func.func @batchNormInference_2D_inner_features(
    %x: tensor<4x256xf32>, %scale: tensor<256xf32>, %offset: tensor<256xf32>,
    %mean: tensor<256xf32>, %variance: tensor<256xf32>)
    -> (tensor<4x256xf32>) {
  // CHECK-DAG: %[[EPS_BCAST:.+]] = mhlo.constant dense<1.001000e-05> : tensor<256xf32>
  // CHECK-DAG: %[[VARIANCE_EPS:.+]] = mhlo.add %[[VARIANCE]], %[[EPS_BCAST]] : tensor<256xf32>

       {
         data_format = "NCHW",
         epsilon = 1.001000e-05 : f32
       } : (tensor<?x256x56x56xf32>, tensor<256xf32>, tensor<256xf32>, tensor<256xf32>, tensor<256xf32>)
        -> (tensor<?x256x56x56xf32>, tensor<256xf32>, tensor<256xf32>, tensor<256xf32>, tensor<256xf32>, tensor<*xf32>)

  // CHECK: %[[BATCH_NORM1:[_a-z0-9]*]], {{.*}} = "tf.FusedBatchNormV3"
  // CHECK-SAME: %[[CONV1]]

func.func @unfuse_batch_norm(
    %x: tensor<4x256xf32>, %scale: tensor<256xf32>, %offset: tensor<256xf32>,
    %mean: tensor<256xf32>, %variance: tensor<256xf32>)
    -> (tensor<4x256xf32>) {
  // CHECK-DAG: %[[EPS_BCAST:.+]] = mhlo.constant dense<1.001000e-05> : tensor<256xf32>
  // CHECK-DAG: %[[VARIANCE_EPS:.+]] = mhlo.add %[[VARIANCE]], %[[EPS_BCAST]] : tensor<256xf32>

  %1 = "tf.BiasAdd"(%0, %cst_0) {data_format = "NHWC"} : (tensor<?x?x?x256xf32>, tensor<256xf32>) -> tensor<?x?x?x256xf32>
  %2 = "tf.Mul"(%1, %cst_1) : (tensor<?x?x?x256xf32>, tensor<256xf32>) -> tensor<?x?x?x256xf32>
  func.return %2 : tensor<?x?x?x256xf32>
}
// CHECK: func @conv2d_with_large_weight_and_mul

  func.return %y : tensor<1x64x28x28xf32>
}

// CHECK-LABEL: bias_add_nchw
func.func @bias_add_nchw(%arg0: tensor<1x256x150x150xf32>, %arg1: tensor<256xf32>) -> tensor<1x256x150x150xf32> {
  // CHECK: (%[[ARG0:.*]]: tensor<1x256x150x150xf32>, %[[ARG1:.*]]: tensor<256xf32>)
  // CHECK: %[[CST:.*]] = "tf.Const"() <{value = dense<[0, 2, 3, 1]> : tensor<4xi64>}>
  // CHECK: %[[R0:.*]] = "tf.Transpose"(%[[ARG0]], %[[CST]])

// CHECK:           %[[VAL_5:.*]] = "tf.Tanh"(%[[VAL_4]]) : (tensor<2xf32>) -> tensor<2xf32>
// CHECK:           %[[VAL_6:.*]] = "tf.Mul"(%[[VAL_5]], %[[VAL_3]]) : (tensor<2xf32>, tensor<2xf32>) -> tensor<2xf32>
// CHECK:           %[[VAL_7:.*]] = "tf.AddV2"(%[[VAL_6]], %[[VAL_3]]) : (tensor<2xf32>, tensor<2xf32>) -> tensor<2xf32>
// CHECK:           return %[[VAL_7]] : tensor<2xf32>

  %1 = "tfl.sum"(%0, %cst) {keep_dims = false} : (tensor<2x2xf32>, tensor<1xi32>) -> tensor<f32>
  %2 = "tfl.sqrt"(%1) : (tensor<f32>) -> tensor<f32>
  %3 = "tfl.div"(%arg0, %2) {fused_activation_function = "NONE"} : (tensor<2x2xf32>, tensor<f32>) -> tensor<2x2xf32>
  func.return %3: tensor<2x2xf32>

  %3 = "tfl.dequantize"(%2) : (tensor<2x3x!quant.uniform<i16:f32, 1.0>>) -> (tensor<2x3xf32>)
  %4 = "tfl.concatenation"(%1, %3) {axis = -1 : i32, fused_activation_function = "NONE"} : (tensor<2x1xf32>, tensor<2x3xf32>) -> tensor<2x4xf32>
  %5 = "tfl.add"(%4, %arg2) {fused_activation_function = "NONE"} : (tensor<2x4xf32>, tensor<2x4xf32>) -> tensor<2x4xf32>
  func.return %5: tensor<2x4xf32>

// CHECK-NEXT:      %17 = mhlo.dynamic_reshape %arg1, %16 : (tensor<4x?x256xf32>, tensor<3xi32>) -> tensor<4x?x256xf32>
// CHECK-NEXT:      %18 = "tfl.batch_matmul"(%8, %17) <{adj_x = false, adj_y = false, asymmetric_quantize_inputs = false}> : (tensor<4x4x?xf32>, tensor<4x?x256xf32>) -> tensor<4x4x256xf32>
// CHECK-NEXT:      %19 = mhlo.reshape %18 : (tensor<4x4x256xf32>) -> tensor<4x4x256xf32>

  // CHECK-DAG: %[[EXP:.*]] = "tf.Exp"(%[[SHIFTED]]) : (tensor<2x3xf32>) -> tensor<2x3xf32>
  // CHECK-DAG: %[[SUM:.*]] = "tf.Sum"(%[[EXP]], %[[AXIS]]) <{keep_dims = true}> : (tensor<2x3xf32>, tensor<1xi64>) -> tensor<2x1xf32>
  // CHECK-DAG: %[[RESULT:.*]] = "tf.Div"(%[[EXP]], %[[SUM]]) : (tensor<2x3xf32>, tensor<2x1xf32>) -> tensor<2x3xf32>
  // CHECK: return %[[RESULT]]