@LaurentPlagne and myself recently noticed the following surprising performances: julia> using BenchmarkTools julia> N = 100_000 julia> x32 = rand(Float32, N); julia> x64 = Float64.(x32); julia> sumExpStd(x) = sum(exp, x) sumExpStd (generic function with 1 method) julia> versioninfo() Julia Versi…

Julia exp does not emit SIMD instructions, when you use SLEEF, simd is probably used making it much faster. Twice as many Float32 fit into a SIMD register compared to Float64.

See some discussion in this answer and thread [image] Fast logsumexp Performance LLVM. GCC will vectorize log/exp/sin/etc with the appropriate optimization flags, but LLVM needs those in addition to -fveclib=SVML or some other vector library. using SIMDPirates, SLEEFPira…

[image] baggepinnen: Julia exp does not emit SIMD instructions, when you use SLEEF, simd is probably used making it much faster Thanks for your answer. This might very well have to do with vectorization, but there remain a few questions: if we consider the Base.exp-based versions to not be…

[image] ffevotte: I would still expect the Float32 version to go faster (not because of SIMD packs of Float32 being twice as large as packs of Float64, but because it is “simpler” to compute te exponential of a Float32 number) The implementation of exp in Julia is described in this PR ( https:…

Performance of Float32 exponential

General Usage Performance

baggepinnen December 21, 2019, 12:23pm 3

See some discussion in this answer and thread

Topic		Replies	Views
Slow arbitrary base exponentiation, a^b Performance	22	2927	June 14, 2020
Exponentiation of floats Performance question	25	1156	July 26, 2024
Comparing exp() performance on Julia versus numpy Performance	38	2189	August 8, 2022
Base.runtests("math") failures on chips older than haswell New to Julia	12	766	August 7, 2021
Slow exp()? Numerics fast-math	12	2190	August 17, 2018

Performance of Float32 exponential

Related topics