This is part 1 of a two posts showing how to SIMD the hell out of a dot-product: [image] Modular: Understanding SIMD: Infinite Complexity of Trivial Problems A deep dive into the complexities of optimizing code for SIMD instruction sets across multiple platforms. …

[image] Dan: Maybe benchmarking the cosine-similarity is the right comparison. Since this is also a BLAS-1-like operation, the compiler does a pretty good job on a handwritten loop with @simd. e.g. this SIMD-vectorizes: function mycos(x, y) s = zero(x[begin]) * zero(y[begin]) nx = n…

Interesting post about SIMD dot product (and cosine similarity)

Offtopic

xiaoxi November 30, 2024, 3:05pm 7

Julia-lang blog used to have this kind of post a long time ago.

I would be a pity. I understand that Modular is trying to make the LLVM project easier for non-clang based languages, so it may benefit Julia.

1 Like

Topic		Replies	Views
Dot function General Usage	45	5693	September 26, 2018
Sum performance for Array{Float64,2} elements Performance	13	2525	May 15, 2018
Cosine seems slow Performance	14	1862	November 27, 2019
Innefficient paralellization? Need some help optimizing a simple dot product Performance question , parallel	32	4902	March 28, 2018
@inbounds: is the compiler now so smart that this is no longer necessary? Performance	33	3018	July 16, 2018

Interesting post about SIMD dot product (and cosine similarity)

Related topics