portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-05-15 15:05:46 +01:00
Code Releases Activity

Bug fix for no accelerator aware MPI, thanks Shuhei for finding it.

Browse Source
This commit is contained in:
Peter Boyle 2025-04-24 11:40:46 -04:00
parent 7af8c77a52
commit 233150d93f
1 changed files with 11 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
Grid/lattice/PaddedCell.h
View File

@ -510,7 +510,6 @@ public:
grid->SendToRecvFromBegin(fwd_req, grid->SendToRecvFromBegin(fwd_req,
(void *)&hsend_buf[d*buffer_size], xmit_to_rank, (void *)&hsend_buf[d*buffer_size], xmit_to_rank,
(void *)&hrecv_buf[d*buffer_size], recv_from_rank, bytes, tag); (void *)&hrecv_buf[d*buffer_size], recv_from_rank, bytes, tag);
acceleratorCopyToDevice(&hrecv_buf[d*buffer_size],&recv_buf[d*buffer_size],bytes);
#endif #endif
t_comms+=usecond()-t; t_comms+=usecond()-t;
} }
@ -531,7 +530,6 @@ public:
grid->SendToRecvFromBegin(bwd_req, grid->SendToRecvFromBegin(bwd_req,
(void *)&hsend_buf[(d+depth)*buffer_size], recv_from_rank, (void *)&hsend_buf[(d+depth)*buffer_size], recv_from_rank,
(void *)&hrecv_buf[(d+depth)*buffer_size], xmit_to_rank, bytes,tag); (void *)&hrecv_buf[(d+depth)*buffer_size], xmit_to_rank, bytes,tag);
acceleratorCopyToDevice(&hrecv_buf[(d+depth)*buffer_size],&recv_buf[(d+depth)*buffer_size],bytes);
#endif #endif
t_comms+=usecond()-t; t_comms+=usecond()-t;
} }
@ -555,8 +553,13 @@ public:
t=usecond(); t=usecond();
grid->CommsComplete(fwd_req); grid->CommsComplete(fwd_req);
#ifndef ACCELERATOR_AWARE_MPI
for ( int d=0;d < depth ; d ++ ) {
acceleratorCopyToDevice(&hrecv_buf[d*buffer_size],&recv_buf[d*buffer_size],bytes);
}
#endif
t_comms+= usecond() - t; t_comms+= usecond() - t;
t=usecond(); t=usecond();
for ( int d=0;d < depth ; d ++ ) { for ( int d=0;d < depth ; d ++ ) {
ScatterSlice(recv_buf,to,nld-depth+d,dimension,plane*buffer_size); plane++; ScatterSlice(recv_buf,to,nld-depth+d,dimension,plane*buffer_size); plane++;
@ -565,6 +568,11 @@ public:
t=usecond(); t=usecond();
grid->CommsComplete(bwd_req); grid->CommsComplete(bwd_req);
#ifndef ACCELERATOR_AWARE_MPI
for ( int d=0;d < depth ; d ++ ) {
acceleratorCopyToDevice(&hrecv_buf[(d+depth)*buffer_size],&recv_buf[(d+depth)*buffer_size],bytes);
}
#endif
t_comms+= usecond() - t; t_comms+= usecond() - t;
t=usecond(); t=usecond();