portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-21 09:12:03 +01:00
Code Releases Activity

merged sycl to feature-gpt

Browse Source
This commit is contained in:
Christoph Lehner
2020-06-12 06:49:38 -04:00
parent 9fcb47ee63 32b2b59be4
commit 7974acff54
131 changed files with 4722 additions and 2571 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
Grid/communicator/SharedMemoryMPI.cc
View File

@ -29,7 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#include <Grid/GridCore.h>
#include <pwd.h>
#ifdef GRID_NVCC
#ifdef GRID_CUDA
#include <cuda_runtime_api.h>
#endif
@ -420,7 +420,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
////////////////////////////////////////////////////////////////////////////////////////////
// Hugetlbfs mapping intended
////////////////////////////////////////////////////////////////////////////////////////////
#ifdef GRID_NVCC
#ifdef GRID_CUDA
void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
{
void * ShmCommBuf ;
@ -440,13 +440,6 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
//////////////////////////////////////////////////////////////////////////////////////////////////////////
// cudaDeviceGetP2PAttribute(&perfRank, cudaDevP2PAttrPerformanceRank, device1, device2);
#ifdef GRID_IBM_SUMMIT
// IBM Jsrun makes cuda Device numbering screwy and not match rank
std::cout << "IBM Summit or similar - NOT setting device to WorldShmRank"<<std::endl;
#else
std::cout << "setting device to WorldShmRank"<<std::endl;
cudaSetDevice(WorldShmRank);
#endif
///////////////////////////////////////////////////////////////////////////////////////////////////////////
// Each MPI rank should allocate our own buffer
///////////////////////////////////////////////////////////////////////////////////////////////////////////
@ -684,7 +677,7 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
/////////////////////////////////////////////////////////////////////////
void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes)
{
#ifdef GRID_NVCC
#ifdef GRID_CUDA
cudaMemset(dest,0,bytes);
#else
bzero(dest,bytes);
@ -692,7 +685,7 @@ void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes)
}
void GlobalSharedMemory::SharedMemoryCopy(void *dest,const void *src,size_t bytes)
{
#ifdef GRID_NVCC
#ifdef GRID_CUDA
cudaMemcpy(dest,src,bytes,cudaMemcpyDefault);
#else
bcopy(src,dest,bytes);