EC2 Default User 10 years ago
parent
commit
b23709173f
  1. 16
      scan/warp/scanWarpShuffle.cuh

16
scan/warp/scanWarpShuffle.cuh

@ -74,6 +74,14 @@ exclusive_scan_warp_shfl(int mysum) @@ -74,6 +74,14 @@ exclusive_scan_warp_shfl(int mysum)
return (lane) ? mysum : 0;
}
template <int levels>
__device__ __forceinline__
int
inclusive_scan_warp_shfl(int mysum)
{
return mysum + exclusive_scan_warp_shfl<levels>(mysum);
}
template <int logBlockSize>
__device__
int
@ -146,4 +154,12 @@ exclusive_scan_block(int val, const unsigned int idx) @@ -146,4 +154,12 @@ exclusive_scan_block(int val, const unsigned int idx)
return warpid ? sPartials[warpid-1] : 0;
}
template <int logBlockSize>
__device__
int
inclusive_scan_block(int val, const unsigned int idx)
{
return val + exclusive_scan_block<logBlockSize>(val,idx);
}
#endif

Loading…
Cancel
Save