var SinNoVec = sin
var SinhNoVec = sinh
var TanhNoVec = tanh
var Log1pNovec = log1p
var AtanhNovec = atanh
var AcosNovec = acos
var AcoshNovec = acosh
var AsinNovec = asin
var AsinhNovec = asinh
var ErfNovec = erf
var ErfcNovec = erfc
var AtanNovec = atan
var Atan2Novec = atan2
var CbrtNovec = cbrt
var LogNovec = log
var TanNovec = tan
var ExpNovec = exp
var Expm1Novec = expm1
var PowNovec = pow

  // CHECK-NEXT: [[r2:%.*]] = tfrt_fallback_async.executeop {{.*}} "tf.Tanh"([[r1]]) {T = f32}
  %4 = "tf.Tanh"(%3) {T = f32, _output_shapes = ["tfshape$dim { size: 3 } dim { size: 3 }"], device = "/device:CPU:0"} : (tensor<3x3xf32>) -> tensor<3x3xf32>
  // CHECK-NOT: tf.Identity

  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)
  elif act == 'TANH':
    return tf.raw_ops.Tanh(x=res)
  else:
    return res

```

Besides defining new ops, composition can be specified for an existing op
for portability. The following code defines the semantics of `AddNOp`:

```python
@Composite('AddNOp')

          (variadic $inputs),
          ConstantStrAttr<StrAttr, "aten.gelu.default">, $attrs, $_, $_),
        (TFL_GeluOp $inputs, ConstBoolAttrTrue),

          "intermediates": [
            25, 26, 27, 28, 29
          ],
          "builtin_options_type": "UnidirectionalSequenceLSTMOptions",
          "builtin_options": {
            "fused_activation_function": "TANH",
            "cell_clip": 50.0
          },
          "mutating_variable_inputs": [
            false,
            false, false, false, false,
            false, false, false, false,

	// Output: 0.00, 1.00
}

func ExampleSinh() {
	fmt.Printf("%.2f", math.Sinh(0))
	// Output: 0.00
}

func ExampleTan() {
	fmt.Printf("%.2f", math.Tan(0))
	// Output: 0.00
}

func ExampleTanh() {
	fmt.Printf("%.2f", math.Tanh(0))
	// Output: 0.00
}

func ExampleSqrt() {
	const (
		a = 3
		b = 4
	)
	c := math.Sqrt(a*a + b*b)
	fmt.Printf("%.1f", c)
	// Output: 5.0
}

def TFL_AFEnum_Relu  : I32EnumAttrCase<"RELU", 1>;
def TFL_AFEnum_Relu1 : I32EnumAttrCase<"RELU_N1_TO_1", 2>;
def TFL_AFEnum_Relu6 : I32EnumAttrCase<"RELU6", 3>;
def TFL_AFEnum_Tanh  : I32EnumAttrCase<"TANH", 4>;
def TFL_AFEnum_Sign  : I32EnumAttrCase<"SIGN_BIT", 5>;

def TFL_AFAttr : TFL_AnyStrAttrOf<[
      TFL_AFEnum_None.symbol,  TFL_AFEnum_Relu.symbol, TFL_AFEnum_Relu1.symbol,

// 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
class GpuBasicSupportedOpNoCost : public TargetHardwareOperation {
  double GetOpCost(mlir::Operation* op) const override { return 0; }

  func.return %0 : tensor<10x10xf32>
}

func.func @testTanh(%arg0: tensor<10x10xf32>) -> tensor<10x10xf32> {
  // CHECK: _arithmetic_count = 6400 : i64
  %0 = "tfl.tanh"(%arg0) : (tensor<10x10xf32>) -> tensor<10x10xf32>
  func.return %0 : tensor<10x10xf32>
}

func.func @testAddN(%arg0: tensor<10x10xf32>, %arg1: tensor<10x10xf32>, %arg2: tensor<10x10xf32>) -> tensor<10x10xf32> {

//

// Atanh returns the inverse hyperbolic tangent of x.
//
// Special cases are:
//
//	Atanh(1) = +Inf
//	Atanh(±0) = ±0
//	Atanh(-1) = -Inf
//	Atanh(x) = NaN if x < -1 or x > 1
//	Atanh(NaN) = NaN
func Atanh(x float64) float64 {
	if haveArchAtanh {
		return archAtanh(x)
	}
	return atanh(x)
}

func atanh(x float64) float64 {
	const NearZero = 1.0 / (1 << 28) // 2**-28
	// special cases