Presentation Notes from Todays Call

Demo https://garretrieger.github.io/ift-demo/

What the demo does:

   -

   Loads Roboto and Noto Sans SC fonts using the new static IFT.
   -

   Roboto is loaded using only shared brotli patches.
   -

   The Noto Sans SC font is loaded using a mix of IFTB and per-table shared
   brotli patches.
   -

      The non-outline tables are augmented using per table shared brotli
      patches.
      -

      While the outline tables (glyf + loca) are augmented using
      independent IFTB patches.
      -

   Which demonstrates the successful mixing of the two patching types.


In order to make this work I had to introduce some new
behaviour/requirements on the client and encoder:

   -

   I added the option to have an additional mapping table (tag ‘IFTX’) in
   addition to the existing mapping table (tag ‘IFT ‘). The entries from the
   two tables are unioned together for processing. The IFTB and shared brotli
   mapping entries are kept in separate tables. This allows us to avoid having
   the shared brotli and IFTB patches modifying the same table. Not the only
   way to solve this problem, but was the simplest for the demo.
   -

   I made changes to how the client selects and resolves patches, the new
   algorithm is:
   1.

      Scan the mapping tables and find all entries which intersect the set
      of new codepoints.
      2.

      Shared brotli patches have higher priority and are applied first. If
      more than one shared brotli patch matched in step a, then select only one
      and apply it.
      3.

      If a shared brotli patch was applied in b, go back to a.
      4.

      Apply all remaining IFTB patches.
      -

   At a high level, this requires all of the shared brotli patches to be
   fully resolved before attempting to apply IFTB patches (downloads can still
   happen in parallel). This ensures that tables such as maxp are in a good
   state before IFTB patches are applied.
   -

   Retain gids must be used to ensure that the gids used by IFTB patches
   are not invalidated.
   -

   IFTB table ordering is enforced after application of per table patches.
   -

   The current IFTB patching implementation doesn’t support changing loca
   size, so I had to force the loca in the initial subset to be the full
   length (and use long offsets).


What I’ve implemented so far:

   -

   An encoder which takes an IFTB font and a partitioning as an input and
   produces a mixed mode IFT font which incorporates the IFTB patches.
   -

   Client library (implementing patch selection and application) and wasm
   js client wrapper that drives the demo.


Lastly, me and Skef chatted earlier and came up with some additional ideas
for future exploration:

   -

   In the IFT mapping table have an old to new gid map. This allows
   relaxing the retain gids requirement and could allow for a more compact
   encoding.
   -

   Supporting variable axis augmentation alongside IFTB patches. This could
   potentially be done by having a shared brotli patch which replaces the
   mapping entries for the IFTB patches to point to new patch URLs which
   contain data for the modified design space (eg. updated gvar deltas). The
   list of patches loaded so far would be reset, so the client would
   redownload IFTB patches for any previously loaded codepoints, which will
   make the font correct for the updated design space.