func.func @add_with_activation_transpose(%arg0: tensor<1x3x3x4xf32>) -> tensor<1x4x3x3xf32> {
  %0 = stablehlo.constant dense<2.000000e+00> : tensor<1x4x3x3xf32>
  %1 = stablehlo.transpose %arg0, dims = [0, 3, 1, 2] : (tensor<1x3x3x4xf32>) -> tensor<1x4x3x3xf32>
  %2 = stablehlo.add %1, %0 : tensor<1x4x3x3xf32>
  return %2 : tensor<1x4x3x3xf32>
}
// CHECK-SAME: (%[[ARG_0:.+]]: tensor<1x3x3x4xf32>) -> tensor<1x4x3x3xf32>

  %custom_2 = "tfl.custom"(%arg0, %dq_w) {custom_code = "CustomTestOp", custom_option = #tfl<const_bytes : "0x">} : (tensor<1x1x1x1xf32>, tensor<1024x1x1x1xf32>) -> tensor<*xf32>
  %custom_3 = "tfl.custom"(%arg0, %dq_w) {custom_code = "CustomTestOp", custom_option = #tfl<const_bytes : "0x">} : (tensor<1x1x1x1xf32>, tensor<1024x1x1x1xf32>) -> tensor<*xf32>

    return %1 : tensor<1x3x4x2xf32>
  }

  func.func private @composite_conv_fn(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> tensor<1x3x4x2xf32> attributes {_from_xla_call_module} {

    %2 = stablehlo.convolution(%1, %0)
      dim_numbers = [b, 0, 1, f]x[0, 1, i, o]->[b, 0, 1, f],
      window = {
        stride = [1, 1], pad = [[0, 0], [1, 1]],
        lhs_dilate = [1, 1],
        rhs_dilate = [1, 1]
      }
      {
        batch_group_count = 1 : i64,
        feature_group_count = 1 : i64
      } : (tensor<1x3x2x3xf32>, tensor<2x3x3x2xf32>)

  %4 = "quantfork.stats"(%3) {
    layerStats = dense<[0.0, 4.0]> : tensor<2xf32>
  } : (tensor<?x1x1x3xf32>) -> tensor<?x1x1x3xf32>
  %5 = "tfl.sqrt"(%4) : (tensor<?x1x1x3xf32>) -> tensor<?x1x1x3xf32>
  %6 = "quantfork.stats"(%5) {
    layerStats = dense<[0.0, 2.0]> : tensor<2xf32>
  } : (tensor<?x1x1x3xf32>) -> tensor<?x1x1x3xf32>

  %1 = mhlo.multiply %0, %cst1 : tensor<1x1x2x4xf32>
  // CHECK:      return %[[RES]] : tensor<1x1x2x4xf32>
  func.return %1 : tensor<1x1x2x4xf32>
}

// CHECK-LABEL: @foldBroadcastInDimBeforeMulOp_bcast_dim_2D_float
func.func @foldBroadcastInDimBeforeMulOp_bcast_dim_2D_float() -> (tensor<1x2x2x3xf32>) {

  %0 = "tfl.pseudo_const"() {value = dense<1.000000e+00> : tensor<16x3x3x3xf32>} : () -> tensor<16x3x3x3xf32>
  %1 = "tfl.pseudo_const"() {value = dense<1.000000e+00> : tensor<16xf32>} : () -> tensor<16xf32>
  %2 = func.call @fold_all_test(%arg0, %0, %1) : (tensor<256x32x32x3xf32>, tensor<16x3x3x3xf32>, tensor<16xf32>) -> tensor<256x30x30x16xf32>
  func.return %2 : tensor<256x30x30x16xf32>
}

  %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>

  %2 = "tfl.select"(%cst_false, %arg0, %arg1) : (tensor<1x2x3x4xi1>, tensor<1x2x3x4xf32>, tensor<1x2x3x4xf32>) -> tensor<1x2x3x4xf32>
  %3 = "tfl.select_v2"(%cst_false, %arg0, %arg1) : (tensor<1x2x3x4xi1>, tensor<1x2x3x4xf32>, tensor<1x2x3x4xf32>) -> tensor<1x2x3x4xf32>
  func.return %0, %1, %2, %3 : tensor<1x2x3x4xf32>, tensor<1x2x3x4xf32>, tensor<1x2x3x4xf32>, tensor<1x2x3x4xf32>

  %0 = stablehlo.constant() {value = dense<3.000000e+00> : tensor<8x8x3x3xf32>} : () -> tensor<8x8x3x3xf32>
  %2 = stablehlo.convolution(%arg0, %0) dim_numbers = [b, f, 0, 1]x[o, i, 0, 1]->[b, f, 0, 1], window = {pad = [[1, 1], [1, 1]]} {batch_group_count = 1 : i64, feature_group_count = 1 : i64} : (tensor<1x8x4x4xf32>, tensor<8x8x3x3xf32>) -> tensor<1x8x4x4xf32>
  return %2 : tensor<1x8x4x4xf32>
}