Re: [Bug-gnubg] gnubg on mac os x

[Joern Thyssen]

On Fri, Jan 24, 2003 at 07:02:13PM +0100, Jim Segrave wrote
> On Fri 24 Jan 2003 (18:00 +0100), Olivier Baur wrote:
> > - add vector-computing for use with G4 processors' built-in superscalar 
> > unit (known as "Velocity Engine" or "AltiVec") -- I have noticed that 
> > during rollouts/evaluations, the CPU would spend 70% of its time 
> > evaluating the neural net, which can be basically thought as a matrix 
> > product between the different layers, which in turn is a perfect 
> > candidate algorithm for vectorization and dramatic performance 
> > improvement
> 
> Interesting idea - I wonder how much that might help for those who
> have vector hardware.

Other hardware may benifit from a rewrite if the neural net evaluations
are expressed in "standard" routines like QGEMM or DAXPY from BLAS:
<URL:http://www.netlib.org/blas/>.

For many platforms these exist in hand-coded assembler optimised for the
particular processor's cache etc. For other platforms it'll be trivial
to provide our own implementation.

But the gain maybe extremely small if most of the time is spend in the
sigmoid function rather than the vector part.

Jørn