[Debian-ppc64-devel] Debian for Cell-styled PPCs

[Sven Luther]

On Wed, Apr 04, 2007 at 03:03:35PM +0200, Felix Domke wrote:
> Hi,
> 
> i'm happily using debian-ppc64 on some of my machines. I'm currently
> also working on machines based on the Cell PPU, i.e. an processor which
> is unable to do out-of-order processing of instructions. Because of
> that, "general" PPC code runs on these machines, but very slow.
> 
> ppu-gcc from the Barcelona Super Computing Center is a hacked gcc which
> special optimizations for the Cell PPU and similiar processors (just to
> clarify this: we are not talking about the SPE "vector units", just
> about the main powerpc core). I know that it usually doesn't make much
> sense to compile distributions optimized for a certain CPU, but in this
> case, it makes a serious difference.
> 
> Some benchmark programs show a 10x speed increase when compiled with the
> ppu-gcc, and i think even general purpose code would see a serious speed
> boost.
> 
> Is it possible to create a flavour of debian-ppc64 which is optimized
> for Cell? If nobody wants to do this, can somebody point me to the
> "proper" way to recompile the entire debian distribution using a custom
> compiler?
> 
> The main targets are gaming consoles at the moment, but they are by far
> the cheapest PPC64 machines you can get today, and i'm very interested
> in providing a good debian-based distribution for these systems.

Hi Felix, ...

The problem of the cell cpu, is that it is not as full fledged as the normal
ppc970 and other power4/5/x cpus, and thus gets an approximately 20%
performance hit compared to his bigger brothers, or so i am told.

This would mean recompiling everything for it, with these optimization flags,
which can probably not be done inside debian proper, for space constraints,
altough there may be a chance for that after etch is out, but i would not hold
my breath for it.

I am planning on doing various powerpc rebuilds of debian in the next days,
and among them we could imagine such a debian-cell distribution. Please
recontact me in 15 days or so, were we will have setup a mailing list to
handle this.

The main problematic is the toolchain support for it, which i guess will be
the first thing to change in order to get cell support, as it was when we
needed the biarch support back then. 

Friendly,

Sven Luther