portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-19 00:07:05 +01:00
Code Releases Activity

Merge branch 'develop' into feature/scidac-wp1

Browse Source
This commit is contained in:
Peter Boyle
2023-10-02 17:24:55 -04:00
parent 36ae6e5aba afc316f501
commit 8a70314f54
90 changed files with 5007 additions and 1615 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
Grid/lattice/Lattice.h
View File

@ -47,3 +47,4 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#include <Grid/lattice/Lattice_transfer.h>
#include <Grid/lattice/Lattice_basis.h>
#include <Grid/lattice/Lattice_crc.h>
#include <Grid/lattice/PaddedCell.h>

4
Grid/lattice/Lattice_ET.h
View File

@ -345,7 +345,9 @@ GridUnopClass(UnaryNot, Not(a));
GridUnopClass(UnaryTrace, trace(a));
GridUnopClass(UnaryTranspose, transpose(a));
GridUnopClass(UnaryTa, Ta(a));
GridUnopClass(UnarySpTa, SpTa(a));
GridUnopClass(UnaryProjectOnGroup, ProjectOnGroup(a));
GridUnopClass(UnaryProjectOnSpGroup, ProjectOnSpGroup(a));
GridUnopClass(UnaryTimesI, timesI(a));
GridUnopClass(UnaryTimesMinusI, timesMinusI(a));
GridUnopClass(UnaryAbs, abs(a));
@ -456,7 +458,9 @@ GRID_DEF_UNOP(operator!, UnaryNot);
GRID_DEF_UNOP(trace, UnaryTrace);
GRID_DEF_UNOP(transpose, UnaryTranspose);
GRID_DEF_UNOP(Ta, UnaryTa);
GRID_DEF_UNOP(SpTa, UnarySpTa);
GRID_DEF_UNOP(ProjectOnGroup, UnaryProjectOnGroup);
GRID_DEF_UNOP(ProjectOnSpGroup, UnaryProjectOnSpGroup);
GRID_DEF_UNOP(timesI, UnaryTimesI);
GRID_DEF_UNOP(timesMinusI, UnaryTimesMinusI);
GRID_DEF_UNOP(abs, UnaryAbs); // abs overloaded in cmath C++98; DON'T do the

59
Grid/lattice/Lattice_trace.h
View File

@ -66,6 +66,65 @@ inline auto TraceIndex(const Lattice<vobj> &lhs) -> Lattice<decltype(traceIndex<
return ret;
};
template<int N, class Vec>
Lattice<iScalar<iScalar<iScalar<Vec> > > > Determinant(const Lattice<iScalar<iScalar<iMatrix<Vec, N> > > > &Umu)
{
GridBase *grid=Umu.Grid();
auto lvol = grid->lSites();
Lattice<iScalar<iScalar<iScalar<Vec> > > > ret(grid);
typedef typename Vec::scalar_type scalar;
autoView(Umu_v,Umu,CpuRead);
autoView(ret_v,ret,CpuWrite);
thread_for(site,lvol,{
Eigen::MatrixXcd EigenU = Eigen::MatrixXcd::Zero(N,N);
Coordinate lcoor;
grid->LocalIndexToLocalCoor(site, lcoor);
iScalar<iScalar<iMatrix<scalar, N> > > Us;
peekLocalSite(Us, Umu_v, lcoor);
for(int i=0;i<N;i++){
for(int j=0;j<N;j++){
scalar tmp= Us()()(i,j);
ComplexD ztmp(real(tmp),imag(tmp));
EigenU(i,j)=ztmp;
}}
ComplexD detD = EigenU.determinant();
typename Vec::scalar_type det(detD.real(),detD.imag());
pokeLocalSite(det,ret_v,lcoor);
});
return ret;
}
template<int N>
Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > Inverse(const Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > &Umu)
{
GridBase *grid=Umu.Grid();
auto lvol = grid->lSites();
Lattice<iScalar<iScalar<iMatrix<vComplexD, N> > > > ret(grid);
autoView(Umu_v,Umu,CpuRead);
autoView(ret_v,ret,CpuWrite);
thread_for(site,lvol,{
Eigen::MatrixXcd EigenU = Eigen::MatrixXcd::Zero(N,N);
Coordinate lcoor;
grid->LocalIndexToLocalCoor(site, lcoor);
iScalar<iScalar<iMatrix<ComplexD, N> > > Us;
iScalar<iScalar<iMatrix<ComplexD, N> > > Ui;
peekLocalSite(Us, Umu_v, lcoor);
for(int i=0;i<N;i++){
for(int j=0;j<N;j++){
EigenU(i,j) = Us()()(i,j);
}}
Eigen::MatrixXcd EigenUinv = EigenU.inverse();
for(int i=0;i<N;i++){
for(int j=0;j<N;j++){
Ui()()(i,j) = EigenUinv(i,j);
}}
pokeLocalSite(Ui,ret_v,lcoor);
});
return ret;
}
NAMESPACE_END(Grid);
#endif

126
Grid/lattice/Lattice_transfer.h
View File

@ -697,8 +697,68 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
for(int d=0;d<nd;d++){
assert(Fg->_processors[d] == Tg->_processors[d]);
}
// the above should guarantee that the operations are local
#if 1
size_t nsite = 1;
for(int i=0;i<nd;i++) nsite *= RegionSize[i];
size_t tbytes = 4*nsite*sizeof(int);
int *table = (int*)malloc(tbytes);
thread_for(idx, nsite, {
Coordinate from_coor, to_coor;
size_t rem = idx;
for(int i=0;i<nd;i++){
size_t base_i = rem % RegionSize[i]; rem /= RegionSize[i];
from_coor[i] = base_i + FromLowerLeft[i];
to_coor[i] = base_i + ToLowerLeft[i];
}
int foidx = Fg->oIndex(from_coor);
int fiidx = Fg->iIndex(from_coor);
int toidx = Tg->oIndex(to_coor);
int tiidx = Tg->iIndex(to_coor);
int* tt = table + 4*idx;
tt[0] = foidx;
tt[1] = fiidx;
tt[2] = toidx;
tt[3] = tiidx;
});
int* table_d = (int*)acceleratorAllocDevice(tbytes);
acceleratorCopyToDevice(table,table_d,tbytes);
typedef typename vobj::vector_type vector_type;
typedef typename vobj::scalar_type scalar_type;
autoView(from_v,From,AcceleratorRead);
autoView(to_v,To,AcceleratorWrite);
accelerator_for(idx,nsite,1,{
static const int words=sizeof(vobj)/sizeof(vector_type);
int* tt = table_d + 4*idx;
int from_oidx = *tt++;
int from_lane = *tt++;
int to_oidx = *tt++;
int to_lane = *tt;
const vector_type* from = (const vector_type *)&from_v[from_oidx];
vector_type* to = (vector_type *)&to_v[to_oidx];
scalar_type stmp;
for(int w=0;w<words;w++){
stmp = getlane(from[w], from_lane);
putlane(to[w], stmp, to_lane);
}
});
acceleratorFreeDevice(table_d);
free(table);
#else
Coordinate ldf = Fg->_ldimensions;
Coordinate rdf = Fg->_rdimensions;
Coordinate isf = Fg->_istride;
@ -738,6 +798,8 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
#endif
}
});
#endif
}
@ -830,6 +892,8 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
}
//Insert subvolume orthogonal to direction 'orthog' with slice index 'slice_lo' from 'lowDim' onto slice index 'slice_hi' of higherDim
//The local dimensions of both 'lowDim' and 'higherDim' orthogonal to 'orthog' should be the same
template<class vobj>
void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int slice_lo,int slice_hi, int orthog)
{
@ -851,6 +915,65 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
}
}
#if 1
size_t nsite = lg->lSites()/lg->LocalDimensions()[orthog];
size_t tbytes = 4*nsite*sizeof(int);
int *table = (int*)malloc(tbytes);
thread_for(idx,nsite,{
Coordinate lcoor(nl);
Coordinate hcoor(nh);
lcoor[orthog] = slice_lo;
hcoor[orthog] = slice_hi;
size_t rem = idx;
for(int mu=0;mu<nl;mu++){
if(mu != orthog){
int xmu = rem % lg->LocalDimensions()[mu]; rem /= lg->LocalDimensions()[mu];
lcoor[mu] = hcoor[mu] = xmu;
}
}
int loidx = lg->oIndex(lcoor);
int liidx = lg->iIndex(lcoor);
int hoidx = hg->oIndex(hcoor);
int hiidx = hg->iIndex(hcoor);
int* tt = table + 4*idx;
tt[0] = loidx;
tt[1] = liidx;
tt[2] = hoidx;
tt[3] = hiidx;
});
int* table_d = (int*)acceleratorAllocDevice(tbytes);
acceleratorCopyToDevice(table,table_d,tbytes);
typedef typename vobj::vector_type vector_type;
typedef typename vobj::scalar_type scalar_type;
autoView(lowDim_v,lowDim,AcceleratorRead);
autoView(higherDim_v,higherDim,AcceleratorWrite);
accelerator_for(idx,nsite,1,{
static const int words=sizeof(vobj)/sizeof(vector_type);
int* tt = table_d + 4*idx;
int from_oidx = *tt++;
int from_lane = *tt++;
int to_oidx = *tt++;
int to_lane = *tt;
const vector_type* from = (const vector_type *)&lowDim_v[from_oidx];
vector_type* to = (vector_type *)&higherDim_v[to_oidx];
scalar_type stmp;
for(int w=0;w<words;w++){
stmp = getlane(from[w], from_lane);
putlane(to[w], stmp, to_lane);
}
});
acceleratorFreeDevice(table_d);
free(table);
#else
// the above should guarantee that the operations are local
autoView(lowDimv,lowDim,CpuRead);
autoView(higherDimv,higherDim,CpuWrite);
@ -866,6 +989,7 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
pokeLocalSite(s,higherDimv,hcoor);
}
});
#endif
}

52
Grid/lattice/PaddedCell.h
View File

@ -26,14 +26,32 @@ Author: Peter Boyle pboyle@bnl.gov
/* END LEGAL */
#pragma once
#include<Grid/cshift/Cshift.h>
NAMESPACE_BEGIN(Grid);
//Allow the user to specify how the C-shift is performed, e.g. to respect the appropriate boundary conditions
template<typename vobj>
struct CshiftImplBase{
virtual Lattice<vobj> Cshift(const Lattice<vobj> &in, int dir, int shift) const = 0;
virtual ~CshiftImplBase(){}
};
template<typename vobj>
struct CshiftImplDefault: public CshiftImplBase<vobj>{
Lattice<vobj> Cshift(const Lattice<vobj> &in, int dir, int shift) const override{ return Grid::Cshift(in,dir,shift); }
};
template<typename Gimpl>
struct CshiftImplGauge: public CshiftImplBase<typename Gimpl::GaugeLinkField::vector_object>{
typename Gimpl::GaugeLinkField Cshift(const typename Gimpl::GaugeLinkField &in, int dir, int shift) const override{ return Gimpl::CshiftLink(in,dir,shift); }
};
class PaddedCell {
public:
GridCartesian * unpadded_grid;
int dims;
int depth;
std::vector<GridCartesian *> grids;
~PaddedCell()
{
DeleteGrids();
@ -78,7 +96,7 @@ public:
}
};
template<class vobj>
inline Lattice<vobj> Extract(Lattice<vobj> &in)
inline Lattice<vobj> Extract(const Lattice<vobj> &in) const
{
Lattice<vobj> out(unpadded_grid);
@ -89,19 +107,19 @@ public:
return out;
}
template<class vobj>
inline Lattice<vobj> Exchange(Lattice<vobj> &in)
inline Lattice<vobj> Exchange(const Lattice<vobj> &in, const CshiftImplBase<vobj> &cshift = CshiftImplDefault<vobj>()) const
{
GridBase *old_grid = in.Grid();
int dims = old_grid->Nd();
Lattice<vobj> tmp = in;
for(int d=0;d<dims;d++){
tmp = Expand(d,tmp); // rvalue && assignment
tmp = Expand(d,tmp,cshift); // rvalue && assignment
}
return tmp;
}
// expand up one dim at a time
template<class vobj>
inline Lattice<vobj> Expand(int dim,Lattice<vobj> &in)
inline Lattice<vobj> Expand(int dim, const Lattice<vobj> &in, const CshiftImplBase<vobj> &cshift = CshiftImplDefault<vobj>()) const
{
GridBase *old_grid = in.Grid();
GridCartesian *new_grid = grids[dim];//These are new grids
@ -112,21 +130,41 @@ public:
if(dim==0) conformable(old_grid,unpadded_grid);
else conformable(old_grid,grids[dim-1]);
// std::cout << " dim "<<dim<<" local "<<local << " padding to "<<plocal<<std::endl;
std::cout << " dim "<<dim<<" local "<<local << " padding to "<<plocal<<std::endl;
double tins=0, tshift=0;
// Middle bit
double t = usecond();
for(int x=0;x<local[dim];x++){
InsertSliceLocal(in,padded,x,depth+x,dim);
}
tins += usecond() - t;
// High bit
shifted = Cshift(in,dim,depth);
t = usecond();
shifted = cshift.Cshift(in,dim,depth);
tshift += usecond() - t;
t=usecond();
for(int x=0;x<depth;x++){
InsertSliceLocal(shifted,padded,local[dim]-depth+x,depth+local[dim]+x,dim);
}
tins += usecond() - t;
// Low bit
shifted = Cshift(in,dim,-depth);
t = usecond();
shifted = cshift.Cshift(in,dim,-depth);
tshift += usecond() - t;
t = usecond();
for(int x=0;x<depth;x++){
InsertSliceLocal(shifted,padded,x,x,dim);
}
tins += usecond() - t;
std::cout << GridLogPerformance << "PaddedCell::Expand timings: cshift:" << tshift/1000 << "ms, insert-slice:" << tins/1000 << "ms" << std::endl;
return padded;
}