[Mesa-dev] [PATCH 00/14] RadeonSI: SDMA texture copy rewrite

[Alex Deucher]

On Thu, May 5, 2016 at 12:17 PM, Marek Olšák <maraeo at gmail.com> wrote:
> On Thu, May 5, 2016 at 5:57 PM, Alex Deucher <alexdeucher at gmail.com> wrote:
>> On Wed, May 4, 2016 at 7:43 PM, Marek Olšák <maraeo at gmail.com> wrote:
>>> Hi,
>>>
>>> This patch series completely rewrites texture copying with SDMA for CIK & VI. It only uses the "partial" copy packets, which makes it a lot simpler (one packet per layered/3D copy). Most of the complexity is in handling hw limitations, but luckily the SDMA path can be used in the majority of cases.
>>>
>>> On top of that, the DMA IB support is improved thanks to extensive testing and benchmarking.
>>>
>>> This is only enabled on VI for now. It can be enabled on CIK after enough testing is done.
>>>
>>> There is one more hardware limitation that can cause VM faults with T2L copies and needs a workaround. The exact workaround is still under discussion, but I think this is good enough for review already.
>>>
>>> On Tonga, it beats the 3D engine in texture upload performance by 0-80%. On APUs, the performance is mixed or slightly worse. The reason may be that the 3D engine uses less bandwidth thanks to DCC.
>>>
>>> The fact that SDMA can't do DCC compression can be detrimental to texturing performance on VI. The impact hasn't been measured yet.
>>>
>>> Texture uploads still use 2 copies (user memory -> linear, linear -> tiled). Merging those 2 copies into 1 is not done in this series.
>>>
>>> Please review and opinions on the DCC issue are welcome,
>>
>> Have you looked at using SDMA for buffer clears as well?
>
> In general, I don't think SDMA is a good match for any buffer-related
> operations if a GFX IB flush is required.
>
> The only area where it might make sense to use SDMA is the initial
> HTILE/DCC/CMASK fill when new textures are being created and there is
> no dependency on the GFX IB.

Yeah, I was thinking it might be useful for UVD and VCE since we do a
lot of clears in the state setup and we don't really have a dependency
on the GFX pipe.

Alex

>
> We already use SDMA for all pipelined buffer uploads. I don't know if
> that is helping us or hurting us. Probably the latter since CP DMA can
> copy through (and thus preload) TC L2, while SDMA can't.
>
> Marek