Multiple Importance Sampling

[wyszolmirski]

Hi,

I noticed that as far as RT GPU is significantly faster than RT CPU, it wasn't as fast when compared to ADV CPU. Is that because of Multiple Importance Sampling (MIS) isn't supported in RT yet ?

If so, are there any plans of making it work in the RT GPU/CPU ?

Best,

Tomasz

[tsvetan.geshev]

Hello,

How do you compare RT/GPU and ADV CPU?
RT always use brute force/brute force, so if you want you have correct results from your tests, you need to set the same GI setup for ADV CPU.
Can you post the exact render times and the models of your CPU and GPU?

[Dariusz Makowski (Dadal)]

Hello,

How do you compare RT/GPU and ADV CPU?
RT always use brute force/brute force, so if you want you have correct results from your tests, you need to set the same GI setup for ADV CPU.
Can you post the exact render times and the models of your CPU and GPU?

Didnt RT use BruteForce/LC ?

[wyszolmirski]

Didnt RT use BruteForce/LC ?

Only RT GPU can use LC when set as production renderer.

I did take this into account. I will quickly redo the test in office and check if I get the same results. Maybe there was issue with drivers on my home workstation. I was pretty sure that difference in denoising was caused by lack of MIS.

Best,

Tomasz

[wyszolmirski]

Now I'm not sure. What's the story.

For all of the tests I've setup noise threshold at 0.01. And here are the results.


ADV CPU BF+BF, 5 min 21 s





RT CPU BF+BF, 8 min 21 s ( noise seems ok but still a bit nosier than Adv):





RT GPU BF+BF, 1 min 34 s ( same threshold but noise seems bigger ):





After that test I don't know anymore what's going on. When I wrote the first post I just didn't compare noise levels at all. Just assumed they are similar. Now when taken this into account, to get something similar to ADV CPU on RT CUDA we would have to drop the noise threshold a bit more, adding to render time.


I'm using I7 3770K and GTX 970 for this test. I've attached the scenes.

[vlado]

To answer your original question, V-Ray RT does use multiple importance sampling very heavily, both on the CPU and on the GPU.

As for why this particular scene comes out noisier, I'm not sure - will look into it.

Best regards,
Vlado

[wyszolmirski]

Hey Vlado,

Thanks for looking into it and answering my first question.

I just got few 970s and ordered Titans for animation renderings and I wonder if there are any improved builds that would include improvement's you've talked about on GTC?

Best,

Tomasz

[vlado]

I just got few 970s and ordered Titans for animation renderings and I wonder if there are any improved builds that would include improvement's you've talked about on GTC?

We will be releasing a new service pack in May that has quite a few improvements in V-Ray RT GPU, including improved speed and noise levels.

Best regards,
Vlado

[3LP]

I noticed few months ago (doing my interiors animations) that the noise threshold between the CPU and GPU is actually quite different.
If Production CPU is set to 0.01, I have to lower down RT GPU as low as 0.003 sometimes to get the same noise, but it's not consistent, sometimes it's 0.005. But I never had RT GPU 0.01 being the same noise as Adv Prod at 0.01

Can't wait to see the new SP Vlado,
I just hope you where able to fit in a multi-loader map of any kind, something as MultiTexture/XMap or "simply" make the VrayMultiSubTex work on GPU.
This is really one of the crucial feature that RT GPU still lacks in my day to day workflow.

Stan

[_martin]

We will be releasing a new service pack in May that has quite a few improvements in V-Ray RT GPU, including improved speed and noise levels.

Best regards,
Vlado

Will this SP be for the 3ds Max or the Maya version? Thanks!

[vlado]

We will have SPs for both 3ds Max and Maya.

Best regards,
Vlado

[vlado]

So I looked at the scene finally; the difference in brightness is because the RT engine traces one GI bounce less for whatever reason, so I had to set the GI bounces to 4 to get the same image (will have to fix this).

The difference in the noise between the production renderer and RT CPU is because the sampling for the area lights seems different - which is odd, I need to look into it.

Finally, the GPU is noisier because it uses more randomness in its calculations. The QMC sampling should help with that for the next service pack.

Best regards,
Vlado

[wyszolmirski]

So I looked at the scene finally; the difference in brightness is because the RT engine traces one GI bounce less for whatever reason, so I had to set the GI bounces to 4 to get the same image (will have to fix this).

The difference in the noise between the production renderer and RT CPU is because the sampling for the area lights seems different - which is odd, I need to look into it.

Finally, the GPU is noisier because it uses more randomness in its calculations. The QMC sampling should help with that for the next service pack.

Best regards,
Vlado

Thanks Vlado for looking into it.

For the last 2 weeks we've been working mainly on RT CUDA and there will be a lot of shots coming out of it. Everyone loves it. And already I can see improvement in shots development speed, so I really like seeing that you're putting a lot of effort in improving it. To me even now it's the best GPU rendering engine available, so finger crossed for the future advancements!

One thing that I miss a bit is some volumetrics support. But I guess that's on the way sooner or later

Keep up good work,

Best,
Tomasz