[Mesa-dev] [PATCH 10/10] vc4/nir: Use the new unified io intrinsics

[Emil Velikov]

Hi Eric,

On 4 December 2015 at 02:03, Eric Anholt <eric at anholt.net> wrote:
> Jason Ekstrand <jason at jlekstrand.net> writes:
>
>> On Dec 3, 2015 3:47 PM, "Eric Anholt" <eric at anholt.net> wrote:
>>>
>>> Jason Ekstrand <jason at jlekstrand.net> writes:
>>>
>>> > Cc: Eric Anholt <eric at anholt.net>
>>>
>>> OK, I've pushed a branch of partial fixes for this series to nir-loads
>>> of my fdo tree, but it's still super broken.  I've spent too much of
>>> today on it, and this series was not ready for review.  There are still
>>> const_index[0] references all over that are obviously from intrinsics
>>> you've removed the const_index on.
>>
>> I'm going to start looking at this again tomorrow.  I've had other things
>> on my plate this week.  Rob already pointed me at some crashing shaders for
>> ir3 and his little standalone compiler. I'll start there.  Is there
>> something similar I can use to debug vc4?
>
> Right now from configure.ac I've got either targeting-hardware mode, or
> targeting-simulator mode.  What I need to add is something in between,
> where we use the ioctls from simulator mode so that people can run the
> driver on i965 systems, and then just don't make the simulator calls at
> the very end of the drawing pipeline since you don't have the simulator.
> Then people could even run shader-db without having the actual hardware,
> and that would get a lot of coverage of compiler changes.
>
> You could knock this together with some #ifdefs around all 7 mentions of
> "simpenrose", but I'll try to get a clean patch sent out.  I'm just
> trying to decide whether this should be a --enable argument or "oh, hey,
> you're not on arm, you must want either simulator or
> everything-but-simulator mode"
>
> I'd also like something a bit nicer than my symlink trick for loading
> the driver (right now I link the built vc4_dri.so to i965_dri.so in a
> subdirectory, and LIBGL_DRIVERS_PATH it). Maybe if I getenv() a name
> From loader.c for computing driver name, I could also skip a bunch of
> the i965 symbols I have to add.
>
What would it take to have things a bit like virgl -> ie. single vc4
gallium/driver but two winsys (vc4/drm and vc4/sim) ? Afaict one can
(should be able to) get things doing with this approach and toggle the
simulator "mode" via an env variable. I might have some time to add
LIBGL_DRIVER_NAME next week, although I'd rather avoid having such a
"hack".

On the standalone compiler thing if you shuffle things ala winsys
style as suggested you should be able to get a nice clean standalone
compiler like nouveau and freedreno.
I'll be glad to help, mostly on the ugly pipe-loader and autohell side
of things.

Thanks
Emil