Evaluating rules is unreasonably slow on first call

[juliadiff]

Evaluating rules is unreasonably slow on first call

[oxinabox]

What happens if we do a top level @nospecialiize on the whole package?

Answer: helps a bit but not much

[iamed2]

What happens if we do a top level @nospecialiize on the whole package?

Is there still a large overhead between first and second call? Or are they about the same?

[oxinabox]

It was below the noise level for both.
Some minor different in allocations but not much.

[iamed2]

You're probably just getting interpreter run time then

[oxinabox]

This feels a lot better when I was trying it after the breakup into two packages,
and when working in julia 1.3,
but I've not done new timing.
I suspect breaking into two packages has made precompilation happier.

[oxinabox]

#35 does indeed fix this.

Test Script

using Pkg: @pkg_str
pkg"activate /Users/oxinabox/JuliaEnvs/ChainRulesWorld/"
using ChainRules
using LinearAlgebra

function main()
    x = [1 2; 3 4]
    @time f, res = rrule(svd, x)
    r = (U = f.U, S=f.S, V=f.V)
    @time res(r)
end
main()

Timing Results

Current Master

rrule: 0.043273 seconds (29.90 k allocations: 1.602 MiB, 14.56% gc time)
res : 0.029336 seconds (59.08 k allocations: 3.668 MiB)

No cassette branch #35

rrule: 0.040967 seconds (29.90 k allocations: 1.602 MiB)
res: 0.000014 seconds (19 allocations: 1.891 KiB)

[nickrobinson251]

pkg"activate ChainRulesWorld"

😆

[oxinabox]

I have a Manifest.toml there
that has all the right branchs of ChainRules, ChainRulesCore, Zygote and FiniteDifferences etc
checked out for what ever I am working on.