// CHECK: tac.cost = 0x4D5C0000
  %0 = "tfl.conv_2d"(%arg0, %arg1, %arg2) {dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, padding = "SAME", stride_h = 1 : i32, stride_w = 1 : i32, fused_activation_function = "RELU6", tac.device = "CPU"} : (tensor<256x32x32x3xf32>, tensor<16x3x3x3xf32>, tensor<16xf32>) -> tensor<256x32x32x16xf32>
  func.return %0 : tensor<256x32x32x16xf32>
}

  // CHECK:          %{{.*}} = "tfl.pseudo_const"() <{value = dense<{{\[\[1, 2\], \[3, 4\], \[5, 6\]\]}}> : tensor<3x2xi32>}>
  // CHECK-NEXT:     [[SUB:%.*]] = tfl.sub %{{.*}}, %{{.*}} {fused_activation_function = "RELU6"} : tensor<3x2xi32>
  // CHECK-NEXT:     [[SCALAR:%.*]] = "tfl.pseudo_const"() <{value = dense<10> : tensor<i32>}> : () -> tensor<i32>

    %0 = tfl.add %arg0, %arg1 {fused_activation_function = "RELU6", tac.device = "GPU"} : tensor<100xf32>
    %1 = tfl.mul %0, %arg2 {fused_activation_function = "RELU6", tac.device = "GPU"} : tensor<100xf32>
    func.return %1 : tensor<100xf32>
  }

// CHECK-DAG: %[[ADD:.*]] = stablehlo.add %[[CONV]], %[[BROADCAST]] : tensor<1x3x2x2xf32>
// CHECK-DAG: %[[RELU6:.*]] = stablehlo.clamp %[[CONST_2]], %[[ADD]], %[[CONST_1]] : (tensor<f32>, tensor<1x3x2x2xf32>, tensor<f32>) -> tensor<1x3x2x2xf32>
// CHECK: return %[[RELU6]] : tensor<1x3x2x2xf32>

// -----

func.func @func_conv_batchnorm_relu6_dynamic(%arg_0: tensor<?x3x4x3xf32>) -> (tensor<?x3x2x2xf32>) {

  %1 = "tfl.mul"(%0, %arg2) {tac.device = "GPU", fused_activation_function = "RELU6", tac.inference_type = "FLOAT"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

    %0 = tfl.add %arg0, %arg1 {fused_activation_function = "RELU6", tac.device = "GPU", tac.inference_type = "FLOAT"} : tensor<1xf32>
    %1 = tfl.mul %0, %arg2 {fused_activation_function = "RELU6", tac.device = "GPU", tac.inference_type = "FLOAT"} : tensor<1xf32>
    func.return %1 : tensor<1xf32>
  }

  %0 = "tfl.add"(%arg0, %arg1) {fused_activation_function = "RELU6",  per_device_costs = {CPU = 5.0 : f32, GPU = 1.0 : f32}, tac.device = "GPU"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
  %1 = "tfl.mul"(%0, %arg2) {fused_activation_function = "RELU6", per_device_costs = {CPU = 5.0 : f32, GPU = 1.0 : f32}, tac.device = "GPU"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

          'input_shape_dynamic': False,
          'enable_per_channel_quantization': False,
          'dilations': [1, 2, 2, 1],
      },
      {
          'testcase_name': 'relu6',
          'activation_fn': nn_ops.relu6,
          'has_bias': False,
          'has_batch_norm': False,
          'target_opset': quant_opts_pb2.TF,
          'input_shape_dynamic': False,

  %cst = arith.constant dense<1.0> : tensor<4xf32>
  %cst_3 = arith.constant dense<2.0> : tensor<4x3x3x3xf32>
  %0 = "tfl.conv_2d"(%arg0, %cst_3, %cst) {dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "RELU6", padding = "VALID", stride_h = 2 : i32, stride_w = 2 : i32} : (tensor<?x19x19x3xf32>, tensor<4x3x3x3xf32>, tensor<4xf32>) -> tensor<?x9x9x4xf32>
  func.return %0 : tensor<?x9x9x4xf32>

    %1 = "tf.BiasAdd"(%0, %cst) {data_format = "NHWC", device = ""} : (tensor<1x2x2x3xf32>, tensor<3xf32>) -> tensor<1x2x2x3xf32>
    %2 = "tf.Relu6"(%1) {device = ""} : (tensor<1x2x2x3xf32>) -> tensor<1x2x2x3xf32>
    %3 = "tf.Identity"(%2) {device = ""} : (tensor<1x2x2x3xf32>) -> tensor<1x2x2x3xf32>
    return %3 : tensor<1x2x2x3xf32>
  }
}