[llvm-branch-commits] [llvm] [mlir] [mlir][GPU][transform] Add gpu_to_rocdl conversion pattern to transfo… (PR #146962)

[Oleksandr Alex Zinenko via llvm-branch-commits]

================
@@ -129,6 +131,42 @@ LogicalResult transform::ApplyGPUSubgroupReduceToNVVMConversionPatternsOp::
   return success();
 }
 
+void transform::ApplyGPUToROCDLConversionPatternsOp::populatePatterns(
+    TypeConverter &typeConverter, RewritePatternSet &patterns) {
+  auto &llvmTypeConverter = static_cast<LLVMTypeConverter &>(typeConverter);
+  populateGpuMemorySpaceAttributeConversions(
+      llvmTypeConverter, [](AddressSpace space) {
+        switch (space) {
+        case AddressSpace::Global:
+          return 1;
+        case AddressSpace::Workgroup:
+          return 3;
+        case AddressSpace::Private:
+          return 5;
----------------
ftynse wrote:

Are you sure these address spaces are correct for ROCDL?

https://github.com/llvm/llvm-project/pull/146962