portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-17 23:37:06 +01:00
Code Releases Activity

Intel DPCPP Gold happy now (compiles all, runs Benchmark_dwf_fp32 )

Browse Source
This commit is contained in:
Peter Boyle
2020-12-03 03:47:11 -08:00
parent 497e7c1c40
commit cf76741ec6
9 changed files with 97 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
Grid/DisableWarnings.h
View File

@ -37,7 +37,9 @@ directory
#endif
//disables and intel compiler specific warning (in json.hpp)
#ifdef __ICC
#pragma warning disable 488
#endif
#ifdef __NVCC__
//disables nvcc specific warning in json.hpp

12
Grid/Makefile.am
View File

@ -21,6 +21,7 @@ if BUILD_HDF5
extra_headers+=serialisation/Hdf5Type.h
endif
all: version-cache Version.h
version-cache:
@ -53,6 +54,17 @@ Version.h: version-cache
include Make.inc
include Eigen.inc
extra_sources+=$(ZWILS_FERMION_FILES)
extra_sources+=$(WILS_FERMION_FILES)
extra_sources+=$(STAG_FERMION_FILES)
if BUILD_GPARITY
extra_sources+=$(GP_FERMION_FILES)
endif
if BUILD_FERMION_REPS
extra_sources+=$(ADJ_FERMION_FILES)
extra_sources+=$(TWOIND_FERMION_FILES)
endif
lib_LIBRARIES = libGrid.a
CCFILES += $(extra_sources)

2
Grid/communicator/SharedMemory.h
View File

@ -102,7 +102,7 @@ public:
///////////////////////////////////////////////////
static void SharedMemoryAllocate(uint64_t bytes, int flags);
static void SharedMemoryFree(void);
static void SharedMemoryCopy(void *dest,const void *src,size_t bytes);
static void SharedMemoryCopy(void *dest,void *src,size_t bytes);
static void SharedMemoryZero(void *dest,size_t bytes);
};

2
Grid/communicator/SharedMemoryMPI.cc
View File

@ -715,7 +715,7 @@ void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes)
bzero(dest,bytes);
#endif
}
void GlobalSharedMemory::SharedMemoryCopy(void *dest,const void *src,size_t bytes)
void GlobalSharedMemory::SharedMemoryCopy(void *dest,void *src,size_t bytes)
{
#ifdef GRID_CUDA
cudaMemcpy(dest,src,bytes,cudaMemcpyDefault);

43
Grid/communicator/SharedMemoryNone.cc
View File

@ -29,6 +29,7 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#include <Grid/GridCore.h>
NAMESPACE_BEGIN(Grid);
#define header "SharedMemoryNone: "
/*Construct from an MPI communicator*/
void GlobalSharedMemory::Init(Grid_MPI_Comm comm)
@ -55,6 +56,38 @@ void GlobalSharedMemory::OptimalCommunicator(const Coordinate &processors,Grid_M
////////////////////////////////////////////////////////////////////////////////////////////
// Hugetlbfs mapping intended, use anonymous mmap
////////////////////////////////////////////////////////////////////////////////////////////
#if 1
void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
{
std::cout << header "SharedMemoryAllocate "<< bytes<< " GPU implementation "<<std::endl;
void * ShmCommBuf ;
assert(_ShmSetup==1);
assert(_ShmAlloc==0);
///////////////////////////////////////////////////////////////////////////////////////////////////////////
// Each MPI rank should allocate our own buffer
///////////////////////////////////////////////////////////////////////////////////////////////////////////
ShmCommBuf = acceleratorAllocDevice(bytes);
if (ShmCommBuf == (void *)NULL ) {
std::cerr << " SharedMemoryNone.cc acceleratorAllocDevice failed NULL pointer for " << bytes<<" bytes " << std::endl;
exit(EXIT_FAILURE);
}
if ( WorldRank == 0 ){
std::cout << WorldRank << header " SharedMemoryNone.cc acceleratorAllocDevice "<< bytes
<< "bytes at "<< std::hex<< ShmCommBuf <<std::dec<<" for comms buffers " <<std::endl;
}
SharedMemoryZero(ShmCommBuf,bytes);
///////////////////////////////////////////////////////////////////////////////////////////////////////////
// Loop over ranks/gpu's on our node
///////////////////////////////////////////////////////////////////////////////////////////////////////////
WorldShmCommBufs[0] = ShmCommBuf;
_ShmAllocBytes=bytes;
_ShmAlloc=1;
}
#else
void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
{
void * ShmCommBuf ;
@ -83,7 +116,15 @@ void GlobalSharedMemory::SharedMemoryAllocate(uint64_t bytes, int flags)
_ShmAllocBytes=bytes;
_ShmAlloc=1;
};
#endif
void GlobalSharedMemory::SharedMemoryZero(void *dest,size_t bytes)
{
acceleratorMemSet(dest,0,bytes);
}
void GlobalSharedMemory::SharedMemoryCopy(void *dest,void *src,size_t bytes)
{
acceleratorCopyToDevice(src,dest,bytes);
}
////////////////////////////////////////////////////////
// Global shared functionality finished
// Now move to per communicator functionality

5
Grid/lattice/Lattice_basis.h
View File

@ -62,7 +62,7 @@ void basisRotate(VField &basis,Matrix& Qt,int j0, int j1, int k0,int k1,int Nm)
basis_v.push_back(basis[k].View(AcceleratorWrite));
}
#if ( (!defined(GRID_SYCL)) && (!defined(GRID_CUDA)) )
#if ( (!defined(GRID_CUDA)) )
int max_threads = thread_max();
Vector < vobj > Bt(Nm * max_threads);
thread_region
@ -164,7 +164,8 @@ void basisRotateJ(Field &result,std::vector<Field> &basis,Eigen::MatrixXd& Qt,in
auto basis_vp=& basis_v[0];
autoView(result_v,result,AcceleratorWrite);
accelerator_for(ss, grid->oSites(),vobj::Nsimd(),{
auto B=coalescedRead(zz);
vobj zzz=Zero();
auto B=coalescedRead(zzz);
for(int k=k0; k<k1; ++k){
B +=Qt_j[k] * coalescedRead(basis_vp[k][ss]);
}