* TTM tt size larger than buffer object?
@ 2021-08-27 13:55 Thomas Hellström
2021-08-27 14:24 ` Christian König
0 siblings, 1 reply; 2+ messages in thread
From: Thomas Hellström @ 2021-08-27 13:55 UTC (permalink / raw)
To: Christian König
Cc: DRI Development, Intel Graphics Development, Daniel Vetter,
Ayaz A Siddiqui
Hi, Christian.
We have a use-case with i915 where the data representation of a buffer
object is larger in system memory than in LMEM/VRAM. Hence we'd like to
create a ttm_tt that is larger than the buffer object itself. Quickly
auditing the TTM code it looks like that should be pretty safe, as
ttm->num_pages is not really much accessed outside the tt code and the
pool code where we're doing the right thing.
The additional data will really only be accessed by the blitter so when
cpu-mapping, mapping just the original buffer object size is correct.
However with swapping the additional data needs to be swapped out and
the code is doing that correctly as well.
Do you think this is an acceptable solution?
/Thomas
^ permalink raw reply [flat|nested] 2+ messages in thread
* Re: TTM tt size larger than buffer object?
2021-08-27 13:55 TTM tt size larger than buffer object? Thomas Hellström
@ 2021-08-27 14:24 ` Christian König
0 siblings, 0 replies; 2+ messages in thread
From: Christian König @ 2021-08-27 14:24 UTC (permalink / raw)
To: Thomas Hellström
Cc: DRI Development, Intel Graphics Development, Daniel Vetter,
Ayaz A Siddiqui
Am 27.08.21 um 15:55 schrieb Thomas Hellström:
> Hi, Christian.
>
> We have a use-case with i915 where the data representation of a buffer
> object is larger in system memory than in LMEM/VRAM. Hence we'd like
> to create a ttm_tt that is larger than the buffer object itself.
> Quickly auditing the TTM code it looks like that should be pretty
> safe, as ttm->num_pages is not really much accessed outside the tt
> code and the pool code where we're doing the right thing.
>
> The additional data will really only be accessed by the blitter so
> when cpu-mapping, mapping just the original buffer object size is
> correct. However with swapping the additional data needs to be swapped
> out and the code is doing that correctly as well.
>
> Do you think this is an acceptable solution?
Yes, I think that should work. Just haven't tested it yet.
I've been working on getting the core TTM code to use byte or rather an
arbitrary size while the tt code should use the num_pages from it's own
object.
The background is that we have a very similar use case on amdgpu which
we might need to support.
Christian.
>
> /Thomas
>
>
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2021-08-27 14:40 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2021-08-27 13:55 TTM tt size larger than buffer object? Thomas Hellström
2021-08-27 14:24 ` Christian König
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).