portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-20 00:36:55 +01:00
Code Releases Activity

Significantly better performance on Aurora without using pipeline mode

Browse Source
This commit is contained in:
Peter Boyle
2025-01-30 16:36:46 +00:00
parent d6b2727f86
commit 94019a922e
12 changed files with 306 additions and 101 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
Grid/threads/Accelerator.h
View File

@ -327,7 +327,10 @@ inline void acceleratorFreeShared(void *ptr){free(ptr,*theGridAccelerator);};
inline void acceleratorFreeDevice(void *ptr){free(ptr,*theGridAccelerator);};
inline void acceleratorCopySynchronise(void) { theCopyAccelerator->wait(); }
inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) { theCopyAccelerator->memcpy(to,from,bytes);}
inline void acceleratorCopyToDeviceAsynch(void *from,void *to,size_t bytes) { theCopyAccelerator->memcpy(to,from,bytes); }
inline void acceleratorCopyFromDeviceAsynch(void *from,void *to,size_t bytes){ theCopyAccelerator->memcpy(to,from,bytes); }
inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes) { theCopyAccelerator->memcpy(to,from,bytes); theCopyAccelerator->wait();}
inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ theCopyAccelerator->memcpy(to,from,bytes); theCopyAccelerator->wait();}
inline void acceleratorMemSet(void *base,int value,size_t bytes) { theCopyAccelerator->memset(base,value,bytes); theCopyAccelerator->wait();}
@ -465,8 +468,7 @@ inline void acceleratorFreeShared(void *ptr){ auto discard=hipFree(ptr);};
inline void acceleratorFreeDevice(void *ptr){ auto discard=hipFree(ptr);};
inline void acceleratorCopyToDevice(void *from,void *to,size_t bytes) { auto discard=hipMemcpy(to,from,bytes, hipMemcpyHostToDevice);}
inline void acceleratorCopyFromDevice(void *from,void *to,size_t bytes){ auto discard=hipMemcpy(to,from,bytes, hipMemcpyDeviceToHost);}
//inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) { hipMemcpy(to,from,bytes, hipMemcpyDeviceToDevice);}
//inline void acceleratorCopySynchronise(void) { }
inline void acceleratorMemSet(void *base,int value,size_t bytes) { auto discard=hipMemset(base,value,bytes);}
inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) // Asynch
@ -483,6 +485,13 @@ inline void acceleratorCopySynchronise(void) { auto discard=hipStreamSynchronize
#endif
inline void acceleratorPin(void *ptr,unsigned long bytes)
{
#ifdef GRID_SYCL
sycl::ext::oneapi::experimental::prepare_for_device_copy(ptr,bytes,theCopyAccelerator->get_context());
#endif
}
//////////////////////////////////////////////
// Common on all GPU targets
//////////////////////////////////////////////