I have a very heavily optimized calculation that runs for weeks. It is making heavy use of SVectors, ForwardDiff.jl for AD, and various other things that tax the compiler, but in turn, it is very fast when compiled. I spent a lot of time making it inferred and mostly allocation-free, so typical @ti…

I’m writing analysis tools right now; I think one important tool will be a scatter plot of “runtime vs compile time.” It might help direct attention to methods that just aren’t worth specializing. See Add tools for analysis of specialization by timholy · Pull Request #169 · timholy/SnoopCompile.jl ·…

https://timholy.github.io/SnoopCompile.jl/stable/snoopc/#macro-snoopc-1 SnoopCompile sounds like it could help you. There’s also possibilities to supress reporting of functions with fast inference, so it should record compile times, but I don’t exactly know how to filter for the parts that take the…

There are big changes afoot in the SnoopCompile world; if you’re really brave see https://github.com/JuliaLang/www.julialang.org/pull/1093 (the preview is here: https://julialang.netlify.app/previews/pr1093/blog/2020/12/package_latency/ ) and try to follow along. Some names have already changed, and …

Much of the focus is on inference, but LoopVectorization’s test suite is about 5% inference and 75% code gen: 640.984369 seconds (379.33 M allocations: 25.899 GiB, 1.58% gc time, 99.55% compilation time) Test Summary: | Pass Total LoopVectorization.jl | 20429 20429 1572.714641 seconds (1.…

Yes, the emphasis is definitely mostly on inference, and you’re right that for some packages inference is a minor component of the time. (Does LoopVectorization’s src/precompile.jl shave much time off? Once it’s precompiled, inference doesn’t have to run again, so the fact that it’s only 5% might be…

Hmm: 654.318623 seconds (380.08 M allocations: 25.946 GiB, 1.55% gc time, 99.56% compilation time) Test Summary: | Pass Total LoopVectorization.jl | 20429 20429 1610.347997 seconds (1.63 G allocations: 108.695 GiB, 2.21% gc time, 99.17% compilation time) ROOT : 0.07 …

Have you seen @aviatesk 's JET.jl https://twitter.com/kdwkshh/status/1337639010533928960 ? It also detects type instabilities. Is there scope to share work/code?

I have, and while I haven’t played with it, it looks wonderful. But it’s a completely different approach to “collecting the data” (JET uses https://github.com/JuliaLang/julia/pull/35845 whereas SnoopCompile uses https://github.com/JuliaLang/julia/pull/37749 ), and so there are going to be some differ…

Is there an easy way to make Julia forget all the work it put into inference and compilation? Restarting the process every time I want to measure things is a minor inconvenience.

For the curious: 90% of excess inference time in my code can be reduced to an (over|ab)use of SArrays. Here is an MWE (all latest package versions, Julia 1.6) using StaticArrays, ForwardDiff, LinearAlgebra, BenchmarkTools # SETUP _dual(x, ::Val{N}) where N = ForwardDiff.Dual(x, ntuple(_ -> x, Val(…

BTW, the bottom line is that the compilation time of the original problem went from 2800 to 7 seconds after working through the code with @snoopi_deep. Yes, you are reading that right. And runtime got 3x faster as an incidental bonus, though it was not targeted. (There is, of course, a downside: I …

Understanding and optimizing compiler time (just a bit)

General Usage

tim.holy January 5, 2021, 3:26pm 16

Topic		Replies	Views
Profiling compilation/inference Performance inference	10	1511	August 24, 2020
Help reducing compilation and inference time Performance compilation , latency	71	1631	July 11, 2024
22 seconds to 3 and now more: Let's fix all of the DifferentialEquations.jl + universe compile times! Performance precompilation	6	5107	August 17, 2021
Understanding precompilation and its limitations (reducing latency) Internals & Design precompilation	13	2769	December 31, 2020
Help diagnosing slow compilation General Usage compilation , profiling	4	217	August 12, 2024

Understanding and optimizing compiler time (just a bit)

Related topics