Segmentation fault (earlier TypeInference failure) with ValueShapes.jl #303
Assignees
Labels
Comments
|
Can you post the full trace to Gist or something similar? |
|
And give current main a try |
|
Looks the same with main, here's the gist: https://gist.github.com/oschulz/94c37f22e1c7afe766f223737196d513 |
vchuravy
changed the title
|
|
This now gives a julia-level error, but is the same error needing fixing. |
|
@vchuravy interestingly this now segfaults. |
|
|
Thanks for looking into this! |
wsmoses
changed the title
|
Reducing: using Enzyme
Enzyme.API.printall!(true)
@noinline function myunaliascopy(V::SubArray{T,N,A,I,LD}) where {T,N,A<:Array,I<:Tuple{Vararg{Union{Real,AbstractRange,Array}}},LD}
dest = Array{T}(undef, Base.index_lengths(V.indices...))
Base.copyto!(dest, V)
SubArray{T,N,A,I,LD}(dest, map(Base._trimmedindex, V.indices), 0, Int(LD))
end
function diffeq_step!(u)
x = @inbounds @view u[1:2]
x′ = Base.unaliascopy(x)
@inbounds Base.setindex!(u, x′[1], 1)
nothing
end
u = rand(2)
diffeq_step!(u)
@show "primal"
Enzyme.autodiff(diffeq_step!, u)
@show "grad" |
|
This seems to be fixed by EnzymeAD/Enzyme#959 which will require a jll bump to be available here. |
|
Thanks Will! |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
This crashes Enzyme:
with
shaped_u.xandshaped_du.xjust produce reshaped array views, so there shouldn't be any AD-unfriendly stuff in the code (well, in principle ... :-) ).System info:
The text was updated successfully, but these errors were encountered: