Re: [eigen] Eigen containers cannot be vectorized

[fayard@xxxxxxxxxxxxx]

I don't have my computer with me tonight (writing from an iPad). But I 
am quite sure that:
#pragma omp simd
should work. As it is also portable, it might be the best solution. But 
this kind of loop could be vectorized automatically by most compilers 
and it is too bad it does not work with Eigen.

Let me explain why I have started this conversation. Many people work on 
numerical analysis in C++. As you know, there is no multidimensional 
array in the standard library (which is a shame, by the way). Many 
people around me ask for such a library and they end up with Linear 
Algebra libraries such as Eigen. Then Tensor module might even push more 
people into this direction. I know that Eigen has not been designed to 
be a multidimensional array library, but it does the job pretty well. It 
also comes with nice pretty printers for gdb, and does not depend on 
Boost. As far as I know, it is the only mature library that fullfill 
those requirements.

Unfortunately, it does play nicely with auto-vectorisation. That's why I 
was suggesting to make this change so make Eigen better on that matter. 
Unfortunately, the problem seems to be much more difficult to handle. As 
I have said, I personaly use my own library, but it would be nice to get 
better auto-vectorisation for those who use Eigen as a multidimensional 
array package.


On 2016-12-08 21:22, Jeff Hammond wrote:
> You might try one of the following pragmas to inform the compiler to
> ignore aliasing in the loop nest.  I haven't tested your code, so I
> apologize if none of these work.
>
> #pragma ivdep (or #pragma GCC ivdep, in the case of GCC) is supported
> by many compilers.  I don't have know the full list
> (https://github.com/jeffhammond/nwchem-tce-triples-kernels/blob/master/src/pragma_vendor.h
> [5] lists all the ones I knew at some point, and Google tells me MSVC
> supports it too).  In any case, it is easy enough to write build
> system tests for this sort of thing.
>
> #pragma simd is an Intel compiler keyword.  See
> https://software.intel.com/en-us/node/524555 [6] for details.  The
> Intel compiler takes this pragma (and the OpenMP equivalent noted
> below) seriously and will generate vector code when you use these,
> even in cases where the heuristics say it isn't beneficial.
>
> #pragma omp simd is part of OpenMP 4 but you can get it and only it
> from GCC and Intel via -fopenmp-simd and -qopenmp-simd, respectively.
> See
> https://gcc.gnu.org/onlinedocs/gcc-5.3.0/gcc/C-Dialect-Options.html
> [7] for details on GCC.
>
> Best,
>
> Jeff