I have been investigating why some piglit tests were failing when VA was
enabled.
What I found was that the problems started with this commit:
http://cgit.freedesktop.org/mesa/mesa/commit/?id=4539f8e20af286d1f521eb016c89c6d9af0b801c
That commit removed the S_0085F0_SMX_ACTION_ENA(1) and S_0085F0_SH_ACTION_ENA(1)
, among other things, from the cache invalidation. So I did some experiments
where
I enabled those and this is what I found.
It didn't see any difference if I put the SMX and SH under
R600_CONTEXT_FLUSH_AND_INV
as they were before the above commit or together with the other caches under
R600_CONTEXT_INVAL_READ_CACHES.
I also tried only SH, only SMX and both together and got fewer failures when
they both
were enabled. At the very least SH needs to be enabled, I'm not equally sure
about SMX,
but it does seem to help on some tests under spec/EXT_framebuffer_multisample.
Since the failures are random it is hard to know exactly what tests are
affected,
there are also unrelated issues that also cause random failures, but these tests
I'm reasonbly sure are fixed:
fast_color_clear/all-colors
fast_color_clear/redundant-clear
spec/!OpenGL 1.1/draw-pixels samples={2,4,6,8}
spec/!OpenGL 1.1/drawbuffer-modes
Then there are a bunch of test under spec/EXT_framebuffer_multisample that if
not
fixed at least have a much greater chance of passing. Those failures seem
unrelated
to VA though.
If my patch is wrong then this information can hopefully help someone else find
the real solution to these problems.
This was tested with a 6950 on mesa master
6d7abafdc86c6d8533bcb798465452c78c2694e8
//Martin
Martin Andersson (1):
r600g: Include SH and SMX when invalidating read caches
src/gallium/drivers/r600/r600_hw_context.c | 2 ++
1 file changed, 2 insertions(+)
--
1.8.3
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/mesa-dev
