Re: [eigen] arm/neon inside eigen internals

[ Thread Index | Date Index | More Archives ]

On Thursday 04 March 2010 09:36:08 Thomas Capricelli wrote:
> (mostly aimed at Konstantinos I guess)
> In the recent patch providing Neon support (Thanks!), i have some questions
>  regarding the use of __ARM_NEON__
> 1) Detection of architecture should be done in Core only, and in other
>  files, we should only use things like EIGEN_VECTORIZE_NEON. This allows,
>  for example, the user to define EIGEN_DONT_VECTORIZE and have indeed his
>  code not making use of simd

well, the use of EIGEN_VECTORIZE_NEON is equivalent to __ARM_NEON__, ie both 
are defined or not defined at the same time -at least right now, I can think 
of no case where one compiles with -mfpu=neon but requires no vectorization, 
one would use -mfpu=vfp instead. So perhaps I can replace __ARM_NEON__ with 
EIGEN_VECTORIZE_NEON, no problem in that.

> 2) For ABI reason, as explained in Eigen/src/Core/util/Macros.h:40, i think
>  the line 42 should use __arm__ instead of __ARM_NEON__ or
> I first thought about patching/cleaning this myself, but i feel
>  uncomfortable doing so because:
> 1) I want to know what Konstantinos thinks about it
> 2) I'm not sure ABI compatibility is such a hard requirement on embedded
>  platforms  such as ARM 3) in some places (like
>  Eigen/src/Core/util/Memory.h:426 or Eigen/src/Core/util/XprHelper.h:92)
>  i'm not sure whether __ARM_NEON__ should be replaced by __arm__ or
>  EIGEN_VECTORIZE_NEON. That is, should this codepath be used on all arm
>  platforms (think 'abi') or only in the neon case.

Regarding use of __arm__ instead: 

These #ifdefs exists only because GCC on ARM cannot guarrantee that 
__attribute__((aligned(16))) will be honoured (sometimes it is, sometimes it 
is not). This is a bug, and I am actually discussing it (via email) right now 
with the GCC ARM developers. Unaligned loads/stores is irrelevant if one is 
not using vectorization so there is no need to force these ifdefs on all arm 
platforms -in fact, I don't think it makes any difference whatsoever. I could 
replace it with EIGEN_VECTORIZE_NEON, though.

Either way, I didn't put that much thought in it, I still have a few issues to 
fix wrt to the NEON port anyway. I'm willing to fix it in anyway Benoit or 
Gael see as more proper.


Mail converted by MHonArc 2.6.19+