I recommend downloading slides and using PowerPoint to see speaker notes and embedded videos. Lots of details in speaker notes. Web viewers don't show this.
-
-
এই থ্রেডটি দেখানধন্যবাদ। আপনার সময়রেখাকে আরো ভালো করে তুলতে টুইটার এটিকে ব্যবহার করবে। পূর্বাবস্থায়পূর্বাবস্থায়
-
-
-
Thanks *so* much for these! I've been looking into 2D GPU sim stuff and your numbers make me quite optimistic. One question: your slides mention both Gauss-Seidel and Jacobi as "failed techniques".. so what are you using to solve collision vs shape-matching results? :)
-
(I'm also curious about whether you have a reference for your integration scheme; wikipedia is telling me that BDF2 is implicit - does that mean you're solving a system during integration (seems expensive with 16k+ particles), or is there a simple "semi-implicit"-style approach?)
-
I can put my integration code in a git snippet tomorrow. IIRC it was pretty straightforward port from some Miles Macklin paper.
-
Shape matching solves all body constraints. First you integrate and displace by SDF collisions (move penetrating particles back to the surface along gradient). Then you shape match every body separately. Works perfectly fine.
-
ah yeah, whoops.. that makes sense. And just 1 iteration of shape-matching per frame?? That's pretty awesome :) (Sadly I'd like to get more general 2D constraints happening, so I guess I'll have to try GS/Jacobi and hope they don't suck for my use-case)
-
GS with graph colorization is very stable. Was very fast too, if we could put whole islands to groupshared memory. 2d has less constraints than 3d. Should scale much better.
-
yep, less constraints and also smaller footprint per particle... I'm optimistic :)
কথা-বার্তা শেষ
নতুন কথা-বার্তা -
-
-
cant download the slides? anything wrong with dropbox ?
-
Hmm. Download seems to work fine on my computer and phone. I’ll retest at work.
-
Strange. I downloaded the slides to work computer too earier today. And to my phone too. Both worked fine.
কথা-বার্তা শেষ
নতুন কথা-বার্তা -
-
-
Tried new stuff from the NV RTX and Titan V? Any perf. bonus?
-
I don’t have TitanV unfortunately.
-
Used two for AI, contemplating on buying a private one, but the price...
-
Will do path tracing experiments soon on top of Claybook. If results are good enough, I will be releasing a high end path traced mode for fast PC GPUs. Would be interesting to test the Nvidia TitanV ML denoiser.
-
I think they're using its tensor cores for denoising; they allegedly got the level of quality equal to a 1000 spp with 1spp+denoiser. That's mind-blowing knowing that at 16 spp vs. 512 spp I get a 70% perf. drop for a few spheres. :)
কথা-বার্তা শেষ
নতুন কথা-বার্তা -
-
-
It was a pleasure meeting you at GDC and attending your talk. Happy to see a game with so much GPU physics simulation!
-
It was nice talking to you too! Our team used your physics engine for 3 titles at Redlynx/Ubisoft: https://en.m.wikipedia.org/wiki/Trials_(series) …
কথা-বার্তা শেষ
নতুন কথা-বার্তা -
-
-
Will this one be on the Vault? Keen to see this even if I don't fully understand any of it!
-
Yes. The presentation was video recorded. Will be on Vault.
কথা-বার্তা শেষ
নতুন কথা-বার্তা -
লোড হতে বেশ কিছুক্ষণ সময় নিচ্ছে।
টুইটার তার ক্ষমতার বাইরে চলে গেছে বা কোনো সাময়িক সমস্যার সম্মুখীন হয়েছে আবার চেষ্টা করুন বা আরও তথ্যের জন্য টুইটারের স্থিতি দেখুন।