New progressive caustics problem

[Jiri.Matys]

Hi,

I want to play with new progressive caustics cause it should be pretty simple and straight forward. But I have a problem, its super slow and I dont know if this is normal or I do something wrong. When I use old caustics on a simple scene, vray precalc caustics for a few seconds and after that I have pretty good visualization how caustic look in a second or so.
When I use new mode, vray doesnt do precalc, but first pass took so long that it is much slower then the old way. Is this normal? Thank you.

[super gnu]

i have a similar impression, i have a super simple scene specifically to test caustics.. without, it renders in 15-20 seconds... with caustics on, it just never seems to finish.. i left it for 15 minutes then got bored and cancelled the render.

[^Lele^]

That's not good.
Did you follow the docs' recommendations?
If you had, and it's still slow, please shoot me a scene.

edit: they do work.

[super gnu]

That's not good.
Did you follow the docs' recommendations?
If you had, and it's still slow, please shoot me a scene.

edit: they do work.

well thats gorgeous... how long did it take to render?!

heres my super basic scene.. this also features only 45% cpu usage on a 3990x, as mentioned here:

https://forums.chaosgroup.com/forum/...but-slow-scene

and, yes, i set the min subdivs to 15 (makes it even slower) , min shading rate to 3, caustics multiplier on both objects raised, and photon emit radius fitted to scene. dont think i missed anything. all else at default.

[fixeighted]

My 3950x averaged 76%.
The only way I could force it to actually finish was to put the noise threshold above 1, then the times kept going down.
The result was effectively identical to one that would theoretically never finish with lower noise values as set in that scene.
So that's odd.
The best time was 2.47 at noise 4.0...anything below that and it starts to slow - anything at all above...I tried 10, 50....time stays the same.

[vlado]

Regarding the CPU usage specifically, it's a known issue that we will be working to resolve.

Best regards,
Vlado

[super gnu]

Is that a general vray 5.1 issue? Because in the test scene above i also get low usage with caustics disabled

[^Lele^]

well thats gorgeous... how long did it take to render?!

A couple of hours on a 1900x, set to 3.5ghz without boost.

In my case, i am over 80% of CPU usage.
The lower CPU usage is unfortunately par for the course with this tech, right now.

This said, your very scene is being profiled, and solutions are being looked into.
We should get (much) better usage in the (hopefully not-too-distant) future.

I may also have cross-posted.

[^Lele^]

Is that a general vray 5.1 issue? Because in the test scene above i also get low usage with caustics disabled

Nah, that is likely because it reaches noise threshold and there are very few pixels left to shade.
Try it on something a bit more articulated.

[super gnu]

Yeah its fine on proper scenes, i suspected it was the unrealistic scene initially..

[fixeighted]

Out of interest Lele, can you maybe try your jewellery scene with the much higher noise threshold as I did?
In theory that render time should fall dramatically.

Even at the scene settings of .01 in the simple scene it took much much longer to reach the stage that it basically stalled, yet the image quality was the same.
I don't understand that, unless what I did was somehow inherently wrong and therefore totally misleading.

[^Lele^]

Yeah its fine on proper scenes, i suspected it was the unrealistic scene initially..

We minnows, with only 8 cores, never quite get to see the problems of you biggies with 64, eh.
But i can imagine how being stinking rich (in cores) might be.

[^Lele^]

Out of interest Lele, can you maybe try your jewellery scene with the much higher noise threshold as I did?
In theory that render time should fall dramatically.

Even at the scene settings of .01 in the simple scene it took much much longer to reach the stage that it basically stalled, yet the image quality was the same.
I don't understand that, unless what I did was somehow inherently wrong and therefore totally misleading.

We're also investigating how caustics influence the noise threshold calculations.
I have personally not had issues, but then i expected (much) extra work to be done for the caustics part, so perhaps i was biased favourably.

It could be the solver doesn't quite stop when it should, or perhaps some other thing (in the wild guesses' department: caustic intensity, invisible noise, etc etc).
Do you have a scene i can try?

[fixeighted]

It's just a bit odd....I just tried the same thing with the simple scene and without caustics, and with the noise at 0.01 it took 1.11 and with the noise at 5.0 it took 52 secs.
That's just not how I imagined it worked and is such a large difference that, yes, it's a good thing to have noticed and could speed stuff up (if it scales), but is pretty unintuitive
and extremely unexpected.

EDIT: I put this together just now. With noise at 0.1 I stopped it as it was not going to finish - time 5.x
Upped the threshold to 5.0 for the second one and it finished as you can see in just under where I randomly stopped it for the first.
The noise is obviously a little different but overall the quality is the same.

[psanitra]

Happy to hear utilization is being looked at, i`m getting 60% on 3970x.