Merge pull request #626 from JuliaDiff/ox/mutabletangent

Introduce MutableTangent

Project.toml

@@ -7,20 +7,17 @@ Compat = "34da2185-b29b-5c13-b0c7-acf172513d20"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 SparseArrays = "2f01184e-e22b-5df5-ae63-d93ebab69eaf"
 
-[weakdeps]
-SparseArrays = "2f01184e-e22b-5df5-ae63-d93ebab69eaf"
-
-[extensions]
-ChainRulesCoreSparseArraysExt = "SparseArrays"
-
 [compat]
 BenchmarkTools = "0.5"
-Compat = "2, 3, 4"
+Compat = "3.40, 4"
 FiniteDifferences = "0.10"
 OffsetArrays = "1"
 StaticArrays = "0.11, 0.12, 1"
 julia = "1.6"
 
+[extensions]
+ChainRulesCoreSparseArraysExt = "SparseArrays"
+
 [extras]
 BenchmarkTools = "6e4b80f9-dd63-53aa-95a3-0cdb28fa8baf"
 FiniteDifferences = "26cc04aa-876d-5657-8c51-4c34ba976000"
@@ -31,3 +28,6 @@ Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 
 [targets]
 test = ["Test", "BenchmarkTools", "FiniteDifferences", "OffsetArrays", "SparseArrays", "StaticArrays"]
+
+[weakdeps]
+SparseArrays = "2f01184e-e22b-5df5-ae63-d93ebab69eaf"

docs/make.jl

@@ -61,6 +61,7 @@ makedocs(;
                 "`@opt_out`" => "rule_author/superpowers/opt_out.md",
                 "`RuleConfig`" => "rule_author/superpowers/ruleconfig.md",
                 "Gradient accumulation" => "rule_author/superpowers/gradient_accumulation.md",
+                "Mutation Support (experimental)" => "rule_author/superpowers/mutation_support.md",
             ],
             "Converting ZygoteRules.@adjoint to rrules" => "rule_author/converting_zygoterules.md",
             "Tips for making your package work with AD" => "rule_author/tips_for_packages.md",

docs/src/api.md

@@ -20,7 +20,7 @@ Modules = [ChainRulesCore]
 Pages = [
     "tangent_types/abstract_zero.jl",
     "tangent_types/one.jl",
-    "tangent_types/tangent.jl",
+    "tangent_types/structural_tangent.jl",
     "tangent_types/thunks.jl",
     "tangent_types/abstract_tangent.jl",
     "tangent_types/notimplemented.jl",

docs/src/rule_author/superpowers/mutation_support.md

@@ -0,0 +1,82 @@
+# Mutation Support
+
+ChainRulesCore.jl offers experimental support for mutation, targeting use in forward mode AD.
+(Mutation support in reverse mode AD is more complicated and will likely require more changes to the interface)
+
+!!! warning "Experimental"
+    This page documents an experimental feature.
+    Expect breaking changes in minor versions while this remains.
+    It is not suitable for general use unless you are prepared to modify how you are using it each minor release.
+    It is thus suggested that if you are using it to use _tilde_ bounds on supported minor versions.
+
+
+## `MutableTangent`
+The [`MutableTangent`](@ref) type is designed to be a partner to the [`Tangent`](@ref) type, with specific support for being mutated in place.
+It is required to be a structural tangent, having one tangent for each field of the primal object.
+
+Technically, not all `mutable struct`s need to use `MutableTangent` to represent their tangents.
+Just like not all `struct`s need to use `Tangent`s.
+Common examples away from this are natural tangent types like for arrays.
+However, if one is setting up to use a custom tangent type for this it is sufficiently off the beaten path that we can not provide much guidance.
+
+## `zero_tangent`
+
+The [`zero_tangent`](@ref) function functions to give you a zero (i.e. additive identity) for any primal value.
+The [`ZeroTangent`](@ref) type also does this.
+The difference is that [`zero_tangent`](@ref) is in general full structural tangent mirroring the structure of the primal.
+To be technical the promise of [`zero_tangent`](@ref) is that it will be a value that supports mutation.
+However, in practice[^1] this is achieved through in a structural tangent
+For mutation support this is important, since it means that there is mutable memory available in the tangent to be mutated when the primal changes.
+To support this you thus need to make sure your zeros are created in various places with [`zero_tangent`](@ref) rather than []`ZeroTangent`](@ref).
+
+
+
+It is also useful for reasons of type stability, since it forces a consistent type (generally a structural tangent) for any given primal type.
+For this reason AD system implementors might chose to use this to create the tangent for all literal values they encounter, mutable or not,
+and to process the output of `frule`s to convert [`ZeroTangent`](@ref) into corresponding [`zero_tangent`](@ref)s.
+
+## Writing a frule for a mutating function
+It is relatively straight forward to write a frule for a mutating function.
+There are a few key points to follow:
+ - There must be a mutable tangent input for every mutated primal input
+ - When the primal value is changed, the corresponding change must be made to its tangent partner
+ - When a value is returned, return its partnered tangent.
+ - If (and only if) primal values alias, then their tangents must also alias.
+
+### Example
+For example, consider the primal function with:
+1. takes two `Ref`s
+2. doubles the first one in place
+3. overwrites the second one's value with the literal 5.0
+4. returns the first one
+
+
+```julia
+function foo!(a::Base.RefValue, b::Base.RefValue)
+    a[] *= 2
+    b[] = 5.0
+    return a
+end
+```
+
+The frule for this would be:
+```julia
+function ChainRulesCore.frule((_, ȧ, ḃ), ::typeof(foo!), a::Base.RefValue, b::Base.RefValue)
+    @assert ȧ isa MutableTangent{typeof(a)}
+    @assert ḃ isa MutableTangent{typeof(b)}
+
+    a[] *= 2
+    ȧ.x *= 2  # `.x` is the field that lives behind RefValues
+
+    b[] = 5.0
+    ḃ.x = zero_tangent(5.0)  # or since we know that the zero for a Float64 is zero could write `ḃ.x = 0.0`
+
+    return a, ȧ
+end
+```
+
+Then assuming the AD system does its part to makes sure you are indeed given mutable values to mutate (i.e. those `@assert`ions are true) then all is well and this rule will make mutation correct.
+
+[^1]:
+    Further, it is hard to achieve this promise of allowing mutation to be supported without returning a structural tangent.
+    Except in the special case of where the struct is not mutable and has no nested fields that are mutable.

src/ChainRulesCore.jl

@@ -2,26 +2,26 @@ module ChainRulesCore
 using Base.Broadcast: broadcasted, Broadcasted, broadcastable, materialize, materialize!
 using Base.Meta
 using LinearAlgebra
-using Compat: hasfield, hasproperty
+using Compat: hasfield, hasproperty, ismutabletype
 
 export frule, rrule  # core function
 # rule configurations
 export RuleConfig, HasReverseMode, NoReverseMode, HasForwardsMode, NoForwardsMode
 export frule_via_ad, rrule_via_ad
 # definition helper macros
 export @non_differentiable, @opt_out, @scalar_rule, @thunk, @not_implemented
-export ProjectTo, canonicalize, unthunk  # tangent operations
+export ProjectTo, canonicalize, unthunk, zero_tangent  # tangent operations
 export add!!, is_inplaceable_destination  # gradient accumulation operations
 export ignore_derivatives, @ignore_derivatives
 # tangents
-export Tangent, NoTangent, InplaceableThunk, Thunk, ZeroTangent, AbstractZero, AbstractThunk
+export StructuralTangent, Tangent, MutableTangent, NoTangent, InplaceableThunk, Thunk, ZeroTangent, AbstractZero, AbstractThunk
 
 include("debug_mode.jl")
 
 include("tangent_types/abstract_tangent.jl")
+include("tangent_types/structural_tangent.jl")
 include("tangent_types/abstract_zero.jl")
 include("tangent_types/thunks.jl")
-include("tangent_types/tangent.jl")
 include("tangent_types/notimplemented.jl")
 
 include("tangent_arithmetic.jl")

src/tangent_arithmetic.jl

@@ -20,7 +20,7 @@ Base.:+(x::NotImplemented, ::NotImplemented) = x
 Base.:*(x::NotImplemented, ::NotImplemented) = x
 LinearAlgebra.dot(x::NotImplemented, ::NotImplemented) = x
 # `NotImplemented` always "wins" +
-for T in (:ZeroTangent, :NoTangent, :AbstractThunk, :Tangent, :Any)
+for T in (:ZeroTangent, :NoTangent, :AbstractThunk, :StructuralTangent, :Any)
     @eval Base.:+(x::NotImplemented, ::$T) = x
     @eval Base.:+(::$T, x::NotImplemented) = x
 end
@@ -33,7 +33,7 @@ for T in (:ZeroTangent, :NoTangent)
     @eval LinearAlgebra.dot(::$T, ::NotImplemented) = $T()
 end
 # `NotImplemented` "wins" * and dot for other types
-for T in (:AbstractThunk, :Tangent, :Any)
+for T in (:AbstractThunk, :StructuralTangent, :Any)
     @eval Base.:*(x::NotImplemented, ::$T) = x
     @eval Base.:*(::$T, x::NotImplemented) = x
     @eval LinearAlgebra.dot(x::NotImplemented, ::$T) = x
@@ -55,7 +55,7 @@ Base.:-(::NoTangent, ::NoTangent) = NoTangent()
 Base.:-(::NoTangent) = NoTangent()
 Base.:*(::NoTangent, ::NoTangent) = NoTangent()
 LinearAlgebra.dot(::NoTangent, ::NoTangent) = NoTangent()
-for T in (:AbstractThunk, :Tangent, :Any)
+for T in (:AbstractThunk, :StructuralTangent, :Any)
     @eval Base.:+(::NoTangent, b::$T) = b
     @eval Base.:+(a::$T, ::NoTangent) = a
     @eval Base.:-(::NoTangent, b::$T) = -b
@@ -95,7 +95,7 @@ Base.:-(::ZeroTangent, ::ZeroTangent) = ZeroTangent()
 Base.:-(::ZeroTangent) = ZeroTangent()
 Base.:*(::ZeroTangent, ::ZeroTangent) = ZeroTangent()
 LinearAlgebra.dot(::ZeroTangent, ::ZeroTangent) = ZeroTangent()
-for T in (:AbstractThunk, :Tangent, :Any)
+for T in (:AbstractThunk, :StructuralTangent, :Any)
     @eval Base.:+(::ZeroTangent, b::$T) = b
     @eval Base.:+(a::$T, ::ZeroTangent) = a
     @eval Base.:-(::ZeroTangent, b::$T) = -b
@@ -126,11 +126,11 @@ for T in (:Tangent, :Any)
     @eval Base.:*(a::$T, b::AbstractThunk) = a * unthunk(b)
 end
 
-function Base.:+(a::Tangent{P}, b::Tangent{P}) where {P}
+function Base.:+(a::StructuralTangent{P}, b::StructuralTangent{P}) where {P}
     data = elementwise_add(backing(a), backing(b))
-    return Tangent{P,typeof(data)}(data)
+    return StructuralTangent{P}(data)
 end
-function Base.:+(a::P, d::Tangent{P}) where {P}
+function Base.:+(a::P, d::StructuralTangent{P}) where {P}
     net_backing = elementwise_add(backing(a), backing(d))
     if debug_mode()
         try
@@ -143,14 +143,14 @@ function Base.:+(a::P, d::Tangent{P}) where {P}
     end
 end
 Base.:+(a::Dict, d::Tangent{P}) where {P} = merge(+, a, backing(d))
-Base.:+(a::Tangent{P}, b::P) where {P} = b + a
+Base.:+(a::StructuralTangent{P}, b::P) where {P} = b + a
 
-Base.:-(tangent::Tangent{P}) where {P} = map(-, tangent)
+Base.:-(tangent::StructuralTangent{P}) where {P} = map(-, tangent)
 
 # We intentionally do not define, `Base.*(::Tangent, ::Tangent)` as that is not meaningful
 # In general one doesn't have to represent multiplications of 2 tangents
 # Only of a tangent and a scaling factor (generally `Real`)
 for T in (:Number,)
-    @eval Base.:*(s::$T, tangent::Tangent) = map(x -> s * x, tangent)
-    @eval Base.:*(tangent::Tangent, s::$T) = map(x -> x * s, tangent)
+    @eval Base.:*(s::$T, tangent::StructuralTangent) = map(x -> s * x, tangent)
+    @eval Base.:*(tangent::StructuralTangent, s::$T) = map(x -> x * s, tangent)
 end

src/tangent_types/abstract_zero.jl

@@ -91,3 +91,90 @@ arguments.
 ```
 """
 struct NoTangent <: AbstractZero end
+
+"""
+    zero_tangent(primal)
+
+This returns an appropriate zero tangent suitable for accumulating tangents of the primal.
+For mutable composites types this is a structural [`MutableTangent`](@ref)
+For `Array`s, it is applied recursively for each element.
+For other types, in particular immutable types, we do not make promises beyond that it will be `iszero`
+and suitable for accumulating against.
+For types without a tangent space (e.g. singleton structs) this returns `NoTangent()`.
+In general, it is more likely to produce a structural tangent.
+
+!!! warning Exprimental
+    `zero_tangent`is an experimental feature, and is part of the mutation support featureset.
+    While this notice remains it may have changes in behavour, and interface in any _minor_ version of ChainRulesCore.
+    Exactly how it should be used (e.g. is it forward-mode only?)
+"""
+function zero_tangent end
+
+zero_tangent(x::Number) = zero(x)
+
+zero_tangent(::Type) = NoTangent()
+
+function zero_tangent(x::MutableTangent{P}) where {P}
+    zb = backing(zero_tangent(backing(x)))
+    return MutableTangent{P}(zb)
+end
+
+function zero_tangent(x::Tangent{P}) where {P}
+    zb = backing(zero_tangent(backing(x)))
+    return Tangent{P,typeof(zb)}(zb)
+end
+
+@generated function zero_tangent(primal)
+    fieldcount(primal) == 0 && return NoTangent()  # no tangent space at all, no need for structural zero.
+    zfield_exprs = map(fieldnames(primal)) do fname
+        fval = :(
+            if isdefined(primal, $(QuoteNode(fname)))
+                zero_tangent(getfield(primal, $(QuoteNode(fname))))
+            else
+                # This is going to be potentially bad, but that's what they get for not giving us a primal
+                # This will never me mutated inplace, rather it will alway be replaced with an actual value first
+                ZeroTangent()
+            end
+        )
+        Expr(:kw, fname, fval)
+    end
+    return if has_mutable_tangent(primal)
+        any_mask = map(fieldnames(primal), fieldtypes(primal)) do fname, ftype
+            # If it is is unassigned, or if it doesn't have a concrete type, let it take any value for its tangent
+            fdef = :(!isdefined(primal, $(QuoteNode(fname))) || !isconcretetype($ftype))
+            Expr(:kw, fname, fdef)
+        end
+        :($MutableTangent{$primal}(
+            $(Expr(:tuple, Expr(:parameters, any_mask...))),
+            $(Expr(:tuple, Expr(:parameters, zfield_exprs...))),
+        ))
+    else
+        :($Tangent{$primal}($(Expr(:parameters, zfield_exprs...))))
+    end
+end
+
+zero_tangent(primal::Tuple) = Tangent{typeof(primal)}(map(zero_tangent, primal)...)
+
+function zero_tangent(x::Array{P,N}) where {P,N}
+    if (isbitstype(P) || all(i -> isassigned(x, i), eachindex(x)))
+        return map(zero_tangent, x)
+    end
+
+    # Now we need to handle nonfully assigned arrays
+    # see discussion at https://github.com/JuliaDiff/ChainRulesCore.jl/pull/626#discussion_r1345235265
+    y = Array{guess_zero_tangent_type(P),N}(undef, size(x)...)
+    @inbounds for n in eachindex(y)
+        if isassigned(x, n)
+            y[n] = zero_tangent(x[n])
+        end
+    end
+    return y
+end
+
+# Sad heauristic methods we need because of unassigned values
+guess_zero_tangent_type(::Type{T}) where {T<:Number} = T
+guess_zero_tangent_type(::Type{T}) where {T<:Integer} = typeof(float(zero(T)))
+function guess_zero_tangent_type(::Type{<:Array{T,N}}) where {T,N}
+    return Array{guess_zero_tangent_type(T),N}
+end
+guess_zero_tangent_type(T::Type) = Any