On Nov 22, 2016 23:49, "Chris Wilson" <<a href="mailto:chris@chris-wilson.co.uk">chris@chris-wilson.co.uk</a>> wrote: > > On Tue, Nov 22, 2016 at 11:32:38PM +0000, Robert Bragg wrote: > > Thanks for sending out. It looked good to me, but testing shows a 'divide > > error'. > > > > I haven't double checked, but I think it's because the max OA exponent > > (31) converted to nanoseconds is > UINT32_MAX with the lower 32bits zero > > and the do_div denominator argument is only 32bit. > > Hmm, I thought do_div() was u64 / u64, but no it is u64 / u32. Looks > like the appropriate function would be div64_u64(). > > > It corresponds to a 5 minute period which is a bit silly, so we could > > reduce the max exponent. A period of UINT32_MAX is about 4 seconds where I > > can't currently think of a good use case for such a low frequency. > > > > Instead of changing the max OA exponent (where the relationship to the > > period changes for gen9 and may become fuzzy if we start training our view > > of the gpu timestamp frequency instead of using constants) maybe we should > > set an early limit on an exponent resulting in a period > UINT32_MAX? > > Seems like picking the right function would help! Or that, yep. Sounds good to me, thanks. - Robert > -Chris > > -- > Chris Wilson, Intel Open Source Technology Centre