[Intel-gfx] [PATCH v4] drm/i915: Execlists small cleanups and micro-optimisations

[Chris Wilson]

On Tue, Mar 01, 2016 at 10:41:26AM +0000, Tvrtko Ursulin wrote:
> What about your patch to move it all to a bottom handler? Are we
> going to progress that one?

I looked for the regressions and found an interesting slowdown in single nop
batches (almost 2x as slow). Another one that should not happen, as along
that idle-execute-sync path the code should be the same (the elsp should
be written directly by the submitter in both cases). CPU time looks
pretty much identical between the two, so I'm left questioning the
scheduler and the impact of using another kthread. The effect is also
only apparent on bsw not big core, perhaps more indication of a
context-switch versus CPU cache? I also found another way to lockup the
interrupt handler thanks to its while(1) looping over GT + display
interrupts.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre