[tmva][sofie] Profiler code generation for SOFIE

tmva/sofie/CMakeLists.txt

@@ -22,6 +22,7 @@ ROOT_STANDARD_LIBRARY_PACKAGE(ROOTTMVASofie
    TMVA/OperatorList.hxx
    TMVA/RModel_Base.hxx
    TMVA/RModel.hxx
+   TMVA/RModelProfiler.hxx
    TMVA/ROperator.hxx
    TMVA/ROperator_BasicUnary.hxx
    TMVA/ROperator_BasicBinary.hxx
@@ -77,6 +78,7 @@ ROOT_STANDARD_LIBRARY_PACKAGE(ROOTTMVASofie
   SOURCES
     src/RModel_Base.cxx
     src/RModel.cxx
+    src/RModelProfiler.cxx
     src/RModel_GNN.cxx
     src/RModel_GraphIndependent.cxx
     src/RFunction.cxx

tmva/sofie/inc/TMVA/RModel.hxx

@@ -11,6 +11,8 @@ namespace SOFIE {
 
 class RModel final : public RModel_Base {
 
+   friend class RModelProfiler;
+
 private:
    bool fIsInitialized = false;
    bool fIsSubGraph = false;
@@ -20,6 +22,8 @@ private:
    size_t fConstantTensorSize = 0; // size  (in Bytes) of the allocated constant tensors
    size_t fWeightsTensorSize = 0;  // size  (in Bytes) of the allocated weight tensors
    size_t fOtherTensorSize = 0;    // size  (in Bytes) of intermediate tensors which are not managed by the memory pool
+   std::string fProfilerGC = "";
+   bool fProfile = false;
 
    OptimizationLevel fOptimizationLevel = OptimizationLevel::kExtended;
 
@@ -151,8 +155,8 @@ public:
 
    void Initialize(int batchSize = -1, bool verbose = false);
    void Initialize(const std::map<std::string,size_t> & inputParams, bool verbose = false);
-
-   void Generate(std::underlying_type_t<Options> options, int batchSize = -1, long pos = 0, bool verbose = false);
+   
+    void Generate(std::underlying_type_t<Options> options, int batchSize = -1, long pos = 0, bool verbose = false);
    void Generate(Options options = Options::kDefault, int batchSize = -1, int pos = 0, bool verbose = false)
    {
       Generate(static_cast<std::underlying_type_t<Options>>(options), batchSize, pos, verbose);

tmva/sofie/inc/TMVA/RModelProfiler.hxx

@@ -0,0 +1,42 @@
+#ifndef TMVA_SOFIE_RMODELPROFILER
+#define TMVA_SOFIE_RMODELPROFILER
+
+#include "TMVA/RModel.hxx"
+
+namespace TMVA {
+namespace Experimental {
+namespace SOFIE {
+
+/// \class RModelProfiler
+/// \brief A helper class to generate profiled inference code for an RModel.
+///
+/// This class instruments the generated C++ code to measure the execution
+/// time of each operator. It is invoked when the RModel::Generate is called
+/// with the Options::kProfile flag. 
+class RModelProfiler {
+private:
+   RModel &fModel;
+
+   void GenerateUtilityFunctions();
+
+public:
+   // The profiler must be constructed with a model to work on.
+   RModelProfiler() = delete;
+   RModelProfiler(RModel &model);
+   ~RModelProfiler() = default;
+
+   // There is no point in copying or moving an RModelProfiler
+   RModelProfiler(const RModelProfiler &other) = delete;
+   RModelProfiler(RModelProfiler &&other) = delete;
+   RModelProfiler &operator=(const RModelProfiler &other) = delete;
+   RModelProfiler &operator=(RModelProfiler &&other) = delete;
+
+   // Main function to generate the profiled code.
+   void Generate();
+};
+
+} // namespace SOFIE
+} // namespace Experimental
+} // namespace TMVA
+
+#endif // TMVA_SOFIE_RMODELPROFILER

tmva/sofie/inc/TMVA/RModel_Base.hxx

@@ -26,6 +26,7 @@ enum class Options {
    kRootBinaryWeightFile = 0x4,
    kGNN = 0x8,
    kGNNComponent = 0x10,
+   kProfile = 0x20,
 };
 
 // Optimization levels inspired by ONNXRuntime.

tmva/sofie/inc/TMVA/ROperator.hxx

@@ -37,6 +37,9 @@ public:
    //virtual void Forward_blas() = 0;
    virtual ~ROperator(){}
 
+   std::string name = "UnnamedOperator";
+   const std::string &GetOperatorName() { return name; };
+
 protected:
 
    const std::string SP = "   ";    ///< space used to correctly indent the generated C++ code

tmva/sofie/src/RModel.cxx

@@ -9,6 +9,7 @@
 #endif
 
 #include "TMVA/RModel.hxx"
+#include "TMVA/RModelProfiler.hxx"
 #include "TMVA/SOFIE_common.hxx"
 
 namespace TMVA {
@@ -941,7 +942,7 @@ void RModel::GenerateSessionCode()
          CheckAndFlushIntermediateMemory(fOperators[op_idx]->GetOpInputTensors(), op_idx);
       }
 
-      // to check remaining unused fragments after memory allocation (lesser the better)
+  // to check remaining unused fragments after memory allocation (lesser the better)
       // for (const auto &it: fIntermediateMemoryInfo.available_stack){
       //    std::cout<<"chunk_idx: "<<it.first<<", chunk_size: "<<it.second<<"\n";
       // }
@@ -969,13 +970,13 @@ void RModel::GenerateSessionCode()
    // Generate code for Session constructor
    if (fUseSession) {
       std::string sessionName = "Session";
-      if (fIsSubGraph)
+      if (fIsSubGraph) 
          sessionName += "_" + fName;
       // add here specific operator code that needs to define session data members
       fGC += "\n";
       for (size_t id = 0; id < fOperators.size(); id++) {
          std::string opName = std::to_string(id);
-         fGC += fOperators[id]->GenerateSessionMembersCode(opName);
+         fGC += fOperators[id]->GenerateSessionMembersCode(opName);        
       }
       fGC += "\n";
       // here add initialization and reading of weight tensors
@@ -1021,23 +1022,28 @@ void RModel::GenerateSessionCode()
       fGC += "}\n\n";
    }
 
-   fGC += doInferSignature + "{\n";
-   fGC += "\n";
+   if (fProfile) {
+      RModelProfiler profiler(*this);
+      profiler.Generate();
+      fGC += fProfilerGC; 
+   } else {
+      fGC += doInferSignature + "{\n";
+      fGC += "\n";
 
-   // generate the inference code
-   if (fVerbose)
-      std::cout << "Generating main inference code for " << fName << std::endl;
+      // generate the inference code
+      if (fVerbose)
+         std::cout << "Generating main inference code for " << fName << std::endl;
 
-   if (fOutputTensorNames.size() == 0)
-      throw std::runtime_error("TMVA-SOFIE: output size=0 are not supported");
+      if (fOutputTensorNames.size() == 0)
+         throw std::runtime_error("TMVA-SOFIE: output size=0 are not supported");
 
-   for (size_t op_idx = 0; op_idx < fOperators.size(); ++op_idx) {
-      if (fVerbose)
+      for (size_t op_idx = 0; op_idx < fOperators.size(); ++op_idx) {
+         if (fVerbose)
          std::cout << "Generating code for operator .... " << op_idx << std::endl;
-      fGC += (fOperators[op_idx]->Generate(std::to_string(op_idx)));
-   }
+         fGC += (fOperators[op_idx]->Generate(std::to_string(op_idx)));
+      }
 
-   fGC += SP + "using TMVA::Experimental::SOFIE::UTILITY::FillOutput;\n\n";
+      fGC += SP + "using TMVA::Experimental::SOFIE::UTILITY::FillOutput;\n\n";
 
    for (std::string const &name : fOutputTensorNames) {
       // need to check is size is the same (don't want to return a vector with
@@ -1048,7 +1054,8 @@ void RModel::GenerateSessionCode()
       fGC += SP + "FillOutput(tensor_" + name + ", output_tensor_" + name + ", " + n + ");\n";
    }
 
-   fGC += "}\n\n";
+      fGC += "}\n\n";
+   }
 
    // generate the inference overload that returns an output struct
    GenerateOutput();
@@ -1061,9 +1068,11 @@ void RModel::GenerateSessionCode()
 
 void RModel::Generate(std::underlying_type_t<Options> options, int batchSize, long pos, bool verbose)
 {
+   bool profile = (options & static_cast<std::underlying_type_t<Options>>(Options::kProfile));
    fVerbose = verbose;
    fBatchSize = batchSize;
    fReadPos = pos;
+   fProfile = profile;
 
    // session flag is used in operator initialize
    if (static_cast<std::underlying_type_t<Options>>(Options::kNoSession) & options) {
@@ -1083,9 +1092,9 @@ void RModel::Generate(std::underlying_type_t<Options> options, int batchSize, lo
          "TMVA-SOFIE: RModel::Generate: cannot use a separate weight file without generating a Session class");
    }
 
-   if (static_cast<std::underlying_type_t<Options>>(Options::kGNN) & options)
+   if (static_cast<std::underlying_type_t<Options>>(Options::kGNN) & options) 
       fIsGNN = true;
-   if (static_cast<std::underlying_type_t<Options>>(Options::kGNNComponent) & options)
+   if (static_cast<std::underlying_type_t<Options>>(Options::kGNNComponent) & options) 
       fIsGNNComponent = true;
 
    // initialize the model including all operators and sub-graphs
@@ -1099,13 +1108,13 @@ void RModel::Generate(std::underlying_type_t<Options> options, int batchSize, lo
 
    // generate first code for the subgraphs
    for (auto &graph : fSubGraphs) {
-      if (fVerbose)
+      if (fVerbose) 
          std::cout << "generate session code for subgraph " << graph->fName << std::endl;
       graph->GenerateSessionCode();
       fGC += graph->fGC;
    }
 
-   if (fVerbose)
+   if (fVerbose) 
       std::cout << "generate Main session code - model  " << fName << std::endl;
 
    // generate main session code