portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-08-15 10:41:54 +01:00
Code Releases Activity

Intel DPCPP Gold happy now (compiles all, runs Benchmark_dwf_fp32 )

Browse Source
This commit is contained in:
Peter Boyle
2020-12-03 03:47:11 -08:00
parent 497e7c1c40
commit cf76741ec6
9 changed files with 97 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
Grid/communicator/SharedMemory.h
View File

@@ -102,7 +102,7 @@ public:
///////////////////////////////////////////////////
static void SharedMemoryAllocate(uint64_t bytes, int flags);
static void SharedMemoryFree(void);
static void SharedMemoryCopy(void *dest,const void *src,size_t bytes);
static void SharedMemoryCopy(void *dest,void *src,size_t bytes);
static void SharedMemoryZero(void *dest,size_t bytes);
};

2
Grid/communicator/SharedMemoryMPI.cc
View File

@@ -715,7 +715,7 @@ void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes)
bzero(dest,bytes);
#endif
}
void GlobalSharedMemory::SharedMemoryCopy(void *dest,const void *src,size_t bytes)
void GlobalSharedMemory::SharedMemoryCopy(void *dest,void *src,size_t bytes)
{
#ifdef GRID_CUDA
cudaMemcpy(dest,src,bytes,cudaMemcpyDefault);

43
Grid/communicator/SharedMemoryNone.cc
View File

@@ -29,6 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#include <Grid/GridCore.h>
NAMESPACE_BEGIN(Grid);
#define header "SharedMemoryNone: "
/*Construct from an MPI communicator*/
void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
@@ -55,6 +56,38 @@ void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_M
////////////////////////////////////////////////////////////////////////////////////////////
// Hugetlbfs mapping intended, use anonymous mmap
////////////////////////////////////////////////////////////////////////////////////////////
#if 1
void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
{
std::cout << header "SharedMemoryAllocate "<< bytes<< " GPU implementation "<<std::endl;
void * ShmCommBuf ;
assert(_ShmSetup==1);
assert(_ShmAlloc==0);
///////////////////////////////////////////////////////////////////////////////////////////////////////////
// Each MPI rank should allocate our own buffer
///////////////////////////////////////////////////////////////////////////////////////////////////////////
ShmCommBuf = acceleratorAllocDevice(bytes);
if (ShmCommBuf == (void *)NULL ) {
std::cerr << " SharedMemoryNone.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl;
exit(EXIT_FAILURE);
}
if ( WorldRank == 0 ){
std::cout << WorldRank << header " SharedMemoryNone.cc acceleratorAllocDevice "<< bytes
<< "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
}
SharedMemoryZero(ShmCommBuf,bytes);
///////////////////////////////////////////////////////////////////////////////////////////////////////////
// Loop over ranks/gpu's on our node
///////////////////////////////////////////////////////////////////////////////////////////////////////////
WorldShmCommBufs[0] = ShmCommBuf;
_ShmAllocBytes=bytes;
_ShmAlloc=1;
}
#else
void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
{
void * ShmCommBuf ;
@@ -83,7 +116,15 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
_ShmAllocBytes=bytes;
_ShmAlloc=1;
};
#endif
void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes)
{
acceleratorMemSet(dest,0,bytes);
}
void GlobalSharedMemory::SharedMemoryCopy(void *dest,void *src,size_t bytes)
{
acceleratorCopyToDevice(src,dest,bytes);
}
////////////////////////////////////////////////////////
// Global shared functionality finished
// Now move to per communicator functionality