%4 = "quantfork.stats"(%3) {layerStats = dense<[0.000000e+00, 0.18486841]> : tensor<2xf32>} : (tensor<*xf32>) -> tensor<*xf32>
    %5 = "tf.Identity"(%4) {device = ""} : (tensor<*xf32>) -> tensor<*xf32>
    %6 = "quantfork.stats"(%5) {layerStats = dense<[0.000000e+00, 0.18486841]> : tensor<2xf32>} : (tensor<*xf32>) -> tensor<*xf32>

    tf_executor.graph {
      %out, %ctl = tf_executor.island wraps "tf.VarHandleOp"() {shared_name = "var_0"} : () -> tensor<!tf_type.resource<tensor<2xf32>>>
      %out_0, %ctl_0 = tf_executor.island wraps "tf.ReadVariableOp"(%out) : (tensor<!tf_type.resource<tensor<2xf32>>>) -> tensor<2xf32>
      %out_1, %ctl_1 = tf_executor.island wraps "tf.Const"() {value = dense<"var_0"> : tensor<1x!tf_type.string>} : () -> tensor<1x!tf_type.string>

    %3 = "tf.XlaCallModule"(%2, %1, %0) <{Sout = [#tf_type.shape<?x2>], module = "", version = 9 : i64}> {_entry_function = @composite_dot_general_with_bias_and_relu6_dynamic_fn_1, _original_entry_function = "composite_dot_general_with_bias_and_relu6_dynamic_fn_1", _tfl_quant_trait = "fully_quantizable"} : (tensor<?x2xf32>, tensor<2x2xf32>, tensor<2xf32>) -> tensor<?x2xf32>

  },  {
    ^bb0(%arg2: tensor<*xi32>, %arg3: tensor<*xf32>, %arg4: tensor<i32>):
      %1 = "tfl.sub"(%arg2, %arg4) {fused_activation_function = "NONE"} :
        (tensor<*xi32>, tensor<i32>) -> tensor<*xi32>
      %2 = tfl.add %arg3, %arg3 {fused_activation_function = "NONE"} : tensor<*xf32>
      "tfl.yield"(%1, %2, %arg4) : (tensor<*xi32>, tensor<*xf32>, tensor<i32>) -> ()
  }) : (tensor<i32>, tensor<1xf32>, tensor<i32>) -> (tensor<i32>, tensor<1xf32>, tensor<i32>)

    %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>) -> ()
    %add_0 = "tf.Identity"(%cst_0) : (tensor<2xf32>) -> tensor<2xf32>

    %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>

    %4 = "quantfork.stats"(%2) {layerStats = dense<[-2795.36523, 4609.57373]> : tensor<2xf32>} : (tensor<*xf32>) -> tensor<*xf32>
    %5 = "tf.BiasAdd"(%4, %3) {data_format = "NHWC", device = ""} : (tensor<*xf32>, tensor<2xf32>) -> tensor<*xf32>
    %6 = "quantfork.stats"(%5) {layerStats = dense<[-2788.31055, 4616.62842]> : tensor<2xf32>} : (tensor<*xf32>) -> tensor<*xf32>

    %outputs_6:2, %control_7 = tf_executor.island wraps "tf.While"(%arg0, %arg1) {body = @body, cond = @cond, is_stateless = false, shape_invariant} : (tensor<i32>, tensor<5xf32>) -> (tensor<i32>, tensor<5xf32>) loc("WhileWithOutputShapes")
    tf_executor.fetch %outputs_2#1, %outputs_4#1, %outputs_6#1 : tensor<5xf32>, tensor<5xf32>, tensor<5xf32>
  }

  } : (tensor<1x!tf_type.qint8>, tensor<f32>, tensor<i32>) -> tensor<1xf32>

  // CHECK: return %[[y]] : tensor<1xf32>
  func.return %0 : tensor<1xf32>
}

// -----

// CHECK-LABEL: func @uniform_quantize_dequantize
func.func @uniform_quantize_dequantize(%arg0: tensor<1xf32>) -> tensor<1xf32>
{
  %scales = "tf.Const"() { value = dense<1.0> : tensor<f32> } : () -> tensor<f32>

    %4 = "quantfork.stats"(%2) {layerStats = dense<[-2795.36523, 4609.57373]> : tensor<2xf32>} : (tensor<*xf32>) -> tensor<*xf32>
    %5 = "tf.BiasAdd"(%4, %3) {data_format = "NHWC", device = ""} : (tensor<*xf32>, tensor<2xf32>) -> tensor<*xf32>
    %6 = "quantfork.stats"(%5) {layerStats = dense<[-2788.31055, 4616.62842]> : tensor<2xf32>} : (tensor<*xf32>) -> tensor<*xf32>