Skip to content

[flang][cuda] Support data transfer with conversion #153242

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
clementval merged 1 commit into from
Aug 13, 2025
+87 −2
Merged

[flang][cuda] Support data transfer with conversion #153242

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
2 changes: 2 additions & 0 deletions flang/include/flang/Lower/CUDA.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -62,6 +62,8 @@ cuf::DataAttributeAttr
translateSymbolCUFDataAttribute(mlir::MLIRContext *mlirContext,
const Fortran::semantics::Symbol &sym);

bool isTransferWithConversion(mlir::Value rhs);

} // end namespace Fortran::lower

#endif // FORTRAN_LOWER_CUDA_H
30 changes: 28 additions & 2 deletions flang/lib/Lower/Bridge.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4827,7 +4827,9 @@ class FirConverter : public Fortran::lower::AbstractConverter {

void genCUDADataTransfer(fir::FirOpBuilder &builder, mlir::Location loc,
const Fortran::evaluate::Assignment &assign,
hlfir::Entity &lhs, hlfir::Entity &rhs) {
hlfir::Entity &lhs, hlfir::Entity &rhs,
bool isWholeAllocatableAssignment,
bool keepLhsLengthInAllocatableAssignment) {
bool lhsIsDevice = Fortran::evaluate::HasCUDADeviceAttrs(assign.lhs);
bool rhsIsDevice = Fortran::evaluate::HasCUDADeviceAttrs(assign.rhs);

Expand Down Expand Up @@ -4892,6 +4894,28 @@ class FirConverter : public Fortran::lower::AbstractConverter {

// host = device
if (!lhsIsDevice && rhsIsDevice) {
if (Fortran::lower::isTransferWithConversion(rhs)) {
mlir::OpBuilder::InsertionGuard insertionGuard(builder);
auto elementalOp =
mlir::dyn_cast<hlfir::ElementalOp>(rhs.getDefiningOp());
assert(elementalOp && "expect elemental op");
auto designateOp =
*elementalOp.getBody()->getOps<hlfir::DesignateOp>().begin();
builder.setInsertionPoint(elementalOp);
// Create a temp to transfer the rhs before applying the conversion.
hlfir::Entity entity{designateOp.getMemref()};
auto [temp, cleanup] = hlfir::createTempFromMold(loc, builder, entity);
auto transferKindAttr = cuf::DataTransferKindAttr::get(
builder.getContext(), cuf::DataTransferKind::DeviceHost);
cuf::DataTransferOp::create(builder, loc, designateOp.getMemref(), temp,
/*shape=*/mlir::Value{}, transferKindAttr);
designateOp.getMemrefMutable().assign(temp);
builder.setInsertionPointAfter(elementalOp);
hlfir::AssignOp::create(builder, loc, elementalOp, lhs,
isWholeAllocatableAssignment,
keepLhsLengthInAllocatableAssignment);
return;
}
auto transferKindAttr = cuf::DataTransferKindAttr::get(
builder.getContext(), cuf::DataTransferKind::DeviceHost);
cuf::DataTransferOp::create(builder, loc, rhsVal, lhsVal, shape,
Expand Down Expand Up @@ -5039,7 +5063,9 @@ class FirConverter : public Fortran::lower::AbstractConverter {
hlfir::Entity rhs = evaluateRhs(localStmtCtx);
hlfir::Entity lhs = evaluateLhs(localStmtCtx);
if (isCUDATransfer && !hasCUDAImplicitTransfer)
genCUDADataTransfer(builder, loc, assign, lhs, rhs);
genCUDADataTransfer(builder, loc, assign, lhs, rhs,
isWholeAllocatableAssignment,
keepLhsLengthInAllocatableAssignment);
else
hlfir::AssignOp::create(builder, loc, rhs, lhs,
isWholeAllocatableAssignment,
Expand Down
9 changes: 9 additions & 0 deletions flang/lib/Lower/CUDA.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -155,3 +155,12 @@ cuf::DataAttributeAttr Fortran::lower::translateSymbolCUFDataAttribute(
Fortran::semantics::GetCUDADataAttr(&sym.GetUltimate());
return cuf::getDataAttribute(mlirContext, cudaAttr);
}

bool Fortran::lower::isTransferWithConversion(mlir::Value rhs) {
if (auto elOp = mlir::dyn_cast<hlfir::ElementalOp>(rhs.getDefiningOp()))
if (llvm::hasSingleElement(elOp.getBody()->getOps<hlfir::DesignateOp>()) &&
llvm::hasSingleElement(elOp.getBody()->getOps<fir::LoadOp>()) == 1 &&
llvm::hasSingleElement(elOp.getBody()->getOps<fir::ConvertOp>()) == 1)
return true;
return false;
}
48 changes: 48 additions & 0 deletions flang/test/Lower/CUDA/cuda-data-transfer.cuf
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -474,3 +474,51 @@ end
! CHECK: cuf.data_transfer %{{.*}} to %{{.*}} {transfer_kind = #cuf.cuda_transfer<device_host>} : !fir.ref<!fir.box<!fir.heap<!fir.array<?xf64>>>>, !fir.box<!fir.array<?xf64>>
! CHECK: hlfir.assign %{{.*}} to %{{.*}} : f64, !fir.ref<f64>
! CHECK: fir.freemem %{{.*}} : !fir.heap<!fir.array<?xf64>>

subroutine sub26(i, j, k)
integer :: i, j, k
real(2), dimension(i,j,k), device :: d
real(4), dimension(i,j,k) :: hd

hd = d
end subroutine

! CHECK-LABEL: func.func @_QPsub26
! CHECK: %[[ALLOC_D:.*]] = cuf.alloc !fir.array<?x?x?xf16>, %{{.*}}, %{{.*}}, %{{.*}} : index, index, index {bindc_name = "d", data_attr = #cuf.cuda<device>, uniq_name = "_QFsub26Ed"} -> !fir.ref<!fir.array<?x?x?xf16>>
! CHECK: %[[D:.*]]:2 = hlfir.declare %[[ALLOC_D]](%{{.*}}) {data_attr = #cuf.cuda<device>, uniq_name = "_QFsub26Ed"} : (!fir.ref<!fir.array<?x?x?xf16>>, !fir.shape<3>) -> (!fir.box<!fir.array<?x?x?xf16>>, !fir.ref<!fir.array<?x?x?xf16>>)
! CHECK: %[[HD:.*]]:2 = hlfir.declare %{{.*}}(%{{.*}}) {uniq_name = "_QFsub26Ehd"} : (!fir.ref<!fir.array<?x?x?xf32>>, !fir.shape<3>) -> (!fir.box<!fir.array<?x?x?xf32>>, !fir.ref<!fir.array<?x?x?xf32>>)
! CHECK: %[[ALLOC:.*]] = fir.allocmem !fir.array<?x?x?xf16>, %8, %13, %18 {bindc_name = ".tmp", uniq_name = ""}
! CHECK: %[[TEMP:.*]]:2 = hlfir.declare %[[ALLOC]](%{{.*}}) {uniq_name = ".tmp"} : (!fir.heap<!fir.array<?x?x?xf16>>, !fir.shape<3>) -> (!fir.box<!fir.array<?x?x?xf16>>, !fir.heap<!fir.array<?x?x?xf16>>)
! CHECK: cuf.data_transfer %[[D]]#0 to %[[TEMP]]#0 {transfer_kind = #cuf.cuda_transfer<device_host>} : !fir.box<!fir.array<?x?x?xf16>>, !fir.box<!fir.array<?x?x?xf16>>
! CHECK: %[[ELE:.*]] = hlfir.elemental %{{.*}} unordered : (!fir.shape<3>) -> !hlfir.expr<?x?x?xf32> {
! CHECK: ^bb0(%{{.*}}: index, %{{.*}}: index, %{{.*}}: index):
! CHECK: %[[DESIGNATE:.*]] = hlfir.designate %[[TEMP]]#0 (%{{.*}}, %{{.*}}, %{{.*}}) : (!fir.box<!fir.array<?x?x?xf16>>, index, index, index) -> !fir.ref<f16>
! CHECK: %[[LOAD:.*]] = fir.load %[[DESIGNATE]] : !fir.ref<f16>
! CHECK: %[[CONV:.*]] = fir.convert %[[LOAD]] : (f16) -> f32
! CHECK: hlfir.yield_element %[[CONV]] : f32
! CHECK: }
! CHECK: hlfir.assign %[[ELE]] to %[[HD]]#0 : !hlfir.expr<?x?x?xf32>, !fir.box<!fir.array<?x?x?xf32>>

subroutine sub27()
real(2), dimension(10, 20, 30), device :: d
real(4), dimension(10, 20, 30) :: hd

hd = d
end subroutine

! CHECK-LABEL: func.func @_QPsub27()
! CHECK: %[[ALLOC_D:.*]] = cuf.alloc !fir.array<10x20x30xf16> {bindc_name = "d", data_attr = #cuf.cuda<device>, uniq_name = "_QFsub27Ed"} -> !fir.ref<!fir.array<10x20x30xf16>>
! CHECK: %[[D:.*]]:2 = hlfir.declare %[[ALLOC_D]](%{{.*}}) {data_attr = #cuf.cuda<device>, uniq_name = "_QFsub27Ed"} : (!fir.ref<!fir.array<10x20x30xf16>>, !fir.shape<3>) -> (!fir.ref<!fir.array<10x20x30xf16>>, !fir.ref<!fir.array<10x20x30xf16>>)
! CHECK: %[[ALLOC_HD:.*]] = fir.alloca !fir.array<10x20x30xf32> {bindc_name = "hd", uniq_name = "_QFsub27Ehd"}
! CHECK: %[[HD:.*]]:2 = hlfir.declare %[[ALLOC_HD]](%{{.*}}) {uniq_name = "_QFsub27Ehd"} : (!fir.ref<!fir.array<10x20x30xf32>>, !fir.shape<3>) -> (!fir.ref<!fir.array<10x20x30xf32>>, !fir.ref<!fir.array<10x20x30xf32>>)
! CHECK: %[[ALLOC_TEMP:.*]] = fir.allocmem !fir.array<10x20x30xf16> {bindc_name = ".tmp", uniq_name = ""}
! CHECK: %[[TEMP:.*]]:2 = hlfir.declare %[[ALLOC_TEMP]](%{{.*}}) {uniq_name = ".tmp"} : (!fir.heap<!fir.array<10x20x30xf16>>, !fir.shape<3>) -> (!fir.heap<!fir.array<10x20x30xf16>>, !fir.heap<!fir.array<10x20x30xf16>>)
! CHECK: cuf.data_transfer %[[D]]#0 to %[[TEMP]]#0 {transfer_kind = #cuf.cuda_transfer<device_host>} : !fir.ref<!fir.array<10x20x30xf16>>, !fir.heap<!fir.array<10x20x30xf16>>
! CHECK: %[[ELE:.*]] = hlfir.elemental %{{.*}} unordered : (!fir.shape<3>) -> !hlfir.expr<10x20x30xf32> {
! CHECK: ^bb0(%{{.*}}: index, %{{.*}}: index, %{{.*}}: index):
! CHECK: %[[DESIGNATE:.*]] = hlfir.designate %[[TEMP]]#0 (%{{.*}}, %{{.*}}, %{{.*}}) : (!fir.heap<!fir.array<10x20x30xf16>>, index, index, index) -> !fir.ref<f16>
! CHECK: %[[LOAD:.*]] = fir.load %[[DESIGNATE]] : !fir.ref<f16>
! CHECK: %[[CONV:.*]] = fir.convert %[[LOAD]] : (f16) -> f32
! CHECK: hlfir.yield_element %[[CONV]] : f32
! CHECK: }
! CHECKL: hlfir.assign %[[ELE]] to %[[HD]]#0 : !hlfir.expr<10x20x30xf32>, !fir.ref<!fir.array<10x20x30xf32>>