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

Merge pull request #17 from paboyle/develop

Browse Source 
sync upstream
This commit is contained in:
Christoph Lehner
2020-10-06 09:01:39 +02:00
committed by GitHub
parent 5534921bee ace9cd64bb
commit 80fd6ab407
20 changed files with 170 additions and 132 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
Grid/threads/Accelerator.cc
View File

@ -59,6 +59,7 @@ void acceleratorInit(void)
printf("AcceleratorCudaInit[%d]: ========================\n",rank);
printf("AcceleratorCudaInit[%d]: Device identifier: %s\n",rank, prop.name);
GPU_PROP_FMT(totalGlobalMem,"%lld");
GPU_PROP(managedMemory);
GPU_PROP(isMultiGpuBoard);
@ -113,20 +114,24 @@ 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);}
printf("world_rank %d has %d devices\n",world_rank,nDevices);
size_t totalDeviceMem=0;
for (int i = 0; i < nDevices; i++) {
#define GPU_PROP_FMT(canMapHostMemory,FMT) printf("AcceleratorHipInit: " #canMapHostMemory ": " FMT" \n",prop.canMapHostMemory);
#define GPU_PROP(canMapHostMemory) GPU_PROP_FMT(canMapHostMemory,"%d");
hipGetDeviceProperties(&gpu_props[i], i);
hipDeviceProp_t prop;
prop = gpu_props[i];
totalDeviceMem = prop.totalGlobalMem;
if ( world_rank == 0) {
hipDeviceProp_t prop;
prop = gpu_props[i];
printf("AcceleratorHipInit: ========================\n");
printf("AcceleratorHipInit: Device Number : %d\n", i);
printf("AcceleratorHipInit: ========================\n");
printf("AcceleratorHipInit: Device identifier: %s\n", prop.name);
GPU_PROP_FMT(totalGlobalMem,"%lu");
// GPU_PROP(managedMemory);
GPU_PROP(isMultiGpuBoard);
GPU_PROP(warpSize);
@ -135,6 +140,7 @@ void acceleratorInit(void)
// GPU_PROP(singleToDoublePrecisionPerfRatio);
}
}
MemoryManager::DeviceMaxBytes = (8*totalDeviceMem)/10; // Assume 80% ours
#undef GPU_PROP_FMT
#undef GPU_PROP
#ifdef GRID_IBM_SUMMIT

8
Grid/threads/Accelerator.h
View File

@ -307,17 +307,13 @@ void LambdaApply(uint64_t numx, uint64_t numy, uint64_t numz, lambda Lambda)
inline void *acceleratorAllocShared(size_t bytes)
{
#if 0
void *ptr=NULL;
auto err = hipMallocManaged((void **)&ptr,bytes);
if( err != hipSuccess ) {
ptr = (void *) NULL;
printf(" hipMallocManaged failed for %d %s \n",bytes,hipGetErrorString(err));
printf(" hipMallocManaged failed for %ld %s \n",bytes,hipGetErrorString(err));
}
return ptr;
#else
return malloc(bytes);
#endif
};
inline int acceleratorIsCommunicable(void *ptr){ return 1; }
@ -327,7 +323,7 @@ inline void *acceleratorAllocDevice(size_t bytes)
auto err = hipMalloc((void **)&ptr,bytes);
if( err != hipSuccess ) {
ptr = (void *) NULL;
printf(" hipMalloc failed for %d %s \n",bytes,hipGetErrorString(err));
printf(" hipMalloc failed for %ld %s \n",bytes,hipGetErrorString(err));
}
return ptr;
};