Allow transposed convolution weights with a non-default dim order

kernels/portable/cpu/util/kernel_ops_util.cpp

@@ -382,8 +382,13 @@ bool check_convolution_args(
   ET_LOG_AND_RETURN_IF_FALSE(tensors_have_same_dtype(in, weight, out));
 
   ET_LOG_AND_RETURN_IF_FALSE(tensor_is_default_or_channels_last_dim_order(in));
-  ET_LOG_AND_RETURN_IF_FALSE(
-      tensor_is_default_or_channels_last_dim_order(weight));
+  // Transposed weights can have a non-default dim order (e.g. out_channels == 1
+  // gives [1, 0, 2, 3]); the kernel indexes them by their strides, so any order
+  // is valid here.
+  if (!transposed) {
+    ET_LOG_AND_RETURN_IF_FALSE(
+        tensor_is_default_or_channels_last_dim_order(weight));
+  }
   ET_LOG_AND_RETURN_IF_FALSE(tensor_is_default_or_channels_last_dim_order(out));
 
   ET_CHECK_OR_RETURN_FALSE(

kernels/test/op_convolution_test.cpp

@@ -785,3 +785,43 @@ TEST_F(OpConvCorrectnessTest, BFloat16TypeSmokeTest) {
       out);
   EXPECT_TENSOR_CLOSE(out, expected);
 }
+
+// Regression test for #20804: a transposed convolution weight with out_channels == 1
+// arrives with a non-default dim order (e.g. [1, 0, 2, 3], since the weight is laid out
+// as (in_channels, out_channels / groups, kH, kW)). The kernel indexes the weight through
+// its strides, so it must run and produce the same result as a default-order weight.
+TEST_F(OpConvCorrectnessTest, TransposedWeightNonDefaultDimOrder) {
+  TensorFactory<ScalarType::Float> tf;
+
+  Tensor input = tf.make({1, 2, 2, 2}, {1, 2, 3, 4, 5, 6, 7, 8});
+
+  // Weight (in_channels=2, out_channels=1, kH=2, kW=2) in dim order [1, 0, 2, 3].
+  Tensor weight = tf.make_with_dimorder(
+      {2, 1, 2, 2}, {1, 2, 3, 4, 5, 6, 7, 8}, /*dim_order=*/{1, 0, 2, 3});
+
+  optional<Tensor> bias;
+  Tensor out = tf.zeros({1, 1, 3, 3});
+  Tensor expected =
+      tf.make({1, 1, 3, 3}, {26, 64, 40, 76, 184, 112, 58, 136, 80});
+
+  int64_t stride[2] = {1, 1};
+  int64_t padding[2] = {0, 0};
+  int64_t dilation[2] = {1, 1};
+  bool transposed = true;
+  int64_t output_padding[2] = {0, 0};
+  int64_t groups = 1;
+
+  op_convolution_out(
+      input,
+      weight,
+      std::optional<Tensor>(bias),
+      executorch::aten::ArrayRef<int64_t>{stride, 2},
+      executorch::aten::ArrayRef<int64_t>{padding, 2},
+      executorch::aten::ArrayRef<int64_t>{dilation, 2},
+      transposed,
+      executorch::aten::ArrayRef<int64_t>{output_padding, 2},
+      groups,
+      out);
+
+  EXPECT_TENSOR_CLOSE(out, expected);
+}