Many people have wanted to know more about how to reduce latency and the internal details of precompilation. I’ve just added a variant of the discussion below to SnoopCompile’s docs ( Snooping on inference: @snoopi · SnoopCompile ), but since it seemed to address an unmet need I decided to cross-post …

Thank you for this nice explanation. I still have a couple of questions: How do you identify functions that need to be looked at in the first place? I have tried snoopcompile a while ago and didn’t find it so easy to navigate the long lists of methods to find critical pieces. Ideally I’d want a so…

[image] jules: You said in the example that fundamentally, vcat was the problem. How did you determine that? None of the methods in the list is vcat. I think in this case it’s simply because some cat methods were invalidated and the only place in the function in question that’s dealing with c…

[image] jules: Ideally I’d want a sorted list of functions with their respective latency contributions so I can work from the top down with maximum efficiency. That’s exactly what @snoopi gives you: a list of (time, methodinstance) tuples, sorted by time. You should start at the end of the li…

Ok yes, the list has the times for each individual call. However, the culprit is not one of those, but a different function above them. So, it might make more sense to sum the times of each component to see if some parent method is the actual problem? Like a flame graph but not for compute time, but…

By the way, it’s also worth emphasizing that it’s much easier to figure these things out if you @snoopi on single calls rather than, say, your package’s entire test suite. What I typically do is use it on runtests.jl to get a sense for the worse offender(s) but then restart my session and tackle the…

Thanks for all your efforts to helps us tackle this despairing problem (spent hours in it and not managed to shave a single fraction of a second). It happens that I also have those annoying cats precompile(grdimage, (GMT.GMTgrid{Float32,2},)) julia> G = gmt("grdmath -R-15/15/-15/15 -I0.5 X Y HYPO…

Sorry it’s proving frustrating. I get You don’t seem to have GMT installed and I don’t currently install it automatically, and installing it looks like more work than I’m willing to do, so I’m afraid I’ll have to limit myself to generic remarks. First, the RadioMenu must have at least two optio…

[image] tim.holy: Sorry it’s proving frustrating. I get You don’t seem to have GMT installed and I don’t currently install it automatically, and installing it looks like more work than I’m willing to do, so I’m afraid I’ll have to limit myself to generic remarks. Yes, I understand that bu…

[image] joa-quim: Cthulhu it’s quite puzzling. The REPL just goes crazy. You saw the link to the video above, right?

Understanding precompilation and its limitations (reducing latency)

Internals & Design

joa-quim December 28, 2020, 7:00pm 12

@sprintf seems to contribute in large to the inference issues. See first post of.
If I replace the @sprintf call here by a call to sprintf in the C lib, it shaves me 1 sec in Time to First Plot.

Topic		Replies	Views
ANN: New package SnoopPrecompile Community package , ttfx , latency	32	5064	September 30, 2022
Julia precompilation limits or are there really any? General Usage precompilation , snoopcompile , jit	20	3518	June 20, 2022
Help reducing compilation and inference time Performance compilation , latency	71	1663	July 11, 2024
Precompiling non-inferrable calls Performance compilation , precompilation , snoopcompile , ttfp , ttfx	2	547	April 23, 2022
Struggling to understand how to use PrecompileTools Performance question	25	515	April 26, 2025

Understanding precompilation and its limitations (reducing latency)

Related topics