Skip to content

[flang][cuda][NFC] Use NVVM operation for thread syncs #166999

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 1 commit into from
Nov 7, 2025
Merged

[flang][cuda][NFC] Use NVVM operation for thread syncs #166999

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
24 changes: 6 additions & 18 deletions flang/lib/Optimizer/Builder/CUDAIntrinsicCall.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1309,34 +1309,22 @@ CUDAIntrinsicLibrary::genThisWarp(mlir::Type resultType,
// THREADFENCE
void CUDAIntrinsicLibrary::genThreadFence(
llvm::ArrayRef<fir::ExtendedValue> args) {
constexpr llvm::StringLiteral funcName = "llvm.nvvm.membar.gl";
mlir::FunctionType funcType =
mlir::FunctionType::get(builder.getContext(), {}, {});
auto funcOp = builder.createFunction(loc, funcName, funcType);
llvm::SmallVector<mlir::Value> noArgs;
fir::CallOp::create(builder, loc, funcOp, noArgs);
assert(args.size() == 0);
mlir::NVVM::MembarOp::create(builder, loc, mlir::NVVM::MemScopeKind::GPU);
}

// THREADFENCE_BLOCK
void CUDAIntrinsicLibrary::genThreadFenceBlock(
llvm::ArrayRef<fir::ExtendedValue> args) {
constexpr llvm::StringLiteral funcName = "llvm.nvvm.membar.cta";
mlir::FunctionType funcType =
mlir::FunctionType::get(builder.getContext(), {}, {});
auto funcOp = builder.createFunction(loc, funcName, funcType);
llvm::SmallVector<mlir::Value> noArgs;
fir::CallOp::create(builder, loc, funcOp, noArgs);
assert(args.size() == 0);
mlir::NVVM::MembarOp::create(builder, loc, mlir::NVVM::MemScopeKind::CTA);
}

// THREADFENCE_SYSTEM
void CUDAIntrinsicLibrary::genThreadFenceSystem(
llvm::ArrayRef<fir::ExtendedValue> args) {
constexpr llvm::StringLiteral funcName = "llvm.nvvm.membar.sys";
mlir::FunctionType funcType =
mlir::FunctionType::get(builder.getContext(), {}, {});
auto funcOp = builder.createFunction(loc, funcName, funcType);
llvm::SmallVector<mlir::Value> noArgs;
fir::CallOp::create(builder, loc, funcOp, noArgs);
assert(args.size() == 0);
mlir::NVVM::MembarOp::create(builder, loc, mlir::NVVM::MemScopeKind::SYS);
}

// TMA_BULK_COMMIT_GROUP
Expand Down
6 changes: 0 additions & 6 deletions flang/test/Lower/CUDA/cuda-device-proc.cuf
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,9 +22,6 @@ attributes(global) subroutine devsub()

call syncthreads()
call syncwarp(1)
call threadfence()
call threadfence_block()
call threadfence_system()
ret = syncthreads_and(1)
res = syncthreads_and(tid > offset)
ret = syncthreads_count(1)
Expand Down Expand Up @@ -106,9 +103,6 @@ end
! CHECK-LABEL: func.func @_QPdevsub() attributes {cuf.proc_attr = #cuf.cuda_proc<global>}
! CHECK: nvvm.barrier0
! CHECK: nvvm.bar.warp.sync %c1{{.*}} : i32
! CHECK: fir.call @llvm.nvvm.membar.gl() fastmath<contract> : () -> ()
! CHECK: fir.call @llvm.nvvm.membar.cta() fastmath<contract> : () -> ()
! CHECK: fir.call @llvm.nvvm.membar.sys() fastmath<contract> : () -> ()
! CHECK: %{{.*}} = fir.call @llvm.nvvm.barrier0.and(%c1{{.*}}) fastmath<contract> : (i32) -> i32
! CHECK: %[[A:.*]] = fir.load %{{.*}} : !fir.ref<i32>
! CHECK: %[[B:.*]] = fir.load %{{.*}} : !fir.ref<i32>
Expand Down
14 changes: 14 additions & 0 deletions flang/test/Lower/CUDA/cuda-synchronization.cuf
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,14 @@
! RUN: bbc -emit-hlfir -fcuda %s -o - | FileCheck %s

! Test CUDA Fortran instrinsics lowerings for synchronization.

attributes(global) subroutine sync()
call threadfence()
call threadfence_block()
call threadfence_system()
end subroutine

! CHECK-LABEL: func.func @_QPsync() attributes {cuf.proc_attr = #cuf.cuda_proc<global>}
! CHECK: nvvm.memory.barrier <gpu>
! CHECK: nvvm.memory.barrier <cta>
! CHECK: nvvm.memory.barrier <sys>