[Mesa-dev] [PATCH] radeonsi: Set declared tessellation LDS size to hardware size.

[Bas Nieuwenhuizen]

On Mon, May 9, 2016 at 9:55 PM, Marek Olšák <maraeo at gmail.com> wrote:
> On Sun, May 8, 2016 at 1:00 PM, Bas Nieuwenhuizen
> <bas at basnieuwenhuizen.nl> wrote:
>> The calculated limit gave problems on SI as it was > 32 KiB
>> and the hardware LDS size on SI is only 32 KiB. It isn't
>> correct anyway when processing multiple patches in a threadgroup.
>>
>> As we potentially have any number of patches such that the
>> used LDS is at most the hardware LDS size, and exact size
>> per patch is not known at compile time, this seems like
>> the only valid bound.
>
> I think NIcolai sent out a similar patch.

Yes, I've sent this patch as a reply.
>
> Would tessellation break if we set the size to 0 or a similar small
> number? It seems like seemingly high LDS usage can trick LLVM now or
> in the future into thinking that it doesn't have to schedule
> instructions for low register usage.

The problem with smaller sizes than total LDS is that we have no
guarantuee that it will be placed at the same position for all shader
stages, and LLVM could introduce new LDS variables (see the promote
alloca pass which promotes alloca to LDS memory, although it currently
has several wrong assumptions for non-compute shaders anyway[1]).

The LDS usage is per threadgroup though, so using the entire LDS does
not put a limit on the number of waves that are in the threadgroup.
Furthermore in the graphics pipeline we have to consider that even
though the current shader has a limit on the number of active waves
due to LDS there might be other stages that do not use LDS and could
use the registers.

Note that neither too small nor too large (as long as it fits in LDS)
is likely to lead to problems at this moment.

- Bas

[1]:
 - It assumes all LDS can be used. However VS as VS can't use any LDS
at all, and PS has to leave some for the inputs.
 - It assumes that the max threadgroup size for non-compute shaders is
a single wave.