def run(self, xrange="log", diff="relative"):
r"""
Compare accuracy of different methods for computing f.
*xrange* is::
log: [10^-3,10^5]
logq: [10^-4, 10^1]
linear: [1,1000]
zoom: [1000,1010]
neg: [-100,100]
*diff* is "relative", "absolute" or "none"
*x_bits* is the precision with which the x values are specified. The
default 23 should reproduce the equivalent of a single precisio
"""
linear = not xrange.startswith("log")
if xrange == "zoom":
lin_min, lin_max, lin_steps = 1000, 1010, 2000
elif xrange == "neg":
lin_min, lin_max, lin_steps = -100.1, 100.1, 2000
elif xrange == "linear":
lin_min, lin_max, lin_steps = 1, 1000, 2000
lin_min, lin_max, lin_steps = 0.001, 2, 2000
elif xrange == "log":
log_min, log_max, log_steps = -3, 5, 400
elif xrange == "logq":
log_min, log_max, log_steps = -4, 1, 400
else:
raise ValueError("unknown range "+xrange)
with mp.workprec(500):
# Note: we make sure that we are comparing apples to apples...
# The x points are set using single precision so that we are
# examining the accuracy of the transformation from x to f(x)
# rather than x to f(nearest(x)) where nearest(x) is the nearest
# value to x in the given precision.
if linear:
lin_min = max(lin_min, self.limits[0])
lin_max = min(lin_max, self.limits[1])
qrf = np.linspace(lin_min, lin_max, lin_steps, dtype='single')
#qrf = np.linspace(lin_min, lin_max, lin_steps, dtype='double')
qr = [mp.mpf(float(v)) for v in qrf]
#qr = mp.linspace(lin_min, lin_max, lin_steps)
else:
log_min = np.log10(max(10**log_min, self.limits[0]))
log_max = np.log10(min(10**log_max, self.limits[1]))
qrf = np.logspace(log_min, log_max, log_steps, dtype='single')
#qrf = np.logspace(log_min, log_max, log_steps, dtype='double')
qr = [mp.mpf(float(v)) for v in qrf]
#qr = [10**v for v in mp.linspace(log_min, log_max, log_steps)]
target = self.call_mpmath(qr, bits=500)
pylab.subplot(121)
self.compare(qr, 'single', target, linear, diff)
pylab.legend(loc='best')
pylab.subplot(122)
self.compare(qr, 'double', target, linear, diff)
pylab.legend(loc='best')
pylab.suptitle(self.name + " compared to 500-bit mpmath")
def ieee_eval(x):
"""
Return the string representation of a floating point value in each prec.
"""
x = mpf(x)
sreps = list()
for prec in ieee_formats:
with mp.workprec(prec):
sreps.append(str(x))
return sreps
def workfloat(bytes_or_bits):
ok = False
bytes_or_bits = int(bytes_or_bits)
for fmt in ieee_formats.values():
if bytes_or_bits in (fmt.bits, fmt.bits // 8):
ok = True
with mp.workprec(fmt.prec):
yield
if not ok:
raise KeyError(bytes_or_bits)
def dist(self, f1, f2):
"""
Return the distance between two values in ULPs.
"""
with mp.workprec(self.prec):
lo = mpf(min(f1, f2))
hi = mpf(max(f1, f2))
dist = 0
while lo != hi:
lo += self.ulp(lo)
dist += 1
return dist
def ulp(self, x):
"""
Return the unit of least precision for a value x.
This is the floating point value which represents the delta between
x and the closest representible number.
"""
from .util import zero, one, two
x = mpf(x)
n = 0
min_exp = abs(self.min_exp(denorm=True))
with mp.workprec(self.prec):
ulp = next_power(x)
while (x + ulp) != x and ulp != zero and n < min_exp:
ulp /= two
n += 1
return ulp * two
def call_mpmath(self, vec, bits=500):
"""
Direct calculation using mpmath extended precision library.
"""
with mp.workprec(bits):
return [self.mp_function(mp.mpf(x)) for x in vec]
def mp_fn(vec, bits=500):
"""
Direct calculation using sympy multiprecision library.
"""
with mp.workprec(bits):
return [_mp_fn(mp.mpf(x)) for x in vec]
