%custom_2 = "tfl.custom"(%arg0, %dq_w) {custom_code = "CustomTestOp", custom_option = #tfl<const_bytes : "0x">} : (tensor<1x1x1x1xf32>, tensor<1024x1x1x1xf32>) -> tensor<*xf32>
  %custom_3 = "tfl.custom"(%arg0, %dq_w) {custom_code = "CustomTestOp", custom_option = #tfl<const_bytes : "0x">} : (tensor<1x1x1x1xf32>, tensor<1024x1x1x1xf32>) -> tensor<*xf32>

    }> {
      _collective_manager_ids = [], device = ""
    } : (tensor<1x2x2x3xf32>) -> tensor<1x2x2x3xf32>
    func.return %0: tensor<1x2x2x3xf32>
  }

  func.func private @some_func(%arg0: tensor<1x2x2x3xf32>) -> tensor<1x2x2x3xf32> {
    return %arg0 : tensor<1x2x2x3xf32>
  }
}

// CHECK: module
// CHECK-NOT: tf.PartitionedCall
// CHECK-NOT: some_func

  %0 = "tf.Conv2D"(%arg0, %cst) {data_format = "NHWC", dilations = [1, 1, 2, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true} : (tensor<1x3x4x3xf32>, tensor<2x3x3x3xf32>) -> tensor<1x3x2x3xf32>
  %1 = "tf.AddV2"(%0, %cst_0) : (tensor<1x3x2x3xf32>, tensor<1x3x2x3xf32>) -> tensor<1x3x2x3xf32>
  func.return %1 : tensor<1x3x2x3xf32>
}

  %2 = "mhlo.dot_general"(%arg2, %arg3) {
    dot_dimension_numbers = #mhlo.dot<
      lhs_contracting_dimensions = [2],
      rhs_contracting_dimensions = [0]
  >} : (tensor<3x1x512xf32>, tensor<512x13xf32>) -> tensor<3x1x13xf32>
  %r = "mhlo.concatenate"(%0, %1, %2) <{dimension = 0 : i64}> : (tensor<1x1x13xf32>, tensor<2x1x13xf32>, tensor<3x1x13xf32>) -> tensor<6x1x13xf32>

}

// CHECK-LABEL: preparePrelu
func.func @preparePrelu(%arg0: tensor<1x10x10x3xf32>) -> tensor<1x10x10x3xf32> {
  %cst = "arith.constant"() {value = dense<[[[1.66394591, 3.61694336, 2.0382936]]]> : tensor<1x1x3xf32>} : () -> tensor<1x1x3xf32>
  %prelu = "tfl.prelu"(%arg0, %cst) : (tensor<1x10x10x3xf32>, tensor<1x1x3xf32>) -> tensor<1x10x10x3xf32>
  func.return %prelu : tensor<1x10x10x3xf32>

func.func @depth_to_space(%arg0: tensor<1x1x1x4xf32>) -> tensor<1x2x2x1xf32> {
  %0 = "tf.DepthToSpace"(%arg0) {block_size = 2: i64,  data_format = "NHWC"}: (tensor<1x1x1x4xf32>) -> tensor<1x2x2x1xf32>
  func.return %0 : tensor<1x2x2x1xf32>
// CHECK: %[[CUSTOM_0:.*]] = "tfl.custom"(%arg0) <{custom_code = "FlexDepthToSpace", custom_option = #tfl<const_bytes : "{{.*}}">}> : (tensor<1x1x1x4xf32>) -> tensor<1x2x2x1xf32>

  func.return %1 : tensor<1x1x1x16xf32>

// CHECK: %[[avgp:.*]] = "tfl.average_pool_2d"(%arg0)
// CHECK: %[[dq:.*]] = "tfl.dequantize"(%[[avgp]]) : (tensor<1x1x1x16x!quant.uniform<u8:f32, 7.812500e-03:128>>) -> tensor<1x1x1x16xf32>
// CHECK: return %[[dq]] : tensor<1x1x1x16xf32>
}

// CHECK-LABEL: QuantizeReshape2D

// CustomOpNotWeightOnly-LABEL: QuantizeCustomOp
func.func @QuantizeCustomOp(%arg0: tensor<1x1x1x1xf32>) -> tensor<*xf32> attributes {tf.entry_function = {inputs = "input", outputs = "custom_op"}} {
  %0 = "quantfork.stats"(%arg0) {layerStats = dense<[0.000000e+00, 2.550000e+02]> : tensor<2xf32>} : (tensor<1x1x1x1xf32>) -> tensor<1x1x1x1xf32>
  %w = arith.constant dense<127.0> : tensor<1024x1x1x1xf32>

  return %2 : tensor<?x2x2x1xi8>
}

// -----

// CHECK-LABEL: func @convolution_add_add_f32
func.func @convolution_add_add_f32(
    %lhs: tensor<?x3x2x1xf32>, %rhs: tensor<2x1x1x1xf32>,
    %zp_offset: tensor<?x2x2x1xf32>, %bias: tensor<1xf32>
  ) -> tensor<?x2x2x1xf32> {
  // CHECK-DAG: %[[conv:.*]] = mhlo.convolution
  // CHECK-DAG: %[[combined:.*]] = chlo.broadcast_add %[[conv:.*]], %[[zp_offset:.*]]

  %1 = "tfl.pseudo_const"() {value = dense<1> : tensor<16x1x1x8xi8>} : () -> tensor<16xf32>
  %2 = "tfl.conv_2d"(%arg0, %0, %1) {dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "NONE", padding = "SAME", stride_h = 1 : i32, stride_w = 1 : i32} : (tensor<1x32x32x8xf32>, tensor<16x1x1x8xf32>, tensor<16xf32>) -> tensor<1x32x32x16xf32>