Update EnzymeCoreExt.jl

[simenhu]

The inactive_noinl failed in broadcast assignment when using Enzyme=0.13.16 and CUDA=5.5.2. Switching to the version that prevents inlining seem to have fixed it. Are there any negative consequences of this?

The issue where the error can be reproduced can be found here: EnzymeAD/Enzyme.jl#2116

[wsmoses]

Can you add your full code as a test?

[ptiede]

This code should test the change

@testset "Reverse sum abs2" begin
    x = CuArray([1.0, 2.0, 3.0, 4.0])
    dx = CuArray([0., 0.0, 0.0, 0.0])
    f(x) = sum(abs2.(x))
    Enzyme.autodiff(Reverse, f, Active, Duplicated(x, dx))
    @test all(dx .≈ 2.*x)
end

[simenhu]

Can you add your full code as a test?

Do you still want me to add the code as an test? Assume that you want me to remove the dependence on OrdinaryDiffEq?

[wsmoses]

yeah I don't think we should add a dependency here, but if you can make a cuda exclusive test, that would be good to keep

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 73.41%. Comparing base (7ff012f) to head (e5e93ba).
Report is 6 commits behind head on master.

Additional details and impacted files

@@             Coverage Diff             @@
##           master    #2565       +/-   ##
===========================================
+ Coverage    9.41%   73.41%   +64.00%     
===========================================
  Files         156      156               
  Lines       14808    14997      +189     
===========================================
+ Hits         1394    11010     +9616     
+ Misses      13414     3987     -9427     

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