[SciPy-user] operations on int8 arrays

Travis Oliphant oliphant at ee.byu.edu
Wed Oct 19 16:53:42 CDT 2005

Gary Pajer wrote:

>>I've personally always been of the opinion that accumulator methods are one 
>>case where automatic upcasting is justified.  Since not doing it is almost 
>>guaranteed to produce incorrect results in most cases (esp. for small bit-size 
>>types), I'm +1 on upcasting on this one.
>>I agree that in general we shouldn't upcast silently, but I think this is a 
>>case of 'practicality beats purity'.
>Would there be a way of preventing upcasting, if desired?
>You wouldn't want to eliminate functionality.
>My $0.02:  getting bitten is easy, but if I'm doing integer arithmetic, 
>I usually want to be in control of what exactly is going on.  Having the 
>size of my integer change without my asking for it is ... well, it's not 
>My feelings on this are significantly influenced  by the way I was 
>"brought up".   Might be old fashioned.
The only thing that would change is what the default reduce type is.  
You can already get the desired behavior by specifying a wider reduce 
type explicitly.  

One issue, though.  Upcasting to a different type will be slower.

Right now the defaults is that the reduce type is the same as the type 
of the object coming in.  We could change this to some "wide" type.   
But should we?  It will carry a speed hit.  But the user can always 
over-ride the default behavior by an explicit rtype= paramter.   

I'd like more opinions...


More information about the SciPy-user mailing list