UnbinnedSplineHandler.cpp

Go to the documentation of this file.

#include "UnbinnedSplineHandler.h"
 
#ifdef MaCh3_CUDA
#include "Splines/gpuSplineUtils.cuh"
#endif
 
#pragma GCC diagnostic ignored "-Wuseless-cast"
 
// *****************************************
//Set everything to NULL or 0
void UnbinnedSplineHandler::Initialise() {
// *****************************************
#ifdef MaCh3_CUDA
  MACH3LOG_INFO("Using GPU version event by event monolith");
  gpu_spline_handler = nullptr;
#endif
 
  cpu_spline_handler = new SplineMonoStruct();
 
  nKnots = 0;
  nTF1coeff = 0;
  NEvents = 0;
  _max_knots = 0;
 
  NSplines_valid = 0;
  NTF1_valid = 0;
 
  cpu_weights_spline_var = nullptr;
  cpu_weights_tf1_var = nullptr;
 
  cpu_total_weights = nullptr;
}
 
// *****************************************
UnbinnedSplineHandler::UnbinnedSplineHandler(std::vector<std::vector<TResponseFunction_red*> > &MasterSpline,
                     const std::vector<RespFuncType> &SplineType,
                     const bool SaveFlatTree,
                     const std::string& _FastSplineName) : SplineBase() {
// *****************************************
  //KS: If true it will save spline monolith into huge ROOT file
  SaveSplineFile = SaveFlatTree;
  FastSplineName = _FastSplineName;
  Initialise();
  MACH3LOG_INFO("-- GPUING WITH arrays and master spline containing TResponseFunction_red");
 
  // Convert the TSpline3 pointers to the reduced form and call the reduced constructor
  PrepareForGPU(MasterSpline, SplineType);
}
 
// *****************************************
// The shared initialiser from constructors of TSpline3 and TSpline3_red
void UnbinnedSplineHandler::PrepareForGPU(std::vector<std::vector<TResponseFunction_red*> > &MasterSpline, const std::vector<RespFuncType> &SplineType) {
// *****************************************
  // Scan for the max number of knots, the number of events (number of splines), and number of parameters
  int maxnSplines = 0;
  ScanMasterSpline(MasterSpline,
                   NEvents,
                   _max_knots,
                   nParams,
                   maxnSplines,
                   NSplines_valid,
                   nKnots,
                   NTF1_valid,
                   nTF1coeff,
                   SplineType);
 
  MACH3LOG_INFO("Found {} events", NEvents);
  MACH3LOG_INFO("Found {} knots at max", _max_knots);
  MACH3LOG_INFO("Found {} parameters", nParams);
  MACH3LOG_INFO("Found {} maximum number of splines in an event", maxnSplines);
  MACH3LOG_INFO("Found total {} knots in all splines", nKnots);
  MACH3LOG_INFO("Number of splines = {}", NSplines_valid);
  MACH3LOG_INFO("Found total {} coeffs in all TF1", nTF1coeff);
  MACH3LOG_INFO("Number of TF1 = {}", NTF1_valid);
 
  unsigned int event_size_max = _max_knots * nParams;
  // Declare the {x}, {y,b,c,d} arrays for all possible splines which the event has
  // We'll filter off the flat and "disabled" (e.g. CCQE event should not have MARES spline) ones in the next for loop, but need to declare these beasts here
 
  // Declare the {y,b,c,d} for each knot
  // float because GPU precision (could change to double, but will incur significant speed reduction on GPU unless you're very rich!)
  cpu_spline_handler->coeff_many.resize(nKnots*_nCoeff_); // *4 because we store y,b,c,d parameters in this array
  //KS: For x coeff we assume that for given dial (MAQE) spacing is identical,
  // here we are sloppy and assume each dial has the same number of knots, not a big problem
  cpu_spline_handler->coeff_x.resize(event_size_max, -999);
 
  //CW: With TF1 we only save the coefficients and the order of the polynomial
  // Makes most sense to have one large monolithic array, but then it becomes impossible to tell apart a coefficient from a "number of points". So have two arrays: one of coefficients and one of number of points
  // Let's first assume all are of _max_knots size
  // Now declare the arrays for each point in the valid splines which the event actually has (i.e. include the splines that the event undergoes)
  // Also make array with the number of points per spline (not per spline point!)
  // float because GPU precision (could change to double, but will incur significant speed reduction on GPU unless you're very rich!)
  cpu_nPoints_arr.resize(NTF1_valid);
  cpu_coeff_TF1_many.resize(nTF1coeff); // *5 because this array holds  a,b,c,d,e parameters
 
  //KS: Map keeping track how many parameters applies to each event, we keep two numbers here {number of splines per event, index where splines start for a given event}
  cpu_nParamPerEvent.resize(2 * NEvents, -1);
  cpu_nParamPerEvent_tf1.resize(2 * NEvents, -1);
 
  // Make array with the number of points per spline (not per spline point!)
  cpu_spline_handler->paramNo_arr.resize(NSplines_valid);
  //KS: And array which tells where each spline stars in a big monolith array, sort of knot map
  cpu_spline_handler->nKnots_arr.resize(NSplines_valid);
  cpu_paramNo_TF1_arr.resize(NTF1_valid);
 
  // Temporary arrays to hold the coefficients for each spline
  // We get one x, one y, one b,... for each point, so only need to be _max_knots big
  //KS: Some params has less splines but this is all right main array will get proper number while this temp will be deleted
  float *x_tmp = new float[_max_knots]();
  float *many_tmp = new float[_max_knots*_nCoeff_]();
  float *temp_coeffs = new float[_nTF1Coeff_]();
 
  // Count the number of events
  unsigned int KnotCounter = 0;
  unsigned int TF1PointsCounter = 0;
  unsigned int NSplinesCounter = 0;
  unsigned int TF1sCounter = 0;
  int ParamCounter = 0;
  int ParamCounterGlobal = 0;
  int ParamCounter_TF1 = 0;
  int ParamCounterGlobalTF1 = 0;
  // Loop over events and extract the spline coefficients
  for(unsigned int EventCounter = 0; EventCounter < MasterSpline.size(); ++EventCounter) {
    // Structure of MasterSpline is std::vector<std::vector<TSpline3*>>
    // A conventional iterator to count which parameter a given spline should be applied to
    for(unsigned int ParamNumber = 0; ParamNumber < MasterSpline[EventCounter].size(); ++ParamNumber) {
      // If nullptr we don't have this spline for the event, so move to next spline
      if (MasterSpline[EventCounter][ParamNumber] == nullptr) continue;
 
      if(SplineType[ParamNumber] == kTSpline3_red)
      {
        //KS: how much knots each spline has
        int nPoints_tmp = 0;
        // Get a pointer to the current spline for this event
        TResponseFunction_red* TespFunc = MasterSpline[EventCounter][ParamNumber];
        TSpline3_red* CurrSpline = static_cast<TSpline3_red*>(TespFunc);
 
        // If the number of knots are greater than 2 the spline is not a dummy and we should extract coefficients to load onto the GPU
        GetSplineCoeff_SepMany(CurrSpline, nPoints_tmp, x_tmp, many_tmp);
 
        //KS: One knot means flat spline so ignore
        if (nPoints_tmp == 1) continue;
        for (int j = 0; j < _max_knots; ++j) {
          cpu_spline_handler->coeff_x[ParamNumber*_max_knots + j] = x_tmp[j];
        }
        //KS: Contrary to X coeff we keep for other coeff only filled knots, there is no much gain for doing so for x coeff
        for (int j = 0; j < nPoints_tmp; ++j) {
          for (int k = 0; k < _nCoeff_; k++) {
            cpu_spline_handler->coeff_many[KnotCounter*_nCoeff_ + j*_nCoeff_ + k] = many_tmp[j*_nCoeff_+k];
          }
        }
        // Set the parameter number for this spline
        cpu_spline_handler->paramNo_arr[NSplinesCounter] = short(ParamNumber);
        //KS: Fill map when each spline starts
        cpu_spline_handler->nKnots_arr[NSplinesCounter] = KnotCounter;
        KnotCounter += nPoints_tmp;
 
        ++ParamCounter;
        // Increment the counter for the number of good splines we have
        ++NSplinesCounter;
      }
      else if (SplineType[ParamNumber] == kTF1_red)
      {
        // Don't actually use this ever -- we give each spline the maximum number of points found in all splines
        int nPoints_tmp = 0;
        // Get a pointer to the current spline for this event
        TF1_red* CurrSpline = dynamic_cast<TF1_red*>(MasterSpline[EventCounter][ParamNumber]);
 
        // If the number of knots are greater than 2 the spline is not a dummy and we should extract coefficients to load onto the GPU
        GetTF1Coeff(CurrSpline, nPoints_tmp, temp_coeffs);
        for (int j = 0; j < _nTF1Coeff_; ++j) {
          cpu_coeff_TF1_many[TF1PointsCounter+j] = temp_coeffs[j];
        }
        // Save the number of points for this spline
        cpu_nPoints_arr[TF1sCounter] = short(nPoints_tmp);
 
        TF1PointsCounter += nPoints_tmp;
        // Set the parameter number for this spline
        cpu_paramNo_TF1_arr[TF1sCounter] = short(ParamNumber);
        ++ParamCounter_TF1;
        // Increment the counter for the number of good splines we have
        ++TF1sCounter;
      }
      //KS: Don't delete in debug
      #ifndef MACH3_DEBUG
      delete MasterSpline[EventCounter][ParamNumber];
      MasterSpline[EventCounter][ParamNumber] = nullptr;
      #endif
    } // End the loop over the parameters in the MasterSpline
    cpu_nParamPerEvent[2*EventCounter] = ParamCounter;
    cpu_nParamPerEvent[2*EventCounter+1] = ParamCounterGlobal;
    ParamCounterGlobal += ParamCounter;
 
    cpu_nParamPerEvent_tf1[2*EventCounter] = ParamCounter_TF1;
    cpu_nParamPerEvent_tf1[2*EventCounter+1] = ParamCounterGlobalTF1;
    ParamCounterGlobalTF1 += ParamCounter_TF1;
 
    ParamCounter = 0;
    ParamCounter_TF1 = 0;
  } // End the loop over the number of events
  delete[] many_tmp;
  delete[] x_tmp;
  delete[] temp_coeffs;
 
  int BadXCounter = 0;
  for (unsigned int j = 0; j < event_size_max; j++) {
    if (cpu_spline_handler->coeff_x[j] == -999) BadXCounter++;
    // Perform checks that all entries have been modified from initial values
    if (cpu_spline_handler->coeff_x[j] == -999 && BadXCounter < 5) {
      MACH3LOG_WARN("***** BAD X !! *****");
      MACH3LOG_WARN("Indicates some parameter doesn't have a single spline");
      MACH3LOG_WARN("j = {}", j);
      //throw MaCh3Exception(__FILE__ , __LINE__ );
    }
    if(BadXCounter == 5) MACH3LOG_WARN("There is more unutilised knots although I will stop spamming");
  }
 
  MACH3LOG_WARN("Found in total {} BAD X", BadXCounter);
  //KS: This is tricky as this variable use both by CPU and GPU, however if use CUDA we use cudaMallocHost
  #ifndef MaCh3_CUDA
  cpu_total_weights = new M3::float_t[NEvents]();
  cpu_weights_spline_var = new float[NSplines_valid]();
  cpu_weights_tf1_var = new float[NTF1_valid]();
  #endif
 
  // Print some info; could probably make this to a separate function
  PrintInitialsiation();
  if(SaveSplineFile) PrepareSplineFile(FastSplineName);
 
  MoveToGPU();
 
  // Can pass the spline segments to the GPU instead of the values
  // Make these here and only refill them for each loop, avoiding unnecessary new/delete on each reconfigure
  SetupSegments();
}
 
// *****************************************
// The shared initialiser from constructors of TSpline3 and TSpline3_red
void UnbinnedSplineHandler::MoveToGPU() {
// *****************************************
  #ifdef MaCh3_CUDA
  unsigned int event_size_max = _max_knots * nParams;
  MACH3LOG_INFO("Total size = {:.2f} MB memory on CPU to move to GPU",
                (double(sizeof(float) * nKnots * _nCoeff_) + double(sizeof(float) * event_size_max) / 1.E6 +
                double(sizeof(short int) * NSplines_valid)) / 1.E6);
  MACH3LOG_INFO("Total TF1 size = {:.2f} MB memory on CPU to move to GPU",
                double(sizeof(float) * NTF1_valid * _nTF1Coeff_) / 1.E6);
  MACH3LOG_INFO("GPU weight array (GPU->CPU every step) = {:.2f} MB", static_cast<double>(sizeof(float)) * (NSplines_valid + NTF1_valid) / 1.0e6);
  MACH3LOG_INFO("Since you are running Total event weight mode then GPU weight array (GPU->CPU every step) = {:.2f} MB",
                double(sizeof(float) * NEvents) / 1.E6);
  MACH3LOG_INFO("Parameter value array (CPU->GPU every step) = {:.4f} MB", double(sizeof(float) * nParams) / 1.E6);
  //CW: With the new set-up we have:   1 coefficient array of size coeff_array_size, all same size
  //                                1 coefficient array of size coeff_array_size*4, holding y,b,c,d in order (y11,b11,c11,d11; y12,b12,c12,d12;...) where ynm is n = spline number, m = spline point. Should really make array so that order is (y11,b11,c11,d11; y21,b21,c21,d21;...) because it will optimise cache hits I think; try this if you have time
  //                                return gpu_weights
 
  gpu_spline_handler = new SplineMonolithGPU();
 
  // The gpu_XY arrays don't actually need initialising, since they are only placeholders for what we'll move onto the GPU. As long as we cudaMalloc the size of the arrays correctly there shouldn't be any problems
  // Can probably make this a bit prettier but will do for now
  // Could be a lot smaller of a function...
  gpu_spline_handler->InitGPU_SplineMonolith(
          &cpu_total_weights,
          NEvents,
          nKnots, // How many entries in coefficient array (*4 for the "many" array)
          NSplines_valid, // What's the number of splines we have (also number of entries in gpu_nPoints_arr)
          NTF1_valid,
          event_size_max //Knots times event number of unique splines
  );
 
  // Move number of splines and spline size to constant GPU memory; every thread does not need a copy...
  // The implementation lives in splines/gpuSplineUtils.cu
  // The GPU splines don't actually need declaring but is good for demonstration, kind of
  // fixed by passing const reference
  gpu_spline_handler->CopyToGPU_SplineMonolith(
          cpu_spline_handler,
 
          // TFI related now
          cpu_coeff_TF1_many,
          cpu_paramNo_TF1_arr,
          NEvents,
          cpu_nParamPerEvent,
          cpu_nParamPerEvent_tf1,
          nParams,
          NSplines_valid,
          _max_knots,
          nKnots,
          NTF1_valid);
 
  // Delete all the coefficient arrays from the CPU once they are on the GPU
  CleanVector(cpu_coeff_TF1_many);
  CleanVector(cpu_paramNo_TF1_arr);
  CleanVector(cpu_nParamPerEvent);
  CleanVector(cpu_nParamPerEvent_tf1);
  delete cpu_spline_handler;
  cpu_spline_handler = nullptr;
  MACH3LOG_INFO("Good GPU loading");
  #endif
}
 
// Need to specify template functions in header
// *****************************************
// Scan the master spline to get the maximum number of knots in any of the TSpline3*
void UnbinnedSplineHandler::ScanMasterSpline(std::vector<std::vector<TResponseFunction_red*> > & MasterSpline,
                                 unsigned int &nEvents,
                                 short int &MaxPoints,
                                 short int &numParams,
                                 int &nSplines,
                                 unsigned int &NSplinesValid,
                                 unsigned int &numKnots,
                                 unsigned int &nTF1Valid,
                                 unsigned int &nTF1_coeff,
                                 const std::vector<RespFuncType> &SplineType) {
// *****************************************
  // Need to extract: the total number of events
  //                  number of parameters
  //                  maximum number of knots
  MaxPoints = 0;
  nEvents   = 0;
  numParams   = 0;
  nSplines = 0;
  numKnots = 0;
  NSplinesValid = 0;
  nTF1Valid = 0;
  nTF1_coeff = 0;
 
  // Check the number of events
  nEvents = int(MasterSpline.size());
 
  // Maximum number of splines one event can have (scan through and find this number)
  int nMaxSplines_PerEvent = 0;
 
  //KS: We later check that each event has the same number of splines so this is fine
  numParams = short(MasterSpline[0].size());
  // Initialise
  SplineInfoArray.resize(numParams);
 
  // Loop over each parameter
  for(unsigned int EventCounter = 0; EventCounter < MasterSpline.size(); ++EventCounter) {
    // Check that each event has each spline saved
    if (numParams > 0) {
      int TempSize = int(MasterSpline[EventCounter].size());
      if (TempSize != numParams) {
        MACH3LOG_ERROR("Found {} parameters for event {}", TempSize, EventCounter);
        MACH3LOG_ERROR("but was expecting {} since that's what I found for the previous event", numParams);
        MACH3LOG_ERROR("Somehow this event has a different number of spline parameters... Please study further!");
        throw MaCh3Exception(__FILE__ , __LINE__ );
      }
    }
    numParams = short(MasterSpline[EventCounter].size());
 
    int nSplines_SingleEvent = 0;
    int nPoints = 0;
    // Loop over each pointer
    for(unsigned int ParamNumber = 0; ParamNumber < MasterSpline[EventCounter].size(); ++ParamNumber) {
      if (MasterSpline[EventCounter][ParamNumber]) {
        if(SplineType[ParamNumber] == kTSpline3_red)
        {
          TResponseFunction_red* TespFunc = MasterSpline[EventCounter][ParamNumber];
          TSpline3_red* CurrSpline = dynamic_cast<TSpline3_red*>(TespFunc);
          if(CurrSpline){
            nPoints = CurrSpline->GetNp();
          }
 
          if (nPoints > MaxPoints) {
            MaxPoints = static_cast<short int>(nPoints);
          }
          numKnots += nPoints;
          nSplines_SingleEvent++;
 
          // Fill the SplineInfoArray entries with information on each splinified parameter
          if (SplineInfoArray[ParamNumber].xPts.size() == 0)
          {
            // Fill the number of points
            SplineInfoArray[ParamNumber].nPts = CurrSpline->GetNp();
 
            // Fill the x points
            SplineInfoArray[ParamNumber].xPts.resize(SplineInfoArray[ParamNumber].nPts);
            for (M3::int_t k = 0; k < SplineInfoArray[ParamNumber].nPts; ++k)
            {
              M3::float_t xtemp = M3::float_t(-999.99);
              M3::float_t ytemp = M3::float_t(-999.99);
              CurrSpline->GetKnot(k, xtemp, ytemp);
              SplineInfoArray[ParamNumber].xPts[k] = xtemp;
            }
          }
          NSplinesValid++;
        }
        else if (SplineType[ParamNumber] == kTF1_red)
        {
          TResponseFunction_red* TespFunc = MasterSpline[EventCounter][ParamNumber];
          TF1_red* CurrSpline = dynamic_cast<TF1_red*>(TespFunc);
          nPoints = CurrSpline->GetSize();
          nTF1_coeff += nPoints;
          nTF1Valid++;
        }
      } else {
        // If NULL we don't have this spline for the event, so move to next spline
        continue;
      }
    }
    if (nSplines_SingleEvent > nMaxSplines_PerEvent) nMaxSplines_PerEvent = nSplines_SingleEvent;
  }
  nSplines = nMaxSplines_PerEvent;
 
  int Counter = 0;
  //KS: Sanity check that everything was set correctly
  for (M3::int_t i = 0; i < numParams; ++i)
  {
    // KS: We don't find segment for TF1, so ignore this
    if (SplineType[i] == kTF1_red) continue;
 
    const M3::int_t nPoints = SplineInfoArray[i].nPts;
    const std::vector<M3::float_t>& xArray = SplineInfoArray[i].xPts;
    if (nPoints == -999 || xArray.size() == 0) {
      Counter++;
      if(Counter < 5) {
        MACH3LOG_WARN("SplineInfoArray[{}] isn't set yet", i);
      }
      continue;
      //throw MaCh3Exception(__FILE__ , __LINE__ );
    }
  }
  MACH3LOG_WARN("In total SplineInfoArray for {} hasn't been initialised", Counter);
}
 
// *****************************************
// Load SplineFile
UnbinnedSplineHandler::UnbinnedSplineHandler(const std::string& FileName)
          : SplineBase() {
// *****************************************
  Initialise();
  MACH3LOG_INFO("-- GPUING WITH {X} and {Y,B,C,D} arrays and master spline containing TSpline3_red");
  // Convert the TSpline3 pointers to the reduced form and call the reduced constructor
  LoadSplineFile(FileName);
}
 
// *****************************************
// Load SplineMonolith from ROOT file
void UnbinnedSplineHandler::LoadSplineFile(std::string FileName) {
// *****************************************
  M3::AddPath(FileName);
  auto SplineFile = std::make_unique<TFile>(FileName.c_str(), "OPEN");
  TTree *Settings = SplineFile->Get<TTree>("Settings");
  TTree *Monolith_TF1 = SplineFile->Get<TTree>("Monolith_TF1");
  TTree *EventInfo = SplineFile->Get<TTree>("EventInfo");
  TTree *SplineTree = SplineFile->Get<TTree>("SplineTree");
 
  unsigned int NEvents_temp;
  short int nParams_temp;
  int _max_knots_temp;
  unsigned int nKnots_temp;
  unsigned int NSplines_valid_temp;
  unsigned int nTF1Valid_temp;
  unsigned int nTF1coeff_temp;
 
  Settings->SetBranchAddress("NEvents", &NEvents_temp);
  Settings->SetBranchAddress("nParams", &nParams_temp);
  Settings->SetBranchAddress("_max_knots", &_max_knots_temp);
  Settings->SetBranchAddress("nKnots", &nKnots_temp);
  Settings->SetBranchAddress("NSplines_valid", &NSplines_valid_temp);
  Settings->SetBranchAddress("NTF1_valid", &nTF1Valid_temp);
  Settings->SetBranchAddress("nTF1coeff", &nTF1coeff_temp);
 
  Settings->GetEntry(0);
 
  NEvents = NEvents_temp;
  nParams = nParams_temp;
  _max_knots = static_cast<short int>(_max_knots_temp);
  nKnots = nKnots_temp;
  NSplines_valid = NSplines_valid_temp;
  NTF1_valid = nTF1Valid_temp;
  nTF1coeff = nTF1coeff_temp;
 
  cpu_nParamPerEvent.resize(2*NEvents);
  cpu_nParamPerEvent_tf1.resize(2*NEvents);
  cpu_coeff_TF1_many.resize(nTF1coeff);
 
  //KS: This is tricky as this variable use both by CPU and GPU, however if use CUDA we use cudaMallocHost
#ifndef MaCh3_CUDA
  cpu_total_weights = new M3::float_t[NEvents]();
  cpu_weights_spline_var = new float[NSplines_valid]();
  cpu_weights_tf1_var = new float[NTF1_valid]();
#endif
 
  SplineTree->SetBranchAddress("SplineObject", &cpu_spline_handler);
  SplineTree->GetEntry(0);
 
  float coeff_tf1 = 0.;
  Monolith_TF1->SetBranchAddress("cpu_coeff_TF1_many", &coeff_tf1);
  for(unsigned int i = 0; i < nTF1coeff; i++)
  {
    Monolith_TF1->GetEntry(i);
    cpu_coeff_TF1_many[i] = coeff_tf1;
  }
 
  unsigned int nParamPerEvent = 0;
  unsigned int nParamPerEvent_tf1 = 0;
 
  EventInfo->SetBranchAddress("cpu_nParamPerEvent", &nParamPerEvent);
  EventInfo->SetBranchAddress("cpu_nParamPerEvent_tf1", &nParamPerEvent_tf1);
  for(unsigned int i = 0; i < 2*NEvents; i++)
  {
    EventInfo->GetEntry(i);
    cpu_nParamPerEvent[i] = nParamPerEvent;
    cpu_nParamPerEvent_tf1[i] = nParamPerEvent_tf1;
  }
 
  LoadFastSplineInfoDir(SplineFile);
 
  SplineFile->Close();
 
  // Print some info; could probably make this to a separate function
  PrintInitialsiation();
 
  MoveToGPU();
 
  SetupSegments();
}
 
// *****************************************
void UnbinnedSplineHandler::SetupSegments() {
// *****************************************
  //KS: Since we are going to copy it each step use fancy CUDA memory allocation
  #ifdef MaCh3_CUDA
  gpu_spline_handler->InitGPU_Segments(&SplineSegments);
  gpu_spline_handler->InitGPU_Vals(&ParamValues);
  #else
  SplineSegments = new short int[nParams]();
  ParamValues = new float[nParams]();
  #endif
  for (M3::int_t j = 0; j < nParams; j++)
  {
    SplineSegments[j] = 0;
    ParamValues[j] = -999;
  }
}
 
// *****************************************
// Save SplineMonolith into ROOT file
void UnbinnedSplineHandler::PrepareSplineFile(std::string FileName) {
// *****************************************
  M3::AddPath(FileName);
 
  auto SplineFile = std::make_unique<TFile>(FileName.c_str(), "recreate");
  TTree *Settings = new TTree("Settings", "Settings");
  TTree *Monolith_TF1 = new TTree("Monolith_TF1", "Monolith_TF1");
  TTree *XKnots = new TTree("XKnots", "XKnots");
  TTree *EventInfo = new TTree("EventInfo", "EventInfo");
 
  unsigned int NEvents_temp = NEvents;
  short int nParams_temp = nParams;
  int _max_knots_temp = _max_knots;
  unsigned int nKnots_temp = nKnots;
  unsigned int NSplines_valid_temp = NSplines_valid;
  unsigned int nTF1Valid_temp = NTF1_valid;
  unsigned int nTF1coeff_temp = nTF1coeff;
 
  Settings->Branch("NEvents", &NEvents_temp, "NEvents/i");
  Settings->Branch("nParams", &nParams_temp, "nParams/S");
  Settings->Branch("_max_knots", &_max_knots_temp, "_max_knots/I");
  Settings->Branch("nKnots", &nKnots_temp, "nKnots/i");
  Settings->Branch("NSplines_valid", &NSplines_valid_temp, "NSplines_valid/i");
  Settings->Branch("NTF1_valid", &nTF1Valid_temp, "NTF1_valid/i");
  Settings->Branch("nTF1coeff", &nTF1coeff_temp, "nTF1coeff/i");
 
  Settings->Fill();
 
  SplineFile->cd();
  Settings->Write();
 
  TTree *SplineTree = new TTree("SplineTree", "SplineTree");
  // Create a branch for the SplineMonoStruct object
  SplineTree->Branch("SplineObject", &cpu_spline_handler);
  SplineTree->Fill();
  SplineTree->Write();
  delete SplineTree;
 
  float coeff_tf1 = 0.;
  Monolith_TF1->Branch("cpu_coeff_TF1_many", &coeff_tf1, "cpu_coeff_TF1_many/F");
  for(unsigned int i = 0; i < nTF1coeff; i++)
  {
    coeff_tf1 = cpu_coeff_TF1_many[i];
    Monolith_TF1->Fill();
  }
  SplineFile->cd();
  Monolith_TF1->Write();
 
  unsigned int nParamPerEvent = 0;
  unsigned int nParamPerEvent_tf1 = 0;
 
  EventInfo->Branch("cpu_nParamPerEvent", &nParamPerEvent, "cpu_nParamPerEvent/i");
  EventInfo->Branch("cpu_nParamPerEvent_tf1", &nParamPerEvent_tf1, "cpu_nParamPerEvent_tf1/i");
 
  for(unsigned int i = 0; i < 2*NEvents; i++)
  {
    nParamPerEvent = cpu_nParamPerEvent[i];
    nParamPerEvent_tf1 = cpu_nParamPerEvent_tf1[i];
    EventInfo->Fill();
  }
  SplineFile->cd();
  EventInfo->Write();
 
  PrepareFastSplineInfoDir(SplineFile);
 
  delete Settings;
  delete Monolith_TF1;
  delete XKnots;
  delete EventInfo;
  SplineFile->Close();
}
 
// *****************************************
// Destructor
// Cleans up the allocated GPU memory
UnbinnedSplineHandler::~UnbinnedSplineHandler() {
// *****************************************
  #ifdef MaCh3_CUDA
  //KS: Since we declared them using CUDA alloc we have to free memory using also cuda functions
  gpu_spline_handler->CleanupPinnedMemory(cpu_total_weights, SplineSegments, ParamValues);
  delete gpu_spline_handler;
  #else
  if(SplineSegments != nullptr) delete[] SplineSegments;
  if(ParamValues != nullptr) delete[] ParamValues;
  if(cpu_total_weights != nullptr) delete[] cpu_total_weights;
  #endif
 
  if(cpu_weights_spline_var != nullptr) delete[] cpu_weights_spline_var;
  if(cpu_weights_tf1_var != nullptr) delete[] cpu_weights_tf1_var;
 
  if(cpu_spline_handler != nullptr) delete cpu_spline_handler;
}
 
// *****************************************
// Get the spline coefficients from the TSpline3 so that we can load ONLY these onto the GPU, not the whole TSpline3 object
// This loads up coefficients into two arrays: one x array and one yabcd array
// This should maximize our cache hits!
void UnbinnedSplineHandler::GetSplineCoeff_SepMany(TSpline3_red* &spl, int &nPoints, float *& xArray, float *& manyArray) const {
// *****************************************
  // Initialise all arrays to 1.0
  for (int i = 0; i < _max_knots; ++i) {
    xArray[i] = 1.0;
    for (int j = 0; j < _nCoeff_; j++) {
      manyArray[i*_nCoeff_+j] = 1.0;
    }
  }
  // Get number of points in spline
  int Np = spl->GetNp();
  // If spline is flat, set number of knots to 1.0,
  // This is used later to expedite the calculations for flat splines
  // tmpArray[0] is number of knots
  nPoints = Np;
  if (Np > _max_knots) {
    MACH3LOG_ERROR("Error, number of points is greater than saved {}", _max_knots);
    MACH3LOG_ERROR("This _WILL_ cause problems with GPU splines and _SHOULD_ be fixed!");
    MACH3LOG_ERROR("nPoints = {}, _max_knots = {}", nPoints, _max_knots);
    throw MaCh3Exception(__FILE__ , __LINE__ );
  }
 
  // The coefficients we're writing to
  M3::float_t x, y, b, c, d;
  // TSpline3 can only take doubles, not floats
  // But our GPU is slow with doubles, so need to cast to float
  for(int i = 0; i < Np; i++) {
    // Get the coefficients from the TSpline3 object
    spl->GetCoeff(i, x, y, b, c, d);
    // Write the arrays
    xArray[i] = float(x);
    manyArray[i*_nCoeff_] = float(y); // 4 because manyArray stores y,b,c,d
    manyArray[i*_nCoeff_+1] = float(b);
    manyArray[i*_nCoeff_+2] = float(c);
    manyArray[i*_nCoeff_+3] = float(d);
    if((xArray[i] == -999) || (manyArray[i*_nCoeff_] == -999) || (manyArray[i*_nCoeff_ +1] == -999) || (manyArray[i*_nCoeff_+2] == -999) || (manyArray[i*_nCoeff_+3] == -999)){
      MACH3LOG_ERROR("*********** Bad params in {} ************", __func__);
      MACH3LOG_ERROR("pre cast to float (x, y, b, c, d) = {:.2f}, {:.2f}, {:.2f}, {:.2f}, {:.2f}", x, y, b, c, d);
      MACH3LOG_ERROR("pre cast to float (x, y, b, c, d) = {:.2f}, {:.2f}, {:.2f}, {:.2f}, {:.2f}", xArray[i], manyArray[i*_nCoeff_], manyArray[i*_nCoeff_+1], manyArray[i*_nCoeff_+2], manyArray[i*_nCoeff_+3]);
      MACH3LOG_ERROR("This will cause problems when preparing for GPU");
      MACH3LOG_ERROR("***************************************************************");
    }
  }
}
 
#ifdef MaCh3_CUDA
// *****************************************
// Tell the GPU to evaluate the weights
// Load up the two x,{y,b,c,d} arrays into memory and have GPU read them with more coalescence instead of one monolithic array
// This should be used when we're using separate x,y,a,b,c,d arrays
// Also pass the segments for the parameter along with their parameter values
// This avoids doing lots of binary searches on the GPU
void UnbinnedSplineHandler::Evaluate() {
// *****************************************
  // There's a parameter mapping that goes from spline parameter to a global parameter index
  // Find the spline segments
  FindSplineSegment();
 
  // The main call to the GPU
  gpu_spline_handler->RunGPU_SplineMonolith(
          cpu_total_weights,
          ParamValues,
          SplineSegments);
}
#else
//If CUDA is not enabled do the same on CPU
// *****************************************
void UnbinnedSplineHandler::Evaluate() {
// *****************************************
  // There's a parameter mapping that goes from spline parameter to a global parameter index
  // Find the spline segments
  FindSplineSegment();
 
  //KS: Huge MP loop over all valid splines
  CalcSplineWeights();
 
  //KS: Huge MP loop over all events calculating total weight per event
  CalcTotalEventWeight();
}
#endif
 
//*********************************************************
void UnbinnedSplineHandler::CalcSplineWeights() {
//*********************************************************
  #ifdef MULTITHREAD
  //KS: Open parallel region
  #pragma omp parallel
  {
  #endif
    //KS: First we calculate
    #ifdef MULTITHREAD
    #pragma omp for simd nowait
    #endif
    for (unsigned int splineNum = 0; splineNum < NSplines_valid; ++splineNum)
    {
      //CW: Which Parameter we are accessing
      const short int Param = cpu_spline_handler->paramNo_arr[splineNum];
 
      //CW: Avoids doing costly binary search on GPU
      const short int segment = SplineSegments[Param];
 
      //KS: Segment for coeff_x is simply parameter*max knots + segment as each parameters has the same spacing
      const short int segment_X = short(Param*_max_knots+segment);
 
      //KS: Find knot position in out monolithical structure
      const unsigned int CurrentKnotPos = cpu_spline_handler->nKnots_arr[splineNum]*_nCoeff_+segment*_nCoeff_;
 
      // We've read the segment straight from CPU and is saved in segment_gpu
      // polynomial parameters from the monolithic splineMonolith
      const float fY = cpu_spline_handler->coeff_many[CurrentKnotPos];
      const float fB = cpu_spline_handler->coeff_many[CurrentKnotPos + 1];
      const float fC = cpu_spline_handler->coeff_many[CurrentKnotPos + 2];
      const float fD = cpu_spline_handler->coeff_many[CurrentKnotPos + 3];
      // The is the variation itself (needed to evaluate variation - stored spline point = dx)
      const float dx = ParamValues[Param] - cpu_spline_handler->coeff_x[segment_X];
 
      //CW: Wooow, let's use some fancy intrinsic and pull down the processing time by <1% from normal multiplication! HURRAY
      cpu_weights_spline_var[splineNum] = fmaf(dx, fmaf(dx, fmaf(dx, fD, fC), fB), fY);
      // Or for the more "easy to read" version:
      //cpu_weights_spline_var[splineNum] = (fY+dx*(fB+dx*(fC+dx*fD)));
    }
 
    #ifdef MULTITHREAD
    #pragma omp for simd
    #endif
    for (unsigned int tf1Num = 0; tf1Num < NTF1_valid; ++tf1Num)
    {
      // The is the variation itself (needed to evaluate variation - stored spline point = dx)
      const float x = ParamValues[cpu_paramNo_TF1_arr[tf1Num]];
 
      // Read the coefficients
      const unsigned int TF1_Index = tf1Num * _nTF1Coeff_;
      const float a = cpu_coeff_TF1_many[TF1_Index];
      const float b = cpu_coeff_TF1_many[TF1_Index + 1];
 
      cpu_weights_tf1_var[tf1Num] = fmaf(a, x, b);
      // cpu_weights_tf1_var[tf1Num] = a*x + b;
      //cpu_weights_tf1_var[splineNum] = 1 + a*x + b*x*x + c*x*x*x + d*x*x*x*x + e*x*x*x*x*x;
    }
  #ifdef MULTITHREAD
  //KS: End parallel region
  }
  #endif
}
 
//*********************************************************
//KS: Calc total event weight on CPU
void UnbinnedSplineHandler::CalcTotalEventWeight() {
//*********************************************************
  #ifdef MULTITHREAD
  #pragma omp parallel for
  #endif
  for (unsigned int EventNum = 0; EventNum < NEvents; ++EventNum)
  {
    float totalWeight = 1.0f; // Initialize total weight for each event
 
    const unsigned int Offset = 2 * EventNum;
 
    // Extract the parameters for the current event
    const unsigned int startIndex = cpu_nParamPerEvent[Offset + 1];
    const unsigned int numParams = cpu_nParamPerEvent[Offset];
 
    // Compute total weight for the current event
    #ifdef MULTITHREAD
    #pragma omp simd reduction(*:totalWeight)
    #endif
    for (unsigned int id = 0; id < numParams; ++id) {
      totalWeight *= cpu_weights_spline_var[startIndex + id];
    }
    //Now TF1
    // Extract the parameters for the current event
    const unsigned int startIndex_tf1 = cpu_nParamPerEvent_tf1[Offset + 1];
    const unsigned int numParams_tf1 = cpu_nParamPerEvent_tf1[Offset];
 
    // Compute total weight for the current event
    #ifdef MULTITHREAD
    #pragma omp simd reduction(*:totalWeight)
    #endif
    for (unsigned int id = 0; id < numParams_tf1; ++id) {
      totalWeight *= cpu_weights_tf1_var[startIndex_tf1 + id];
    }
 
    // Store the total weight for the current event
    cpu_total_weights[EventNum] = static_cast<M3::float_t>(totalWeight);
  }
}
 
//*********************************************************
//KS: Print info about how much knots etc has been initialised
void UnbinnedSplineHandler::PrintInitialsiation() const {
//*********************************************************
  unsigned int event_size_max = _max_knots * nParams;
 
  MACH3LOG_INFO("--- INITIALISED Spline Monolith ---");
  MACH3LOG_INFO("{} events with {} splines", NEvents, NSplines_valid);
  MACH3LOG_INFO("On average {:.2f} splines per event ({}/{})", float(NSplines_valid)/float(NEvents), NSplines_valid, NEvents);
  MACH3LOG_INFO("Size of x array = {:.4f} MB", double(sizeof(float)*event_size_max)/1.E6);
  MACH3LOG_INFO("Size of coefficient (y,b,c,d) array = {:.2f} MB", double(sizeof(float)*nKnots*_nCoeff_)/1.E6);
  MACH3LOG_INFO("Size of parameter # array = {:.2f} MB", double(sizeof(short int)*NSplines_valid)/1.E6);
 
  MACH3LOG_INFO("On average {:.2f} TF1 per event ({}/{})", float(NTF1_valid)/float(NEvents), NTF1_valid, NEvents);
  MACH3LOG_INFO("Size of TF1 coefficient (a,b,c,d,e) array = {:.2f} MB", double(sizeof(float)*NTF1_valid*_nTF1Coeff_)/1.E6);
}
 
//*********************************************************
//KS: After calculations are done on GPU we copy memory to CPU. This operation is asynchronous meaning while memory is being copied some operations are being carried. Memory must be copied before actual reweight. This function make sure all has been copied.
void UnbinnedSplineHandler::SynchroniseMemTransfer() const {
//*********************************************************
  #ifdef MaCh3_CUDA
  SynchroniseSplines();
  CudaCheckError();
  #endif
}