NewtonTreeCollisionEndBuild crash

[bmwop]

Hi!

the NewtonTreeCollisionEndBuild() crashes in 2.24 (it was okay in 2.09)

test code:

Code: Select all

  NewtonWorld *nWorld = NewtonCreate(/*NULL, NULL*/);
    NewtonSetThreadsCount(nWorld, 2);

    {
        NewtonWorldCriticalSectionLock(nWorld);
        NewtonCollision* collision = NewtonCreateTreeCollision(nWorld, roadID/*, NULL*/);
        FILE* f = fopen("last_road_points", "r");
        NewtonTreeCollisionBeginBuild(collision);

        bool canRead = true;
        float vArray[9];

        printf("read points\n");
        while (canRead)
        {
            for (int wi = 0; wi < 9 && canRead; wi++)
            {
                int ret = fscanf(f, "%f\n", &vArray[wi]);
                if (ret < 1) canRead = false;
            }
            if (canRead)
            {
                NewtonTreeCollisionAddFace(collision, 3, (float*)vArray, 3 * sizeof(float), 1);
            }
        }
        fclose(f);
        printf("collisionendbuild\n");
        NewtonTreeCollisionEndBuild(collision, 0);
        printf("collisionendbuild done\n");
    }

for the test the following file contains the test points:
http://bmw.homelinux.org/newton/last_road_points

Am i doing something wrong?

[Julio Jerez]

why are you using a critical section?
critical section are meant to be used in callbacks.

and if you use it you are not closing the section, whi mean teh prgam will frezze or crash if it enter another section.

[bmwop]

I copied the code from the original place, so the CS is not needed here, but anyway it will crash, so it does not matter.

here is the backtrace from the gdb:
#0 0xb6c21d8c in memcpy () from /lib/libc.so.6
#1 0xb7bad0e8 in _ZN15InternalGeoUtilL12SortVerticesEPfiiiif () from ./lib/libNewton.so
#2 0xb7badcd5 in dgVertexListToIndexList () from ./lib/libNewton.so
#3 0xb7bc4ef1 in dgAABBPolygonSoup::Create () from ./lib/libNewton.so
#4 0xb7c4c0ec in dgCollisionBVH::EndBuild () from ./lib/libNewton.so
#5 0xb7cba854 in NewtonTreeCollisionEndBuild () from ./lib/libNewton.so

[Julio Jerez]

Ok I pasted that code in the test demo, and you are correct it does crash.
Your mesh is one of those pathological arrays of points in which qsort algorithm result into a n O(n^2)
In 2.22 I replaced the standard qsort with a Template based sort that does not allocate memory
And is non recursive, this is much faster than the standard one.
but the stack pool is too small for pathological arrays.
The pool was set to 128 levels deep, but your mesh goes around 500+ deep.

I changed the stack pool to 1024, this is an extreme rare cituation that should never happen, so a stack of 1024 should be good for anything.

I fixed already, Get replacement library here
http://www.newtondynamics.com/downloads ... n-2.25.rar

[bmwop]

Thanks for the reply and the explanation. I checked that, and it was okay. (but my raycast car's tires behave strangely with this)

[Julio Jerez]

what do you mean that?
the fix did not change anything else. how was working before?

Is the debug display correct?

I see you are using libraries with .so extension. are you in Linux?
the fix in only in the PC build.
I am confused. how do you get a .so library?

[bmwop]

Here are two videos, one from the old (good) newton:
http://bmw.homelinux.org/newton/good.avi

and the other from the new 2.25 (wrong) newton:
http://bmw.homelinux.org/newton/wrong.avi

This car is mostly build on the DGRaycastCar.
I am using both linux and windows (with gcc).

[Julio Jerez]

I can not see the videos it just play the audio, can you post it youtube or
better can you post and executable of the windows vertion?
Maybe something else is wrong

[bmwop]

the fraps uses a strange avi format to record video, i can play it with mplayer but cannot convert it, so i made a video with my mobile:
http://bmw.homelinux.org/newton/wrong.wmv
http://bmw.homelinux.org/newton/wrong.mov

it is only from the wrong case, the tires should move totally straight on the road.

[Julio Jerez]

It look like Raycast is boosted?
Are you using the joint librray from 2.25?, dot not use the library only the Newton libraries.

also can you make the video with the debug collision, jus to make sure the polygons are there?

[bmwop]

It look like Raycast is boosted?
Are you using the joint librray from 2.25?, dot not use the library only the Newton libraries.

no, i am using some older with modifications, but i compiled it myself, so i don't use the jointlibrary.dll.

also can you make the video with the debug collision, jus to make sure the polygons are there?

Do you mean the NewtonStaticCollisionSetDebugCallback?

I'm not sure that a raycast object really generates a collision callback in this case, but maybe i can drop a real object on it to see how it behaves. Do you have a linux (32bit) build of the 2.25? Because i can try it only on linux in these days.

Anyway the car does the same thing on a collision tree and the height field object.

[Julio Jerez]

But the 2.25 version I posted with the Fix was compiled on Windows, how did you go around the crash to make it work?

[bmwop]

I can use both linux and windows. The linux is more convenient and i can use during work too

[Julio Jerez]

for me I have to use Window because the Linux build has not being updated.
I ususally build Mac and Linux after I complete a milestone, and nwo I am in teh middle of more than one,
otherwise it is too much work debuging in more than one OS.

can you post a copy of teh Linux Build for debugging please? one that uses the DLL

[bmwop]

can you post a copy of teh Linux Build for debugging please? one that uses the DLL

i think there is a small misunderstanding here, i have tested the 2.25 only on windows and i don't have linux build with the 2.25. Or what linux build do you need?

Anyway it is not a serious problem for me now, so if you don't have much time for debug this than i understand. And i read in some earlier post you are working on some new raycast car.