor_clipped_relu)
+IS_HPVM_CALL(tensor_tanh)
+IS_HPVM_CALL(tensor_sigmoid)
+IS_HPVM_CALL(tensor_softmax)
+
// Return the constant integer represented by value V
static unsigned getNumericValue(Value *V) {
assert(isa<ConstantInt>(V) &&
@@ -264,32 +279,32 @@ bool GenHPVM::runOnModule(Module &M) {
// Load Runtime API Module
SMDiagnostic Err;
- std::string runtimeAPI = std::string(LLVM_BUILD_DIR_STR) +
- "/tools/hpvm/projects/hpvm-rt/hpvm-rt.bc";
+ //std::string runtimeAPI = std::string(LLVM_BUILD_DIR_STR) +
+ // "/tools/hpvm/projects/hpvm-rt/hpvm-rt.bc";
- std::unique_ptr<Module> runtimeModule =
- parseIRFile(runtimeAPI, Err, M.getContext());
+ //std::unique_ptr<Module> runtimeModule =
+ // parseIRFile(runtimeAPI, Err, M.getContext());
- if (runtimeModule == NULL) {
- DEBUG(errs() << Err.getMessage() << " " << runtimeAPI << "\n");
- assert(false && "couldn't parse runtime");
- } else
- DEBUG(errs() << "Successfully loaded hpvm-rt API module\n");
+ //if (runtimeModule == NULL) {
+ //DEBUG(errs() << Err.getMessage() << " " << runtimeAPI << "\n");
+ // assert(false && "couldn't parse runtime");
+ //} else
+ // DEBUG(errs() << "Successfully loaded hpvm-rt API module\n");
- llvm_hpvm_initializeTimerSet = M.getOrInsertFunction(
- "llvm_hpvm_initializeTimerSet",
- runtimeModule->getFunction("llvm_hpvm_initializeTimerSet")
- ->getFunctionType());
+ //llvm_hpvm_initializeTimerSet = M.getOrInsertFunction(
+ // "llvm_hpvm_initializeTimerSet",
+ // runtimeModule->getFunction("llvm_hpvm_initializeTimerSet")
+ // ->getFunctionType());
// DEBUG(errs() << *llvm_hpvm_initializeTimerSet);
- llvm_hpvm_switchToTimer = M.getOrInsertFunction(
- "llvm_hpvm_switchToTimer",
- runtimeModule->getFunction("llvm_hpvm_switchToTimer")->getFunctionType());
+ //llvm_hpvm_switchToTimer = M.getOrInsertFunction(
+ // "llvm_hpvm_switchToTimer",
+ // runtimeModule->getFunction("llvm_hpvm_switchToTimer")->getFunctionType());
// DEBUG(errs() << *llvm_hpvm_switchToTimer);
- llvm_hpvm_printTimerSet = M.getOrInsertFunction(
- "llvm_hpvm_printTimerSet",
- runtimeModule->getFunction("llvm_hpvm_printTimerSet")->getFunctionType());
+ //llvm_hpvm_printTimerSet = M.getOrInsertFunction(
+ // "llvm_hpvm_printTimerSet",
+ // runtimeModule->getFunction("llvm_hpvm_printTimerSet")->getFunctionType());
// DEBUG(errs() << *llvm_hpvm_printTimerSet);
// Insert init context in main
@@ -298,16 +313,16 @@ bool GenHPVM::runOnModule(Module &M) {
assert(VI->getNumUses() == 1 && "__hpvm__init should only be used once");
Instruction *I = cast<Instruction>(*VI->user_begin());
- DEBUG(errs() << "Initialize Timer Set\n");
- initializeTimerSet(I);
- switchToTimer(hpvm_TimerID_NONE, I);
+ //DEBUG(errs() << "Initialize Timer Set\n");
+ //initializeTimerSet(I);
+ //switchToTimer(hpvm_TimerID_NONE, I);
// Insert print instruction at hpvm exit
DEBUG(errs() << "Locate __hpvm__cleanup()\n");
Function *VC = M.getFunction("__hpvm__cleanup");
assert(VC->getNumUses() == 1 && "__hpvm__cleanup should only be used once");
I = cast<Instruction>(*VC->user_begin());
- printTimerSet(I);
+ //printTimerSet(I);
DEBUG(errs() << "-------- Searching for launch sites ----------\n");
@@ -711,6 +726,42 @@ bool GenHPVM::runOnModule(Module &M) {
if (isHPVMCall_cos(I)) {
ReplaceCallWithIntrinsic(I, Intrinsic::cos, &toBeErased);
}
+ if (isHPVMCall_tensor_convolution(I)) {
+ ReplaceCallWithIntrinsic(I, Intrinsic::hpvm_tensor_convolution, &toBeErased);
+ }
+ if (isHPVMCall_tensor_group_convolution(I)) {
+ ReplaceCallWithIntrinsic(I, Intrinsic::hpvm_tensor_group_convolution, &toBeErased);
+ }
+ if (isHPVMCall_tensor_add(I)) {
+ ReplaceCallWithIntrinsic(I, Intrinsic::hpvm_tensor_add, &toBeErased);
+ }
+ if (isHPVMCall_tensor_batchnorm(I)) {
+ ReplaceCallWithIntrinsic(I, Intrinsic::hpvm_tensor_batchnorm, &toBeErased);
+ }
+ if (isHPVMCall_tensor_mul(I)) {
+ ReplaceCallWithIntrinsic(I, Intrinsic::hpvm_tensor_mul, &toBeErased);
+ }
+ if (isHPVMCall_tensor_pool_max(I)) {
+ ReplaceCallWithIntrinsic(I, Intrinsic::hpvm_tensor_pool_max, &toBeErased);
+ }
+ if (isHPVMCall_tensor_pool_min(I)) {
+ ReplaceCallWithIntrinsic(I, Intrinsic::hpvm_tensor_pool_min, &toBeErased);
+ }
+ if (isHPVMCall_tensor_pool_mean(I)) {
+ ReplaceCallWithIntrinsic(I, Intrinsic::hpvm_tensor_pool_mean, &toBeErased);
+ }
+ if (isHPVMCall_tensor_relu(I)) {
+ ReplaceCallWithIntrinsic(I, Intrinsic::hpvm_tensor_relu, &toBeErased);
+ }
+ if (isHPVMCall_tensor_tanh(I)) {
+ ReplaceCallWithIntrinsic(I, Intrinsic::hpvm_tensor_tanh, &toBeErased);
+ }
+ if (isHPVMCall_tensor_clipped_relu(I)) {
+ ReplaceCallWithIntrinsic(I, Intrinsic::hpvm_tensor_clipped_relu, &toBeErased);
+ }
+ if (isHPVMCall_tensor_softmax(I)) {
+ ReplaceCallWithIntrinsic(I, Intrinsic::hpvm_tensor_softmax, &toBeErased);
+ }
}
// Erase the __hpvm__node calls
diff --git a/hpvm/lib/Transforms/InPlaceDFG/InPlaceDFGAnalysis.cpp b/hpvm/lib/Transforms/InPlaceDFG/InPlaceDFGAnalysis.cpp
index a45e6e3645..db5a1f5fe0 100644
--- a/hpvm/lib/Transforms/InPlaceDFG/InPlaceDFGAnalysis.cpp
+++ b/hpvm/lib/Transforms/InPlaceDFG/InPlaceDFGAnalysis.cpp
@@ -17,8 +17,8 @@
#define DEBUG_TYPE "InPlaceDFGAnalysis"
#include "llvm/Support/SourceMgr.h"
-#include "llvm/InPlaceDFG/InPlaceDFGAnalysis.h"
-#include "llvm/SupportVISC/DFG2LLVM.h"
+#include "InPlaceDFG/InPlaceDFGAnalysis.h"
+#include "SupportHPVM/DFG2LLVM.h"
using namespace llvm;
using namespace builddfg;
@@ -279,7 +279,7 @@ void AT_OCL::codeGen(DFLeafNode* N) {
CallInst *CI = dyn_cast<CallInst>(OutValues[i]);
assert(CI &&
"Expected return value to be the result of a call instruction\n");
- assert ((CI->getCalledFunction()->getName()).startswith("llvm.visc.tensor") &&
+ assert ((CI->getCalledFunction()->getName()).startswith("llvm.hpvm.tensor") &&
"Node output must be the result of an HPVM tensor intrinsic\n");
}
}
@@ -316,3 +316,4 @@ static RegisterPass<InPlaceDFGAnalysisWrapper> X("inplace",
} // End of namespace
+
--
GitLab