[Numpy-discussion] [SciPy-user] Arithmetic Errors
Travis Oliphant
oliphant.travis at ieee.org
Mon Jul 17 19:10:36 CDT 2006
John Lawless wrote:
> Travis,
>
> Thanks!
>
> 1). I haven't found any documentation on dtype='O'. (I purchased your
> trelgol book but it hasn't arrived yet.) Does 'O' guarantee no
> wrong answers?
>
The object data-type uses Python objects instead of low-level C-types
for the calculations. So, it gives the same calculations that Python
would do (but of course it's much slower).
> 2). My actual code was more complex than the example I posted. It was
> giving correct answers until I increased the dataset size. Then,
> luckily, the result became obviously wrong. I can go through a
> code and try to coerce everything to double but, when debugging a
> large code, how can one ever be sure that all types are coerced
> correctly if no errors are generated?
>
NumPy uses c data-types for calculations. It is therefore, *much*
faster, but you have to take precautions about overflowing on integer
operations.
> 3). AFAIK, checking for overflows should take no CPU time whatsoever
> unless an exception is actually generated.
This is true for floating point operations, but you were doing integer
multiplication. There is no support for hardware multiply overflow in
NumPy (is there even such a thing?).
Python checks for overflow on integer arithmetic by doing some
additional calculations. It would be possible to add slower, integer
over-flow checking ufuncs to NumPy if this was desired and you could
replace the standard non-checking functions pretty easily.
-Travis
More information about the Numpy-discussion
mailing list