Hi there, I am in need to optimize as much as possible the function I will present here. I moved it from C to Julia for convenience since the operations around it become much easier but I am getting significantly worst performance in the performance critical function: using SpecialFunctions utili…

Profiling shows that findmax in Julia does a lot more work then the simple loop for e.g. nan handling… Translating the C version 1:1 without findmax leads for me to a 12x speedup: function assignment_probabilities_sorted_gauss!(utilities, cutoffs, mean, stddev) last_prob = 1.0 expectedU = …

The next thing I would investigate is the dependence of the performance on the compiler flags, on both sides. (And maybe expand the findmax into a loop as in the C code)

Oh wow… this is surprising since I also found findmax to be the largest bottleneck in the profiling and then I compared the performance of the C and Julia findmax parts and they performed the same. C_code = """ #include <stddef.h> #include <math.h> double find_max(size_t n, double *X) { double…

And inbounds there doesn’t make a difference? ( @sdanisch used it)

not in my benchmarks. just tried again. it seems like the speed-up is coming from the findmax circumvent.

temp = rand(500) function naive_findmax(w) x = @fastmath foldl(max, w) i = findfirst(==(x), w)::Int x, i end julia> @btime naive_findmax($(temp)) 507.223 ns (0 allocations: 0 bytes) (0.9960969077821132, 422) Thanks a lot! It does make a significant differen…

it’s a shame really, because this is on average looping over the array 1.5 times. Concretely, we can improve this by having @fastmath findmax(w) to ignore IEEE754 NaN rules. Related if you only want value not the index: Move all platforms to use llvm.minimum/llvm.maximum for fmin/fmax by gbarald…

[image] miguelborrero: Oh wow… this is surprising since I also found findmax to be the largest bottleneck in the profiling and then I compared the performance of the C and Julia findmax parts and they performed the same. [image] miguelborrero: But you are right and you’re version gets the…

[image] jling: we can improve this by having @fastmath findmax(w) to ignore IEEE754 NaN rules. We could. Note that the case @fastmath maximum(w) was added in Julia 1.10 . There is also a PR for extrema and some others, not including findmax or argmax.

Hello, can you expand a bit on this? I am a bit surpriced that a 5% slower inner function results in almost 500% slower outer function…

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

General Usage Performance

jling March 3, 2025, 1:21pm 13

there’s a shorter way to write faster (but no NaN handling) findmin and findmax: Faster `findmin` without LoopVectorization.jl - #3 by matthias314

Topic		Replies	Views
Performance of findmax vs. raw loop General Usage	12	921	July 19, 2020
Faster `min`,`max` for `Float64\|32` Internals & Design performance	38	2386	August 17, 2021
Poor performance due to memory allocations? Performance memory-allocation	17	2967	January 15, 2019
Julia Beginner (from Python): Numba outperforms Julia in rewrite. Any tips to improve performance? Performance benchmark , python , tullio , loopvectorization	56	6712	August 18, 2021
Replicate @tturbo performance Performance	23	2564	August 23, 2022

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

Related topics