fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-07 00:38:48 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 4

pan/va: don't merge workgroups when subgroups are used

Browse source 
Vulkan guarantees that all subgroup invocations will be part of the same
workgroup, so we need to disable merging workgroups for shaders where
the subgroup layout is observable.

Signed-off-by: Olivia Lee <olivia.lee@collabora.com>
Reviewed-by: Eric R. Smith <eric.smith@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/38586>
This commit is contained in:
Olivia Lee 2025-11-20 02:54:04 -08:00 • committed by Marge Bot
parent a0f6c6d84d
commit c42e124a66
1 changed files with 27 additions and 3 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

30
src/panfrost/compiler/bifrost/bifrost_nir.c
View file

@ -576,9 +576,15 @@ bi_lower_subgroups(nir_builder *b, nir_intrinsic_instr *intr, void *data)
}
/* Workgroups may be merged if the structure of the workgroup is not software
* visible. This is true if neither shared memory nor BARRIER instructions are
* used. The hardware may be able to optimize compute shaders that set this
* flag. */
* visible. This is true if neither shared memory nor BARRIER instructions nor
* subgroups are used. The hardware may be able to optimize compute shaders
* that set this flag.
*
* From the vulkan spec version 1.4.317 section 9.25.8:
*
* "For shaders that have defined workgroups, each invocation in a subgroup
* must be in the same local workgroup."
*/
bool
valhall_can_merge_workgroups(nir_shader *nir)
{
@ -599,6 +605,24 @@ valhall_can_merge_workgroups(nir_shader *nir)
if (intrin->intrinsic == nir_intrinsic_barrier &&
nir_intrinsic_execution_scope(intrin) == SCOPE_WORKGROUP)
return false;
/* This is in nir->info.uses_wide_subgroups, but we don't want to
* force an extra nir_gather_shader_info call. */
if (nir_intrinsic_has_semantic(intrin, NIR_INTRINSIC_SUBGROUP))
return false;
/* load_subgroup_invocation allows observing merged workgroups
* because the first thread in the workgroup may have a nonzero
* subgroup invocation and so on. We don't have to care about
* load_subgroup_id, because we implement it by dividing the local
* invocation id, so it doesn't care what the actual subgroup
* layout is in hw.
*
* Note that these intrinsics do not have NIR_INTRINSIC_SUBGROUP
* because they do not perform any communication with other
* subgroup threads. */
if (intrin->intrinsic == nir_intrinsic_load_subgroup_invocation)
return false;
}
}
}