[Debian-med-packaging] SSE3 issue with iqtree when trying to enable i386 (Was: Help with new version of iqtree needed)

[Bui Quang Minh]

Hi all,

Thanks for informative discussions! To be clear, what we can do is following:

1. We require to compile with gcc 4.6 at least (the version that supports AVX). We would not support older gcc, as this needs too much work, and the binary code would run much slower than compiled with newer gcc.

2. At runtime, IQ-TREE will detect the instruction set, and decide which kernel to use. Right now IQ-TREE selects between SSE3 or AVX kernel. What is missing currently, is that it does not switch to non-SSE kernel (though this kernel is already there) if the machine does not support SSE3. Thus, what we can do, is to implement this switch.

What do you think about this solution?

Cheers, Minh

> On Mar 11, 2016, at 10:46 AM, Tung Nguyen <tung.nguyen at univie.ac.at> wrote:
> 
> Hi Andreas, 
> 
> Such a flag doesn't exist yet but we will provide it quickly.
> 
> Cheers
> Tung
> 
> 
> On Fri, Mar 11, 2016, 10:03 Andreas Tille <tille at debian.org <mailto:tille at debian.org>> wrote:
> Dear Tung,
> 
> On Fri, Mar 11, 2016 at 08:43:03AM +0000, Tung Nguyen wrote:
> > Then we will provide a compiler flag to remove the SSE requirement. We also
> > have non-sse code for the same functions, so this should be easy.
> 
> Sounds good.  Am I understand you correctly that this is possible even
> right now?  If so, what exact flags do I need to use?
> 
> Kind regards
> 
>       Andreas.
> 
> > On Thu, Mar 10, 2016, 22:02 Andreas Tille <tille at debian.org <mailto:tille at debian.org>> wrote:
> >
> > > Hi,
> > >
> > > On Thu, Mar 10, 2016 at 09:56:53PM +0100, Gert Wollny wrote:
> > > > Hi all,
> > > >
> > > > On Thu, 2016-03-10 at 19:20 +0000, Tung Nguyen wrote:
> > > > > Maybe we will provide a flag to disable the SSE instructions
> > > > > completely for use on i386. However, if the patch from Gert works
> > > > > well then it would be the best solution.
> > > >
> > > > The patch only provides an SSE2 code path for that one SSE3 instruction
> > > > that was forced in the code, so it does not really help to get the code
> > > > compile for the the lowest spec i386 hardware supported by Debian.
> > > >
> > > > Providing a code path without any SSE could also be interesting if you
> > > > think of getting the code ported to other architectures (ARM, PowerPC,
> > > > ...).
> > >
> > > If you could really come up without any requirement for SSE this would
> > > be great since the goal of Debian is to support all those architectures
> > > if possible.
> > >
> > > If you could confirm this attempt I might go for an amd64 package
> > > exclusively for the moment and wide the number of architectures later
> > > in a clean way.
> > >
> > > Kind regards
> > >
> > >        Andreas.
> > >
> > > --
> > > http://fam-tille.de <http://fam-tille.de/>
> > >
> 
> --
> http://fam-tille.de <http://fam-tille.de/>

--
Bui Quang Minh
Center for Integrative Bioinformatics Vienna (CIBIV)
Campus Vienna Biocenter 5, VBC5, Ebene 1
A-1030 Vienna, Austria
Phone: ++43 1 4277 74326
Email: minh.bui (AT) univie.ac.at







-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.alioth.debian.org/pipermail/debian-med-packaging/attachments/20160311/9f3b5d2e/attachment.html>