Re: [eigen] alignment bug - or where is our non sse inversion gone?

[ Thread Index | Date Index | More lists.tuxfamily.org/eigen Archives ]


hm sorry but I don't see it. Here is what I have for ei_compute_inverse_size4:

template<int Arch, typename Scalar, typename MatrixType, typename ResultType>
struct ei_compute_inverse_size4;

Anyway, actually I'm doing a different fix which consist in doing
unaligned loads/stores if the input and/or output do not have the
AlignedBit flags. Here doing unaligned accesses is fine.

gael.

On Tue, Aug 24, 2010 at 12:19 PM, Hauke Heibel
<hauke.heibel@xxxxxxxxxxxxxx> wrote:
> But we already have this parameter, see ei_compute_inverse_size4 in
> Inverse.h line 223 and then line 249.
>
> This is why I was thinking we need to adapt the Architecture structure.
>
> - Hauke
>
> On Tue, Aug 24, 2010 at 12:13 PM, Gael Guennebaud
> <gael.guennebaud@xxxxxxxxx> wrote:
>> Thanks for finding this bug.
>>
>> Actually no need to define EIGEN_DONT_ALIGN_STATICALLY to hit it, simply use:
>>
>> Matrix<float,4,4,DontAlign>
>>
>> The fix is to do the same than for ei_quat_product, i.e., add an
>> Aligned/Unaligned template parameter to ei_compute_inverse_size4.
>>
>> will do it.
>>
>> gael
>>
>> On Tue, Aug 24, 2010 at 12:06 PM, Hauke Heibel
>> <hauke.heibel@xxxxxxxxxxxxxx> wrote:
>>> On Tue, Aug 24, 2010 at 9:24 AM, Hauke Heibel
>>> <hauke.heibel@xxxxxxxxxxxxxx> wrote:
>>>> I just stumbled over a problem when using EIGEN_DONT_ALIGN_STATICALLY
>>>> is defined.
>>>
>>> Uups, did I mention that what is happening is an unaligned memory
>>> access and followed by a segfault??
>>>
>>> - Hauke
>>>
>>>
>>>
>>
>>
>>
>
>
>



Mail converted by MHonArc 2.6.19+ http://listengine.tuxfamily.org/