Performance regression in Julia v0.5

[MikaelSlevinsky]

This isn't officially an ApproxFun issue, but the following is a performance anomaly in v0.5. Without loading ApproxFun, the simple function f evaluates fast and allocation-free:

julia> versioninfo()
Julia Version 0.5.1-pre+2
Commit f0d40ec* (2016-09-20 03:34 UTC)
Platform Info:
  System: Darwin (x86_64-apple-darwin16.0.0)
  CPU: Intel(R) Core(TM) i7-4980HQ CPU @ 2.80GHz
  WORD_SIZE: 64
  BLAS: libopenblas (USE64BITINT DYNAMIC_ARCH NO_AFFINITY Haswell)
  LAPACK: libopenblas64_
  LIBM: libopenlibm
  LLVM: libLLVM-3.7.1 (ORCJIT, haswell)

julia> f(x::Complex{Float64}) = exp(im*(1.0/x+x-x^3/12.0))::Complex{Float64}
f (generic function with 1 method)

julia> z = 1.0+im
1.0 + 1.0im

julia> @time f(z)
  0.042085 seconds (37.53 k allocations: 1.641 MB, 34.28% gc time)
-0.06858891931151362 + 0.7132409685234511im

julia> @time f(z)
  0.000004 seconds (5 allocations: 192 bytes)
-0.06858891931151362 + 0.7132409685234511im

We try the same thing with the development branch of ApproxFun loaded first, and f is significantly slower and allocates:

julia> using ApproxFun

julia> f(x::Complex{Float64}) = exp(im*(1.0/x+x-x^3/12.0))::Complex{Float64}
f (generic function with 1 method)

julia> z = 1.0+im
1.0 + 1.0im

julia> @time f(z)
  0.037003 seconds (59.92 k allocations: 2.544 MB)
-0.06858891931151362 + 0.7132409685234511im

julia> @time f(z)
  0.000064 seconds (12 allocations: 320 bytes)
-0.06858891931151362 + 0.7132409685234511im

The type-assertion is irrelevant (unless one cares to look at the result of @code_warntype). It would be great for this to be demystified. CC: @yuyichao @tkelman, Julia experts.

The problem is magnified when one vectorizes f.

[yuyichao]

Ref JuliaLang/julia#18465

[MikaelSlevinsky]

Does this mean ApproxFun breaks type-stability of Complex{Float64}?

[yuyichao]

I don't know what exactly happens for either case but they look similar.

[dlfivefifty]

@yuyichao So is the solution to stick to 0.4 until JuliaLang/julia#18465 is resolved?

[yuyichao]

I've heard that it is what a few people does. I didn't hit it myself and I didn't look into it (@vtjnash should have) so I don't really have more to say other than pointing to the issue.....

[MikaelSlevinsky]

This appears not to be an issue in 0.6. Without ApproxFun:

julia> versioninfo()
Julia Version 0.6.0-pre.alpha.173
Commit 93cddfc* (2017-03-18 17:19 UTC)
Platform Info:
  OS: macOS (x86_64-apple-darwin16.4.0)
  CPU: Intel(R) Core(TM) i7-4980HQ CPU @ 2.80GHz
  WORD_SIZE: 64
  BLAS: libopenblas (USE64BITINT DYNAMIC_ARCH NO_AFFINITY Haswell)
  LAPACK: libopenblas64_
  LIBM: libopenlibm
  LLVM: libLLVM-3.9.1 (ORCJIT, haswell)

julia> f(x::Complex{Float64}) = exp(im*(1.0/x+x-x^3/12.0))::Complex{Float64}
f (generic function with 1 method)

julia> z = 1.0+im
1.0 + 1.0im

julia> @time f(z)
  0.070867 seconds (86.66 k allocations: 3.526 MiB)
-0.06858891931151362 + 0.7132409685234511im

julia> @time f(z)
  0.000005 seconds (5 allocations: 192 bytes)
-0.06858891931151362 + 0.7132409685234511im

With ApproxFun, the run-time and allocations are back to normal, but the compile time is higher.

julia> using ApproxFun

julia> f(x::Complex{Float64}) = exp(im*(1.0/x+x-x^3/12.0))::Complex{Float64}
f (generic function with 1 method)

julia> z = 1.0+im
1.0 + 1.0im

julia> @time f(z)
  0.167900 seconds (183.24 k allocations: 7.191 MiB)
-0.06858891931151362 + 0.7132409685234511im

julia> @time f(z)
  0.000004 seconds (5 allocations: 192 bytes)
-0.06858891931151362 + 0.7132409685234511im

[dlfivefifty]

It should also be fixed in 0.5.1

…

Sent from my iPhone

On 20 Mar 2017, at 08:15, Richard Mikael Slevinsky ***@***.***> wrote: This appears not to be an issue in 0.6. Without ApproxFun: julia> versioninfo() Julia Version 0.6.0-pre.alpha.173 Commit 93cddfc* (2017-03-18 17:19 UTC) Platform Info: OS: macOS (x86_64-apple-darwin16.4.0) CPU: Intel(R) Core(TM) i7-4980HQ CPU @ 2.80GHz WORD_SIZE: 64 BLAS: libopenblas (USE64BITINT DYNAMIC_ARCH NO_AFFINITY Haswell) LAPACK: libopenblas64_ LIBM: libopenlibm LLVM: libLLVM-3.9.1 (ORCJIT, haswell) julia> f(x::Complex{Float64}) = exp(im*(1.0/x+x-x^3/12.0))::Complex{Float64} f (generic function with 1 method) julia> z = 1.0+im 1.0 + 1.0im julia> @time f(z) 0.070867 seconds (86.66 k allocations: 3.526 MiB) -0.06858891931151362 + 0.7132409685234511im julia> @time f(z) 0.000005 seconds (5 allocations: 192 bytes) -0.06858891931151362 + 0.7132409685234511im With ApproxFun, the run-time and allocations are back to normal, but the compile time is higher. julia> using ApproxFun julia> f(x::Complex{Float64}) = exp(im*(1.0/x+x-x^3/12.0))::Complex{Float64} f (generic function with 1 method) julia> z = 1.0+im 1.0 + 1.0im julia> @time f(z) 0.167900 seconds (183.24 k allocations: 7.191 MiB) -0.06858891931151362 + 0.7132409685234511im julia> @time f(z) 0.000004 seconds (5 allocations: 192 bytes) -0.06858891931151362 + 0.7132409685234511im — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.