Package: python-numpy
Version: 1:1.8.2-2
(This problem report was written by Yaroslav Bulatov. I've confirmed it on
Debian
sid chroot. My computer gave a 30X improvement.)
Default numpy install uses inferior BLAS, and is very slow. Matrix
multiplication benchmark below gets me 1.26 G items/second with default
install on my Xeon 6-core 3.2 Ghz. When I install OpenBLAS, it goes up 186
G items/second. That's 150x improvement in speed which should be the
default for numpy.
benchmark
----
import time
import numpy as np
import numpy.random as random
size = 512
iters = 500
a = random.rand(size, size).astype(np.float32)
b = random.rand(size, size).astype(np.float32)
start_time = time.time()
for i in range(iters):
np.dot(a, b)
end_time = time.time()
# size**3 multiplies, (size - 1) * size ** 2 adds
num_ops = size ** 3 + (size - 1) * size ** 2
print "Did %d multiplications of %d x %d matrices in %.1f seconds" %
(iters, size, size, end_time - start_time)
print "%.4f G items/sec" % (num_ops * float(iters) / (end_time -
start_time) / 10 ** 9)
fixing it with openblas:
-----
$ sudo apt-get install libopenblas-dev virtualenv python-dev
$ mkdir ~/tmp
$ cd ~/tmp
$ virtualenv env
$ source env/bin/activate
$ pip install numpy
$ which python
~/tmp/env/bin/python
$ python ...
