func.func @depthwise_conv_2d_bf16(%arg0 : tensor<256x32x32x3xbf16>, %arg1 : tensor<3x3x3x4xf32>, %arg2 : tensor<256x3x32x32xf32>) -> tensor<256x30x30x12xbf16> {
  %0 = "tf.DepthwiseConv2dNative"(%arg0, %arg1) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", dilations = [1, 2, 3, 1], padding = "SAME", strides = [1, 4, 5, 1]} : (tensor<256x32x32x3xbf16>, tensor<3x3x3x4xf32>) -> tensor<256x30x30x12xbf16>
  func.return %0 : tensor<256x30x30x12xbf16>

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

  %q_weight = "quantfork.qcast"(%weight) : (tensor<2x3x3x2xf32>) -> tensor<2x3x3x2x!quant.uniform<i8:f32, 0.074855112561992565:-1>>
  %dq_weight = "quantfork.dcast"(%q_weight) : (tensor<2x3x3x2x!quant.uniform<i8:f32, 0.074855112561992565:-1>>) -> tensor<2x3x3x2xf32>
  %q_bias = "quantfork.qcast"(%bias) : (tensor<2xf32>) -> tensor<2x!quant.uniform<i32:f32, 0.044022349891595126>>

  %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 private @quantize_conv_fn(%arg0: tensor<1x3x4x3xf32>) -> tensor<1x3x4x2xf32> attributes {tf._original_func_name = "main_0"} {
    %cst = "tf.Const"() {value = dense<3.00000000e-1> : tensor<2x3x3x2xf32>} : () -> tensor<2x3x3x2xf32>
    %0 = "quantfork.stats"(%arg0) {layerStats = dense<[6.00000000e-6, 9.00000000e-1]> : tensor<2xf32>} : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xf32>
    %1 = "tf.XlaCallModule"(%0, %cst) {

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

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

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>

    %cst_0 = "tf.Const"() {device = "", value = dense<[[[[0.211401448, 0.205456927], [0.418355644, -0.314615548]], [[0.493921608, -0.101286061], [-0.16083248, -0.0546654463]], [[-0.157245964, 0.419805884], [-0.0499645844, 3.726310e-01]]],...

  %w = arith.constant dense<2.0> : tensor<3x3x3x3xf32>
  %q = "tfl.quantize"(%w) {qtype = tensor<3x3x3x3x!quant.uniform<i8<-127:127>:f32:0,{1.0,2.0,3.0}>>} : (tensor<3x3x3x3xf32>) -> tensor<3x3x3x3x!quant.uniform<i8<-127:127>:f32:0,{1.0,2.0,3.0}>>
  %dq = "tfl.dequantize"(%q) : (tensor<3x3x3x3x!quant.uniform<i8<-127:127>:f32:0,{1.0,2.0,3.0}>>) -> tensor<3x3x3x3xf32>

  %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>)...