[SciPy-Dev] suitesparse vs scipy
Wed Aug 31 08:33:04 CDT 2011
On Tue, Aug 30, 2011 at 03:02:48PM +0100, Ben Willmore wrote:
> Hi Jack,
> > Pauli,
> > On x86_64-apple-darwin11, building scypy 0.9.0 against python2.7 using
> > the clang compilers and gfortran from FSF gcc 4.6.1, I get the following testsuite
> > results...
> I'm a scipy novice, probably hanging out on the wrong list, but I know a little bit about most of the errors you're seeing.
> There are some differences in calling conventions between Apple's veclib and scipy's expectations which mean that various single-precision functions fail, see here:
> This can be fixed by setting 'export FFLAGS=-ff2c' before running/compiling numpy and scipy (*). For me, this fixed all test_blas failures, using a recent git clone. As far as I can tell, it doesn't introduce other problems (i.e. most tests pass).
> I tried this with all the standard Lion C compilers, and a couple of different versions of gfortran, with similar results (though llvm has it's own problems (**), and you shouldn't use it). So this is not a question of choosing the right compiler.
This doesn't make any sense for numpy. I tried you approach of using...
/sw/bin/python2.7 setup.py build --fcompiler=gfortran
and while I see..
Found executable /sw/bin/gfortran
customize Gnu95FCompiler using config
at no point during the numpy build is anything compiled with gfortran. The resulting numpy still fails the
test_special_values tests. Note that these are claimed to be non-critical and should be disabled on darwin...
> (*) The results of my experiments are here:
> (**) For instance, ifft(fft(signal)) != signal when using scipy.fftpack
> SciPy-Dev mailing list
More information about the SciPy-Dev