// CHECK:  return
}

func.func @topk_4(%arg0: tensor<1x2x3x4xf32>) -> (tensor<1x2x3x2xf32>, tensor<1x2x3x2xi32>) {
  %0 = "tf.Const"() { value = dense<2> : tensor<i32> } : () -> tensor<i32>
  %1:2 = "tf.TopKV2"(%arg0, %0) : (tensor<1x2x3x4xf32>, tensor<i32>) -> (tensor<1x2x3x2xf32>, tensor<1x2x3x2xi32>)
  func.return %1#0, %1#1: tensor<1x2x3x2xf32>, tensor<1x2x3x2xi32>

// CHECK-LABEL: topk_4
// CHECK:  "tfl.topk_v2"(%arg0, %cst)

  %cst_1 = "tf.Const"() {value = dense<[0.1, 0.2]> : tensor<2xf32>} : () -> tensor<2xf32>
  %cst_2 = "tf.Const"() {value = dense<[0.3, 0.4]> : tensor<2xf32>} : () -> tensor<2xf32>

    return %1 : tensor<1x3x4x2xf32>
  }

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

    %cst_1 = "tf.Const"() {value = dense<3.000000e+00> : tensor<2x3x3x1xf32>} : () -> tensor<2x3x3x1xf32>
    %cst_2 = "tf.Const"() {value = dense<3.000000e+00> : tensor<2x3x3x2xf32>} : () -> tensor<2x3x3x2xf32>
    %0 = "tf.PartitionedCall"(%arg0, %cst_1) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_depthwise_conv2d_fn} : (tensor<1x3x4x3xf32>, tensor<2x3x3x1xf32>) -> tensor<*xf32>

  %0 = "quantfork.qcast"(%cst_1) : (tensor<2x3x3x2xf32>) -> tensor<2x3x3x2x!quant.uniform<i8<-127:127>:f32:3, {0.003937007874015748,0.003937007874015748}>>
  %1 = "quantfork.dcast"(%0) : (tensor<2x3x3x2x!quant.uniform<i8<-127:127>:f32:3, {0.003937007874015748,0.003937007874015748}>>) -> tensor<2x3x3x2xf32>

    %0 = "tf.PartitionedCall"(%arg0, %cst_1) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_depthwise_conv2d_fn} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<*xf32>

  return %4 : tensor<1x3x2x2xf32>
}

// CHECK: func @cast_bf16_conv_with_bias_to_fp32
// CHECK-DAG: %[[cst:.*]] = "tf.Const"() <{value = dense<1.000000e+00> : tensor<2x3x3x2xf32>}> : () -> tensor<2x3x3x2xf32>
// CHECK-DAG: %[[cst_0:.*]] = "tf.Const"() <{value = dense<1.000000e+00> : tensor<2xf32>}> : () -> tensor<2xf32>
// CHECK: %[[conv:.*]] = "tf.Conv2D"(%arg0, %[[cst]])

    %0 = "tf.PartitionedCall"(%arg0, %cst_1) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_depthwise_conv2d_fn} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<*xf32>

  %2 = "tfl.select"(%cst_false, %arg0, %arg1) : (tensor<1x2x3x4xi1>, tensor<1x2x3x4xf32>, tensor<1x2x3x4xf32>) -> tensor<1x2x3x4xf32>
  %3 = "tfl.select_v2"(%cst_false, %arg0, %arg1) : (tensor<1x2x3x4xi1>, tensor<1x2x3x4xf32>, tensor<1x2x3x4xf32>) -> tensor<1x2x3x4xf32>
  func.return %0, %1, %2, %3 : tensor<1x2x3x4xf32>, tensor<1x2x3x4xf32>, tensor<1x2x3x4xf32>, tensor<1x2x3x4xf32>

    %0 = "tf.Cast"(%arg0) {Truncate = false, device = ""} : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xf32>
    %1 = "tf.Conv2D"(%0, %cst) {data_format = "NHWC", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true}