[Dune] state of Debian packages for dune
Ansgar Burchardt
ansgar.burchardt at iwr.uni-heidelberg.de
Tue Apr 10 15:07:10 CEST 2012
On 04/09/2012 10:26 PM, Carsten Gräser wrote:
>> It works for me on amd64 as well and Debian's buildd for powerpc also
>> built the package successfully. In addition the numbers look the same:
>>
>> geomCopy.global(pos) != glob: 0.0083333333333333332177 != 0.0083333333333333332177
>>
>> I tried rebuilding with `-fpmath=sse' and the tests pass. So this might
>> be caused by the 80bit precision the 387 FPU uses internally (as far as
>> I know this causes numeric differences if a value is stored in the main
>> memory and then loaded into a FPU register again).
>
> to check if this is really the issue, you could also assign
> the results to variables and compile with '-ffloat-store'
> Normally the variables could be optimized away, but this flag
> ensures that all floats are really stored to memory before accessing
> them again.
I applied a tiny patch for this (attached) and rebuilt the package.
Without -ffloat-store the test fails, but when I add the flag the test
passes. So I assume the test should check the difference (of the norm)
instead. I just filed #1061 for this issue.
Regards,
Ansgar
-------------- next part --------------
A non-text attachment was scrubbed...
Name: float.patch
Type: text/x-patch
Size: 487 bytes
Desc: not available
URL: <https://lists.dune-project.org/pipermail/dune/attachments/20120410/e71ceb99/attachment.bin>
More information about the Dune
mailing list