func.func @mean_4d_keepdim(%arg0: tensor<1x48x48x512xf32>) -> tensor<1x1x1x512xf32> {
  %cst = arith.constant dense<[1, 2]> : tensor<2xi32>
  %0 = "tfl.mean"(%arg0, %cst) {keep_dims = true} : (tensor<1x48x48x512xf32>, tensor<2xi32>) -> tensor<1x1x1x512xf32>
  func.return %0 : tensor<1x1x1x512xf32>
}

// CHECK:       func @mean_4d_keepdim([[VAL_0:%.*]]: tensor<1x48x48x512xf32>) -> tensor<1x1x1x512xf32> {

// CHECK-LABEL: lift_float_matmul
func.func @lift_float_matmul(%arg0: tensor<1x12x12x512xf32>) -> (tensor<*xf32>, tensor<*xf32>) {
  %cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<512x512xf32>} : () -> tensor<512x512xf32>
  %out_1 = "tf.MatMul"(%arg0, %cst) {
    device = "", transpose_a = false, transpose_b = false
  } : (tensor<1x12x12x512xf32>, tensor<512x512xf32>) -> tensor<*xf32>
  %out_2 = "tf.MatMul"(%arg0, %arg0) {

// CHECK:  [[VAL_1:%.*]] = "tfl.reshape"(%2, %[[CST]]) {tac.device = "GPU",  tac.inference_type = "FLOAT"} : (tensor<1xf32>, tensor<4xi32>) -> tensor<1x1x1x1xf32>
// CHECK:  [[VAL_2:%.*]] = "tfl.concatenation"([[VAL_0]], [[VAL_1]]) <{axis = 3 : i32, fused_activation_function = "NONE"}> {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<1x1x1x1xf32>, tensor<1x1x1x1xf32>) -> tensor<1x1x1x2xf32>

    %2 = "tf.BiasAdd"(%1, %cst_0) {data_format = "NHWC", device = ""} : (tensor<*xf32>, tensor<2xf32>) -> tensor<*xf32>
    func.return %2: tensor<*xf32>
  }
  func.func private @composite_conv2d_fn_1(%arg0: tensor<1x3x4x512xf32>, %arg1: tensor<2x3x512x2xf32>) -> tensor<*xf32> attributes {tf_quant.composite_function} {

    %1 = "tf.BiasAdd"(%0, %cst_0) {data_format = "NHWC", device = ""} : (tensor<*xf32>, tensor<2xf32>) -> tensor<*xf32>
    func.return %1: tensor<*xf32>
  }
  func.func private @composite_depthwise_conv2d_fn(%arg0: tensor<1x3x4x512xf32>, %arg1: tensor<2x3x3x512xf32>) -> tensor<*xf32> attributes {tf_quant.composite_function} {

func.func @main(tensor<1x6x6x16xf32>) -> tensor<1x1x1x16xf32> {
^bb0(%arg0: tensor<1x6x6x16xf32>):
  // CHECK: "tfl.average_pool_2d"(%{{.*}}) <{filter_height = 3 : i32, filter_width = 6 : i32, fused_activation_function = "NONE", padding = "VALID", stride_h = 3 : i32, stride_w = 1 : i32}> : (tensor<1x6x6x16xf32>) -> tensor<1x1x1x16xf32>

  func.return %0 : tensor<1x1x1x16xf32>

    return %15 : tensor<1x2240x1120x512xi8>
  }

// CHECK-LABEL: func @conv_with_filter_larger_than_1GB
// CHECK-DAG: %[[CONST:.*]] = "tf.Const"() <{value = dense<-237772800> : tensor<1x1x1x512xi32>}> : () -> tensor<1x1x1x512xi32>
// CHECK: %[[PADV2_0:.*]] = "tf.PadV2"
// CHECK: %[[XLACONVV2_0:.*]] = "tf.XlaConvV2"(%[[PADV2_0]]
// CHECK: %[[SUB_0:.*]] = "tf.Sub"(%[[XLACONVV2_0]], %[[CONST]])

  func.return %result : tensor<1x1x512xf32>