Use PEP 757 API for fmpz <-> int conversions

[skirpichev]

Closes #159

This lacks implementation for Python < 3.14. Not sure how to do this better, conditional compilation seems to be deprecated by Cython.

To my surprise, #324 approach seems to be better. Here my benchmarks.

Export

Benchmark	ref	pr324	patch
1<<7	166 ns	169 ns: 1.02x slower	169 ns: 1.02x slower
1<<38	216 ns	not significant	219 ns: 1.01x slower
1<<300	3.06 us	923 ns: 3.31x faster	787 ns: 3.88x faster
1<<3000	15.9 us	2.17 us: 7.34x faster	3.70 us: 4.29x faster
1<<10000	49.4 us	5.93 us: 8.33x faster	11.6 us: 4.27x faster
Geometric mean	(ref)	2.88x faster	2.33x faster

Import

Benchmark	ref	pr324	patch
1<<7	454 ns	451 ns: 1.01x faster	459 ns: 1.01x slower
1<<38	471 ns	466 ns: 1.01x faster	475 ns: 1.01x slower
1<<300	4.80 us	1.91 us: 2.52x faster	1.10 us: 4.37x faster
1<<3000	23.6 us	3.07 us: 7.68x faster	4.17 us: 5.65x faster
1<<10000	73.0 us	6.78 us: 10.76x faster	12.5 us: 5.86x faster
Geometric mean	(ref)	2.92x faster	2.69x faster

Details

# bench-export.py

import os

import pyperf

_T = os.getenv('_T')
if _T == "gmpy2.mpz":
    from gmpy2 import mpz
elif _T == "gmp.mpz":
    from gmp import mpz
else:
    from flint import fmpz as mpz

cases = ['1<<7', '1<<38', '1<<300', '1<<3000', '1<<10000']
runner = pyperf.Runner()
for c in cases:
    i = eval(c)
    m = mpz(i)
    runner.bench_func(c, int, m)

# bench-import.py

import os

import pyperf

_T = os.getenv('_T')
if _T == "gmpy2.mpz":
    from gmpy2 import mpz
elif _T == "gmp.mpz":
    from gmp import mpz
else:
    from flint import fmpz as mpz

cases = ['1<<7', '1<<38', '1<<300', '1<<3000', '1<<10000']
runner = pyperf.Runner()
for c in cases:
    i = eval(c)
    runner.bench_func(c, mpz, i)

[oscarbenjamin]

conditional compilation seems to be deprecated by Cython.

The way to do it is to move the conditional compilation to C e.g.:

python-flint/src/flint/flintlib/types/fmpz_mod_mat_compat.pxd

Lines 4 to 14 in b2b71f4

	cdef extern from *:
	"""
	/*
	* fmpz_mod_mat function signatures were changed in FLINT 3.1.0
	*/
	#if __FLINT_RELEASE >= 30100 /* Flint 3.1.0 or later */
	#define compat_fmpz_mod_mat_init(mat, rows, cols, ctx) fmpz_mod_mat_init(mat, rows, cols, ctx)
	#define compat_fmpz_mod_mat_init_set(mat, src, ctx) fmpz_mod_mat_init_set(mat, src, ctx)
	#define compat_fmpz_mod_mat_clear(mat, ctx) fmpz_mod_mat_clear(mat, ctx)
	#define compat_fmpz_mod_mat_set(A, B, ctx) fmpz_mod_mat_set(A, B, ctx)

[oscarbenjamin]

To my surprise, #324 approach seems to be better.

Is there anything that can improve this either in CPython or GMP?

Or should we just conclude that for python-flint the approach in gh-324 is best?

The approach in gh-324 seems simpler and works for all supported CPython versions without needing any conditional compilation.

What are gmpy2 or python-gmp going to use?