Re: [PATCH 2/2] tests/kms_fbc_crc: Fill entire target framebuffer to adds crc FIXMES

From: "Ville Syrjälä" <ville.syrjala@linux.intel.com>
To: Daniel Vetter <daniel@ffwll.ch>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>,
	Intel Graphics Development <intel-gfx@lists.freedesktop.org>
Subject: Re: [PATCH 2/2] tests/kms_fbc_crc: Fill entire target framebuffer to adds crc FIXMES
Date: Thu, 26 Mar 2015 15:27:40 +0200	[thread overview]
Message-ID: <20150326132740.GO17419@intel.com> (raw)
In-Reply-To: <20150326131439.GC24805@phenom.ffwll.local>

On Thu, Mar 26, 2015 at 02:14:39PM +0100, Daniel Vetter wrote:
> On Thu, Mar 26, 2015 at 01:20:50PM +0200, Ville Syrjälä wrote:
> > On Thu, Mar 26, 2015 at 11:03:55AM +0100, Daniel Vetter wrote:
> > > On Thu, Mar 26, 2015 at 09:35:10AM +0200, Ville Syrjälä wrote:
> > > > On Wed, Mar 25, 2015 at 06:47:21PM -0300, Paulo Zanoni wrote:
> > > > > 2015-03-25 17:15 GMT-03:00 Daniel Vetter <daniel.vetter@ffwll.ch>:
> > > > > > And use the same colors for both flip and fills so that we can reuse
> > > > > > crcs. Some details:
> > > > > > - For the flip_and_foo tests flip twice so that we again start with
> > > > > >   the black framebuffer and hence have a real change when painting it
> > > > > >   white.
> > > > > 
> > > > > But you don't really check the CRC after the first flip, so if it's
> > > > > completely ignored by the display engine, we won't know. That still
> > > > > looks like a regression from the previous code we had.
> > > > > 
> > > > > > - The upload for the rendercopy source isn't fast, but hey I'm lazy.
> > > > > 
> > > > > You're also changing everything form single-pixel write to full-fb
> > > > > write, which is not necessarily a bad thing (although slower), but
> > > > > could at least be on the changelog.
> > > > 
> > > > The reason I riginally used very small operations was to make sure the
> > > > hardware catches precicesly such small operations. If we just blast
> > > > away the entire thing we can't tell if the hardware would even notice
> > > > a small change. Although since hardware tracking was declared to be
> > > > the evil that may not matter so much, except people still want to use
> > > > the GTT tracking for some reason.
> > > 
> > > We're still using the gtt cpu write tracking, which is the only hw
> > > tracking bit that works per-line-block. I guess if we see that one fail we
> > > can add a specific testcase for the pattern, but for now I kinda trust the
> > > hw in that regard actually. At least I haven't seen bugs like that.
> > 
> > We don't even test panning, which is where I think the gtt tracking code
> > is broken. And you can't catch bugs like that if you paint the entire
> > fb.
> 
> You could pan an entire width from an all black to an all white
> framebuffer. At least I don't see a fundamental difference in panning 1
> pixel and in panning by 1600.

We don't want to test invalidate due to panning (that's a flip which
means the hw will nuke the entire thing anyway), instead we want to
test invalidating when scanning out a panned fb and there's a gtt
write into the visible part.

-- 
Ville Syrjälä
Intel OTC
_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/intel-gfx