[clang] bde0a8c - [Clang][HIP][CUDA] Add `__cluster_dims__` and `__no_cluster__` attribute (#156686)
via cfe-commits
cfe-commits at lists.llvm.org
Fri Oct 17 08:39:46 PDT 2025
Author: Shilei Tian
Date: 2025-10-17T11:39:43-04:00
New Revision: bde0a8cd632ec714ea26e04dea0b7bd0f3231cfa
URL: https://github.com/llvm/llvm-project/commit/bde0a8cd632ec714ea26e04dea0b7bd0f3231cfa
DIFF: https://github.com/llvm/llvm-project/commit/bde0a8cd632ec714ea26e04dea0b7bd0f3231cfa.diff
LOG: [Clang][HIP][CUDA] Add `__cluster_dims__` and `__no_cluster__` attribute (#156686)
This PR adds basic frontend support for `__cluster_dims__` and
`__no_cluster__` attribute.
In CUDA/HIP programming, the ``__cluster_dims__`` attribute can be
applied to a kernel function to set the dimensions of a thread block
cluster. The ``__no_cluster__`` attribute can be applied to a kernel
function to indicate that the thread block cluster feature will not be
enabled at both compile time and kernel launch time. Note that
`__no_cluster__` is a LLVM/Clang only attribute.
Co-authored-by: Yaxun (Sam) Liu <yaxun.liu at amd.com>
Co-authored-by: Jay Foad <jay.foad at amd.com>
Added:
clang/test/CodeGenCUDA/cluster_dims.cu
clang/test/SemaCUDA/cluster_dims.cu
Modified:
clang/include/clang/Basic/Attr.td
clang/include/clang/Basic/AttrDocs.td
clang/include/clang/Basic/DiagnosticSemaKinds.td
clang/include/clang/Sema/Sema.h
clang/lib/CodeGen/Targets/AMDGPU.cpp
clang/lib/Headers/__clang_hip_runtime_wrapper.h
clang/lib/Sema/SemaDeclAttr.cpp
clang/lib/Sema/SemaTemplateInstantiateDecl.cpp
clang/test/CodeGenCUDA/Inputs/cuda.h
clang/test/Misc/pragma-attribute-supported-attributes-list.test
clang/test/SemaCUDA/Inputs/cuda.h
Removed:
################################################################################
diff --git a/clang/include/clang/Basic/Attr.td b/clang/include/clang/Basic/Attr.td
index 22e60aa9fe312..eb48a0c01fd1e 100644
--- a/clang/include/clang/Basic/Attr.td
+++ b/clang/include/clang/Basic/Attr.td
@@ -1572,6 +1572,23 @@ def HIPManaged : InheritableAttr {
let Documentation = [HIPManagedAttrDocs];
}
+def CUDAClusterDims : InheritableAttr {
+ let Spellings = [GNU<"cluster_dims">];
+ let Args = [ExprArgument<"X">, ExprArgument<"Y", /*opt=*/1>, ExprArgument<"Z", /*opt=*/1>];
+ let Subjects = SubjectList<[ObjCMethod, FunctionLike]>;
+ let LangOpts = [CUDA];
+ let Documentation = [CUDAClusterDimsAttrDoc];
+}
+
+def CUDANoCluster : InheritableAttr {
+ let Spellings = [GNU<"no_cluster">];
+ let Subjects = SubjectList<[ObjCMethod, FunctionLike]>;
+ let LangOpts = [CUDA];
+ let Documentation = [CUDANoClusterAttrDoc];
+}
+
+def : MutualExclusions<[CUDAClusterDims, CUDANoCluster]>;
+
def CUDAInvalidTarget : InheritableAttr {
let Spellings = [];
let Subjects = SubjectList<[Function]>;
diff --git a/clang/include/clang/Basic/AttrDocs.td b/clang/include/clang/Basic/AttrDocs.td
index e0bbda083b5cf..2fdd041c1b46e 100644
--- a/clang/include/clang/Basic/AttrDocs.td
+++ b/clang/include/clang/Basic/AttrDocs.td
@@ -7545,6 +7545,45 @@ A managed variable can be accessed in both device and host code.
}];
}
+def CUDAClusterDimsAttrDoc : Documentation {
+ let Category = DocCatDecl;
+ let Content = [{
+In CUDA/HIP programming, the ``cluster_dims`` attribute, conventionally exposed as the
+``__cluster_dims__`` macro, can be applied to a kernel function to set the dimensions of a
+thread block cluster, which is an optional level of hierarchy and made up of thread blocks.
+``__cluster_dims__`` defines the cluster size as ``(X, Y, Z)``, where each value is the number
+of thread blocks in that dimension. The ``cluster_dims`` and `no_cluster`` attributes are
+mutually exclusive.
+
+.. code::
+
+ __global__ __cluster_dims__(2, 1, 1) void kernel(...) {
+ ...
+ }
+
+ }];
+}
+
+def CUDANoClusterAttrDoc : Documentation {
+ let Category = DocCatDecl;
+ let Content = [{
+In CUDA/HIP programming, a kernel function can still be launched with the cluster feature enabled
+at runtime, even without being annotated with ``__cluster_dims__``. The LLVM/Clang-exclusive
+``no_cluster`` attribute, conventionally exposed as the ``__no_cluster__`` macro, can be applied to
+a kernel function to explicitly indicate that the cluster feature will not be enabled either at
+compile time or at kernel launch time. This allows the compiler to apply certain optimizations
+without assuming that clustering could be enabled at runtime. It is undefined behavior to launch a
+kernel annotated with ``__no_cluster__`` if the cluster feature is enabled at runtime.
+The ``cluster_dims`` and ``no_cluster`` attributes are mutually exclusive.
+
+.. code::
+
+ __global__ __no_cluster__ void kernel(...) {
+ ...
+ }
+ }];
+}
+
def LifetimeOwnerDocs : Documentation {
let Category = DocCatDecl;
let Content = [{
diff --git a/clang/include/clang/Basic/DiagnosticSemaKinds.td b/clang/include/clang/Basic/DiagnosticSemaKinds.td
index 12fd7b08683e1..22de85d90a3cf 100644
--- a/clang/include/clang/Basic/DiagnosticSemaKinds.td
+++ b/clang/include/clang/Basic/DiagnosticSemaKinds.td
@@ -13070,6 +13070,12 @@ def warn_cuda_maxclusterrank_sm_90 : Warning<
"maxclusterrank requires sm_90 or higher, CUDA arch provided: %0, ignoring "
"%1 attribute">, InGroup<IgnoredAttributes>;
+def err_cluster_attr_not_supported : Error<
+ "%0 is not supported for this GPU architecture">;
+
+def err_cluster_dims_too_large : Error<
+ "cluster does not support more than %0 thread blocks; %1 provided">;
+
// VTable pointer authentication errors
def err_non_polymorphic_vtable_pointer_auth : Error<
"cannot set vtable pointer authentication on monomorphic type %0">;
diff --git a/clang/include/clang/Sema/Sema.h b/clang/include/clang/Sema/Sema.h
index e2444bc03327f..e3faaac7046e8 100644
--- a/clang/include/clang/Sema/Sema.h
+++ b/clang/include/clang/Sema/Sema.h
@@ -5010,6 +5010,14 @@ class Sema final : public SemaBase {
void AddLaunchBoundsAttr(Decl *D, const AttributeCommonInfo &CI,
Expr *MaxThreads, Expr *MinBlocks, Expr *MaxBlocks);
+ /// Add a cluster_dims attribute to a particular declaration.
+ CUDAClusterDimsAttr *createClusterDimsAttr(const AttributeCommonInfo &CI,
+ Expr *X, Expr *Y, Expr *Z);
+ void addClusterDimsAttr(Decl *D, const AttributeCommonInfo &CI, Expr *X,
+ Expr *Y, Expr *Z);
+ /// Add a no_cluster attribute to a particular declaration.
+ void addNoClusterAttr(Decl *D, const AttributeCommonInfo &CI);
+
enum class RetainOwnershipKind { NS, CF, OS };
UuidAttr *mergeUuidAttr(Decl *D, const AttributeCommonInfo &CI,
diff --git a/clang/lib/CodeGen/Targets/AMDGPU.cpp b/clang/lib/CodeGen/Targets/AMDGPU.cpp
index 0fcbf7e458a34..16d5919d62cbb 100644
--- a/clang/lib/CodeGen/Targets/AMDGPU.cpp
+++ b/clang/lib/CodeGen/Targets/AMDGPU.cpp
@@ -402,6 +402,26 @@ void AMDGPUTargetCodeGenInfo::setFunctionDeclAttributes(
F->addFnAttr("amdgpu-max-num-workgroups", AttrVal.str());
}
+
+ if (auto *Attr = FD->getAttr<CUDAClusterDimsAttr>()) {
+ auto GetExprVal = [&](const auto &E) {
+ return E ? E->EvaluateKnownConstInt(M.getContext()).getExtValue() : 1;
+ };
+ unsigned X = GetExprVal(Attr->getX());
+ unsigned Y = GetExprVal(Attr->getY());
+ unsigned Z = GetExprVal(Attr->getZ());
+ llvm::SmallString<32> AttrVal;
+ llvm::raw_svector_ostream OS(AttrVal);
+ OS << X << ',' << Y << ',' << Z;
+ F->addFnAttr("amdgpu-cluster-dims", AttrVal.str());
+ }
+
+ // OpenCL doesn't support cluster feature.
+ const TargetInfo &TTI = M.getContext().getTargetInfo();
+ if ((IsOpenCLKernel &&
+ TTI.hasFeatureEnabled(TTI.getTargetOpts().FeatureMap, "clusters")) ||
+ FD->hasAttr<CUDANoClusterAttr>())
+ F->addFnAttr("amdgpu-cluster-dims", "0,0,0");
}
void AMDGPUTargetCodeGenInfo::setTargetAttributes(
diff --git a/clang/lib/Headers/__clang_hip_runtime_wrapper.h b/clang/lib/Headers/__clang_hip_runtime_wrapper.h
index da1e39ac7270e..fb0ece96e1418 100644
--- a/clang/lib/Headers/__clang_hip_runtime_wrapper.h
+++ b/clang/lib/Headers/__clang_hip_runtime_wrapper.h
@@ -25,6 +25,8 @@
#define __constant__ __attribute__((constant))
#define __managed__ __attribute__((managed))
+#define __cluster_dims__(...) __attribute__((cluster_dims(__VA_ARGS__)))
+
#if !defined(__cplusplus) || __cplusplus < 201103L
#define nullptr NULL;
#endif
diff --git a/clang/lib/Sema/SemaDeclAttr.cpp b/clang/lib/Sema/SemaDeclAttr.cpp
index e6f8748db7644..9475b8a684082 100644
--- a/clang/lib/Sema/SemaDeclAttr.cpp
+++ b/clang/lib/Sema/SemaDeclAttr.cpp
@@ -5676,6 +5676,114 @@ static void handleLaunchBoundsAttr(Sema &S, Decl *D, const ParsedAttr &AL) {
AL.getNumArgs() > 2 ? AL.getArgAsExpr(2) : nullptr);
}
+static std::pair<Expr *, int>
+makeClusterDimsArgExpr(Sema &S, Expr *E, const CUDAClusterDimsAttr &AL,
+ const unsigned Idx) {
+ if (!E || S.DiagnoseUnexpandedParameterPack(E))
+ return {};
+
+ // Accept template arguments for now as they depend on something else.
+ // We'll get to check them when they eventually get instantiated.
+ if (E->isInstantiationDependent())
+ return {E, 1};
+
+ std::optional<llvm::APSInt> I = E->getIntegerConstantExpr(S.Context);
+ if (!I) {
+ S.Diag(E->getExprLoc(), diag::err_attribute_argument_n_type)
+ << &AL << Idx << AANT_ArgumentIntegerConstant << E->getSourceRange();
+ return {};
+ }
+ // Make sure we can fit it in 4 bits.
+ if (!I->isIntN(4)) {
+ S.Diag(E->getExprLoc(), diag::err_ice_too_large)
+ << toString(*I, 10, false) << 4 << /*Unsigned=*/1;
+ return {};
+ }
+ if (*I < 0) {
+ S.Diag(E->getExprLoc(), diag::warn_attribute_argument_n_negative)
+ << &AL << Idx << E->getSourceRange();
+ }
+
+ return {ConstantExpr::Create(S.getASTContext(), E, APValue(*I)),
+ I->getZExtValue()};
+}
+
+CUDAClusterDimsAttr *Sema::createClusterDimsAttr(const AttributeCommonInfo &CI,
+ Expr *X, Expr *Y, Expr *Z) {
+ CUDAClusterDimsAttr TmpAttr(Context, CI, X, Y, Z);
+
+ auto [NewX, ValX] = makeClusterDimsArgExpr(*this, X, TmpAttr, /*Idx=*/0);
+ auto [NewY, ValY] = makeClusterDimsArgExpr(*this, Y, TmpAttr, /*Idx=*/1);
+ auto [NewZ, ValZ] = makeClusterDimsArgExpr(*this, Z, TmpAttr, /*Idx=*/2);
+
+ if (!NewX || (Y && !NewY) || (Z && !NewZ))
+ return nullptr;
+
+ int FlatDim = ValX * ValY * ValZ;
+ const llvm::Triple TT =
+ (!Context.getLangOpts().CUDAIsDevice && Context.getAuxTargetInfo())
+ ? Context.getAuxTargetInfo()->getTriple()
+ : Context.getTargetInfo().getTriple();
+ int MaxDim = 1;
+ if (TT.isNVPTX())
+ MaxDim = 8;
+ else if (TT.isAMDGPU())
+ MaxDim = 16;
+ else
+ return nullptr;
+
+ // A maximum of 8 thread blocks in a cluster is supported as a portable
+ // cluster size in CUDA. The number is 16 for AMDGPU.
+ if (FlatDim > MaxDim) {
+ Diag(CI.getLoc(), diag::err_cluster_dims_too_large) << MaxDim << FlatDim;
+ return nullptr;
+ }
+
+ return CUDAClusterDimsAttr::Create(Context, NewX, NewY, NewZ, CI);
+}
+
+void Sema::addClusterDimsAttr(Decl *D, const AttributeCommonInfo &CI, Expr *X,
+ Expr *Y, Expr *Z) {
+ if (auto *Attr = createClusterDimsAttr(CI, X, Y, Z))
+ D->addAttr(Attr);
+}
+
+void Sema::addNoClusterAttr(Decl *D, const AttributeCommonInfo &CI) {
+ D->addAttr(CUDANoClusterAttr::Create(Context, CI));
+}
+
+static void handleClusterDimsAttr(Sema &S, Decl *D, const ParsedAttr &AL) {
+ const TargetInfo &TTI = S.Context.getTargetInfo();
+ OffloadArch Arch = StringToOffloadArch(TTI.getTargetOpts().CPU);
+ if ((TTI.getTriple().isNVPTX() && Arch < clang::OffloadArch::SM_90) ||
+ (TTI.getTriple().isAMDGPU() &&
+ !TTI.hasFeatureEnabled(TTI.getTargetOpts().FeatureMap, "clusters"))) {
+ S.Diag(AL.getLoc(), diag::err_cluster_attr_not_supported) << AL;
+ return;
+ }
+
+ if (!AL.checkAtLeastNumArgs(S, /*Num=*/1) ||
+ !AL.checkAtMostNumArgs(S, /*Num=*/3))
+ return;
+
+ S.addClusterDimsAttr(D, AL, AL.getArgAsExpr(0),
+ AL.getNumArgs() > 1 ? AL.getArgAsExpr(1) : nullptr,
+ AL.getNumArgs() > 2 ? AL.getArgAsExpr(2) : nullptr);
+}
+
+static void handleNoClusterAttr(Sema &S, Decl *D, const ParsedAttr &AL) {
+ const TargetInfo &TTI = S.Context.getTargetInfo();
+ OffloadArch Arch = StringToOffloadArch(TTI.getTargetOpts().CPU);
+ if ((TTI.getTriple().isNVPTX() && Arch < clang::OffloadArch::SM_90) ||
+ (TTI.getTriple().isAMDGPU() &&
+ !TTI.hasFeatureEnabled(TTI.getTargetOpts().FeatureMap, "clusters"))) {
+ S.Diag(AL.getLoc(), diag::err_cluster_attr_not_supported) << AL;
+ return;
+ }
+
+ S.addNoClusterAttr(D, AL);
+}
+
static void handleArgumentWithTypeTagAttr(Sema &S, Decl *D,
const ParsedAttr &AL) {
if (!AL.isArgIdent(0)) {
@@ -7141,6 +7249,12 @@ ProcessDeclAttribute(Sema &S, Scope *scope, Decl *D, const ParsedAttr &AL,
case ParsedAttr::AT_CUDALaunchBounds:
handleLaunchBoundsAttr(S, D, AL);
break;
+ case ParsedAttr::AT_CUDAClusterDims:
+ handleClusterDimsAttr(S, D, AL);
+ break;
+ case ParsedAttr::AT_CUDANoCluster:
+ handleNoClusterAttr(S, D, AL);
+ break;
case ParsedAttr::AT_Restrict:
handleRestrictAttr(S, D, AL);
break;
diff --git a/clang/lib/Sema/SemaTemplateInstantiateDecl.cpp b/clang/lib/Sema/SemaTemplateInstantiateDecl.cpp
index 4ace03dd1d130..4863b4522a92d 100644
--- a/clang/lib/Sema/SemaTemplateInstantiateDecl.cpp
+++ b/clang/lib/Sema/SemaTemplateInstantiateDecl.cpp
@@ -707,6 +707,23 @@ static void instantiateDependentAMDGPUMaxNumWorkGroupsAttr(
S.AMDGPU().addAMDGPUMaxNumWorkGroupsAttr(New, Attr, XExpr, YExpr, ZExpr);
}
+static void instantiateDependentCUDAClusterDimsAttr(
+ Sema &S, const MultiLevelTemplateArgumentList &TemplateArgs,
+ const CUDAClusterDimsAttr &Attr, Decl *New) {
+ EnterExpressionEvaluationContext Unevaluated(
+ S, Sema::ExpressionEvaluationContext::ConstantEvaluated);
+
+ auto SubstElt = [&S, &TemplateArgs](Expr *E) {
+ return E ? S.SubstExpr(E, TemplateArgs).get() : nullptr;
+ };
+
+ Expr *XExpr = SubstElt(Attr.getX());
+ Expr *YExpr = SubstElt(Attr.getY());
+ Expr *ZExpr = SubstElt(Attr.getZ());
+
+ S.addClusterDimsAttr(New, Attr, XExpr, YExpr, ZExpr);
+}
+
// This doesn't take any template parameters, but we have a custom action that
// needs to happen when the kernel itself is instantiated. We need to run the
// ItaniumMangler to mark the names required to name this kernel.
@@ -929,6 +946,11 @@ void Sema::InstantiateAttrs(const MultiLevelTemplateArgumentList &TemplateArgs,
*this, TemplateArgs, *AMDGPUMaxNumWorkGroups, New);
}
+ if (const auto *CUDAClusterDims = dyn_cast<CUDAClusterDimsAttr>(TmplAttr)) {
+ instantiateDependentCUDAClusterDimsAttr(*this, TemplateArgs,
+ *CUDAClusterDims, New);
+ }
+
if (const auto *ParamAttr = dyn_cast<HLSLParamModifierAttr>(TmplAttr)) {
instantiateDependentHLSLParamModifierAttr(*this, TemplateArgs, ParamAttr,
Tmpl, New);
diff --git a/clang/test/CodeGenCUDA/Inputs/cuda.h b/clang/test/CodeGenCUDA/Inputs/cuda.h
index dc85eae0c5178..e7ad784335027 100644
--- a/clang/test/CodeGenCUDA/Inputs/cuda.h
+++ b/clang/test/CodeGenCUDA/Inputs/cuda.h
@@ -13,6 +13,8 @@
#endif
#define __launch_bounds__(...) __attribute__((launch_bounds(__VA_ARGS__)))
#define __grid_constant__ __attribute__((grid_constant))
+#define __cluster_dims__(...) __attribute__((cluster_dims(__VA_ARGS__)))
+#define __no_cluster__ __attribute__((no_cluster))
#else
#define __constant__
#define __device__
@@ -22,6 +24,8 @@
#define __managed__
#define __launch_bounds__(...)
#define __grid_constant__
+#define __cluster_dims__(...)
+#define __no_cluster__
#endif
struct dim3 {
diff --git a/clang/test/CodeGenCUDA/cluster_dims.cu b/clang/test/CodeGenCUDA/cluster_dims.cu
new file mode 100644
index 0000000000000..00635e3572a7f
--- /dev/null
+++ b/clang/test/CodeGenCUDA/cluster_dims.cu
@@ -0,0 +1,38 @@
+// RUN: %clang_cc1 -triple amdgcn-amd-amdhsa -target-cpu gfx1250 -fcuda-is-device -emit-llvm -x hip -o - %s | FileCheck %s
+// RUN: %clang_cc1 -triple x86_64-pc-linux-gnu -aux-triple amdgcn-amd-amdhsa -emit-llvm -x hip -o - %s | FileCheck --check-prefix=HOST %s
+
+#include "Inputs/cuda.h"
+
+const int constint = 4;
+
+// HOST-NOT: "amdgpu-cluster-dims"
+
+// CHECK: "amdgpu-cluster-dims"="2,2,2"
+__global__ void __cluster_dims__(2, 2, 2) test_literal_3d() {}
+
+// CHECK: "amdgpu-cluster-dims"="2,2,1"
+__global__ void __cluster_dims__(2, 2) test_literal_2d() {}
+
+// CHECK: "amdgpu-cluster-dims"="4,1,1"
+__global__ void __cluster_dims__(4) test_literal_1d() {}
+
+// CHECK: "amdgpu-cluster-dims"="4,2,1"
+__global__ void __cluster_dims__(constint, constint / 2, 1) test_constant() {}
+
+// CHECK: "amdgpu-cluster-dims"="0,0,0"
+__global__ void __no_cluster__ test_no_cluster() {}
+
+// CHECK: "amdgpu-cluster-dims"="7,1,1"
+template<unsigned a>
+__global__ void __cluster_dims__(a) test_template_1d() {}
+template __global__ void test_template_1d<7>();
+
+// CHECK: "amdgpu-cluster-dims"="2,6,1"
+template<unsigned a, unsigned b>
+__global__ void __cluster_dims__(a, b) test_template_2d() {}
+template __global__ void test_template_2d<2, 6>();
+
+// CHECK: "amdgpu-cluster-dims"="1,2,3"
+template<unsigned a, unsigned b, unsigned c>
+__global__ void __cluster_dims__(a, b, c) test_template_3d() {}
+template __global__ void test_template_3d<1, 2, 3>();
diff --git a/clang/test/Misc/pragma-attribute-supported-attributes-list.test b/clang/test/Misc/pragma-attribute-supported-attributes-list.test
index 73d4cb1769ed5..ab4153a64f028 100644
--- a/clang/test/Misc/pragma-attribute-supported-attributes-list.test
+++ b/clang/test/Misc/pragma-attribute-supported-attributes-list.test
@@ -35,6 +35,7 @@
// CHECK-NEXT: CFUnknownTransfer (SubjectMatchRule_function)
// CHECK-NEXT: CPUDispatch (SubjectMatchRule_function)
// CHECK-NEXT: CPUSpecific (SubjectMatchRule_function)
+// CHECK-NEXT: CUDAClusterDims (SubjectMatchRule_objc_method, SubjectMatchRule_hasType_functionType)
// CHECK-NEXT: CUDAConstant (SubjectMatchRule_variable)
// CHECK-NEXT: CUDADevice (SubjectMatchRule_function, SubjectMatchRule_variable)
// CHECK-NEXT: CUDADeviceBuiltinSurfaceType (SubjectMatchRule_record)
@@ -43,6 +44,7 @@
// CHECK-NEXT: CUDAGridConstant (SubjectMatchRule_variable_is_parameter)
// CHECK-NEXT: CUDAHost (SubjectMatchRule_function)
// CHECK-NEXT: CUDALaunchBounds (SubjectMatchRule_objc_method, SubjectMatchRule_hasType_functionType)
+// CHECK-NEXT: CUDANoCluster (SubjectMatchRule_objc_method, SubjectMatchRule_hasType_functionType)
// CHECK-NEXT: CUDAShared (SubjectMatchRule_variable)
// CHECK-NEXT: CXX11NoReturn (SubjectMatchRule_function)
// CHECK-NEXT: CallableWhen (SubjectMatchRule_function_is_member)
diff --git a/clang/test/SemaCUDA/Inputs/cuda.h b/clang/test/SemaCUDA/Inputs/cuda.h
index 10db947d8246c..2bf45e03d91c7 100644
--- a/clang/test/SemaCUDA/Inputs/cuda.h
+++ b/clang/test/SemaCUDA/Inputs/cuda.h
@@ -13,6 +13,8 @@
#define __managed__ __attribute__((managed))
#define __grid_constant__ __attribute__((grid_constant))
#define __launch_bounds__(...) __attribute__((launch_bounds(__VA_ARGS__)))
+#define __cluster_dims__(...) __attribute__((cluster_dims(__VA_ARGS__)))
+#define __no_cluster__ __attribute__((no_cluster))
struct dim3 {
unsigned x, y, z;
diff --git a/clang/test/SemaCUDA/cluster_dims.cu b/clang/test/SemaCUDA/cluster_dims.cu
new file mode 100644
index 0000000000000..dcb8737a51006
--- /dev/null
+++ b/clang/test/SemaCUDA/cluster_dims.cu
@@ -0,0 +1,64 @@
+// RUN: %clang_cc1 -triple amdgcn-amd-amdhsa -fcuda-is-device -ast-print -x hip -verify=NS,all %s
+// RUN: %clang_cc1 -triple nvptx-nvidia-cuda -fcuda-is-device -ast-print -x hip -verify=NS,all %s
+// RUN: %clang_cc1 -triple amdgcn-amd-amdhsa -target-cpu gfx1250 -fcuda-is-device -ast-print -x hip -verify=amd,common,all %s | FileCheck -check-prefixes=CHECK %s
+// RUN: %clang_cc1 -triple nvptx-nvidia-cuda -target-cpu sm_90 -fcuda-is-device -ast-print -x hip -verify=cuda,common,all %s | FileCheck -check-prefixes=CHECK %s
+// RUN: %clang_cc1 -triple x86_64-pc-linux-gnu -aux-triple amdgcn-amd-amdhsa -ast-print -x hip -verify=amd,common,all %s | FileCheck -check-prefixes=CHECK %s
+
+#include "Inputs/cuda.h"
+
+const int constint = 4;
+
+// CHECK: __attribute__((global)) __attribute__((cluster_dims(2, 2, 2))) void test_literal_3d()
+__global__ void __cluster_dims__(2, 2, 2) test_literal_3d() {} //NS-error {{'cluster_dims' is not supported for this GPU architecture}}
+
+// CHECK: __attribute__((global)) __attribute__((cluster_dims(2, 2))) void test_literal_2d()
+__global__ void __cluster_dims__(2, 2) test_literal_2d() {} //NS-error {{'cluster_dims' is not supported for this GPU architecture}}
+
+// CHECK: __attribute__((global)) __attribute__((cluster_dims(4))) void test_literal_1d()
+__global__ void __cluster_dims__(4) test_literal_1d() {} //NS-error {{'cluster_dims' is not supported for this GPU architecture}}
+
+// CHECK: __attribute__((global)) __attribute__((cluster_dims(constint, constint / 4, 1))) void test_constant()
+__global__ void __cluster_dims__(constint, constint / 4, 1) test_constant() {} //NS-error {{'cluster_dims' is not supported for this GPU architecture}}
+
+// CHECK: template <int x, int y, int z> void test_template() __attribute__((cluster_dims(x, y, z)))
+template <int x, int y, int z> void test_template(void) __cluster_dims__(x, y, z){} //NS-error {{'cluster_dims' is not supported for this GPU architecture}}
+
+// CHECK: template <int x, int y, int z> void test_template_expr() __attribute__((cluster_dims(x + constint, y, z)))
+template <int x, int y, int z> void test_template_expr(void) __cluster_dims__(x + constint, y, z) {} //NS-error {{'cluster_dims' is not supported for this GPU architecture}}
+
+//NS-error at +1 {{'cluster_dims' is not supported for this GPU architecture}}
+__global__ void __cluster_dims__(32, 2, 4) test_too_large_dim_0() {} // common-error {{integer constant expression evaluates to value 32 that cannot be represented in a 4-bit unsigned integer type}}
+
+// cuda-error at +2 {{cluster does not support more than 8 thread blocks; 64 provided}}
+// amd-error at +1 {{cluster does not support more than 16 thread blocks; 64 provided}}
+__global__ void __cluster_dims__(4, 4, 4) test_too_large_dim_1() {} // NS-error {{'cluster_dims' is not supported for this GPU architecture}}
+
+// cuda-error at +3 {{cluster does not support more than 8 thread blocks; 64 provided}}
+// amd-error at +2 {{cluster does not support more than 16 thread blocks; 64 provided}}
+template<unsigned a, unsigned b, unsigned c>
+__global__ void __cluster_dims__(a, b, c) test_too_large_dim_template() {} // NS-error {{'cluster_dims' is not supported for this GPU architecture}}
+template __global__ void test_too_large_dim_template<4, 4, 4>(); // common-note {{in instantiation of function template specialization 'test_too_large_dim_template<4U, 4U, 4U>' requested here}}
+
+int none_const_int = 4;
+
+//NS-error at +1 {{'cluster_dims' is not supported for this GPU architecture}}
+__global__ void __cluster_dims__(none_const_int, 2, 4) test_non_constant_0() {} // common-error {{'cluster_dims' attribute requires parameter 0 to be an integer constant}}
+
+//NS-error at +1 {{'cluster_dims' is not supported for this GPU architecture}}
+__global__ void __cluster_dims__(8, none_const_int / 2, 4) test_non_constant_1() {} // common-error {{'cluster_dims' attribute requires parameter 1 to be an integer constant}}
+
+//NS-error at +1 {{'cluster_dims' is not supported for this GPU architecture}}
+__global__ void __cluster_dims__(8, 2, none_const_int / 4) test_non_constant_2() {} // common-error {{'cluster_dims' attribute requires parameter 2 to be an integer constant}}
+
+//NS-error at +1 {{'no_cluster' is not supported for this GPU architecture}}
+__global__ void __no_cluster__ test_no_cluster() {}
+
+//NS-error at +2 {{'no_cluster' is not supported for this GPU architecture}}
+//NS-error at +1 {{'cluster_dims' is not supported for this GPU architecture}}
+__global__ void __no_cluster__ __cluster_dims__(2,2,2) test_have_both() {} // common-error {{'cluster_dims' and 'no_cluster' attributes are not compatible}} common-note {{conflicting attribute is here}}
+
+template <int... args>
+__cluster_dims__(args) void test_template_variadic_args(void) {} // all-error {{expression contains unexpanded parameter pack 'args'}}
+
+template <int... args>
+__cluster_dims__(1, args) void test_template_variadic_args_2(void) {} // all-error {{expression contains unexpanded parameter pack 'args'}}
More information about the cfe-commits
mailing list