portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-08-21 13:41:53 +01:00
Code Releases Activity

Deprecate UVM

Browse Source
This commit is contained in:
Peter Boyle
2024-09-17 13:34:27 +00:00
parent 11be10d2c0
commit 066544281f
75 changed files with 668 additions and 1082 deletions
Download Patch File Download Diff File Expand all files Collapse all files

61
Grid/lattice/Lattice_basis.h
View File

@@ -53,36 +53,19 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
typedef decltype(basis[0]) Field;
typedef decltype(basis[0].View(AcceleratorRead)) View;
Vector<View> basis_v; basis_v.reserve(basis.size());
typedef typename std::remove_reference<decltype(basis_v[0][0])>::type vobj;
hostVector<View> h_basis_v(basis.size());
deviceVector<View> d_basis_v(basis.size());
typedef typename std::remove_reference<decltype(h_basis_v[0][0])>::type vobj;
typedef typename std::remove_reference<decltype(Qt(0,0))>::type Coeff_t;
GridBase* grid = basis[0].Grid();
for(int k=0;k<basis.size();k++){
basis_v.push_back(basis[k].View(AcceleratorWrite));
h_basis_v[k] = basis[k].View(AcceleratorWrite);
acceleratorPut(d_basis_v[k],h_basis_v[k]);
}
#if ( !(defined(GRID_CUDA) || defined(GRID_HIP) || defined(GRID_SYCL)) )
int max_threads = thread_max();
Vector < vobj > Bt(Nm * max_threads);
thread_region
{
vobj* B = &Bt[Nm * thread_num()];
thread_for_in_region(ss, grid->oSites(),{
for(int j=j0; j<j1; ++j) B[j]=0.;
for(int j=j0; j<j1; ++j){
for(int k=k0; k<k1; ++k){
B[j] +=Qt(j,k) * basis_v[k][ss];
}
}
for(int j=j0; j<j1; ++j){
basis_v[j][ss] = B[j];
}
});
}
#else
View *basis_vp = &basis_v[0];
View *basis_vp = &d_basis_v[0];
int nrot = j1-j0;
if (!nrot) // edge case not handled gracefully by Cuda
@@ -91,17 +74,19 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
uint64_t oSites =grid->oSites();
uint64_t siteBlock=(grid->oSites()+nrot-1)/nrot; // Maximum 1 additional vector overhead
Vector <vobj> Bt(siteBlock * nrot);
deviceVector <vobj> Bt(siteBlock * nrot);
auto Bp=&Bt[0];
// GPU readable copy of matrix
Vector<Coeff_t> Qt_jv(Nm*Nm);
hostVector<Coeff_t> h_Qt_jv(Nm*Nm);
deviceVector<Coeff_t> Qt_jv(Nm*Nm);
Coeff_t *Qt_p = & Qt_jv[0];
thread_for(i,Nm*Nm,{
int j = i/Nm;
int k = i%Nm;
Qt_p[i]=Qt(j,k);
h_Qt_jv[i]=Qt(j,k);
});
acceleratorCopyToDevice(&h_Qt_jv[0],Qt_p,Nm*Nm*sizeof(Coeff_t));
// Block the loop to keep storage footprint down
for(uint64_t s=0;s<oSites;s+=siteBlock){
@@ -137,9 +122,8 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
coalescedWrite(basis_vp[jj][sss],coalescedRead(Bp[ss*nrot+j]));
});
}
#endif
for(int k=0;k<basis.size();k++) basis_v[k].ViewClose();
for(int k=0;k<basis.size();k++) h_basis_v[k].ViewClose();
}
// Extract a single rotated vector
@@ -152,16 +136,19 @@ void basisRotateJ(Field &result,std::vector<Field> &basis,Eigen::MatrixXd& Qt,in
result.Checkerboard() = basis[0].Checkerboard();
Vector<View> basis_v; basis_v.reserve(basis.size());
hostVector<View> h_basis_v(basis.size());
deviceVector<View> d_basis_v(basis.size());
for(int k=0;k<basis.size();k++){
basis_v.push_back(basis[k].View(AcceleratorRead));
h_basis_v[k]=basis[k].View(AcceleratorRead);
acceleratorPut(d_basis_v[k],h_basis_v[k]);
}
vobj zz=Zero();
Vector<double> Qt_jv(Nm);
double * Qt_j = & Qt_jv[0];
for(int k=0;k<Nm;++k) Qt_j[k]=Qt(j,k);
auto basis_vp=& basis_v[0];
vobj zz=Zero();
deviceVector<double> Qt_jv(Nm);
double * Qt_j = & Qt_jv[0];
for(int k=0;k<Nm;++k) acceleratorPut(Qt_j[k],Qt(j,k));
auto basis_vp=& d_basis_v[0];
autoView(result_v,result,AcceleratorWrite);
accelerator_for(ss, grid->oSites(),vobj::Nsimd(),{
vobj zzz=Zero();
@@ -171,7 +158,7 @@ void basisRotateJ(Field &result,std::vector<Field> &basis,Eigen::MatrixXd& Qt,in
}
coalescedWrite(result_v[ss], B);
});
for(int k=0;k<basis.size();k++) basis_v[k].ViewClose();
for(int k=0;k<basis.size();k++) h_basis_v[k].ViewClose();
}
template<class Field>