%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>

    %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>

    %1 = stablehlo.constant dense<1.000000e+03> : tensor<1x1x1x1xf32>  // Input inverse scale.
    %2 = stablehlo.constant dense<-128> : tensor<1x1x1x1xi8>  // Input zero point.
    %3 = stablehlo.constant dense<1> : tensor<3x3x4x4xi8>  // Quantized filter tensor.
    %4 = stablehlo.constant dense<3.000000e+03> : tensor<1x1x1x4xf32>
    %5 = stablehlo.constant dense<4.000000e+03> : tensor<1x1x1x1xf32>  // Output inverse scale.

  %cst = arith.constant dense<2.0> : tensor<1x1x1x1x2xf32>
  %shape = arith.constant dense<[1, 1, 1, 1, 2]> : tensor<5xi32>
  %1 = "tfl.reshape"(%arg0, %shape) : (tensor<2x1x1x1x1xf32>, tensor<5xi32>) -> tensor<1x1x1x1x2xf32>
  %2 = "tfl.add"(%1, %cst) {fused_activation_function = "NONE"} : (tensor<1x1x1x1x2xf32>, tensor<1x1x1x1x2xf32>) -> tensor<1x1x1x1x2xf32>

  %r = "mhlo.concatenate"(%0, %1, %2) <{dimension = 0 : i64}> : (tensor<1x1x512xf32>, tensor<1x1x512xf32>, tensor<1x1x512xf32>) -> tensor<3x1x512xf32>
  func.return %r : tensor<3x1x512xf32>

// CHECK: return %arg0 : tensor<3x1x512xf32>
}

// -----

// CHECK-LABEL: testConvertReshapeDotRhsToBatchedDot

    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]])

  %cst_0 = "tf.Const"() {value = dense<0.400000e+00> : tensor<1x1x1x2xf32>} : () -> tensor<1x1x1x2xf32>
  %cst_1 = "tf.Const"() {value = dense<0.200000e+00> : tensor<1x1x1x2xf32>} : () -> tensor<1x1x1x2xf32>
  %cst_2 = "tf.Const"() {value = dense<0.300000e+00> : tensor<1x1x1x2xf32>} : () -> tensor<1x1x1x2xf32>

// 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) {