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 #18 from paboyle/develop

Browse Source 
Sync
This commit is contained in:
Christoph Lehner
2020-11-20 15:36:50 +01:00
committed by GitHub
parent 80fd6ab407 e49b7f2f88
commit 4ea8d128c2
170 changed files with 2504 additions and 564 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
Grid/threads/Accelerator.cc
View File

@ -52,7 +52,7 @@ void acceleratorInit(void)
prop = gpu_props[i];
totalDeviceMem = prop.totalGlobalMem;
if ( world_rank == 0) {
#ifndef GRID_IBM_SUMMIT
#ifndef GRID_DEFAULT_GPU
if ( i==rank ) {
printf("AcceleratorCudaInit[%d]: ========================\n",rank);
printf("AcceleratorCudaInit[%d]: Device Number : %d\n", rank,i);
@ -77,11 +77,17 @@ void acceleratorInit(void)
#undef GPU_PROP_FMT
#undef GPU_PROP
#ifdef GRID_IBM_SUMMIT
#ifdef GRID_DEFAULT_GPU
// 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
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);
#endif
if ( world_rank == 0 ) printf("AcceleratorCudaInit: ================================================\n");
@ -143,11 +149,18 @@ void acceleratorInit(void)
MemoryManager::DeviceMaxBytes = (8*totalDeviceMem)/10; // Assume 80% ours
#undef GPU_PROP_FMT
#undef GPU_PROP
#ifdef GRID_IBM_SUMMIT
// IBM Jsrun makes cuda Device numbering screwy and not match rank
if ( world_rank == 0 ) printf("AcceleratorHipInit: IBM Summit or similar - NOT setting device to node rank\n");
#ifdef GRID_DEFAULT_GPU
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
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);
#endif
if ( world_rank == 0 ) printf("AcceleratorHipInit: ================================================\n");

25
Grid/threads/Accelerator.h
View File

@ -166,15 +166,18 @@ inline void *acceleratorAllocDevice(size_t bytes)
inline void acceleratorFreeShared(void *ptr){ cudaFree(ptr);};
inline void acceleratorFreeDevice(void *ptr){ cudaFree(ptr);};
inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes) { cudaMemcpy(to,from,bytes, cudaMemcpyHostToDevice);}
inline void acceleratorCopyDeviceToDevice(void *from,void *to,size_t bytes) { cudaMemcpy(to,from,bytes, cudaMemcpyDeviceToDevice);}
inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ cudaMemcpy(to,from,bytes, cudaMemcpyDeviceToHost);}
inline void acceleratorMemSet(void *base,int value,size_t bytes) { cudaMemset(base,value,bytes);}
inline int acceleratorIsCommunicable(void *ptr)
{
int uvm;
auto
cuerr = cuPointerGetAttribute( &uvm, CU_POINTER_ATTRIBUTE_IS_MANAGED, (CUdeviceptr) ptr);
assert(cuerr == cudaSuccess );
if(uvm) return 0;
else return 1;
// int uvm=0;
// auto
// cuerr = cuPointerGetAttribute( &uvm, CU_POINTER_ATTRIBUTE_IS_MANAGED, (CUdeviceptr) ptr);
// assert(cuerr == cudaSuccess );
// if(uvm) return 0;
// else return 1;
return 1;
}
#endif
@ -229,8 +232,10 @@ inline void *acceleratorAllocShared(size_t bytes){ return malloc_shared(bytes,*t
inline void *acceleratorAllocDevice(size_t bytes){ return malloc_device(bytes,*theGridAccelerator);};
inline void acceleratorFreeShared(void *ptr){free(ptr,*theGridAccelerator);};
inline void acceleratorFreeDevice(void *ptr){free(ptr,*theGridAccelerator);};
inline void acceleratorCopyDeviceToDevice(void *from,void *to,size_t bytes) { theGridAccelerator->memcpy(to,from,bytes); theGridAccelerator->wait();}
inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes) { theGridAccelerator->memcpy(to,from,bytes); theGridAccelerator->wait();}
inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ theGridAccelerator->memcpy(to,from,bytes); theGridAccelerator->wait();}
inline void acceleratorMemSet(void *base,int value,size_t bytes) { theGridAccelerator->memset(base,value,bytes); theGridAccelerator->wait();}
inline int acceleratorIsCommunicable(void *ptr)
{
#if 0
@ -328,10 +333,12 @@ inline void *acceleratorAllocDevice(size_t bytes)
return ptr;
};
inline void acceleratorFreeShared(void *ptr){ free(ptr);};
inline void acceleratorFreeShared(void *ptr){ hipFree(ptr);};
inline void acceleratorFreeDevice(void *ptr){ hipFree(ptr);};
inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes) { hipMemcpy(to,from,bytes, hipMemcpyHostToDevice);}
inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ hipMemcpy(to,from,bytes, hipMemcpyDeviceToHost);}
inline void acceleratorCopyDeviceToDevice(void *from,void *to,size_t bytes) { hipMemcpy(to,from,bytes, hipMemcpyDeviceToDevice);}
inline void acceleratorMemSet(void *base,int value,size_t bytes) { hipMemset(base,value,bytes);}
#endif
@ -369,8 +376,10 @@ inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ hipMemc
accelerator_inline int acceleratorSIMTlane(int Nsimd) { return 0; } // CUDA specific
inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes) { memcpy(to,from,bytes);}
inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ memcpy(to,from,bytes);}
inline void acceleratorCopyDeviceToDevice(void *from,void *to,size_t bytes) { memcpy(to,from,bytes);}
inline int acceleratorIsCommunicable(void *ptr){ return 1; }
inline void acceleratorMemSet(void *base,int value,size_t bytes) { memset(base,value,bytes);}
#ifdef HAVE_MM_MALLOC_H
inline void *acceleratorAllocShared(size_t bytes){return _mm_malloc(bytes,GRID_ALLOC_ALIGN);};
inline void *acceleratorAllocDevice(size_t bytes){return _mm_malloc(bytes,GRID_ALLOC_ALIGN);};
@ -393,6 +402,8 @@ inline void *acceleratorAllocCpu(size_t bytes){return memalign(GRID_ALLOC_ALIGN,
inline void acceleratorFreeCpu (void *ptr){free(ptr);};
#endif
///////////////////////////////////////////////////
// Synchronise across local threads for divergence resynch
///////////////////////////////////////////////////