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

PeekLocal needs a view to keep thread safe.

Browse Source 
ALLOCATION_CACHEE reenable
This commit is contained in:
Peter Boyle 2020-06-19 17:13:27 -04:00
parent 11bc1aeadc
commit b949cf6b12
11 changed files with 125 additions and 96 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
Grid/algorithms/FFT.h
View File

@ -230,14 +230,15 @@ public:
result = source;
int pc = processor_coor[dim];
for(int p=0;p<processors[dim];p++) {
autoView(r_v,result,CpuRead);
autoView(p_v,pgbuf,CpuWrite);
thread_for(idx, sgrid->lSites(),{
Coordinate cbuf(Nd);
sobj s;
sgrid->LocalIndexToLocalCoor(idx,cbuf);
peekLocalSite(s,result,cbuf);
cbuf[dim]+=((pc+p) % processors[dim])*L;
// cbuf[dim]+=p*L;
pokeLocalSite(s,pgbuf,cbuf);
peekLocalSite(s,r_v,cbuf);
acbuf[dim]+=((pc+p) % processors[dim])*L;
pokeLocalSite(s,p_v,cbuf);
});
if (p != processors[dim] - 1) {
result = Cshift(result,dim,L);

24
Grid/allocator/MemoryManager.cc
View File

@ -86,23 +86,33 @@ void MemoryManager::Init(void)
Ncache[AccSmall]=Nc;
}
}
std::cout << "MemoryManager::Init() setting up"<<std::endl;
std::cout << GridLogMessage<< "MemoryManager::Init() setting up"<<std::endl;
#ifdef ALLOCATION_CACHE
std::cout << "MemoryManager::Init() cache pool for recent allocations: SMALL "<<Ncache[CpuSmall]<<" LARGE "<<Ncache[Cpu]<<std::endl;
std::cout << GridLogMessage<< "MemoryManager::Init() cache pool for recent allocations: SMALL "<<Ncache[CpuSmall]<<" LARGE "<<Ncache[Cpu]<<std::endl;
#endif
#ifdef GRID_UVM
std::cout << "MemoryManager::Init() Unified memory space"<<std::endl;
std::cout << GridLogMessage<< "MemoryManager::Init() Unified memory space"<<std::endl;
#ifdef GRID_CUDA
std::cout << "MemoryManager::Init() Using cudaMallocManaged"<<std::endl;
std::cout << GridLogMessage<< "MemoryManager::Init() Using cudaMallocManaged"<<std::endl;
#endif
#ifdef GRID_HIP
std::cout << "MemoryManager::Init() Using hipMallocManaged"<<std::endl;
std::cout << GridLogMessage<< "MemoryManager::Init() Using hipMallocManaged"<<std::endl;
#endif
#ifdef GRID_SYCL
std::cout << "MemoryManager::Init() Using SYCL malloc_shared"<<std::endl;
std::cout << GridLogMessage<< "MemoryManager::Init() Using SYCL malloc_shared"<<std::endl;
#endif
#else
std::cout << "MemoryManager::Init() Non unified: Caching accelerator data in dedicated memory"<<std::endl;
std::cout << GridLogMessage<< "MemoryManager::Init() Non unified: Caching accelerator data in dedicated memory"<<std::endl;
#ifdef GRID_CUDA
std::cout << GridLogMessage<< "MemoryManager::Init() Using cudaMalloc"<<std::endl;
#endif
#ifdef GRID_HIP
std::cout << GridLogMessage<< "MemoryManager::Init() Using hipMalloc"<<std::endl;
#endif
#ifdef GRID_SYCL
std::cout << GridLogMessage<< "MemoryManager::Init() Using SYCL malloc_device"<<std::endl;
#endif
#endif
}

2
Grid/allocator/MemoryManager.h
View File

@ -34,7 +34,7 @@ NAMESPACE_BEGIN(Grid);
// Move control to configure.ac and Config.h?
#undef ALLOCATION_CACHE
#define ALLOCATION_CACHE
#define GRID_ALLOC_ALIGN (2*1024*1024)
#define GRID_ALLOC_SMALL_LIMIT (4096)

24
Grid/lattice/Lattice_peekpoke.h
View File

@ -154,17 +154,18 @@ void peekSite(sobj &s,const Lattice<vobj> &l,const Coordinate &site){
//////////////////////////////////////////////////////////
// Peek a scalar object from the SIMD array
//////////////////////////////////////////////////////////
// Must be CPU read view
template<class vobj,class sobj>
inline void peekLocalSite(sobj &s,const Lattice<vobj> &l,Coordinate &site)
inline void peekLocalSite(sobj &s,const LatticeView<vobj> &l,Coordinate &site)
{
GridBase *grid = l.Grid();
GridBase *grid = l.getGrid();
assert(l.mode==CpuRead);
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
int Nsimd = grid->Nsimd();
assert( l.Checkerboard()== l.Grid()->CheckerBoard(site));
assert( l.Checkerboard()== grid->CheckerBoard(site));
assert( sizeof(sobj)*Nsimd == sizeof(vobj));
static const int words=sizeof(vobj)/sizeof(vector_type);
@ -172,8 +173,7 @@ inline void peekLocalSite(sobj &s,const Lattice<vobj> &l,Coordinate &site)
idx= grid->iIndex(site);
odx= grid->oIndex(site);
autoView( l_v , l, CpuRead);
scalar_type * vp = (scalar_type *)&l_v[odx];
scalar_type * vp = (scalar_type *)&l[odx];
scalar_type * pt = (scalar_type *)&s;
for(int w=0;w<words;w++){
@ -182,18 +182,19 @@ inline void peekLocalSite(sobj &s,const Lattice<vobj> &l,Coordinate &site)
return;
};
// Must be CPU write view
template<class vobj,class sobj>
inline void pokeLocalSite(const sobj &s,Lattice<vobj> &l,Coordinate &site)
inline void pokeLocalSite(const sobj &s,LatticeView<vobj> &l,Coordinate &site)
{
GridBase *grid=l.Grid();
GridBase *grid=l.getGrid();
assert(l.mode==CpuWrite);
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
int Nsimd = grid->Nsimd();
assert( l.Checkerboard()== l.Grid()->CheckerBoard(site));
assert( l.Checkerboard()== grid->CheckerBoard(site));
assert( sizeof(sobj)*Nsimd == sizeof(vobj));
static const int words=sizeof(vobj)/sizeof(vector_type);
@ -201,8 +202,7 @@ inline void pokeLocalSite(const sobj &s,Lattice<vobj> &l,Coordinate &site)
idx= grid->iIndex(site);
odx= grid->oIndex(site);
autoView( l_v , l, CpuWrite);
scalar_type * vp = (scalar_type *)&l_v[odx];
scalar_type * vp = (scalar_type *)&l[odx];
scalar_type * pt = (scalar_type *)&s;
for(int w=0;w<words;w++){
vp[idx+w*Nsimd] = pt[w];

32
Grid/lattice/Lattice_transfer.h
View File

@ -520,15 +520,17 @@ void localConvert(const Lattice<vobj> &in,Lattice<vvobj> &out)
assert(ig->lSites() == og->lSites());
}
autoView(in_v,in,CpuRead);
autoView(out_v,out,CpuWrite);
thread_for(idx, ig->lSites(),{
sobj s;
ssobj ss;
Coordinate lcoor(ni);
ig->LocalIndexToLocalCoor(idx,lcoor);
peekLocalSite(s,in,lcoor);
peekLocalSite(s,in_v,lcoor);
ss=s;
pokeLocalSite(ss,out,lcoor);
pokeLocalSite(ss,out_v,lcoor);
});
}
@ -588,8 +590,6 @@ void localCopyRegion(const Lattice<vobj> &From,Lattice<vobj> & To,Coordinate Fro
for(int w=0;w<words;w++){
tp[idx_t+w*Nsimd] = fp[idx_f+w*Nsimd]; // FIXME IF RRII layout, type pun no worke
}
// peekLocalSite(s,From,Fcoor);
// pokeLocalSite(s,To ,Tcoor);
}
});
}
@ -620,6 +620,8 @@ void InsertSlice(const Lattice<vobj> &lowDim,Lattice<vobj> & higherDim,int slice
}
// the above should guarantee that the operations are local
autoView(lowDimv,lowDim,CpuRead);
autoView(higherDimv,higherDim,CpuWrite);
thread_for(idx,lg->lSites(),{
sobj s;
Coordinate lcoor(nl);
@ -632,8 +634,8 @@ void InsertSlice(const Lattice<vobj> &lowDim,Lattice<vobj> & higherDim,int slice
hcoor[d]=lcoor[ddl++];
}
}
peekLocalSite(s,lowDim,lcoor);
pokeLocalSite(s,higherDim,hcoor);
peekLocalSite(s,lowDimv,lcoor);
pokeLocalSite(s,higherDimv,hcoor);
});
}
@ -661,6 +663,8 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
}
}
// the above should guarantee that the operations are local
autoView(lowDimv,lowDim,CpuWrite);
autoView(higherDimv,higherDim,CpuRead);
thread_for(idx,lg->lSites(),{
sobj s;
Coordinate lcoor(nl);
@ -673,8 +677,8 @@ void ExtractSlice(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int slic
hcoor[d]=lcoor[ddl++];
}
}
peekLocalSite(s,higherDim,hcoor);
pokeLocalSite(s,lowDim,lcoor);
peekLocalSite(s,higherDimv,hcoor);
pokeLocalSite(s,lowDimv,lcoor);
});
}
@ -702,6 +706,8 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
}
// the above should guarantee that the operations are local
autoView(lowDimv,lowDim,CpuRead);
autoView(higherDimv,higherDim,CpuWrite);
thread_for(idx,lg->lSites(),{
sobj s;
Coordinate lcoor(nl);
@ -710,8 +716,8 @@ void InsertSliceLocal(const Lattice<vobj> &lowDim, Lattice<vobj> & higherDim,int
if( lcoor[orthog] == slice_lo ) {
hcoor=lcoor;
hcoor[orthog] = slice_hi;
peekLocalSite(s,lowDim,lcoor);
pokeLocalSite(s,higherDim,hcoor);
peekLocalSite(s,lowDimv,lcoor);
pokeLocalSite(s,higherDimv,hcoor);
}
});
}
@ -739,6 +745,8 @@ void ExtractSliceLocal(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int
}
// the above should guarantee that the operations are local
autoView(lowDimv,lowDim,CpuWrite);
autoView(higherDimv,higherDim,CpuRead);
thread_for(idx,lg->lSites(),{
sobj s;
Coordinate lcoor(nl);
@ -747,8 +755,8 @@ void ExtractSliceLocal(Lattice<vobj> &lowDim,const Lattice<vobj> & higherDim,int
if( lcoor[orthog] == slice_lo ) {
hcoor=lcoor;
hcoor[orthog] = slice_hi;
peekLocalSite(s,higherDim,hcoor);
pokeLocalSite(s,lowDim,lcoor);
peekLocalSite(s,higherDimv,hcoor);
pokeLocalSite(s,lowDimv,lcoor);
}
});
}

2
Grid/lattice/Lattice_view.h
View File

@ -43,6 +43,8 @@ public:
if (grid) conformable(grid, _grid);
else grid = _grid;
};
// Host only
GridBase * getGrid(void) const { return _grid; };
};
/////////////////////////////////////////////////////////////////////////////////////////

15
Grid/qcd/action/fermion/DomainWallVec5dImpl.h
View File

@ -114,19 +114,22 @@ public:
U = adj(Cshift(U, mu, -1));
PokeIndex<LorentzIndex>(Uadj, U, mu);
}
for (int lidx = 0; lidx < GaugeGrid->lSites(); lidx++) {
autoView(Umu_v,Umu,CpuRead);
autoView(Uadj_v,Uadj,CpuRead);
autoView(Uds_v,Uds,CpuWrite);
thread_for( lidx, GaugeGrid->lSites(), {
Coordinate lcoor;
GaugeGrid->LocalIndexToLocalCoor(lidx, lcoor);
peekLocalSite(ScalarUmu, Umu, lcoor);
peekLocalSite(ScalarUmu, Umu_v, lcoor);
for (int mu = 0; mu < 4; mu++) ScalarUds(mu) = ScalarUmu(mu);
peekLocalSite(ScalarUmu, Uadj, lcoor);
peekLocalSite(ScalarUmu, Uadj_v, lcoor);
for (int mu = 0; mu < 4; mu++) ScalarUds(mu + 4) = ScalarUmu(mu);
pokeLocalSite(ScalarUds, Uds, lcoor);
}
pokeLocalSite(ScalarUds, Uds_v, lcoor);
});
}
inline void InsertForce4D(GaugeField &mat, FermionField &Btilde,FermionField &A, int mu)

15
Grid/qcd/action/fermion/StaggeredVec5dImpl.h
View File

@ -113,20 +113,7 @@ public:
inline void InsertGaugeField(DoubledGaugeField &U_ds,const GaugeLinkField &U,int mu)
{
GridBase *GaugeGrid = U_ds.Grid();
thread_for(lidx, GaugeGrid->lSites(),{
SiteScalarGaugeLink ScalarU;
SiteDoubledGaugeField ScalarUds;
Coordinate lcoor;
GaugeGrid->LocalIndexToLocalCoor(lidx, lcoor);
peekLocalSite(ScalarUds, U_ds, lcoor);
peekLocalSite(ScalarU, U, lcoor);
ScalarUds(mu) = ScalarU();
});
assert(0);
}
inline void DoubleStore(GridBase *GaugeGrid,
DoubledGaugeField &UUUds, // for Naik term

53
Grid/qcd/action/fermion/implementation/WilsonCloverFermionImplementation.h
View File

@ -98,32 +98,35 @@ void WilsonCloverFermion<Impl>::ImportGauge(const GaugeField &_Umu)
Coordinate lcoor;
typename SiteCloverType::scalar_object Qx = Zero(), Qxinv = Zero();
for (int site = 0; site < lvol; site++)
{
grid->LocalIndexToLocalCoor(site, lcoor);
EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
peekLocalSite(Qx, CloverTerm, lcoor);
Qxinv = Zero();
//if (csw!=0){
for (int j = 0; j < Ns; j++)
for (int k = 0; k < Ns; k++)
for (int a = 0; a < DimRep; a++)
for (int b = 0; b < DimRep; b++){
auto zz = Qx()(j, k)(a, b);
EigenCloverOp(a + j * DimRep, b + k * DimRep) = std::complex<double>(zz);
}
// if (site==0) std::cout << "site =" << site << "\n" << EigenCloverOp << std::endl;
EigenInvCloverOp = EigenCloverOp.inverse();
//std::cout << EigenInvCloverOp << std::endl;
for (int j = 0; j < Ns; j++)
for (int k = 0; k < Ns; k++)
for (int a = 0; a < DimRep; a++)
for (int b = 0; b < DimRep; b++)
Qxinv()(j, k)(a, b) = EigenInvCloverOp(a + j * DimRep, b + k * DimRep);
// if (site==0) std::cout << "site =" << site << "\n" << EigenInvCloverOp << std::endl;
// }
pokeLocalSite(Qxinv, CloverTermInv, lcoor);
autoView(CTv,CloverTerm,CpuRead);
autoView(CTIv,CloverTermInv,CpuWrite);
for (int site = 0; site < lvol; site++) {
grid->LocalIndexToLocalCoor(site, lcoor);
EigenCloverOp = Eigen::MatrixXcd::Zero(Ns * DimRep, Ns * DimRep);
peekLocalSite(Qx, CTv, lcoor);
Qxinv = Zero();
//if (csw!=0){
for (int j = 0; j < Ns; j++)
for (int k = 0; k < Ns; k++)
for (int a = 0; a < DimRep; a++)
for (int b = 0; b < DimRep; b++){
auto zz = Qx()(j, k)(a, b);
EigenCloverOp(a + j * DimRep, b + k * DimRep) = std::complex<double>(zz);
}
// if (site==0) std::cout << "site =" << site << "\n" << EigenCloverOp << std::endl;
EigenInvCloverOp = EigenCloverOp.inverse();
//std::cout << EigenInvCloverOp << std::endl;
for (int j = 0; j < Ns; j++)
for (int k = 0; k < Ns; k++)
for (int a = 0; a < DimRep; a++)
for (int b = 0; b < DimRep; b++)
Qxinv()(j, k)(a, b) = EigenInvCloverOp(a + j * DimRep, b + k * DimRep);
// if (site==0) std::cout << "site =" << site << "\n" << EigenInvCloverOp << std::endl;
// }
pokeLocalSite(Qxinv, CTIv, lcoor);
}
}
// Separate the even and odd parts

34
Grid/qcd/action/fermion/implementation/WilsonFermion5DImplementation.h
View File

@ -580,16 +580,21 @@ void WilsonFermion5D<Impl>::MomentumSpacePropagatorHt_5d(FermionField &out,const
cosha = (one + W*W + sk) / (abs(W)*2.0);
// FIXME Need a Lattice acosh
for(int idx=0;idx<_grid->lSites();idx++){
Coordinate lcoor(Nd);
Tcomplex cc;
// RealD sgn;
_grid->LocalIndexToLocalCoor(idx,lcoor);
peekLocalSite(cc,cosha,lcoor);
assert((double)real(cc)>=1.0);
assert(fabs((double)imag(cc))<=1.0e-15);
cc = ScalComplex(::acosh(real(cc)),0.0);
pokeLocalSite(cc,a,lcoor);
{
autoView(cosha_v,cosha,CpuRead);
autoView(a_v,a,CpuWrite);
for(int idx=0;idx<_grid->lSites();idx++){
Coordinate lcoor(Nd);
Tcomplex cc;
// RealD sgn;
_grid->LocalIndexToLocalCoor(idx,lcoor);
peekLocalSite(cc,cosha_v,lcoor);
assert((double)real(cc)>=1.0);
assert(fabs((double)imag(cc))<=1.0e-15);
cc = ScalComplex(::acosh(real(cc)),0.0);
pokeLocalSite(cc,a_v,lcoor);
}
}
Wea = ( exp( a) * abs(W) );
@ -775,17 +780,20 @@ void WilsonFermion5D<Impl>::MomentumSpacePropagatorHt(FermionField &out,const Fe
cosha = (one + W*W + sk) / (abs(W)*2.0);
// FIXME Need a Lattice acosh
{
autoView(cosha_v,cosha,CpuRead);
autoView(a_v,a,CpuWrite);
for(int idx=0;idx<_grid->lSites();idx++){
Coordinate lcoor(Nd);
Tcomplex cc;
// RealD sgn;
_grid->LocalIndexToLocalCoor(idx,lcoor);
peekLocalSite(cc,cosha,lcoor);
peekLocalSite(cc,cosha_v,lcoor);
assert((double)real(cc)>=1.0);
assert(fabs((double)imag(cc))<=1.0e-15);
cc = ScalComplex(::acosh(real(cc)),0.0);
pokeLocalSite(cc,a,lcoor);
}
pokeLocalSite(cc,a_v,lcoor);
}}
Wea = ( exp( a) * abs(W) );
Wema= ( exp(-a) * abs(W) );

11
Grid/util/Init.cc
View File

@ -286,8 +286,6 @@ void Grid_init(int *argc,char ***argv)
//////////////////////////////////////////////////////////
acceleratorInit(); // Must come first to set device prior to MPI init due to Omnipath Driver
MemoryManager::Init();
if( GridCmdOptionExists(*argv,*argv+*argc,"--shm") ){
int MB;
arg= GridCmdOptionPayload(*argv,*argv+*argc,"--shm");
@ -358,6 +356,15 @@ void Grid_init(int *argc,char ***argv)
std::cout << GridLogMessage << "MPI is initialised and logging filters activated "<<std::endl;
std::cout << GridLogMessage << "================================================ "<<std::endl;
//////////////////////////////////////////////////////////
// Memory manager
//////////////////////////////////////////////////////////
MemoryManager::Init();
/////////////////////////////////////////////////////////
// Reporting
/////////////////////////////////////////////////////////
std::cout << GridLogMessage << "Requested "<< GlobalSharedMemory::MAX_MPI_SHM_BYTES <<" byte stencil comms buffers "<<std::endl;
if ( GlobalSharedMemory::Hugepages) {
std::cout << GridLogMessage << "Mapped stencil comms buffers as MAP_HUGETLB "<<std::endl;