[clang] [clang][CodeGen][SPIRV] Translate `amdgpu_flat_work_group_size` into `reqd_work_group_size`. (PR #116820)

Matt Arsenault via cfe-commits cfe-commits at lists.llvm.org
Thu Dec 19 01:20:43 PST 2024


================
@@ -245,6 +247,41 @@ SPIRVTargetCodeGenInfo::getGlobalVarAddressSpace(CodeGenModule &CGM,
   return DefaultGlobalAS;
 }
 
+void SPIRVTargetCodeGenInfo::setTargetAttributes(
+    const Decl *D, llvm::GlobalValue *GV, CodeGen::CodeGenModule &M) const {
+  if (!M.getLangOpts().HIP ||
+      M.getTarget().getTriple().getVendor() != llvm::Triple::AMD)
+    return;
+  if (GV->isDeclaration())
+    return;
+
+  auto F = dyn_cast<llvm::Function>(GV);
+  if (!F)
+    return;
+
+  auto FD = dyn_cast_or_null<FunctionDecl>(D);
+  if (!FD)
+    return;
+  if (!FD->hasAttr<CUDAGlobalAttr>())
+    return;
+
+  unsigned N = M.getLangOpts().GPUMaxThreadsPerBlock;
+  if (auto FlatWGS = FD->getAttr<AMDGPUFlatWorkGroupSizeAttr>())
+    N = FlatWGS->getMax()->EvaluateKnownConstInt(M.getContext()).getExtValue();
+
+  // We encode the maximum flat WG size in the first component of the 3D
+  // max_work_group_size attribute, which will get reverse translated into the
+  // original AMDGPU attribute when targeting AMDGPU.
+  auto Int32Ty = llvm::IntegerType::getInt32Ty(M.getLLVMContext());
+  llvm::Metadata *AttrMDArgs[] = {
+      llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(Int32Ty, N)),
+      llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(Int32Ty, 1)),
+      llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(Int32Ty, 1))};
+
+  F->setMetadata("max_work_group_size",
----------------
arsenm wrote:

Why metadata? I know the OpenCL stuff uses metadata, but I think that's because it predates arbitrary string attributes. This is also "setTargetAttributes".

https://github.com/llvm/llvm-project/pull/116820


More information about the cfe-commits mailing list