AMDGPU Tuning Tests With DRI2/DRI3, PowerPlay, Semaphores, Scheduler

[phoronix]

Phoronix: AMDGPU Tuning Tests With DRI2/DRI3, PowerPlay, Semaphores, Scheduler

Complementing yesterday's AMDGPU tests with the new DRM-Next code that has PowerPlay support where the speed of this latest open-source driver code was compared to the proprietary driver, here are some tests showing the AMDGPU driver performance under a few different scenarios.

http://www.phoronix.com/vr.php?view=22610

[agd5f]

Semaphores are only used when the scheduler is disabled. The scheduler eliminates the need for them.

[FireBurn]

Semaphores are only used when the scheduler is disabled. The scheduler eliminates the need for them.

So if they're manually enabled does that disable the scheduler, is it just ignored or some undefined result?

[artivision]

When i play for example DIRT_SHOWDOWN on Gallium_Nine i get 100% CPU utilization on my i5-AVX-quad and only 30% GPU utilization on my 4TFlops GCN Radeon. On Win10 i get 60% CPU and 100% GPU. The overhead is 5x on Linux, what can do that?

[agd5f]

So if they're manually enabled does that disable the scheduler, is it just ignored or some undefined result?

If the scheduler is enabled, they are not used regardless of how the semaphore module parameter is set. The semaphore parameter is only relevant if the scheduler is disabled.

[sudden6]

If the scheduler is enabled, they are not used regardless of how the semaphore module parameter is set. The semaphore parameter is only relevant if the scheduler is disabled.

So basically the performance with PP + DRI3 should be the same as with PP + DRI3 + Semaphores?

Now the question is, why are they different in the benchmarks?^^

[Michael]

So basically the performance with PP + DRI3 should be the same as with PP + DRI3 + Semaphores?

Now the question is, why are they different in the benchmarks?^^

In only a few benchmarks was there semaphore difference, but yes.

(as a reminder before people say it's noise or benchmarks are invalid, keep in mind PTS runs each of these graphics tests a minimum of three times and will dynamically increase the run count if there is much variation in the results, as shown by the standard deviation on the graphs)

[agd5f]

The other thing is that semaphores have a hw bug on tonga, Fiji, and Iceland, so even if you enable semaphores in the non-scheduler case, the driver still uses fence waits for the gfx ring.

[agd5f]

Maybe variation in the clocks with pp enabled due to different thermal conditions at the time of each test. Might be worth comparing with pp disabled or forced to the high or low state to take clock variance out of the picture.