Serious regression of warp!

[jianghaizhu]

Some serious memory allocation and regression in Julia 0.7. Here is a MWE.

julia> img = rand(100, 100);
julia> tfm = recenter(RotMatrix(-pi/4), center(img));

In Julia 0.6,

julia> @btime warp(img, tfm);
  422.876 μs (5 allocations: 157.72 KiB)

In Julia 0.7,

julia> @btime warp(img, tfm);
  8.478 ms (261626 allocations: 9.07 MiB)

Looking at the code of warp!,

function warp!(out, img::AbstractExtrapolation, tform)
    @inbounds for I in CartesianIndices(axes(out))
        out[I] = _getindex(img, tform(SVector(I.I)))
    end
    out       <==== should not return anything
end

Obviously, warp! should not return out. But eliminating the return only helps marginally. But I cannot dig any further. Might be some problem with _getindex.

[kmsquire]

Returning out is not a problem. It doesn't allocate any additional memory, and allows the result to be used pipelines. (Unlike, e.g., pandas, there's no convention that functions which modify their parameters in place return nothing.)

[jianghaizhu]

Yes. Returning is not a problem. The question here is what the problem is.

[Evizero]

Thanks for the report. I'll investigate soon when I find some time

[RalphAS]

Profiling points to getindex, which we hoped was fixed by JuliaLang/julia#28707. But the OP's example is even worse on a recent nightly of Julia (after that PR).
Is @timholy's overhaul of Interpolations.jl likely to make compiler optimization any more effective for interpolation objects, or do we need to pray for more magic from Keno?

[timholy]

The real issue is inferring length, which is special-cased for Tuples. Relevant: JuliaLang/julia#27434 (comment)

[RalphAS]

I'm glad that @Evizero caught my mistake. The sad thing is that I knew better last year.

Is it the case that splatting (almost?) anything other than a Tuple leads to dynamic dispatch, which can foul up inference in various (hard-to-predict) ways?

[Evizero]

No worries, these things happen to us all. If anything I should have caught it in the review. Your contributions are a clear net positive and we are lucky to have you. The occasional regression is to be expected and nothing we can't fix