CPU vs GPU rendering modes

Juju · Post by **Juju** » Tue Nov 20, 2012 3:37 am

From the manual I understand the following w.r.t. rendering modes:

- CPU rendering: path tracing, bi-directional path tracing, path tracing with MLT, bi-directional path tracing with MLT.
- GPU rendering: path tracing, not sure if bi-direction path tracing can also be done.
- speed wise: GPU > CPU to factor of about 3x (depending on hardware, etc.)

When will MLT (and bi-directional path tracing?) be able to be done on the GPU?

I had a 1920 x 1080 render of an external scene bake for 20000 SPP and the CUDA accelerated version still had a lot of fire flies that wouldn't resolve, so I re-rendered with GPU acceleration disabled and it resolved very well, compared 20000 SPP versions and eventually stopped it at 22000 SPP as it was not clearing up any more visibly.

Any suggestions? I don't particularly like post processing renders much.

Headroom · Post by **Headroom** » Tue Nov 20, 2012 4:19 am

20k SPP seems very high, particularly for an exterior scene. Please post an image if the scene if you can.

Generally I don't think we'll see MLT on a GPU any time soon. It's a more complicated algorithm that is hard to implement on a GPU architecture.

Bidirectional may be on the horizon however.

Zom-B · Post by **Zom-B** » Tue Nov 20, 2012 5:08 am

Juju wrote:Any suggestions? I don't particularly like post processing renders much.

And we don't particularly like to scratch our heads without a image to look at! There are a lot of possible reasons for slow rendertimes, there is a very good chance that the problem is manmade

Juju · Post by **Juju** » Wed Nov 21, 2012 10:01 pm

Here are the two images.

Zom-B · Post by **Zom-B** » Wed Nov 21, 2012 10:23 pm

Juju wrote:Here are the two images.

Now we can Talk

Regarding the Fireflies you refer to the glass situation in the car to the left, right?!

I see two issues in your scene:

1) A evergreen that I repeat like a mantra around here in the Forum ^^
Your maximum RGB values should be R204 B204 G204 for! Otherwise that ultra white materials will reflect every ray and cause extra calculation time! Also for colors the 204 rule is valid... same for textures!
This reflecting of rays causes long bounce rates and longer computation times

2) as you can see in the car to the left in the MLT image the window looks like a "magnifying glass", so we come to the second place of user caused issues: Glass needs to have a proper volume!! Don't use single sided meshes as glass (or water or whatever). Your glass NEEDS physical thickness with normals facing out!

3) in your scene Glass Acceleration could maybe help if you have issues with noise behind windows, but since your glass is so reflective it doesn't seem to be needed here

(don't use if not needed, since it slightly slows down overall performance!)

Juju · Post by **Juju** » Wed Nov 21, 2012 11:49 pm

Thanks for the feedback. The fireflies in the glass of the car in the foreground are the post noticable, I've marked the ones I noticed.

Zom-B · Post by **Zom-B** » Thu Nov 22, 2012 12:32 am

How exactly do you light the scene, how is the cloudy sky done? In a HDRI workflow fireflies on the sky are quite impossible!
The ringing around the Fireflies seems to be PostPro AD related.

I have a hard time to find Fireflies on the walls of the building and that rolling door.

Because of your messed up glass meshes in the cars there is a good possibility that light enters that glass inner of the car and bounces out in a total different direction, so some caustic based fireflies can be created in strange spots of the scene.
Pure PT isn't very good in rendering caustics, you know!

Fix the glass in your scene first and rerender then! Should be way better results (regarding the FFs) in just 1 to 2 hours

Juju · Post by **Juju** » Thu Nov 22, 2012 1:03 am

No HDRi scene used, the sky in the background is an image billboard. Natural (simulated) sun.

Some of the fireflies became accentuated when applying the Indigo AD.

I'll keep the glass thickness thing in mind, see if / when I can look into that and re-render.

Zom-B · Post by **Zom-B** » Thu Nov 22, 2012 6:09 am

Juju wrote:No HDRi scene used, the sky in the background is an image billboard. Natural (simulated) sun.

Such stuff is usually done in PostPro! You can't put a 25 meters tall billboard into your scene and think that it doesn't interact with the light that indigo throws into it. At first it simply slows down rendering and second you loose all flexibility for further tweaking that (fake) sky!
A quickly rendered alpha foreground render can help you to mask indigos sky out and place some image in the BG during PostPro.

Juju wrote:Some of the fireflies became accentuated when applying the Indigo AD.

Yes, Indigos PPAD is sensitive to fireflies :/
But in your scene its quite only the Fireflies that react to the AD, so you can also keep it off ^^

Juju wrote:I'll keep the glass thickness thing in mind, see if / when I can look into that and re-render.

That quite a big deal and a general way of handling (physical) render engines, you should handle that this way for the most engines.
If Time is an issue, I would suggest simply to fix the glass and do a region rendering of that car windows, mixing the result up in PS or Gimp with the old output. The few other hotspots can be fixt using the stemp tool very quickly

Juju · Post by **Juju** » Thu Nov 29, 2012 3:55 am

Zom-B wrote:
Juju wrote:No HDRi scene used, the sky in the background is an image billboard. Natural (simulated) sun.
Such stuff is usually done in PostPro! You can't put a 25 meters tall billboard into your scene and think that it doesn't interact with the light that indigo throws into it. At first it simply slows down rendering and second you loose all flexibility for further tweaking that (fake) sky!
A quickly rendered alpha foreground render can help you to mask indigos sky out and place some image in the BG during PostPro.

Juju wrote:Some of the fireflies became accentuated when applying the Indigo AD.
Yes, Indigos PPAD is sensitive to fireflies :/
But in your scene its quite only the Fireflies that react to the AD, so you can also keep it off ^^

Juju wrote:I'll keep the glass thickness thing in mind, see if / when I can look into that and re-render.
That quite a big deal and a general way of handling (physical) render engines, you should handle that this way for the most engines.
If Time is an issue, I would suggest simply to fix the glass and do a region rendering of that car windows, mixing the result up in PS or Gimp with the old output. The few other hotspots can be fixt using the stemp tool very quickly

yeah, so I extruded all glass in the scene so it has a thickness, 6mm throughout, not making a difference it seems in either fireflies or magnifying effect in the car (left front) back window...

Zom-B · Post by **Zom-B** » Thu Nov 29, 2012 4:42 am

Juju wrote:yeah, so I extruded all glass in the scene so it has a thickness, 6mm throughout, not making a difference it seems in either fireflies or magnifying effect in the car (left front) back window...

Then you did something wrong, since I trust a proven raytracer more then your skill here

Just kidding

Double check that the glass meshes have all normals facing out of the glass center.
Otherwise maybe upload just your car windows as sketchup file, so somebody from the SK camp can take a look

CTZn · Post by **CTZn** » Thu Nov 29, 2012 4:49 am

The car window AND what it is attached to, since how a glass pane is interpenetrating other surfaces or not may have an impact on the fireflies generation.

You didn't mention wether you tried Glass Acceleration as Zom-B suggested.