[Mesa-dev] [PATCH 0/5] ac, radeonsi: workarounds for various cubemap issues

[Nicolai Hähnle]

Hi all,

Clearly, the Vulkan CTS inherited the thoroughness of dEQP's ES tests
in this area, because the first two fixes were already done for radv.

The last fix is for a quite amusing issue. One of the dEQP tests
chooses a viewport pseudo-randomly for rendering a sqare. With some of
the pseudo-random choices, one pixel at the corner of one of the
triangles ends up in a quad with three helper invocations, one of
which lies outside the triangle in such a way that its layer index
becomes smaller than -0.5. So its layer is rounded to -1.

This triggers a hardware bug that exists on <= VI, where clamping
of a negative cube map layer number causes the face number to be
corrupted. This in turn messes with the automatic LOD calculation.
We end up choosing the wrong mip level and failing the test.

Wave VGPR dumping definitely helped to make sense of that one :)

Please review!

Thanks,
Nicolai
--
 src/amd/common/ac_llvm_build.c               |  42 ++++++-
 src/amd/common/ac_llvm_build.h               |  10 +-
 src/amd/common/ac_nir_to_llvm.c              |  13 +-
 src/gallium/drivers/radeonsi/si_pipe.c       |   1 -
 src/gallium/drivers/radeonsi/si_pipe.h       |   1 -
 src/gallium/drivers/radeonsi/si_shader.c     |   3 +-
 .../drivers/radeonsi/si_shader_tgsi_mem.c    | 107 ++++++++++++++++-
 .../drivers/radeonsi/si_shader_tgsi_setup.c  |   2 +-
 8 files changed, 152 insertions(+), 27 deletions(-)