I wrote a naive convolution function for some personal purpose and tested it against Python numpy’s function (which does not use FFT). On my computer the Julia script is about 7x slower than Python numpy’s version. Could anyone help me understand why it’s slow or how to accelerate the code? Here i…

The naive loop is faster than numpy, just make sure it gets SIMD’ed: function naive_convol_full!(w, u, v) if length(u) < length(v) return naive_convol_full!(w, v, u) end n = length(u) m = length(v) for i = 1:n+m-1 s = zero(eltype(u)) @simd for j = max(i-m…

Try to add @inbounds or @simd calls to the hot loops. Currently your implementation is checking if the indices are valid and that slow things down for example.

Also, it looks like your current implementation was copied from a language with 0 indexing. function naive_convol_full!(w, u, v) if length(u)<length(v) return naive_convol_full!(w, v, u) end n=length(u) m=length(v) @inbounds @fastmath for i=1:n+m-1 w[i]=zero(eltype(u)) for j=max(i…

why would numpy version be this double for loop and not fft based (for example) ?

[image] romainvieme: Python numpy’s function If you drill down in the Numpy source code, I think its convolution ends up at multiarraymodule.c:_pyarray_correlate . It looks like it just does a sequence of dot calls, which probably is some optimized SIMD-based function, and possibly uses thre…

Thanks, didn’t know this trick ! Doesn’t make much of a difference though… function naive_convol_full!(w, u, v) if length(u)<length(v) naive_convol_full!(v, u, w) else n=length(u) m=length(v) @simd for i=0:n+m-2 k=max(i+1-m, 0) l=min(n-1,i) @inbounds w[i+1]=zero(eltype(u)) @…

Sorry, not copied, I wrote it myself :sweat_smile: Thanks for the code though ! Just ran it, I gained a few milliseconds but not enough to make the difference julia> function naive_convol_full!(w, u, v) if length(u)<length(v) return naive_convol_full!(w, v, u)…

I don’t actually know, but that’s what’s said in the docs !

I think @stevengj answer goes into that direction

Thanks, I couldn’t figure out how to use @turbo from LoopVectorization but I got a 3x speedup by replacing the innermost loop by a dot product (following your suggestion). That’ll have to do for now ! The numpy call is still about 2.2x faster than the Julia one though. Thank you all for your help —…

romainvieme , without actually checking your code, may I point you to a discussion which followed this post: [image] General questions from Python user Performance In the course of my weekend procrastinations I contrieved an example to check my statement: some home-made im…

Performance of naive convolution against Python Numpy

New to Julia

Eben60 February 1, 2022, 10:20pm 11

romainvieme, without actually checking your code, may I point you to a discussion which followed this post:

Topic		Replies	Views
General questions from Python user Performance	59	4652	March 8, 2021
Numpy 10x faster than Julia ?! What am I doing wrong ?! [solved - julia faster now] Performance question	37	11320	October 15, 2019
Vector addition in Julia slower than numpy in Linux Performance	21	1455	May 1, 2020
Why is this simple function twice as slow as its Python version Performance question	97	4955	April 12, 2021
Comparing Python, Julia, and C++ Performance broadcast , python	21	34328	November 1, 2018

Performance of naive convolution against Python Numpy

Related topics