Re: [Numpy-discussion] A reimplementation of MaskedArray

[Tim H. <tim...@ie...>]

Christopher Barker wrote:
> Tim Hochberg wrote:
>   
>> I've been told that operations on NANs are slow because they aren't 
>> always implemented in the FPU hardware. Instead they are trapped and 
>> implemented software or firmware or something or other. 
>>     
>
> which still doesn't make sense -- doesn't ANY operation with a NaN 
> return NaN? how hard could that be?
>   
Well, it's not free. You still have to recognize a particular bit 
pattern corresponding to QNaN and then copy the result in that case. All 
that probably burns more silicon for a case that most people don't care 
about. Do NaNs make Half Life run any faster? No I didn't think so.

I suspect that instead they just look for the case where the exponent is 
all 1s, in which case the result is either +-Inf, SNaN or QNaN and then 
drop into software to handle those infrequent cases.
> I share in the disappointment that NaN's are not first class citizens in 
> common hardware....
>   
It would be nice to see IEEE 754 supported more faithfully, but there 
seems to be more of a problem on the software side than on the hardware 
side in practice.

I still can't help feeling that using NaNs for mask values would be 
abuse even if it weren't slow and eventually it would reach around and 
bite you someplace tender.

-tim