From b9838d8c90b99c7ab7d8b10dea8bc174a6e14f35 Mon Sep 17 00:00:00 2001
From: Hashim Sharif <hsharif3@tyler.cs.illinois.edu>
Date: Fri, 6 Dec 2019 00:17:03 -0600
Subject: [PATCH] Removing ApproxHPVM passes from ./lib/Transforms
---
.../ApproxScheduler/ApproxScheduler.cpp | 275 ---
.../Transforms/ApproxScheduler/CMakeLists.txt | 12 -
.../Transforms/ApproxScheduler/LLVMBuild.txt | 21 -
.../Transforms/DFG2LLVM_CUDNN/CMakeLists.txt | 12 -
.../DFG2LLVM_CUDNN/DFG2LLVM_CUDNN.cpp | 609 -------
.../DFG2LLVM_CUDNN/DFG2LLVM_CUDNN.exports | 0
.../Transforms/DFG2LLVM_CUDNN/LLVMBuild.txt | 21 -
.../DFG2LLVM_PROMISE/CMakeLists.txt | 12 -
.../DFG2LLVM_PROMISE/DFG2LLVM_PROMISE.cpp | 1283 --------------
.../DFG2LLVM_PROMISE/DFG2LLVM_PROMISE.exports | 0
.../Transforms/DFG2LLVM_PROMISE/LLVMBuild.txt | 21 -
.../DFG2LLVM_WrapperAPI/CMakeLists.txt | 12 -
.../DFG2LLVM_WrapperAPI.cpp | 1530 -----------------
.../DFG2LLVM_WrapperAPI.exports | 0
.../DFG2LLVM_WrapperAPI/LLVMBuild.txt | 21 -
.../FuseHPVMTensorNodes/CMakeLists.txt | 12 -
.../FuseHPVMTensorNodes.cpp | 971 -----------
.../FuseHPVMTensorNodes.exports | 0
.../FuseHPVMTensorNodes/LLVMBuild.txt | 21 -
.../InsertApproxInfo/CMakeLists.txt | 12 -
.../InsertApproxInfo/InsertApproxInfo.cpp | 498 ------
.../Transforms/InsertApproxInfo/LLVMBuild.txt | 21 -
22 files changed, 5364 deletions(-)
delete mode 100644 llvm/lib/Transforms/ApproxScheduler/ApproxScheduler.cpp
delete mode 100644 llvm/lib/Transforms/ApproxScheduler/CMakeLists.txt
delete mode 100644 llvm/lib/Transforms/ApproxScheduler/LLVMBuild.txt
delete mode 100644 llvm/lib/Transforms/DFG2LLVM_CUDNN/CMakeLists.txt
delete mode 100644 llvm/lib/Transforms/DFG2LLVM_CUDNN/DFG2LLVM_CUDNN.cpp
delete mode 100644 llvm/lib/Transforms/DFG2LLVM_CUDNN/DFG2LLVM_CUDNN.exports
delete mode 100644 llvm/lib/Transforms/DFG2LLVM_CUDNN/LLVMBuild.txt
delete mode 100644 llvm/lib/Transforms/DFG2LLVM_PROMISE/CMakeLists.txt
delete mode 100644 llvm/lib/Transforms/DFG2LLVM_PROMISE/DFG2LLVM_PROMISE.cpp
delete mode 100644 llvm/lib/Transforms/DFG2LLVM_PROMISE/DFG2LLVM_PROMISE.exports
delete mode 100644 llvm/lib/Transforms/DFG2LLVM_PROMISE/LLVMBuild.txt
delete mode 100644 llvm/lib/Transforms/DFG2LLVM_WrapperAPI/CMakeLists.txt
delete mode 100644 llvm/lib/Transforms/DFG2LLVM_WrapperAPI/DFG2LLVM_WrapperAPI.cpp
delete mode 100644 llvm/lib/Transforms/DFG2LLVM_WrapperAPI/DFG2LLVM_WrapperAPI.exports
delete mode 100644 llvm/lib/Transforms/DFG2LLVM_WrapperAPI/LLVMBuild.txt
delete mode 100644 llvm/lib/Transforms/FuseHPVMTensorNodes/CMakeLists.txt
delete mode 100644 llvm/lib/Transforms/FuseHPVMTensorNodes/FuseHPVMTensorNodes.cpp
delete mode 100644 llvm/lib/Transforms/FuseHPVMTensorNodes/FuseHPVMTensorNodes.exports
delete mode 100644 llvm/lib/Transforms/FuseHPVMTensorNodes/LLVMBuild.txt
delete mode 100644 llvm/lib/Transforms/InsertApproxInfo/CMakeLists.txt
delete mode 100644 llvm/lib/Transforms/InsertApproxInfo/InsertApproxInfo.cpp
delete mode 100644 llvm/lib/Transforms/InsertApproxInfo/LLVMBuild.txt
diff --git a/llvm/lib/Transforms/ApproxScheduler/ApproxScheduler.cpp b/llvm/lib/Transforms/ApproxScheduler/ApproxScheduler.cpp
deleted file mode 100644
index 7537b517bc..0000000000
--- a/llvm/lib/Transforms/ApproxScheduler/ApproxScheduler.cpp
+++ /dev/null
@@ -1,275 +0,0 @@
-//===------------------------ InPlaceDFGAnalysis.cpp ----------------------===//
-//
-//
-//
-// The LLVM Compiler Infrastructure
-//
-//
-//
-// This file is distributed under the University of Illinois Open Source
-//
-// License. See LICENSE.TXT for details.
-//
-//
-//
-//===----------------------------------------------------------------------===//
-
-#define DEBUG_TYPE "ApproxScheduler"
-
-#include "llvm/IR/Module.h"
-#include "llvm/Pass.h"
-#include "llvm/Support/SourceMgr.h"
-#include "llvm/InPlaceDFG/InPlaceDFGAnalysis.h"
-#include "llvm/SupportVISC/DFG2LLVM.h"
-#include "llvm/IR/InstrTypes.h"
-#include <unordered_map>
-#include <dirent.h>
-#include <stdio.h>
-#include <sstream>
-#include <fstream>
-
-
-using namespace llvm;
-using namespace builddfg;
-using namespace dfg2llvm;
-using namespace inplacedfg;
-
-
-namespace {
-
-static cl::opt<std::string> category_input("category", cl::desc(" Hardware-agnostic ranking cateogy {log, linear, quad} "));
-static cl::opt<int> rank_input("rank", cl::desc(" Hardware-agostic rank given by autotuner "));
-
-
-struct ApproxMetrics{
- std::string op_name;
- std::string category;
- unsigned int rank; // rank given by autotuner
- double approx_level;
- // Relative L-norm metrics
- double relative_l1;
- double relative_l2;
- double relative_linf;
- // Mean L-norm metrics
- double mean_l1;
- double mean_l2;
- double mean_linf;
-};
-
-
-
-
-struct ApproxSchedulerWrapperPass : public ModulePass {
- static char ID; // Pass identification, replacement for typeid
- ApproxSchedulerWrapperPass() : ModulePass(ID) {}
-
-public:
- // Functions
- bool runOnModule(Module &M);
- void getAnalysisUsage(AnalysisUsage &AU) const;
-};
-
-
-// Visitor for Code generation traversal (tree traversal for now)
-class ApproxScheduler : public CodeGenTraversal {
-
-private:
-
- int rank; // Rank to use for scheduling - ranks added in operand bundles
- std::string category; // category = {log, linear, quad}
-
- // Virtual Functions
- void init() {}
- void initRuntimeAPI() {}
- void codeGen(DFInternalNode* N);
- void codeGen(DFLeafNode* N);
- bool rankMatches(OperandBundleUse opBundle, std::string category, int rank);
- ApproxMetrics* getApproxInfo(Instruction* I);
- ApproxMetrics* loadApproxMetrics(OperandBundleUse opBundle);
-
- // Tracks the id of the tensor op processed
- unsigned int currentID;
-
-public:
- // Constructor
- ApproxScheduler(Module &_M, BuildDFG &_DFG, std::string category, int rank);
- void run();
-
-};
-
-
-
-void ApproxSchedulerWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
- AU.addRequired<BuildDFG>();
- AU.addPreserved<BuildDFG>();
-}
-
-
-bool ApproxSchedulerWrapperPass::runOnModule(Module &M) {
-
- BuildDFG &DFG = getAnalysis<BuildDFG>();
-
- std::string category = category_input.getValue();
- int rank = rank_input.getValue();
-
- ApproxScheduler scheduler(M, DFG, category, rank);
- scheduler.run();
-
- return true;
-}
-
-
-
-ApproxScheduler::ApproxScheduler(Module &_M, BuildDFG &_DFG, std::string category, int rank) :
- CodeGenTraversal(_M, _DFG){
-
- this->category = category;
- this->rank = rank;
-}
-
-
-void ApproxScheduler::run() {
-
- errs() << "\n NOTE: Approximation-based scheduling transform \n";
- std::vector<DFInternalNode*> Roots = DFG.getRoots();
-
- // Iterate over all the DFGs
- for (auto rootNode: Roots) {
- this->visit(rootNode);
- }
-
- return;
-}
-
-
-/*** Analysis of internal node ***/
-void ApproxScheduler::codeGen(DFInternalNode* N) {
- DEBUG(errs() << "Analysing Node: " << N->getFuncPointer()->getName() << "\n");
-}
-
-
-ApproxMetrics* ApproxScheduler::loadApproxMetrics(OperandBundleUse opBundle){
-
- ApproxMetrics* approx_metrics = new ApproxMetrics;
- for(unsigned int j = 0; j < opBundle.Inputs.size(); j = j + 2){
-
- GlobalVariable* gv = dyn_cast<GlobalVariable>(opBundle.Inputs[j].get());
- ConstantDataArray* constString = dyn_cast<ConstantDataArray>(gv->getInitializer());
- std::string metric = std::string(constString->getAsCString().data());
-
- if(metric == "rel_l1"){
- double norm_value = dyn_cast<ConstantFP>(opBundle.Inputs[j+1].get())->getValueAPF().convertToDouble();
- approx_metrics->relative_l1 = norm_value;
- errs()<<"***relative_l1 = "<<approx_metrics->relative_l1<<"\n";
- }
-
- if(metric == "rel_l2"){
- double norm_value = dyn_cast<ConstantFP>(opBundle.Inputs[j+1].get())->getValueAPF().convertToDouble();
- approx_metrics->relative_l2 = norm_value;
- }
- }
-
-}
-
-
-bool ApproxScheduler::rankMatches(OperandBundleUse opBundle, std::string category_in, int rank_in){
-
- // Extracting value of the 'category' attribute
- GlobalVariable* gv = dyn_cast<GlobalVariable>(opBundle.Inputs[1].get());
- ConstantDataArray* constString = dyn_cast<ConstantDataArray>(gv->getInitializer());
- std::string category = std::string(constString->getAsCString().data());
- errs()<<"*category = "<<category<<"\n";
-
- int rank = dyn_cast<ConstantInt>(opBundle.Inputs[3].get())->getZExtValue();
- errs()<<"-rank = "<<rank<<"\n";
-
- if(category == category_in && rank == rank_in)
- return true;
- else
- return false;
-
-}
-
-
-ApproxMetrics* ApproxScheduler::getApproxInfo(Instruction* I){
-
- CallSite* CS = new CallSite(I);
- if(CS->hasOperandBundles()){
- errs()<<"CallSite has OperandBundles \n";
-
- for(unsigned int i = 0; i < CS->getNumOperandBundles(); i++){
- OperandBundleUse bundleUse = CS->getOperandBundleAt(i);
- errs()<<"bundleUse -> getTagName() = "<<bundleUse.getTagName()<<"\n";
-
- if(rankMatches(bundleUse, category, rank)){
- return loadApproxMetrics(bundleUse);
- }
-
- /*for(unsigned int j = 0; j < bundleUse.Inputs.size(); j++){
- Value* bundleVal = bundleUse.Inputs[j].get();
- errs()<<"Val = "<<*bundleVal<<"\n";
- }
- */
-
- }
- }
- else{
- errs()<<"DOES NOT have OperandBundles \n";
- }
-
- assert("No Bundle Matched the provided rank and Category! \n");
-
-}
-
-
-/*** Analysis of leaf node ***/
-void ApproxScheduler::codeGen(DFLeafNode* N) {
- DEBUG(errs() << "Analysing Node: " << N->getFuncPointer()->getName() << "\n");
-
- // Skip code generation if it is a dummy node
- if(N->isDummyNode()) {
- DEBUG(errs() << "Skipping dummy node\n");
- return;
- }
-
- // Abort code generation if it is an allocation node
- if(N->isAllocationNode()) {
- assert(false && "Allocation Node not expected in ApproxHPVM");
- return;
- }
-
- Function *F = N->getFuncPointer();
- Module* M = F->getParent();
-
- std::vector<ApproxMetrics*> metrics_list;
- /**** Reading all tensor operations in the DFG Leaf Node ****/
- for (inst_iterator i = inst_begin(F), e = inst_end(F); i != e; ++i) {
- Instruction *I = &(*i);
- errs()<<*I<<"\n";
-
- if (BuildDFG::isViscIntrinsic(I)) {
- IntrinsicInst* II = dyn_cast<IntrinsicInst>(I);
- // FIXME: The assumption of only tensor instrinsics is restrictive
- assert((II->getCalledFunction()->getName()).startswith("llvm.visc.tensor")
- && "Only HPVM tensor intrinsics allowed in ApproxHPVM leaf nodes\n");
-
- // NOTE: Get tensorOp name - the scheduling decisions are made per-operation type
- std::string intrinsic_id = std::string(II->getCalledFunction()->getName().data());
- ApproxMetrics* approx_metrics = getApproxInfo(I);
- metrics_list.push_back(approx_metrics);
- }
-
- }
-
-}
-
-char ApproxSchedulerWrapperPass::ID = 0;
-static RegisterPass<ApproxSchedulerWrapperPass> X("approx-scheduler",
- "Select target compute unit based on aprroximation metrics",
- false /* does not modify the CFG */,
- false /* not transformation, just analysis */);
-
-
-
-} // End of namespace
-
diff --git a/llvm/lib/Transforms/ApproxScheduler/CMakeLists.txt b/llvm/lib/Transforms/ApproxScheduler/CMakeLists.txt
deleted file mode 100644
index 267ad1d859..0000000000
--- a/llvm/lib/Transforms/ApproxScheduler/CMakeLists.txt
+++ /dev/null
@@ -1,12 +0,0 @@
-if(WIN32 OR CYGWIN)
- set(LLVM_LINK_COMPONENTS Core Support)
-endif()
-
-add_llvm_loadable_module( ApproxScheduler
- ApproxScheduler.cpp
-
- DEPENDS
- intrinsics_gen
- PLUGIN_TOOL
- opt
- )
diff --git a/llvm/lib/Transforms/ApproxScheduler/LLVMBuild.txt b/llvm/lib/Transforms/ApproxScheduler/LLVMBuild.txt
deleted file mode 100644
index ccd8479c2e..0000000000
--- a/llvm/lib/Transforms/ApproxScheduler/LLVMBuild.txt
+++ /dev/null
@@ -1,21 +0,0 @@
-;===- ./lib/Transforms/LocalMem/LLVMBuild.txt ------------------*- Conf -*--===;
-;
-; The LLVM Compiler Infrastructure
-;
-; This file is distributed under the University of Illinois Open Source
-; License. See LICENSE.TXT for details.
-;
-;===------------------------------------------------------------------------===;
-;
-; This is an LLVMBuild description file for the components in this subdirectory.
-;
-; For more information on the LLVMBuild system, please see:
-;
-; http://llvm.org/docs/LLVMBuild.html
-;
-;===------------------------------------------------------------------------===;
-
-[component_0]
-type = Library
-name = ApproxScheduler
-parent = Transforms
diff --git a/llvm/lib/Transforms/DFG2LLVM_CUDNN/CMakeLists.txt b/llvm/lib/Transforms/DFG2LLVM_CUDNN/CMakeLists.txt
deleted file mode 100644
index dc98faafec..0000000000
--- a/llvm/lib/Transforms/DFG2LLVM_CUDNN/CMakeLists.txt
+++ /dev/null
@@ -1,12 +0,0 @@
-if(WIN32 OR CYGWIN)
- set(LLVM_LINK_COMPONENTS Core Support)
-endif()
-
-add_llvm_loadable_module( LLVMDFG2LLVM_CUDNN
- DFG2LLVM_CUDNN.cpp
-
- DEPENDS
- intrinsics_gen
- PLUGIN_TOOL
- opt
- )
diff --git a/llvm/lib/Transforms/DFG2LLVM_CUDNN/DFG2LLVM_CUDNN.cpp b/llvm/lib/Transforms/DFG2LLVM_CUDNN/DFG2LLVM_CUDNN.cpp
deleted file mode 100644
index abc4e9ef89..0000000000
--- a/llvm/lib/Transforms/DFG2LLVM_CUDNN/DFG2LLVM_CUDNN.cpp
+++ /dev/null
@@ -1,609 +0,0 @@
-//=== DFG2LLVM_CUDNN.cpp ===//
-//
-// The LLVM Compiler Infrastructure
-//
-// This file is distributed under the University of Illinois Open Source
-// License. See LICENSE.TXT for details.
-//
-//===----------------------------------------------------------------------===//
-#define ENABLE_ASSERTS
-
-#define DEBUG_TYPE "DFG2LLVM_CUDNN"
-#include "llvm/IR/DataLayout.h"
-#include "llvm/IR/IRBuilder.h"
-#include "llvm/IR/Module.h"
-#include "llvm/Pass.h"
-#include "llvm/IR/InstIterator.h"
-#include "llvm/Transforms/Utils/ValueMapper.h"
-#include "llvm/Transforms/Utils/BasicBlockUtils.h"
-#include "llvm/Transforms/Utils/Cloning.h"
-#include "llvm/IRReader/IRReader.h"
-#include "llvm/Linker/Linker.h"
-#include "llvm/Support/SourceMgr.h"
-#include "llvm/Support/FileSystem.h"
-#include "llvm/IR/Attributes.h"
-#include "llvm-c/Core.h"
-#include "llvm/SupportVISC/VISCTimer.h"
-#include "llvm/SupportVISC/DFG2LLVM.h"
-#include "llvm/InPlaceDFG/InPlaceDFGAnalysis.h"
-#include <sstream>
-
-using namespace llvm;
-using namespace builddfg;
-using namespace dfg2llvm;
-
-using namespace inplacedfg;
-
-namespace {
-// Helper class declarations
-
-// DFG2LLVM_CUDNN - The first implementation.
-
-struct DFG2LLVM_CUDNN : public DFG2LLVM {
- static char ID; // Pass identification, replacement for typeid
- DFG2LLVM_CUDNN() : DFG2LLVM(ID) {}
-private:
-
-public:
-
- void getAnalysisUsage(AnalysisUsage &AU) const {
- AU.addRequired<BuildDFG>();
- AU.addRequired<InPlaceDFGAnalysisWrapper>();
- AU.addPreserved<BuildDFG>();
- AU.addPreserved<InPlaceDFGAnalysisWrapper>();
- }
-
- bool runOnModule(Module &M);
-};
-
-// Visitor for Code generation traversal (tree traversal for now)
-class CGT_CUDNN : public CodeGenTraversal {
-
-private:
- //Member variables
- InPlaceDFGAnalysis::InPlaceDFGParameter *IPP;
-
- // VISC Runtime API and Tensor runtime API
- Constant* llvm_hpvm_initTensorRt;
- Constant* llvm_hpvm_cleanupTensorRt;
- Constant* hpvm_request_tensor;
-
- // Functions
- bool isValidOperandForInPlaceOperation(Value *Op, Function *Fgen, DFNode *N);
-
-
-
- // Virtual Functions
- void init();
- void initRuntimeAPI();
- void codeGen(DFInternalNode* N);
- void codeGen(DFLeafNode* N);
-
-public:
-
- // Constructor
- CGT_CUDNN(Module &_M, BuildDFG &_DFG, InPlaceDFGAnalysis::InPlaceDFGParameter &_IPP)
- : CodeGenTraversal(_M, _DFG), IPP(&_IPP) {
- initRuntimeAPI();
- }
-
-};
-
-bool CGT_CUDNN::isValidOperandForInPlaceOperation(Value *Op,
- Function *Fgen,
- DFNode *N) {
-
- if (Argument *Arg = dyn_cast<Argument>(Op)) {
- DEBUG(errs() << *Arg << "\t: argument, candidate for in place\n");
- assert((Arg->getParent() == Fgen) &&
- "Extra Parameter in body of Function\n");
- // Candidae parameter is a function argument
- // In this case, consult the result of in place analysis
- // Find position in arg list
- unsigned pos = Arg->getArgNo();
- // If this parameter cannot be used for in place operation
- // code gen cannot continue
- if (IPP->at(N)[pos]) {
- DEBUG(errs() << *Arg << "\t: argument, suitable for in place\n");
- return true;
- } else {
- DEBUG(errs() << *Arg << "\t: argument, not suitable for in place\n");
- return false;
- }
- }
- else {
- // If it is not an argument, then it needs to be the result of
- // another intrinsic. These are new objects that are allocated,
- // and consumed by next intrinsic.
- DEBUG(errs() << *Op << "\t: Test for result of intrinsic operation\n");
- if (dyn_cast<IntrinsicInst>(Op)) {
- DEBUG(errs() << *Arg << "\t: local, suitable for in place\n");
- return true;
- } else {
- DEBUG(errs() << *Arg << "\t: local, not suitable for in place\n");
- return false;
- }
- }
-}
-
-
-void CGT_CUDNN::init() {
-}
-
-// Initialize the VISC runtime API. This makes it easier to insert these calls
-void CGT_CUDNN::initRuntimeAPI() {
-
- // 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!\n");
-
- // FIXME: set correct path
- Twine llvmSrcRoot = LLVM_SRC_ROOT;
- Twine runtimeAPI = llvmSrcRoot+"/projects/hpvm-tensor-rt/lib/tensor_runtime.ll";
- runtimeModule = parseIRFile(runtimeAPI.str(), Err, M.getContext());
- if(runtimeModule == nullptr)
- DEBUG(errs() << Err.getMessage());
- else
- DEBUG(errs() << "Successfully loaded hpvm-tensor-rt API module\n");
-
- // Get or insert Global declarations for
- // - initialization
- // - cleanup
- // - request a tensor
- DECLARE(llvm_hpvm_initTensorRt);
- DECLARE(llvm_hpvm_cleanupTensorRt);
- DECLARE(hpvm_request_tensor);
-
- // Find visc.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");
- InitCall = cast<Instruction>(*VI->user_begin());
- CallInst::Create(llvm_hpvm_initTensorRt,
- 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());
- CallInst::Create(llvm_hpvm_cleanupTensorRt, ArrayRef<Value*>(), "", CleanupCall);
-
-}
-
-void CGT_CUDNN::codeGen(DFInternalNode* N) {
- errs () << "Inside node: " << N->getFuncPointer()->getName() << "\n";
- errs () << "Skipping internal node\n";
-}
-
-
-void CGT_CUDNN::codeGen(DFLeafNode* N) {
-
- // Skip code generation if it is a dummy node
- if(N->isDummyNode()) {
- DEBUG(errs() << "Skipping dummy node\n");
- return;
- }
-
- // Abort code generation if it is an allocation node
- if(N->isAllocationNode()) {
- assert(false && "Allocation Node not expected in ApproxHPVM");
- return;
- }
-
- // Generate code only if it has the right hint
- if (!checkPreferredTarget(N, visc::CUDNN_TARGET)) {
- errs() << "Skipping node: "<< N->getFuncPointer()->getName() << "\n";
- return;
- }
-
- // Get the function associated with the dataflow node
- Function *F = N->getFuncPointer();
- errs()<<"function name = "<< F->getName()<<"\n";
-
- /* Removing HPVM in/out/inout function attributes */
- for(Function::arg_iterator ai = F->arg_begin(), ae = F->arg_end(); ai != ae; ai++){
- Argument *Arg = &*ai;
- if(Arg->hasAttribute(Attribute::In))
- Arg->removeAttr(Attribute::In);
- if(Arg->hasAttribute(Attribute::Out))
- Arg->removeAttr(Attribute::Out);
- if(Arg->hasAttribute(Attribute::InOut))
- Arg->removeAttr(Attribute::InOut);
- }
-
- // 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_cudnn = N->getGenFuncForTarget(visc::CUDNN_TARGET);
-
- assert((F_cudnn == NULL) &&
- "Error: Visiting a node for which code already generated");
-
- // Clone the function
- ValueToValueMapTy VMap;
- std::string FName(F->getName().data());
- F_cudnn = CloneFunction(F, VMap);
- F_cudnn->setName(FName + "_cudnn");
- errs()<<"Cloned function name2 = "<<F_cudnn->getName()<<"\n";
- F_cudnn->removeFromParent();
- M.getFunctionList().push_back(F_cudnn);
-
- N->addGenFunc(F_cudnn, visc::CUDNN_TARGET, true);
-
- // Adding nounwind to generated function : FIXME: needed?
- DEBUG(errs() << "Adding nounwind to generated function\n");
- F_cudnn->addAttribute(AttributeSet::FunctionIndex, Attribute::NoUnwind);
-
- // Add llvm_visc_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_cudnn->getEntryBlock().begin());
-
- // In this backend, the target device is GPU, represented by i32 1.
- ConstantInt *TargetDeviceID =
- ConstantInt::get(Type::getInt32Ty(M.getContext()), 1);
-
- for (Function::arg_iterator ai = F_cudnn->arg_begin(),
- ae = F_cudnn->arg_end(); ai != ae; ++ai) {
- Argument* Arg = &*ai;
- if (Arg->getType()->isPointerTy()) {
- Value *Args[] = {Arg, TargetDeviceID};
- CallInst::Create(hpvm_request_tensor,
- ArrayRef<Value*>(Args, 2),
- "", FI);
- }
- }
-
- std::vector<IntrinsicInst *> IItoRemove;
-
- for (inst_iterator i = inst_begin(F_cudnn), e = inst_end(F_cudnn); i != e; ++i) {
- Instruction *I = &(*i);
-
- 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");
-
- /********************* Handle VISC Tensor intrinsics ********************/
- switch (II->getIntrinsicID()) {
-
- case Intrinsic::visc_tensor_convolution:
- { /* llvm.hpvm.tensor.mul */
- // Tensor mul is not in place.
- DEBUG(errs() << F_cudnn->getName() << "\t: Handling tensor convolution \n");
-
- // Argument list for the runtime call
- std::vector<Value*> Args;
- Args.push_back(II->getOperand(0));
- Args.push_back(II->getOperand(1));
- Args.push_back(II->getOperand(2));
- Args.push_back(II->getOperand(3));
- Args.push_back(II->getOperand(4));
- Args.push_back(II->getOperand(5));
-
- Constant* conv_mode = ConstantInt::get(Type::getInt32Ty(M.getContext()), 1);
- Constant* conv_precision = ConstantInt::get(Type::getInt32Ty(M.getContext()), 0);
-
- Args.push_back(conv_mode);
- Args.push_back(conv_precision);
-
- // Create cudnn runtime function call
- Constant* tensorConvolution;
- DECLARE(tensorConvolution);
-
- CallInst* CI = CallInst::Create(tensorConvolution,
- Args, "", II);
- // We can replace the call to hpvm.tensor.mul with the runtime call
- II->replaceAllUsesWith(CI);
-
- // Mark to remove at the end
- IItoRemove.push_back(II);
- }
- break;
-
- case Intrinsic::visc_tensor_group_convolution:
- { /* llvm.hpvm.tensor.mul */
- // Tensor mul is not in place.
- DEBUG(errs() << F_cudnn->getName() << "\t: Handling tensor convolution \n");
-
- // Argument list for the runtime call
- std::vector<Value*> Args;
- Args.push_back(II->getOperand(0));
- Args.push_back(II->getOperand(1));
- Args.push_back(II->getOperand(2));
- Args.push_back(II->getOperand(3));
- Args.push_back(II->getOperand(4));
- Args.push_back(II->getOperand(5));
-
- Constant* conv_mode = ConstantInt::get(Type::getInt32Ty(M.getContext()), 1);
-
- Args.push_back(conv_mode);
- Args.push_back(II->getOperand(7));
-
- // Create cudnn runtime function call
- Constant* tensorConvolution;
- DECLARE(tensorConvolution);
-
- CallInst* CI = CallInst::Create(tensorConvolution,
- Args, "", II);
- // We can replace the call to hpvm.tensor.mul with the runtime call
- II->replaceAllUsesWith(CI);
-
- // Mark to remove at the end
- IItoRemove.push_back(II);
- }
- break;
-
- case Intrinsic::visc_tensor_batchnorm:
- { /* llvm.hpvm.tensor.batchnorm */
- // Tensor batchnorm is in place.
- // FIXME: Add Check for InPlace Analysis
- DEBUG(errs() << F_cudnn->getName() << "\t: Handling tensor batch normalization \n");
-
- // Argument list for the runtime call
- std::vector<Value*> Args;
- Args.push_back(II->getOperand(0));
- Args.push_back(II->getOperand(1));
- Args.push_back(II->getOperand(2));
- Args.push_back(II->getOperand(3));
- Args.push_back(II->getOperand(4));
- Args.push_back(II->getOperand(5));
-
- // Create cudnn runtime function call
- Constant* tensorBatchNorm;
- DECLARE(tensorBatchNorm);
-
- CallInst* CI = CallInst::Create(tensorBatchNorm,
- Args, "", II);
- // We can replace the call to hpvm.tensor.batchnorm with the TensorRT call
- II->replaceAllUsesWith(CI);
-
- // Mark to remove at the end
- IItoRemove.push_back(II);
- }
- break;
-
-
- case Intrinsic::visc_tensor_mul:
- { /* llvm.hpvm.tensor.mul */
- // Tensor mul is not in place.
- DEBUG(errs() << F_cudnn->getName() << "\t: Handling tensor mul\n");
-
- // Argument list for the runtime call
- std::vector<Value*> Args;
- Args.push_back(II->getOperand(0));
- Args.push_back(II->getOperand(1));
-
- // Create cudnn runtime function call
- Constant* tensorGemmGPU;
- DECLARE(tensorGemmGPU);
-
- CallInst* CI = CallInst::Create(tensorGemmGPU,
- Args, "", II);
- // We can replace the call to hpvm.tensor.mul with the runtime call
- II->replaceAllUsesWith(CI);
-
- // Mark to remove at the end
- IItoRemove.push_back(II);
- }
- break;
- case Intrinsic::visc_tensor_add:
- { /* llvm.hpvm.tensor.add */
- DEBUG(errs() << F_cudnn->getName() << "\t: Handling tensor add\n");
- // Tensor add(a,b) is in place for argument a.
- Value *Op = II->getOperand(0);
-
- // Test the intrinsic operand for in place operation.
- bool inplace = isValidOperandForInPlaceOperation(Op, F_cudnn, N);
- // Code generation cannot continue if this is false, because the target
- // only provides an in place operation
-
- // FIXME: remove this comment - must check for in-place
- //assert(inplace &&
- // "Operand not valid for in place operation. Code gen aborted.\n");
-
- // Argument list for the runtime call
- std::vector<Value*> Args;
- Args.push_back(II->getOperand(0));
- Args.push_back(II->getOperand(1));
-
- // Create cudnn runtime function call
- Constant* tensorAdd;
- DECLARE(tensorAdd);
- CallInst::Create(tensorAdd, Args, "", II);
- // We can replace the call to hpvm.tensor.add with the 1st argument
- // that, due to in place operation, now contains the result
- II->replaceAllUsesWith(II->getOperand(0));
-
- // Mark to remove at the end
- IItoRemove.push_back(II);
- }
- break;
- case Intrinsic::visc_tensor_pool_max:
- case Intrinsic::visc_tensor_pool_mean:
- { /* 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,
- // vertical_stride, horizontal_stride);
- std::vector<Value*> Args;
- Args.push_back(II->getOperand(0));
-
- int pool_type = 0;
- if (II->getIntrinsicID() == Intrinsic::visc_tensor_pool_max){
- pool_type = 0;
- }
- if (II->getIntrinsicID() == Intrinsic::visc_tensor_pool_mean){
- pool_type = 1;
- }
-
- Constant* constPoolType = ConstantInt::get(Type::getInt32Ty(M.getContext()), pool_type);
- Args.push_back(constPoolType); // ID for max pool. Min/Avg have different IDs (non-zero)
- Args.push_back(II->getOperand(1));
- Args.push_back(II->getOperand(2));
- Args.push_back(II->getOperand(3));
- Args.push_back(II->getOperand(4));
- Args.push_back(II->getOperand(5));
- Args.push_back(II->getOperand(6));
-
- // Create cudnn runtime function call
- Constant* tensorPooling;
- DECLARE(tensorPooling);
- CallInst* CI = CallInst::Create(tensorPooling, Args, "", II);
-
- // Replacing intrinsic result uses with the result of the tensor runtime operation
- II->replaceAllUsesWith(CI);
-
- // Mark to remove at the end
- IItoRemove.push_back(II);
- }
- break;
-
- case Intrinsic::visc_tensor_relu:
- case Intrinsic::visc_tensor_clipped_relu:
- case Intrinsic::visc_tensor_tanh:
- { /* llvm.visc.tensor.relu */
- DEBUG(errs() << F_cudnn->getName() << "\t: Handling tensor activation functions \n");
- // Tensor relu(a) is in place for argument a.
- Value *Op = II->getOperand(0);
-
- // Test the intrinsic operand for in place operation.
- bool inplace = isValidOperandForInPlaceOperation(Op, F_cudnn, N);
- // Code generation cannot continue if this is false, because the target
- // only provides an in place operation
- assert(inplace &&
- "Operand not valid for in place operation. Code gen aborted.\n");
-
- // Argument list for the runtime call
- std::vector<Value*> Args;
- Args.push_back(II->getOperand(0));
-
- if (II->getIntrinsicID() == Intrinsic::visc_tensor_relu){
- // Create cudnn runtime function call
- Constant* tensorRelu;
- DECLARE(tensorRelu);
- CallInst::Create(tensorRelu, Args, "", II);
- }
- else if (II->getIntrinsicID() == Intrinsic::visc_tensor_clipped_relu){
- // Create cudnn runtime function call
- //-- Constant* tensorClippedRelu;
- Constant* tensorRelu2;
- DECLARE(tensorRelu2);
- CallInst::Create(tensorRelu2, Args, "", II);
- }
- else if (II->getIntrinsicID() == Intrinsic::visc_tensor_tanh){
- // Create cudnn runtime function call
- Constant* tensorTanh;
- errs()<<"tensorTanh Call = \n\n";
- DECLARE(tensorTanh);
- //errs()<<"tensorTanh Call = "<<*tensorTanh<<"\l";
- CallInst::Create(tensorTanh, Args, "", II);
- }
-
- // We can replace the call to hpvm.tensor.relu with the 1st argument
- // that, due to in place operation, now contains the result
- II->replaceAllUsesWith(II->getOperand(0));
-
- // Mark to remove at the end
- IItoRemove.push_back(II);
- }
- break;
- case Intrinsic::visc_tensor_softmax:
- { /* llvm.visc.tensor.softmax */
- DEBUG(errs() << F_cudnn->getName() << "\t: Handling tensor softmax\n");
- // Tensor relu(a) is in place for argument a.
- Value *Op = II->getOperand(0);
-
- // Test the intrinsic operand for in place operation.
- bool inplace = isValidOperandForInPlaceOperation(Op, F_cudnn, N);
- // Code generation cannot continue if this is false, because the target
- // only provides an in place operation
- assert(inplace &&
- "Operand not valid for in place operation. Code gen aborted.\n");
-
- // Argument list for the runtime call
- std::vector<Value*> Args;
- Args.push_back(II->getOperand(0));
-
- // Create cudnn runtime function call
- Constant* tensorSoftmax;
- DECLARE(tensorSoftmax);
- CallInst::Create(tensorSoftmax, Args, "", II);
- // We can replace the call to hpvm.tensor.softmax with the 1st argument
- // that, due to in place operation, now contains the result
- II->replaceAllUsesWith(II->getOperand(0));
-
- // Mark to remove at the end
- IItoRemove.push_back(II);
- }
- break;
- default:
- llvm_unreachable("Unknown VISC Intrinsic!");
- break;
- }
- }
- }
-
- //--- errs()<<"IIToRemove.size() = "<<IItoRemove.size()<<"\n\n";
-
- // We need to do this explicitly: DCE pass may not remove them.
- // Traverse the vector backwards, otherwise definitions are deleted while
- // their subsequent uses are still around.
- for (std::vector<IntrinsicInst *>::reverse_iterator ri = IItoRemove.rbegin(),
- re = IItoRemove.rend(); ri != re; ++ri) {
- DEBUG(errs() << "Erasing: " << **ri << "\n");
- errs() << "Erasing: " << **ri << "\n";
- (*ri)->eraseFromParent();
- }
-
- return;
-}
-
-bool DFG2LLVM_CUDNN::runOnModule(Module &M) {
- errs() << "\nDFG2LLVM_CUDNN PASS\n";
-
- // Get the BuildDFG Analysis Results:
- // - Dataflow graph
- BuildDFG &DFG = getAnalysis<BuildDFG>();
-
- // 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
- CGT_CUDNN *CGTVisitor = new CGT_CUDNN(M, DFG, IPP);
-
- // Iterate over all the DFGs and produce code for each one of them
- for (auto rootNode: Roots) {
- // Initiate code generation for root DFNode
- CGTVisitor->visit(rootNode);
- }
-
- //TODO: Edit module epilogue to remove the VISC intrinsic declarations
- delete CGTVisitor;
-
- return true;
-}
-
-
-/******************************************************************************
- * Helper functions *
- ******************************************************************************/
-
-
-} // End of namespace
-
-char DFG2LLVM_CUDNN::ID = 0;
-static RegisterPass<DFG2LLVM_CUDNN> X("dfg2llvm-cudnn",
- "Dataflow Graph to LLVM for CUDNN Pass",
- false /* does not modify the CFG */,
- true /* transformation, *
- * not just analysis */);
-
diff --git a/llvm/lib/Transforms/DFG2LLVM_CUDNN/DFG2LLVM_CUDNN.exports b/llvm/lib/Transforms/DFG2LLVM_CUDNN/DFG2LLVM_CUDNN.exports
deleted file mode 100644
index e69de29bb2..0000000000
diff --git a/llvm/lib/Transforms/DFG2LLVM_CUDNN/LLVMBuild.txt b/llvm/lib/Transforms/DFG2LLVM_CUDNN/LLVMBuild.txt
deleted file mode 100644
index 1579b2fc47..0000000000
--- a/llvm/lib/Transforms/DFG2LLVM_CUDNN/LLVMBuild.txt
+++ /dev/null
@@ -1,21 +0,0 @@
-;===- ./lib/Transforms/DFG2LLVM_NVPTX/LLVMBuild.txt ------------*- Conf -*--===;
-;
-; The LLVM Compiler Infrastructure
-;
-; This file is distributed under the University of Illinois Open Source
-; License. See LICENSE.TXT for details.
-;
-;===------------------------------------------------------------------------===;
-;
-; This is an LLVMBuild description file for the components in this subdirectory.
-;
-; For more information on the LLVMBuild system, please see:
-;
-; http://llvm.org/docs/LLVMBuild.html
-;
-;===------------------------------------------------------------------------===;
-
-[component_0]
-type = Library
-name = DFG2LLVM_CUDNN
-parent = Transforms
diff --git a/llvm/lib/Transforms/DFG2LLVM_PROMISE/CMakeLists.txt b/llvm/lib/Transforms/DFG2LLVM_PROMISE/CMakeLists.txt
deleted file mode 100644
index 5b5d2677d0..0000000000
--- a/llvm/lib/Transforms/DFG2LLVM_PROMISE/CMakeLists.txt
+++ /dev/null
@@ -1,12 +0,0 @@
-if(WIN32 OR CYGWIN)
- set(LLVM_LINK_COMPONENTS Core Support)
-endif()
-
-add_llvm_loadable_module( LLVMDFG2LLVM_PROMISE
- DFG2LLVM_PROMISE.cpp
-
- DEPENDS
- intrinsics_gen
- PLUGIN_TOOL
- opt
- )
diff --git a/llvm/lib/Transforms/DFG2LLVM_PROMISE/DFG2LLVM_PROMISE.cpp b/llvm/lib/Transforms/DFG2LLVM_PROMISE/DFG2LLVM_PROMISE.cpp
deleted file mode 100644
index 184f92910a..0000000000
--- a/llvm/lib/Transforms/DFG2LLVM_PROMISE/DFG2LLVM_PROMISE.cpp
+++ /dev/null
@@ -1,1283 +0,0 @@
-//=== DFG2LLVM_PROMISE.cpp ===//
-//
-// The LLVM Compiler Infrastructure
-//
-// This file is distributed under the University of Illinois Open Source
-// License. See LICENSE.TXT for details.
-//
-//===----------------------------------------------------------------------===//
-#define ENABLE_ASSERTS
-
-#define DEBUG_TYPE "DFG2LLVM_PROMISE"
-#include "llvm/IR/DataLayout.h"
-#include "llvm/IR/IRBuilder.h"
-#include "llvm/IR/Module.h"
-#include "llvm/Pass.h"
-#include "llvm/IR/InstIterator.h"
-#include "llvm/Transforms/Utils/ValueMapper.h"
-#include "llvm/Transforms/Utils/BasicBlockUtils.h"
-#include "llvm/Transforms/Utils/Cloning.h"
-#include "llvm/IRReader/IRReader.h"
-#include "llvm/Linker/Linker.h"
-#include "llvm/Support/SourceMgr.h"
-#include "llvm/Support/FileSystem.h"
-#include "llvm/IR/Attributes.h"
-#include "llvm-c/Core.h"
-#include "llvm/SupportVISC/VISCTimer.h"
-#include "llvm/SupportVISC/DFG2LLVM.h"
-#include "llvm/InPlaceDFG/InPlaceDFGAnalysis.h"
-#include <sstream>
-#include <fstream>
-
-using namespace llvm;
-using namespace builddfg;
-using namespace dfg2llvm;
-
-namespace {
-
-cl::opt<std::string> QuantizationInputsFilename(
- "quantization-levels-filename",
- cl::desc("<PROMISE quantization levels input file (path)>"),
- cl::value_desc("filename"),
- cl::Required);
-
-// Helper class declarations
-
-// State machine definition for pattern identification
-
-/* An assumption is made for the PROMISE simulator: *
- * a leaf node will contain consequtive operations that will map to a *
- * single PROMISE simulator call *
-
- * To alleviate that, the states that correspond to valid patterns *
- * - (FullyConnectedLayer_(2,3,x, ConvilutionLayer_(2,3,4,x)) *
- * can invoke codeGen when detecting the beginning of a new pattern, then *
- * clear the collected IIs and Args, then go to initial and invoke its *
- * transition. */
-
-class AbstractState;
-
-class CodeGenStateMachine {
-private:
- Module *M;
- Module *RtM;
-
- std::ifstream &qin; // Quantization levels input stream reference
- std::vector<Value*> Args;
- std::vector<IntrinsicInst*> IIs;
- AbstractState *current;
-
-public:
- CodeGenStateMachine(Module *, Module *, std::ifstream &);
-
- void setCurrent(AbstractState *s) {
- current = s;
- }
-
- void transition(IntrinsicInst *II);
-
- Module *getModule() {
- return M;
- }
-
- void getNextQuantizationLevel(float &ql) {
- qin >> ql;
- }
-
- void addArgument(Value *Arg) {
- Args.push_back(Arg);
- }
-
- void addIntrinsicInst(IntrinsicInst *II) {
- IIs.push_back(II);
- }
-
- IntrinsicInst *getIntrinsicInstAt(unsigned idx) {
- return IIs[idx];
- }
-
- void codeGen();
-
-};
-
-class AbstractState {
-public:
- enum ID
- {
- INITIAL_STATE,
- FULLY_CONNECTED_LAYER_1,
- FULLY_CONNECTED_LAYER_2,
- FULLY_CONNECTED_LAYER_3,
- FULLY_CONNECTED_LAYER,
- CONVOLUTION_LAYER_1,
- CONVOLUTION_LAYER_2,
- CONVOLUTION_LAYER_3,
- CONVOLUTION_LAYER_4,
- CONVOLUTION_LAYER,
- NO_PATTERN,
- };
-
-protected:
- enum ID StateID;
-
-public:
- enum ID getStateID() {
- return StateID;
- }
-
- virtual void transition(CodeGenStateMachine *Mch, IntrinsicInst *II) = 0;
- virtual ~AbstractState() {}
-};
-
-class InitialState : public AbstractState {
-public:
- InitialState() {
- StateID = ID::INITIAL_STATE;
- DEBUG(errs() << "new InitialState\n");
- }
- ~InitialState() {}
-
- void transition(CodeGenStateMachine *Mch, IntrinsicInst *II) override;
-};
-
-class FullyConnectedLayer_1 : public AbstractState {
-public:
- FullyConnectedLayer_1() {
- StateID = ID::FULLY_CONNECTED_LAYER_1;
- DEBUG(errs() << "new FullyConnectedLayer_1\n");
- }
- ~FullyConnectedLayer_1() {}
-
- void transition(CodeGenStateMachine *Mch, IntrinsicInst *II) override;
-};
-
-class FullyConnectedLayer_2 : public AbstractState {
-public:
- FullyConnectedLayer_2() {
- StateID = ID::FULLY_CONNECTED_LAYER_2;
- DEBUG(errs() << "new FullyConnectedLayer_2\n");
- }
- ~FullyConnectedLayer_2() {}
-
- void transition(CodeGenStateMachine *Mch, IntrinsicInst *II) override;
-};
-
-class FullyConnectedLayer_3 : public AbstractState {
-public:
- FullyConnectedLayer_3() {
- StateID = ID::FULLY_CONNECTED_LAYER_3;
- DEBUG(errs() << "new FullyConnectedLayer_3\n");
- }
- ~FullyConnectedLayer_3() {}
-
- void transition(CodeGenStateMachine *Mch, IntrinsicInst *II) override;
-};
-
-class FullyConnectedLayer : public AbstractState {
-public:
- FullyConnectedLayer() {
- StateID = ID::FULLY_CONNECTED_LAYER;
- DEBUG(errs() << "new FullyConnectedLayer\n");
- }
- ~FullyConnectedLayer() {}
-
- void transition(CodeGenStateMachine *Mch, IntrinsicInst *II) override;
-};
-
-class ConvolutionLayer_1 : public AbstractState {
-public:
- ConvolutionLayer_1() {
- StateID = ID::CONVOLUTION_LAYER_1;
- DEBUG(errs() << "new ConvolutionLayer_1\n");
- }
- ~ConvolutionLayer_1() {}
-
- void transition(CodeGenStateMachine *Mch, IntrinsicInst *II) override;
-};
-
-class ConvolutionLayer_2 : public AbstractState {
-public:
- ConvolutionLayer_2() {
- StateID = ID::CONVOLUTION_LAYER_2;
- DEBUG(errs() << "new ConvolutionLayer_2\n");
- }
- ~ConvolutionLayer_2() {}
-
- void transition(CodeGenStateMachine *Mch, IntrinsicInst *II) override;
-};
-
-class ConvolutionLayer_3 : public AbstractState {
-public:
- ConvolutionLayer_3() {
- StateID = ID::CONVOLUTION_LAYER_3;
- DEBUG(errs() << "new ConvolutionLayer_3\n");
- }
- ~ConvolutionLayer_3() {}
-
- void transition(CodeGenStateMachine *Mch, IntrinsicInst *II) override;
-};
-
-class ConvolutionLayer_4 : public AbstractState {
-public:
- ConvolutionLayer_4() {
- StateID = ID::CONVOLUTION_LAYER_4;
- DEBUG(errs() << "new ConvolutionLayer_4\n");
- }
- ~ConvolutionLayer_4() {}
-
- void transition(CodeGenStateMachine *Mch, IntrinsicInst *II) override;
-};
-
-class ConvolutionLayer : public AbstractState {
-public:
- ConvolutionLayer() {
- StateID = ID::CONVOLUTION_LAYER;
- DEBUG(errs() << "new ConvolutionLayer\n");
- }
- ~ConvolutionLayer() {}
-
- void transition(CodeGenStateMachine *Mch, IntrinsicInst *II) override;
-};
-
-class NoPattern : public AbstractState {
-public:
- NoPattern() {
- StateID = ID::NO_PATTERN;
- DEBUG(errs() << "new NoPattern\n");
- }
- ~NoPattern() {}
-
- void transition(CodeGenStateMachine *Mch, IntrinsicInst *II) override;
-};
-
-void InitialState::transition(CodeGenStateMachine *Mch, IntrinsicInst *II) {
- if (II) { // Not end of instruction stream
- switch (II->getIntrinsicID()) {
- case Intrinsic::visc_tensor_convolution:
- {
- Mch->addIntrinsicInst(II);
- Mch->addArgument(II->getOperand(0)); // conv input
-
- // Read quantization levels for input
- float i_min, i_max;
- Mch->getNextQuantizationLevel(i_min);
- Mch->getNextQuantizationLevel(i_max);
- errs() << "i_min: " << i_min << "\n";
- errs() << "i_max: " << i_max << "\n";
-
- // Create associated arguments for the quantization levels
- Constant *IminC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) i_min);
-// errs() << "IminC : "
-// << dyn_cast<ConstantFP>(IminC)->getValueAPF().convertToFloat()
-// << "\n";
- Mch->addArgument(IminC);
- Constant *ImaxC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) i_max);
- Mch->addArgument(ImaxC);
-
- Mch->addArgument(II->getOperand(1)); // conv kernel
-
- // Read quantization levels for filter
- float w_min, w_max;
- Mch->getNextQuantizationLevel(w_min);
- Mch->getNextQuantizationLevel(w_max);
- errs() << "w_min: " << w_min << "\n";
- errs() << "w_max: " << w_max << "\n";
- Constant *WminC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) w_min);
- Mch->addArgument(WminC);
- Constant *WmaxC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) w_max);
- Mch->addArgument(WmaxC);
-
- Mch->setCurrent(new ConvolutionLayer_1());
- }
- break;
- case Intrinsic::visc_tensor_mul:
- {
- Mch->addIntrinsicInst(II);
- Mch->addArgument(II->getOperand(0)); // 1st gemm input
-
- // Read quantization levels for input
- float i_min, i_max;
- Mch->getNextQuantizationLevel(i_min);
- Mch->getNextQuantizationLevel(i_max);
- errs() << "i_min: " << i_min << "\n";
- errs() << "i_max: " << i_max << "\n";
-
- Constant *IminC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) i_min);
- Mch->addArgument(IminC);
- Constant *ImaxC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) i_max);
- Mch->addArgument(ImaxC);
-
- Mch->addArgument(II->getOperand(1)); // 2nd gemm input
-
- // Read quantization levels for weight
- float w_min, w_max;
- Mch->getNextQuantizationLevel(w_min);
- Mch->getNextQuantizationLevel(w_max);
- errs() << "w_min: " << w_min << "\n";
- errs() << "w_max: " << w_max << "\n";
-
- Constant *WminC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) w_min);
- Mch->addArgument(WminC);
- Constant *WmaxC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) w_max);
- Mch->addArgument(WmaxC);
-
- Mch->setCurrent(new FullyConnectedLayer_1());
- }
- break;
- default: // Other HPVM intrinsic
- Mch->setCurrent(new NoPattern());
- break;
- }
- delete this;
- } // else {} // No HPVM intrinsic received. Remain at initial
-}
-
-void FullyConnectedLayer_1::transition(CodeGenStateMachine *Mch,
- IntrinsicInst *II) {
- if (II) { // Not end of instruction stream
- switch (II->getIntrinsicID()) {
- case Intrinsic::visc_tensor_add:
- {
- IntrinsicInst *MulII = Mch->getIntrinsicInstAt(0);
- assert((MulII == II->getOperand(0)) &&
- "Output of mul must be used as 1st operand of add");
- Mch->addIntrinsicInst(II);
-
- Mch->addArgument(II->getOperand(1)); // bias
-
- // Read quantization levels for input
- float b_min, b_max;
- Mch->getNextQuantizationLevel(b_min);
- Mch->getNextQuantizationLevel(b_max);
- errs() << "b_min: " << b_min << "\n";
- errs() << "b_max: " << b_max << "\n";
-
- Constant *BminC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) b_min);
- Mch->addArgument(BminC);
- Constant *BmaxC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) b_max);
- Mch->addArgument(BmaxC);
-
- Mch->setCurrent(new FullyConnectedLayer_2());
- }
- break;
- default:
- Mch->setCurrent(new NoPattern());
- break;
- }
- } else {
- Mch->setCurrent(new NoPattern());
- }
- delete this;
-}
-
-void FullyConnectedLayer_2::transition(CodeGenStateMachine *Mch,
- IntrinsicInst *II) {
- if (II) { // Not end of instruction stream
- switch (II->getIntrinsicID()) {
- case Intrinsic::visc_tensor_tanh:
- {
- // Type of activation : TanH
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 0));
-
- // Read quantization levels for output
- float out_min, out_max;
- Mch->getNextQuantizationLevel(out_min);
- Mch->getNextQuantizationLevel(out_max);
- errs() << "out_min: " << out_min << "\n";
- errs() << "out_max: " << out_max << "\n";
-
- Constant *OutminC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) out_min);
- Mch->addArgument(OutminC);
- Constant *OutmaxC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) out_max);
- Mch->addArgument(OutmaxC);
-
- Mch->addIntrinsicInst(II);
-
- Mch->setCurrent(new FullyConnectedLayer_3());
- }
- break;
- case Intrinsic::visc_tensor_relu:
- {
- // Type of activation : ReLU
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 1));
-
- // Read quantization levels for output
- float out_min, out_max;
- Mch->getNextQuantizationLevel(out_min);
- Mch->getNextQuantizationLevel(out_max);
- errs() << "out_min: " << out_min << "\n";
- errs() << "out_max: " << out_max << "\n";
-
- Constant *OutminC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) out_min);
- Mch->addArgument(OutminC);
- Constant *OutmaxC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) out_max);
- Mch->addArgument(OutmaxC);
-
- Mch->addIntrinsicInst(II);
-
- Mch->setCurrent(new FullyConnectedLayer_3());
- }
- break;
- case Intrinsic::visc_tensor_clipped_relu:
- {
- // Type of activation : Clipped ReLU
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 2));
-
- // Read quantization levels for output
- float out_min, out_max;
- Mch->getNextQuantizationLevel(out_min);
- Mch->getNextQuantizationLevel(out_max);
- errs() << "out_min: " << out_min << "\n";
- errs() << "out_max: " << out_max << "\n";
-
- Constant *OutminC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) out_min);
- Mch->addArgument(OutminC);
- Constant *OutmaxC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) out_max);
- Mch->addArgument(OutmaxC);
-
- Mch->addIntrinsicInst(II);
-
- Mch->setCurrent(new FullyConnectedLayer_3());
- }
- break;
- default: // No activation, but HPVM intrinsic
- Mch->setCurrent(new NoPattern());
- break;
- }
- } else { // End of instruction stream
- // No activation
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), -1));
-
- // Read quantization levels for output
- float out_min, out_max;
- Mch->getNextQuantizationLevel(out_min);
- Mch->getNextQuantizationLevel(out_max);
- errs() << "out_min: " << out_min << "\n";
- errs() << "out_max: " << out_max << "\n";
-
- Constant *OutminC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) out_min);
- Mch->addArgument(OutminC);
- Constant *OutmaxC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) out_max);
- Mch->addArgument(OutmaxC);
-
- Mch->setCurrent(new FullyConnectedLayer());
- }
- delete this;
-}
-
-void FullyConnectedLayer_3::transition(CodeGenStateMachine *Mch,
- IntrinsicInst *II) {
- if (!II) { // End of instruction stream
- Mch->setCurrent(new FullyConnectedLayer());
- } else {
- Mch->setCurrent(new NoPattern());
- }
- delete this;
-}
-
-void FullyConnectedLayer::transition(CodeGenStateMachine *Mch,
- IntrinsicInst *II) {
- if (II) { // Not end of instruction stream
- Mch->setCurrent(new NoPattern());
- delete this;
- }
-}
-
-void ConvolutionLayer_1::transition(CodeGenStateMachine *Mch,
- IntrinsicInst *II) {
- if (II) { // Not end of instruction stream
- switch (II->getIntrinsicID()) {
- case Intrinsic::visc_tensor_add:
- {
- IntrinsicInst *ConvII = Mch->getIntrinsicInstAt(0);
- assert((ConvII == II->getOperand(0)) &&
- "Output of conv must be used as 1st operand of add");
- Mch->addIntrinsicInst(II);
-
- Mch->addArgument(II->getOperand(1)); // bias
- // Read quantization levels for bias
- float b_min, b_max;
- Mch->getNextQuantizationLevel(b_min);
- Mch->getNextQuantizationLevel(b_max);
- errs() << "b_min: " << b_min << "\n";
- errs() << "b_max: " << b_max << "\n";
-
- Constant *BminC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) b_min);
- Mch->addArgument(BminC);
- Constant *BmaxC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) b_max);
- Mch->addArgument(BmaxC);
-
- Mch->addArgument(ConvII->getOperand(2)); // 1st numeric arg of conv
- 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->setCurrent(new ConvolutionLayer_2());
- }
- break;
- default:
- Mch->setCurrent(new NoPattern());
- break;
- }
- } else {
- // No addition
- Mch->addArgument(ConstantPointerNull::get(
- Type::getInt8PtrTy(Mch->getModule()->getContext())));
- // Still need to add the quantization constants - and remove them from file
- float b_min, b_max;
- Mch->getNextQuantizationLevel(b_min);
- Mch->getNextQuantizationLevel(b_max);
- errs() << "b_min: " << b_min << "\n";
- errs() << "b_max: " << b_max << "\n";
- Constant *BminC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) b_min);
- Mch->addArgument(BminC);
- Constant *BmaxC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) b_max);
- Mch->addArgument(BmaxC);
-
- // Zero for all convolution numeric arguments FIXME???
- IntrinsicInst *ConvII = Mch->getIntrinsicInstAt(0);
- Mch->addArgument(ConvII->getOperand(2)); // 1st numeric arg of conv
- 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));
-
- // No pooling
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 0));
- // 0 for unused pool argument
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 0));
- // No activation
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), -1));
-
- // Read quantization levels for output
- float out_min, out_max;
- Mch->getNextQuantizationLevel(out_min);
- Mch->getNextQuantizationLevel(out_max);
- errs() << "out_min: " << out_min << "\n";
- errs() << "out_max: " << out_max << "\n";
-
- Constant *OutminC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) out_min);
- Mch->addArgument(OutminC);
- Constant *OutmaxC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) out_max);
- Mch->addArgument(OutmaxC);
-
- Mch->setCurrent(new ConvolutionLayer());
- }
- delete this;
-}
-
-void ConvolutionLayer_2::transition(CodeGenStateMachine *Mch,
- IntrinsicInst *II) {
- if (II) { // Not end of instruction stream
- switch (II->getIntrinsicID()) {
- case Intrinsic::visc_tensor_tanh:
- {
- // Type of activation : TanH
-// Mch->addArgument(ConstantInt::get(
-// Type::getInt32Ty(Mch->getModule()->getContext()), 0));
- Mch->addIntrinsicInst(II);
-
- Mch->setCurrent(new ConvolutionLayer_3());
- }
- break;
- case Intrinsic::visc_tensor_relu:
- {
- // Type of activation : ReLU
-// Mch->addArgument(ConstantInt::get(
-// Type::getInt32Ty(Mch->getModule()->getContext()), 1));
- Mch->addIntrinsicInst(II);
-
- Mch->setCurrent(new ConvolutionLayer_3());
- }
- break;
- case Intrinsic::visc_tensor_clipped_relu:
- {
- // Type of activation : Clipped ReLU
-// Mch->addArgument(ConstantInt::get(
-// Type::getInt32Ty(Mch->getModule()->getContext()), 2));
- Mch->addIntrinsicInst(II);
-
- Mch->setCurrent(new ConvolutionLayer_3());
- }
- break;
- case Intrinsic::visc_tensor_pool_max:
- {
- // pool max
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 0));
- // poolSize
- Mch->addArgument(II->getOperand(1));
- // No activation
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), -1));
- Mch->addIntrinsicInst(II);
-
- // Read quantization levels for output
- float out_min, out_max;
- Mch->getNextQuantizationLevel(out_min);
- Mch->getNextQuantizationLevel(out_max);
- errs() << "out_min: " << out_min << "\n";
- errs() << "out_max: " << out_max << "\n";
-
- Constant *OutminC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) out_min);
- Mch->addArgument(OutminC);
- Constant *OutmaxC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) out_max);
- Mch->addArgument(OutmaxC);
-
- Mch->setCurrent(new ConvolutionLayer_4());
- }
- break;
- case Intrinsic::visc_tensor_pool_min:
- {
- // pool min FIXME: 2: supported?
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 2));
- // poolSize
- Mch->addArgument(II->getOperand(1));
- // No activation
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), -1));
- Mch->addIntrinsicInst(II);
-
- // Read quantization levels for output
- float out_min, out_max;
- Mch->getNextQuantizationLevel(out_min);
- Mch->getNextQuantizationLevel(out_max);
- errs() << "out_min: " << out_min << "\n";
- errs() << "out_max: " << out_max << "\n";
-
- Constant *OutminC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) out_min);
- Mch->addArgument(OutminC);
- Constant *OutmaxC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) out_max);
- Mch->addArgument(OutmaxC);
-
- Mch->setCurrent(new ConvolutionLayer_4());
- }
- break;
- case Intrinsic::visc_tensor_pool_mean:
- {
- // pool mean
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 1));
- // poolSize
- Mch->addArgument(II->getOperand(1));
- // No activation
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), -1));
- Mch->addIntrinsicInst(II);
-
- // Read quantization levels for output
- float out_min, out_max;
- Mch->getNextQuantizationLevel(out_min);
- Mch->getNextQuantizationLevel(out_max);
- errs() << "out_min: " << out_min << "\n";
- errs() << "out_max: " << out_max << "\n";
-
- Constant *OutminC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) out_min);
- Mch->addArgument(OutminC);
- Constant *OutmaxC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) out_max);
- Mch->addArgument(OutmaxC);
-
- Mch->setCurrent(new ConvolutionLayer_4());
- }
- break;
- default: // No activation, No pooling, but HPVM intrinsic
- Mch->setCurrent(new NoPattern());
- break;
- }
- } else { // End of instruction stream
- // No pooling
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 0));
- // 0 for unused pool argument
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 0));
- // No activation
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), -1));
-
- // Read quantization levels for output
- float out_min, out_max;
- Mch->getNextQuantizationLevel(out_min);
- Mch->getNextQuantizationLevel(out_max);
- errs() << "out_min: " << out_min << "\n";
- errs() << "out_max: " << out_max << "\n";
-
- Constant *OutminC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) out_min);
- Mch->addArgument(OutminC);
- Constant *OutmaxC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) out_max);
- Mch->addArgument(OutmaxC);
-
- Mch->setCurrent(new ConvolutionLayer());
- }
- delete this;
-}
-
-void ConvolutionLayer_3::transition(CodeGenStateMachine *Mch,
- IntrinsicInst *II) {
- if (II) { // Not end of instruction stream
- switch (II->getIntrinsicID()) {
- case Intrinsic::visc_tensor_pool_max:
- {
- // pool max
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 0));
- // poolSize
- Mch->addArgument(II->getOperand(1));
- Mch->addIntrinsicInst(II);
-
- // Revisit last intrinsic, to add argument for activation operation
- 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) {
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 0));
- } else if (ActIID == Intrinsic::visc_tensor_relu) {
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 1));
- } else { //ActIID == Intrinsic::visc_tensor_clipped_relu
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 2));
- }
-
- // Read quantization levels for output
- float out_min, out_max;
- Mch->getNextQuantizationLevel(out_min);
- Mch->getNextQuantizationLevel(out_max);
- errs() << "out_min: " << out_min << "\n";
- errs() << "out_max: " << out_max << "\n";
-
- Constant *OutminC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) out_min);
- Mch->addArgument(OutminC);
- Constant *OutmaxC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) out_max);
- Mch->addArgument(OutmaxC);
-
- Mch->setCurrent(new ConvolutionLayer_4());
- }
- break;
- case Intrinsic::visc_tensor_pool_min:
- {
- // pool min FIXME: 2: supported?
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 2));
- // poolSize
- Mch->addArgument(II->getOperand(1));
- Mch->addIntrinsicInst(II);
-
- // Revisit last intrinsic, to add argument for activation operation
- 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) {
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 0));
- } else if (ActIID == Intrinsic::visc_tensor_relu) {
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 1));
- } else { //ActIID == Intrinsic::visc_tensor_clipped_relu
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 2));
- }
-
- // Read quantization levels for output
- float out_min, out_max;
- Mch->getNextQuantizationLevel(out_min);
- Mch->getNextQuantizationLevel(out_max);
- errs() << "out_min: " << out_min << "\n";
- errs() << "out_max: " << out_max << "\n";
-
- Constant *OutminC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) out_min);
- Mch->addArgument(OutminC);
- Constant *OutmaxC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) out_max);
- Mch->addArgument(OutmaxC);
-
- Mch->setCurrent(new ConvolutionLayer_4());
- }
- break;
- case Intrinsic::visc_tensor_pool_mean:
- {
- // pool max
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 1));
- // poolSize
- Mch->addArgument(II->getOperand(1));
- Mch->addIntrinsicInst(II);
-
- // Revisit last intrinsic, to add argument for activation operation
- 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) {
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 0));
- } else if (ActIID == Intrinsic::visc_tensor_relu) {
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 1));
- } else { //ActIID == Intrinsic::visc_tensor_clipped_relu
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 2));
- }
-
- // Read quantization levels for output
- float out_min, out_max;
- Mch->getNextQuantizationLevel(out_min);
- Mch->getNextQuantizationLevel(out_max);
- errs() << "out_min: " << out_min << "\n";
- errs() << "out_max: " << out_max << "\n";
-
- Constant *OutminC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) out_min);
- Mch->addArgument(OutminC);
- Constant *OutmaxC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) out_max);
- Mch->addArgument(OutmaxC);
-
- Mch->setCurrent(new ConvolutionLayer_4());
- }
- break;
- default: // No pooling, but HPVM intrinsic
- Mch->setCurrent(new NoPattern());
- break;
- }
- } else { // End of instruction stream
- // No pooling
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 0));
- // 0 for unused pool argument
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 0));
-
- // Revisit last intrinsic, to add argument for activation operation
- 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) {
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 0));
- } else if (ActIID == Intrinsic::visc_tensor_relu) {
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 1));
- } else { //ActIID == Intrinsic::visc_tensor_clipped_relu
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 2));
- }
-
- // Read quantization levels for output
- float out_min, out_max;
- Mch->getNextQuantizationLevel(out_min);
- Mch->getNextQuantizationLevel(out_max);
- errs() << "out_min: " << out_min << "\n";
- errs() << "out_max: " << out_max << "\n";
-
- Constant *OutminC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) out_min);
- Mch->addArgument(OutminC);
- Constant *OutmaxC = ConstantFP::get(Type::getFloatTy(Mch->getModule()->getContext()),
- (double) out_max);
- Mch->addArgument(OutmaxC);
-
- Mch->setCurrent(new ConvolutionLayer());
- }
- delete this;
-}
-
-void ConvolutionLayer_4::transition(CodeGenStateMachine *Mch,
- IntrinsicInst *II) {
- if (!II) { // End of instruction stream
- Mch->setCurrent(new ConvolutionLayer());
- } else {
- Mch->setCurrent(new NoPattern());
- }
- delete this;
-}
-
-void ConvolutionLayer::transition(CodeGenStateMachine *Mch,
- IntrinsicInst *II) {
- if (II) { // Not end of instruction stream
- Mch->setCurrent(new NoPattern());
- delete this;
- }
-}
-
-void NoPattern::transition(CodeGenStateMachine *Mch, IntrinsicInst *II) {}
-
-CodeGenStateMachine::CodeGenStateMachine(Module *_M, Module *_RtM, std::ifstream &_qin) :
- M(_M), RtM(_RtM), qin(_qin) {
- current = new InitialState();
-}
-
-void CodeGenStateMachine::transition(IntrinsicInst *II) {
- current->transition(this, II);
-}
-
-void CodeGenStateMachine::codeGen() {
-
- if ((current->getStateID() != AbstractState::ID::FULLY_CONNECTED_LAYER) &&
- (current->getStateID() != AbstractState::ID::CONVOLUTION_LAYER)) {
- // Not a valid instruction sequence.
- assert(false && "Unsupported instruction sequence by PROMISE simulator\n");
- }
-
- // We have a valid instruction sequence.
- // Make sure that the instruction sequence can be traslated:
- // each instruction's result must be used only by the next one in sequence.
- for (unsigned p = 0; p < IIs.size()-1; p++) {
- IntrinsicInst *II = IIs[p];
- assert((II->hasOneUse()) &&
- "Instruction sequence does not fit expected pattern: not single use\n");
-
- Value::user_iterator ui = II->user_begin(); // The only use
- assert((*ui == IIs[p+1]) &&
- "Instruction sequence does not fit expected pattern: not used by next instruction\n");
- }
-
- // Create corresponding PROMISE simulator call
- CallInst *CI;
- switch (current->getStateID()) {
- case AbstractState::ID::CONVOLUTION_LAYER:
- {
- Constant* ConvLayer_PROMISE =
- M->getOrInsertFunction(StringRef("ConvLayer_PROMISE"),
- RtM->getFunction(StringRef("ConvLayer_PROMISE"))->getFunctionType());
- DEBUG(errs() << *ConvLayer_PROMISE);
-
- // FIXME: get last argument from some intrinsic. For now, 7
- Args.push_back(ConstantInt::get(Type::getInt32Ty(M->getContext()), 7));
- // Create PROMISE simulator function call
- CI = CallInst::Create(ConvLayer_PROMISE, Args, "");
- }
- break;
- case AbstractState::ID::FULLY_CONNECTED_LAYER:
- {
- Constant* FCLayer_PROMISE =
- M->getOrInsertFunction(StringRef("FCLayer_PROMISE"),
- RtM->getFunction(StringRef("FCLayer_PROMISE"))->getFunctionType());
- DEBUG(errs() << *FCLayer_PROMISE);
-
- // FIXME: get last argument from some intrinsic. For now, 7
- Args.push_back(ConstantInt::get(Type::getInt32Ty(M->getContext()), 7));
- // Create PROMISE simulator function call
- CI = CallInst::Create(FCLayer_PROMISE, Args, "");
- }
- break;
- default:
- llvm_unreachable("Unexpected CodeGenStateMachine State\n");
- break;
- }
-
- // Insert new call and replace all uses of pattern result with
- // the PROMISE simulator call
- IntrinsicInst *IIlast = *(IIs.rbegin());
- CI->insertBefore(IIlast);
- IIlast->replaceAllUsesWith(CI);
-
- // Remove the instructions we translated to the simulator call.
- // Traverse the vector backwards, otherwise definitions are deleted while
- // their subsequent uses are still around.
- for (std::vector<IntrinsicInst *>::reverse_iterator ri = IIs.rbegin(),
- re = IIs.rend(); ri != re; ++ri) {
- DEBUG(errs() << "Erasing: " << **ri << "\n");
- (*ri)->eraseFromParent();
- }
-errs() << "****** GenF:\n" << *(CI->getParent()->getParent());
-
-}
-
-// DFG2LLVM_PROMISE - The first implementation.
-
-struct DFG2LLVM_PROMISE : public DFG2LLVM {
- static char ID; // Pass identification, replacement for typeid
- DFG2LLVM_PROMISE() : DFG2LLVM(ID) {}
-private:
-
-public:
-
- void getAnalysisUsage(AnalysisUsage &AU) const {
- AU.addRequired<BuildDFG>();
- AU.addPreserved<BuildDFG>();
- }
-
- bool runOnModule(Module &M);
-};
-
-// Visitor for Code generation traversal (tree traversal for now)
-class CGT_PROMISE : public CodeGenTraversal {
-
-private:
- //Member variables
- std::ifstream qin;
-
- // VISC Runtime API and Tensor runtime API
- Constant* llvm_hpvm_initTensorRt;
- Constant* llvm_hpvm_cleanupTensorRt;
- Constant* hpvm_request_tensor;
-
- // Functions
-
- // Virtual Functions
- void init();
- void initRuntimeAPI();
- void codeGen(DFInternalNode* N);
- void codeGen(DFLeafNode* N);
-
-public:
-
- // Constructor
- CGT_PROMISE(Module &_M, BuildDFG &_DFG, std::string &_str) : CodeGenTraversal(_M, _DFG) {
- qin.open(_str.c_str());
- assert(qin && "Failed to open quantization levels input file\n");
- initRuntimeAPI();
- }
-
- ~CGT_PROMISE() {
- qin.close();
- }
-
-};
-
-void CGT_PROMISE::init() {
- // FIXME: what to do here? If anything?
-}
-
-// Initialize the VISC runtime API. This makes it easier to insert these calls
-void CGT_PROMISE::initRuntimeAPI() {
-
- // 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!\n");
-
- // FIXME: set correct path
- Twine llvmSrcRoot = LLVM_SRC_ROOT;
- Twine runtimeAPI = llvmSrcRoot+"/projects/hpvm-tensor-rt/lib/tensor_runtime.ll";
- runtimeModule = parseIRFile(runtimeAPI.str(), Err, M.getContext());
- if(runtimeModule == nullptr)
- DEBUG(errs() << Err.getMessage());
- else
- DEBUG(errs() << "Successfully loaded hpvm-tensor-rt API module\n");
-
- // Get or insert Global declarations for
- // - initialization
- // - cleanup
- // - request a tensor
- DECLARE(llvm_hpvm_initTensorRt);
- DECLARE(llvm_hpvm_cleanupTensorRt);
- DECLARE(hpvm_request_tensor);
-
- // Find visc.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");
- InitCall = cast<Instruction>(*VI->user_begin());
- CallInst::Create(llvm_hpvm_initTensorRt,
- 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());
- CallInst::Create(llvm_hpvm_cleanupTensorRt, ArrayRef<Value*>(), "", CleanupCall);
-
-}
-
-void CGT_PROMISE::codeGen(DFInternalNode* N) {
- errs () << "Inside node: " << N->getFuncPointer()->getName() << "\n";
- errs () << "Skipping internal node\n";
-}
-
-void CGT_PROMISE::codeGen(DFLeafNode* N) {
-
- // Skip code generation if it is a dummy node
- if(N->isDummyNode()) {
- DEBUG(errs() << "Skipping dummy node\n");
- return;
- }
-
- // Abort code generation if it is an allocation node
- if(N->isAllocationNode()) {
- assert(false && "Allocation Node not expected in ApproxHPVM");
- return;
- }
-
- // Generate code only if it has the right hint
- if (!checkPreferredTarget(N, visc::PROMISE_TARGET)) {
- errs() << "Skipping node: "<< N->getFuncPointer()->getName() << "\n";
- return;
- }
-
- // Get the function associated with the dataflow node
- Function *F = N->getFuncPointer();
-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.
- Function *F_promise = N->getGenFuncForTarget(visc::PROMISE_TARGET);
-
- assert((F_promise == NULL) &&
- "Error: Visiting a node for which code already generated");
-
- // Clone the function
- ValueToValueMapTy VMap;
- std::string FName(F->getName().data());//Twine FName = F->getName();
- F_promise = CloneFunction(F, VMap);
- F_promise->setName(FName+"_promise");
- F_promise->removeFromParent();
- M.getFunctionList().push_back(F_promise);
-
- N->addGenFunc(F_promise, visc::PROMISE_TARGET, true);
-
- /* Removing HPVM in/out/inout function attributes */
- for(Function::arg_iterator ai = F_promise->arg_begin(), ae = F_promise->arg_end();
- ai != ae; ai++){
- Argument *Arg = &*ai;
- if(Arg->hasAttribute(Attribute::In))
- Arg->removeAttr(Attribute::In);
- if(Arg->hasAttribute(Attribute::Out))
- Arg->removeAttr(Attribute::Out);
- if(Arg->hasAttribute(Attribute::InOut))
- Arg->removeAttr(Attribute::InOut);
- }
-
- // Adding nounwind to generated function : FIXME: needed?
- DEBUG(errs() << "Adding nounwind to generated function\n");
- F_promise->addAttribute(AttributeSet::FunctionIndex, Attribute::NoUnwind);
-
- // Add llvm_visc_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_promise->getEntryBlock().begin());
-
- // FIXME: verify that we want 0 as a target device
- // In this backend, the target device is CPU, represented by i32 0.
- ConstantInt *TargetDeviceID =
- ConstantInt::get(Type::getInt32Ty(M.getContext()), 0);
-
- for (Function::arg_iterator ai = F_promise->arg_begin(),
- ae = F_promise->arg_end(); ai != ae; ++ai) {
- Argument* Arg = &*ai;
- if (Arg->getType()->isPointerTy()) {
- Value *Args[] = {Arg, TargetDeviceID};
- CallInst::Create(hpvm_request_tensor,
- ArrayRef<Value*>(Args, 2),
- "", FI);
- }
- }
-
- CodeGenStateMachine CGM(&M, runtimeModule.get(), qin);
-
- /* An assumption is made for the PROMISE simulator: *
- * a leaf node will contain consequtive operations that will map to a *
- * single PROMISE simulator call */
-
- for (inst_iterator i = inst_begin(F_promise), e = inst_end(F_promise);
- i != e; ++i) {
- Instruction *I = &(*i);
- CGM.transition(dyn_cast<IntrinsicInst>(I));
- }
-
- CGM.codeGen();
-
-//errs() << "-----------------------------------\n";
-//errs() << *F_promise << "\n";
-
- return;
-}
-
-bool DFG2LLVM_PROMISE::runOnModule(Module &M) {
- errs() << "\nDFG2LLVM_PROMISE PASS\n";
-
- errs() << QuantizationInputsFilename << "\n";
-
-// std::ifstream qin(quantizationInputsFilename_cstr);
-// std::ifstream qin;
-// qin.open(QuantizationInputsFilename.c_str());
-// qin.open(QuantizationInputsFilename.c_str(), std::ifstream::in);
-
- // Get the BuildDFG Analysis Results:
- // - Dataflow graph
- BuildDFG &DFG = getAnalysis<BuildDFG>();
-
- std::vector<DFInternalNode*> Roots = DFG.getRoots();
-
- // Visitor for Code Generation Graph Traversal
- CGT_PROMISE *CGTVisitor = new CGT_PROMISE(M, DFG, QuantizationInputsFilename);
-
- // Iterate over all the DFGs and produce code for each one of them
- for (auto rootNode: Roots) {
- // Initiate code generation for root DFNode
- CGTVisitor->visit(rootNode);
- }
-
- //TODO: Edit module epilogue to remove the VISC intrinsic declarations
- delete CGTVisitor;
-
- return true;
-}
-
-
-/******************************************************************************
- * Helper functions *
- ******************************************************************************/
-
-} // End of namespace
-
-char DFG2LLVM_PROMISE::ID = 0;
-static RegisterPass<DFG2LLVM_PROMISE> X("dfg2llvm-promise",
- "Dataflow Graph to LLVM for PROMISE Pass",
- false /* does not modify the CFG */,
- true /* transformation, *
- * not just analysis */);
-
diff --git a/llvm/lib/Transforms/DFG2LLVM_PROMISE/DFG2LLVM_PROMISE.exports b/llvm/lib/Transforms/DFG2LLVM_PROMISE/DFG2LLVM_PROMISE.exports
deleted file mode 100644
index e69de29bb2..0000000000
diff --git a/llvm/lib/Transforms/DFG2LLVM_PROMISE/LLVMBuild.txt b/llvm/lib/Transforms/DFG2LLVM_PROMISE/LLVMBuild.txt
deleted file mode 100644
index 714ad14f18..0000000000
--- a/llvm/lib/Transforms/DFG2LLVM_PROMISE/LLVMBuild.txt
+++ /dev/null
@@ -1,21 +0,0 @@
-;===- ./lib/Transforms/DFG2LLVM_NVPTX/LLVMBuild.txt ------------*- Conf -*--===;
-;
-; The LLVM Compiler Infrastructure
-;
-; This file is distributed under the University of Illinois Open Source
-; License. See LICENSE.TXT for details.
-;
-;===------------------------------------------------------------------------===;
-;
-; This is an LLVMBuild description file for the components in this subdirectory.
-;
-; For more information on the LLVMBuild system, please see:
-;
-; http://llvm.org/docs/LLVMBuild.html
-;
-;===------------------------------------------------------------------------===;
-
-[component_0]
-type = Library
-name = DFG2LLVM_PROMISE
-parent = Transforms
diff --git a/llvm/lib/Transforms/DFG2LLVM_WrapperAPI/CMakeLists.txt b/llvm/lib/Transforms/DFG2LLVM_WrapperAPI/CMakeLists.txt
deleted file mode 100644
index 22c219d0a1..0000000000
--- a/llvm/lib/Transforms/DFG2LLVM_WrapperAPI/CMakeLists.txt
+++ /dev/null
@@ -1,12 +0,0 @@
-if(WIN32 OR CYGWIN)
- set(LLVM_LINK_COMPONENTS Core Support)
-endif()
-
-add_llvm_loadable_module( LLVMDFG2LLVM_WrapperAPI
- DFG2LLVM_WrapperAPI.cpp
-
- DEPENDS
- intrinsics_gen
- PLUGIN_TOOL
- opt
- )
diff --git a/llvm/lib/Transforms/DFG2LLVM_WrapperAPI/DFG2LLVM_WrapperAPI.cpp b/llvm/lib/Transforms/DFG2LLVM_WrapperAPI/DFG2LLVM_WrapperAPI.cpp
deleted file mode 100644
index c54dd9ef3b..0000000000
--- a/llvm/lib/Transforms/DFG2LLVM_WrapperAPI/DFG2LLVM_WrapperAPI.cpp
+++ /dev/null
@@ -1,1530 +0,0 @@
-//=== DFG2LLVM_WrapperAPI.cpp ===//
-//
-// The LLVM Compiler Infrastructure
-//
-// This file is distributed under the University of Illinois Open Source
-// License. See LICENSE.TXT for details.
-//
-//===----------------------------------------------------------------------===//
-#define ENABLE_ASSERTS
-
-#define DEBUG_TYPE "DFG2LLVM_WrapperAPI"
-#include "llvm/IR/DataLayout.h"
-#include "llvm/IR/IRBuilder.h"
-#include "llvm/IR/Module.h"
-#include "llvm/Pass.h"
-#include "llvm/IR/InstIterator.h"
-#include "llvm/Transforms/Utils/ValueMapper.h"
-#include "llvm/Transforms/Utils/BasicBlockUtils.h"
-#include "llvm/Transforms/Utils/Cloning.h"
-#include "llvm/IRReader/IRReader.h"
-#include "llvm/Linker/Linker.h"
-#include "llvm/Support/SourceMgr.h"
-#include "llvm/Support/FileSystem.h"
-#include "llvm/IR/Attributes.h"
-#include "llvm-c/Core.h"
-#include "llvm/SupportVISC/VISCTimer.h"
-#include "llvm/SupportVISC/DFG2LLVM.h"
-#include "llvm/InPlaceDFG/InPlaceDFGAnalysis.h"
-#include <sstream>
-#include <fstream>
-
-using namespace llvm;
-using namespace builddfg;
-using namespace dfg2llvm;
-
-using namespace inplacedfg;
-
-namespace {
-
-cl::opt<std::string> QuantizationInputsFilename(
- "quantization-levels-filename",
- cl::desc("<PROMISE quantization levels input file (path)>"),
- cl::value_desc("filename"),
- cl::Required);
-
-cl::opt<std::string> ConfigurationInputsFilename(
- "configuration-inputs-filename",
- cl::desc("<Autotuner configurations input file (path)>"),
- cl::value_desc("filename"),
- cl::Required);
-
-// Helper function declarations
-bool isValidOperandForInPlaceOperation(Value *, Function *, DFNode *,
- InPlaceDFGAnalysis::InPlaceDFGParameter &);
-
-// Helper class declarations
-
-// State machine definition for pattern identification
-
-/* An assumption is made for the Wrapper API input: *
- * a leaf node will contain consequtive operations that will map to a *
- * single convolution or fully connected layer, or a single tensor operation. *
-
- * FullyConnectedLayer: Multiply, Add, [Activation] *
- * ConvolutionLayer: Convolution, [Add], [Activation], [Pooling] */
-
-class AbstractState;
-
-class CodeGenStateMachine {
-private:
- Module *M;
- Module *RtM;
-
- std::vector<Value*> Args;
- std::vector<IntrinsicInst*> IIs;
- AbstractState *current;
-
-public:
- CodeGenStateMachine(Module *, Module *);
-
- void setCurrent(AbstractState *s) {
- current = s;
- }
-
- void transition(IntrinsicInst *II);
-
- Module *getModule() {
- return M;
- }
-
- void addArgument(Value *Arg) {
- Args.push_back(Arg);
- }
-
- void addIntrinsicInst(IntrinsicInst *II) {
- IIs.push_back(II);
- }
-
- IntrinsicInst *getIntrinsicInstAt(unsigned idx) {
- return IIs[idx];
- }
-
- void codeGen(DFNode *, Function * , const StringRef &,
- InPlaceDFGAnalysis::InPlaceDFGParameter &);
-
-};
-
-class AbstractState {
-public:
- enum ID
- {
- INITIAL_STATE,
- FULLY_CONNECTED_LAYER_1,
- FULLY_CONNECTED_LAYER_2,
- FULLY_CONNECTED_LAYER_3,
- FULLY_CONNECTED_LAYER,
- CONVOLUTION_LAYER_1,
- CONVOLUTION_LAYER_2,
- CONVOLUTION_LAYER_3,
- CONVOLUTION_LAYER_4,
- CONVOLUTION_LAYER,
- SINGLE_TENSOR_OPERATION,
- NO_PATTERN,
- };
-
-protected:
- enum ID StateID;
-
-public:
- enum ID getStateID() {
- return StateID;
- }
-
- virtual void transition(CodeGenStateMachine *Mch, IntrinsicInst *II) = 0;
- virtual ~AbstractState() {}
-};
-
-class InitialState : public AbstractState {
-public:
- InitialState() {
- StateID = ID::INITIAL_STATE;
- DEBUG(errs() << "new InitialState\n");
- }
- ~InitialState() {}
-
- void transition(CodeGenStateMachine *Mch, IntrinsicInst *II) override;
-};
-
-class FullyConnectedLayer_1 : public AbstractState {
-public:
- FullyConnectedLayer_1() {
- StateID = ID::FULLY_CONNECTED_LAYER_1;
- DEBUG(errs() << "new FullyConnectedLayer_1\n");
- }
- ~FullyConnectedLayer_1() {}
-
- void transition(CodeGenStateMachine *Mch, IntrinsicInst *II) override;
-};
-
-class FullyConnectedLayer_2 : public AbstractState {
-public:
- FullyConnectedLayer_2() {
- StateID = ID::FULLY_CONNECTED_LAYER_2;
- DEBUG(errs() << "new FullyConnectedLayer_2\n");
- }
- ~FullyConnectedLayer_2() {}
-
- void transition(CodeGenStateMachine *Mch, IntrinsicInst *II) override;
-};
-
-class FullyConnectedLayer_3 : public AbstractState {
-public:
- FullyConnectedLayer_3() {
- StateID = ID::FULLY_CONNECTED_LAYER_3;
- DEBUG(errs() << "new FullyConnectedLayer_3\n");
- }
- ~FullyConnectedLayer_3() {}
-
- void transition(CodeGenStateMachine *Mch, IntrinsicInst *II) override;
-};
-
-class FullyConnectedLayer : public AbstractState {
-public:
- FullyConnectedLayer() {
- StateID = ID::FULLY_CONNECTED_LAYER;
- DEBUG(errs() << "new FullyConnectedLayer\n");
- }
- ~FullyConnectedLayer() {}
-
- void transition(CodeGenStateMachine *Mch, IntrinsicInst *II) override;
-};
-
-class ConvolutionLayer_1 : public AbstractState {
-public:
- ConvolutionLayer_1() {
- StateID = ID::CONVOLUTION_LAYER_1;
- DEBUG(errs() << "new ConvolutionLayer_1\n");
- }
- ~ConvolutionLayer_1() {}
-
- void transition(CodeGenStateMachine *Mch, IntrinsicInst *II) override;
-};
-
-class ConvolutionLayer_2 : public AbstractState {
-public:
- ConvolutionLayer_2() {
- StateID = ID::CONVOLUTION_LAYER_2;
- DEBUG(errs() << "new ConvolutionLayer_2\n");
- }
- ~ConvolutionLayer_2() {}
-
- void transition(CodeGenStateMachine *Mch, IntrinsicInst *II) override;
-};
-
-class ConvolutionLayer_3 : public AbstractState {
-public:
- ConvolutionLayer_3() {
- StateID = ID::CONVOLUTION_LAYER_3;
- DEBUG(errs() << "new ConvolutionLayer_3\n");
- }
- ~ConvolutionLayer_3() {}
-
- void transition(CodeGenStateMachine *Mch, IntrinsicInst *II) override;
-};
-
-class ConvolutionLayer_4 : public AbstractState {
-public:
- ConvolutionLayer_4() {
- StateID = ID::CONVOLUTION_LAYER_4;
- DEBUG(errs() << "new ConvolutionLayer_4\n");
- }
- ~ConvolutionLayer_4() {}
-
- void transition(CodeGenStateMachine *Mch, IntrinsicInst *II) override;
-};
-
-class ConvolutionLayer : public AbstractState {
-public:
- ConvolutionLayer() {
- StateID = ID::CONVOLUTION_LAYER;
- DEBUG(errs() << "new ConvolutionLayer\n");
- }
- ~ConvolutionLayer() {}
-
- void transition(CodeGenStateMachine *Mch, IntrinsicInst *II) override;
-};
-
-class SingleTensorOperation : public AbstractState {
-public:
- SingleTensorOperation() {
- StateID = ID::SINGLE_TENSOR_OPERATION;
- DEBUG(errs() << "new SingleTensorOperation\n");
- }
- ~SingleTensorOperation() {}
-
- void transition(CodeGenStateMachine *Mch, IntrinsicInst *II) override;
-};
-
-class NoPattern : public AbstractState {
-public:
- NoPattern() {
- StateID = ID::NO_PATTERN;
- DEBUG(errs() << "new NoPattern\n");
- }
- ~NoPattern() {}
-
- void transition(CodeGenStateMachine *Mch, IntrinsicInst *II) override;
-};
-
-void InitialState::transition(CodeGenStateMachine *Mch, IntrinsicInst *II) {
- if (II) { // Not end of instruction stream
- switch (II->getIntrinsicID()) {
- case Intrinsic::visc_tensor_convolution:
- {
- Mch->addIntrinsicInst(II);
- Mch->addArgument(II->getOperand(0)); // conv input
- Mch->addArgument(II->getOperand(1)); // conv kernel
-
- Mch->setCurrent(new ConvolutionLayer_1());
- }
- break;
- case Intrinsic::visc_tensor_mul:
- {
- Mch->addIntrinsicInst(II);
- Mch->addArgument(II->getOperand(0)); // 1st gemm input
- Mch->addArgument(II->getOperand(1)); // 2nd gemm input
-
- Mch->setCurrent(new FullyConnectedLayer_1());
- }
- break;
- default: // Other HPVM intrinsic
- {
- Mch->addIntrinsicInst(II);
- Mch->setCurrent(new SingleTensorOperation());
- }
- break;
- }
- delete this;
- } // else {} // No HPVM intrinsic received. Remain at initial
-}
-
-void SingleTensorOperation::transition(CodeGenStateMachine *Mch,
- IntrinsicInst *II) {
- if (II) { // Not end of instruction stream
- Mch->setCurrent(new NoPattern());
- delete this;
- }
-}
-
-void FullyConnectedLayer_1::transition(CodeGenStateMachine *Mch,
- IntrinsicInst *II) {
- if (II) { // Not end of instruction stream
- switch (II->getIntrinsicID()) {
- case Intrinsic::visc_tensor_add:
- {
- IntrinsicInst *MulII = Mch->getIntrinsicInstAt(0);
- assert((MulII == II->getOperand(0)) &&
- "Output of mul must be used as 1st operand of add");
- Mch->addIntrinsicInst(II);
-
- Mch->addArgument(II->getOperand(1)); // bias
-
- Mch->setCurrent(new FullyConnectedLayer_2());
- }
- break;
- default:
- Mch->setCurrent(new NoPattern());
- break;
- }
- } else {
- Mch->setCurrent(new NoPattern());
- }
- delete this;
-}
-
-void FullyConnectedLayer_2::transition(CodeGenStateMachine *Mch,
- IntrinsicInst *II) {
- if (II) { // Not end of instruction stream
- switch (II->getIntrinsicID()) {
- case Intrinsic::visc_tensor_tanh:
- {
- // Type of activation : TanH
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 0));
-
- Mch->addIntrinsicInst(II);
-
- Mch->setCurrent(new FullyConnectedLayer_3());
- }
- break;
- case Intrinsic::visc_tensor_relu:
- {
- // Type of activation : ReLU
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 1));
-
- Mch->addIntrinsicInst(II);
-
- Mch->setCurrent(new FullyConnectedLayer_3());
- }
- break;
- case Intrinsic::visc_tensor_clipped_relu:
- {
- // Type of activation : Clipped ReLU
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 2));
-
- Mch->addIntrinsicInst(II);
-
- Mch->setCurrent(new FullyConnectedLayer_3());
- }
- break;
- default: // No activation, but HPVM intrinsic
- Mch->setCurrent(new NoPattern());
- break;
- }
- } else { // End of instruction stream
- // No activation
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), -1));
-
- Mch->setCurrent(new FullyConnectedLayer());
- }
- delete this;
-}
-
-void FullyConnectedLayer_3::transition(CodeGenStateMachine *Mch,
- IntrinsicInst *II) {
- if (!II) { // End of instruction stream
- Mch->setCurrent(new FullyConnectedLayer());
- } else {
- Mch->setCurrent(new NoPattern());
- }
- delete this;
-}
-
-void FullyConnectedLayer::transition(CodeGenStateMachine *Mch,
- IntrinsicInst *II) {
- if (II) { // Not end of instruction stream
- Mch->setCurrent(new NoPattern());
- delete this;
- }
-}
-
-void ConvolutionLayer_1::transition(CodeGenStateMachine *Mch,
- IntrinsicInst *II) {
- if (II) { // Not end of instruction stream
- switch (II->getIntrinsicID()) {
- case Intrinsic::visc_tensor_add:
- {
- IntrinsicInst *ConvII = Mch->getIntrinsicInstAt(0);
- assert((ConvII == II->getOperand(0)) &&
- "Output of conv must be used as 1st operand of add");
- Mch->addIntrinsicInst(II);
-
- Mch->addArgument(II->getOperand(1)); // bias
-
- Mch->addArgument(ConvII->getOperand(2)); // 1st numeric arg of conv
- 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->setCurrent(new ConvolutionLayer_2());
- }
- break;
- default:
- Mch->setCurrent(new NoPattern());
- break;
- }
- } else {
- // No addition
- Mch->addArgument(ConstantPointerNull::get(
- Type::getInt8PtrTy(Mch->getModule()->getContext())));
-
- // Zero for all convolution numeric arguments FIXME???
- IntrinsicInst *ConvII = Mch->getIntrinsicInstAt(0);
- Mch->addArgument(ConvII->getOperand(2)); // 1st numeric arg of conv
- 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));
-
- // No pooling
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 0));
- // 0 for unused pool argument
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 0));
- // No activation
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), -1));
-
- Mch->setCurrent(new ConvolutionLayer());
- }
- delete this;
-}
-
-void ConvolutionLayer_2::transition(CodeGenStateMachine *Mch,
- IntrinsicInst *II) {
- if (II) { // Not end of instruction stream
- switch (II->getIntrinsicID()) {
- case Intrinsic::visc_tensor_tanh:
- {
- // Type of activation : TanH
-// Mch->addArgument(ConstantInt::get(
-// Type::getInt32Ty(Mch->getModule()->getContext()), 0));
- Mch->addIntrinsicInst(II);
-
- Mch->setCurrent(new ConvolutionLayer_3());
- }
- break;
- case Intrinsic::visc_tensor_relu:
- {
- // Type of activation : ReLU
-// Mch->addArgument(ConstantInt::get(
-// Type::getInt32Ty(Mch->getModule()->getContext()), 1));
- Mch->addIntrinsicInst(II);
-
- Mch->setCurrent(new ConvolutionLayer_3());
- }
- break;
- case Intrinsic::visc_tensor_clipped_relu:
- {
- // Type of activation : Clipped ReLU
-// Mch->addArgument(ConstantInt::get(
-// Type::getInt32Ty(Mch->getModule()->getContext()), 2));
- Mch->addIntrinsicInst(II);
-
- Mch->setCurrent(new ConvolutionLayer_3());
- }
- break;
- case Intrinsic::visc_tensor_pool_max:
- {
- // pool max
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 0));
- // poolSize
- Mch->addArgument(II->getOperand(1));
- // No activation
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), -1));
- Mch->addIntrinsicInst(II);
-
- Mch->setCurrent(new ConvolutionLayer_4());
- }
- break;
- case Intrinsic::visc_tensor_pool_min:
- {
- // pool min FIXME: 2: supported?
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 2));
- // poolSize
- Mch->addArgument(II->getOperand(1));
- // No activation
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), -1));
- Mch->addIntrinsicInst(II);
-
- Mch->setCurrent(new ConvolutionLayer_4());
- }
- break;
- case Intrinsic::visc_tensor_pool_mean:
- {
- // pool mean
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 1));
- // poolSize
- Mch->addArgument(II->getOperand(1));
- // No activation
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), -1));
- Mch->addIntrinsicInst(II);
-
- Mch->setCurrent(new ConvolutionLayer_4());
- }
- break;
- default: // No activation, No pooling, but HPVM intrinsic
- Mch->setCurrent(new NoPattern());
- break;
- }
- } else { // End of instruction stream
- // No pooling
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 0));
- // 0 for unused pool argument
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 0));
- // No activation
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), -1));
-
- Mch->setCurrent(new ConvolutionLayer());
- }
- delete this;
-}
-
-void ConvolutionLayer_3::transition(CodeGenStateMachine *Mch,
- IntrinsicInst *II) {
- if (II) { // Not end of instruction stream
- switch (II->getIntrinsicID()) {
- case Intrinsic::visc_tensor_pool_max:
- {
- // pool max
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 0));
- // poolSize
- Mch->addArgument(II->getOperand(1));
- Mch->addIntrinsicInst(II);
-
- // Revisit last intrinsic, to add argument for activation operation
- 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) {
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 0));
- } else if (ActIID == Intrinsic::visc_tensor_relu) {
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 1));
- } else { //ActIID == Intrinsic::visc_tensor_clipped_relu
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 2));
- }
-
- Mch->setCurrent(new ConvolutionLayer_4());
- }
- break;
- case Intrinsic::visc_tensor_pool_min:
- {
- // pool min FIXME: 2: supported?
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 2));
- // poolSize
- Mch->addArgument(II->getOperand(1));
- Mch->addIntrinsicInst(II);
-
- // Revisit last intrinsic, to add argument for activation operation
- 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) {
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 0));
- } else if (ActIID == Intrinsic::visc_tensor_relu) {
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 1));
- } else { //ActIID == Intrinsic::visc_tensor_clipped_relu
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 2));
- }
-
- Mch->setCurrent(new ConvolutionLayer_4());
- }
- break;
- case Intrinsic::visc_tensor_pool_mean:
- {
- // pool max
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 1));
- // poolSize
- Mch->addArgument(II->getOperand(1));
- Mch->addIntrinsicInst(II);
-
- // Revisit last intrinsic, to add argument for activation operation
- 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) {
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 0));
- } else if (ActIID == Intrinsic::visc_tensor_relu) {
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 1));
- } else { //ActIID == Intrinsic::visc_tensor_clipped_relu
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 2));
- }
-
- Mch->setCurrent(new ConvolutionLayer_4());
- }
- break;
- default: // No pooling, but HPVM intrinsic
- Mch->setCurrent(new NoPattern());
- break;
- }
- } else { // End of instruction stream
- // No pooling
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 0));
- // 0 for unused pool argument
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 0));
-
- // Revisit last intrinsic, to add argument for activation operation
- 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) {
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 0));
- } else if (ActIID == Intrinsic::visc_tensor_relu) {
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 1));
- } else { //ActIID == Intrinsic::visc_tensor_clipped_relu
- Mch->addArgument(ConstantInt::get(
- Type::getInt32Ty(Mch->getModule()->getContext()), 2));
- }
-
- Mch->setCurrent(new ConvolutionLayer());
- }
- delete this;
-}
-
-void ConvolutionLayer_4::transition(CodeGenStateMachine *Mch,
- IntrinsicInst *II) {
- if (!II) { // End of instruction stream
- Mch->setCurrent(new ConvolutionLayer());
- } else {
- Mch->setCurrent(new NoPattern());
- }
- delete this;
-}
-
-void ConvolutionLayer::transition(CodeGenStateMachine *Mch,
- IntrinsicInst *II) {
- if (II) { // Not end of instruction stream
- Mch->setCurrent(new NoPattern());
- delete this;
- }
-}
-
-void NoPattern::transition(CodeGenStateMachine *Mch, IntrinsicInst *II) {}
-
-CodeGenStateMachine::CodeGenStateMachine(Module *_M, Module *_RtM) :
- M(_M), RtM(_RtM) {
- current = new InitialState();
-}
-
-void CodeGenStateMachine::transition(IntrinsicInst *II) {
- current->transition(this, II);
-}
-
-void CodeGenStateMachine::codeGen(DFNode *N, Function *F, const StringRef &strRef,
- InPlaceDFGAnalysis::InPlaceDFGParameter &IPP) {
-
- 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" );
-
- 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.
- // Make sure that the instruction sequence can be traslated:
- // each instruction's result must be used only by the next one in sequence.
-
- for (unsigned p = 0; p < IIs.size()-1; p++) {
- IntrinsicInst *II = IIs[p];
- assert((II->hasOneUse()) &&
- "Instruction sequence does not fit pattern: not single use\n");
-
- Value::user_iterator ui = II->user_begin(); // The only use
- assert((*ui == IIs[p+1]) &&
- "Instruction sequence does not fit pattern: not used by next instruction\n");
- }
-
- // Create corresponding wrapper API call
- CallInst *CI;
- switch (current->getStateID()) {
- case AbstractState::ID::CONVOLUTION_LAYER:
- {
- Constant* wrapper_ConvLayer =
- M->getOrInsertFunction(StringRef("wrapper_ConvLayer"),
- RtM->getFunction(StringRef("wrapper_ConvLayer"))->getFunctionType());
- DEBUG(errs() << *wrapper_ConvLayer);
-
- // FIXME: get last (float) arguments from clipped relu intrinsic. For now, 0
- Args.push_back(ConstantFP::get(Type::getFloatTy(M->getContext()), (double) 0));
- Args.push_back(ConstantFP::get(Type::getFloatTy(M->getContext()), (double) 0));
-
-
- // 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);
-
- std::vector<Value*> UpdatedArgs;
- UpdatedArgs.push_back(GEPConst);
- for (unsigned i = 0; i < Args.size(); i++) {
- UpdatedArgs.push_back(Args[i]);
- }
- // Create wrapper API function call
- CI = CallInst::Create(wrapper_ConvLayer, UpdatedArgs, "");
- }
- break;
- case AbstractState::ID::FULLY_CONNECTED_LAYER:
- {
- Constant* 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));
- Args.push_back(ConstantFP::get(Type::getFloatTy(M->getContext()), (double) 0));
-
- // 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);
-
- std::vector<Value*> UpdatedArgs;
- UpdatedArgs.push_back(GEPConst);
- for (unsigned i = 0; i < Args.size(); i++) {
- UpdatedArgs.push_back(Args[i]);
- }
-
- // Create wrapper API function call
- CI = CallInst::Create(wrapper_FCLayer, UpdatedArgs, "");
- }
- break;
- default:
- llvm_unreachable("Unexpected CodeGenStateMachine State\n");
- break;
- }
-
- // Insert new call and replace all uses of pattern result with
- // the wrapper API call
- IntrinsicInst *IIlast = *(IIs.rbegin());
- CI->insertBefore(IIlast);
- IIlast->replaceAllUsesWith(CI);
-
- } 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";
-
- switch (TensorII->getIntrinsicID()) {
- case Intrinsic::visc_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");
-
- // 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);
- 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);
-
- 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));
-
- // Create wrapper API runtime function call
- 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.mul with the runtime call
- TensorII->replaceAllUsesWith(CI);
- }
- break;
-
- case Intrinsic::visc_tensor_batchnorm:
- { /* llvm.hpvm.tensor.batchnorm */
- // Tensor batchnorm is not in place.
- // FIXME: Add Check for InPlace Analysis
- DEBUG(errs() << F->getName() << "\t: Handling tensor batch normalization \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);
- 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);
-
- 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));
-
- // Create wrapper API runtime function call
- Constant* wrapper_tensorBatchNorm =
- M->getOrInsertFunction(StringRef("wrapper_tensorBatchNorm"),
- RtM->getFunction(StringRef("wrapper_tensorBatchNorm"))->getFunctionType());
- CallInst* CI = CallInst::Create(wrapper_tensorBatchNorm,
- Args, "", TensorII);
- // We can replace the call to hpvm.tensor.batchnorm with the wrapper API call
- TensorII->replaceAllUsesWith(CI);
- }
- break;
-
- 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);
-
- // 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)
- // FIXME: remove this comment - must check for in-place
-// 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);
- 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);
-
- Args.push_back(TensorII->getOperand(0));
- Args.push_back(TensorII->getOperand(1));
-
- // Create wrapper API runtime function call
- Constant* wrapper_tensorAdd =
- M->getOrInsertFunction(StringRef("wrapper_tensorAdd"),
- RtM->getFunction(StringRef("wrapper_tensorAdd"))->getFunctionType());
- CallInst::Create(wrapper_tensorAdd, Args, "", TensorII);
- // We can replace the call to hpvm.tensor.add with the 1st argument
- // that, due to in place operation, now contains the result
- TensorII->replaceAllUsesWith(TensorII->getOperand(0));
- }
- break;
-
- case Intrinsic::visc_tensor_pool_max:
- case Intrinsic::visc_tensor_pool_mean:
- case Intrinsic::visc_tensor_pool_min:
- {
- DEBUG(errs() << F->getName() << "\t: Handling tensor pooling functions\n");
-
- // Argument list for tensor pooling:
- // input, poolFunction, window_height, window_width,
- // vertical_pad, horizontal_pad, vertical_stride, horizontal_stride
-
- // 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);
-
- Args.push_back(TensorII->getOperand(0));
-
- int pool_type = 0;
- if (TensorII->getIntrinsicID() == Intrinsic::visc_tensor_pool_max) {
- pool_type = 0;
- }
- if (TensorII->getIntrinsicID() == Intrinsic::visc_tensor_pool_mean) {
- pool_type = 1;
- }
- if (TensorII->getIntrinsicID() == Intrinsic::visc_tensor_pool_min) {
- pool_type = 2;
- }
-
- Constant *constPoolType =
- ConstantInt::get(Type::getInt32Ty(M->getContext()), pool_type);
- Args.push_back(constPoolType);
-
- 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));
- Args.push_back(TensorII->getOperand(6));
-
- // Create wrapper API runtime function call
- Constant* 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
- TensorII->replaceAllUsesWith(CI);
- }
- break;
-
- case Intrinsic::visc_tensor_relu:
- case Intrinsic::visc_tensor_clipped_relu:
- case Intrinsic::visc_tensor_tanh:
- {
- DEBUG(errs() << F->getName() << "\t: Handling tensor activation functions\n");
-
- // Tensor relu(a) (and others) 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);
- 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);
-
- Args.push_back(TensorII->getOperand(0));
-
- if (TensorII->getIntrinsicID() == Intrinsic::visc_tensor_relu) {
- // Create wrapper API runtime function call
- Constant* 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) {
- // Create wrapper API runtime function call
- Constant* 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) {
- // Create wrapper API runtime function call
- Constant* wrapper_tensorTanh =
- M->getOrInsertFunction(StringRef("wrapper_tensorTanh"),
- RtM->getFunction(StringRef("wrapper_tensorTanh"))->getFunctionType());
- DEBUG(errs() << *wrapper_tensorTanh);
- CallInst::Create(wrapper_tensorTanh, Args, "", TensorII);
- }
-
- // We can replace the call to hpvm.tensor.{relu,clipped relu, tanh}
- // with the 1st argument that, due to in place operation,
- // now contains the result
- TensorII->replaceAllUsesWith(TensorII->getOperand(0));
- }
- break;
-
- case Intrinsic::visc_tensor_softmax:
- { /* llvm.visc.tensor.softmax */
-
- DEBUG(errs() << F->getName() << "\t: Handling tensor softmax\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);
- 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);
-
- Args.push_back(TensorII->getOperand(0));
-
- // Create wrapper API runtime function call
- Constant* 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
- 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;
- }
-
- } // No other case exists, since assertion passed
-
-
- // Remove the instructions we translated to the simulator call.
- // Traverse the vector backwards, otherwise definitions are deleted while
- // their subsequent uses are still around.
- for (std::vector<IntrinsicInst *>::reverse_iterator ri = IIs.rbegin(),
- re = IIs.rend(); ri != re; ++ri) {
- DEBUG(errs() << "Erasing: " << **ri << "\n");
- (*ri)->eraseFromParent();
- }
-
-}
-
-// DFG2LLVM_WrapperAPI - The first implementation.
-
-struct DFG2LLVM_WrapperAPI : public DFG2LLVM {
- static char ID; // Pass identification, replacement for typeid
- DFG2LLVM_WrapperAPI() : DFG2LLVM(ID) {}
-private:
-
-public:
-
- void getAnalysisUsage(AnalysisUsage &AU) const {
- AU.addRequired<BuildDFG>();
- AU.addRequired<InPlaceDFGAnalysisWrapper>();
- AU.addPreserved<BuildDFG>();
- AU.addPreserved<InPlaceDFGAnalysisWrapper>();
- }
-
- bool runOnModule(Module &M);
-};
-
-// Visitor for Code generation traversal (tree traversal for now)
-class CGT_WrapperAPI : public CodeGenTraversal {
-
-private:
- //Member variables
- unsigned nodeID; // Used as a node identifier
-
- std::string QuantizationInputsFilenameStr;
- std::string ConfigurationInputsFilenameStr;
-
- InPlaceDFGAnalysis::InPlaceDFGParameter *IPP;
-
- // VISC Runtime API and Tensor runtime API
- Constant* llvm_hpvm_initApproxhpvmRt;
- Constant* llvm_hpvm_cleanupApproxhpvmRt;
- Constant* hpvm_request_tensor;
-
- Constant* llvm_hpvm_initializeRuntimeController;
- Constant* llvm_hpvm_clearRuntimeController;
-
- // Functions
-
- // Virtual Functions
- void init();
- void initRuntimeAPI();
- void codeGen(DFInternalNode* N);
- void codeGen(DFLeafNode* N);
-
-public:
-
- // Constructor
- CGT_WrapperAPI(Module &_M, BuildDFG &_DFG,
- InPlaceDFGAnalysis::InPlaceDFGParameter &_IPP,
- std::string &_QuantizationInputsFilenameStr,
- std::string &_ConfigurationInputsFilenameStr)
- : CodeGenTraversal(_M, _DFG), IPP(&_IPP),
- QuantizationInputsFilenameStr(_QuantizationInputsFilenameStr),
- ConfigurationInputsFilenameStr(_ConfigurationInputsFilenameStr) {
- nodeID = 0;
- initRuntimeAPI();
- }
-
-};
-
-
-void CGT_WrapperAPI::init() {
- // FIXME: what to do here? If anything?
-}
-
-// Initialize the VISC runtime API. This makes it easier to insert these calls
-void CGT_WrapperAPI::initRuntimeAPI() {
-
- // 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!\n");
-
- // FIXME: set correct path
- Twine llvmSrcRoot = LLVM_SRC_ROOT;
- Twine runtimeAPI = llvmSrcRoot+"/projects/hpvm-tensor-rt/lib/tensor_runtime.ll";
- runtimeModule = parseIRFile(runtimeAPI.str(), Err, M.getContext());
- if(runtimeModule == nullptr)
- DEBUG(errs() << Err.getMessage());
- else
- DEBUG(errs() << "Successfully loaded hpvm-tensor-rt API module\n");
-
- // Get or insert Global declarations for
- // - initialization
- // - cleanup
- // - request a tensor
- DECLARE(llvm_hpvm_initApproxhpvmRt);
- DECLARE(llvm_hpvm_cleanupApproxhpvmRt);
- DECLARE(hpvm_request_tensor);
-
- DECLARE(llvm_hpvm_initializeRuntimeController);
- DECLARE(llvm_hpvm_clearRuntimeController);
-
- // Find visc.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");
- InitCall = cast<Instruction>(*VI->user_begin());
- CallInst::Create(llvm_hpvm_initApproxhpvmRt,
- ArrayRef<Value*>(ConstantInt::get(Type::getInt32Ty(M.getContext()), 0)),
- "", InitCall);
-
- StringRef QRangesStrRef = StringRef(QuantizationInputsFilenameStr);
- // Create string for node name, as first argument for wrapper API call
- Constant *ConstArray1 = ConstantDataArray::getString(M.getContext(),
- QRangesStrRef, true);
- GlobalVariable *GV1 = new GlobalVariable(M,ConstArray1->getType(),
- true, GlobalValue::ExternalLinkage, ConstArray1, "");
- // Create GEP expression to access it
- Constant* Int_0 = ConstantInt::get(Type::getInt32Ty(M.getContext()), 0);
- Constant* GEPIndices[] = { Int_0, Int_0 };
- Constant* QRangesGEPConst =
- ConstantExpr::getGetElementPtr(GV1->getType()->getPointerElementType(),
- GV1, GEPIndices);
-
- StringRef ConfsStrRef = StringRef(ConfigurationInputsFilenameStr);
- // Create string for node name, as first argument for wrapper API call
- Constant *ConstArray2 = ConstantDataArray::getString(M.getContext(),
- ConfsStrRef, true);
- GlobalVariable *GV2 = new GlobalVariable(M,ConstArray2->getType(),
- true, GlobalValue::ExternalLinkage, ConstArray2, "");
- Constant* ConfsGEPConst =
- ConstantExpr::getGetElementPtr(GV2->getType()->getPointerElementType(),
- GV2, GEPIndices);
- 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");
- CleanupCall = cast<Instruction>(*VC->user_begin());
- CallInst::Create(llvm_hpvm_cleanupApproxhpvmRt, ArrayRef<Value*>(), "", CleanupCall);
- CallInst::Create(llvm_hpvm_clearRuntimeController, ArrayRef<Value*>(), "", CleanupCall);
-
-}
-
-void CGT_WrapperAPI::codeGen(DFInternalNode* N) {
- errs () << "Inside node: " << N->getFuncPointer()->getName() << "\n";
- errs () << "Skipping internal node\n";
-}
-
-void CGT_WrapperAPI::codeGen(DFLeafNode* N) {
-
- // Skip code generation if it is a dummy node
- if(N->isDummyNode()) {
- DEBUG(errs() << "Skipping dummy node\n");
- return;
- }
-
- // Abort code generation if it is an allocation node
- if(N->isAllocationNode()) {
- assert(false && "Allocation Node not expected in ApproxHPVM");
- 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";
- // 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);
-
- assert((F_wrapper_api == NULL) &&
- "Error: Visiting a node for which code already generated");
-
- // 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();
- M.getFunctionList().push_back(F_wrapper_api);
-
- N->addGenFunc(F_wrapper_api, visc::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();
- ai != ae; ai++){
- Argument *Arg = &*ai;
- if(Arg->hasAttribute(Attribute::In))
- Arg->removeAttr(Attribute::In);
- if(Arg->hasAttribute(Attribute::Out))
- Arg->removeAttr(Attribute::Out);
- if(Arg->hasAttribute(Attribute::InOut))
- Arg->removeAttr(Attribute::InOut);
- }
-
- // Adding nounwind to generated function : FIXME: needed?
- DEBUG(errs() << "Adding nounwind to generated function\n");
- F_wrapper_api->addAttribute(AttributeSet::FunctionIndex, Attribute::NoUnwind);
-
- // Add llvm_visc_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());
-
- // FIXME: verify that we want 1 as a target device
- // In this backend, the target device is GPU, represented by i32 1.
- ConstantInt *TargetDeviceID =
- ConstantInt::get(Type::getInt32Ty(M.getContext()), 1);
-
- for (Function::arg_iterator ai = F_wrapper_api->arg_begin(),
- ae = F_wrapper_api->arg_end(); ai != ae; ++ai) {
- Argument* Arg = &*ai;
- if (Arg->getType()->isPointerTy()) {
- Value *Args[] = {Arg, TargetDeviceID};
- CallInst::Create(hpvm_request_tensor,
- ArrayRef<Value*>(Args, 2),
- "", FI);
- }
- }
-
- CodeGenStateMachine CGM(&M, runtimeModule.get());
-
- for (inst_iterator i = inst_begin(F_wrapper_api), e = inst_end(F_wrapper_api);
- i != e; ++i) {
- Instruction *I = &(*i);
- CGM.transition(dyn_cast<IntrinsicInst>(I));
- }
-
- errs() << "Node ID string: "<< StringRef(std::to_string(nodeID)) << "\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";
-
- // Get the BuildDFG Analysis Results:
- // - Dataflow graph
- BuildDFG &DFG = getAnalysis<BuildDFG>();
-
- // 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
- CGT_WrapperAPI *CGTVisitor = new CGT_WrapperAPI(M, DFG, IPP,
- QuantizationInputsFilename,
- ConfigurationInputsFilename);
-
- // Iterate over all the DFGs and produce code for each one of them
- for (auto rootNode: Roots) {
- // Initiate code generation for root DFNode
- CGTVisitor->visit(rootNode);
- }
-
- //TODO: Edit module epilogue to remove the VISC intrinsic declarations
- delete CGTVisitor;
-
- return true;
-}
-
-
-/******************************************************************************
- * Helper functions *
- ******************************************************************************/
-
-/* Method needs to be called as part of an analysis pre-step, before code *
- * generation is run on a node function, so that the HPVM intrinsics are still *
- * in place. */
-bool isValidOperandForInPlaceOperation(Value *Op, Function *Fgen, DFNode *N,
- InPlaceDFGAnalysis::InPlaceDFGParameter &IPP) {
-
- if (Argument *Arg = dyn_cast<Argument>(Op)) {
- DEBUG(errs() << *Arg << "\t: argument, candidate for in place\n");
- assert((Arg->getParent() == Fgen) &&
- "Extra Parameter in body of Function\n");
- // Candidate parameter is a function argument
- // In this case, consult the result of in place analysis
- // Find position in arg list
- unsigned pos = Arg->getArgNo();
- // If this parameter cannot be used for in place operation
- // code gen cannot continue
- if (IPP.at(N)[pos]) {
- DEBUG(errs() << *Arg << "\t: argument, suitable for in place\n");
- return true;
- } else {
- DEBUG(errs() << *Arg << "\t: argument, not suitable for in place\n");
- return false;
- }
- }
- else {
- // If it is not an argument, then it needs to be the result of
- // another intrinsic. These are new objects that are allocated,
- // and consumed by next intrinsic.
- DEBUG(errs() << *Op << "\t: Test for result of intrinsic operation\n");
- if (dyn_cast<IntrinsicInst>(Op)) {
- DEBUG(errs() << *Arg << "\t: local, suitable for in place\n");
- return true;
- } else {
- DEBUG(errs() << *Arg << "\t: local, not suitable for in place\n");
- return false;
- }
- }
-}
-
-} // End of namespace
-
-char DFG2LLVM_WrapperAPI::ID = 0;
-static RegisterPass<DFG2LLVM_WrapperAPI> X("dfg2llvm-wrapperapi",
- "Dataflow Graph to LLVM for WrapperAPI Pass",
- false /* does not modify the CFG */,
- true /* transformation, *
- * not just analysis */);
-
diff --git a/llvm/lib/Transforms/DFG2LLVM_WrapperAPI/DFG2LLVM_WrapperAPI.exports b/llvm/lib/Transforms/DFG2LLVM_WrapperAPI/DFG2LLVM_WrapperAPI.exports
deleted file mode 100644
index e69de29bb2..0000000000
diff --git a/llvm/lib/Transforms/DFG2LLVM_WrapperAPI/LLVMBuild.txt b/llvm/lib/Transforms/DFG2LLVM_WrapperAPI/LLVMBuild.txt
deleted file mode 100644
index b4ebb8019d..0000000000
--- a/llvm/lib/Transforms/DFG2LLVM_WrapperAPI/LLVMBuild.txt
+++ /dev/null
@@ -1,21 +0,0 @@
-;===- ./lib/Transforms/DFG2LLVM_WrapperAPI/LLVMBuild.txt -------*- Conf -*--===;
-;
-; The LLVM Compiler Infrastructure
-;
-; This file is distributed under the University of Illinois Open Source
-; License. See LICENSE.TXT for details.
-;
-;===------------------------------------------------------------------------===;
-;
-; This is an LLVMBuild description file for the components in this subdirectory.
-;
-; For more information on the LLVMBuild system, please see:
-;
-; http://llvm.org/docs/LLVMBuild.html
-;
-;===------------------------------------------------------------------------===;
-
-[component_0]
-type = Library
-name = DFG2LLVM_WrapperAPI
-parent = Transforms
diff --git a/llvm/lib/Transforms/FuseHPVMTensorNodes/CMakeLists.txt b/llvm/lib/Transforms/FuseHPVMTensorNodes/CMakeLists.txt
deleted file mode 100644
index 374f3b26f1..0000000000
--- a/llvm/lib/Transforms/FuseHPVMTensorNodes/CMakeLists.txt
+++ /dev/null
@@ -1,12 +0,0 @@
-if(WIN32 OR CYGWIN)
- set(LLVM_LINK_COMPONENTS Core Support)
-endif()
-
-add_llvm_loadable_module( LLVMFuseHPVMTensorNodes
- FuseHPVMTensorNodes.cpp
-
- DEPENDS
- intrinsics_gen
- PLUGIN_TOOL
- opt
- )
diff --git a/llvm/lib/Transforms/FuseHPVMTensorNodes/FuseHPVMTensorNodes.cpp b/llvm/lib/Transforms/FuseHPVMTensorNodes/FuseHPVMTensorNodes.cpp
deleted file mode 100644
index d9a3c588b5..0000000000
--- a/llvm/lib/Transforms/FuseHPVMTensorNodes/FuseHPVMTensorNodes.cpp
+++ /dev/null
@@ -1,971 +0,0 @@
-//=== FuseHPVMTensorNodes.cpp ===//
-//
-// The LLVM Compiler Infrastructure
-//
-// This file is distributed under the University of Illinois Open Source
-// License. See LICENSE.TXT for details.
-//
-//===----------------------------------------------------------------------===//
-
-#define DEBUG_TYPE "FuseTensorNodes"
-
-#include "llvm/IR/ValueMap.h"
-#include "llvm/Transforms/Utils/BasicBlockUtils.h"
-
-#include "llvm/FuseHPVMTensorNodes/FuseHPVMTensorNodes.h"
-#include "llvm/SupportVISC/DFG2LLVM.h"
-#include "llvm/SupportVISC/VISCUtils.h"
-
-using namespace llvm;
-using namespace builddfg;
-using namespace dfg2llvm;
-using namespace viscUtils;
-
-namespace tensorfuse {
-/*** Classes ***/
-
-/*** Helper Functions ***/
-
-/* Return the constant integer represented by value V */
-static unsigned getNumericValue(Value* V) {
- assert(isa<ConstantInt>(V)
- && "Value indicating the number of arguments should be a constant integer");
- return cast<ConstantInt>(V)->getZExtValue();
-}
-
-/* Query the kind of edge described by a createEdge intrinsic IIe *
- * with respect to node handle IIn */
-static bool isIncomingEdgeIntrinsic(IntrinsicInst* IIe, IntrinsicInst* IIn) {
- Value* Src = IIe->getArgOperand(1);
- IntrinsicInst* ArgII = cast<IntrinsicInst>(Src);
- assert(ArgII && "First argument of createEdge is not an intrinsic");
- return (ArgII == IIn);
-}
-static bool isOutgoingEdgeIntrinsic(IntrinsicInst* IIe, IntrinsicInst* IIn) {
- Value* Src = IIe->getArgOperand(0);
- IntrinsicInst* ArgII = cast<IntrinsicInst>(Src);
- assert(ArgII && "First argument of createEdge is not an intrinsic");
- return (ArgII == IIn);
-}
-
-/* Populates vector with all incoming edge intrinsics to node II */
-static void getIncomingEdgeIntrinsicList(IntrinsicInst *II,
- std::vector<IntrinsicInst*> &EdgeList) {
- for(Value::user_iterator ui = II->user_begin(),
- ue = II->user_end(); ui!=ue; ++ui) {
- IntrinsicInst* useI = dyn_cast<IntrinsicInst>(*ui);
- assert(useI &&
- "HPVM graph intrinsic used in non HPVM intrinsic instruction\n");
- if (useI->getIntrinsicID() != Intrinsic::visc_createEdge)
- continue; // Skip all non edge intrinsics
-
- // For edge intrinsics, test the descination operand
- if (useI->getOperand(1) == II) { // Argument is the destination
- EdgeList.push_back(useI);
- }
- }
- return;
-}
-
-/* Returns true if argument at position argno is coming from a dataflow edge *
- * in the vector EdgeList */
-static bool isIncomingEdgeArgument(unsigned argno,
- std::vector<IntrinsicInst*> &EdgeList) {
- for (IntrinsicInst *ii : EdgeList) {
- if (getNumericValue(ii->getOperand(4)) == argno)
- return true;
- }
- 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");
- 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,
-// and DstN a single predeccessor, SrcN (other than the Root node)
-static DFNode *findNextNodeInSequence(DFNode *SrcN) {
-
- DFNode *DstN = NULL;
-
- for (DFNode::successor_iterator si = SrcN->successors_begin(),
- se = SrcN->successors_end(); si != se; ++si) {
- DFNode *N = *si;
- if (N->isDummyNode()) {
- continue;
- }
- if (!DstN)
- DstN = N;
- if (DstN != N) {
- errs() << "Found different destination nodes: no node sequence.\n";
- return NULL;
- }
- }
-
- // If we reach this point, DstN is the unique successor of SrcN
-
- // Now, test that the DstN has a single predeccessor except Root (dummy)
- for (DFNode::indfedge_iterator eb = DstN->indfedge_begin(),
- ee = DstN->indfedge_end(); eb != ee; ++eb) {
- DFNode *SN = (*eb)->getSourceDF();
- if ((SN != SrcN) && (!(SN->isDummyNode()))) {
- // Does not satisfy requirement
- return NULL;
- }
- }
-
- return DstN;
-}
-
-/*** Methods ***/
-
-/* Create an identical bind (in or out, depending on the argument intrinsic) *
- * with different src (true) or dst (false) port */
-IntrinsicInst* FuseHPVMTensorNodes::createIdenticalBindWithDifferentPort(
- IntrinsicInst* II, unsigned port, bool srcport) {
- // Argument of the function to be called
- ConstantInt* PortConstant =
- ConstantInt::get(Type::getInt32Ty(II->getContext()), port);
- Value* SrcPort = (srcport) ? PortConstant: II->getArgOperand(1);
- Value* DstPort = (srcport) ? II->getArgOperand(2): PortConstant;
-
- Value* BindArgs[] = {II->getArgOperand(0),
- SrcPort,
- DstPort,
- II->getArgOperand(3)
- };
- Function* BindF = II->getCalledFunction();
- CallInst* BindInst = CallInst::Create(BindF,
- ArrayRef<Value*>(BindArgs, 4),
- "");
- IntrinsicInst* newII = dyn_cast<IntrinsicInst>(BindInst);
-
- return newII;
-}
-
-/* Given two createNode intrinsics describing connected nodes, this function *
- * returns the argument list type of the fused function */
-void FuseHPVMTensorNodes::createArgTypes(IntrinsicInst* II1,
- IntrinsicInst* II2,
- std::vector<Type*> &ArgTypes) {
- Function* F1 = cast<Function>((II1->getOperand(0))->stripPointerCasts());
- Function* F2 = cast<Function>((II2->getOperand(0))->stripPointerCasts());
-
- // Arguments of the first node are simply added
- for(auto& arg: F1->getArgumentList()) {
- DEBUG(errs() << arg << "\n");
- ArgTypes.push_back(arg.getType());
- }
-
- // Arguments of the second node are added only if they are not the output of
- // the previous node
-
- // Find all incoming edges.
- std::vector<IntrinsicInst *> IncomingEdgeList;
- getIncomingEdgeIntrinsicList(II2, IncomingEdgeList);
-
- // Their source must be the first fusion node, otherwise they would not have
- // been fusion candidates
- for (IntrinsicInst *ii : IncomingEdgeList) {
- assert((ii->getOperand(0) == II1) && "Unexpected source operand\n");
- }
-
- // Add argument type to the new function only if it is not incoming from
- // an edge
- for(auto& arg: F2->getArgumentList()) {
- DEBUG(errs() << arg << "\n");
- unsigned inport = arg.getArgNo();
- if (isIncomingEdgeArgument(inport, IncomingEdgeList))
- continue;
- ArgTypes.push_back(arg.getType());
- }
-}
-
-/* Get the return type of the function for fused node II1-II2 */
-StructType* FuseHPVMTensorNodes::createReturnType(IntrinsicInst* II1,
- IntrinsicInst* II2) {
- Function* F1 = cast<Function>((II1->getOperand(0))->stripPointerCasts());
- Function* F2 = cast<Function>((II2->getOperand(0))->stripPointerCasts());
-
- // Based on the HPVM tensor node assumptions and the patterns we want to
- // support, when two nodes are fused the result will always be the result
- // of the second node.
- StructType* F1RetTy = dyn_cast<StructType>(F1->getReturnType());
- assert(F1RetTy && "Return Type must always be a struct");
- StructType* F2RetTy = dyn_cast<StructType>(F2->getReturnType());
- assert(F2RetTy && "Return Type must always be a struct");
-
- return F2RetTy;
-}
-
-/* Copy argument names, from functions of II1 and II2 to F */
-void FuseHPVMTensorNodes::copyArgumentNames(IntrinsicInst* II1,
- IntrinsicInst* II2,
- Function* F) {
- Function* F1 = cast<Function>((II1->getOperand(0))->stripPointerCasts());
- Function* F2 = cast<Function>((II2->getOperand(0))->stripPointerCasts());
-
- Function::arg_iterator dest_it = F->arg_begin();
-
- // Argument names of the first node are simply copied
- for(auto& arg: F1->getArgumentList()) {
- dest_it->setName("s_" + arg.getName());
- dest_it++;
- }
-
- // For the second node, we ignore those arguments that are incoming edges
- // (from II1)
- // Find all incoming edges.
- std::vector<IntrinsicInst *> IncomingEdgeList;
- getIncomingEdgeIntrinsicList(II2, IncomingEdgeList);
-
- // Their source must be the first fusion node, otherwise they would not have
- // been fusion candidates
- for (IntrinsicInst *ii : IncomingEdgeList) {
- assert((ii->getOperand(0) == II1) && "Unexpected source operand\n");
- }
-
- // Copy argument name to the new function only if it is not incoming from
- // an edge
- for(auto& arg: F2->getArgumentList()) {
- DEBUG(errs() << arg << "\n");
- unsigned inport = arg.getArgNo();
- if (isIncomingEdgeArgument(inport, IncomingEdgeList))
- continue;
-
- dest_it->setName("d_" + arg.getName());
- dest_it++;
- }
- assert((dest_it == F->arg_end()) &&
- "Argument list of fused function not fully traversed\n");
- return;
-}
-
-/* Copy attributes, from functions of II1 and II2 to F */
-void FuseHPVMTensorNodes::copyAttrList(IntrinsicInst* II1,
- IntrinsicInst* II2,
- Function* F) {
- Function* F1 = cast<Function>((II1->getOperand(0))->stripPointerCasts());
- Function* F2 = cast<Function>((II2->getOperand(0))->stripPointerCasts());
-
- Function::arg_iterator f1_ai = F1->arg_begin(), f1_ae = F1->arg_end();
- Function::arg_iterator f2_ai = F2->arg_begin(), f2_ae = F2->arg_end();
- Function::arg_iterator f_ai = F->arg_begin(), f_ae = F->arg_end();
-
- // For the second node, we have to ignore the arguments that are incoming
- // edges (from II1)
- // Find all incoming edges.
- std::vector<IntrinsicInst *> IncomingEdgeList;
- getIncomingEdgeIntrinsicList(II2, IncomingEdgeList);
-
- // Their source must be the first fusion node, otherwise they would not have
- // been fusion candidates
- for (IntrinsicInst *ii : IncomingEdgeList) {
- assert((ii->getOperand(0) == II1) && "Unexpected source operand\n");
- }
-
- // Copy attributes of F1
- for(; f1_ai != f1_ae && f_ai != f_ae; ++f1_ai, ++f_ai) {
- AttributeSet AS = F1->getAttributes();
- DEBUG(errs() << "Copying attributes from "
- << F1->getName() << " at " << f1_ai->getArgNo() << "\n");
- AttrBuilder AB(AS, f1_ai->getArgNo()+1);
- AttributeSet argAS = AttributeSet::get(F1->getContext(),
- f_ai->getArgNo()+1, AB);
- F->addAttributes(f_ai->getArgNo()+1, argAS);
- }
-
- // Copy needed attributes of F2
- for(; f2_ai != f2_ae && f_ai != f_ae; ++f2_ai) {
- unsigned inport = f2_ai->getArgNo();
- if (isIncomingEdgeArgument(inport, IncomingEdgeList))
- continue;
-
- AttributeSet AS = F2->getAttributes();
- DEBUG(errs() << "Copying attributes from "
- << F2->getName() << " at " << f2_ai->getArgNo() << "\n");
- AttrBuilder AB(AS, f2_ai->getArgNo()+1);
- AttributeSet argAS = AttributeSet::get(F2->getContext(),
- f_ai->getArgNo()+1, AB);
- F->addAttributes(f_ai->getArgNo()+1, argAS);
- ++f_ai;;
- }
- return;
-}
-
-/* Creates and inserts an empty function of the rght type for the fused node */
-Function* FuseHPVMTensorNodes::createEmptyDFNodeFunction(IntrinsicInst* II1,
- IntrinsicInst* II2,
- Module &M) {
- Function* F1 = cast<Function>((II1->getOperand(0))->stripPointerCasts());
- Function* F2 = cast<Function>((II2->getOperand(0))->stripPointerCasts());
-
- DEBUG(errs () << "Constructing argument list\n");
- // Construct argument list
- std::vector<Type*> ArgTypes;
- createArgTypes(II1, II2, ArgTypes);
-
- DEBUG(errs () << "Constructing return type\n");
- // Construct return type
- StructType* FRetTy = createReturnType(II1, II2);
-
- FunctionType* FTy = FunctionType::get(FRetTy, ArgTypes, false);
- // Create a function with the new type
- Function* F = Function::Create(FTy, F1->getLinkage(),
- F1->getName()+"_"+F2->getName(), &M);
-
- DEBUG(errs () << "Copying argument names\n");
- // Copy argument names from original functions
- copyArgumentNames(II1, II2, F);
- // Copy argument attributes from original functions
- copyAttrList(II1, II2, F);
-
- return F;
-}
-
-/* Inline first node function, updating required mappings *
- * - F1: first node function *
- * - M: module containing the node function *
- * - Ffused: fused node function *
- * - VMap: maps values used in the body of F1 to those that mst be used in *
- 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) {
-
- ReturnInst *RI = cast<ReturnInst>(Ffused->getEntryBlock().getTerminator());
-
- inst_iterator f1_i = inst_begin(F1);
- // First, we copy the HPVM intrinsics of F1 into Ffused, applying the mapping
- for (inst_iterator f1_e = inst_end(F1); f1_i != f1_e; ++f1_i) {
- Instruction *I = &(*f1_i);
- if (!(BuildDFG::isViscIntrinsic(I))) {
- // We are done with the node computation
- break;
- }
-
- IntrinsicInst* II = dyn_cast<IntrinsicInst>(I);
- assert((II->getCalledFunction()->getName()).startswith("llvm.visc.tensor")
- && "Only HPVM tensor intrinsics allowed in ApproxHPVM leaf nodes\n");
-
- std::vector<Value*> Args;
- for(unsigned i = 0; i < II->getNumArgOperands(); i++) {
- Value *V = II->getArgOperand(i);
- if (isa<Constant>(V)) { // Constants can be reused
- Args.push_back(V);
- } else {
- assert((VMap.find(V) != VMap.end()) &&
- "Attempted to use value without existing mapping in VMap");
- Args.push_back(VMap[V]);
- }
- }
- Function *F = Intrinsic::getDeclaration(&M, II->getIntrinsicID());
- CallInst* CI =
- CallInst::Create(F, Args,
- F->getReturnType()->isVoidTy()? "" : "s_"+II->getName(), RI);
- // Update the map with the newly created value
- VMap[II] = CI;
- }
-
- // We continue with gathering information about the return values
- for (inst_iterator f1_e = inst_end(F1); f1_i != f1_e; ++f1_i) {
- Instruction *I = &(*f1_i);
- InsertValueInst* IV = dyn_cast<InsertValueInst>(I);
- if (!IV) {
- // End of insertvalue instructions. This should be a return statement
- assert((dyn_cast<ReturnInst>(I)) && "Unexpected Instruction\n");
- break; // Done processing this function
- }
- OutVs.push_back(IV->getOperand(1));
- }
- return;
-}
-
-/* Inline second node function, updating required mappings *
- * - F2: second node function *
- * - M: module containing the node function *
- * - Ffused: fused node function *
- * - VMap: maps values used in the body of F2 to those that mst be used in *
- the body of the fused function instead */
-void FuseHPVMTensorNodes::inlineSecondNodeFunction(Module &M, Function *F2,
- Function *Ffused, ValueMap<Value*, Value*> &VMap) {
-
- ReturnInst *RI = cast<ReturnInst>(Ffused->getEntryBlock().getTerminator());
-
- // Copy the body of F2 into Ffused, applying the mapping
- inst_iterator f2_i = inst_begin(F2);
- for (inst_iterator f2_e = inst_end(F2); f2_i != f2_e; ++f2_i) {
- Instruction *I = &(*f2_i);
- 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");
-
- std::vector<Value*> Args;
- for(unsigned i = 0; i < II->getNumArgOperands(); i++) {
- Value *V = II->getArgOperand(i);
- if (isa<Constant>(V)) { // Constants can be reused
- Args.push_back(V);
- } else {
- assert((VMap.find(V) != VMap.end()) &&
- "Attempted to use value without existing mapping in VMap");
- Args.push_back(VMap[V]);
- }
- }
- Function *F = Intrinsic::getDeclaration(&M, II->getIntrinsicID());
- CallInst* CI =
- CallInst::Create(F, Args,
- F->getReturnType()->isVoidTy()? "" : II->getName(),
- RI);
- // Update the map with the newly created value
- VMap[II] = CI;
- } else if (InsertValueInst* IV = dyn_cast<InsertValueInst>(I)) {
- Value *AggOp = IV->getAggregateOperand();
- Value *InsOp = IV->getInsertedValueOperand();
- assert(((VMap.find(AggOp) != VMap.end()) ||
- (isa<Constant>(AggOp)) ) &&
- "Attempted to use value without existing mapping in VMap");
- assert(((VMap.find(InsOp) != VMap.end()) ||
- (isa<Constant>(InsOp))) &&
- "Attempted to use value without existing mapping in VMap");
- InsertValueInst* IVI = InsertValueInst::Create(
- (isa<Constant>(AggOp)) ? AggOp : VMap[AggOp],
- (isa<Constant>(InsOp)) ? InsOp : VMap[InsOp],
- IV->getIndices(),
- IV->getName(),
- RI);
- // Update the map with the newly created value
- VMap[IV] = IVI;
- } else {
- ReturnInst* RetI = dyn_cast<ReturnInst>(I);
- assert(RetI && "Unexpected Instruction\n");
- Value *RetVal = RetI->getOperand(0);
- ReturnInst *newRI = ReturnInst::Create(Ffused->getContext(),
- VMap[RetVal]);
- ReplaceInstWithInst(RI, newRI);
- }
- }
- return;
-}
-
-/* Create function of leaf node after fusion *
- * - create type *
- * - create empty function of the type *
- * - inline body of first function (applying and updating appropriate *
- * mappings) *
- * - inline body of second function (applying and updating appropriate *
- * mappings) */
-Function* FuseHPVMTensorNodes::createLeafDFNodeFunction(IntrinsicInst* II1,
- IntrinsicInst* II2,
- Module &M) {
- DEBUG(errs () << "Creating function signature\n");
-
- /* Create empty node function of the correct type */
- Function* Ffused = createEmptyDFNodeFunction(II1, II2, M);
-
- // Get return type, needed for building the assignmens to the return struct
- StructType* FfusedRetTy = cast<StructType>(Ffused->getReturnType());
-
- /* Mapping information required for using the correct values in the body of *
- * the fused node function */
-
- // This map maps the values used in the original function bodies with
- // the ones that need to be used in the fused function body.
- ValueMap<Value*, Value*> FusedValueMap;
-
- // Intemediate information saved for return values of first node function
- // This maps the output port to the value returned through the outgoing edge
- std::vector<Value*> OutValues;
-
- DEBUG(errs () << "Creating function body\n");
-
- // Add a basic block to the new, empty function
- BasicBlock *BB = BasicBlock::Create(M.getContext(), "entry", Ffused);
- ReturnInst::Create(M.getContext(), UndefValue::get(FfusedRetTy), BB);
-
- // Get the node functions
- Function* F1 = cast<Function>((II1->getOperand(0))->stripPointerCasts());
- Function* F2 = cast<Function>((II2->getOperand(0))->stripPointerCasts());
-
- // Initially, update FusedValueMap: it is populated with the arguments of F1
- Function::arg_iterator fused_arg_it = Ffused->arg_begin();
- // Argument names of the first node are simply copied
- for(auto& arg: F1->getArgumentList()) {
- FusedValueMap[&arg] = &*fused_arg_it;
- ++fused_arg_it;
- }
-
-// 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);
-
- // Compute mapping between inputs of F2 and outputs of F1
- std::vector<IntrinsicInst *> IncomingEdgeList;
- getIncomingEdgeIntrinsicList(II2, IncomingEdgeList);
- std::vector<unsigned> PortMap(IncomingEdgeList.size(), 0);
- for (IntrinsicInst * ii : IncomingEdgeList) {
- unsigned srcPort = getNumericValue(ii->getOperand(3));
- unsigned dstPort = getNumericValue(ii->getOperand(4));
- PortMap[dstPort] = srcPort;
- }
-
- // FusedValueMap is now populated with the arguments of F2 as well
- for(auto& arg: F2->getArgumentList()) {
- DEBUG(errs() << arg << "\n");
- unsigned inport = arg.getArgNo();
- if (isIncomingEdgeArgument(inport, IncomingEdgeList)) {
- // Get the mappings of the return values of F1 if incoming edge argument
- Value *V = OutValues[PortMap[inport]];
- FusedValueMap[&arg] = (isa<Constant>(V)) ? V: FusedValueMap[V];
- }
- else {
- // Get new argument otherwise
- FusedValueMap[&arg] = &*fused_arg_it;
- ++fused_arg_it;
- }
- }
-
- // Invoke function that inlines F2 into Ffused, using and updating mappings
- inlineSecondNodeFunction(M, F2, Ffused, FusedValueMap);
-
- // Done with fused node function
- return Ffused;
-}
-
-/* Updates parent of fused nodes to use the new node intrinsic */
-void FuseHPVMTensorNodes::updateParentNodeFunction(IntrinsicInst* II1,
- IntrinsicInst* II2,
- IntrinsicInst* IInew) {
-
- // Compute the required shifting of positions for edges/binds to the second
- // fusion node. No shifting is required for the first fusion node.
- Function* F1 = cast<Function>((II1->getOperand(0))->stripPointerCasts());
- Function* F2 = cast<Function>((II2->getOperand(0))->stripPointerCasts());
- std::vector<unsigned> ShiftMap(F2->getFunctionType()->getNumParams(), 0);
- unsigned shiftCount = F1->getFunctionType()->getNumParams();
-
- // Find all incoming edges.
- std::vector<IntrinsicInst *> IncomingEdgeList;
- getIncomingEdgeIntrinsicList(II2, IncomingEdgeList);
- // Their source must be the first fusion node, otherwise they would not have
- // been fusion candidates
- for (IntrinsicInst *ii : IncomingEdgeList) {
- assert((ii->getOperand(0) == II1) && "Unexpected source operand\n");
- }
-
- // Compute shift map for n2: maps position in F2 arg list to Ffused arg list
- for(auto& arg: F2->getArgumentList()) {
- DEBUG(errs() << arg << "\n");
- unsigned inport = arg.getArgNo();
- if (isIncomingEdgeArgument(inport, IncomingEdgeList))
- continue;
-
- ShiftMap[inport] = shiftCount;
- shiftCount++;
- }
-
- std::vector<IntrinsicInst*> IItoRemove;
-
- // First, iterate over uses of the first node's createNode intrinsic
- for (Value::user_iterator i = II1->user_begin(), ie = II1->user_end();
- i != ie; ++i) {
- Instruction *VI = dyn_cast<Instruction>(*i);
- IntrinsicInst* II = dyn_cast<IntrinsicInst>(VI);
- assert(II && "Use of a node handle outside of a visc intrinsic\n");
-
- switch(II->getIntrinsicID()) {
- case Intrinsic::visc_createEdge:
- {
- if (isOutgoingEdgeIntrinsic(II,II1)) {
- assert(isIncomingEdgeIntrinsic(II,II2) &&
- "Outgoing edge of node 1 should only go to node 2\n");
- IItoRemove.push_back(II);
- }
- }
- break;
- case Intrinsic::visc_bind_input:
- {
- }
- break;
- case Intrinsic::visc_bind_output:
- {
- assert(false &&
- "Source node of node fusion not expected in bind.out\n");
- }
- break;
- default:
- llvm_unreachable("Unknown use of HPVM createNode handle\n");
- break;
- }
- }
-
- // Delete gathered instructions - they are the edges between n1-n2
- for (std::vector<IntrinsicInst *>::iterator ib = IItoRemove.begin(),
- ie = IItoRemove.end(); ib != ie; ++ib) {
- DEBUG(errs() << "Erasing: " << **ib << "\n");
- (*ib)->eraseFromParent();
- }
- II1->replaceAllUsesWith(IInew);
- II1->eraseFromParent();
-
- IItoRemove.clear();
-
- // Then, iterate over uses of the second node's createNode intrinsic
- for (Value::user_iterator i = II2->user_begin(), ie = II2->user_end();
- i != ie; ++i) {
- Instruction *VI = dyn_cast<Instruction>(*i);
- IntrinsicInst* II = dyn_cast<IntrinsicInst>(VI);
- assert(II && "Use of a node handle outside of a visc intrinsic\n");
-
- switch(II->getIntrinsicID()) {
- case Intrinsic::visc_createEdge:
- {
- assert(isOutgoingEdgeIntrinsic(II,II2) &&
- "Node 2 is expected to have only outgoing edges at this point\n");
- }
- break;
- case Intrinsic::visc_bind_input:
- {
- /* The index must be updated to the matching argument position of *
- * the fused functionm using ShiftMap */
- unsigned dstPos = cast<ConstantInt>(II->getOperand(2))->getZExtValue();
- IntrinsicInst *newII =
- createIdenticalBindWithDifferentPort(II,
- ShiftMap[dstPos],
- false);
- newII->insertBefore(II);
- IItoRemove.push_back(II);
- }
- break;
- case Intrinsic::visc_bind_output:
- {
- assert(false &&
- "Source node of node fusion not expected in bind.out\n");
- }
- break;
- default:
- llvm_unreachable("Unknown use of HPVM createNode handle\n");
- break;
- }
- }
-
- // Delete gathered instructions - they are the old bindings of n2
- for (std::vector<IntrinsicInst *>::iterator ib = IItoRemove.begin(),
- ie = IItoRemove.end(); ib != ie; ++ib) {
- DEBUG(errs() << "Erasing: " << **ib << "\n");
- (*ib)->eraseFromParent();
- }
- II2->replaceAllUsesWith(IInew);
- II2->eraseFromParent();
-
- return;
-}
-
-/* Performs all operations required at the IR level for fusion of HPVM tensor *
- * nodes with intrinsic instructions II1 and II2 *
- * - Creates fused node function *
- * - Creates createNode intrinsic for it and returns it *
- * - Updates parent function: *
- * - - adds new intrinsic *
- * - - edges and binds consistently use the new intrinsic *
- * - Removes old functions */
-IntrinsicInst* FuseHPVMTensorNodes::FuseHPVMTensorNodesStep(IntrinsicInst* II1,
- IntrinsicInst* II2,
- Module &M) {
- // Get the node functions
- Function* F1 = cast<Function>((II1->getOperand(0))->stripPointerCasts());
- Function* F2 = cast<Function>((II2->getOperand(0))->stripPointerCasts());
-
- // Create fused node function
- Function *Ffused = createLeafDFNodeFunction(II1, II2, M);
- addHint(Ffused, getPreferredTarget(F1));
-
- // FIX PARENT DFNode'S FUNCTION
-
- // Generate createNode Intrinsic for fused node and insert it
- Function* CreateNodeF = Intrinsic::getDeclaration(&M,
- Intrinsic::visc_createNode);
- Constant* Fp = ConstantExpr::getPointerCast(Ffused,
- Type::getInt8PtrTy(M.getContext()));
- CallInst *CI = CallInst::Create(CreateNodeF,
- ArrayRef<Value*>(Fp),
- Ffused->getName()+".node");
- IntrinsicInst* CreateNodeII = cast<IntrinsicInst>(CI);
- CreateNodeII->insertBefore(II1);
-
- // By the assumptions about the fusion pattern structure, all edges that have
- // II1 as source will have II2 as destination and vice versa.
- // We can simply delete them.
-
- // All createEdge intrinsics with destination argument = II1 need to use
- // CreateNodeII instead.
- // Similarly with bind.in
-
- // All createEdge intrinsics with source argument = II1 need to use
- // CreateNodeII instead
- // Similarly with bind.out
-
- // By the assumptions about the fusion pattern structure, the first node
- // cannot be the argument of a bind.out
- // The second node can be the argument of a bind.in.
- // For the bind.in, we need to adjust the destination port.
- updateParentNodeFunction(II1, II2, CreateNodeII);
-
- // Remove old node functions
- removeHint(F1, getPreferredTarget(F1));
- removeHint(F2, getPreferredTarget(F2));
- F1->replaceAllUsesWith(UndefValue::get(F1->getType()));
- F1->eraseFromParent();
- F2->replaceAllUsesWith(UndefValue::get(F2->getType()));
- F2->eraseFromParent();
-
- return CreateNodeII;
-}
-
-/* Fuse node sequence described by creaetNode intrinsics in IIs. *
- * Contents of IIs are cleared. */
-void FuseHPVMTensorNodes::FuseHPVMTensorNodeSequence(
- std::vector<IntrinsicInst*> &IIs, Module &M) {
- for (IntrinsicInst *II : IIs) {
- assert((II->getIntrinsicID() == Intrinsic::visc_createNode) &&
- "Expected createNode intrinsic in fuse intrinsic sequence\n");
- }
-
- if (IIs.size() < 2) {
- errs() << "Warning: Attempted to fuse fewer than 2 nodes\n";
- return;
- }
-
- for (unsigned i = 0; i + 1 < IIs.size(); i++) {
- IntrinsicInst *II1 = IIs[i];
- IntrinsicInst *II2 = IIs[i+1];
- IIs[i+1] = FuseHPVMTensorNodesStep(II1, II2, M);
- }
- IIs.clear();
- return;
-}
-
-/* Run method for FuseHPVMTensorNodes class, simply invokes fusion of all the *
- * sequenses in member variable FTs. */
-void FuseHPVMTensorNodes::run(Module &M, FusionTargets &FTs) {
- for (unsigned i = 0; i < FTs.size(); i++) {
- FuseHPVMTensorNodeSequence(FTs[i], M);
- }
- return;
-}
-
-// Print Fusion Targets. The argument vector contains createNode intrinsics
-// of nodes to be fused).
-void FuseHPVMTensorNodes::printFusionTargets(FusionTargets &FTs) {
- errs() << "Print Fusion Targets\n";
- errs() << "Found " << FTs.size() << " targets\n";
- for (FuseHPVMTensorNodes::FusionTargets::iterator ii = FTs.begin(),
- ie = FTs.end(); ii != ie ; ++ii) {
- errs() << "Target:\n";
- std::vector<IntrinsicInst*> IIv = *ii;
- for (std::vector< IntrinsicInst*>::iterator pi = IIv.begin(),
- pe = IIv.end(); pi != pe; ++pi) {
- errs() << "\t" << *((*pi)->getOperand(0)) << "\n";
- }
- }
- return;
-}
-
-void FindFusionTargetsTraversal::codeGen(DFInternalNode *N) {
- DEBUG(errs() << "Skipping Internal Node: "
- << N->getFuncPointer()->getName() << "\n");
- return;
-}
-
-void FindFusionTargetsTraversal::codeGen(DFLeafNode *N) {
- DEBUG(errs() << "Inside leaf node: "
- << N->getFuncPointer()->getName() << "\n");
-
- // Skip fusion check if it is a dummy node
- if(N->isDummyNode()) {
- DEBUG(errs() << "Skipping dummy node\n");
- return;
- }
-
-// if(N->getTargetHint() != visc::PROMISE_TARGET) {
- if(!preferredTargetIncludes(N, visc::PROMISE_TARGET)) {
- // Only fuse if we plan to target PROMISE
- // 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: "
- << N->getFuncPointer()->getName() << "\n");
- return;
- }
-
- visc::Target StartNodePreferredTarget = getPreferredTarget(N);
- // Make sure that this is a valid HPVM Tensor Node
- // Find first instruction, and check that it is an HPVM tensor intrinsic
- IntrinsicInst *II = isValidHPVMTensorNode(N);
-
- std::vector<IntrinsicInst*> CurrentNodeSequence;
-
- switch(II->getIntrinsicID()) {
- case Intrinsic::visc_tensor_convolution:
- { // Found beginning of pattern conv-bias-activation-pooling.
- // Look for the rest
- CurrentNodeSequence.push_back(N->getInstruction());
-
- // Look for bias
- DFNode *SN = findNextNodeInSequence(N);
- if (!SN) {
- return; // Did not find a node sequence starting at N. Simpy return.
- }
- if (getPreferredTarget(SN) != StartNodePreferredTarget) {
- return; // Node in sequence has different hint. Simpy return.
- }
- IntrinsicInst *SII = isValidHPVMTensorNode(SN);
- if (SII->getIntrinsicID() != Intrinsic::visc_tensor_add) {
- // Successor is not the bias operation, thus does not fit the pattern.
- return;
- }
- // Otherwise, push this node to the current sequence
- CurrentNodeSequence.push_back(SN->getInstruction());
-
- // This is a valid sequence.
- // We still need to fuse activation and/or pooling if we find them
- // Continue with next node, looking for activation (relu, clipped relu, tanh)
- SN = findNextNodeInSequence(SN);
- if (!SN) {
- // Did not find a node sequence starting at N.Use current sequence.
- break;
- }
- if (getPreferredTarget(SN) != StartNodePreferredTarget) {
- break; // Node in sequence has different hint. Use current sequence.
- }
- SII = isValidHPVMTensorNode(SN);
-
- if ((SII->getIntrinsicID() == Intrinsic::visc_tensor_clipped_relu) ||
- (SII->getIntrinsicID() == Intrinsic::visc_tensor_relu) ||
- (SII->getIntrinsicID() == Intrinsic::visc_tensor_tanh)) {
- // Successor is activation. Push this node to the current sequence.
- CurrentNodeSequence.push_back(SN->getInstruction());
-
- // Will continue, looking for pooling in the next node
- SN = findNextNodeInSequence(SN);
- if (!SN) {
- break; // No node in sequence. Use currently found sequence.
- }
- if (getPreferredTarget(SN) != StartNodePreferredTarget) {
- break; // Node in sequence has different hint. Use current sequence.
- }
- SII = isValidHPVMTensorNode(SN);
- } //else {} // Look for pooling in this node
-
- if ((SII->getIntrinsicID() == Intrinsic::visc_tensor_pool_max) ||
- (SII->getIntrinsicID() == Intrinsic::visc_tensor_pool_min) ||
- (SII->getIntrinsicID() == Intrinsic::visc_tensor_pool_mean)) {
- // Successor is a pool operation. Use currently found sequence.
- CurrentNodeSequence.push_back(SN->getInstruction());
- }
- }
- break;
- case Intrinsic::visc_tensor_mul:
- { // Found beginning of pattern gemm-bias-activation. Look for the rest
- CurrentNodeSequence.push_back(N->getInstruction());
- // Look for bias
- DFNode *SN = findNextNodeInSequence(N);
- if (!SN) {
- return; // Did not find a node sequence starting at N. Simpy return.
- }
- if (getPreferredTarget(SN) != StartNodePreferredTarget) {
- return; // Node in sequence has different hint. Simpy return.
- }
- IntrinsicInst *SII = isValidHPVMTensorNode(SN);
- if (SII->getIntrinsicID() != Intrinsic::visc_tensor_add) {
- // Successor is not the bias operation, thus does not fit the pattern.
- return;
- }
- // Otherwise, push this node to the current sequence
- CurrentNodeSequence.push_back(SN->getInstruction());
- // This is a possible fuse target, gemm-add.
- // We need to reach the end of the function, where the found sequence
- // is added.
-
- // If the next operation is activation, we fuse that as well.
- // Continue with next node, looking for activation (relu, clipped relu, tanh)
- SN = findNextNodeInSequence(SN);
- if (SN) {
- if (getPreferredTarget(SN) == StartNodePreferredTarget) {
- SII = isValidHPVMTensorNode(SN);
- if ((SII->getIntrinsicID() == Intrinsic::visc_tensor_clipped_relu) ||
- (SII->getIntrinsicID() == Intrinsic::visc_tensor_relu) ||
- (SII->getIntrinsicID() == Intrinsic::visc_tensor_tanh)) {
- // We found activation in sequence. Push in vector as well.
- CurrentNodeSequence.push_back(SN->getInstruction());
- }
- }
- }
- }
- break;
- default:
- DEBUG(errs() << "No pattern begins at this node\n");
- break;
- }
-
- if (CurrentNodeSequence.size() != 0) {
- // A sequence was found. Store the node sequence in FTs.
- FTs.push_back(CurrentNodeSequence);
- }
-
- return;
-}
-
-bool FuseHPVMTensorNodesWrapper::runOnModule(Module &M) {
- errs() << "\nFUSE HPVM TENSOR NODES PASS\n";
-
-// Get the BuildDFG Analysis Results:
- // - Dataflow graph
- BuildDFG &DFG = getAnalysis<BuildDFG>();
-
- std::vector<DFInternalNode*> Roots = DFG.getRoots();
- // Visitor for Fuse Target Detection Graph Traversal
- FindFusionTargetsTraversal *FTTVisitor =
- new FindFusionTargetsTraversal(M, DFG);
-
- errs() << "Find targets\n";
- // Iterate over all the DFGs and produce code for each one of them
- for (auto rootNode: Roots) {
- // Initiate code generation for root DFNode
- FTTVisitor->visit(rootNode);
- }
-
- FuseHPVMTensorNodes::FusionTargets &FTs = FTTVisitor->getFusionTargets();
-
- FuseHPVMTensorNodes Fuse;
-// Fuse.printFusionTargets(FTs);
-
- Fuse.run(M, FTs);
-
- delete FTTVisitor;
-
- return true;
-}
-
-char FuseHPVMTensorNodesWrapper::ID = 0;
-static RegisterPass<FuseHPVMTensorNodesWrapper> X("hpvm-fuse",
- "Fuse HPVM Tensor Nodes Pass",
- false /* does not modify the CFG */,
- true /* transformation, not just analysis */);
-
-} // End of namespace
-
diff --git a/llvm/lib/Transforms/FuseHPVMTensorNodes/FuseHPVMTensorNodes.exports b/llvm/lib/Transforms/FuseHPVMTensorNodes/FuseHPVMTensorNodes.exports
deleted file mode 100644
index e69de29bb2..0000000000
diff --git a/llvm/lib/Transforms/FuseHPVMTensorNodes/LLVMBuild.txt b/llvm/lib/Transforms/FuseHPVMTensorNodes/LLVMBuild.txt
deleted file mode 100644
index 55a6ee5150..0000000000
--- a/llvm/lib/Transforms/FuseHPVMTensorNodes/LLVMBuild.txt
+++ /dev/null
@@ -1,21 +0,0 @@
-;===- ./lib/Transforms/LocalMem/LLVMBuild.txt ------------------*- Conf -*--===;
-;
-; The LLVM Compiler Infrastructure
-;
-; This file is distributed under the University of Illinois Open Source
-; License. See LICENSE.TXT for details.
-;
-;===------------------------------------------------------------------------===;
-;
-; This is an LLVMBuild description file for the components in this subdirectory.
-;
-; For more information on the LLVMBuild system, please see:
-;
-; http://llvm.org/docs/LLVMBuild.html
-;
-;===------------------------------------------------------------------------===;
-
-[component_0]
-type = Library
-name = FuseHPVMTensorNodes
-parent = Transforms
diff --git a/llvm/lib/Transforms/InsertApproxInfo/CMakeLists.txt b/llvm/lib/Transforms/InsertApproxInfo/CMakeLists.txt
deleted file mode 100644
index 2b6d41bd70..0000000000
--- a/llvm/lib/Transforms/InsertApproxInfo/CMakeLists.txt
+++ /dev/null
@@ -1,12 +0,0 @@
-if(WIN32 OR CYGWIN)
- set(LLVM_LINK_COMPONENTS Core Support)
-endif()
-
-add_llvm_loadable_module( InsertApproxInfo
- InsertApproxInfo.cpp
-
- DEPENDS
- intrinsics_gen
- PLUGIN_TOOL
- opt
- )
diff --git a/llvm/lib/Transforms/InsertApproxInfo/InsertApproxInfo.cpp b/llvm/lib/Transforms/InsertApproxInfo/InsertApproxInfo.cpp
deleted file mode 100644
index bde4ef8907..0000000000
--- a/llvm/lib/Transforms/InsertApproxInfo/InsertApproxInfo.cpp
+++ /dev/null
@@ -1,498 +0,0 @@
-//===------------------------ InPlaceDFGAnalysis.cpp ----------------------===//
-//
-//
-//
-// The LLVM Compiler Infrastructure
-//
-//
-//
-// This file is distributed under the University of Illinois Open Source
-//
-// License. See LICENSE.TXT for details.
-//
-//
-//
-//===----------------------------------------------------------------------===//
-
-#define DEBUG_TYPE "InsertApproxInfo"
-
-#include "llvm/IR/Module.h"
-#include "llvm/Pass.h"
-#include "llvm/Support/SourceMgr.h"
-#include "llvm/InPlaceDFG/InPlaceDFGAnalysis.h"
-#include "llvm/SupportVISC/DFG2LLVM.h"
-#include "llvm/IR/InstrTypes.h"
-#include <unordered_map>
-#include <dirent.h>
-#include <stdio.h>
-#include <sstream>
-#include <fstream>
-
-
-using namespace llvm;
-using namespace builddfg;
-using namespace dfg2llvm;
-using namespace inplacedfg;
-
-
-namespace {
-
-static cl::opt<std::string> dir_name("results-dir", cl::desc(" Name of directory with Autotuner results "));
-
-
-struct ApproxMetrics{
- std::string op_name;
- std::string category;
- unsigned int rank; // rank given by autotuner
- double approx_level;
- // Relative L-norm metrics
- double relative_l1;
- double relative_l2;
- double relative_linf;
- // Mean L-norm metrics
- double mean_l1;
- double mean_l2;
- double mean_linf;
-};
-
-
-
-struct InsertApproxInfoWrapperPass : public ModulePass {
- static char ID; // Pass identification, replacement for typeid
- InsertApproxInfoWrapperPass() : ModulePass(ID) {}
-
-public:
- // Functions
- bool runOnModule(Module &M);
- void getAnalysisUsage(AnalysisUsage &AU) const;
-};
-
-
-// Visitor for Code generation traversal (tree traversal for now)
-class InsertApproxInfo : public CodeGenTraversal {
-
-private:
- // Virtual Functions
- void init() {}
- void initRuntimeAPI() {}
- void codeGen(DFInternalNode* N);
- void codeGen(DFLeafNode* N);
- void loadTrainedApproxMetrics(std::string dir_path);
- void loadMetricsFromFile(std::string dir_path, std::string file_path, std::string category);
- void loadMetricsFromDir(std::string dir_path, std::string category);
- void readApproxValues(const std::string line, ApproxMetrics* approx_metrics);
- void initIntrinsicNames();
- void initGlobalStrings();
-
- // private data
- std::unordered_map<std::string, std::string> intrinsics_map;
- std::unordered_map<std::string, std::vector<ApproxMetrics*>> operation_metrics;
- GlobalVariable* rank_str;
- GlobalVariable* category_str;
- GlobalVariable* mean_l1_str;
- GlobalVariable* mean_l2_str;
- GlobalVariable* mean_linf_str;
- GlobalVariable* rel_l1_str;
- GlobalVariable* rel_l2_str;
- GlobalVariable* rel_linf_str;
-
-
- // Tracks the id of the tensor op processed
- unsigned int currentID;
-
-public:
- // Constructor
- InsertApproxInfo(Module &_M, BuildDFG &_DFG);
-
- //void run(Module &M, BuildDFG &DFG);
- void run(std::string dir_path);
-
-};
-
-
-
-void InsertApproxInfo::initIntrinsicNames(){
-
- intrinsics_map["llvm.visc.tensor.convolution"] = "tensorConv";
- intrinsics_map["llvm.visc.tensor.mul"] = "tensorGemm";
- intrinsics_map["llvm.visc.tensor.add"] = "tensorAdd";
- intrinsics_map["llvm.visc.tensor.pool.max"] = "tensorPooling";
- intrinsics_map["llvm.visc.tensor.tanh"] = "tensorTanh";
-}
-
-
-void InsertApproxInfo::initGlobalStrings(){
-
- /**** Creating global constant strings for each approximation metric type *******/
-
- std::string rank_string = "rank";
- Constant* stringConst = ConstantDataArray::getString(M.getContext(), StringRef(rank_string.c_str()), true);
- rank_str = new GlobalVariable(M, stringConst->getType(), true,
- GlobalValue::ExternalLinkage, stringConst, "");
-
- std::string category_string = "category";
- stringConst = ConstantDataArray::getString(M.getContext(), StringRef(category_string.c_str()), true);
- category_str = new GlobalVariable(M, stringConst->getType(), true,
- GlobalValue::ExternalLinkage, stringConst, "");
-
- // Mean l-norm metrics
- std::string metric_string = "mean_l1";
- stringConst = ConstantDataArray::getString(M.getContext(), StringRef(metric_string.c_str()), true);
- mean_l1_str = new GlobalVariable(M, stringConst->getType(), true,
- GlobalValue::ExternalLinkage, stringConst, "");
-
- metric_string = "mean_l2";
- stringConst = ConstantDataArray::getString(M.getContext(), StringRef(metric_string.c_str()), true);
- mean_l2_str = new GlobalVariable(M, stringConst->getType(), true,
- GlobalValue::ExternalLinkage, stringConst, "");
-
- metric_string = "mean_linf";
- stringConst = ConstantDataArray::getString(M.getContext(), StringRef(metric_string.c_str()), true);
- mean_linf_str = new GlobalVariable(M, stringConst->getType(), true,
- GlobalValue::ExternalLinkage, stringConst, "");
-
- // Relative l-norm metrics
- metric_string = "rel_l1";
- stringConst = ConstantDataArray::getString(M.getContext(), StringRef(metric_string.c_str()), true);
- rel_l1_str = new GlobalVariable(M, stringConst->getType(), true,
- GlobalValue::ExternalLinkage, stringConst, "");
-
- metric_string = "rel_l2";
- stringConst = ConstantDataArray::getString(M.getContext(), StringRef(metric_string.c_str()), true);
- rel_l2_str = new GlobalVariable(M, stringConst->getType(), true,
- GlobalValue::ExternalLinkage, stringConst, "");
-
- metric_string = "rel_linf";
- stringConst = ConstantDataArray::getString(M.getContext(), StringRef(metric_string.c_str()), true);
- rel_linf_str = new GlobalVariable(M, stringConst->getType(), true,
- GlobalValue::ExternalLinkage, stringConst, "");
-
-}
-
-
-InsertApproxInfo::InsertApproxInfo(Module &_M, BuildDFG &_DFG) :
- CodeGenTraversal(_M, _DFG){
-
- currentID = 1;
-
- initIntrinsicNames();
- initGlobalStrings();
-}
-
-
-void InsertApproxInfoWrapperPass::getAnalysisUsage(AnalysisUsage &AU) const {
- AU.addRequired<BuildDFG>();
- AU.addPreserved<BuildDFG>();
-}
-
-
-bool InsertApproxInfoWrapperPass::runOnModule(Module &M) {
-
- std::string dir_path = dir_name.getValue();
- // Get the BuildDFG Analysis Results:
- // - Dataflow graph
- BuildDFG &DFG = getAnalysis<BuildDFG>();
-
- InsertApproxInfo IApprox(M, DFG);
- IApprox.run(dir_path);
-
- return false;
-}
-
-
-void InsertApproxInfo::readApproxValues(const std::string line, ApproxMetrics* approx_metrics){
-
- std::istringstream in(line);
- std::string op_name;
-
- float approx_level;
-
- float mean_l1;
- float mean_l2;
- float mean_linf;
-
- float relative_l1;
- float relative_l2;
- float relative_linf;
-
- in >> op_name;
- in >> approx_level;
-
- in >> mean_l1;
- in >> mean_l2;
- in >> mean_linf;
-
- in >> relative_l1;
- in >> relative_l2;
- in >> relative_linf;
-
- printf("\n *** op_name = %s \n", op_name.c_str());
- printf("approx_level = %f \n", approx_level);
- printf("relative_l1 = %f \n", relative_l1);
- printf("relative_l2 = %f \n", relative_l2);
- printf("relative_linf = %f \n", relative_linf);
- printf("mean_l1 = %f \n", mean_l1);
- printf("mean_l2 = %f \n", mean_l2);
- printf("mean_linf = %f \n", mean_linf);
-
- approx_metrics->op_name = op_name;
- approx_metrics->approx_level = approx_level;
- approx_metrics->mean_l1 = mean_l1;
- approx_metrics->mean_l2 = mean_l2;
- approx_metrics->mean_linf = mean_linf;
- approx_metrics->relative_l1 = relative_l1;
- approx_metrics->relative_l2 = relative_l2;
- approx_metrics->relative_linf = relative_linf;
-
-}
-
-
-unsigned int getFileRank(std::string file_path){
-
- char file_name[100]; // Assuming no file names greater than 100 chars
- strcpy(file_name, file_path.c_str());
-
- char* pch = strtok(file_name, "_");
- char* last_pch;
- while(pch != NULL){
- last_pch = pch;
- pch = strtok(NULL, "_");
- }
-
- printf("NOTE: ****** last_pch = %s \n", last_pch);
-
- size_t sz;
- int rank = std::stoi(last_pch, &sz);
-
- return rank + 1; // NOTE: Adding 1 to start ranks with '1'
-}
-
-
-
-void InsertApproxInfo::loadMetricsFromFile(std::string dir_path, std::string file_path, std::string category){
-
- std::string full_path = dir_path + "/" + file_path;
- printf("full_path = %s \n", full_path.c_str());
- std::ifstream infile(full_path.c_str());
- std::string line;
-
- unsigned int it_count = 0;
- while(std::getline(infile, line)){
-
- // Skip first line with confidence information
- if(it_count > 0){
- std::vector<float> approx_values;
- ApproxMetrics* approx_metrics = new ApproxMetrics;
- readApproxValues(line, approx_metrics);
-
- approx_metrics->category = category;
- unsigned int rank = getFileRank(file_path);
- approx_metrics->rank = rank;
-
- std::string unique_op_name = approx_metrics->op_name + std::to_string(it_count);
- operation_metrics[unique_op_name].push_back(approx_metrics);
- printf("\n ** unique_op_name = %s \n", unique_op_name.c_str());
- }
-
- it_count++;
- }
-
-}
-
-
-
-void InsertApproxInfo::loadMetricsFromDir(std::string dir_path, std::string category){
-
- struct dirent* entry;
- dir_path = dir_path + category;
-
- DIR* dir = opendir(dir_path.c_str());
- if(dir == NULL){
- printf("Directory %s not found . Aborting ... \n\n ", dir_path.c_str());
- abort();
- }
-
- while((entry = readdir(dir)) != NULL){
- printf("f_name = %s \n", entry->d_name);
- std::string f_name = entry->d_name;
- loadMetricsFromFile(dir_path, f_name, category);
- }
-}
-
-
-
-void InsertApproxInfo::loadTrainedApproxMetrics(std::string dir_path){
-
- std::string root_path = dir_path + "/high_confidence/";
- loadMetricsFromDir(root_path, "linear");
- loadMetricsFromDir(root_path, "log");
- loadMetricsFromDir(root_path, "quad");
-}
-
-
-/*** Methods of InPlaceDFGAnalysis ***/
-void InsertApproxInfo::run(std::string dir_path) {
-
- loadTrainedApproxMetrics(dir_path);
-
- errs() << "\n NOTE: ApproxInfo INSERTION TRANSFORM \n";
- std::vector<DFInternalNode*> Roots = DFG.getRoots();
-
- // Iterate over all the DFGs
- // Analyse the edges for parameters that are valid to be used in place
- for (auto rootNode: Roots) {
- //ATVisitor->visit(rootNode);
- this->visit(rootNode);
- }
-
- //delete ATVisitor;
- return;
-}
-
-/*** Analysis of internal node ***/
-void InsertApproxInfo::codeGen(DFInternalNode* N) {
- DEBUG(errs() << "Analysing Node: " << N->getFuncPointer()->getName() << "\n");
-}
-
-/*** Analysis of leaf node ***/
-void InsertApproxInfo::codeGen(DFLeafNode* N) {
- DEBUG(errs() << "Analysing Node: " << N->getFuncPointer()->getName() << "\n");
-
- // Skip code generation if it is a dummy node
- if(N->isDummyNode()) {
- DEBUG(errs() << "Skipping dummy node\n");
- return;
- }
-
- // Abort code generation if it is an allocation node
- if(N->isAllocationNode()) {
- assert(false && "Allocation Node not expected in ApproxHPVM");
- return;
- }
-
- Function *F = N->getFuncPointer();
- Module* M = F->getParent();
- std::vector<IntrinsicInst *> IItoRemove;
-
-
- /**** Adding operand bundles for each tensor operation in the HPVM DFG Leaf Node ****/
- for (inst_iterator i = inst_begin(F), e = inst_end(F); i != e; ++i) {
- Instruction *I = &(*i);
- errs()<<*I<<"\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");
-
- std::string intrinsic_id = std::string(II->getCalledFunction()->getName().data());
- std::string runtime_func_name = intrinsics_map[intrinsic_id];
- std::string unique_name = runtime_func_name + std::to_string(currentID);
- printf("\n ---- unique_name = %s \n ", unique_name.c_str());
- std::vector<ApproxMetrics*> approx_metrics;
- if(operation_metrics.find(unique_name) != operation_metrics.end()){
- approx_metrics = operation_metrics[unique_name];
- }
- else{
- errs()<<"Intrinsic Name NOT found in the map - Unexpected Error. Aborting ... \n\n";
- abort();
- }
-
-
- unsigned int num_configs = approx_metrics.size();
- std::vector<OperandBundleDef> conf_bundles;
- for(unsigned int i = 0; i < num_configs; i++){
- std::vector<Value*> norm_vals;
-
- norm_vals.push_back(category_str);
- Constant* categoryConst = ConstantDataArray::getString(M->getContext(), StringRef(approx_metrics[i]->category.c_str()), true);
- GlobalVariable* category_value = new GlobalVariable(*M, categoryConst->getType(), true,
- GlobalValue::ExternalLinkage, categoryConst, "");
- norm_vals.push_back(category_value);
-
- norm_vals.push_back(rank_str);
- Constant* constIntVal = ConstantInt::get(Type::getInt32Ty(M->getContext()), approx_metrics[i]->rank);
- norm_vals.push_back(constIntVal);
-
- // Adding mean l-norm metrics
- norm_vals.push_back(mean_l1_str);
- Constant* constFPVal = ConstantFP::get(Type::getDoubleTy(M->getContext()), approx_metrics[i]->mean_l1);
- norm_vals.push_back(constFPVal);
-
- norm_vals.push_back(mean_l2_str);
- constFPVal = ConstantFP::get(Type::getDoubleTy(M->getContext()), approx_metrics[i]->mean_l2);
- norm_vals.push_back(constFPVal);
-
- norm_vals.push_back(mean_linf_str);
- constFPVal = ConstantFP::get(Type::getDoubleTy(M->getContext()), approx_metrics[i]->mean_linf);
- norm_vals.push_back(constFPVal);
-
- // Relative l-norm Metrics
- norm_vals.push_back(rel_l1_str);
- constFPVal = ConstantFP::get(Type::getDoubleTy(M->getContext()), approx_metrics[i]->relative_l1);
- norm_vals.push_back(constFPVal);
-
- norm_vals.push_back(rel_l2_str);
- constFPVal = ConstantFP::get(Type::getDoubleTy(M->getContext()), approx_metrics[i]->relative_l2);
- norm_vals.push_back(constFPVal);
-
- norm_vals.push_back(rel_linf_str);
- constFPVal = ConstantFP::get(Type::getDoubleTy(M->getContext()), approx_metrics[i]->relative_linf);
- norm_vals.push_back(constFPVal);
-
-
- std::string config_name = "config_" + std::to_string(i+1);
- OperandBundleDef norm_bundle(config_name, norm_vals);
-
- conf_bundles.push_back(norm_bundle);
- }
-
- ArrayRef<OperandBundleDef> bundle_arr(conf_bundles);
-
- /*** Creating new Intrinsic call with Operand Bundles attached **/
- Function* calledFunction = II->getCalledFunction();
- unsigned num_args = II->getNumArgOperands();
- std::vector<Value*> args;
- for(unsigned i = 0; i < num_args; i++){
- Value* argValue = II->getArgOperand(i);
- args.push_back(argValue);
- }
-
- CallInst* CI = CallInst::Create(calledFunction,
- args, bundle_arr, "", II);
-
- errs()<<"NOTE: New CallInst = "<<*CI<<"\n";
-
- II->replaceAllUsesWith(CI);
- // Mark to remove at the end
- IItoRemove.push_back(II);
-
- // Increment counter of op processed
- currentID++;
- }
- }
-
-
- for (std::vector<IntrinsicInst *>::reverse_iterator ri = IItoRemove.rbegin(),
- re = IItoRemove.rend(); ri != re; ++ri) {
- DEBUG(errs() << "Erasing: " << **ri << "\n");
- errs() << "Erasing: " << **ri << "\n";
- (*ri)->eraseFromParent();
- }
-
-
-}
-
-char InsertApproxInfoWrapperPass::ID = 0;
-static RegisterPass<InsertApproxInfoWrapperPass> X("insert-approxinfo",
- "Pass to add approximation information (l-norm metrics) in the ApproxHPVM DFG",
- false /* does not modify the CFG */,
- false /* not transformation, just analysis */);
-
-
-
-
-
-} // End of namespace
-
diff --git a/llvm/lib/Transforms/InsertApproxInfo/LLVMBuild.txt b/llvm/lib/Transforms/InsertApproxInfo/LLVMBuild.txt
deleted file mode 100644
index e9cf5afd4a..0000000000
--- a/llvm/lib/Transforms/InsertApproxInfo/LLVMBuild.txt
+++ /dev/null
@@ -1,21 +0,0 @@
-;===- ./lib/Transforms/LocalMem/LLVMBuild.txt ------------------*- Conf -*--===;
-;
-; The LLVM Compiler Infrastructure
-;
-; This file is distributed under the University of Illinois Open Source
-; License. See LICENSE.TXT for details.
-;
-;===------------------------------------------------------------------------===;
-;
-; This is an LLVMBuild description file for the components in this subdirectory.
-;
-; For more information on the LLVMBuild system, please see:
-;
-; http://llvm.org/docs/LLVMBuild.html
-;
-;===------------------------------------------------------------------------===;
-
-[component_0]
-type = Library
-name = InsertApproxInfo
-parent = Transforms
--
GitLab