Mesa (main): aco/scheduler: allow moving down VMEM stores to below VMEM loads

[GitLab Mirror]

Module: Mesa
Branch: main
Commit: b23a9dd1f63684d1a19f58b808a1fe3c3dbca7ab
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=b23a9dd1f63684d1a19f58b808a1fe3c3dbca7ab

Author: Rhys Perry <pendingchaos02 at gmail.com>
Date:   Thu Aug  5 11:42:39 2021 +0100

aco/scheduler: allow moving down VMEM stores to below VMEM loads

fossil-db (Vega10):
Totals from 93 (0.06% of 150305) affected shaders:
SGPRs: 4832 -> 4768 (-1.32%)
VGPRs: 4084 -> 4144 (+1.47%)
CodeSize: 316080 -> 317208 (+0.36%); split: -0.11%, +0.47%
MaxWaves: 589 -> 580 (-1.53%)
Instrs: 60229 -> 60511 (+0.47%); split: -0.15%, +0.61%
Latency: 636477 -> 540029 (-15.15%); split: -15.26%, +0.10%
InvThroughput: 293027 -> 283043 (-3.41%); split: -4.21%, +0.80%
VClause: 2557 -> 2716 (+6.22%); split: -0.35%, +6.57%
SClause: 1381 -> 1395 (+1.01%); split: -0.14%, +1.16%
Copies: 9424 -> 9728 (+3.23%); split: -0.74%, +3.97%

fossil-db (Sienna Cichlid):
Totals from 88 (0.06% of 150170) affected shaders:
VGPRs: 3840 -> 3872 (+0.83%)
CodeSize: 300544 -> 300960 (+0.14%); split: -0.09%, +0.23%
Instrs: 53714 -> 53871 (+0.29%); split: -0.05%, +0.35%
Latency: 489854 -> 462001 (-5.69%); split: -6.30%, +0.61%
InvThroughput: 100307 -> 95142 (-5.15%); split: -5.50%, +0.35%
VClause: 2322 -> 2564 (+10.42%); split: -0.39%, +10.81%
SClause: 1345 -> 1358 (+0.97%); split: -0.15%, +1.12%
Copies: 4113 -> 4351 (+5.79%); split: -0.66%, +6.44%

Signed-off-by: Rhys Perry <pendingchaos02 at gmail.com>
Reviewed-by: Daniel Schürmann <daniel at schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12211>

---

 src/amd/compiler/aco_scheduler.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/src/amd/compiler/aco_scheduler.cpp b/src/amd/compiler/aco_scheduler.cpp
index 9b4c9ffa48c..2f72ba0cc96 100644
--- a/src/amd/compiler/aco_scheduler.cpp
+++ b/src/amd/compiler/aco_scheduler.cpp
@@ -826,7 +826,7 @@ schedule_VMEM(sched_ctx& ctx, Block* block, std::vector<RegisterDemand>& registe
       }
 
       /* if current depends on candidate, add additional dependencies and continue */
-      bool can_move_down = !is_vmem || part_of_clause;
+      bool can_move_down = !is_vmem || part_of_clause || candidate->definitions.empty();
 
       HazardResult haz =
          perform_hazard_query(part_of_clause ? &clause_hq : &indep_hq, candidate.get(), false);