: (tensor<*x!tf_type.resource>) -> (tensor<3x3x1x32xf32>, tensor<3x3x1x32xf32>)
  "tf_device.launch"() ({
    "tf.TPUCompileSucceededAssert"(%compile#0) : (tensor<!tf_type.string>) -> ()
    tf_device.return
  }) {device = "/device:CPU:0"} : () -> ()
  %execute0 = "tf_device.launch"() ({
    %3 = "tf.TPUExecute"(%2#0, %2#1, %compile#1)
      : (tensor<3x3x1x32xf32>, tensor<3x3x1x32xf32>, tensor<2x!tf_type.string>) -> tensor<i32>

    return %5 : tensor<1x3x1x1xf32>
  }
  func.func private @composite_gather_fn_1(%arg0: tensor<1x3x1x1xf32>, %arg1: tensor<1xi32>, %arg2: tensor<i32>) -> tensor<1x3x1x1xf32> attributes {tf_quant.composite_function} {

  %fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%in, %mini, %maxi) {num_bits = 5, narrow_range = false} : (tensor<3x3x3x16xf32>, tensor<16xf32>, tensor<16xf32>) -> tensor<3x3x3x16xf32>
  %rst = "tf.Conv2D"(%arg, %fq) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", dilations = [1, 2, 3, 1], padding = "SAME", strides = [1, 4, 5, 1]} : (tensor<256x32x32x3xf32>, tensor<3x3x3x16xf32>) -> tensor<256x8x7x16xf32>
  func.return %rst : tensor<256x8x7x16xf32>

  %fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%in, %mini, %maxi) {num_bits = 3, narrow_range = false} : (tensor<3x3x3x16xf32>, tensor<16xf32>, tensor<16xf32>) -> tensor<3x3x3x16xf32>
  %rst = "tf.Conv2D"(%arg, %fq) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", dilations = [1, 2, 3, 1], padding = "SAME", strides = [1, 4, 5, 1]} : (tensor<256x32x32x3xf32>, tensor<3x3x3x16xf32>) -> tensor<256x8x7x16xf32>
  func.return %rst : tensor<256x8x7x16xf32>

  // CHECK-SAME: %[[ARG1:.*]]: tensor<*x!tf_type.resource<tensor<f32>>> {tf.device = "/device:TPU:1"},
  // CHECK-SAME: %[[ARG2:.*]]: tensor<*x!tf_type.resource<tensor<3x3x1x32xf32>>> {tf.device = "/device:TPU:0"},
  // CHECK-SAME: %[[ARG3:.*]]: tensor<*x!tf_type.resource<tensor<3x3x1x32xf32>>> {tf.device = "/device:TPU:1"})
  func.func @main(%arg0: !tf_res_f32 {tf.device = "/device:TPU:0"},
             %arg1: !tf_res_f32 {tf.device = "/device:TPU:1"},

  %9 = "tfl.quantize"(%8) {qtype = tensor<1x3x1x1x!quant.uniform<i8:f32, 1.0:2>>, volatile} : (tensor<1x3x1x1xf32>) -> tensor<1x3x1x1x!quant.uniform<i8:f32, 1.0:2>>
  %10 = "tfl.dequantize"(%9) : (tensor<1x3x1x1x!quant.uniform<i8:f32, 1.0:2>>) -> tensor<1x3x1x1xf32>

// CHECK-LABEL: conv2d_backprop_input_with_add
func.func @conv2d_backprop_input_with_add(%arg0: tensor<4xi32>, %arg1: tensor<3x3x1x32xf32>, %arg2: tensor<15x14x14x32xf32>) -> tensor<15x28x28x1xf32> {
  %0 = "tf.Conv2DBackpropInput"(%arg0, %arg1, %arg2) {strides = [1, 2, 2, 1], padding="SAME", dilations=[1, 1, 1, 1]}: (tensor<4xi32>, tensor<3x3x1x32xf32>, tensor<15x14x14x32xf32>) -> tensor<15x28x28x1xf32>

  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

    %3 = "tf.Identity"(%2) {device = ""} : (tensor<1x3x1x1xf32>) -> tensor<1x3x1x1xf32>
    return %3 : tensor<1x3x1x1xf32>
  }

// CHECK-LABEL: func @multiple_quantizable_ops_in_graph

  %w = arith.constant dense<[[[[0.0]]], [[[127.0]]], [[[-127.0]]]]> : tensor<3x1x1x1xf32>
  %b = arith.constant dense<0.0> : tensor<3xf32>
  %conv = "tfl.conv_2d"(%arg0, %w, %b) {dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "RELU", padding = "SAME", stride_h = 1 : i32, stride_w = 1 : i32} : (tensor<1x5x5x1xf32>, tensor<3x1x1x1xf32>, tensor<3xf32>) -> tensor<1x5x5x3xf32>
  func.return %conv : tensor<1x5x5x3xf32>