[Mesa-dev] Ivybridge support for ARB_transform_feedback2

[Paul Berry]

On 22 October 2013 10:07, Ian Romanick <idr at freedesktop.org> wrote:

> On 10/21/2013 11:48 AM, Kenneth Graunke wrote:
> > On 10/21/2013 08:40 AM, Ian Romanick wrote:
> >> On 10/17/2013 11:09 PM, Kenneth Graunke wrote:
> >>> Here's my implementation of ARB_transform_feedback2.  I believe it's
> >>> complete; it passes all of our Piglit tests and a lot of Intel's
> >>> oglconform tests.
> >>>
> >>> This should work out of the box on Ivybridge and Baytrail.  It won't
> >>> work on Haswell at the moment, due to restrictions on register writes
> >>> (to be solved in a future kernel version).  Patch 9 will need to be
> >>> replaced with something that detects whether or not we can write
> >>> registers from userspace batchbuffers.
> >>>
> >>> In the meantime, I figured I'd send out the rest for review.
> >>>
> >>> Porting this back to Sandybridge is probably doable, but annoying.
> >>> Sandybridge doesn't have the MI_LOAD_REGISTER_MEM command, so we'd have
> >>> to map the buffers and use MI_LOAD_REGISTER_IMM.  Seems pretty gross.
> >>> Plus, transform feedback is done very differently pre-Ivybridge.  I'm
> >>> not sure it's worth it, seeing as it's a GL 4.0 feature.
> >>
> >> I assume this is just to support glDrawTransformFeedback?
> >
> > No, it's to support glResumeTransformFeedback.
>
> This is also part of OpenGL ES 3.0, so we need to fix it. :(
>

Wow, you're right.  For some reason I had convinced myself that OpenGL ES
3.0 didn't support BindTransformFeedback(), therefore it didn't have to
support switching transform feedback objects while paused.  But I just
checked the spec and I was totally wrong about that.

So yeah, we need to fix this even for Sandy Bridge.


>
> > glDrawTransformFeedback actually just reads pipeline statistics counters
> > and leaves them free-running.
> >
> >> Can you add that information to
> http://dri.freedesktop.org/wiki/I965Todo/ ?
> >
> > Actually, I'm probably wrong...on Gen7 we use MI_LOAD_REGISTER_MEM to
> > copy offsets into the SO_WRITE_OFFSET(n) registers.  But on Sandybridge,
> > XFB is done using the geometry shader, so it works entirely differently.
> >  I don't think there is a register to load.
>

There is a register on Sandy Bridge, but instead of writing to it using
MI_LOAD_REGISTER_*, you're supposed to write to it using
3DSTATE_GS_SVB_INDEX (see brw_begin_transform_feedback() in gen6_sol.c).
That's kind of unfortunate for pause/resume because 3DSTATE_GS_SVB_INDEX
doesn't support anything akin to MI_LOAD_REGISTER_MEM.  It's possible that
you can *also* write to those registers directly using
MI_LOAD_REGISTER_MEM, but I haven't investigated that.


> >
> > I'll just let whoever looks into it figure it out.  Not much insight
> anyway.
> >
> > --Ken
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/mesa-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.freedesktop.org/archives/mesa-dev/attachments/20131022/6705b0d7/attachment.html>