On Tue, Aug 26, 2025 at 5:31 PM Jan Hubicka <hubi...@ucw.cz> wrote:
>
> > > > > > In general we should have a look at register pressure, I
> > > > > > suppose issue_rate / m_num_reductions ensures we're never
> > > > > > getting close to this in practice.
> > > > >
> > > > > Bootstrapped and regtested on x86_64-pc-linu-gnu{-m32,}.
> > > >
> > > > This looks reasonable from my side now. Please give Honza the
> > > > chance to chime in.
> > >
> > > Any comments Honza?
> > I'm doing to check in the patch.
>
> It looks reasonable to me, too. I am curious how it will do in
I observed at most 5% improvement on some real workload due to unroll
+ partial_sum of vectorized fma.
> practice. Sorry for late reply - I was off email for 2 weeks of
> vacation and I am catching up now.
>
> Honza
--
BR,
Hongtao
