%0 = tfl.add %arg0, %arg1 {fused_activation_function = "NONE"} : tensor<*xf32>
  func.return %0 : tensor<*xf32>
}

func.func @cond_false(%arg0: tensor<*xf32>, %arg1: tensor<*xf32>) -> tensor<*xf32> {
  %0 = tfl.mul %arg0, %arg1 {fused_activation_function = "NONE"} : tensor<*xf32>
  func.return %0 : tensor<*xf32>

  %0 = "tfl.add"(%arg0, %arg1) {fused_activation_function = "RELU", tac.device = "CPU"} : (tensor<10x10x10xf32>, tensor<10xf32>) -> tensor<10x10x10xf32>
  %1 = "tfl.mul"(%0, %arg1) {fused_activation_function = "RELU", tac.device = "CPU"} : (tensor<10x10x10xf32>, tensor<10xf32>) -> tensor<10x10x10xf32>
  func.return %1 : tensor<10x10x10xf32>
}

// -----

// CHECK: tac.cost = 0x4B673001

	b, _ := new(big.Int).SetString("180692823610368451951102211649591374573781973061758082626801", 10)
	n := new(big.Int).Mul(a, b)

	N, _ := NewModulusFromBig(n)
	A := NewNat().setBig(a).ExpandFor(N)
	B := NewNat().setBig(b).ExpandFor(N)

	if A.Mul(B, N).IsZero() != 1 {
		t.Error("a * b mod (a * b) != 0")
	}

	i := new(big.Int).ModInverse(a, b)
	N, _ = NewModulusFromBig(b)

  %1 = "tfl.mul"(%0, %cst) {fused_activation_function = "NONE"} : (tensor<256x8x7x3xf32>, tensor<3xf32>) -> tensor<256x8x7x3xf32>
  func.return %1 : tensor<256x8x7x3xf32>

  AbstractTensorHandle* expand_dims_outputs;
  TF_RETURN_IF_ERROR(
      ops::ExpandDims(ctx, vec, dim.get(), &expand_dims_outputs, "ExpandDims"));
  TF_RETURN_IF_ERROR(
      ops::Mul(ctx, expand_dims_outputs, mat, &outputs[0], "Mul"));
  expand_dims_outputs->Unref();
  return absl::OkStatus();
}

class SparseSoftmaxCrossEntropyWithLogitsGradientFunction
    : public GradientFunction {
 public:

	var p Point
	p.ScalarMult(dalekScalar, B)
	if dalekScalarBasepoint.Equal(&p) != 1 {
		t.Error("Scalar mul does not match dalek")
	}
	checkOnCurve(t, &p)
}

func TestBaseMultVsDalek(t *testing.T) {
	var p Point
	p.ScalarBaseMult(dalekScalar)
	if dalekScalarBasepoint.Equal(&p) != 1 {
		t.Error("Scalar mul does not match dalek")
	}
	checkOnCurve(t, &p)
}

func TestVarTimeDoubleBaseMultVsDalek(t *testing.T) {

	return e
}

// Sub sets e = t1 - t2, and returns e.
func (e *P384Element) Sub(t1, t2 *P384Element) *P384Element {
	p384Sub(&e.x, &t1.x, &t2.x)
	return e
}

// Mul sets e = t1 * t2, and returns e.
func (e *P384Element) Mul(t1, t2 *P384Element) *P384Element {
	p384Mul(&e.x, &t1.x, &t2.x)
	return e
}

// Square sets e = t * t, and returns e.
func (e *P384Element) Square(t *P384Element) *P384Element {

	return e
}

// Sub sets e = t1 - t2, and returns e.
func (e *P521Element) Sub(t1, t2 *P521Element) *P521Element {
	p521Sub(&e.x, &t1.x, &t2.x)
	return e
}

// Mul sets e = t1 * t2, and returns e.
func (e *P521Element) Mul(t1, t2 *P521Element) *P521Element {
	p521Mul(&e.x, &t1.x, &t2.x)
	return e
}

// Square sets e = t * t, and returns e.
func (e *P521Element) Square(t *P521Element) *P521Element {

    // CHECK: "tf.Mul"(%[[v1]], %[[v2]])
    %0 = "tf.VarHandleOp"() {container = "", shared_name = "v"} : () -> tensor<!tf_type.resource<tensor<10xf32>>>
    %1 = "tf.ReadVariableOp"(%0) : (tensor<!tf_type.resource<tensor<10xf32>>>) -> tensor<10xf32>
    %2 = "tf.ReadVariableOp"(%0) : (tensor<!tf_type.resource<tensor<10xf32>>>) -> tensor<10xf32>
    %3 = "tf.Mul"(%1, %2) : (tensor<10xf32>, tensor<10xf32>) -> tensor<10xf32>

                      std::vector<Output>* grad_outputs) {
  auto softmax = Exp(scope, op.output(0));
  auto sum = Sum(scope, grad_inputs[0], {1}, Sum::KeepDims(true));
  auto mul = Mul(scope, sum, softmax);
  auto dx = Sub(scope, grad_inputs[0], mul);
  grad_outputs->push_back(dx);
  return scope.status();
}
REGISTER_GRADIENT_OP("LogSoftmax", LogSoftmaxGrad);

Status ReluGradHelper(const Scope& scope, const Operation& op,