[Piglit] [PATCH] drawoverhead: new microbenchmark

Timothy Arceri tarceri at itsqueeze.com
Sun May 14 23:01:41 UTC 2017

On 12/05/17 05:49, Eric Anholt wrote:
> Marek Olšák <maraeo at gmail.com> writes:
>> From: Marek Olšák <marek.olsak at amd.com>
>> Based on a benchmark from mesa/demos, but rewritten and extended.
>> It's a benchmark expected to be run separately, not a piglit test.
>> So why piglit? Because it's a good framework for writing apps like this.
>> mesa_glthread won't show an improvement here, because there is no app
>> overhead.
>> This is what the output looks like. The percentage is relative to
>> the first test of the given draw call.
>> The obvious thing there is that enabled vertex attribs decrease
>> Mesa performance even if there are no state changes.
> Since nobody else has replied,
> I think piglit is the wrong place for this.  I agree that it's sorta
> convenient, but mesa-demos or glmark2 are the right place.

I'm not totally against this. It would be nice to leverage more of the 
piglit framework/tools to be able to run a group of perf tests and 
generate some kind of comparison automatically. That way way we could 
have an easy tools for comparing releases etc. I know Intel has an 
internal setup for this type of thing, but having something in the test 
suite all mesa devs use would be handy.

Anyway I've given this benchmark a run with my KHR_no_error changes and 
I'm seeing a nice jump in a bunch of the tests.

The biggest jump is:

DrawArrays(16 VBOs, 0 UBOs,  0 Tex) w/ no state change: 5.41 million (49.3%)

DrawArrays(16 VBOs, 0 UBOs,  0 Tex) w/ no state change: 16.33 million 

But there is a general increase across the majority of tests which is 
encouraging to see as I still haven't been able to measure much change 
in the games I've been trying. I suspect I'll need to have coverage 
across most of the api a game uses before I see much change.

More information about the Piglit mailing list