Page 1 of 1

Dyverso - confused :?: unexpected results

Posted: Sat Aug 01, 2015 4:16 am
by hmercado
Hello,

I ran my first test with the Dyverso solver and the results got me really confused :roll:

Please see the following notes:

I tried to simulate around 1 million particles just to test how fast Dyverso performed.
Stepping and Liquid - PBD solver at default settings.

System spec: dual Xeon E5-2680 v2 - GTX 780 Ti - GTX TITAN Z
I have the GTX 780 Ti as my primary GPU and Dyverso will ONLY run on the TITAN Z GPU 0 regardless of which one I choose in the "Simulation options"

Here is a link to the video results: http://youtu.be/UVVhMcpjM9Q

TEST-01
Dyverso Domain - Circle emitter setup with speed
CPU

- On the CPU the simulation seem to run very fast as it was expected, but then when the fluid was starting to settle it exploded upwards for no apparent reason.

GPU
- On the GPU I was expecting to see a faster simulation since it was announced that Dyverso was GPU accelerated right? wrong! :roll: the same scene took an hour and a half :shock: :shock:
Then again towards the end of the time range the fluid also exploded upwards.

TEST-02
Dyverso Domain - square emitter setup as volume
CPU

-Again on the CPU only Dyverso seems to simulate faster, but then the particle volume got compressed towards the end of the simulation :?:

GPU
- The simulation once again its running really slow :roll:

Any Idea what is going on?

Thanks

Re: Dyverso - confused :?: unexpected results

Posted: Mon Aug 03, 2015 1:23 pm
by ivalduan
Hello hmercado,

The problem that you can't select the GPU card to use during simulation seems like a bug. We tested the selection feature with two ATi cards and it worked well, how are you measuring that Dyverso is always using the Titan Z?

About simulation results CPU vs GPU, you have a really high-performance CPU platform with 40 virtual-cores, Dyverso solvers are GPU accelerated but other parts of RealFlow like Daemons still run on CPU, all those transfers of information make it difficult to beat 40-cores with any GPU in the market until we move the entire RealFlow to work inside GPU. Here we measure something between 3x/2x extra performance with 8-core/12-core machines, for now I will recommend to stay with the 40-cores in that platform, simulations are going to be really fast with that.

About jumps in that PBD simulations try to increase the number of minimum substeps or modify the PBD solver parameters to let the fluid converge as Hybrido. We will have some tutorials online soon about how simulation parameters affect PBD-fluid behavior.

Cheers.
Ivan.

Re: Dyverso - confused :?: unexpected results

Posted: Mon Aug 03, 2015 2:17 pm
by hmercado
Hello Ivan

I can select any GPU in the simulation options but when the simulation start it is only running on the GPU when the one selected is the TITAN Z id 0. If I select the TITAN Z id 1 or the GTX 780ti the simulation runs only on CPU. It's like it's ignoring the other GPUs

I can see which GPU is been used with GPU-Z, EVGA precision X or the NVIDIA windows desktop widget.

The scene I was testing was a very simple one with just a Dyverso domain, a single emitter and a single deamon (gravity)

Same thing as with Hybrido in 2014, with GPU selected it ran actually much slower and no GPU was actually use during the simulation. I haven't tested it on 2015 yet.

Are you saying you have plans to make the whole realflow code to run on GPU ? :)

I am looking forward to more tutorials about simulation parameters :)

On a side note I think I have found a couple of bugs while testing 2015.

- Geometry text option "uniform triangles" crashes RealFlow

This two I tried after running the GPU tests so it may or may not be related. Closing an opening RealFlow fixed it.
- dspline deamon created with no visual display. Unable to enter edit mode in the nodes parameters palette. Entering edit mode with the button in the left toolbar was possible but was not able to add any points
- Object emitters (standard and dyverso) unable to select object in the nodes parameters palette

Thanks

Re: Dyverso - confused :?: unexpected results

Posted: Tue Aug 04, 2015 9:32 am
by Alex
Hi there,

just a quick question regarding one of the bugs you found:
Geometry text option "uniform triangles" crashes RealFlow
Can you provide an exact text sequence, font and size which causes crashes in your system?

Thanks!

Re: Dyverso - confused :?: unexpected results

Posted: Tue Aug 04, 2015 2:39 pm
by hmercado
Hi,

Always crashes when I turn on "uniform triangles" :roll:
Not that I need that option (don't even know what exactly is supposed to do), just checked it to try an see what it would do and it just crashed.

Font Arial, regular, size 12.
All other text creation options at default values; just "uniform triangles" on
First time I tried simply "RF"; just tried with "realflow" and some other random text sequence and it still crashes.
If I type "A" the text object is created but not visible in the viewport. If I type "a" real flow crashes.

Thanks