masked loads?

[Matthias Kretz]

On Samstag, 16. Juli 2016 08:38:42 CEST Kay F. Jahnke wrote:
> I'm looking for masked load operations in Vc. It looks like AVX has
> them, like _mm_maskload_ps, and it would be nice to find them in Vc -
> for symmetry reasons, since it does offer masked stores. My current
> workaround is to use a masked gather, but this is ugly:
> 
> Vector<something> ( mem , index_type::IndexesFromZero() , mask ) ;
> 
> compared to the masked store operation
> 
> Vector<something>.store ( mem , mask ) ;
> 
> When I tried
> 
> Vector<something> ( mem , mask ) ;
> 
> This seemed to work, but after some time I noticed that the mask is
> misinterpreted as a set of indices for an unmasked gather operation,
> resulting in unexpected results.

Ouch. Yes, the requirement for the index type is just that it has a subscript 
operator and can produce integral offsets. (0 or 1 in the mask case) Therefore 
you're calling a gather constructor. This is a feature, though unexpected. 
I'll have to think about it. This is an unfortunate clash with the masked 
store syntax...

The reason there is no masked load is that the functionality in hardware is 
fairly new. Feedback from the C++ committee subgroup on concurrency (including 
an Intel rep) actually was to omit masked loads in the interface. One major 
reason is the question whether we want to specify the masked load to emit 
faults or not for the masked off parts. A proper masked load certainly should 
not emit faults. An efficient implementation where hardware support is missing 
would emit faults, though.

> Or am I missing something?

Vector<T> x = {};
x(mask) = Vector<T>(mem);

i.e. initialize a null vector object. Then do a masked assignment from an 
unmasked load. This is not as efficient as _mm_maskload_ps (unless the 
compiler does the optimization).

Cheers,
  Matthias
-- 
──────────────────────────────────────────────────────────────────────────
 Dr. Matthias Kretz                                https://kretzfamily.de
 GSI Helmholtzzentrum für Schwerionenforschung             https://gsi.de
 SIMD easy and portable                     https://github.com/VcDevel/Vc
──────────────────────────────────────────────────────────────────────────