[Mlir-commits] [mlir] [mlir][XeGPU] Fix matrix op lowering for a payload that is not contiguous (PR #228638)

[Artem Kroviakov]

================
@@ -920,11 +907,156 @@ class CreateMemDescOpPattern final
   }
 };
 
+/// Returns the shape of the largest piece of a matrix op payload that one LLVM
+/// access can move, as an access only moves consecutive elements. Memory runs
+/// along the dimension whose stride is 1, so on a row-major tile a [1, 4]
+/// payload is four neighbours and moves in one access, while [2, 1] is two
+/// elements a row apart and needs one access each.
+static SmallVector<int64_t>
+getContiguousPayloadShape(ArrayRef<int64_t> payloadShape, size_t numOffsets,
+                          xegpu::MemDescType mdescTy) {
+  SmallVector<int64_t> shape(payloadShape);
+  int64_t rank = shape.size();
+  if (rank != static_cast<int64_t>(numOffsets))
+    return shape;
+
+  int64_t fastestDim = 0;
+  for (auto [dim, strideAttr] :
+       llvm::enumerate(mdescTy.getStrideAttr().getValue())) {
+    if (cast<IntegerAttr>(strideAttr).getInt() == 1) {
+      fastestDim = dim;
+      break;
+    }
+  }
+
+  bool spansFastestDim =
----------------
akroviakov wrote:

Does this utility assume a row-major walk order?

https://github.com/llvm/llvm-project/pull/228638