portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-04-10 06:00:45 +01:00
Code Releases Activity

Merge branch 'develop' of https://github.com/paboyle/Grid into feature-aurora

Browse Source
This commit is contained in:
Christoph Lehner 2025-03-13 07:29:55 +00:00
commit d15a6c5933
6 changed files with 6 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
Grid/allocator/AlignedAllocator.h
View File

@ -69,7 +69,7 @@ public:
} }
// FIXME: hack for the copy constructor: it must be avoided to avoid single thread loop // FIXME: hack for the copy constructor: it must be avoided to avoid single thread loop
void construct(pointer __p, const _Tp& __val) { assert(0);}; void construct(pointer __p, const _Tp& __val) { };
void construct(pointer __p) { }; void construct(pointer __p) { };
void destroy(pointer __p) { }; void destroy(pointer __p) { };
}; };

3
Grid/allocator/MemoryManagerCache.cc
View File

@ -234,6 +234,9 @@ void *MemoryManager::ViewOpen(void* _CpuPtr,size_t bytes,ViewMode mode,ViewAdvis
} }
void MemoryManager::EvictVictims(uint64_t bytes) void MemoryManager::EvictVictims(uint64_t bytes)
{ {
if(bytes>=DeviceMaxBytes) {
printf("EvictVictims bytes %ld DeviceMaxBytes %ld\n",bytes,DeviceMaxBytes);
}
assert(bytes<DeviceMaxBytes); assert(bytes<DeviceMaxBytes);
while(bytes+DeviceLRUBytes > DeviceMaxBytes){ while(bytes+DeviceLRUBytes > DeviceMaxBytes){
if ( DeviceLRUBytes > 0){ if ( DeviceLRUBytes > 0){

3
Grid/communicator/Communicator_mpi3.cc
View File

@ -759,9 +759,6 @@ void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsReque
int nreq=MpiRequests.size(); int nreq=MpiRequests.size();
std::cout << GridLogMessage << " StencilSendToRevFromComplete "<<nreq<<" Mpi Requests"<<std::endl;
if (nreq>0) { if (nreq>0) {
status.resize(MpiRequests.size()); status.resize(MpiRequests.size());
int ierr = MPI_Waitall(MpiRequests.size(),&MpiRequests[0],&status[0]); // Sends are guaranteed in order. No harm in not completing. int ierr = MPI_Waitall(MpiRequests.size(),&MpiRequests[0],&status[0]); // Sends are guaranteed in order. No harm in not completing.

5
Grid/cshift/Cshift_mpi.h
View File

@ -126,8 +126,8 @@ template<class vobj> void Cshift_comms(Lattice<vobj> &ret,const Lattice<vobj> &r
static deviceVector<vobj> send_buf; send_buf.resize(buffer_size); static deviceVector<vobj> send_buf; send_buf.resize(buffer_size);
static deviceVector<vobj> recv_buf; recv_buf.resize(buffer_size); static deviceVector<vobj> recv_buf; recv_buf.resize(buffer_size);
#ifndef ACCELERATOR_AWARE_MPI #ifndef ACCELERATOR_AWARE_MPI
static hostVector<vobj> hsend_buf; hsend_buf.resize(buffer_size); static hostVector<vobj> hsend_buf; hsend_buf.resize(buffer_size);
static hostVector<vobj> hrecv_buf; hrecv_buf.resize(buffer_size); static hostVector<vobj> hrecv_buf; hrecv_buf.resize(buffer_size);
#endif #endif
int cb= (cbmask==0x2)? Odd : Even; int cb= (cbmask==0x2)? Odd : Even;
@ -244,7 +244,6 @@ template<class vobj> void Cshift_comms_simd(Lattice<vobj> &ret,const Lattice<vo
scalar_object * recv_buf_extract_mpi; scalar_object * recv_buf_extract_mpi;
scalar_object * send_buf_extract_mpi; scalar_object * send_buf_extract_mpi;
for(int s=0;s<Nsimd;s++){ for(int s=0;s<Nsimd;s++){
send_buf_extract[s].resize(buffer_size); send_buf_extract[s].resize(buffer_size);
recv_buf_extract[s].resize(buffer_size); recv_buf_extract[s].resize(buffer_size);

1
Grid/qcd/action/fermion/WilsonCompressor.h
View File

@ -485,7 +485,6 @@ public:
assert(this->u_comm_offset==this->_unified_buffer_size); assert(this->u_comm_offset==this->_unified_buffer_size);
accelerator_barrier(); accelerator_barrier();
#ifdef NVLINK_GET #ifdef NVLINK_GET
#warning "NVLINK_GET"
this->_grid->StencilBarrier(); // He can now get mu local gather, I can get his this->_grid->StencilBarrier(); // He can now get mu local gather, I can get his
// Synch shared memory on a single nodes; could use an asynchronous barrier here and defer check // Synch shared memory on a single nodes; could use an asynchronous barrier here and defer check
// Or issue barrier AFTER the DMA is running // Or issue barrier AFTER the DMA is running

1
Grid/stencil/Stencil.h
View File

@ -518,7 +518,6 @@ public:
} }
accelerator_barrier(); // All my local gathers are complete accelerator_barrier(); // All my local gathers are complete
#ifdef NVLINK_GET #ifdef NVLINK_GET
#warning "NVLINK_GET"
_grid->StencilBarrier(); // He can now get mu local gather, I can get his _grid->StencilBarrier(); // He can now get mu local gather, I can get his
// Synch shared memory on a single nodes; could use an asynchronous barrier here and defer check // Synch shared memory on a single nodes; could use an asynchronous barrier here and defer check
// Or issue barrier AFTER the DMA is running // Or issue barrier AFTER the DMA is running