Flex: FlexBox For Reason

This is a native Reason implementation of CSS Flexbox layout. It is a port of the Yoga project.

It can also be compiled to JavaScript.

Building and Installing

Install esy

Esy is "npm for native Reason". You install dependencies and make sure they're prepared via the esy install, and esy build commands.

npm install -g esy

Clone, and Build

git clone git+ssh://git@github.com/jordwalke/flex.git
cd flex
esy install
esy build

Installing may take several minutes if this is your first time doing native Reason development. Next time you create a project it will be fast.

Rebuilding

Edit source and then just run:

esy build

Running The Tests:

There are known failures where we are not compliant with browser's current implementations - for some of those our implementation may be more compliant with the W3C spec.

./test              # Run the test in native. 
./test byte         # Run the tests in compiled bytecode. 
./test jsc          # Run tests in JavaScriptCore. 
./test jscWithJit   # Run tests in JavaScriptCore with JIT. 
./test v8           # Run tests in v8. 

The tests only run the layout tests (not the measurement test) so that we can run the tests as a way to approximate startup time. (Otherwise JS engines would be dually punished since they'd have to start up twice, whereas native targets don't really have a "startup" process).

Benchmarking:

While benchmarking, close any programs that are running so they do not interfere with benchmarks (especially Chrome, or music/media apps).

After running esy build:

./bench              # Run the test in native. 
./bench byte         # Run the tests in compiled bytecode. 
./bench jsc          # Run tests in JavaScriptCore. 
./bench jscWithJit   # Run tests in JavaScriptCore with JIT. 
./bench v8           # Run tests in v8. 

It will print the mean and median running time of the layout test runs. The median is not reliably on any of the JS targets due to lack of ubiquitous high precision timers in JS environments, so the mean is the only universally comparable metric - still, the median is useful when comparing against the C implementation of Yoga, and

More Debuggable JavaScript

If you need to debug an issue in the JS behavior, generate more readable JavaScript by explicitly asking for --dev versions of a JS file.

esy build jbuilder build --dev ./src/layout-test-fixed-encoding/LayoutTestFixedEncoding.bc.js
node ./src/layout-test-fixed-encoding/LayoutTestFixedEncoding.bc.js

After doing so, and before you run any benchmarks on the native/byte versions, completely rm _build to make sure the --dev flag didn't cause deoptimized native binaries.

Crosscompiling to armv7 and x86

This was as simple as 3 steps, but it needs to be brought up to date and retested. It generates a pure C API that any library can consume.

1. Install docker
2. Run build_android.sh
3. Artifacts will be generated in _dist directory

Building and running cstub tests

cd stub_test
make gtest
cd ../
npm run stubtest

Some sample benchmark results (for those curious):

Your mileage may vary, but here's an example of the different performance you can observe when compiling Reason layout to either native assembly, byte code, and various JS engines. Comparison against a pure-C implementation is also included.

This test uses the FixedPoint encoding of layout data. native currently performs better on FixedPoint encoding (Float encoding will be improved shortly - so that neither is significantly advantaged).

Update: Retesting on the newer compiler shows a greater difference between ocamlopt and ocamlc. I will investigate further once I have the C version rebuilt on the same machine.

Comparison to C Implementation.

The native C implementation in the table above uses the official Yoga implementation, but with a benchmark test suite that matches the flex benchmark tests. You can run the C benchmark yourself by cloning the official Yoga project, checking out the revision at the time flex was ported from Yoga, and replacing the official benchmark with this. It's important to check out the revision at the time this was ported to Reason because only then are the algorithms identical (this cross-language benchmark attempts as much as possible to only change one thing at a time - the language of implementation).

Here are several datapoints gathered from running the flex test suite, and comparing it with those same tests executed in the C implementation. We include both the fixed point and the floating point encodings. All tests performed on ocamlopt(native Reason).

The Reason implementation of layout hasn't been optimized at all. Cleaning up the code to be less imperative will likely make it even faster. Upgrading to F-lambda will also likely make both fixed point, and especially floating point, faster.

    ┌────────────────────────────┬────────────────────────────┬────────────────────────────┐
    │    Reason Layout Float     |      C Implementation      |   ReasonLayout FixedPoint  |
    ├────────────────────────────┼────────────────────────────┼────────────────────────────┤
    │                            |                            |                            |
    │   ┌──────────┬─────────┐   |   ┌──────────┬─────────┐   |   ┌──────────┬─────────┐   |
    │   │ AverageMs│ MedianMs│   |   │ AverageMs│ MedianMs│   |   │ AverageMs│ MedianMs│   |
    │   ├──────────┼─────────┤   |   ├──────────┼─────────┤   |   ├──────────┼─────────┤   |
    │   │ 0.32469  │ 0.296   │   |   │ 0.39579  │ 0.350   │   |   │ 0.27288  │ 0.245   │   |
    │   │ 0.36458  │ 0.328   │   |   │ 0.33136  │ 0.299   │   |   │ 0.25957  │ 0.238   │   |
    │   │ 0.36345  │ 0.329   │   |   │ 0.33733  │ 0.325   │   |   │ 0.26478  │ 0.241   │   |
    │   │ 0.31851  │ 0.265   │   |   │ 0.31649  │ 0.289   │   |   │ 0.25594  │ 0.234   │   |
    │   │ 0.3742   │ 0.331   │   |   │ 0.29206  │ 0.281   │   |   │ 0.27518  │ 0.241   │   |
    │   │ 0.33161  │ 0.307   │   |   │ 0.34415  │ 0.319   │   |   │ 0.23987  │ 0.216   │   |
    │   │ 0.34044  │ 0.303   │   |   │ 0.32270  │ 0.282   │   |   │ 0.25471  │ 0.233   │   |
    │   │ 0.38866  │ 0.341   │   |   │ 0.35345  │ 0.332   │   |   │ 0.32457  │ 0.297   │   |  
    │   └──────────┴─────────┘   |   └──────────┴─────────┘   |   └──────────┴─────────┘   |
    │                            |                            |                            | 
    └────────────────────────────┴────────────────────────────┴────────────────────────────┘

Interpreting the Data

This benchmark is very useful because it tests the core compiler and runtimes of various language ecosystems - knowing that we are compiling a single exact algorithm to different runtimes. In general the JS that is compiled here (by js_of_ocaml), is somewhat idiomatic JS, and generally performs slightly better than what you would write by hand in JS. In this case, because of how the original code is written in Reason (which was ported from C and is not very functional), many of js_of_ocaml's optimization opportunities are lost, so it's likely that this JS output is slightly slower than what you'd write by hand (I'd estimate 20% max). We can easily fix that in the Reason source, and not only will the JS output's performance benefit, but likely the native/byte targets as well.

Still, regardless of what we do to improve the JS output, it's likely not going to recover the order(s) of magnitude.

One interesting fact is that flex is compiles the byte version using ocamlc, which produces a Virtual Machine byte code, but this VM is interesting in that it does not use any runtime JIT to achieve decent performance. If a JIT is out of the question for you (or you just don't want to wait for JIT warmup at startup time), then byte is a good option for you. If raw performance and startup time are important to you, then compiling to native is the best option.

All of the JS benchmarks are compiled with js_of_ocaml, and --opt 3. You can adjust the jsoo flags in package.json. v8's performance is especially impacted by --opt 3 vs. --opt 1. Note that v8 only has the option to enable the JIT (for now).

Theories

In past benchmark experiments I've performed with jsc/v8/ocamlopt, I've found that jsc with a JIT can be competitive with ocamlopt, under the right circumstances, but those other experiments were very allocation heavy. The layout algorithm is very computationally heavy, and not allocation heavy. It seems that ocamlopt does well in a wide variety of cases (elegant allocation heavy function style, or dirty imperative systems work/computation).

Startup Time

Different runtimes have different startup time characteristics. The following benchmarks give a very rough idea of how long each of the respective language runtimes take to startup. This isn't measuring how fast the languages initialize their environments, but rather, the languages along with their containing ecosystems (such as Node setup, and the JSC harness). v8's startup time is at an inherent disadvantage because it includes to initialize the VM along with many built in libraries. Still, this is a good idea of what you could expect your relative startup time overhead to be for a very small (1000 line) app in these respective environments.

One thing not accounted for here, is how the startup time grows (or doesn't) with the amount of code added. JS engines must parse their code and generate some intermediate representation at startup time, so that will cause large apps to slow down during the startup phase. This is much less of an issue for natively compiled code, where all compilation has been done ahead of time.

Measurements use the time command line program (real).

To test startup time, change the number numIterations to 1 in LayoutTestFixedEncoding.re, rerun esy build, then do ./bench or ./bench byte etc.

Multiple layout representations:

There are two implementations of layout data encodings, one uses fixed point with explicit rounding, and the other uses floating point. The default is currently fixed point. You can toggle between the two by doing the following:

• To test the floating point representation, open ./src/LayoutValue.re and uncomment the upper half, and comment out the lower half.
• Open package.json and replace src/LayoutTestFixedEncoding.re with src/LayoutTestFloatEncoding.re.
• Run npm run build, npm run test, npm run bench, etc.

The performance in ocamlopt would decrease by about 20% when switching to floating point representation. That can be fixed in the floating point representation without having to use fixed point representation, but I just haven't gotten to that yet.

More accurate benchmarks for native compilation

flex depends on Core_bench, which allows much better isolation of benchmarks, and ensures that various batch sizes are tested to attempt to eliminate misleading measurements caused by convenient (or inconvenient) batch sizes.

Since it uses native hooks, Core_bench won't work in byte or any of the JS modes, to enable Core_bench:

• Inside of src/LayoutTestFixedEncoding.re and src/LayoutTestFloatEncoding.re:
• Comment out include FakeCore; below.
• Uncomment open Core_bench.Std;
• Delete the two targets in package.json (byteTarget, jsTarget)
• Run npm run build, then npm run bench

Adding Tests

Instead of manually writing a test which ensures parity with web implementations of flexbox you can run gentest/gentest.sh to generated a test for you. After running gentest/gentest.sh a editor window should pop open (make sure you have $EDITOR env variable exported). Here you can write html which you want to verify in CSSLayout, such as the following.

Then put something like this in the editor that pops up:

<div style="width: 100px; height: 100px; align-items: center;">
  <div style="width: 50px; height: 50px;"></div>
</div>

Once saving and exiting the editor window the script will open a browser window. From here open the developer console and you should see two buttons, that will copy tests cases to your clipboard. Copy the fixed point test and paste it into ./src/LayoutTestFixedEncoding.re, and copy/paste the floating point test into ./src/LayoutTestFloatEncoding.re. Run npm run build and npm run test.

The ASCII output paints a pseudo-accurate picture of any broken layouts.

See the Yoga's README/docs for more information about the limitations and special defaults of flex. flex is a direct port of that project from C to Reason on ocamlopt.

Profiling on Mac:

This command will build the artifacts with the proper profiling/debug symbol flags, and will change YGAlignContentTest to wrap one test case in a large for loop. It should be suitable for either mac or linux.

npm run stubtestbindingsperf

This command is useful for generating profiling traces on mac:

Warning: You currently must clean the build before switching between npm run stubtest and npm run stubtestbindingsperf.

sudo instruments -v -t 'Time Profiler'  -D ~/Desktop/yourTrace.trace _build/test/test
# To open in Instruments, need to restore perms. 
sudo chown -R you:staff ~/Desktop/yourTrace.trace
# Now open Instruments.app and open the trace. 

Then open Instruments.app (GUI), and open file test.trace.

• Enable "High Frequency" and disable "Hide System Libraries" in the GUI to see all the content.
• Click on the gear, then click "Invert Call Tree" and "Top Functions".
• Right click on the table view header and enable Self # Samples and # Samples. It's important to open these to get the full picture in the abscense of perfect profiling symbols.
• Click on "Self #Samples" to sort so that the highest numbers are first.

The result is that you have the bottlenecks at the top. When you drill down in the tree, it tells you who's calling that bottleneck. Many of the symbols are still obfuscated (4.03 improves that). Open up all the .s files that ocamlopt dumped so that when you see obfuscated symbols in the Instruments GUI, you can search for that symbol in all the .s files you have open. Double click on the actual symbol in instruments and it will show you what the assembly looks like and you can match it to the symbol you have open in your editor. There may be multiple .L123 across all the files so you can cross reference Instruments (double click) which shows you the assembly. 4.03 is said to greatly improve the source locations when profiling so it tells you the function names instead of assembly locations more often than not.

Thanks

A very special Thank You goes out to Yehor Lvivski, who was kind enough to donate his npm package name flex. You may still install the previous package flex under its respective versions. All versions 1.0.0 and greater will refer to this package - the flex box layout computation algorithm written in Reason.