GMPY_CFFI mpz speed

gmpy_cffi is slow

Today I released version 0.1 of gmpy_cffi, a PyPy compatible implementation of gmpy. The problem is that my implementation is quite a bit slower.

gmpy_mpz.png

As you can see, the time taken to initialize an mpz instance is about 10x slower under gmpy_cffi when compared to gmpy.

Explanation

The gmpy mpz initialisation code uses the CPython internals to access the raw bits storing the value of a python integer (or long). In comparison, gmpy_cffi converts the number to a hex string and then uses mpz_set_str with base 16 to set the GMP mpz value

def _pyint_to_mpz(n, a):
    """
    Set `a` from `n`.
    :type n: int,long
    :type a: mpz_t
    """
    if -sys.maxsize - 1 <= n <= sys.maxsize:
        gmp.mpz_set_si(a, n)
    elif sys.maxsize < n <= MAX_UI:
        gmp.mpz_set_ui(a, n)
    else:
        gmp.mpz_set_str(a, hex(n).rstrip('L').encode('UTF-8'), 0)

I've tried other methods such as using bit twiddling to break the python int into an array of C ints and then using mpz_import, but it was slower than using hex

def _pyint_to_mpz(n, a):
    """
    Set `a` from `n`.
    :type n: int,long
    :type a: mpz_t
    """
    neg = n < 0
    n = abs(n)
    tmp = array.array('L')
    size = tmp.itemsize
    numb = (8 * size)
    mask = ~(~0 << numb)
    while n:
        v = n & mask
        n = n >> numb
        tmp.append(v)
    addr, count = tmp.buffer_info()
    gmp.mpz_import(a, count, -1, size, 0, 0, ffi.cast('void *', addr))
    if neg:
        gmp.mpz_neg(a, a)

Conclusion

The conversion from python long to gmp mpz is by far the main reason why initializing mpzs in gmpy_cffi is so slow, but there isn't really an obvious way to improve on this.

Perhaps cffi could provide a fast method for converting python longs to an array of C ints, but I doubt there's a huge demand for that right now.

We're within a factor of 10 of evil low level C hackery. For now that will have to do.

social