Last updated on: 2008, June 22; for GMPY release: 1.03 Go to Google Code gmpy page

GMPY Project goals and strategies

The General Multiprecision PYthon project (GMPY) focuses on Python-usable modules providing multiprecision arithmetic functionality to Python programmers. The project mission includes both C and C++ Python-modules (for speed) and pure Python modules (for flexibility and convenience); it potentially includes integral, rational and floating-point arithmetic in any base. Only cross-platform functionality is of interest, at least for now.

As there are many good existing free C and C++ libraries that address these issues, it is expected that most of the work of the GMPY project will involve wrapping, and exposing to Python, exactly these existing libraries (possibly with additional "convenience" wrappers written in Python itself). For starters, we've focused on the popular (and excellent) GNU Multiple Precision library, GMP, exposing its functionality through module gmpy.

The GMPY Module

Python modules older than GMPY exposes a subset of the integral-MP (MPZ) functionality of earlier releases of the GMP library. The first GMPY goal (currently nearly reached) was to develop the gmpy module into a complete exposure of MPZ, MPF (floating-point), and MPQ (rational) functionality of current GMP (release 4), as well as auxiliary functionality such as random number generation, with full support for current Python releases (2.3 and up) and the Python 'distutils' (and also support for a "C API" allowing some level of interoperation with other C-written extension modules for Python).

Note: the module's ability to be used as a "drop-in replacement" for Python's own implementation of longs, to rebuild Python from sources in a version using GMP, was a characteristic of the gmp-module we started from, but is not a target of the gmpy project, and we have no plans to support it.

This first GMPY module is called gmpy, just like the whole project.

The extended MP floating-point facilities of MPFR will later also be considered for inclusion in gmpy (either within the same module, or as a further, separate add-on module). [[ Rooting for MPFR to be merged with GMP so we can avoid some awkwardness (but seeing no movement on this front so far) ]].

Mutability... but not for now

Early tests have shown that supporting Python 2's "in-place operation" functionality (by making MPZ, MPF and MPQ Python objects mutable) would offer a substantial performance boost.

Despite this, widespread feeling among Python cognoscenti appears to be against exposing such "mutable numbers". As a consequence, our current aim is for a first release of GMPY without mutability, to be followed at some later time by one which will also fully support in-place-mutable versions of number objects (as well as the default immutable ones), but only when explicitly and deliberately requested by a user (who can then be presumed to know what he or she is doing). Meanwhile, caching strategies are used to ameliorate performance issues, and appear to be reasonably worthwhile (so far, only MPZ and MPQ objects are subject to this caching).

We've tended to solve other debatable design issues in a similar vein, i.e., by trying to work "like Python's built-in numbers" when there was a choice and two or more alternatives made sense.

Project Status and near-future plans

The gmpy module's current release (latest current release as of 2008/06/22: 1.03) is available for download in both source and Windows-binary form, and Mac-binary form too. gmpy 1.03 exposes all of the mpz, mpq and mpf functionality that was already available in GMP 3.1, and most of the random-number generation functionality (there are no current plans to extend gmpy to expose other such functionality, although the currently experimental way in which it is architected is subject to possible future changes).

On most platforms, you will need to separately procure and install the GMP library itself to be able to build and use GMPY. Note that 4.0.1 or better is needed; take care: some Linux releases come bundled with older GMP versions, such as GMP 3, and you may have to install the latest GMP version instead -- beware also of /usr/lib vs /usr/local/lib issues.

Please read the file "windows_build.txt" for detailed instructions on compiling GMP and GMPY using the freely available MinGW tools.

[[ OLD: The exception to this need is under (32-bit) Windows, where binary-accompanied releases are the norm, and builds of GMP usable with MS VC++ 6 (the main C compiler used for Python on this platform) are traditionally hard to come by.

We started the GMPY project using a VC++ port of GMP.LIB "just found on the net", and later a port by Jorgen Lundman, but are currently relying on other volunteers to build Windows binaries since we don't have any Windows machine any more. This site does appear to offer all needed files and instructions for Windows developers who want to re-build the gmpy module from sources; the gmpy project itself just supplies a separate 'binary module' package is supplied, containing only the pre-built GMPY.PYD, for those who do not want to re-build from sources. ]]

Do note, however, that all gmpy users should download the gmpy source-package as well, as currently that is the one including gmpy documentation and unit-tests!

Currently-open issues

A still-weakish point is with the output-formatting of mpf numbers; sometimes, this formatting ends up providing a few more digits than a given number's accuracy would actually warrant (a few noise digits after a long string of trailing '0' or '9' digits), particularly when the mpf number is built from a Python float -- the accuracy situation is quite a bit better when the mpf number is built from a string.

Because of this, since release 0.6, gmpy introduced an optional 'floating-conversion format string' module-level setting: if present, float->mpf conversion goes through an intermediate formatted string (by default, it still proceeds directly, at least for now); this does ameliorate things a bit, as does the better tracking done (since 0.6, with further enhancements in 0.7) of the 'requested' precision for an mpf (as opposed to the precision the underlying GMP actually 'assigns' to it); but the issue cannot yet be considered fully solved, and may well still need some design changes in the output formatting functionality.

Unit tests are not considered a weak point any more; the over 1000 unit-tests now being run provide a decent cover of 93+% SLOC for gmpy.c, up from 72% in 0.7. The non-covered SLOCs (about 370 of gmpy.c's current 5820) are mostly disaster-tests to handle out-of-memory situations, code providing optional debugging output (for debugging the gmpy module -- see the set_debug facility), a smattering of 'defensive programming' cases (to handle situations that 'should never happen, but'...) and some cases of the new experimental 'callbacks' facility (mostly provided for the specific use of PySymbolic, and intended to be tested by that package). We'll have to do better, eventually, but, for now, this can be considered OK.

In the attempt to make gmpy as useful as can be for both stand-alone use, and also in the context of PySymbolic, a tad too many design decisions have been delayed/postponed by introducing module-level flags, letting us 'have it both ways' in the current beta gmpy; this has produced a somewhat unwieldy mix of module-level flag-setting and call-back functions. This whole area's architecture will neet to be revisited, including other such design-decisions yet.

Near-future plans

Future releases may have changes including: re-architecting the module-level setting functions; more elegantly formatted documentation; more timing-measurement scripts and usage-examples. Some of the currently experimental 'callbacks' will also be removed, having been proven unnecessary. All relevant GMP 4 functionality will be exposed.

No predictions on timing, though. gmpy 1.03 meets all current needs of the main author, so his motivation to work more on it is low:-). So, don't hold your breath (pitching in and helping it happen, on the other hand, _might_ be advisable, and will surely yield results:-).

Project page on Go to Google Code gmpy page