Question about Hybrid rendering

[Wingman_mn]

Just a simple question:
In the 0.9 dev (v. Sept. 10, 2011) of LuxRender for Windows, there are some new settings labeled "Show OpenCL options". I was just wondering what each of the new icons mean (i.e. Ray buffer size, State Buffer count, etc) and how they each effect render times.

Thanks for your help!
Wingman_mn

[Dade]

Some information is available here: viewtopic.php?f=8&t=6413

If the values shown in the picture are the default, they aren't particularly good:

- Ray Buffer Size: is the number of rays sent to the gpu in a single shot. More is better for performances (however, at some point, it will too much and the performance will start to decrease because it will not feet anymore inside caches). More is bad for memory usage. 2048 is really a too small value, I use 16384 on my hardware.

- State Buffer Count: it is used to overlap the CPU and GPU computation. There will be no overlap with only 1 buffer. Some overlap with 2. 3 should offer a complete overlap. Again, more is bad for memory usage. 4 is a good value but 3 should offer the same performance with less memory usage.

- OpenCL workgroup size: it is the wavefront size using CUDA terminology. It is a bit hard to explain but 64 is the best value for about any GPU available today (the only exception could be the NVIDIA 460 GTX). 0 is a particularly bad value because it leave to OpenCL the freedom to choose the size and drivers select a bad value 99.99% of times.

To recap, I use the following settings:

Ray Buffer Size=16384
State Buffer Count=2
OpenCL workgroup size=64

[J the Ninja]

If the defaults are bad, why not just change them? I could change the defaults here pretty easily, but they don't actually export unless the user enables "Show OpenCL Options", so perhaps it would be best to change them in the core?

[luxRoot]

Is there a lag time when Hybrid Bi-dir mode is supposed to "kick in" ?

I used this (viewtopic.php?f=15&t=2977) scene to do some testing with these settings

Ray Buffer Size=16384
State Buffer Count=2
OpenCL workgroup size=64

Hybrid Bi-dir 1 min of rendering time (6.7 s/p)
Bi-dir 1 min of rendering time (6 s/p)

kC/s and kS/s differed pretty much the same percentage as s/p

Intel 720QM / GTX 230 / 4GB / Win7 x64 / Blender 2.60 off release / Lux 0.9 dev latest testing build.

I'm very happy my PC didn't freeze this time in RAM starvation, I probably overkilled the settings when doing testing last time

[Lord Crc]

Is there a lag time when Hybrid Bi-dir mode is supposed to "kick in" ?
[...]
Intel 720QM / GTX 230

I suspect it will be when you upgrade that CPU What's the GPU utilization you get?

[J the Ninja]

Is there a lag time when Hybrid Bi-dir mode is supposed to "kick in" ?
[...]
Intel 720QM / GTX 230

I suspect it will be when you upgrade that CPU What's the GPU utilization you get?

Assuming he meant the GT 230M, i doubt it. The 230M seems to just be a higher clocked version of the 9600M GT in my MacBook. And it can easily keep the 9600 AND the 9400 IGP busy...with its 2.4Ghz Core 2 Duo

[luxRoot]

Is there a lag time when Hybrid Bi-dir mode is supposed to "kick in" ?
[...]
Intel 720QM / GTX 230

I suspect it will be when you upgrade that CPU What's the GPU utilization you get?

30-40%, and yes i have GTM 230, Funny thing but CPU loads were 50% at it's max as well

[J the Ninja]

Is there a lag time when Hybrid Bi-dir mode is supposed to "kick in" ?
[...]
Intel 720QM / GTX 230

I suspect it will be when you upgrade that CPU What's the GPU utilization you get?

30-40%, and yes i have GTM 230, Funny thing but CPU loads were 50% at it's max as well

The CPU was at 50%? Yeah, it can keep your GPU maxed out with power to spare.

[Wingman_mn]

My goodness... thanks Dade, your settings decreased my render times by ~30%.

Only problem I have now is that the render keeps rendering after the haltspp is met. My latest test reached the haltspp at 16 min but continued on until I stopped it 49 minutes later... for animations one could say this is unacceptable.

I'm assuming that this is because the renderer is sending all these rays to the gpu and won't stop until all the rays in that 'bundle' are traced, is this correct? Is there anyway of forcing a stop to happen short of closing luxrender?

Thanks for you help you guys!
Wingman_mn

[luxRoot]

The CPU was at 50%? Yeah, it can keep your GPU maxed out with power to spare.

No, the problem is that ssems like both GPU + CPU have resources to spare, but they don't.

On the screenshots you can see that, when using Hybrid, RAM get's drained like hell, but both CPU and GPU are being two lazy lads making me wait longer, on the Other hand non-hybrid bidir is jamming that CPU throttle bar to it's fullest, not getting so hard on the RAM.

Do I need more RAM to fully load CPU+GPU or some setting tweaks might help ? Setting State Buffer Count to 3 freezes the PC and I can't increase Ray Buffer Size as it's already maxed.