[Mlir-commits] [mlir] [MLIR][AMDGPU] Add amdgpu.global_transpose_load op for RDNA4 global memory transpose loads (PR #195287)

[Krzysztof Drewniak]

================
@@ -1079,6 +1079,41 @@ LogicalResult TransposeLoadOp::verify() {
   return success();
 }
 
+//===----------------------------------------------------------------------===//
+// GlobalTransposeLoadOp
+//===----------------------------------------------------------------------===//
+
+LogicalResult GlobalTransposeLoadOp::verify() {
+  MemRefType srcType = cast<MemRefType>(getSrc().getType());
+
+  if (!hasGlobalMemorySpace(srcType.getMemorySpace()))
+    return emitOpError("source memory address space must be Global");
+
+  auto resultType = cast<VectorType>(getType());
+  size_t numElements = resultType.getNumElements();
+  size_t elementTypeSize =
+      resultType.getElementType().getIntOrFloatBitWidth();
+
+  // ElementSize -> NumElements (matches ISA-documented global_load_tr variants)
+  const llvm::SmallDenseMap<size_t, size_t> kValidLoadSizeMap = {
+      {8, 8},  // global_load_tr_b64
----------------
krzysz00 wrote:

Missing cases as above

https://github.com/llvm/llvm-project/pull/195287