Skip to content

[flang][cuda] Do not use address cast for src and dst in TMA bulk load #170564

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 2 commits into from
Dec 3, 2025
Merged

[flang][cuda] Do not use address cast for src and dst in TMA bulk load #170564

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
9319733
[flang][cuda] Do not use address cast for src and dst in TMA bulk load
clementval Dec 3, 2025
1217781
Merge branch 'main' into cuf_tma_load_llvmotr
clementval Dec 3, 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
6 changes: 2 additions & 4 deletions flang/lib/Optimizer/Builder/CUDAIntrinsicCall.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1506,10 +1506,8 @@ static void genTMABulkLoad(fir::FirOpBuilder &builder, mlir::Location loc,
auto llvmPtrTy = mlir::LLVM::LLVMPointerType::get(builder.getContext());
barrier = builder.createConvert(loc, llvmPtrTy, barrier);
setAlignment(dst, 16);
dst = convertPtrToNVVMSpace(builder, loc, dst,
mlir::NVVM::NVVMMemorySpace::Shared);
src = convertPtrToNVVMSpace(builder, loc, src,
mlir::NVVM::NVVMMemorySpace::Shared);
dst = builder.createConvert(loc, llvmPtrTy, dst);
src = builder.createConvert(loc, llvmPtrTy, src);
mlir::NVVM::InlinePtxOp::create(
builder, loc, mlir::TypeRange{}, {dst, src, size, barrier}, {},
"cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%0], "
Expand Down
14 changes: 7 additions & 7 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 @@ -543,7 +543,7 @@ end subroutine
! CHECK: %[[ELEM_SIZE:.*]] = arith.constant 8 : i32
! CHECK: %[[SIZE:.*]] = arith.muli %[[COUNT]], %[[ELEM_SIZE]] : i32
! CHECK: %[[BARRIER_PTR:.*]] = fir.convert %[[BARRIER]]#0 : (!fir.ref<i64>) -> !llvm.ptr
! CHECK: nvvm.inline_ptx "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%0], [%1], %2, [%3];" ro(%{{.*}}, %{{.*}}, %[[SIZE]], %[[BARRIER_PTR]] : !llvm.ptr<3>, !llvm.ptr<3>, i32, !llvm.ptr)
! CHECK: nvvm.inline_ptx "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%0], [%1], %2, [%3];" ro(%{{.*}}, %{{.*}}, %[[SIZE]], %[[BARRIER_PTR]] : !llvm.ptr, !llvm.ptr, i32, !llvm.ptr)
! CHECK: nvvm.inline_ptx "mbarrier.expect_tx.relaxed.cta.shared::cta.b64 [%0], %1;" ro(%[[BARRIER_PTR]], %[[SIZE]] : !llvm.ptr, i32)

attributes(global) subroutine test_tma_bulk_load_c8(a, n)
Expand All @@ -563,7 +563,7 @@ end subroutine
! CHECK: %[[ELEM_SIZE:.*]] = arith.constant 16 : i32
! CHECK: %[[SIZE:.*]] = arith.muli %[[COUNT]], %[[ELEM_SIZE]] : i32
! CHECK: %[[BARRIER_PTR:.*]] = fir.convert %[[BARRIER]]#0 : (!fir.ref<i64>) -> !llvm.ptr
! CHECK: nvvm.inline_ptx "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%0], [%1], %2, [%3];" ro(%{{.*}}, %{{.*}}, %[[SIZE]], %[[BARRIER_PTR]] : !llvm.ptr<3>, !llvm.ptr<3>, i32, !llvm.ptr)
! CHECK: nvvm.inline_ptx "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%0], [%1], %2, [%3];" ro(%{{.*}}, %{{.*}}, %[[SIZE]], %[[BARRIER_PTR]] : !llvm.ptr, !llvm.ptr, i32, !llvm.ptr)
! CHECK: nvvm.inline_ptx "mbarrier.expect_tx.relaxed.cta.shared::cta.b64 [%0], %1;" ro(%[[BARRIER_PTR]], %[[SIZE]] : !llvm.ptr, i32)

attributes(global) subroutine test_tma_bulk_load_i4(a, n)
Expand All @@ -583,7 +583,7 @@ end subroutine
! CHECK: %[[ELEM_SIZE:.*]] = arith.constant 4 : i32
! CHECK: %[[SIZE:.*]] = arith.muli %[[COUNT]], %[[ELEM_SIZE]] : i32
! CHECK: %[[BARRIER_PTR:.*]] = fir.convert %[[BARRIER]]#0 : (!fir.ref<i64>) -> !llvm.ptr
! CHECK: nvvm.inline_ptx "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%0], [%1], %2, [%3];" ro(%{{.*}}, %{{.*}}, %[[SIZE]], %[[BARRIER_PTR]] : !llvm.ptr<3>, !llvm.ptr<3>, i32, !llvm.ptr)
! CHECK: nvvm.inline_ptx "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%0], [%1], %2, [%3];" ro(%{{.*}}, %{{.*}}, %[[SIZE]], %[[BARRIER_PTR]] : !llvm.ptr, !llvm.ptr, i32, !llvm.ptr)
! CHECK: nvvm.inline_ptx "mbarrier.expect_tx.relaxed.cta.shared::cta.b64 [%0], %1;" ro(%[[BARRIER_PTR]], %[[SIZE]] : !llvm.ptr, i32)

attributes(global) subroutine test_tma_bulk_load_i8(a, n)
Expand All @@ -603,7 +603,7 @@ end subroutine
! CHECK: %[[ELEM_SIZE:.*]] = arith.constant 8 : i32
! CHECK: %[[SIZE:.*]] = arith.muli %[[COUNT]], %[[ELEM_SIZE]] : i32
! CHECK: %[[BARRIER_PTR:.*]] = fir.convert %[[BARRIER]]#0 : (!fir.ref<i64>) -> !llvm.ptr
! CHECK: nvvm.inline_ptx "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%0], [%1], %2, [%3];" ro(%{{.*}}, %{{.*}}, %[[SIZE]], %[[BARRIER_PTR]] : !llvm.ptr<3>, !llvm.ptr<3>, i32, !llvm.ptr)
! CHECK: nvvm.inline_ptx "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%0], [%1], %2, [%3];" ro(%{{.*}}, %{{.*}}, %[[SIZE]], %[[BARRIER_PTR]] : !llvm.ptr, !llvm.ptr, i32, !llvm.ptr)
! CHECK: nvvm.inline_ptx "mbarrier.expect_tx.relaxed.cta.shared::cta.b64 [%0], %1;" ro(%[[BARRIER_PTR]], %[[SIZE]] : !llvm.ptr, i32)

attributes(global) subroutine test_tma_bulk_load_r2(a, n)
Expand All @@ -623,7 +623,7 @@ end subroutine
! CHECK: %[[ELEM_SIZE:.*]] = arith.constant 2 : i32
! CHECK: %[[SIZE:.*]] = arith.muli %[[COUNT]], %[[ELEM_SIZE]] : i32
! CHECK: %[[BARRIER_PTR:.*]] = fir.convert %[[BARRIER]]#0 : (!fir.ref<i64>) -> !llvm.ptr
! CHECK: nvvm.inline_ptx "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%0], [%1], %2, [%3];" ro(%{{.*}}, %{{.*}}, %[[SIZE]], %[[BARRIER_PTR]] : !llvm.ptr<3>, !llvm.ptr<3>, i32, !llvm.ptr)
! CHECK: nvvm.inline_ptx "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%0], [%1], %2, [%3];" ro(%{{.*}}, %{{.*}}, %[[SIZE]], %[[BARRIER_PTR]] : !llvm.ptr, !llvm.ptr, i32, !llvm.ptr)
! CHECK: nvvm.inline_ptx "mbarrier.expect_tx.relaxed.cta.shared::cta.b64 [%0], %1;" ro(%[[BARRIER_PTR]], %[[SIZE]] : !llvm.ptr, i32)

attributes(global) subroutine test_tma_bulk_load_r4(a, n)
Expand All @@ -643,7 +643,7 @@ end subroutine
! CHECK: %[[ELEM_SIZE:.*]] = arith.constant 4 : i32
! CHECK: %[[SIZE:.*]] = arith.muli %[[COUNT]], %[[ELEM_SIZE]] : i32
! CHECK: %[[BARRIER_PTR:.*]] = fir.convert %[[BARRIER]]#0 : (!fir.ref<i64>) -> !llvm.ptr
! CHECK: nvvm.inline_ptx "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%0], [%1], %2, [%3];" ro(%{{.*}}, %{{.*}}, %[[SIZE]], %[[BARRIER_PTR]] : !llvm.ptr<3>, !llvm.ptr<3>, i32, !llvm.ptr)
! CHECK: nvvm.inline_ptx "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%0], [%1], %2, [%3];" ro(%{{.*}}, %{{.*}}, %[[SIZE]], %[[BARRIER_PTR]] : !llvm.ptr, !llvm.ptr, i32, !llvm.ptr)
! CHECK: nvvm.inline_ptx "mbarrier.expect_tx.relaxed.cta.shared::cta.b64 [%0], %1;" ro(%[[BARRIER_PTR]], %[[SIZE]] : !llvm.ptr, i32)

attributes(global) subroutine test_tma_bulk_load_r8(a, n)
Expand All @@ -663,7 +663,7 @@ end subroutine
! CHECK: %[[ELEM_SIZE:.*]] = arith.constant 8 : i32
! CHECK: %[[SIZE:.*]] = arith.muli %[[COUNT]], %[[ELEM_SIZE]] : i32
! CHECK: %[[BARRIER_PTR:.*]] = fir.convert %[[BARRIER]]#0 : (!fir.ref<i64>) -> !llvm.ptr
! CHECK: nvvm.inline_ptx "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%0], [%1], %2, [%3];" ro(%{{.*}}, %{{.*}}, %[[SIZE]], %[[BARRIER_PTR]] : !llvm.ptr<3>, !llvm.ptr<3>, i32, !llvm.ptr)
! CHECK: nvvm.inline_ptx "cp.async.bulk.shared::cluster.global.mbarrier::complete_tx::bytes [%0], [%1], %2, [%3];" ro(%{{.*}}, %{{.*}}, %[[SIZE]], %[[BARRIER_PTR]] : !llvm.ptr, !llvm.ptr, i32, !llvm.ptr)
! CHECK: nvvm.inline_ptx "mbarrier.expect_tx.relaxed.cta.shared::cta.b64 [%0], %1;" ro(%[[BARRIER_PTR]], %[[SIZE]] : !llvm.ptr, i32)

attributes(global) subroutine test_tma_bulk_store_c4(c, n)
Expand Down