jlchan / fluxdiffutils.jl Goto Github PK

Would like the matrices returned by the hadamard_* functions (in particular hadamard_jacobian) to be compatible with the vector of vectors, U, used to construct them.

Should write directly to nzval in Jacobian blocks instead of using A[i,j] access. The latter is slow, see MWE

function copy_sparse!(A,B)
    rows = rowvals(B)
    vals = nonzeros(B)
    for j = 1:size(B,2)
        for row_id in nzrange(B,j)
            i = rows[row_id]
            # modify B_ij = f(B_ij,i,j) ...
            A[i,j] = vals[row_id]
            # A.nzval[row_id] = vals[row_id] # faster
        end
    end
end

Running with N=1000 matrices (both dense/sparse) shows sparse writes are much slower.

julia> @btime copy_sparse!($Asparse,$B);
  11.801 μs (0 allocations: 0 bytes)

julia> @btime copy_sparse!($Adense,$B);
  2.049 μs (0 allocations: 0 bytes)

Fix so that users don't need to specify "scale = +/- 1", just the matrix type (symmetric, skew-symmetric, or general).

Line 75 of hadamard_jacobian.jl causes the function hadamard_jacobian(A_list, product_type...) to be type unstable. The instability can be recreated with the following snippet of code:

    function makeBlockJ(U)
        Nfields = length(U)
        n = length(U[1])
        Z = spzeros(n,n)
        J = SMatrix{Nfields,Nfields}( [Z for i in 1:Nfields, j in 1:Nfields])
    end
    @code_warntype makeBlockJ(U)

This issue is used to trigger TagBot; feel free to unsubscribe.

If you haven't already, you should update your TagBot.yml to include issue comment triggers.
Please see this post on Discourse for instructions and more details.

If you'd like for me to do this for you, comment TagBot fix on this issue.
I'll open a PR within a few hours, please be patient!

Passing in sparse matrices to hadamard_sum_ATr! seems slower than passing in dense matrices.

Thinking of changing name to FluxDiffUtils.jl, since the new hadamard_sum for dense operators is general purpose.

Matrices are no longer sparse after the broadcasted call totranspose in hadamard_sum and hadamard_sum!(rhs, A_list,...).