Skip to content

[flang][cuda] Do not consider PINNED as device attribute #95988

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
Jun 19, 2024
Merged

[flang][cuda] Do not consider PINNED as device attribute #95988

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
4e7a0aa
[flang][cuda] Do not consider PINNED as device attribute
clementval Jun 18, 2024
e23e007
Add missing pinned
clementval Jun 18, 2024
522fcc8
Avoid implicit data_transfer for PINNED variable
clementval Jun 18, 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
15 changes: 9 additions & 6 deletions flang/include/flang/Evaluate/tools.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1231,12 +1231,13 @@ bool CheckForCoindexedObject(parser::ContextualMessages &,
const std::string &argName);

// Get the number of distinct symbols with CUDA attribute in the expression.
template <typename A> inline int GetNbOfCUDASymbols(const A &expr) {
template <typename A> inline int GetNbOfCUDADeviceSymbols(const A &expr) {
semantics::UnorderedSymbolSet symbols;
for (const Symbol &sym : CollectSymbols(expr)) {
if (const auto *details =
sym.GetUltimate().detailsIf<semantics::ObjectEntityDetails>()) {
if (details->cudaDataAttr()) {
if (details->cudaDataAttr() &&
*details->cudaDataAttr() != common::CUDADataAttr::Pinned) {
symbols.insert(sym);
}
}
Expand All @@ -1246,8 +1247,8 @@ template <typename A> inline int GetNbOfCUDASymbols(const A &expr) {

// Check if any of the symbols part of the expression has a CUDA data
// attribute.
template <typename A> inline bool HasCUDAAttrs(const A &expr) {
return GetNbOfCUDASymbols(expr) > 0;
template <typename A> inline bool HasCUDADeviceAttrs(const A &expr) {
return GetNbOfCUDADeviceSymbols(expr) > 0;
}

/// Check if the expression is a mix of host and device variables that require
Expand All @@ -1258,7 +1259,8 @@ inline bool HasCUDAImplicitTransfer(const Expr<SomeType> &expr) {
for (const Symbol &sym : CollectSymbols(expr)) {
if (const auto *details =
sym.GetUltimate().detailsIf<semantics::ObjectEntityDetails>()) {
if (details->cudaDataAttr()) {
if (details->cudaDataAttr() &&
*details->cudaDataAttr() != common::CUDADataAttr::Pinned) {
++deviceSymbols;
} else {
if (sym.owner().IsDerivedType()) {
Expand All @@ -1267,7 +1269,8 @@ inline bool HasCUDAImplicitTransfer(const Expr<SomeType> &expr) {
.GetSymbol()
->GetUltimate()
.detailsIf<semantics::ObjectEntityDetails>()) {
if (details->cudaDataAttr()) {
if (details->cudaDataAttr() &&
*details->cudaDataAttr() != common::CUDADataAttr::Pinned) {
++deviceSymbols;
}
}
Expand Down
11 changes: 6 additions & 5 deletions flang/lib/Lower/Bridge.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4107,8 +4107,8 @@ 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) {
bool lhsIsDevice = Fortran::evaluate::HasCUDAAttrs(assign.lhs);
bool rhsIsDevice = Fortran::evaluate::HasCUDAAttrs(assign.rhs);
bool lhsIsDevice = Fortran::evaluate::HasCUDADeviceAttrs(assign.lhs);
bool rhsIsDevice = Fortran::evaluate::HasCUDADeviceAttrs(assign.rhs);

auto getRefIfLoaded = [](mlir::Value val) -> mlir::Value {
if (auto loadOp =
Expand Down Expand Up @@ -4177,7 +4177,8 @@ class FirConverter : public Fortran::lower::AbstractConverter {
if (const auto *details =
sym.GetUltimate()
.detailsIf<Fortran::semantics::ObjectEntityDetails>()) {
if (details->cudaDataAttr()) {
if (details->cudaDataAttr() &&
*details->cudaDataAttr() != Fortran::common::CUDADataAttr::Pinned) {
if (sym.owner().IsDerivedType() && IsAllocatable(sym.GetUltimate()))
TODO(loc, "Device resident allocatable derived-type component");
// TODO: This should probably being checked in semantic and give a
Expand Down Expand Up @@ -4229,8 +4230,8 @@ class FirConverter : public Fortran::lower::AbstractConverter {
fir::FirOpBuilder &builder = getFirOpBuilder();

bool isInDeviceContext = isDeviceContext(builder);
bool isCUDATransfer = (Fortran::evaluate::HasCUDAAttrs(assign.lhs) ||
Fortran::evaluate::HasCUDAAttrs(assign.rhs)) &&
bool isCUDATransfer = (Fortran::evaluate::HasCUDADeviceAttrs(assign.lhs) ||
Fortran::evaluate::HasCUDADeviceAttrs(assign.rhs)) &&
!isInDeviceContext;
bool hasCUDAImplicitTransfer =
Fortran::evaluate::HasCUDAImplicitTransfer(assign.rhs);
Expand Down
4 changes: 2 additions & 2 deletions flang/lib/Semantics/check-cuda.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -548,8 +548,8 @@ void CUDAChecker::Enter(const parser::AssignmentStmt &x) {
return;
}

int nbLhs{evaluate::GetNbOfCUDASymbols(assign->lhs)};
int nbRhs{evaluate::GetNbOfCUDASymbols(assign->rhs)};
int nbLhs{evaluate::GetNbOfCUDADeviceSymbols(assign->lhs)};
int nbRhs{evaluate::GetNbOfCUDADeviceSymbols(assign->rhs)};

// device to host transfer with more than one device object on the rhs is not
// legal.
Expand Down
29 changes: 28 additions & 1 deletion 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 @@ -180,7 +180,6 @@ end subroutine
! 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>>>>


subroutine sub8(a, b, n)
integer :: n
integer, device :: a(n)
Expand All @@ -195,3 +194,31 @@ end subroutine
! CHECK: %[[A:.*]]:2 = hlfir.declare %[[ARG0]](%{{.*}}) dummy_scope %{{.*}} {data_attr = #cuf.cuda<device>, uniq_name = "_QFsub8Ea"} : (!fir.ref<!fir.array<?xi32>>, !fir.shape<1>, !fir.dscope) -> (!fir.box<!fir.array<?xi32>>, !fir.ref<!fir.array<?xi32>>)
! CHECK: cuf.data_transfer %[[A]]#0 to %[[B]]#0 {transfer_kind = #cuf.cuda_transfer<device_host>} : !fir.box<!fir.array<?xi32>>, !fir.ref<!fir.array<10xi32>>
! CHECK: cuf.data_transfer %[[B]]#0 to %[[A]]#0 {transfer_kind = #cuf.cuda_transfer<host_device>} : !fir.ref<!fir.array<10xi32>>, !fir.box<!fir.array<?xi32>>

subroutine sub9(a)
integer, pinned, allocatable :: a(:)
do concurrent (i = 1 : 10)
a(i) = 10 + a(i)
end do
end subroutine

! CHECK-LABEL: func.func @_QPsub9
! CHECK-NOT: cuf.data_transfer

subroutine sub10(a, b)
integer, device :: a
integer, allocatable, pinned :: b
integer :: res

res = a + b
end subroutine



! CHECK-LABEL: func.func @_QPsub10(
! CHECK-SAME: %[[ARG0:.*]]: !fir.ref<i32> {cuf.data_attr = #cuf.cuda<device>, fir.bindc_name = "a"}

! CHECK: %[[A:.*]]:2 = hlfir.declare %[[ARG0]] dummy_scope %1 {data_attr = #cuf.cuda<device>, uniq_name = "_QFsub10Ea"} : (!fir.ref<i32>, !fir.dscope) -> (!fir.ref<i32>, !fir.ref<i32>)
! CHECK: cuf.data_transfer %[[A]]#1 to %{{.*}}#0 {transfer_kind = #cuf.cuda_transfer<device_host>} : !fir.ref<i32>, !fir.ref<i32>
! CHECK-NOT: cuf.data_transfer

Loading