shadowfacts/fonttools
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[varStore] use the same sorting for both input/output

Browse Source 
When the VarStore.optimize() algorithm was improved with the addition of a priority queue where pairs of encodings are sorted by decreasing gain (from merging one into the other), specifically with this commit https://github.com/fonttools/fonttools/commit/47ec18f788, the pre-sorting step of the todo list (before the queue itself was populated) was kept to make sure the algorithm produces stable results no matter the order of the inputs.
The optimizer's output was itself sorted, again for stability, but using a different key function with_sort_key() from the one used on the input todo list gain_sort_key().
The rationale for a distinct gain_sort_key, where encodings get sorted by maximum theoretical gain (decreasing, initally, when reverse=True was set, then incrasing as reverse was somehow dropped), is no longer needed now that a priority queue is used (which sorts by actual gains from merging specific pairs of encodings) and is a remnant of the previous algorithm.

I propose we keep some pre-sorting to ensure stability (in case the priority queue initially contains multiple pairs with the same gain), but we use the same width_sort_key that is used at the end.

Note this doesn't change the overall level of compression achieved by the optimizer, but makes the algorithm a bit less complicated, and easier to match in our alternative Rust implementation.
This commit is contained in:
Cosimo Lupo 2025-02-06 19:01:31 +00:00
parent 307b31250d
commit 548ffda37b
1 changed files with 7 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

19
Lib/fontTools/varLib/varStore.py
View File

@ -554,7 +554,11 @@ def VarStore_optimize(self, use_NO_VARIATION_INDEX=True, quantization=1):
# #
# - Put all encodings into a "todo" list. # - Put all encodings into a "todo" list.
# #
# - Sort todo list by decreasing gain (for stability). # - Sort todo list (for stability) by width_sort_key(), which is a tuple
# of the following items:
# * The "width" of the encoding.
# * The characteristic bitmap of the encoding, with higher-numbered
# columns compared first.
# #
# - Make a priority-queue of the gain from combining each two # - Make a priority-queue of the gain from combining each two
# encodings in the todo list. The priority queue is sorted by # encodings in the todo list. The priority queue is sorted by
@ -575,16 +579,7 @@ def VarStore_optimize(self, use_NO_VARIATION_INDEX=True, quantization=1):
# #
# The output is then sorted for stability, in the following way: # The output is then sorted for stability, in the following way:
# - The VarRegionList of the input is kept intact. # - The VarRegionList of the input is kept intact.
# - All encodings are sorted before the main algorithm, by # - The VarData is sorted by the same width_sort_key() used at the beginning.
# gain_key_sort(), which is a tuple of the following items:
# * The gain of the encoding.
# * The characteristic bitmap of the encoding, with higher-numbered
# columns compared first.
# - The VarData is sorted by width_sort_key(), which is a tuple
# of the following items:
# * The "width" of the encoding.
# * The characteristic bitmap of the encoding, with higher-numbered
# columns compared first.
# - Within each VarData, the items are sorted as vectors of numbers. # - Within each VarData, the items are sorted as vectors of numbers.
# #
# Finally, each VarData is optimized to remove the empty columns and # Finally, each VarData is optimized to remove the empty columns and
@ -626,7 +621,7 @@ def VarStore_optimize(self, use_NO_VARIATION_INDEX=True, quantization=1):
front_mapping[(major << 16) + minor] = row front_mapping[(major << 16) + minor] = row
# Prepare for the main algorithm. # Prepare for the main algorithm.
todo = sorted(encodings.values(), key=_Encoding.gain_sort_key) todo = sorted(encodings.values(), key=_Encoding.width_sort_key)
del encodings del encodings
# Repeatedly pick two best encodings to combine, and combine them. # Repeatedly pick two best encodings to combine, and combine them.