[Numpy-discussion] Matlab/Numeric/numarray benchmarks

[Simon Burton]

Todd Miller wrote:

>Hi Simon,
>
>I found a benchmark bug which explains the performance difference in +. 
>Here are my times with the modified benchmark (Python-2.4 gcc version
>3.2.2 20030222 (Red Hat Linux 3.2.2-5) on 1.7 GHz P-IV w/ 2G):
>
>numarray + : 0.0540893316269
>numarray matrixmultiply : 16.9448821545
>numarray eigenvalues : 9.67254910469
>
>Numeric + : 0.0653991508484
>Numeric matrixmultiply : 33.0565470934
>Numeric eigenvalues : 9.44225819111
>
>So, for large arrays with a simple to install / built-in linear algebra
>system,  numarray is doing just fine.  
>
>Looking at your results,  I think you may have been comparing numarray
>built using a built-in blas_lite versus Numeric using ATLAS.  There,  I
>think numarray *is* behind but that is fixable with some effort.  The
>key is porting and integrating the Numeric dotblas package with
>numarray.  I've been looking at that some today...  err, yesterday,
>apparently I forgot to hit "send".
>  
>

Wow, those results look great, Todd.

I have double checked my install. However, the numarray multiply is 
still x10 slower.

This is set in addons:
 lapack_libs = ['lapack', 'f77blas', 'cblas', 'atlas', 'blas']

and, at runtime, python has loaded:
40771000-40cb6000 r-xp 00000000 00:0c 783245     
/usr/lib/atlas/liblapack.so.3.0
40cb6000-40cb9000 rw-p 00545000 00:0c 783245     
/usr/lib/atlas/liblapack.so.3.0
40cb9000-40dbd000 rw-p 00000000 00:00 0
40dbd000-40dd7000 r-xp 00000000 00:0c 783242     /usr/lib/libf77blas.so.3.0
40dd7000-40dd8000 rw-p 00019000 00:0c 783242     /usr/lib/libf77blas.so.3.0
40dd8000-40df7000 r-xp 00000000 00:0c 783241     /usr/lib/libcblas.so.3.0
40df7000-40df8000 rw-p 0001e000 00:0c 783241     /usr/lib/libcblas.so.3.0
40df8000-4110b000 r-xp 00000000 00:0c 783240     /usr/lib/libatlas.so.3.0
4110b000-4110f000 rw-p 00312000 00:0c 783240     /usr/lib/libatlas.so.3.0
4110f000-41454000 r-xp 00000000 00:0c 783244     
/usr/lib/atlas/libblas.so.3.0
41454000-41458000 rw-p 00345000 00:0c 783244     
/usr/lib/atlas/libblas.so.3.0
41458000-41472000 r-xp 00000000 00:0c 17227      /usr/lib/libg2c.so.0.0.0
41472000-41473000 rw-p 0001a000 00:0c 17227      /usr/lib/libg2c.so.0.0.0
( as well as lapack_lite2.so )

So I assumed that since the eigenvalues came in fast ATLAS was alive and 
well.
Also, the above libs are exactly what Numeric uses:
4033b000-40880000 r-xp 00000000 00:0c 783245     
/usr/lib/atlas/liblapack.so.3.0
40880000-40883000 rw-p 00545000 00:0c 783245     
/usr/lib/atlas/liblapack.so.3.0
40883000-40987000 rw-p 00000000 00:00 0
40987000-409a6000 r-xp 00000000 00:0c 783241     /usr/lib/libcblas.so.3.0
409a6000-409a7000 rw-p 0001e000 00:0c 783241     /usr/lib/libcblas.so.3.0
409a7000-409c1000 r-xp 00000000 00:0c 783242     /usr/lib/libf77blas.so.3.0
409c1000-409c2000 rw-p 00019000 00:0c 783242     /usr/lib/libf77blas.so.3.0
409c2000-40cd5000 r-xp 00000000 00:0c 783240     /usr/lib/libatlas.so.3.0
40cd5000-40cd9000 rw-p 00312000 00:0c 783240     /usr/lib/libatlas.so.3.0
40cd9000-40cf3000 r-xp 00000000 00:0c 17227      /usr/lib/libg2c.so.0.0.0
40cf3000-40cf4000 rw-p 0001a000 00:0c 17227      /usr/lib/libg2c.so.0.0.0
40cf4000-40cf7000 rw-p 00000000 00:00 0
40cf7000-4103c000 r-xp 00000000 00:0c 783244     
/usr/lib/atlas/libblas.so.3.0
4103c000-41040000 rw-p 00345000 00:0c 783244     
/usr/lib/atlas/libblas.so.3.0


Any ideas ? It's not even clear to me where the matrixmultiply is taking 
place. I couldn't find it in my lapack_lite2.so even though there seems 
to be a lite version of dgemm in the source. But then dgemm is not 
referenced anywhere else in the numarray source. Comeing from the other 
end, I traced matrixmultiply to an _ipFloat64. But then the trail went 
cold again :) Flumoxed.

Simon.



-- 
Simon Burton, B.Sc.
Licensed PO Box 8066
ANU Canberra 2601
Australia
Ph. 61 02 6249 6940
http://arrowtheory.com