Re: [PATCH 0/6] Radeon memory management improvements

["Christian König" <deathsimple@vodafone.de>]

Hi Marek,

Some minor comments on patch 1, 2 and 5, but nothing serious. Patch 3, 4 
and 6 are Reviewed-by: christian König <christian.koenig@amd.com>

See below for a few in line comments.

Am 24.02.2014 16:20, schrieb Marek Olšák:
> This series improves performance for the cases when there is not enough VRAM for all buffers.
>
> First of all, I'd like to mention that if you set both VRAM and GTT domains for a buffer, you pretty much say you don't care where the buffer ends up. It usually makes the performance even worse.
>
> This work was largely benchmark-driven and I tried a lot of ideas before I found out which ones work. The patches describe what they do and they're quite simple, so I'll just share the results here.
>
>
> Card: Evergreen Redwood (HD 5670), 512 MB of VRAM
> Test: Unigine Heaven 4.0, High settings
>
> 1) 1280x720, 4x MSAA, need 525 MB of VRAM
>
> Without patches: 16.6 FPS
> With patches: 16.6 FPS
> Improvement: 0 %
>
> 2) 1600x900, 4x MSAA, need 642 MB of VRAM
>
> Without patches: 7.1 FPS
> With patches: 9.7 FPS
> Improvement: 36 %
>
> 3) 1920x1080, 4x MSAA, need 743 MB of VRAM
>
> Without patches: 3.7 FPS
> With patches: 5.6 FPS
> Improvement: 51 %
>
> 4) 1600x900, 8x MSAA, need 838 MB of VRAM
> Without patches: 2.9 FPS
> With patches: 4.6 FPS
> Improvement: 58 %
>
> These results don't change if you run the benchmark several times, which proves the improvement is stable.
>
>
> To conclude this, here are ideas for future work:
>
> 1) Add virtual memory support for VRAM. Our GPUs support virtual memory, which not only solves fragmentation issues, but it also allows each buffer to be partially in VRAM and partially in GTT, which becomes more important with large buffers like 100 MB. Moving whole buffers back and forth between VRAM and GTT is inefficient if you can do it at page granularity. Also, due to fragmentation, we can never really use all of VRAM, but only about 90-95%.

Yeah, I'm also thinking about this for quite some time now. The basic 
problem is that while our GPUs support VM they don't support faulting 
pages in and continuing (at least nobody got that working reliable so 
far). E.g. when you hit a page fault you can't relocate the page and 
then continue.

Support for partially resident textures on newer hardware currently 
works by splitting the buffer up into smaller buffers in userspace and 
then actively checking in the shader if we hit a buffer that's not 
currently in memory, but that's not really applicable in the general use 
case (to much shader overhead).

> 2) Add support for uncached GTT. I think it should improve performance for dGPUs under memory pressure, but some testing needs to be done to confirm that. Uncached GTT doesn't seem to work for me on Evergreen, but it's said to be working on some later chips.

Did you try to make the whole GTT uncached or just evicted BOs? Making 
the whole GTT uncached probably won't work out of the box, but avoiding 
setting the "SNOOPED" flag on those pages might get us better 
performance while swapping them into VRAM again.

Christian.

>
> The patches for Mesa will follow later today. Please review.
>
> Marek
> _______________________________________________
> dri-devel mailing list
> dri-devel@lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/dri-devel