Mesa (main): aco: make clause-forming depend on the number of moved instructions

[GitLab Mirror]

Module: Mesa
Branch: main
Commit: 7e1faf9349b04d05deb8900deefa4014e1ccaa5f
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=7e1faf9349b04d05deb8900deefa4014e1ccaa5f

Author: Daniel Schürmann <daniel at schuermann.dev>
Date:   Wed May 19 11:25:18 2021 +0200

aco: make clause-forming depend on the number of moved instructions

This allows more aggressive clause-forming in presence of
larger def-use distances. To compensate for the effect,
VMEM_CLAUSE_MAX_GRAB_DIST was decreased.

Totals from 5788 (3.85% of 150170) affected shaders: (GFX10.3)
VGPRs: 483960 -> 475272 (-1.80%); split: -1.82%, +0.02%
CodeSize: 59661240 -> 59669084 (+0.01%); split: -0.01%, +0.02%
MaxWaves: 70408 -> 71450 (+1.48%); split: +1.51%, -0.03%
Instrs: 11222417 -> 11224479 (+0.02%); split: -0.01%, +0.03%
Latency: 349397104 -> 349298602 (-0.03%); split: -0.03%, +0.00%
InvThroughput: 88584832 -> 87762262 (-0.93%); split: -0.93%, +0.00%
VClause: 168905 -> 177089 (+4.85%); split: -0.48%, +5.32%
SClause: 375795 -> 375767 (-0.01%); split: -0.01%, +0.01%
Copies: 840298 -> 840231 (-0.01%); split: -0.04%, +0.03%
Branches: 373265 -> 373278 (+0.00%); split: -0.00%, +0.00%

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, 2 insertions(+), 2 deletions(-)

diff --git a/src/amd/compiler/aco_scheduler.cpp b/src/amd/compiler/aco_scheduler.cpp
index ebf5fedfdd7..42b830826a4 100644
--- a/src/amd/compiler/aco_scheduler.cpp
+++ b/src/amd/compiler/aco_scheduler.cpp
@@ -37,7 +37,7 @@
 #define SMEM_MAX_MOVES      (64 - ctx.num_waves * 4)
 #define VMEM_MAX_MOVES      (256 - ctx.num_waves * 16)
 /* creating clauses decreases def-use distances, so make it less aggressive the lower num_waves is */
-#define VMEM_CLAUSE_MAX_GRAB_DIST (ctx.num_waves * 8)
+#define VMEM_CLAUSE_MAX_GRAB_DIST (ctx.num_waves * 4)
 #define POS_EXP_MAX_MOVES         512
 
 namespace aco {
@@ -822,7 +822,7 @@ schedule_VMEM(sched_ctx& ctx, Block* block, std::vector<RegisterDemand>& registe
          /* We can't easily tell how much this will decrease the def-to-use
           * distances, so just use how far it will be moved as a heuristic. */
          part_of_clause =
-            grab_dist < clause_max_grab_dist && should_form_clause(current, candidate.get());
+            grab_dist < clause_max_grab_dist + k && should_form_clause(current, candidate.get());
       }
 
       /* if current depends on candidate, add additional dependencies and continue */