LutNetworkFixedPointRegression2Outputs.h

Go to the documentation of this file.

/*
 * LutNetworkFixedPointRegression2Outputs.h
 *
 *  Created on: April 13, 2021
 *      Author: Karol Bunkowski, kbunkow@cern.ch
 */

#ifndef L1Trigger_L1TMuonOverlapPhase2_LutNetworkFixedPointRegression2Outputs_h
#define L1Trigger_L1TMuonOverlapPhase2_LutNetworkFixedPointRegression2Outputs_h

#include "L1Trigger/L1TMuonOverlapPhase2/interface/LutNeuronLayerFixedPoint.h"

#include "ap_int.h"

#include "FWCore/MessageLogger/interface/MessageLogger.h"

#include <cmath>

#include <boost/property_tree/ptree.hpp>
#include <boost/property_tree/xml_parser.hpp>

namespace lutNN {

  //_I - number of integer bits in the ap_ufixed, _F - number of fractional bits in the ap_ufixed
  //the network has two outputs, and since each output can have different range, the LUTs in the last layer have different I and F
  template <int input_I,
            int input_F,
            std::size_t inputSize,
            int layer1_lut_I,
            int layer1_lut_F,
            int layer1_neurons,
            int layer1_output_I,  //to the layer1 output the bias is added to make the layer1 input
            int layer2_input_I,
            int layer2_lut_I,
            int layer2_lut_F,
            int layer2_neurons,
            int layer3_input_I,
            int layer3_0_inputCnt,
            int layer3_0_lut_I,
            int layer3_0_lut_F,
            int output0_I,
            int output0_F,
            int layer3_1_inputCnt,
            int layer3_1_lut_I,
            int layer3_1_lut_F,
            int output1_I,
            int output1_F>
  class LutNetworkFixedPointRegression2Outputs : public LutNetworkFixedPointRegressionBase {
  public:
    LutNetworkFixedPointRegression2Outputs() {
      static_assert(layer2_neurons == (layer3_0_inputCnt + layer3_1_inputCnt));

      std::cout << "LutNetworkFixedPoint" << std::endl;
      lutLayer1.setName("lutLayer1");
      lutLayer2.setName("lutLayer2");
      lutLayer3_0.setName("lutLayer3_0");
      lutLayer3_1.setName("lutLayer3_1");
    };

    ~LutNetworkFixedPointRegression2Outputs() override {}

    typedef LutNeuronLayerFixedPoint<input_I, input_F, inputSize, layer1_lut_I, layer1_lut_F, layer1_neurons, layer1_output_I>
        LutLayer1;
    LutLayer1 lutLayer1;

    static constexpr unsigned int noHitCntShift = layer1_output_I;  //FIXME should be layer1_output_I ???

    static constexpr int layer2_input_F = layer1_lut_F;

    typedef LutNeuronLayerFixedPoint<layer2_input_I,
                                     layer2_input_F,
                                     layer1_neurons,
                                     layer2_lut_I,
                                     layer2_lut_F,
                                     layer2_neurons,
                                     layer3_input_I>
        LutLayer2;
    LutLayer2 lutLayer2;

    static constexpr int layer3_input_F = layer2_lut_F;

    typedef LutNeuronLayerFixedPoint<layer3_input_I,
                                     layer3_input_F,
                                     layer3_0_inputCnt,
                                     layer3_0_lut_I,
                                     layer3_0_lut_F,
                                     1,
                                     output0_I>
        LutLayer3_0;
    LutLayer3_0 lutLayer3_0;  //"lutLayer3_0"

    typedef LutNeuronLayerFixedPoint<layer3_input_I,
                                     layer3_input_F,
                                     layer3_1_inputCnt,
                                     layer3_1_lut_I,
                                     layer3_1_lut_F,
                                     1,
                                     output1_I>
        LutLayer3_1;
    LutLayer3_1 lutLayer3_1;  //"lutLayer3_1"

    void runWithInterpolation() {
      lutLayer1.runWithInterpolation(inputArray);
      auto& layer1Out = lutLayer1.getOutWithOffset();

      std::array<ap_ufixed<layer2_input_I + layer2_input_F, layer2_input_I, AP_TRN, AP_SAT>, layer1_neurons>
          layer1OutWithBias;
      for (unsigned int i = 0; i < layer1Out.size(); i++) {
        layer1OutWithBias[i] = layer1Out[i] + layer1Bias;
      }

      lutLayer2.runWithInterpolation(layer1OutWithBias);
      auto& layer2Out = lutLayer2.getOutWithOffset();

      typename LutLayer3_0::inputArrayType lutLayer3_0_input;
      std::copy(layer2Out.begin(), layer2Out.begin() + lutLayer3_0_input.size(), lutLayer3_0_input.begin());

      typename LutLayer3_1::inputArrayType lutLayer3_1_input;
      std::copy(layer2Out.begin() + lutLayer3_0_input.size(), layer2Out.end(), lutLayer3_1_input.begin());

      lutLayer3_0.runWithInterpolation(lutLayer3_0_input);
      lutLayer3_1.runWithInterpolation(lutLayer3_1_input);
    }

    void run(std::vector<float>& inputs, float noHitVal, std::vector<double>& nnResult) override {
      unsigned int noHitsCnt = 0;
      for (unsigned int iInput = 0; iInput < inputs.size(); iInput++) {
        inputArray[iInput] = inputs[iInput];
        if (inputs[iInput] == noHitVal)
          noHitsCnt++;
      }

      unsigned int bias = (noHitsCnt << noHitCntShift);

      //layer1Bias switches the input of the layer2 (i.e. output of the layer1) do different regions in the LUTs
      //depending on the  number of layers without hits
      layer1Bias = bias;

      runWithInterpolation();

      //output0_I goes to the declaration of the lutLayer3_0, but it does not matter, as it is used only for the outputArray
      //auto layer3_0_out = ap_ufixed<output0_I+output0_F, output0_I, AP_RND_CONV, AP_SAT>(lutLayer3_0.getLutOutSum()[0]); //TODO should be AP_RND_CONV rather, but it affect the rate
      //auto layer3_1_out = ap_fixed <output1_I+output1_F, output1_I, AP_RND_CONV, AP_SAT>(lutLayer3_1.getLutOutSum()[0]); //here layer3_0_out has size 1
      auto layer3_0_out = lutLayer3_0.getLutOutSum()[0];  //here layer3_0_out has size 1
      auto layer3_1_out = lutLayer3_1.getLutOutSum()[0];  //here layer3_0_out has size 1

      nnResult[0] = layer3_0_out.to_float();
      nnResult[1] = layer3_1_out.to_float();
      LogTrace("l1tOmtfEventPrint") << "layer3_0_out[0] " << layer3_0_out[0] << " layer3_1_out[0] " << layer3_1_out[0]
                                    << std::endl;
    }

    //pt in the hardware scale, ptGeV = (ptHw -1) / 2
    int getCalibratedHwPt() override {
      auto lutAddr = ap_ufixed<output0_I + output0_F + output0_F, output0_I + output0_F, AP_RND_CONV, AP_SAT>(
          lutLayer3_0.getLutOutSum()[0]);
      lutAddr = lutAddr << output0_F;
      //std::cout<<"lutLayer3_0.getLutOutSum()[0] "<<lutLayer3_0.getLutOutSum()[0]<<" lutAddr.to_uint() "<<lutAddr.to_uint()<<" ptCalibrationArray[lutAddr] "<<ptCalibrationArray[lutAddr.to_uint()]<<std::endl;
      return ptCalibrationArray[lutAddr.to_uint()].to_uint();
    }

    void save(const std::string& filename) override {
      // Create an empty property tree object.
      boost::property_tree::ptree tree;

      PUT_VAR(tree, name, output0_I)
      PUT_VAR(tree, name, output0_F)
      PUT_VAR(tree, name, output1_I)
      PUT_VAR(tree, name, output1_F)

      lutLayer1.save(tree, name);
      lutLayer2.save(tree, name);
      lutLayer3_0.save(tree, name);
      lutLayer3_1.save(tree, name);

      int size = ptCalibrationArray.size();
      std::string key = "LutNetworkFixedPointRegression2Outputs.ptCalibrationArray";
      PUT_VAR(tree, key, size)
      std::ostringstream ostr;
      for (auto& a : ptCalibrationArray) {
        ostr << a.to_uint() << ", ";
      }
      tree.put(key + ".values", ostr.str());

      boost::property_tree::write_xml(filename,
                                      tree,
                                      std::locale(),
                                      boost::property_tree::xml_parser::xml_writer_make_settings<std::string>(' ', 2));
    }

    void load(const std::string& filename) override {
      // Create an empty property tree object.
      boost::property_tree::ptree tree;

      boost::property_tree::read_xml(filename, tree);

      CHECK_VAR(tree, name, output0_I)
      CHECK_VAR(tree, name, output0_F)
      CHECK_VAR(tree, name, output1_I)
      CHECK_VAR(tree, name, output1_F)

      lutLayer1.load(tree, name);
      lutLayer2.load(tree, name);
      lutLayer3_0.load(tree, name);
      lutLayer3_1.load(tree, name);

      std::string key = "LutNetworkFixedPointRegression2Outputs.ptCalibrationArray";
      int size = ptCalibrationArray.size();
      CHECK_VAR(tree, key, size)

      auto str = tree.get<std::string>(key + ".values");

      std::stringstream ss(str);
      std::string item;

      for (auto& a : ptCalibrationArray) {
        if (std::getline(ss, item, ',')) {
          a = std::stoul(item, nullptr, 10);
        } else {
          throw std::runtime_error(
              "LutNetworkFixedPointRegression2Outputs::read: number of items get from file is smaller than lut size");
        }
      }
    }

    auto& getPtCalibrationArray() { return ptCalibrationArray; }

  private:
    std::array<ap_ufixed<LutLayer1::input_W, input_I, AP_TRN, AP_SAT>, inputSize> inputArray;
    ap_uint<layer2_input_I> layer1Bias;

    //ptCalibrationArray size should be 1024, the LSB of the input 0.25 GeV,
    //the output is int, with range 0...511, the LSB of output 0.5 GeV
    std::array<ap_uint<9>, 1 << (output0_I + output0_F)> ptCalibrationArray;

    std::string name = "LutNetworkFixedPointRegression2Outputs";
  };

} /* namespace lutNN */

#endif /* L1Trigger_L1TMuonOverlapPhase2_LutNetworkFixedPointRegression2Outputs_h */