Messages in this thread | | | Date | Thu, 12 Jun 2014 09:53:57 -0700 | Subject | Re: [PATCH 3/3] clocksource: exynos_mct: Optimize register reads with ldmia | From | Doug Anderson <> |
| |
Thomas,
On Wed, Jun 4, 2014 at 11:05 AM, Thomas Gleixner <tglx@linutronix.de> wrote: > On Wed, 4 Jun 2014, Doug Anderson wrote: > >> As we saw in (clocksource: exynos_mct: cache mct upper count), the >> time spent reading the MCT shows up fairly high in real-world >> profiles. That means that it's worth some optimization. >> >> We get a roughly 10% speedup in userspace gettimeofday() by using an >> ldmia to read the two halfs of the MCT. That seems like a worthwhile >> thing to do. >> >> Before: 1173084 us for 1000000 gettimeofday in userspace >> After: 1045674 us for 1000000 gettimeofday in userspace >> >> NOTE: we could actually do better than this if we really wanted to. >> Technically we could register the clocksource as a 32-bit timer and >> only use the "lower" half. Doing so brings us down to 1014429 us for >> 1000000 gettimeofday in userspace (and doesn't even require assembly >> code). That would be an alternative to this change. > > I was about to ask exactly that question: What's the advantage of the > 64 bit dance there? AFAICT nothing.
Are you explicitly naking the 64-bit version of these patches? If so I'll send out the 32-bit version right away. If nothing else we should get the ftrace fix (patch 1 in this series) landed ASAP since that fixes a regression. I'd like to make the 32-bit decision first though since fixing the regression is easier in the 32-bit version.
Roughly: with the 64-bit version there's no question that we're not regressing anyone's functionality or performance and it's nearly as fast as the 32-bit version. The 32-bit version is simpler / faster but has the potential to cause (unknown) problems.
-Doug
| |