portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2026-09-03 00:09:36 +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/communicator/Communicator_base.h
+7 -7
View File
@@ -108,7 +108,7 @@ public:
// very VERY rarely (Log, serial RNG) we need world without a grid
////////////////////////////////////////////////////////////////////////////////
static int RankWorld(void) ;
static void BroadcastWorld(int root,void* data, int bytes);
static void BroadcastWorld(int root,void* data, uint64_t bytes);
static void BarrierWorld(void);
////////////////////////////////////////////////////////////
@@ -175,27 +175,27 @@ public:
int dest,
void *recv,
int from,
int bytes,int dir);
uint64_t bytes,int dir);
void SendToRecvFrom(void *xmit,
int xmit_to_rank,
void *recv,
int recv_from_rank,
int bytes);
uint64_t bytes);
int IsOffNode(int rank);
double StencilSendToRecvFrom(void *xmit,
int xmit_to_rank,int do_xmit,
void *recv,
int recv_from_rank,int do_recv,
int bytes,int dir);
uint64_t bytes,int dir);
double StencilSendToRecvFromPrepare(std::vector<CommsRequest_t> &list,
void *xmit,
int xmit_to_rank,int do_xmit,
void *recv,
int recv_from_rank,int do_recv,
int xbytes,int rbytes,int dir);
uint64_t xbytes,uint64_t rbytes,int dir);
// Could do a PollHtoD and have a CommsMerge dependence
void StencilSendToRecvFromPollDtoH (std::vector<CommsRequest_t> &list);
@@ -206,7 +206,7 @@ public:
int xmit_to_rank,int do_xmit,
void *recv,void *recv_comp,
int recv_from_rank,int do_recv,
int xbytes,int rbytes,int dir);
uint64_t xbytes,uint64_t rbytes,int dir);
void StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int i);
@@ -220,7 +220,7 @@ public:
////////////////////////////////////////////////////////////
// Broadcast a buffer and composite larger
////////////////////////////////////////////////////////////
void Broadcast(int root,void* data, int bytes);
void Broadcast(int root,void* data, uint64_t bytes);
////////////////////////////////////////////////////////////
// All2All down one dimension
Grid/communicator/Communicator_mpi3.cc
+21 -21
View File
@@ -342,23 +342,23 @@ void CartesianCommunicator::SendToRecvFromBegin(std::vector<MpiCommsRequest_t> &
int dest,
void *recv,
int from,
int bytes,int dir)
uint64_t bytes,int dir)
{
MPI_Request xrq;
MPI_Request rrq;
GRID_ASSERT(dest != _processor);
GRID_ASSERT(from != _processor);
GRID_ASSERT(bytes/(sizeof(int32_t))<= 2*1024*1024*1024);
int tag;
tag= dir+from*32;
int ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,tag,communicator,&rrq);
int ierr=MPI_Irecv(recv,(int)( bytes/sizeof(int32_t)), MPI_INT32_T,from,tag,communicator,&rrq);
GRID_ASSERT(ierr==0);
list.push_back(rrq);
tag= dir+_processor*32;
ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,tag,communicator,&xrq);
ierr =MPI_Isend(xmit,(int)(bytes/sizeof(int32_t)), MPI_INT32_T,dest,tag,communicator,&xrq);
GRID_ASSERT(ierr==0);
list.push_back(xrq);
}
@@ -379,7 +379,7 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
int dest,
void *recv,
int from,
int bytes)
uint64_t bytes)
{
std::vector<MpiCommsRequest_t> reqs(0);
@@ -392,8 +392,8 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
// Give the CPU to MPI immediately; can use threads to overlap optionally
// printf("proc %d SendToRecvFrom %d bytes Sendrecv \n",_processor,bytes);
ierr=MPI_Sendrecv(xmit,bytes,MPI_CHAR,dest,myrank,
recv,bytes,MPI_CHAR,from, from,
ierr=MPI_Sendrecv(xmit,(int)(bytes/sizeof(int32_t)),MPI_INT32_T,dest,myrank,
recv,(int)(bytes/sizeof(int32_t)),MPI_INT32_T,from, from,
communicator,MPI_STATUS_IGNORE);
GRID_ASSERT(ierr==0);
@@ -403,7 +403,7 @@ double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
int dest, int dox,
void *recv,
int from, int dor,
int bytes,int dir)
uint64_t bytes,int dir)
{
std::vector<CommsRequest_t> list;
double offbytes = StencilSendToRecvFromPrepare(list,xmit,dest,dox,recv,from,dor,bytes,bytes,dir);
@@ -426,7 +426,7 @@ double CartesianCommunicator::StencilSendToRecvFromPrepare(std::vector<CommsRequ
int dest,int dox,
void *recv,
int from,int dor,
int xbytes,int rbytes,int dir)
uint64_t xbytes,uint64_t rbytes,int dir)
{
return 0.0; // Do nothing -- no preparation required
}
@@ -435,7 +435,7 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
int dest,int dox,
void *recv,void *recv_comp,
int from,int dor,
int xbytes,int rbytes,int dir)
uint64_t xbytes,uint64_t rbytes,int dir)
{
int ncomm =communicator_halo.size();
int commdir=dir%ncomm;
@@ -458,7 +458,7 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
tag= dir+from*32;
// std::cout << " StencilSendToRecvFrom "<<dir<<" MPI_Irecv "<<std::hex<<recv<<std::dec<<std::endl;
ierr=MPI_Irecv(recv_comp, rbytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
ierr=MPI_Irecv(recv_comp,(int)(rbytes/sizeof(int32_t)), MPI_INT32_T,from,tag,communicator_halo[commdir],&rrq);
GRID_ASSERT(ierr==0);
list.push_back(rrq);
off_node_bytes+=rbytes;
@@ -476,7 +476,7 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
if (dox) {
if ( (gdest == MPI_UNDEFINED) || Stencil_force_mpi ) {
tag= dir+_processor*32;
ierr =MPI_Isend(xmit_comp, xbytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
ierr =MPI_Isend(xmit_comp,(int)(xbytes/sizeof(int32_t)), MPI_INT32_T,dest,tag,communicator_halo[commdir],&xrq);
GRID_ASSERT(ierr==0);
list.push_back(xrq);
off_node_bytes+=xbytes;
@@ -543,7 +543,7 @@ double CartesianCommunicator::StencilSendToRecvFromPrepare(std::vector<CommsRequ
int dest,int dox,
void *recv,
int from,int dor,
int xbytes,int rbytes,int dir)
uint64_t xbytes,uint64_t rbytes,int dir)
{
/*
* Bring sequence from Stencil.h down to lower level.
@@ -584,7 +584,7 @@ double CartesianCommunicator::StencilSendToRecvFromPrepare(std::vector<CommsRequ
if ( (gfrom ==MPI_UNDEFINED) || Stencil_force_mpi ) {
tag= dir+from*32;
host_recv = this->HostBufferMalloc(rbytes);
ierr=MPI_Irecv(host_recv, rbytes, MPI_CHAR,from,tag,communicator_halo[commdir],&rrq);
ierr=MPI_Irecv(host_recv,(int)(rbytes/sizeof(int32_t)), MPI_INT32_T,from,tag,communicator_halo[commdir],&rrq);
GRID_ASSERT(ierr==0);
CommsRequest_t srq;
srq.PacketType = InterNodeRecv;
@@ -686,7 +686,7 @@ void CartesianCommunicator::StencilSendToRecvFromPollDtoH(std::vector<CommsReque
if ( acceleratorEventIsComplete(list[idx].ev) ) {
void *host_xmit = list[idx].host_buf;
uint32_t xbytes = list[idx].bytes;
uint64_t xbytes = list[idx].bytes;
int dest = list[idx].dest;
int tag = list[idx].tag;
int commdir = list[idx].commdir;
@@ -697,7 +697,7 @@ void CartesianCommunicator::StencilSendToRecvFromPollDtoH(std::vector<CommsReque
// std::cout << " DtoH is complete for index "<<idx<<" calling MPI_Isend "<<std::endl;
MPI_Request xrq;
int ierr =MPI_Isend(host_xmit, xbytes, MPI_CHAR,dest,tag,communicator_halo[commdir],&xrq);
int ierr =MPI_Isend(host_xmit, (int)(xbytes/sizeof(int32_t)), MPI_INT32_T,dest,tag,communicator_halo[commdir],&xrq);
GRID_ASSERT(ierr==0);
list[idx].req = xrq; // Update the MPI request in the list
@@ -718,7 +718,7 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
int dest,int dox,
void *recv,void *recv_comp,
int from,int dor,
int xbytes,int rbytes,int dir)
uint64_t xbytes,uint64_t rbytes,int dir)
{
int ncomm =communicator_halo.size();
int commdir=dir%ncomm;
@@ -884,11 +884,11 @@ void CartesianCommunicator::Barrier(void)
int ierr = MPI_Barrier(communicator);
GRID_ASSERT(ierr==0);
}
void CartesianCommunicator::Broadcast(int root,void* data, int bytes)
void CartesianCommunicator::Broadcast(int root,void* data,uint64_t bytes)
{
FlightRecorder::StepLog("Broadcast");
int ierr=MPI_Bcast(data,
bytes,
(int)bytes,
MPI_BYTE,
root,
communicator);
@@ -904,11 +904,11 @@ void CartesianCommunicator::BarrierWorld(void){
int ierr = MPI_Barrier(communicator_world);
GRID_ASSERT(ierr==0);
}
void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes)
void CartesianCommunicator::BroadcastWorld(int root,void* data, uint64_t bytes)
{
FlightRecorder::StepLog("BroadcastWorld");
int ierr= MPI_Bcast(data,
bytes,
(int)bytes,
MPI_BYTE,
root,
communicator_world);
Grid/communicator/Communicator_none.cc
+7 -7
View File
@@ -89,7 +89,7 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
int dest,
void *recv,
int from,
int bytes)
uint64_t bytes)
{
GRID_ASSERT(0);
}
@@ -99,7 +99,7 @@ void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &lis
int dest,
void *recv,
int from,
int bytes,int dir)
uint64_t bytes,int dir)
{
GRID_ASSERT(0);
}
@@ -115,8 +115,8 @@ void CartesianCommunicator::AllToAll(void *in,void *out,uint64_t words,uint64_t
int CartesianCommunicator::RankWorld(void){return 0;}
void CartesianCommunicator::Barrier(void){}
void CartesianCommunicator::Broadcast(int root,void* data, int bytes) {}
void CartesianCommunicator::BroadcastWorld(int root,void* data, int bytes) { }
void CartesianCommunicator::Broadcast(int root,void* data, uint64_t bytes) {}
void CartesianCommunicator::BroadcastWorld(int root,void* data, uint64_t bytes) { }
void CartesianCommunicator::BarrierWorld(void) { }
int CartesianCommunicator::RankFromProcessorCoor(Coordinate &coor) { return 0;}
void CartesianCommunicator::ProcessorCoorFromRank(int rank, Coordinate &coor){ coor = _processor_coor; }
@@ -132,7 +132,7 @@ double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
int xmit_to_rank,int dox,
void *recv,
int recv_from_rank,int dor,
int bytes, int dir)
uint64_t bytes, int dir)
{
return 2.0*bytes;
}
@@ -143,7 +143,7 @@ double CartesianCommunicator::StencilSendToRecvFromPrepare(std::vector<CommsRequ
int xmit_to_rank,int dox,
void *recv,
int recv_from_rank,int dor,
int xbytes,int rbytes, int dir)
uint64_t xbytes,uint64_t rbytes, int dir)
{
return 0.0;
}
@@ -152,7 +152,7 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
int xmit_to_rank,int dox,
void *recv, void *recv_comp,
int recv_from_rank,int dor,
int xbytes,int rbytes, int dir)
uint64_t xbytes,uint64_t rbytes, int dir)
{
return xbytes+rbytes;
}