[PATCH 6/7] etnaviv: apply non-alpha workaround for reduced masks

[Lucas Stach]

Am Dienstag, den 22.11.2016, 13:16 +0000 schrieb Russell King:
> On Tue, Nov 22, 2016 at 01:53:58PM +0100, Lucas Stach wrote:
> > Am Dienstag, den 22.11.2016, 12:17 +0000 schrieb Russell King:
> > > On Tue, Nov 22, 2016 at 12:44:19PM +0100, Lucas Stach wrote:
> > > > If accel_reduce_mask() succeeded in replacing the mask by constant
> > > > alpha the source format still needs to be switched to an alpha one
> > > > to get correct swizzles. We only want to avoid clobbering the alpha
> > > > mode setup done by accel_reduce_mask().
> > > 
> > > Definitely not.  You're correct that the test order is important, and
> > > the test order is intentionally the way it is, so that
> > > etnaviv_accel_reduce_mask() works.
> > > 
> > > The effect that your patch will have is to corrupt all the state which
> > > etnaviv_accel_reduce_mask() has computed to effect the combining of the
> > > constant A8 mask with the source - overriding its alpha with a constant
> > > 255.
> > > 
> > No. We are only switching the format to the alpha one in
> > etnaviv_workaround_nonalpha(). Replacing the alpha channel with constant
> > 255 is only done if etnaviv_blend_src_alpha_normal() returns true, which
> > isn't the case if etnaviv_accel_reduce_mask() changed the alpha op
> > already.
> 
> No.
> 
> If we have a constant A8 alpha mask, then etnaviv_accel_reduce_mask()
> sets the source alpha mode according to whether the source has alpha
> or not.
> 
> If the source has no alpha, we switch to global source alpha mode,
> which means we override whatever alpha was obtained from the source.
> This override happens _after_ the buggy hardware - which is the
> fetching of the alpha.
> 
> If the source has alpha, then we use scaled source alpha mode, so
> that the constant alpha scales the source alpha which will be correctly
> read by the GPU.
> 
> This logic is here to handle exactly the case you are purportedly
> trying to fix - it's been tested to work.
> 
> If you think otherwise, please supply a test case which fails, so that
> I can reproduce it here.
> 
I don't have individual testcases for each patch, but the rendercheck
commandline I'm using is:

rendercheck -f a8r8g8b8,x8r8g8b8,b8g8r8a8,b8g8r8x8,r8g8b8,b8g8r8,r5g5b5,b5g5r5 -t composite

This fails with the current unstable-devel branch. With the alpha
workaround patches applied this test passes.

Regards,
Lucas