root/TMVA/ROperator_Tile.hxx

0001 #ifndef TMVA_SOFIE_ROPERATOR_Tile
0002 #define TMVA_SOFIE_ROPERATOR_Tile
0003
0004 #include "TMVA/SOFIE_common.hxx"
0005 #include "TMVA/ROperator.hxx"
0006 #include "TMVA/RModel.hxx"
0007
0008 #include <sstream>
0009
0010 namespace TMVA{
0011 namespace Experimental{
0012 namespace SOFIE{
0013
0014 template <typename T>
0015 class ROperator_Tile final : public ROperator
0016 {
0017
0018 private:
0019
0020    std::string fNRepeats;
0021    std::string fNInput;
0022    std::string fNY;
0023    std::vector<size_t>fShapeInput;
0024    std::vector<size_t> fShapeY;
0025
0026 public:
0027    ROperator_Tile(){}
0028    ROperator_Tile(std::string nameRepeat, std::string nameInput, std::string nameY):
0029       fNRepeats(UTILITY::Clean_name(nameRepeat)),fNInput(UTILITY::Clean_name(nameInput)), fNY(UTILITY::Clean_name(nameY)){
0030          fInputTensorNames = { fNRepeats, fNInput };
0031          fOutputTensorNames = { fNY };
0032       }
0033
0034    std::vector<ETensorType> TypeInference(std::vector<ETensorType> input) override {
0035       return input;
0036    }
0037
0038    std::vector<std::vector<size_t>> ShapeInference(std::vector<std::vector<size_t>> input) override {
0039       std::vector<size_t> ret = input[0];
0040
0041       for(size_t i=0; i < input[1].size(); i++) {
0042             ret[i]=ret[i]*input[1][i];
0043       }
0044       return {ret};
0045    }
0046
0047    void Initialize(RModel& model) override {
0048        //input must be a graph input, or already initialized intermediate tensor
0049       if (model.CheckIfTensorAlreadyExist(fNInput) == false){
0050         throw std::runtime_error("TMVA SOFIE Tile Op Input Tensor is not found in model");
0051       }
0052       if (model.CheckIfTensorAlreadyExist(fNRepeats) == false){
0053         throw std::runtime_error("TMVA SOFIE Tile Op Input Tensor is not found in model");
0054       }
0055       fShapeInput=model.GetTensorShape(fNInput);
0056
0057       // if repeats vector is not initialized we cannot deduce shape of output
0058       // not support for time being this case
0059       if (!model.IsInitializedTensor(fNRepeats)) {
0060          throw std::runtime_error("TMVA SOFIE Tile Op: non-initialized repeats input is not supported");
0061       }
0062
0063       // Retrieve the data pointer for the repeats tensor
0064       auto repptr = model.GetInitializedTensorData(fNRepeats);
0065       // Cast the raw pointer to the appropriate type (size_t*)
0066       auto repeats_data = static_cast<int64_t*>(repptr.get());
0067       if (repeats_data == nullptr) {
0068         throw std::runtime_error("Failed to retrieve the data for the repeats tensor.");
0069       }
0070       // Get the shape of the repeats tensor to determine the number of elements
0071       auto repeats_shape = model.GetTensorShape(fNRepeats);
0072       // Ensure the repeats tensor is 1D and get the number of elements
0073       if (repeats_shape.size() != 1) {
0074          throw std::runtime_error("Repeats tensor is not 1D.");
0075       }
0076       size_t num_elements = repeats_shape[0];
0077       // Convert the data to a vector of size_t
0078       std::vector<size_t> repeats_vector(num_elements);
0079       std::copy(repeats_data, repeats_data + num_elements, repeats_vector.begin());
0080
0081
0082       fShapeY = ShapeInference({fShapeInput,repeats_vector})[0];
0083
0084       model.AddIntermediateTensor(fNY, model.GetTensorType(fNInput), fShapeY);
0085
0086       if (model.Verbose())
0087          std::cout <<  "Tile: " << fNInput << " " << ConvertShapeToString(fShapeInput) << " -> " << fNY << " with shape " << ConvertShapeToString(fShapeY)
0088             << " given repeats " << ConvertShapeToString(repeats_vector) << std::endl;
0089    }
0090
0091    std::string Generate(std::string OpName) override {
0092       OpName = "op_" + OpName;
0093       if (fShapeInput.empty() || fShapeY.empty()) {
0094             throw std::runtime_error("TMVA SOFIE Tile Op called to Generate without being initialized first");
0095       }
0096
0097       //size_t input_length = ConvertShapeToLength(fShapeInput);
0098       //size_t output_length = ConvertShapeToLength(fShapeY);
0099
0100
0101       std::stringstream out;
0102       std::string input = "tensor_" + fNInput;
0103       std::string output = "tensor_" + fNY;
0104       out << "///-------- Tile operator\n";
0105       out << "{\n"; // add scope to re-use same names
0106       out << "const int input_shape[" << fShapeInput.size() << "] = " << ConvertShapeToString(fShapeInput) << ";\n";
0107
0108       out << "int inputLength = " << ConvertShapeToLength(fShapeInput) << ";\n";
0109       out << "int s = 1;\n";
0110       // loop from inverse dim order
0111       out << "for (int i = " << fShapeInput.size()-1 << "; i >=0; i--) {\n";
0112       out << SP << "int r = tensor_" << fNRepeats << "[i];\n";
0113       // we cannot exclude case where repeats=1 since we need offset
0114       //out << SP << "if (r == 1 && i < " << fShapeInput.size()-1 <<  ") continue;\n";
0115       out << SP << "int i_offset = 0, o_offset = 0;\n";
0116       out << SP << "s = s * input_shape[i];\n";
0117       // case we have first copy
0118       out << SP << "if (i == " << fShapeInput.size()-1 <<  ") {\n";
0119       out << SP << SP <<  "for (int j = 0; j < inputLength/s ; j++) {\n";
0120       out << SP << SP << SP << "for (int k = 0; k < r ; k++) {\n";
0121       out << SP << SP << SP << SP << "std::copy(" << input << "+ i_offset, "
0122                                     << input << "+ i_offset + s, " << output << "+ o_offset);\n";
0123       out << SP << SP << SP << SP << "o_offset += s;\n";
0124       out << SP << SP << SP << "}\n"; // end k loop
0125       out << SP << SP << SP << "i_offset += s;\n";
0126       out << SP << SP << "}\n"; // end j loop
0127       out << SP << "} else {\n";  // second copy we do from output to output
0128       // and we need to loop on j from reverse order to avoir re-writing in output tensor
0129       out << SP << SP << "for (int j = inputLength/s - 1 ; j>=0; j--) {\n";
0130       out << SP << SP << SP << "o_offset = j*s*r;\n";
0131       out << SP << SP << SP << "i_offset = j*s;\n";
0132       out << SP << SP << SP << "for (int k = 0; k < r ; k++) {\n";
0133       out << SP << SP << SP << SP << "std::copy(" << output << "+ i_offset, "
0134                                     << output << "+ i_offset + s, " << output << "+ o_offset);\n";
0135       out << SP << SP << SP << SP << "o_offset += s;\n";
0136       out << SP << SP << SP << "}\n"; // end k loop
0137       out << SP << SP << "}\n"; // end j loop
0138       out << SP << "}\n"; // end if
0139       out << SP << "s *= r;\n";
0140       out << SP << "inputLength *= r;\n";
0141       out << "}\n"; // end i loop
0142       out << "}\n";  // end of scope
0143       return out.str();
0144    }
0145 };
0146
0147 }//SOFIE
0148 }//Experimental
0149 }//TMVA
0150
0151 #endif //TMVA_SOFIE_ROPERATOR_Tile