[Mesa-dev] [PATCH 05/18] i965: Reuse our VBO for streaming fast-clear vertices

[Martin Peres]

On 08/07/15 12:01, Chris Wilson wrote:
> On Tue, Jul 07, 2015 at 01:42:47PM +0300, Martin Peres wrote:
>> On 06/07/15 19:43, Kenneth Graunke wrote:
>>> On Monday, July 06, 2015 11:33:10 AM Chris Wilson wrote:
>>>> Rather than allocating a fresh page every time we clear a buffer, keep
>>>> that page around between invocations by tracking the last used offset
>>>> and only allocating a fresh page when we wrap.
>>>>
>>>> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
>>>> ---
>>>>   src/mesa/drivers/dri/i965/brw_meta_fast_clear.c | 17 ++++++++++++++---
>>>>   1 file changed, 14 insertions(+), 3 deletions(-)
>>> This looks okay to me.  Do you have any performance data to justify the
>>> extra complexity?
>> I actually get a negative performance improvement on a customer
>> benchmark (-1.3%). Could it be because we are waiting on the VBO at
>> some point?
>>
>> What benchmark did you try to get a perf improvement?
> For this something as simple and cpu bound as windowed glxgears:
> 14550 -> 15400 (+-30fps both sides).
> -Chris

Thanks!

I re-run the tests on the same kernel (4.0.7-2-ARCH) and another one 
(drm-intel-nighly-06-07-2015) and I can't reproduce the hit anymore... 
The hit was there, visible over the 30 rounds of the test, so this is 
weird. Let's just ignore this.

Martin