Hi, On Thu, Jul 15, 2010 at 4:49 AM, Ian Ozsvald <i...@ianozsvald.com> wrote: > One question that's raised from my post - how come double precision > CPU math is faster than single precision CPU math? I hadn't expected > that result to drop out of the test and it has been a while since I > did any good speed tests on CPUs. Is it generally the case that double > precision math (both in C and Python?) on x86 is faster than single > precision now?
Cpu and gpu are still faster in float then in double. That is caused by the hardware. I have already see float being slower then double. It was the exp fct in gcc. If I change the implementation to use the one in amblibm, the difference disappeared (and it was faster). So I suppose that is also your case. Don't forget that some implementation will spend more time optimizing the double case as it is more general. Frédéric Bastien _______________________________________________ PyCUDA mailing list PyCUDA@tiker.net http://lists.tiker.net/listinfo/pycuda