12 May
2011
12 May
'11
5:20 p.m.
Phil Endecott wrote:
OK, but the issue is that I don't trust that you can accurately measure execution times of the order of microseconds. The overhead of, for example, the system call to read the clock will be significant. If you believe that you can accurately measure such short periods, you should justify your methodology. Otherwise, repeat a few thousand times.
http://www.fftw.org/cycle.h But I agree, measuring something on the order of 1 microsecond won't be reliable, even with this non-portable code. Not sure whether fftw uses this itself when measuring which algorithm to use for a specific setup at runtime. Regards, Thomas