%fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%in, %mini, %maxi) {num_bits = 4, narrow_range = true} : (tensor<3x3x3x4xf32>, tensor<4xf32>, tensor<4xf32>) -> tensor<3x3x3x4xf32>
  %rst = "tf.Conv2D"(%arg, %fq) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", dilations = [1, 2, 3, 1], padding = "SAME", strides = [1, 4, 5, 1]} : (tensor<256x32x32x3xf32>, tensor<3x3x3x4xf32>) -> tensor<256x8x7x4xf32>
  func.return %rst : tensor<256x8x7x4xf32>

  %0 = "tf.Const"() {value = dense<0.111111112> : tensor<3x3x1x1xf32>} : () -> tensor<3x3x1x1xf32>
  %1 = "tf.Const"() {value = dense<1.000000e+00> : tensor<1x520x520x1xf32>} : () -> tensor<1x520x520x1xf32>
  %2 = "tf.DepthwiseConv2dNative"(%1, %0) {data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 1, 1]} : (tensor<1x520x520x1xf32>, tensor<3x3x1x1xf32>) -> tensor<1x520x520x1xf32>

// CHECK-LABEL: conv_valid_padding
func.func @conv_valid_padding(%arg0: tensor<1x4x5x1xf32>, %arg1: tensor<3x3x1x1xf32>) -> tensor<1x2x3x1xf32> {
  // CHECK: mhlo.convolution(%arg0, %arg1)

  %0 = "tf.Conv2D"(%arg0, %arg1) {data_format = "NHWC", dilations = [1, 1, 1, 1], padding = "VALID", strides = [1, 1, 1, 1]} : (tensor<1x4x5x1xf32>, tensor<3x3x1x1xf32>) -> tensor<1x2x3x1xf32>
  func.return %0 : tensor<1x2x3x1xf32>
}

// -----

// CHECK:           %[[VAL_12:.*]] = "tf.Reshape"(%[[VAL_10]], %[[VAL_11]]) : (tensor<3x5x4xf32>, tensor<4xi64>) -> tensor<3x5x1x4xf32>
// CHECK:           return %[[VAL_12]] : tensor<3x5x1x4xf32>
// CHECK:         }
func.func @convert_dot_general(%arg0: tensor<3x2x6x5x1xf32>, %arg1: tensor<3x2x4x6xf32>) -> tensor<3x5x1x4xf32> {
  %0 = "mhlo.dot_general"(%arg0, %arg1) {
    dot_dimension_numbers = #mhlo.dot<