[Numpy-discussion] how to run the tests.
Thu Apr 5 15:28:42 CDT 2007
Keith Goodman wrote:
> On 4/5/07, Robert Kern <email@example.com> wrote:
>> There have been significant improvements to the Core
>> Duo 2 code in ATLAS releases [snip]
> What kind of speed up are people seeing with Core 2 Duo aware ATLAS?
>From the announcements on math-atlas-devel:
3.7.17 is out. It's mainly for Core2Duo. I've got a new double precision
gemm kernel that bumps up performance to roughly 77% of peak for large
GEMM. This kernel also seems faster on Pentium-4 chips, but I haven't
updated any other architecture defaults. I've got a few more experiments
that might speed up the Core2Duo kernel, and I'll wait until the new kernel
is more set before updateing other defaults.
Anyway, if you have a Core2Duo, grab this release for a noticable double
Well, there's a couple of days I'll never see again. However, the new config
can finally (at least for me) successfully build 32/64 bit libs when you
override the compiler default on the configure line (using -b 32 or -b 64).
I think the last of the assembly needed to build the 64-bit Core2 defaults
under OS X have been adapted as well.
I've just released 3.7.29. The big deal is that it has new algorithms
(for real precisions only) that essentially doubles the performance (at least
on the Core2Duo, but should work many places) for long skinny matrices.
I.e., shapes like M=56, N=56, K=200000.
OK, it's been a long time coming, but 3.7.30 is out. In it, I have finished
adding the just-in-time copy techniqe to complex GEMM. Some long-K speedup
is possible on some platforms. On some architectures, asymptotic complex
GEMM performance goes up as well. Essentially, if you are on a platform where
real GEMM is faster than complex GEMM, there's a good chance you'll see a
speedup here, though you may have to recompute [c,z]Xover.h to make it
happen (I've updated only the Core2Duo [6% ZGEMM speedup] and Athlon-64
[no speedup] arch defs).
"I have come to believe that the whole world is an enigma, a harmless enigma
that is made terrible by our own mad attempt to interpret it as though it had
an underlying truth."
-- Umberto Eco
More information about the Numpy-discussion