Big julia-0.6 performance regression when indexing with indexes in views of (sparse) matrices.

[mkborregaard]

I have been trying to get around the very slow execution on reducedim operations on views into large sparse matrices (see https://discourse.julialang.org/t/slow-arithmetic-on-views-of-sparse-matrices/3644), and see also https://github.com/JuliaLang/julia/issues/21515 and https://github.com/JuliaLang/julia/issues/21796). To do this, I use a custom rowsum function to do sum(x,2) on views into sparse matrices like this:

rowsum(x::SubArray{T,2,P}) where {T,P<:SparseMatrixCSC} = (x.parent * sparse(x.indexes[2],ones(Int,length(x.indexes[2])), ones(Int,length(x.indexes[2])), size(x.parent,2),1))[x.indexes[1]]

On 0.6 this function is >100 times slower for matrices of the size I work with (roughly 10000 x 10000 with 0.01 sparsity), when x is a SubArray that has only been sliced along the columns (e.g. x = view(parentmatrix, :, some_index). The regression is due to the final getindex ([x.indexes[1]]), because x.indexes[1] has a different type on 0.5 (Colon) and 0.6 (Base.Slice{OneTo}), and the Slice is much slower than Colon though they refer to the same elements.

Here is an MWE (not sensible in itself, but the use case is given above:

parentmatrix = sprand(1000,1000, 0.1)
rowview = view(parentmatrix, : , rand(1:1000, 500))
sparsevec = sparse(rand(1000,1))
using BenchmarkTools

#julia 0.5
@benchmark $(sparsevec)[$(rowview.indexes[1])]
  BenchmarkTools.Trial:
    memory estimate:  15.91 KiB
    allocs estimate:  3
    --------------
    minimum time:     4.837 μs (0.00% GC)
    median time:      8.594 μs (0.00% GC)
    mean time:        29.137 μs (19.64% GC)
    maximum time:     5.003 ms (99.44% GC)
    --------------
    samples:          10000
    evals/sample:     7

versioninfo()
Julia Version 0.5.2
Commit f4c6c9d4bb (2017-05-06 16:34 UTC)
Platform Info:
  OS: macOS (x86_64-apple-darwin13.4.0)
  CPU: Intel(R) Core(TM) i7-5557U CPU @ 3.10GHz
  WORD_SIZE: 64
  BLAS: libopenblas (USE64BITINT DYNAMIC_ARCH NO_AFFINITY Haswell)
  LAPACK: libopenblas64_
  LIBM: libopenlibm
  LLVM: libLLVM-3.7.1 (ORCJIT, broadwell)


#julia 0.6
@benchmark $(sparsevec)[$(rowview.indexes[1])]
  BenchmarkTools.Trial:
    memory estimate:  15.92 KiB
    allocs estimate:  4
    --------------
    minimum time:     247.825 μs (0.00% GC)
    median time:      472.646 μs (0.00% GC)
    mean time:        1.549 ms (1.57% GC)
    maximum time:     38.920 ms (98.90% GC)
    --------------
    samples:          3159
    evals/sample:     1

versioninfo()
Julia Version 0.6.0-rc2.0
Commit 68e911be53 (2017-05-18 02:31 UTC)
Platform Info:
  OS: macOS (x86_64-apple-darwin13.4.0)
  CPU: Intel(R) Core(TM) i7-5557U CPU @ 3.10GHz
  WORD_SIZE: 64
  BLAS: libopenblas (USE64BITINT DYNAMIC_ARCH NO_AFFINITY Haswell)
  LAPACK: libopenblas64_
  LIBM: libopenlibm
  LLVM: libLLVM-3.9.1 (ORCJIT, broadwell)