Optimize GlobeSurfaceShaderSet.getShaderProgram

[pjcozzi]

For 3K fullscreen view of Mount Everest, it was 2.3% to 2.5% of the profile (one of the top hot-spots).

Now it is ~0.07%.

@kring at this point, we might want to rename GlobeSurfaceShaderSet to something else and make it just a function.

[kring]

The steady-state speedup is nice, but this increases tile load time a bit and it makes tiles much larger. I think I can fix both of those without impacting the steady-state performance. I'm working on it now.

[pjcozzi]

Sounds good, thanks. Why does it increase load time? Web worker transfer?

As for tile size, my only concern would be L2 cache pollution (I actually think I am running into this elsewhere for setting uniforms) given the tile itself is tiny compared to the renderer resources.

[kring]

Why does it increase load time?

Because the first time we see a tile, we're now building up a giant shader string and then looking that up in the shader cache. Previously we built a smaller key and looked that up. It should be possible to avoid string lookups entirely.

my only concern would be L2 cache pollution

Yeah, exactly. We end up looking at a lot more tiles during tile selection than we actually end up rendering. I don't know if the size increase ends up having a measurable impact on the performance due to cache pollution, but as long as we're optimizing this up we might as well optimize it a bit better, since it's not much harder.

[pjcozzi]

Ah, I almost keep the older cache too, but a non-string lookup sounds good.

As for L2, I've thought about - in general - organizing objects so that the properties needed for traversal are tightly defined with a pointer to another object with all the other properties. I believe this is a common old-school game technique for fitting as many objects as possible into cache lines. I haven't tried this because we have bigger fish to fry, but keeping the tile tight for now sounds good. Just pack them into an int or two or whatever.

[kring]

Ok this should be in good shape now. New tiles can now get an existing shader without any string concatenations or comparisons. GlobeSurfaceTile only has one new field. Verifying that a tile's existing shader is still good requires three comparisons plus a bunch of bit operations, which should be better than a bunch of comparisons. Also, enabling/disabling lighting on the fly should actually work now. I believe it was broken in the previous version of this pull request.

[kring]

Someone else should probably merge this now that I've changed it pretty significantly.

[pjcozzi]

All looks good, thanks! By the way, is it possible to use the squared distance instead for computeDistanceToTile? That is another hot spot.

[kring]

By the way, is it possible to use the squared distance instead for computeDistanceToTile?

Yes, almost certainly. Actually the entire screenSpaceError computation can be optimized quite a bit. I think Insight3D precomputes the switching-distance-squared for each tile and just compares that, rather than doing any math per frame. I'm sure we can do the same. When I originally wrote it, the profiler showed that a negligible amount of time was spent there, so I didn't spend any time optimizing it. That has changed since, maybe because WebGL implementations have gotten faster or something.

[pjcozzi]

Want to add it to the roadmp (#526) - or just do it! 👍