portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-10-31 20:14:32 +00:00
Code Releases Activity

Offload more loops

Browse Source
This commit is contained in:
Peter Boyle
2020-06-10 12:57:55 -04:00
parent e97f3688db
commit 8720aecb80
3 changed files with 29 additions and 108 deletions
Download Patch File Download Diff File Expand all files Collapse all files

37
Grid/qcd/action/fermion/WilsonCloverFermion.h
View File

@@ -258,15 +258,16 @@ private:
CloverFieldType CloverTermDagEven, CloverTermDagOdd; // Clover term Dag EO
CloverFieldType CloverTermInvDagEven, CloverTermInvDagOdd; // Clover term Inv Dag EO
public:
// eventually these can be compressed into 6x6 blocks instead of the 12x12
// using the DeGrand-Rossi basis for the gamma matrices
CloverFieldType fillCloverYZ(const GaugeLinkField &F)
{
CloverFieldType T(F.Grid());
T = Zero();
autoView(T_v,T,CpuWrite);
autoView(F_v,F,CpuRead);
thread_for(i, CloverTerm.Grid()->oSites(),
autoView(T_v,T,AcceleratorWrite);
autoView(F_v,F,AcceleratorRead);
accelerator_for(i, CloverTerm.Grid()->oSites(),1,
{
T_v[i]()(0, 1) = timesMinusI(F_v[i]()());
T_v[i]()(1, 0) = timesMinusI(F_v[i]()());
@@ -282,9 +283,9 @@ private:
CloverFieldType T(F.Grid());
T = Zero();
autoView(T_v, T,CpuWrite);
autoView(F_v, F,CpuRead);
thread_for(i, CloverTerm.Grid()->oSites(),
autoView(T_v, T,AcceleratorWrite);
autoView(F_v, F,AcceleratorRead);
accelerator_for(i, CloverTerm.Grid()->oSites(),1,
{
T_v[i]()(0, 1) = -F_v[i]()();
T_v[i]()(1, 0) = F_v[i]()();
@@ -300,9 +301,9 @@ private:
CloverFieldType T(F.Grid());
T = Zero();
autoView(T_v,T,CpuWrite);
autoView(F_v,F,CpuRead);
thread_for(i, CloverTerm.Grid()->oSites(),
autoView(T_v,T,AcceleratorWrite);
autoView(F_v,F,AcceleratorRead);
accelerator_for(i, CloverTerm.Grid()->oSites(),1,
{
T_v[i]()(0, 0) = timesMinusI(F_v[i]()());
T_v[i]()(1, 1) = timesI(F_v[i]()());
@@ -318,9 +319,9 @@ private:
CloverFieldType T(F.Grid());
T = Zero();
autoView( T_v , T, CpuWrite);
autoView( F_v , F, CpuRead);
thread_for(i, CloverTerm.Grid()->oSites(),
autoView( T_v , T, AcceleratorWrite);
autoView( F_v , F, AcceleratorRead);
accelerator_for(i, CloverTerm.Grid()->oSites(),1,
{
T_v[i]()(0, 1) = timesI(F_v[i]()());
T_v[i]()(1, 0) = timesI(F_v[i]()());
@@ -336,9 +337,9 @@ private:
CloverFieldType T(F.Grid());
T = Zero();
autoView( T_v ,T,CpuWrite);
autoView( F_v ,F,CpuRead);
thread_for(i, CloverTerm.Grid()->oSites(),
autoView( T_v ,T,AcceleratorWrite);
autoView( F_v ,F,AcceleratorRead);
accelerator_for(i, CloverTerm.Grid()->oSites(),1,
{
T_v[i]()(0, 1) = -(F_v[i]()());
T_v[i]()(1, 0) = (F_v[i]()());
@@ -355,9 +356,9 @@ private:
T = Zero();
autoView( T_v , T,CpuWrite);
autoView( F_v , F,CpuRead);
thread_for(i, CloverTerm.Grid()->oSites(),
autoView( T_v , T,AcceleratorWrite);
autoView( F_v , F,AcceleratorRead);
accelerator_for(i, CloverTerm.Grid()->oSites(),1,
{
T_v[i]()(0, 0) = timesI(F_v[i]()());
T_v[i]()(1, 1) = timesMinusI(F_v[i]()());

16
Grid/qcd/action/fermion/WilsonImpl.h
View File

@@ -106,10 +106,10 @@ public:
const _SpinorField & phi,
int mu)
{
autoView( out_v, out, CpuWrite);
autoView( phi_v, phi, CpuRead);
autoView( Umu_v, Umu, CpuRead);
thread_for(sss,out.Grid()->oSites(),{
autoView( out_v, out, AcceleratorWrite);
autoView( phi_v, phi, AcceleratorRead);
autoView( Umu_v, Umu, AcceleratorRead);
accelerator_for(sss,out.Grid()->oSites(),1,{
multLink(out_v[sss],Umu_v[sss],phi_v[sss],mu);
});
}
@@ -192,10 +192,10 @@ public:
GaugeLinkField tmp(mat.Grid());
tmp = Zero();
{
autoView( tmp_v , tmp, CpuWrite);
autoView( Btilde_v , Btilde, CpuRead);
autoView( Atilde_v , Atilde, CpuRead);
thread_for(sss,tmp.Grid()->oSites(),{
autoView( tmp_v , tmp, AcceleratorWrite);
autoView( Btilde_v , Btilde, AcceleratorRead);
autoView( Atilde_v , Atilde, AcceleratorRead);
accelerator_for(sss,tmp.Grid()->oSites(),1,{
int sU=sss;
for(int s=0;s<Ls;s++){
int sF = s+Ls*sU;

84
Grid/qcd/action/fermion/implementation/WilsonFermionImplementation.h
View File

@@ -467,32 +467,7 @@ void WilsonFermion<Impl>::ContractConservedCurrent(PropagatorField &q_in_1,
conformable(_grid, q_in_1.Grid());
conformable(_grid, q_in_2.Grid());
conformable(_grid, q_out.Grid());
#if 0
PropagatorField tmp1(_grid), tmp2(_grid);
q_out = Zero();
// Forward, need q1(x + mu), q2(x). Backward, need q1(x), q2(x + mu).
// Inefficient comms method but not performance critical.
tmp1 = Cshift(q_in_1, mu, 1);
tmp2 = Cshift(q_in_2, mu, 1);
autoView( tmp1_v , tmp1, CpuWrite);
autoView( tmp2_v , tmp2, CpuWrite);
autoView( q_in_1_v,q_in_1, CpuRead);
autoView( q_in_2_v,q_in_2, CpuRead);
autoView( q_out_v , q_out, CpuRead);
autoView( Umu_v , Umu, CpuRead);
thread_for(sU, Umu.Grid()->oSites(),{
Kernels::ContractConservedCurrentSiteFwd(tmp1_v[sU],
q_in_2_v[sU],
q_out_v[sU],
Umu_v, sU, mu);
Kernels::ContractConservedCurrentSiteBwd(q_in_1_v[sU],
tmp2_v[sU],
q_out_v[sU],
Umu_v, sU, mu);
});
#else
#endif
assert(0);
}
@@ -508,62 +483,7 @@ void WilsonFermion<Impl>::SeqConservedCurrent(PropagatorField &q_in,
{
conformable(_grid, q_in.Grid());
conformable(_grid, q_out.Grid());
#if 0
// Lattice<iSinglet<Simd>> ph(_grid), coor(_grid);
Complex i(0.0,1.0);
PropagatorField tmpFwd(_grid), tmpBwd(_grid), tmp(_grid);
unsigned int tshift = (mu == Tp) ? 1 : 0;
unsigned int LLt = GridDefaultLatt()[Tp];
q_out = Zero();
LatticeInteger coords(_grid);
LatticeCoordinate(coords, Tp);
// Need q(x + mu) and q(x - mu).
tmp = Cshift(q_in, mu, 1);
tmpFwd = tmp*lattice_cmplx;
tmp = lattice_cmplx*q_in;
tmpBwd = Cshift(tmp, mu, -1);
autoView( coords_v , coords, CpuRead);
autoView( tmpFwd_v , tmpFwd, CpuRead);
autoView( tmpBwd_v , tmpBwd, CpuRead);
autoView( Umu_v , Umu, CpuRead);
autoView( q_out_v , q_out, CpuWrite);
thread_for(sU, Umu.Grid()->oSites(), {
// Compute the sequential conserved current insertion only if our simd
// object contains a timeslice we need.
vPredicate t_mask;
t_mask() = ((coords_v[sU] >= tmin) && (coords_v[sU] <= tmax));
Integer timeSlices = Reduce(t_mask());
if (timeSlices > 0) {
Kernels::SeqConservedCurrentSiteFwd(tmpFwd_v[sU],
q_out_v[sU],
Umu_v, sU, mu, t_mask);
}
// Repeat for backward direction.
t_mask() = ((coords_v[sU] >= (tmin + tshift)) &&
(coords_v[sU] <= (tmax + tshift)));
//if tmax = LLt-1 (last timeslice) include timeslice 0 if the time is shifted (mu=3)
unsigned int t0 = 0;
if((tmax==LLt-1) && (tshift==1)) t_mask() = (t_mask() || (coords_v[sU] == t0 ));
timeSlices = Reduce(t_mask());
if (timeSlices > 0) {
Kernels::SeqConservedCurrentSiteBwd(tmpBwd_v[sU],
q_out_v[sU],
Umu_v, sU, mu, t_mask);
}
});
#else
#endif
assert(0);
}
NAMESPACE_END(Grid);