I just pushed a little benchmark program based on the eprun data files. I also pushed a corresponding benchmark of ICU (although it gives a slightly unfair advantage to ICU by preallocating a huge buffer for the output, whereas utf8proc figures out the output size dynamically).

Results on my machine:

$ cat bench.out 
Deutsch_.txt: 0.00390216
Japanese_.txt: 0.00458185
Korean_.txt: 0.00171503
Vietnamese_.txt: 0.0036274

$ cat icu.out 
Deutsch_.txt: 0.00011213
Japanese_.txt: 0.00026891
Korean_.txt: 9.649e-05
Vietnamese_.txt: 0.00151726

So, unless I am messing something up, ICU is significantly faster, albeit with a far more painful API.

Just to make sure that ICU is not "cheating" by doing some kind of caching that helps it for repeated normalization of the same string (the above benchmark loops 100x), I tried normalizing a single long string formed by concatenating the above files a few dozen times, and got 0.679s for utf8proc and 0.104s for ICU.

(Compiling with -O3 instead of -O2 makes only a slight (< 5%) difference.)

Would be nice to also benchmark against GNU libunistring and perhaps unf_ext from above.

benchmarking/profiling #12

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions