## GPU renderer Battle

Post and discuss your competition entries here!

Forum rules
1. All competition entries must be a new original work unless otherwise specified during a competition.
2. All submissions must be uploaded to the corresponding competition gallery no later than the date specified in the competition post.

### Re: GPU renderer Battle

When you post the latest version of the scene, I can run it with the under development micro-kernel edition of BIASPATHOCL. I have a GTX980 and a R9 290X where I can run the test.

Posts: 8404
Joined: Sat Apr 19, 2008 6:04 pm
Location: Italy

### Re: GPU renderer Battle

here's the scene
FStormSceneLux.7z

short note:
There's just too much difference between the two engines and the scene: i.e. FStorm is RGB, few lights (in the scene) don't contribute to GI (direct only), have different falloff (?), scene is lit by HDRI with portal on the balcony (why upper part/ceiling is lit orange and lower blueish), different shading system(?), more/better(?) postFX, tonemapping options (which consequently make 'the look' pop up)... otherwise are very close considering render times (yes, it all comes down to the time artist has to invest, so it really matters to have efficient workflow without tinkering, waiting for trial&error results (compiling on nvidia ), clicking/seeking all over the menus to get to proper settings in question).
... for now am done messing with my head and vision... sounds appearing

& here are raw render comparisons (1024samples)
burnin

Posts: 287
Joined: Mon May 03, 2010 8:04 pm

### Re: GPU renderer Battle

Ok, I did some tests with the GTX980 (because it is comparable to your Quadro):

- PATHOCL => 8:48secs

- New BIASPATHOCL with micro-kernels => 6:18secs

- I have then tuned some parameter. New BIASPATHOCL with

Code: Select all
accelerator.type = BVHaccelerator.instances.enable = 0

because you should use MBVH and true instance support only if you run out of memory => 5:20secs

- Cut path depths (i.e. reduce a bit the quality of rendering) with

Code: Select all
biaspath.pathdepth.total = 4biaspath.pathdepth.diffuse = 3biaspath.pathdepth.glossy = 2biaspath.pathdepth.specular = 2

result => 4:01secs

And I could continue removing volumes and replacing (arch)glasses with transparencies, etc.

P.S. this is one of the reason why render engine comparisons are so hard (if not always unfair), I can cut here, push there and pretty much get any time you want.

Posts: 8404
Joined: Sat Apr 19, 2008 6:04 pm
Location: Italy

### Re: GPU renderer Battle

Nice

Yup, my bad for leaving volumes untouched.
Archglass was intentionally there (wanted to see how it feels with interior scene, cuz in preview is brighter than background - but it's fast and subtle reflections help with the look)
- New BIASPATHOCL with micro-kernels => 6:18secs

Yes!
Is it coming soon?
- I have then tuned some parameter. New BIASPATHOCL with... because you should use MBVH and true instance support only if you run out of memory

didn't know that, thanks
- Cut path depths (i.e. reduce a bit the quality of rendering) with

Done that too, but wasn't satisfied
P.S. this is one of the reason why render engine comparisons are so hard (if not always unfair), I can cut here, push there and pretty much get any time you want.

Same thoughts here
Also FStorm uses post on benchmark scene, which makes it even harder to reproduce in detail.
Thank you, for everything.

BTW, did you experience any lag while compiling?
burnin

Posts: 287
Joined: Mon May 03, 2010 8:04 pm

### Re: GPU renderer Battle

burnin wrote:Nice

Yup, my bad for leaving volumes untouched.
Archglass was intentionally there (wanted to see how it feels with interior scene, cuz in preview is brighter than background - but it's fast and subtle reflections help with the look)
- New BIASPATHOCL with micro-kernels => 6:18secs

Yes!
Is it coming soon?

I've quickly made an experimental build for your convenience (viewtopic.php?f=30&t=13054&p=123962#p123962), but I can't gurantee that I've built it properly. It's marked as purely for testing since this is not yet completed by the developer and even if I did the compiling part right, things may not work as expected (or at all).

If testing with BIASPATHOCL, make sure to use sampler.type = "BIASPATHSAMPLER" or you will get an error message telling you to set it.

In fact, I'm slightly curious about these stats:

Tonga is a 'bit' more powerful (r9 380 vs r7 260x) and also working on a faster PCI-E lane.

Here are the stats when using PathOCL:

Here the contributions seem relative to the hardware power (similar with sobol as well). This may be just how biaspath works and expected, but I'd like someone more knowledgable to estimate.

biaspathocl Results after 500 sec/8.3 min (manually stopped, thus the 'box' parts):
Attachments
kalel

Posts: 258
Joined: Sun Jul 17, 2016 6:35 am

### Re: GPU renderer Battle

kalel wrote:Here the contributions seem relative to the hardware power (similar with sobol as well). This may be just how biaspath works and expected, but I'd like someone more knowledgable to estimate.

BIASPATHOCL performance is all about using a tile size large enough to keep the GPU busy, what have you used ? I have used a 1024x1024 tile with 1 sample per pixel in my tests.

Posts: 8404
Joined: Sat Apr 19, 2008 6:04 pm
Location: Italy

### Re: GPU renderer Battle

kalel wrote:Here the contributions seem relative to the hardware power (similar with sobol as well). This may be just how biaspath works and expected, but I'd like someone more knowledgable to estimate.

BIASPATHOCL performance is all about using a tile size large enough to keep the GPU busy, what have you used ? I have used a 1024x1024 tile with 1 sample per pixel in my tests.

Thanks, you mentioned something important to try out. Normally I try various tile sizes when setting up scenes to see what works best, but this time I was using whatever was set in the blend file (16 I guess), in order to quickly start rendering.

After changing the AA size to 1, I was able to not get out of memory errors when using 1024 tile size. The stats in this case are much better for every device, with Bonaire still contributing a somewhat larger percentage than with PathCL. CPU performance improved significantly as well.

The only issue with 1024 (considering that means the entire image) is that we're not getting some benefits from biaspath (making more samples on those tiles that are more noisy), but perhaps it's necessary to do this in order to utilize the GPU properly.
Attachments
Stats
kalel

Posts: 258
Joined: Sun Jul 17, 2016 6:35 am

### Re: GPU renderer Battle

Thanks for the build, kalel.
Gloomy sunday here, so testing mode on...
burnin

Posts: 287
Joined: Mon May 03, 2010 8:04 pm

### Re: GPU renderer Battle

burnin wrote:Thanks for the build, kalel.
Gloomy sunday here, so testing mode on...

You're welcome. Will be happy to see what you do. Sunny sunday here, but I'm not feeling well enough to use it properly.
kalel

Posts: 258
Joined: Sun Jul 17, 2016 6:35 am

### Re: GPU renderer Battle

Im kind of late trying out this render test scene.
Render time = 36 minutes
Using 2 x RX 480 + i7-6700K 4.0GHz CPU
Using LuxRender 1.7 x64 OpenCL Setup
Here is my result:
Attachments
david57

Posts: 208
Joined: Sun Aug 23, 2009 3:33 pm

Previous