%cst = "tf.Const"() {value = dense<1.000000e+00> : tensor<2x3x3x3xf32>} : () -> tensor<2x3x3x3xf32>
  %cst_0 = "tf.Const"() {value = dense<0.500000e+00> : tensor<1x3x2x3xf32>} : () -> tensor<1x3x2x3xf32>
  %0 = "tf.Conv2D"(%arg0, %cst) {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<2x3x3x3xf32>) -> tensor<1x3x2x3xf32>

    func.return %result : tensor<10x24x24x64xf32>
  }

  // CHECK-LABEL: @max_pool_grad_same
  func.func @max_pool_grad_same(%orig_input: tensor<2x13x25x7xf32>, %orig_output: tensor<2x4x7x7xf32>, %grad: tensor<2x4x7x7xf32>) -> tensor<2x13x25x7xf32> {
    // CHECK: padding = dense<{{\[\[}}0, 0], [0, 1], [1, 1], [0, 0]]> : tensor<4x2xi64>
    %result = "tf.MaxPoolGrad"(%orig_input, %orig_output, %grad) {
      data_format = "NHWC",

// CHECK:           %[[VAL_6:.*]] = arith.constant dense<[0, 3, 1, 2]> : tensor<4xi32>
// CHECK:           %[[VAL_7:.*]] = "tfl.transpose"(%[[VAL_5]], %[[VAL_6]]) : (tensor<1x4x4x3xf32>, tensor<4xi32>) -> tensor<1x3x4x4xf32>

  %arg1: tensor<f32>
) -> tensor<2x4x3x3xf32> {
  %0 = "stablehlo.reduce_window"(%arg0, %arg1) ({
  ^bb0(%arg2: tensor<f32>, %arg3: tensor<f32>):
    %1 = stablehlo.maximum %arg2, %arg3 : tensor<f32>
    stablehlo.return %1 : tensor<f32>
  }) {window_dimensions = array<i64: 1, 3, 4, 1>, window_strides = array<i64: 1, 2, 3, 1>} : (tensor<2x9x10x3xf32>, tensor<f32>) -> tensor<2x4x3x3xf32>
  return %0 : tensor<2x4x3x3xf32>
}

}

// -----

// CHECK-LABEL: testSpaceToBatchND
func.func @testSpaceToBatchND(%arg0 : tensor<1x4x4x3xf32>, %arg1 : tensor<2xi32>, %arg2 : tensor<2x2xi32>) -> tensor<?xf32> {
  // CHECK: "tfl.space_to_batch_nd"(%arg0, %arg1, %arg2)
  %0 = "tfl.space_to_batch_nd"(%arg0, %arg1, %arg2) : (tensor<1x4x4x3xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<?xf32>
  func.return %0 : tensor<?xf32>
}

// -----

  // CHECK: %0 = "tfl.conv_2d"(%arg0, %arg1, %cst)
}

// CHECK-LABEL: fuse4DAddIntoConv2d
func.func @fuse4DAddIntoConv2d(%arg0: tensor<256x32x32x3xf32>, %arg1: tensor<2x3x3x3xf32>) -> tensor<256x32x32x2xf32> {
  %cst = arith.constant dense<[[[[1.0, 2.0]]]]> : tensor<1x1x1x2xf32>
  %cst_0 = arith.constant dense<[1.0, 2.0]> : tensor<2xf32>
  %0 = "tfl.conv_2d"(%arg0, %arg1, %cst_0) {