[Mlir-commits] [mlir] [mlir][vector] Fix multi-reduction unrolling with a lower rank target shape (PR #216799)

[Andrzej Warzyński]

================
@@ -864,3 +864,29 @@ func.func @deinterleave_2d(%v: vector<4x8xi32>) -> (vector<4x4xi32>, vector<4x4x
 // CHECK:   {{.*}} = vector.insert_strided_slice {{.*}}, {{.*}} {offsets = [2, 0], strides = [1, 1]} : vector<2x4xi32> into vector<4x4xi32>
 // CHECK:   {{.*}} = vector.insert_strided_slice {{.*}}, {{.*}} {offsets = [2, 0], strides = [1, 1]} : vector<2x4xi32> into vector<4x4xi32>
 // CHECK:   return {{.*}}, {{.*}} : vector<4x4xi32>, vector<4x4xi32>
+
+// -----
+
+// The native unroll shape {2, 2} has fewer dimensions than the source vector.
----------------
banach-space wrote:

I don't follow this comment, the source vectors in this example are  `vector<2x2x4xf32>` and `vector<2x2xf32>` - that's rank 3 and rank 2, wheres `{2, 2}` means rank 2. What dimensions are missing here?

https://github.com/llvm/llvm-project/pull/216799