[Mesa-dev] [PATCH 3/4] radeonsi: add more unlikely() uses into si_draw_vbo

[Michel Dänzer]

On 07/09/16 11:01 PM, Nicolai Hähnle wrote:
> On 07.09.2016 12:23, Marek Olšák wrote:
>> On Wed, Sep 7, 2016 at 11:47 AM, Michel Dänzer <michel at daenzer.net>
>> wrote:
>>> On 07/09/16 06:02 PM, Marek Olšák wrote:
>>>>
>>>> Based on the comments so far, it looks like all annotations in the
>>>> patch are very well placed, so I don't know what the fuss is about.
>>>
>>> As I said, there's no question about the first three annotations, but
>>> for the last two we can't know whether they help or actually hurt
>>> without measuring it. That's the whole point.
>>
>> OK. The last two are so rare that their likeliness is below 1%. The
>> conditions are true when discarding CMASK or DCC, changing the micro
>> tile mode of a texture, reallocating texture storage, and degrading
>> the tile mode to linear. Some of those are caused by SDMA doing a
>> whole-layer overwrite, a texture transfer, or a buffer handle export.
>> Some of them can occur only once during the lifetime of a texture.
>> They are very rare during normal gaming, which just consists of state
>> changes and draw calls and nothing in between. I don't know about
>> microbenchmarks.
> 
> FWIW, though I haven't said so explicitly (other than sending my R-b
> while this discussion was already going on), I also think those last two
> cases are fine. We probably shouldn't go overboard with the (un)likely,
> but really, the patch is fine as is.

Okay, thanks.

FWIW, my concern wasn't only about this patch itself but also about not
wasting time on similar changes in the future.


-- 
Earthling Michel Dänzer               |               http://www.amd.com
Libre software enthusiast             |             Mesa and X developer