[IR2Vec] Scale vocab

Author: svkeerthy

llvm/include/llvm/Analysis/IR2Vec.h

@@ -108,6 +108,7 @@ struct Embedding {
   /// Arithmetic operators
   Embedding &operator+=(const Embedding &RHS);
   Embedding &operator-=(const Embedding &RHS);
+  Embedding &operator*=(double Factor);
 
   /// Adds Src Embedding scaled by Factor with the called Embedding.
   /// Called_Embedding += Src * Factor
@@ -116,6 +117,8 @@ struct Embedding {
   /// Returns true if the embedding is approximately equal to the RHS embedding
   /// within the specified tolerance.
   bool approximatelyEquals(const Embedding &RHS, double Tolerance = 1e-6) const;
+
+  void print(raw_ostream &OS) const;
 };
 
 using InstEmbeddingsMap = DenseMap<const Instruction *, Embedding>;
@@ -234,6 +237,8 @@ class IR2VecVocabResult {
 class IR2VecVocabAnalysis : public AnalysisInfoMixin<IR2VecVocabAnalysis> {
   ir2vec::Vocab Vocabulary;
   Error readVocabulary();
+  Error parseVocabSection(const char *Key, const json::Value ParsedVocabValue,
+                          ir2vec::Vocab &TargetVocab, unsigned &Dim);
   void emitError(Error Err, LLVMContext &Ctx);
 
 public:
@@ -249,14 +254,23 @@ class IR2VecVocabAnalysis : public AnalysisInfoMixin<IR2VecVocabAnalysis> {
 /// functions.
 class IR2VecPrinterPass : public PassInfoMixin<IR2VecPrinterPass> {
   raw_ostream &OS;
-  void printVector(const ir2vec::Embedding &Vec) const;
 
 public:
   explicit IR2VecPrinterPass(raw_ostream &OS) : OS(OS) {}
   PreservedAnalyses run(Module &M, ModuleAnalysisManager &MAM);
   static bool isRequired() { return true; }
 };
 
+/// This pass prints the embeddings in the vocabulary
+class IR2VecVocabPrinterPass : public PassInfoMixin<IR2VecVocabPrinterPass> {
+  raw_ostream &OS;
+
+public:
+  explicit IR2VecVocabPrinterPass(raw_ostream &OS) : OS(OS) {}
+  PreservedAnalyses run(Module &M, ModuleAnalysisManager &MAM);
+  static bool isRequired() { return true; }
+};
+
 } // namespace llvm
 
 #endif // LLVM_ANALYSIS_IR2VEC_H

llvm/lib/Analysis/IR2Vec.cpp

@@ -85,6 +85,12 @@ Embedding &Embedding::operator-=(const Embedding &RHS) {
   return *this;
 }
 
+Embedding &Embedding::operator*=(double Factor) {
+  std::transform(this->begin(), this->end(), this->begin(),
+                 [Factor](double Elem) { return Elem * Factor; });
+  return *this;
+}
+
 Embedding &Embedding::scaleAndAdd(const Embedding &Src, float Factor) {
   assert(this->size() == Src.size() && "Vectors must have the same dimension");
   for (size_t Itr = 0; Itr < this->size(); ++Itr)
@@ -101,6 +107,13 @@ bool Embedding::approximatelyEquals(const Embedding &RHS,
   return true;
 }
 
+void Embedding::print(raw_ostream &OS) const {
+  OS << " [";
+  for (const auto &Elem : Data)
+    OS << " " << format("%.2f", Elem) << " ";
+  OS << "]\n";
+}
+
 // ==----------------------------------------------------------------------===//
 // Embedder and its subclasses
 //===----------------------------------------------------------------------===//
@@ -196,18 +209,12 @@ void SymbolicEmbedder::computeEmbeddings(const BasicBlock &BB) const {
   for (const auto &I : BB.instructionsWithoutDebug()) {
     Embedding InstVector(Dimension, 0);
 
-    const auto OpcVec = lookupVocab(I.getOpcodeName());
-    InstVector.scaleAndAdd(OpcVec, OpcWeight);
-
     // FIXME: Currently lookups are string based. Use numeric Keys
     // for efficiency.
-    const auto Type = I.getType();
-    const auto TypeVec = getTypeEmbedding(Type);
-    InstVector.scaleAndAdd(TypeVec, TypeWeight);
-
+    InstVector += lookupVocab(I.getOpcodeName());
+    InstVector += getTypeEmbedding(I.getType());
     for (const auto &Op : I.operands()) {
-      const auto OperandVec = getOperandEmbedding(Op.get());
-      InstVector.scaleAndAdd(OperandVec, ArgWeight);
+      InstVector += getOperandEmbedding(Op.get());
     }
     InstVecMap[&I] = InstVector;
     BBVector += InstVector;
@@ -251,6 +258,46 @@ bool IR2VecVocabResult::invalidate(
   return !(PAC.preservedWhenStateless());
 }
 
+Error IR2VecVocabAnalysis::parseVocabSection(const char *Key,
+                                             const json::Value ParsedVocabValue,
+                                             ir2vec::Vocab &TargetVocab,
+                                             unsigned &Dim) {
+  assert(Key && "Key cannot be null");
+
+  json::Path::Root Path("");
+  const json::Object *RootObj = ParsedVocabValue.getAsObject();
+  if (!RootObj)
+    return createStringError(errc::invalid_argument,
+                             "JSON root is not an object");
+
+  const json::Value *SectionValue = RootObj->get(Key);
+  if (!SectionValue)
+    return createStringError(errc::invalid_argument,
+                             "Missing '" + std::string(Key) +
+                                 "' section in vocabulary file");
+  if (!json::fromJSON(*SectionValue, TargetVocab, Path))
+    return createStringError(errc::illegal_byte_sequence,
+                             "Unable to parse '" + std::string(Key) +
+                                 "' section from vocabulary");
+
+  Dim = TargetVocab.begin()->second.size();
+  if (Dim == 0)
+    return createStringError(errc::illegal_byte_sequence,
+                             "Dimension of '" + std::string(Key) +
+                                 "' section of the vocabulary is zero");
+
+  if (!std::all_of(TargetVocab.begin(), TargetVocab.end(),
+                   [Dim](const std::pair<StringRef, Embedding> &Entry) {
+                     return Entry.second.size() == Dim;
+                   }))
+    return createStringError(
+        errc::illegal_byte_sequence,
+        "All vectors in the '" + std::string(Key) +
+            "' section of the vocabulary are not of the same dimension");
+
+  return Error::success();
+};
+
 // FIXME: Make this optional. We can avoid file reads
 // by auto-generating a default vocabulary during the build time.
 Error IR2VecVocabAnalysis::readVocabulary() {
@@ -259,32 +306,40 @@ Error IR2VecVocabAnalysis::readVocabulary() {
     return createFileError(VocabFile, BufOrError.getError());
 
   auto Content = BufOrError.get()->getBuffer();
-  json::Path::Root Path("");
+
   Expected<json::Value> ParsedVocabValue = json::parse(Content);
   if (!ParsedVocabValue)
     return ParsedVocabValue.takeError();
 
-  bool Res = json::fromJSON(*ParsedVocabValue, Vocabulary, Path);
-  if (!Res)
-    return createStringError(errc::illegal_byte_sequence,
-                             "Unable to parse the vocabulary");
+  ir2vec::Vocab OpcodeVocab, TypeVocab, ArgVocab;
+  unsigned OpcodeDim, TypeDim, ArgDim;
+  if (auto Err = parseVocabSection("Opcodes", *ParsedVocabValue, OpcodeVocab,
+                                   OpcodeDim))
+    return Err;
 
-  if (Vocabulary.empty())
-    return createStringError(errc::illegal_byte_sequence,
-                             "Vocabulary is empty");
+  if (auto Err =
+          parseVocabSection("Types", *ParsedVocabValue, TypeVocab, TypeDim))
+    return Err;
 
-  unsigned Dim = Vocabulary.begin()->second.size();
-  if (Dim == 0)
+  if (auto Err =
+          parseVocabSection("Arguments", *ParsedVocabValue, ArgVocab, ArgDim))
+    return Err;
+
+  if (!(OpcodeDim == TypeDim && TypeDim == ArgDim))
     return createStringError(errc::illegal_byte_sequence,
-                             "Dimension of vocabulary is zero");
+                             "Vocabulary sections have different dimensions");
 
-  if (!std::all_of(Vocabulary.begin(), Vocabulary.end(),
-                   [Dim](const std::pair<StringRef, Embedding> &Entry) {
-                     return Entry.second.size() == Dim;
-                   }))
-    return createStringError(
-        errc::illegal_byte_sequence,
-        "All vectors in the vocabulary are not of the same dimension");
+  auto scaleVocabSection = [](ir2vec::Vocab &Vocab, double Weight) {
+    for (auto &Entry : Vocab)
+      Entry.second *= Weight;
+  };
+  scaleVocabSection(OpcodeVocab, OpcWeight);
+  scaleVocabSection(TypeVocab, TypeWeight);
+  scaleVocabSection(ArgVocab, ArgWeight);
+
+  Vocabulary.insert(OpcodeVocab.begin(), OpcodeVocab.end());
+  Vocabulary.insert(TypeVocab.begin(), TypeVocab.end());
+  Vocabulary.insert(ArgVocab.begin(), ArgVocab.end());
 
   return Error::success();
 }
@@ -304,7 +359,7 @@ void IR2VecVocabAnalysis::emitError(Error Err, LLVMContext &Ctx) {
 IR2VecVocabAnalysis::Result
 IR2VecVocabAnalysis::run(Module &M, ModuleAnalysisManager &AM) {
   auto Ctx = &M.getContext();
-  // FIXME: Scale the vocabulary once. This would avoid scaling per use later.
+
   // If vocabulary is already populated by the constructor, use it.
   if (!Vocabulary.empty())
     return IR2VecVocabResult(std::move(Vocabulary));
@@ -323,16 +378,9 @@ IR2VecVocabAnalysis::run(Module &M, ModuleAnalysisManager &AM) {
 }
 
 // ==----------------------------------------------------------------------===//
-// IR2VecPrinterPass
+// Printer Passes
 //===----------------------------------------------------------------------===//
 
-void IR2VecPrinterPass::printVector(const Embedding &Vec) const {
-  OS << " [";
-  for (const auto &Elem : Vec)
-    OS << " " << format("%.2f", Elem) << " ";
-  OS << "]\n";
-}
-
 PreservedAnalyses IR2VecPrinterPass::run(Module &M,
                                          ModuleAnalysisManager &MAM) {
   auto IR2VecVocabResult = MAM.getResult<IR2VecVocabAnalysis>(M);
@@ -353,15 +401,15 @@ PreservedAnalyses IR2VecPrinterPass::run(Module &M,
 
     OS << "IR2Vec embeddings for function " << F.getName() << ":\n";
     OS << "Function vector: ";
-    printVector(Emb->getFunctionVector());
+    Emb->getFunctionVector().print(OS);
 
     OS << "Basic block vectors:\n";
     const auto &BBMap = Emb->getBBVecMap();
     for (const BasicBlock &BB : F) {
       auto It = BBMap.find(&BB);
       if (It != BBMap.end()) {
         OS << "Basic block: " << BB.getName() << ":\n";
-        printVector(It->second);
+        It->second.print(OS);
       }
     }
 
@@ -373,10 +421,24 @@ PreservedAnalyses IR2VecPrinterPass::run(Module &M,
         if (It != InstMap.end()) {
           OS << "Instruction: ";
           I.print(OS);
-          printVector(It->second);
+          It->second.print(OS);
         }
       }
     }
   }
   return PreservedAnalyses::all();
 }
+
+PreservedAnalyses IR2VecVocabPrinterPass::run(Module &M,
+                                              ModuleAnalysisManager &MAM) {
+  auto IR2VecVocabResult = MAM.getResult<IR2VecVocabAnalysis>(M);
+  assert(IR2VecVocabResult.isValid() && "IR2Vec Vocabulary is invalid");
+
+  auto Vocab = IR2VecVocabResult.getVocabulary();
+  for (const auto &Entry : Vocab) {
+    OS << "Key: " << Entry.first << ": ";
+    Entry.second.print(OS);
+  }
+
+  return PreservedAnalyses::all();
+}