Merge pull request #259 from avik-pal/ap/known_sparsity

Allow specifying colorvecs directly

Author: ChrisRackauckas

Project.toml

@@ -1,7 +1,7 @@
 name = "SparseDiffTools"
 uuid = "47a9eef4-7e08-11e9-0b38-333d64bd3804"
 authors = ["Pankaj Mishra <pankajmishra1511@gmail.com>", "Chris Rackauckas <contact@chrisrackauckas.com>"]
-version = "2.5.2"
+version = "2.6.0"
 
 [deps]
 ADTypes = "47edcb42-4c32-4615-8424-f2b9edc5f35b"

src/SparseDiffTools.jl

@@ -88,8 +88,9 @@ export update_coefficients, update_coefficients!, value!
 # High Level Interface: sparse_jacobian
 export AutoSparseEnzyme
 
-export NoSparsityDetection,
-    SymbolicsSparsityDetection, JacPrototypeSparsityDetection, AutoSparsityDetection
+export NoSparsityDetection, SymbolicsSparsityDetection, JacPrototypeSparsityDetection,
+    PrecomputedJacobianColorvec, AutoSparsityDetection
 export sparse_jacobian, sparse_jacobian_cache, sparse_jacobian!
+export init_jacobian
 
 end # module

src/highlevel/coloring.jl

@@ -16,8 +16,16 @@ struct NoMatrixColoring end
 # Prespecified Jacobian Structure
 function (alg::JacPrototypeSparsityDetection)(ad::AbstractSparseADType, args...; kwargs...)
     J = alg.jac_prototype
-    reverse_mode = ad isa AbstractSparseReverseMode
-    colorvec = matrix_colors(J, alg.alg; partition_by_rows = reverse_mode)
+    colorvec = matrix_colors(J, alg.alg;
+        partition_by_rows = ad isa AbstractSparseReverseMode)
+    (nz_rows, nz_cols) = ArrayInterface.findstructralnz(J)
+    return MatrixColoringResult(colorvec, J, nz_rows, nz_cols)
+end
+
+# Prespecified Colorvecs
+function (alg::PrecomputedJacobianColorvec)(ad::AbstractSparseADType, args...; kwargs...)
+    colorvec = _get_colorvec(alg, ad)
+    J = alg.jac_prototype
     (nz_rows, nz_cols) = ArrayInterface.findstructralnz(J)
     return MatrixColoringResult(colorvec, J, nz_rows, nz_cols)
 end

src/highlevel/common.jl

@@ -9,18 +9,110 @@ abstract type AbstractSparsityDetection <: AbstractMaybeSparsityDetection end
 
 struct NoSparsityDetection <: AbstractMaybeSparsityDetection end
 
+"""
+    SymbolicsSparsityDetection(; alg = GreedyD1Color())
+
+Use Symbolics to compute the sparsity pattern of the Jacobian. This requires `Symbolics.jl`
+to be explicitly loaded.
+
+## Keyword Arguments
+
+    - `alg`: The algorithm used for computing the matrix colors
+
+See Also: [JacPrototypeSparsityDetection](@ref), [PrecomputedJacobianColorvec](@ref)
+"""
 Base.@kwdef struct SymbolicsSparsityDetection{A <: ArrayInterface.ColoringAlgorithm} <:
                    AbstractSparsityDetection
     alg::A = GreedyD1Color()
 end
 
+"""
+    JacPrototypeSparsityDetection(; jac_prototype, alg = GreedyD1Color())
+
+Use a pre-specified `jac_prototype` to compute the matrix colors of the Jacobian.
+
+## Keyword Arguments
+
+    - `jac_prototype`: The prototype Jacobian used for computing the matrix colors
+    - `alg`: The algorithm used for computing the matrix colors
+
+See Also: [SymbolicsSparsityDetection](@ref), [PrecomputedJacobianColorvec](@ref)
+"""
 Base.@kwdef struct JacPrototypeSparsityDetection{
-    J, A <: ArrayInterface.ColoringAlgorithm,
-} <: AbstractSparsityDetection
+    J, A <: ArrayInterface.ColoringAlgorithm} <: AbstractSparsityDetection
     jac_prototype::J
     alg::A = GreedyD1Color()
 end
 
+"""
+    PrecomputedJacobianColorvec(jac_prototype, row_colorvec, col_colorvec)
+
+Use a pre-specified `colorvec` which can be directly used for sparse differentiation. Based
+on whether a reverse mode or forward mode or finite differences is used, the corresponding
+`row_colorvec` or `col_colorvec` is used. Atmost one of them can be set to `nothing`.
+
+## Arguments
+
+    - `jac_prototype`: The prototype Jacobian used for computing structural nonzeros
+    - `row_colorvec`: The row colorvec of the Jacobian
+    - `col_colorvec`: The column colorvec of the Jacobian
+
+See Also: [SymbolicsSparsityDetection](@ref), [JacPrototypeSparsityDetection](@ref)
+"""
+struct PrecomputedJacobianColorvec{J, RC, CC} <: AbstractSparsityDetection
+    jac_prototype::J
+    row_colorvec::RC
+    col_colorvec::CC
+end
+
+"""
+    PrecomputedJacobianColorvec(; jac_prototype, partition_by_rows::Bool = false,
+        colorvec = missing, row_colorvec = missing, col_colorvec = missing)
+
+Use a pre-specified `colorvec` which can be directly used for sparse differentiation. Based
+on whether a reverse mode or forward mode or finite differences is used, the corresponding
+`row_colorvec` or `col_colorvec` is used. Atmost one of them can be set to `nothing`.
+
+## Keyword Arguments
+
+    - `jac_prototype`: The prototype Jacobian used for computing structural nonzeros
+    - `partition_by_rows`: Whether to partition the Jacobian by rows or columns (row
+      partitioning is used for reverse mode AD)
+    - `colorvec`: The colorvec of the Jacobian. If `partition_by_rows` is `true` then this
+      is the row colorvec, otherwise it is the column colorvec
+    - `row_colorvec`: The row colorvec of the Jacobian
+    - `col_colorvec`: The column colorvec of the Jacobian
+
+See Also: [SymbolicsSparsityDetection](@ref), [JacPrototypeSparsityDetection](@ref)
+"""
+function PrecomputedJacobianColorvec(; jac_prototype, partition_by_rows::Bool = false,
+    colorvec = missing, row_colorvec = missing, col_colorvec = missing)
+    if colorvec === missing
+        @assert row_colorvec !== missing||col_colorvec !== missing "Either `colorvec` or `row_colorvec` and `col_colorvec` must be specified!"
+        row_colorvec = row_colorvec === missing ? nothing : row_colorvec
+        col_colorvec = col_colorvec === missing ? nothing : col_colorvec
+        return PrecomputedJacobianColorvec(jac_prototype, row_colorvec, col_colorvec)
+    else
+        @assert row_colorvec === missing&&col_colorvec === missing "Specifying `colorvec` is incompatible with specifying `row_colorvec` or `col_colorvec`!"
+        row_colorvec = partition_by_rows ? colorvec : nothing
+        col_colorvec = partition_by_rows ? nothing : colorvec
+        return PrecomputedJacobianColorvec(jac_prototype, row_colorvec, col_colorvec)
+    end
+end
+
+function _get_colorvec(alg::PrecomputedJacobianColorvec, ad)
+    cvec = alg.col_colorvec
+    @assert cvec!==nothing "`col_colorvec` is nothing, but Forward Mode AD or Finite Differences is being used!"
+    return cvec
+end
+
+function _get_colorvec(alg::PrecomputedJacobianColorvec, ::AbstractReverseMode)
+    cvec = alg.row_colorvec
+    @assert cvec!==nothing "`row_colorvec` is nothing, but Reverse Mode AD is being used!"
+    return cvec
+end
+
+# No one should be using this currently
 Base.@kwdef struct AutoSparsityDetection{A <: ArrayInterface.ColoringAlgorithm} <:
                    AbstractSparsityDetection
     alg::A = GreedyD1Color()
@@ -41,7 +133,8 @@ Inplace update the matrix `J` with the Jacobian of `f` at `x` using the AD backe
 function sparse_jacobian! end
 
 """
-    sparse_jacobian_cache(ad::AbstractADType, sd::AbstractSparsityDetection, f, x; fx=nothing)
+    sparse_jacobian_cache(ad::AbstractADType, sd::AbstractSparsityDetection, f, x;
+        fx=nothing)
     sparse_jacobian_cache(ad::AbstractADType, sd::AbstractSparsityDetection, f!, fx, x)
 
 Takes the underlying AD backend `ad`, sparsity detection algorithm `sd`, function `f`,
@@ -67,7 +160,7 @@ with the same cache to compute the jacobian.
 function sparse_jacobian(ad::AbstractADType, sd::AbstractMaybeSparsityDetection, args...;
     kwargs...)
     cache = sparse_jacobian_cache(ad, sd, args...; kwargs...)
-    J = __init_𝒥(cache)
+    J = init_jacobian(cache)
     return sparse_jacobian!(J, ad, cache, args...)
 end
 
@@ -80,7 +173,7 @@ Jacobian at every function call
 """
 function sparse_jacobian(ad::AbstractADType, cache::AbstractMaybeSparseJacobianCache,
     args...)
-    J = __init_𝒥(cache)
+    J = init_jacobian(cache)
     return sparse_jacobian!(J, ad, cache, args...)
 end
 
@@ -106,7 +199,20 @@ function __gradient end
 function __gradient! end
 function __jacobian! end
 
-function __init_𝒥 end
+"""
+    init_jacobian(cache::AbstractMaybeSparseJacobianCache)
+
+Initialize the Jacobian based on the cache. Uses sparse jacobians if possible.
+
+!!! note
+    This function doesn't alias the provided jacobian prototype. It always initializes a
+    fresh jacobian that can be mutated without any side effects.
+"""
+function init_jacobian end
+
+# Never thought this was a useful function externally, but I ended up using it in quite a
+# few places. Keeping this till I remove uses of those.
+const __init_𝒥 = init_jacobian
 
 # Misc Functions
 __chunksize(::AutoSparseForwardDiff{C}) where {C} = C
@@ -123,12 +229,12 @@ end
     return :(nothing)
 end
 
-function __init_𝒥(c::AbstractMaybeSparseJacobianCache)
+function init_jacobian(c::AbstractMaybeSparseJacobianCache)
     T = promote_type(eltype(c.fx), eltype(c.x))
-    return __init_𝒥(__getfield(c, Val(:jac_prototype)), T, c.fx, c.x)
+    return init_jacobian(__getfield(c, Val(:jac_prototype)), T, c.fx, c.x)
 end
-__init_𝒥(::Nothing, ::Type{T}, fx, x) where {T} = similar(fx, T, length(fx), length(x))
-__init_𝒥(J, ::Type{T}, _, _) where {T} = similar(J, T, size(J, 1), size(J, 2))
+init_jacobian(::Nothing, ::Type{T}, fx, x) where {T} = similar(fx, T, length(fx), length(x))
+init_jacobian(J, ::Type{T}, _, _) where {T} = similar(J, T, size(J, 1), size(J, 2))
 
 __maybe_copy_x(_, x) = x
 __maybe_copy_x(_, ::Nothing) = nothing

test/test_sparse_jacobian.jl

@@ -26,9 +26,12 @@ J_true = ForwardDiff.jacobian(fdiff, x);
 
 # SparseDiffTools High-Level API
 J_sparsity = Symbolics.jacobian_sparsity(fdiff, similar(x), x);
+row_colorvec = SparseDiffTools.matrix_colors(J_sparsity; partition_by_rows = true)
+col_colorvec = SparseDiffTools.matrix_colors(J_sparsity; partition_by_rows = false)
 
-SPARSITY_DETECTION_ALGS = [JacPrototypeSparsityDetection(jac_prototype = J_sparsity),
-    SymbolicsSparsityDetection(), NoSparsityDetection()]
+SPARSITY_DETECTION_ALGS = [JacPrototypeSparsityDetection(; jac_prototype = J_sparsity),
+    SymbolicsSparsityDetection(), NoSparsityDetection(),
+    PrecomputedJacobianColorvec(; jac_prototype = J_sparsity, row_colorvec, col_colorvec)]
 
 @testset "High-Level API" begin
     @testset "Sparsity Detection: $(nameof(typeof(sd)))" for sd in SPARSITY_DETECTION_ALGS
@@ -40,7 +43,7 @@ SPARSITY_DETECTION_ALGS = [JacPrototypeSparsityDetection(jac_prototype = J_spars
             AutoSparseFiniteDiff(), AutoFiniteDiff(), AutoEnzyme(), AutoSparseEnzyme())
             @testset "Cache & Reuse" begin
                 cache = sparse_jacobian_cache(difftype, sd, fdiff, x)
-                J = SparseDiffTools.__init_𝒥(cache)
+                J = init_jacobian(cache)
 
                 sparse_jacobian!(J, difftype, cache, fdiff, x)
 
@@ -74,7 +77,7 @@ SPARSITY_DETECTION_ALGS = [JacPrototypeSparsityDetection(jac_prototype = J_spars
                 @info "$(nameof(typeof(difftype)))() `sparse_jacobian` (complete) time: $(t₁)s"
 
                 cache = sparse_jacobian_cache(difftype, sd, fdiff, x)
-                J = SparseDiffTools.__init_𝒥(cache)
+                J = init_jacobian(cache)
 
                 sparse_jacobian!(J, difftype, sd, fdiff, x)
 
@@ -95,7 +98,7 @@ SPARSITY_DETECTION_ALGS = [JacPrototypeSparsityDetection(jac_prototype = J_spars
             cache = sparse_jacobian_cache(difftype, sd, fdiff, y, x)
 
             @testset "Cache & Reuse" begin
-                J = SparseDiffTools.__init_𝒥(cache)
+                J = init_jacobian(cache)
                 sparse_jacobian!(J, difftype, cache, fdiff, y, x)
 
                 @test J ≈ J_true
@@ -126,7 +129,7 @@ SPARSITY_DETECTION_ALGS = [JacPrototypeSparsityDetection(jac_prototype = J_spars
                 t₁ = @elapsed sparse_jacobian(difftype, sd, fdiff, y, x)
                 @info "$(nameof(typeof(difftype)))() `sparse_jacobian` (complete) time: $(t₁)s"
 
-                J = SparseDiffTools.__init_𝒥(cache)
+                J = init_jacobian(cache)
 
                 sparse_jacobian!(J, difftype, sd, fdiff, y, x)
 
@@ -142,7 +145,7 @@ SPARSITY_DETECTION_ALGS = [JacPrototypeSparsityDetection(jac_prototype = J_spars
             AutoZygote())
             y = similar(x)
             cache = sparse_jacobian_cache(difftype, sd, fdiff, y, x)
-            J = SparseDiffTools.__init_𝒥(cache)
+            J = init_jacobian(cache)
 
             @testset "Cache & Reuse" begin
                 @test_throws Exception sparse_jacobian!(J, difftype, cache, fdiff, y, x)