On Saturday, May 16, 2020 6:01 AM, kate via Std-Proposals <std-proposals_at_[hidden]> wrote:

> I found a
> novel approach
> to implement similar interfaces. The idea is very simple: unroll the
> loops introduced
> by vector operations at compile time and leave the rest work, i.e.
> generating code using
> SIMD instructions, to compilers. Then the vectorization is done. It
> relies on compilers'
> optimization ability to generate vectorized instructions from unrolled
> code. For modern
> compiler, it's should be easy.
>

I think many of us did this manually before,
e.g. tweaking the code until it gets auto-
vectorized. Having this accomplished via
a library is attractive for some uses.

But I can see a few things this approach
having trouble with:

1. Generated SIMD instructions in Debug build.
   This approach couples uses of SIMD
   algorithms with optimization. So in Debug
   build the program may be too slow, or it
   can produce different numerical results due
   to change of instructions.

2. Control over ISA.
   This approach does not abstract out ISA,
   therefore the use of ISA is determined by
   the build rather than from the code. Not
   all SIMD libraries can switch between ISAs
   or even mixing ISAs, but at least they
   a chance to do so.

3. Fine control over the access pattern.
   Loop unrolling mandates a particular access
   pattern. It may contribute to a small
   fraction of all the possible algorithms we
   need.

--
Zhihao Yuan, ID lichray
The best way to predict the future is to invent it.
_______________________________________________