Skip to content

Commit

Permalink
feat: use contexts from DifferentiationInterface.jl
Browse files Browse the repository at this point in the history
  • Loading branch information
@avik-pal
avik-pal committed Sep 25, 2024
1 parent c2706b8 commit 153244f
Show file tree
Hide file tree
Showing 3 changed files with 26 additions and 38 deletions.
2 changes: 1 addition & 1 deletion lib/SciMLJacobianOperators/Project.toml
View file
Original file line number Diff line number Diff line change
Expand Up @@ -19,7 +19,7 @@ ADTypes = "1.8.1"
Aqua = "0.8.7"
ConcreteStructs = "0.2.3"
ConstructionBase = "1.5"
DifferentiationInterface = "0.5"
DifferentiationInterface = "0.6"
Enzyme = "0.12, 0.13"
EnzymeCore = "0.7, 0.8"
ExplicitImports = "1.9.0"
Expand Down
56 changes: 22 additions & 34 deletions lib/SciMLJacobianOperators/src/SciMLJacobianOperators.jl
View file
Original file line number Diff line number Diff line change
@@ -1,9 +1,9 @@
module SciMLJacobianOperators

using ADTypes: ADTypes, AutoSparse, AutoEnzyme
using ADTypes: ADTypes, AutoSparse
using ConcreteStructs: @concrete
using ConstructionBase: ConstructionBase
using DifferentiationInterface: DifferentiationInterface
using DifferentiationInterface: DifferentiationInterface, Constant
using EnzymeCore: EnzymeCore
using FastClosures: @closure
using LinearAlgebra: LinearAlgebra
Expand Down Expand Up @@ -112,10 +112,10 @@ function JacobianOperator(prob::AbstractNonlinearProblem, fu, u; jvp_autodiff =
iip = SciMLBase.isinplace(prob)
T = promote_type(eltype(u), eltype(fu))

vjp_autodiff = set_function_as_const(get_dense_ad(vjp_autodiff))
vjp_autodiff = get_dense_ad(vjp_autodiff)
vjp_op = prepare_vjp(skip_vjp, prob, f, u, fu; autodiff = vjp_autodiff)

jvp_autodiff = set_function_as_const(get_dense_ad(jvp_autodiff))
jvp_autodiff = get_dense_ad(jvp_autodiff)
jvp_op = prepare_jvp(skip_jvp, prob, f, u, fu; autodiff = jvp_autodiff)

output_cache = fu isa Number ? T(fu) : similar(fu, T)
Expand Down Expand Up @@ -295,23 +295,21 @@ function prepare_vjp(::Val{false}, prob::AbstractNonlinearProblem,

@assert autodiff!==nothing "`vjp_autodiff` must be provided if `f` doesn't have \
analytic `vjp` or `jac`."
# TODO: Once DI supports const params we can use `p`
fₚ = SciMLBase.JacobianWrapper{SciMLBase.isinplace(f)}(f, prob.p)
if SciMLBase.isinplace(f)
@assert DI.check_twoarg(autodiff) "Backend: $(autodiff) doesn't support in-place \
problems."
@assert DI.check_inplace(autodiff) "Backend: $(autodiff) doesn't support in-place \
problems."
fu_cache = copy(fu)
v_fake = copy(fu)
di_extras = DI.prepare_pullback(fₚ, fu_cache, autodiff, u, v_fake)
di_extras = DI.prepare_pullback(f, fu_cache, autodiff, u, (fu,), Constant(prob.p))
return @closure (vJ, v, u, p) -> begin
DI.pullback!(fₚ, fu_cache, reshape(vJ, size(u)), autodiff,
u, reshape(v, size(fu_cache)), di_extras)
DI.pullback!(f, fu_cache, (reshape(vJ, size(u)),), di_extras, autodiff,
u, (reshape(v, size(fu_cache)),), Constant(p))
return
end
else
di_extras = DI.prepare_pullback(fₚ, autodiff, u, fu)
di_extras = DI.prepare_pullback(f, autodiff, u, (fu,), Constant(prob.p))
return @closure (v, u, p) -> begin
return DI.pullback(fₚ, autodiff, u, reshape(v, size(fu)), di_extras)
return only(DI.pullback(
f, di_extras, autodiff, u, (reshape(v, size(fu)),), Constant(p)))
end
end
end
Expand Down Expand Up @@ -342,23 +340,21 @@ function prepare_jvp(::Val{false}, prob::AbstractNonlinearProblem,

@assert autodiff!==nothing "`jvp_autodiff` must be provided if `f` doesn't have \
analytic `vjp` or `jac`."
# TODO: Once DI supports const params we can use `p`
fₚ = SciMLBase.JacobianWrapper{SciMLBase.isinplace(f)}(f, prob.p)
if SciMLBase.isinplace(f)
@assert DI.check_twoarg(autodiff) "Backend: $(autodiff) doesn't support in-place \
problems."
@assert DI.check_inplace(autodiff) "Backend: $(autodiff) doesn't support in-place \
problems."
fu_cache = copy(fu)
di_extras = DI.prepare_pushforward(fₚ, fu_cache, autodiff, u, u)
di_extras = DI.prepare_pushforward(f, fu_cache, autodiff, u, (u,), Constant(prob.p))
return @closure (Jv, v, u, p) -> begin
DI.pushforward!(
fₚ, fu_cache, reshape(Jv, size(fu_cache)),
autodiff, u, reshape(v, size(u)), di_extras)
DI.pushforward!(f, fu_cache, (reshape(Jv, size(fu_cache)),), di_extras,
autodiff, u, (reshape(v, size(u)),), Constant(p))
return
end
else
di_extras = DI.prepare_pushforward(fₚ, autodiff, u, u)
di_extras = DI.prepare_pushforward(f, autodiff, u, (u,), Constant(prob.p))
return @closure (v, u, p) -> begin
return DI.pushforward(fₚ, autodiff, u, reshape(v, size(u)), di_extras)
return only(DI.pushforward(
f, di_extras, autodiff, u, (reshape(v, size(u)),), Constant(p)))
end
end
end
Expand All @@ -371,10 +367,8 @@ function prepare_scalar_op(::Val{false}, prob::AbstractNonlinearProblem,

@assert autodiff!==nothing "`autodiff` must be provided if `f` doesn't have \
analytic `vjp` or `jvp` or `jac`."
# TODO: Once DI supports const params we can use `p`
fₚ = Base.Fix2(f, prob.p)
di_extras = DI.prepare_derivative(fₚ, autodiff, u)
return @closure (v, u, p) -> DI.derivative(fₚ, autodiff, u, di_extras) * v
di_extras = DI.prepare_derivative(f, autodiff, u, Constant(prob.p))
return @closure (v, u, p) -> DI.derivative(f, di_extras, autodiff, u, Constant(p)) * v
end

get_dense_ad(::Nothing) = nothing
Expand All @@ -386,12 +380,6 @@ function get_dense_ad(ad::AutoSparse)
return dense_ad
end

# In our case we know that it is safe to mark the function as const
set_function_as_const(ad) = ad
function set_function_as_const(ad::AutoEnzyme{M, Nothing}) where {M}
return AutoEnzyme(; ad.mode, function_annotation = EnzymeCore.Const)
end

export JacobianOperator, VecJacOperator, JacVecOperator
export StatefulJacobianOperator
export StatefulJacobianNormalFormOperator
Expand Down
6 changes: 3 additions & 3 deletions lib/SciMLJacobianOperators/test/core_tests.jl
View file
Original file line number Diff line number Diff line change
Expand Up @@ -7,7 +7,7 @@
AutoEnzyme(),
AutoEnzyme(; mode = Enzyme.Reverse),
AutoZygote(),
AutoReverseDiff(),
# AutoReverseDiff(), # FIXME: https://github.com/gdalle/DifferentiationInterface.jl/issues/503
AutoTracker(),
AutoFiniteDiff()
]
Expand Down Expand Up @@ -91,7 +91,7 @@ end
reverse_ADs = [
AutoEnzyme(),
AutoEnzyme(; mode = Enzyme.Reverse),
AutoReverseDiff(),
# AutoReverseDiff(), # FIXME: https://github.com/gdalle/DifferentiationInterface.jl/issues/503
AutoFiniteDiff()
]

Expand Down Expand Up @@ -182,7 +182,7 @@ end
AutoEnzyme(; mode = Enzyme.Reverse),
AutoZygote(),
AutoTracker(),
AutoReverseDiff(),
# AutoReverseDiff(), # FIXME: https://github.com/gdalle/DifferentiationInterface.jl/issues/503
AutoFiniteDiff()
]

Expand Down

2 comments on commit 153244f

@avik-pal
Copy link
Member Author

Choose a reason for hiding this comment

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

@JuliaRegistrator register subdir=lib/SciMLJacobianOperators

@JuliaRegistrator
Copy link

Choose a reason for hiding this comment

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

Registration pull request updated: JuliaRegistries/General/115880

Tip: Release Notes

Did you know you can add release notes too? Just add markdown formatted text underneath the comment after the text
"Release notes:" and it will be added to the registry PR, and if TagBot is installed it will also be added to the
release that TagBot creates. i.e.

@JuliaRegistrator register

Release notes:

## Breaking changes

- blah

To add them here just re-invoke and the PR will be updated.

Tagging

After the above pull request is merged, it is recommended that a tag is created on this repository for the registered package version.

This will be done automatically if the Julia TagBot GitHub Action is installed, or can be done manually through the github interface, or via:

git tag -a SciMLJacobianOperators-v0.1.0 -m "<description of version>" 153244f5ebac15f17572aeeedb8e181f439ee48a
git push origin SciMLJacobianOperators-v0.1.0

Please sign in to comment.