Adshir's LocalRay mobile-friendly realtime raytracing solution *spawn

Isn't this software RT?

Maybe it's a library that can take advantage of HW features.

 

According to the article, it's definitely doesn't require RT HW to work.

Question remains as to whether it can use RT HW if present.

 

Well, skimmed the presentation video, just looking on the slides and ignoring any cloud related stuff.
What we see is sharp reflections only, same is in CryTeks demo. The scenes have lower geometric complexity, but it's 60 fps not 30. (Interesting: They mention 5700 XT, which i guess they have used for the shown robot demo.)

He mentions 'free skinning'. This makes me think he uses bounding spheres instead boxes. I do this too and AS update cost is indeed negligible because bounds need no refitting with animation.
Problem here is spheres a re a bad fit - still worse than bounding boxes in their worst cases. Solution can be to subdivide the spheres further so they better approximate a flat surface.
But i would not call this alone an 'invention'. Still, it's rarely used so maybe you can sell it as that if you keep details secret.
Pure speculation of mine... i'll look up his patents later the day...

I'm also doubtful Sony would sell this as 'HW RT'. I don't think it can compete with RTX, but it would be enough to be used in games as well. Still, NVs DXR fallback seems at least as powerful, from what we see.

 

i Wonder if this could be used on Xbox one PS4.
I guess so ?

 

Ok, after browsing patents i see nothing special. Good work but nothing exciting. Thing is, Sony could do this themselves easily, so i don't believe there is a connection.

 

Interesting tidbit from that link:

 

Performance is also not great for real games as the demos are simple demos. The 'mobile' example is 4 animated characters running at 40 fps on a Surface (Pro?) with only hard shadows. The high-end PC version is a single character. There's no evidence of it handling full-flavour (area lights, rough surfaces, etc) ray-tracing. Very reminiscent of the early Cell RT demos in simply being classic RT image generation.

 

Agree, but maybe there is some more hidden. The article says 20 people working on this for 7 years. That's a lot of manpower and time.
But i can't see soft shadows either. Maybe below the torus knot but don't think so. And no reflections of reflections.
Curious: Why a need to 'port' a game to LocalRay? What could cause a need to do significant changes to be compatible?

 

You mean running on the CPU only, no compute?

 

How do spherical bounding boxes equal free skinning? Problem with skinning is not just the high level bounding box which has to cover the entire potential range, but foremost the lower level parts as any structure not pre-baked for a specific pose can't properly bin triangles.

Or do you have to use "mid level" bounding boxes grouped by rig hierarchy rather than raw model space triangle binning? I suppose in that case you can actually estimate worst case bounding boxes for each group, and then refine / transform inner bonding volumes by current rig pose. The AS has to be aware of the rig a time of evaluation in order to be able to support "free skinning", and has to be baked with the full possible range of motion in mind.

But how would that be "for free"? You'd have to store the active rig inside the AS in order to avoid additional loads, you have to perform additional math for transforming effective bounding boxes according to current rig, etc.

 

In my case, using a hierarchy of surfels, each surfel also has a bounding sphere. So for characters i only need to ensure the bounds are large enough to bound children under any animation. (Ensuring large enough is a matter of precalculation.)
Animation is then just transforming the surfels, which i need to do anyways, so the BVH update is free for me. There is no extra complexity like mid levels or taking character bones into account.

The same would work for triangles. You would only need to transform the BVH using the same skinning system as for the mesh itself, but no bottom up dependencies.
It would also work for bounding boxes, in case you do not need to guarantee the parent box bounds child boxes, and the guarantee to bound only the triangles would suffice (which is the case for RT).

I think this will become the norm. Tracing speed suffers, but FF HW will compensate and dynamic geometry has no more extra cost.
So no refitting, just transform the bottom AS and rebuild the top levels as usual.
It might be worth to shrink the bounds so they bound triangles exactly to prevent the hit on tracing speed for the lowest levels of the tree.

 

It's just software, who cares. There's a 98% chance whatever neat thing it is, it's replicable in a convenient way that gets around whatever stupid patents they came up with. The asshole "inventor" that came up with it would've done better to just publish it as a paper than sit there and pretend he's going to get money off it, rendering patents never seem worth much beyond annoyance.