https://github.com/chapel-lang/chapel/pull/24787 will add +, min and max reductions. These are probably the most common reduction kinds, but more importantly, CUB/hipCUB has direct support for those.
minloc and maxloc reductions are also supported in CUB via ArgMin and ArgMax functions. We'll need to wire some more things from the compiler into the runtime. I expect this work to be relatively straightforward implementation in the compiler and the runtime.
For other reduction types, we need engineer a way to use CUB's generic Reduce interface where we pass a function of ours into CUB to handle the reduction. IOW, CUB is supposed to call the accumulate function of the user-defined reduction implementation. Needless to say, the priority should be for those reduction kinds that we already have and not any kind of generic user-defined reductions.
https://github.com/chapel-lang/chapel/pull/24787 will add
+,minandmaxreductions. These are probably the most common reduction kinds, but more importantly, CUB/hipCUB has direct support for those.minlocandmaxlocreductions are also supported in CUB viaArgMinandArgMaxfunctions. We'll need to wire some more things from the compiler into the runtime. I expect this work to be relatively straightforward implementation in the compiler and the runtime.For other reduction types, we need engineer a way to use CUB's generic
Reduceinterface where we pass a function of ours into CUB to handle the reduction. IOW, CUB is supposed to call theaccumulatefunction of the user-defined reduction implementation. Needless to say, the priority should be for those reduction kinds that we already have and not any kind of generic user-defined reductions.