So.. I've done some digging towards what we could show.
Both NumPy and SciPy have benchmarks, which we can use to generate results with all the different backends. They use a tool called Airspeed Velocity for this (EasyConfig for 2025b coming...), to compare results of different revisions. Though it is not really meant for just a single version, we can still use this to generate a few nice overviews with different systems and architectures.

I've tried that with NumPy, replacing their default config with:

diff --git a/benchmarks/asv.conf.json b/benchmarks/asv.conf.json
index 7c7542b1ec..3d9b94a26f 100644
--- a/benchmarks/asv.conf.json
+++ b/benchmarks/asv.conf.json
@@ -46,9 +46,9 @@
     // list indicates to just test against the default (latest)
     // version.
     "matrix": {
-        "Cython": [],
-        "build": [],
-        "packaging": []
+        "env": {
+            "FLEXIBLAS": ["NETLIB", "AOCL_MT", "IMKL", "IMKL_SEQ", "BLIS", "OPENBLAS"],
+        },
     },

     // The directory (relative to the current directory) that benchmarks are

Then, I ran the linalg benchmarks with:

$ asv run --python=same -b "bench_linalg" --set-commit-hash bc5e4f811db9487a9ea1618ffb77a33b3919bb8e
$ asv publish
$ asv preview

This gives me this (on an AMD Ryzen 7800X3D):

A blog post could give a short overview of FlexiBLAS and why it exists, shortly introduce asv so that users understand what we're using (and I think this tool could be interesting for some developers in general...) and then present results of different architectures, and also mention that results vary on the test one is running.

post on changing BLAS/LAPACK library at runtime via FlexiBLAS #12

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions