portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-15 14:27:06 +01:00
Code Releases Activity

Adding reproducibility tests

Browse Source
This commit is contained in:
Guido Cossu
2016-11-21 09:52:07 +00:00
parent 036ec31c48
commit f1908c7bc9
2 changed files with 85 additions and 45 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
lib/algorithms/iterative/ConjugateGradient.h
View File

@ -33,17 +33,21 @@ directory
namespace Grid {
struct CG_state{
struct CG_state {
bool do_repro;
std::vector<RealD> residuals;
CG_state(){
CG_state() {
do_repro = false;
residuals.clear();}
residuals.clear();
}
void reset(){
do_repro = false;
residuals.clear();
}
};
/////////////////////////////////////////////////////////////
// Base classes for iterative processes based on operators
// single input vec, single output vec.
@ -57,7 +61,7 @@ class ConjugateGradient : public OperatorFunction<Field> {
RealD Tolerance;
Integer MaxIterations;
bool ReproTest;
CG_state CGState;
CG_state CGState;//to check reproducibility by repeating the CG
ConjugateGradient(RealD tol, Integer maxit, bool err_on_no_conv = true,
bool ReproducibilityTest = false)
@ -199,8 +203,7 @@ class ConjugateGradient : public OperatorFunction<Field> {
}
// Clear state
CGState.residuals.clear();
CGState.do_repro = false;
CGState.reset();
return;
}
}

43
lib/lattice/Lattice_reduction.h
View File

@ -31,6 +31,25 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#define GRID_LATTICE_REDUCTION_H
namespace Grid {
template <class T>
struct ReproducibilityState {
int n_call;
bool do_check;
std::vector<std::vector<T, alignedAllocator<T> > > th_states;
void reset(){
th_states.clear();
do_check = false;
n_call = 0;
}
ReproducibilityState(){
reset();
}
};
#ifdef GRID_WARN_SUBOPTIMAL
#warning "Optimisation alert all these reduction loops are NOT threaded "
#endif
@ -43,6 +62,9 @@ namespace Grid {
return std::real(nrm);
}
template<class vobj>
inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right)
{
@ -57,7 +79,9 @@ namespace Grid {
sumarray[i]=zero;
}
PARALLEL_FOR_LOOP
// accumulation done in the same precision ad vobj...
// may need to froce higher precision
PARALLEL_FOR_LOOP
for(int thr=0;thr<grid->SumArraySize();thr++){
int nwork, mywork, myoff;
GridThread::GetWork(left._grid->oSites(),thr,mywork,myoff);
@ -69,6 +93,19 @@ PARALLEL_FOR_LOOP
sumarray[thr]=TensorRemove(vnrm) ;
}
/*
if (repr.do_check){
if (sumarray!=repr.th_states[n_call]){
std::cout << GridLogMessage << "Reproducibility failure on node " << grid->ThisRank() << std::endl;
exit(1);
}
}
} else {
repr.th_states.push_back(sumarray);//save threads state
repr.n_call +=1;
}
*/
vector_type vvnrm; vvnrm=zero; // sum across threads
for(int i=0;i<grid->SumArraySize();i++){
vvnrm = vvnrm+sumarray[i];
@ -114,7 +151,7 @@ PARALLEL_FOR_LOOP
sumarray[i]=zero;
}
PARALLEL_FOR_LOOP
PARALLEL_FOR_LOOP
for(int thr=0;thr<grid->SumArraySize();thr++){
int nwork, mywork, myoff;
GridThread::GetWork(grid->oSites(),thr,mywork,myoff);
@ -146,7 +183,7 @@ PARALLEL_FOR_LOOP
template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<typename vobj::scalar_object> &result,int orthogdim)
{
{
typedef typename vobj::scalar_object sobj;
GridBase *grid = Data._grid;
assert(grid!=NULL);