%1 = "tf.BiasAdd"(%0, %arg2) {data_format = "NHWC", device = ""} : (tensor<1x3x4x2xf32>, tensor<2xf32>) -> tensor<1x3x4x2xf32>
    %2 = "tf.Relu6"(%1) {device = ""} : (tensor<1x3x4x2xf32>) -> tensor<1x3x4x2xf32>
    func.return %2 : tensor<1x3x4x2xf32>
  }

// CHECK-LABEL: func @conv_with_per_channel_and_tensor_weight

  %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 = stablehlo.constant() {value = dense<3> : tensor<3x3x4x2xi8>} : () -> tensor<3x3x4x2x!quant.uniform<i8:f32, 3.000000e-01:-5>>
  %1 = stablehlo.convolution(%arg0, %0) dim_numbers = [b, 0, 1, f]x[0, 1, i, o]->[b, 0, 1, f], window = {pad = [[1, 1], [1, 1]]} {batch_group_count = 1 : i64, feature_group_count = 1 : i64} : (tensor<1x3x3x4xf32>, tensor<3x3x4x2x!quant.uniform<i8:f32, 3.000000e-01:-5>>) -> tensor<1x3x3x2xf32>

}

// -----

// CHECK-LABEL: lift_float_conv
func.func @lift_float_conv(%arg0: tensor<1x3x4x3xf32>) -> (tensor<*xf32>, tensor<*xf32>) {
  %cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<2xf32>} : () -> tensor<2xf32>
  %cst_1 = "tf.Const"() {value = dense<3.000000e+00> : tensor<2x3x3x2xf32>} : () -> tensor<2x3x3x2xf32>
  %0 = "tf.Conv2D"(%arg0, %cst_1) {

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

  %output, %min, %max, %histogram = "tf.CustomAggregator"(%arg0) <{calibration_method = 5 : i32, id = "0", 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>)

        version = 5 : i64
      } : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>, tensor<2xf32>) -> tensor<1x3x4x2xf32>
    %2 = "quantfork.stats"(%1) {layerStats = dense<[5.00000000e-6, 7.00000000e-1]> : tensor<2xf32>} : (tensor<1x3x4x2xf32>) -> tensor<1x3x4x2xf32>
    return %2 : tensor<1x3x4x2xf32>
  }
// CHECK: func.func private @quantize_conv_with_bias_1d_fn(%[[ARG_0:.+]]: tensor<1x3x4x3xf32>) -> tensor<1x3x4x2xf32> attributes {tf._original_func_name = "main_0"}

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>
  }

// CHECK:           %[[VAL_6:.*]] = arith.constant dense<[0, 3, 1, 2]> : tensor<4xi32>
// CHECK:           %[[VAL_7:.*]] = "tfl.transpose"(%[[VAL_5]], %[[VAL_6]]) : (tensor<1x4x4x3xf32>, tensor<4xi32>) -> tensor<1x3x4x4xf32>
// CHECK:           %[[VAL_8:.*]] = "tf.Identity"(%[[VAL_7]]) {device = ""} : (tensor<1x3x4x4xf32>) -> tensor<*xf32>