LAPACK Archives

[Lapack] Query re LAPACK testing


Hi Shandong,

While [c|d|s|z]gd.out contains many error output like this:

   Matrix order=    2, type=26, seed=2749,1691, 257,2573, result  5 is 
8.389E+06

they are still marked as passed.

This is OK. Or kind of OK. We should probably removed these tests or do 
something about this.

We actually do expect these to fail with an order of 1e6. The matrix pencils 
used for testing the
code are ill-conditionned and the test #4 tests the forward error and so it is 
OK if it is at the 1e+6 level. 
As you can see the tests #1 and #2 are OK. (They check the backward error. 
These are the tests
which matter.

Bottom line: we should probably remove test #4. Or leave it only for the 
well-conditionned problems. 
Or divide by the result by the condition number of the problem. 

Another example is [c|d|s|z]gbak.out. On different plat form the largest 
test errors can be very different:

   value of largest test error                  =    .105E-03
   value of largest test error                  =   0.524E+00

They are all said to be passed. How can we judge the correctness of the 
tests?

Order of three magnitude is kind of scary from one platform to the other,  I 
agree.
FYI: I also have .524E+00 on my machine.

I just read the testing of DGEBAL. We start by inputting an initial matrix, and 
then
we input the balanced matrix and the scaling coefficient (so the expected 
output). 
Then we check that the output of DGEBAL matches the output we expect. Since
we are in relative errors here, (and we do not divide by EPS,) .524 seems high. 
We could look at this as well indeed. 

Cheers, 
Julien.




Could you please help me to understand the test results?

Thank you,
Shandong

On 12/13/11 15:12, GREGORY TARSY wrote:
Dear Julie,

We would be pleased if you could add an Oracle supported platform with 
either Solaris or Oracle Linux and using
Oracle Solaris Studio compilers. Studio is freely available for 
download as is, I believe, Oracle Linux.

The Oracle Performance Library, which is included in Studio, contains 
tuned versions of LAPACK and the BLAS, so testing with our compilers 
would really aid us in upgrading to new versions of LAPACK.

Regards,
Gregory

On 11/22/11 3:07 PM, julie langou wrote:
Dear Gregory,
We have nightly builds of LAPACK here.
http://my.cdash.org/index.php?project=LAPACK

Please let us know if you want us to add a specific configuration.
We are trying our best  to test the LAPACK library on very various 
platform.

Our release process is date based actually?one major release usually 
for SuperComputing with new interface, etc.. and as many bug fix 
release as needed during the year.
We would be glad to get some feedback from your team and improve our 
release process if you have any suggestions.
Sincerely,
Julie Langou
On Nov 22, 2011, at 12:22 PM, GREGORY TARSY wrote:

Would you please explain the procedure for testing LAPACK releases 
before they are made available.

In particular, which platforms and which compilers are used?
Is only the distributed test-suite used?
Also what is the standard for deciding that a release is of 
sufficient quality?

Thanks,

-- 
Gregory Tarsy | Senior Manager | Advanced Math Libraries
Phone: +1 4082760055 <tel:+1%204082760055> | Mobile: +1 8318182198 
<tel:+1%208318182198>
ORACLE Solaris Developer Tools
4180 Network Circle | Santa Clara, CA 95054
_______________________________________________
Lapack mailing list
Lapack@Domain.Removed <mailto:Lapack@Domain.Removed>
http://lists.eecs.utk.edu/mailman/listinfo/lapack


-- 
Gregory Tarsy | Senior Manager | Advanced Math Libraries
Phone: +1 4082760055 <tel:+1%204082760055> | Mobile: +1 8318182198 
<tel:+1%208318182198>
ORACLE Solaris Developer Tools
4180 Network Circle | Santa Clara, CA 95054
_______________________________________________
Lapack mailing list
Lapack@Domain.Removed
http://lists.eecs.utk.edu/mailman/listinfo/lapack


<Prev in Thread] Current Thread [Next in Thread>


For additional information you may use the LAPACK/ScaLAPACK Forum.
Or one of the mailing lists, or