[SciPy-Dev] [SciPy-User] ANN: scipy 0.8.0 release candidate 3

[Derek Homeier]

Hi Ralf,

>
> Notably, I never got trouble-free builds with 2.7 - both the numpy  
> and scipy test suites fail
> with a bus error - I already tried compiling with gcc 4.2 instead of  
> 4.0, but to no avail.
>
> Yes, there are issues with 2.7. Compiling against numpy 1.4.1  
> doesn't work, against trunk also has some issues on OS X. As Vincent  
> pointed out, with a 64-bit Python 2.7 built with gcc-4.2 (python.org  
> binaries are gcc-4.0) it does work without problems.

well, as I said I did try the gcc-4.2 build, but probably missed that  
Vincent was compiling numpy trunk -
I probably should try again on my 10.6 64-bit installation.
Anyway, compiling both numpy 1.4.1 and scipy 0.8.0rc3 works fine both  
on i386/ppc 32-bit and and x86_64,
using the fink build system. And it is running as well, so far, except  
for those tests - btw.
numpy.test(verbose=5) crashes at
test_multiarray.TestIO.test_ascii ... Bus error

and scipy.test() at
Testing that kmeans2 init methods work. ... Bus error

>
> These should have been fixed by r6520:
> -        self.check_cephes_vs_amos(iv, iv, rtol=1e-12, atol=1e-305)
> +        self.check_cephes_vs_amos(iv, iv, rtol=5e-9, atol=1e-305)
>
>          data(gammaincinv, 'gamma_inv_big_data_ipp- 
> gamma_inv_big_data',
> -             (0,1), 2, rtol=5e-12),
> +             (0,1), 2, rtol=1e-11),
>
> Can you check if you have these changes, and with what accuracy the  
> tests pass?
>
I have (rc3 tarball), but it's a different couple of tests.
'gamma_inv_big_data_ipp-gamma_inv_big_data' actually still passes with  
rtol=2.1e-12.

These are the minimal accuracy changes needed:

--- /sw/lib/python2.6/site-packages/scipy/special/tests/test_data.py	 
2010-07-11 17:25:24.000000000 +0200
+++ test_data.py	2010-07-17 21:13:07.000000000 +0200
@@ -85,7 +85,7 @@
          data(gamma, 'test_gamma_data_ipp-near_1', 0, 1),
          data(gamma, 'test_gamma_data_ipp-near_2', 0, 1),
          data(gamma, 'test_gamma_data_ipp-near_m10', 0, 1),
-        data(gamma, 'test_gamma_data_ipp-near_m55', 0, 1),
+        data(gamma, 'test_gamma_data_ipp-near_m55', 0, 1, rtol=7e-12),
          data(gamma, 'test_gamma_data_ipp-near_0', 0j, 1, rtol=2e-9),
          data(gamma, 'test_gamma_data_ipp-near_1', 0j, 1, rtol=2e-9),
          data(gamma, 'test_gamma_data_ipp-near_2', 0j, 1, rtol=2e-9),
--- /sw/lib/python2.6/site-packages/scipy/special/tests/test_basic.py	 
2010-07-11 17:25:24.000000000 +0200
+++ test_basic.py	2010-07-17 21:28:07.000000000 +0200
@@ -1618,7 +1618,7 @@

          # Most error apparently comes from AMOS and not our  
implementation;
          # there are some problems near integer orders there
-        assert dc[k] < 1e-9, (v[k], x[k], iv(v[k], x[k]), iv(v[k],  
x[k]+0j))
+        assert dc[k] < 1.9e-7, (v[k], x[k], iv(v[k], x[k]), iv(v[k],  
x[k]+0j))

      def test_kv_cephes_vs_amos(self):
          #self.check_cephes_vs_amos(kv, kn, rtol=1e-9, atol=1e-305)

Of course nearly 2e-7 seems a quite high tolerance - don't know if  
that's acceptable.

>
> Both standard and full test pass on i386 and x86_64, but I noticed  
> the following strange behaviour:
> when running the test suite twice in a row, the second run produces  
> these failures:
>
> This is because warnings are only raised once from the same code, so  
> the check if they're raised a second time fails. So no problem.
>

Ah, that makes sense, thanks!

Cheers,
							Derek