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

Merge branch 'feature/dirichlet' of https://github.com/paboyle/Grid into feature/dirichlet

Browse Source
This commit is contained in:
Peter Boyle 2022-05-29 11:08:09 -04:00
commit 6a1a198144
4 changed files with 20 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
Grid/communicator/Communicator_mpi3.cc
View File

@ -372,7 +372,7 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
double off_node_bytes=0.0;
int tag;
if ( dox ) {
if ( dor ) {
if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
tag= dir+from*32;
ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
@ -382,7 +382,7 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
}
}
if (dor) {
if (dox) {
if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
tag= dir+_processor*32;
ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);

6
Grid/lattice/Lattice_peekpoke.h
View File

@ -125,6 +125,12 @@ void pokeSite(const sobj &s,Lattice<vobj> &l,const Coordinate &site){
//////////////////////////////////////////////////////////
// Peek a scalar object from the SIMD array
//////////////////////////////////////////////////////////
template<class vobj>
typename vobj::scalar_object peekSite(const Lattice<vobj> &l,const Coordinate &site){
typename vobj::scalar_object s;
peekSite(s,l,site);
return s;
}
template<class vobj,class sobj>
void peekSite(sobj &s,const Lattice<vobj> &l,const Coordinate &site){

18
Grid/stencil/Stencil.h
View File

@ -665,7 +665,7 @@ public:
this->_comms_recv[ii] = comm_dim;
if ( block && comm_dim ) {
assert(abs(displacement) < ld );
// Quiesce communication across block boundaries
if( displacement > 0 ) {
// High side, low side
// | <--B--->|
@ -730,7 +730,7 @@ public:
int gd = _grid->_gdimensions[dimension];
int fd = _grid->_fdimensions[dimension];
int pd = _grid->_processors [dimension];
int ld = gd/pd;
// int ld = gd/pd;
int rd = _grid->_rdimensions[dimension];
int pc = _grid->_processor_coor[dimension];
this->_permute_type[point]=_grid->PermuteType(dimension);
@ -871,12 +871,14 @@ public:
for(int x=0;x<rd;x++){
int permute_type=grid->PermuteType(dimension);
int permute_slice;
int sx = (x+sshift)%rd;
int offnode = 0;
if ( simd_layout > 1 ) {
permute_slice=1;
for(int i=0;i<Nsimd;i++){
int inner_bit = (Nsimd>>(permute_type+1));
@ -893,6 +895,7 @@ public:
} else {
int comm_proc = ((x+sshift)/rd)%pd;
offnode = (comm_proc!= 0);
permute_slice=0;
}
int wraparound=0;
@ -906,19 +909,18 @@ public:
// Wrap locally dirichlet support case OR node local
if ( offnode==0 ) {
int permute_slice=0;
permute_slice=0;
CopyPlane(point,dimension,x,sx,cbmask,permute_slice,wraparound);
} else {
if ( comms_recv==0 ) {
if ( comms_recv ) {
int permute_slice=1;
CopyPlane(point,dimension,x,sx,cbmask,permute_slice,wraparound);
ScatterPlane(point,dimension,x,cbmask,_unified_buffer_size,wraparound); // permute/extract/merge is done in comms phase
} else {
ScatterPlane(point,dimension,x,cbmask,_unified_buffer_size,wraparound); // permute/extract/merge is done in comms phase
CopyPlane(point,dimension,x,sx,cbmask,permute_slice,wraparound);
}
@ -1208,7 +1210,7 @@ public:
face_table[face_idx].size()*sizeof(face_table_host[0]));
}
if ( comms_send )
if ( comms_send || comms_recv )
Gather_plane_exchange_table(face_table[face_idx],rhs,spointers,dimension,sx,cbmask,compress,permute_type);
face_idx++;

4
systems/Perlmutter/dwf4.slurm
View File

@ -19,8 +19,8 @@ export MPICH_GPU_EAGER_REGISTER_HOST_MEM=0
export MPICH_GPU_NO_ASYNC_MEMCPY=0
#export MPICH_SMP_SINGLE_COPY_MODE=CMA
OPT="--comms-overlap --shm-mpi 1"
VOL=64.64.32.32
OPT="--comms-sequential --shm-mpi 1"
VOL=64.64.64.64
srun ./benchmarks/Benchmark_dwf_fp32 --mpi 2.2.1.1 --grid $VOL --accelerator-threads 8 --shm 2048 $OPT
#srun ./benchmarks/Benchmark_dwf_fp32 --mpi 2.1.1.4 --grid $VOL --accelerator-threads 8 --shm 2048 $OPT
#srun ./benchmarks/Benchmark_dwf_fp32 --mpi 1.1.1.8 --grid $VOL --accelerator-threads 8 --shm 2048 $OPT