Luoshuang's GPULightmass

[Luoshuang]

Do you have comparision screenshots from cpu vs gpu versions. This is really interesting.

While I'm planning on it, I don't want to "advertise" much now since it's pretty premature

[Norman3D]

Would you like to provide your scene to me for debugging?

Do you have some setups like precomputed visibility which can only be done on CPU currently?

Also I remember that currently the calculation of volumetric lightmaps doesn't have a proper progression display, so if you see the GPU is not idle it may be working on that.

I think I'm getting closer to the problem. The GPU calculation gets stuck, the GPU stops working, but Lightmass.exe sits at 62% CPU usage. This only happens when I'm using "Sparse Volume Lighting Samples". Are these supported?

[Luoshuang]

I think I'm getting closer to the problem. The GPU calculation gets stuck, the GPU stops working, but Lightmass.exe sits at 62% CPU usage. This only happens when I'm using "Sparse Volume Lighting Samples". Are these supported?

Sparse Volume Lighting Samples are computed entirely on CPU now. I believe it should be the problem.

[Norman3D]

Sparse Volume Lighting Samples are computed entirely on CPU now. I believe it should be the problem.

I see! Does that mean the lighting needs to be recalculated entirely on CPU for the Sparse Volume Lighting Samples to work? I'm just trying to get an idea of how long it should take, as it seems to take forever and there is no output on the log giving me an idea that progress is indeed happening.

By the way, the results are amazing!!!!

[Luoshuang]

I see! Does that mean the lighting needs to be recalculated entirely on CPU for the Sparse Volume Lighting Samples to work? I'm just trying to get an idea of how long it should take, as it seems to take forever and there is no output on the log giving me an idea that progress is indeed happening.

By the way, the results are amazing!!!!

Unfortunately, yes. In fact the recalculation is not implemented so currently Sparse Volume Lighting Samples won't compute properly (they'll simply turn black, or have only direct lighting).

Edit: maybe I can come up with a quick fix to move their caluclation onto GPU

Do you have screenshots for your amazing results? =)

[Norman3D]

Unfortunately, yes. In fact the recalculation is not implemented so currently Sparse Volume Lighting Samples won't compute properly (they'll simply turn black, or only have direct lighting).

Do you have screenshots for your amazing results? =)

I can't share them here, but I can PM you!

EDIT: Maybe your inbox is full?

[Luoshuang]

I can't share them here, but I can PM you!

EDIT: Maybe your inbox is full?

I've already sent you my email addr a while ago. Didn't receive?

[Norman3D]

I've already sent you my email addr a while ago. Didn't receive?

Got it!

[Raildex_]

Is it planned to implemented the features currently not supported? It lacks a lot of features of Lightmass.

[Luoshuang]

Is it planned to implemented the features currently not supported? It lacks a lot of features of Lightmass.

Currently I don't have a plan to add new features since I'm refactoring it completely...

[StrangeLove13]

This looks great! Is there any chance for an AMD supported version in the future?

[Luoshuang]

Got it!

I've updated GPULightmass to support sparse volume lighting samples. Please redownload to get the newer version. Note that the samples don't get proper progress display so you can only tell by looking at GPU load.

This looks great! Is there any chance for an AMD supported version in the future?

Probable not very soon. To make it work on AMD I'll probably need to learn OpenCL. While OpenCL is somehow similar to CUDA I still need to rewrite the whole program from scratch.

[Luoshuang]

I've updated the Adaptive Hemisphere Sampler of GPULightmass's Brute Force GI Engine. Please redownload.

Also there are some screenshots of Brute Force (GPULightmass) vs Irradiance Caching (Lightmass, default production quality):
2min50s (GTX1080) vs 8min10s (i7-6700k), Simplified SunTemple
If you look at the timings only then it is not very impressive. However the quality has a big difference. If you want to tweak Lightmass parameters so that it prduces same quality images as GPULightmass then I believe it'll be at least 10x slower:





The downside is GPULightmass has noise while Irradiance Caching has 'built-in' smoothing. While a denoiser is definitely on the plan, I'm still working on the sampling algorithm to get better quality samples.

Edit 04/14/2018: I made some changes to the sampler to suppress the artifacts on the wall:

[Bladerskb]

This looks good and promising!

[thejturner]

Looks amazing! The quality of the shadows is much higher and I notice that you don't seem to get the variations induced by multi-threading. Look forward to trying it out and hope Epic adopts this method officially.