Created on 2014-09-24 20:40 by scoder, last changed 2020-01-16 10:02 by vstinner. This issue is now closed.
fractions.gcd() is required for normalising numerator and denominator of the Fraction data type. Some speed improvements were applied to Fraction in issue 22464, now the gcd() function takes up about half of the instantiation time in the benchmark in issue 22458, which makes it quite a heavy part of the overall Fraction computation time. The current implementation is def gcd(a, b): while b: a, b = b, a%b return a Reimplementing it in C would provide for much faster calculations. Here is a Cython version that simply drops the calculation loop into C as soon as the numbers are small enough to fit into a C long long int: def _gcd(a, b): # Try doing all computation in C space. If the numbers are too large # at the beginning, retry until they are small enough. cdef long long ai, bi while b: try: ai, bi = a, b except OverflowError: pass else: # switch to C loop while bi: ai, bi = bi, ai%bi return ai a, b = b, a%b return a It's substantially faster already because the values will either be small enough right from the start or quickly become so after a few iterations with Python objects. Further improvements should be possible with a dedicated PyLong implementation based on Lehmer's GCD algorithm: https://en.wikipedia.org/wiki/Lehmer_GCD_algorithm
In case it's useful, see issue #1682 for my earlier Lehmer gcd implementation. At the time, that approach was dropped as being premature optimisation.
If Python grows an optimized implementation, how about exposing it in the math module?
> If Python grows an optimized implementation, how about exposing it in the math module? +1.
That's what the patch does anyway. +1
Hmm... which patch?
http://bugs.python.org/file9486/lehmer_gcd.patch (see #1682)
Here is updated Mark's patch from issue1682. It is ported to 3.5, slightly simplified and optimized (I did not touched the main algorithm still), utilized in the fractions module, added tests and documentation. It speeds up Stefan's fractions benchmark about 20%.
The problem is that this changes the behaviour of fractions.gcd() w.r.t. negative numbers. It's a public function, so we should keep it for backwards compatibility reasons, *especially* when adding a new function in the math module.
Oh, and thanks for working on it, Serhiy! :)
see issue22477 for a discussion of whether the behavior of fractions.gcd should be changed or not
sorry, forgot to format the link: issue<22477>
The thing is, if we add something new in a substantially more exposed place (the math module), then why break legacy code *in addition*? Just leaving it the way it is won't harm anyone, really.
I wasn't arguing for or against anything, just providing a link to the relevant discussion.
Well, here is a patch which keeps the same weird behavior of fractions.gcd().
I am inclined to think that a maths.gcd() makes sense as this would be where I would go first to find this function. And the prospect of better performance is attractive since the gcd is an important operation in work with number theory algorithms. Would it co-exist with fractions.gcd(), with identical semantics? Or would it co-exist with fractions.gcd(), with the 'less surprising' semantics that are under discussion in the 'GCD in Fractions' thread? Would it take on the suggestion of operating on one or more input parameters?
> Or would it co-exist with fractions.gcd(), with the 'less surprising' > semantics that are under discussion in the 'GCD in Fractions' thread? Yes, exactly. math.gcd will always give a nonnegative result. The output of fractions.gcd remains unchanged for integer inputs, for backwards compatibility.
Serhiy: thank you! I've been meaning to update that patch for a long time, but hadn't found the courage or time to face the inevitable bitrot.
Now I spent more time on the patch. Changes in updated patch: * Removed code duplication for odd and even k. * Temporary buffers c and d no longer allocated on every iteration. * Long result now compacted. No longer unused allocated size. * Added checks for results of long_abs() (it can fail). * Merged _PyLong_GCD and long_gcd. Fast path for small negative integers no longer need to copy long objects in long_abs(). * Added tests for large negative numbers and for case Py_SIZE(a) - Py_SIZE(b) > 3.
Thanks, Serhiy. However, something is wrong with the implementation. The benchmark runs into an infinite loop (it seems). And so do the previous patches. Does it work for you?
I compiled it with 30 bit digits, in case that's relevant. (It might be.)
It works to me (compiled with 15-bit digits). Cold you please add debugging prints (before and after the call of math.gcd()) and find which operation is looping (math.gcd() itself, and for what arguments, or some Python code)?
This is what hangs for me: math.gcd(1216342683557601535506311712, 436522681849110124616458784) "a" and "b" keep switching between both values, but otherwise, the loop just keeps running. The old fractions.gcd() gives 32 for them.
I can confirm that it works with 15 bit digits.
To avoid regressions, please can we leave the old `fractions.gcd` exactly as it was? For example, the current `fractions.gcd` *does* work for Fraction instances [1]. That's certainly not its intended use, but I wouldn't be surprised if there's code out there that uses it in that way. It also just happens to work for nonnegative finite float inputs, because a % b gives exact results when a and b are positive floats, so no error is introduced at any point. I'd also worry about breaking existing uses involving integer-like objects (instances of numpy.int64, for example) in place of instances of ints. [1] By "works", I mean that if a and b are Fractions then gcd(a, b) returns a Fraction such that (1) a and b are integer multiples of gcd(a, b), and (2) gcd(a, b) is an integer multiple of any other number with this property.
> This is what hangs for me: Uh-oh. Sounds like I screwed up somewhere. I'll take a look this weekend, unless Serhiy beats me too it.
> too it. Bah. "to it". Stupid fingers.
Thank you Stefan. I confirm that it hangs with 30-bit digits. One existing bug is in the use of PyLong_AsLong() before simple Euclidean loop. It should be PyLong_AsLongLong() if the long is not enough for two digits. But there is another bug in inner loop...
Here is fixed patch. There was integer overflow. In C short*short is extended to int, but int*int results int.
And for comparison here is simpler patch with Euclidean algorithm.
Patch 7 works for me. Why are the two Py_ABS() calls at the end needed when we start off the algorithm with long_abs()? The Lehmer code is complex (I guess that's why you added the pure Euclidean implementation), but it's the right algorithm to use here, so I'd say we should. It's 4% faster than the Euclidean code for the fractions benchmark when using 30 bit digits, but (surprisingly enough) about the same speed with 15 bit digits. There is no major difference to expect here as the numbers are perpetually normalised in Fractions and thus kept small (usually small enough to fit into a 64bit integer), i.e. Euclid should do quite well on them. The difference for big numbers is substantial though: Euclid: $ ./python -m timeit -s 'from math import gcd; a = 2**123 + 3**653 + 5**23 + 7**49; b = 2**653 + 2**123 + 5**23 + 11**34' 'gcd(a,b)' 10000 loops, best of 3: 71 usec per loop Lehmer: $ ./python -m timeit -s 'from math import gcd; a = 2**123 + 3**653 + 5**23 + 7**49; b = 2**653 + 2**123 + 5**23 + 11**34' 'gcd(a,b)' 100000 loops, best of 3: 11.6 usec per loop
> Why are the two Py_ABS() calls at the end needed when we start off the > algorithm with long_abs()? Because long_abs()'s are omitted for small enough numbers (common case). So we avoid a copying for negative numbers or int subclasses. > I guess that's why you added the pure Euclidean implementation Euclidean algorithm is required step at the end of Lehmer algorithm. > It's 4% faster than the Euclidean code for the fractions benchmark > when using 30 bit digits, but (surprisingly enough) about the same speed > with 15 bit digits. May be because Lehmer code uses 64-bit computation for 30-bit digits, and Euclidean code always uses 32-bit computation. > The difference for big numbers is substantial though: 1000-bit integers are big, but can be encountered in real word (e.g. in cryptography). So may be there is need in Lehmer algorithm.
My personal take is: if there is an implementation in the stdlib, it should be the one that's most widely applicable. And that includes large numbers. We have a working implementation that is algorithmically faster for large numbers, so I can't see why we should drop it unused. I'm for merging patch 7.
Any objections to merging the last patch?
> Any objections to merging the last patch? Yes! Please don't make these changes to `Fractions.gcd`: they'll cause regressions for existing uses of `Fractions.gcd` with objects not of type `int`.
There are not such changes in patch 7. The fractions.gcd() function is unchanged but no longer used by the Fraction type, which now uses math.gcd() internally instead.
Ah, I misread; thanks. What happens with this patch if a Fraction has been created with Integrals that aren't of type int? (E.g., with NumPy int32 instances, for example?)
Why patching fraction.Fraction constructor instead of fractions.gcd()? I don't like the idea of having two functions, math.gcd and fractions.gcd, which do almost the same, but one is slow, whereas the other is fast. It's harder to write efficient code working on Python < 3.5 (use fractions) and Python >= 3.5 (use math or fractions?). I suggest to modify fractions.gcd() to use math.gcd() if the two parameters are int. We just have to adjust the sign: if the second parameter is negative, return -math.gcd(a, b). (I guess that we have unit tests for fractions.gcd checking the 4 cases for signed parameters.)
> I suggest to modify fractions.gcd() to use math.gcd() if the two parameters are int. Sounds fine to me, so long as the code (both fractions.gcd and the fractions.Fraction implementation) continues to function as before for objects that don't have exact type int.
+1 I mean, there is already such a type check in Fraction.__init__(), but I can see a case for also optimising fraction.gcd() for exact ints.
One other suggestion: I think math.gcd should work with arbitrary Python objects implementing __index__, and not just with instances of int.
> I mean, there is already such a type check in Fraction.__init__() That type-check doesn't protect us from non-int Integrals, though, as far as I can tell. It looks to me as though doing `Fraction(numpy.int32(3), numpy.int32(2))` would fail with a TypeError after this patch. (It works in Python 3.4.)
I suggest just add deprecation warning in fractions.gcd(). Or at least add notes which recommend math.gcd() in the docstring and the documentation of fractions.gcd(). > One other suggestion: I think math.gcd should work with arbitrary Python > objects implementing __index__, and not just with instances of int. Agree.
What's the status of this issue? See also the issue #22477.
Here is a patch which addresses both Mark's suggestions. * math.gcd() now work with arbitrary Python objects implementing __index__. * fractions.gcd() and Fraction's constructor now use math.gcd() if both arguments are int, but also support non-ints (e.g. Fractions or floats). * fractions.gcd() now is deprecated. But before committing I want to experiment with simpler implementation and compare it with current complex implementation. If the difference will be not too large, we could use simpler implementation.
Any more comments on this? The deadlines for new features in Py3.5 are getting closer. It seems we're just discussing details here, but pretty much everyone wants this feature. So, what are the things that still need to be done? Serhiy submitted working patches months ago.
New changeset 34648ce02bd4 by Serhiy Storchaka in branch 'default': Issue #22486: Added the math.gcd() function. The fractions.gcd() function now is https://hg.python.org/cpython/rev/34648ce02bd4
New changeset 4691a2f2a2b8174a6c958ce6976ed5f3354c9504 by Victor Stinner in branch 'master': bpo-39350: Remove deprecated fractions.gcd() (GH-18021) https://github.com/python/cpython/commit/4691a2f2a2b8174a6c958ce6976ed5f3354c9504
messages: + msg227664
messages: + msg227663
messages: + msg227570
messages: + msg227537
messages:
+ msg227529
stage: patch review