Is it possible to build reductions over individual vector elements with FLoops.jl (or any other package? I only know of FLoops for now)? I'm trying to get a solution to a question I asked some time ago, https://discourse.julialang.org/t/problem-converting-serial-code-to-parallel-code-with-floops/87610, and this seems to be the issue with it. The actual example I'm interested in is for parallelising some finite volume code https://github.com/DanielVandH/FiniteVolumeMethod.jl/blob/main/src/fvm.jl#L793-L902.
Is it possible to build reductions over individual vector elements
Can you give a minimal example of what this means?
Well, actually I'm not 100% sure what I mean there other than that I wanted to use FLoops.@reduce
for this, applying it to expressions like du[j] += 100
so that @reduce du[j] += 100
works.
In the linked discourse post, I have the code:
using FLoops, Random
function update_u!(du, u, k, q, j)
v = [k, (k + 1) % length(u) + 1, (k + 7) % length(u) + 1]
q .= [u[v[1]], u[v[3]]]
du[v[j]] += q' * q
du[v[(j%3)+1]] -= sin.(q)' * sin.(q)
nothing
end
function do_parallel(du::AbstractVector{T}, u, elements) where {T}
q = zeros(2, length(elements), 3)
@floop for k ∈ elements
for j in 1:3
@views update_u!(du, u, k, q[:, k, j], j)
end
end
nothing
end
This somewhat captures what I want. @reduce
would (I believe) avoid the data-race issues with updating du
, but it doesn't seem to work on expressions like that.
The floop macro is just a wrapper around the functional primitives, so if it's not working you should still be able to write it with Folds or Transducers
I'll look into those packages. I had previously considered those but couldn't follow the documentation. Thanks
Last updated: Nov 06 2024 at 04:40 UTC