[Mesa-dev] [PATCH v4 0/4] Fix OpenGL 1.3 big-endian support in r600g

[Oded Gabbay]

Hi,

v4:
Per marek's remarks I:

- removed all changes to evergreen_init_color_surface_rat(), as it is used
  only in compute and don't handle DB surfaces.

- Set V_0280A0_COLOR_16_16 and V_0280A0_COLOR_16_16_FLOAT to always do 8IN16 
  endian swap (array formats that still need to do byte swap).

v3 cover letter:

In v3 I removed all changes to gallium and mesa/st as they are no longer needed.
All the work to recognize if and when we need to do endian swapping is done 
inside the r600g driver and no additional flag was added to any structure.

This magic is done by understanding that the only use case where endian swapping
is *not* needed in a big-endian machine is for textures that are handled by the
DB unit in the GPU, because DB can't do swapping. Therefore, if a texture is 
marked as depth and not flushing, I disable the swap (treat it with GPU 
endianess). In all other cases, I configure endian swap. Credit is due to 
Marek Olšák who gave me this information. 

That is not the entire picture though. There are exceptions, such as array 
formats. There are also cases where it is not enough to do byte-swapping and 
I need to do component swapping or change the swizzling configuration. But all
that is done strictly in the r600g translation functions.

As a result, and because two patches that were approved have already been 
pushed to master, the patch-set got a lot smaller.

The results got a bit better than v2. Still there are some difference vs. 
x86-64 but the majority of cases were fixed.

Tests were done using piglit and by visual observance (I often looked at
the screen to verify the colors are correct).

For BE, I used a POWER7 machine with a CEDAR GL card (Evergreen).
For LE, I used a skylake machine with the same card.

I fixed most of the major tests (and some minor ones) - readpixsanity, 
draw-pixels, texwrap, copy-pixels and more. I started with 222 failed tests and
finished with 118 failed tests. Most of the tests I fixed were already working
in llvmpipe BE, so I knew that at least the core mesa code is working. In 
accordance to that, most of the remaining failed tests also don't work in 
llvmpipe BE, so there is probably something more fundamentaly wrong there.

Results of piglit tests are found in my fdo web folder:
https://people.freedesktop.org/~gabbayo/piglit_results/r600g-big-endian/

There are three comparisons, all of them done after reducing OpenGL version to 
1.3:

1. r600g-1.3-base-le-vs-be -
   Diff between LE and BE with upstream code to see the current status

2. r600g-1.3-le-vs-be -
   Diff between LE and BE with upstream + patch-set to see the remaining tests 
   to be fixed

3. r600g-be-1.3-all-fixes - 
   Diff between upstream code and upstream + patch-set, both running in BE. 
   This shows all the tests my patch-set fixed (and no regressions)

In addition, I of course checked for regressions on LE, without reducing OpenGL
version.

Thanks,

   Oded

Oded Gabbay (4):
  r600g/radeonsi: send endian info to format translation functions
  r600g: set endianess of 16/32-bit buffers according to do_endian_swap
  r600g: use do_endian_swap in color swapping functions
  r600g: use do_endian_swap in texture swapping function

 src/gallium/drivers/r600/evergreen_state.c    | 33 ++++++++------
 src/gallium/drivers/r600/r600_pipe.h          |  8 ++--
 src/gallium/drivers/r600/r600_state.c         | 27 ++++++++----
 src/gallium/drivers/r600/r600_state_common.c  | 62 +++++++++++++++++++++++----
 src/gallium/drivers/radeon/r600_pipe_common.h |  2 +-
 src/gallium/drivers/radeon/r600_texture.c     | 24 +++++++----
 src/gallium/drivers/radeonsi/si_state.c       |  4 +-
 7 files changed, 115 insertions(+), 45 deletions(-)

-- 
2.5.5