Missing methods on mpz c.f. int about gmpy HOT 11 CLOSED

It will be awesome, I currently cast mpz to python int to get bytes: int(x).to_bytes(32,'big') and it is pretty slow...

from gmpy.

from gmpy.

BTW, here some simple benchmarks (including my draft patch for this bug):

$ python -m timeit -s 'from math import factorial' -s 'a = factorial(57)' 'a.to_bytes(32)'
1000000 loops, best of 5: 281 nsec per loop
$ python -m timeit -s 'from gmpy2 import fac' -s 'a = fac(57)' 'a.to_bytes(32)'
500000 loops, best of 5: 772 nsec per loop
$ python -m timeit -s 'from gmpy2 import fac' -s 'a = fac(57)' 'int(a).to_bytes(32)'
500000 loops, best of 5: 863 nsec per loop
$ python -m timeit -s 'from gmpy2 import fac, to_binary' -s 'a = fac(57)' 'to_binary(a)'
500000 loops, best of 5: 444 nsec per loop

from gmpy.

For clarification, to_binary converts all gmpy2 types such that they can be recovered by from_binary. It should be (somewhat) platform neutral. 64-bit systems should be able to read binary blobs created on 32-bit system and vice-versa.

I'll review the pull request.

from gmpy.

@skirpichev, yes but to_bytes() and to_binary() have different meaning.
to_binary() takes care of a portable import/export and so adds a header. It needs a lot of operations to get a result similar to the python int.to_bytes()

Why mpz.to_bytes() is so slow vs int.to_bytes() ?

from gmpy.

Why mpz.to_bytes() is so slow vs int.to_bytes() ?

For above tests it seems to be connected with the argument parsing and (?) METH_FASTCALL calling convention for int.to_bytes (I'll try to build the gmpy2 with this patch for the v3.6 of CPython to verify that.).

from gmpy.

Here is v3.6 on same host:

$ python -m timeit -s 'from math import factorial' -s 'a = factorial(57)' 'a.to_bytes(32, "big")'
1000000 loops, best of 3: 0.776 usec per loop

from gmpy.

Wow, METH_FASTCALL worth it!

from gmpy.

Some comments on METH_FASTCALL. I converted _mpmath_normalize to use the vectorcall protocol. _mpmath_normalize requires exactly 6 positional arguments. There are no keyword arguments. METH_FASTCALL decreased the function call overhear by ~30ns.

to_bytes/from_bytes take keyword arguments so the savings may be better. I plan to do some experimentation this weekend.

from gmpy.

from gmpy.

Here is an initial version with METH_FASTCALL:
https://github.com/skirpichev/gmpy/tree/fix-357

$ python -m timeit -s 'from math import factorial' -s 'a = factorial(57)' 'a.to_bytes(32)'
1000000 loops, best of 5: 276 nsec per loop
$ python -m timeit -s 'from gmpy2 import fac' -s 'a = fac(57)' 'a.to_bytes(32)'
500000 loops, best of 5: 485 nsec per loop
$ python -m timeit -s 'from gmpy2 import fac' -s 'a = fac(57)' 'int(a).to_bytes(32)'
500000 loops, best of 5: 810 nsec per loop

from gmpy.