add memory handling for DKS/GPU. No fitting possible, yet.

src/classes/PFitterFcnDKS.cpp

@@ -81,6 +81,17 @@ Double_t PFitterFcnDKS::operator()(const std::vector<Double_t>& par) const
 {
   Double_t value = 0.0;
 
+  // write parameter to GPU
+  Int_t ierr = fDKS.writeParams(&par[0], par.size());
+
+  // loop over all data sets
+  for (UInt_t i=0; i<fRunListCollection->GetNoOfSingleHisto(); i++) {
+    // set N0, Nbkg
+    // set fun values
+    // set map values
+    // calc chisq/log-mlh
+  }
+
   return value;
 }
 
@@ -122,6 +133,8 @@ void PFitterFcnDKS::CalcExpectedChiSquare(const std::vector<Double_t> &par, Doub
  */
 void PFitterFcnDKS::InitDKS(const UInt_t dksTag)
 {
+  Int_t ierr = 0;
+
   // if any device was allocated before, free the device resources
   FreeDKS();
 
@@ -142,16 +155,89 @@ void PFitterFcnDKS::InitDKS(const UInt_t dksTag)
 
   // init chisq buffer on the GPU
 
+  // 1) calculated the maximum size for the data needed.
+  Int_t maxSize = -1, size = -1;
+  Int_t parSize = -1, mapSize = -1, funSize = -1;
+  for (UInt_t i=0; i<fRunListCollection->GetNoOfSingleHisto(); i++) {
+    size = fRunListCollection->GetSingleHisto(i)->GetValue()->size();
+    if (maxSize < size)
+      maxSize = size;
+  }
+  for (UInt_t i=0; i<fRunListCollection->GetNoOfAsymmetry(); i++) {
+    size = fRunListCollection->GetAsymmetry(i)->GetValue()->size();
+    if (maxSize < size)
+      maxSize = size;
+  }
+  if (maxSize == -1) {
+    cerr << ">> PFitterFcnDKS::InitDKS: **ERROR** failed to get data size to be fitted." << endl;
+    fValid = false;
+    return;
+  }
+
+  // 2) get number of parameters / functions / maps
+  parSize = fRunListCollection->GetNoOfParameters();
+  if (parSize == -1) {
+    cerr << ">> PFitterFcnDKS::InitDKS: **ERROR** failed to get number of fit parameters." << endl;
+    fValid = false;
+    return;
+  }
+  funSize = fRunListCollection->GetNoOfFunctions();
+  if (funSize == -1) {
+    cerr << ">> PFitterFcnDKS::InitDKS: **ERROR** failed to get number of functions." << endl;
+    fValid = false;
+    return;
+  }
+  mapSize = fRunListCollection->GetNoOfMaps();
+  if (mapSize == -1) {
+    cerr << ">> PFitterFcnDKS::InitDKS: **ERROR** failed to get number of maps." << endl;
+    fValid = false;
+    return;
+  }
+  cout << "debug> parSize=" << parSize << ", funSize=" << funSize << ", mapSize=" << mapSize << endl;
+
+  // now ready to init the chisq buffer on the GPU
+  ierr = fDKS.initChiSquare(maxSize, parSize, funSize, mapSize);
+  if (ierr != 0) {
+    cerr << ">> PFitterFcnDKS::InitDKS: **ERROR** failed to allocate the necessary chisq buffer on the GPU." << endl;
+    fValid = false;
+    return;
+  }
+
   // allocate memory for the data on the GPU/CPU and transfer the data sets
+  fMemData.resize(fRunListCollection->GetNoOfSingleHisto());
+  PRunData *runData=0;
+  for (UInt_t i=0; i<fRunListCollection->GetNoOfSingleHisto(); i++) {
+    runData = fRunListCollection->GetSingleHisto(i);
+    if (runData == 0) {
+      cerr << ">> PFitterFcnDKS::InitDKS: **ERROR** failed to get data set (i=" << i << ") from fRunListCollection." << endl;
+      fValid = false;
+      return;
+    }
+    fMemData[i] = fDKS.allocateMemory<Double_t>(runData->GetValue()->size(), ierr);
+    if (ierr != 0) {
+      cerr << ">> PFitterFcnDKS::InitDKS: **ERROR** failed to allocated data set memory (i=" << i << ") on the GPU" << endl;
+      fValid = false;
+      return;
+    }
+    fDKS.writeData<double>(fMemData[i], runData->GetValue(), runData->GetValue()->size());
+  }
 
   // set the function string and compile the program
-  Int_t ierr = fDKS.callCompileProgram(fTheoStr, !fUseChi2);
+  ierr = fDKS.callCompileProgram(fTheoStr, !fUseChi2);
   if (ierr != 0) {
     cerr << ">> PFitterFcnDKS::InitDKS: **ERROR** failed to compile theory!" << endl;
     fValid = false;
     return;
   }
 
+  // keep the N0 and Nbkg for each data set
+  fNidx.resize(fRunListCollection->GetNoOfSingleHisto());
+  for (UInt_t i=0; i<fRunListCollection->GetNoOfSingleHisto(); i++) {
+    fNidx[i].fN0   = fRunListCollection->GetN0Idx(i);
+    fNidx[i].fNbkg = fRunListCollection->GetNbkgIdx(i);
+    cout << "debug> runNo=" << i << ", N0 idx=" << fNidx[i].fN0 << ", Nbkg idx=" << fNidx[i].fNbkg << endl;
+  }
+
   // checks device properties if openCL
   ierr = fDKS.checkMuSRKernels();
   if (ierr != 0) {
@@ -172,5 +258,10 @@ void PFitterFcnDKS::InitDKS(const UInt_t dksTag)
  */
 void PFitterFcnDKS::FreeDKS()
 {
-
+  PRunData *runData=0;
+  for (UInt_t i=0; i<fMemData.size(); i++) {
+    runData = fRunListCollection->GetSingleHisto(i);
+    fDKS.freeMemory<Double_t>(fMemData[i], runData->GetValue()->size());
+  }
+  fMemData.clear();
 }