This is 100K objects, 10substeps, real gravity. First frame(sending objects geo. to bullet) takes 3sec. Average time(200frames) is 4sec. And quality is better than bulletDOP. It doesnt update obj. properties every frame(only init), but Its around (1-5)% slower If it updates.

So I am still thinking that I made some mistake(s) in DOP setting.


edit:
If I extend attachment file to 100K objs, First frame takes 1m44s, and 2-4frame average around 1m30s. Other problem is, that with bulletSOP I can have(and I have it) different size of boxes with +-same sim time, but for this I used “RBD point object”, but If I use “RBD fracture object” it crashes, because out-of-memory(16GB). With bulletSOP I can have millions of different boxes!

edit2: no multithreading, no GPU

edit3: Sorry, I miscalculated the time of my bulletSOP sim. Its not 7.2s its only 4s!!! :shock:

Awesome sim times. Release the storm!

TimeShifted Christmas celebration !!?

yeah ..
.. better to be ‘consumed’ long before 12AM of 21 December 2012 .


–


impressive performance cybermax ! bravo .

.cheers

thx,
I want to release it before Xmas(early December, next weekend maybe). Plugin is already done, but We are still working on few bugs, and finding speed issue.

Otherwise, I rewrited completely all source code. So everything is faster. For eg. MultiConstraint takes only 1s in my test scene, Old Multijoint took 1h in same scene. Other issue, what I have solved, was problem with updating(changing) data. So It doesnt use feedbackSOP or fileSOP nodes anymore, but I wrote special node, which read data from previous frame, so Its very easy to use! Better than using new solverSOP, which came with H12.

This is great news, can't wait…..!!!!

This is amazing. Thanks a lot.
Any chance we can see some of your source code for Christmas present
I would be really interested to see some of your bindings for bullet even from the older release.

sabotage3d
Any chance we can see some of your source code for Christmas present

If everything goes well, you will be more surprised, that you can imagine :shock:

new video: https://vimeo.com/54185240 [vimeo.com]

right now, I am computing 500K objs simulation and after 40frames its still under 40sec/frame, but until I will have video, I can not say anything about quality :?

Other news: I will build version for Win64, Linux64 and OSX64

Is it possible to use all CPU cores for your solver? It would be great I think.
Could you make test with more complex shapes than boxes? For example voronoi pieces.

almatea
Is it possible to use all CPU cores for your solver? It would be great I think.
Could you make test with more complex shapes than boxes? For example voronoi pieces.

Bullet isn't fully mullti threaded yet, it's pretty fast tho. What can be multithreaded is the process of adding back the transformation to each RBD chunk, if what he's getting back from bullet is the centroid points with the transformations for each object.

-Rick

Ratmann

almatea
Is it possible to use all CPU cores for your solver? It would be great I think.
Could you make test with more complex shapes than boxes? For example voronoi pieces.

Bullet isn't fully mullti threaded yet, it's pretty fast tho. What can be multithreaded is the process of adding back the transformation to each RBD chunk, if what he's getting back from bullet is the centroid points with the transformations for each object.

-Rick

Yes mostly Bullet was not multi-thread before. But I saw some methods for multi-thread workflow that did not work in early versions of Bullet. Do it work now may be? It would be great to check it.
And OpenCL: http://youtu.be/8jGZv1YYe2c [youtu.be] of 2.80 is great improvement.

I think It will be better to wait on Bullet 3.0. I tried some multi-threading stuff, but its not so stable, but its faster. I have 4cores with intel HT, so normally its using 12%, but I can get around 30% If I use different classes. Other things is turn of HT in bios. I havent tried yet, but It can add some speed too. OpenCL code has not been tested, yet.

Otherwise, We have few big problems left(btGImpactMeshShape , …), but I want to release at least demo to Xmas day

great Milan!

keep it rolling!

buki
great Milan!
keep it rolling!

thx


I tried turn off Intel HT => 4cores only(not 8virtual cores) and …. It has same sim time!!! Very, very strange :shock:

Anyway, for this “more than 1milion boxes scene”, takes 25sec to load all objects to bullet and around 30secs for sim of one frame(tested on first 15frames, so It will be slower when it starts to collide)

Sneak peek: https://vimeo.com/57193837 [vimeo.com]

This weekend I am releasing first tutorial and beta version will come during next week!

PS: Sorry for my english in video, I know, its bad, but I think its still better, than no sound, subtitles or background music

You rule man!

buki
You rule man!

thx man


ok, this is absolutely awesome! I installed the lastest production build(185) and everything is working fine even plugin is compiled for previous production build(125). Can somebody tell me more about this? For eg. If it still true that it is better recompile it for every version?, ….


edit: from other news, I have compiled version for Linux and OSX, but there is still little bug(but basically its working), so First beta will be for Win64 and in next few days other OS will come!

BulletSOP beta for win64 will be public tomorrow! Other OS will come in next few days.

I sent request on OdForce forum for creating subforum for this plugin in HDK directory(there is “Houdini Ocean Toolkit”). I'm waiting for answer since saturday, so I will release it here. It will be a little bit chaos to respond on more questions in one thread, but I dont want to create new forum, because you would have to register, …

I'll get a sub-forum made for you over at od right now./…


this has been done:
http://forums.odforce.net/index.php?/forum/69-bullet/ [forums.odforce.net]

oho ! cannt wait ..

great idea to open its dedicated section on odForce too !

it shd fit better there , imho .




.cheers