Re: [eigen] Re: Eigen beta test errors

[Gael Guennebaud <gael.guennebaud@xxxxxxxxx>]

I still don't understand why it worked on other platforms....

gael

On Mon, Jul 5, 2010 at 3:40 PM, Benoit Jacob <jacob.benoit.1@xxxxxxxxx> wrote:
> What was the error? Was it really 32bit specific or was it something
> that  a suitable unit test could detect on other platforms?
>
> Benoit
>
> 2010/7/5 Gael Guennebaud <gael.guennebaud@xxxxxxxxx>:
>> product_syrk_1 is fixed.
>>
>> I can also reproduce packetmath_1 by running it a very large number of
>> times, but there is nothing critical here... it fails on stuff like:
>> 0.00128718 is not approx to 0.00128716.... so that's fine.
>>
>> Same for SVD.
>>
>> It is "normal" that some tests randomly fail. Well, these tests should
>> be fixed, but the problem is not in Eigen itself.
>>
>> gael
>>
>> On Mon, Jul 5, 2010 at 6:07 AM, Benoit Jacob <jacob.benoit.1@xxxxxxxxx> wrote:
>>> On linux x86-64 with gcc 4.4.4, I can't reproduce any of these issues.
>>>
>>> So they are probably 32bit-specific. Can anybody try these tests on
>>> linux x86-32 and confirm? I don't have multilib working (one of many
>>> fedora annoyances) so I can't try.
>>>
>>> I tried -mfpmath=387 and that didn't make any difference: I still
>>> couldn't reproduce the failures.
>>>
>>> Can you also check that you are testing the latest revision from hg
>>> (hg diff should say nothing and hg parent should return the latest
>>> revision).
>>>
>>> Benoit
>>>
>>> 2010/7/4 Carlos Becker <carlosbecker@xxxxxxxxx>:
>>>> I could now reproduce the error in packetmath_1:
>>>> a[3]: 0.00128718, b[3]: 0.00128716
>>>> Test packetmath<float>() failed in "/home/cjb/eigenThings
>>>>     areApprox(ref, data2, PacketSize) && "ei_preduxp"
>>>> and in svd_1 after some trials:
>>>> Test svd(Matrix3f()) failed in "/home/cjb/eigenThings/eigen/test/svd.cpp"
>>>> (63)
>>>>     test_ei_isApprox(a * x, b)
>>>> but I still have to get the right values.
>>>> The most important error is in product_syrk_1, where I just get a segfault
>>>> :S, anyone getting the same behaviour?
>>>>
>>>> On Sun, Jul 4, 2010 at 10:47 PM, Carlos Becker <carlosbecker@xxxxxxxxx>
>>>> wrote:
>>>>>
>>>>> Hi everyone, these are the errors I get with OpenMP enabled, vectorization
>>>>> enabled (up to SSE3) and row major alignment. GCC v4.4.1, ubuntu x32
>>>>>
>>>>> 99% tests passed, 4 tests failed out of 452
>>>>>
>>>>> The following tests FAILED:
>>>>>       14 - packetmath_1 (Failed)
>>>>>      194 - product_syrk_1 (Failed)
>>>>>      214 - stable_norm_1 (Failed)
>>>>>      319 - svd_1 (Failed)
>>>>> Errors while running CTest
>>>>>
>>>>> However, I ran it a second time and only stable_norm_1 and product_syrk_1
>>>>> failed! Now, ran it a third time and got these two failing again PLUS
>>>>> redux_8 !!! I really don't know what is happening, and I tried running
>>>>> packetmath_1 several times on my own and could not make it fail. The same
>>>>> happened with svd_1.
>>>>>
>>>>> However, I can make redux_8 fail: Test vectorRedux(ArrayXf(33)) failed in
>>>>> "/home/cjb/eigenThings/eigen/test/redux.cpp" (88)
>>>>>     test_ei_isApprox(s, v.head(i).sum())
>>>>>
>>>>> and printing some values:  s : 0.000287235, v.head(i).sum(): 0.000286877,
>>>>> although I don't know if this makes sense really.
>>>>>
>>>>> Does this all make sense somehow? I am starting to believe that something
>>>>> is wrong with my processor, or I am a very special case.
>>>>>
>>>>> Carlos
>>>>
>>>
>>>
>>>
>>
>>
>>
>
>
>