Vector Operations

Michel Bardiaux mbardiaux@peaktime.be
Fri, 11 Jul 2003 11:02:14 +0200


Francois G. Dorais wrote:
> I was wondering if there has ever been interest in adding optimized basic
> vector operations to gmp. It seems to me that the performance of vector
> operations on mp types could be greatly improved by low level
> optimizations especially on vector and other parallel architectures.
> 
> Surely something like an optimized BLAS implementation for mpf/mpfr
> would be interesting for those who use these types extensively. I for one
> would appreciate similar routines for mpz and maybe mpq. A direct
> translation of BLAS wouldn't be very appropriate for mpz and mpq since
> since stability considerations are very different, but something very
> similar would do.
> 
> However, it may be that the design of such optimized routines is not as
> "straightforward" as for the basic types because of the greater variety of
> architectural considerations to take care of? It could also be that such
> extensions don't fit in the "spirit" of gmp for some reason or another?
> 
> In any case, I would like to hear what people have to say on the interest
> and feasibility of such extensions to the gmp family.
> 
> Thanks,
> F. G. Dorais
> 

I'm no expert on the internals of gmp, but it seems to me efficient 
vectorization/parallelization would be possible only if all components 
of the vectors have the same number of limbs. Am I correct?

-- 
Michel Bardiaux
Peaktime Belgium S.A.  Bd. du Souverain, 191  B-1160 Bruxelles
Tel : +32 2 790.29.41