PEP: | 754 |
---|---|
Title: | IEEE 754 Floating Point Special Values |
Version: | 928634d80fc6 |
Last-Modified: | 2008-01-18 08:36:51 +0000 (Fri, 18 Jan 2008) |
Author: | Gregory R. Warnes <gregory_r_warnes at groton.pfizer.com> (Pfizer, Inc.) |
Status: | Rejected |
Type: | Standards Track |
Content-Type: | text/x-rst |
Created: | 28-Mar-2003 |
Python-Version: | 2.3 |
Post-History: |
Contents
This PEP has been rejected. After sitting open for four years, it hasfailed to generate sufficient community interest.
Several ideas of this PEP were implemented for Python 2.6. float('inf')and repr(float('inf')) are now guaranteed to work on every supportedplatform with IEEE 754 semantics. However the eval(repr(float('inf')))roundtrip is still not supported unless you define inf and nan yourself:
>>> inf = float('inf')>>> inf, 1E400(inf, inf)>>> neginf = float('-inf')>>> neginf, -1E400(-inf, -inf)>>> nan = float('nan')>>> nan, inf * 0.(nan, nan)
The math and the sys module also have gained additional features,sys.float_info, math.isinf, math.isnan, math.copysign.
This PEP proposes an API and a provides a reference module thatgenerates and tests for IEEE 754 double-precision special values:positive infinity, negative infinity, and not-a-number (NaN).
The IEEE 754 standard defines a set of binary representations andalgorithmic rules for floating point arithmetic. Included in thestandard is a set of constants for representing special values,including positive infinity, negative infinity, and indeterminate ornon-numeric results (NaN). Most modern CPUs implement theIEEE 754 standard, including the (Ultra)SPARC, PowerPC, and x86processor series.
Currently, the handling of IEEE 754 special values in Python dependson the underlying C library. Unfortunately, there is littleconsistency between C libraries in how or whether these values arehandled. For instance, on some systems "float('Inf')" will properlyreturn the IEEE 754 constant for positive infinity. On many systems,however, this expression will instead generate an error message.
The output string representation for an IEEE 754 special value alsovaries by platform. For example, the expression "float(1e3000)",which is large enough to generate an overflow, should return a stringrepresentation corresponding to IEEE 754 positive infinity. Python2.1.3 on x86 Debian Linux returns "inf". On Sparc Solaris 8 withPython 2.2.1, this same expression returns "Infinity", and onMS-Windows 2000 with Active Python 2.2.1, it returns "1.#INF".
Adding to the confusion, some platforms generate one string onconversion from floating point and accept a different string forconversion to floating point. On these systems
float(str(x))
will generate an error when "x" is an IEEE special value.
In the past, some have recommended that programmers use expressionslike:
PosInf = 1e300**2NaN = PosInf/PosInf
to obtain positive infinity and not-a-number constants. However, thefirst expression generates an error on current Python interpreters. Apossible alternative is to use:
PosInf = 1e300000NaN = PosInf/PosInf
While this does not generate an error with current Pythoninterpreters, it is still an ugly and potentially non-portable hack.In addition, defining NaN in this way does solve the problem ofdetecting such values. First, the IEEE 754 standard provides for anentire set of constant values for Not-a-Number. Second, the standardrequires that
NaN != X
for all possible values of X, including NaN. As a consequence
NaN == NaN
should always evaluate to false. However, this behavior also is notconsistently implemented. [e.g. Cygwin Python 2.2.2]
Due to the many platform and library inconsistencies in handling IEEEspecial values, it is impossible to consistently set or detect IEEE754 floating point values in normal Python code without resorting todirectly manipulating bit-patterns.
This PEP proposes a standard Python API and provides a referencemodule implementation which allows for consistent handling of IEEE 754special values on all supported platforms.
(Run under Python 2.2.1 on Solaris 8.)
>>> import fpconst>>> val = 1e30000 # should be cause overflow and result in "Inf">>> valInfinity>>> fpconst.isInf(val)1>>> fpconst.PosInfInfinity>>> nval = val/val # should result in NaN>>> nvalNaN>>> fpconst.isNaN(nval)1>>> fpconst.isNaN(val)0
The reference implementation is provided in the module "fpconst" [1],which is written in pure Python by taking advantage of the "struct"standard module to directly set or test for the bit patterns thatdefine IEEE 754 special values. Care has been taken to generateproper results on both big-endian and little-endian machines. Thecurrent implementation is pure Python, but some efficiency could begained by translating the core routines into C.
Patch 1151323 "New fpconst module" [2] on SourceForge adds thefpconst module to the Python standard library.
See http://babbage.cs.qc.edu/courses/cs341/IEEE-754references.html forreference material on the IEEE 754 floating point standard.
[1] | Further information on the reference package is available at http://research.warnes.net/projects/rzope/fpconst/ |
[2] | http://sourceforge.net/tracker/?func=detail&aid=1151323&group_id=5470&atid=305470 |
This document has been placed in the public domain.
联系客服