1
0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-13 12:47:05 +01:00

Pushed the overlap comms tweaks

This commit is contained in:
paboyle
2016-01-11 06:34:22 -08:00
parent dafc74020c
commit fc6ad65751
6 changed files with 96 additions and 18 deletions

View File

@ -217,5 +217,50 @@ void merge(vobj &vec,std::vector<typename vobj::scalar_object *> &extracted,int
}
}
}
template<class vobj> inline
void merge1(vobj &vec,std::vector<typename vobj::scalar_object *> &extracted,int offset)
{
typedef typename vobj::scalar_type scalar_type ;
typedef typename vobj::vector_type vector_type ;
const int Nsimd=vobj::vector_type::Nsimd();
const int words=sizeof(vobj)/sizeof(vector_type);
scalar_type *pointer;
scalar_type *vp = (scalar_type *)&vec;
// assert( (((uint64_t)vp)&(sizeof(scalar_type)-1)) == 0);
for(int i=0;i<Nsimd;i++){
pointer=(scalar_type *)&extracted[i][offset];
for(int w=0;w<words;w++){
vp[w*Nsimd+i] = pointer[w];
}
}
}
template<class vobj> inline
void merge2(vobj &vec,std::vector<typename vobj::scalar_object *> &extracted,int offset)
{
typedef typename vobj::scalar_type scalar_type ;
typedef typename vobj::vector_type vector_type ;
const int Nsimd=vobj::vector_type::Nsimd();
const int words=sizeof(vobj)/sizeof(vector_type);
scalar_type *pointer;
scalar_type *vp = (scalar_type *)&vec;
// assert( (((uint64_t)vp)&(sizeof(scalar_type)-1)) == 0);
for(int w=0;w<words;w++){
for(int i=0;i<Nsimd;i++){
pointer=(scalar_type *)&extracted[i][offset];
vp[w*Nsimd+i] =pointer[w];
}
}
}
}
#endif