portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-17 15:27:06 +01:00
Code Releases Activity

Better SIMD usage/coalescence

Browse Source
This commit is contained in:
Peter Boyle
2021-02-26 17:51:41 +01:00
parent 1ac13ec3a7
commit f9b1f240f6
5 changed files with 143 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

62
Grid/tensors/Tensor_SIMT.h
View File

@ -64,6 +64,68 @@ void coalescedWriteNonTemporal(vobj & __restrict__ vec,const vobj & __restrict__
}
#else
#ifndef GRID_SYCL
// Use the scalar as our own complex on GPU
template<class vsimd,IfSimd<vsimd> = 0> accelerator_inline
typename vsimd::scalar_type
coalescedRead(const vsimd & __restrict__ vec,int lane=acceleratorSIMTlane(vsimd::Nsimd()))
{
typedef typename vsimd::scalar_type S;
S * __restrict__ p=(S *)&vec;
return p[lane];
}
template<int ptype,class vsimd,IfSimd<vsimd> = 0> accelerator_inline
typename vsimd::scalar_type
coalescedReadPermute(const vsimd & __restrict__ vec,int doperm,int lane=acceleratorSIMTlane(vsimd::Nsimd()))
{
typedef typename vsimd::scalar_type S;
S * __restrict__ p=(S *)&vec;
int mask = vsimd::Nsimd() >> (ptype + 1);
int plane= doperm ? lane ^ mask : lane;
return p[plane];
}
template<class vsimd,IfSimd<vsimd> = 0> accelerator_inline
void coalescedWrite(vsimd & __restrict__ vec,
const typename vsimd::scalar_type & __restrict__ extracted,
int lane=acceleratorSIMTlane(vsimd::Nsimd()))
{
typedef typename vsimd::scalar_type S;
S * __restrict__ p=(S *)&vec;
p[lane]=extracted;
}
#else
template<class vsimd,IfSimd<vsimd> = 0> accelerator_inline
typename vsimd::vector_type::datum
coalescedRead(const vsimd & __restrict__ vec,int lane=acceleratorSIMTlane(vsimd::Nsimd()))
{
typedef typename vsimd::vector_type::datum S;
S * __restrict__ p=(S *)&vec;
return p[lane];
}
template<int ptype,class vsimd,IfSimd<vsimd> = 0> accelerator_inline
typename vsimd::vector_type::datum
coalescedReadPermute(const vsimd & __restrict__ vec,int doperm,int lane=acceleratorSIMTlane(vsimd::Nsimd()))
{
typedef typename vsimd::vector_type::datum S;
S * __restrict__ p=(S *)&vec;
int mask = vsimd::Nsimd() >> (ptype + 1);
int plane= doperm ? lane ^ mask : lane;
return p[plane];
}
template<class vsimd,IfSimd<vsimd> = 0> accelerator_inline
void coalescedWrite(vsimd & __restrict__ vec,
const typename vsimd::vector_type::datum & __restrict__ extracted,
int lane=acceleratorSIMTlane(vsimd::Nsimd()))
{
typedef typename vsimd::vector_type::datum S;
S * __restrict__ p=(S *)&vec;
p[lane]=extracted;
}
#endif
//////////////////////////////////////////
// Extract and insert slices on the GPU
//////////////////////////////////////////