Date: Tue, 6 Aug 2019 10:06:01 +0100
> Granted - but I think this is off-topic. By a significant degree-
> I'd be quite interested in a proposal of the type you're describing,
> particularly one that did a good job of explaining the "what, where and
> why" to slum-dog programmers, such as myself ;)
A prerequisite is P1031 Low level file i/o, or some other virtual memory
and page fault management layer. So there is no rush in bringing these
to the committee, given that they are unimplementable in standard C++
currently.
> But in terms of my stuff, can we focus on whether or not SIMD is (a)
> worth focussing on (this talk suggests it is:
> https://www.youtube.com/watch?v=8khWb-Bhhvs), (B) whether gather-scatter
> SIMD is currently worth the time on most modern hardware and
For anything exceeding tight loops of a few hundred opcodes, SIMD is
usually worth it. *IF* you can lay out your data for SIMD. Often
repacking the data layout exceeds the gain from SIMD, it's very
frustrating at times.
>
(C) whether
> I should be exposing the internals of my container.
That's the same debate as the intrusive-vs-not container debate. There
is no one answer, it "depends".
I've often wondered if it would be possible to make a container which
can be both, but I've never tried. The value_type would need to CRTP
from a container metadata template, I think.
Niall
> I'd be quite interested in a proposal of the type you're describing,
> particularly one that did a good job of explaining the "what, where and
> why" to slum-dog programmers, such as myself ;)
A prerequisite is P1031 Low level file i/o, or some other virtual memory
and page fault management layer. So there is no rush in bringing these
to the committee, given that they are unimplementable in standard C++
currently.
> But in terms of my stuff, can we focus on whether or not SIMD is (a)
> worth focussing on (this talk suggests it is:
> https://www.youtube.com/watch?v=8khWb-Bhhvs), (B) whether gather-scatter
> SIMD is currently worth the time on most modern hardware and
For anything exceeding tight loops of a few hundred opcodes, SIMD is
usually worth it. *IF* you can lay out your data for SIMD. Often
repacking the data layout exceeds the gain from SIMD, it's very
frustrating at times.
>
(C) whether
> I should be exposing the internals of my container.
That's the same debate as the intrusive-vs-not container debate. There
is no one answer, it "depends".
I've often wondered if it would be possible to make a container which
can be both, but I've never tried. The value_type would need to CRTP
from a container metadata template, I think.
Niall
Received on 2019-08-06 04:08:03