portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-17 15:27:06 +01:00
Code Releases Activity

Make view specify where and drive data motion - first cut.

Browse Source 
This is a compile tiime option --enable-unified=yes/no
This commit is contained in:
Peter Boyle
2020-05-21 16:13:16 -04:00
parent ebb60330c9
commit 7860a50f70
48 changed files with 688 additions and 718 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
Grid/algorithms/iterative/BiCGSTAB.h
View File

@ -122,9 +122,9 @@ class BiCGSTAB : public OperatorFunction<Field>
LinearCombTimer.Start();
bo = beta * omega;
auto p_v = p.View();
auto r_v = r.View();
auto v_v = v.View();
auto p_v = p.View(AcceleratorWrite);
auto r_v = r.View(AcceleratorWrite);
auto v_v = v.View(AcceleratorWrite);
accelerator_for(ss, p_v.size(), Field::vector_object::Nsimd(),{
coalescedWrite(p_v[ss], beta*p_v(ss) - bo*v_v(ss) + r_v(ss));
});
@ -142,13 +142,13 @@ class BiCGSTAB : public OperatorFunction<Field>
alpha = rho / Calpha.real();
LinearCombTimer.Start();
auto h_v = h.View();
auto psi_v = psi.View();
auto h_v = h.View(AcceleratorWrite);
auto psi_v = psi.View(AcceleratorWrite);
accelerator_for(ss, h_v.size(), Field::vector_object::Nsimd(),{
coalescedWrite(h_v[ss], alpha*p_v(ss) + psi_v(ss));
});
auto s_v = s.View();
auto s_v = s.View(AcceleratorWrite);
accelerator_for(ss, s_v.size(), Field::vector_object::Nsimd(),{
coalescedWrite(s_v[ss], -alpha*v_v(ss) + r_v(ss));
});
@ -166,7 +166,7 @@ class BiCGSTAB : public OperatorFunction<Field>
omega = Comega.real() / norm2(t);
LinearCombTimer.Start();
auto t_v = t.View();
auto t_v = t.View(AcceleratorWrite);
accelerator_for(ss, psi_v.size(), Field::vector_object::Nsimd(),{
coalescedWrite(psi_v[ss], h_v(ss) + omega * s_v(ss));
coalescedWrite(r_v[ss], -omega * t_v(ss) + s_v(ss));

6
Grid/algorithms/iterative/ConjugateGradient.h
View File

@ -140,9 +140,9 @@ public:
b = cp / c;
LinearCombTimer.Start();
auto psi_v = psi.View();
auto p_v = p.View();
auto r_v = r.View();
auto psi_v = psi.View(AcceleratorWrite);
auto p_v = p.View(AcceleratorWrite);
auto r_v = r.View(AcceleratorWrite);
accelerator_for(ss,p_v.size(), Field::vector_object::Nsimd(),{
coalescedWrite(psi_v[ss], a * p_v(ss) + psi_v(ss));
coalescedWrite(p_v[ss] , b * p_v(ss) + r_v (ss));

14
Grid/algorithms/iterative/ImplicitlyRestartedLanczos.h
View File

@ -57,17 +57,17 @@ void basisOrthogonalize(std::vector<Field> &basis,Field &w,int k)
template<class Field>
void basisRotate(std::vector<Field> &basis,Eigen::MatrixXd& Qt,int j0, int j1, int k0,int k1,int Nm)
{
typedef decltype(basis[0].View()) View;
auto tmp_v = basis[0].View();
typedef decltype(basis[0].View(CpuWrite)) View;
auto tmp_v = basis[0].View(CpuWrite);
Vector<View> basis_v(basis.size(),tmp_v);
View *basis_vp = &basis_v[0];
typedef typename Field::vector_object vobj;
GridBase* grid = basis[0].Grid();
for(int k=0;k<basis.size();k++){
basis_v[k] = basis[k].View();
basis_v[k] = basis[k].View(CpuWrite);
}
#if 0
#if 1
std::vector < vobj , commAllocator<vobj> > Bt(thread_max() * Nm); // Thread private
thread_region
{
@ -149,16 +149,16 @@ void basisRotate(std::vector<Field> &basis,Eigen::MatrixXd& Qt,int j0, int j1, i
template<class Field>
void basisRotateJ(Field &result,std::vector<Field> &basis,Eigen::MatrixXd& Qt,int j, int k0,int k1,int Nm)
{
typedef decltype(basis[0].View()) View;
typedef decltype(basis[0].View(AcceleratorWrite)) View;
typedef typename Field::vector_object vobj;
GridBase* grid = basis[0].Grid();
result.Checkerboard() = basis[0].Checkerboard();
auto result_v=result.View();
auto result_v=result.View(AcceleratorWrite);
Vector<View> basis_v(basis.size(),result_v);
View * basis_vp = &basis_v[0];
for(int k=0;k<basis.size();k++){
basis_v[k] = basis[k].View();
basis_v[k] = basis[k].View(AcceleratorRead);
}
Vector<double> Qt_jv(Nm);
double * Qt_j = & Qt_jv[0];