-
由 Tvrtko Ursulin 提交于
Because it is based on jiffies, current implementation releases the forcewake at any time between straight away and between 1ms and 10ms, depending on the kernel configuration (CONFIG_HZ). This is probably not what has been desired, since the dynamics of keeping parts of the GPU awake should not be correlated with this kernel configuration parameter. Change the auto-release mechanism to use hrtimers and set the timeout to 1ms with a 1ms of slack. This should make the GPU power consistent across kernel configs, and timer slack should enable some timer coalescing where multiple force-wake domains exist, or with unrelated timers. For GlBench/T-Rex this decreases the number of forcewake releases from ~480 to ~300 per second, and for a heavy combined OGL/OCL test from ~670 to ~360 (HZ=1000 kernel). Even though this reduction can be attributed to the average release period extending from 0-1ms to 1-2ms, as discussed above, it will make the forcewake timeout consistent for different CONFIG_HZ values. Real life measurements with the above workload has shown that, with this patch, both manage to auto-release the forcewake between 2-4 times per 10ms, even though the number of forcewake gets is dramatically different. T-Rex requests between 5-10 explicit gets and 5-10 implict gets in each 10ms period, while the OGL/OCL test requests 250 and 380 times in the same period. The two data points together suggest that the nature of the forwake accesses is bursty and that further changes and potential timeout extensions, or moving the start of timeout from the first to the last automatic forcewake grab, should be carefully measured for power and performance effects. v2: * Commit spelling. (Dave Gordon) * More discussion on numbers in the commit. (Chris Wilson) Signed-off-by: NTvrtko Ursulin <tvrtko.ursulin@intel.com> Reviewed-by: NDave Gordon <david.s.gordon@intel.com> Cc: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: NChris Wilson <chris@chris-wilson.co.uk>
a57a4a67