Experimental: precompile functions

[IainNZ]

Test script:

start_t = time_ns()
println("start ", (time_ns()-start_t)/10^9); start_t = time_ns()
using JuMP
println("using ", (time_ns()-start_t)/10^9); start_t = time_ns()
m = Model()
println("model ", (time_ns()-start_t)/10^9); start_t = time_ns()
N=10000
@defVar(m, x[1:N])
println("defva ", (time_ns()-start_t)/10^9); start_t = time_ns()
@addConstraint(m, sum{i*x[i],i=1:N} >= N)
println("addcn ", (time_ns()-start_t)/10^9); start_t = time_ns()

[IainNZ]

Seems to help way more in global scope than in a function, not sure why.

[joehuchette]

I'm pretty sure the printing code takes forever to compile; what happens if you precompile some of that? Or use ; in global scope...

[IainNZ]

Yeah printing takes forever, that was gonna my next challenge if this worked at all.

To be clear, I was running that code as a script (with precompiled JuMP), so its not doing any printing (I also run with coverage to identity what was running).

Most of the benefit came from the addToExpression_reorder, the rest from the parse* methods. The others didn't do anything as far as I can tell.

[joehuchette]

These are v0.4 only

[IainNZ]

Yep

[joehuchette]

Cool, looks like it's helping a lot with the constraint

[mlubin]

Conditional include on VERSION > v"0.4-"?

[IainNZ]

If we merge this, yeah

[IainNZ]

@iaindunning: have you measured a benefit to doing that? unless you use it in sysimg/userimg, precompile doesn't actually save machine code.

slackbot [11:51 PM] 
spell-check "precompile": did you mean "magic"?

----- Today August 15th, 2015 -----

iaindunning [12:24 AM] 
@ihnorton: I'm seeing something

iaindunning [12:24 AM]
https://github.com/JuliaOpt/JuMP.jl/pull/531

GitHub
Experimental: precompile functions by IainNZ · Pull Request #531 · JuliaOpt/JuMP.jl · GitHub
Test script: start_t = time_ns() println("start ", (time_ns()-start_t)/10^9); start_t = time_ns() using JuMP println("using ", (time_ns()-start_t)/10^9); start_t = time_ns() m = Model() println("m...

iaindunning [12:24 AM]
i don't know how to explain it, if thats the case

iaindunning [12:24 AM]
maybe something to do with the generation function

ihnorton [12:29 AM] 
ah, yeah... with so much code generation that makes sense as precompiling should save you the fully expanded versions of everything

ihnorton [12:29 AM]
in lowered form

iaindunning [12:29 AM] 
the timing change was non-trivial, like 2x for the thing i was focused on

iaindunning [12:29 AM]
ahhh ok

iaindunning [12:29 AM]
so code with not involving macros would not likely see the benefit?

ihnorton [12:31 AM] 
jameson was talking about it earlier this evening and said (IIUC) that the benefit was minimal in the code he had profiled, but I think that was inference (which does a lot of things, but thankfully doesn't use @generated yet)

[IainNZ]

I was looking around and saw that PyCall was exploring similar ideas (JuliaPy/PyCall.jl#167)

[IainNZ]

Closing in favour of #541