portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-04-16 08:45:55 +01:00
Code Releases Activity

NVLINK GET/PUT on cuda aware mpi

Browse Source
This commit is contained in:
Peter Boyle 2025-04-04 16:43:36 -04:00
parent ebbd015c5c
commit adc90d3a86
1 changed files with 11 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
Grid/communicator/Communicator_mpi3.cc
View File

@ -438,8 +438,15 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
list.push_back(rrq);
off_node_bytes+=rbytes;
}
#ifdef NVLINK_GET
else {
void *shm = (void *) this->ShmBufferTranslate(from,xmit);
assert(shm!=NULL);
acceleratorCopyDeviceToDeviceAsynch(shm,recv,rbytes);
}
#endif
}
// This is a NVLINK PUT
if (dox) {
if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
tag= dir+_processor*32;
@ -448,9 +455,11 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
list.push_back(xrq);
off_node_bytes+=xbytes;
} else {
#ifndef NVLINK_GET
void *shm = (void *) this->ShmBufferTranslate(dest,recv);
assert(shm!=NULL);
acceleratorCopyDeviceToDeviceAsynch(xmit,shm,xbytes);
#endif
}
}
return off_node_bytes;
@ -459,7 +468,7 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)
{
int nreq=list.size();
/*finishes Get/Put*/
acceleratorCopySynchronise();
if (nreq==0) return;