portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-18 07:47:06 +01:00
Code Releases Activity

Priintinig and device memory size detection

Browse Source
This commit is contained in:
Peter Boyle
2020-06-04 14:58:03 -04:00
parent 0d95805cde
commit f39c2a240b
4 changed files with 19 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
Grid/threads/Accelerator.cc
View File

@ -31,6 +31,7 @@ void acceleratorInit(void)
if ((localRankStr = getenv(ENV_RANK_OMPI )) != NULL) { world_rank = atoi(localRankStr);}
if ((localRankStr = getenv(ENV_RANK_MVAPICH)) != NULL) { world_rank = atoi(localRankStr);}
size_t totalDeviceMem=0;
for (int i = 0; i < nDevices; i++) {
#define GPU_PROP_FMT(canMapHostMemory,FMT) printf("AcceleratorCudaInit: " #canMapHostMemory ": " FMT" \n",prop.canMapHostMemory);
@ -45,14 +46,18 @@ void acceleratorInit(void)
printf("AcceleratorCudaInit: ========================\n");
printf("AcceleratorCudaInit: Device identifier: %s\n", prop.name);
GPU_PROP_FMT(totalGlobalMem,"%lld");
GPU_PROP(managedMemory);
GPU_PROP(isMultiGpuBoard);
GPU_PROP(warpSize);
totalDeviceMem = prop.totalGlobalMem;
// GPU_PROP(unifiedAddressing);
// GPU_PROP(l2CacheSize);
// GPU_PROP(singleToDoublePrecisionPerfRatio);
}
}
MemoryManager::DeviceMaxBytes = (8*totalDeviceMem)/10; // Assume 80% ours
#ifdef GRID_IBM_SUMMIT
// IBM Jsrun makes cuda Device numbering screwy and not match rank
if ( world_rank == 0 ) printf("AcceleratorCudaInit: IBM Summit or similar - NOT setting device to node rank\n");