z.Square(x)
	z.Mul(x, z)
	t0.Square(z)
	for s := 1; s < 2; s++ {
		t0.Square(t0)
	}
	z.Mul(z, t0)
	t0.Square(z)
	for s := 1; s < 4; s++ {
		t0.Square(t0)
	}
	z.Mul(z, t0)
	t0.Square(z)
	for s := 1; s < 8; s++ {
		t0.Square(t0)
	}
	z.Mul(z, t0)
	t0.Square(z)
	for s := 1; s < 16; s++ {
		t0.Square(t0)
	}
	z.Mul(z, t0)
	for s := 0; s < 32; s++ {

	// 386:"SHLL\t[$]5",-"IMULL"
	// arm:"SLL\t[$]5",-"MUL"
	// arm64:"LSL\t[$]5",-"MUL"
	// ppc64x:"SLD\t[$]5",-"MUL"
	a := n1 * 32

	// amd64:"SHLQ\t[$]6",-"IMULQ"
	// 386:"SHLL\t[$]6",-"IMULL"
	// arm:"SLL\t[$]6",-"MUL"
	// arm64:`NEG\sR[0-9]+<<6,\sR[0-9]+`,-`LSL`,-`MUL`
	// ppc64x:"SLD\t[$]6","NEG\\sR[0-9]+,\\sR[0-9]+",-"MUL"
	b := -64 * n2

	return a, b
}

func Mul_96(n int) int {

	}
	i := new(big.Int).Lsh(h, 1)
	i.Mul(i, i)
	j := new(big.Int).Mul(h, i)

	s1 := new(big.Int).Mul(y1, z2)
	s1.Mul(s1, z2z2)
	s1.Mod(s1, curve.P)
	s2 := new(big.Int).Mul(y2, z1)
	s2.Mul(s2, z1z1)
	s2.Mod(s2, curve.P)
	r := new(big.Int).Sub(s2, s1)
	if r.Sign() == -1 {
		r.Add(r, curve.P)
	}
	yEqual := r.Sign() == 0
	if xEqual && yEqual {
		return curve.doubleJacobian(x1, y1, z1)
	}

	MUL	x2, x2, t0
	ADCS	t0, acc4, acc4
	UMULH	x2, x2, t1
	ADCS	t1, acc5, acc5

	MUL	x3, x3, t0
	ADCS	t0, acc6, acc6
	UMULH	x3, x3, t1
	ADC	t1, acc7, acc7
	// First reduction step
	MUL	acc0, hlp1, hlp0

	MUL	const0, hlp1, t0
	ADDS	t0, acc0, acc0
	UMULH	const0, hlp0, t1

	MUL	const1, hlp0, t0
	ADCS	t0, acc1, acc1
	UMULH	const1, hlp0, y0

	MUL	const2, hlp0, t0
	ADCS	t0, acc2, acc2

    // CHECK: [[ACCUM_MOMENTUM:%.*]] = "tf.Mul"([[ACCUM]], [[MOMENTUM]])
    // CHECK: [[ACCUM_NEW:%.*]] = "tf.AddV2"([[ACCUM_MOMENTUM]], [[GRAD]])
    // CHECK: "tf.AssignVariableOp"([[ACCUM_HANDLE]], [[ACCUM_NEW]])
    // CHECK: [[GRAD_LR:%.*]] = "tf.Mul"([[GRAD]], [[LR]])
    // CHECK: [[MOMENTUM_LR:%.*]] = "tf.Mul"([[MOMENTUM]], [[LR]])
    // CHECK: [[ACCUM_NEW_MOMENTUM_LR:%.*]] = "tf.Mul"([[ACCUM_NEW]], [[MOMENTUM_LR]])

        %cast = "tf.Cast"(%output) : (tensor<1024x24x24x3xi32>) -> tensor<1024x24x24x3xf32>
        %mul = "tf.Mul"(%cast, %scale) : (tensor<1024x24x24x3xf32>, tensor<f32>) -> tensor<1024x24x24x3xf32>
        func.return %mul : tensor<1024x24x24x3xf32>
      }
    }
  )mlir";

  OwningOpRef<ModuleOp> module_op_ref = ParseModuleOpString(kModuleCode);
  const auto test_func =

                        ZerosLike(scope, x));
    auto gy_1 = Mul(scope, Mul(scope, grad, z), log_x);
    return BinaryGradCommon(scope, op, grad_outputs, gx_1, gy_1);
  } else {
    // There's no sensible real value to return if x < 0, so return 0
    auto log_x = Where3(scope, Greater(scope, x, zero), Log(scope, x),
                        ZerosLike(scope, x));
    auto gy_1 = Mul(scope, Mul(scope, grad, z), log_x);

    %0 = "tf.Mul"(%arg1, %arg0) : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
    func.return %0 : tensor<1xf32>
  }
// CHECK: func private @mul1(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>) -> tensor<1xf32> attributes {tf.entry_function = {inputs = "mul1_y:0,mul1_x:0", outputs = "PartitionedCall:0"}}
// CHECK:   %[[MUL_0:.*]] = "tf.Mul"(%arg1, %arg0) : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

  func.return %2 : tensor<4xf32>

// CHECK-NEXT: %[[CST:.*]] = arith.constant dense<1.000000e+00> : tensor<4xf32>
// CHECK-NEXT: %[[MUL:.*]] = tfl.mul %arg0, %[[CST]] {fused_activation_function = "NONE"} : tensor<4xf32>
// CHECK-NEXT: %[[CUSTOM_1:.*]] = "tfl.custom_tf"(%[[MUL]], %[[CST]]) ({
// CHECK-NEXT: ^bb0(%arg1: tensor<4xf32>, %arg2: tensor<4xf32>):

	t1.Mul(t4, y3)                            // t1 := t4 * Y3
	t2.Mul(t0, y3)                            // t2 := t0 * Y3
	y3.Mul(x3, z3)                            // Y3 := X3 * Z3
	y3.Add(y3, t2)                            // Y3 := Y3 + t2
	x3.Mul(t3, x3)                            // X3 := t3 * X3
	x3.Sub(x3, t1)                            // X3 := X3 - t1
	z3.Mul(t4, z3)                            // Z3 := t4 * Z3