A Monday 05 July 2010 15:32:51 Isaac Gouy escrigué: > Sturla Molden <sturla <at> molden.no> writes: > > It is also the kind of tasks where NumPy would help. It would be nice to > > get NumPy into the shootout. At least for the sake of advertising > > http://shootout.alioth.debian.org/u32/program.php?test=spectralnorm&lang=py > thon&id=2
Let's join the game :-) If I run the above version on my desktop computer (Intel E8600 Duo @ 3 GHz, DDR2 @ 800 MHz memory) I get: $ time python -OO spectralnorm-numpy.py 5500 1.274224153 real 0m9.724s user 0m9.295s sys 0m0.269s which should correspond to the 12.86s in shootout (so my machine is around 30% faster). But, if I use ATLAS (3.9.25) so as to accelerate linear algebra: $ python -OO spectralnorm-numpy.py 5500 1.274224153 real 0m5.862s user 0m5.566s sys 0m0.225s Then, my profile said that building M matrix took a lot of time. After using numexpr to improve this (see attached script), I get: $ python -OO spectralnorm-numpy-numexpr.py 5500 1.274224153 real 0m3.333s user 0m3.071s sys 0m0.163s Interestingly, memory consumption also dropped from 480 MB to 255 MB. Finally, if using Intel's MKL for taking advantage of my 2 cores: $ python -OO spectralnorm-numpy-numexpr.py 5500 1.274224153 real 0m2.785s user 0m4.117s sys 0m0.139s which is a 3.5x improvement over the initial version. Also, this seems faster (around ~25%), and consumes similar memory than the fastest version written in pure C in "interesting alternatives" section: http://shootout.alioth.debian.org/u32/performance.php?test=spectralnorm#about I suppose that, provided that Matlab also have a JIT and supports Intel's MKL, it could beat this mark too. Any Matlab user would accept the challenge? -- Francesc Alted
# The Computer Language Benchmarks Game # http://shootout.alioth.debian.org/ # # Contributed by Sebastien Loisel # Fixed by Isaac Gouy # Sped up by Josh Goldfoot # Dirtily sped up by Simon Descarpentries # Sped up with numpy by Kittipong Piyawanno # More speed-up with Numexpr by Francesc Alted from sys import argv import numpy as np import numexpr as ne def spectralnorm(n): u = np.matrix(np.ones(n)) j = np.arange(n, dtype=np.float64) M = np.empty(n*n, dtype=np.float64).reshape(n,n) expr = "1.0 / ((i + j) * (i + j + 1) / 2 + i + 1)" for i in xrange(n): M[i] = ne.evaluate(expr) MT = M.T for i in xrange(10): v = (u*MT)*M u = (v*MT)*M print "%0.9f" % ((u*v.T).sum() / (v*v.T).sum())**0.5 spectralnorm(int(argv[1]))
_______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@scipy.org http://mail.scipy.org/mailman/listinfo/numpy-discussion
