[flang-commits] [flang] [flang][cuda] Flatten memref descriptors in GPU kernel argument packing (PR #193651)

[Zhen Wang via flang-commits]

================
@@ -33,16 +34,41 @@ using namespace Fortran::runtime;
 
 namespace {
 
+// Build the kernel argument array used for the CUDA kernel launch.
+//
+// For each operand:
+//   - memref operands are unpacked into their descriptor scalar fields so the
+//     host-side parameter list matches the NVVM-lowered device kernel signature
+//     (gpu-to-nvvm expands each memref into 3 + 2*rank scalar parameters).
+//     We delegate to MemRefDescriptor::unpack so we follow the canonical memref
+//     descriptor layout owned by the LLVMCommon library;
+//   - all other operands are passed through unchanged.
+//
+// The flattened values are materialized on the stack in a single struct
+// (preserving argument order), and a companion pointer array is populated with
+// the address of each field. That pointer array is what the CUDA launch
+// interface expects as kernelParams.
 static mlir::Value createKernelArgArray(mlir::Location loc,
-                                        mlir::ValueRange operands,
+                                        mlir::ValueRange origOperands,
+                                        mlir::ValueRange adaptedOperands,
                                         mlir::PatternRewriter &rewriter) {
 
   auto *ctx = rewriter.getContext();
-  llvm::SmallVector<mlir::Type> structTypes(operands.size(), nullptr);
 
-  for (auto [i, arg] : llvm::enumerate(operands))
-    structTypes[i] = arg.getType();
+  llvm::SmallVector<mlir::Value, 8> flatValues;
+  flatValues.reserve(adaptedOperands.size());
+  for (auto [origArg, adaptedArg] :
+       llvm::zip_equal(origOperands, adaptedOperands)) {
+    if (auto memrefTy = mlir::dyn_cast<mlir::MemRefType>(origArg.getType())) {
+      mlir::MemRefDescriptor::unpack(rewriter, loc, adaptedArg, memrefTy,
----------------
wangzpgi wrote:

Good catch. Today useBarePtrCallConv is off in the CUDA Fortran pipeline, so MemRefDescriptor::unpack happens to match. But rather than depending on that, I will switch to LLVMTypeConverter::promoteOperands — it internally branches on useBarePtrCallConv and also handles UnrankedMemRefType, so host-side packing stays consistent with whatever the type converter is configured to do.

https://github.com/llvm/llvm-project/pull/193651