[SciPy-user] Getting the right numerical libraries for scipy

William K. Coulter wcoulter@berkeley....
Fri Apr 3 10:32:43 CDT 2009


Thanks all for your replies.

While this benchmark is interesting, it is not the one that I was 
trying.  I'm interested in computations where just *one* of the matrices 
is actually sparse.  I'm looking at something more like the following in 
your script's language -- I'm guessing on the syntax for Timer:

c = np.random.random((s,s))     # c is dense
b = np.random.random((s, s))
b[b > d] = 0                    # b sparse, but dense representation
b_s = ss.csc_matrix(b)          # b_s sparse, with sparse rep.

timer_dense = Timer("np.dot(b,c)", "import __main__ as m")    # test1
timer_sparse = Timer("b_s.matmat(c)", "import __main__ as m") # test2

Could you look at those curves?  Would you not expect the multiplication 
using the sparse version of b (b_s) to be faster?

On matlab, the equivalent to test2 runs far faster (20x) than test1, 
whereas in python, I'm seeing completely the opposite.  Since there are 
questions about the dense/sparse ratio, I get speedup in matlab on using 
sparse representation for one matrix for ratios (d) of 0.5 and lower.

Any ideas for the performance difference?

-- Will

 From Stéfan van der Walt on 4/3/2009 6.55:
> 2009/4/3 David Cournapeau <david@ar.media.kyoto-u.ac.jp>:
>> Maybe a more useful benchmark would be the dense/sparse ratio as a
>> function of density for a given size,
> 
> http://mentat.za.net/refer/bench_ratio.png
> 
> Cheers
> Stéfan
> 
> 
> ------------------------------------------------------------------------
> 
> _______________________________________________
> SciPy-user mailing list
> SciPy-user@scipy.org
> http://mail.scipy.org/mailman/listinfo/scipy-user

-- 
  William K. Coulter                          wcoulter@berkeley.edu
  Graduate Student, Helen Wills Neuroscience Institute, UC Berkeley


More information about the SciPy-user mailing list