portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-10 07:55:35 +00:00
Code Releases Activity

Use of default GPU is behaviour, not a system property. Move Summit specific to configure.ac

Browse Source
This commit is contained in:
Peter Boyle 2020-11-13 03:50:16 +01:00
parent b13d1f7238
commit 6e313575be
1 changed files with 20 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

27
Grid/threads/Accelerator.cc
View File

@ -48,7 +48,7 @@ void acceleratorInit(void)
prop = gpu_props[i]; prop = gpu_props[i];
totalDeviceMem = prop.totalGlobalMem; totalDeviceMem = prop.totalGlobalMem;
if ( world_rank == 0) { if ( world_rank == 0) {
#ifndef GRID_IBM_SUMMIT #ifndef GRID_DEFAULT_GPU
if ( i==rank ) { if ( i==rank ) {
printf("AcceleratorCudaInit[%d]: ========================\n",rank); printf("AcceleratorCudaInit[%d]: ========================\n",rank);
printf("AcceleratorCudaInit[%d]: Device Number : %d\n", rank,i); printf("AcceleratorCudaInit[%d]: Device Number : %d\n", rank,i);
@ -73,11 +73,17 @@ void acceleratorInit(void)
#undef GPU_PROP_FMT #undef GPU_PROP_FMT
#undef GPU_PROP #undef GPU_PROP
#ifdef GRID_IBM_SUMMIT #ifdef GRID_DEFAULT_GPU
// IBM Jsrun makes cuda Device numbering screwy and not match rank // IBM Jsrun makes cuda Device numbering screwy and not match rank
if ( world_rank == 0 ) printf("AcceleratorCudaInit: IBM Summit or similar - use default device\n"); if ( world_rank == 0 ) {
printf("AcceleratorCudaInit: using default device \n");
printf("AcceleratorCudaInit: assume user either uses a) IBM jsrun, or \n");
printf("AcceleratorCudaInit: b) invokes through a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding \n");
printf("AcceleratorCudaInit: Configure options --enable-summit, --enable-select-gpu=no \n");
}
#else #else
printf("AcceleratorCudaInit: rank %d setting device to node rank %d\n",world_rank,rank); printf("AcceleratorCudaInit: rank %d setting device to node rank %d\n",world_rank,rank);
printf("AcceleratorCudaInit: Configure options --enable-select-gpu=yes \n");
cudaSetDevice(rank); cudaSetDevice(rank);
#endif #endif
if ( world_rank == 0 ) printf("AcceleratorCudaInit: ================================================\n"); if ( world_rank == 0 ) printf("AcceleratorCudaInit: ================================================\n");
@ -139,11 +145,18 @@ void acceleratorInit(void)
MemoryManager::DeviceMaxBytes = (8*totalDeviceMem)/10; // Assume 80% ours MemoryManager::DeviceMaxBytes = (8*totalDeviceMem)/10; // Assume 80% ours
#undef GPU_PROP_FMT #undef GPU_PROP_FMT
#undef GPU_PROP #undef GPU_PROP
#ifdef GRID_IBM_SUMMIT
// IBM Jsrun makes cuda Device numbering screwy and not match rank #ifdef GRID_DEFAULT_GPU
if ( world_rank == 0 ) printf("AcceleratorHipInit: IBM Summit or similar - NOT setting device to node rank\n"); if ( world_rank == 0 ) {
printf("AcceleratorHipInit: using default device \n");
printf("AcceleratorHipInit: assume user either uses a wrapping script to set CUDA_VISIBLE_DEVICES, UCX_NET_DEVICES, and numa binding \n");
printf("AcceleratorHipInit: Configure options --enable-summit, --enable-select-gpu=no \n");
}
#else #else
if ( world_rank == 0 ) printf("AcceleratorHipInit: setting device to node rank\n"); if ( world_rank == 0 ) {
printf("AcceleratorHipInit: rank %d setting device to node rank %d\n",world_rank,rank);
printf("AcceleratorHipInit: Configure options --enable-select-gpu=yes \n");
}
hipSetDevice(rank); hipSetDevice(rank);
#endif #endif
if ( world_rank == 0 ) printf("AcceleratorHipInit: ================================================\n"); if ( world_rank == 0 ) printf("AcceleratorHipInit: ================================================\n");