I have been doing some experiments with the great LoopVectorization for convolution, and I must admit that in spite of my efforts to manually refactor my code, I have been unable to even approach the speed that this package is giving. Consider the following code: using LoopVectorization, OffsetArr…

It looks like the bigger problem is that the function which does the lion’s share of work (_turbo_!) is not inlined. As such, you’d have to use Cthulhu.jl to get to the Julia or LLVM IR for most of the loop code. That doesn’t make the massive amount of IR any easier to sift through, but it’s more i…

[image] Marco_Lombardi: kernel = OffsetArray(rand(31,21)) For completeness, we should mention that for so large kernels, it is advantageous to compute the convolution in the Fourier domain. Of course, your intention here is to test loop vectorizaton.

@pitsiannis true, FFT is faster (not by much: a factor ~30% over LV on my laptop), but my point was to understand what is LV doing on my back.

@photor there is a deprecation warning, which is part of my problem… I am currently using Julia v1.10 to avoid any possible issues. Incidentally, I hope someone knowledgeable enough takes over the maintenance of LV, as it is really a fantastic package.

No that does not help too much, as LV is doing a lot of work in internal (generated) functions, and it is really (at least for me) very difficult to follow all these internal routines. Anyway, if you wish to try and find any hints, I would really appreciate it. I tried @code_llvm instead on the LV-…

An intermediate-level (no pun intended) tool between @macroexpand and @code_llvm would be @code_warntype/@code_typed. Those will include the IR all those generated functions output. IIRC LoopVectorization can do some fancy transforms like loop reordering, and these reflection macros would let you ca…

Actually LV does a lot of deep magic and does not just generate Julia code. It compiles parts of the code by itself AFAIK. So @code_warntype and the like are not helpful. Here is an excerpt from the function from above (with some linebreak added by me): julia> @code_typed f0!(rand(10,10), rand(10,1…

Yes, indeed. From the LV manual, I understand that one of the first passes performed by LV is to convert the (nested) loops into an internal representation based on these OperationStruct that we see in the excerpt that @abroemer provided. So, unless one knows or can figure out what LV internally is,…

Thank you @ToucheSir , I was reaching independently the same conclusions! And I now believe SIMD is actually used with LLVM intrinsics within _turbo_! (for example, one of the first steps is to call LoopVectorization.pick_vector_width, which computes the SIMD vector size for my processor). I will now…

Thank you @RoyiAvital , correct me if I am wrong, but these two posts just conclude that @turbo gives the optimal performance, without further analysis of the reasons for that (i.e.: what is happening behind @turbo?).

Experiments with LoopVectorization and convolutions

General Usage Performance

RoyiAvital November 29, 2024, 6:13pm 14

You may find Optimizing Direct 1D Convolution Code and Optimizing Direct 2D Convolution Code useful.

Topic		Replies	Views
Tullio seems two times slower than basic LoopVectorization Performance question , tullio , loopvectorization	3	1113	April 10, 2022
LoopVec, Tullio losing to Matrix multiplication Performance	9	834	July 25, 2024
How to make LoopVectorization work with computing Euclidean distance for every column? New to Julia tullio , loopvectorization	13	1215	August 21, 2020
[ANN] LoopVectorization Package Announcements	157	24344	May 27, 2020
ANN: LoopVectorization 0.12: multithreading and better handling of discontiguous memory accesses Performance	16	2286	March 17, 2021

Experiments with LoopVectorization and convolutions

Related topics