[Cython] Cython array type: Summary, introducing CEP 518

[Dag Sverre Seljebotn]

Stefan Behnel wrote:
> Dag Sverre Seljebotn wrote:
>>> # Some ways of multiplying all elements with 2
>>> x *= 2
>>> x[...] *= 2
>>> x[:,:] *= 2
>>> x += x
>>> x[...] += x
>> OK I'll make an exception here -- I'm willing to discuss whether we 
>> should depart from NumPy semantics here and let
>>
>> x2 = x
>> x *= 2
>>
>> allocate new memory, so that x2 is not modified, being consistent with a 
>> direct transformation to "x = x * 2". One can always write
>>
>> x[...] *= 2
>>
>> if one wishes to modify original memory.
> 
> What's wrong with
> 
> 	x = 2 * x2
> 
> for doing a copy ?
> 
> x *= 2
> 
> pretty clearly states that I want to modify x in place.

The problem is that if you do

y = x
x *= 2

then in current NumPy, y and x will still point to the same memory and 
reference the same values (in fact, be the exact same view object).

Usual Python semantics seems to imply that what NumPy *should* have done 
is let the latter line mean "x = x * 2", where a new array is allocated 
and the value of x*2 copied into the new array, so that x and y points 
to different memory after the operation.

BTW, something I forgot to say about the == operator is that a common 
pattern is to do

x[x == 2] = 0

to replace 2 with 0 everywhere in x, that's one of its primary uses.

To make it more clear:

xeq2mask = (x == 2) # array of booleans
x[xeq2mask] = 0

(That is connected with the boolean indexing thing, which I'm vary about 
letting into Cython, although an exception can safely be made in the 
case of __setitem__).

-- 
Dag Sverre