'input_': input_,
        'filter_': filter_,
        'bias': bias,
        'stride_w': 2,
        'stride_h': 2,
        'dilation_w': 1,
        'dilation_h': 1,
        'padding': 'SAME',
        'act': 'RELU'
    }

    self._assertOpAndComposite([input_, filter_, bias],
                               tf.function(gen_mnist_ops.new_conv2d),
                               ops_defs._composite_conv_add_relu, kwargs)

    attrs=['act: {"", "RELU", "RELU6", "TANH"} = ""'],
    derived_attrs=['T: {float, int8}'],
    outputs=['o: T'])
def _composite_fully_connected(input_, filter_, bias, act):
  res = tf.raw_ops.MatMul(
      a=input_, b=filter_, transpose_a=False, transpose_b=True)
  res = tf.raw_ops.Add(x=res, y=bias)
  if act == 'RELU':
    return tf.raw_ops.Relu(features=res)
  elif act == 'RELU6':
    return tf.raw_ops.Relu6(features=res)

    .SetIsAggregate();

REGISTER_OP("MyBiasedDense")
    .Input("input: T")
    .Input("weight: T")
    .Input("bias: T")
    .Output("out: T")
    .Attr("T: {float, int8}")
    .Attr("act: {'', 'relu', 'relu6'} = ''");

 Conv2DBackpropInput 52
 DivNoNan 1
 Equal 1
 FusedBatchNorm 53
 FusedBatchNormGrad 53
 Identity 2
 MatMul 3
 MaxPool 1
 MaxPoolGrad 1
 Mean 1
 Mul 164
 Pad 1
 ReadVariableOp 646
 Relu 49
 ReluGrad 49
 Reshape 2
 ResourceApplyKerasMomentum 161
 ShapeN 50
 Softmax 1
 SparseSoftmaxCrossEntropyWithLogits 1
 Square 55
 Squeeze 1
 Sub 106
 Sum 57
 Tile 1

def MaximumOfZeroToRelu : Pat<
  (TF_MaximumOp:$maximum_op $x, $y), (TF_ReluOp:$dest $x),
  [(IsConstantValueOf<0> $y), (IsDeviceCompatible $maximum_op)],
  [(CopyAttrs $maximum_op, $dest)]>;

//===----------------------------------------------------------------------===//
// Canonicalize tf.Relu of Minimul six to tf.Relu6

// CHECK-NEXT:      outputs: [ 5, 6, 7, 8 ],
// CHECK-NEXT:      builtin_options_type: LSTMOptions,
// CHECK-NEXT:      builtin_options: {
// CHECK-NEXT:        fused_activation_function: RELU,
// CHECK-NEXT:        cell_clip: 1.0,
// CHECK-NEXT:        proj_clip: 2.0,
// CHECK-NEXT:        kernel_type: BASIC
// CHECK-NEXT:      },
// CHECK-NEXT:      intermediates: [ ]
// CHECK-NEXT:    } ],

 Conv2D 53
 Conv2DBackpropFilter 53
 Conv2DBackpropInput 52
 Equal 1
 FusedBatchNormGradV2 53
 FusedBatchNormV2 53
 MatMul 3
 MaxPool 1
 MaxPoolGrad 1
 Mean 1
 Mul 218
 Pad 2
 ReadVariableOp 538
 Relu 49
 ReluGrad 49
 Reshape 2
 ResourceApplyKerasMomentum 161
 Slice 1
 Softmax 1
 SparseSoftmaxCrossEntropyWithLogits 1
 Squeeze 1
 Sum 1
 Tile 1
 Transpose 1
cluster 1 size 815
 AddN 1

  // CHECK: tfrt_fallback_async.executeop key(0) cost(100) device("cpu") "tf.Relu"(%{{.*}}) {T = f32} : 1
  %2 = tfrt_fallback_async.executeop key(0) cost(100) device("cpu") "tf.Relu"(%1) {T = f32} : 1
  // CHECK: tfrt_fallback_async.fallback_tensor_to_corert_tensorhandle

  %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>
   // CHECK: tac.device = "GPU", tac.inference_type = "FLOAT"
  %2 = "tfl.relu"(%arg0) : (tensor<1xf32>) -> tensor<1xf32>

      {"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"},