Open serge-sans-paille opened 8 years ago
arggg, I've do many bitwise operations on uint8. Say me if there is something I can do to help.
Did you have a chance to improve speed for np.uint8 & np.uint8 ?
On Sat, Jan 07, 2017 at 07:48:24AM -0800, Luis wrote:
Did you have a chance to improve speed for np.uint8 & np.uint8 ?
No, but thanks to your reminder, that will be my next task! Keep in touch :-)
@LuisBL just to be sure : any reason why not nusing a uint64 ?
yes it's because I use numpy.unpackbits
to store 8 0/1 in only one byte.
Cf. numpy.bitwise_and
and numpy.unpackbits
I've got the fix but SIMD instructions are not used for comparison operators, I'll push it soon.
Expressions like
are slow in pythran if
a.dtype is np.uint8
. It's no longer the case for uint64, so I guess we should introduce a representation change when all operations are bitwise.