NewtonCollisionCollideContinue missing edge collisions

[Julio Jerez]

try again

[pHySiQuE]

And here is your demo. This uses a convex hull modifier as the primitive for the convex cast:
http://www.mediafire.com/?i1j0a173fsss6wn

[Julio Jerez]

I do not underntand the demo. I see a big black box spinning around and I can not control
spins even when I do not touch anything.

[Julio Jerez]

Now I check in a perfect parallet solver.

The solve is perfect is teh sence that it does not generate a worse solution than the sequencial solver in any case.
it is also simd optimized.

nwo teh parallet solve is slower than teh sequcncil solver if it no usin muticores, because it work on teh same data, but it has to run a preconditioning of the joint graph.
and the phase is and overhead. teh overehead is morginal but not negegible.

righ now teh precondition phase is naive, but thsi weekend I will make a high optimize one that will almos eleminate all of the threa contingeies,
even wit the naive precondinioner the solve is abput 30 % faste than the sequacial when usin flor threads and simd on buts.

then is everythong goes right, I will make this solve an AVX, and I beleive I should get faster preformance than using OpenCL on CPU.
the cool thing of this solver is that I can be coverted easslly to OpenCL. but I am not going to loose sleep thinking of that.

There one and only one reason why I would like to try OpenGL, and that is that GPUs has a zero eatancy task swithing mechanism,


ha also when you get the laste version try not uncomnet teh thread priorio define, I beeive I fix that isseiu with teh semphore and Waitforsingleobject function.
In fact sime the treads are so high performce and they take zero cpu time when tehy are not running I can even use a highe priority, but for now I will let as it is.
The cool thing is that now the engine runs concurrent, and it takes zero time from the main thread when is not running.

but I believe a also solved that problem for the PC mutithreaded version by using the priority mechanism for children threads. since the micro thread are Kerners that run for a very short period of time.

[Julio Jerez]

Th eparalle solve is so stable than In the pyramid I have to restle the engine by pick a box from the bottom and keep moving it,
if I want to compare the difference in performance with the sequencial solve.
if I do not do that, It bring the entire pyramid to sleep righ away. even in the middle of me moving the box, it is awesome.

If is was because it add soem extrat over head I woudl leave as the default solver, but no everyone is using level with high stacking plus many people use the determinisic mode.
anyway now I only nee to fix the memory trash bug, and that will be all for this solver.

another thing I can do nwo than teh solve is so good, is that I can cut the interation copun to half and double amost dopuble teh performace at the spcence of less acuracy.
we do that for Iphone , and it works fine, so maybe we can do that on PC too, and then newton will be in the same ball part of the comercial engines Havok and PhysX but with far superior accuracy and stability.

[pHySiQuE]

I do not underntand the demo. I see a big black box spinning around and I can not control
spins even when I do not touch anything.

What? This is what it looks like. It only requires OpenGL 2.1 support:

Untitled.jpg (62.05 KiB) Viewed 1917 times