Hi all, I’m rewriting a function from R to Julia that reduces an array across the 4th dimension via LogSumExp, and I’m currently trying to optimise the speed. I’ve made a bit of progress through the Tullio package, but I’m sure there is more speed to be got. Here are my attempts so far with toy da…

Hi @ericphanson - thanks for letting me know about that package I was unaware. So I just tested it too - relatively slow I’m afraid: julia> @btime logsumexp($A, dims = 4); 87.023 ms (12 allocations: 36.62 MiB)

[image] dlakelan: The 4th version using tullio for everything seems like it’s quite good. fast and succinct. I’m not sure you’re going to get much better. Ah that is surprising its basically my second time writing a Julia function (I’m familiar with numpy.einsum in Python though so that helps…

You could very easily get a factor of 2 to 3 by using Float32, but that may not be accurate enough for some use cases.

It might be. I’ve unit testing set up on the R side so if it causes issues I should detect it straight away. How do I do that is it as simply as changing the array declaration code to Float32 ?

[image] jroon: overhead of moving data from R to Julia Yes, but you clearly are getting the hang of Julia. Perhaps you can move somewhat more of your computation out of R. Ive been using R since before version 1, back in 1999 or so. And while it’s an impressive piece of software, it doesn’t …

[image] jroon: I think then I’ll have to stick with my R version. Unless there is more improvement to be found here, the overhead of moving data from R to Julia which isn’t captured in my R benchmarks above that erases small benefits In general, you won’t get much benefit from Julia from movi…

[image] dlakelan: Yes, but you clearly are getting the hang of Julia. Perhaps you can move somewhat more of your computation out of R. Ive been using R since before version 1, back in 1999 or so. And while it’s an impressive piece of software, it doesn’t hold a candle to Julia’s elegance and po…

[image] jroon: its a short 7 line function then perhaps you can post the R code here that you’re trying to replace? Also maybe the context in which this function is called? Perhaps what’s needed is to replace the next higher up level of the computation?

Sure I can post the R code for the function - although it may not be very readable because its written for speed. But the context is this sits in nested loops. I don’t want to convert any higher up stuff at present time, because its complicated. Besides this function is the slow part and marginal ga…

Try adding using LoopVectorization before defining your functions. Without: julia> A = rand(20, 100, 40, 5); julia> @btime logsumexp4D1($A); 4.307 ms (39 allocations: 5.49 MiB) julia> @btime logsumexp4D2($A); 2.129 ms (270 allocations: 2.46 MiB) julia> @btime logsumexp4D3($A); 995.302 μs …

Fast LogSumExp over 4th dimension

General Usage Performance

stevengj July 7, 2021, 2:05pm 9

In general, you won’t get much benefit from Julia from moving a single vectorized function to Julia. See also this post: Comparing Numba and Julia for a complex matrix computation - #3 by stevengj

Topic		Replies	Views
Speeding up my logsumexp function Performance performance , array , tullio , loopvectorization	35	5339	October 21, 2022
How to improve performance of sum() Performance question , fast-math , tullio , loopvectorization , sum	19	5017	July 19, 2021
Tullio seems two times slower than basic LoopVectorization Performance question , tullio , loopvectorization	3	1113	April 10, 2022
Fast 4D argmax Performance tullio	26	2277	April 6, 2021
Understanding Tullio performances Performance tullio , jax	12	547	June 10, 2025

Fast LogSumExp over 4th dimension

Related topics