portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2026-05-21 17:44:16 +01:00
Code Releases Activity

Merge remote-tracking branch 'upstream' into gauge_action_deriv

Browse Source
This commit is contained in:
Mashy Green
2025-03-10 08:37:36 +00:00
parent 0000d2e558 3d014864e2
commit 1a1fe85428
13 changed files with 389 additions and 73 deletions
Download Patch File Download Diff File Expand all files Collapse all files
Grid/communicator/Communicator_mpi3.cc
+19 -11
View File
@@ -746,26 +746,34 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
}
void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &list,int dir)
{
// int nreq=list.size();
acceleratorCopySynchronise(); // Complete all pending copy transfers D2D
// if (nreq==0) return;
// std::vector<MPI_Status> status(nreq);
// std::vector<MPI_Request> MpiRequests(nreq);
std::vector<MPI_Status> status;
std::vector<MPI_Request> MpiRequests;
for(int r=0;r<list.size();r++){
// Must check each Send buf is clear to reuse
if ( list[r].PacketType == InterNodeXmitISend ) MpiRequests.push_back(list[r].req);
// if ( list[r].PacketType == InterNodeRecv ) MpiRequests.push_back(list[r].req); // Already "Test" passed
}
// for(int r=0;r<nreq;r++){
// MpiRequests[r] = list[r].req;
// }
int nreq=MpiRequests.size();
std::cout << GridLogMessage << " StencilSendToRevFromComplete "<<nreq<<" Mpi Requests"<<std::endl;
if (nreq>0) {
status.resize(MpiRequests.size());
int ierr = MPI_Waitall(MpiRequests.size(),&MpiRequests[0],&status[0]); // Sends are guaranteed in order. No harm in not completing.
assert(ierr==0);
}
// int ierr = MPI_Waitall(nreq,&MpiRequests[0],&status[0]); // Sends are guaranteed in order. No harm in not completing.
// assert(ierr==0);
// for(int r=0;r<nreq;r++){
// if ( list[r].PacketType==InterNodeRecv ) {
// acceleratorCopyToDeviceAsynch(list[r].host_buf,list[r].device_buf,list[r].bytes);
// }
// }
acceleratorCopySynchronise(); // Complete all pending copy transfers D2D
list.resize(0); // Delete the list
this->HostBufferFreeAll(); // Clean up the buffer allocs
Grid/communicator/Communicator_none.cc
+1 -1
View File
@@ -91,7 +91,7 @@ void CartesianCommunicator::SendToRecvFrom(void *xmit,
{
assert(0);
}
void CartesianCommunicator::CommsComplete(std::vector<CommsRequest_t> &list){ assert(0);}
void CartesianCommunicator::CommsComplete(std::vector<CommsRequest_t> &list){ assert(list.size()==0);}
void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void *xmit,
int dest,
Grid/threads/Accelerator.h
+21 -21
View File
@@ -245,12 +245,12 @@ inline void *acceleratorAllocDevice(size_t bytes)
inline void acceleratorFreeShared(void *ptr){ cudaFree(ptr);};
inline void acceleratorFreeDevice(void *ptr){ cudaFree(ptr);};
inline void acceleratorFreeHost(void *ptr){ cudaFree(ptr);};
inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes) { cudaMemcpy(to,from,bytes, cudaMemcpyHostToDevice);}
inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ cudaMemcpy(to,from,bytes, cudaMemcpyDeviceToHost);}
inline void acceleratorCopyToDeviceAsync(void *from, void *to, size_t bytes, cudaStream_t stream = copyStream) { cudaMemcpyAsync(to,from,bytes, cudaMemcpyHostToDevice, stream);}
inline void acceleratorCopyFromDeviceAsync(void *from, void *to, size_t bytes, cudaStream_t stream = copyStream) { cudaMemcpyAsync(to,from,bytes, cudaMemcpyDeviceToHost, stream);}
inline void acceleratorCopyToDevice(const void *from,void *to,size_t bytes) { cudaMemcpy(to,from,bytes, cudaMemcpyHostToDevice);}
inline void acceleratorCopyFromDevice(const void *from,void *to,size_t bytes){ cudaMemcpy(to,from,bytes, cudaMemcpyDeviceToHost);}
inline void acceleratorCopyToDeviceAsync(const void *from, void *to, size_t bytes, cudaStream_t stream = copyStream) { cudaMemcpyAsync(to,from,bytes, cudaMemcpyHostToDevice, stream);}
inline void acceleratorCopyFromDeviceAsync(const void *from, void *to, size_t bytes, cudaStream_t stream = copyStream) { cudaMemcpyAsync(to,from,bytes, cudaMemcpyDeviceToHost, stream);}
inline void acceleratorMemSet(void *base,int value,size_t bytes) { cudaMemset(base,value,bytes);}
inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) // Asynch
inline void acceleratorCopyDeviceToDeviceAsynch(const void *from,void *to,size_t bytes) // Asynch
{
cudaMemcpyAsync(to,from,bytes, cudaMemcpyDeviceToDevice,copyStream);
}
@@ -359,12 +359,12 @@ inline int acceleratorEventIsComplete(acceleratorEvent_t ev)
return (ev.get_info<sycl::info::event::command_execution_status>() == sycl::info::event_command_status::complete);
}
inline acceleratorEvent_t acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) { return theCopyAccelerator->memcpy(to,from,bytes);}
inline acceleratorEvent_t acceleratorCopyToDeviceAsynch(void *from,void *to,size_t bytes) { return theCopyAccelerator->memcpy(to,from,bytes); }
inline acceleratorEvent_t acceleratorCopyFromDeviceAsynch(void *from,void *to,size_t bytes) { return theCopyAccelerator->memcpy(to,from,bytes); }
inline acceleratorEvent_t acceleratorCopyDeviceToDeviceAsynch(const void *from,void *to,size_t bytes) { return theCopyAccelerator->memcpy(to,from,bytes);}
inline acceleratorEvent_t acceleratorCopyToDeviceAsynch(const void *from,void *to,size_t bytes) { return theCopyAccelerator->memcpy(to,from,bytes); }
inline acceleratorEvent_t acceleratorCopyFromDeviceAsynch(const void *from,void *to,size_t bytes) { return theCopyAccelerator->memcpy(to,from,bytes); }
inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes) { theCopyAccelerator->memcpy(to,from,bytes); theCopyAccelerator->wait();}
inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ theCopyAccelerator->memcpy(to,from,bytes); theCopyAccelerator->wait();}
inline void acceleratorCopyToDevice(const void *from,void *to,size_t bytes) { theCopyAccelerator->memcpy(to,from,bytes); theCopyAccelerator->wait();}
inline void acceleratorCopyFromDevice(const void *from,void *to,size_t bytes){ theCopyAccelerator->memcpy(to,from,bytes); theCopyAccelerator->wait();}
inline void acceleratorMemSet(void *base,int value,size_t bytes) { theCopyAccelerator->memset(base,value,bytes); theCopyAccelerator->wait();}
inline int acceleratorIsCommunicable(void *ptr)
@@ -511,19 +511,19 @@ inline void *acceleratorAllocDevice(size_t bytes)
inline void acceleratorFreeHost(void *ptr){ auto discard=hipFree(ptr);};
inline void acceleratorFreeShared(void *ptr){ auto discard=hipFree(ptr);};
inline void acceleratorFreeDevice(void *ptr){ auto discard=hipFree(ptr);};
inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes) { auto discard=hipMemcpy(to,from,bytes, hipMemcpyHostToDevice);}
inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ auto discard=hipMemcpy(to,from,bytes, hipMemcpyDeviceToHost);}
inline void acceleratorCopyToDevice(const void *from,void *to,size_t bytes) { auto discard=hipMemcpy(to,from,bytes, hipMemcpyHostToDevice);}
inline void acceleratorCopyFromDevice(const void *from,void *to,size_t bytes){ auto discard=hipMemcpy(to,from,bytes, hipMemcpyDeviceToHost);}
inline void acceleratorMemSet(void *base,int value,size_t bytes) { auto discard=hipMemset(base,value,bytes);}
inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) // Asynch
inline void acceleratorCopyDeviceToDeviceAsynch(const void *from,void *to,size_t bytes) // Asynch
{
auto discard=hipMemcpyDtoDAsync(to,from,bytes, copyStream);
}
inline void acceleratorCopyToDeviceAsync(void *from, void *to, size_t bytes, hipStream_t stream = copyStream) {
inline void acceleratorCopyToDeviceAsync(const void *from, void *to, size_t bytes, hipStream_t stream = copyStream) {
auto r = hipMemcpyAsync(to,from,bytes, hipMemcpyHostToDevice, stream);
}
inline void acceleratorCopyFromDeviceAsync(void *from, void *to, size_t bytes, hipStream_t stream = copyStream) {
inline void acceleratorCopyFromDeviceAsync(const void *from, void *to, size_t bytes, hipStream_t stream = copyStream) {
auto r = hipMemcpyAsync(to,from,bytes, hipMemcpyDeviceToHost, stream);
}
inline void acceleratorCopySynchronise(void) { auto discard=hipStreamSynchronize(copyStream); };
@@ -583,9 +583,9 @@ inline void acceleratorMem(void)
accelerator_inline int acceleratorSIMTlane(int Nsimd) { return 0; } // CUDA specific
inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes) { thread_bcopy(from,to,bytes); }
inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ thread_bcopy(from,to,bytes);}
inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) { thread_bcopy(from,to,bytes);}
inline void acceleratorCopyToDevice(const void *from,void *to,size_t bytes) { thread_bcopy(from,to,bytes); }
inline void acceleratorCopyFromDevice(const void *from,void *to,size_t bytes){ thread_bcopy(from,to,bytes);}
inline void acceleratorCopyDeviceToDeviceAsynch(const void *from,void *to,size_t bytes) { thread_bcopy(from,to,bytes);}
inline void acceleratorCopySynchronise(void) {};
inline int acceleratorIsCommunicable(void *ptr){ return 1; }
@@ -668,15 +668,15 @@ accelerator_inline void acceleratorFence(void)
return;
}
inline void acceleratorCopyDeviceToDevice(void *from,void *to,size_t bytes)
inline void acceleratorCopyDeviceToDevice(const void *from,void *to,size_t bytes)
{
acceleratorCopyDeviceToDeviceAsynch(from,to,bytes);
acceleratorCopySynchronise();
}
template<class T> void acceleratorPut(T& dev,T&host)
template<class T> void acceleratorPut(T& dev,const T&host)
{
acceleratorCopyToDevice(&host,&dev,sizeof(T));
acceleratorCopyToDevice((void *)&host,&dev,sizeof(T));
}
template<class T> T acceleratorGet(T& dev)
{
Grid/threads/Threads.h
+3 -3
View File
@@ -73,9 +73,9 @@ Author: paboyle <paboyle@ph.ed.ac.uk>
#define thread_critical DO_PRAGMA(omp critical)
#ifdef GRID_OMP
inline void thread_bcopy(void *from, void *to,size_t bytes)
inline void thread_bcopy(const void *from, void *to,size_t bytes)
{
uint64_t *ufrom = (uint64_t *)from;
const uint64_t *ufrom = (const uint64_t *)from;
uint64_t *uto = (uint64_t *)to;
assert(bytes%8==0);
uint64_t words=bytes/8;
@@ -84,7 +84,7 @@ inline void thread_bcopy(void *from, void *to,size_t bytes)
});
}
#else
inline void thread_bcopy(void *from, void *to,size_t bytes)
inline void thread_bcopy(const void *from, void *to,size_t bytes)
{
bcopy(from,to,bytes);
}
Grid/util/Init.cc
+9 -1
View File
@@ -509,7 +509,14 @@ void Grid_init(int *argc,char ***argv)
Grid_default_latt,
Grid_default_mpi);
if( GridCmdOptionExists(*argv,*argv+*argc,"--flightrecorder") ){
std::cout << GridLogMessage <<" Enabling flight recorder " <<std::endl;
FlightRecorder::SetLoggingMode(FlightRecorder::LoggingModeRecord);
FlightRecorder::PrintEntireLog = 1;
FlightRecorder::ChecksumComms = 1;
FlightRecorder::ChecksumCommsSend=1;
}
if( GridCmdOptionExists(*argv,*argv+*argc,"--decomposition") ){
std::cout<<GridLogMessage<<"Grid Default Decomposition patterns\n";
std::cout<<GridLogMessage<<"\tOpenMP threads : "<<GridThread::GetThreads()<<std::endl;
@@ -651,3 +658,4 @@ void Grid_debug_handler_init(void)
}
NAMESPACE_END(Grid);
Grid/util/Lexicographic.h
+1 -1
View File
@@ -50,7 +50,7 @@ namespace Grid{
int64_t index64;
IndexFromCoorReversed(coor,index64,dims);
if ( index64>=2*1024*1024*1024LL ){
std::cout << " IndexFromCoorReversed " << coor<<" index " << index64<< " dims "<<dims<<std::endl;
// std::cout << " IndexFromCoorReversed " << coor<<" index " << index64<< " dims "<<dims<<std::endl;
}
assert(index64<2*1024*1024*1024LL);
index = (int) index64;