portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-10 15:55:37 +00:00
Code Releases Activity

Better extract merge for GPU. Let the SIMD header files define the pointer type for

Browse Source 
access. GPU redirects through builtin float2, double2 for complex
This commit is contained in:
Peter Boyle 2018-07-05 07:05:13 -04:00
parent 4730d4692a
commit 19b527e83f
1 changed files with 12 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

19
lib/tensors/Tensor_extract_merge.h
View File

@ -113,32 +113,37 @@ void merge(vobj &vec,ExtractBuffer<sobj> &extracted)
template<class vobj> accelerator_inline
typename vobj::scalar_object extractLane(int lane, const vobj & __restrict__ vec)
{
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::scalar_object scalar_object;
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
typedef typename vobj::vector_type vector_type;
typedef typename ExtractTypeMap<scalar_type>::extract_type extract_type;
typedef extract_type * pointer;
constexpr int words=sizeof(vobj)/sizeof(vector_type);
constexpr int Nsimd=vector_type::Nsimd();
scalar_object extracted;
scalar_type * __restrict__ sp = (scalar_type *)&extracted; // Type pun
scalar_type * __restrict__ vp = (scalar_type *)&vec;
pointer __restrict__ sp = (pointer)&extracted; // Type pun
pointer __restrict__ vp = (pointer)&vec;
for(int w=0;w<words;w++){
sp[w]=vp[w*Nsimd+lane];
}
return extracted;
}
template<class vobj> accelerator_inline
void insertLane(int lane, vobj & __restrict__ vec,const typename vobj::scalar_object & __restrict__ extracted)
{
typedef typename vobj::scalar_type scalar_type;
typedef typename vobj::vector_type vector_type;
typedef typename vector_type::scalar_type scalar_type;
typedef typename ExtractTypeMap<scalar_type>::extract_type extract_type;
typedef extract_type * pointer;
constexpr int words=sizeof(vobj)/sizeof(vector_type);
constexpr int Nsimd=vector_type::Nsimd();
scalar_type * __restrict__ sp = (scalar_type *)&extracted;
scalar_type * __restrict__ vp = (scalar_type *)&vec;
pointer __restrict__ sp = (pointer)&extracted;
pointer __restrict__ vp = (pointer)&vec;
for(int w=0;w<words;w++){
vp[w*Nsimd+lane]=sp[w];
}