Re: [Gretl-devel] matrix_perf
On Mon, 16 Jun 2014, Artur T. wrote:
Am 16.06.2014 21:10, schrieb Allin Cottrell:
> There's quite a lot going on in gretl CVS right now. Some of the new
> things will require more time to document, here I'll just mention a
> few quick things.
Nice to hear that new features are implemented.
> * There's a new function package on the server, matrix_perf, which
> runs some tests of matrix multiplication performance. It would be
> interesting to hear what results people get from this -- to see if
> there are any surprises. The package requires current CVS or
> snapshot.
I ran it on a 5 year old AMD machine with 2 cores on ubuntu 13.10. See
the output below. Please note that there seems to be an error in the
"printvec" function.
Thanks for testing. Sorry, the error was due to a version mix-up;
there's now a fixed version (0.3) on the server.
<OUTPUT>
? matrix_perf(1234)
dgemm experiment 1, variant 1, speed in Gflops
m n k vanilla openmp netlib
128 128 128 1.0131 1.6128 1.2421
128 128 256 1.0918 1.6047 2.8438
128 128 512 1.1172 1.5613 2.9728
128 128 1024 0.56444 0.90985 2.9641
128 128 2048 0.55022 0.81304 3.0153
result: netlib dominates for mnk >= 4194304
This is interesting. I suspect that what gretl is calling "netlib"
here must actually be an optimized BLAS. Could you post the results
of
ldd /path/to/gretl
along with any information you have about libblas on your system?
Allin