z3.Mul(z3, z3)
	z3.Sub(z3, gamma)
	if z3.Sign() == -1 {
		z3.Add(z3, curve.P)
	}
	z3.Sub(z3, delta)
	if z3.Sign() == -1 {
		z3.Add(z3, curve.P)
	}
	z3.Mod(z3, curve.P)

	beta.Lsh(beta, 2)
	beta.Sub(beta, x3)
	if beta.Sign() == -1 {
		beta.Add(beta, curve.P)
	}
	y3 := alpha.Mul(alpha, beta)

	gamma.Mul(gamma, gamma)
	gamma.Lsh(gamma, 3)
	gamma.Mod(gamma, curve.P)

    return row * pool.w_out + col;
  };

  std::vector<float> correction_data(out_shaped_type.getNumElements(), 1.0);

  const float kern_size = pool.kh * pool.kw;

  // LEMMA 1: Changing the rounding mode from floor to ceil will increase an
  // output dimension by at most 1 (see `ComputeSpatialOutDim`). This is because
  // for any `x`, `ceil(x) - floor(x) <= 1`.