[Mesa-dev] r600g winsys backend rework

[Alex Deucher]

On Mon, Aug 8, 2011 at 7:56 PM, Alex Deucher <alexdeucher at gmail.com> wrote:
> On Sun, Aug 7, 2011 at 5:25 PM, Marek Olšák <maraeo at gmail.com> wrote:
>> Hi,
>>
>> I have been recently trying to get thread offloading of the CS ioctl
>> into r600g in order to reduce the impact of kernel overhead on fps.
>> That, unfortunately, requires whole winsys/radeon to be used, because
>> even the buffer management (bo_map, bo_wait, bo_busy) must take into
>> account that a CS ioctl may be in progress. Besides that, there are
>> several possible race conditions in r600g, so instead of rewriting
>> r600g and trying to do what winsys/radeon is doing, I decided to
>> simply use winsys/radeon.
>>
>> What's new in r600g:
>> - Thread offloading of the DRM CS ioctl. I expect 0-15% increase in
>> performance from that in CPU-bound apps.
>> - The new GEM_WAIT ioctl is used to avoid waiting for a buffer when
>> possible. (e.g. Mesa may map an index buffer to compute index bounds,
>> which shouldn't cause unnecessary waiting now) I have sent the DRM
>> patches which add the ioctl to dri-devel.
>> - Thread-safety: There are several possible races in r600g. I
>> especially don't like radeon_bo::reloc, which may cause pretty ugly
>> races if a resource is shared and relocated in multiple contexts.
>> winsys/radeon doesn't have that race and also fixes a couple more.
>> Hopefully this thread-safeness won't cause performance regressions.
>>
>> winsys/radeon can do space checking as well, but we don't use that in r600g yet.
>>
>> Performance improvements - I have been able to find a difference with
>> these apps:
>>
>> Unigine Heaven
>> Before: 7.3 fps
>> After: 7.6 fps
>>
>> Torcs
>> Before: 29 fps
>> After: 34 fps
>>
>> Note that every commit in the r600winsys2 branch has been committed
>> without piglit regressions, so that we can bisect through it if
>> needed. The net loss is a little over 900 lines of code in r600g. The
>> fenced cache buffer manager in r600g has turned out not to be superior
>> to pb_cache_bufmgr with the is_buffer_busy hook set, so I removed the
>> former too, as the latter is way simpler.
>>
>> This new work has been pushed into a new branch r600winsys2 in the
>> main Mesa repository, please review/test.
>>
>
> Looks good.  I'll try and run some tests on various hw later this week.

I ran this on a few cards today and all seems well.

Alex

>
> Reviewed-by: Alex Deucher <alexander.deucher at amd.com>
>
>
>> Marek
>> _______________________________________________
>> mesa-dev mailing list
>> mesa-dev at lists.freedesktop.org
>> http://lists.freedesktop.org/mailman/listinfo/mesa-dev
>>
>