Mesa (main): aco: do not clause NSA instructions
GitLab Mirror
gitlab-mirror at kemper.freedesktop.org
Thu Jun 3 04:04:15 UTC 2021
Module: Mesa
Branch: main
Commit: 81162265b1d7f976c3810b420754ff8403f3aa18
URL: http://cgit.freedesktop.org/mesa/mesa/commit/?id=81162265b1d7f976c3810b420754ff8403f3aa18
Author: Rhys Perry <pendingchaos02 at gmail.com>
Date: Wed May 19 14:36:43 2021 +0100
aco: do not clause NSA instructions
According to LLVM, this has "unpredictable results on GFX10.1".
https://reviews.llvm.org/D102211
fossil-db (Navi10):
Totals from 26690 (17.81% of 149839) affected shaders:
CodeSize: 167935160 -> 167706280 (-0.14%); split: -0.14%, +0.00%
Instrs: 31801427 -> 31744142 (-0.18%); split: -0.18%, +0.00%
Latency: 732672435 -> 732622463 (-0.01%)
InvThroughput: 163361435 -> 163357838 (-0.00%); split: -0.00%, +0.00%
VClause: 546131 -> 546903 (+0.14%); split: -0.00%, +0.14%
Signed-off-by: Rhys Perry <pendingchaos02 at gmail.com>
Reviewed-by: Timur Kristóf <timur.kristof at gmail.com>
Fixes: c353895c922 ("aco: use non-sequential addressing")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/10898>
---
src/amd/compiler/README-ISA.md | 4 ++++
src/amd/compiler/aco_form_hard_clauses.cpp | 14 +++++++++-----
src/amd/compiler/tests/helpers.cpp | 3 +++
src/amd/compiler/tests/test_isel.cpp | 2 +-
4 files changed, 17 insertions(+), 6 deletions(-)
diff --git a/src/amd/compiler/README-ISA.md b/src/amd/compiler/README-ISA.md
index 678759ff25c..a790522ba4f 100644
--- a/src/amd/compiler/README-ISA.md
+++ b/src/amd/compiler/README-ISA.md
@@ -246,3 +246,7 @@ or vice versa: DS instruction, then a branch, then a VMEM/GLOBAL/SCRATCH instruc
Mitigated by:
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"
diff --git a/src/amd/compiler/aco_form_hard_clauses.cpp b/src/amd/compiler/aco_form_hard_clauses.cpp
index 13e9dfadc3d..223a4e2478a 100644
--- a/src/amd/compiler/aco_form_hard_clauses.cpp
+++ b/src/amd/compiler/aco_form_hard_clauses.cpp
@@ -73,14 +73,18 @@ void form_hard_clauses(Program *program)
aco_ptr<Instruction>& instr = block.instructions[i];
clause_type type = clause_other;
- if (instr->isVMEM() && !instr->operands.empty())
+ if (instr->isVMEM() && !instr->operands.empty()) {
+ if (program->chip_class == GFX10 && instr->isMIMG() && get_mimg_nsa_dwords(instr.get()) > 0)
+ type = clause_other;
+ else
+ type = clause_vmem;
+ } else if (instr->isScratch() || instr->isGlobal()) {
type = clause_vmem;
- else if (instr->isScratch() || instr->isGlobal())
- type = clause_vmem;
- else if (instr->isFlat())
+ } else if (instr->isFlat()) {
type = clause_flat;
- else if (instr->isSMEM() && !instr->operands.empty())
+ } else if (instr->isSMEM() && !instr->operands.empty()) {
type = clause_smem;
+ }
if (type != current_type || num_instrs == 64 ||
(num_instrs && !should_form_clause(current_instrs[0].get(), instr.get()))) {
diff --git a/src/amd/compiler/tests/helpers.cpp b/src/amd/compiler/tests/helpers.cpp
index 52f93e09ac4..7f3202fe3eb 100644
--- a/src/amd/compiler/tests/helpers.cpp
+++ b/src/amd/compiler/tests/helpers.cpp
@@ -265,6 +265,9 @@ VkDevice get_vk_device(enum chip_class chip_class)
case GFX10:
family = CHIP_NAVI10;
break;
+ case GFX10_3:
+ family = CHIP_SIENNA_CICHLID;
+ break;
default:
family = CHIP_UNKNOWN;
break;
diff --git a/src/amd/compiler/tests/test_isel.cpp b/src/amd/compiler/tests/test_isel.cpp
index 6434553eadf..a6aae159064 100644
--- a/src/amd/compiler/tests/test_isel.cpp
+++ b/src/amd/compiler/tests/test_isel.cpp
@@ -137,7 +137,7 @@ BEGIN_TEST(isel.gs.no_verts)
END_TEST
BEGIN_TEST(isel.sparse.clause)
- for (unsigned i = GFX10; i <= GFX10; i++) {
+ for (unsigned i = GFX10_3; i <= GFX10_3; i++) {
if (!set_variant((chip_class)i))
continue;
More information about the mesa-commit
mailing list