It was quite a pain - ArrayBuffers / Typed Arrays are fussy about byte alignment - but seems to be worth it.
Parsing times are down (modest gains for small objects / Firefox, bigger gains for large objects / Chrome, performance gains from 10% - 200%).
Also as a positive side effect, changing interleaved face arrays into uniform arrays seems to have helped a bit with gzip compression (about 6% smaller size for gzipped 526K Lucy).
Work-in-progress => do not merge yet.
OBJ converter, JSONLoader, BinaryLoader and some examples are done. Still need to do Blender exporter, SceneLoader and more examples.
Advantages:
- faster parse times
- bypass array initializer size limit in Firefox
- no more crashes in Chrome 15 (caused by too many concurrently active workers)
Disadvantages:
- more browser UI freezing
- multiple models tend to appear all at once with the last one (instead of popping up progressively)
Problem was having too many concurrent workers opened at once. Garbage collection apparently works very slowly on workers. Solution was to close them manually, after model data is sent to main thread.
Also some progress with Blender scene exporter. Some scenes now export correctly, though monkey test scene got worse, even with less hacky code.
If you want to get edges, run converter like with "-e" flag:
python convert_obj_three.py -i model.obj -o model.js -e
Output seems to match Blender exporter.
If you want to use binary format, convert OBJ models using "convert_obj_threejs_slim.py" with "-t binary" option.
This will now create two files:
JS part with materials
BIN part with binary buffers
Binary models are loaded like this:
loader.loadBinary( 'obj/lucy/Lucy250k_bin.js', function( geometry ) { createScene( geometry, s ) }, "obj/lucy" );
(difference to ascii format is that url root must be always specified so loader can find binary buffers file, like it already was for textures)
Good news:
- raw binary files are quite smaller than raw ascii files (about half size)
- loading times went down about 20-25%
- now also Firefox can handle 250k triangle mesh (it just bugs about long running script), with ascii format it threw "array initialiser too large" exception
Mixed news:
- gzipped binary files are only about 25% smaller than gzipped ascii files so actual benefits are smaller,
also it's more likely server will have JS gzipping on by default, while you need to set it up for
other formats (could be hacked around by naming binary files with JS extension instead of BIN?)
Bad news:
- browser blocking is back :( Not that it ever went completely away (for large models object creation is more costly than loading and it is blocking), but it was slightly better with ascii loader (where data is loaded as worker's JS body).
- this comes from having to use Ajax to load binary data
- loading binary data by Ajax from within worker didn't really help, it was in fact slightly slower and browser did freeze :(
- can't easily embed binary data in JSON (would need to encode it which would make it bigger, defying the purpose)
- Loader got fatter
Also in this commit: renamed Loader.loadAsync() => Loader.loadAsciiOld() and Loader.loadWorker() = Loader.loadAscii() so that names correspond to model formats instead of underlying implementation.
loadAsciiOld - JS exported by Blender and old OBJ converter
loadAscii - JSON created by slim OBJ converter (-t ascii)
loadBinary - JSON + BIN created by slim OBJ converter (-t binary)
TODO:
- look into UV coordinates, should be indexed the same way as vertices, normals and faces are (instead of unrolled per each face)
- look into HTML5 File API, binary blobs could help
The idea is that later there will be more loaders which would load different formats (like OBJLoader, ColladaLoader).
Usage (async JS):
var loader = new THREE.Loader();
loader.loadAsync( "obj/torus/Torus.js", function() { createScene( new Torus() ) } );
Usage (web worker):
var loader = new THREE.Loader();
loader.loadWorker( "obj/torus/Torus_slim.js", function( geometry ) { createScene( geometry ) } );
Web worker loader is useful for large meshes, where it allows browser to stay responsive for longer time and also it can handle larger meshes than async JS loader.
Web worker loader needs a simpler format of the model. Web workers can communicate with the main application only via message passing, where messages are JSON objects.
There is a new version of OBJ -> Three.js converter (convert_obj_threejs_slim.py) which can produce model in format needed for web workers.
All examples which were using models from OBJ converter were refactored to use Loader.
Except large mesh example, all examples are using just async JS loading. Web worker loading is there, it's just commented out, as it's a bit pain for local development.
Chrome doesn't allow to run web workers from pages accessed via file://, so you need either to run it with "--allow-file-access-from-files" flag, or access examples via local server (http://localhost/example.html).
alteredq
Mr.doob