Valentin Churavy

Might be fixed by #2694

``` L37: ; preds = %L11 %19 = addrspacecast {} addrspace(10)* %getfield4 to {} addrspace(11)*, !dbg !1268 %20 = call nonnull "enzyme_type"="{[-1]:Pointer, [-1,0]:Pointer, [-1,0,-1]:Float@double, [-1,8]:Integer, [-1,9]:Integer, [-1,10]:Integer, [-1,11]:Integer, [-1,12]:Integer, [-1,13]:Integer,...

Backtrace is: ``` [79106] signal (6.-6): Aborted in expression starting at /home/vchuravy/src/Enzyme/test/integration/MPI/multi_request.jl:56 unknown function (ip: 0x7f3c4209894c) gsignal at /usr/lib/libc.so.6 (unknown line) abort at /usr/lib/libc.so.6 (unknown line) unknown function (ip: 0x7f3c420254e2)...

I am thinking this is due to: ``` julia> Enzyme.Compiler.guaranteed_const(MPI.API.MPI_Request) true julia> MPI.API.MPI_Request Int32 ``` But also true for OpenMPI where this is a `Ptr{Cvoid}`. @wsmoses is there a way...

There is a related issue with a code that doesn't use `MultiRequest` ``` MPI.Init() comm = MPI.COMM_WORLD function ring_no_mr(token, comm) rank = MPI.Comm_rank(comm) N = MPI.Comm_size(comm) reqs = Vector{MPI.Request}(undef, 2)...

Thankfully: ``` function ring_no_mr_shadow_req(token, comm, reqs) rank = MPI.Comm_rank(comm) N = MPI.Comm_size(comm) buf = Ref(token) if rank != 0 reqs[2] = MPI.Irecv!(buf, comm; source = rank - 1) MPI.Wait(reqs[2]) end...

The failing tests are the same as mentioned by @maleadt here https://github.com/pocl/pocl/issues/1871#issuecomment-2769083558

Looking at the IR and the native code with Paul the Crux is that atomicrmw lowers to a cmpxchg loop for floating point operations. On the GPU we have dedicated...

I believe @cncastillo identified the correct fix, so would either of you be up to opening a PR?

@albertomercurio ideally you first check that the behavior of your ode function is correct. https://docs.sciml.ai/SciMLSensitivity/stable/faq/nothing/SciMLSensitivity/stable/faq/#How-do-I-isolate-potential-gradient-issues-and-improve-performance? Has some tips how to extract that function. > However dp is not updated Why...