[pypy-dev] Shift and typed array

[Tuom Larsen]

Alright, thanks again!

On Mon, Apr 4, 2016 at 3:58 PM, Maciej Fijalkowski <fijall at gmail.com> wrote:
> you can write a if in_pypy wrapper that uses __pypy__.intops or
> something, but for exactly the same code is hard to do better. We can
> think about adding some optimizations that do that btw
>
>
> On Mon, Apr 4, 2016 at 3:53 PM, Tuom Larsen <tuom.larsen at gmail.com> wrote:
>> So other than numpy, do you think:
>>
>>     a[0] = (2**63 << 1) & (2**64 - 1)
>>
>> is the best way I can do?
>>
>> On Mon, Apr 4, 2016 at 3:52 PM, Tuom Larsen <tuom.larsen at gmail.com> wrote:
>>> Yes, I just want to write code which is portable (as for example that
>>> `& ...`) but still favors PyPy.
>>>
>>> On Mon, Apr 4, 2016 at 3:49 PM, Maciej Fijalkowski <fijall at gmail.com> wrote:
>>>> right, so there is no way to get wrap-around arithmetics in cpython
>>>> without modifications
>>>>
>>>> On Mon, Apr 4, 2016 at 3:48 PM, Tuom Larsen <tuom.larsen at gmail.com> wrote:
>>>>> I would like to avoid numpy, if possible. Even if it might be bundled
>>>>> with PyPy (is it?) I still would like the code to run in CPython with
>>>>> no dependencies.
>>>>>
>>>>> On Mon, Apr 4, 2016 at 3:45 PM, Maciej Fijalkowski <fijall at gmail.com> wrote:
>>>>>> so numpy64 will give you wrap-around arithmetics. What else are you
>>>>>> looking for? :-)
>>>>>>
>>>>>> On Mon, Apr 4, 2016 at 3:38 PM, Tuom Larsen <tuom.larsen at gmail.com> wrote:
>>>>>>> You mean I should first store the result into numpy's `int64`, and
>>>>>>> then to `array.array`? Like:
>>>>>>>
>>>>>>>     x = int64(2**63 << 1)
>>>>>>>     a[0] = x
>>>>>>>
>>>>>>> Or:
>>>>>>>
>>>>>>>     x = int64(2**63)
>>>>>>>     x[0] = x << 1
>>>>>>>
>>>>>>> What the "real types" goes, is this the only option?
>>>>>>>
>>>>>>> Thanks in any case!
>>>>>>>
>>>>>>>
>>>>>>> On Mon, Apr 4, 2016 at 3:32 PM, Maciej Fijalkowski <fijall at gmail.com> wrote:
>>>>>>>> one option would be to use integers from _numpypy module:
>>>>>>>>
>>>>>>>> from numpy import int64 after installing numpy.
>>>>>>>>
>>>>>>>> There are obscure ways to get it without installing numpy. Another
>>>>>>>> avenue would be to use __pypy__.intop.int_mul etc.
>>>>>>>>
>>>>>>>> Feel free to complain "no, I want real types that I can work with" :-)
>>>>>>>>
>>>>>>>> Cheers,
>>>>>>>> fijal
>>>>>>>>
>>>>>>>> On Mon, Apr 4, 2016 at 3:10 PM, Tuom Larsen <tuom.larsen at gmail.com> wrote:
>>>>>>>>> Hello!
>>>>>>>>>
>>>>>>>>> Suppose I'm on 64-bit machine and there is an `a = arrar.array('L',
>>>>>>>>> [0])` (item size is 8 bytes). In Python, when an integer does not fit
>>>>>>>>> machine width it gets promoted to "long" integer of arbitrary size. So
>>>>>>>>> this will fail:
>>>>>>>>>
>>>>>>>>>     a[0] = 2**63 << 1
>>>>>>>>>
>>>>>>>>> To fix this, one could instead write:
>>>>>>>>>
>>>>>>>>>     a[0] = (2**63 << 1) & (2**64 - 1)
>>>>>>>>>
>>>>>>>>> My question is, when I know that the result will be stored in
>>>>>>>>> `array.array` anyway, how to prevent the promotion to long integers?
>>>>>>>>> What is the most performat way to perform such calculations? Is PyPy
>>>>>>>>> able to optimize away that `& (2**64 - 1)` when I use `'L'` typecode?
>>>>>>>>>
>>>>>>>>> I mean, in C I wouldn't have to worry about it as everything above the
>>>>>>>>> 63rd bit will be simply cut off. I would like to help PyPy to generate
>>>>>>>>> the best possible code, does anyone have some suggestions please?
>>>>>>>>>
>>>>>>>>> Thanks!
>>>>>>>>> _______________________________________________
>>>>>>>>> pypy-dev mailing list
>>>>>>>>> pypy-dev at python.org
>>>>>>>>> https://mail.python.org/mailman/listinfo/pypy-dev