[Intel-gfx] [PATCH 1/6] drm/i915: Eliminate vmap overhead for cmd parser

[Chris Wilson]

On Tue, Oct 06, 2015 at 02:44:22PM +0200, Daniel Vetter wrote:
> On Thu, Oct 01, 2015 at 12:57:10PM +0100, Chris Wilson wrote:
> > With a little complexity to handle cmds straddling page boundaries, we
> > can completely avoiding having to vmap the batch and the shadow batch
> > objects whilst running the command parser.
> > 
> > On ivb i7-3720MQ:
> > 
> > x11perf -dot before 54.3M, after 53.2M (max 203M)
> > glxgears before 7110 fps, after 7300 fps (max 7860 fps)
> > 
> > Before:
> > Time to blt 16384 bytes x      1:	 12.400µs, 1.2GiB/s
> > Time to blt 16384 bytes x   4096:	  3.055µs, 5.0GiB/s
> > 
> > After:
> > Time to blt 16384 bytes x      1:	  8.600µs, 1.8GiB/s
> > Time to blt 16384 bytes x   4096:	  2.456µs, 6.2GiB/s
> 
> Numbers for the overall series (or individual patches would be even
> better) are needed. I thought you have this neat script now to do that for
> an entire series?

Note that numbers on a patch are for the patch unless otherwise stated.
Double so since these are numbers from when I first posted it and didn't
have anything else to boost cmdparser perf.

I do and I even added the benchmark to demonstrate one case. Then I
forgot to enable the cmdparser in mesa and so its numbers are bunk.
Fancy scripts still can't eliminate pebkac.

However, it did show that we still get the throughput improvement from
killing the vmap even with the temporary copy.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre