portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-10 07:55:35 +00:00
Code Releases Activity

Simplify the compressor interface again.

Browse Source
This commit is contained in:
Peter Boyle 2016-02-17 18:16:45 -06:00
parent c650bb3f3d
commit c9fadf97a5
8 changed files with 228 additions and 46 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
lib/AlignedAllocator.h
View File

@ -108,6 +108,14 @@ public:
#endif
#endif
_Tp tmp;
#undef FIRST_TOUCH_OPTIMISE
#ifdef FIRST_TOUCH_OPTIMISE
#pragma omp parallel for
for(int i=0;i<__n;i++){
ptr[i]=tmp;
}
#endif
return ptr;
}

37
lib/Stencil.h
View File

@ -77,7 +77,7 @@ namespace Grid {
int _around_the_world;
};
template<class vobj,class cobj, class compressor>
template<class vobj,class cobj>
class CartesianStencil { // Stencil runs along coordinate axes only; NO diagonal fill in.
public:
@ -580,6 +580,7 @@ PARALLEL_FOR_LOOP
}
template<class compressor>
std::thread HaloExchangeBegin(const Lattice<vobj> &source,compressor &compress) {
Mergers.resize(0);
Packets.resize(0);
@ -587,6 +588,7 @@ PARALLEL_FOR_LOOP
return std::thread([&] { this->Communicate(); });
}
template<class compressor>
void HaloExchange(const Lattice<vobj> &source,compressor &compress)
{
auto thr = HaloExchangeBegin(source,compress);
@ -601,20 +603,9 @@ PARALLEL_FOR_LOOP
jointime+=usecond();
}
void HaloGather(const Lattice<vobj> &source,compressor &compress)
template<class compressor>
void HaloGatherDir(const Lattice<vobj> &source,compressor &compress,int point)
{
// conformable(source._grid,_grid);
assert(source._grid==_grid);
halogtime-=usecond();
assert (comm_buf.size() == _unified_buffer_size );
u_comm_offset=0;
// Gather all comms buffers
for(int point = 0 ; point < _npoints; point++) {
compress.Point(point);
int dimension = _directions[point];
int displacement = _distances[point];
@ -662,12 +653,29 @@ PARALLEL_FOR_LOOP
}
}
}
}
template<class compressor>
void HaloGather(const Lattice<vobj> &source,compressor &compress)
{
// conformable(source._grid,_grid);
assert(source._grid==_grid);
halogtime-=usecond();
assert (comm_buf.size() == _unified_buffer_size );
u_comm_offset=0;
// Gather all comms buffers
for(int point = 0 ; point < _npoints; point++) {
compress.Point(point);
HaloGatherDir(source,compress,point);
}
assert(u_comm_offset==_unified_buffer_size);
halogtime+=usecond();
}
template<class compressor>
void Gather(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor & compress)
{
typedef typename cobj::vector_type vector_type;
@ -728,6 +736,7 @@ PARALLEL_FOR_LOOP
}
template<class compressor>
void GatherSimd(const Lattice<vobj> &rhs,int dimension,int shift,int cbmask,compressor &compress)
{
const int Nsimd = _grid->Nsimd();

2
lib/algorithms/CoarsenedMatrix.h
View File

@ -200,7 +200,7 @@ namespace Grid {
////////////////////
Geometry geom;
GridBase * _grid;
CartesianStencil<siteVector,siteVector,SimpleCompressor<siteVector> > Stencil;
CartesianStencil<siteVector,siteVector> Stencil;
std::vector<CoarseMatrix> A;

17
lib/cshift/Cshift_common.h
View File

@ -35,7 +35,7 @@ class SimpleCompressor {
public:
void Point(int) {};
vobj operator() (const vobj &arg,int dimension,int plane,int osite,GridBase *grid) {
vobj operator() (const vobj &arg) {
return arg;
}
};
@ -63,7 +63,7 @@ PARALLEL_NESTED_LOOP2
for(int b=0;b<e2;b++){
int o = n*rhs._grid->_slice_stride[dimension];
int bo = n*rhs._grid->_slice_block[dimension];
buffer[off+bo+b]=compress(rhs._odata[so+o+b],dimension,plane,so+o+b,rhs._grid);
buffer[off+bo+b]=compress(rhs._odata[so+o+b]);
}
}
} else {
@ -73,7 +73,7 @@ PARALLEL_NESTED_LOOP2
int o = n*rhs._grid->_slice_stride[dimension];
int ocb=1<<rhs._grid->CheckerBoardFromOindex(o+b);// Could easily be a table lookup
if ( ocb &cbmask ) {
buffer[off+bo++]=compress(rhs._odata[so+o+b],dimension,plane,so+o+b,rhs._grid);
buffer[off+bo++]=compress(rhs._odata[so+o+b]);
}
}
}
@ -97,16 +97,17 @@ Gather_plane_extract(const Lattice<vobj> &rhs,std::vector<typename cobj::scalar_
int e1=rhs._grid->_slice_nblock[dimension];
int e2=rhs._grid->_slice_block[dimension];
int n1=rhs._grid->_slice_stride[dimension];
int n2=rhs._grid->_slice_block[dimension];
if ( cbmask ==0x3){
PARALLEL_NESTED_LOOP2
for(int n=0;n<e1;n++){
for(int b=0;b<e2;b++){
int o = n*rhs._grid->_slice_stride[dimension];
int offset = b+n*rhs._grid->_slice_block[dimension];
int o = n*n1;
int offset = b+n*n2;
cobj temp =compress(rhs._odata[so+o+b],dimension,plane,so+o+b,rhs._grid);
cobj temp =compress(rhs._odata[so+o+b]);
extract<cobj>(temp,pointers,offset);
}
@ -121,7 +122,7 @@ PARALLEL_NESTED_LOOP2
int offset = b+n*rhs._grid->_slice_block[dimension];
if ( ocb & cbmask ) {
cobj temp =compress(rhs._odata[so+o+b],dimension,plane,so+o+b,rhs._grid);
cobj temp =compress(rhs._odata[so+o+b]);
extract<cobj>(temp,pointers,offset);
}
}

4
lib/qcd/action/fermion/FermionOperatorImpl.h
View File

@ -130,7 +130,7 @@ namespace Grid {
typedef WilsonCompressor<SiteHalfSpinor,SiteSpinor> Compressor;
typedef WilsonImplParams ImplParams;
typedef CartesianStencil<SiteSpinor,SiteHalfSpinor,Compressor> StencilImpl;
typedef WilsonStencil<SiteSpinor,SiteHalfSpinor> StencilImpl;
ImplParams Params;
@ -205,7 +205,7 @@ PARALLEL_FOR_LOOP
typedef Lattice<SiteDoubledGaugeField> DoubledGaugeField;
typedef WilsonCompressor<SiteHalfSpinor,SiteSpinor> Compressor;
typedef CartesianStencil<SiteSpinor,SiteHalfSpinor,Compressor> StencilImpl;
typedef WilsonStencil<SiteSpinor,SiteHalfSpinor> StencilImpl;
typedef GparityWilsonImplParams ImplParams;

174
lib/qcd/action/fermion/WilsonCompressor.h
View File

@ -48,12 +48,7 @@ namespace QCD {
mu=p;
};
virtual SiteHalfSpinor operator () (const SiteSpinor &in,int dim,int plane,int osite,GridBase *grid) {
return spinproject(in);
}
SiteHalfSpinor spinproject(const SiteSpinor &in)
{
inline SiteHalfSpinor operator () (const SiteSpinor &in) {
SiteHalfSpinor ret;
int mudag=mu;
if (!dag) {
@ -92,6 +87,173 @@ namespace QCD {
}
};
/////////////////////////
// optimised versions
/////////////////////////
template<class SiteHalfSpinor,class SiteSpinor>
class WilsonXpCompressor {
public:
inline SiteHalfSpinor operator () (const SiteSpinor &in) {
SiteHalfSpinor ret;
spProjXp(ret,in);
return ret;
}
};
template<class SiteHalfSpinor,class SiteSpinor>
class WilsonYpCompressor {
public:
inline SiteHalfSpinor operator () (const SiteSpinor &in) {
SiteHalfSpinor ret;
spProjYp(ret,in);
return ret;
}
};
template<class SiteHalfSpinor,class SiteSpinor>
class WilsonZpCompressor {
public:
inline SiteHalfSpinor operator () (const SiteSpinor &in) {
SiteHalfSpinor ret;
spProjZp(ret,in);
return ret;
}
};
template<class SiteHalfSpinor,class SiteSpinor>
class WilsonTpCompressor {
public:
inline SiteHalfSpinor operator () (const SiteSpinor &in) {
SiteHalfSpinor ret;
spProjTp(ret,in);
return ret;
}
};
template<class SiteHalfSpinor,class SiteSpinor>
class WilsonXmCompressor {
public:
inline SiteHalfSpinor operator () (const SiteSpinor &in) {
SiteHalfSpinor ret;
spProjXm(ret,in);
return ret;
}
};
template<class SiteHalfSpinor,class SiteSpinor>
class WilsonYmCompressor {
public:
inline SiteHalfSpinor operator () (const SiteSpinor &in) {
SiteHalfSpinor ret;
spProjYm(ret,in);
return ret;
}
};
template<class SiteHalfSpinor,class SiteSpinor>
class WilsonZmCompressor {
public:
inline SiteHalfSpinor operator () (const SiteSpinor &in) {
SiteHalfSpinor ret;
spProjZm(ret,in);
return ret;
}
};
template<class SiteHalfSpinor,class SiteSpinor>
class WilsonTmCompressor {
public:
inline SiteHalfSpinor operator () (const SiteSpinor &in) {
SiteHalfSpinor ret;
spProjTm(ret,in);
return ret;
}
};
// Fast comms buffer manipulation which should inline right through (avoid direction
// dependent logic that prevents inlining
template<class vobj,class cobj>
class WilsonStencil : public CartesianStencil<vobj,cobj> {
public:
WilsonStencil(GridBase *grid,
int npoints,
int checkerboard,
const std::vector<int> &directions,
const std::vector<int> &distances) : CartesianStencil<vobj,cobj> (grid,npoints,checkerboard,directions,distances)
{ };
template < class compressor>
std::thread HaloExchangeOptBegin(const Lattice<vobj> &source,compressor &compress) {
this->Mergers.resize(0);
this->Packets.resize(0);
this->HaloGatherOpt(source,compress);
return std::thread([&] { this->Communicate(); });
}
template < class compressor>
void HaloExchangeOpt(const Lattice<vobj> &source,compressor &compress)
{
auto thr = this->HaloExchangeOptBegin(source,compress);
this->HaloExchangeOptComplete(thr);
}
void HaloExchangeOptComplete(std::thread &thr)
{
this->CommsMerge(); // spins
this->jointime-=usecond();
thr.join();
this->jointime+=usecond();
}
template < class compressor>
void HaloGatherOpt(const Lattice<vobj> &source,compressor &compress)
{
// conformable(source._grid,_grid);
assert(source._grid==this->_grid);
this->halogtime-=usecond();
assert (this->comm_buf.size() == this->_unified_buffer_size );
this->u_comm_offset=0;
int dag = compress.dag;
static std::vector<int> dirs(8);
for(int mu=0;mu<4;mu++){
if ( dag ) {
dirs[mu] =mu;
dirs[mu+4]=mu+4;
} else {
dirs[mu] =mu+4;
dirs[mu+4]=mu;
}
}
WilsonXpCompressor<cobj,vobj> XpCompress;
this->HaloGatherDir(source,XpCompress,dirs[0]);
WilsonYpCompressor<cobj,vobj> YpCompress;
this->HaloGatherDir(source,YpCompress,dirs[1]);
WilsonZpCompressor<cobj,vobj> ZpCompress;
this->HaloGatherDir(source,ZpCompress,dirs[2]);
WilsonTpCompressor<cobj,vobj> TpCompress;
this->HaloGatherDir(source,TpCompress,dirs[3]);
WilsonXmCompressor<cobj,vobj> XmCompress;
this->HaloGatherDir(source,XmCompress,dirs[4]);
WilsonYmCompressor<cobj,vobj> YmCompress;
this->HaloGatherDir(source,YmCompress,dirs[5]);
WilsonZmCompressor<cobj,vobj> ZmCompress;
this->HaloGatherDir(source,ZmCompress,dirs[6]);
WilsonTmCompressor<cobj,vobj> TmCompress;
this->HaloGatherDir(source,TmCompress,dirs[7]);
assert(this->u_comm_offset==this->_unified_buffer_size);
this->halogtime+=usecond();
}
};
}} // namespace close
#endif

8
lib/qcd/action/fermion/WilsonFermion5D.cc
View File

@ -304,8 +304,8 @@ void WilsonFermion5D<Impl>::DhopInternalCommsThenCompute(StencilImpl & st, Lebes
int nwork = U._grid->oSites();
commtime -=usecond();
auto handle = st.HaloExchangeBegin(in,compressor);
st.HaloExchangeComplete(handle);
auto handle = st.HaloExchangeOptBegin(in,compressor);
st.HaloExchangeOptComplete(handle);
commtime +=usecond();
jointime -=usecond();
@ -440,7 +440,7 @@ void WilsonFermion5D<Impl>::DhopInternalCommsOverlapCompute(StencilImpl & st, Le
int nwork = U._grid->oSites();
commtime -=usecond();
auto handle = st.HaloExchangeBegin(in,compressor);
auto handle = st.HaloExchangeOptBegin(in,compressor);
commtime +=usecond();
// Dhop takes the 4d grid from U, and makes a 5d index for fermion
@ -498,7 +498,7 @@ PARALLEL_FOR_LOOP
dslashtime +=usecond();
jointime -=usecond();
st.HaloExchangeComplete(handle);
st.HaloExchangeOptComplete(handle);
jointime +=usecond();
local = false;

24
lib/tensors/Tensor_extract_merge.h
View File

@ -44,8 +44,8 @@ template<class vsimd,class scalar>
inline void extract(typename std::enable_if<!isGridTensor<vsimd>::value, const vsimd >::type * y,
std::vector<scalar *> &extracted,int offset){
// FIXME: bounce off memory is painful
static const int Nsimd=vsimd::Nsimd();
int Nextr=extracted.size();
int Nsimd=vsimd::Nsimd();
int s=Nsimd/Nextr;
scalar*buf = (scalar *)y;
@ -59,8 +59,8 @@ inline void extract(typename std::enable_if<!isGridTensor<vsimd>::value, const v
template<class vsimd,class scalar>
inline void merge(typename std::enable_if<!isGridTensor<vsimd>::value, vsimd >::type * y,
std::vector<scalar *> &extracted,int offset){
static const int Nsimd=vsimd::Nsimd();
int Nextr=extracted.size();
int Nsimd=vsimd::Nsimd();
int s=Nsimd/Nextr; // can have sparse occupation of simd vector if simd_layout does not fill it
// replicate n-fold. Use to allow Integer masks to
// predicate floating point of various width assignments and maintain conformable.
@ -85,6 +85,7 @@ inline void extract(typename std::enable_if<!isGridTensor<vsimd>::value, const v
scalar *buf = (scalar *)&y;
for(int i=0;i<Nextr;i++){
extracted[i]=buf[i*s];
#ifdef PARANOID
for(int ii=1;ii<s;ii++){
if ( buf[i*s]!=buf[i*s+ii] ){
std::cout<<GridLogMessage << " SIMD extract failure splat = "<<s<<" ii "<<ii<<" " <<Nextr<<" "<< Nsimd<<" "<<std::endl;
@ -96,6 +97,7 @@ inline void extract(typename std::enable_if<!isGridTensor<vsimd>::value, const v
}
assert(buf[i*s]==buf[i*s+ii]);
}
#endif
}
};
@ -106,7 +108,7 @@ inline void extract(typename std::enable_if<!isGridTensor<vsimd>::value, const v
template<class vsimd,class scalar>
inline void merge(typename std::enable_if<!isGridTensor<vsimd>::value, vsimd >::type &y,std::vector<scalar> &extracted){
int Nextr=extracted.size();
int Nsimd=vsimd::Nsimd();
static const int Nsimd=vsimd::Nsimd();
int s=Nsimd/Nextr;
scalar *buf = (scalar *)&y;
@ -125,9 +127,9 @@ template<class vobj> inline void extract(const vobj &vec,std::vector<typename vo
typedef typename vobj::scalar_type scalar_type ;
typedef typename vobj::vector_type vector_type ;
const int Nsimd=vobj::vector_type::Nsimd();
static const int Nsimd=vobj::vector_type::Nsimd();
static const int words=sizeof(vobj)/sizeof(vector_type);
int Nextr=extracted.size();
const int words=sizeof(vobj)/sizeof(vector_type);
int s=Nsimd/Nextr;
std::vector<scalar_type *> pointers(Nextr);
@ -148,8 +150,8 @@ void extract(const vobj &vec,std::vector<typename vobj::scalar_object *> &extrac
typedef typename vobj::scalar_type scalar_type ;
typedef typename vobj::vector_type vector_type ;
const int words=sizeof(vobj)/sizeof(vector_type);
const int Nsimd=vobj::vector_type::Nsimd();
static const int words=sizeof(vobj)/sizeof(vector_type);
static const int Nsimd=vobj::vector_type::Nsimd();
int Nextr=extracted.size();
int s = Nsimd/Nextr;
@ -172,8 +174,8 @@ void merge(vobj &vec,std::vector<typename vobj::scalar_object> &extracted)
typedef typename vobj::scalar_type scalar_type ;
typedef typename vobj::vector_type vector_type ;
const int Nsimd=vobj::vector_type::Nsimd();
const int words=sizeof(vobj)/sizeof(vector_type);
static const int Nsimd=vobj::vector_type::Nsimd();
static const int words=sizeof(vobj)/sizeof(vector_type);
int Nextr = extracted.size();
int splat=Nsimd/Nextr;
@ -224,8 +226,8 @@ void merge1(vobj &vec,std::vector<typename vobj::scalar_object *> &extracted,int
typedef typename vobj::scalar_type scalar_type ;
typedef typename vobj::vector_type vector_type ;
const int Nsimd=vobj::vector_type::Nsimd();
const int words=sizeof(vobj)/sizeof(vector_type);
static const int Nsimd=vobj::vector_type::Nsimd();
static const int words=sizeof(vobj)/sizeof(vector_type);
scalar_type *vp = (scalar_type *)&vec;