[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: gemv driving me crazy


>I found my problem on the P3/P4. I was using aligned acces on X, and that
>was causing a problem. I thougth the X vector was copied to aligned
>continous storage?

If atlas does a data copy, it copies to an ATL_Cachelen aligned vector.
However, if the user's increment is already 1, and we don't have to apply
alpha, ATLAS leaves the vector alone (so no known alignment) . . .

Is there a big performance hit for non-aligned X?  If so, I guess you could
produce two versions, one assuming well-aligned X, and one not, and do test
and call the correct one.  Later, some #ifdef's could make ATLAS do copying
most of the time (when matrix was large enough) if we determined it was worth
it . . .