%biasadd = "tf.BiasAdd"(%conv, %dq_bias) {data_format = "NHWC", device = ""} : (tensor<*xf32>, tensor<2xf32>) -> tensor<*xf32>
  %res = "tf.Relu6"(%biasadd) : (tensor<*xf32>) -> tensor<*xf32>
  %q_res = "quantfork.qcast"(%res) : (tensor<*xf32>) -> tensor<*x!quant.uniform<i8:f32, 0.023529411764705882:-128>>

    %cst_1 = "tf.Const"() {value = dense<3.000000e+00> : tensor<2x3x3x512xf32>} : () -> tensor<2x3x3x512xf32>
    %0 = "tf.PartitionedCall"(%arg0, %cst_1) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_depthwise_conv2d_fn} : (tensor<1x3x4x512xf32>, tensor<2x3x3x512xf32>) -> tensor<*xf32>

    %outputs_2, %control_2 = tf_executor.island wraps "tf.Identity"(%outputs_1) {T = "tfdtype$DT_FLOAT", _tpu_input_identity = true, _tpu_replicate = "cluster", device = "", name = "replicated_input_1"} : (tensor<0xf32>) -> tensor<0xf32>

    %cst_0 = "tf.Const"() {value = dense<1.000000e+00> : tensor<2xf32>} : () -> tensor<2xf32>
    %var_0 = "tf.VarHandleOp"() {shared_name = "var_0"} : () -> tensor<!tf_type.resource<tensor<2xf32>>>
    "tf.AssignVariableOp"(%var_0, %cst_0) : (tensor<!tf_type.resource<tensor<2xf32>>>, tensor<2xf32>) -> ()
    %cst_1 = "tf.Const"() {value = dense<2.000000e+00> : tensor<3xf32>} : () -> tensor<3xf32>

    func.func @main(%arg0: tensor<1x2xf32>) -> tensor<1x3xf32> {
      %cst = "tf.Const"() {value = dense<3.00000000e-1> : tensor<2x3xf32>} : () -> tensor<2x3xf32>
      %0 = "quantfork.stats"(%arg0) {layerStats = dense<[6.00000000e-6, 9.00000000e-1]> : tensor<2xf32>} : (tensor<1x2xf32>) -> tensor<1x2xf32>

^bb0(%arg0: tensor<40x37xf32>, %arg1: tensor<40x37xf32>):
  %cst = arith.constant dense<1.0> : tensor<40xf32>
  %0:2 = "tfl.fully_connected"(%arg0, %arg1, %cst) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<40x37xf32>, tensor<40x37xf32>, tensor<40xf32>) -> (tensor<40x40xf32>, tensor<40x40xf32>)
  func.return %0 : tensor<40x40xf32>

// CHECK-LABEL: @main
// CHECK: %[[RES0:.*]] = "quantfork.stats"(%arg0) <{layerStats = dense<[0.000000e+00, 2.550000e+02]> : tensor<2xf32>}> : (tensor<1x512x672x8xf32>) -> tensor<1x512x672x8xf32>

  %0 = "quantfork.stats"(%arg0) {layerStats = dense<[0.000000e+00, 2.550000e+02]> : tensor<2xf32>} : (tensor<1x512x672x8xf32>) -> tensor<1x512x672x8xf32>
  func.return %0 : tensor<1x512x672x8xf32>

      %a, %control1 = tf_executor.island wraps "tf.A"() {_class = ["loc:@class"]} : () -> (tensor<2xf32>)
      %s:2, %control2 = tf_executor.island wraps "tf.Split"(%c, %a) {num_split = 2 : i32} : (tensor<i32>, tensor<2xf32>) -> (tensor<1xf32>, tensor<1xf32>)
      tf_executor.fetch
    }
    func.return
  }

  %2 = "tf.Relu6"(%1) {device = ""} : (tensor<*xf32>) -> tensor<*xf32>
  func.return %2 : tensor<*xf32>
}

// CHECK-LABEL: func @conv_with_non_constant_filter
// CHECK-DAG: %[[CONST_0:.*]] = "tf.Const"() <{value = dense<0.000000e+00> : tensor<2xf32>}> : () -> tensor<2xf32>
// CHECK: %[[PARTITIONEDCALL_0:.*]] = "tf.PartitionedCall"(%arg0, %arg1, %[[CONST_0]])
// CHECK-SAME: f = @composite_conv2d_with_bias_and_relu6_fn_1

    %cst_1 = "tf.Const"() {value = dense<3.000000e+00> : tensor<2x3x3x512xf32>} : () -> tensor<2x3x3x512xf32>
    %0 = "tf.PartitionedCall"(%arg0, %cst_1) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_depthwise_conv2d_fn} : (tensor<1x3x4x512xf32>, tensor<2x3x3x512xf32>) -> tensor<*xf32>