diff --git a/hpvm/lib/Transforms/BuildDFG/BuildDFG.cpp b/hpvm/lib/Transforms/BuildDFG/BuildDFG.cpp
index a8c807704a89853d54cc6b7c01efbc98deef78f2..a21bba52177a803018df501ccd5d7476f8155c15 100644
--- a/hpvm/lib/Transforms/BuildDFG/BuildDFG.cpp
+++ b/hpvm/lib/Transforms/BuildDFG/BuildDFG.cpp
@@ -32,20 +32,20 @@ using namespace llvm;
namespace builddfg {
bool BuildDFG::runOnModule(Module &M) {
- DEBUG(errs() << "\nBUILDDFG PASS\n");
- DEBUG(errs() << "-------- Searching for launch sites ----------\n");
+ //DEBUG(errs() << "\nBUILDDFG PASS\n");
+ //DEBUG(errs() << "-------- Searching for launch sites ----------\n");
IntrinsicInst *II;
// Iterate over all functions in the module
for (auto &Func : M) {
Function *F = &Func;
- DEBUG(errs() << "Function: " << F->getName() << "\n");
+ //DEBUG(errs() << "Function: " << F->getName() << "\n");
for (inst_iterator i = inst_begin(F), e = inst_end(F); i != e; ++i) {
Instruction *I = &*i; // Grab pointer to Instruction
if (isHPVMLaunchIntrinsic(I)) {
- DEBUG(errs() << "------------ Found launch site --------------\n");
+ // DEBUG(errs() << "------------ Found launch site --------------\n");
II = cast<IntrinsicInst>(I);
assert(II && "Launch intrinsic not recognized.");
@@ -53,8 +53,8 @@ bool BuildDFG::runOnModule(Module &M) {
// Intrinsic Instruction has been initialized from this point on.
Function *F = cast<Function>(II->getOperand(0)->stripPointerCasts());
Root = DFInternalNode::Create(II, F, hpvmUtils::getPreferredTarget(F));
- errs() << "INTRINSIC: " << II << "\n";
- errs() << "ROOT NODE" << Root << "\n";
+ //errs() << "INTRINSIC: " << II << "\n";
+ //errs() << "ROOT NODE" << Root << "\n";
Roots.push_back(Root);
BuildGraph(Root, F);
@@ -62,14 +62,14 @@ bool BuildDFG::runOnModule(Module &M) {
e = Root->getChildGraph()->end();
i != e; i++) {
DFNode *N = *i;
- DEBUG(errs() << "\t" << N->getFuncPointer()->getName() << "\n");
+ //DEBUG(errs() << "\t" << N->getFuncPointer()->getName() << "\n");
}
Root->getChildGraph()->sortChildren();
for (DFGraph::children_iterator i = Root->getChildGraph()->begin(),
e = Root->getChildGraph()->end();
i != e; i++) {
DFNode *N = *i;
- DEBUG(errs() << "\t" << N->getFuncPointer()->getName() << "\n");
+ //DEBUG(errs() << "\t" << N->getFuncPointer()->getName() << "\n");
}
viewDFGraph(Root->getChildGraph());
}
@@ -176,9 +176,9 @@ bool BuildDFG::isTypeCongruent(Type *L, Type *R) {
// Handles all the createNodeXX hpvm intrinsics.
void BuildDFG::handleCreateNode(DFInternalNode *N, IntrinsicInst *II) {
- errs() << "************ HANDLE CREATE NODE *********\n";
- II->print(errs());
- errs() << "\n";
+ //errs() << "************ HANDLE CREATE NODE *********\n";
+ //II->print(errs());
+ //errs() << "\n";
bool isInternalNode = false;
Function *F = cast<Function>((II->getOperand(0))->stripPointerCasts());
@@ -211,7 +211,7 @@ void BuildDFG::handleCreateNode(DFInternalNode *N, IntrinsicInst *II) {
// dataflow graph
DFInternalNode *childDFNode = DFInternalNode::Create(
II, F, hpvmUtils::getPreferredTarget(F), N, numOfDim, dimLimits);
- errs() << "INTERNAL NODE: " << childDFNode << "\n";
+ //errs() << "INTERNAL NODE: " << childDFNode << "\n";
N->addChildToDFGraph(childDFNode);
HandleToDFNodeMap[II] = childDFNode;
BuildGraph(childDFNode, F);
@@ -219,26 +219,26 @@ void BuildDFG::handleCreateNode(DFInternalNode *N, IntrinsicInst *II) {
// Create Leaf DFnode and add it to the map.
DFLeafNode *childDFNode = DFLeafNode::Create(
II, F, hpvmUtils::getPreferredTarget(F), N, numOfDim, dimLimits);
- errs() << "LEAF NODE: " << childDFNode << "\n";
+ //errs() << "LEAF NODE: " << childDFNode << "\n";
N->addChildToDFGraph(childDFNode);
HandleToDFNodeMap[II] = childDFNode;
}
}
void BuildDFG::handleCreateEdge(DFInternalNode *N, IntrinsicInst *II) {
- errs() << "************ HANDLE CREATE EDGE *********\n";
- II->print(errs());
- errs() << "\n";
+ //errs() << "************ HANDLE CREATE EDGE *********\n";
+ //II->print(errs());
+ //errs() << "\n";
// The DFNode structures must be in the map before the edge is processed
HandleToDFNode::iterator DFI = HandleToDFNodeMap.find(II->getOperand(0));
assert(DFI != HandleToDFNodeMap.end());
DFI = HandleToDFNodeMap.find(II->getOperand(1));
assert(DFI != HandleToDFNodeMap.end());
- errs() << "NODE TO MAP OPERAND 0: " << II->getOperand(0) << "\n";
- errs() << "NODE TO MAP OPERAND 1: " << II->getOperand(1) << "\n";
- errs() << "SRC NODE: " << HandleToDFNodeMap[II->getOperand(0)] << "\n";
- errs() << "DEST NODE: " << HandleToDFNodeMap[II->getOperand(1)] << "\n";
+ //errs() << "NODE TO MAP OPERAND 0: " << II->getOperand(0) << "\n";
+ //errs() << "NODE TO MAP OPERAND 1: " << II->getOperand(1) << "\n";
+ //errs() << "SRC NODE: " << HandleToDFNodeMap[II->getOperand(0)] << "\n";
+ //errs() << "DEST NODE: " << HandleToDFNodeMap[II->getOperand(1)] << "\n";
DFNode *SrcDF = HandleToDFNodeMap[II->getOperand(0)];
DFNode *DestDF = HandleToDFNodeMap[II->getOperand(1)];
@@ -272,23 +272,23 @@ void BuildDFG::handleCreateEdge(DFInternalNode *N, IntrinsicInst *II) {
DestPosition, DestTy, isStreaming);
HandleToDFEdgeMap[II] = newDFEdge;
- errs() << "NEW EDGE: " << newDFEdge << "\n";
+ //errs() << "NEW EDGE: " << newDFEdge << "\n";
// Add Edge to the dataflow graph associated with the parent node
N->addEdgeToDFGraph(newDFEdge);
}
void BuildDFG::handleBindInput(DFInternalNode *N, IntrinsicInst *II) {
- errs() << "************ HANDLE BIND INPUT *********\n";
- II->print(errs());
- errs() << "\n";
+ //errs() << "************ HANDLE BIND INPUT *********\n";
+ //II->print(errs());
+ //errs() << "\n";
// The DFNode structures must be in the map before the edge is processed
HandleToDFNode::iterator DFI = HandleToDFNodeMap.find(II->getOperand(0));
assert(DFI != HandleToDFNodeMap.end());
- errs() << "NODE TP MAP: " << II->getOperand(0) << "\n";
- errs() << "SRC NODE: " << N->getChildGraph()->getEntry() << "\n";
- errs() << "DEST NODE: " << HandleToDFNodeMap[II->getOperand(0)] << "\n";
+ //errs() << "NODE TP MAP: " << II->getOperand(0) << "\n";
+ //errs() << "SRC NODE: " << N->getChildGraph()->getEntry() << "\n";
+ //errs() << "DEST NODE: " << HandleToDFNodeMap[II->getOperand(0)] << "\n";
DFNode *SrcDF = N->getChildGraph()->getEntry();
DFNode *DestDF = HandleToDFNodeMap[II->getOperand(0)];
@@ -319,23 +319,23 @@ void BuildDFG::handleBindInput(DFInternalNode *N, IntrinsicInst *II) {
DestPosition, DestTy, isStreaming);
HandleToDFEdgeMap[II] = newDFEdge;
- errs() << "NEW EDGE: " << newDFEdge << "\n";
+ // errs() << "NEW EDGE: " << newDFEdge << "\n";
// Add Edge to the dataflow graph associated with the parent node
N->addEdgeToDFGraph(newDFEdge);
}
void BuildDFG::handleBindOutput(DFInternalNode *N, IntrinsicInst *II) {
- errs() << "************ HANDLE BIND OUTPUT *********\n";
- II->print(errs());
- errs() << "\n";
+ //errs() << "************ HANDLE BIND OUTPUT *********\n";
+ //II->print(errs());
+ //errs() << "\n";
// The DFNode structures must be in the map before the edge is processed
HandleToDFNode::iterator DFI = HandleToDFNodeMap.find(II->getOperand(0));
assert(DFI != HandleToDFNodeMap.end());
- errs() << "NODE TP MAP: " << II->getOperand(0) << "\n";
- errs() << "SRC NODE: " << HandleToDFNodeMap[II->getOperand(0)] << "\n";
- errs() << "DEST NODE: " << N->getChildGraph()->getExit() << "\n";
+ //errs() << "NODE TP MAP: " << II->getOperand(0) << "\n";
+ //errs() << "SRC NODE: " << HandleToDFNodeMap[II->getOperand(0)] << "\n";
+ //errs() << "DEST NODE: " << N->getChildGraph()->getExit() << "\n";
DFNode *SrcDF = HandleToDFNodeMap[II->getOperand(0)];
DFNode *DestDF = N->getChildGraph()->getExit();
@@ -366,14 +366,14 @@ void BuildDFG::handleBindOutput(DFInternalNode *N, IntrinsicInst *II) {
DestPosition, DestTy, isStreaming);
HandleToDFEdgeMap[II] = newDFEdge;
- errs() << "NEW EDGE: " << newDFEdge << "\n";
+ //errs() << "NEW EDGE: " << newDFEdge << "\n";
// Add Edge to the dataflow graph associated with the parent node
N->addEdgeToDFGraph(newDFEdge);
}
void BuildDFG::BuildGraph(DFInternalNode *N, Function *F) {
- DEBUG(errs() << "FUNCTION: " << F->getName() << "\n");
+ //DEBUG(errs() << "FUNCTION: " << F->getName() << "\n");
// TODO: Place checks for valid hpvm functions. For example one of the
// check can be that any function that contains hpvm dataflow graph
// construction intrinsics should not have other llvm IR statements.
@@ -382,10 +382,10 @@ void BuildDFG::BuildGraph(DFInternalNode *N, Function *F) {
// intrinsics.
for (inst_iterator i = inst_begin(F), e = inst_end(F); i != e; ++i) {
Instruction *I = &*i; // Grab pointer to Instruction
- DEBUG(errs() << *I << "\n");
+ // DEBUG(errs() << *I << "\n");
if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
- DEBUG(errs() << "IntrinsicID = " << II->getIntrinsicID() << ": "
- << II->getCalledFunction()->getName() << "\n");
+ // DEBUG(errs() << "IntrinsicID = " << II->getIntrinsicID() << ": "
+ // << II->getCalledFunction()->getName() << "\n");
switch (II->getIntrinsicID()) {
case Intrinsic::hpvm_createNode:
case Intrinsic::hpvm_createNode1D:
@@ -406,15 +406,15 @@ void BuildDFG::BuildGraph(DFInternalNode *N, Function *F) {
// TODO: Reconsider launch within a dataflow graph (recursion?)
case Intrinsic::hpvm_wait:
case Intrinsic::hpvm_launch:
- DEBUG(errs()
- << "Error: Launch/wait intrinsic used within a dataflow graph\n\t"
- << *II << "\n");
+ // DEBUG(errs()
+ // << "Error: Launch/wait intrinsic used within a dataflow graph\n\t"
+ // << *II << "\n");
break;
default:
- DEBUG(
- errs() << "Error: Invalid HPVM Intrinsic inside Internal node!\n\t"
- << *II << "\n");
+ //DEBUG(
+ // errs() << "Error: Invalid HPVM Intrinsic inside Internal node!\n\t"
+ // << *II << "\n");
break;
}
continue;
diff --git a/hpvm/lib/Transforms/DFG2LLVM_CPU/DFG2LLVM_CPU.cpp b/hpvm/lib/Transforms/DFG2LLVM_CPU/DFG2LLVM_CPU.cpp
index de9c025c0e7e996b6abfaa8748adf6688d04d10d..349614f01948db52911c6760e1e99117c06e1bf5 100644
--- a/hpvm/lib/Transforms/DFG2LLVM_CPU/DFG2LLVM_CPU.cpp
+++ b/hpvm/lib/Transforms/DFG2LLVM_CPU/DFG2LLVM_CPU.cpp
@@ -13,7 +13,9 @@
//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "DFG2LLVM_CPU"
+
#include "SupportHPVM/DFG2LLVM.h"
+
#include "llvm/IR/Constant.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/InstIterator.h"
@@ -1426,14 +1428,56 @@ void CGT_CPU::codeGen(DFLeafNode *N) {
<< " : skipping it\n");
switch (N->getTag()) {
- case hpvm::GPU_TARGET:
- // A leaf node should not have an cpu function for GPU
- // by design of DFG2LLVM_OpenCL backend
- assert(!(N->hasCPUGenFuncForTarget(hpvm::GPU_TARGET)) &&
+ case hpvm::GPU_TARGET:
+ {
+ // A leaf node should not have an cpu function for GPU
+ // by design of DFG2LLVM_OpenCL backend
+ assert(!(N->hasCPUGenFuncForTarget(hpvm::GPU_TARGET)) &&
"Leaf node not expected to have GPU GenFunc");
- break;
- default:
- break;
+ break;
+ }
+ case hpvm::CUDNN_TARGET:
+ {
+ errs() << "CUDNN hint found. Store CUDNN function as CPU funtion.\n";
+ // Make sure there is a generated CPU function for cudnn
+ assert(N->getGenFuncForTarget(hpvm::CUDNN_TARGET) && "");
+ assert(N->hasCPUGenFuncForTarget(hpvm::CUDNN_TARGET) && "");
+ // Store the CUDNN x86 function as the CPU generated function
+ Function *Ftmp = N->getGenFuncForTarget(N->getTag());
+ // after adding the required number of arguments
+ if (!N->getParent()->isChildGraphStreaming()) {
+ Ftmp = addIdxDimArgs(Ftmp);
+ }
+
+ N->removeGenFuncForTarget(hpvm::CUDNN_TARGET);
+ N->setTag(hpvm::None);
+ N->addGenFunc(Ftmp, hpvm::CPU_TARGET, true);
+ N->setTag(hpvm::CPU_TARGET);
+ break;
+ }
+ case hpvm::PROMISE_TARGET:
+ {
+ errs() << "Promise hint found. Store PROMISE function as CPU funtion.\n";
+ // Make sure there is a generated x86 function for promise
+ assert(N->getGenFuncForTarget(hpvm::PROMISE_TARGET) && "");
+ assert(N->hasCPUGenFuncForTarget(hpvm::PROMISE_TARGET) && "");
+ // Store the PROMISE x86 function as the CPU generated function
+ Function *Ftmp = N->getGenFuncForTarget(N->getTag());
+ // after adding the required number of arguments
+ if (!N->getParent()->isChildGraphStreaming()) {
+ Ftmp = addIdxDimArgs(Ftmp);
+ }
+
+ N->setTag(hpvm::None);
+ N->removeGenFuncForTarget(hpvm::PROMISE_TARGET);
+ N->addGenFunc(Ftmp, hpvm::CPU_TARGET, true);
+ N->setTag(hpvm::CPU_TARGET);
+ break;
+ }
+ default:
+ {
+ break;
+ }
}
return;
diff --git a/hpvm/lib/Transforms/DFG2LLVM_CUDNN/DFG2LLVM_CUDNN.cpp b/hpvm/lib/Transforms/DFG2LLVM_CUDNN/DFG2LLVM_CUDNN.cpp
index dce9f25319f6fd75d72c16cd847e98fe44b5a793..8b2570fdad9f43be73fa6682d4954413375a5041 100644
--- a/hpvm/lib/Transforms/DFG2LLVM_CUDNN/DFG2LLVM_CUDNN.cpp
+++ b/hpvm/lib/Transforms/DFG2LLVM_CUDNN/DFG2LLVM_CUDNN.cpp
@@ -143,7 +143,7 @@ void CGT_CUDNN::initRuntimeAPI() {
// FIXME: set correct path
Twine llvmSrcRoot = LLVM_SRC_ROOT;
- Twine runtimeAPI = llvmSrcRoot+"/projects/hpvm-tensor-rt/lib/tensor_runtime.ll";
+ Twine runtimeAPI = llvmSrcRoot+"/tools/hpvm/projects/hpvm-tensor-rt/lib/tensor_runtime.ll";
runtimeModule = parseIRFile(runtimeAPI.str(), Err, M.getContext());
if(runtimeModule == nullptr)
DEBUG(errs() << Err.getMessage());
diff --git a/hpvm/lib/Transforms/DFG2LLVM_WrapperAPI/DFG2LLVM_WrapperAPI.cpp b/hpvm/lib/Transforms/DFG2LLVM_WrapperAPI/DFG2LLVM_WrapperAPI.cpp
index aa310291a13acafdbad38a4269a848470c892bc2..294f9ac5742d561e79b8b81b3d88c22d3b156bcd 100644
--- a/hpvm/lib/Transforms/DFG2LLVM_WrapperAPI/DFG2LLVM_WrapperAPI.cpp
+++ b/hpvm/lib/Transforms/DFG2LLVM_WrapperAPI/DFG2LLVM_WrapperAPI.cpp
@@ -40,7 +40,7 @@ namespace {
cl::opt<std::string> QuantizationInputsFilename(
"quantization-levels-filename",
- cl::desc("<CPU quantization levels input file (path)>"),
+ cl::desc("<PROMISE quantization levels input file (path)>"),
cl::value_desc("filename"),
cl::Required);
@@ -280,6 +280,7 @@ public:
void InitialState::transition(CodeGenStateMachine *Mch, IntrinsicInst *II) {
if (II) { // Not end of instruction stream
+ errs() << "INITIAL STATE\n";
switch (II->getIntrinsicID()) {
case Intrinsic::hpvm_tensor_convolution:
{
@@ -288,6 +289,7 @@ void InitialState::transition(CodeGenStateMachine *Mch, IntrinsicInst *II) {
Mch->addArgument(II->getOperand(1)); // conv kernel
Mch->setCurrent(new ConvolutionLayer_1());
+ errs() << "TO CONVOLUTION LAYER 1\n";
}
break;
case Intrinsic::hpvm_tensor_mul:
@@ -297,6 +299,7 @@ void InitialState::transition(CodeGenStateMachine *Mch, IntrinsicInst *II) {
Mch->addArgument(II->getOperand(1)); // 2nd gemm input
Mch->setCurrent(new FullyConnectedLayer_1());
+ errs() << "TO FULLY CONNECTED LAYER 1\n";
}
break;
@@ -321,6 +324,7 @@ void InitialState::transition(CodeGenStateMachine *Mch, IntrinsicInst *II) {
Mch->addIntrinsicToRemove(II);
Mch->setCurrent(new InitialState());
+ errs() << "TO INIT STATE\n";
}
break;
@@ -328,24 +332,30 @@ void InitialState::transition(CodeGenStateMachine *Mch, IntrinsicInst *II) {
{
Mch->addIntrinsicInst(II);
Mch->setCurrent(new SingleTensorOperation());
+ errs() << "TO SINGLE OP\n";
}
break;
}
delete this;
- } // else {} // No HPVM intrinsic received. Remain at initial
+ } // else {} // No HPVM intrinsic received. Remain at initial
+ errs() << "TO NO CHANGE\n";
}
void SingleTensorOperation::transition(CodeGenStateMachine *Mch,
IntrinsicInst *II) {
if (II) { // Not end of instruction stream
+ errs() << "SINGLE TENSOR OP\n";
Mch->setCurrent(new NoPattern());
+ errs() << "TO NO PATTERN\n";
delete this;
}
+ errs() << "NO CHANGE\n";
}
void FullyConnectedLayer_1::transition(CodeGenStateMachine *Mch,
IntrinsicInst *II) {
if (II) { // Not end of instruction stream
+ errs() << "FULLY CONNECTED LAYER 1\n";
switch (II->getIntrinsicID()) {
case Intrinsic::hpvm_tensor_add:
{
@@ -357,14 +367,17 @@ void FullyConnectedLayer_1::transition(CodeGenStateMachine *Mch,
Mch->addArgument(II->getOperand(1)); // bias
Mch->setCurrent(new FullyConnectedLayer_2());
+ errs() << "TO FULLY CONNECTED LAYER 2\n";
}
break;
default:
Mch->setCurrent(new NoPattern());
+ errs() << "TO NO PATERN\n";
break;
}
} else {
Mch->setCurrent(new NoPattern());
+ errs() << "TO NO PATERN\n";
}
delete this;
}
@@ -372,6 +385,7 @@ void FullyConnectedLayer_1::transition(CodeGenStateMachine *Mch,
void FullyConnectedLayer_2::transition(CodeGenStateMachine *Mch,
IntrinsicInst *II) {
if (II) { // Not end of instruction stream
+ errs() << "FULLY CONNECTED LAYER 2\n";
switch (II->getIntrinsicID()) {
case Intrinsic::hpvm_tensor_tanh:
{
@@ -382,6 +396,7 @@ void FullyConnectedLayer_2::transition(CodeGenStateMachine *Mch,
Mch->addIntrinsicInst(II);
Mch->setCurrent(new FullyConnectedLayer_3());
+ errs() << "TO FULLY CONNECTED LAYER 3\n";
}
break;
case Intrinsic::hpvm_tensor_relu:
@@ -393,6 +408,7 @@ void FullyConnectedLayer_2::transition(CodeGenStateMachine *Mch,
Mch->addIntrinsicInst(II);
Mch->setCurrent(new FullyConnectedLayer_3());
+ errs() << "TO FULLY CONNECTED LAYER 3\n";
}
break;
case Intrinsic::hpvm_tensor_clipped_relu:
@@ -404,10 +420,12 @@ void FullyConnectedLayer_2::transition(CodeGenStateMachine *Mch,
Mch->addIntrinsicInst(II);
Mch->setCurrent(new FullyConnectedLayer_3());
+ errs() << "TO FULLY CONNECTED LAYER 3\n";
}
break;
default: // No activation, but HPVM intrinsic
Mch->setCurrent(new NoPattern());
+ errs() << "TO NO PATTERN\n";
break;
}
} else { // End of instruction stream
@@ -416,6 +434,7 @@ void FullyConnectedLayer_2::transition(CodeGenStateMachine *Mch,
Type::getInt32Ty(Mch->getModule()->getContext()), -1));
Mch->setCurrent(new FullyConnectedLayer());
+ errs() << "TO FULLY CONNECTED LAYER\n";
}
delete this;
}
@@ -423,9 +442,12 @@ void FullyConnectedLayer_2::transition(CodeGenStateMachine *Mch,
void FullyConnectedLayer_3::transition(CodeGenStateMachine *Mch,
IntrinsicInst *II) {
if (!II) { // End of instruction stream
+ errs() << "FULLY CONNECTED LAYER 3\n";
Mch->setCurrent(new FullyConnectedLayer());
+ errs() << "TO FULLY CONNECTED LAYER\n";
} else {
Mch->setCurrent(new NoPattern());
+ errs() << "TO NO PATTERN\n";
}
delete this;
}
@@ -433,14 +455,18 @@ void FullyConnectedLayer_3::transition(CodeGenStateMachine *Mch,
void FullyConnectedLayer::transition(CodeGenStateMachine *Mch,
IntrinsicInst *II) {
if (II) { // Not end of instruction stream
+ errs() << "FULLY CONNECTED LAYER\n";
Mch->setCurrent(new NoPattern());
+ errs() << "TO NO PATTERN\n";
delete this;
}
+ errs() << "TO NO CHANGE\n";
}
void ConvolutionLayer_1::transition(CodeGenStateMachine *Mch,
IntrinsicInst *II) {
if (II) { // Not end of instruction stream
+ errs() << "CONVOLUTION LAYER 1\n";
switch (II->getIntrinsicID()) {
case Intrinsic::hpvm_tensor_add:
{
@@ -457,10 +483,12 @@ void ConvolutionLayer_1::transition(CodeGenStateMachine *Mch,
Mch->addArgument(ConvII->getOperand(5)); // 4th numeric arg of conv
Mch->setCurrent(new ConvolutionLayer_2());
+ errs() << "TO CONVOLUTION LAYER 2\n";
}
break;
default:
Mch->setCurrent(new NoPattern());
+ errs() << "TO NO PATTERN\n";
break;
}
} else {
@@ -497,6 +525,7 @@ void ConvolutionLayer_1::transition(CodeGenStateMachine *Mch,
Type::getInt32Ty(Mch->getModule()->getContext()), -1));
Mch->setCurrent(new ConvolutionLayer());
+ errs() << "TO CONVOLUTION LAYER\n";
}
delete this;
}
@@ -504,6 +533,7 @@ void ConvolutionLayer_1::transition(CodeGenStateMachine *Mch,
void ConvolutionLayer_2::transition(CodeGenStateMachine *Mch,
IntrinsicInst *II) {
if (II) { // Not end of instruction stream
+ errs() << "CONVOLUTION LAYER 2\n";
switch (II->getIntrinsicID()) {
case Intrinsic::hpvm_tensor_tanh:
{
@@ -513,6 +543,7 @@ void ConvolutionLayer_2::transition(CodeGenStateMachine *Mch,
Mch->addIntrinsicInst(II);
Mch->setCurrent(new ConvolutionLayer_3());
+ errs() << "TO CONVOLUTION LAYER 3\n";
}
break;
case Intrinsic::hpvm_tensor_relu:
@@ -523,6 +554,7 @@ void ConvolutionLayer_2::transition(CodeGenStateMachine *Mch,
Mch->addIntrinsicInst(II);
Mch->setCurrent(new ConvolutionLayer_3());
+ errs() << "TO CONVOLUTION LAYER 3\n";
}
break;
case Intrinsic::hpvm_tensor_clipped_relu:
@@ -533,6 +565,7 @@ void ConvolutionLayer_2::transition(CodeGenStateMachine *Mch,
Mch->addIntrinsicInst(II);
Mch->setCurrent(new ConvolutionLayer_3());
+ errs() << "TO CONVOLUTION LAYER 3\n";
}
break;
case Intrinsic::hpvm_tensor_pool_max:
@@ -551,6 +584,7 @@ void ConvolutionLayer_2::transition(CodeGenStateMachine *Mch,
Mch->addIntrinsicInst(II);
Mch->setCurrent(new ConvolutionLayer_4());
+ errs() << "TO CONVOLUTION LAYER 4\n";
}
break;
case Intrinsic::hpvm_tensor_pool_min:
@@ -569,6 +603,7 @@ void ConvolutionLayer_2::transition(CodeGenStateMachine *Mch,
Mch->addIntrinsicInst(II);
Mch->setCurrent(new ConvolutionLayer_4());
+ errs() << "TO CONVOLUTION LAYER 4\n";
}
break;
case Intrinsic::hpvm_tensor_pool_mean:
@@ -587,10 +622,12 @@ void ConvolutionLayer_2::transition(CodeGenStateMachine *Mch,
Mch->addIntrinsicInst(II);
Mch->setCurrent(new ConvolutionLayer_4());
+ errs() << "TO CONVOLUTION LAYER 4\n";
}
break;
default: // No activation, No pooling, but HPVM intrinsic
Mch->setCurrent(new NoPattern());
+ errs() << "TO NO PATTERN\n";
break;
}
} else { // End of instruction stream
@@ -607,6 +644,7 @@ void ConvolutionLayer_2::transition(CodeGenStateMachine *Mch,
Type::getInt32Ty(Mch->getModule()->getContext()), -1));
Mch->setCurrent(new ConvolutionLayer());
+ errs() << "TO CONVOLUTION LAYER\n";
}
delete this;
}
@@ -614,6 +652,7 @@ void ConvolutionLayer_2::transition(CodeGenStateMachine *Mch,
void ConvolutionLayer_3::transition(CodeGenStateMachine *Mch,
IntrinsicInst *II) {
if (II) { // Not end of instruction stream
+ errs() << "CONVOLUTION LAYER 3\n";
switch (II->getIntrinsicID()) {
case Intrinsic::hpvm_tensor_pool_max:
{
@@ -643,6 +682,7 @@ void ConvolutionLayer_3::transition(CodeGenStateMachine *Mch,
}
Mch->setCurrent(new ConvolutionLayer_4());
+ errs() << "TO CONVOLUTION LAYER 4\n";
}
break;
case Intrinsic::hpvm_tensor_pool_min:
@@ -674,6 +714,7 @@ void ConvolutionLayer_3::transition(CodeGenStateMachine *Mch,
}
Mch->setCurrent(new ConvolutionLayer_4());
+ errs() << "TO CONVOLUTION LAYER 4\n";
}
break;
case Intrinsic::hpvm_tensor_pool_mean:
@@ -704,10 +745,12 @@ void ConvolutionLayer_3::transition(CodeGenStateMachine *Mch,
}
Mch->setCurrent(new ConvolutionLayer_4());
+ errs() << "TO CONVOLUTION LAYER 4\n";
}
break;
default: // No pooling, but HPVM intrinsic
Mch->setCurrent(new NoPattern());
+ errs() << "TO NO PATTERN\n";
break;
}
} else { // End of instruction stream
@@ -736,6 +779,7 @@ void ConvolutionLayer_3::transition(CodeGenStateMachine *Mch,
}
Mch->setCurrent(new ConvolutionLayer());
+ errs() << "TO CONVOLUTION LAYER\n";
}
delete this;
}
@@ -743,9 +787,12 @@ void ConvolutionLayer_3::transition(CodeGenStateMachine *Mch,
void ConvolutionLayer_4::transition(CodeGenStateMachine *Mch,
IntrinsicInst *II) {
if (!II) { // End of instruction stream
+ errs() << "CONVOLUTION LAYER 4\n";
Mch->setCurrent(new ConvolutionLayer());
+ errs() << "TO CONVOLUTION LAYER\n";
} else {
Mch->setCurrent(new NoPattern());
+ errs() << "TO NO PATTERN\n";
}
delete this;
}
@@ -753,9 +800,12 @@ void ConvolutionLayer_4::transition(CodeGenStateMachine *Mch,
void ConvolutionLayer::transition(CodeGenStateMachine *Mch,
IntrinsicInst *II) {
if (II) { // Not end of instruction stream
+ errs() << "CONVOLUTION LAYER\n";
Mch->setCurrent(new NoPattern());
+ errs() << "TO NO PATTERN\n";
delete this;
}
+ errs() << "NO CHANGE\n";
}
void NoPattern::transition(CodeGenStateMachine *Mch, IntrinsicInst *II) {}
@@ -772,6 +822,7 @@ void CodeGenStateMachine::transition(IntrinsicInst *II) {
void CodeGenStateMachine::codeGen(DFNode *N, Function *F, const StringRef &strRef,
InPlaceDFGAnalysis::InPlaceDFGParameter &IPP) {
+ errs() << "TRANSITIONTED TO: " << std::to_string(current->getStateID()) << "\n";
assert( ( (current->getStateID() == AbstractState::ID::FULLY_CONNECTED_LAYER) ||
(current->getStateID() == AbstractState::ID::CONVOLUTION_LAYER) ||
(current->getStateID() == AbstractState::ID::SINGLE_TENSOR_OPERATION) ) &&
@@ -1174,7 +1225,7 @@ void CodeGenStateMachine::codeGen(DFNode *N, Function *F, const StringRef &strRe
default:
- llvm_unreachable("Unknown VISC Intrinsic!");
+ llvm_unreachable("Unknown HPVM Intrinsic!");
break;
}
@@ -1280,7 +1331,7 @@ void CGT_WrapperAPI::initRuntimeAPI() {
// FIXME: set correct path
Twine llvmSrcRoot = LLVM_SRC_ROOT;
- Twine runtimeAPI = llvmSrcRoot+"/projects/hpvm-tensor-rt/lib/tensor_runtime.ll";
+ Twine runtimeAPI = llvmSrcRoot+"/tools/hpvm/projects/hpvm-tensor-rt/lib/tensor_runtime.ll";
runtimeModule = parseIRFile(runtimeAPI.str(), Err, M.getContext());
if(runtimeModule == nullptr)
DEBUG(errs() << Err.getMessage());
@@ -1363,6 +1414,7 @@ void CGT_WrapperAPI::codeGen(DFLeafNode* N) {
// Increment the node ID, for current node.
++nodeID;
+ errs() << "Node ID string: "<< StringRef(std::to_string(nodeID)) << "\n";
// Get the function associated with the dataflow node
Function *F = N->getFuncPointer();
@@ -1370,7 +1422,7 @@ void CGT_WrapperAPI::codeGen(DFLeafNode* N) {
// Look up if we have visited this function before. If we have, then just
// get the cloned function pointer from DFNode. Otherwise, create the cloned
// function and add it to the DFNode GenFunc.
- Function *F_wrapper_api = N->getGenFuncForTarget(hpvm::CPU_TARGET);
+ Function *F_wrapper_api = N->getGenFuncForTarget(hpvm::PROMISE_TARGET);
assert((F_wrapper_api == NULL) &&
"Error: Visiting a node for which code already generated");
@@ -1385,7 +1437,7 @@ void CGT_WrapperAPI::codeGen(DFLeafNode* N) {
F_wrapper_api->removeFromParent();
M.getFunctionList().push_back(F_wrapper_api);
- N->addGenFunc(F_wrapper_api, hpvm::CPU_TARGET, true);
+ N->addGenFunc(F_wrapper_api, hpvm::PROMISE_TARGET, true);
/* Removing HPVM in/out/inout function attributes */
for(Function::arg_iterator ai = F_wrapper_api->arg_begin(), ae = F_wrapper_api->arg_end();
@@ -1429,10 +1481,11 @@ void CGT_WrapperAPI::codeGen(DFLeafNode* N) {
for (inst_iterator i = inst_begin(F_wrapper_api), e = inst_end(F_wrapper_api);
i != e; ++i) {
Instruction *I = &(*i);
+ errs() << "PRINT INST: " << *I << "\n";
CGM.transition(dyn_cast<IntrinsicInst>(I));
}
-
- errs() << "Node ID string: "<< StringRef(std::to_string(nodeID)) << "\n";
+ errs() << "CLONED FUNCTION: " << *F_wrapper_api << "\n";
+ // 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);
diff --git a/hpvm/lib/Transforms/FuseHPVMTensorNodes/FuseHPVMTensorNodes.cpp b/hpvm/lib/Transforms/FuseHPVMTensorNodes/FuseHPVMTensorNodes.cpp
index 2ddc77ff4bcbfa1bd03ab5fc8fba0d35202980f2..e0032eaaf4cb073a99912a5eef08f1433f942234 100644
--- a/hpvm/lib/Transforms/FuseHPVMTensorNodes/FuseHPVMTensorNodes.cpp
+++ b/hpvm/lib/Transforms/FuseHPVMTensorNodes/FuseHPVMTensorNodes.cpp
@@ -823,18 +823,19 @@ void FindFusionTargetsTraversal::codeGen(DFInternalNode *N) {
void FindFusionTargetsTraversal::codeGen(DFLeafNode *N) {
DEBUG(errs() << "Inside leaf node: "
<< N->getFuncPointer()->getName() << "\n");
-
+ errs() << "FUSE TARGETS AT LEAF NODE\n";
// Skip fusion check if it is a dummy node
if(N->isDummyNode()) {
DEBUG(errs() << "Skipping dummy node\n");
return;
}
-
-
- if(!preferredTargetIncludes(N, hpvm::CPU_TARGET)) {
+ errs() << "THIS IS NOT A DUMMY NODE\n";
+ errs() << "INTRINSIC: " << *isValidHPVMTensorNode(N) << "\n";
+ if(!preferredTargetIncludes(N, hpvm::PROMISE_TARGET)) {
// Only fuse if we plan to target PROMISE/Layers API
// The CUDNN backend would be able to generate calls for the fused node,
// but not the other way around
+ errs() << "NO PROMISE HINT. SKIPPING NODE.\n";
DEBUG(errs() << "No PROMISE hint. Skipping node: "
<< N->getFuncPointer()->getName() << "\n");
return;
@@ -857,23 +858,29 @@ void FindFusionTargetsTraversal::codeGen(DFLeafNode *N) {
*/
case Intrinsic::hpvm_tensor_convolution:
- { // Found beginning of pattern conv-bias-activation-pooling.
+ { errs() << "INSTRUCTION: " << *II << "\n";
+
+ // 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) {
+ errs() << "DID NOT FIND ADD IN NODE SEQUENCE\n";
return; // Did not find a node sequence starting at N. Simpy return.
}
if (getPreferredTarget(SN) != StartNodePreferredTarget) {
+ errs() << "NODE IN SEQUENCE HAS DIFFERENT HINT\n";
return; // Node in sequence has different hint. Simpy return.
}
IntrinsicInst *SII = isValidHPVMTensorNode(SN);
if (SII->getIntrinsicID() != Intrinsic::hpvm_tensor_add) {
+ errs() << "SUCCESSOR IS NOT A BIAS OPERATION\n";
// Successor is not the bias operation, thus does not fit the pattern.
return;
}
+ errs() << "SUCCESSOR IS A BIAS OPERATION\n";
// Otherwise, push this node to the current sequence
CurrentNodeSequence.push_back(SN->getInstruction());
@@ -882,12 +889,15 @@ void FindFusionTargetsTraversal::codeGen(DFLeafNode *N) {
// Continue with next node, looking for activation (relu, clipped relu, tanh)
SN = findNextNodeInSequence(SN);
if (!SN) {
+ errs() << "DID NOT FIND POOLING AND ACTIVATION NODE SEQUENCE\n";
// Did not find a node sequence starting at N.Use current sequence.
break;
}
if (getPreferredTarget(SN) != StartNodePreferredTarget) {
+ errs() << "NODE IN SEQUENCE HAS DIFFERENT HINT\n";
break; // Node in sequence has different hint. Use current sequence.
}
+ errs() << "SUCCESSOR IS A ACTIVATION OR POOLING OPERATION\n";
SII = isValidHPVMTensorNode(SN);
if ((SII->getIntrinsicID() == Intrinsic::hpvm_tensor_clipped_relu) ||
@@ -895,13 +905,15 @@ void FindFusionTargetsTraversal::codeGen(DFLeafNode *N) {
(SII->getIntrinsicID() == Intrinsic::hpvm_tensor_tanh)) {
// Successor is activation. Push this node to the current sequence.
CurrentNodeSequence.push_back(SN->getInstruction());
-
+ errs() << "SUCCESSOR IS AN ACTIVATION OPERATION\n";
// Will continue, looking for pooling in the next node
SN = findNextNodeInSequence(SN);
if (!SN) {
+ errs() << "DID NOT FIND POOLING NODE SEQUENCE\n";
break; // No node in sequence. Use currently found sequence.
}
if (getPreferredTarget(SN) != StartNodePreferredTarget) {
+ errs() << "NODE IN SEQUENCE HAS DIFFERENT HINT\n";
break; // Node in sequence has different hint. Use current sequence.
}
SII = isValidHPVMTensorNode(SN);
@@ -910,6 +922,7 @@ void FindFusionTargetsTraversal::codeGen(DFLeafNode *N) {
if ((SII->getIntrinsicID() == Intrinsic::hpvm_tensor_pool_max) ||
(SII->getIntrinsicID() == Intrinsic::hpvm_tensor_pool_min) ||
(SII->getIntrinsicID() == Intrinsic::hpvm_tensor_pool_mean)) {
+ errs() << "SUCCESSOR IS A POOLING OPERATION\n";
// Successor is a pool operation. Use currently found sequence.
CurrentNodeSequence.push_back(SN->getInstruction());
}
@@ -921,16 +934,20 @@ void FindFusionTargetsTraversal::codeGen(DFLeafNode *N) {
// Look for bias
DFNode *SN = findNextNodeInSequence(N);
if (!SN) {
+ errs() << "DID NOT FIND ADD IN NODE SEQUENCE\n";
return; // Did not find a node sequence starting at N. Simpy return.
}
if (getPreferredTarget(SN) != StartNodePreferredTarget) {
+ errs() << "HINT DO NOT MATCH IN NODE SEQUENCE\n";
return; // Node in sequence has different hint. Simpy return.
}
IntrinsicInst *SII = isValidHPVMTensorNode(SN);
if (SII->getIntrinsicID() != Intrinsic::hpvm_tensor_add) {
+ errs() << "SUCCESSOR IS NOT IS BIAS OPERATION\n";
// Successor is not the bias operation, thus does not fit the pattern.
return;
}
+ errs() << "SUCCESSOR IS BIAS OPERATION\n";
// Otherwise, push this node to the current sequence
CurrentNodeSequence.push_back(SN->getInstruction());
// This is a possible fuse target, gemm-add.
@@ -946,6 +963,7 @@ void FindFusionTargetsTraversal::codeGen(DFLeafNode *N) {
if ((SII->getIntrinsicID() == Intrinsic::hpvm_tensor_clipped_relu) ||
(SII->getIntrinsicID() == Intrinsic::hpvm_tensor_relu) ||
(SII->getIntrinsicID() == Intrinsic::hpvm_tensor_tanh)) {
+ errs() << "SUCCESSOR IS ACTIVATION OPERATION\n";
// We found activation in sequence. Push in vector as well.
CurrentNodeSequence.push_back(SN->getInstruction());
}
diff --git a/hpvm/lib/Transforms/GenHPVM/GenHPVM.cpp b/hpvm/lib/Transforms/GenHPVM/GenHPVM.cpp
index 00063d6dc85c2de2ec609fce29024da037d84f12..cac468a704990746806684d5bad0b3054ccfcc9d 100644
--- a/hpvm/lib/Transforms/GenHPVM/GenHPVM.cpp
+++ b/hpvm/lib/Transforms/GenHPVM/GenHPVM.cpp
@@ -381,7 +381,7 @@ bool GenHPVM::runOnModule(Module &M) {
assert(isa<ConstantInt>(CI->getArgOperand(0)) &&
"Argument to hint must be constant integer!");
ConstantInt *hint = cast<ConstantInt>(CI->getArgOperand(0));
-
+ errs() << "HINT INSTRUCTION: " << *I << "\n";
hpvm::Target t = (hpvm::Target)hint->getZExtValue();
addHint(CI->getParent()->getParent(), t);
DEBUG(errs() << "Found hpvm hint call: " << *CI << "\n");