aco/gfx11.5: skip dealloc_vgprs for stages with exports

fossil-db (gfx1150):
Totals from 72997 (91.94% of 79395) affected shaders:
Instrs: 35047397 -> 34679136 (-1.05%)
CodeSize: 182414292 -> 180934440 (-0.81%)
Latency: 225452700 -> 225045411 (-0.18%)
InvThroughput: 40416910 -> 40395608 (-0.05%); split: -0.06%, +0.00%
Branches: 604968 -> 601660 (-0.55%)

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/30241>
This commit is contained in:
Rhys Perry 2024-07-17 14:49:01 +01:00 committed by Marge Bot
parent 492d99fc6a
commit 3b732340ec

View file

@ -1399,6 +1399,14 @@ dealloc_vgprs(Program* program)
if (uses_scratch(program))
return false;
/* If we insert the sendmsg on GFX11.5, the export priority workaround will require us to insert
* a wait after exports. There might still be pending VMEM stores for PS parameter exports,
* except NGG lowering usually inserts a memory barrier. This means there is unlikely to be any
* pending VMEM stores or exports if we insert the sendmsg for these stages. */
if (program->gfx_level == GFX11_5 && (program->stage.hw == AC_HW_NEXT_GEN_GEOMETRY_SHADER ||
program->stage.hw == AC_HW_PIXEL_SHADER))
return false;
Block& block = program->blocks.back();
/* don't bother checking if there is a pending VMEM store or export: there almost always is */