Am 25.03.2017 um 19:14 schrieb Mike Horvath:
> On 3/25/2017 11:30 AM, clipka wrote:
> while on modern Windows systems we expect a noticeable
>> gain.
> 
> 
> Compared to when? And by how much?

Hm, guess I misread that.

"Compared to when?" is the wrong question, since "modern" can't just be
defined in terms of some parameter as simple as the date of purchase.

The most important factor is the extended instruction sets your CPU
supports, but also how well those instruction sets are implemented:

Modern:

- AMD Ryzen machines (expect a lot of gain).
- Intel AVX2/FMA3-capable machines (expect a lot of gain).
- AMD AVX/FMA4-capable machines (expect a bit of gain).
- Other Intel AVX-capable machines (expect moderate gain).

Not modern:

- Other AMD AVX-capable machines (expect some loss*).
- Non-AVX-capable machines (all bets are off).

All of the above information pertains to Windows machines, comparing
3.7.1-beta.5 to 3.7.0 official release.

As for "by how much?" see my other post for ballpark figures.


(*) We haven't been able to perform tests of our own to confirm this
behaviour, so need to rely on 3rd party feedback. Judging from that, it
looks as if old AMD machines supporting AVX but not FMA4 exhibit poor
AVX performance, and actually suffer from enabling AVX-specific
optimizations; we'll have to see how we can address that problem. Maybe
we'll have to restrict the AVX-specific optimizations to genuine Intel
CPUs, and/or provide a setting to override the automatic choice of
optimized code.

Another performance-related ToDo item is making the alternative
codepaths available on Linux machines as well.

Post a reply to this message