[Mesa-dev] [PATCH] etnaviv: bugfix: Don't do resolve-in-place without valid TS

[Lucas Stach]

Am Mittwoch, den 01.11.2017, 12:34 +0100 schrieb Wladimir J. van der Laan:
> On Wed, Nov 01, 2017 at 11:52:55AM +0100, Lucas Stach wrote:
> > Am Mittwoch, den 01.11.2017, 11:17 +0100 schrieb Wladimir J. van der Laan:
> > > GC3000 resolve-in-place assumes that the TS state is configured.
> > > If it is not, this will result in MMU errors. This is especially
> > > apparent when using glGenMipmaps().
> > > 
> > > Fixes a problem introduced in 78ade659569ee6fe9bd244170956139f19dd8c6c.
> > > 
> > > > Signed-off-by: Wladimir J. van der Laan <laanwj at gmail.com>
> > > 
> > > ---
> > >  src/gallium/drivers/etnaviv/etnaviv_clear_blit.c | 4 ++++
> > >  src/gallium/drivers/etnaviv/etnaviv_emit.c       | 4 ++++
> > >  src/gallium/drivers/etnaviv/etnaviv_rs.c         | 1 +
> > >  src/gallium/drivers/etnaviv/etnaviv_rs.h         | 2 ++
> > >  4 files changed, 11 insertions(+)
> > > 
> > > Ooops. This seems like an obvious oversight but I hadn't thought we would get
> > > into this path at all when there is no TS to "flush".
> > 
> > And that's probably what we should fix. The self-resolve cases on
> > resource flush and sampler update don't check the TS status, but they
> > are only useful if there is a valid TS.
> > 
> > With the change you did here we are still wasting bandwidth for a no-op 
> > blit on older cores like GC880 when generating mipmaps.
> 
> Yes, just a bugfix (my original commit did not introduce the higher-level
> behavior). This particular case should not result in a MMU error. If we fix the
> higher level, then this could be replaced with an assertion instead.
> 
> On the longer run I'd personally prefer to make "Flush resource level TS" a
> separate, explicit operation, for example a method on the context, and not make
> it go through the blit path with source==destination. It's a hardware operation
> implemented differently on GC3000 and GC7000, that just happens to use the RS
> blit on <GC3000.

Agreed. I'll make sure this commit lands in 17.3 and will work on a
follow up change to fix the high-level behavior.

Regards,
Lucas