Adaptive DMC sampler theory ?

[-SeB-]

Hi,

I have tried these last days to learn/understand the theory behind the adaptive DMC sampler and I would like to have your opinion to be sure that I correctly understand it.
I have used the amazing documentation of Toni Bratincevic "Demystifying V-Ray DMC Sampler" to understand the DMC but still with it, it took me times to figured it out how it works.

My first question/interrogation is :
If I understand well, the DMC min/max subdivs are the number of samples send for one EYE RAY
and the subdivs of shaders, lights,... are the number of samples send for one SECONDARY RAY ? right ?

If I use the same example in the documentation ("Demystifying V-Ray DMC Sampler"):
Let's say I have a scene with the DMC subdivs min 1 - max 5(5*5= 25samples), a glossy shader with 40 subdivs and a rectangle light with 10 subdivs and the adaptive amount to 0.85.

- the glossy shader subdivs will be divided by the DMC max subdivs 40/5 = 8 subdivs (8*8=64 samples)
- the rectangle light subdivs will also be divided by the DMC max subdivs 10/5 = 2 subdivs (2*2=4 samples)

So if a pixel need the max DMC subdivs (5 subdivs), DMC knows that he can use more subdivs for the glossy shader if he need it!
my question is :
Is the DMC get 8 more subdivs to render the glossy shader (8subdivs plus 5DMC subdivs) or is it get only 3 more subdivs (8subdiv minus 5DMC subdiv) ??

I am a bit more confuse about the rectangle light subdivs!
if a pixel need the max DMC subdivs (5 subdivs), is the DMC can use 2 more subdivs for the lights ? (5 + 2subdivs)

Now if the max DMC is not reach I suppose that the sudivs of the shader and the lights will not be use ! right ?

Concerning the adpative amount:
if it is set to 0.85, the DMC will be adaptive at 85% so 15% will be fixed
It means that the DMC will use 0.75 fixed subdivs on the 5 max subdivs and then check if he need more samples. Right ?


Thanks.

-SeB-

[joconnell]

Heya Seb - yeah you're totally right on how you're thinking about DMC. The big thing that I've understood recently is that yes, the AA will divide the light and material subdivs by whatever number of subdivs it's using on a pixel and sometimes this can be a bad thing. The aa sampler is always trying to use the smallest number it can to reach the noise threshold. At aa rates where vray is using closer to min subdivs, it means that it's using more subdivs from the lights and material settings, and as it goes gradually up to higher Max subdivs in your aa, it starts to use less and less material and light subdivs as you've described above. The problem for me is that if your noise in the scene is being caused by a light or a material then you're far better off not trying to use more AA to fix it, you're better off trying to find out what material or light is causing the issue and then raising the subdivs on that. I think of AA sampling as texture, bump and geometry sampling only (also motion blur) and I use the sample rate element to see if I've got enough aa quality to give me clean edges and texture detail in my images. After that I use the raw lighting, raw GI, reflection and specular passes to see if I have enough quality in my materials and lights. The two are very similar in some ways as the noise on a glossy material can be fixed by making that material use higher subdivs, or if there's a light hitting it, making the light use more subdivs so that it's less noisy when it hits the material.

[-SeB-]

Thanks for the answer ! I have same ideas than you
But I'm still confuse on this part

my question is :
Is the DMC get 8 more subdivs to render the glossy shader (8subdivs plus 5DMC subdivs) or is it get only 3 more subdivs (8subdiv minus 5DMC subdiv) ??

I am a bit more confuse about the rectangle light subdivs!
if a pixel need the max DMC subdivs (5 subdivs), is the DMC can use 2 more subdivs for the lights ? (5 + 2subdivs)

Now if the max DMC is not reach I suppose that the sudivs of the shader and the lights will not be use ! right ?

-SeB-

[joconnell]

Nope - what it means is at aa 5, vray will fire 5 camera rays into a pixel. For each of those 5 rays, it will take 2 samples from each light, and if you stick with your 40 material subdivs it'll take 8 samples.

[-SeB-]

okayyy, it's more clear now !! Thanks
So,
- First eye ray + 8 shader subdivs + 2 light subdivs
- second eye ray + 8 shader subdivs + 2 light subdivs
- ...
- until the fifth eye ray!

-SeB-

[joconnell]

At aa level 1 - 1 camera rays per pixel, 40 material subdivs, 10 light samples
At aa level 2 - 2 camera rays per pixel, 20 material subdivs per ray, 5 light samples per ray
At aa level 3 - 3 camera rays per pixel, 13 material subdivs per ray, 3 light samples per ray

It's always material or light samples divided by the AA level.

[-SeB-]

are you sure of that ?
I think it's the shader subdivs divided by the MAX rate. so 40/5 = 8samples .
AA1 = 1 eye ray + 8 subdivs
AA2 = 2 eye ray + 8 subdivs
AA3 = 3 eye ray + 8 subdivs
AA4 = 4 eye ray + 8 subdivs
AA5 = 5 eye ray + 8 subdivs

so at the end you finish with 8*5 = 40 subdivs for the material

with your theory if the max dmc of 5 is reached you have a total of 40+20+13+10+8 = 91 subdivs for the material

but I agree if you set 1 in max dmc you'll have 40 subdivs for the material

what do you think ?

-SeB-

[joconnell]

Seb you're totally right - I got a confirmation from Svetlozar on this that the division is only done once at the start, and based on your max aa number so your calculations above are totally right.

Here's the discussion, hopefully lots of clarity coming from it!

http://www.chaosgroup.com/forums/vbu...-setting/page2