huge performance hit in 0.6: negative of a sparse matrix

[StephenVavasis]

There appears to be a huge performance problem with taking the negative of a sparse matrix in 0.6. Observe the following timings for the code below (64-bit Windows 10 machine):

0.4.6:

julia> VERSION
v"0.4.6"

julia> @time test_minus_sparse.tms(4,10000)
  0.024588 seconds (106 allocations: 28.906 MB, 17.33% gc time)
-7.818594853651364

0.6.0-dev:

julia> VERSION
v"0.6.0-dev.1378"

julia> @time test_minus_sparse.tms(4,10000)
 10.796751 seconds (140 allocations: 28.757 MB, 0.11% gc time)
-7.818594853651364

Code:

module test_minus_sparse

function tms(ntrial,n)
    is = zeros(Int64,0)
    js = zeros(Int64,0)
    es = zeros(0)
    for i = 1 : n
        for k = max(1,i-10) : min(n, i+10)
            push!(is, i)
            push!(js, k)
            push!(es, sin(Float64(i)) * cos(Float64(k)))
        end
    end
    s = sparse(is,js,es,n,n)
    sum_e = 0.0
    for j = 1 : ntrial
        t = -s
        sum_e += t[1,1]
        s[1,1] += 1.0
    end
    sum_e
end
end

[JaredCrean2]

I looked into this a little bit and found:

on 0.4

julia> s = sprand(100, 100, 0.1)
julia> @which -(s)
-(A::SparseMatrixCSC{Tv,Ti<:Integer}) at sparse/sparsematrix.jl:674

on 0.6

julia> s = sprand(100, 100, 0.1)
julia> @which -(s)
-(A::AbstractArray) in Base at arraymath.jl:22

which looks like it iterates over all the entries, rather than the non-zero ones. @Sacha0 has been working on the sparse matrix code recently, he might know more.

[Sacha0]

Seems I missed reinjecting unary minus as a short child of broadcast in #17265. Time allowing I will fix this issue tomorrow. (Edit: Apologies for the delay, prioritized #19518.) Thanks for catching this!