Example from #220:
val e = depFun((n: Nat) => fun(n`.`f32)(in =>
in |> split(32) |> mapWorkGroup(fun(x =>
zip(x)(x) |> mapLocal(fun(y =>
makePair(fst(y) |> toLocal)(snd(y) |> toGlobal)
)) |> gather(reverse) |>
mapLocal(fun(p => fst(p) + snd(p)))
))
))
In many cases we have been able to rewrite such Lift programs differently to make them work in Rise with access annotations. However, in this case it does not seem possible.
What should we do about it?
Should we make toMem more flexible, maybe by accepting a memory layout: toMem: {t: DataType} -> (a: AddrSpaceLayout[t]) -> exp[t, wr] -> exp[t, rd]?
Should patterns like mapSeq/mapLocal be allowed to be read -> read instead of read -> write? Would that require changing how memory is allocated? (i.e hoisting allocations outside of loops)
cc @bastian-koepcke
Example from #220:
In many cases we have been able to rewrite such Lift programs differently to make them work in Rise with access annotations. However, in this case it does not seem possible.
What should we do about it?
Should we make
toMemmore flexible, maybe by accepting a memory layout:toMem: {t: DataType} -> (a: AddrSpaceLayout[t]) -> exp[t, wr] -> exp[t, rd]?Should patterns like
mapSeq/mapLocalbe allowed to beread -> readinstead ofread -> write? Would that require changing how memory is allocated? (i.e hoisting allocations outside of loops)cc @bastian-koepcke