%1 = "tf.Identity"(%0) {device = ""} : (tensor<1x3x6x6xf32>) -> tensor<*xf32>
  %2 = "tf.Identity"(%1) {device = ""} : (tensor<*xf32>) -> tensor<*xf32>
  return %2 : tensor<*xf32>
}
func.func private @XlaCallModule_aten.avg_pool2d.default.impl_1(%arg0: tensor<1x3x6x6xf32>) -> tensor<1x3x6x6xf32> {

    %3 = "tf.Identity"(%2) {device = ""} : (tensor<1x3x4x2xf32>) -> tensor<1x3x4x2xf32>
    %4 = "quantfork.stats"(%3) {layerStats = dense<[3.50919247, 6.000000e+00]> : tensor<2xf32>} : (tensor<1x3x4x2xf32>) -> tensor<1x3x4x2xf32>
    func.return %4 : tensor<1x3x4x2xf32>
  }

//   func.return %0#0 : tensor<1x1x2x8xf32>
// }
// COM: CHECK: func.func @main() -> tensor<1x1x2x8xf32>
// COM: CHECK-DAG: %[[CONST:.*]] = stablehlo.constant dense<{{.*}}> : tensor<1x1x2x8xf32>

  %0 = stablehlo.constant dense<2.000000e+00> : tensor<1x3x2x4xf32>
  %1 = stablehlo.transpose %arg0, dims = [0, 2, 1, 3] : (tensor<1x2x3x4xf32>) -> tensor<1x3x2x4xf32>
  %2 = stablehlo.add %1, %0 : tensor<1x3x2x4xf32>
  return %2 : tensor<1x3x2x4xf32>
}
// CHECK: %[[TRANSPOSE_0:.+]] = stablehlo.transpose
// CHECK: %[[ADD_0:.+]] = stablehlo.add %[[TRANSPOSE_0]], {{.*}}
// CHECK: return %[[ADD_0]]

// -----

    %1 = "tf.Conv2D"(%0, %cst) {data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x3x2x2xf32>
    %2 = "tf.Relu"(%1) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>

func.func @move_across_double_transpose(%arg0: tensor<1x4x4x8xf32>, %arg1: tensor<1x4x4x8xf32>) -> tensor<1x4x8x4xf32> {

  // CHECK: %[[RES_PERM:.*]] = "tf.Const"() <{value = dense<[0, 3, 1, 2]> : tensor<4xi32>}>
  // CHECK: %[[ADD:[0-9]*]] = "tf.AddV2"(%arg0, %arg1) : (tensor<1x4x4x8xf32>, tensor<1x4x4x8xf32>) -> tensor<1x4x4x8xf32>
  // CHECK: %[[RES_TRANSPOSE_0:[0-9]*]] = "tf.Transpose"(%[[ADD]], %[[RES_PERM]])

        _stablehlo_module_attrs = {}, _tfl_quant_trait = "fully_quantizable",
        device = ""
      } : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x3x4x2xf32>
    return %0 : tensor<1x3x4x2xf32>
  }

  // CHECK: func.func private @qdq_for_conv_weight_per_channel_default(%[[ARG0:.+]]: tensor<1x3x4x3xf32>)
  // CHECK: %[[CST:.+]] = "tf.Const"() <{value = dense<3.000000e-01> : tensor<2x3x3x2xf32>}> : () -> tensor<2x3x3x2xf32>

    return %1 : tensor<1x3x4x2xf32>
  }

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

// CHECK: return %[[identity]] : tensor<1x3x2x2xf32>

func.func @cast_bf16_avg_pool_to_fp32(%arg0: tensor<1x3x4x3xf32>) -> (tensor<1x3x2x2xf32>) {
  %cst = "tf.Const"() {device = "", value = dense<1.000000e+00> : tensor<2x3x3x2xbf16>} : () -> tensor<2x3x3x2xbf16>
  %0 = "tf.Cast"(%arg0) {Truncate = false, device = ""} : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xbf16>

// Remove TPU related ops.
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>
}