Hybrid Bidir

[J the Ninja]

I know that it isn't rendering correct images right now, but does hybrid bidir support all the features that classic does?

Same limitations as hybrid path, aka no instancing, no non-mesh shapes (meaning no microdisplacement, no sphere-lights)

(I think)

[J the Ninja]

Ok, glorious test results!

I fired things up with my old "locked tower" scene (~180k tris + a good deal of textures + env map).

30kS/s vs 57kS/s, and this is with a 5770! (that's over 85% faster!)

RAM use was 1.36GB for sampler, 2.14GB for hybrid. Btw, Dade, did you already implement all the stuff you mentioned for lowering hybrid's RAM use, or is that expected to be lower in the future?

Sampler:

Hybrid:

[Dade]

Btw, Dade, did you already implement all the stuff you mentioned for lowering hybrid's RAM use, or is that expected to be lower in the future?

The memory usage should be the same in the future too but keep in mind it is a fixed cost: for instance, if you render a 8GB scene, it will require "only" 8+1.2GB of memory (i.e. the fixed memory cost doesn't scale up with the scene complexity).

[J the Ninja]

Getting closer. Still some weird dark areas and eff is still wrong, but the color basically is correct with the latest change.

Also, RAM use is down to 1.95GB, and sampling rate is up a tiny bit. It crossed over "double the speed sampler got". Interestingly, a cube on a plane takes 2.8GB of RAM. So the "hybrid tax" is lower on more complex scenes?

[Eros]

Hi All

I was playing around with a scene, and naturally liked to try of the hybrid bi-dir

I did see a significant increase in speed, however found a few issues with the render.

Traditional CPU

This is the Normal Traditional

and here is the same with hybrid bi-dir

Hybrid

The main differences are the fireflies and the absence of the nice light patch due to the sun shining directly into the room... for info the room is completely sealed with arch windows and portals. outside there is a sunsky... tone mapping of the two is linear and using exactly the same settings
Thought it could be interesting or useful for analysis...

[Edit]
I actually managed to fix it by removing the arch glass, the difference i discovered between the two windows was that the smaller one (the one that worked) didnt have glass in the window... still interesting however that with identical setups one works and one doesnt... sky still looks a bit odd in the fixed version.... being grey rather than blue

fixed with volumetric scattering added

[jeanphi]

Hi,

If the weighting has not been fixed, then the observed behaviour is pretty much expected.

Jeanphi

[Dade]

The weight of paths is still not correct. I mean it is supposed to be now but I have a problem between direct light contribution and "indirect" light: the "indirect" part seems a bit too low so there must be a bug somewhere. There is also to note that I have implemented only uniform path weights (i.e. multiple importance sampling not yet available) so the rendering may have a bit more noise than normal BiDir.

@Eros: All material requiring to continue the tracing of shadow rays are note yes supported too (i.e. architectural glass, null material, etc.).

On the bright side, the speed up seems pretty solid across different scenes.

[Dade]

Also, RAM use is down to 1.95GB, and sampling rate is up a tiny bit. It crossed over "double the speed sampler got". Interestingly, a cube on a plane takes 2.8GB of RAM. So the "hybrid tax" is lower on more complex scenes?

This is somewhat true: hybrid uses and adaptive number of path states. It starts with one and increase the pool of states up to the point required to fill a buffer of ray to send to the GPU. This mean that a scene with long eye/light paths will require very few states to fill the buffer (i.e. remember the there is a quadratic relation between the average length of eye/light paths and the number of rays generated to connect all path vertex, this is exactly the strength of BiDir compared to Path).

A cube over a plane is nearly a worst case scenario (i.e. average eye/light paths length is 0 or 1).

You can see the adaptive changes to the path state pool size by running Lux with "-V" option.

[Eros]

Thanks for the reply there, makes alot of sense now... great work so far though! The boost in speed is very very welcome even if it is just in testing, makes it all the more exciting.

[J the Ninja]

On the bright side, the speed up seems pretty solid across different scenes.

To continue the encouraging news, I tested my block fort scene as well. 61kS/s for sampler vs 124kS/s for hybrid. However, several of my scenes ran out of memory, probably due to the lack of instancing with hybrid.