portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-17 07:17:06 +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

42
Grid/algorithms/CoarsenedMatrix.h
View File

@ -186,10 +186,10 @@ public:
hermop.HermOp(*Tn,y);
auto y_v = y.View();
auto Tn_v = Tn->View();
auto Tnp_v = Tnp->View();
auto Tnm_v = Tnm->View();
auto y_v = y.View(AcceleratorWrite);
auto Tn_v = Tn->View(AcceleratorWrite);
auto Tnp_v = Tnp->View(AcceleratorWrite);
auto Tnm_v = Tnm->View(AcceleratorWrite);
const int Nsimd = CComplex::Nsimd();
accelerator_forNB(ss, FineGrid->oSites(), Nsimd, {
coalescedWrite(y_v[ss],xscale*y_v(ss)+mscale*Tn_v(ss));
@ -264,12 +264,12 @@ public:
Stencil.HaloExchange(in,compressor);
comms_usec += usecond();
auto in_v = in.View();
auto out_v = out.View();
auto in_v = in.View(AcceleratorRead);
auto out_v = out.View(AcceleratorWrite);
typedef LatticeView<Cobj> Aview;
Vector<Aview> AcceleratorViewContainer;
for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer.push_back(A[p].View());
for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer.push_back(A[p].View(AcceleratorRead));
Aview *Aview_p = & AcceleratorViewContainer[0];
const int Nsimd = CComplex::Nsimd();
@ -343,11 +343,11 @@ public:
typedef LatticeView<Cobj> Aview;
Vector<Aview> AcceleratorViewContainer;
for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer.push_back(A[p].View());
for(int p=0;p<geom.npoint;p++) AcceleratorViewContainer.push_back(A[p].View(AcceleratorRead));
Aview *Aview_p = & AcceleratorViewContainer[0];
auto out_v = out.View();
auto in_v = in.View();
auto out_v = out.View(AcceleratorWrite);
auto in_v = in.View(AcceleratorRead);
const int Nsimd = CComplex::Nsimd();
typedef decltype(coalescedRead(in_v[0])) calcVector;
@ -542,10 +542,10 @@ public:
blockMaskedInnerProduct(oZProj,omask,Subspace.subspace[j],Mphi);
auto iZProj_v = iZProj.View() ;
auto oZProj_v = oZProj.View() ;
auto A_p = A[p].View();
auto A_self = A[self_stencil].View();
auto iZProj_v = iZProj.View(AcceleratorRead) ;
auto oZProj_v = oZProj.View(AcceleratorRead) ;
auto A_p = A[p].View(AcceleratorWrite);
auto A_self = A[self_stencil].View(AcceleratorWrite);
accelerator_for(ss, Grid()->oSites(), Fobj::Nsimd(),{ coalescedWrite(A_p[ss](j,i),oZProj_v(ss)); });
// if( disp!= 0 ) { accelerator_for(ss, Grid()->oSites(), Fobj::Nsimd(),{ coalescedWrite(A_p[ss](j,i),oZProj_v(ss)); });}
@ -563,11 +563,11 @@ public:
mult(tmp,phi,oddmask ); linop.Op(tmp,Mphio);
{
auto tmp_ = tmp.View();
auto evenmask_ = evenmask.View();
auto oddmask_ = oddmask.View();
auto Mphie_ = Mphie.View();
auto Mphio_ = Mphio.View();
auto tmp_ = tmp.View(AcceleratorWrite);
auto evenmask_ = evenmask.View(AcceleratorRead);
auto oddmask_ = oddmask.View(AcceleratorRead);
auto Mphie_ = Mphie.View(AcceleratorRead);
auto Mphio_ = Mphio.View(AcceleratorRead);
accelerator_for(ss, FineGrid->oSites(), Fobj::Nsimd(),{
coalescedWrite(tmp_[ss],evenmask_(ss)*Mphie_(ss) + oddmask_(ss)*Mphio_(ss));
});
@ -575,8 +575,8 @@ public:
blockProject(SelfProj,tmp,Subspace.subspace);
auto SelfProj_ = SelfProj.View();
auto A_self = A[self_stencil].View();
auto SelfProj_ = SelfProj.View(AcceleratorRead);
auto A_self = A[self_stencil].View(AcceleratorWrite);
accelerator_for(ss, Grid()->oSites(), Fobj::Nsimd(),{
for(int j=0;j<nbasis;j++){

3
Grid/algorithms/FFT.h
View File

@ -1,4 +1,3 @@
/*************************************************************************************
Grid physics library, www.github.com/paboyle/Grid
@ -191,7 +190,7 @@ public:
typedef typename sobj::scalar_type scalar;
Lattice<sobj> pgbuf(&pencil_g);
auto pgbuf_v = pgbuf.View();
auto pgbuf_v = pgbuf.View(CpuWrite);
typedef typename FFTW<scalar>::FFTW_scalar FFTW_scalar;
typedef typename FFTW<scalar>::FFTW_plan FFTW_plan;

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];