Hello all
I managed to get NumPy to compile with the free for non-commercial use Linux
versions of the Intel C compiler, version 9.1 and Intel MKL 8.1.
Instructions at the bottom of this page:
http://www.scipy.org/Installing_SciPy/Linux
I ran the NumPy test suite, and it turned up a few test failures in
test_ufunclike -- it seems the Intel C/MKL combo comes up with different
answers for the sign of NaN.
Compiling with -Wall turned up a few warnings, but nothing too serious. I
submitted a ticket for the ones that seem easily fixable and might warrant a
quick look:
http://projects.scipy.org/scipy/numpy/ticket/366
I recently started building some of my other C code on Windows with the
Intel compiler, and the speed increases in my application are quite
dramatic. The Intel code is about 4 times faster than the code produced by
Visual Studio .NET 2003.
I suspect the speed increase is due to some loop vectorization. When
compiling NumPy on Linux, some interesting messages pop up:
numpy/core/src/arraytypes.inc.src(658) : (col. 9) remark: LOOP WAS
VECTORIZED.
numpy.distutils doesn't seem to support the Intel compiler on Windows, but
if you have other C code lying around, you can easily use SCons to compile
it with the Intel compiler (Intel offers 30-day evaluation versions of most
of their products).
Have fun!
Cheers,
Albert
|
