Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

ensure sum(f,x) works on GPU #1004

Merged
merged 5 commits into from
Jun 21, 2021
Merged

ensure sum(f,x) works on GPU #1004

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
8f4354f
add tests to see if sum(f,x) is broken on GPU
oxinabox Jun 20, 2021
075f530
add back old way of doing sum(f, xs) for CuArrays only
oxinabox Jun 20, 2021
22602b3
Qualified names
oxinabox Jun 20, 2021
7776fdd
don't test gradient of abs at 0
oxinabox Jun 20, 2021
bd8c5fb
Bump version
oxinabox Jun 20, 2021
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 8 additions & 1 deletion src/lib/broadcast.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -254,7 +254,14 @@ end
placeholder = similar(xs)
sum(xs, dims = dims), Δ -> (placeholder .= Δ,)
end


# Make sure sum(f, ::CuArray) uses broadcase through forward-mode defined above
# Not the ChainRules.rrule which will use the Zygote.Context and thus not be GPU compatible
@adjoint function sum(f, xs::CUDA.CuArray; kws...)
@assert !haskey(kws, :init) # TODO add init support (julia 1.6)
return pullback(__context__, (f, xs) -> sum(f.(xs); kws...), f, xs)
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

You probably don't need the context manager here

Copy link
Member Author

@oxinabox oxinabox Jun 21, 2021 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is exactly the old code.
I don't really know what it is doing returning the context<

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think keeping it is better.
It should end up hitting the same code either way, but I can imagine if some specialization are added then the errors we would get are more confusing if we don't include the __context__

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Might be slightly tidier to call broadcast_forward directly, instead of going back to pullback. But this is a quick fix to repair an accidental break so the most direct restoration is fine too.

end

@adjoint function Base.convert(::Type{T}, xs::Array) where {T<:CUDA.CuArray}
Base.convert(T, xs), Δ -> (nothing, Base.convert(Array, Δ),)
end
Expand Down
11 changes: 11 additions & 0 deletions test/cuda.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,6 +26,17 @@ end
@test g_gpu |> collect ≈ g
end

@testset "sum(f, x)" begin
a = Float32.(-4:4)
oxinabox marked this conversation as resolved.
Show resolved Hide resolved
a_gpu = a |> cu

f(x) = sum(abs, x)
g = gradient(f, a)[1]
g_gpu = gradient(f, a_gpu)[1]
@test g_gpu isa CuArray
@test g_gpu |> collect ≈ g
end

@testset "jacobian" begin
v1 = cu(collect(1:3f0))

Expand Down