mirror of
https://github.com/paboyle/Grid.git
synced 2025-04-05 11:45:56 +01:00
Coalesced write
This commit is contained in:
parent
4c3225412b
commit
d6ffadb33b
@ -69,6 +69,11 @@ void coalescedWrite(vobj & __restrict__ vec,const vobj & __restrict__ extracted,
|
||||
// vstream(vec, extracted);
|
||||
vec = extracted;
|
||||
}
|
||||
template<class vobj> accelerator_inline
|
||||
void coalescedWriteNonTemporal(vobj & __restrict__ vec,const vobj & __restrict__ extracted,int lane=0)
|
||||
{
|
||||
vstream(vec, extracted);
|
||||
}
|
||||
#else
|
||||
accelerator_inline int SIMTlane(int Nsimd) { return threadIdx.y; } // CUDA specific
|
||||
|
||||
@ -92,6 +97,11 @@ void coalescedWrite(vobj & __restrict__ vec,const typename vobj::scalar_object &
|
||||
{
|
||||
insertLane(lane,vec,extracted);
|
||||
}
|
||||
template<class vobj> accelerator_inline
|
||||
void coalescedWriteNonTemporal(vobj & __restrict__ vec,const vobj & __restrict__ extracted,int lane=0)
|
||||
{
|
||||
insertLane(lane,vec,extracted);
|
||||
}
|
||||
#endif
|
||||
|
||||
|
||||
|
Loading…
x
Reference in New Issue
Block a user