Test scences GPU Path vs CPU BiDir

Post your tests, experiments and unfinished renderings here.
FarbigeWelt
Posts: 101
Joined: Sun Jul 01, 2018 12:07 pm
Location: Switzerland
Contact:

Re: Test scences GPU Path vs CPU BiDir

Post by FarbigeWelt » Mon Jul 16, 2018 11:03 am

B.Y.O.B. wrote:
Mon Jul 16, 2018 8:22 am
Edit: by the way, your light sources are very high poly. Consider using less triangles for them.
Currently you have 4536 light sources (emitting tris) in total.
Since they are all in one mesh, LuxCore will automatically switch off direct light sampling for those light sources (this happens with any emitting mesh with more than 256 triangles). This is done because usually, meshes with many triangles are not the main light source of the scene and would steal away computing power from the other lights, your scene is a special case.
Thank you. This explains the dark lights in combination with env. light. Did not know that this is intended behaviour.
Exactly, a warning helps in this case. Frist I‘ve used rectanlges. But this did not fit well to the scene. Looking forward to installing the next release. I will apply the array and separate the lights by loose parts. Reflector needs some improvements anyway.
Instagrammer, please join #luxcorerender for your renderings.

TryToRemember
Posts: 10
Joined: Tue Jul 03, 2018 3:35 am

Re: Test scences GPU Path vs CPU BiDir

Post by TryToRemember » Mon Jul 16, 2018 1:26 pm

Yep, thought it was something like that. Double windows at night. of course. shiny upon shiny. Perfectly perpendicular to the specular floor and the specular reflectors and the very many emitter sources is a rendering task, to be sure! But is also a textbook recipe for "snow". Commonly accepted terminology be damned, I like snow more than "fireflies" when it's everywhere.

I'd cut the emitters down to just rectangles in a cross-section of the tube, and make one quad face, duplicate and rotate 180°, per lamp. (it's a emitter, no offset necessary, they can occupy the same exact space.) That is where I'd start. From whatever it was (4536??!) to 24...

What's brilliant, if you let it chug along (and with maxdepth high enough) that kind of scene does eventually clear up! With all the sources it will take awhile. After a few thousand samples, if it's not clear then you might want to consider post-processing "cheats" like hot pixel removal or clamping. I'm guessing in the scenes you like to build --; from what? 4 pages of materials / samplers testing over ~2 weeks now? --; You seem more the type of person who would let his all-mirrors-and-lenses scene, all with heavy dispersion, pinlight illuminated through multiple, nested scattering volumetrics, run all day to get 100/spp. :-D

Please keep sharing your experiments, I love seeing and reading them.

FarbigeWelt
Posts: 101
Joined: Sun Jul 01, 2018 12:07 pm
Location: Switzerland
Contact:

Re: Test scences GPU Path vs CPU BiDir

Post by FarbigeWelt » Mon Jul 16, 2018 4:13 pm

Ever tried to simulate snow with particles physics in blender? Or a pendulum with chains? Fludics simulations with the molecule addon? Well this takes all its time.. I‘used to wait for simple matte sphere half a day once. It dosn‘t matter how fast your system is, capabilities grow faster.

My rented appartment has a very nice, smooth wooden floor. I‘ve modelled all the rooms with some furniture without messure them. Proportions are therfore a bit out of line and I may start again from scratch. There are some million triangles because of a few plants. The stairs iare a impression of the whole scene. One tree consumed 2 GB, I had to remove it. It was a nice one with polygonal volumetric leafes.

Maybe I will do a few experiments for optimizing Samples/min approaches. There are for sure some generic topics to learn. Let‘ see.
Instagrammer, please join #luxcorerender for your renderings.

FarbigeWelt
Posts: 101
Joined: Sun Jul 01, 2018 12:07 pm
Location: Switzerland
Contact:

Bump map increases Sample Count

Post by FarbigeWelt » Mon Jul 16, 2018 10:07 pm

First I setup a new scene with a very simple setup. Four cubes, one point lamp. The biggest cube is a room with the three cubes inside. Because all white is not much of contrast I've colored the litte cubes. Material used for the first test is matte, what else? Interesting to notice: sigma 1.0 leads to higher sample count than sigma 0.0, sigma 0.5 lies somewhere between the two. With my Methusalem i7 I get for HD resolution and default settings without denoiser in the best case 905 kS/s. Glossy by the way reaches only 630 kS/s. Now to the miracle, I've added a bump map to the room cube and got 1.2 MS/s. I tried another one and got sa similar result. The images used for the bump map are png epxorts with 32 bit. I tried a black and a white image saved as 24 bit png and used them. The odd thing was, they both run only at 780 kS/s. Whereas a white 32 bit png reaches 1.1 MS/s. The pattern with the concentric rings keeps on top of todays performance test.
All sigma 1.0 cpath 60s 905kSps.png
All sigma 1.0 cpath 60s 905kSps
RGB Cubes (2).zip
Blender Scence
(184.8 KiB) Downloaded 7 times
Matte with bump cpath 60s 1.2 MSps.png
Matte with bump cpath 60s 1.2 MSps
Maybe counter counts different for bump maps if not there is a good portion of performace gain there waiting to be released.
Instagrammer, please join #luxcorerender for your renderings.

FarbigeWelt
Posts: 101
Joined: Sun Jul 01, 2018 12:07 pm
Location: Switzerland
Contact:

Light number: CPU BiDir is, CPU Path is not

Post by FarbigeWelt » Tue Jul 17, 2018 10:03 pm

I'd cut the emitters down to just rectangles in a cross-section of the tube, and make one quad face, duplicate and rotate 180°, per lamp. (it's a emitter, no offset necessary, they can occupy the same exact space.) That is where I'd start. From whatever it was (4536??!) to 24...
Are you sure about that?

I've tested the scene above with 1, 234, 1170 light areas (9 cylinders with 130 areas and matte material with emission).
It did not matter for CPU path render speed but it decreased CPU BiDir 4%.
CPU BiDir decreases further if cylinders are smothed or the small cubes have a bump map (10%).
CPU path kept rendering at a speed of 1.2 MS/s.

Conclusion: Use nice bodies as lamps with currently less than 255 areas if you want. There are other, more CPU time consuming matters.
Instagrammer, please join #luxcorerender for your renderings.

FarbigeWelt
Posts: 101
Joined: Sun Jul 01, 2018 12:07 pm
Location: Switzerland
Contact:

Newton Pendulum

Post by FarbigeWelt » Fri Jul 20, 2018 9:16 pm

Some time ago I've been quite busy to simulate a Newton Pendulum with Blender rigid body physics. There are pretty easy to follow tuorials for that purpose. But these samples used here or there a trick and did not convince me. I could not get the idea of rattling chains out of my mind. Modelling haven't been any issue. But then everything happend to the chain but hanging. The chain broke, the parts blew away, the bowl hovered rotating throu space. The physics behaved strange. It required a good portion patience to learn how to deal with the parameters for rigid bodies. Just because I mentioned the scene a few posts earlier I've scanned for its blender file. Old Luxrender scences don't run with the 2.x addon. At least not this scene. Everything I got was some material node error message. And the only way to render the scene with LuxCoreRender was to append objects and groups only to a new Blender file. After struggling a while with well known simulation instability I've got the solution.
Chained.jpg
Newton Pendulum, CPU Path, 100 samples
And the animation:
https://www.youtube.com/watch?v=V1LMsq0o-gs
Blender 2.79b, LuxCoreRender addon 2.1 alpha2, CPU Path, 1:30 min., ~25 samples, video filtered and rendered in Photoshop.
Last edited by FarbigeWelt on Fri Jul 20, 2018 9:35 pm, edited 1 time in total.
Instagrammer, please join #luxcorerender for your renderings.

User avatar
B.Y.O.B.
Developer
Posts: 1312
Joined: Mon Dec 04, 2017 10:08 pm
Location: Germany
Contact:

Re: Test scences GPU Path vs CPU BiDir

Post by B.Y.O.B. » Fri Jul 20, 2018 9:21 pm

FarbigeWelt wrote:
Fri Jul 20, 2018 9:16 pm
Old Luxrender scences don't run with the 2.x addon.
Yes, this is true in general.
It's a completely new addon and there is no converter, you will have to re-create all material and light setup of old scenes when you switch to the new addon.
However, it should normally not be required to append anything into a new Blender file.

Nice physics sim!
Are you using motion blur for the chains?
By the way, in the shadows on the right you can see that the chains are not connected to the spheres.
Support LuxCoreRender project with salts and bounties

FarbigeWelt
Posts: 101
Joined: Sun Jul 01, 2018 12:07 pm
Location: Switzerland
Contact:

Re: Test scences GPU Path vs CPU BiDir

Post by FarbigeWelt » Fri Jul 20, 2018 9:42 pm

B.Y.O.B. wrote:
Fri Jul 20, 2018 9:21 pm
However, it should normally not be required to append anything into a new Blender file.

Nice physics sim!
Are you using motion blur for the chains?
By the way, in the shadows on the right you can see that the chains are not connected to the spheres.
I used to render the scence with cylces too. Maybe this is a reason.

Thank you.
Yes, I use motion blur for this scene.

Really? I've never noticed. Do you mean the hook? The sphere is connected with a hook to the last chain.
Instagrammer, please join #luxcorerender for your renderings.

User avatar
B.Y.O.B.
Developer
Posts: 1312
Joined: Mon Dec 04, 2017 10:08 pm
Location: Germany
Contact:

Re: Test scences GPU Path vs CPU BiDir

Post by B.Y.O.B. » Fri Jul 20, 2018 10:05 pm

FarbigeWelt wrote:
Fri Jul 20, 2018 9:42 pm
I used to render the scence with cylces too. Maybe this is a reason.
No that also should not cause problems.
If you can provide the exact error message (and the scene, if possible), that would be great.
FarbigeWelt wrote:
Fri Jul 20, 2018 9:42 pm
Really? I've never noticed. Do you mean the hook? The sphere is connected with a hook to the last chain.
Yes, the hook is "hovering" in a certain distance from the sphere.
FarbigeWelt wrote:
Fri Jul 20, 2018 9:42 pm
Yes, I use motion blur for this scene.
Glad that it works :D
Support LuxCoreRender project with salts and bounties

FarbigeWelt
Posts: 101
Joined: Sun Jul 01, 2018 12:07 pm
Location: Switzerland
Contact:

Re: Test scences GPU Path vs CPU BiDir

Post by FarbigeWelt » Sat Jul 21, 2018 7:27 am

B.Y.O.B. wrote:
Fri Jul 20, 2018 10:05 pm
FarbigeWelt wrote:
Fri Jul 20, 2018 9:42 pm
I used to render the scence with cylces too. Maybe this is a reason.
No that also should not cause problems.
If you can provide the exact error message (and the scene, if possible), that would be great.
Okay, I opened a new issue and uploaded the old scence and noted some step by step reproduction notes.
B.Y.O.B. wrote:
Fri Jul 20, 2018 10:05 pm
FarbigeWelt wrote:
Fri Jul 20, 2018 9:42 pm
Really? I've never noticed. Do you mean the hook? The sphere is connected with a hook to the last chain.
Yes, the hook is "hovering" in a certain distance from the sphere.
If the hook were not connected to the chains the sphere would follow its own impuls. The simulation is very fragile. If collision is not detected objects pass throu each other. By the way, sphere and hook are one mesh object.
Instagrammer, please join #luxcorerender for your renderings.

Post Reply