Slowdown in Julia 0.2

[magistere]

On Julia 0.2 release I got significant slowdown in my old test script comparing with Julia 0.2-prerelease-d6f7c7c. Run time increased from 5s to 40s.

import Base.div

const width = 640
const height = 480
const iters = 260

immutable Pixel
  r::Uint8
  g::Uint8
  b::Uint8
end

+(p1::Pixel, p2::Pixel) = Pixel(uint8(p1.r+p2.r),uint8(p1.g+p2.g),uint8(p1.b+p2.b))
div(p::Pixel, n::Int) = Pixel(uint8(div(p.r,n)),uint8(div(p.g,n)),uint8(div(p.b,n)))

function avg!(a::Array{Pixel,2}, b::Array{Pixel,2})
  for j = 2:height-1, i = 2:width-1
    b[i,j] = div(a[i-1,j] + a[i+1,j] + a[i,j-1] + a[i,j+1], 4)
  end
end

function test()
  in = Array(Pixel,width,height)
  out = Array(Pixel,width,height)
  tic()
  for i = 1:iters
    avg!(in,out)
    avg!(out,in)
  end
  toc()
end

test()

[kmsquire]

Hi @magistere, since you know a good commit and bad commit, are you up to doing a git bisect? #1689 (comment)

[magistere]

Thanks, Kevin. I've found the commit 9489e6a that causes this slowdown.

[kmsquire]

cc: @JeffBezanson

[StefanKarpinski]

Maybe add this to our performance benchmarks?

On Monday, December 2, 2013, Kevin Squire wrote:

cc: @JeffBezanson https://github.com/JeffBezanson

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/5011#issuecomment-29673966
.

[vtjnash]

this is improved in 0.3 master, but still far from the original:

julia> @time test()
elapsed time: 14.814826383 seconds
elapsed time: 15.471788592 seconds (9518721684 bytes allocated)
14.814826383

julia> VERSION
v"0.2.1+13"

julia> @time test()
elapsed time: 6.039069812 seconds
elapsed time: 6.300719424 seconds (4157816 bytes allocated)
6.039069812

julia> VERSION
v"0.3.0-prerelease+2742"

julia> VERSION
v"0.2.0-prerelease+3768"

julia> @time test()
elapsed time: 1.596434471 seconds
elapsed time: 1.690605844 seconds (3749388 bytes allocated)
1.596434471

[ViralBShah]

This sort of stuff should certainly go into the perf benchmarks.

[vtjnash]

i think we may need to make inline_worthy a little smarter: changing the threshold there from current value of 40 to 240 (tested on my jn/6599 bugfix branch) results in net smaller code generation (current testing seems to show that that number corresponds roughly to the number of assembly bytes!) and results in a net speed increase over pre-0.2 best timing.

[ViralBShah]

I am tempted to close this as a 0.2 thing, but running the code fragment here takes 15 seconds for me.

[simonster]

Yes, this is still an issue. See @vtjnash's comments here.