Roadmap for Newton 4

[FSA]

Is there a roadmap for newton 4? Like when will the core (collision/solver) be done, when the particles, then fluids, clothes, and other stuff.
Just a rough estimation in what time what feature will be available.
It doesn't need to be very accurate, but I need some Info for future planning.

Thanks

[Julio Jerez]

Road map is a dev tool used by manager to associate budgets to development time.
I am not on a budget, nor I am running a commercial enterprise.
Plus as much as I would like too, I can't afford to dedicate to this full time,
many factors had conspire to make that an impossibility.

I can give you an idea of what I have in my radar.

1-get the particle system functional and practical.
2-polish the fracture effect.
3-polish the vehicle and add the dynamic fully articulated character.

All these effects shows that they require massive volume of objected manipulation that are way beyond of what cpu can do. Even multicore cps can't deal with that kind of work load, since memory bandwidth and software multitasking become serious limitation bottleneck.

Newton has solved the problem of parallelism.
But it is clean that those other two issues are the dominant factor.
For example, with the rigid body solver, it is very good, accurate and stable, but once it pass few thousand joints, I the profiler shows that most time, all the time to one is spent on task switching. There is not way to go around that onless task switching is moved to the hardware.
GPUs do that at the hardware level so it is possible work work to achieve that, for CPUs this is an impossible barrier.

Another problem is memory bandwidth, a cpu can only move memory in small chunks, for example in the fluid simulation, the much expensive part is ready the particle ID. Even when that code is organized so that particle are coalesced, but the cpu can only get 4 at a time, in a gpu coalescing is a integral part of the architecture.

For these reasons and few other, the next step to solve them is to move to high perforce computing.
ther is not point is try to optimize code for CPU as it is was a GPU when the benefit is some time a regression.

I am given it another shot at open cl. I tried about three or four years ago, and found it to be in a state of anarchy among vendors, so I decided to way until is more stable.
I am trying again but all indications are that the only progress made is that Khronos made two or three more pie on the sky specifications that has been partially adopted or totally ignored and the result is that the state of affairs of open cl is now even worse.
I am giving it another shot, but if it keeps disappointing, as much as I dislike nvidia, I will probably have to hold my nose and try Cuda.

[FSA]

"Roadmap is often used as a synonym for a time schedule of different parts from a project."
Also I explained just in the next line what I mean:

Just a rough estimation in what time what feature will be available.

So it should be absolute clear, that it is not a "[...] dev tool used by manager to associate budgets to development time".

What I don't get: You get task switching (or context switching) mostly when running too many threads. Are you creating one thread per Joint//Body or why is task switching a bottleneck for you?

[JoeJ]

Road map is a dev tool used by manager to associate budgets to development
2-polish the fracture effect.

What are your plans for this?
AFAIK this was the plan:
Preprocessing step:
Have a mesh (manifold? convex?), provide a set of points. Newton than cuts the mesh into voronoi cells.
Runtime:
Calculate damage, probably evolving through a graph of joints, and break stuff. (Some multi level hierarchy with some fixed limits?)

As you talk about massive stuff on GPU, what numbers do you have in mind?
My far fetched goals on world creation would be procedural but high quality terrain, nature, and architecture. And then, maybe, simulate an earthquake and destruct a whole city. So i'm asking with offline content generation in mind.

I worked a lot on damage propagation myself recently, within the MPM fluid and also with another simulator more like SPH to increase detail. At some quantity it becomes a fluid simulation problem, but that's not what we expect from game runtimes, so just asking.