%40 = mhlo.subtract %39, %37 : tensor<1x64x32x32xf32>
  %41 = mhlo.multiply %40, %28 : tensor<1x64x32x32xf32>
  %42 = mhlo.add %37, %41 : tensor<1x64x32x32xf32>
  %43 = mhlo.subtract %42, %30 : tensor<1x64x32x32xf32>
  %44 = "mhlo.broadcast_in_dim"(%0) <{broadcast_dimensions = dense<[2, 3]> : tensor<2xi64>}> : (tensor<32x1xf32>) -> tensor<1x64x32x1xf32>

// CHECK-LABEL: func @transposeConv2D
func.func @transposeConv2D(%arg0: tensor<1x3x32x32xf32>, %arg1: tensor<1x1x3x8xf32>) -> tensor<1x8x32x32xf32> {

  // Convert input: NCHW -> NHWC
  %0 = "tf.Const"() {value = dense<[0, 2, 3, 1]> : tensor<4xi32>} : () -> tensor<4xi32>
  %1 = "tf.Transpose"(%arg0, %0) : (tensor<1x3x32x32xf32>, tensor<4xi32>) -> tensor<1x32x32x3xf32>

  // Compute in NHWC
  %2 = "tf.Conv2D"(%1, %arg1)
    {

  // CHECK-SAME: dilations = [1, 4, 2, 3]
  // CHECK-SAME: explicit_paddings = [1, 2, 7, 8, 3, 4, 5, 6]
  // CHECK-SAME: padding = "EXPLICIT"
  // CHECK-SAME: strides = [5, 8, 6, 7]
  // CHECK-SAME: (tensor<1x3x32x32xf32>, tensor<4xi32>, tensor<1x8x32x32xf32>)
  // CHECK-SAME: -> tensor<1x1x3x8xf32>

  // CHECK: return %[[CONV2D_BACKPROP]]

  %0 = "tf.Conv2DBackpropFilter"(%input, %filter_sizes, %out_backprop)
       {
         data_format = "NHWC",

    "tfl.yield"(%4, %5, %6, %7) : (tensor<1x32x32x32xf32>, tensor<1x32x32x32xf32>, tensor<1x32x32x32xf32>, tensor<1x32x32x32xf32>) -> ()

  func.return %0 : tensor<1x64x84x32xf32>

  %0 = "tfl.custom"(%arg0, %arg1, %arg2) {custom_code = "Convolution2DTransposeBias", custom_option = #tfl<const_bytes : "0x010000000200000002000000">} : (tensor<32x4x4x128xf32>, tensor<1x32x42x128xf32>, tensor<4xi32>) -> tensor<1x64x84x32xf32>
  func.return %0 : tensor<1x64x84x32xf32>
}
// CHECK-LABEL: main

  %0 = "tfl.custom"(%arg0, %arg1, %arg2) {custom_code = "Convolution2DTransposeBias", custom_option = #tfl<const_bytes : "0x010000000200000002000000">} : (tensor<32x4x4x128xf32>, tensor<1x32x42x128xf32>, tensor<4xi32>) -> tensor<1x64x84x32xf32>
  func.return %0 : tensor<1x64x84x32xf32>
}
// CHECK-LABEL: main

// dilations, etc...). This test only verifies that changing convolution data
// layout will update all the attributes.

// CHECK-LABEL: func @transposeConv2D
func.func @transposeConv2D(%input: tensor<1x3x32x32xf32>, %filter: tensor<1x1x3x8xf32>) -> tensor<1x8x7x6xf32> {

  // CHECK: %[[ARG_PERM:.*]] = "tf.Const"() <{value = dense<[0, 2, 3, 1]> : tensor<4xi64>}>
  // CHECK: %[[ARG_TRANSPOSE:[0-9]*]] = "tf.Transpose"(%arg0, %[[ARG_PERM]])

  %cst = "tfl.no_value"() {value = unit} : () -> none
  // CHECK: _arithmetic_count = 176160768 : i64
  %0 = "tfl.transpose_conv"(%arg0, %arg1, %arg2, %cst) {padding = "SAME", stride_h = 2 : i32, stride_w = 2 : i32, fused_activation_function = "NONE"} : (tensor<4xi32>, tensor<32x4x4x128xf32>, tensor<1x32x42x128xf32>, none) -> tensor<1x64x84x32xf32>
  func.return %0 : tensor<1x64x84x32xf32>
}

  %1 = "tfl.add"(%0, %cst) {fused_activation_function = "NONE"} : (tensor<1x64x84x32xf32>, tensor<32xf32>) -> tensor<1x64x84x32xf32>
  func.return %1 : tensor<1x64x84x32xf32>

  // CHECK-DAG: %[[SHAPE:.*]] = arith.constant dense<[1, 64, 84, 32]> : tensor<4xi32>