Skip to content

[mlir][AMDGPU] Use LDS-only MMRA fences for lds_barrier #157919

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Sep 23, 2025
Merged

[mlir][AMDGPU] Use LDS-only MMRA fences for lds_barrier #157919

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
b414fff
[mlir][AMDGPU] Use LDS-only MMRA fences for lds_barrier
krzysz00 Sep 10, 2025
9706ccc
Move to a workgroup fence from a workgroup-one-as fence because thing…
krzysz00 Sep 11, 2025
6639112
Rewrap comment
krzysz00 Sep 12, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 9 additions & 1 deletion mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -543,7 +543,15 @@ struct LDSBarrierOpLowering : public ConvertOpToLLVMPattern<LDSBarrierOp> {

Attribute mmra =
rewriter.getAttr<LLVM::MMRATagAttr>("amdgpu-synchronize-as", "local");
StringRef scope = "workgroup-one-as";
// Note: while there *is* a workgroup-one-as scope, this, when combined with
// the MMRA, will lead to the fence having no effect. This is because
// the codepaths for an atomic load or store will observe that a
// one-address-space atomic to LDS requires no synchronization because
// operations on LDS are totally ordered with respect to each other,
// and so will not emit the correct waitcnt operations that these fences
// are intended to produce. Therefore, we use a broader type of fence
// and rely on the MMRA to relax it to the semantics we want.
StringRef scope = "workgroup";

auto relFence = LLVM::FenceOp::create(rewriter, loc,
LLVM::AtomicOrdering::release, scope);
Expand Down
4 changes: 2 additions & 2 deletions mlir/test/Conversion/AMDGPUToROCDL/amdgpu-to-rocdl.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -416,7 +416,7 @@ func.func @amdgpu_raw_buffer_atomic_cmpswap_v2f16(%src : vector<2xf16>, %cmp : v

// CHECK-LABEL: func @lds_barrier
func.func @lds_barrier() {
// CHECK: llvm.fence syncscope("workgroup-one-as") release {llvm.mmra = #[[$MMRA_TAG]]}
// CHECK: llvm.fence syncscope("workgroup") release {llvm.mmra = #[[$MMRA_TAG]]}
// GFX908: llvm.inline_asm has_side_effects asm_dialect = att
// GFX908-SAME: ";;;WARNING: BREAKS DEBUG WATCHES\0As_barrier"
// GFX90A-NEXT: rocdl.s.barrier
Expand All @@ -425,7 +425,7 @@ func.func @lds_barrier() {
// GFX11-NEXT: rocdl.s.barrier
// GFX12-NEXT: rocdl.s.barrier.signal -1
// GFX12-NEXT: rocdl.s.barrier.wait -1
// CHECK-NEXT: llvm.fence syncscope("workgroup-one-as") acquire {llvm.mmra = #[[$MMRA_TAG]]}
// CHECK-NEXT: llvm.fence syncscope("workgroup") acquire {llvm.mmra = #[[$MMRA_TAG]]}
amdgpu.lds_barrier
func.return
}
Expand Down
Loading