The Vamana index build has several different L2 comparator methods based on the dataset dimension and datatype. For large dimensions, each thread loads several values and accumulates the total to increase ILP, but each value is loaded independently. For the __half and float types we can improve performance by using the half2 and float2 types and using the __hfma2, __hsub2, __fmaf2, and __fsub2 operators to calculate and accumulate the L2 distnace.
The Vamana index build has several different L2 comparator methods based on the dataset dimension and datatype. For large dimensions, each thread loads several values and accumulates the total to increase ILP, but each value is loaded independently. For the
__halfandfloattypes we can improve performance by using thehalf2andfloat2types and using the__hfma2,__hsub2,__fmaf2, and__fsub2operators to calculate and accumulate the L2 distnace.