[Numpy-discussion] Problems when using ACML with numpy
thomas_unterthiner at web.de
Fri May 11 18:54:47 EDT 2012
I'm having troubles getting numpy to work with ACML. I'm running Ubuntu
12.04 on an x86-64 system and use acml 5.1.0.
On my first try, I installed numpy/scipy from the official ubuntu
repository, then just changed the symlink to the blas/lapack libraries
of my system to use acml. i.e. I did:
ln -s /usr/lib/liblapack.so.3gf
ln -s /usr/lib/libblas.so.3gf /opt/acml5.1.0/gfortran64_fma4/lib/libacml.so
This method worked perfectly fine for Octave. However with numpy the
following code will always run into what seems to be an endless loop:
import numpy as np
a = np.random.randn(5000, 5000)
t0 = time.clock()
b = np.dot(a, a)
t1 = time.clock()
print t1 - t0
The process will have 100% CPU usage and will not show any activity
under strace. A gdb backtrace looks as follows:
#0 0x00007fdcc000e524 in ?? ()
#1 0x00007fdcc008bcb9 in ?? ()
#2 0x00007fdcc00a304d in ?? ()
#3 0x000000000042a485 in PyEval_EvalFrameEx ()
#4 0x00000000004317f2 in PyEval_EvalCodeEx ()
#5 0x000000000054bd50 in PyRun_InteractiveOneFlags ()
#6 0x000000000054c045 in PyRun_InteractiveLoopFlags ()
#7 0x000000000054ce9f in Py_Main ()
#8 0x00007fdcc0e7976d in __libc_start_main ()
Curiously enough, when changing the matrix-dimensions from 5000x5000 to
500x500, all is good and the code executes in 0.32 seconds. (As a
reference: the 5000x5000 matrix multiply takes 65 seconds in octave with
ATLAS and 10 seconds with ACML, so the problem is not that the
matrix-multiply just takes too long).
I then tried compiling numpy-1.6.1 myself, doing:
tar xzf numpy-1.6.1.tar.gz
export CFLAGS="-O3 -march=native"
export CXXFLAGS="-O3 -march=native"
export FFLAGS="-O3 -march=native"
export FCFLAGS="-O3 -march=native"
python setup.py build
This worked (apart from a missing '-shared' flag when linking
lapack_lite.so, which I then had to link by hand by adding the flag),
however the error persisted. The same with numpy-1.6.2rc1.
Fromt here I don't know how to proceed. Any help would be greatly
More information about the NumPy-Discussion