Skip to content

[flang][cuda] Enable data transfer for descriptors #92804

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 4 commits into from
May 21, 2024
Merged

[flang][cuda] Enable data transfer for descriptors #92804

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
0b06f72
[flang][cuda] Enable data transfer for descriptor
clementval May 20, 2024
1205ace
Use dyn_cast_or_null
clementval May 20, 2024
8edd8f1
Add comment about descriptor updates
clementval May 20, 2024
d52edab
Merge branch 'main' into cuda_data_transfer_desc
clementval May 21, 2024
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: 8 additions & 2 deletions flang/include/flang/Optimizer/Dialect/CUF/CUFOps.td
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -152,15 +152,21 @@ def cuf_DataTransferOp : cuf_Op<"data_transfer", []> {
a = adev ! transfer device to host
bdev = adev ! transfer device to device
```

When the data transfer is done on data hold by descriptors, the LHS data
hold by the descriptor are updated. When required, the LHS decriptor is also
updated.
}];

let arguments = (ins Arg<AnyReferenceLike, "", [MemRead]>:$src,
Arg<AnyReferenceLike, "", [MemWrite]>:$dst,
let arguments = (ins Arg<AnyRefOrBoxType, "", [MemRead]>:$src,
Arg<AnyRefOrBoxType, "", [MemWrite]>:$dst,
cuf_DataTransferKindAttr:$transfer_kind);

let assemblyFormat = [{
$src `to` $dst attr-dict `:` type(operands)
}];

let hasVerifier = 1;
}

def cuf_KernelLaunchOp : cuf_Op<"kernel_launch", [CallOpInterface,
Expand Down
35 changes: 18 additions & 17 deletions flang/lib/Lower/Bridge.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3782,8 +3782,16 @@ class FirConverter : public Fortran::lower::AbstractConverter {
hlfir::Entity &lhs, hlfir::Entity &rhs) {
bool lhsIsDevice = Fortran::evaluate::HasCUDAAttrs(assign.lhs);
bool rhsIsDevice = Fortran::evaluate::HasCUDAAttrs(assign.rhs);
if (rhs.isBoxAddressOrValue() || lhs.isBoxAddressOrValue())
TODO(loc, "CUDA data transfler with descriptors");

auto getRefIfLoaded = [](mlir::Value val) -> mlir::Value {
if (auto loadOp =
mlir::dyn_cast_or_null<fir::LoadOp>(val.getDefiningOp()))
return loadOp.getMemref();
return val;
};

mlir::Value rhsVal = getRefIfLoaded(rhs.getBase());
mlir::Value lhsVal = getRefIfLoaded(lhs.getBase());

// device = host
if (lhsIsDevice && !rhsIsDevice) {
Expand All @@ -3792,11 +3800,12 @@ class FirConverter : public Fortran::lower::AbstractConverter {
if (!rhs.isVariable()) {
auto associate = hlfir::genAssociateExpr(
loc, builder, rhs, rhs.getType(), ".cuf_host_tmp");
builder.create<cuf::DataTransferOp>(loc, associate.getBase(), lhs,
builder.create<cuf::DataTransferOp>(loc, associate.getBase(), lhsVal,
transferKindAttr);
builder.create<hlfir::EndAssociateOp>(loc, associate);
} else {
builder.create<cuf::DataTransferOp>(loc, rhs, lhs, transferKindAttr);
builder.create<cuf::DataTransferOp>(loc, rhsVal, lhsVal,
transferKindAttr);
}
return;
}
Expand All @@ -3805,26 +3814,18 @@ class FirConverter : public Fortran::lower::AbstractConverter {
if (!lhsIsDevice && rhsIsDevice) {
auto transferKindAttr = cuf::DataTransferKindAttr::get(
builder.getContext(), cuf::DataTransferKind::DeviceHost);
if (!rhs.isVariable()) {
// evaluateRhs loads scalar. Look for the memory reference to be used in
// the transfer.
if (mlir::isa_and_nonnull<fir::LoadOp>(rhs.getDefiningOp())) {
auto loadOp = mlir::dyn_cast<fir::LoadOp>(rhs.getDefiningOp());
builder.create<cuf::DataTransferOp>(loc, loadOp.getMemref(), lhs,
transferKindAttr);
return;
}
} else {
builder.create<cuf::DataTransferOp>(loc, rhs, lhs, transferKindAttr);
}
builder.create<cuf::DataTransferOp>(loc, rhsVal, lhsVal,
transferKindAttr);
return;
}

// device = device
if (lhsIsDevice && rhsIsDevice) {
assert(rhs.isVariable() && "CUDA Fortran assignment rhs is not legal");
auto transferKindAttr = cuf::DataTransferKindAttr::get(
builder.getContext(), cuf::DataTransferKind::DeviceDevice);
builder.create<cuf::DataTransferOp>(loc, rhs, lhs, transferKindAttr);
builder.create<cuf::DataTransferOp>(loc, rhsVal, lhsVal,
transferKindAttr);
return;
}
llvm_unreachable("Unhandled CUDA data transfer");
Expand Down
13 changes: 13 additions & 0 deletions flang/lib/Optimizer/Dialect/CUF/CUFOps.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -89,6 +89,19 @@ mlir::LogicalResult cuf::AllocateOp::verify() {
return mlir::success();
}

//===----------------------------------------------------------------------===//
// DataTransferOp
//===----------------------------------------------------------------------===//

mlir::LogicalResult cuf::DataTransferOp::verify() {
mlir::Type srcTy = getSrc().getType();
mlir::Type dstTy = getDst().getType();
if (fir::isa_ref_type(srcTy) && fir::isa_ref_type(dstTy) ||
fir::isa_box_type(srcTy) && fir::isa_box_type(dstTy))
return mlir::success();
return emitOpError("expect src and dst to be both references or descriptors");
}

//===----------------------------------------------------------------------===//
// DeallocateOp
//===----------------------------------------------------------------------===//
Expand Down
19 changes: 19 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 @@ -159,3 +159,22 @@ end subroutine

! CHECK-LABEL: func.func @_QPsub6
! CHECK: cuf.data_transfer

subroutine sub7(a, b, c)
integer, device, allocatable :: a(:), c(:)
integer, allocatable :: b(:)
b = a

a = b

c = a
end subroutine

! CHECK-LABEL: func.func @_QPsub7(
! CHECK-SAME: %[[ARG0:.*]]: !fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>> {cuf.data_attr = #cuf.cuda<device>, fir.bindc_name = "a"}, %[[ARG1:.*]]: !fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>> {fir.bindc_name = "b"}, %[[ARG2:.*]]: !fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>> {cuf.data_attr = #cuf.cuda<device>, fir.bindc_name = "c"}) {
! CHECK: %[[A:.*]]:2 = hlfir.declare %[[ARG0]] dummy_scope %{{.*}} {data_attr = #cuf.cuda<device>, fortran_attrs = #fir.var_attrs<allocatable>, uniq_name = "_QFsub7Ea"} : (!fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>>, !fir.dscope) -> (!fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>>, !fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>>)
! CHECK: %[[B:.*]]:2 = hlfir.declare %[[ARG1]] dummy_scope %{{.*}} {fortran_attrs = #fir.var_attrs<allocatable>, uniq_name = "_QFsub7Eb"} : (!fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>>, !fir.dscope) -> (!fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>>, !fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>>)
! CHECK: %[[C:.*]]:2 = hlfir.declare %[[ARG2]] dummy_scope %0 {data_attr = #cuf.cuda<device>, fortran_attrs = #fir.var_attrs<allocatable>, uniq_name = "_QFsub7Ec"} : (!fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>>, !fir.dscope) -> (!fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>>, !fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>>)
! CHECK: cuf.data_transfer %[[A]]#0 to %[[B]]#0 {transfer_kind = #cuf.cuda_transfer<device_host>} : !fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>>, !fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>>
! CHECK: cuf.data_transfer %[[B]]#0 to %[[A]]#0 {transfer_kind = #cuf.cuda_transfer<host_device>} : !fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>>, !fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>>
! CHECK: cuf.data_transfer %[[A]]#0 to %[[C]]#0 {transfer_kind = #cuf.cuda_transfer<device_device>} : !fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>>, !fir.ref<!fir.box<!fir.heap<!fir.array<?xi32>>>>
Loading