Mesa (main): aco: stop scheduling if clause-forming fails

[GitLab Mirror]

Module: Mesa
Branch: main
Commit: 903999c11964e3b871c5da3e7ec81902144739c6
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=903999c11964e3b871c5da3e7ec81902144739c6

Author: Daniel Schürmann <daniel at schuermann.dev>
Date:   Tue May 18 22:33:30 2021 +0200

aco: stop scheduling if clause-forming fails

This avoids unintended reordering of VMEM instructions.
It is also highly unlikely that we find more independent
instructions before previous clause-related instructions.

Totals from 1921 (1.28% of 150170) affected shaders: (GFX10.3)
VGPRs: 103832 -> 103736 (-0.09%); split: -0.10%, +0.01%
CodeSize: 8695560 -> 8706000 (+0.12%); split: -0.03%, +0.15%
Instrs: 1643752 -> 1646349 (+0.16%); split: -0.04%, +0.20%
Latency: 26755527 -> 26614645 (-0.53%); split: -0.67%, +0.14%
InvThroughput: 7226604 -> 7204809 (-0.30%); split: -0.39%, +0.08%
VClause: 46536 -> 46201 (-0.72%); split: -0.81%, +0.09%
SClause: 47910 -> 47769 (-0.29%); split: -0.43%, +0.14%
Copies: 94647 -> 94558 (-0.09%); split: -0.26%, +0.17%
Branches: 36843 -> 36847 (+0.01%); split: -0.00%, +0.01%

Reviewed-by: Rhys Perry <pendingchaos02 at gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10896>

---

 src/amd/compiler/aco_scheduler.cpp | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/src/amd/compiler/aco_scheduler.cpp b/src/amd/compiler/aco_scheduler.cpp
index 2f72ba0cc96..ebf5fedfdd7 100644
--- a/src/amd/compiler/aco_scheduler.cpp
+++ b/src/amd/compiler/aco_scheduler.cpp
@@ -838,6 +838,8 @@ schedule_VMEM(sched_ctx& ctx, Block* block, std::vector<RegisterDemand>& registe
          break;
 
       if (!can_move_down) {
+         if (part_of_clause)
+            break;
          add_to_hazard_query(&indep_hq, candidate.get());
          add_to_hazard_query(&clause_hq, candidate.get());
          ctx.mv.downwards_skip(cursor);
@@ -847,6 +849,8 @@ schedule_VMEM(sched_ctx& ctx, Block* block, std::vector<RegisterDemand>& registe
       Instruction* candidate_ptr = candidate.get();
       MoveResult res = ctx.mv.downwards_move(cursor, part_of_clause);
       if (res == move_fail_ssa || res == move_fail_rar) {
+         if (part_of_clause)
+            break;
          add_to_hazard_query(&indep_hq, candidate.get());
          add_to_hazard_query(&clause_hq, candidate.get());
          ctx.mv.downwards_skip(cursor);