Dear Dirk and Johannes,
I'm appealing to you as the packagers who have shown the greatest
interest in gretl over the years. I think you both know about the
"gretl AVX" issue, but then I'm sure you have lots of things to think
about so let me briefly restate it, as a lead-up to asking your
Since version 1.9.14 of November 2013, gretl's configure script has by
default checked the build host for support for the SSE2 and AVX
instruction sets, and has enabled these instructions if they're
supported. However, there are configure switches available to disable
The current default is fine for individuals building gretl for their
own use, but not necessarily fine for packagers. SSE2 is not much of a
issue, I think -- it has been in Intel CPUs since 2001 and AMD CPUs
since 2003. AVX, on the other hand, can be a problem. It has been in
new Intel CPUs since Q1 of 2011 and AMD CPUs since Q3 of 2011, so
there are still plenty of PCs out there that don't support AVX.
So here's the key point: If you build gretl on an AVX host without
using the --disable-avx flag with configure, you'll get a binary that
will not run on a non-AVX machine (it will crash with "Illegal
instruction" on start-up). And we're seeing this happen in userland.
How do you think we should handle this? Do you reckon we should revise
our configure script to enable AVX only if it's explicitly requested?
Or is it enough if packagers are aware on the issue?
One more question: since we're getting nasty reports from people with
non-AVX machines on both Debian-derived distros and Fedora, we need to
do something about it soon. How would you feel about making available
new non-AVX builds of recent gretl? (It seems that packagers
sometimes put out a second build of given sources, though I don't know
the usual criteria for doing that.)
Show replies by date