def compensated_hd1(x, coeffs):
"""Performs the compensated ``HD`` algorithm when ``k = 1``.
.. _JGH+13: https://dx.doi.org/10.1016/j.cam.2012.11.008
See the `JGH+13`_ paper for more details on the ``HD`` algorithm and the
``CompHD`` algorithm.
Here ``HD`` stands for "Horner derivative".
"""
y1 = 0.0
y2 = coeffs[0]
e1 = 0.0 # y1_hat = y1 + e1
e2 = 0.0 # y2_hat = y2 + e2
for coeff in coeffs[1:-1]:
# Update ``y1`` and ``e1``.
prod, pi = eft.multiply_eft(x, y1)
y1, sigma = eft.add_eft(prod, y2)
e1 = x * e1 + e2 + (pi + sigma)
# Update ``y2`` and ``e2``.
prod, pi = eft.multiply_eft(x, y2)
y2, sigma = eft.add_eft(prod, coeff)
e2 = x * e2 + (pi + sigma)
# Perform one last update of ``y1`` and ``e1``.
prod, pi = eft.multiply_eft(x, y1)
y1, sigma = eft.add_eft(prod, y2)
e1 = x * e1 + e2 + (pi + sigma)
# Return the compensated form of ``y1``.
return y1 + e1
def compensated_residual(s, coeffs1, t, coeffs2):
x1, dx1 = de_casteljau._compensated_k(s, coeffs1[0, :], 2)
y1, dy1 = de_casteljau._compensated_k(s, coeffs1[1, :], 2)
x2, dx2 = de_casteljau._compensated_k(t, coeffs2[0, :], 2)
y2, dy2 = de_casteljau._compensated_k(t, coeffs2[1, :], 2)
dx, sigma = eft.add_eft(x1, -x2)
tau = (dx1 - dx2) + sigma
dx += tau
dy, sigma = eft.add_eft(y1, -y2)
tau = (dy1 - dy2) + sigma
dy += tau
return np.array([[dx], [dy]])
def local_error_eft(errors, rho, delta_b):
r"""Perform an error-free transformation for computing :math:`\ell`.
This assumes, but does not check, that there are at least two
``errors``.
"""
num_errs = len(errors)
new_errors = [None] * (num_errs + 1)
l_hat, new_errors[0] = eft.add_eft(errors[0], errors[1])
for j in range(2, num_errs):
l_hat, new_errors[j - 1] = eft.add_eft(l_hat, errors[j])
prod, new_errors[num_errs - 1] = eft.multiply_eft(rho, delta_b)
l_hat, new_errors[num_errs] = eft.add_eft(l_hat, prod)
return new_errors, l_hat
def count_add_eft():
parent = operation_count.Computation()
val1 = operation_count.Float(1.5, parent)
val2 = operation_count.Float(0.5 + 0.5**52, parent)
sum_, error = eft.add_eft(val1, val2)
assert sum_.value == 2.0
assert error.value == 0.5**52
assert parent.count == 6
print(" add_eft(): {}".format(parent.display))
def pre_compensated_derivative(coeffs):
degree = len(coeffs) - 1
pk = []
err_k = []
for k in range(degree):
delta_b, sigma = eft.add_eft(coeffs[k + 1], -coeffs[k])
pk.append(delta_b)
err_k.append(sigma)
return pk, err_k
def _compensated_derivative(s, pk, err_k):
r, rho = eft.add_eft(1.0, -s)
degree = len(pk)
for k in range(1, degree):
new_pk = []
new_err_k = []
for j in range(degree - k):
# new_pk.append(r * pk[j] + s * pk[j + 1])
prod1, d_pi1 = eft.multiply_eft(pk[j], r)
prod2, d_pi2 = eft.multiply_eft(pk[j + 1], s)
new_dp, d_sigma = eft.add_eft(prod1, prod2)
new_pk.append(new_dp)
d_ell = d_pi1 + d_pi2 + d_sigma + pk[j] * rho
new_err = d_ell + s * err_k[j + 1] + r * err_k[j]
new_err_k.append(new_err)
# Update the "current" values.
pk = new_pk
err_k = new_err_k
return degree, pk[0], err_k[0]
def _compensated(x, coeffs):
if not coeffs:
return 0.0, [], []
p = coeffs[0]
e_pi = []
e_sigma = []
for coeff in coeffs[1:]:
prod, e1 = eft.multiply_eft(p, x)
p, e2 = eft.add_eft(prod, coeff)
e_pi.append(e1)
e_sigma.append(e2)
return p, e_pi, e_sigma
def _compensated_k(s, coeffs, K):
r"""Performs a K-compensated de Casteljau.
.. _JLCS10: https://doi.org/10.1016/j.camwa.2010.05.021
Note that the order of operations exactly matches the `JLCS10`_ paper.
For example, :math:`\widehat{\partial b}_j^{(k)}` is computed as
.. math::
\widehat{ell}_{1, j}^{(k)} \oplus \left(s \otimes
\widehat{\partial b}_{j + 1}^{(k + 1)}\right) \oplus
\left(\widehat{r} \otimes \widehat{\partial b}_j^{(k + 1)}\right)
instead of "typical" order
.. math::
\left(\widehat{r} \otimes \widehat{\partial b}_j^{(k + 1)}\right)
\oplus \left(s \otimes \widehat{\partial b}_{j + 1}^{(k + 1)}
\right) \oplus \widehat{ell}_{1, j}^{(k)}.
This is so that the term
.. math::
\widehat{r} \otimes \widehat{\partial b}_j^{(k + 1)}
only has to be in one sum. We avoid an extra sum because
:math:`\widehat{r}` already has round-off error.
"""
r, rho = eft.add_eft(1.0, -s)
degree = len(coeffs) - 1
bk = {0: list(coeffs)}
# NOTE: This will be shared, but is read only.
all_zero = (0.0, ) * (degree + 1)
for F in range(1, K - 1 + 1):
bk[F] = all_zero
for k in range(degree):
new_bk = {F: [] for F in range(K - 1 + 1)}
for j in range(degree - k):
# Update the "level 0" stuff.
P1, pi1 = eft.multiply_eft(r, bk[0][j])
P2, pi2 = eft.multiply_eft(s, bk[0][j + 1])
S3, sigma3 = eft.add_eft(P1, P2)
new_bk[0].append(S3)
errors = [pi1, pi2, sigma3]
delta_b = bk[0][j]
for F in range(1, K - 2 + 1):
new_errors, l_hat = local_error_eft(errors, rho, delta_b)
P1, pi1 = eft.multiply_eft(s, bk[F][j + 1])
S2, sigma2 = eft.add_eft(l_hat, P1)
P3, pi3 = eft.multiply_eft(r, bk[F][j])
S, sigma4 = eft.add_eft(S2, P3)
new_bk[F].append(S)
new_errors.extend([pi1, sigma2, pi3, sigma4])
errors = new_errors
delta_b = bk[F][j]
# Update the "level 2" stuff.
l_hat = local_error(errors, rho, delta_b)
new_bk[K - 1].append(l_hat + s * bk[K - 1][j + 1] +
r * bk[K - 1][j])
# Update the "current" values.
bk = new_bk
return tuple(bk[F][0] for F in range(K - 1 + 1))
