fonttools subset performance

[Jijun]

hi,
i want use fonttools to subbset our chinese truetype fonts, but i found it is so slowly for realtime web request.
the flavor are woff and woff2 for webpage。some code snippets as below:

from fontTools.subset import Options, Subsetter
from fontTools.ttLib import TTFont
import io

drop_tables_extra = ['GSUB', 'prep', 'vhea', 'vmtx', 'FFTM', 'GDEF', 'GPOS', 'kern', 'hdmx', 'VDMX', 'LTSH', 'DSIG', 'EBLC', 'EBDT', 'morx', 'mort']

from enum import Enum
class TTFormat(str, Enum):
    woff = "woff"
    woff2 = "woff2"
    ttf = "ttf"
fontDir='/tmp'
def subset_font(fontFamily: str, text : str, format : TTFormat):
    
    fontPath = f"{fontDir}/{fontFamily}.ttf"
    options = Options()
    options.flavor = format
    #options.desubroutinize = True
    options.hinting = False
    options.drop_tables += drop_tables_extra
    try:
        font = TTFont(fontPath, recalcTimestamp=False, lazy=False)
        subset_font = font
        if format ==TTFormat.woff or format == TTFormat.woff2:
            subset_font.flavor = format
            
        subsetter = Subsetter(options = options)
        subsetter.populate(text = text)
        subsetter.subset(subset_font)

        return subset_font
    except FileNotFoundError:
        pass
    except AssertionError:
        pass
    return None

if __name__ == '__main__' :
    font = subset_font('fangzheng_ktjt', 'Javatest', 'woff2')
    if font:
        buf = io.BytesIO()
        font.save(buf)
        buf.seek(0)
        print(buf)
        font.close()

so i use cProfile to find the problem.

   Ordered by: cumulative time

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
        1    0.000    0.000   19.195   19.195 FontService.py:22(subset_font)
        1    0.000    0.000   19.192   19.192 __init__.py:2720(subset)
    26/18    0.000    0.000   19.166    1.065 ttFont.py:369(__getitem__)
        1    0.000    0.000   19.163   19.163 __init__.py:2500(_prune_pre_subset)
        1    0.030    0.030   19.157   19.157 _g_l_y_f.py:57(decompile)
     8107    0.040    0.000   19.059    0.002 _g_l_y_f.py:531(expand)
     8100    1.156    0.000   18.909    0.002 _g_l_y_f.py:692(decompileCoordinates)
  3628674    3.245    0.000   12.489    0.000 _g_l_y_f.py:1452(_checkFloat)
  2419116    1.784    0.000   10.414    0.000 _g_l_y_f.py:1482(__setitem__)
  7257349    4.079    0.000    8.544    0.000 {built-in method builtins.any}
     8100    0.558    0.000    5.778    0.001 _g_l_y_f.py:1517(relativeToAbsolute)
     8100    0.016    0.000    5.044    0.001 _g_l_y_f.py:1464(zeros)
     8100    0.011    0.000    5.028    0.001 _g_l_y_f.py:1435(__init__)
     8100    0.486    0.000    5.017    0.001 _g_l_y_f.py:1504(extend)
 10886022    2.074    0.000    2.906    0.000 _g_l_y_f.py:1458(<genexpr>)
 10886022    1.558    0.000    1.558    0.000 _g_l_y_f.py:1455(<genexpr>)
     8100    1.156    0.000    1.537    0.000 _g_l_y_f.py:746(decompileCoordinatesRaw)
 10974219    1.227    0.000    1.227    0.000 {built-in method builtins.isinstance}
  3628674    0.700    0.000    0.700    0.000 _g_l_y_f.py:1443(isFloat)
  1209558    0.379    0.000    0.379    0.000 {method 'extend' of 'array.array' objects}
  1209566    0.233    0.000    0.325    0.000 py23.py:65(byteord)
  1217658    0.173    0.000    0.173    0.000 _g_l_y_f.py:744(<genexpr>)
     8129    0.016    0.000    0.103    0.000 sstruct.py:101(unpack2)

you can see some truetype check use most of the time.

anyone know how to optimze it? thanks.
fangzheng_ktjt.ttf.zip
.

[justvanrossum]

I would suggest to try hb-subset instead (part of the HarfBuzz project). (I haven't used it myself, and I'm assuming it has similar functionality, yet it's written in C++, so should be a lot faster.)