[PATCH] std::experimental::simd

Thomas Rodgers trodgers@redhat.com
Mon Feb 10 16:49:00 GMT 2020


Catching up on this...(and since I'm not in Prague I won't be able to do so in person).

I finally got a chance to talk with Jeff Law recently about a more extensive testing strategy using Red Hat internal resources, and I am going to start working on
adapting his CI build scripts that drive checking out machines from our lab environment starting later this week.

Jonathan indicated you may have a newer version of your <simd> patch forthcoming? Is that something that's likely to happen before the end of this month? I'd
like to work on getting start working on getting <simd> in during stage4 (Jonathan is willing to accept it because it is experimental as long as it is in
good shape to commit).

Thanks,
Tom.


----- Original Message -----
From: "Dr. Matthias Kretz" <m.kretz@gsi.de>
To: trodgers@redhat.com
Cc: libstdc++@gcc.gnu.org
Sent: Thursday, October 24, 2019 1:26:47 AM
Subject: Re: [PATCH] std::experimental::simd

On Montag, 14. Oktober 2019 14:12:12 CEST Matthias Kretz wrote:
> This patch is without documentation and testsuite. I can add them on request
> but would prefer a follow-up patch after getting this one right.

Regarding tests, here's what I think a "consumer test" should cover:

1. compiler flags
  - `-march=native -O2`

2. element types
  - char
  - uint8_t
  - int
  - float
  - double

3. operations
  - broadcast ctor
  - generator ctor
  - all operators
  - compares
  - mask reductions
  - non-converting loads & stores
  - non-converting masked loads & stores
  - horizontal reductions
  - math: 1024 input values, comparing fun(T) against fun(simd<T>)
  - conversions:
    * char <-> int
    * uint8_t <-> int
    * uint8_t <-> float
    * int <-> float
    * int <-> double
    * float <-> double

4. ABI tags:
  - scalar
  - fixed_size<{2, 3, 4, 6, 8, 12, 16, 32}>
  - deduce_t<T, {2, 3, 4, 6, 8, 12, 16, 32}> (if different from fixed_size)
  - deduce_t<T, 64> for AVX512BW targets
  - compatible (if different from above)
  - native (if different from above)

Note that 1-4 are orthogonal and already span a huge space. I'm sure we need 
to reduce this list, not expand.

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



More information about the Libstdc++ mailing list