[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: Benchmarking GSL
From: |
Ludovic Courtès |
Subject: |
Re: Benchmarking GSL |
Date: |
Mon, 30 Nov 2020 17:18:30 +0100 |
User-agent: |
Gnus/5.13 (Gnus v5.13) Emacs/27.1 (gnu/linux) |
Hi Patrick,
Patrick Alken <patrick.alken@Colorado.EDU> skribis:
> Is there a particular part of the GSL library you want to benchmark?
> Some routines have been optimized to run as fast as possible while
> others have not. The overall focus of GSL algorithms is on robustness
> and reliability, not necessarily speed. So you will probably find quite
> different results depending which algorithms you are benchmarking.
I’m interested in benchmarking routines that benefit from vectorization
(AVX2, AVX512, etc.).
Based on GCC’s ‘-fopt-info-vec’, this could be any of the following
components:
--8<---------------cut here---------------start------------->8---
bspline
bst
cdf
cheb
combination
complex
dht
diff
eigen
fft
filter
fit
histogram
integration
interpolation
linalg
matrix
min
monte
movstat
multifit
multifit_nlinear
multifit
multilarge_nlinear
multilarge
multimin
multiroots
multiset
ode-initval2
ode-initval
permutation
poly
qrng
randist
rng
roots
rstat
sort
spblas
specfunc
splinalg
spmatrix
statistics
vector
wavelet
--8<---------------cut here---------------end--------------->8---
This is quite a long list, but really benchmarks for a couple of these
would already be helpful to me.
For background, I’m currently looking into automatic “function
multi-versioning” where code is generated for several ISA extensions and
the right one is selected at load time¹.
Thanks in advance!
Ludo’.
PS: Please keep me Cc’d as I’m not subscribed.
¹ https://gitlab.inria.fr/guix-hpc/function-multi-versioning