Mark

I have a use case for this where I solve a sparse linear system a bunch of times and need to be careful about allocating memory. In case helpful to...

Thanks, @j-fu!

+1 for this -- I've been having the same issue: reading a Linux-generated .feather file using R on Windows causes R to fail.

FWIW, here are a couple of work-arounds ```julia using DataFrames using RCall using RData using Test df = DataFrame( a = collect(1:3), b = [ "a", "b", "c",] ) FILEPATH...

See notebook [logsumexp-speedtests.pdf](https://github.com/JuliaSIMD/LoopVectorization.jl/files/9687244/logsumexp-speedtests.pdf)

HTML rendering of notebook (strip off .txt) [logsumexp-speedtests.jl.html.txt](https://github.com/JuliaSIMD/LoopVectorization.jl/files/9687265/logsumexp-speedtests.jl.html.txt) Pluto notebook (strip off .txt) [logsumexp-speedtests.jl.txt](https://github.com/JuliaSIMD/LoopVectorization.jl/files/9687266/logsumexp-speedtests.jl.txt)

I was able to get a bit faster for Dual numbers by pirating `vexp` and `log_fast`, though the relative speedup (2x) is still less than what `@turbo` does for `Float64`...

This also worked, though wasn't quite as fast ```julia "using base SIMD loops with LoopVectorization tricks" function logsumexp_turbo2!(Vbar, tmp_max, X) m,n = size(X) maximum!(tmp_max, X) fill!(Vbar, 0) @turbo safe=false warn_check_args=false...

Thank you for the quick response, @chriselrod -- really appreciate it. I had a bit of a hard time understanding the code you referenced, but it looked to me that...

Hi -- Thanks for the helpful work-around. I'd also much appreciate the ability to call remote-specific functions. I'm currently computing a likelihood in parallel and am using sets of pre-allocated...