[vchuravy]

This adds a package extension for MPI.jl, so that FixedSizeArrays, and SubArrays of them, can be used directly as communication buffers (MPI.Send, MPI.Recv!, MPI.Bcast!, MPI.Gatherv!, MPI.Allreduce!, ...).

Motivation

Trixi.jl is replacing unsafe_wrap(Array, pointer(v), dims) views of resize!able vectors with FixedSizeArrays.wrap(FixedSizeArray, v, dims) (#196). Some of those arrays are passed to MPI. With the released MPI.jl (v0.20.27):

• MPI.Buffer has methods only for Array, Ref and some SubArrays, so e.g. MPI.Gatherv!(vec(a[1, .., :]), ...) or MPI.Send(a, ...) with a FixedSizeArray throws a MethodError.
• Base.cconvert(MPI.MPIPtr, x) is defined only for Ptr, Array, Ref and SubArray. A non-fast SubArray (e.g. view(a, 1:1, 2:3, :)) is sent with a subarray datatype, and MPI.Buffer stores its parent, which then also fails for a FixedSizeArray parent.

What the extension defines

All methods involve FixedSizeArray, so there is no type piracy.

• Base.cconvert(::Type{MPIPtr}, a) and Base.unsafe_convert(::Type{MPIPtr}, a) for FixedSizeArrays backed by host memory (Vector, and Memory on Julia v1.11+). cconvert returns the FixedSizeArray itself, which ccall keeps rooted and which keeps its parent alive. Delegating to the parent does not work here, since MPI.jl v0.20.27 has no MPIPtr conversion for Memory, the default storage type.
• For any other storage type (e.g. GPU arrays), cconvert delegates to the parent, so MPI.jl's own support for that array type is used.
• MPI.Buffer(a::FixedSizeArray), sending length(a) elements of MPI.Datatype(eltype(a)).

UBuffer, VBuffer and RBuffer already accept any AbstractArray with unit stride, so they need nothing extra.

Relation to MPI.jl upstream

MPI.jl master (unreleased, JuliaParallel/MPI.jl#941) makes the MPIPtr conversion generic for AbstractArrays, but MPI.Buffer is still Array-only there. I'm opening a draft PR to MPI.jl to accept any DenseArray in MPI.Buffer (link to follow). This extension is still needed for all released MPI v0.20.x, and its methods are strictly more specific than the upstream ones, so it stays correct afterwards. The full test suite also passes against MPI.jl master.

Tests

A new "MPI" testset runs single-process (MPI.COMM_WORLD, no mpiexec), for Memory and Vector storage, element types Float64, Int, ComplexF64 and a padded isbits struct, and 1-, 2- and 3-dimensional arrays. It covers Buffer and pointer conversion, point-to-point to self, Bcast!, Gatherv!, Allgather!, Scatterv!, Allreduce!/Allreduce/Reduce! (out of place and in place), fast and non-fast strided views (send, receive into, scatter into), and delegation for a non-host DenseVector parent. Reductions are skipped for the struct type, since user-defined reduction operators are not supported on all platforms.

This adds MPI.jl (and MPICH_jll, ~30 MB) as a test dependency; the testset takes about 20 s. Happy to trim it if that's too heavy.

Locally, Pkg.test() (including Aqua) passes on Julia 1.10, 1.11, 1.12 and 1.13-rc1 with MPI.jl v0.20.27, and on Julia 1.12 with MPI.jl master. I also ran a separate 2-rank mpiexec script (Send/Recv!, non-fast views, Gatherv!, Scatterv!, Allreduce, Reduce!, Bcast! on 4-dimensional arrays) with MPI.jl v0.20.0, v0.20.27 and master; it is not part of this PR. I did not test Windows (MicrosoftMPI) or 32-bit Linux locally, and GPU-backed FixedSizeArrays are untested.

This PR was prepared with substantial assistance from Claude Code (Anthropic's AI coding assistant). I reviewed the changes.

🤖 Generated with Claude Code

https://claude.ai/code/session_01MEJYUviAUshj5fZt8LTe8A

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 93.47%. Comparing base (b038629) to head (6b71071).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #198      +/-   ##
==========================================
+ Coverage   93.39%   93.47%   +0.08%     
==========================================
  Files           5        6       +1     
  Lines         227      230       +3     
==========================================
+ Hits          212      215       +3     
  Misses         15       15              

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[giordano]

And we could also drop this with JuliaParallel/MPI.jl#996? But ok, this shouldn't hold this PR, we can remove it afterwards.

[vchuravy]

Yeah it will still be correct after that lands.