// -----

func.func @ensureBiasForConv2d(%arg0: tensor<128x32x32x3xf32>, %arg1: tensor<32x1x1x3xf32>) -> tensor<128x32x32x32xf32> {
  %cst = "tfl.no_value"() {value = unit} : () -> none

^bb0(%arg0: tensor<256x32x32x3xf32>, %arg1: tensor<16x3x3x3xf32>, %arg2: tensor<16xf32>):
  // CHECK: _arithmetic_count = 230686720 : i64

  %cst = "tf.Const"() {device = "", value = dense_resource<__elided__> : tensor<2x3x3x2xbf16>} : () -> tensor<2x3x3x2xbf16>
  %0 = "tf.Cast"(%arg0) {Truncate = false, device = ""} : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xbf16>
  %1 = "tf.Conv2D"(%0, %cst) {data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true} : (tensor<1x3x4x3xbf16>, tensor<2x3x3x2xbf16>) -> tensor<1x3x2x2xbf16>

  // Input: [N, H, W, C] for Conv2D or [N, D, H, W, C] for Conv3D.
  dnums.set_input_batch_dimension(0);
  dnums.set_input_feature_dimension(num_dims - 1);
  // Kernel: [K, K, I, O] for Conv2D or [K, K, K, I, O] for Conv3D.
  dnums.set_kernel_input_feature_dimension(num_dims - 2);
  dnums.set_kernel_output_feature_dimension(num_dims - 1);
  // Output: [N, H, W, C] for Conv2D or [N, D, H, W, C] for Conv3D.

// -----

func.func @conv_with_non_constant_filter(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> tensor<*xf32> {
  %cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<2xf32>} : () -> tensor<2xf32>
  %0 = "tf.Conv2D"(%arg0, %arg1) {data_format = "NHWC", dilations = [1, 1, 2, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<*xf32>

// CHECK: return %0 : tensor<*xf32>
}

// -----

module {
  func.func @conv2d(%arg0: tensor<1x3x4x512xf32>) -> (tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<2xf32>} : () -> tensor<2xf32>

  const auto& subgraph = model_.subgraphs[0];
  auto conv_op = subgraph->operators[0].get();
  const int input_tensor_idx = 0;
  const int weights_tensor_idx = 1;
  const int bias_tensor_index = 2;
  const int output_tensor_idx = 0;
  const auto bias_tensor =
      subgraph->tensors[conv_op->inputs[bias_tensor_index]].get();
  const auto input_tensor =
      subgraph->tensors[conv_op->inputs[input_tensor_idx]].get();

}

func Conv2(v interface{}) (string, bool) {
	return conv2[string](v)
}

func conv2[T any](v interface{}) (T, bool) {
	x, ok := v.(T)
	return x, ok
}

func Conv3(v interface{}) string {
	return conv3[string](v)
}

func conv3[T any](v interface{}) T {
	switch v := v.(type) {
	case T:
		return v
	default:
		var z T
		return z
	}
}

type Mystring string

func (Mystring) Foo() {
}

      %outputs_2, %control_3 = tf_executor.island(%control_1) wraps "tf.Const"() {value = dense<0.000000e+00> : tensor<5x5x32x16xf32>} : () -> tensor<5x5x32x16xf32>
      %outputs_4, %control_5 = tf_executor.island wraps "tf.Conv2D"(%outputs, %outputs_2) {data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "VALID", strides = [1, 1, 1, 1], use_cudnn_on_gpu = true} : (tensor<*xf32>, tensor<5x5x32x16xf32>) -> tensor<*xf32>

  %cst_0 = "tf.Const"() {value = dense<-1.000000e+00> : tensor<f32>} : () -> tensor<f32>
  %cst_1 = "tf.Const"() {value = dense<1.000000e+00> : tensor<f32>} : () -> tensor<f32>
  %0 = "tf.Conv2D"(%arg0, %cst) {data_format = "NHWC", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 1, 1]} : (tensor<1x3x4x3xf32>, tensor<1x1x3x2xf32>) -> tensor<1x3x4x2xf32>