Srajan Garg Computer Scientist in the making

Flint Polynomials


This week has not been as interesting as the last. Most of my work this week involved wrapping Flint Polynomials, for use in SymEngine. I will discuss their internal implementation and how they were integrated to the library.

Flint Polynomials

Flint’s integer polynomial class is the fmpz_poly class. Flint, being a C library also provides a C++ wrapper for most of it’s classes. In our case, it’s the fmpz_polyxx class. As discussed last week, the arithmetic operators for the class have already been overloaded and it can be seamlessly used with the base class which is already set up.

Internally, flint uses a dense representation of polynomials. What this means is that the coefficient for each degree in the polynomial are stored in a vector. This is in contrast to a sparse approach which deals stores only non-zero coefficients of a polynomial in a dictionary. Flint also uses it’s own integer class as the coefficient type. Flint’s integers are basically similar to GMP integers, and also follow the same naming conventions. (eg. fmpz_t is analogous to mpz_t) They have internal optimizations to make them faster for use within the flint library itself. Flint does not provide the option to use a custom integer class as the type of coefficients like Piranha does.

All the work related to flint polynomials can be seen in #971. The majority of the PR concerns introducing a new class UIntPolyFlint and all the necessary functionality like __hash__, eval and compare. Test cases have also been added, to see if the polynomial manipulation is working as intended. There is some additional minor work done here apart from this class introduction.

The polynomial base class now has three methods :

static Poly from_container()
static Poly from_dict()
static Poly from_vec()

These methods are, as indicated from their names, used to construct the polynomial from any of the three sources. So, for example to construct a UIntPolyFlint, you can :

PolyFlint a = PolyFlint::from_container(fmpz_polyxx A);
PolyFlint b = PolyFlint::from_vec(vector<integer_class> B);
PolyFlint a = PolyFlint::from_dict(map<uint, integer_class> C);

integer_class Conversions

Another issue was that methods like eval must return integers of integer_class. Thus, we needed functions to convert fmpzxx to any of the 4 possible integer classes used by SymEngine. The 4 integers used by SymEngine are :

1. mpz_class         The default C++ class provided by GMP
2. mpz_wrapper       SymEngine's wrapper around GMP's mpz_t
3. piranha::integer  Piranha's integer class
3. fmpz_wrapper      SymEngine's wrapper around Flint's fmpz_t

We, might even require such conversion for Piranha polynomials. So, I generalized this and made to_integer_class() which takes in either flint::fmpzxx or piranha::integer and returns an integer_class. There was a small discussion on where these functions should be placed, with Isuru which was insightful. Basically, what I was doing earlier caused a considerable increase in compilation time, which was later corrected. An example of one such function is :

inline integer_class to_integer_class(const flint::fmpzxx &i)
    integer_class x;
    fmpz_get_mpz(x.get_mpz_t(), i._data().inner);
    return x;

A get_coeff() function was also added to the polynomial base is a much needed function for any polynomial manipulation.

Miscellaneous Work

  • There were some changes done to the polynomial structure before the flint polynomials were introduced. Constructor of the polynomials from vectors was removed, as it wasn’t required. And a from_vec was added for the SymEngine polynomial dictionaries UIntDict and UExprDict. The changes can be seen here #965.

  • In Travis CI builds, SymEngine was using flint-2.4.4 from a static location. There’s a bug that I encountered using this version, which dealt with incorrect header declaration in one of the flint files. Now, we clone the flint repository from wbhart/flint2 to use in the CI builds. The change is in #973.

I’m working on wrapping Piranha polynomials next week. After that is done, I plan to start getting higher level functionality like gcd and factorize ready.