Skip to content

[OpenACC][CIR] Lower 'num_workers' for parallel/kernels #136578

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
erichkeane merged 2 commits into from
Apr 21, 2025
Merged

[OpenACC][CIR] Lower 'num_workers' for parallel/kernels #136578

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
e12a501
[OpenACC][CIR] Lower 'num_workers' for parallel/kernels
erichkeane Apr 16, 2025
40766ab
Correct device-type for loops
erichkeane Apr 21, 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
63 changes: 63 additions & 0 deletions clang/lib/CIR/CodeGen/CIRGenStmtOpenACC.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -46,10 +46,27 @@ class OpenACCClauseCIREmitter final
// diagnostics are gone.
SourceLocation dirLoc;

const OpenACCDeviceTypeClause *lastDeviceTypeClause = nullptr;

void clauseNotImplemented(const OpenACCClause &c) {
cgf.cgm.errorNYI(c.getSourceRange(), "OpenACC Clause", c.getClauseKind());
}

mlir::Value createIntExpr(const Expr *intExpr) {
mlir::Value expr = cgf.emitScalarExpr(intExpr);
mlir::Location exprLoc = cgf.cgm.getLoc(intExpr->getBeginLoc());

mlir::IntegerType targetType = mlir::IntegerType::get(
&cgf.getMLIRContext(), cgf.getContext().getIntWidth(intExpr->getType()),
intExpr->getType()->isSignedIntegerOrEnumerationType()
? mlir::IntegerType::SignednessSemantics::Signed
: mlir::IntegerType::SignednessSemantics::Unsigned);

auto conversionOp = builder.create<mlir::UnrealizedConversionCastOp>(
exprLoc, targetType, expr);
return conversionOp.getResult(0);
}

// 'condition' as an OpenACC grammar production is used for 'if' and (some
// variants of) 'self'. It needs to be emitted as a signless-1-bit value, so
// this function emits the expression, then sets the unrealized conversion
Expand Down Expand Up @@ -109,6 +126,7 @@ class OpenACCClauseCIREmitter final
}

void VisitDeviceTypeClause(const OpenACCDeviceTypeClause &clause) {
lastDeviceTypeClause = &clause;
Copy link
Contributor

@andykaylor andykaylor Apr 21, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What's the lifetime of this object? Who owns it?

Copy link
Collaborator Author

@erichkeane erichkeane Apr 21, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Its owned by the AST, so its lifetime lasts until the AST is destroyed, which is after frontend codegen.

if constexpr (isOneOfTypes<OpTy, InitOp, ShutdownOp>) {
llvm::SmallVector<mlir::Attribute> deviceTypes;
std::optional<mlir::ArrayAttr> existingDeviceTypes =
Expand Down Expand Up @@ -136,6 +154,51 @@ class OpenACCClauseCIREmitter final
if (!clause.getArchitectures().empty())
operation.setDeviceType(
decodeDeviceType(clause.getArchitectures()[0].getIdentifierInfo()));
} else if constexpr (isOneOfTypes<OpTy, ParallelOp, SerialOp, KernelsOp>) {
// Nothing to do here, these constructs don't have any IR for these, as
// they just modify the other clauses IR. So setting of `lastDeviceType`
// (done above) is all we need.
} else {
return clauseNotImplemented(clause);
}
}

void VisitNumWorkersClause(const OpenACCNumWorkersClause &clause) {
if constexpr (isOneOfTypes<OpTy, ParallelOp, KernelsOp>) {
// Collect the 'existing' device-type attributes so we can re-create them
// and insert them.
llvm::SmallVector<mlir::Attribute> deviceTypes;
mlir::ArrayAttr existingDeviceTypes =
operation.getNumWorkersDeviceTypeAttr();

if (existingDeviceTypes) {
for (const mlir::Attribute &Attr : existingDeviceTypes)
Copy link
Member

@bcardosolopes bcardosolopes Apr 21, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Should this be by value because value semantics?

Copy link
Collaborator Author

@erichkeane erichkeane Apr 21, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yeah, you're probably right. I'll fix this here and 1 other place.

deviceTypes.push_back(mlir::acc::DeviceTypeAttr::get(
builder.getContext(),
cast<mlir::acc::DeviceTypeAttr>(Attr).getValue()));
}

// Insert 1 version of the 'int-expr' to the NumWorkers list per-current
// device type.
mlir::Value intExpr = createIntExpr(clause.getIntExpr());
if (lastDeviceTypeClause) {
for (const DeviceTypeArgument &arg :
lastDeviceTypeClause->getArchitectures()) {
deviceTypes.push_back(mlir::acc::DeviceTypeAttr::get(
builder.getContext(), decodeDeviceType(arg.getIdentifierInfo())));
operation.getNumWorkersMutable().append(intExpr);
}
} else {
// Else, we just add a single for 'none'.
deviceTypes.push_back(mlir::acc::DeviceTypeAttr::get(
builder.getContext(), mlir::acc::DeviceType::None));
operation.getNumWorkersMutable().append(intExpr);
}

operation.setNumWorkersDeviceTypeAttr(
mlir::ArrayAttr::get(builder.getContext(), deviceTypes));
} else if constexpr (isOneOfTypes<OpTy, SerialOp>) {
llvm_unreachable("num_workers not valid on serial");
} else {
return clauseNotImplemented(clause);
}
Expand Down
52 changes: 52 additions & 0 deletions clang/test/CIR/CodeGenOpenACC/kernels.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -106,5 +106,57 @@ void acc_kernels(int cond) {
// CHECK-NEXT: acc.terminator
// CHECK-NEXT: } loc

#pragma acc kernels num_workers(cond)
{}
// CHECK-NEXT: %[[COND_LOAD:.*]] = cir.load %[[COND]] : !cir.ptr<!s32i>, !s32i
// CHECK-NEXT: %[[CONV_CAST:.*]] = builtin.unrealized_conversion_cast %[[COND_LOAD]] : !s32i to si32
// CHECK-NEXT: acc.kernels num_workers(%[[CONV_CAST]] : si32) {
// CHECK-NEXT: acc.terminator
// CHECK-NEXT: } loc

#pragma acc kernels num_workers(cond) device_type(nvidia) num_workers(2u)
{}
// CHECK-NEXT: %[[COND_LOAD:.*]] = cir.load %[[COND]] : !cir.ptr<!s32i>, !s32i
// CHECK-NEXT: %[[CONV_CAST:.*]] = builtin.unrealized_conversion_cast %[[COND_LOAD]] : !s32i to si32
// CHECK-NEXT: %[[TWO_LITERAL:.*]] = cir.const #cir.int<2> : !u32i
// CHECK-NEXT: %[[TWO_CAST:.*]] = builtin.unrealized_conversion_cast %[[TWO_LITERAL]] : !u32i to ui32
// CHECK-NEXT: acc.kernels num_workers(%[[CONV_CAST]] : si32, %[[TWO_CAST]] : ui32 [#acc.device_type<nvidia>]) {
// CHECK-NEXT: acc.terminator
// CHECK-NEXT: } loc

#pragma acc kernels num_workers(cond) device_type(nvidia, host) num_workers(2) device_type(radeon) num_workers(3)
{}
// CHECK-NEXT: %[[COND_LOAD:.*]] = cir.load %[[COND]] : !cir.ptr<!s32i>, !s32i
// CHECK-NEXT: %[[CONV_CAST:.*]] = builtin.unrealized_conversion_cast %[[COND_LOAD]] : !s32i to si32
// CHECK-NEXT: %[[TWO_LITERAL:.*]] = cir.const #cir.int<2> : !s32i
// CHECK-NEXT: %[[TWO_CAST:.*]] = builtin.unrealized_conversion_cast %[[TWO_LITERAL]] : !s32i to si32
// CHECK-NEXT: %[[THREE_LITERAL:.*]] = cir.const #cir.int<3> : !s32i
// CHECK-NEXT: %[[THREE_CAST:.*]] = builtin.unrealized_conversion_cast %[[THREE_LITERAL]] : !s32i to si32
// CHECK-NEXT: acc.kernels num_workers(%[[CONV_CAST]] : si32, %[[TWO_CAST]] : si32 [#acc.device_type<nvidia>], %[[TWO_CAST]] : si32 [#acc.device_type<host>], %[[THREE_CAST]] : si32 [#acc.device_type<radeon>]) {
// CHECK-NEXT: acc.terminator
// CHECK-NEXT: } loc

#pragma acc kernels num_workers(cond) device_type(nvidia) num_workers(2) device_type(radeon, multicore) num_workers(3)
{}
// CHECK-NEXT: %[[COND_LOAD:.*]] = cir.load %[[COND]] : !cir.ptr<!s32i>, !s32i
// CHECK-NEXT: %[[CONV_CAST:.*]] = builtin.unrealized_conversion_cast %[[COND_LOAD]] : !s32i to si32
// CHECK-NEXT: %[[TWO_LITERAL:.*]] = cir.const #cir.int<2> : !s32i
// CHECK-NEXT: %[[TWO_CAST:.*]] = builtin.unrealized_conversion_cast %[[TWO_LITERAL]] : !s32i to si32
// CHECK-NEXT: %[[THREE_LITERAL:.*]] = cir.const #cir.int<3> : !s32i
// CHECK-NEXT: %[[THREE_CAST:.*]] = builtin.unrealized_conversion_cast %[[THREE_LITERAL]] : !s32i to si32
// CHECK-NEXT: acc.kernels num_workers(%[[CONV_CAST]] : si32, %[[TWO_CAST]] : si32 [#acc.device_type<nvidia>], %[[THREE_CAST]] : si32 [#acc.device_type<radeon>], %[[THREE_CAST]] : si32 [#acc.device_type<multicore>]) {
// CHECK-NEXT: acc.terminator
// CHECK-NEXT: } loc

#pragma acc kernels device_type(nvidia) num_workers(2) device_type(radeon) num_workers(3)
{}
// CHECK-NEXT: %[[TWO_LITERAL:.*]] = cir.const #cir.int<2> : !s32i
// CHECK-NEXT: %[[TWO_CAST:.*]] = builtin.unrealized_conversion_cast %[[TWO_LITERAL]] : !s32i to si32
// CHECK-NEXT: %[[THREE_LITERAL:.*]] = cir.const #cir.int<3> : !s32i
// CHECK-NEXT: %[[THREE_CAST:.*]] = builtin.unrealized_conversion_cast %[[THREE_LITERAL]] : !s32i to si32
// CHECK-NEXT: acc.kernels num_workers(%[[TWO_CAST]] : si32 [#acc.device_type<nvidia>], %[[THREE_CAST]] : si32 [#acc.device_type<radeon>]) {
// CHECK-NEXT: acc.terminator
// CHECK-NEXT: } loc

// CHECK-NEXT: cir.return
}
52 changes: 52 additions & 0 deletions clang/test/CIR/CodeGenOpenACC/parallel.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -105,5 +105,57 @@ void acc_parallel(int cond) {
// CHECK-NEXT: acc.yield
// CHECK-NEXT: } loc

#pragma acc parallel num_workers(cond)
{}
// CHECK-NEXT: %[[COND_LOAD:.*]] = cir.load %[[COND]] : !cir.ptr<!s32i>, !s32i
// CHECK-NEXT: %[[CONV_CAST:.*]] = builtin.unrealized_conversion_cast %[[COND_LOAD]] : !s32i to si32
// CHECK-NEXT: acc.parallel num_workers(%[[CONV_CAST]] : si32) {
// CHECK-NEXT: acc.yield
// CHECK-NEXT: } loc

#pragma acc parallel num_workers(cond) device_type(nvidia) num_workers(2u)
{}
// CHECK-NEXT: %[[COND_LOAD:.*]] = cir.load %[[COND]] : !cir.ptr<!s32i>, !s32i
// CHECK-NEXT: %[[CONV_CAST:.*]] = builtin.unrealized_conversion_cast %[[COND_LOAD]] : !s32i to si32
// CHECK-NEXT: %[[TWO_LITERAL:.*]] = cir.const #cir.int<2> : !u32i
// CHECK-NEXT: %[[TWO_CAST:.*]] = builtin.unrealized_conversion_cast %[[TWO_LITERAL]] : !u32i to ui32
// CHECK-NEXT: acc.parallel num_workers(%[[CONV_CAST]] : si32, %[[TWO_CAST]] : ui32 [#acc.device_type<nvidia>]) {
// CHECK-NEXT: acc.yield
// CHECK-NEXT: } loc

#pragma acc parallel num_workers(cond) device_type(nvidia, host) num_workers(2) device_type(radeon) num_workers(3)
{}
// CHECK-NEXT: %[[COND_LOAD:.*]] = cir.load %[[COND]] : !cir.ptr<!s32i>, !s32i
// CHECK-NEXT: %[[CONV_CAST:.*]] = builtin.unrealized_conversion_cast %[[COND_LOAD]] : !s32i to si32
// CHECK-NEXT: %[[TWO_LITERAL:.*]] = cir.const #cir.int<2> : !s32i
// CHECK-NEXT: %[[TWO_CAST:.*]] = builtin.unrealized_conversion_cast %[[TWO_LITERAL]] : !s32i to si32
// CHECK-NEXT: %[[THREE_LITERAL:.*]] = cir.const #cir.int<3> : !s32i
// CHECK-NEXT: %[[THREE_CAST:.*]] = builtin.unrealized_conversion_cast %[[THREE_LITERAL]] : !s32i to si32
// CHECK-NEXT: acc.parallel num_workers(%[[CONV_CAST]] : si32, %[[TWO_CAST]] : si32 [#acc.device_type<nvidia>], %[[TWO_CAST]] : si32 [#acc.device_type<host>], %[[THREE_CAST]] : si32 [#acc.device_type<radeon>]) {
// CHECK-NEXT: acc.yield
// CHECK-NEXT: } loc

#pragma acc parallel num_workers(cond) device_type(nvidia) num_workers(2) device_type(radeon, multicore) num_workers(4)
{}
// CHECK-NEXT: %[[COND_LOAD:.*]] = cir.load %[[COND]] : !cir.ptr<!s32i>, !s32i
// CHECK-NEXT: %[[CONV_CAST:.*]] = builtin.unrealized_conversion_cast %[[COND_LOAD]] : !s32i to si32
// CHECK-NEXT: %[[TWO_LITERAL:.*]] = cir.const #cir.int<2> : !s32i
// CHECK-NEXT: %[[TWO_CAST:.*]] = builtin.unrealized_conversion_cast %[[TWO_LITERAL]] : !s32i to si32
// CHECK-NEXT: %[[FOUR_LITERAL:.*]] = cir.const #cir.int<4> : !s32i
// CHECK-NEXT: %[[FOUR_CAST:.*]] = builtin.unrealized_conversion_cast %[[FOUR_LITERAL]] : !s32i to si32
// CHECK-NEXT: acc.parallel num_workers(%[[CONV_CAST]] : si32, %[[TWO_CAST]] : si32 [#acc.device_type<nvidia>], %[[FOUR_CAST]] : si32 [#acc.device_type<radeon>], %[[FOUR_CAST]] : si32 [#acc.device_type<multicore>]) {
// CHECK-NEXT: acc.yield
// CHECK-NEXT: } loc

#pragma acc parallel device_type(nvidia) num_workers(2) device_type(radeon) num_workers(3)
{}
// CHECK-NEXT: %[[TWO_LITERAL:.*]] = cir.const #cir.int<2> : !s32i
// CHECK-NEXT: %[[TWO_CAST:.*]] = builtin.unrealized_conversion_cast %[[TWO_LITERAL]] : !s32i to si32
// CHECK-NEXT: %[[THREE_LITERAL:.*]] = cir.const #cir.int<3> : !s32i
// CHECK-NEXT: %[[THREE_CAST:.*]] = builtin.unrealized_conversion_cast %[[THREE_LITERAL]] : !s32i to si32
// CHECK-NEXT: acc.parallel num_workers(%[[TWO_CAST]] : si32 [#acc.device_type<nvidia>], %[[THREE_CAST]] : si32 [#acc.device_type<radeon>]) {
// CHECK-NEXT: acc.yield
// CHECK-NEXT: } loc

// CHECK-NEXT: cir.return
}
Loading