From: Andrey Semashev (andrey.semashev_at_[hidden])
Date: 2024-12-10 21:00:09
On 12/10/24 21:15, Peter Dimov via Boost wrote:
> Ruben Perez wrote:
>>
>> Is it expected to reach similar levels of performance in the future using Hash2?
>> Or is this not in scope for the library?
>
> I don't see why it wouldn't be in scope; performance optimizations are never
> ruled out.
>
> Taking advantage of SHA-NI when it's statically known to be available (under
> GCC/Clang -march=native, for instance) should be relatively easy to add.
Requiring SHA instructions unconditionally sets a rather high bar on the
CPU requirements.
> Dynamic dispatch is a bit more convoluted because of constexpr but should
> be possible.
SIMD intrinsics (at least x86 ones) are not constexpr-friendly as they
require reinterpret_cast to access memory. Only AVX-512 added load/store
intrinsics taking void pointers. Maybe gcc's vector extensions are
better, but I'm not sure if SHA intrinsics will work with them. I
suspect the constexpr version will have to always use the generic code,
and the runtime version may use CPU detection and vectorized code.