On Fri, Jul 27, 2012 at 5:15 AM, Charles R Harris <charlesr.har...@gmail.com> wrote: > I would support accumulating in 64 bits but, IIRC, the function will need to > be rewritten so that it works by adding 32 bit floats to the accumulator to > save space. There are also more stable methods that could also be > investigated. There is a nice little project there for someone to cut their > teeth on.
So the obvious solution here would be to make the ufunc reduce loop smart enough that x = np.zeros(2 ** 30, dtype=float32) np.sum(x, dtype=float64) does not upcast 'x' to float64's as a whole. This shouldn't be too terrible to implement -- iterate over the float32 array, and only upcast each inner-loop "buffer" as you go, instead of upcasting the whole thing. In fact, nditer might do this already? Then using a wide accumulator by default would just take a few lines of code in numpy.core._methods._mean to select the proper dtype and downcast the result. -n _______________________________________________ NumPy-Discussion mailing list NumPy-Discussion@scipy.org http://mail.scipy.org/mailman/listinfo/numpy-discussion
