{"quantized_ops": ["${main_op}", "Relu"], "act_func": "internal_requantize_and_relu_fn", "output_type": "i8"},
      {"quantized_ops": ["${main_op}", "Relu6"], "act_func": "internal_requantize_and_relu6_fn", "output_type": "i8"},
      {"quantized_ops": ["${main_op}"], "act_func": "internal_dequantize_no_activation_fn", "output_type": "f32"},

  HASHTABLE_LOOKUP = 10,
  L2_NORMALIZATION = 11,
  L2_POOL_2D = 12,
  LOCAL_RESPONSE_NORMALIZATION = 13,
  LOGISTIC = 14,
  LSH_PROJECTION = 15,
  LSTM = 16,
  MAX_POOL_2D = 17,
  MUL = 18,
  RELU = 19,
  // NOTE(aselle): RELU_N1_TO_1 used to be called RELU1, but it was renamed
  // since different model developers use RELU1 in different ways. Never
  // create another op called RELU1.
  RELU_N1_TO_1 = 20,

    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<1x5x5x2xf32>, tensor<3x1x1x2xf32>, tensor<3xf32>) -> tensor<1x5x5x3xf32>
  %conv2 = "tfl.conv_2d"(%0, %w, %b2) {
    dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "RELU",
    padding = "SAME", stride_h = 1 : i32, stride_w = 1 : i32

    // Currently, GPU only supports Conv2D+BiasAdd+Relu fusion.
    if (IsGpuDevice(conv)) {
      auto activation = GetActivation(bias_add);
      if (!activation || activation->getName().stripDialect() != "Relu" ||
          !bias_add.getOutput().hasOneUse()) {
        (void)rewriter.notifyMatchFailure(conv, [&](Diagnostic &diag) {
          diag << "GPU only supports Conv2D+BiasAdd+Relu fusion";
        });
        return false;

  HASHTABLE_LOOKUP = 10,
  L2_NORMALIZATION = 11,
  L2_POOL_2D = 12,
  LOCAL_RESPONSE_NORMALIZATION = 13,
  LOGISTIC = 14,
  LSH_PROJECTION = 15,
  LSTM = 16,
  MAX_POOL_2D = 17,
  MUL = 18,
  RELU = 19,
  // NOTE(aselle): RELU_N1_TO_1 used to be called RELU1, but it was renamed
  // since different model developers use RELU1 in different ways. Never
  // create another op called RELU1.
  RELU_N1_TO_1 = 20,

  return %2 : tensor<*xf32>
}

// CHECK-LABEL: gelu_aten
// CHECK: %0 = "tfl.gelu"(%arg0) <{approximate = false}> : (tensor<5x10xf32>) -> tensor<5x10xf32>

func.func private @gelu_decomp_2(%arg0: tensor<5x10xf32>) -> tensor<5x10xf32>
func.func @gelu_aten_approximate(%arg0: tensor<5x10xf32>) -> (tensor<*xf32>) {

    %14 = "tfl.relu"(%10#1) : (tensor<4x2xf32>) -> tensor<4x2xf32>
    %15 = "tfl.logistic"(%10#0) : (tensor<4x2xf32>) -> tensor<4x2xf32>
    %16 = tfl.mul %15, %14 {fused_activation_function = "NONE"} : tensor<4x2xf32>
    %17 = tfl.add %13, %16 {fused_activation_function = "NONE"} : tensor<4x2xf32>
    %18 = "tfl.relu"(%17) : (tensor<4x2xf32>) -> tensor<4x2xf32>

               (MHLO_ConstantOp:$one (GetScalarOfType<1> $features)),
               (BinBroadcastDimensions $one, $features))))>;

//===----------------------------------------------------------------------===//
// Relu op patterns.
//===----------------------------------------------------------------------===//

// TODO(hinsu): Make these patterns to TF to TF lowering. Relu6 lowering will

  %broad2_s = "quantfork.stats"(%broad2) {layerStats = dense<[0.000000e+00, 1.000000e+01]> : tensor<2xf32>} : (tensor<?x26x26x26x16xf32>) -> tensor<?x26x26x26x16xf32>
  %add = "tfl.add"(%broad1_s, %broad2_s) {fused_activation_function = "RELU"} : (tensor<?x26x26x26x16xf32>, tensor<?x26x26x26x16xf32>) -> tensor<?x26x26x26x16xf32>

  // CHECK-NEXT: tf_mlrt.executeop.device{{.*}}op: \22Relu\22
  // CHECK-NEXT: tf_mlrt_tpu.compile_and_execute
  %0 = "tf.Cast"(%arg0) {__op_key = 0: i32, device = "/device:CPU:0"} : (tensor<i32>) -> tensor<f32>
  %1 = "tf.Relu"(%0) {__op_key = 1: i32, device = "/device:CPU:0"} : (tensor<f32>) -> (tensor<f32>)