[Mesa-dev] [PATCH 00/14] isl: Add support for auxiliary surfaces

Jason Ekstrand jason at jlekstrand.net
Sat Jul 9 19:17:17 UTC 2016

This set of patches adds support in isl for the various types of auxiliary
surfaces.  Chad and I have been fiercely arguing back and forth for about
the last two weeks on how we want to handle all of the different aspects of
auxiliary surfaces.  This little series is the result of that discussion.
At this point, I think both of us are happy with the solution proposed

The basic idea is that an auxiliary surface stores the data in a sort of
compressed format.  The auxiliary surface is its own surface that has a
carefully chosen format and tiling that both accurately represents what is
going on in the compression and yields the correct surface layout
calculations.  For instance, with SKL color compression, every two bits in
the CCS corresponds to a cache-line pair in main surface and each cache
line in the CCS corresponds to a 16x16 block of cache line pairs.
Therefore, the CCS surface has a 2bpb compressed format that covers one
cache line pair and a tiling that has a logical size of 128el x 128el.

One of the primary changes required to make this work is to alter the
concept of the size of a tile.  Instead of a single tile size, we now have
both a logical size in surface elements and a physical size in bytes and
rows.  The logical size is used when carving the image up into tiles, but
the physical size is what is used when calculating the surface pitch.  This
seems to much better model how the hardware works where different types of
surfaces have different tile dimensions but most tiles are 128B x 32rows.
Not only does this help when defining auxiliary surfaces but it also takes
care of the oddity of W-tiled surfaces where you have to multiply the pitch
by 2.  Instead, we simply say that a W-tile is logically 64el x 64el but
has a physical size of 128B x 32rows.

Because CCS surfaces are 1 or 2 bits per block, we have to bring back the
isl_format_layout::bpb field.  There are 4 patches in this series to do so.
I'd like to commit the first three, but the other one (patch 9) I don't
much care about.  I included it in case we really want to get rid of
redundant information, but I'm ok with having both fields.

This series can be found on my freedesktop.org cgit here:


Jason Ekstrand (14):
  isl: Get rid of tiling_get_extent
  isl: Stop multiplying height by block size
  isl: Rework the way we handle surface padding
  isl: Rework the way we define tile sizes.
  isl: Change the physical size of a W-tile to 128x32
  isl: Bring back isl_format_layout::bpb
  isl: Use bpb in a few places where it's more natural than bs
  isl: Take bpb rather than bs in get_format_layout
  isl: Kill off isl_format_layout::bs
  isl: Add support for HiZ surfaces
  isl: Add support for multisample compression surfaces
  isl: Add support for color control surfaces
  isl: Add an auxiliary surface usage enum
  isl/state: Add support for handling color control surfaces

 src/intel/isl/gen_format_layout.py                 |   4 +-
 src/intel/isl/isl.c                                | 286 ++++++++++++---------
 src/intel/isl/isl.h                                | 116 ++++++++-
 src/intel/isl/isl_format_layout.csv                |  14 +
 src/intel/isl/isl_gen6.c                           |  10 +-
 src/intel/isl/isl_gen7.c                           |  23 +-
 src/intel/isl/isl_gen8.c                           |  16 +-
 src/intel/isl/isl_gen9.c                           |  26 +-
 src/intel/isl/isl_storage_image.c                  |   8 +-
 src/intel/isl/isl_surface_state.c                  |  54 +++-
 src/intel/vulkan/anv_formats.c                     |   4 +-
 src/intel/vulkan/anv_image.c                       |   4 +-
 src/intel/vulkan/anv_meta_blit2d.c                 |   2 +-
 src/intel/vulkan/anv_meta_copy.c                   |   4 +-
 src/intel/vulkan/genX_cmd_buffer.c                 |   7 +-
 .../drivers/dri/i965/brw_fs_surface_builder.cpp    |   4 +-
 16 files changed, 406 insertions(+), 176 deletions(-)


More information about the mesa-dev mailing list