[Mesa-dev] [PATCH 2/2] i965: Implement ARB_compute_variable_group_size.

[Plamena Manolova]

On Mon, 4 Jun 2018 at 21:36, Ilia Mirkin <imirkin at alum.mit.edu> wrote:

> On Mon, Jun 4, 2018 at 4:30 PM, Plamena Manolova
> <plamena.n.manolova at gmail.com> wrote:
> > Thank you for the review Ilia!
> >
> > On Fri, 1 Jun 2018 at 23:44, Ilia Mirkin <imirkin at alum.mit.edu> wrote:
> >>
> >> On Fri, Jun 1, 2018 at 6:21 PM, Plamena Manolova
> >> <plamena.n.manolova at gmail.com> wrote:
> >> > +      /*
> >> > +       * If the local work group size is variable we have to use a
> >> > dispatch
> >> > +       * width of 32 here, since at this point we don't know the
> actual
> >> > size of
> >> > +       * the workload.
> >> > +       */
> >> > +      min_dispatch_width = 32;
> >>
> >> Is that a good idea? You are able to specify a different maximum when
> >> using a variable size (MAX_COMPUTE_VARIABLE_GROUP_INVOCATIONS_ARB)
> >> s.t. this is 16 (or even 8, although that may be too few for practical
> >> use) -- that way you would just set the max to 768 or whatever on
> >> gen8+.
> >
> >
> > That's a good point, MAX_COMPUTE_VARIABLE_GROUP_INVOCATIONS_ARB is the
> > same on all platforms, so it makes sense to use simd16 instead. Thank you
> > for noticing that.
>
> Well IIRC gen7/gen7.5 can do something like 1536 invocations with
> SIMD16, while gen8 are down to 768. Either way, not forcing SIMD32 may
> be nice -- but that's something for you Intel folk to decide. I just
> wanted to point out that you could have different max's for "regular"
> compute (where the min max is 1024) and variable-size groups.
>

Ah, I see what you mean now. I took a closer look at
MAX_COMPUTE_VARIABLE_GROUP_INVOCATIONS_ARB and it's
related to the maximum number of CS threads (which of course varies).


>   -ilia
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20180604/8fde2021/attachment.html>