@graeme-a-stewart has this faster alternative of Base.findmin: julia> using LoopVectorization, Chairmarks julia> function fast_findmin(dij, n) best = 1 @inbounds dij_min = dij[1] @turbo for here in 2:n newmin = dij[here] < dij_min best…

[image] matthias314: @b This reports minimal time, can you try be

The medians suggest that naive_findmin is even faster than what I said before. Do you get something similar on your machine? julia> @be sample(basic_findmin) Benchmark: 57 samples with 1 evaluation min 1.035 ms (2800 allocs: 2.715 MiB) median 1.290 ms (2826 allocs: 2.762 MiB) mean 1.644 ms …

alright, so I think the more realistic use case is: graeme-a-stewart/JetReconstructionBenchmarks.jl/blob/6da3c419382e639dc543d20a3ea0e1dfc8a2ca22/src/findmin-test.jl#L57-L59 @benchmark for j in array_size:-1:1 fast_findmin(x, j) end …

[image] matthias314: function naive_findmin(w, N...) x = @fastmath foldl(min, w) i = findfirst(==(x), w)::Int x, i end Since this iterates over the collection twice, it must necessarily be suboptimal.

Of course, that’s why it’s called “naive”. But it vectorizes, so there is a good chance that it’s faster than a method that slowly iterates once. One could iterate over batches of suitable size to see if the minimum changes and then iterate a second time (to find the position) only in case it matter…

Is that like copy/ modifying LLVM output of LV.jl? Or was it entirely manually written…

[image] DNF: Since this iterates over the collection twice, it must necessarily be suboptimal. indeed, I started looking into SIMD.jl and currently have: using SIMD function sample(f::T) where T s, i = 0.0, 0 N = 512 v = rand(N) for n in N:…

Taking my implementation from Renderer/src/RayTracingInOneWeekend.jl at master · Zentrik/Renderer · GitHub , using SIMD function hor_min(x::SIMD.Vec{8, F}) where F @fastmath a = shufflevector(x, Val((4, 5, 6, 7, :undef, :undef, :undef, :undef))) # high half @fastmath b = min(a, x) @fast…

[image] Zentrik: hor_min(x::SIMD.Vec{8, F}) where F is this faster than the SIMD.minimum()?

Faster `findmin` without LoopVectorization.jl

General Usage Performance

Elrod October 26, 2024, 12:37am 11

If you want to avoid LoopVectorization.jl, you could always take FindFirstFunctions.jl’s approach.

C implementation of function being ~4 times faster even absence of allocs

Topic		Replies	Views
Replicate @tturbo performance Performance	23	2564	August 23, 2022
Why is minimum so much faster than argmin? Performance	11	2127	August 23, 2021
`minimum()` 3x - 6x slower than `numpy.min()` Performance question	16	617	November 2, 2024
Min/max swap using SIMD Performance simd	18	412	March 12, 2025
Slow argmin? General Usage	3	127	November 12, 2024

Faster `findmin` without LoopVectorization.jl

Related topics