From 6ea9443726fff6921d548faae3c1afa8cedc6a4c Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Daniel=20Sch=C3=BCrmann?= <daniel@schuermann.dev>
Date: Wed, 24 Jul 2024 11:48:47 +0200
Subject: [PATCH] aco/scheduler: stop rounding down the target number of waves
 on GFX10+

This way, it can make use of uneven wave numbers.

Totals from 4078 (5.14% of 79395) affected shaders: (Navi21)
MaxWaves: 58715 -> 65460 (+11.49%); split: +11.49%, -0.01%
Instrs: 5033684 -> 5048244 (+0.29%); split: -0.09%, +0.38%
CodeSize: 26833884 -> 26898780 (+0.24%); split: -0.07%, +0.32%
VGPRs: 302360 -> 265312 (-12.25%); split: -12.26%, +0.01%
Latency: 34636448 -> 36044242 (+4.06%); split: -0.08%, +4.14%
InvThroughput: 7999403 -> 7662697 (-4.21%); split: -4.55%, +0.34%
VClause: 105403 -> 111996 (+6.26%); split: -0.40%, +6.66%
SClause: 132996 -> 133460 (+0.35%); split: -0.81%, +1.16%
Copies: 297036 -> 308122 (+3.73%); split: -0.64%, +4.37%
Branches: 89376 -> 89390 (+0.02%); split: -0.00%, +0.02%
VALU: 3477621 -> 3488510 (+0.31%); split: -0.05%, +0.36%
SALU: 484211 -> 484191 (-0.00%); split: -0.08%, +0.08%

Totals from 1840 (2.32% of 79395) affected shaders: (Navi31)

MaxWaves: 30714 -> 34182 (+11.29%)
Instrs: 3102955 -> 3131001 (+0.90%); split: -0.05%, +0.95%
CodeSize: 16160564 -> 16273100 (+0.70%); split: -0.04%, +0.74%
VGPRs: 174540 -> 150600 (-13.72%)
Latency: 23521914 -> 24515055 (+4.22%); split: -0.07%, +4.29%
InvThroughput: 4373397 -> 4202912 (-3.90%); split: -4.40%, +0.50%
VClause: 59087 -> 64091 (+8.47%); split: -0.24%, +8.71%
SClause: 74844 -> 75366 (+0.70%); split: -0.53%, +1.22%
Copies: 184396 -> 197747 (+7.24%); split: -0.25%, +7.49%
Branches: 46015 -> 46028 (+0.03%); split: -0.00%, +0.03%
VALU: 1929286 -> 1942709 (+0.70%); split: -0.02%, +0.71%
SALU: 216126 -> 215983 (-0.07%); split: -0.18%, +0.12%
VOPD: 1216 -> 1217 (+0.08%); split: +1.40%, -1.32%
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33644>
---
 src/amd/compiler/aco_scheduler.cpp | 13 ++++---------
 1 file changed, 4 insertions(+), 9 deletions(-)

diff --git a/src/amd/compiler/aco_scheduler.cpp b/src/amd/compiler/aco_scheduler.cpp
index dfab79f5d85..896452e9d6c 100644
--- a/src/amd/compiler/aco_scheduler.cpp
+++ b/src/amd/compiler/aco_scheduler.cpp
@@ -1246,7 +1246,6 @@ schedule_program(Program* program)
    /* Allowing the scheduler to reduce the number of waves to as low as 5
     * improves performance of Thrones of Britannia significantly and doesn't
     * seem to hurt anything else. */
-   // TODO: account for possible uneven num_waves on GFX10+
    unsigned wave_fac = program->dev.physical_vgprs / 256;
    if (program->num_waves <= 5 * wave_fac)
       ctx.num_waves = program->num_waves;
@@ -1260,17 +1259,13 @@ schedule_program(Program* program)
    ctx.num_waves = std::min<uint16_t>(ctx.num_waves, program->num_waves);
    ctx.num_waves = max_suitable_waves(program, ctx.num_waves);
 
+   assert(ctx.num_waves >= program->min_waves);
+   ctx.mv.max_registers = {int16_t(get_addr_vgpr_from_waves(program, ctx.num_waves) - 2),
+                           int16_t(get_addr_sgpr_from_waves(program, ctx.num_waves))};
+
    /* VMEM_MAX_MOVES and such assume pre-GFX10 wave count */
    ctx.num_waves = std::max<uint16_t>(ctx.num_waves / wave_fac, 1);
 
-   assert(ctx.num_waves > 0);
-   ctx.mv.max_registers = {
-      int16_t(get_addr_vgpr_from_waves(
-                 program, std::max<uint16_t>(ctx.num_waves * wave_fac, program->min_waves)) -
-              2),
-      int16_t(get_addr_sgpr_from_waves(
-         program, std::max<uint16_t>(ctx.num_waves * wave_fac, program->min_waves)))};
-
    /* NGG culling shaders are very sensitive to position export scheduling.
     * Schedule less aggressively when early primitive export is used, and
     * keep the position export at the very bottom when late primitive export is used.