aco: don't create 4 and 5 dword NSA instructions on GFX10

"stability issues", apparently: https://reviews.llvm.org/D103348

fossil-db (Navi10):
Totals from 4512 (3.01% of 149839) affected shaders:
VGPRs: 221516 -> 223308 (+0.81%); split: -0.07%, +0.88%
CodeSize: 23000080 -> 23070672 (+0.31%); split: -0.08%, +0.39%
MaxWaves: 107718 -> 107496 (-0.21%); split: +0.11%, -0.32%
Instrs: 4321890 -> 4362822 (+0.95%); split: -0.00%, +0.95%
Latency: 71495710 -> 71581476 (+0.12%); split: -0.07%, +0.19%
InvThroughput: 11858568 -> 11938960 (+0.68%); split: -0.00%, +0.68%
VClause: 76575 -> 76585 (+0.01%); split: -0.05%, +0.07%
SClause: 168771 -> 168709 (-0.04%); split: -0.06%, +0.02%
Copies: 182305 -> 221948 (+21.75%); split: -0.00%, +21.75%
PreVGPRs: 194657 -> 195635 (+0.50%); split: -0.00%, +0.50%

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Fixes: c353895c92 ("aco: use non-sequential addressing")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10898>
This commit is contained in:
Rhys Perry 2021-06-02 16:30:35 +01:00 committed by Marge Bot
parent bb52484df5
commit 903f814b78
2 changed files with 10 additions and 1 deletions

View File

@ -250,3 +250,8 @@ Only `s_waitcnt_vscnt null, 0`. Needed even if the first instruction is a load.
### NSAClauseBug
"MIMG-NSA in a hard clause has unpredictable results on GFX10.1"
### NSAMaxSize5
NSA MIMG instructions should be limited to 3 dwords before GFX10.3 to avoid
stability issues: https://reviews.llvm.org/D103348

View File

@ -5544,7 +5544,11 @@ static MIMG_instruction *emit_mimg(Builder& bld, aco_opcode op,
unsigned wqm_mask=0,
Operand vdata=Operand(v1))
{
if (bld.program->chip_class < GFX10) {
/* Limit NSA instructions to 3 dwords on GFX10 to avoid stability issues. */
unsigned max_nsa_size = bld.program->chip_class >= GFX10_3 ? 13 : 5;
bool use_nsa = bld.program->chip_class >= GFX10 && coords.size() <= max_nsa_size;
if (!use_nsa) {
Temp coord = coords[0];
if (coords.size() > 1) {
coord = bld.tmp(RegType::vgpr, coords.size());