On Tue, 2007-09-04 at 14:19 +0000, Cesare Tirabassi wrote:
> Thanks Mark, super job!
> Unfortunately this means, unless there is some #ifdefine in ATLAS which 
> branches as a function of the compiler, that this is a compiler bug.
> Would I be asking too much, if I'd ask you to check with the latest gcc 
> (version 4.1.3 20070831 (prerelease) (Ubuntu 4.1.2-16ubuntu1))?
> 

Unfortunately the problem just seems to have got a tad more complex.
Firstly there was a build problem with the atlas, the atlas distribution
only makes the static libraries, and debuild fails with  

        "make: *** No rule to make target
        `debian/tmp/usr/lib/atlas/liblapack.so', needed by `build'.
        Stop."
        
The complexity of the atlas make system places this problem beyond my
ability to solve. However, the previous results relating to locally
built atlas are almost certainly wrong.


A second, possibly more serious, complicating factor, is that I now have
a matrix whose behaviour is complementary to the original one, i.e. it
passes when the original fails and vice-versa. 

This rather tends to suggest that the cause of the problem does, after
all, lie with scipy.

Another aspect that might shed light on the issue is that the dimension
of the kernel is too large (compared a against a reference of calculated
with Gauss-Jordan and arbitrary precision rationals).

In the summary of results below, matrix null(A) had dimension 125 when
calculated by Gauss-Jordan, and 130 via svd, for matrix B the dimensions
were 126 and 130 respectively. 

Updated results, matrices and scripts can be found here:
http://mudshark.brookes.ac.uk/mark/SciTest/

warty/opteron

+atlas (3.6.0-13)
matrix A : max= 3.62882467131e-16 min= -3.61473004307e-16
matrix B : max= 0.0242918127535 min= -0.0329155144634

-atlas
Can't do, breaks dependencies


dapper/pentium

+atlas (3.6.0-20)
matrix A : max= 0.030491255128 min= -0.249297772779
matrix B : max= 5.63011418035e-17 min= -7.21035472861e-17

-atlas
matrix A : max= 2.28048070053e-16 min= -4.583786292e-16
matrix B : max= 1.25730556444e-16 min= -1.65316267349e-16


dapper/amd64

+atlas
matrix A : max= 0.0599265087162 min= -0.13143867524
matrix B : max= 2.16840434497e-16 min= -1.21864324187e-16


-atlas
matrix A : max= 3.62882467131e-16 min= -3.61473004307e-16
matrix B : max= 0.0242918127535 min= -0.0329155144634


dapper/opteron - identical to dapper/amd64


edgy/amd64

+atlas
matrix A : max = 0.97897895303 min= -0.935360320766
matrix B : max = 3.05137859424e-15 min= -1.76247905159e-15

-atlas
matrix A : max = 1.26382736741e-14 min= -1.76067046684e-14
matrix B : max = 0.502358991218 min= -0.792848865561

fiesty/amd64 - identical to edgy/amd64

-- 
Error in linalg.svd for large matrix
https://bugs.launchpad.net/bugs/134277
You received this bug notification because you are a member of Ubuntu
Bugs, which is the bug contact for Ubuntu.

-- 
ubuntu-bugs mailing list
ubuntu-bugs@lists.ubuntu.com
https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs

Reply via email to