Depth-of-field and ribbons examples performance regress

[alteredq]

These two examples are a canaries for detecting performance changes as they have tons of objects and materials.

Seems something got messed up again - performance in DOF example went down from 14 to 8 fps (Chrome ANGLE) and from 33 to 15 fps (Chrome OpenGL).

Ribbons went down from 28 to 20 fps in ANGLE and from 38 to 29 fps in OpenGL.

(compared to original versions http://alteredqualia.com/three/examples/webgl_postprocessing_dof.html and http://alteredqualia.com/three/examples/ribbons_gl.html)

(all fps from Chrome 11 stable)

[alteredq]

I tried to profile this and it seems culprit is finish - takes 70 - 80 % of time!

Any particular reason why finish was kept? Seems like it's causing the same synchronization trouble we had with lens flares (and it didn't help with ANGLE texture update bug).

Even after commenting it out all examples seem to work fine (and performance is back to normal, also other examples are faster, including new ones - webgl_interactive_cubes went from 26 to 47 fps).

[mrdoob]

Copied from the commit thread:

Aha! So it was that one. A friend from Mandreel told me we were supposed to be calling this for sync reasons. You lose performance but you ensure sync. Which it used to be more obvious before, we can probably have that one commented until we spot something on these lines.

[alteredq]

Were there any noticeable problems coming from lack of synchronization before?

If there are use-cases where it makes some worse problems than significant performance degradation, there could be for example parameter to WebGLRenderer to enable it.

Also, wouldn't flush be better than finish for this purpose?

http://www.khronos.org/opengles/sdk/docs/man/xhtml/glFlush.xml

http://www.khronos.org/opengles/sdk/docs/man/xhtml/glFinish.xml

[mrdoob]

Some months ago the lack of sync was noticeable on the example that uses canvas/svg/webgl. Nowadays that example seems ok.

No idea about flush/finish... :S