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 branch 'develop' into feature/dirichlet

Browse Source
This commit is contained in:
Peter Boyle
2022-08-31 19:05:57 -04:00
parent 442bfb3d42 913fbca74a
commit 1177b8f661
15 changed files with 167 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
Grid/threads/Accelerator.cc
View File

@ -201,12 +201,15 @@ void acceleratorInit(void)
#ifdef GRID_SYCL
cl::sycl::queue *theGridAccelerator;
cl::sycl::queue *theCopyAccelerator;
void acceleratorInit(void)
{
int nDevices = 1;
cl::sycl::gpu_selector selector;
cl::sycl::device selectedDevice { selector };
theGridAccelerator = new sycl::queue (selectedDevice);
// theCopyAccelerator = new sycl::queue (selectedDevice);
theCopyAccelerator = theGridAccelerator; // Should proceed concurrenlty anyway.
#ifdef GRID_SYCL_LEVEL_ZERO_IPC
zeInit(0);

29
Grid/threads/Accelerator.h
View File

@ -248,7 +248,6 @@ inline int acceleratorIsCommunicable(void *ptr)
//////////////////////////////////////////////
// SyCL acceleration
//////////////////////////////////////////////
#ifdef GRID_SYCL
NAMESPACE_END(Grid);
#include <CL/sycl.hpp>
@ -263,6 +262,7 @@ NAMESPACE_END(Grid);
NAMESPACE_BEGIN(Grid);
extern cl::sycl::queue *theGridAccelerator;
extern cl::sycl::queue *theCopyAccelerator;
#ifdef __SYCL_DEVICE_ONLY__
#define GRID_SIMT
@ -290,7 +290,7 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
cgh.parallel_for( \
cl::sycl::nd_range<3>(global,local), \
[=] (cl::sycl::nd_item<3> item) /*mutable*/ \
[[intel::reqd_sub_group_size(8)]] \
[[intel::reqd_sub_group_size(16)]] \
{ \
auto iter1 = item.get_global_id(0); \
auto iter2 = item.get_global_id(1); \
@ -299,19 +299,19 @@ accelerator_inline int acceleratorSIMTlane(int Nsimd) {
}); \
});
#define accelerator_barrier(dummy) 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 acceleratorCopyDeviceToDeviceAsynch(void *from,void *to,size_t bytes) {
theGridAccelerator->memcpy(to,from,bytes);
}
inline void acceleratorCopySynchronise(void) { theGridAccelerator->wait(); std::cout<<"acceleratorCopySynchronise() wait "<<std::endl; }
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 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 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();}
inline int acceleratorIsCommunicable(void *ptr)
{
#if 0
@ -514,7 +514,16 @@ inline void *acceleratorAllocCpu(size_t bytes){return memalign(GRID_ALLOC_ALIGN,
inline void acceleratorFreeCpu (void *ptr){free(ptr);};
#endif
//////////////////////////////////////////////
// Fencing needed ONLY for SYCL
//////////////////////////////////////////////
#ifdef GRID_SYCL
inline void acceleratorFenceComputeStream(void){ accelerator_barrier();};
#else
// Ordering within a stream guaranteed on Nvidia & AMD
inline void acceleratorFenceComputeStream(void){ };
#endif
///////////////////////////////////////////////////
// Synchronise across local threads for divergence resynch