[Mesa-dev] [PATCH v2] nir: Handle variables dependent on the local work group size.

[Karol Herbst]

 an)
On Mon, Nov 12, 2018 at 12:37 AM Jason Ekstrand <jason at jlekstrand.net> wrote:
>
> On November 11, 2018 16:36:16 Karol Herbst <kherbst at redhat.com> wrote:
>
> > On Sun, Nov 11, 2018 at 10:48 PM Jason Ekstrand <jason at jlekstrand.net> wrote:
> >>
> >> On Sun, Nov 11, 2018 at 3:35 PM Plamena Manolova
> >> <plamena.n.manolova at gmail.com> wrote:
> >>>
> >>>
> >>> Lowering shader variables which depend on the local work group
> >>> size being available in nir_lower_system_values is only possible
> >>> if the local work group size isn't variable, otherwise this should
> >>> be handled by the native driver (if it supports
> >>> ARB_compute_variable_group_size).
> >>>
> >>>
> >>> Signed-off-by: Plamena Manolova <plamena.n.manolova at gmail.com>
> >>> ---
> >>> src/compiler/nir/nir_lower_system_values.c | 22 ++++++++++++++++++++++
> >>> 1 file changed, 22 insertions(+)
> >>>
> >>>
> >>> diff --git a/src/compiler/nir/nir_lower_system_values.c
> >>> b/src/compiler/nir/nir_lower_system_values.c
> >>> index bde7eb1180..6fdf9f9c51 100644
> >>> --- a/src/compiler/nir/nir_lower_system_values.c
> >>> +++ b/src/compiler/nir/nir_lower_system_values.c
> >>> @@ -77,6 +77,15 @@ convert_block(nir_block *block, nir_builder *b)
> >>>    *    "The value of gl_GlobalInvocationID is equal to
> >>>    *    gl_WorkGroupID * gl_WorkGroupSize + gl_LocalInvocationID"
> >>>    */
> >>> +
> >>> +         /*
> >>> +          * If the local work group size is variable we can't lower the global
> >>> +          * invocation id here.
> >>
> >>
> >> Why not?
> >>
> >>>
> >>>
> >>> +          */
> >>> +         if (b->shader->info.cs.local_size_variable) {
> >>
> >>
> >> I didn't realize we'd added a bit for this.  At one point in time, Karol
> >> had patches which had it keyed off of the size being zero.  Having a
> >> separate bit is probably fine, it just surpised me.
> >
> >
> > yeah.. I guess I choose that, because I had nothing better. But I
> > guess having the size being 0 is good enough as long as we sure it is
> > 0 in relevant cases.
>
> It's fine. I just thought we'd chosen a different convention bit there's
> nothing wrong with it. I was just surprised. That's all.
>
> >
> >>>
> >>>
> >>> +            break;
> >>> +         }
> >>> +
> >>>   nir_ssa_def *group_size = build_local_group_size(b);
> >>>   nir_ssa_def *group_id = nir_load_work_group_id(b);
> >>>   nir_ssa_def *local_id = nir_load_local_invocation_id(b);
> >>> @@ -115,6 +124,11 @@ convert_block(nir_block *block, nir_builder *b)
> >>> }
> >>>
> >>>
> >>> case SYSTEM_VALUE_LOCAL_GROUP_SIZE: {
> >>> +         /* If the local work group size is variable we can't lower it here */
> >>> +         if (b->shader->info.cs.local_size_variable) {
> >>> +            break;
> >>> +         }
> >>> +
> >>>   sysval = build_local_group_size(b);
> >>>   break;
> >>> }
> >>> @@ -189,6 +203,14 @@ convert_block(nir_block *block, nir_builder *b)
> >>>   break;
> >>>
> >>>
> >>> case SYSTEM_VALUE_GLOBAL_GROUP_SIZE: {
> >>> +         /*
> >>> +          * If the local work group size is variable we can't lower the global
> >>> +          * group size here.
> >>> +          */
> >>> +         if (b->shader->info.cs.local_size_variable) {
> >>> +            break;
> >>> +         }
> >>
> >>
> >> Why can't we lower the global size?  It seems like we would want the below
> >> multiplication regardless of whether the local size is constant.
> >
> > well I am not sure about ARB_compute_variable_group_size, but at least
> > in CL you know nothing about it at compile time as you specify
> > everything when you enqueue the kernel. Could be that the number of
> > work_groups is fixed with ARB_compute_variable_group_size though
>
> Why can't you multiply two non-constant things together? Maybe the driver
> can do something now efficient but it's not clear to me that this isn't
> what we want.
>

oh, you meant it that way. Mhh, I actually wrote the patch adding
build_local_group_size to handle that inside that function (lower to
constants if known, intrinsics otherwise) as we need the local_size
for three different system values. I have a patch for that for CL, but
the patch itself is a bit hacky right now (as in, it doesn't handle it
inside build_local_group_size).

And I think that patch misses handling for
SYSTEM_VALUE_GLOBAL_GROUP_SIZE anyhow.

For CL we even want to add a global_invocation_id_offset system value
and add it when handling SYSTEM_VALUE_GLOBAL_INVOCATION_ID, but that
we can ignore for now.

Anyhow, I think it would make sense to move the check into
build_local_group_sized and either return a constant value or an
intrinsic loading the value from the driver. I just remember vaguely
that I think somebody brought that up months ago and there was a good
reason not to do that for every driver?

> >
> >>>
> >>>
> >>> +
> >>>   nir_ssa_def *group_size = build_local_group_size(b);
> >>>   nir_ssa_def *num_work_groups = nir_load_num_work_groups(b);
> >>>   sysval = nir_imul(b, group_size, num_work_groups);
> >>> --
> >>> 2.11.0
> >>>
> >>>
> >>> _______________________________________________
> >>> mesa-dev mailing list
> >>> mesa-dev at lists.freedesktop.org
> >>> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
>
>
>