LMU/musrfit
5
0
Fork 0
Code Issues 6 Pull Requests Actions Packages Projects Releases 4 Activity

Fixed a bug in the LF calculations and parallelized the chi^2 calculations in single-histogram and asymmetry fits

Browse Source
This commit is contained in:
Bastian M. Wojek
2011-05-21 14:11:00 +00:00
parent c9c8f83971
commit 1b5c1df691
5 changed files with 101 additions and 29 deletions
Download Patch File Download Diff File Expand all files Collapse all files

40
src/classes/PRunSingleHisto.cpp
View File

@ -29,6 +29,14 @@
* 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. *
***************************************************************************/
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif
#ifdef HAVE_GOMP
#include <omp.h>
#endif
#include <iostream>
#include <fstream>
using namespace std;
@ -139,8 +147,19 @@ Double_t PRunSingleHisto::CalcChiSquare(const std::vector<Double_t>& par)
}
// calculate chi square
Double_t time;
for (UInt_t i=0; i<fData.GetValue()->size(); i++) {
Double_t time(1.0);
Int_t i, N(static_cast<Int_t>(fData.GetValue()->size()));
// Calculate the theory function once to ensure one function evaluation for the current set of parameters.
// This is needed for the LF and user functions where some non-thread-save calculations only need to be calculated once
// for a given set of parameters---which should be done outside of the parallelized loop.
// For all other functions it means a tiny and acceptable overhead.
time = fTheory->Func(time, par, fFuncValues);
#ifdef HAVE_GOMP
#pragma omp parallel for default(shared) private(i,time,diff) schedule(dynamic,N/(2*omp_get_num_procs())) reduction(+:chisq)
#endif
for (i=0; i < N; ++i) {
time = fData.GetDataTimeStart() + (Double_t)i*fData.GetDataTimeStep();
if ((time>=fFitStartTime) && (time<=fFitEndTime)) {
diff = fData.GetValue()->at(i) -
@ -212,12 +231,25 @@ Double_t PRunSingleHisto::CalcMaxLikelihood(const std::vector<Double_t>& par)
// calculate maximum log likelihood
Double_t theo;
Double_t data;
Double_t time;
Double_t time(1.0);
Int_t i, N(static_cast<Int_t>(fData.GetValue()->size()));
// norm is needed since there is no simple scaling like in chisq case to get the correct Max.Log.Likelihood value when normlizing N(t) to 1/ns
Double_t normalizer = 1.0;
if (fScaleN0AndBkg)
normalizer = fRunInfo->GetPacking() * (fTimeResolution * 1.0e3);
for (UInt_t i=0; i<fData.GetValue()->size(); i++) {
// Calculate the theory function once to ensure one function evaluation for the current set of parameters.
// This is needed for the LF and user functions where some non-thread-save calculations only need to be calculated once
// for a given set of parameters---which should be done outside of the parallelized loop.
// For all other functions it means a tiny and acceptable overhead.
time = fTheory->Func(time, par, fFuncValues);
#ifdef HAVE_GOMP
#pragma omp parallel for default(shared) private(i,time,theo,data) schedule(dynamic,N/(2*omp_get_num_procs())) reduction(-:mllh)
#endif
for (i=0; i < N; ++i) {
time = fData.GetDataTimeStart() + (Double_t)i*fData.GetDataTimeStep();
if ((time>=fFitStartTime) && (time<=fFitEndTime)) {
// calculate theory for the given parameter set