NewtonWaitForUpdateToFinish problem

[JernejL]

NewtonWaitForUpdateToFinish doesn't seem to work in latest build anymore.

When i call NewtonWaitForUpdateToFinish it seems like i get callbacks called after that, if i replace call with just NewtonUpdate, everything starts to work properly again - that is, all callbacks are done within that call.

I tested this by setting a global variable before calling NewtonUpdateAsync and un-setting it after NewtonWaitForUpdateToFinish, i put assert into transform callback and it shows it being called after NewtonWaitForUpdateToFinish, and this breaks a lot of thread-unsafe variables that rely on newton calls being between newton async + wait calls, and as consequence gives me very strange data corruption problems.

[Julio Jerez]

That would be a really bad bug. I have not touched that in a long time.

can you edit the test demo we're I put you mesh to reproduce the problem.

Edit:
Actually I will add that test to the sandbox this way it will show if it happens on any demo.

[JernejL]

My test case is pretty simple.

Just set a true / false flag before NewtonUpdateAsync and clear it after NewtonWaitForUpdateToFinish, then check it in your transform callbacks.

Also a question: when is continous collision coming back? i have really big problems with bodies tunneling due to this that did not happen in previous versions of newton

[Julio Jerez]

I am adding that sync check to the engine side.

then let us see about the CCD.

[JernejL]

Here is a video showing NewtonWaitForUpdateToFinish bug:

https://www.youtube.com/watch?v=Y8vzvD0I-h0

Please note.. call stack is broken in lazarus (and shows all newton functions as "newtoncollisionaggregateselfcollision" ), because it does not work well with PDB symbols.

[Julio Jerez]

I can't really read the text is too burry, and is set to highest quality 480 already.

I added the check to make sure is working as expected from the low level.
the check is here: file ../sdk\dgPhysics\dgWorld.cpp line 965
dgAssert(m_testAsynUpdate);

please sync and see if the sdk assert on your, I tested and seem to be fine with the sandbox.
My guess is that you are misusing it. but first les us see if the SDK assert.

[JernejL]

It's impractical to use debug compiled newton because there are so many other asserts that are triggered that are still in the code that i'd need to remove first (an example of assert that is unneccesary is dgAssert in NewtonMaterialSetContactNormalAcceleration function).

I'm not sure what youtube did to the video but it made it awful.. i'll try to make another one later without remote desktop involved :/

My setup is as following:

Code: Select all

                in_update:= true;

                //NewtonUpdate(Newton_ext.NWorld, 0.1 / 6);

                NewtonUpdateAsync(Newton_ext.NWorld, 0.1 / 6);
                NewtonWaitForUpdateToFinish(Newton_ext.NWorld); // sync physics thread

                in_update:= false;


in my forcetorque and transform callbacks i check in_update value with assert, and if i use NewtonUpdateAsync + NewtonWaitForUpdateToFinish the assert is triggering, indicating that callbacks were triggered after NewtonWaitForUpdateToFinish.

[Dave Gravel]

You can build the debug dll with this flag DG_DISABLE_ASSERT for remove the assert in debug mode.
It is not good for all situations but maybe it can help you for debug from pascal side.

[Julio Jerez]

It's impractical to use debug compiled newton because there are so many other asserts that are triggered that are still in the code that i'd need to remove first
(an example of assert that is unneccesary is dgAssert in NewtonMaterialSetContactNormalAcceleration function).

the asserts are a sign something is not quite right. yes some are unnecessary but I added them because at the time I did not have a way to test the functionality.
I removed that one below.

Code: Select all

void NewtonMaterialSetContactNormalAcceleration(const NewtonMaterial* const materialHandle, dFloat accel)
{
   TRACE_FUNCTION(__FUNCTION__);
   dgContactMaterial* const material = (dgContactMaterial*) materialHandle;
   dgAssert(0);
   material->m_normal_Force.m_force = accel;
   material->m_flags |= dgContactMaterial::m_overrideNormalAccel;
}


in my forcetorque and transform callbacks i check in_update value with assert, and if i use NewtonUpdateAsync + NewtonWaitForUpdateToFinish the assert is triggering, indicating that callbacks were triggered after NewtonWaitForUpdateToFinish.

what variable is in_update?
the one added is called dgAssert(m_testAsynUpdate);
please sync again let us see if we can clear the assert, I can run most the time in debug without assert triggering, you should be able to do the same.

[Julio Jerez]

I assume you meant 1.0/60.0
I will add the same set up to the sandbox demo and see if I reproduce the same bug.

Code: Select all

in_update:= true;
                //NewtonUpdate(Newton_ext.NWorld, 0.1 / 6);

                NewtonUpdateAsync(Newton_ext.NWorld, 0.1 / 6);
                NewtonWaitForUpdateToFinish(Newton_ext.NWorld); // sync physics thread

                in_update:= false;

but I am not trying to be pedantry, the asserts are important.

[Julio Jerez]

I just added your test and you are right the assert happen the force and torque callback.
It seems it should not but is does.
I am debugging it.

[JernejL]

Glad to be able to help finding bugs let me know when a patch is made for this so i can re-compile newton dll and re-test.

[Julio Jerez]

actually this broke about a year ago when I try to make the Engine thread work in both mode synchronous and asynchronous. that's always a big mistake. Ii added too much complexity to the state machine logic and in the end it do not work.

I re stated the old synchronous thread and for the time been I made the Async update call teh SynUpdate this make the system work, and late tonight I will add the AsyncThread.
Please sync when you have time and tell my is thsi work, even is not Async it just pretend it is.

This also allows us to address teh continues collsion problem.

[JernejL]

Ok, i've synced it and now it works well - no more asserts.

I can now make tests with continous collision when you have fixes ready

[Julio Jerez]

if you sync before this morning please do it again, I committed the Async functionality late last night.
I tested and worked fine here.
also on this: //NewtonUpdate(Newton_ext.NWorld, 0.1 / 6);

I assume you mean //NewtonUpdate(Newton_ext.NWorld, 0.1 / 60.0f);

also of the CCD, here si oen thong you can do that will make the entire Game work much mich better
no just the CCD. if you are no doing it already, but I believe your aren't
after you create the newton word call function NewtonSetNumberOfSubsteps like this

Code: Select all

   // create the newton world
   m_world = NewtonCreate();

   // set the number of sub steps
   NewtonSetNumberOfSubsteps (m_world, 2;

The improvement is quite dramatic and the cost is marginal.
in you case you can even try 3,

please try that and tell how it goes.