[Mlir-commits] [mlir] [mlir][nvgpu] Use the strides of the memref descriptor to construct the TMA descriptor (PR #85403)

[Guray Ozen]

================
@@ -427,13 +427,21 @@ namespace {
 
 template <int rank>
 void mgpuGetMemRefDataAndShape(void *raw_descriptor, char **addr,
-                               uint64_t *globalDim) {
+                               uint64_t *globalDim, uint64_t *globalStrides,
+                               const CUtensorMapDataType tensorDataType) {
   auto descriptor =
       reinterpret_cast<StridedMemRefType<char, rank> *>(raw_descriptor);
   *addr = descriptor->data;
   for (int i = 0; i < rank; ++i) {
     globalDim[i] = static_cast<uint64_t>(descriptor->sizes[rank - i - 1]);
   }
+  static constexpr int elementSizeInBytes[] = {1, 2, 4, 4, 8, 8, 2,
+                                               4, 8, 2, 4, 4, 4};
+  // TODO(grypp): Check that the minormost stride is equal to the element size.
----------------
grypp wrote:

LLVM doesn't use TODO with name. Let's just keep this as TODO

https://github.com/llvm/llvm-project/pull/85403