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

MultiRHS solver improvements with slice operations moved into lattice and sped up.

Browse Source 
Block solver requires a lot of performance work.
This commit is contained in:
paboyle
2017-04-18 10:51:55 +01:00
parent 3141ebac10
commit 8e161152e4
9 changed files with 366 additions and 285 deletions
Download Patch File Download Diff File Expand all files Collapse all files

457
lib/lattice/Lattice_reduction.h
View File

@ -44,6 +44,7 @@ template<class vobj> inline RealD norm2(const Lattice<vobj> &arg){
ComplexD nrm = innerProduct(arg,arg);
return std::real(nrm);
}
// Double inner product
template<class vobj>
inline ComplexD innerProduct(const Lattice<vobj> &left,const Lattice<vobj> &right)
@ -101,7 +102,6 @@ inline auto sum(const LatticeTrinaryExpression<Op,T1,T2,T3> & expr)
return sum(closure(expr));
}
// FIXME precision promoted summation
template<class vobj>
inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
{
@ -141,14 +141,22 @@ inline typename vobj::scalar_object sum(const Lattice<vobj> &arg)
return ssum;
}
//////////////////////////////////////////////////////////////////////////////////////////////////////////////
// sliceSum, sliceInnerProduct, sliceAxpy, sliceNorm etc...
//////////////////////////////////////////////////////////////////////////////////////////////////////////////
template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<typename vobj::scalar_object> &result,int orthogdim)
{
///////////////////////////////////////////////////////
// FIXME precision promoted summation
// may be important for correlation functions
// But easily avoided by using double precision fields
///////////////////////////////////////////////////////
typedef typename vobj::scalar_object sobj;
GridBase *grid = Data._grid;
assert(grid!=NULL);
// FIXME
// std::cout<<GridLogMessage<<"WARNING ! SliceSum is unthreaded "<<grid->SumArraySize()<<" threads "<<std::endl;
const int Nd = grid->_ndimension;
const int Nsimd = grid->Nsimd();
@ -163,18 +171,27 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
std::vector<sobj> lsSum(ld,zero); // sum across these down to scalars
std::vector<sobj> extracted(Nsimd); // splitting the SIMD
result.resize(fd); // And then global sum to return the same vector to every node for IO to file
result.resize(fd); // And then global sum to return the same vector to every node
for(int r=0;r<rd;r++){
lvSum[r]=zero;
}
std::vector<int> coor(Nd);
int e1= grid->_slice_nblock[orthogdim];
int e2= grid->_slice_block [orthogdim];
int stride=grid->_slice_stride[orthogdim];
// sum over reduced dimension planes, breaking out orthog dir
for(int ss=0;ss<grid->oSites();ss++){
Lexicographic::CoorFromIndex(coor,ss,grid->_rdimensions);
int r = coor[orthogdim];
lvSum[r]=lvSum[r]+Data._odata[ss];
// Parallel over orthog direction
parallel_for(int r=0;r<rd;r++){
int so=r*grid->_ostride[orthogdim]; // base offset for start of plane
for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){
int ss= so+n*stride+b;
lvSum[r]=lvSum[r]+Data._odata[ss];
}
}
}
// Sum across simd lanes in the plane, breaking out orthog dir.
@ -212,32 +229,6 @@ template<class vobj> inline void sliceSum(const Lattice<vobj> &Data,std::vector<
}
}
template<class vobj>
static void sliceInnerProductVectorSlow( std::vector<ComplexD> & vec, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog)
{
// FIXME: Implementation is slow
// Look at localInnerProduct implementation,
// and do inside a site loop with block strided iterators
typedef typename vobj::scalar_object sobj;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
typedef typename vobj::tensor_reduced scalar;
typedef typename scalar::scalar_object scomplex;
int Nblock = lhs._grid->GlobalDimensions()[Orthog];
vec.resize(Nblock);
std::vector<scomplex> sip(Nblock);
Lattice<scalar> IP(lhs._grid);
IP=localInnerProduct(lhs,rhs);
sliceSum(IP,sip,Orthog);
for(int ss=0;ss<Nblock;ss++){
vec[ss] = TensorRemove(sip[ss]);
}
}
template<class vobj>
static void sliceInnerProductVector( std::vector<ComplexD> & result, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int orthogdim)
{
@ -247,8 +238,6 @@ static void sliceInnerProductVector( std::vector<ComplexD> & result, const Latti
assert(grid!=NULL);
conformable(grid,rhs._grid);
// FIXME
// std::cout<<GridLogMessage<<"WARNING ! SliceSum is unthreaded "<<grid->SumArraySize()<<" threads "<<std::endl;
const int Nd = grid->_ndimension;
const int Nsimd = grid->Nsimd();
@ -268,16 +257,18 @@ static void sliceInnerProductVector( std::vector<ComplexD> & result, const Latti
lvSum[r]=zero;
}
// sum over reduced dimension planes, breaking out orthog dir
PARALLEL_REGION {
std::vector<int> coor(Nd);
vector_type vv;
PARALLEL_FOR_LOOP_INTERN
for(int ss=0;ss<grid->oSites();ss++){
Lexicographic::CoorFromIndex(coor,ss,grid->_rdimensions);
int r = coor[orthogdim];
vv = TensorRemove(innerProduct(lhs._odata[ss],rhs._odata[ss]));
PARALLEL_CRITICAL { // ouch slow rfo thrashing atomic fp add
int e1= grid->_slice_nblock[orthogdim];
int e2= grid->_slice_block [orthogdim];
int stride=grid->_slice_stride[orthogdim];
parallel_for(int r=0;r<rd;r++){
int so=r*grid->_ostride[orthogdim]; // base offset for start of plane
for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){
int ss= so+n*stride+b;
vector_type vv = TensorRemove(innerProduct(lhs._odata[ss],rhs._odata[ss]));
lvSum[r]=lvSum[r]+vv;
}
}
@ -287,7 +278,8 @@ static void sliceInnerProductVector( std::vector<ComplexD> & result, const Latti
std::vector<int> icoor(Nd);
for(int rt=0;rt<rd;rt++){
iScalar<vector_type> temp; temp._internal = lvSum[rt];
iScalar<vector_type> temp;
temp._internal = lvSum[rt];
extract(temp,extracted);
for(int idx=0;idx<Nsimd;idx++){
@ -317,176 +309,9 @@ static void sliceInnerProductVector( std::vector<ComplexD> & result, const Latti
result[t]=gsum;
}
}
#if 0
template<class vobj>
static void sliceInnerProductVector( std::vector<ComplexD> & vec, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog)
static void sliceNorm (std::vector<RealD> &sn,const Lattice<vobj> &rhs,int Orthog)
{
// FIXME: Implementation is slow
// Look at sliceSum implementation,
// and do inside a site loop with block strided iterators
typedef typename vobj::scalar_object sobj;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
typedef typename vobj::tensor_reduced scalar;
typedef typename scalar::scalar_object scomplex;
GridBase * grid = lhs._grid;
const int Nd = grid->_ndimension;
const int Nsimd = grid->Nsimd();
assert(orthogdim >= 0);
assert(orthogdim < Nd);
int fd=grid->_fdimensions[orthogdim];
int ld=grid->_ldimensions[orthogdim];
int rd=grid->_rdimensions[orthogdim];
int Nblock = grid->GlobalDimensions()[Orthog];
int Nrblock = grid->_rdimensions[Orthog];
int Nthr = grid->SumArraySize();
std::vector<vector_type,alignedAllocator<vector_type> > sumarray(Nrblock*Nthr);
parallel_for(int thr=0;thr<grid->SumArraySize();thr++){
int nwork, mywork, myoff;
for(int rb=0;rb<Nrblock;rb++){
GridThread::GetWork((left._grid->oSites()/Nrblock),thr,mywork,myoff);
int off = rb * grid->_slice_
vector_type vnrm=zero; // private to thread; sub summation
for(int ss=myoff;ss<mywork+myoff; ss++){
vnrm = vnrm + TensorRemove(innerProductD(left._odata[ss],right._odata[ss]));
}
}
sumarray[thr+Nthr*rb]=vnrm ;
}
vec.resize(Nblock);
std::vector<scomplex> sip(Nblock);
Lattice<scalar> IP(lhs._grid);
IP=localInnerProduct(lhs,rhs);
sliceSum(IP,sip,Orthog);
for(int ss=0;ss<Nblock;ss++){
vec[ss] = TensorRemove(sip[ss]);
}
}
#endif
inline GridBase *makeSubSliceGrid(const GridBase *BlockSolverGrid,int Orthog)
{
int NN = BlockSolverGrid->_ndimension;
int nsimd = BlockSolverGrid->Nsimd();
std::vector<int> latt_phys(0);
std::vector<int> simd_phys(0);
std::vector<int> mpi_phys(0);
for(int d=0;d<NN;d++){
if( d!=Orthog ) {
latt_phys.push_back(BlockSolverGrid->_fdimensions[d]);
simd_phys.push_back(BlockSolverGrid->_simd_layout[d]);
mpi_phys.push_back(BlockSolverGrid->_processors[d]);
}
}
return (GridBase *)new GridCartesian(latt_phys,simd_phys,mpi_phys);
}
//////////////////////////////////////////////////////////////////////////////////////////////////////////////
// Need to move sliceInnerProduct, sliceAxpy, sliceNorm etc... into lattice sector along with sliceSum
//////////////////////////////////////////////////////////////////////////////////////////////////////////////
template<class vobj>
static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,const Lattice<vobj> &Y,int Orthog,RealD scale=1.0)
{
typedef typename vobj::scalar_object sobj;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
int Nblock = X._grid->GlobalDimensions()[Orthog];
GridBase *FullGrid = X._grid;
GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
Lattice<vobj> Xslice(SliceGrid);
Lattice<vobj> Rslice(SliceGrid);
// FIXME: Implementation is slow
// If we based this on Cshift it would work for spread out
// but it would be even slower
//
// Repeated extract slice is inefficient
//
// Best base the linear combination by constructing a
// set of vectors of size grid->_rdimensions[Orthog].
for(int i=0;i<Nblock;i++){
ExtractSlice(Rslice,Y,i,Orthog);
for(int j=0;j<Nblock;j++){
ExtractSlice(Xslice,X,j,Orthog);
Rslice = Rslice + Xslice*(scale*aa(j,i));
}
InsertSlice(Rslice,R,i,Orthog);
}
};
template<class vobj>
static void sliceMaddVector (Lattice<vobj> &R,std::vector<RealD> &a,const Lattice<vobj> &X,const Lattice<vobj> &Y,
int Orthog,RealD scale=1.0)
{
// FIXME: Implementation is slow
// Best base the linear combination by constructing a
// set of vectors of size grid->_rdimensions[Orthog].
typedef typename vobj::scalar_object sobj;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
int Nblock = X._grid->GlobalDimensions()[Orthog];
GridBase *FullGrid = X._grid;
GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
Lattice<vobj> Xslice(SliceGrid);
Lattice<vobj> Rslice(SliceGrid);
// If we based this on Cshift it would work for spread out
// but it would be even slower
for(int i=0;i<Nblock;i++){
ExtractSlice(Rslice,Y,i,Orthog);
ExtractSlice(Xslice,X,i,Orthog);
Rslice = Rslice + Xslice*(scale*a[i]);
InsertSlice(Rslice,R,i,Orthog);
}
};
template<class vobj>
static void sliceInnerProductMatrix( Eigen::MatrixXcd &mat, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog)
{
// FIXME: Implementation is slow
// Not sure of best solution.. think about it
typedef typename vobj::scalar_object sobj;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
GridBase *FullGrid = lhs._grid;
GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
int Nblock = FullGrid->GlobalDimensions()[Orthog];
Lattice<vobj> Lslice(SliceGrid);
Lattice<vobj> Rslice(SliceGrid);
mat = Eigen::MatrixXcd::Zero(Nblock,Nblock);
for(int i=0;i<Nblock;i++){
ExtractSlice(Lslice,lhs,i,Orthog);
for(int j=0;j<Nblock;j++){
ExtractSlice(Rslice,rhs,j,Orthog);
mat(i,j) = innerProduct(Lslice,Rslice);
}
}
return;
}
template<class vobj>
static void sliceNorm (std::vector<RealD> &sn,const Lattice<vobj> &rhs,int Orthog) {
typedef typename vobj::scalar_object sobj;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
@ -499,9 +324,207 @@ template<class vobj>
for(int ss=0;ss<Nblock;ss++){
sn[ss] = real(ip[ss]);
}
};
};
template<class vobj>
static void sliceMaddVector(Lattice<vobj> &R,std::vector<RealD> &a,const Lattice<vobj> &X,const Lattice<vobj> &Y,
int orthogdim,RealD scale=1.0)
{
typedef typename vobj::scalar_object sobj;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
typedef typename vobj::tensor_reduced tensor_reduced;
GridBase *grid = X._grid;
int Nsimd =grid->Nsimd();
int Nblock =grid->GlobalDimensions()[orthogdim];
int fd =grid->_fdimensions[orthogdim];
int ld =grid->_ldimensions[orthogdim];
int rd =grid->_rdimensions[orthogdim];
int e1 =grid->_slice_nblock[orthogdim];
int e2 =grid->_slice_block [orthogdim];
int stride =grid->_slice_stride[orthogdim];
std::vector<int> icoor;
for(int r=0;r<rd;r++){
int so=r*grid->_ostride[orthogdim]; // base offset for start of plane
vector_type av;
for(int l=0;l<Nsimd;l++){
grid->iCoorFromIindex(icoor,l);
int ldx =r+icoor[orthogdim]*rd;
scalar_type *as =(scalar_type *)&av;
as[l] = scalar_type(a[ldx])*scale;
}
tensor_reduced at; at=av;
parallel_for_nest2(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){
int ss= so+n*stride+b;
R._odata[ss] = at*X._odata[ss]+Y._odata[ss];
}
}
}
};
/*
template<class vobj>
static void sliceMaddVectorSlow (Lattice<vobj> &R,std::vector<RealD> &a,const Lattice<vobj> &X,const Lattice<vobj> &Y,
int Orthog,RealD scale=1.0)
{
// FIXME: Implementation is slow
// Best base the linear combination by constructing a
// set of vectors of size grid->_rdimensions[Orthog].
typedef typename vobj::scalar_object sobj;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
int Nblock = X._grid->GlobalDimensions()[Orthog];
GridBase *FullGrid = X._grid;
GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
Lattice<vobj> Xslice(SliceGrid);
Lattice<vobj> Rslice(SliceGrid);
// If we based this on Cshift it would work for spread out
// but it would be even slower
for(int i=0;i<Nblock;i++){
ExtractSlice(Rslice,Y,i,Orthog);
ExtractSlice(Xslice,X,i,Orthog);
Rslice = Rslice + Xslice*(scale*a[i]);
InsertSlice(Rslice,R,i,Orthog);
}
};
template<class vobj>
static void sliceInnerProductVectorSlow( std::vector<ComplexD> & vec, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog)
{
// FIXME: Implementation is slow
// Look at localInnerProduct implementation,
// and do inside a site loop with block strided iterators
typedef typename vobj::scalar_object sobj;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
typedef typename vobj::tensor_reduced scalar;
typedef typename scalar::scalar_object scomplex;
int Nblock = lhs._grid->GlobalDimensions()[Orthog];
vec.resize(Nblock);
std::vector<scomplex> sip(Nblock);
Lattice<scalar> IP(lhs._grid);
IP=localInnerProduct(lhs,rhs);
sliceSum(IP,sip,Orthog);
for(int ss=0;ss<Nblock;ss++){
vec[ss] = TensorRemove(sip[ss]);
}
}
*/
//////////////////////////////////////////////////////////////////////////////////////////
// FIXME: Implementation is slow
// If we based this on Cshift it would work for spread out
// but it would be even slower
//
// Repeated extract slice is inefficient
//
// Best base the linear combination by constructing a
// set of vectors of size grid->_rdimensions[Orthog].
//////////////////////////////////////////////////////////////////////////////////////////
inline GridBase *makeSubSliceGrid(const GridBase *BlockSolverGrid,int Orthog)
{
int NN = BlockSolverGrid->_ndimension;
int nsimd = BlockSolverGrid->Nsimd();
std::vector<int> latt_phys(0);
std::vector<int> simd_phys(0);
std::vector<int> mpi_phys(0);
for(int d=0;d<NN;d++){
if( d!=Orthog ) {
latt_phys.push_back(BlockSolverGrid->_fdimensions[d]);
simd_phys.push_back(BlockSolverGrid->_simd_layout[d]);
mpi_phys.push_back(BlockSolverGrid->_processors[d]);
}
}
return (GridBase *)new GridCartesian(latt_phys,simd_phys,mpi_phys);
}
template<class vobj>
static void sliceMaddMatrix (Lattice<vobj> &R,Eigen::MatrixXcd &aa,const Lattice<vobj> &X,const Lattice<vobj> &Y,int Orthog,RealD scale=1.0)
{
typedef typename vobj::scalar_object sobj;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
int Nblock = X._grid->GlobalDimensions()[Orthog];
GridBase *FullGrid = X._grid;
GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
Lattice<vobj> Xslice(SliceGrid);
Lattice<vobj> Rslice(SliceGrid);
for(int i=0;i<Nblock;i++){
ExtractSlice(Rslice,Y,i,Orthog);
for(int j=0;j<Nblock;j++){
ExtractSlice(Xslice,X,j,Orthog);
Rslice = Rslice + Xslice*(scale*aa(j,i));
}
InsertSlice(Rslice,R,i,Orthog);
}
};
template<class vobj>
static void sliceInnerProductMatrix( Eigen::MatrixXcd &mat, const Lattice<vobj> &lhs,const Lattice<vobj> &rhs,int Orthog)
{
// FIXME: Implementation is slow
// Not sure of best solution.. think about it
typedef typename vobj::scalar_object sobj;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
GridBase *FullGrid = lhs._grid;
GridBase *SliceGrid = makeSubSliceGrid(FullGrid,Orthog);
int Nblock = FullGrid->GlobalDimensions()[Orthog];
Lattice<vobj> Lslice(SliceGrid);
Lattice<vobj> Rslice(SliceGrid);
mat = Eigen::MatrixXcd::Zero(Nblock,Nblock);
for(int i=0;i<Nblock;i++){
ExtractSlice(Lslice,lhs,i,Orthog);
for(int j=0;j<Nblock;j++){
ExtractSlice(Rslice,rhs,j,Orthog);
mat(i,j) = innerProduct(Lslice,Rslice);
}
}
#undef FORCE_DIAG
#ifdef FORCE_DIAG
for(int i=0;i<Nblock;i++){
for(int j=0;j<Nblock;j++){
if ( i != j ) mat(i,j)=0.0;
}
}
#endif
return;
}
} /*END NAMESPACE GRID*/
#endif