Compiler optimization for Geode/Floating Point pipeline

[Adrian Chadd]

(I'll stay lurking for the most part, I thought I'd just put in my 2c)

On Mon, Oct 29, 2007, Rob Savoye wrote:
> On Mon, Oct 29, 2007 at 03:33:22PM -0700, Brian Carnes wrote:
> 
> > What aspects of this issue/request for help are still open?  I'll go take
> > a look at the OLPC build system tonight to see what is being used (late
> > versions of GCC do have some Geode -mtune/-march modes), but would love to
> > be hooked into whatever project is addressing this
>  
>   There is more info on geode optimizations for GCC and glibc at:
> http://wiki.gnashdev.org/wiki/index.php/Building_OLPC_Tools. There is some
> minor tweaking of the optimizer for the geode, it's mostly treated as a
> pentium class processor. While it could be better, with these patches it's
> still better than the default of pentium. Anyway, I'd love to work with folks
> that are also into shaving cycles whereever possible.

Since you are lucky and target a fixed platform you could go the SGI/MacOSX
type route - provide the math functions above in a set of libraries and
then provide different libraries for CPU types - one for generic, one
optimised for the current processor of choice, etc.

You may have more luck with this over trying to rely on the compiler. There
may also be hand-optimised versions of general math functions, if they're
not included in the applications already.

(Of course, the first place to start is profiling..)

2c



Adrian