Skip to content

[flang][OpenMP] Lower target .. private(..) to omp.private ops #94195

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
ergawy merged 10 commits into from
Jun 7, 2024
Merged

[flang][OpenMP] Lower target .. private(..) to omp.private ops #94195

Show file tree
Hide file tree
Changes from 8 commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
75291f1
[flang][OpenMP] Lower `target .. private(..)` to `omp.private` ops
ergawy Jun 3, 2024
9e73ff4
add test descriptions
ergawy Jun 3, 2024
02c4c39
Merge remote-tracking branch 'upstream/main' into target_private_delayed
ergawy Jun 5, 2024
8ce67bb
abstract and comment shared code
ergawy Jun 5, 2024
7d326f5
Merge remote-tracking branch 'upstream/main' into target_private_delayed
ergawy Jun 5, 2024
52ebe3e
add more extensive test
ergawy Jun 5, 2024
591a341
add more check lines
ergawy Jun 7, 2024
e8a754b
Merge remote-tracking branch 'upstream/main' into target_private_delayed
ergawy Jun 7, 2024
6ab5849
Merge remote-tracking branch 'upstream/main' into target_private_delayed
ergawy Jun 7, 2024
0d2a3df
fix typo
ergawy Jun 7, 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
21 changes: 7 additions & 14 deletions flang/lib/Lower/OpenMP/DataSharingProcessor.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -48,14 +48,13 @@ DataSharingProcessor::DataSharingProcessor(
}

void DataSharingProcessor::processStep1(
mlir::omp::PrivateClauseOps *clauseOps,
llvm::SmallVectorImpl<const semantics::Symbol *> *privateSyms) {
mlir::omp::PrivateClauseOps *clauseOps) {
collectSymbolsForPrivatization();
collectDefaultSymbols();
collectImplicitSymbols();
collectPreDeterminedSymbols();

privatize(clauseOps, privateSyms);
privatize(clauseOps);

insertBarrier();
}
Expand Down Expand Up @@ -415,16 +414,14 @@ void DataSharingProcessor::collectPreDeterminedSymbols() {
preDeterminedSymbols);
}

void DataSharingProcessor::privatize(
mlir::omp::PrivateClauseOps *clauseOps,
llvm::SmallVectorImpl<const semantics::Symbol *> *privateSyms) {
void DataSharingProcessor::privatize(mlir::omp::PrivateClauseOps *clauseOps) {
for (const semantics::Symbol *sym : allPrivatizedSymbols) {
if (const auto *commonDet =
sym->detailsIf<semantics::CommonBlockDetails>()) {
for (const auto &mem : commonDet->objects())
doPrivatize(&*mem, clauseOps, privateSyms);
doPrivatize(&*mem, clauseOps);
} else
doPrivatize(sym, clauseOps, privateSyms);
doPrivatize(sym, clauseOps);
}
}

Expand All @@ -441,9 +438,8 @@ void DataSharingProcessor::copyLastPrivatize(mlir::Operation *op) {
}
}

void DataSharingProcessor::doPrivatize(
const semantics::Symbol *sym, mlir::omp::PrivateClauseOps *clauseOps,
llvm::SmallVectorImpl<const semantics::Symbol *> *privateSyms) {
void DataSharingProcessor::doPrivatize(const semantics::Symbol *sym,
mlir::omp::PrivateClauseOps *clauseOps) {
if (!useDelayedPrivatization) {
cloneSymbol(sym);
copyFirstPrivateSymbol(sym);
Expand Down Expand Up @@ -548,9 +544,6 @@ void DataSharingProcessor::doPrivatize(
clauseOps->privateVars.push_back(hsb.getAddr());
}

if (privateSyms)
privateSyms->push_back(sym);

symToPrivatizer[sym] = privatizerOp;
}

Expand Down
18 changes: 9 additions & 9 deletions flang/lib/Lower/OpenMP/DataSharingProcessor.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -105,18 +105,15 @@ class DataSharingProcessor {
void collectDefaultSymbols();
void collectImplicitSymbols();
void collectPreDeterminedSymbols();
void privatize(mlir::omp::PrivateClauseOps *clauseOps,
llvm::SmallVectorImpl<const semantics::Symbol *> *privateSyms);
void privatize(mlir::omp::PrivateClauseOps *clauseOps);
void defaultPrivatize(
mlir::omp::PrivateClauseOps *clauseOps,
llvm::SmallVectorImpl<const semantics::Symbol *> *privateSyms);
void implicitPrivatize(
mlir::omp::PrivateClauseOps *clauseOps,
llvm::SmallVectorImpl<const semantics::Symbol *> *privateSyms);
void
doPrivatize(const semantics::Symbol *sym,
mlir::omp::PrivateClauseOps *clauseOps,
llvm::SmallVectorImpl<const semantics::Symbol *> *privateSyms);
void doPrivatize(const semantics::Symbol *sym,
mlir::omp::PrivateClauseOps *clauseOps);
void copyLastPrivatize(mlir::Operation *op);
void insertLastPrivateCompare(mlir::Operation *op);
void cloneSymbol(const semantics::Symbol *sym);
Expand Down Expand Up @@ -147,15 +144,18 @@ class DataSharingProcessor {
// Step2 performs the copying for lastprivates and requires knowledge of the
// MLIR operation to insert the last private update. Step2 adds
// dealocation code as well.
void processStep1(
mlir::omp::PrivateClauseOps *clauseOps = nullptr,
llvm::SmallVectorImpl<const semantics::Symbol *> *privateSyms = nullptr);
void processStep1(mlir::omp::PrivateClauseOps *clauseOps = nullptr);
void processStep2(mlir::Operation *op, bool isLoop);

void setLoopIV(mlir::Value iv) {
assert(!loopIV && "Loop iteration variable already set");
loopIV = iv;
}

const llvm::SetVector<const semantics::Symbol *> &
getAllSymbolsToPrivatize() const {
return allPrivatizedSymbols;
}
};

} // namespace omp
Expand Down
118 changes: 94 additions & 24 deletions flang/lib/Lower/OpenMP/OpenMP.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -456,6 +456,33 @@ markDeclareTarget(mlir::Operation *op, lower::AbstractConverter &converter,
declareTargetOp.setDeclareTarget(deviceType, captureClause);
}

/// For an operation that takes `omp.private` values as region args, this util
/// merges the private vars info into the region arguments list.
///
/// \tparam OMPOP - the OpenMP op that takes `omp.private` inputs.
/// \tparam InfoTy - the type of private info we want to merge; e.g. mlir::Type
/// or mlir::Location fields of the private var list.
///
/// \param [in] op - the op accpeting `omp.private` inputs.
/// \param [in] currentList - the current list of region info that we
/// want to merge private info with. For example this could be the list of types
/// or locations of previous arguments to \op's region.
/// \param [in] infoAccessor - for a private variable, this returns the
/// data we want to merge: type or location.
/// \param [out] allRegionArgsInfo - the merged list of region info.
template <typename OMPOp, typename InfoTy>
static void
mergePrivateVarsInfo(OMPOp op, llvm::ArrayRef<InfoTy> currentList,
llvm::function_ref<InfoTy(mlir::Value)> infoAccessor,
llvm::SmallVectorImpl<InfoTy> &allRegionArgsInfo) {
mlir::OperandRange privateVars = op.getPrivateVars();

llvm::transform(currentList, std::back_inserter(allRegionArgsInfo),
[](InfoTy i) { return i; });
llvm::transform(privateVars, std::back_inserter(allRegionArgsInfo),
infoAccessor);
}

//===----------------------------------------------------------------------===//
// Op body generation helper structures and functions
//===----------------------------------------------------------------------===//
Expand Down Expand Up @@ -758,15 +785,28 @@ genBodyOfTargetOp(lower::AbstractConverter &converter, lower::SymMap &symTable,
llvm::ArrayRef<const semantics::Symbol *> mapSyms,
llvm::ArrayRef<mlir::Location> mapSymLocs,
llvm::ArrayRef<mlir::Type> mapSymTypes,
DataSharingProcessor &dsp,
const mlir::Location &currentLocation,
const ConstructQueue &queue, ConstructQueue::iterator item) {
assert(mapSymTypes.size() == mapSymLocs.size());

fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder();
mlir::Region &region = targetOp.getRegion();

auto *regionBlock =
firOpBuilder.createBlock(&region, {}, mapSymTypes, mapSymLocs);
llvm::SmallVector<mlir::Type> allRegionArgTypes;
mergePrivateVarsInfo(targetOp, mapSymTypes,
llvm::function_ref<mlir::Type(mlir::Value)>{
[](mlir::Value v) { return v.getType(); }},
allRegionArgTypes);

llvm::SmallVector<mlir::Location> allRegionArgLocs;
mergePrivateVarsInfo(targetOp, mapSymLocs,
llvm::function_ref<mlir::Location(mlir::Value)>{
[](mlir::Value v) { return v.getLoc(); }},
allRegionArgLocs);

auto *regionBlock = firOpBuilder.createBlock(&region, {}, allRegionArgTypes,
allRegionArgLocs);

// Clones the `bounds` placing them inside the target region and returns them.
auto cloneBound = [&](mlir::Value bound) {
Expand Down Expand Up @@ -830,6 +870,20 @@ genBodyOfTargetOp(lower::AbstractConverter &converter, lower::SymMap &symTable,
});
}

for (auto [argIndex, argSymbol] :
llvm::enumerate(dsp.getAllSymbolsToPrivatize())) {
argIndex = mapSyms.size() + argIndex;

const mlir::BlockArgument &arg = region.getArgument(argIndex);
converter.bindSymbol(*argSymbol,
hlfir::translateToExtendedValue(
currentLocation, firOpBuilder, hlfir::Entity{arg},
/*contiguousHint=*/
evaluate::IsSimplyContiguous(
*argSymbol, converter.getFoldingContext()))
.first);
}

// Check if cloning the bounds introduced any dependency on the outer region.
// If so, then either clone them as well if they are MemoryEffectFree, or else
// copy them to a new temporary and add them to the map and block_argument
Expand Down Expand Up @@ -907,6 +961,8 @@ genBodyOfTargetOp(lower::AbstractConverter &converter, lower::SymMap &symTable,
} else {
genNestedEvaluations(converter, eval);
}

dsp.processStep2(targetOp, /*isLoop=*/false);
}

template <typename OpTy, typename... Args>
Expand Down Expand Up @@ -1048,15 +1104,18 @@ static void genTargetClauses(
devicePtrSyms);
cp.processMap(loc, stmtCtx, clauseOps, &mapSyms, &mapLocs, &mapTypes);
cp.processThreadLimit(stmtCtx, clauseOps);
// TODO Support delayed privatization.

if (processHostOnlyClauses)
cp.processNowait(clauseOps);

cp.processTODO<clause::Allocate, clause::Defaultmap, clause::Firstprivate,
clause::InReduction, clause::Private, clause::Reduction,
clause::InReduction, clause::Reduction,
clause::UsesAllocators>(loc,
llvm::omp::Directive::OMPD_target);

// `target private(..)` is only supported in delayed privatization mode.
if (!enableDelayedPrivatization)
cp.processTODO<clause::Private>(loc, llvm::omp::Directive::OMPD_target);
}

static void genTargetDataClauses(
Expand Down Expand Up @@ -1289,7 +1348,6 @@ genParallelOp(lower::AbstractConverter &converter, lower::SymMap &symTable,
fir::FirOpBuilder &firOpBuilder = converter.getFirOpBuilder();
lower::StatementContext stmtCtx;
mlir::omp::ParallelClauseOps clauseOps;
llvm::SmallVector<const semantics::Symbol *> privateSyms;
llvm::SmallVector<mlir::Type> reductionTypes;
llvm::SmallVector<const semantics::Symbol *> reductionSyms;
genParallelClauses(converter, semaCtx, stmtCtx, item->clauses, loc,
Expand Down Expand Up @@ -1319,34 +1377,35 @@ genParallelOp(lower::AbstractConverter &converter, lower::SymMap &symTable,
/*useDelayedPrivatization=*/true, &symTable);

if (privatize)
dsp.processStep1(&clauseOps, &privateSyms);
dsp.processStep1(&clauseOps);

auto genRegionEntryCB = [&](mlir::Operation *op) {
auto parallelOp = llvm::cast<mlir::omp::ParallelOp>(op);

llvm::SmallVector<mlir::Location> reductionLocs(
clauseOps.reductionVars.size(), loc);

mlir::OperandRange privateVars = parallelOp.getPrivateVars();
mlir::Region &region = parallelOp.getRegion();
llvm::SmallVector<mlir::Type> allRegionArgTypes;
mergePrivateVarsInfo(parallelOp, llvm::ArrayRef(reductionTypes),
llvm::function_ref<mlir::Type(mlir::Value)>{
[](mlir::Value v) { return v.getType(); }},
allRegionArgTypes);

llvm::SmallVector<mlir::Type> privateVarTypes = reductionTypes;
privateVarTypes.reserve(privateVarTypes.size() + privateVars.size());
llvm::transform(privateVars, std::back_inserter(privateVarTypes),
[](mlir::Value v) { return v.getType(); });
llvm::SmallVector<mlir::Location> allRegionArgLocs;
mergePrivateVarsInfo(parallelOp, llvm::ArrayRef(reductionLocs),
llvm::function_ref<mlir::Location(mlir::Value)>{
[](mlir::Value v) { return v.getLoc(); }},
allRegionArgLocs);

llvm::SmallVector<mlir::Location> privateVarLocs = reductionLocs;
privateVarLocs.reserve(privateVarLocs.size() + privateVars.size());
llvm::transform(privateVars, std::back_inserter(privateVarLocs),
[](mlir::Value v) { return v.getLoc(); });

firOpBuilder.createBlock(&region, /*insertPt=*/{}, privateVarTypes,
privateVarLocs);
mlir::Region &region = parallelOp.getRegion();
firOpBuilder.createBlock(&region, /*insertPt=*/{}, allRegionArgTypes,
allRegionArgLocs);

llvm::SmallVector<const semantics::Symbol *> allSymbols = reductionSyms;
allSymbols.append(privateSyms);
allSymbols.append(dsp.getAllSymbolsToPrivatize().begin(),
dsp.getAllSymbolsToPrivatize().end());

for (auto [arg, prv] : llvm::zip_equal(allSymbols, region.getArguments())) {
fir::ExtendedValue hostExV = converter.getSymbolExtendedValue(*arg);
converter.bindSymbol(*arg, hlfir::translateToExtendedValue(
loc, firOpBuilder, hlfir::Entity{prv},
/*contiguousHint=*/
Expand Down Expand Up @@ -1541,11 +1600,22 @@ genTargetOp(lower::AbstractConverter &converter, lower::SymMap &symTable,
deviceAddrLocs, deviceAddrTypes, devicePtrSyms,
devicePtrLocs, devicePtrTypes);

llvm::SmallVector<const semantics::Symbol *> privateSyms;
DataSharingProcessor dsp(converter, semaCtx, item->clauses, eval,
/*shouldCollectPreDeterminedSymbols=*/
lower::omp::isLastItemInQueue(item, queue),
/*useDelayedPrivatization=*/true, &symTable);
dsp.processStep1(&clauseOps);

// 5.8.1 Implicit Data-Mapping Attribute Rules
// The following code follows the implicit data-mapping rules to map all the
// symbols used inside the region that have not been explicitly mapped using
// the map clause.
// symbols used inside the region that do not have explicit data-environment
// attribute clauses (neither data-sharing; e.g. `private`, nor `map`
// clauses).
auto captureImplicitMap = [&](const semantics::Symbol &sym) {
if (dsp.getAllSymbolsToPrivatize().contains(&sym))
return;

if (llvm::find(mapSyms, &sym) == mapSyms.end()) {
mlir::Value baseOp = converter.getSymbolAddress(sym);
if (!baseOp)
Expand Down Expand Up @@ -1632,7 +1702,7 @@ genTargetOp(lower::AbstractConverter &converter, lower::SymMap &symTable,

auto targetOp = firOpBuilder.create<mlir::omp::TargetOp>(loc, clauseOps);
genBodyOfTargetOp(converter, symTable, semaCtx, eval, targetOp, mapSyms,
mapLocs, mapTypes, loc, queue, item);
mapLocs, mapTypes, dsp, loc, queue, item);
return targetOp;
}

Expand Down
71 changes: 71 additions & 0 deletions flang/test/Lower/OpenMP/DelayedPrivatization/target-private-allocatable.f90
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,71 @@
! Tests delayed privatization for `targets ... private(..)` for allocatables.

! RUN: %flang_fc1 -emit-hlfir -fopenmp -mmlir --openmp-enable-delayed-privatization \
! RUN: -o - %s 2>&1 | FileCheck %s
! RUN: bbc -emit-hlfir -fopenmp --openmp-enable-delayed-privatization -o - %s 2>&1 \
! RUN: | FileCheck %s

subroutine target_allocatable
implicit none
integer, allocatable :: alloc_var

!$omp target private(alloc_var)
alloc_var = 10
!$omp end target
end subroutine target_allocatable

! CHECK-LABEL: omp.private {type = private}
! CHECK-SAME: @[[VAR_PRIVATIZER_SYM:.*]] :
! CHECK-SAME: [[TYPE:!fir.ref<!fir.box<!fir.heap<i32>>>]] alloc {
! CHECK: ^bb0(%[[PRIV_ARG:.*]]: [[TYPE]]):
! CHECK: %[[PRIV_ALLOC:.*]] = fir.alloca !fir.box<!fir.heap<i32>> {bindc_name = "alloc_var", {{.*}}}

! CHECK-NEXT: %[[PRIV_ARG_VAL:.*]] = fir.load %[[PRIV_ARG]] : !fir.ref<!fir.box<!fir.heap<i32>>>
! CHECK-NEXT: %[[PRIV_ARG_BOX:.*]] = fir.box_addr %[[PRIV_ARG_VAL]] : (!fir.box<!fir.heap<i32>>) -> !fir.heap<i32>
! CHECK-NEXT: %[[PRIV_ARG_ADDR:.*]] = fir.convert %[[PRIV_ARG_BOX]] : (!fir.heap<i32>) -> i64
! CHECK-NEXT: %[[C0:.*]] = arith.constant 0 : i64
! CHECK-NEXT: %[[ALLOC_COND:.*]] = arith.cmpi ne, %[[PRIV_ARG_ADDR]], %[[C0]] : i64

! CHECK-NEXT: fir.if %[[ALLOC_COND]] {
! CHECK: %[[PRIV_ALLOCMEM:.*]] = fir.allocmem i32 {fir.must_be_heap = true, {{.*}}}
! CHECK-NEXT: %[[PRIV_ALLOCMEM_BOX:.*]] = fir.embox %[[PRIV_ALLOCMEM]] : (!fir.heap<i32>) -> !fir.box<!fir.heap<i32>>
! CHECK-NEXT: fir.store %[[PRIV_ALLOCMEM_BOX]] to %[[PRIV_ALLOC]] : !fir.ref<!fir.box<!fir.heap<i32>>>
! CHECK-NEXT: } else {
! CHECK-NEXT: %[[ZERO_BITS:.*]] = fir.zero_bits !fir.heap<i32>
! CHECK-NEXT: %[[ZERO_BOX:.*]] = fir.embox %[[ZERO_BITS]] : (!fir.heap<i32>) -> !fir.box<!fir.heap<i32>>
! CHECK-NEXT: fir.store %[[ZERO_BOX]] to %[[PRIV_ALLOC]] : !fir.ref<!fir.box<!fir.heap<i32>>>
! CHECK-NEXT: }

! CHECK-NEXT: %[[PRIV_DECL:.*]]:2 = hlfir.declare %[[PRIV_ALLOC]]
! CHECK-NEXT: omp.yield(%[[PRIV_DECL]]#0 : [[TYPE]])

! CHECK-NEXT: } dealloc {
! CHECK-NEXT: ^bb0(%[[PRIV_ARG:.*]]: [[TYPE]]):

! CHECK-NEXT: %[[PRIV_VAL:.*]] = fir.load %[[PRIV_ARG]]
! CHECK-NEXT: %[[PRIV_ADDR:.*]] = fir.box_addr %[[PRIV_VAL]]
! CHECK-NEXT: %[[PRIV_ADDR_I64:.*]] = fir.convert %[[PRIV_ADDR]]
! CHECK-NEXT: %[[C0:.*]] = arith.constant 0 : i64
! CHECK-NEXT: %[[PRIV_NULL_COND:.*]] = arith.cmpi ne, %[[PRIV_ADDR_I64]], %[[C0]] : i64

! CHECK-NEXT: fir.if %[[PRIV_NULL_COND]] {
! CHECK: %[[PRIV_VAL_2:.*]] = fir.load %[[PRIV_ARG]]
! CHECK-NEXT: %[[PRIV_ADDR_2:.*]] = fir.box_addr %[[PRIV_VAL_2]]
! CHECK-NEXT: fir.freemem %[[PRIV_ADDR_2]]
! CHECK-NEXT: %[[ZEROS:.*]] = fir.zero_bits
! CHECK-NEXT: %[[ZEROS_BOX:.*]] = fir.embox %[[ZEROS]]
! CHECK-NEXT: fir.store %[[ZEROS_BOX]] to %[[PRIV_ARG]]
! CHECK-NEXT: }

! CHECK-NEXT: omp.yield
! CHECK-NEXT: }


! CHECK-LABEL: func.func @_QPtarget_allocatable() {

! CHECK: %[[VAR_ALLOC:.*]] = fir.alloca !fir.box<!fir.heap<i32>>
! CHECK-SAME: {bindc_name = "alloc_var", {{.*}}}
! CHECK: %[[VAR_DECL:.*]]:2 = hlfir.declare %[[VAR_ALLOC]]

! CHECK: omp.target private(
! CHECK-SAME: @[[VAR_PRIVATIZER_SYM]] %[[VAR_DECL]]#0 -> %{{.*}} : [[TYPE]]) {
Loading