305% performance improvement with python bindings

[tayloraswift]

(Addresses #287)

There are probably many better ways of doing this than adding a field to hb_glyph_position_t but as a proof of concept, it is possible to achieve a roughly 300% increase in speed by having the glyph info/position lookups resolved in C rather than in Python through GObject.

def shape_new_buffer(extractor=lambda b: b):
    buf = hb.buffer_create ()
    hb.buffer_add_utf32 (buf, array.array('I', text.encode('utf-32')), 0, -1)
    hb.buffer_guess_segment_properties (buf)
    hb.shape (font, buf, [])
    return extractor(buf)

def unpack_hb_buffer1(HBB):
    return [(N.cluster, N.codepoint, P.x_advance, P.x_offset) for N, P in zip(hb.buffer_get_glyph_infos(HBB), hb.buffer_get_glyph_positions(HBB))]

def unpack_hb_buffer2(HBB):
    hb.buffer_compact_glyphs(HBB)
    return [P.compact for P in hb.buffer_get_glyph_positions(HBB)]

import timeit
t1 = timeit.timeit("shape_new_buffer(unpack_hb_buffer1)", number=1000, setup="from __main__ import shape_new_buffer, unpack_hb_buffer1")

t2 = timeit.timeit("shape_new_buffer(unpack_hb_buffer2)", number=1000, setup="from __main__ import shape_new_buffer, unpack_hb_buffer2")

print(t1)
print(t2)
print(t1/t2)

shaped_buffer = shape_new_buffer()

t3 = timeit.timeit("unpack_hb_buffer1(shaped_buffer)", number=1000, setup="from __main__ import shaped_buffer, unpack_hb_buffer1")

t4 = timeit.timeit("unpack_hb_buffer2(shaped_buffer)", number=1000, setup="from __main__ import shaped_buffer, unpack_hb_buffer2")

print(t3)
print(t4)
print(t3/t4)

0.5077813390016672
0.1881202860022313
2.6992375452567856 (overall stack performance improvement factor)

0.5054081650014268
0.16407813999830978
3.0802894584655407 (glyph extraction performance improvement factor)

This reduces the number of python attribute lookups required from 4 to 1 per glyph. Since sticking the values into the hb_glyph_position_t struct probably isn’t the best way to go, if there was a way to generate the array independently, testing shows it would probably achieve close to a 1000% percent performance improvement (since the bottleneck is still the one python attribute lookup).

[behdad]

You can convert the glyph positions and glyph infos into Python array.array objects and access those. Let me make it absolutely clear: no change to the library is going to be accepted for this purpose.

[tayloraswift]

@behdad how to convert <HarfBuzz.glyph_info_t object> into python array?