This looks great! Are the calculations done in the GPU?

Not the example code I pasted into the blogpost.

To be honest, its not an obvious candidate for offloading the the GPU. Its a light task, the result is probably wanted by the CPU, and these days we probably have spare CPU cores that could be utilized while the GPU is busy.

However, with the move towards UMA for CPUs and GPUs - even the newly-announced ARM cores with Mali GPUs have coherent caches between CPU and GPU - and a move towards Vulkan, then the cost will change. As the cost of getting the result back to where its used for game logic diminishes, then this could increasingly be a candidate for GPU crunching :)

(In my code there's a divide-per-grid-square that could be tackled using scaling perhaps.)

> the result is probably wanted by the CPU -- I think this is a good reason to do it in the CPU.

> its not an obvious candidate for offloading the the GPU -- How? Shooting a ray to every vertex comes to my mind.

> its not an obvious candidate for offloading the the GPU -- How? Shooting a ray to every vertex comes to my mind.

I'm not following your meaning here ;)

My blog post describes how to do it in a sweep rather than 'shooting lines', for the performance reasons given in the blog post.

'Shooting lines' is pretty poor for performance because of data locality and cache pressure.

In general, locality is the big performance problem with 'ray tracing' in general. All attempts at speeding up ray tracing are about trying to make 'bundles' of adjacent rays flying in close formation that can be combined or computed together so as to try and give some locality to the problem.

My scanline approach (line as in cache array of adjacent memory, not line as in line-of-sight between eye and obstacle) is good for CPUs and good for GPUs. If you want to offload viewshed computation to the GPU, you ought strive for a scanline approach rather than 'shooting a ray to every vertex' too. My code ought be straightforward to port to a shader.

My bad! Of course ray tracing in the GPU would be costly; I dunno where I got that. I think my mind is still stuck on 2D after this, need to go back to 3D :)

When I had to do picking on a 3D terrain, I did a scanline approach too (https://bitbucket.org/rmsundaram/tryouts/src/master/CG/Terra...)

I did something even worse with WebGL, shooting a ray from every light source and view point to every pixel on the screen in the pixel shader (so something has to be in LOS and and lit to be seen).

http://i.imgur.com/ZJhr4Ll.png

The "player" is sitting on the "wall" of a room, to the right of it there are stairs going "downwards", it's all just a heightmap.

Needless to say it's kinda slow, but I was surprised how neat it came out considering how little I put in. I'm sure someone who knows what they're doing (which isn't me) could do the same a lot faster and prettier, since I really did it in the most plump way possible.

Me too. Though it's just rendered as a 2D raster: http://monkeybrains.net/viewshed

I'm trying to build something that can be used across the DSM we have of SF for better planning our wireless links.

Oh, interesting!

Shouldn't it be possible to then feed that into a fitness function and randomly "evolve" the "perfect" distribution, or is that silly?

Totally possible. Though getting data out of WebGL can be hairy. And actually getting on the rooftops of the ideal buildings is a whole nother story.

Make ease of access (and how much it's wanted there, and so on) part of the fitness function, too :)

And future configurations could also take into account how many things you'd have to move, or whatnot (I have no idea about what you're doing, I just imagine that's not something you do once and never change, right?).