[Numpy-discussion] allocated memory cache for numpy

Julian Taylor jtaylor.debian at googlemail.com
Tue Feb 18 14:19:04 EST 2014

On 18.02.2014 16:21, Julian Taylor wrote:
> On Mon, Feb 17, 2014 at 9:42 PM, Nathaniel Smith <njs at pobox.com> wrote:
>> On 17 Feb 2014 15:17, "Sturla Molden" <sturla.molden at gmail.com> wrote:
>>> Julian Taylor <jtaylor.debian at googlemail.com> wrote:
>>>> When an array is created it tries to get its memory from the cache and
>>>> when its deallocated it returns it to the cache.
> ...
>> Another optimization we should consider that might help a lot in the same
>> situations where this would help: for code called from the cpython eval
>> loop, it's afaict possible to determine which inputs are temporaries by
>> checking their refcnt. In the second call to __add__ in '(a + b) + c', the
>> temporary will have refcnt 1, while the other arrays will all have refcnt
>>> 1. In such cases (subject to various sanity checks on shape, dtype, etc) we
>> could elide temporaries by reusing the input array for the output. The risk
>> is that there may be some code out there that calls these operations
>> directly from C with non-temp arrays that nonetheless have refcnt 1, but we
>> should at least investigate the feasibility. E.g. maybe we can do the
>> optimization for tp_add but not PyArray_Add.
> this seems to be a really good idea, I experimented a bit and it
> solves the temporary problem for this types of arithmetic nicely.
> Its simple to implement, just change to inplace in
> array_{add,sub,mul,div} handlers for the python slots. Doing so does
> not fail numpy, scipy and pandas testsuite so it seems save.
> Performance wise, besides the simple page zeroing limited benchmarks
> (a+b+c), it also it brings the laplace out of place benchmark to the
> same speed as the inplace benchmark [0]. This is very nice as the
> inplace variant is significantly harder to read.
> Does anyone see any issue we might be overlooking in this refcount ==
> 1 optimization for the python api?
> I'll post a PR with the change shortly.

here is the PR:

probably still lacking some checks, but I think it can be tested

More information about the NumPy-Discussion mailing list