KernelFunctions* HistogramOnGrid::getKernelAndNeighbors( std::vector<double>& point, unsigned& num_neigh, std::vector<unsigned>& neighbors ) const {
if( discrete ) {
plumed_assert( getType()=="flat" );
num_neigh=1;
for(unsigned i=0; i<dimension; ++i) point[i] += 0.5*dx[i];
neighbors[0] = getIndex( point );
return NULL;
} else if( getType()=="flat" ) {
KernelFunctions* kernel = new KernelFunctions( point, bandwidths, kerneltype, false, 1.0, true );
getNeighbors( kernel->getCenter(), nneigh, num_neigh, neighbors );
return kernel;
} else {
num_neigh = getNumberOfPoints();
if( neighbors.size()!=getNumberOfPoints() ) neighbors.resize( getNumberOfPoints() );
for(unsigned i=0; i<getNumberOfPoints(); ++i) neighbors[i]=i;
}
return NULL;
}
void BiasRepresentation::pushKernel( IFile *ifile ){
KernelFunctions *kk;
// here below the reading of the kernel is completely hidden
if(histosigma.size()==0){
ifile->allowIgnoredFields();
kk=KernelFunctions::read(ifile,names) ;
}else{
// when doing histogram assume gaussian with a given diagonal sigma
// and neglect all the rest
kk=readFromPoint(ifile) ;
}
hills.push_back(kk);
// the bias factor is not something about the kernels but
// must be stored to keep the bias/free energy duality
string dummy; double dummyd;
if(ifile->FieldExist("biasf")){
ifile->scanField("biasf",dummy);
Tools::convert(dummy,dummyd);
}else{dummyd=1.0;}
biasf.push_back(dummyd);
// the domain does not pertain to the kernel but to the values here defined
string mins,maxs,minv,maxv,mini,maxi;mins="min_";maxs="max_";
for(int i=0 ; i<ndim; i++){
if(values[i]->isPeriodic()){
ifile->scanField(mins+names[i],minv);
ifile->scanField(maxs+names[i],maxv);
// verify that the domain is correct
values[i]->getDomain(mini,maxi);
plumed_massert(mini==minv,"the input periodicity in hills and in value definition does not match" );
plumed_massert(maxi==maxv,"the input periodicity in hills and in value definition does not match" );
}
}
// if grid is defined then it should be added on the grid
//cerr<<"now with "<<hills.size()<<endl;
if(hasgrid){
vector<unsigned> nneighb=kk->getSupport(BiasGrid_->getDx());
vector<unsigned> neighbors=BiasGrid_->getNeighbors(kk->getCenter(),nneighb);
vector<double> der(ndim);
vector<double> xx(ndim);
if(mycomm.Get_size()==1){
for(unsigned i=0;i<neighbors.size();++i){
unsigned ineigh=neighbors[i];
for(int j=0;j<ndim;++j){der[j]=0.0;}
BiasGrid_->getPoint(ineigh,xx);
// assign xx to a new vector of values
for(int j=0;j<ndim;++j){values[j]->set(xx[j]);}
double bias=kk->evaluate(values,der,true);
if(rescaledToBias){
double f=(biasf.back()-1.)/(biasf.back());
bias*=f;
for(int j=0;j<ndim;++j){der[j]*=f;}
}
BiasGrid_->addValueAndDerivatives(ineigh,bias,der);
}
} else {
unsigned stride=mycomm.Get_size();
unsigned rank=mycomm.Get_rank();
vector<double> allder(ndim*neighbors.size(),0.0);
vector<double> allbias(neighbors.size(),0.0);
vector<double> tmpder(ndim);
for(unsigned i=rank;i<neighbors.size();i+=stride){
unsigned ineigh=neighbors[i];
BiasGrid_->getPoint(ineigh,xx);
for(int j=0;j<ndim;++j){values[j]->set(xx[j]);}
allbias[i]=kk->evaluate(values,tmpder,true);
if(rescaledToBias){
double f=(biasf.back()-1.)/(biasf.back());
allbias[i]*=f;
for(int j=0;j<ndim;++j){tmpder[j]*=f;}
}
// this solution with the temporary vector is rather bad, probably better to take
// a pointer of double as it was in old gaussian
for(int j=0;j<ndim;++j){ allder[ndim*i+j]=tmpder[j];tmpder[j]=0.;}
}
mycomm.Sum(allbias);
mycomm.Sum(allder);
for(unsigned i=0;i<neighbors.size();++i){
unsigned ineigh=neighbors[i];
for(int j=0;j<ndim;++j){der[j]=allder[ndim*i+j];}
BiasGrid_->addValueAndDerivatives(ineigh,allbias[i],der);
}
}
}
}
