func.func @softmax(%arg0: tensor<8x16xf32>) -> tensor<8x16xf32> {
  %0 = "tf.Softmax"(%arg0) : (tensor<8x16xf32>) -> tensor<8x16xf32>
  func.return %0 : tensor<8x16xf32>
// CHECK: %[[CUSTOM_0:.*]] = "tfl.custom"(%arg0) <{custom_code = "FlexSoftmax", custom_option = #tfl<const_bytes : "0x07536F66746D617800161207536F66746D61781A002A070A0154120230013200000221191414042801">}> : (tensor<8x16xf32>) -> tensor<8x16xf32>

  %2 = "tfl.pseudo_const"() {value = dense<[[0.2]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
  %3 = "tfl.pseudo_const"() {value = dense<[[0.3]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
  %4 = "tfl.pseudo_const"() {value = dense<[[0.4]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
  %5 = "tfl.pseudo_const"() {value = dense<[[0.5]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>

func.func private @gelu_decomp_2(%arg0: tensor<5x10xf32>) -> tensor<5x10xf32>
func.func @gelu_aten_approximate(%arg0: tensor<5x10xf32>) -> (tensor<*xf32>) {
  %0 = mhlo.composite "aten.gelu.default" %arg0 {composite_attributes = {approximate = "tanh"}, decomposition = @gelu_decomp_2} : (tensor<5x10xf32>) -> tensor<5x10xf32>

    %4:4 = "tf.CustomAggregator"(%3) {calibration_method = 1 : i32, id = "1", max_percentile = 0.000000e+00 : f32, min_percentile = 0.000000e+00 : f32, num_bins = 0 : i32} : (tensor<3x11xf32>) -> (tensor<3x11xf32>, tensor<f32>, tensor<f32>, tensor<*xi64>)
    %5 = stablehlo.add %4#0, %1 : tensor<3x11xf32>
    %6 = stablehlo.multiply %5, %1 : tensor<3x11xf32>

// CHECK-NEXT:    return [[RES]], [[MATRES]] : tensor<1x1xf32>, tensor<1x1xf32>
//
module {
  func.func @serving_default(%arg0: tensor<1x3xf32>) -> (tensor<1x1xf32>, tensor<1x1xf32>) {
    %0 = "tf.VarHandleOp"() <{container = "", shared_name = "y"}> : () -> tensor<!tf_type.resource<tensor<3x1xf32>>>
    %2 = "tf.ReadVariableOp"(%0) : (tensor<!tf_type.resource<tensor<3x1xf32>>>) -> tensor<3x1xf32>

  %outputs_2 = "tf.add"(%outputs_0, %outputs_1): (tensor<1x1xf32>, tensor<1x1xf32>) -> tensor<1x1xf32>
  return %outputs_2 : tensor<1x1xf32>
}

// CHECK-LABEL: @_func
func.func private @_func(%arg0: tensor<1x3xf32>, %arg1: tensor<3x1xf32>) -> (tensor<1x1xf32>) {

    } : (tensor<?x100352xf32>, tensor<100352x10xf32>) -> tensor<?x10xf32>
    %3 = "tf.Identity"(%2) {device = ""} : (tensor<?x10xf32>) -> tensor<?x10xf32>
    return %3 : tensor<?x10xf32>
  }

  // CHECK-LABEL: func.func private @composite_dot_general_fn_1
  func.func private @composite_dot_general_fn_1(%arg0: tensor<?x100352xf32>, %arg1: tensor<100352x10xf32>) -> tensor<?x10xf32> {
    // CHECK-NOT: tf.CustomAggregator

    // CHECK-NEXT: return %[[DIV]] : tensor<4x10xf32>
    %0 = "tf.Softsign"(%arg0) : (tensor<4x10xf32>) -> tensor<4x10xf32>
    func.return %0 : tensor<4x10xf32>
}

// -----

// CHECK-LABEL: func @softsign_grad
func.func @softsign_grad(%arg0: tensor<4x10xf32>, %arg1: tensor<4x10xf32>) -> tensor<4x10xf32> {

    // CHECK-NEXT: %[[ONE:.*]] = mhlo.constant dense<1.000000e+00> : tensor<f32>

  // CHECK-DAG: %[[LOG_SOFTMAX_LOG:.*]] = "tf.Log"(%[[LOG_SOFTMAX_SUM]]) : (tensor<2x1xf32>) -> tensor<2x1xf32>
  // CHECK-DAG: %[[LOG_SOFTMAX:.*]] = "tf.Sub"(%[[LOG_SOFTMAX_SHIFTED]], %[[LOG_SOFTMAX_LOG]]) : (tensor<2x3xf32>, tensor<2x1xf32>) -> tensor<2x3xf32>

  %w = arith.constant dense<127.0> : tensor<4x12xf32>
  %b = arith.constant dense<0.0> : tensor<4xf32>
  %fc = "tfl.fully_connected"(%arg0, %w, %b) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<1x224x224x3xf32>, tensor<4x12xf32>, tensor<4xf32>) -> tensor<1x112x112x4xf32>
  func.return %fc : tensor<1x112x112x4xf32>