Re: [eigen] Problem with GCC<4.5+amd64+vectorization (Was: On lazy evaluation and temporary matrices)

[Jose Luis Blanco <joseluisblancoc@xxxxxxxxx>]

I don't think many would find this problem, but after two month I
finally fixed this issue and wanted to leave here a (dirty) solution
for the joy of future readers of the list archive ;-)

It turned out that just for one specific part of my code, an Eigen
matrix wasn't computed OK. This only happened with GCC older than 4.5
and for the amd64 architecture. In all other cases, everthing works
nicely. Since valgrind doesn't complain of anything, I wasn't able to
generate an isolated test case and the code actually works for all
other situations, I'm forced to think it may be some intrincated bug
in GCC 4.4 (?).

My solution has been to:

1) Disable vectorization for GCC <4.5 with EIGEN_DONT_VECTORIZE
(thanks Gael!) in the file with the problematic code, and
2) move to GCC 4.5 !!

JL

On Thu, Nov 25, 2010 at 8:25 AM, Gael Guennebaud
<gael.guennebaud@xxxxxxxxx> wrote:
> Hi,
>
> first make sure to test it in "debug mode", i.e., without NDEBUG or
> EIGEN_DEBUG defined to enable the detection of some some alignment
> issues. If that does not help, try to disable vectorization by
> defining EIGEN_DONT_VECTORIZE.
>
> Is your code multithreaded?
>
> Also in your case the .eval() is superfluous.
>
> gael
>
> On Thu, Nov 25, 2010 at 2:45 AM, Jose Luis Blanco
> <joseluisblancoc@xxxxxxxxx> wrote:
>> Hi Thomas,
>>
>> Thanks, I added the EIGEN_ALIGN16 macro to all the number arrays, it
>> definitively makes sense... but it behaves exactly the same :-(
>>
>> Best,
>> JL
>>
>> PS: And yes, it's beta3*, actually an hg checkout from a few days ago...
>>
>> On Thu, Nov 25, 2010 at 2:22 AM, Thomas Capricelli
>> <orzel@xxxxxxxxxxxxxxx> wrote:
>>> It looks like an alignement problem. You should ensure "numbers" is aligned,
>>> see bottom of http://eigen.tuxfamily.org/dox-devel/
>>>
>>> (I assume you use the devel branch.. 3.0beta*)
>>>
>>> regards,
>>>
>>> --
>>>
>>> Thomas Capricelli <orzel@xxxxxxxxxxxxxxx>
>>>
>>> http://www.freehackers.org/thomas
>>>
>>> On Thursday 25 November 2010 02:01:48 Jose Luis Blanco wrote:
>>>
>>>> Hi guys,
>>>
>>>>
>>>
>>>> I need the help of someone knowing the inner workings of Eigen. I've
>>>
>>>> spent a few days stuck debugging an error in some silly computations
>>>
>>>> with Eigen matrices, and want to rule out potential errors...
>>>
>>>> Consider this code:
>>>
>>>>
>>>
>>>> ======================
>>>
>>>> void func( Matrix<...> &m_out )
>>>
>>>> {
>>>
>>>> Matrix<...> I = ....
>>>
>>>> const double numbers[] = { 1.0, x(), -y()*x() , whatever, .... };
>>>
>>>> m_out = I * Matrix<...>(numbers);
>>>
>>>> }
>>>
>>>>
>>>
>>>> int main()
>>>
>>>> {
>>>
>>>> Matrix<...> M;
>>>
>>>> func(M);
>>>
>>>> cout << M << endl;
>>>
>>>> }
>>>
>>>> ======================
>>>
>>>>
>>>
>>>> Is it a problem for the temporary object "Matrix<...>(numbers)" to be
>>>
>>>> created with a pointer to memory on the stack, while the evaluation of
>>>
>>>> the object for the matrix product is actually evaluated *out* of that
>>>
>>>> function (where obviously that pointer is invalid)??
>>>
>>>>
>>>
>>>> I ask because I'm having a very erratic behavior in code like above
>>>
>>>> (on GCC doesn't work, it does on MSVC; adding a "cout << ..." for an
>>>
>>>> unrelated matrix changes the result, etc...).
>>>
>>>> Is it necessary to use instead:
>>>
>>>>
>>>
>>>> ( I * Matrix<...>(numbers) ).eval();
>>>
>>>>
>>>
>>>> or not even with this .eval() is my code correct?
>>>
>>>> I'm still having random wrong matrix elements with and without eval(),
>>>
>>>> so I wanted to be sure on this part before keep on debugging...
>>>
>>>>
>>>
>>>> Thanks for your time!
>>>
>>>> JL
>>>