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

Fencing on a stream in SYCL is needed. Didn't know that ... gulp

Browse Source
This commit is contained in:
Peter Boyle 2022-08-02 07:58:04 -07:00
parent 2ab1af5754
commit d32b923b6c
3 changed files with 13 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
Grid/qcd/action/fermion/implementation/WilsonKernelsImplementation.h
View File

@ -498,6 +498,7 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st, DoubledGaugeField
#ifndef GRID_CUDA
if (Opt == WilsonKernelsStatic::OptInlineAsm ) { ASM_CALL(AsmDhopSiteDag); return;}
#endif
acceleratorFenceComputeStream();
} else if( interior ) {
if (Opt == WilsonKernelsStatic::OptGeneric ) { KERNEL_CALL(GenericDhopSiteDagInt); return;}
if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagInt); return;}
@ -505,11 +506,13 @@ void WilsonKernels<Impl>::DhopKernel(int Opt,StencilImpl &st, DoubledGaugeField
if (Opt == WilsonKernelsStatic::OptInlineAsm ) { ASM_CALL(AsmDhopSiteDagInt); return;}
#endif
} else if( exterior ) {
acceleratorFenceComputeStream();
if (Opt == WilsonKernelsStatic::OptGeneric ) { KERNEL_CALL(GenericDhopSiteDagExt); return;}
if (Opt == WilsonKernelsStatic::OptHandUnroll ) { KERNEL_CALL(HandDhopSiteDagExt); return;}
#ifndef GRID_CUDA
if (Opt == WilsonKernelsStatic::OptInlineAsm ) { ASM_CALL(AsmDhopSiteDagExt); return;}
#endif
acceleratorFenceComputeStream();
}
assert(0 && " Kernel optimisation case not covered ");
}

3
Grid/threads/Accelerator.cc
View File

@ -202,7 +202,8 @@ void acceleratorInit(void)
cl::sycl::gpu_selector selector;
cl::sycl::device selectedDevice { selector };
theGridAccelerator = new sycl::queue (selectedDevice);
theCopyAccelerator = new sycl::queue (selectedDevice);
// theCopyAccelerator = new sycl::queue (selectedDevice);
theCopyAccelerator = theGridAccelerator; // Should proceed concurrenlty anyway.
#ifdef GRID_SYCL_LEVEL_ZERO_IPC
zeInit(0);

11
Grid/threads/Accelerator.h
View File

@ -247,7 +247,12 @@ inline int acceleratorIsCommunicable(void *ptr)
//////////////////////////////////////////////
// SyCL acceleration
//////////////////////////////////////////////
#ifdef GRID_SYCL
inline void acceleratorFenceComputeStream(void){ accelerator_barrier();};
#else
// Ordering within a stream guaranteed on Nvidia & AMD
inline void acceleratorFenceComputeStream(void){ };
#endif
#ifdef GRID_SYCL
NAMESPACE_END(Grid);
#include <CL/sycl.hpp>
@ -299,15 +304,15 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
}); \
});
#define accelerator_barrier(dummy) { printf(" theGridAccelerator::wait()\n"); ; theGridAccelerator->wait(); }
#define accelerator_barrier(dummy) { printf(" theGridAccelerator::wait()\n"); theGridAccelerator->wait(); }
inline void *acceleratorAllocShared(size_t bytes){ return malloc_shared(bytes,*theGridAccelerator);};
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 acceleratorCopySynchronise(void) { printf(" theCopyAccelerator::wait()\n"); theCopyAccelerator->wait(); }
inline void acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) { theCopyAccelerator->memcpy(to,from,bytes);}
inline void acceleratorCopySynchronise(void) { theCopyAccelerator->wait(); }
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();}