portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2026-08-19 08:59:36 +01:00
Code Releases Activity

FFT offload to GPU and MUCH faster comms.

Browse Source 
40x speed up on Frontier
This commit is contained in:
Peter Boyle
2025-08-21 16:45:38 -04:00
parent 76c0ada1e1
commit fe0db53842
8 changed files with 443 additions and 176 deletions
Download Patch File Download Diff File Expand all files Collapse all files
Grid/cshift/Cshift_mpi.h
+63
View File
@@ -49,6 +49,20 @@ template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension
// Map to always positive shift modulo global full dimension.
shift = (shift+fd)%fd;
if( shift ==0 ) {
ret = rhs;
return ret;
}
//
// Potential easy fast cases:
// Shift is a multiple of the local lattice extent.
// Then need only to shift whole subvolumes
int L = rhs.Grid()->_ldimensions[dimension];
if ( (shift%L )==0 && !rhs.Grid()->CheckerBoarded(dimension) ) {
Cshift_simple(ret,rhs,dimension,shift);
return ret;
}
ret.Checkerboard() = rhs.Grid()->CheckerBoardDestination(rhs.Checkerboard(),shift,dimension);
// the permute type
@@ -73,6 +87,55 @@ template<class vobj> Lattice<vobj> Cshift(const Lattice<vobj> &rhs,int dimension
return ret;
}
template<class vobj> void Cshift_simple(Lattice<vobj>& ret,const Lattice<vobj> &rhs,int dimension,int shift)
{
GridBase *grid=rhs.Grid();
int comm_proc, xmit_to_rank, recv_from_rank;
int fd = rhs.Grid()->_fdimensions[dimension];
int rd = rhs.Grid()->_rdimensions[dimension];
int ld = rhs.Grid()->_ldimensions[dimension];
int pd = rhs.Grid()->_processors[dimension];
int simd_layout = rhs.Grid()->_simd_layout[dimension];
int comm_dim = rhs.Grid()->_processors[dimension] >1 ;
comm_proc = ((shift)/ld)%pd;
grid->ShiftedRanks(dimension,comm_proc,xmit_to_rank,recv_from_rank);
if(comm_dim) {
int64_t bytes = sizeof(vobj) * grid->oSites();
autoView(rhs_v , rhs, AcceleratorRead);
autoView(ret_v , ret, AcceleratorWrite);
void *send_buf = (void *)&rhs_v[0];
void *recv_buf = (void *)&ret_v[0];
#ifdef ACCELERATOR_AWARE_MPI
grid->SendToRecvFrom(send_buf,
xmit_to_rank,
recv_buf,
recv_from_rank,
bytes);
#else
static hostVector<vobj> hrhs; hrhs.resize(grid->oSites());
static hostVector<vobj> hret; hret.resize(grid->oSites());
void *hsend_buf = (void *)&hrhs[0];
void *hrecv_buf = (void *)&hret[0];
acceleratorCopyFromDevice(&send_buf[0],&hsend_buf[0],bytes);
grid->SendToRecvFrom(hsend_buf,
xmit_to_rank,
hrecv_buf,
recv_from_rank,
bytes);
acceleratorCopyToDevice(&hrecv_buf[0],&recv_buf[0],bytes);
#endif
}
}
template<class vobj> void Cshift_comms(Lattice<vobj>& ret,const Lattice<vobj> &rhs,int dimension,int shift)
{
int sshift[2];