[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
R Clint Whaley wrote:
> Since you are not a fan of compiling from source, I have posted the 256K L2
> PIII ATLAS library at:
> This guy peaks at around 78% of peak for DGEMM instead of the 72% that the
> 512K library you used for your ATLAS timings gets. My guess is this should
> make a big difference, particularly in the large LU timings . . .
> Let me know when you update your ATLAS timings; I'll scope out ITX timings
> myself after the forthcoming ATLAS release when I have a little more time,
> and let you and Greg know the results.
> P.S.: GPL is the Gnu Public License; you mention using it yourself for
> some forthcoming code, but I think you may want to examine the license
> a little more carefully before saying that again :) You seem to be indicating
> you want to keep your assembler core binary-only, which is illegal if you
> are GPL-ing the source. You may be able to get around this by packaging
> them in two seperate libs and and using the LGPL instead of the GPL, but
> all the people who would be glad you are using GPL would be similarly
> outraged that you violate its spirit in this way, and would probably raise
> a stink, if you care about such things.
I see no significant change in the difference between ATLAS and ITXGEMM when
linking with this new version. Notice that we consider "peak" to be much less
important than how the routines perform for all matrix sizes.
(See our web page under "peak is for sissies").
I will do an exhaustive run tonight. Do you prefer m,n,k=40:40:520 or
or what increments favor this version of ATLAS?
As for licensing: this is an issue for people who are interested in packaging
marketing purposes. We are merely interested in verifying some theoretical
results. The fact that we make our resulting libraries available to the
is just a bonus for the community, regardless of how the license reads.