ROperator_RNN.hxx

Go to the documentation of this file.

#ifndef TMVA_SOFIE_ROPERATOR_RNN
#define TMVA_SOFIE_ROPERATOR_RNN
 
#include "TMVA/RModel.hxx"
#include "TMVA/ROperator.hxx"
#include "TMVA/SOFIE_common.hxx"
 
#include <memory>
#include <sstream>
#include <vector>
 
namespace TMVA::Experimental::SOFIE {
 
/*! \brief Recurrent Neural Network operator
 *
 * Inference code generation for one-layer vanilla RNN. Supports forward, reverse and bidirectional RNNs.
 * See the <a href="https://github.com/onnx/onnx/blob/master/docs/Operators.md#RNN">ONNX documentation</a>
 * for details about the supported RNN architectures.
 */
template <typename T> class ROperator_RNN final : public ROperator {
 private:
   std::vector<float> fAttrActivationAlpha;   ///< Scaling values used by some activation functions
   std::vector<float> fAttrActivationBeta;    ///< Scaling values used by some activation functions
   std::vector<std::string> fAttrActivations; ///< Activation functions
   float fAttrClip;                           ///< Clip threshold
   std::string fAttrDirection;                ///< Direction of processing
   size_t fAttrHiddenSize;                    ///< Number of the hidden layers
   size_t fAttrLayout;                        ///< Data layout
 
   std::string fNX;                           ///< Name of the input
   std::string fNW;                           ///< Name of the weights
   std::string fNR;                           ///< Name of the recurrence
   std::string fNB;                           ///< Name of the bias
   std::string fNSequence_lens;               ///< Name of the length of the sequences
   std::string fNInitial_h;                   ///< Name of the initial value of the hidden states
   std::string fNY;                           ///< Name of the output
   std::string fNY_h;                         ///< Name of the last sequence of the output
 
   std::vector<size_t> fShapeX;               ///< Shape of the input
   std::vector<size_t> fShapeW;               ///< Shape of the weights
   std::vector<size_t> fShapeR;               ///< Shape of the recurrence
   std::vector<size_t> fShapeB;               ///< Shape of the bias
   std::vector<size_t> fShapeSequence_lens;   ///< Shape of the length of the sequences
   std::vector<size_t> fShapeInitial_h;       ///< Shape of the initial value of the hidden states
   std::vector<size_t> fShapeY;               ///< Shape of the output
   std::vector<size_t> fShapeY_h;             ///< Shape of the last sequence of the output
 
   std::string fType; ///< Type of the tensors
 
 public:
   /*! Default constructor of ROperator_RNN */
   ROperator_RNN() {}
 
   /*! \brief Constructor of ROperator_RNN from the attributes
    *
    * \param activation_alpha scaling values used by some activation functions
    * \param activation_beta scaling values used by some activation functions
    * \param activations activation functions
    * \param clip clip threshold
    * \param direction direction of processing of the sequneces
    * \param hidden_size number of hidden layers
    * \param layout data layout
    * \param nameX name of the input tensor
    * \param nameW name of the weight tensor
    * \param nameR name of the recurrence tensor
    * \param nameB name of the bias tensor
    * \param nameSequence_lens name of the length of the sequences
    * \param nameInitial_h name of the initial value of the hidden states
    * \param nameY name of the output
    * \param nameY_h name of the last sequence of the output
    */
   ROperator_RNN(std::vector<float> activation_alpha,
                 std::vector<float> activation_beta,
                 std::vector<std::string> activations, float clip,
                 std::string direction, size_t hidden_size, size_t layout,
                 std::string nameX, std::string nameW, std::string nameR,
                 std::string nameB, std::string nameSequence_lens,
                 std::string nameInitial_h, std::string nameY,
                 std::string nameY_h)
       : fAttrActivationAlpha(activation_alpha),
         fAttrActivationBeta(activation_beta), fAttrActivations(activations),
         fAttrClip(clip), fAttrDirection(direction),
         fAttrHiddenSize(hidden_size), fAttrLayout(layout),
         fNX(UTILITY::Clean_name(nameX)), fNW(UTILITY::Clean_name(nameW)),
         fNR(UTILITY::Clean_name(nameR)), fNB(UTILITY::Clean_name(nameB)),
         fNSequence_lens(UTILITY::Clean_name(nameSequence_lens)),
         fNInitial_h(UTILITY::Clean_name(nameInitial_h)),
         fNY(UTILITY::Clean_name(nameY)), fNY_h(UTILITY::Clean_name(nameY_h)) {
      if (std::is_same<T, float>::value) {
         fType = "float";
      } else {
         throw std::runtime_error(
             "TMVA SOFIE Encountered unsupported type parsing a RNN operator");
      }
 
      fInputTensorNames = { fNX, fNW, fNR };
      if(!fNB.empty()){
         fInputTensorNames.emplace_back(fNB);
      }
      if(!fNSequence_lens.empty()){
         fInputTensorNames.emplace_back(fNSequence_lens);
      }
      if(!fNInitial_h.empty()){
         fInputTensorNames.emplace_back(fNInitial_h);
      }
 
      fOutputTensorNames = { };
      if(!fNY.empty()){
         fOutputTensorNames.emplace_back(fNY);
      }
      if(!fNY_h.empty()){
         fOutputTensorNames.emplace_back(fNY_h);
      }
   }
 
   /*! \brief Infers the type of the output tensors
    *
    * \param input type of the input tensors
    */
   std::vector<ETensorType> TypeInference(std::vector<ETensorType> input) override;
 
   /*! \brief Infers the shape of the output tensors
    *
    * \param input shape of the input tensors
    */
   std::vector<std::vector<size_t>>
   ShapeInference(std::vector<std::vector<size_t>> input) override;
 
   /*! \brief Initialize the model
    *
    * \param model Model
    */
   void Initialize(RModel &) override;
 
   /*! \brief Generates the inference code
    *
    * \param OpName name of the operator
    */
   std::string Generate(std::string OpName) override;
 
   // generate code for Session data members (e.g. internal vectors)
   std::string GenerateSessionMembersCode(std::string opName) override;
 
   /*! \brief Returns the blas routines needed to compile the generated code
    */
   std::vector<std::string> GetBlasRoutines() override { return { std::string("Gemm"), std::string("Axpy") }; }
};
 
template <typename T>
auto ROperator_RNN<T>::TypeInference(std::vector<ETensorType> input) -> std::vector<ETensorType>
{
   ETensorType out = input[0];
   return {out, out};
}
 
template <typename T>
auto ROperator_RNN<T>::ShapeInference(std::vector<std::vector<size_t>> input) -> std::vector<std::vector<size_t>>
{
   size_t num_directions = input[1][0];
   size_t hidden_size = input[1][1];
   if (fAttrLayout == 0) {
      size_t seq_length = input[0][0];
      size_t batch_size = input[0][1];
      std::vector<std::vector<size_t>> ret(
         {{seq_length, num_directions, batch_size, hidden_size}, {num_directions, batch_size, hidden_size}});
      return ret;
   } else {
      size_t batch_size = input[0][0];
      size_t seq_length = input[0][1];
      std::vector<std::vector<size_t>> ret(
         {{batch_size, seq_length, num_directions, hidden_size}, {batch_size, num_directions, hidden_size}});
      return ret;
   }
}
 
template <typename T>
auto ROperator_RNN<T>::Initialize(RModel &model) -> void
{
   fUseSession = model.UseSession();
   // Check the input and output tensors
   if (!model.CheckIfTensorAlreadyExist(fNX)) {
      throw std::runtime_error("TMVA SOFIE RNN Op input tensor " + fNX + "  is not found in model.");
   }
   fShapeX = model.GetTensorShape(fNX);
   if (fShapeX.size() != 3) {
      throw std::runtime_error("TMVA SOFIE RNN Op input tensor " + fNX + " is not of 3 dimensions.");
   }
   if (!model.CheckIfTensorAlreadyExist(fNW)) {
      throw std::runtime_error("TMVA SOFIE RNN Op input tensor " + fNW + "  is not found in model.");
   }
   fShapeW = model.GetTensorShape(fNW);
   if (fShapeW.size() != 3) {
      throw std::runtime_error("TMVA SOFIE RNN Op input tensor " + fNW + " is not of 3 dimensions.");
   }
   if (!model.CheckIfTensorAlreadyExist(fNR)) {
      throw std::runtime_error("TMVA SOFIE RNN Op input tensor " + fNR + "  is not found in model.");
   }
   fShapeR = model.GetTensorShape(fNR);
   if (fShapeR.size() != 3) {
      throw std::runtime_error("TMVA SOFIE RNN Op input tensor " + fNR + " is not of 3 dimensions.");
   }
   if (!fNB.empty()) {
      if (!model.CheckIfTensorAlreadyExist(fNB)) {
         throw std::runtime_error("TMVA SOFIE RNN op input tensor " + fNB + " is not  found in model.");
      }
      fShapeB = model.GetTensorShape(fNB);
      if (fShapeB.size() != 2 && fShapeB.size() != 4) {
         throw std::runtime_error("TMVA SOFIE RNN op input tensor " + fNB + " is not of 2 or 4 dimensions.");
      }
      if (fShapeB.size() == 2) {
         // Broadcasting the bias
         auto original_data = model.GetInitializedTensorData(fNB);
         size_t num_directions = fShapeW[0];
         size_t seq_length = (fAttrLayout == 0) ? fShapeX[0] : fShapeX[1];
         size_t batch_size = (fAttrLayout == 0) ? fShapeX[1] : fShapeX[0];
         if (fType == "float") {
            float *original_bias = static_cast<float *>(original_data.get());
            float *new_bias = new float[num_directions * seq_length * batch_size * fAttrHiddenSize];
            std::vector<float> sum(fAttrHiddenSize);
            for (size_t direction = 0; direction < num_directions; direction++) {
               for (size_t h = 0; h < fAttrHiddenSize; h++) {
                  sum[h] = original_bias[direction * 2 * fAttrHiddenSize + h] +
                           original_bias[(2 * direction + 1) * fAttrHiddenSize + h];
               }
               for (size_t seq = 0; seq < seq_length; seq++) {
                  for (size_t batch = 0; batch < batch_size; batch++) {
                     size_t bias_offset = direction * seq_length * batch_size * fAttrHiddenSize +
                                          seq * batch_size * fAttrHiddenSize + batch * fAttrHiddenSize;
                     std::copy(sum.begin(), sum.end(), new_bias + bias_offset);
                  }
               }
            }
            std::vector<size_t> new_bias_shape = {num_directions, seq_length, batch_size, fAttrHiddenSize};
            std::shared_ptr<void> new_bias_ptr(new_bias, std::default_delete<float[]>());
            model.UpdateInitializedTensor(fNB, model.GetTensorType(fNB), new_bias_shape, new_bias_ptr);
            fShapeB = model.GetTensorShape(fNB);
         }
      }
   }
   if (!fNSequence_lens.empty()) {
      if (!model.CheckIfTensorAlreadyExist(fNSequence_lens)) {
         throw std::runtime_error("TMVA SOFIE RNN Op input tensor " + fNSequence_lens + "is not found in model.");
      }
      fShapeSequence_lens = model.GetTensorShape(fNSequence_lens);
      if (fShapeSequence_lens.size() != 1) {
         throw std::runtime_error("TMVA SOFIE RNN Op input tensor " + fNSequence_lens + " is not of 1 dimension.");
      }
   }
   if (!fNInitial_h.empty()) {
      if (!model.CheckIfTensorAlreadyExist(fNInitial_h)) {
         throw std::runtime_error("TMVA SOFIE RNN Op input tensor " + fNInitial_h + " is not found in model.");
      }
      fShapeInitial_h = model.GetTensorShape(fNInitial_h);
      if (fShapeInitial_h.size() != 3) {
         throw std::runtime_error("TMVA SOFIE RNN Op input tensor " + fNInitial_h + " is not of 3 dimensions.");
      }
   }
   if (!fNY.empty()) {
      fShapeY = ShapeInference({fShapeX, fShapeW})[0];
      if (!model.CheckIfTensorAlreadyExist(fNY)) {
         model.AddIntermediateTensor(fNY, model.GetTensorType(fNX), fShapeY);
      }
   }
   if (!fNY_h.empty()) {
      fShapeY_h = ShapeInference({fShapeX, fShapeW})[1];
      if (!model.CheckIfTensorAlreadyExist(fNY_h)) {
         model.AddIntermediateTensor(fNY_h, model.GetTensorType(fNX), fShapeY_h);
      }
   }
   // Check the attributes
   for (auto &activation : fAttrActivations) {
      if (activation != "Relu" && activation != "Tanh" && activation != "Sigmoid" && activation != "Affine" &&
          activation != "LeakyRelu" && activation != "ThresholdRelu" && activation != "ScaledTanh" &&
          activation != "HardSigmoid" && activation != "Elu" && activation != "Softsign" && activation != "Softplus") {
         throw std::runtime_error("TMVA SOFIE - Activation function " + activation + " not implemented");
      }
   }
   if (fAttrDirection != "forward" && fAttrDirection != "backward" && fAttrDirection != "bidirectional") {
      throw std::runtime_error("TMVA SOFIE - Invalid RNN direction fAttrDirection = " + fAttrDirection);
   }
   if (fAttrHiddenSize != fShapeW[1]) {
      throw std::runtime_error("TMVA SOFIE - fAttrHiddenSize must be equal to " + std::to_string(fShapeW[1]));
   }
   if (fAttrLayout > 1) {
      throw std::runtime_error("TMVA SOFIE - Layout fAttrLayout = " + std::to_string(fAttrLayout) +
                               " must be 0 (timewise) or 1 (batchwise)");
   }
   if (fAttrActivations.empty()) {
      if (fAttrDirection == "bidirectional") {
         fAttrActivations = {"Tanh", "Tanh"};
      } else {
         fAttrActivations = {"Tanh"};
      }
   }
   // Add needed standard library headers
   model.AddNeededStdLib("cmath");
}
 
// generate code for Session data members (e.g. internal vectors)
template <typename T>
std::string ROperator_RNN<T>::GenerateSessionMembersCode(std::string opName)
{
   opName = "op_" + opName;
   std::stringstream out;
 
   size_t num_directions = fShapeW[0];
   size_t seq_length = (fAttrLayout == 0) ? fShapeX[0] : fShapeX[1];
   size_t batch_size = (fAttrLayout == 0) ? fShapeX[1] : fShapeX[0];
   size_t input_size = fShapeX[2];
 
   struct Block {
      std::string name;
      size_t size;
   };
 
   std::vector<Block> blocks;
 
   if (fAttrLayout != 0) {
      blocks.push_back({"input", seq_length * batch_size * input_size});
      blocks.push_back({"initial_hidden_state", num_directions * batch_size * fAttrHiddenSize});
   }
   blocks.push_back({"feedforward", seq_length * batch_size * fAttrHiddenSize});
   if (fAttrLayout != 0 || fNY.empty()) {
      blocks.push_back({"hidden_state", seq_length * num_directions * batch_size * fAttrHiddenSize});
   }
 
   // Compute total size
   size_t total_size = 0;
   for (const auto &b : blocks) {
      total_size += b.size;
   }
 
   // Emit backing storage
   out << "std::vector<" << fType << "> fVec_" << opName << "_buffer = std::vector<" << fType << ">(" << total_size
       << ");\n";
 
   // Emit pointers
   std::size_t offset = 0;
   for (const auto &b : blocks) {
      out << fType << "* fVec_" << opName << "_" << b.name << " = fVec_" << opName << "_buffer.data() + " << offset
          << ";\n";
      offset += b.size;
   }
 
   out << "\n";
 
   return out.str();
}
 
//////////////////////////////////////////////////////////////////////////////////////////////////
template <typename T>
auto ROperator_RNN<T>::Generate(std::string OpName) -> std::string
{
   OpName = "op_" + OpName;
   std::stringstream out;
 
   size_t seq_length = (fAttrLayout == 0) ? fShapeX[0] : fShapeX[1];
   size_t batch_size = (fAttrLayout == 0) ? fShapeX[1] : fShapeX[0];
   size_t input_size = fShapeX[2];
   size_t num_directions = fShapeW[0];
 
   // set the input
   if (fAttrLayout == 0) {
      if (fType == "float") {
         out << SP << "float const*" << OpName << "_input = tensor_" << fNX << ";\n";
      }
   } else {
      if (fUseSession)
         out << SP << fType << " * " << OpName << "_input = this->fVec_" << OpName << "_input;\n";
      else
         out << SP << fType << " " << OpName << "_input[" << seq_length * batch_size * input_size << "];\n";
      out << SP << "for(size_t seq = 0; seq < " << seq_length << "; seq++) {\n";
      out << SP << SP << "for(size_t batch = 0; batch < " << batch_size << "; batch++) {\n";
      out << SP << SP << SP << "for(size_t i = 0; i < " << input_size << "; i++) {\n";
      out << SP << SP << SP << SP << OpName << "_input[seq * " << batch_size * input_size << " + batch * " << input_size
          << " + i] = " << "tensor_" << fNX << "[batch * " << seq_length * input_size << " + seq * " << input_size
          << " + i];\n";
      out << SP << SP << SP << "}\n";
      out << SP << SP << "}\n";
      out << SP << "}\n";
   }
 
   // Set the initial hidden state
   if (!fNInitial_h.empty()) {
      if (fAttrLayout == 0) {
         out << SP << fType << " *" << OpName << "_initial_hidden_state = " << " tensor_" << fNInitial_h << ";\n";
      } else {
         if (fUseSession)
            out << SP << fType << " * " << OpName << "_initial_hidden_state = this->fVec_" << OpName
                << "_initial_hidden_state;\n";
         else
            out << fType << " " << OpName << "_initial_hidden_state[" << num_directions * batch_size * fAttrHiddenSize
                << "] = {0};\n";
 
         for (size_t direction = 0; direction < num_directions; direction++) {
            out << SP << "for(size_t batch = 0; batch < " << batch_size << "; batch++) {\n";
            out << SP << SP << "for(size_t h = 0; h < " << fAttrHiddenSize << "; h++) {\n";
            out << SP << SP << SP << OpName << "_initial_hidden_state[" << direction * batch_size * fAttrHiddenSize
                << " + batch * " << fAttrHiddenSize << " + h] = tensor_" << fNInitial_h << "[batch * "
                << num_directions * fAttrHiddenSize << " + " << direction * fAttrHiddenSize << " + h];\n";
            out << SP << SP << "}\n";
            out << SP << "}\n";
         }
      }
   }
 
   if (fUseSession)
      out << SP << fType << " * " << OpName << "_feedforward = this->fVec_" << OpName << "_feedforward;\n";
   else
      out << SP << fType << " " << OpName << "_feedforward[" << seq_length * batch_size * fAttrHiddenSize
          << "] = {0};\n";
 
   // Set the hidden state
   if (fAttrLayout == 0 && !fNY.empty()) {
      out << SP << fType << " *" << OpName << "_hidden_state = tensor_" << fNY << ";\n";
   } else {
      if (fUseSession)
         out << SP << fType << " * " << OpName << "_hidden_state = this->fVec_" << OpName << "_hidden_state;\n";
      else
         out << SP << fType << " " << OpName << "_hidden_state["
             << seq_length * num_directions * batch_size * fAttrHiddenSize << "] = {0};\n";
   }
 
   out << SP << "char " << OpName << "_transA = 'N';\n";
   out << SP << "char " << OpName << "_transB = 'T';\n";
   out << SP << "int " << OpName << "_m = " << seq_length * batch_size << ";\n";
   out << SP << "int " << OpName << "_n = " << fAttrHiddenSize << ";\n";
   out << SP << "int " << OpName << "_k = " << input_size << ";\n";
   if (fType == "float") {
      out << SP << "float " << OpName << "_alpha = 1.;\n";
      out << SP << "float " << OpName << "_beta = .0;\n";
   }
   if (!fNB.empty()) {
      out << SP << "int " << OpName << "_bias_size = " << seq_length * batch_size * fAttrHiddenSize << ";\n";
      out << SP << "int " << OpName << "_incx = 1;\n";
      out << SP << "int " << OpName << "_incy = 1;\n";
   }
 
   for (size_t direction = 0; direction < num_directions; direction++) {
      // feedforward = input * W^T + bias
      if (fType == "float") {
         if (direction == 0) {
            out << SP << "BLAS::sgemm_(&" << OpName << "_transB, &" << OpName << "_transA, &" << OpName << "_n, &"
                << OpName << "_m, &" << OpName << "_k, &" << OpName << "_alpha, tensor_" << fNW << ", &" << OpName
                << "_k, " << OpName << "_input, &" << OpName << "_k, &" << OpName << "_beta, " << OpName
                << "_feedforward, &" << OpName << "_n);\n";
         } else {
            out << SP << "size_t " << OpName << "_w_offset = " << fAttrHiddenSize * input_size << ";\n";
            out << SP << "BLAS::sgemm_(&" << OpName << "_transB, &" << OpName << "_transA, &" << OpName << "_n, &"
                << OpName << "_m, &" << OpName << "_k, &" << OpName << "_alpha, tensor_" << fNW << " + " << OpName
                << "_w_offset, &" << OpName << "_k, " << OpName << "_input, &" << OpName << "_k, &" << OpName
                << "_beta, " << OpName << "_feedforward, &" << OpName << "_n);\n";
         }
      }
      // Add the bias
      if (!fNB.empty()) {
         if (fType == "float") {
            if (direction == 0) {
               out << SP << "BLAS::saxpy_(&" << OpName << "_bias_size, &" << OpName << "_alpha, tensor_" << fNB << ", &"
                   << OpName << "_incx, " << OpName << "_feedforward, &" << OpName << "_incy);\n";
            } else {
               out << SP << "size_t " << OpName << "_bias_offset = " << seq_length * batch_size * fAttrHiddenSize
                   << ";\n";
               out << SP << "BLAS::saxpy_(&" << OpName << "_bias_size, &" << OpName << "_alpha, tensor_" << fNB << " + "
                   << OpName << "_bias_offset, &" << OpName << "_incx, " << OpName << "_feedforward, &" << OpName
                   << "_incy);\n";
            }
         }
      }
 
      // Copy feedforward into hidden state
      out << SP << "for (size_t seq = 0; seq < " << seq_length << "; seq++) {\n";
      out << SP << SP << "size_t offset = seq * " << batch_size * fAttrHiddenSize << ";\n";
      out << SP << SP << "size_t size = " << batch_size * fAttrHiddenSize << ";\n";
      out << SP << SP << "size_t h_offset = seq * " << num_directions * batch_size * fAttrHiddenSize << " + "
          << direction * batch_size * fAttrHiddenSize << ";\n";
      out << SP << SP << "std::copy(" << OpName << "_feedforward + offset, " << OpName
          << "_feedforward + offset + size, " << OpName << "_hidden_state + h_offset);\n";
      out << SP << "}\n";
 
      out << SP << "for (size_t seq = 0; seq < " << seq_length << "; seq++) {\n";
      if (fAttrDirection == "backward" || direction == 1) {
         out << SP << SP << "size_t index = " << seq_length - 1 << " - seq;\n";
      } else {
         out << SP << SP << "size_t index = seq;\n";
      }
 
      out << SP << SP << "int m2 = " << batch_size << ";\n";
      out << SP << SP << "size_t offset = index * " << num_directions * batch_size * fAttrHiddenSize << " + "
          << direction * batch_size * fAttrHiddenSize << ";\n";
      out << SP << SP << "size_t size = " << batch_size * fAttrHiddenSize << ";\n";
      out << SP << SP << "if (seq == 0) {\n";
      if (!fNInitial_h.empty()) {
         // hidden_state = hidden_state + initial_hidden_state * R^T
         out << SP << SP << SP << "size_t r_offset = " << direction * fAttrHiddenSize * fAttrHiddenSize << ";\n";
         out << SP << SP << SP << "size_t initial_hidden_state_offset = " << direction * batch_size * fAttrHiddenSize
             << ";\n";
         if (fType == "float") {
            out << SP << SP << SP << "BLAS::sgemm_(&" << OpName << "_transB, &" << OpName << "_transA, &" << OpName
                << "_n, &m2, &" << OpName << "_n, &" << OpName << "_alpha, tensor_" << fNR << " + r_offset, &" << OpName
                << "_n, " << OpName << "_initial_hidden_state + initial_hidden_state_offset, &" << OpName << "_n, &"
                << OpName << "_alpha, " << OpName << "_hidden_state + offset, &" << OpName << "_n);\n";
         }
      }
      out << SP << SP << "} else {\n";
      // hidden_state = hidden_state + previous_hidden_state * R^T
      out << SP << SP << SP << "size_t r_offset = " << direction * fAttrHiddenSize * fAttrHiddenSize << ";\n";
      if (fAttrDirection == "backward" || direction == 1) {
         out << SP << SP << SP << "size_t previous_offset = (index + 1) * "
             << num_directions * batch_size * fAttrHiddenSize << " + " << direction * batch_size * fAttrHiddenSize
             << ";\n";
      } else {
         out << SP << SP << SP << "size_t previous_offset = (seq - 1) * "
             << num_directions * batch_size * fAttrHiddenSize << " + " << direction * batch_size * fAttrHiddenSize
             << ";\n";
      }
      if (fType == "float") {
         out << SP << SP << SP << "BLAS::sgemm_(&" << OpName << "_transB, &" << OpName << "_transA, &" << OpName
             << "_n, &m2, &" << OpName << "_n, &" << OpName << "_alpha, tensor_" << fNR << " + r_offset, &" << OpName
             << "_n, " << OpName << "_hidden_state + previous_offset, &" << OpName << "_n, &" << OpName << "_alpha, "
             << OpName << "_hidden_state + offset, &" << OpName << "_n);\n";
      }
      out << SP << SP << "}\n";
 
      // Clip the elements of the hidden state into the range [-fAttrClip, fAttrClip]
      if (fAttrClip > .0) {
         out << SP << SP << "for (size_t i = offset; i < offset + size; i++) {\n";
         if (fType == "float") {
            out << SP << SP << SP << "float x = (" << OpName << "_hidden_state[i] > " << -fAttrClip << ") ? " << OpName
                << "_hidden_state[i] : " << -fAttrClip << ";\n";
         }
         out << SP << SP << SP << OpName << "_hidden_state[i] = (x < " << fAttrClip << ") ? x : " << fAttrClip << ";\n";
         out << SP << SP << "}\n";
      }
 
      // Apply the activation function to the hidden state
      if (fAttrActivations[direction] == "Relu") {
         out << SP << SP << "for (size_t i = offset; i < offset + size; i++) {\n";
         out << SP << SP << SP << "if (" << OpName << "_hidden_state[i] < 0.)\n";
         out << SP << SP << SP << SP << OpName << "_hidden_state[i] = 0.;\n";
         out << SP << SP << "}\n";
      } else if (fAttrActivations[direction] == "Tanh") {
         out << SP << SP << "for (size_t i = offset; i < offset + size; i++) {\n";
         if (fType == "float") {
            out << SP << SP << SP << "float ex = std::exp(-2 * " << OpName << "_hidden_state[i]);\n";
         }
         out << SP << SP << SP << SP << OpName << "_hidden_state[i] = (1. - ex) / (1. + ex);\n";
         out << SP << SP << "}\n";
      } else if (fAttrActivations[direction] == "Sigmoid") {
         out << SP << SP << "for (size_t i = offset; i < offset + size; i++) {\n";
         out << SP << SP << SP << SP << OpName << "_hidden_state[i] = 1. / (1. + std::exp(-" << OpName
             << "_hidden_state[i]));\n";
         out << SP << SP << "}\n";
      } else if (fAttrActivations[direction] == "Affine") {
         out << SP << SP << "for (size_t i = offset; i < offset + size; i++) {\n";
         out << SP << SP << SP << SP << OpName << "_hidden_state[i] = " << fAttrActivationAlpha[direction] << " * "
             << OpName << "_hidden_state[i] + " << fAttrActivationBeta[direction] << ";\n";
         out << SP << SP << "}\n";
      } else if (fAttrActivations[direction] == "ScaledTanh") {
         out << SP << SP << "for (size_t i = offset; i < offset + size; i++) {\n";
         if (fType == "float") {
            out << SP << SP << SP << "float ex = std::exp(-2 * " << fAttrActivationBeta[direction] << " * " << OpName
                << "_hidden_state[i]);\n";
         }
         out << SP << SP << SP << SP << OpName << "_hidden_state[i] = " << fAttrActivationAlpha[direction]
             << " * (1. - ex) / (1. + ex);\n";
         out << SP << SP << "}\n";
      } else if (fAttrActivations[direction] == "HardSigmoid") {
         out << SP << SP << "for (size_t i = offset; i < offset + size; i++) {\n";
         if (fType == "float") {
            out << SP << SP << SP << "float a = " << fAttrActivationAlpha[direction] << " * " << OpName
                << "_hidden_state[i] + " << fAttrActivationBeta[direction] << ";\n";
            out << SP << SP << SP << "float b = (a > 0.) ? a : 0.;\n";
         }
         out << SP << SP << SP << SP << OpName << "_hidden_state[i] = (b < 1.) ? b : 1.;\n";
         out << SP << SP << "}\n";
      } else if (fAttrActivations[direction] == "LeakyRelu") {
         out << SP << SP << "for (size_t i = offset; i < offset + size; i++) {\n";
         out << SP << SP << SP << "if (" << OpName << "_hidden_state[i] < 0.)\n";
         out << SP << SP << SP << SP << OpName << "_hidden_state[i] = " << fAttrActivationAlpha[direction] << " * "
             << OpName << "_hidden_state[i];\n";
         out << SP << SP << "}\n";
      } else if (fAttrActivations[direction] == "ThresholdRelu") {
         out << SP << SP << "for (size_t i = offset; i < offset + size; i++) {\n";
         out << SP << SP << SP << "if (" << OpName << "_hidden_state[i] < " << fAttrActivationAlpha[direction] << ")\n";
         out << SP << SP << SP << SP << OpName << "_hidden_state[i] = 0.;\n";
         out << SP << SP << "}";
      } else if (fAttrActivations[direction] == "Elu") {
         out << SP << SP << "for (size_t i = offset; i < offset + size; i++) {\n";
         out << SP << SP << SP << "if (" << OpName << "_hidden_state[i] < 0.)\n";
         out << SP << SP << SP << SP << OpName << "_hidden_state[i] = " << fAttrActivationAlpha[direction]
             << " * std::exp(" << OpName << "_hidden_state[i] - 1.);\n";
         out << SP << SP << "}\n";
      } else if (fAttrActivations[direction] == "Softsign") {
         out << SP << SP << "for (size_t i = offset; i < offset + size; i++) {\n";
         out << SP << SP << SP << SP << OpName << "_hidden_state[i] = " << OpName << "_hidden_state[i] / (1. + abs("
             << OpName << "_hidden_state[i]));\n";
         out << SP << SP << "}\n";
      } else { // fAttrActivations[direction] = Softplus
         out << SP << SP << "for (size_t i = offset; i < offset + size; i++) {\n";
         out << SP << SP << SP << SP << OpName << "_hidden_state[i] = log(1. + std::exp(" << OpName
             << "_hidden_state[i]));\n";
         out << SP << SP << "}\n";
         out << SP << "}\n";
      }
      out << SP << "}\n";
   }
 
   // Padding the hidden state for RNN with different sequence lengths
   if (!fNSequence_lens.empty()) {
      out << SP << "for (size_t seq = 0; seq < " << seq_length << "; seq++) {\n";
      out << SP << SP << "for (size_t batch = 0; batch < " << batch_size << "; batch++) {\n";
      out << SP << SP << SP << "if (seq >= tensor_" << fNSequence_lens << "[batch]) {\n";
      out << SP << SP << SP << SP << "for (size_t h = 0; h < " << fAttrHiddenSize << "; h++) {\n";
      if (num_directions == 1) {
         out << SP << SP << SP << SP << SP << OpName << "_hidden_state[seq * "
             << num_directions * batch_size * fAttrHiddenSize << " + batch * " << fAttrHiddenSize << " + h] = 0.;\n";
      } else {
         out << SP << SP << SP << SP << SP << OpName << "_hidden_state[seq * "
             << num_directions * batch_size * fAttrHiddenSize << " + batch * " << fAttrHiddenSize << " + h] = 0.;\n";
         out << SP << SP << SP << SP << SP << OpName << "_hidden_state[seq * "
             << num_directions * batch_size * fAttrHiddenSize << " + " << batch_size * fAttrHiddenSize << " + batch * "
             << fAttrHiddenSize << " + h] = 0.;\n";
      }
      out << SP << SP << SP << SP << "}\n";
      out << SP << SP << SP << "}\n";
      out << SP << SP << "}\n";
      out << SP << "}\n";
   }
 
   // Copy the hidden state into y and y_h
   if (fAttrLayout == 0) {
      if (!fNY_h.empty()) {
         if (fNSequence_lens.empty()) {
            size_t yh_size = batch_size * fAttrHiddenSize;
            if (fAttrDirection == "backward") {
               out << SP << "std::copy(" << OpName << "_hidden_state, " << OpName << "_hidden_state + " << yh_size
                   << ", tensor_" << fNY_h << ");\n";
            } else {
               size_t offset = (seq_length - 1) * num_directions * batch_size * fAttrHiddenSize;
               out << SP << "std::copy(" << OpName << "_hidden_state + " << offset << ", " << OpName
                   << "_hidden_state + " << offset << " + " << yh_size << ", tensor_" << fNY_h << ");\n";
            }
            if (num_directions == 2) {
               out << SP << "std::copy(" << OpName << "_hidden_state + " << yh_size << ", " << OpName
                   << "_hidden_state + " << 2 * yh_size << ", tensor_" << fNY_h << " + " << yh_size << ");\n";
            }
         } else { // RNN with different sequence lengths
            if (fAttrDirection == "backward") {
               out << SP << "for (size_t batch = 0; batch < " << batch_size << "; batch++) {\n";
               out << SP << SP << "size_t offset = batch * " << fAttrHiddenSize << ";\n";
               out << SP << SP << "std::copy(" << OpName << "_hidden_state + offset, " << OpName
                   << "_hidden_state + offset + " << fAttrHiddenSize << ", tensor_" << fNY_h << " + offset);\n";
               out << SP << "}\n";
            } else {
               out << SP << "for (size_t batch = 0; batch < " << batch_size << "; batch++) {\n";
               out << SP << SP << "size_t seq = " << "tensor_" << fNSequence_lens << "[batch] - 1;\n";
               out << SP << SP << "size_t offset = seq * " << num_directions * batch_size * fAttrHiddenSize
                   << " + batch * " << fAttrHiddenSize << ";\n";
               out << SP << SP << "size_t yh_offset = batch * " << fAttrHiddenSize << ";\n";
               out << SP << SP << "std::copy(" << OpName << "_hidden_state + offset, " << OpName
                   << "_hidden_state + offset + " << fAttrHiddenSize << ", tensor_" << fNY_h << " + yh_offset);\n";
               out << SP << "}\n";
            }
            if (num_directions == 2) {
               out << SP << "for (size_t batch = 0; batch < " << batch_size << "; batch++) {\n";
               out << SP << SP << "size_t offset = " << batch_size * fAttrHiddenSize << " + batch * " << fAttrHiddenSize
                   << ";\n";
               out << SP << SP << "size_t yh_offset = " << batch_size * fAttrHiddenSize << " + batch * "
                   << fAttrHiddenSize << ";\n";
               out << SP << SP << "std::copy(" << OpName << "_hidden_state + offset, " << OpName
                   << "_hidden_state + offset + " << fAttrHiddenSize << ", tensor_" << fNY_h << " + yh_offset);\n";
               out << SP << "}\n";
            }
         }
      }
   } else { // fAttrLayout=1
      if (!fNY.empty()) {
         for (size_t direction = 0; direction < num_directions; direction++) {
            out << SP << "for (size_t seq = 0; seq < " << seq_length << "; seq++) {\n";
            out << SP << SP << "for (size_t batch = 0; batch < " << batch_size << "; batch++) {\n";
            out << SP << SP << SP << "size_t offset = seq * " << num_directions * batch_size * fAttrHiddenSize << " + "
                << direction * batch_size * fAttrHiddenSize << " + batch * " << fAttrHiddenSize << ";\n";
            out << SP << SP << SP << "size_t y_offset = batch * " << seq_length * num_directions * fAttrHiddenSize
                << " + seq * " << num_directions * fAttrHiddenSize << " + " << direction * fAttrHiddenSize << ";\n";
            out << SP << SP << SP << "std::copy(" << OpName << "_hidden_state + offset, " << OpName
                << "_hidden_state + offset + " << fAttrHiddenSize << ", tensor_" << fNY << " + y_offset);\n";
            out << SP << SP << "}\n";
            out << SP << "}\n";
         }
      }
      if (!fNY_h.empty()) {
         if (fAttrDirection == "backward") {
            out << SP << "for (size_t batch = 0; batch < " << batch_size << "; batch++) {\n";
            out << SP << SP << "size_t offset = batch * " << fAttrHiddenSize << ";\n";
            out << SP << SP << "size_t yh_offset = batch * " << num_directions * fAttrHiddenSize << ";\n";
            out << SP << SP << "std::copy(" << OpName << "_hidden_state + offset, " << OpName
                << "_hidden_state + offset + " << fAttrHiddenSize << ", tensor_" << fNY_h << " + yh_offset);\n";
            out << SP << "}\n";
         } else {
            out << SP << "for (size_t batch = 0; batch < " << batch_size << "; batch++) {\n";
            if (fNSequence_lens.empty()) {
               out << SP << SP << "size_t seq = " << seq_length - 1 << ";\n";
            } else {
               out << SP << SP << "size_t seq = " << "tensor_" << fNSequence_lens << "[batch] - 1;\n";
            }
            out << SP << SP << "size_t offset = seq * " << num_directions * batch_size * fAttrHiddenSize
                << " + batch * " << fAttrHiddenSize << ";\n";
            out << SP << SP << "size_t yh_offset = batch * " << num_directions * fAttrHiddenSize << ";\n";
            out << SP << SP << "std::copy(" << OpName << "_hidden_state + offset, " << OpName
                << "_hidden_state + offset + " << fAttrHiddenSize << ", tensor_" << fNY_h << " + yh_offset);\n";
            out << SP << "}\n";
         }
         if (num_directions == 2) {
            out << SP << "for (size_t batch = 0; batch < " << batch_size << "; batch++) {\n";
            out << SP << SP << "size_t offset = " << batch_size * fAttrHiddenSize << " + batch * " << fAttrHiddenSize
                << ";\n";
            out << SP << SP << "size_t yh_offset = batch * " << num_directions * fAttrHiddenSize << " + "
                << fAttrHiddenSize << ";\n";
            out << SP << SP << "std::copy(" << OpName << "_hidden_state + offset, " << OpName
                << "_hidden_state + offset + " << fAttrHiddenSize << ", tensor_" << fNY_h << " + yh_offset);\n";
            out << SP << "}\n";
         }
      }
   }
 
   return out.str();
}
 
} // namespace TMVA::Experimental::SOFIE
 
#endif