[Mesa-dev] ARB_arrays_of_arrays GLSL ES

[Timothy Arceri]

On Thu, 2015-06-25 at 10:07 -0700, Jason Ekstrand wrote:
> On Thu, Jun 25, 2015 at 1:19 AM, Timothy Arceri <
> t_arceri at yahoo.com.au> wrote:
> > On Wed, 2015-06-24 at 11:17 -0700, Jason Ekstrand wrote:
> > > On Sat, Jun 20, 2015 at 5:32 AM, Timothy Arceri <
> > > t_arceri at yahoo.com.au> wrote:
> > > > Hi all,
> > > > 
> > > > The restrictions in ES make the extension easier to implement 
> > > > so
> > > > I thought I'd try get this stuff reviewed an committed before
> > > > finishing
> > > > up the full extension.
> > > > The bits that I'm still working on for the desktop version are 
> > > > AoA
> > > > inputs
> > > > outputs, and interface blocks.
> > > > 
> > > > The only thing I know is definatly missing in this series for 
> > > > ES is
> > > > support for indirect indexing of samplers, but that didn't seem
> > > > like
> > > > something that should hold up the series.
> > > > 
> > > > Once the SSBO series lands (with a patch that restricts unsized
> > > > arrays)
> > > > then all the AoA ES conformance tests will pass.
> > > > 
> > > > There are already a bunch of piglit tests in git but I've just 
> > > > sent
> > > > a
> > > > series with all the patches still waiting review here:
> > > > http://lists.freedesktop.org/archives/piglit/2015
> > > > -June/016312.html
> > > > 
> > > > I haven't made a patch marking this as done yet because 
> > > > currently
> > > > the i965 backend takes a very long time trying to optimise some 
> > > > of
> > > > the
> > > > conformance tests. They still pass but they are taking 15
> > > > -minutes+
> > > > just
> > > > to compile so this really needs to be sorted out first. If 
> > > > someone
> > > > with
> > > > more knowledge in this area than me wants to take a look at 
> > > > this I
> > > > would
> > > > be greatful for being pointed in the right direction.
> > > 
> > > Can you try it with this patch:
> > > 
> > > http://lists.freedesktop.org/archives/mesa-dev/2015
> > > -June/086011.html
> > 
> > Hi Jason,
> > 
> > I tried that patch it didn't seem to make any difference, then I
> > noticed its for fs. The slowdown is currently happening in
> > vec4_live_variables. I've also noticed other large slowdowns with 
> > some
> > piglit tests I've been working on this time in the register 
> > allocate
> > code.
> 
> I just sent an equivalent patch for vec4:
> 
> http://patchwork.freedesktop.org/patch/52801/
> 
> I would love to know if it helps something so that I can put a good
> justification in the commit message beyond "the same as for FS".
> --Jason
> 

I ran the shortest of the two tests:

ES31-CTS.arrays_of_arrays.InteractionFunctionCalls1

With your patch: 3:31.40 elapsed

Without your patch: 6:49.62 elapsed

> > Once I finish fixing up some bugs with my current patchset I'll 
> > start
> > digging deeper, just thought since it's so noticeably slow it might 
> > be
> > easy for someone with good knowledge of the backend to point out 
> > where
> > I should start looking, or to say your doing this wrong.
> > 
> > One thing the slow shaders have in common is the use of arrays of
> > arrays in nested loops so maybe something funny is going on when 
> > they
> > go through the optimisation paths, I haven't spent much time 
> > looking at
> > any of these passes yet so its highly likely they don't work as
> > expected for arrays of arrays.
> > 
> > Thanks anyway,
> > Tim
> > 
> > 
> > > 
> > > > If useful the series is in the 'gles4' branch of the repo here:
> > > > https://github.com/tarceri/Mesa_arrays_of_arrays.git
> > > > 
> > > > Thanks,
> > > > Tim
> > > > _______________________________________________
> > > > mesa-dev mailing list
> > > > mesa-dev at lists.freedesktop.org
> > > > http://lists.freedesktop.org/mailman/listinfo/mesa-dev