%1 = "tf.BiasAdd"(%0, %arg2) {data_format = "NHWC", device = ""} : (tensor<*xf32>, tensor<2xf32>) -> tensor<*xf32>
    %2 = "tf.Relu6"(%1) : (tensor<*xf32>) -> tensor<*xf32>
    func.return %2 : tensor<*xf32>
  }

}

// -----

func.func @testAddReluPack(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>) {
   // CHECK: tac.device = "GPU", tac.inference_type = "FLOAT"
  %0 = "tfl.add"(%arg0, %arg1) {fused_activation_function = "RELU6"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
   // CHECK: tac.device = "GPU", tac.inference_type = "FLOAT"
  %1 = "tfl.add"(%arg0, %0) {fused_activation_function = "RELU"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

// tfl.log / tfl.logistic / tfl.max_pool_2d / tfl.mirror_pad / tfl.maximum /
// tfl.custom / tfl.mean / tfl.minimum / tfl.pad / tfl.pow / tfl.prelu /
// tfl.relu / tfl.relu6 / tfl.rsqrt / tfl.sin / tfl.slice / tfl.softmax /
// tfl.space_to_depth / tfl.sqrt / tfl.square / tfl.squared_difference /
// tfl.strided_slice / tfl.tanh / tfl.transpose / tfl.transpose_conv

    %1 = "tf.BiasAdd"(%0, %arg2) {data_format = "NHWC", device = ""} : (tensor<*xf32>, tensor<2xf32>) -> tensor<*xf32>
    %2 = "tf.Relu6"(%1) : (tensor<*xf32>) -> tensor<*xf32>
    func.return %2 : tensor<*xf32>
  }

// CHECK-LABEL: func @conv_with_single_layer

// CHECK: %[[quantize:.*]] = "tf.PartitionedCall"(%arg0
// CHECK-SAME: f = @quantize_i8

using ops::FusedBatchNormV3;
using ops::L2Loss;
using ops::LogSoftmax;
using ops::LRN;
using ops::MaxPool;
using ops::MaxPool3D;
using ops::MaxPoolV2;
using ops::Placeholder;
using ops::Relu;
using ops::Relu6;
using ops::Selu;
using ops::Softmax;
using ops::Softplus;
using ops::Softsign;

class NNGradTest : public ::testing::Test {
 protected:
  NNGradTest() : scope_(Scope::NewRootScope()) {}

  %biasadd = "tf.BiasAdd"(%conv, %dq_bias) {data_format = "NHWC", device = ""} : (tensor<*xf32>, tensor<2xf32>) -> tensor<*xf32>
  %res = "tf.Relu6"(%biasadd) : (tensor<*xf32>) -> tensor<*xf32>
  %q_res = "quantfork.qcast"(%res) : (tensor<*xf32>) -> tensor<*x!quant.uniform<i8:f32, 0.023529411764705882:-128>>

    %6 = "quantfork.stats"(%5) {layerStats = dense<[-2788.31055, 4616.62842]> : tensor<2xf32>} : (tensor<*xf32>) -> tensor<*xf32>
    %7 = "tf.Relu6"(%6) {device = ""} : (tensor<*xf32>) -> tensor<*xf32>
    %8 = "quantfork.stats"(%7) {layerStats = dense<[0.000000e+00, 6.000000e+00]> : tensor<2xf32>} : (tensor<*xf32>) -> tensor<*xf32>
    return %8 : tensor<*xf32>
  }

// CHECK: %[[conv:.*]] = "tf.Conv2D"(%[[dq_input]], %[[dq_weight]])
// CHECK: %[[bias:.*]] = "tf.BiasAdd"(%[[conv]], %[[cst_0]]) <{data_format = "NHWC"}>
// CHECK: %[[relu6:.*]] = "tf.Relu6"(%[[bias]])

// -----

func.func @remove_check_numerics_op(%arg0: tensor<*xf32>) -> tensor<*xf32> {
  %0 = "tf.CheckNumerics"(%arg0) {device = "", message = "transformer"} : (tensor<*xf32>) -> tensor<*xf32>

    [(IsNotInLiftedFunc $res),
     (IsEinsumSupportedByXlaDotV2 $equation)],
    [], (addBenefit 10)>;

}
defm : LiftCompositeOpsWithActivation<TF_ReluOp, "relu">;
defm : LiftCompositeOpsWithActivation<TF_Relu6Op, "relu6">;

def LiftGather : Pat<
  (TF_GatherV2Op:$res $params, $indices, $axis, $batch_dims),
  (LiftAsTFPartitionedCall<"composite_gather_fn">
    (ArgumentList $params, $indices, $axis),
    (ResultList $res),

  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,
  RELU6 = 21,
  RESHAPE = 22,
  RESIZE_BILINEAR = 23,
  RNN = 24,
  SOFTMAX = 25,
  SPACE_TO_DEPTH = 26,
  SVDF = 27,
  TANH = 28,
  CONCAT_EMBEDDINGS = 29,
  SKIP_GRAM = 30,
  CALL = 31,