[SYCL] Introduce min_work_groups_per_cu and max_work_groups_per_mp (#11192)

The attributes match to CUDA's launch bounds
`minBlocksPerMultiprocessor`
and `maxBlocksPerCluster` respectively. See:
https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#cluster-dimension-directives-maxclusterrank
and

https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#launch-bounds
for details.

Author: Jakub Chlanda

clang/include/clang/Basic/Attr.td

@@ -1587,6 +1587,22 @@ def SYCLIntelMaxWorkGroupSize : InheritableAttr {
   let SupportsNonconformingLambdaSyntax = 1;
 }
 
+def SYCLIntelMinWorkGroupsPerComputeUnit : InheritableAttr {
+  let Spellings = [CXX11<"intel", "min_work_groups_per_cu">];
+  let Args = [ExprArgument<"Value">];
+  let LangOpts = [SYCLIsDevice, SilentlyIgnoreSYCLIsHost];
+  let Subjects = SubjectList<[Function], ErrorDiag>;
+  let Documentation = [SYCLIntelMinWorkGroupsPerComputeUnitAttrDocs];
+}
+
+def SYCLIntelMaxWorkGroupsPerMultiprocessor : InheritableAttr {
+  let Spellings = [CXX11<"intel", "max_work_groups_per_mp">];
+  let Args = [ExprArgument<"Value">];
+  let LangOpts = [SYCLIsDevice, SilentlyIgnoreSYCLIsHost];
+  let Subjects = SubjectList<[Function], ErrorDiag>;
+  let Documentation = [SYCLIntelMaxWorkGroupsPerMultiprocessorDocs];
+}
+
 def SYCLIntelMaxGlobalWorkDim : InheritableAttr {
   let Spellings = [CXX11<"intel", "max_global_work_dim">];
   let Args = [ExprArgument<"Value">];

clang/include/clang/Basic/AttrDocs.td

@@ -3017,6 +3017,65 @@ In SYCL 2020 mode, the attribute is not propagated to the kernel.
   }];
 }
 
+def SYCLIntelMinWorkGroupsPerComputeUnitAttrDocs: Documentation {
+  let Category = DocCatFunction;
+  let Heading = "intel::min_work_groups_per_cu";
+  let Content = [{
+Applies to a device function/lambda function. Indicates the desired minimum
+number of resident work_groups per multiprocessor. It complies to the
+.minnctapersm PTX directive.
+
+.. code-block:: c++
+
+  [[intel::min_work_groups_per_cu(2)]] void foo() {}
+
+  class Foo {
+  public:
+    [[intel::min_work_groups_per_cu(2)]] void operator()() const {}
+  };
+
+  template <int N>
+  class Functor {
+  public:
+    [[intel::min_work_groups_per_cu(N)]] void operator()() const {}
+  };
+
+  template <int N>
+  [[intel::min_work_groups_per_cu(N)]] void func() {}
+
+  }];
+}
+
+def SYCLIntelMaxWorkGroupsPerMultiprocessorDocs: Documentation {
+  let Category = DocCatFunction;
+  let Heading = "intel::max_work_groups_per_mp";
+  let Content = [{
+Applies to a device function/lambda function. Indicates the desired maximum
+number work_groups per cluster with which the application will ever launch. It
+complies to the .maxclusterrank PTX directive. Note, that the feature requires
+SM_90 or higher.
+
+.. code-block:: c++
+
+  [[intel::max_work_groups_per_mp(2)]] void foo() {}
+
+  class Foo {
+  public:
+    [[intel::max_work_groups_per_mp(2)]] void operator()() const {}
+  };
+
+  template <int N>
+  class Functor {
+  public:
+    [[intel::max_work_groups_per_mp(N)]] void operator()() const {}
+  };
+
+  template <int N>
+  [[intel::max_work_groups_per_mp(N)]] void func() {}
+
+  }];
+}
+
 def SYCLIntelMaxGlobalWorkDimAttrDocs : Documentation {
   let Category = DocCatFunction;
   let Heading = "intel::max_global_work_dim";

clang/include/clang/Basic/DiagnosticSemaKinds.td

@@ -12016,9 +12016,12 @@ def warn_sycl_kernel_return_type : Warning<
 def err_sycl_special_type_num_init_method : Error<
   "types with 'sycl_special_class' attribute must have one and only one '__init' "
   "method defined">;
+def warn_launch_bounds_is_cuda_specific : Warning<
+  "%0 attribute ignored, only applicable when targeting Nvidia devices">,
+  InGroup<IgnoredAttributes>;
 
 def warn_cuda_maxclusterrank_sm_90 : Warning<
-  "maxclusterrank requires sm_90 or higher, CUDA arch provided: %0, ignoring "
+  "'maxclusterrank' requires sm_90 or higher, CUDA arch provided: %0, ignoring "
   "%1 attribute">, InGroup<IgnoredAttributes>;
 
 def err_bit_int_bad_size : Error<"%select{signed|unsigned}0 _BitInt must "

clang/include/clang/Sema/Sema.h

@@ -11432,6 +11432,16 @@ class Sema final {
   SYCLIntelMaxGlobalWorkDimAttr *
   MergeSYCLIntelMaxGlobalWorkDimAttr(Decl *D,
                                      const SYCLIntelMaxGlobalWorkDimAttr &A);
+  void AddSYCLIntelMinWorkGroupsPerComputeUnitAttr(
+      Decl *D, const AttributeCommonInfo &CI, Expr *E);
+  SYCLIntelMinWorkGroupsPerComputeUnitAttr *
+  MergeSYCLIntelMinWorkGroupsPerComputeUnitAttr(
+      Decl *D, const SYCLIntelMinWorkGroupsPerComputeUnitAttr &A);
+  void AddSYCLIntelMaxWorkGroupsPerMultiprocessorAttr(
+      Decl *D, const AttributeCommonInfo &CI, Expr *E);
+  SYCLIntelMaxWorkGroupsPerMultiprocessorAttr *
+  MergeSYCLIntelMaxWorkGroupsPerMultiprocessorAttr(
+      Decl *D, const SYCLIntelMaxWorkGroupsPerMultiprocessorAttr &A);
   void AddSYCLIntelBankWidthAttr(Decl *D, const AttributeCommonInfo &CI,
                                  Expr *E);
   SYCLIntelBankWidthAttr *

clang/lib/CodeGen/CodeGenFunction.cpp

@@ -758,6 +758,24 @@ void CodeGenFunction::EmitKernelMetadata(const FunctionDecl *FD,
                     llvm::MDNode::get(Context, AttrMDArgs));
   }
 
+  auto attrAsMDArg = [&](Expr *E) {
+    const auto *CE = cast<ConstantExpr>(E);
+    std::optional<llvm::APSInt> ArgVal = CE->getResultAsAPSInt();
+    return llvm::ConstantAsMetadata::get(
+        Builder.getInt32(ArgVal->getSExtValue()));
+  };
+
+  if (const auto *A = FD->getAttr<SYCLIntelMinWorkGroupsPerComputeUnitAttr>()) {
+    Fn->setMetadata("min_work_groups_per_cu",
+                    llvm::MDNode::get(Context, {attrAsMDArg(A->getValue())}));
+  }
+
+  if (const auto *A =
+          FD->getAttr<SYCLIntelMaxWorkGroupsPerMultiprocessorAttr>()) {
+    Fn->setMetadata("max_work_groups_per_mp",
+                    llvm::MDNode::get(Context, {attrAsMDArg(A->getValue())}));
+  }
+
   if (const SYCLIntelMaxWorkGroupSizeAttr *A =
           FD->getAttr<SYCLIntelMaxWorkGroupSizeAttr>()) {
 

clang/lib/CodeGen/Targets/NVPTX.cpp

@@ -245,6 +245,31 @@ void NVPTXTargetCodeGenInfo::setTargetAttributes(
       // And kernel functions are not subject to inlining
       F->addFnAttr(llvm::Attribute::NoInline);
     }
+    if (const auto *MWGS = FD->getAttr<SYCLIntelMaxWorkGroupSizeAttr>()) {
+      auto MaxThreads = (*MWGS->getZDimVal()).getExtValue() *
+                        (*MWGS->getYDimVal()).getExtValue() *
+                        (*MWGS->getXDimVal()).getExtValue();
+      if (MaxThreads > 0)
+        addNVVMMetadata(F, "maxntidx", MaxThreads);
+
+      auto attrValue = [&](Expr *E) {
+        const auto *CE = cast<ConstantExpr>(E);
+        std::optional<llvm::APInt> Val = CE->getResultAsAPSInt();
+        return Val->getZExtValue();
+      };
+
+      if (const auto *MWGPCU =
+              FD->getAttr<SYCLIntelMinWorkGroupsPerComputeUnitAttr>()) {
+        // The value is guaranteed to be > 0, pass it to the metadata.
+        addNVVMMetadata(F, "minnctapersm", attrValue(MWGPCU->getValue()));
+
+        if (const auto *MWGPMP =
+                FD->getAttr<SYCLIntelMaxWorkGroupsPerMultiprocessorAttr>()) {
+          // The value is guaranteed to be > 0, pass it to the metadata.
+          addNVVMMetadata(F, "maxclusterrank", attrValue(MWGPMP->getValue()));
+        }
+      }
+    }
   }
 
   // Perform special handling in CUDA mode.

clang/lib/Sema/SemaDecl.cpp

@@ -2999,6 +2999,12 @@ static bool mergeDeclAttribute(Sema &S, NamedDecl *D,
     NewAttr = S.MergeSYCLIntelInitiationIntervalAttr(D, *A);
   else if (const auto *A = dyn_cast<SYCLWorkGroupSizeHintAttr>(Attr))
     NewAttr = S.MergeSYCLWorkGroupSizeHintAttr(D, *A);
+  else if (const auto *A =
+               dyn_cast<SYCLIntelMinWorkGroupsPerComputeUnitAttr>(Attr))
+    NewAttr = S.MergeSYCLIntelMinWorkGroupsPerComputeUnitAttr(D, *A);
+  else if (const auto *A =
+               dyn_cast<SYCLIntelMaxWorkGroupsPerMultiprocessorAttr>(Attr))
+    NewAttr = S.MergeSYCLIntelMaxWorkGroupsPerMultiprocessorAttr(D, *A);
   else if (const auto *A = dyn_cast<SYCLIntelMaxGlobalWorkDimAttr>(Attr))
     NewAttr = S.MergeSYCLIntelMaxGlobalWorkDimAttr(D, *A);
   else if (const auto *BTFA = dyn_cast<BTFDeclTagAttr>(Attr))