Valve actually used cdak (http://www.pouet.net/prod.php?which=55758) to demo their prototype headset.

It was a specially optimized version to be able to run at full framerate on the hardware of that time. I have looked for it but it doesn't seem to be available anywhere.

There is a Shadertoy version but without the optimization and last time I tried I actually had to tweak it a little to make it compile.

What overhead is that? Raymarching does rely on raw GPU throughput, but the common methods of doing it work by issuing extremely few API calls, and leaving the GPU alone to do the work.

API overhead caused by WebGL is absolutely minuscule.

Hmmm. I need to go back and check but I'm fairly sure I've had better performance from Shadertoy shaders ported to native than I've even had running them via WebVR.

Why would raymarching and VR be different from anything else and VR? What is the overhead of webgl you are talking about?

Raymarching is heavily shader dependent which shouldn't be bottlenecked by webgl at all.

I thought there were some differences in how shaders were compiled. I'm no expert on this but my understanding was always that WebGl couldn't get the same raw performance as native apps - even at the shader level.

Some of the slides here seem to support this: https://www.khronos.org/assets/uploads/developers/library/20...

What part of the slides support that? All it says is 'loss of performance on some mobile devices'.