On Mon, Jan 19, 2009 at 8:09 AM, Mathias Gaunard < mathias.gaunard@ens-lyon.org> wrote:
Patrick Mihelich wrote:
uBlas doesn't do any explicit vectorization.
Than that's an implementation issue. It could be solved by submitting a patch.
Well, that's a pretty glib statement. I wonder if you have any idea what such a "patch" (I would say rewrite) would entail? Eigen2 was written from the ground up around the packet concept, which allows it to portably leverage SIMD instruction sets when available. uBlas I think would need to be restructured similarly to take advantage of SSE, Altivec, etc. in a manageable way. This is certainly doable, but there's a reason no one has actually done it yet: writing SIMD-enabled code for multiple architectures is currently really unpleasant. A SIMD wrapper library would solve this unpleasantness, and IMO is a prerequisite for a SIMD-accelerated uBlas.