%output, %min, %max, %histogram = "tf.CustomAggregator"(%arg0) <{calibration_method = 5 : i32, id = "skipping_id", num_bins = 32 : i32, max_percentile = 0.000000e+00 : f32, min_percentile = 0.000000e+00 : f32}> : (tensor<1x3x4x3xf32>) -> (tensor<1x3x4x3xf32>, tensor<f32>, tensor<f32>, tensor<512xi64>)

    %1 = "tf.Conv2D"(%0, %cst) {data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x3x2x2xf32>
    %2 = "tf.Relu"(%1) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>

  %3 = "tf.BiasAdd"(%2, %cst_0) {data_format = "NHWC", device = ""} : (tensor<1x3x2x2xf32>, tensor<2xf32>) -> tensor<1x3x2x2xf32>
  %4 = "tf.Relu"(%3) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>
  %5 = "quantfork.qcast"(%4) : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2x!quant.uniform<i8:f32, 0.0027450981093387976:-19>>
  %6 = "quantfork.dcast"(%5) : (tensor<1x3x2x2x!quant.uniform<i8:f32, 0.0027450981093387976:-19>>) -> tensor<1x3x2x2xf32>

    %0 = "quantfork.stats"(%arg0) {layerStats = dense<[1.27501142, 149.824783]> : tensor<2xf32>} : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xf32>

func.func private @conv(%input: tensor<1x3x4x3xf32> {tf._user_specified_name = "input_tensor"}) -> tensor<*xf32> attributes {tf._construction_context = "kEagerRuntime", tf._input_shapes = [#tf_type.shape<1x3x4x3>]} {
  %weight = arith.constant dense_resource<__elided__> : tensor<2x3x3x2xf32>
  %bias = arith.constant dense<[7.11401462, 7.05456924]> : tensor<2xf32>

module {
  func.func @matmul(%arg0: tensor<1x2x2x3xf32>) -> (tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<2x1024xf32>} : () -> tensor<2x1024xf32>
    %1 = "tf.PartitionedCall"(%arg0, %cst_0) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_matmul_fn} : (tensor<1x2x2x3xf32>, tensor<2x1024xf32>) -> tensor<*xf32>
    func.return %1: tensor<*xf32>
  }

  %cst_0 = arith.constant dense<1.000000e+00> : tensor<f32>
  %cst_1 = arith.constant dense<0.000000e+00> : tensor<8xf32>
  %cst_2 = arith.constant dense<0.000000e+00> : tensor<8x3x3x3xf32>
  %0 = "tfl.sub"(%arg0, %cst_0) {fused_activation_function = "NONE"} : (tensor<3x15x14x3xf32>, tensor<f32>) -> tensor<3x15x14x3xf32>

func.func @fake_quant_conv(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> tensor<*xf32> {
  %cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<2xf32>} : () -> tensor<2xf32>
  %0 = "tf.FakeQuantWithMinMaxArgs"(%arg1) {device = "", max = 2.000000e+00 : f32, min = -1.000000e+00 : f32, narrow_range = false, num_bits = 8 : i64} : (tensor<2x3x3x2xf32>) -> tensor<*xf32>

// CHECK: return %[[RES0]] : tensor<256x32x32x16xf32>

{
  "version": 3,
  "operator_codes": [
    {
      "builtin_code": "CONV_2D"
    }
  ],
  "subgraphs": [
    {

  %cst2 = arith.constant dense<[1.0, 2.0]> : tensor<2xf32>
  %0 = "tf.Conv2D"(%arg0, %cst0) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", dilations = [1, 2, 3, 1], padding = "SAME", strides = [1, 4, 5, 1]} : (tensor<1x112x112x3xf32>, tensor<1x3x3x2xf32>) -> tensor<1x28x23x2xf32>