Closed GoogleCodeExporter closed 9 years ago
This problem happens only with MultiThreaded setup.
Original comment by nharisha...@gmail.com
on 9 Oct 2012 at 7:42
Can you re-create the problem in a modified Bullet demo and attach it to this
issue?
Thanks a lot,
Erwin
Original comment by erwin.coumans
on 9 Oct 2012 at 3:54
I have seen this issue before, this was due to
badly generated convex hulls in our case, from single triangle meshes
or completly flat meshes.
Original comment by majestik...@gmail.com
on 10 Oct 2012 at 2:57
We used v2.80 and it was working fine so far. Once we replaced the sources with
2.81, we had this issue.
I'll try to make a standalone example to reproduce this problem.
Original comment by nharisha...@gmail.com
on 10 Oct 2012 at 3:12
The _maxdot_large is a new SIMD optimized routine (contributed by Apple)
It requires the data to be 16-byte aligned. There might be some limitations of
the routine that I'm not aware of yet. How many vertices does your convex have?
Can you provide more information about the line number/location of the crash,
in debug mode?
Original comment by erwin.coumans
on 10 Oct 2012 at 3:34
Looking at the code, it might be that for a concave trimesh, the vertices are
not 16-byte aligned, causing the crash.
If it is a showstopper, you can always disable the SIMD optimization
in LinearMath/btVector3.h check out btVector3::maxDot and change the line
if( array_count < scalar_cutoff )
into
if (1)
Still, we should fix it so I hope for an easy reproduction case.
Original comment by erwin.coumans
on 10 Oct 2012 at 3:44
One more possibility for unaligned data:
can you change
class spuNodeCallback : public btNodeOverlapCallback
into
ATTRIBUTE_ALIGNED16(class) spuNodeCallback : public btNodeOverlapCallback
in file
bullet\src\BulletMultiThreaded\SpuNarrowPhaseCollisionTask\SpuGatheringCollision
Task.cpp around line 294 and see if that helps?
Original comment by erwin.coumans
on 10 Oct 2012 at 3:53
@Erwin
The last suggestions fixed the problem. The first suggestion didn't help. I'll
try to test it in few more levels today to confirm this fix. :)
Thanks a lot.
Original comment by nharisha...@gmail.com
on 10 Oct 2012 at 4:05
Thanks for the feedback!
I'm baffled that the first suggestions didn't help, changing the line
if( array_count < scalar_cutoff )
into
if (1)
because in that case the _maxdot_large method should not be called.
Original comment by erwin.coumans
on 10 Oct 2012 at 4:17
I am wondering if this is the line producing the problem.
https://code.google.com/p/bullet/source/browse/trunk/src/LinearMath/btVector3.h#
994
Original comment by nharisha...@gmail.com
on 10 Oct 2012 at 4:40
Ah, indeed. In case the data is not aligned, you need to disable both _maxdot
and _mindot.
I checked in the ATTRIBUTE_ALIGNED16 fix, can you let me know if you see any
other crashes?
Original comment by erwin.coumans
on 10 Oct 2012 at 4:42
I tried running the fixed code on different ConvexHull and TriMesh Combinations
and there was no crashes so far. I'll try MacOSX and IOS port of the same and
let you know later today.
Original comment by nharisha...@gmail.com
on 10 Oct 2012 at 4:47
On the same Win32 project after I rebuilt all projects, I still have the same
crash, when objects collide. The ATTRIBUTE_ALIGNED16 fix didn't help.
I just run the code in debug mode:
.exe!_maxdot_large(const float * vv, const float * vec, unsigned long count,
float * dotResult) Line 187 C++
.exe!convexHullSupport(const btVector3 & localDirOrg, const btVector3 * points,
int numPoints, const btVector3 & localScaling) Line 113 + 0x12d bytes C++
.exe!btConvexShape::localGetSupportVertexWithoutMarginNonVirtual(const
btVector3 & localDir) Line 276 + 0x1d bytes C++
.exe!btGjkPairDetector::getClosestPointsNonVirtual(const
btDiscreteCollisionDetectorInterface::ClosestPointInput & input,
btDiscreteCollisionDetectorInterface::Result & output, btIDebugDraw *
debugDraw) Line 147 C++
.exe!btGjkPairDetector::getClosestPoints(const
btDiscreteCollisionDetectorInterface::ClosestPointInput & input,
btDiscreteCollisionDetectorInterface::Result & output, btIDebugDraw *
debugDraw, bool swapResults) Line 77 C++
.exe!ProcessSpuConvexConvexCollision(SpuCollisionPairInput * wuInput,
CollisionTask_LocalStoreMemory * lsMemPtr, SpuContactResult & spuContacts)
Line 781 C++
.exe!handleCollisionPair(SpuCollisionPairInput & collisionPairInput,
CollisionTask_LocalStoreMemory & lsMem, SpuContactResult & spuContacts,
unsigned int collisionShape0Ptr, void * collisionShape0Loc, unsigned int
collisionShape1Ptr, void * collisionShape1Loc, bool dmaShapes) Line 861 C++
.exe!processCollisionTask(void * userPtr, void * lsMemPtr) Line 1385 + 0x32
bytes C++
.exe!Thread_no_1(void * lpParam) Line 67 C++
Original comment by nharisha...@gmail.com
on 10 Oct 2012 at 5:29
I reproduced it using the multi-threaded benchmark, and fixed it in latest
trunk:
https://code.google.com/p/bullet/source/detail?r=2616
thanks for the report!
Original comment by erwin.coumans
on 10 Oct 2012 at 6:36
[deleted comment]
I have similar problem in 2.8.1
when I run ios demo, err: undefined symbol. _maxdot_large && _mindot_large
After My search. I solved This problem.
Checking LinearMath--> btVector3.cpp
line 42&43
long _maxdot_large( const float *vv, const float *vec, unsigned long count,
float *dotResult );
long _maxdot_large( const float *vv, const float *vec, unsigned long count,
float *dotResult )
I found function declare at cpp.
I just move declare to .h file.
And everything goes right.
Hopes can help.
Original comment by dinosaur...@gmail.com
on 30 Sep 2013 at 2:34
Original issue reported on code.google.com by
nharisha...@gmail.com
on 9 Oct 2012 at 7:40