Hi Vincent, On Tue, Oct 19, 2010 at 2:28 AM, Vincent Favre-Nicolin <vincent.favre-nico...@cea.fr> wrote: > I'm not sure what is happening exactly, but there is no indication that the > random numbers are *repeating themselves* > > However you seem to hit a floating point issue *when computing the sum* => if > you replace numpy.sum(a)/size by numpy.sum(a.astype(numpy.float64))/size, you > get correct results for float32 (always around 0.5). > > It is actually not suprising: when you sum 2**25 times numbers which are on > average 0.5, you get ~ 2**24. And in single precision floating point, the > mantissa is indeed stored using 24 bits => when you try adding the next > numbers, they are simply too small to register in the mantissa, and are not > added. e.g. try: float32(2**24)+float32(.5)= float32(2**24) ! > > Funny example of the limits of single precision floating point...
Yes, it seems that you are right. When sum reaches 2 ** 24, every next term is not big enough to change it. It never occurred to me that I am limited not only by exponent, but by mantissa too... Best regards, Bogdan _______________________________________________ PyCUDA mailing list PyCUDA@tiker.net http://lists.tiker.net/listinfo/pycuda
