Fast conversion of wrapped types to SparseMatrixCSC (#30552)

* Fast conversion of wrapped types to SparseMatrixCSC
* moved sparse conversions to sparseconvert.jl

stdlib/SparseArrays/src/SparseArrays.jl

@@ -36,8 +36,11 @@ export AbstractSparseArray, AbstractSparseMatrix, AbstractSparseVector,
     issparse, nonzeros, nzrange, rowvals, sparse, sparsevec, spdiagm,
     sprand, sprandn, spzeros, nnz, permute, findnz
 
+export unwrap, iswrsparse
+
 include("abstractsparse.jl")
 include("sparsematrix.jl")
+include("sparseconvert.jl")
 include("sparsevector.jl")
 include("higherorderfns.jl")
 include("linalg.jl")

stdlib/SparseArrays/src/sparseconvert.jl

@@ -0,0 +1,288 @@
+# This file is a part of Julia. License is MIT: https://julialang.org/license
+
+import LinearAlgebra: AbstractTriangular
+
+"""
+    SparseMatrixCSCSymmHerm
+
+`Symmetric` or `Hermitian` of a `SparseMatrixCSC` or `SparseMatrixCSCView`.
+"""
+const SparseMatrixCSCSymmHerm{Tv,Ti} = Union{Symmetric{Tv,<:SparseMatrixCSCUnion{Tv,Ti}},
+                                            Hermitian{Tv,<:SparseMatrixCSCUnion{Tv,Ti}}}
+
+const AbstractTriangularSparse{Tv,Ti} = AbstractTriangular{Tv,<:SparseMatrixCSCUnion{Tv,Ti}}
+
+# converting Symmetric/Hermitian/AbstractTriangular/SubArray of SparseMatrixCSC
+# and Transpose/Adjoint of AbstractTriangular of SparseMatrixCSC to SparseMatrixCSC
+for wr in (Symmetric, Hermitian, Transpose, Adjoint,
+           UpperTriangular, LowerTriangular, UnitUpperTriangular, UnitLowerTriangular,
+           SubArray)
+
+    @eval SparseMatrixCSC(A::$wr) = _sparsem(A)
+    @eval SparseMatrixCSC{Tv}(A::$wr{Tv}) where Tv = _sparsem(A)
+    @eval SparseMatrixCSC{Tv}(A::$wr) where Tv = SparseMatrixCSC{Tv}(_sparsem(A))
+    @eval SparseMatrixCSC{Tv,Ti}(A::$wr) where {Tv,Ti} = SparseMatrixCSC{Tv,Ti}(_sparsem(A))
+end
+
+"""
+    iswrsparse(::S)
+    iswrsparse(::Type{S})
+
+Returns `true` if type `S` is backed by a sparse array, and `false` otherwise.
+"""
+iswrsparse(::T) where T<:AbstractArray = iswrsparse(T)
+iswrsparse(::Type) = false
+iswrsparse(::Type{T}) where T<:AbstractSparseArray = true
+
+"""
+    depth(::Type{S})
+
+Returns 0 for unwrapped S, and nesting depth for wrapped (nested) abstract arrays.
+"""
+depth(::T) where T = depth(T)
+depth(::Type{T}) where T<:AbstractArray = 0
+
+for wr in (Symmetric, Hermitian,
+           LowerTriangular, UnitLowerTriangular, UpperTriangular, UnitUpperTriangular,
+           Transpose, Adjoint, SubArray,
+           Diagonal, Bidiagonal, Tridiagonal, SymTridiagonal)
+
+    pl = wr === SubArray ? :($wr{<:Any,<:Any,T}) : :($wr{<:Any,T})
+    @eval iswrsparse(::Type{<:$pl}) where T = iswrsparse(T)
+    @eval depth(::Type{<:$pl}) where T = depth(T) + 1
+end
+
+# convert parent and re-wrap in same wrapper
+_sparsewrap(A::Symmetric) = Symmetric(_sparsem(parent(A)), A.uplo == 'U' ? :U : :L)
+_sparsewrap(A::Hermitian) = Hermitian(_sparsem(parent(A)), A.uplo == 'U' ? :U : :L)
+_sparsewrap(A::SubArray) = SubArray(_sparsem(parent(A)), A.indices)
+for ty in ( LowerTriangular, UnitLowerTriangular,
+            UpperTriangular, UnitUpperTriangular,
+            Transpose, Adjoint)
+
+    @eval _sparsewrap(A::$ty) = $ty(_sparsem(parent(A)))
+end
+function _sparsewrap(A::Union{Diagonal,Bidiagonal,Tridiagonal,SymTridiagonal})
+    dropzeros!(sparse(A))
+end
+
+"""
+    unwrap(A::AbstractMatrix)
+
+In case A is a wrapper type (`SubArray, Symmetric, Adjoint, SubArray, Triangular, Tridiagonal`, etc.)
+convert to `Matrix` or `SparseMatrixCSC`, depending on final storage type of A.
+For other types return A itself.
+"""
+unwrap(A::Any) = A
+unwrap(A::AbstractMatrix) = iswrsparse(A) ? convert(SparseMatrixCSC, A) : convert(Array, A)
+
+import Base.copy
+copy(A::SubArray{T,2}) where T = getindex(unwrap(parent(A)), A.indices...)
+
+# For pure sparse matrices and vectors return A.
+# For wrapped sparse matrices or vectors convert to SparseMatrixCSC.
+# Handle nested wrappers properly.
+# Use abstract matrix fallback if A is not sparse.
+function _sparsem(@nospecialize A::AbstractArray{Tv}) where Tv
+    if iswrsparse(A)
+        if depth(A) >= 1
+            _sparsem(_sparsewrap(A))
+        else
+            A
+        end
+    else
+        # explicitly call abstract matrix fallback using getindex(A,...)
+        invoke(SparseMatrixCSC{Tv,Int}, Tuple{AbstractMatrix}, A)
+    end
+end
+
+_sparsem(A::AbstractSparseMatrix) = A
+_sparsem(A::AbstractSparseVector) = A
+
+# Transpose/Adjoint of sparse vector (returning sparse matrix)
+function _sparsem(A::Union{Transpose{<:Any,<:AbstractSparseVector},Adjoint{<:Any,<:AbstractSparseVector}})
+    B = parent(A)
+    n = length(B)
+    Ti = eltype(B.nzind)
+    fadj = A isa Transpose ? transpose : adjoint
+    colptr = fill!(Vector{Ti}(undef, n + 1), 0)
+    colptr[1] = 1
+    colptr[B.nzind .+ 1] .= 1
+    cumsum!(colptr, colptr)
+    rowval = fill!(similar(B.nzind), 1)
+    nzval = fadj.(nonzeros(B))
+    SparseMatrixCSC(1, n, colptr, rowval, nzval)
+end
+
+function _sparsem(A::Union{Transpose{<:Any,<:SparseMatrixCSC},Adjoint{<:Any,<:SparseMatrixCSC}})
+    ftranspose(parent(A), A isa Transpose ? transpose : adjoint)
+end
+
+# Symmetric/Hermitian of sparse matrix
+_sparsem(A::SparseMatrixCSCSymmHerm) = _sparsem(A.uplo == 'U' ? nzrangeup : nzrangelo, A)
+# Triangular of sparse matrix
+_sparsem(A::UpperTriangular{T,<:AbstractSparseMatrix}) where T = triu(A.data)
+_sparsem(A::LowerTriangular{T,<:AbstractSparseMatrix}) where T = tril(A.data)
+# view of sparse matrix
+_sparsem(S::SubArray{<:Any,2,<:SparseMatrixCSC}) = getindex(S.parent,S.indices...)
+
+# 4 cases: (Symmetric|Hermitian) variants (:U|:L)
+function _sparsem(fnzrange::Function, sA::SparseMatrixCSCSymmHerm{Tv}) where {Tv}
+    A = sA.data
+    rowval = rowvals(A)
+    nzval = nonzeros(A)
+    m, n = size(A)
+    Ti = eltype(rowval)
+    fadj = sA isa Symmetric ? transpose : adjoint
+    newcolptr = Vector{Ti}(undef, n+1)
+    diagmap = fadj == transpose ? identity : real
+
+    newcolptr[1] = 1
+    colrange = fnzrange === nzrangeup ? (1:n) : (n:-1:1)
+    @inbounds for j = colrange
+        r = fnzrange(A, j); r1 = r.start; r2 = r.stop
+        newcolptr[j+1] = r2 - r1 + 1
+        for k = r1:r2
+            row = rowval[k]
+            if row != j
+                newcolptr[row+1] += 1
+            end
+        end
+    end
+    cumsum!(newcolptr, newcolptr)
+    nz = newcolptr[n+1] - 1
+    newrowval = Vector{Ti}(undef, nz)
+    newnzval = Vector{Tv}(undef, nz)
+    @inbounds for j = 1:n
+        newk = newcolptr[j]
+        for k = fnzrange(A, j)
+            i = rowval[k]
+            nzv = nzval[k]
+            if i != j
+                newrowval[newk] = i
+                newnzval[newk] = nzv
+                newk += 1
+                ni = newcolptr[i]
+                newrowval[ni] = j
+                newnzval[ni] = fadj(nzv)
+                newcolptr[i] = ni + 1
+            else
+                newrowval[newk] = i
+                newnzval[newk] = diagmap(nzv)
+                newk += 1
+            end
+        end
+        newcolptr[j] = newk
+    end
+    _sparse_gen(m, n, newcolptr, newrowval, newnzval)
+end
+
+# 2 cases: Unit(Upper|Lower)Triangular{Tv,SparseMatrixCSC}
+function _sparsem(A::AbstractTriangularSparse{Tv}) where Tv
+    S = A.data
+    rowval = rowvals(S)
+    nzval = nonzeros(S)
+    m, n = size(S)
+    Ti = eltype(rowval)
+    fnzrange = A isa Union{UpperTriangular,UnitUpperTriangular} ? nzrangeup : nzrangelo
+    unit = A isa Union{UnitUpperTriangular,UnitLowerTriangular}
+    newcolptr = Vector{Ti}(undef, n+1)
+    newrowval = Vector{Ti}(undef, nnz(S))
+    newnzval = Vector{Tv}(undef, nnz(S))
+    newcolptr[1] = 1
+    uplo = fnzrange == nzrangeup
+    newk = 1
+    @inbounds for j = 1:n
+        newkk = newk
+        if unit
+            newk += !uplo
+        end
+        r = fnzrange(S, j); r1 = r.start; r2 = r.stop
+        for k = r1:r2
+            i = rowval[k]
+            if i != j || i == j && !unit
+                newrowval[newk] = i
+                newnzval[newk] = nzval[k]
+                newk += 1
+            end
+        end
+        if unit
+            uplo && (newkk = newk)
+            newrowval[newkk] = j
+            newnzval[newkk] = one(Tv)
+            newk += uplo
+        end
+        newcolptr[j+1] = newk
+    end
+    nz = newcolptr[n+1] - 1
+    resize!(newrowval, nz)
+    resize!(newnzval, nz)
+    SparseMatrixCSC(m, n, newcolptr, newrowval, newnzval)
+end
+
+# 8 cases: (Transpose|Adjoint){Tv,[Unit](Upper|Lower)Triangular}
+function _sparsem(taA::Union{Transpose{Tv,<:AbstractTriangularSparse},
+                             Adjoint{Tv,<:AbstractTriangularSparse}}) where {Tv}
+
+    sA = taA.parent
+    A = sA.data
+    rowval = rowvals(A)
+    nzval = nonzeros(A)
+    m, n = size(A)
+    Ti = eltype(rowval)
+    fnzrange = sA isa Union{UpperTriangular,UnitUpperTriangular} ? nzrangeup : nzrangelo
+    fadj = taA isa Transpose ? transpose : adjoint
+    unit = sA isa Union{UnitUpperTriangular,UnitLowerTriangular}
+    uplo = A isa Union{UpperTriangular,UnitUpperTriangular}
+
+    newcolptr = Vector{Ti}(undef, n+1)
+    fill!(newcolptr, 1unit)
+    newcolptr[1] = 1
+    @inbounds for j = 1:n
+        for k = fnzrange(A, j)
+            i = rowval[k]
+            if i != j || i == j && !unit
+                newcolptr[i+1] += 1
+            end
+        end
+    end
+    cumsum!(newcolptr, newcolptr)
+    nz = newcolptr[n+1] - 1
+    newrowval = Vector{Ti}(undef, nz)
+    newnzval = Vector{Tv}(undef, nz)
+
+    @inbounds for j = 1:n
+        if !uplo && unit
+            ni = newcolptr[j]
+            newrowval[ni] = j
+            newnzval[ni] = fadj(one(Tv))
+            newcolptr[j] = ni + 1
+        end
+        for k = fnzrange(A, j)
+            i = rowval[k]
+            nzv = nzval[k]
+            if i != j || i == j && !unit
+                ni = newcolptr[i]
+                newrowval[ni] = j
+                newnzval[ni] = fadj(nzv)
+                newcolptr[i] = ni + 1
+            end
+        end
+        if uplo && unit
+            ni = newcolptr[j]
+            newrowval[ni] = j
+            newnzval[ni] = fadj(one(Tv))
+            newcolptr[j] = ni + 1
+        end
+    end
+    _sparse_gen(n, m, newcolptr, newrowval, newnzval)
+end
+
+function _sparse_gen(m, n, newcolptr, newrowval, newnzval)
+    @inbounds for j = n:-1:1
+        newcolptr[j+1] = newcolptr[j]
+    end
+    newcolptr[1] = 1
+    SparseMatrixCSC(m, n, newcolptr, newrowval, newnzval)
+end
+

stdlib/SparseArrays/test/sparse.jl

@@ -2510,6 +2510,42 @@ end
     end
 end
 
+@testset "wrappers of sparse" begin
+    m = n = 10
+    A = spzeros(ComplexF64, m, n)
+    A[:,1] = 1:m
+    A[:,2] = [1 3 0 0 0 0 0 0 0 0]'
+    A[:,3] = [2 4 0 0 0 0 0 0 0 0]'
+    A[:,4] = [0 0 0 0 5 3 0 0 0 0]'
+    A[:,5] = [0 0 0 0 6 2 0 0 0 0]'
+    A[:,6] = [0 0 0 0 7 4 0 0 0 0]'
+    A[:,7:n] = rand(ComplexF64, m, n-6)
+    B = Matrix(A)
+    dowrap(wr, A) = wr(A)
+    dowrap(wr::Tuple, A) = (wr[1])(A, wr[2:end]...)
+
+    @testset "sparse($wr(A))" for wr in (
+                        Symmetric, (Symmetric, :L), Hermitian, (Hermitian, :L),
+                        Transpose, Adjoint,
+                        UpperTriangular, LowerTriangular,
+                        UnitUpperTriangular, UnitLowerTriangular,
+                        (view, 3:6, 2:5))
+
+        @test SparseMatrixCSC(dowrap(wr, A)) == Matrix(dowrap(wr, B))
+    end
+
+    @testset "sparse($at($wr))" for at = (Transpose, Adjoint), wr =
+        (UpperTriangular, LowerTriangular,
+         UnitUpperTriangular, UnitLowerTriangular)
+
+        @test SparseMatrixCSC(at(wr(A))) == Matrix(at(wr(B)))
+    end
+
+    @test sparse([1,2,3,4,5]') == SparseMatrixCSC([1 2 3 4 5])
+    @test sparse(UpperTriangular(A')) == UpperTriangular(B')
+    @test sparse(Adjoint(UpperTriangular(A'))) == Adjoint(UpperTriangular(B'))
+end
+
 @testset "Ti cannot store all potential values #31024" begin
     @test_throws ArgumentError SparseMatrixCSC(128, 1, [Int8(1), Int8(1)], Int8[], Int[])
     @test_throws ArgumentError SparseMatrixCSC(12, 12, [Int8(1), Int8(1)], Int8[], Int[])