portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-19 00:07:05 +01:00
Code Releases Activity

summit mem

Browse Source
This commit is contained in:
Christoph Lehner
2020-05-18 17:52:26 -04:00
parent 32fbdf4fb1
commit a7635fd5ba
6 changed files with 197 additions and 27 deletions
Download Patch File Download Diff File Expand all files Collapse all files

39
Grid/lattice/Lattice_base.h
View File

@ -97,33 +97,14 @@ public:
else grid = _grid;
};
accelerator_inline void Advise(int advise) {
#ifdef GRID_NVCC
#ifndef __CUDA_ARCH__ // only on host
if (advise & AdviseInfrequentUse) {
cudaMemAdvise(_odata,_odata_size*sizeof(vobj),cudaMemAdviseSetPreferredLocation,cudaCpuDeviceId);
}
if (advise & AdviseReadMostly) {
cudaMemAdvise(_odata,_odata_size*sizeof(vobj),cudaMemAdviseSetReadMostly,-1);
}
#endif
#endif
};
accelerator_inline void AcceleratorPrefetch(int accessMode = ViewReadWrite) { // will use accessMode in future
#ifdef GRID_NVCC
#ifndef __CUDA_ARCH__ // only on host
int target;
cudaGetDevice(&target);
cudaMemPrefetchAsync(_odata,_odata_size*sizeof(vobj),target);
#endif
#endif
gridAcceleratorPrefetch(_odata,_odata_size*sizeof(vobj));
};
accelerator_inline void HostPrefetch(int accessMode = ViewReadWrite) { // will use accessMode in future
#ifdef GRID_NVCC
#ifndef __CUDA_ARCH__ // only on host
cudaMemPrefetchAsync(_odata,_odata_size*sizeof(vobj),cudaCpuDeviceId);
//cudaMemPrefetchAsync(_odata,_odata_size*sizeof(vobj),cudaCpuDeviceId);
#endif
#endif
};
@ -246,13 +227,27 @@ private:
dealloc();
this->_odata_size = size;
if ( size )
if ( size )
this->_odata = alloc.allocate(this->_odata_size);
else
this->_odata = nullptr;
}
}
public:
void Advise(int advise) {
#ifdef GRID_NVCC
#ifndef __CUDA_ARCH__ // only on host
if (advise & AdviseInfrequentUse) {
gridMoveToHost((void**)&this->_odata);
}
if (advise & AdviseReadMostly) {
//cudaMemAdvise(_odata,_odata_size*sizeof(vobj),cudaMemAdviseSetReadMostly,-1);
}
#endif
#endif
};
/////////////////////////////////////////////////////////////////////////////////
// Return a view object that may be dereferenced in site loops.
// The view is trivially copy constructible and may be copied to an accelerator device

2
Grid/lattice/Lattice_transfer.h
View File

@ -96,7 +96,7 @@ accelerator_inline void convertType(ComplexF & out, const std::complex<float> &
out = in;
}
#ifdef __CUDA_ARCH__
#ifdef GRID_NVCC
accelerator_inline void convertType(vComplexF & out, const ComplexF & in) {
((ComplexF*)&out)[SIMTlane(vComplexF::Nsimd())] = in;
}