Rework vs Postcss speed bench

[MoOx]

I'm curious to know if speed diff with rework is it's just because of keeping use whitespace & coding style.
https://gist.github.com/MoOx/1b0d7d2bd987e0735e0e
Any comment ?

[ai]

Yeap, now PostCSS made few RegExp match on every node parsing to split spaces correctly. I think with tokenaize step before I can reduce RegExp and allow to autodetect at-rule content #41

[ai]

PostCSS will not be faster that Rework, but tokenizer get some speed boost.

Current:

PostCSS:     775 ms
CSSOM:       163 ms (4.8 times faster)
Rework:      314 ms (2.5 times faster)
Gonzales:    776 ms (1.0 times slower)
Gonzales PE: 761 ms (1.0 times faster)

My private tokenizer branch:

PostCSS:     480 ms
CSSOM:       224 ms (2.1 times faster)
Rework:      379 ms (1.3 times faster)
Gonzales:    735 ms (1.5 times slower)
Gonzales PE: 698 ms (1.5 times slower)

[MoOx]

Awesome news !
Great work :)

[yisibl]

💯

[ai]

Good news, everyone!

@Termina1 made really amazing work and speed up new tokenizer branch. PostCSS 3.0 will be fatser, that Rework.

Current PostCSS 2.x:

PostCSS:     426 ms
CSSOM:       107 ms (4.0 times faster)
Rework:      194 ms (2.2 times faster)
Gonzales:    468 ms (1.1 times slower)
Gonzales PE: 470 ms (1.1 times slower)

New tokenizer branch with @Termina1 fixes:

PostCSS:     139 ms
CSSOM:       99 ms  (1.4 times faster)
Rework:      181 ms (1.3 times slower)
Gonzales:    416 ms (3.0 times slower)
Gonzales PE: 406 ms (2.9 times slower)

[kizu]

That's some really awesome news!

[MoOx]

👍 👍 👍

[magsout]

Very good job

[andreypopp]

🚀 🚀 🚀

[ai]

Unfortunately, es6-transpiler has side effects and as I found out it slow down some JS code (mostly CSSOM and Rework). Currect tokenizer branch uses 6to5 and has more honest benchmark:

CSSOM:       27 ms  (2.3 times faster)
PostCSS:     62 ms  
Rework:      63 ms  (1.0 times slower)
Gonzales PE: 163 ms (2.6 times slower)
Gonzales:    170 ms (2.7 times slower)
Stylecow:    219 ms (3.5 times slower)

Also I start to use Bootstrap CSS, because it has whitespaces and comments indead of GitHub CSS. And now benchmark contains Stylecow.

[ai]

BTW, I suggested to @brainopia and @Termina1 2 boutles of Ruby porto award if they will increase PostCSS perfomance for 2 times (to 30-40 ms for parsing Bootstrap). Maybe somebody want to join this competition? ;)

[ai]

I disable node.source generation if user doesn’t want to add source maps and it increase perfomance up to 30%:

CSSOM:       27 ms  (1.8 times faster)
Mensch:      47 ms  (1.1 times faster)
PostCSS:     49 ms  
Rework:      64 ms  (1.3 times slower)
Gonzales PE: 171 ms (3.5 times slower)
Gonzales:    179 ms (3.6 times slower)
Stylecow:    206 ms (4.2 times slower)

[ai]

You will not belive, but @brainopia won perfomance contest! The new tokenizer is incredibly fast!

PostCSS 3 become the fastest CSS parser, written on JS! We are about 20 % faster, that even CSSOM and about 3 times faster, that Rework (don’t forget, that we still parse and store whitespaces).

PostCSS 3:   23 ms
CSSOM:       28 ms  (1.2 times slower)
Mensch:      47 ms  (2.0 times slower)
Rework:      62 ms  (2.7 times slower)
Stylecow:    122 ms (5.3 times slower)
PostCSS 2:   141 ms (6.1 times slower)
Gonzales PE: 162 ms (7.0 times slower)
Gonzales:    175 ms (7.5 times slower)

[MoOx]

This is so fucking cool. Thank you guys !!

[lionelB]

you guys rocks ! congrats @brainopia

[kizu]

Awesome! 🚢

Although, I guess PostCSS is still partial parser (got this definition from the Stoyan's talk here), as it won't parse all the stuff (all the simpleselectors, contents of :not, attr selectors, values like gradients, transforms etc, etc.), while parsers like Gonzales are trying to parse just everything.

But whatever. Impressive work is going on there!

[ai]

@kizu yeap, Gonzales parses more :).

[ai]

PostCSS 3.0 with new parser was released today. Please, test :).

[ai]

I add benchmark with preprocessors parsers: dce2f3e

PostCSS is few times faster even that libsass:

PostCSS:   19 ms
libsass:   106 ms  (5.5 times slower)
Less:      150 ms  (7.7 times slower)
Stylus:    278 ms  (14.3 times slower)
Ruby Sass: 1078 ms (55.5 times slower)

[MoOx]

This kind of bench should be done with some plugins added. For example you should consider use cssnext to compare sass & friends ;)

[ai]

@MoOx there is no cssnext features in Sass ;). Maybe when I add Sass style variables and @mixin we can compare with some logic inside.

[MoOx]

I mean that making bench with a parser (postss) & parser + features (sass, less etc) doesn't really make sense. cssnext that use postcss add some features like vars, math, custom media etc. so it's a bit more logic to compare old preproc to cssnext, more that comparing them to postcss only.

[ai]

@MoOx I agree, but cssnext vs Sass doesn’t make sense too ;). Correct benchmark should use same features.

[ai]

@MoOx hm. But maybe you are right. Even if test CSS will not contains cssnext features, cssnext AST lookups will make benchmark little better.

[MoOx]

Even if cssnext doesn't have mixins/if/loops, you have similar features like vars, math, color manipulations. It's better than nothing ^^

[MoOx]

This kind of bench are just sort of "just to let you know" :D

[ai]

Yeap, I updated benchmark: 7b0e467

PostCSS:   36 ms
libsass:   110 ms  (3.0 times slower)
Less:      157 ms  (4.3 times slower)
Stylus:    286 ms  (7.9 times slower)
Ruby Sass: 1092 ms (30.1 times slower)

[MoOx]

You should add the output to a file like BENCHMARKS.md or something like that & update it on each release to make this even more cool :)

[ai]

@MoOx maybe yu know some service, that can run benchmarks on every commit and output graph of perfomance?

[MoOx]

we can probably update your task to save the output + run the task on the prepublish npm script ?

[ai]

and where we will store results and print graph?

[MoOx]

I think the evolution is not the most interesting thing.
The current state on a release is the interesting thing.
The result (output) should be easily retrieved from a child process.

[ai]

@MoOx I want this graph for PostCSS developers to prevent perfomance regression :)

[MoOx]

Oh yeah. But that would be another problem. Can be drawn from collecting git history values on the versionned bench output ?