I’m trying to find out the exact clock cycles each math function takes to execute (e.g. sinf, cosf…). In the programming guide, it says “__sinf(x), __cosf(x), __expf(x) take 32 clock cycles. sinf(x), cosf(x), tanf(x), sincosf(x) are much more expensive and even more so (i.e. about an order of magnitude slower) if the absolute value of x is greater than 48039”. But it doesn’t say exactly how many clock cycles. Can anybody point me to any document(s) for this information?
You can probably see for yourself by checking the implementation. It is in a header file in your cuda installation. As far as I remember, they execute the __sinf function a few times to have a sinf for large x.