return %1 : tensor<1x3x4x2xf32>
  }

  func.func private @composite_conv_fn(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> tensor<1x3x4x2xf32> attributes {_from_xla_call_module} {

      -> tensor<1x2x2x2xf32>
    %3 = "quantfork.stats"(%arg2) {layerStats = dense<[7.05456924, 7.11401462]> : tensor<2xf32>} : (tensor<2xf32>) -> tensor<2xf32>
    %4 = "quantfork.stats"(%2) {layerStats = dense<[-1.36523, 3.57373]> : tensor<2xf32>} : (tensor<1x2x2x2xf32>) -> tensor<1x2x2x2xf32>
    %5 = "chlo.broadcast_add"(%4, %3) : (tensor<1x2x2x2xf32>, tensor<2xf32>) -> tensor<1x2x2x2xf32>

    %2 = "tf.Relu"(%1) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>
    %3 = "tf.FakeQuantWithMinMaxArgs"(%2) {device = "", max = 4.000000e-01 : f32, min = -3.000000e-01 : f32, narrow_range = false, num_bits = 8 : i64} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>

// CHECK-DAG: %[[BROADCAST:.*]] = stablehlo.broadcast_in_dim %[[CONST_0]], dims = [3] : (tensor<2xf32>) -> tensor<1x3x2x2xf32>
// CHECK-DAG: %[[ADD:.*]] = stablehlo.add %[[CONV]], %[[BROADCAST]] : tensor<1x3x2x2xf32>
// CHECK-DAG: %[[RELU6:.*]] = stablehlo.clamp %[[CONST_2]], %[[ADD]], %[[CONST_1]] : (tensor<f32>, tensor<1x3x2x2xf32>, tensor<f32>) -> tensor<1x3x2x2xf32>

  %2 = "tf.Cast"(%1) {Truncate = false} : (tensor<1x3x2x2xbf16>) -> tensor<1x3x2x2xf32>
  %3 = "tf.IdentityN"(%2) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>
  return %3 : tensor<1x3x2x2xf32>
// CHECK: tf.Conv2D
// CHECK-SAME: loc("Model/conv2d")
}

func.func @conv2d_with_valid_loc(%arg0: tensor<1x3x4x3xf32>) -> (tensor<1x3x2x2xf32>) {

func.func @tpu_conv(%arg0: tensor<1x3x4x3xf32>) -> tensor<1x3x2x2xf32> {
  %0 = "tf.TPUOrdinalSelector"() {device = ""} : () -> tensor<?xi32>
  %1 = "tf.TPUPartitionedCall"(%arg0, %0) {autotuner_thresh = 0 : i64, device = "", f = @tpu_func_0_optim0} : (tensor<1x3x4x3xf32>, tensor<?xi32>) -> tensor<1x3x2x2xf32>
  %2 = "tf.IdentityN"(%1) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>
  func.return %2 : tensor<1x3x2x2xf32>
}

  %output_1, %min_2, %max_3, %histogram_4 = "tf.CustomAggregator"(%0) <{calibration_method = 5 : i32, id = "keeping_id", num_bins = 32 : i32, max_percentile = 0.000000e+00 : f32, min_percentile = 0.000000e+00 : f32}> : (tensor<1x2x2x2xf32>) -> (tensor<1x2x2x2xf32>, tensor<f32>, tensor<f32>, tensor<512xi64>)
  %1 = "tf.Identity"(%output_1) {device = ""} : (tensor<1x2x2x2xf32>) -> tensor<1x2x2x2xf32>

  } : (tensor<1x6x6x3xf32>, tensor<2x3x3x2x!tf_type.qint8>, tensor<f32>, tensor<i32>) -> tensor<1x4x1x2xf32>
  func.return %0 : tensor<1x4x1x2xf32>
}

// -----

// CHECK-LABEL: func @uniform_quantized_convolution_hybrid_same
func.func @uniform_quantized_convolution_hybrid_same(%input: tensor<1x2x2x3xf32>) -> tensor<1x2x1x2xf32> {

  // CHECK-DAG: %[[CST:.+]] = mhlo.constant dense<[1.000000e-01, 2.000000e-01]> : tensor<2xf32>
  // CHECK-DAG: %[[CST_BCAST:.+]] = "mhlo.broadcast_in_dim"(%[[CST]]) <{broadcast_dimensions = dense<3> : tensor<1xi64>}> : (tensor<2xf32>) -> tensor<1x1x3x2xf32>
  // CHECK-DAG: %[[NEW_FILTER:.+]] =  mhlo.multiply %[[CST_BCAST]], %[[FILTER]] : tensor<1x1x3x2xf32>

// CHECK: return %[[DEQUANTIZED]] : tensor<*xf32>
}

// -----

module {
  func.func @not_quantize_matmul_without_const(%arg0: tensor<1x2x2x2xf32>, %arg1: tensor<2x1024xf32>) -> (tensor<*xf32>) {
    %arg0_identity = "tf.Identity"(%arg0) {device = ""} : (tensor<1x2x2x2xf32>) -> tensor<1x2x2x2xf32>
    %arg1_identity = "tf.Identity"(%arg1) {device = ""} : (tensor<2x1024xf32>) -> tensor<2x1024xf32>