1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
|
"""
Machine arithmetics - determine the parameters of the
floating-point arithmetic system
"""
# Author: Pearu Peterson, September 2003
__all__ = ['MachAr']
from numpy.core.fromnumeric import any
from numpy.core.numeric import seterr
# Need to speed this up...especially for longfloat
class MachAr(object):
"""Diagnosing machine parameters.
The following attributes are available:
ibeta - radix in which numbers are represented
it - number of base-ibeta digits in the floating point mantissa M
machep - exponent of the smallest (most negative) power of ibeta that,
added to 1.0,
gives something different from 1.0
eps - floating-point number beta**machep (floating point precision)
negep - exponent of the smallest power of ibeta that, substracted
from 1.0, gives something different from 1.0
epsneg - floating-point number beta**negep
iexp - number of bits in the exponent (including its sign and bias)
minexp - smallest (most negative) power of ibeta consistent with there
being no leading zeros in the mantissa
xmin - floating point number beta**minexp (the smallest (in
magnitude) usable floating value)
maxexp - smallest (positive) power of ibeta that causes overflow
xmax - (1-epsneg)* beta**maxexp (the largest (in magnitude)
usable floating value)
irnd - in range(6), information on what kind of rounding is done
in addition, and on how underflow is handled
ngrd - number of 'guard digits' used when truncating the product
of two mantissas to fit the representation
epsilon - same as eps
tiny - same as xmin
huge - same as xmax
precision - int(-log10(eps))
resolution - 10**(-precision)
Reference:
Numerical Recipies.
"""
def __init__(self, float_conv=float,int_conv=int,
float_to_float=float,
float_to_str = lambda v:'%24.16e' % v,
title = 'Python floating point number'):
"""
float_conv - convert integer to float (array)
int_conv - convert float (array) to integer
float_to_float - convert float array to float
float_to_str - convert array float to str
title - description of used floating point numbers
"""
# We ignore all errors here because we are purposely triggering
# underflow to detect the properties of the runninng arch.
saverrstate = seterr(under='ignore')
try:
self._do_init(float_conv, int_conv, float_to_float, float_to_str, title)
finally:
seterr(**saverrstate)
def _do_init(self, float_conv, int_conv, float_to_float, float_to_str, title):
max_iterN = 10000
msg = "Did not converge after %d tries with %s"
one = float_conv(1)
two = one + one
zero = one - one
# Do we really need to do this? Aren't they 2 and 2.0?
# Determine ibeta and beta
a = one
for _ in xrange(max_iterN):
a = a + a
temp = a + one
temp1 = temp - a
if any(temp1 - one != zero):
break
else:
raise RuntimeError, msg % (_, one.dtype)
b = one
for _ in xrange(max_iterN):
b = b + b
temp = a + b
itemp = int_conv(temp-a)
if any(itemp != 0):
break
else:
raise RuntimeError, msg % (_, one.dtype)
ibeta = itemp
beta = float_conv(ibeta)
# Determine it and irnd
it = -1
b = one
for _ in xrange(max_iterN):
it = it + 1
b = b * beta
temp = b + one
temp1 = temp - b
if any(temp1 - one != zero):
break
else:
raise RuntimeError, msg % (_, one.dtype)
betah = beta / two
a = one
for _ in xrange(max_iterN):
a = a + a
temp = a + one
temp1 = temp - a
if any(temp1 - one != zero):
break
else:
raise RuntimeError, msg % (_, one.dtype)
temp = a + betah
irnd = 0
if any(temp-a != zero):
irnd = 1
tempa = a + beta
temp = tempa + betah
if irnd==0 and any(temp-tempa != zero):
irnd = 2
# Determine negep and epsneg
negep = it + 3
betain = one / beta
a = one
for i in range(negep):
a = a * betain
b = a
for _ in xrange(max_iterN):
temp = one - a
if any(temp-one != zero):
break
a = a * beta
negep = negep - 1
# Prevent infinite loop on PPC with gcc 4.0:
if negep < 0:
raise RuntimeError, "could not determine machine tolerance " \
"for 'negep', locals() -> %s" % (locals())
else:
raise RuntimeError, msg % (_, one.dtype)
negep = -negep
epsneg = a
# Determine machep and eps
machep = - it - 3
a = b
for _ in xrange(max_iterN):
temp = one + a
if any(temp-one != zero):
break
a = a * beta
machep = machep + 1
else:
raise RuntimeError, msg % (_, one.dtype)
eps = a
# Determine ngrd
ngrd = 0
temp = one + eps
if irnd==0 and any(temp*one - one != zero):
ngrd = 1
# Determine iexp
i = 0
k = 1
z = betain
t = one + eps
nxres = 0
for _ in xrange(max_iterN):
y = z
z = y*y
a = z*one # Check here for underflow
temp = z*t
if any(a+a == zero) or any(abs(z)>=y):
break
temp1 = temp * betain
if any(temp1*beta == z):
break
i = i + 1
k = k + k
else:
raise RuntimeError, msg % (_, one.dtype)
if ibeta != 10:
iexp = i + 1
mx = k + k
else:
iexp = 2
iz = ibeta
while k >= iz:
iz = iz * ibeta
iexp = iexp + 1
mx = iz + iz - 1
# Determine minexp and xmin
for _ in xrange(max_iterN):
xmin = y
y = y * betain
a = y * one
temp = y * t
if any(a+a != zero) and any(abs(y) < xmin):
k = k + 1
temp1 = temp * betain
if any(temp1*beta == y) and any(temp != y):
nxres = 3
xmin = y
break
else:
break
else:
raise RuntimeError, msg % (_, one.dtype)
minexp = -k
# Determine maxexp, xmax
if mx <= k + k - 3 and ibeta != 10:
mx = mx + mx
iexp = iexp + 1
maxexp = mx + minexp
irnd = irnd + nxres
if irnd >= 2:
maxexp = maxexp - 2
i = maxexp + minexp
if ibeta == 2 and not i:
maxexp = maxexp - 1
if i > 20:
maxexp = maxexp - 1
if any(a != y):
maxexp = maxexp - 2
xmax = one - epsneg
if any(xmax*one != xmax):
xmax = one - beta*epsneg
xmax = xmax / (xmin*beta*beta*beta)
i = maxexp + minexp + 3
for j in range(i):
if ibeta==2:
xmax = xmax + xmax
else:
xmax = xmax * beta
self.ibeta = ibeta
self.it = it
self.negep = negep
self.epsneg = float_to_float(epsneg)
self._str_epsneg = float_to_str(epsneg)
self.machep = machep
self.eps = float_to_float(eps)
self._str_eps = float_to_str(eps)
self.ngrd = ngrd
self.iexp = iexp
self.minexp = minexp
self.xmin = float_to_float(xmin)
self._str_xmin = float_to_str(xmin)
self.maxexp = maxexp
self.xmax = float_to_float(xmax)
self._str_xmax = float_to_str(xmax)
self.irnd = irnd
self.title = title
# Commonly used parameters
self.epsilon = self.eps
self.tiny = self.xmin
self.huge = self.xmax
import math
self.precision = int(-math.log10(float_to_float(self.eps)))
ten = two + two + two + two + two
resolution = ten ** (-self.precision)
self.resolution = float_to_float(resolution)
self._str_resolution = float_to_str(resolution)
def __str__(self):
return '''\
Machine parameters for %(title)s
---------------------------------------------------------------------
ibeta=%(ibeta)s it=%(it)s iexp=%(iexp)s ngrd=%(ngrd)s irnd=%(irnd)s
machep=%(machep)s eps=%(_str_eps)s (beta**machep == epsilon)
negep =%(negep)s epsneg=%(_str_epsneg)s (beta**epsneg)
minexp=%(minexp)s xmin=%(_str_xmin)s (beta**minexp == tiny)
maxexp=%(maxexp)s xmax=%(_str_xmax)s ((1-epsneg)*beta**maxexp == huge)
---------------------------------------------------------------------
''' % self.__dict__
if __name__ == '__main__':
print MachAr()
|