Skip to content

A fast 2D and 3D game engine — TypeScript SDK, C++/WebAssembly core, visual editor. Ship one project to web, desktop, WeChat MiniGames, playable ads, and native Android / iOS.

97.9 KB
Raw Blame
// Copyright (c) 2001-2004 NovodeX AG. All rights reserved.
// Copyright (c) 2004-2008 AGEIA Technologies, Inc. All rights reserved.
// SPDX-FileCopyrightText: Copyright (c) 2008-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
// SPDX-License-Identifier: Apache-2.0

#include "DyDynamics.h"

#include "common/PxProfileZone.h"
#include "DyBodyCoreIntegrator.h"
#include "DySolverControl.h"
#include "DyConstraintPrep.h"
#include "DyConstraintPartition.h"
#include "CmFlushPool.h"
#include "PxsSimpleIslandManager.h"
#include "PxvSimStats.h"
#include "DyContactPrepShared.h"
#include "DySleep.h"
#include "DyIslandManager.h"
#if PGS_SUPPORT_COMPOUND_CONSTRAINTS
	#include "PxsMaterialManager.h"
	#include "DyContactReduction.h"
	#include "DyThreadContext.h"
#endif

//KS - used to turn on/off batched SIMD constraints.
#define DY_BATCH_CONSTRAINTS 1
//KS - used to specifically turn on/off batches 1D SIMD constraints.
#define DY_BATCH_1D 1

static const bool gMergePartitionAndFinalizeConstraintsTasks = true;

namespace physx
{
namespace Dy
{
// PT: TODO: what is const and what is not?
struct SolverIslandObjects
{
	PxsRigidBody**				bodies;
	FeatherstoneArticulation**	articulations;
	PxsIndexedContactManager*	contactManagers;	// PT: points to DynamicsContext::mContactList

	const IG::IslandId*			islandIds;
	PxU32						numIslands;
	PxU32*						bodyRemapTable;
	PxU32*						nodeIndexArray;

	PxSolverConstraintDesc*		constraintDescs;
	PxSolverConstraintDesc*		orderedConstraintDescs;
	PxSolverConstraintDesc*		tempConstraintDescs;
	PxConstraintBatchHeader*	constraintBatchHeaders;
	Cm::SpatialVector*			motionVelocities;
	PxsBodyCore**				bodyCoreArray;

	SolverIslandObjects() : bodies(NULL), articulations(NULL), 
		contactManagers(NULL), islandIds(NULL), numIslands(0), nodeIndexArray(NULL), constraintDescs(NULL), orderedConstraintDescs(NULL), 
		tempConstraintDescs(NULL), constraintBatchHeaders(NULL), motionVelocities(NULL), bodyCoreArray(NULL)
	{
	}
};

Context* createDynamicsContext(	PxcNpMemBlockPool* memBlockPool, Cm::FlushPool& taskPool, PxvSimStats& simStats,
								Cm::VirtualAllocatorCallback& allocator, PxsMaterialManager* materialManager,
								IG::SimpleIslandManager& islandManager, PxU64 contextID, PxReal maxBiasCoefficient,
								PxReal lengthScale, PxSceneFlags sceneFlags)
{
	return PX_NEW(DynamicsContext)(memBlockPool, taskPool, simStats, allocator, materialManager, islandManager, contextID, maxBiasCoefficient, lengthScale, sceneFlags);
}

void DynamicsContext::destroy()
{
	this->~DynamicsContext();
	PX_FREE_THIS;
}

DynamicsContext::DynamicsContext(	PxcNpMemBlockPool* memBlockPool,
									Cm::FlushPool& taskPool,
									PxvSimStats& simStats,
									Cm::VirtualAllocatorCallback& allocator,
									PxsMaterialManager* materialManager,
									IG::SimpleIslandManager& islandManager,
									PxU64 contextID,
									PxReal maxBiasCoefficient,
									PxReal lengthScale,
									PxSceneFlags sceneFlags) :
	DynamicsContextBase				(memBlockPool, taskPool, simStats, allocator, materialManager, islandManager, contextID, maxBiasCoefficient, lengthScale, sceneFlags),
	mSolveFrictionEveryIteration	(sceneFlags & PxSceneFlag::eENABLE_FRICTION_EVERY_ITERATION)
{
	mWorldSolverBody.linearVelocity = PxVec3(0.0f);
	mWorldSolverBody.angularState = PxVec3(0.0f);
	mWorldSolverBodyData.invMass = 0;
	mWorldSolverBodyData.sqrtInvInertia = PxMat33(PxZero);
	mWorldSolverBodyData.nodeIndex = PX_INVALID_NODE;
	mWorldSolverBodyData.reportThreshold = PX_MAX_REAL;
	mWorldSolverBodyData.penBiasClamp = -PX_MAX_REAL;
	mWorldSolverBodyData.maxContactImpulse = PX_MAX_REAL;
	mWorldSolverBody.solverProgress = MAX_PERMITTED_SOLVER_PROGRESS;
	mWorldSolverBody.maxSolverNormalProgress = MAX_PERMITTED_SOLVER_PROGRESS;
	mWorldSolverBody.maxSolverFrictionProgress = MAX_PERMITTED_SOLVER_PROGRESS;
	mWorldSolverBodyData.linearVelocity = mWorldSolverBodyData.angularVelocity = PxVec3(0.0f);
	mWorldSolverBodyData.body2World = PxTransform(PxIdentity);

	mBiasCoefficients.set(false, false, 0);  // note: for PGS the bias coefficients are independent of the number of position iterations
}

#if PX_ENABLE_SIM_STATS
void DynamicsContext::addThreadStats(const ThreadContext::ThreadSimStats& stats)
{
	mSimStats.mNbActiveConstraints += stats.numActiveConstraints;
	mSimStats.mNbActiveDynamicBodies += stats.numActiveDynamicBodies;
	mSimStats.mNbActiveKinematicBodies += stats.numActiveKinematicBodies;
	mSimStats.mNbAxisSolverConstraints += stats.numAxisSolverConstraints;
}
#else
	PX_CATCH_UNDEFINED_ENABLE_SIM_STATS
#endif

PX_COMPILE_TIME_ASSERT(PxsIndexedInteraction::eBODY == 0);
PX_COMPILE_TIME_ASSERT(PxsIndexedInteraction::eKINEMATIC == 1);

void DynamicsContext::setDescFromIndices_Contacts(PxSolverConstraintDesc& desc, const IG::IslandSim& islandSim, const PxsIndexedInteraction& constraint, PxU32 solverBodyOffset)
{
	const PxU32 offsetMap[] = {solverBodyOffset, 0};
	//const PxU32 offsetMap[] = {mKinematicCount, 0};

	if(constraint.indexType0 == PxsIndexedInteraction::eARTICULATION)
	{
		const PxNodeIndex nodeIndex0(constraint.articulation0);
		const IG::Node& node0 = islandSim.getNode(nodeIndex0);
		desc.articulationA = node0.mObject;
		desc.linkIndexA = nodeIndex0.articulationLinkId();
	}
	else
	{
		desc.linkIndexA = PxSolverConstraintDesc::RIGID_BODY;
		//desc.articulationALength = 0; //this is unioned with bodyADataIndex
		desc.bodyA = constraint.indexType0 == PxsIndexedInteraction::eWORLD ? &mWorldSolverBody
																			: &mSolverBodyPool[PxU32(constraint.solverBody0) + offsetMap[constraint.indexType0]];
		desc.bodyADataIndex = constraint.indexType0 == PxsIndexedInteraction::eWORLD ? 0
																			: PxU32(constraint.solverBody0) + 1 + offsetMap[constraint.indexType0];
	}

	if(constraint.indexType1 == PxsIndexedInteraction::eARTICULATION)
	{
		const PxNodeIndex nodeIndex1(constraint.articulation1);
		const IG::Node& node1 = islandSim.getNode(nodeIndex1);
		desc.articulationB = node1.mObject;
		desc.linkIndexB = nodeIndex1.articulationLinkId();// PxTo8(getLinkIndex(constraint.articulation1));
	}
	else
	{
		desc.linkIndexB = PxSolverConstraintDesc::RIGID_BODY;
		//desc.articulationBLength = 0; //this is unioned with bodyBDataIndex
		desc.bodyB = constraint.indexType1 == PxsIndexedInteraction::eWORLD ? &mWorldSolverBody
																			: &mSolverBodyPool[PxU32(constraint.solverBody1) + offsetMap[constraint.indexType1]];
		desc.bodyBDataIndex = constraint.indexType1 == PxsIndexedInteraction::eWORLD ? 0
																			: PxU32(constraint.solverBody1) + 1 + offsetMap[constraint.indexType1];
	}
}

void DynamicsContext::setDescFromIndices_Constraints(PxSolverConstraintDesc& desc, const IG::IslandSim& islandSim, IG::EdgeIndex edgeIndex, const PxU32* bodyRemap, PxU32 solverBodyOffset)
{
	const PxNodeIndex node1 = islandSim.mCpuData.getNodeIndex1(edgeIndex);
	if (node1.isStaticBody())
	{
		desc.bodyA = &mWorldSolverBody;
		desc.bodyADataIndex = 0;
		desc.linkIndexA = PxSolverConstraintDesc::RIGID_BODY;
	}
	else
	{
		const IG::Node& node = islandSim.getNode(node1);
		if (node.getNodeType() == IG::Node::eARTICULATION_TYPE)
		{
			desc.articulationA = getArticulationFromIG(islandSim, node1);
			desc.linkIndexA = node1.articulationLinkId();
		}
		else
		{
			const PxU32 activeIndex = islandSim.getActiveNodeIndex(node1);
			const PxU32 index = node.isKinematic() ? activeIndex : bodyRemap[activeIndex] + solverBodyOffset;
			desc.bodyA = &mSolverBodyPool[index];
			desc.bodyADataIndex = index + 1;
			desc.linkIndexA = PxSolverConstraintDesc::RIGID_BODY;
		}
	}

	const PxNodeIndex node2 = islandSim.mCpuData.getNodeIndex2(edgeIndex);
	if (node2.isStaticBody())
	{
		desc.bodyB = &mWorldSolverBody;
		desc.bodyBDataIndex = 0;
		desc.linkIndexB = PxSolverConstraintDesc::RIGID_BODY;
	}
	else
	{
		const IG::Node& node = islandSim.getNode(node2);
		if (node.getNodeType() == IG::Node::eARTICULATION_TYPE)
		{
			desc.articulationB = getArticulationFromIG(islandSim, node2);
			desc.linkIndexB = node2.articulationLinkId();
		}
		else
		{
			const PxU32 activeIndex = islandSim.getActiveNodeIndex(node2);
			const PxU32 index = node.isKinematic() ? activeIndex : bodyRemap[activeIndex] + solverBodyOffset;
			desc.bodyB = &mSolverBodyPool[index];
			desc.bodyBDataIndex = index + 1;
			desc.linkIndexB = PxSolverConstraintDesc::RIGID_BODY;
		}
	}
}

class PxsPreIntegrateTask : public Cm::Task
{
	PxsPreIntegrateTask& operator=(const PxsPreIntegrateTask&);
public:
	PxsPreIntegrateTask(	DynamicsContext&	context,
							PxsBodyCore*const*	bodyArray,
							PxsRigidBody*const*	originalBodyArray,
							PxU32 const*		nodeIndexArray,
							PxSolverBodyData*	solverBodyDataPool,
							PxF32				dt,
							PxU32				numBodies,
							volatile PxU32*		maxSolverPositionIterations,
							volatile PxU32*		maxSolverVelocityIterations,
							PxU32				startIndex,
							PxU32				numToIntegrate,
							const PxVec3&		gravity) :
		Cm::Task					(context.getContextId()),
		mContext					(context),
		mBodyArray					(bodyArray),
		mOriginalBodyArray			(originalBodyArray),
		mNodeIndexArray				(nodeIndexArray),
		mSolverBodyDataPool			(solverBodyDataPool),
		mDt							(dt),
		mNumBodies					(numBodies),
		mMaxSolverPositionIterations(maxSolverPositionIterations),
		mMaxSolverVelocityIterations(maxSolverVelocityIterations),
		mStartIndex					(startIndex),
		mNumToIntegrate				(numToIntegrate),
		mGravity					(gravity)
	{}

	virtual void runInternal() PX_OVERRIDE;

	virtual const char* getName() const PX_OVERRIDE
	{
		return "PxsDynamics.preIntegrate";
	}

public:
	DynamicsContext&		mContext;
	PxsBodyCore*const*		mBodyArray;
	PxsRigidBody*const*		mOriginalBodyArray;
	PxU32 const*			mNodeIndexArray;
	PxSolverBody*			mSolverBodies;
	PxSolverBodyData*		mSolverBodyDataPool;
	const PxF32				mDt;
	const PxU32				mNumBodies;
	volatile PxU32*			mMaxSolverPositionIterations;
	volatile PxU32*			mMaxSolverVelocityIterations;
	const PxU32				mStartIndex;
	const PxU32				mNumToIntegrate;
	const PxVec3			mGravity;
};

class PxsParallelSolverTask : public Cm::Task
{
	PxsParallelSolverTask& operator=(PxsParallelSolverTask&);
public:

	PxsParallelSolverTask(SolverIslandParams& params, DynamicsContext& context, IG::IslandSim& islandSim)
		: Cm::Task(context.getContextId()), mParams(params), mContext(context), mIslandSim(islandSim)
	{
	}

	virtual void runInternal() PX_OVERRIDE
	{
		solveParallel(mContext, mParams, mIslandSim);
	}

	virtual const char* getName() const PX_OVERRIDE
	{
		return "PxsDynamics.parallelSolver";
	}

	SolverIslandParams&	mParams;
	DynamicsContext&	mContext;
	IG::IslandSim&		mIslandSim;
};

#if PGS_SUPPORT_COMPOUND_CONSTRAINTS
#define PX_CONTACT_REDUCTION 1

class PxsSolverConstraintPostProcessTask : public Cm::Task
{
	PxsSolverConstraintPostProcessTask& operator=(const PxsSolverConstraintPostProcessTask&);
public:

	PxsSolverConstraintPostProcessTask(DynamicsContext& context,
		ThreadContext& threadContext,
		const SolverIslandObjects& objects,				  
		PxU32 solverBodyOffset,
		PxU32 startIndex,
		PxU32 stride,
		PxsMaterialManager* materialManager,
		PxsContactManagerOutputIterator& iterator) :
		Cm::Task			(context.getContextId()),
		mContext			(context), 
		mThreadContext		(threadContext),
		mObjects			(objects),
		mSolverBodyOffset	(solverBodyOffset),
		mStartIndex			(startIndex),
		mStride				(stride),
		mMaterialManager	(materialManager),
		mOutputs			(iterator)
	{}

	void mergeContacts(CompoundContactManager& header, ThreadContext& threadContext)
	{
		PxContactBuffer& buffer = threadContext.mContactBuffer;
		PxsMaterialInfo materialInfo[PxContactBuffer::MAX_CONTACTS];
		PxU32 size = 0;

		for(PxU32 a = 0; a < header.mStride; ++a)
		{
			PxsContactManager* manager = mThreadContext.orderedContactList[a+header.mStartIndex]->contactManager;
			PxcNpWorkUnit& unit = manager->getWorkUnit();
			const PxsContactManagerOutput& output = mOutputs.getContactManagerOutput(unit.mNpIndex);
			PxContactStreamIterator iter(output.contactPatches, output.contactPoints, output.getInternalFaceIndice(), output.nbPatches, output.nbContacts);

			PxU32 origSize = size;
			PX_UNUSED(origSize);
			if(!iter.forceNoResponse)
			{
				while(iter.hasNextPatch())
				{
					iter.nextPatch();
					while(iter.hasNextContact() && (size < PxContactBuffer::MAX_CONTACTS))
					{
						PX_ASSERT(size < PxContactBuffer::MAX_CONTACTS);
						iter.nextContact();
						PxsMaterialInfo& info = materialInfo[size];
						PxContactPoint& point = buffer.contacts[size++];
						point.dynamicFriction = iter.getDynamicFriction();
						point.staticFriction = iter.getStaticFriction();
						point.restitution = iter.getRestitution();
						point.internalFaceIndex1 = iter.getFaceIndex1();
						PX_ASSERT(iter.getMaterialFlags() <= PX_MAX_U8);
						point.materialFlags = PxU8(iter.getMaterialFlags());
						point.maxImpulse = iter.getMaxImpulse();
						point.targetVel = iter.getTargetVel();
						point.normal = iter.getContactNormal();
						point.point = iter.getContactPoint();
						point.separation = iter.getSeparation();
						info.mMaterialIndex0 = iter.getMaterialIndex0();
						info.mMaterialIndex1 = iter.getMaterialIndex1();
					}
				}
				PX_ASSERT(output.nbContacts == (size - origSize));
			}
		}

		PxU32 origSize = size;
#if PX_CONTACT_REDUCTION
		ContactReduction<6, 6> reduction(buffer.contacts, materialInfo, size);
		reduction.reduceContacts();
		//OK, now we write back the contacts...

		PxU8 histo[PxContactBuffer::MAX_CONTACTS];
		PxMemZero(histo, sizeof(histo));

		size = 0;
		for(PxU32 a = 0; a < reduction.mNumPatches; ++a)
		{
			ReducedContactPatch<6>& patch = reduction.mPatches[a];
			for(PxU32 b = 0; b < patch.numContactPoints; ++b)
			{
				histo[patch.contactPoints[b]] = 1;
				++size;
			}
		}
#endif

		// TODO AD: this probably needs an overflow check as well.
		PxU16* PX_RESTRICT data = reinterpret_cast<PxU16*>(threadContext.mConstraintBlockStream.reserve(size * sizeof(PxU16), mThreadContext.mConstraintBlockManager));

		header.forceBufferList = data;
		
#if PX_CONTACT_REDUCTION
		const PxU32 reservedSize = size;
		PX_UNUSED(reservedSize);
		size = 0;
		for(PxU32 a = 0; a < origSize; ++a)
		{
			if(histo[a])
			{
				if(size != a)
				{
					buffer.contacts[size] = buffer.contacts[a];
					materialInfo[size] = materialInfo[a];
				}
				data[size] = PxTo16(a);
				size++;
			}
		}
		PX_ASSERT(reservedSize >= size);
#else
		for(PxU32 a = 0; a < size; ++a)
			data[a] = a;
#endif

		PxU32 contactForceByteSize = size * sizeof(PxReal);

		PxsContactManagerOutput& output = mOutputs.getContactManagerOutput(header.unit->mNpIndex);

		PxU16 compressedContactSize;

		physx::writeCompressedContact(buffer.contacts, size, NULL, output.nbContacts, output.contactPatches, output.contactPoints, compressedContactSize,
			reinterpret_cast<PxReal*&>(output.contactForces), contactForceByteSize,
			output.frictionPatches, NULL,
			mMaterialManager, false,
			false, materialInfo, output.nbPatches, 0, &mThreadContext.mConstraintBlockManager, &threadContext.mConstraintBlockStream, false);
	}

	virtual void runInternal() PX_OVERRIDE
	{
		PX_PROFILE_ZONE("ConstraintPostProcess", mContextID);

		PxU32 endIndex = mStartIndex + mStride;

		ThreadContext* threadContext = mContext.getThreadContext();
		//TODO - we need to do this somewhere else
		//threadContext->mContactBlockStream.reset();
		threadContext->mConstraintBlockStream.reset();

		for(PxU32 a = mStartIndex; a < endIndex; ++a)
		{
			mergeContacts(mThreadContext.compoundConstraints[a], *threadContext);
		}
		mContext.putThreadContext(threadContext);
	}

	virtual const char* getName() const PX_OVERRIDE { return "PxsDynamics.solverConstraintPostProcess"; }

	DynamicsContext&			mContext;
	ThreadContext&				mThreadContext;
	const SolverIslandObjects	mObjects;
	const PxU32					mSolverBodyOffset;
	const PxU32					mStartIndex;
	const PxU32					mStride;
	PxsMaterialManager*			mMaterialManager;
	PxsContactManagerOutputIterator& mOutputs;
};
#endif

class PxsForceThresholdTask  : public Cm::Task
{
	DynamicsContext&		mDynamicsContext;

	PxsForceThresholdTask& operator=(const PxsForceThresholdTask&);
public:

	PxsForceThresholdTask(DynamicsContext& context): Cm::Task(context.getContextId()), mDynamicsContext(context) 
	{
	}

	void createForceChangeThresholdStream()
	{
		ThresholdStream& thresholdStream = mDynamicsContext.getThresholdStream();
		//bool haveThresholding = thresholdStream.size()!=0;

		ThresholdTable& thresholdTable = mDynamicsContext.getThresholdTable();
		thresholdTable.build(thresholdStream.begin(), thresholdStream.size());

		//generate current force exceeded threshold stream
		ThresholdStream& curExceededForceThresholdStream = mDynamicsContext.mExceededForceThresholdStream[mDynamicsContext.mCurrentIndex];
		ThresholdStream& preExceededForceThresholdStream = mDynamicsContext.mExceededForceThresholdStream[1 - mDynamicsContext.mCurrentIndex];
		curExceededForceThresholdStream.forceSize_Unsafe(0);

		//fill in the currrent exceeded force threshold stream
		for(PxU32 i=0; i<thresholdTable.mPairsSize; ++i)
		{
			ThresholdTable::Pair& pair = thresholdTable.mPairs[i];
			ThresholdStreamElement& elem = thresholdStream[pair.thresholdStreamIndex];
			if(pair.accumulatedForce > elem.threshold * mDynamicsContext.mDt)
			{
				elem.accumulatedForce = pair.accumulatedForce;
				curExceededForceThresholdStream.pushBack(elem);
			}
		}

		//It's okay to allocate ThresholdStreamMapped here without OOM handling, since this is a
		//a CPU pipeline function, and the buffer is backed by pageable memory.
		ThresholdStreamMapped& forceChangeThresholdStream = mDynamicsContext.getForceChangedThresholdStream();
		forceChangeThresholdStream.forceSize_Unsafe(0);
		PxArray<PxU32>& forceChangeMask = mDynamicsContext.mExceededForceThresholdStreamMask;

		const PxU32 nbPreExceededForce = preExceededForceThresholdStream.size();
		const PxU32 nbCurExceededForce = curExceededForceThresholdStream.size();

		//generate force change thresholdStream
		if(nbPreExceededForce)
		{
			thresholdTable.build(preExceededForceThresholdStream.begin(), preExceededForceThresholdStream.size());

			//set force change mask
			const PxU32 nbTotalExceededForce = nbPreExceededForce + nbCurExceededForce;
			forceChangeMask.reserve(nbTotalExceededForce);
			forceChangeMask.forceSize_Unsafe(nbTotalExceededForce);
			
			//initialize the forceChangeMask
			for (PxU32 i = 0; i < nbTotalExceededForce; ++i)
				forceChangeMask[i] = 1;

			for(PxU32 i=0; i< nbCurExceededForce; ++i)
			{
				ThresholdStreamElement& curElem = curExceededForceThresholdStream[i];
				
				PxU32 pos;
				if(thresholdTable.check(preExceededForceThresholdStream.begin(), preExceededForceThresholdStream.size(), curElem, pos))
				{
					forceChangeMask[pos] = 0;
					forceChangeMask[i + nbPreExceededForce] = 0;
				}
			}

			//create force change threshold stream
			for(PxU32 i=0; i<nbTotalExceededForce; ++i)
			{
				const PxU32 hasForceChange = forceChangeMask[i];
				if(hasForceChange)
				{
					bool lostPair = (i < nbPreExceededForce);
					ThresholdStreamElement& elem = lostPair ?  preExceededForceThresholdStream[i] : curExceededForceThresholdStream[i - nbPreExceededForce];
					ThresholdStreamElement elt;
					elt = elem;
					elt.accumulatedForce = lostPair ? 0.f : elem.accumulatedForce;
					forceChangeThresholdStream.pushBack(elt);
				}
				else
				{
					//persistent pair
					if (i < nbPreExceededForce)
					{
						ThresholdStreamElement& elem = preExceededForceThresholdStream[i];
						ThresholdStreamElement elt;
						elt = elem;
						elt.accumulatedForce = elem.accumulatedForce;
						forceChangeThresholdStream.pushBack(elt);
					}
				}
			}
		}
		else
		{
			forceChangeThresholdStream.reserve(nbCurExceededForce);
			forceChangeThresholdStream.forceSize_Unsafe(nbCurExceededForce);
			PxMemCopy(forceChangeThresholdStream.begin(), curExceededForceThresholdStream.begin(), sizeof(ThresholdStreamElement) * nbCurExceededForce);
		}
	}

	virtual void runInternal() PX_OVERRIDE
	{
		mDynamicsContext.getThresholdStream().forceSize_Unsafe(PxU32(mDynamicsContext.mThresholdStreamOut));
		createForceChangeThresholdStream();
	}

	virtual const char* getName() const PX_OVERRIDE { return "PxsDynamics.createForceChangeThresholdStream"; }
};

#if PGS_SUPPORT_COMPOUND_CONSTRAINTS
struct ArticulationSortPredicate
{
	bool operator()(const PxsIndexedContactManager*& left, const PxsIndexedContactManager*& right) const
	{
		return left->contactManager->getIndex() < right->contactManager->getIndex();
	}
};
#endif

class SolverArticulationUpdateTask : public Cm::Task
{
	ThreadContext&				mIslandThreadContext;
	FeatherstoneArticulation**	mArticulations;
	const PxU32					mNbToProcess;
	Dy::DynamicsContext&		mContext;

public:

	static const PxU32 NbArticulationsPerTask = 32;

	SolverArticulationUpdateTask(ThreadContext& islandThreadContext, FeatherstoneArticulation** articulations, PxU32 nbToProcess, Dy::DynamicsContext& context) :
		Cm::Task(context.getContextId()), mIslandThreadContext(islandThreadContext), mArticulations(articulations), mNbToProcess(nbToProcess), mContext(context)
	{
	}

	virtual const char* getName() const PX_OVERRIDE { return "SolverArticulationUpdateTask"; }

	virtual void runInternal() PX_OVERRIDE
	{
		ThreadContext& threadContext = *mContext.getThreadContext();

		threadContext.mConstraintBlockStream.reset(); //Clear in case there's some left-over memory in this context, for which the block has already been freed 
		PxU32 maxVelIters = 0;
		PxU32 maxPosIters = 0;
		PxU32 maxLinks = 0;

		for (PxU32 i = 0; i < mNbToProcess; i++)
			maxLinks = PxMax(maxLinks, mArticulations[i]->getBodyCount());

		threadContext.mDeltaV.forceSize_Unsafe(0);
		threadContext.mDeltaV.reserve(maxLinks);
		threadContext.mDeltaV.forceSize_Unsafe(maxLinks);

		const PxReal invLengthScale = 1.f/mContext.getLengthScale();

		for(PxU32 i=0;i<mNbToProcess; i++)
		{
			FeatherstoneArticulation* a = mArticulations[i];

			FeatherstoneArticulation::computeUnconstrainedVelocities(a, mContext.mDt,
				mContext.getGravity(), invLengthScale, false, false);

			const PxU16 iterWord = a->getIterationCounts();
			maxVelIters = PxMax<PxU32>(PxU32(iterWord >> 8),	maxVelIters);
			maxPosIters = PxMax<PxU32>(PxU32(iterWord & 0xff),	maxPosIters);
		}

		PxAtomicMax(reinterpret_cast<PxI32*>(&mIslandThreadContext.mMaxSolverPositionIterations), PxI32(maxPosIters));
		PxAtomicMax(reinterpret_cast<PxI32*>(&mIslandThreadContext.mMaxSolverVelocityIterations), PxI32(maxVelIters));
		PxAtomicMax(reinterpret_cast<PxI32*>(&mIslandThreadContext.mMaxArticulationLinks), PxI32(maxLinks));

		mContext.putThreadContext(&threadContext);
	}

private:
	PX_NOCOPY(SolverArticulationUpdateTask)
};

class PxsSolverStartTask : public Cm::Task
{
	PxsSolverStartTask& operator=(const PxsSolverStartTask&);
public:

	PxsSolverStartTask(DynamicsContext& context,
		IslandContext& islandContext,
		const SolverIslandObjects& objects,
		PxU32 solverBodyOffset,
		PxU32 kinematicCount,
		IG::SimpleIslandManager& islandManager,
		PxU32* bodyRemapTable,
		PxsMaterialManager* materialManager,
		PxsContactManagerOutputIterator& iterator,
		bool enhancedDeterminism
		) :
		Cm::Task				(context.getContextId()),
		mContext				(context), 
		mIslandContext			(islandContext),
		mObjects				(objects),
		mSolverBodyOffset		(solverBodyOffset),
		mKinematicCount			(kinematicCount),
		mIslandManager			(islandManager),
		mBodyRemapTable			(bodyRemapTable),
		mMaterialManager		(materialManager),
		mOutputs				(iterator),
		mEnhancedDeterminism	(enhancedDeterminism)
	{}

	void setupThreadContext()
	{
		PX_PROFILE_ZONE("setupThreadContext", mContextID);

		ThreadContext* threadContext = mContext.getThreadContext();

		mIslandContext.mThreadContext = threadContext;

		threadContext->mMaxSolverPositionIterations	= 0;
		threadContext->mMaxSolverVelocityIterations	= 0;
		threadContext->mAxisConstraintCount			= 0;
		threadContext->mNumDifferentBodyConstraints	= 0;
		threadContext->mNumStaticConstraints		= 0;
		threadContext->numContactConstraintBatches	= 0;
		threadContext->contactDescArraySize			= 0;
		threadContext->mMaxArticulationLinks		= 0;

		threadContext->contactConstraintDescArray		= mObjects.constraintDescs;
		threadContext->orderedContactConstraints		= mObjects.orderedConstraintDescs;
		threadContext->mContactDescPtr					= mObjects.constraintDescs;
		threadContext->tempConstraintDescArray			= mObjects.tempConstraintDescs;
		threadContext->contactConstraintBatchHeaders	= mObjects.constraintBatchHeaders;
		threadContext->motionVelocityArray				= mObjects.motionVelocities;
		threadContext->mBodyCoreArray					= mObjects.bodyCoreArray;
		threadContext->mRigidBodyArray					= mObjects.bodies;
		threadContext->mArticulationArray				= mObjects.articulations;
		threadContext->bodyRemapTable					= mObjects.bodyRemapTable;
		threadContext->mNodeIndexArray					= mObjects.nodeIndexArray;

		threadContext->resizeArrays(mIslandContext.mCounts.articulations);
	}

	// PT: TODO: rename to "prepareBodiesAndConstraints" like in TGS?
	void startTasks()
	{
		PX_PROFILE_ZONE("Dynamics.solveGroup", mContextID);

		const ThreadContext& threadContext = *mIslandContext.mThreadContext;

		const PxU32 nbIslands = mObjects.numIslands;
		const IG::IslandId* const islandIds = mObjects.islandIds;
		PxU32* PX_RESTRICT bodyRemapTable = threadContext.bodyRemapTable;

		mContext.iterateIslandsNodes(mIslandContext.mCounts, nbIslands, islandIds, threadContext.mBodyCoreArray, threadContext.mRigidBodyArray, threadContext.mArticulationArray, bodyRemapTable, threadContext.mNodeIndexArray);

		mIslandContext.mCounts.contactManagers = mContext.iterateIslandsContactEdges(mIslandContext.mCounts, nbIslands, islandIds, mObjects.contactManagers, bodyRemapTable);
	}

	void integrate()
	{
		ThreadContext& threadContext = *mIslandContext.mThreadContext;
		PxSolverBody* solverBodies = mContext.mSolverBodyPool.begin() + mSolverBodyOffset;
		PxSolverBodyData* solverBodyData = mContext.mSolverBodyDataPool.begin() + mSolverBodyOffset;

		{
			PX_PROFILE_ZONE("Dynamics.updateVelocities", mContextID);

			mContext.preIntegrationParallel(
				mContext.mDt,
				threadContext.mBodyCoreArray,
				mObjects.bodies,
				threadContext.mNodeIndexArray,
				mIslandContext.mCounts.bodies,
				solverBodies,
				solverBodyData,
				threadContext.motionVelocityArray,
				threadContext.mMaxSolverPositionIterations,
				threadContext.mMaxSolverVelocityIterations,
				*mCont
				);
		}
	}

	void articulationTask()
	{
		PX_PROFILE_ZONE("SetupArticulationTasks", mContextID);

		ThreadContext& threadContext = *mIslandContext.mThreadContext;

		for(PxU32 i=0;i<mIslandContext.mCounts.articulations; i+= SolverArticulationUpdateTask::NbArticulationsPerTask)
		{
			SolverArticulationUpdateTask* task = PX_PLACEMENT_NEW(mContext.getTaskPool().allocate(sizeof(SolverArticulationUpdateTask)), SolverArticulationUpdateTask)(threadContext, 
				&mObjects.articulations[i], PxMin(SolverArticulationUpdateTask::NbArticulationsPerTask, mIslandContext.mCounts.articulations - i), mContext);

			task->setContinuation(mCont);
			task->removeReference();
		}
	}

	void setupDescTask()
	{
		PX_PROFILE_ZONE("SetupDescs", mContextID);

		ThreadContext& threadContext = *mIslandContext.mThreadContext;
		PxSolverConstraintDesc* contactDescPtr = threadContext.mContactDescPtr;

		//PxU32 constraintCount = mCounts.constraints + mCounts.contactManagers;

		PxU32 nbIslands = mObjects.numIslands;
		const IG::IslandId* const islandIds = mObjects.islandIds;

		const IG::IslandSim& islandSim = mIslandManager.getAccurateIslandSim();

		{
			// PT: TODO: refactor with TGS
			PX_PROFILE_ZONE("IterateIslandsConstraintEdges", mContextID);

			for(PxU32 i = 0; i < nbIslands; ++i)
			{
				const IG::Island& island = islandSim.getIsland(islandIds[i]);

				START_ENUMERATING_ISLAND_EDGES(IG::Edge::eCONSTRAINT)
				{
					GET_CURRENT_ISLAND_EDGE

					PxSolverConstraintDesc& desc = *contactDescPtr;
				
					Dy::Constraint* constraint = mIslandManager.getConstraint(edgeId);
					mContext.setDescFromIndices_Constraints(desc, islandSim, edgeId, mBodyRemapTable, mSolverBodyOffset);
					desc.constraint = reinterpret_cast<PxU8*>(constraint);
					desc.constraintType = DY_SC_TYPE_RB_1D;
					contactDescPtr++;

					GET_NEXT_ISLAND_EDGE
				}
			}
		}

		{
			PX_PROFILE_ZONE("PxSort", mContextID);
			PxSort(threadContext.mContactDescPtr, PxU32(contactDescPtr - threadContext.mContactDescPtr), ConstraintLess());
		}

		const PxU32 nbCms = mIslandContext.mCounts.contactManagers;

#if PGS_SUPPORT_COMPOUND_CONSTRAINTS
		threadContext.orderedContactList.forceSize_Unsafe(0);
		threadContext.orderedContactList.reserve(nbCms);
		threadContext.orderedContactList.forceSize_Unsafe(nbCms);
		threadContext.tempContactList.forceSize_Unsafe(0);
		threadContext.tempContactList.reserve(nbCms);
		threadContext.tempContactList.forceSize_Unsafe(nbCms);

		const PxsIndexedContactManager** constraints = threadContext.orderedContactList.begin();

		//OK, we sort the orderedContactList 

		threadContext.compoundConstraints.forceSize_Unsafe(0);
		if(nbCms)
		{
			{
				threadContext.sortIndexArray.forceSize_Unsafe(0);

				const PxI32 offsetMap[] = {PxI32(mContext.mKinematicCount), 0};

				const PxU32 totalBodies = mContext.mKinematicCount + mIslandContext.mCounts.bodies+1;

				threadContext.sortIndexArray.reserve(totalBodies);
				threadContext.sortIndexArray.forceSize_Unsafe(totalBodies);
				PxMemZero(threadContext.sortIndexArray.begin(), totalBodies * 4);

				//Iterate over the array based on solverBodyDatapool, creating a list of sorted constraints (in order of body pair)
				//We only do this with contacts. It's important that this is done this way because we don't want to break our rules that all joints
				//appear before all contacts in the constraint list otherwise we will lose all guarantees about sorting joints.

				for(PxU32 a = 0; a < nbCms; ++a)
				{
					PX_ASSERT(mObjects.contactManagers[a].indexType0 != PxsIndexedInteraction::eWORLD);
					//Index first body...
					PxU8 indexType = mObjects.contactManagers[a].indexType0;
					if(indexType != PxsIndexedInteraction::eARTICULATION && mObjects.contactManagers[a].indexType1 != PxsIndexedInteraction::eARTICULATION)
					{
						PX_ASSERT((indexType == PxsIndexedInteraction::eBODY) || (indexType == PxsIndexedInteraction::eKINEMATIC));

						PxI32 index = PxI32(mObjects.contactManagers[a].solverBody0 + offsetMap[indexType]);
						PX_ASSERT(index >= 0);
						threadContext.sortIndexArray[PxU32(index)]++;
					}
				}

				PxU32 accumulatedCount = 0;

				for(PxU32 a = threadContext.sortIndexArray.size(); a > 0; --a)
				{
					PxU32 ind = a - 1;
					PxU32 val = threadContext.sortIndexArray[ind];
					threadContext.sortIndexArray[ind] = accumulatedCount;
					accumulatedCount += val;
				}

				//OK, now copy across data to orderedConstraintDescs, pushing articulations to the end...
				for(PxU32 a = 0; a < nbCms; ++a)
				{
					//Index first body...
					PxU8 indexType = mObjects.contactManagers[a].indexType0;
					if(indexType != PxsIndexedInteraction::eARTICULATION && mObjects.contactManagers[a].indexType1 != PxsIndexedInteraction::eARTICULATION)
					{
						PX_ASSERT((indexType == PxsIndexedInteraction::eBODY) || (indexType == PxsIndexedInteraction::eKINEMATIC));

						PxI32 index = PxI32(mObjects.contactManagers[a].solverBody0 + offsetMap[indexType]);
						PX_ASSERT(index >= 0);
						threadContext.tempContactList[threadContext.sortIndexArray[PxU32(index)]++] = &mObjects.contactManagers[a];
					}
					else
					{
						threadContext.tempContactList[accumulatedCount++] = &mObjects.contactManagers[a];
					}
				}

				//Now do the same again with bodyB, being careful not to overwrite the joints
				PxMemZero(threadContext.sortIndexArray.begin(), totalBodies * 4);

				for(PxU32 a = 0; a < nbCms; ++a)
				{
					//Index first body...
					PxU8 indexType = threadContext.tempContactList[a]->indexType1;
					if(indexType != PxsIndexedInteraction::eARTICULATION && mObjects.contactManagers[a].indexType0 != PxsIndexedInteraction::eARTICULATION)
					{
						PX_ASSERT((indexType == PxsIndexedInteraction::eBODY) || (indexType == PxsIndexedInteraction::eKINEMATIC) || (indexType == PxsIndexedInteraction::eWORLD));

						PxI32 index = (indexType == PxsIndexedInteraction::eWORLD) ? 0 : PxI32(threadContext.tempContactList[a]->solverBody1 + offsetMap[indexType]);
						PX_ASSERT(index >= 0);
						threadContext.sortIndexArray[PxU32(index)]++;
					}
				}

				accumulatedCount = 0;
				for(PxU32 a = threadContext.sortIndexArray.size(); a > 0; --a)
				{
					PxU32 ind = a - 1;
					PxU32 val = threadContext.sortIndexArray[ind];
					threadContext.sortIndexArray[ind] = accumulatedCount;
					accumulatedCount += val;
				}

				PxU32 articulationStartIndex = accumulatedCount;

				//OK, now copy across data to orderedConstraintDescs, pushing articulations to the end...
				for(PxU32 a = 0; a < nbCms; ++a)
				{
					//Index first body...
					PxU8 indexType = threadContext.tempContactList[a]->indexType1;
					if(indexType != PxsIndexedInteraction::eARTICULATION && mObjects.contactManagers[a].indexType0 != PxsIndexedInteraction::eARTICULATION)
					{
						PX_ASSERT((indexType == PxsIndexedInteraction::eBODY) || (indexType == PxsIndexedInteraction::eKINEMATIC) || (indexType == PxsIndexedInteraction::eWORLD));

						PxI32 index = (indexType == PxsIndexedInteraction::eWORLD) ? 0 : PxI32(threadContext.tempContactList[a]->solverBody1 + offsetMap[indexType]);
						PX_ASSERT(index >= 0);
						constraints[threadContext.sortIndexArray[PxU32(index)]++] = threadContext.tempContactList[a];
					}
					else
					{
						constraints[accumulatedCount++] = threadContext.tempContactList[a];
					}
				}

				{
					PX_PROFILE_ZONE("PxSort", mContextID);
					PxSort(constraints + articulationStartIndex, accumulatedCount - articulationStartIndex, ArticulationSortPredicate());
				}
			}

			{
				PX_PROFILE_ZONE("Create compound constraints", mContextID);

				threadContext.compoundConstraints.reserve(1024);
				threadContext.compoundConstraints.forceSize_Unsafe(0);
				//threadContext.compoundConstraints.forceSize_Unsafe(mCounts.contactManagers);

				PxSolverConstraintDesc* startDesc = contactDescPtr;
				mContext.setDescFromIndices_Contacts(*startDesc, islandSim, *constraints[0], mSolverBodyOffset);
				startDesc->constraint = reinterpret_cast<PxU8*>(constraints[0]->contactManager);
				startDesc->constraintType = DY_SC_TYPE_RB_CONTACT;

				PxsContactManagerOutput* startManagerOutput = &mOutputs.getContactManagerOutput(constraints[0]->contactManager->getWorkUnit().mNpIndex);
				PxU32 contactCount = startManagerOutput->nbContacts;
				PxU32 startIndex = 0;
				PxU32 numHeaders = 0;

				const bool gDisableConstraintWelding = false;

				for(PxU32 a = 1; a < nbCms; ++a)
				{
					PxSolverConstraintDesc& desc = *(contactDescPtr+1);
					mContext.setDescFromIndices_Contacts(desc, islandSim, *constraints[a], mSolverBodyOffset);

					PxsContactManager* manager = constraints[a]->contactManager;
					PxsContactManagerOutput& output = mOutputs.getContactManagerOutput(manager->getWorkUnit().mNpIndex);

					desc.constraint = reinterpret_cast<PxU8*>(constraints[a]->contactManager);
					desc.constraintType = DY_SC_TYPE_RB_CONTACT;

					if (contactCount == 0)
					{
						//This is the first object in the pair
						*startDesc = *(contactDescPtr + 1);
						startIndex = a;
						startManagerOutput = &output;
					}
				
					if(startDesc->bodyA != desc.bodyA || startDesc->bodyB != desc.bodyB 
						|| startDesc->linkIndexA != PxSolverConstraintDesc::RIGID_BODY || startDesc->linkIndexB != PxSolverConstraintDesc::RIGID_BODY
						|| contactCount + output.nbContacts > PxContactBuffer::MAX_CONTACTS
						|| manager->isChangeable()
						|| gDisableConstraintWelding
						) //If this is the first thing and no contacts...then we skip
					{
						const PxU32 stride = a - startIndex;
						if(contactCount > 0)
						{
							if(stride > 1)
							{
								++numHeaders;
								CompoundContactManager& header = threadContext.compoundConstraints.insert();
								header.mStartIndex = startIndex;
								header.mStride = PxTo16(stride);
								header.mReducedContactCount = PxTo16(contactCount);
								PxsContactManager* manager1 = constraints[startIndex]->contactManager;
								PxcNpWorkUnit& unit = manager1->getWorkUnit();

								PX_ASSERT(startManagerOutput == &mOutputs.getContactManagerOutput(unit.mNpIndex));

								header.unit = &unit;
								header.cmOutput = startManagerOutput;
								header.originalContactPatches = startManagerOutput->contactPatches;
								header.originalContactPoints = startManagerOutput->contactPoints;
								header.originalContactCount = startManagerOutput->nbContacts;
								header.originalPatchCount = startManagerOutput->nbPatches;
								header.originalForceBuffer = reinterpret_cast<PxReal*>(startManagerOutput->contactForces);
								header.originalStatusFlags = startManagerOutput->statusFlag;
								header.originalFrictionPatches = startManagerOutput->frictionPatches;
							}
							startDesc = ++contactDescPtr;
						}
						else
						{
							//Copy back next contactDescPtr
							*startDesc = *(contactDescPtr+1);
						}
						contactCount = 0;
						startIndex = a;
						startManagerOutput = &output;
					}
					contactCount += output.nbContacts;
				}

				const PxU32 stride = nbCms - startIndex;
				if(contactCount > 0)
				{
					if(stride > 1)
					{
						++numHeaders;
						CompoundContactManager& header = threadContext.compoundConstraints.insert();
						header.mStartIndex = startIndex;
						header.mStride = PxTo16(stride);
						header.mReducedContactCount = PxTo16(contactCount);
						PxsContactManager* manager = constraints[startIndex]->contactManager;
						PxcNpWorkUnit& unit = manager->getWorkUnit();
						header.unit = &unit;
						header.cmOutput = startManagerOutput;
						header.originalContactPatches = startManagerOutput->contactPatches;
						header.originalContactPoints = startManagerOutput->contactPoints;
						header.originalContactCount = startManagerOutput->nbContacts;
						header.originalPatchCount = startManagerOutput->nbPatches;
						header.originalForceBuffer = reinterpret_cast<PxReal*>(startManagerOutput->contactForces);
						header.originalStatusFlags = startManagerOutput->statusFlag;
						header.originalFrictionPatches = startManagerOutput->frictionPatches;
					}
					contactDescPtr++;
				}

				if(numHeaders)
				{
					const PxU32 unrollSize = 8;
					for(PxU32 a = 0; a < numHeaders; a+= unrollSize)
					{
						PxsSolverConstraintPostProcessTask* postProcessTask = PX_PLACEMENT_NEW( mContext.getTaskPool().allocate(sizeof(PxsSolverConstraintPostProcessTask)), 
							PxsSolverConstraintPostProcessTask)(mContext, threadContext, mObjects, mSolverBodyOffset, a, PxMin(unrollSize, numHeaders - a), mMaterialManager,
							mOutputs);
						postProcessTask->setContinuation(mCont);
						postProcessTask->removeReference();
					}
				}
			}
		}
		threadContext.contactDescArraySize = PxU32(contactDescPtr - threadContext.contactConstraintDescArray);
		threadContext.mContactDescPtr = contactDescPtr;
#else
		if (nbCms)
		{
			// PT: TODO: refactor with TGS
			PX_PROFILE_ZONE("setDescFromIndices_contacts", mContextID);

			for (PxU32 a = 0; a < nbCms; ++a)
			{
				PxSolverConstraintDesc& desc = *contactDescPtr;
				mContext.setDescFromIndices_Contacts(desc, islandSim, mObjects.contactManagers[a], mSolverBodyOffset);
				desc.constraint = reinterpret_cast<PxU8*>(mObjects.contactManagers[a].contactManager);
				desc.constraintType = DY_SC_TYPE_RB_CONTACT;
				contactDescPtr++;
			}
		}
		threadContext.contactDescArraySize = PxU32(contactDescPtr - mObjects.constraintDescs);
#endif
	}

	virtual void runInternal() PX_OVERRIDE
	{
		setupThreadContext();
		startTasks();
		integrate();
		setupDescTask();
		articulationTask();
	}

	virtual const char* getName() const PX_OVERRIDE
	{
		return "PxsDynamics.solverStart";
	}

private:
	DynamicsContext&			mContext;
	IslandContext&				mIslandContext;
	const SolverIslandObjects	mObjects;
	const PxU32					mSolverBodyOffset;
	const PxU32					mKinematicCount;
	IG::SimpleIslandManager&	mIslandManager;
	PxU32*						mBodyRemapTable;
	PxsMaterialManager*			mMaterialManager;
	PxsContactManagerOutputIterator& mOutputs;
	const bool					mEnhancedDeterminism;
};

static void partitionConstraints(IslandContext& islandContext, PxSolverBody* solverBodies, PxU64 contextID)
{
	PX_PROFILE_ZONE("PartitionConstraints", contextID);
	PX_UNUSED(contextID);

	ThreadContext& threadContext = *islandContext.mThreadContext;

	//Compact articulation pairs...
	const PxSolverConstraintDesc* descBegin = threadContext.contactConstraintDescArray;
	const PxU32 descCount = threadContext.contactDescArraySize;

	threadContext.mNumDifferentBodyConstraints = descCount;
	threadContext.mNumDifferentBodyConstraints = 0;
	threadContext.mNumStaticConstraints = 0;

	if(descCount > 0)
	{
		const PxU32 numArticulations = threadContext.mNbArticulations;
		FeatherstoneArticulation** eaArticulations = threadContext.mArticulationArray;

		// PT: maxPartitions = 64 in TGS
		const ConstraintPartitionIn in(	reinterpret_cast<PxU8*>(solverBodies), islandContext.mCounts.bodies, sizeof(PxSolverBody),
										eaArticulations, numArticulations, descBegin, descCount, PX_MAX_U32, false);
				
		ConstraintPartitionOut out(threadContext.orderedContactConstraints, threadContext.tempConstraintDescArray, &threadContext.mConstraintsPerPartition);

		threadContext.mMaxPartitions = partitionContactConstraints(out, in);
		threadContext.mNumDifferentBodyConstraints = out.mNumDifferentBodyConstraints;
		threadContext.mNumStaticConstraints = out.mNumStaticConstraints;
	}
	else
	{
		PxMemZero(threadContext.mConstraintsPerPartition.begin(), sizeof(PxU32) * threadContext.mConstraintsPerPartition.capacity());
	}

	PX_ASSERT((threadContext.mNumDifferentBodyConstraints + threadContext.mNumStaticConstraints) == descCount);
}

class PxsSolverConstraintPartitionTask : public Cm::Task
{
	PxsSolverConstraintPartitionTask& operator=(const PxsSolverConstraintPartitionTask&);
public:

	PxsSolverConstraintPartitionTask(DynamicsContext& context, IslandContext& islandContext, PxU32 solverBodyOffset) :
		Cm::Task			(context.getContextId()),
		mContext			(context), 
		mIslandContext		(islandContext),
		mSolverBodyOffset	(solverBodyOffset)
	{}

	virtual void runInternal() PX_OVERRIDE
	{
		partitionConstraints(mIslandContext, mContext.mSolverBodyPool.begin() + mSolverBodyOffset, mContext.getContextId());
	}

	virtual const char* getName() const PX_OVERRIDE { return "PxsDynamics.solverConstraintPartition"; }

	DynamicsContext&	mContext;
	IslandContext&		mIslandContext;
	const PxU32			mSolverBodyOffset;
};

static void integrate(	const IG::IslandSim& islandSim, PxSolverBodyData* PX_RESTRICT solverBodyData, PxsRigidBody** PX_RESTRICT rigidBodies,
						Cm::SpatialVector* PX_RESTRICT motionVelocityArray, PxSolverBody* PX_RESTRICT solverBodies, 
						PxU32 count, PxF32 dt, bool enableStabilization, bool isSleepingDisabled)
{
	for(PxU32 i=0; i<count; i++)
	{
		PxSolverBodyData& data = solverBodyData[i];
		PxsRigidBody& rBody = *rigidBodies[i];
		PxsBodyCore& core = rBody.getCore();

		// PT: note that only PGS uses "integrateCore", TGS actually uses "integrateCoreStep".
		integrateCore(motionVelocityArray[i].linear, motionVelocityArray[i].angular, solverBodies[i], data, dt, core.lockFlags);

		rBody.mLastTransform = core.body2World;
		core.body2World = data.body2World;
		core.linearVelocity = data.linearVelocity;
		core.angularVelocity = data.angularVelocity;

		if(!isSleepingDisabled)
		{
			const PxU32 hasStaticTouch = islandSim.getIslandStaticTouchCount(PxNodeIndex(data.nodeIndex));
			sleepCheck(rigidBodies[i], dt, enableStabilization, motionVelocityArray[i], hasStaticTouch);
		}
	}
}

class PxsSolverSetupSolveTask : public Cm::Task
{
	PxsSolverSetupSolveTask& operator=(const PxsSolverSetupSolveTask&);
public:

	PxsSolverSetupSolveTask(DynamicsContext& context, IslandContext& islandContext, const SolverIslandObjects& objects, PxU32 solverBodyOffset, IG::IslandSim& islandSim) :
		Cm::Task			(context.getContextId()),
		mContext			(context), 
		mIslandContext		(islandContext),
		mObjects			(objects),
		mSolverBodyOffset	(solverBodyOffset),
		mIslandSim			(islandSim)
	{}

	virtual void runInternal() PX_OVERRIDE
	{
		ThreadContext& threadContext = *mIslandContext.mThreadContext;

		//On PS3, self-constraints will be bumped to the end of the constraint list
		//and processed separately. On PC/360, they will be mixed in the array and
		//classed as "different body" constraints regardless of the fact that they're self-constraints.
		//PxU32 numBatches = threadContext.numDifferentBodyBatchHeaders;
		// TODO: maybe replace with non-null joints from end of the array

		// PT: TODO: share with TGS

		PxU32 j = 0, i = 0;
		PxU32 numBatches = 0;
		PxU32 currIndex = 0;

		PxSolverConstraintDesc* PX_RESTRICT contactDescBegin = threadContext.orderedContactConstraints;
		PxConstraintBatchHeader* PX_RESTRICT headers = threadContext.contactConstraintBatchHeaders;

		for(PxU32 a = 0; a < threadContext.mConstraintsPerPartition.size(); ++a)
		{
			const PxU32 endIndex = currIndex + threadContext.mConstraintsPerPartition[a];

			PxU32 numBatchesInPartition = 0;
			for(PxU32 b = currIndex; b < endIndex; ++b)
			{
				PxConstraintBatchHeader& _header = headers[b];
				PxU16 stride = _header.stride, newStride = _header.stride;
				PxU32 startIndex = j;
				for(PxU16 c = 0; c < stride; ++c)
				{
					if(getConstraintLength(contactDescBegin[i]) == 0)
					{
						newStride--;
						i++;
					}
					else
					{
						if(i != j)
							contactDescBegin[j] = contactDescBegin[i];
						i++;
						j++;
					}
				}

				if(newStride != 0)
				{
					headers[numBatches].startIndex = startIndex;
					headers[numBatches].stride = newStride;
					PxU8 type = *contactDescBegin[startIndex].constraint;
					if(type == DY_SC_TYPE_STATIC_CONTACT)
					{
						//Check if any block of constraints is classified as type static (single) contact constraint.
						//If they are, iterate over all constraints grouped with it and switch to "dynamic" contact constraint
						//type if there's a dynamic contact constraint in the group.
						for(PxU32 c = 1; c < newStride; ++c)
						{
							if(*contactDescBegin[startIndex + c].constraint == DY_SC_TYPE_RB_CONTACT)
							{
								type = DY_SC_TYPE_RB_CONTACT;
							}
						}
					}

					headers[numBatches].constraintType = type;
					numBatches++;
					numBatchesInPartition++;
				}
			}

			currIndex += threadContext.mConstraintsPerPartition[a];
			threadContext.mConstraintsPerPartition[a] = numBatchesInPartition;
		}

		threadContext.mNumDifferentBodyConstraints = j;		

		threadContext.numContactConstraintBatches = numBatches;
		threadContext.mOrderedContactDescCount = j;

		{
			{
				PX_PROFILE_ZONE("Dynamics.solver", mContextID);

				PxSolverConstraintDesc* contactDescs = threadContext.orderedContactConstraints;

				PxSolverBody* solverBodies = mContext.mSolverBodyPool.begin() + mSolverBodyOffset;
				PxSolverBodyData* solverBodyDatas = mContext.mSolverBodyDataPool.begin();

				PxI32* thresholdPairsOut = &mContext.mThresholdStreamOut;

				SolverIslandParams& params = *reinterpret_cast<SolverIslandParams*>(mContext.getTaskPool().allocate(sizeof(SolverIslandParams)));
				params.positionIterations = threadContext.mMaxSolverPositionIterations;
				params.velocityIterations = threadContext.mMaxSolverVelocityIterations;
				params.bodyListStart = solverBodies;
				params.bodyDataList = solverBodyDatas;
				params.solverBodyOffset = mSolverBodyOffset;
				params.bodyListSize = mIslandContext.mCounts.bodies;
				params.articulationListStart = threadContext.mArticulationArray;
				params.articulationListSize = threadContext.mNbArticulations;
				params.constraintList = contactDescs;
				params.constraintIndex = 0;
				params.constraintIndexCompleted = 0;
				params.bodyListIndex = 0;
				params.bodyListIndexCompleted = 0;
				params.articSolveIndex = 0;
				params.articSolveIndexCompleted = 0;
				params.bodyIntegrationListIndex = 0;
				params.thresholdStream = mContext.getThresholdStream().begin();
				params.thresholdStreamLength = mContext.getThresholdStream().size();
				params.outThresholdPairs = thresholdPairsOut;
				params.motionVelocityArray = threadContext.motionVelocityArray;
				params.numObjectsIntegrated = 0;
				params.constraintBatchHeaders = threadContext.contactConstraintBatchHeaders;
				params.numConstraintHeaders = threadContext.numContactConstraintBatches;
				params.headersPerPartition = threadContext.mConstraintsPerPartition.begin();
				params.nbPartitions = threadContext.mConstraintsPerPartition.size();
				params.rigidBodies = const_cast<PxsRigidBody**>(mObjects.bodies);
				params.mMaxArticulationLinks = threadContext.mMaxArticulationLinks;
				params.dt = mContext.mDt;
				params.invDt = mContext.mInvDt;

				const PxU32 unrollSize = 8;
				const PxU32 denom = PxMax(1u, (threadContext.mMaxPartitions*unrollSize));
				const PxU32 maxTasks = getTaskManager()->getCpuDispatcher()->getWorkerCount();
				// PT: small improvement: if there's no contacts, use the number of bodies instead.
				// That way the integration work still benefits from multiple tasks.
				const PxU32 numWorkItems = threadContext.numContactConstraintBatches ? threadContext.numContactConstraintBatches : mIslandContext.mCounts.bodies;
				const PxU32 idealThreads = (numWorkItems+denom-1)/denom;
				const PxU32 numTasks = PxMax(1u, PxMin(idealThreads, maxTasks));
				
				if(numTasks > 1)
				{
					const PxU32 idealBatchSize = PxMax(unrollSize, idealThreads*unrollSize/(numTasks*2));

					params.batchSize = idealBatchSize; //assigning ideal batch size for the solver to grab work at. Only needed by the multi-threaded island solver.

					for(PxU32 a = 1; a < numTasks; ++a)
					{
						void* tsk = mContext.getTaskPool().allocate(sizeof(PxsParallelSolverTask));
						PxsParallelSolverTask* pTask = PX_PLACEMENT_NEW(tsk, PxsParallelSolverTask)(
							params, mContext, mIslandSim);

						//Force to complete before merge task!
						pTask->setContinuation(mCont);						
						pTask->removeReference();
					}

					//Avoid kicking off one parallel task when we can do the work inline in this function
					{
						PX_PROFILE_ZONE("Dynamics.parallelSolve", mContextID);

						solveParallel(mContext, params, mIslandSim);
					}
					const PxI32 numBodiesPlusArtics = PxI32( mIslandContext.mCounts.bodies + mIslandContext.mCounts.articulations );

					PxI32* numObjectsIntegrated = &params.numObjectsIntegrated;

					WAIT_FOR_PROGRESS_NO_TIMER(numObjectsIntegrated, numBodiesPlusArtics);
				}
				else
				{
					threadContext.mDeltaV.forceSize_Unsafe(0);
					threadContext.mDeltaV.reserve(threadContext.mMaxArticulationLinks);
					threadContext.mDeltaV.forceSize_Unsafe(threadContext.mMaxArticulationLinks);

					params.deltaV = threadContext.mDeltaV.begin();

					//Only one task - a small island so do a sequential solve (avoid the atomic overheads)
					solveV_Blocks(params, mContext.getBiasCoefficients().articulation, mContext.solveFrictionEveryIteration(), mContext.mSolveArticulationContactLast);

					PxSolverBodyData* solverBodyData2 = solverBodyDatas + mSolverBodyOffset + 1;
					integrate(mIslandSim, solverBodyData2, mObjects.bodies, threadContext.motionVelocityArray, solverBodies, mIslandContext.mCounts.bodies, mContext.mDt, mContext.mEnableStabilization, mContext.isSleepingDisabled());

					for(PxU32 cnt=0; cnt<mIslandContext.mCounts.articulations; cnt++)
					{
						//PX_PROFILE_ZONE("Articulations.integrate", mContextID);
						FeatherstoneArticulation::updateBodies(threadContext.mArticulationArray[cnt], threadContext.mDeltaV.begin(), mContext.getDt());
					}
				}
			}
		}
	}

	virtual const char* getName() const PX_OVERRIDE { return "PxsDynamics.solverSetupSolve"; }

	DynamicsContext&			mContext;
	IslandContext&				mIslandContext;
	const SolverIslandObjects	mObjects;
	const PxU32					mSolverBodyOffset;
	IG::IslandSim&				mIslandSim;
};

class PxsSolverEndTask : public Cm::Task
{
	PxsSolverEndTask& operator=(const PxsSolverEndTask&);
public:

	PxsSolverEndTask(DynamicsContext& context, IslandContext& islandContext, const SolverIslandObjects& objects, PxU32 solverBodyOffset, PxsContactManagerOutputIterator& cmOutputs) :
		Cm::Task			(context.getContextId()),
		mContext			(context), 
		mIslandContext		(islandContext),
		mObjects			(objects),
		mSolverBodyOffset	(solverBodyOffset),
		mOutputs			(cmOutputs)
	{}

	virtual void runInternal() PX_OVERRIDE
	{
		PX_PROFILE_ZONE("Dynamics.endTask", getContextId());
		ThreadContext& threadContext = *mIslandContext.mThreadContext;
#if PX_ENABLE_SIM_STATS
		threadContext.getSimStats().numAxisSolverConstraints += threadContext.mAxisConstraintCount;
#else
		PX_CATCH_UNDEFINED_ENABLE_SIM_STATS
#endif

#if PGS_SUPPORT_COMPOUND_CONSTRAINTS
		//Patch up the contact managers (TODO - fix up force writeback)
		PxU32 numCompoundConstraints = threadContext.compoundConstraints.size();
		for(PxU32 i = 0; i < numCompoundConstraints; ++i)
		{
			CompoundContactManager& manager = threadContext.compoundConstraints[i];
			PxsContactManagerOutput* cmOutput = manager.cmOutput;

			PxReal* contactForces = reinterpret_cast<PxReal*>(cmOutput->contactForces);
			PxU32 contactCount = cmOutput->nbContacts;

			cmOutput->contactPatches = manager.originalContactPatches;
			cmOutput->contactPoints = manager.originalContactPoints;
			cmOutput->nbContacts = manager.originalContactCount;
			cmOutput->nbPatches = manager.originalPatchCount;
			cmOutput->statusFlag = manager.originalStatusFlags;
			cmOutput->contactForces = manager.originalForceBuffer;
			cmOutput->frictionPatches = manager.originalFrictionPatches;
			
			for(PxU32 a = 1; a < manager.mStride; ++a)
			{
				PxsContactManager* pManager = threadContext.orderedContactList[manager.mStartIndex + a]->contactManager;
				pManager->getWorkUnit().mFrictionDataPtr = manager.unit->mFrictionDataPtr;
				pManager->getWorkUnit().mFrictionPatchCount = manager.unit->mFrictionPatchCount;
				//pManager->getWorkUnit().prevFrictionPatchCount = manager.unit->prevFrictionPatchCount;
			}

			//This is a stride-based contact force writer. The assumption is that we may have skipped certain unimportant contacts reported by the 
			//discrete narrow phase
			if(contactForces)
			{
				PxU32 currentContactIndex = 0;
				PxU32 currentManagerIndex = manager.mStartIndex;
				PxU32 currentManagerContactIndex = 0;

				for(PxU32 a = 0; a < contactCount; ++a)
				{
					PxU32 index = manager.forceBufferList[a];
					PxsContactManager* pManager = threadContext.orderedContactList[currentManagerIndex]->contactManager;
					const PxsContactManagerOutput* output = &mOutputs.getContactManagerOutput(pManager->getWorkUnit().mNpIndex);
					while(currentContactIndex < index || output->nbContacts == 0)
					{
						//Step forwards...first in this manager...
						
						PxU32 numToStep = PxMin(index - currentContactIndex, PxU32(output->nbContacts) - currentManagerContactIndex);
						currentContactIndex += numToStep;
						currentManagerContactIndex += numToStep;
						if(currentManagerContactIndex == output->nbContacts)
						{
							currentManagerIndex++;
							currentManagerContactIndex = 0;
							pManager = threadContext.orderedContactList[currentManagerIndex]->contactManager;
							output = &mOutputs.getContactManagerOutput(pManager->getWorkUnit().mNpIndex);
						}
					}
					if(output->nbContacts > 0 && output->contactForces)
						output->contactForces[currentManagerContactIndex] = contactForces[a];
				}
			}
		}

		threadContext.compoundConstraints.forceSize_Unsafe(0);
#endif
		threadContext.mConstraintBlockManager.reset();

		mContext.putThreadContext(&threadContext);
	}

	virtual const char* getName() const PX_OVERRIDE
	{
		return "PxsDynamics.solverEnd";
	}

	DynamicsContext&					mContext;	
	IslandContext&						mIslandContext;
	const SolverIslandObjects			mObjects;
	const PxU32							mSolverBodyOffset;
	PxsContactManagerOutputIterator&	mOutputs;
};

class PxsSolverCreateFinalizeConstraintsTask : public Cm::Task
{
	PxsSolverCreateFinalizeConstraintsTask& operator=(const PxsSolverCreateFinalizeConstraintsTask&);
public:

	PxsSolverCreateFinalizeConstraintsTask(DynamicsContext& context, IslandContext& islandContext, PxU32 solverDataOffset, PxsContactManagerOutputIterator& outputs, bool enhancedDeterminism) : 
		Cm::Task				(context.getContextId()),
		mContext				(context),
		mIslandContext			(islandContext),
		mSolverDataOffset		(solverDataOffset),
		mOutputs				(outputs),
		mEnhancedDeterminism	(enhancedDeterminism)
	{
	}
	
	virtual void runInternal() PX_OVERRIDE;

	virtual const char* getName() const PX_OVERRIDE { return "PxsDynamics.solverCreateFinalizeConstraints"; }

	DynamicsContext&					mContext;
	IslandContext&						mIslandContext;
	const PxU32							mSolverDataOffset;
	PxsContactManagerOutputIterator&	mOutputs;
	const bool							mEnhancedDeterminism;
};

static void createFinalizeConstraints(DynamicsContext& context, IslandContext& islandContext, PxSolverBodyData* solverBodyData, PxsContactManagerOutputIterator& outputs, PxBaseTask* continuation, PxU32 numThreads, PxU64 contextID, bool enhancedDeterminism);

class PxsSolverPartitionAndCreateFinalizeConstraintsTask : public Cm::Task
{
	PxsSolverPartitionAndCreateFinalizeConstraintsTask& operator=(const PxsSolverPartitionAndCreateFinalizeConstraintsTask&);
public:

	PxsSolverPartitionAndCreateFinalizeConstraintsTask(DynamicsContext& context, IslandContext& islandContext, PxsContactManagerOutputIterator& outputs, PxU32 solverBodyOffset, bool enhancedDeterminism) :
		Cm::Task			(context.getContextId()),
		mContext			(context), 
		mIslandContext		(islandContext),
		mOutputs			(outputs),
		mSolverBodyOffset	(solverBodyOffset),
		mEnhancedDeterminism(enhancedDeterminism)
	{
	}

	virtual void runInternal() PX_OVERRIDE
	{
		partitionConstraints(mIslandContext, mContext.mSolverBodyPool.begin() + mSolverBodyOffset, mContext.getContextId());

		const PxU32 numThreads = getTaskManager()->getCpuDispatcher()->getWorkerCount();
		createFinalizeConstraints(mContext, mIslandContext, mContext.mSolverBodyDataPool.begin(), mOutputs, mCont, numThreads, mContextID, mEnhancedDeterminism);
	}

	virtual const char* getName() const PX_OVERRIDE { return "PxsDynamics.partitionAndCreateFinalizeConstraintsTask"; }

	DynamicsContext&					mContext;
	IslandContext&						mIslandContext;
	PxsContactManagerOutputIterator&	mOutputs;
	const PxU32							mSolverBodyOffset;
	const bool							mEnhancedDeterminism;
};

// helper function to join two tasks together and ensure ref counts are correct
static void chainTasks(PxLightCpuTask* first, PxLightCpuTask* next)
{
	first->setContinuation(next);
	next->removeReference();
}

static void createSolverTaskChain(	DynamicsContext& dynamicContext,
									const SolverIslandObjects& objects,
									const PxsIslandIndices& counts,
									PxU32 solverBodyOffset,
									IG::SimpleIslandManager& islandManager,
									PxU32* bodyRemapTable, PxsMaterialManager* materialManager, PxBaseTask* continuation,
									PxsContactManagerOutputIterator& iterator, bool useEnhancedDeterminism)
{
	Cm::FlushPool& taskPool = dynamicContext.getTaskPool();
	taskPool.lock();

	IslandContext* islandContext = reinterpret_cast<IslandContext*>(taskPool.allocate(sizeof(IslandContext)));
	islandContext->mThreadContext = NULL;
	islandContext->mCounts = counts;

	// PT:
	// "X -> Y" means X has Y as continuation
	// "X => Y" means X spawns Y task(s)
	//
	// Chain is: startTask -> partitionConstraintsTask -> createFinalizeConstraintsTask -> setupSolveTask -> endTask -> continuation
	//
	// startTask [PxsSolverStartTask]:
	// - setupThreadContext()
	// - startTasks()
	// - integrate() => PxsPreIntegrateTask -> PxsSolverConstraintPartitionTask
	// - setupDescTask()
	// - articulationTask() => SolverArticulationUpdateTask -> PxsSolverConstraintPartitionTask
	//
	// partitionConstraintsTask [PxsSolverConstraintPartitionTask]:
	// -
	//
	// createFinalizeConstraintsTask [PxsSolverCreateFinalizeConstraintsTask]:
	// => PxsCreateFinalizeContactsTask -> PxsSolverSetupSolveTask
	// => PxsCreateArticConstraintsTask -> PxsSolverSetupSolveTask
	//
	// setupSolveTask [PxsSolverSetupSolveTask]:
	// => PxsParallelSolverTask -> PxsSolverEndTask
	//
	// endTask [PxsSolverEndTask]:
	// -
	//
	// ========================================================================
	//
	// The above notes indicate that we could merge PxsSolverConstraintPartitionTask and PxsSolverCreateFinalizeConstraintsTask,
	// as PxsSolverConstraintPartitionTask only does local work and does not spawn new tasks.
	//
	// Without the compound constraints PxsSolverEndTask also does not do much, and the small cleanup there could potentially
	// be moved to the continuation task.
	//
	// It is also a question whether PxsPreIntegrateTask and SolverArticulationUpdateTask could jump over PxsSolverConstraintPartitionTask.
	// In which case we should keep that task independent.

	// PT: TODO: would it be more efficient to allocate all of them at once from the task pool?

	if(!gMergePartitionAndFinalizeConstraintsTasks)
	{
		// create lead task
		PxsSolverStartTask* startTask = PX_PLACEMENT_NEW(taskPool.allocateNotThreadSafe(sizeof(PxsSolverStartTask)), PxsSolverStartTask)(dynamicContext, *islandContext, objects, solverBodyOffset, dynamicContext.getKinematicCount(), 
			islandManager, bodyRemapTable, materialManager, iterator, useEnhancedDeterminism);
		PxsSolverEndTask* endTask = PX_PLACEMENT_NEW(taskPool.allocateNotThreadSafe(sizeof(PxsSolverEndTask)), PxsSolverEndTask)(dynamicContext, *islandContext, objects, solverBodyOffset, iterator);

		PxsSolverCreateFinalizeConstraintsTask* createFinalizeConstraintsTask = PX_PLACEMENT_NEW(taskPool.allocateNotThreadSafe(sizeof(PxsSolverCreateFinalizeConstraintsTask)), PxsSolverCreateFinalizeConstraintsTask)(dynamicContext, *islandContext, solverBodyOffset, iterator, useEnhancedDeterminism);
		PxsSolverSetupSolveTask* setupSolveTask = PX_PLACEMENT_NEW(taskPool.allocateNotThreadSafe(sizeof(PxsSolverSetupSolveTask)), PxsSolverSetupSolveTask)(dynamicContext, *islandContext, objects, solverBodyOffset, islandManager.getAccurateIslandSim());

		PxsSolverConstraintPartitionTask* partitionConstraintsTask = PX_PLACEMENT_NEW(taskPool.allocateNotThreadSafe(sizeof(PxsSolverConstraintPartitionTask)), PxsSolverConstraintPartitionTask)(dynamicContext, *islandContext, solverBodyOffset);

		taskPool.unlock();

		endTask->setContinuation(continuation);

		// set up task chain in reverse order
		chainTasks(setupSolveTask, endTask);
		chainTasks(createFinalizeConstraintsTask, setupSolveTask);
		chainTasks(partitionConstraintsTask, createFinalizeConstraintsTask);
		chainTasks(startTask, partitionConstraintsTask);

		startTask->removeReference();
	}
	else
	{
		// create lead task
		PxsSolverStartTask* startTask = PX_PLACEMENT_NEW(taskPool.allocateNotThreadSafe(sizeof(PxsSolverStartTask)), PxsSolverStartTask)(dynamicContext, *islandContext, objects, solverBodyOffset, dynamicContext.getKinematicCount(), 
			islandManager, bodyRemapTable, materialManager, iterator, useEnhancedDeterminism);
		PxsSolverEndTask* endTask = PX_PLACEMENT_NEW(taskPool.allocateNotThreadSafe(sizeof(PxsSolverEndTask)), PxsSolverEndTask)(dynamicContext, *islandContext, objects, solverBodyOffset, iterator);

		PxsSolverSetupSolveTask* setupSolveTask = PX_PLACEMENT_NEW(taskPool.allocateNotThreadSafe(sizeof(PxsSolverSetupSolveTask)), PxsSolverSetupSolveTask)(dynamicContext, *islandContext, objects, solverBodyOffset, islandManager.getAccurateIslandSim());

		PxsSolverPartitionAndCreateFinalizeConstraintsTask* partitionAndCreateFinalizeConstraintsTask = PX_PLACEMENT_NEW(taskPool.allocateNotThreadSafe(sizeof(PxsSolverPartitionAndCreateFinalizeConstraintsTask)), PxsSolverPartitionAndCreateFinalizeConstraintsTask)
					(dynamicContext, *islandContext, iterator, solverBodyOffset, useEnhancedDeterminism);

		taskPool.unlock();

		endTask->setContinuation(continuation);

		// set up task chain in reverse order
		chainTasks(setupSolveTask, endTask);
		chainTasks(partitionAndCreateFinalizeConstraintsTask, setupSolveTask);
		chainTasks(startTask, partitionAndCreateFinalizeConstraintsTask);

		startTask->removeReference();
	}
}

namespace
{
class UpdateContinuationTask : public Cm::Task
{
	DynamicsContext& mContext;
	IG::SimpleIslandManager& mSimpleIslandManager;
	PxBaseTask* mLostTouchTask;
	PxU32 mMaxArticulationLinks;

	PX_NOCOPY(UpdateContinuationTask)
public:

	UpdateContinuationTask(DynamicsContext& context,
		IG::SimpleIslandManager& simpleIslandManager,
		PxBaseTask* lostTouchTask, PxU64 contextID, PxU32 maxLinks) :
		Cm::Task(contextID), mContext(context), mSimpleIslandManager(simpleIslandManager),
		mLostTouchTask(lostTouchTask), mMaxArticulationLinks(maxLinks)
	{
	}

	virtual const char* getName() const PX_OVERRIDE { return "UpdateContinuationTask"; }

	virtual void runInternal() PX_OVERRIDE
	{
		mContext.updatePostKinematic(mSimpleIslandManager, mCont, mLostTouchTask, mMaxArticulationLinks);
		//Allow lost touch task to run once all tasks have be scheduled
		mLostTouchTask->removeReference();
	}
};

class KinematicCopyTask : public Cm::Task
{
	const PxNodeIndex* const	mKinematicIndices;
	const PxU32					mNbKinematics;
	const IG::IslandSim&		mIslandSim;
	PxSolverBodyData*			mBodyData;

	PX_NOCOPY(KinematicCopyTask)

public:

	static const PxU32 NbKinematicsPerTask = 1024;

	KinematicCopyTask(const PxNodeIndex* const kinematicIndices,
		PxU32 nbKinematics, const IG::IslandSim& islandSim, 
		PxSolverBodyData* datas, PxU64 contextID) : Cm::Task(contextID),
		mKinematicIndices(kinematicIndices), mNbKinematics(nbKinematics),
		mIslandSim(islandSim), mBodyData(datas)
	{
	}

	virtual const char* getName() const PX_OVERRIDE { return "KinematicCopyTask"; }

	virtual void runInternal() PX_OVERRIDE
	{
		for (PxU32 i = 0; i<mNbKinematics; i++)
		{
			PxsRigidBody* rigidBody = getRigidBodyFromIG(mIslandSim, mKinematicIndices[i]);
			const PxsBodyCore& core = rigidBody->getCore();
			copyToSolverBodyData(core.linearVelocity, core.angularVelocity, core.inverseMass, core.inverseInertia, core.body2World, core.maxPenBias,
				core.maxContactImpulse, mKinematicIndices[i].index(), core.contactReportThreshold, mBodyData[i + 1], core.lockFlags, 0.f,
				core.mFlags & PxRigidBodyFlag::eENABLE_GYROSCOPIC_FORCES);
			rigidBody->saveLastCCDTransform();
		}
	}
};
}

void DynamicsContext::update(Cm::FlushPool& /*flushPool*/, PxBaseTask* continuation, PxBaseTask* /*postPartitioningTask*/, PxBaseTask* lostTouchTask,
	PxvNphaseImplementationContext* nphase, PxU32 /*maxPatches*/, PxU32 maxArticulationLinks,
	PxReal dt, const PxVec3& gravity, Cm::PinnableBitMap& /*changedHandleMap*/)
{
	if(!updateShared(nphase, dt, gravity))
		return;

	PX_PROFILE_ZONE("Dynamics.solverQueueTasks", mContextID);

	//Block to make sure it doesn't run before stage2 of update!
	lostTouchTask->addReference();

	UpdateContinuationTask* task = PX_PLACEMENT_NEW(mTaskPool.allocate(sizeof(UpdateContinuationTask)), UpdateContinuationTask)
		(*this, mIslandManager, lostTouchTask, mContextID, maxArticulationLinks);

	task->setContinuation(continuation);

	//KS - test that world solver body's velocities are finite and 0, then set it to 0.
	//Technically, the velocity should always be 0 but can be stomped if a NAN creeps into the simulation.
	PX_ASSERT(mWorldSolverBody.linearVelocity == PxVec3(0.0f));
	PX_ASSERT(mWorldSolverBody.angularState == PxVec3(0.0f));
	PX_ASSERT(mWorldSolverBody.linearVelocity.isFinite());
	PX_ASSERT(mWorldSolverBody.angularState.isFinite());

	mWorldSolverBody.linearVelocity = mWorldSolverBody.angularState = PxVec3(0.f);

	const IG::IslandSim& islandSim = mIslandManager.getAccurateIslandSim();
	const PxU32 kinematicCount = islandSim.getNbActiveKinematics();
	const PxNodeIndex* const kinematicIndices = islandSim.getActiveKinematics();
	mKinematicCount = kinematicCount;

	{
		const PxU32 bodyCount = islandSim.getNbActiveNodes(IG::Node::eRIGID_BODY_TYPE);

		if (kinematicCount + bodyCount > mSolverBodyPool.capacity())
		{
			PX_PROFILE_ZONE("reserve", mContextID);

			mSolverBodyPool.reserve((kinematicCount + bodyCount + 31) & ~31); // pad out to 32 * 128 = 4k to prevent alloc churn
			mSolverBodyDataPool.reserve((kinematicCount + bodyCount + 31 + 1) & ~31); // pad out to 32 * 128 = 4k to prevent alloc churn
			mSolverBodyRemapTable.reserve((kinematicCount + bodyCount + 31 + 1) & ~31);
		}

		{
			PX_PROFILE_ZONE("resize", mContextID);

			PxSolverBody emptySolverBody;
			PxMemZero(&emptySolverBody, sizeof(PxSolverBody));
			mSolverBodyPool.resize(kinematicCount + bodyCount, emptySolverBody);
			PxSolverBodyData emptySolverBodyData;
			PxMemZero(&emptySolverBodyData, sizeof(PxSolverBodyData));
			mSolverBodyDataPool.resize(kinematicCount + bodyCount + 1, emptySolverBodyData);
			mSolverBodyRemapTable.resize(bodyCount);
		}

		// integrate and copy all the kinematics - overkill, since not all kinematics
		// need solver bodies

		mSolverBodyDataPool[0] = mWorldSolverBodyData;

		if(kinematicCount)
		{
			PX_PROFILE_ZONE("Dynamics.updateKinematics", mContextID);

			PxMemZero(mSolverBodyPool.begin(), kinematicCount * sizeof(PxSolverBody));
			for (PxU32 i = 0; i < kinematicCount; i+= KinematicCopyTask::NbKinematicsPerTask)
			{
				const PxU32 nbToProcess = PxMin(KinematicCopyTask::NbKinematicsPerTask, kinematicCount - i);

				KinematicCopyTask* copyTask = PX_PLACEMENT_NEW(mTaskPool.allocate(sizeof(KinematicCopyTask)), KinematicCopyTask)
					(&kinematicIndices[i], nbToProcess, islandSim, &mSolverBodyDataPool[i], mContextID);
				
				copyTask->setContinuation(task);

				copyTask->removeReference();
			}
		}
	}

	{
		PX_PROFILE_ZONE("reserve2", mContextID);

		const PxU32 totalConstraintCount = reserveSharedSolverConstraintsArrays(islandSim, maxArticulationLinks);

		mSolverConstraintDescPool.forceSize_Unsafe(0);
		mSolverConstraintDescPool.reserve((totalConstraintCount + 63) & (~63));
		mSolverConstraintDescPool.forceSize_Unsafe(totalConstraintCount);

		mOrderedSolverConstraintDescPool.forceSize_Unsafe(0);
		mOrderedSolverConstraintDescPool.reserve((totalConstraintCount + 63) & (~63));
		mOrderedSolverConstraintDescPool.forceSize_Unsafe(totalConstraintCount);

		mTempSolverConstraintDescPool.forceSize_Unsafe(0);
		mTempSolverConstraintDescPool.reserve((totalConstraintCount + 63) & (~63));
		mTempSolverConstraintDescPool.forceSize_Unsafe(totalConstraintCount);
	}

	task->removeReference();
}

void DynamicsContext::updatePostKinematic(IG::SimpleIslandManager& simpleIslandManager, PxBaseTask* /*continuation*/, PxBaseTask* lostTouchTask, PxU32 maxLinks)
{
	const IG::IslandSim& islandSim = simpleIslandManager.getAccurateIslandSim();

	const PxU32 islandCount = islandSim.getNbActiveIslands();

	PxU32 constraintIndex = 0;

	const PxU32 solverBatchMax = mSolverBatchSize;
	const PxU32 articulationBatchMax = mSolverArticBatchSize;
	const PxU32 minimumConstraintCount = 1;

	//create force threshold tasks to produce force change events
	PxsForceThresholdTask* forceThresholdTask = PX_PLACEMENT_NEW(getTaskPool().allocate(sizeof(PxsForceThresholdTask)), PxsForceThresholdTask)(*this);
	forceThresholdTask->setContinuation(lostTouchTask);

	const IG::IslandId*const islandIds = islandSim.getActiveIslands();

	PxU32 currentIsland = 0;
	PxU32 currentBodyIndex = 0;
	PxU32 currentArticulation = 0;
	PxU32 currentContact = 0;

	while(currentIsland < islandCount)
	{
		SolverIslandObjects objectStarts;
		objectStarts.articulations				= mArticulationArray.begin() + currentArticulation;
		objectStarts.bodies						= mRigidBodyArray.begin() + currentBodyIndex;
		objectStarts.contactManagers			= mContactList.begin() + currentContact;
		objectStarts.constraintDescs			= mSolverConstraintDescPool.begin() + constraintIndex;
		objectStarts.orderedConstraintDescs		= mOrderedSolverConstraintDescPool.begin() + constraintIndex;
		objectStarts.tempConstraintDescs		= mTempSolverConstraintDescPool.begin() + constraintIndex;
		objectStarts.constraintBatchHeaders		= mContactConstraintBatchHeaders.begin() + constraintIndex;
		objectStarts.motionVelocities			= mMotionVelocityArray.begin() + currentBodyIndex;
		objectStarts.bodyCoreArray				= mBodyCoreArray.begin() + currentBodyIndex;
		objectStarts.islandIds					= islandIds + currentIsland;
		objectStarts.bodyRemapTable				= mSolverBodyRemapTable.begin();
		objectStarts.nodeIndexArray				= mNodeIndexArray.begin() + currentBodyIndex;

		PxU32 startIsland = currentIsland;
		PxU32 constraintCount = 0;

		PxU32 nbArticulations = 0;
		PxU32 nbBodies = 0;
		PxU32 nbConstraints = 0;
		PxU32 nbContactManagers =0;

		// islandSim.checkInternalConsistency();

		//KS - logic is a bit funky here. We will keep rolling the island together provided currentIsland < islandCount AND either we haven't exceeded the max number of bodies or we have
		//zero constraints AND we haven't exceeded articulation batch counts (it's still currently beneficial to keep articulations in separate islands but this is only temporary).
		while((currentIsland < islandCount && (nbBodies < solverBatchMax || constraintCount < minimumConstraintCount)) && nbArticulations < articulationBatchMax)
		{
			const IG::Island& island = islandSim.getIsland(islandIds[currentIsland]);
			nbBodies += island.mNodeCount[IG::Node::eRIGID_BODY_TYPE];
			nbArticulations += island.mNodeCount[IG::Node::eARTICULATION_TYPE];
			nbConstraints += island.mEdges.getCount(IG::Edge::eCONSTRAINT);
			nbContactManagers += island.mEdges.getCount(IG::Edge::eCONTACT_MANAGER);
			constraintCount = nbConstraints + nbContactManagers;
			currentIsland++;

			if(mUseEnhancedDeterminism)
				break;
		}

		objectStarts.numIslands = currentIsland - startIsland;

		constraintIndex += nbArticulations * maxLinks;

		PxsIslandIndices counts;
		counts.bodies			= nbBodies;
		counts.articulations	= nbArticulations;
		counts.contactManagers	= nbContactManagers;
		counts.constraints		= nbConstraints;

		if(counts.articulations + counts.bodies > 0)
		{
			createSolverTaskChain(*this, objectStarts, counts, 
				mKinematicCount + currentBodyIndex, simpleIslandManager, mSolverBodyRemapTable.begin(), mMaterialManager,
				forceThresholdTask, mOutputIterator, mUseEnhancedDeterminism);
		}

		currentBodyIndex += nbBodies;
		currentArticulation += nbArticulations;
		currentContact += nbContactManagers;

		constraintIndex += constraintCount;
	}

	//kick off forceThresholdTask
	forceThresholdTask->removeReference();
}

void DynamicsContext::mergeResults()
{
	PX_PROFILE_ZONE("Dynamics.solverMergeResults", mContextID);
	//OK. Sum up sim stats here...

#if PX_ENABLE_SIM_STATS	
	{
		PxcThreadCoherentCacheIterator<ThreadContext, PxcNpMemBlockPool> threadContextIt(mThreadContextPool);
		ThreadContext* threadContext = threadContextIt.getNext();

		while (threadContext != NULL)
		{
			ThreadContext::ThreadSimStats& threadStats = threadContext->getSimStats();
			addThreadStats(threadStats);
			threadStats.clear();
			threadContext = threadContextIt.getNext();
		}
	}
#else
	PX_CATCH_UNDEFINED_ENABLE_SIM_STATS
#endif
}

static void preIntegrationParallel(
	PxF32 dt,
	PxsBodyCore*const* bodyArray,					// INOUT: core body attributes
	PxsRigidBody*const* originalBodyArray,			// IN: original bodies
	PxU32 const* nodeIndexArray,					// IN: island node index
	PxU32 bodyCount,								// IN: body count
	PxSolverBodyData* solverBodyDataPool,			// IN: solver body data pool (space preallocated)
	volatile PxU32* maxSolverPositionIterations,
	volatile PxU32* maxSolverVelocityIterations,
	const PxVec3& gravity)
{
	PxU32 localMaxPosIter = 0;
	PxU32 localMaxVelIter = 0;

	for(PxU32 i = 0; i < bodyCount; ++i)
	{
		PxsBodyCore& core = *bodyArray[i];
		const PxsRigidBody& rBody = *originalBodyArray[i];
		
		const PxU16 iterWord = core.solverIterationCounts;
		localMaxPosIter = PxMax<PxU32>(PxU32(iterWord & 0xff), localMaxPosIter);
		localMaxVelIter = PxMax<PxU32>(PxU32(iterWord >> 8), localMaxVelIter);

		bodyCoreComputeUnconstrainedVelocity(gravity, dt, core.linearDamping, core.angularDamping, rBody.mAccelScale, core.maxLinearVelocitySq, core.maxAngularVelocitySq,
			core.linearVelocity, core.angularVelocity, core.disableGravity!=0);

		copyToSolverBodyData(core.linearVelocity, core.angularVelocity, core.inverseMass, core.inverseInertia, core.body2World, core.maxPenBias, core.maxContactImpulse, nodeIndexArray[i], 
			core.contactReportThreshold, solverBodyDataPool[i + 1], core.lockFlags, dt, core.mFlags & PxRigidBodyFlag::eENABLE_GYROSCOPIC_FORCES);
	}

	physx::PxAtomicMax(reinterpret_cast<volatile PxI32*>(maxSolverPositionIterations), PxI32(localMaxPosIter));
	physx::PxAtomicMax(reinterpret_cast<volatile PxI32*>(maxSolverVelocityIterations), PxI32(localMaxVelIter));
}

void PxsPreIntegrateTask::runInternal()
{
	PX_PROFILE_ZONE("PreIntegration", mContextID);

	preIntegrationParallel(mDt, mBodyArray + mStartIndex, mOriginalBodyArray + mStartIndex, mNodeIndexArray + mStartIndex, mNumToIntegrate,
						mSolverBodyDataPool + mStartIndex, mMaxSolverPositionIterations, mMaxSolverVelocityIterations, mGravity);
}

void DynamicsContext::preIntegrationParallel(
	PxF32 dt,
	PxsBodyCore*const* bodyArray,					// INOUT: core body attributes
	PxsRigidBody*const* originalBodyArray,			// IN: original bodies
	PxU32 const* nodeIndexArray,					// IN: island node index
	PxU32 bodyCount,								// IN: body count
	PxSolverBody* solverBodyPool,					// IN: solver body pool (space preallocated)
	PxSolverBodyData* solverBodyDataPool,			// IN: solver body data pool (space preallocated)
	Cm::SpatialVector* /*motionVelocityArray*/,		// OUT: motion velocities
	PxU32& maxSolverPositionIterations,
	PxU32& maxSolverVelocityIterations,
	PxBaseTask& task
	)
{
	//TODO - make this based on some variables so we can try different configurations
	const PxU32 IntegrationPerThread = 256;

	const PxU32 numTasks = ((bodyCount + IntegrationPerThread-1)/IntegrationPerThread);
	const PxU32 taskBatchSize = 64;

	for(PxU32 i = 0; i < numTasks; i+=taskBatchSize)
	{
		const PxU32 nbTasks = PxMin(numTasks - i, taskBatchSize);
		PxsPreIntegrateTask* tasks = reinterpret_cast<PxsPreIntegrateTask*>(getTaskPool().allocate(sizeof(PxsPreIntegrateTask)*nbTasks));
		for(PxU32 a = 0; a < nbTasks; ++a)
		{
			PxU32 startIndex = (i+a)*IntegrationPerThread;
			PxU32 nbToIntegrate = PxMin((bodyCount-startIndex), IntegrationPerThread);
			PxsPreIntegrateTask* pTask = PX_PLACEMENT_NEW(&tasks[a], PxsPreIntegrateTask)(*this, bodyArray,
							originalBodyArray, nodeIndexArray, solverBodyDataPool, dt, bodyCount,
							&maxSolverPositionIterations, &maxSolverVelocityIterations, startIndex, 
							nbToIntegrate, mGravity);

			pTask->setContinuation(&task);
			pTask->removeReference();
		}
	}

	PxMemZero(solverBodyPool, bodyCount * sizeof(PxSolverBody));
}

void solveParallel(SOLVER_PARALLEL_METHOD_ARGS)
{
	Dy::ThreadContext& threadContext = *context.getThreadContext();

	threadContext.mDeltaV.forceSize_Unsafe(0);
	threadContext.mDeltaV.reserve(params.mMaxArticulationLinks);
	threadContext.mDeltaV.forceSize_Unsafe(params.mMaxArticulationLinks);

	context.solveParallel(params, islandSim, threadContext.mDeltaV.begin());

	context.putThreadContext(&threadContext);
}

void DynamicsContext::solveParallel(SolverIslandParams& params, IG::IslandSim& islandSim, 
	Cm::SpatialVectorF* deltaV)
{
	solveVParallelAndWriteBack(params, deltaV, getBiasCoefficients().articulation,
		mSolveFrictionEveryIteration, mSolveArticulationContactLast);
	integrateCoreParallel(params, deltaV, islandSim);
}

void DynamicsContext::integrateCoreParallel(SolverIslandParams& params, Cm::SpatialVectorF* deltaV, IG::IslandSim& islandSim)
{
	const PxI32 unrollCount = 128;

	PxI32* bodyIntegrationListIndex = &params.bodyIntegrationListIndex;

	PxI32 index = physx::PxAtomicAdd(bodyIntegrationListIndex, unrollCount) - unrollCount;

	const PxI32 numBodies = PxI32(params.bodyListSize);
	const PxI32 numArtics = PxI32(params.articulationListSize);

	Cm::SpatialVector* PX_RESTRICT motionVelocityArray = params.motionVelocityArray;
	PxsRigidBody** PX_RESTRICT rigidBodies = params.rigidBodies;
	FeatherstoneArticulation** PX_RESTRICT articulationListStart = params.articulationListStart;

	PxI32 numIntegrated = 0;

	PxI32 bodyRemainder = unrollCount;

	while(index < numArtics)
	{
		const PxI32 remainder = PxMin(numArtics - index, unrollCount);
		bodyRemainder -= remainder;

		for(PxI32 a = 0; a < remainder; ++a, index++)
		{
			const PxI32 i = index;
			{
				//PX_PROFILE_ZONE("Articulations.integrate", mContextID);

				FeatherstoneArticulation::updateBodies(articulationListStart[i], deltaV, mDt);
			}

			++numIntegrated;
		}
		if(bodyRemainder == 0)
		{
			index = physx::PxAtomicAdd(bodyIntegrationListIndex, unrollCount) - unrollCount;
			bodyRemainder = unrollCount;
		}
	}

	index -= numArtics;

	const PxI32 unrollPlusArtics = unrollCount + numArtics;

	PxSolverBody* PX_RESTRICT solverBodies = const_cast<PxSolverBody*>(params.bodyListStart);
	PxSolverBodyData* PX_RESTRICT solverBodyData = params.bodyDataList + params.solverBodyOffset+1;

	while(index < numBodies)
	{
		const PxI32 remainder = PxMin(numBodies - index, bodyRemainder);
		bodyRemainder -= remainder;
		integrate(islandSim, solverBodyData + index, rigidBodies + index, motionVelocityArray + index, solverBodies + index, remainder, mDt, mEnableStabilization, isSleepingDisabled());
		numIntegrated += remainder;

		{
			index = physx::PxAtomicAdd(bodyIntegrationListIndex, unrollCount) - unrollPlusArtics;
			bodyRemainder = unrollCount;
		}
	}

	PxMemoryBarrier();
	physx::PxAtomicAdd(&params.numObjectsIntegrated, numIntegrated);
}

static PxU32 createFinalizeContacts_Parallel(PxSolverBodyData* solverBodyData, ThreadContext& threadContext, DynamicsContext& context,
									  PxU32 startIndex, PxU32 endIndex, PxsContactManagerOutputIterator& outputs)
{
	PX_PROFILE_ZONE("createFinalizeContacts_Parallel", context.getContextId());
	const PxReal correlationDist = context.getCorrelationDistance();
	const PxReal bounceThreshold = context.getBounceThreshold();
	const PxReal frictionOffsetThreshold = context.getFrictionOffsetThreshold();
	const PxReal dt = context.getDt();
	const PxReal invDt = PxMin(context.getMaxBiasCoefficient(), context.getInvDt());

	PxSolverConstraintDesc* contactDescPtr = threadContext.orderedContactConstraints;

	PxConstraintBatchHeader* headers = threadContext.contactConstraintBatchHeaders;
	
	PxI32 axisConstraintCount = 0;
	ThreadContext* threadContextLocal = context.getThreadContext();
	threadContextLocal->mConstraintBlockStream.reset(); //ensure there's no left-over memory that belonged to another island

	threadContextLocal->mZVector.forceSize_Unsafe(0);
	threadContextLocal->mZVector.reserve(threadContext.mMaxArticulationLinks);
	threadContextLocal->mZVector.forceSize_Unsafe(threadContext.mMaxArticulationLinks);

	//threadContextLocal->mDeltaV.forceSize_Unsafe(0);
	//threadContextLocal->mDeltaV.reserve(threadContext.mMaxArticulationLinks);
	//threadContextLocal->mDeltaV.forceSize_Unsafe(threadContext.mMaxArticulationLinks);

	Cm::SpatialVectorF* Z = threadContextLocal->mZVector.begin();

	const PxTransform idt(PxIdentity);

	BlockAllocator blockAllocator(threadContext.mConstraintBlockManager, threadContextLocal->mConstraintBlockStream, threadContextLocal->mFrictionPatchStreamPair, threadContextLocal->mConstraintSize);

	const PxReal ccdMaxSeparation = context.getCCDSeparationThreshold();

	const PxReal rigidContactBiasCoefficient = context.getBiasCoefficients().rigidContact;
	const PxReal jointBiasCoefficient = context.getBiasCoefficients().joint;

	for(PxU32 a = startIndex; a < endIndex; ++a)
	{
		PxConstraintBatchHeader& header = headers[a];

		if(contactDescPtr[header.startIndex].constraintType == DY_SC_TYPE_RB_CONTACT)
		{
			PxSolverContactDesc blockDescs[4];
			PxsContactManagerOutput* cmOutputs[4];
			PxsContactManager* cms[4];
			for (PxU32 i = 0; i < header.stride; ++i)
			{
				PxSolverConstraintDesc& desc = contactDescPtr[header.startIndex + i];
				PxSolverContactDesc& blockDesc = blockDescs[i];
				PxsContactManager* cm = reinterpret_cast<PxsContactManager*>(desc.constraint);

				cms[i] = cm;

				PxcNpWorkUnit& unit = cm->getWorkUnit();

				cmOutputs[i] = &outputs.getContactManagerOutput(unit.mNpIndex);

				PxSolverBodyData& data0 = desc.linkIndexA != PxSolverConstraintDesc::RIGID_BODY ? solverBodyData[0] : solverBodyData[desc.bodyADataIndex];
				PxSolverBodyData& data1 = desc.linkIndexB != PxSolverConstraintDesc::RIGID_BODY ? solverBodyData[0] : solverBodyData[desc.bodyBDataIndex];

				blockDesc.data0 = &data0;
				blockDesc.data1 = &data1;

				PxU8 flags = unit.mRigidCore0->mFlags;
				if (unit.mRigidCore1)
					flags |= PxU8(unit.mRigidCore1->mFlags);

				blockDesc.bodyFrame0 = unit.mRigidCore0->body2World;
				blockDesc.bodyFrame1 = unit.mRigidCore1 ? unit.mRigidCore1->body2World : idt;
				blockDesc.shapeInteraction = cm->getShapeInteraction();
				blockDesc.contactForces = cmOutputs[i]->contactForces;
				blockDesc.desc = &desc;
				blockDesc.body0 = desc.bodyA;
				blockDesc.body1 = desc.bodyB;
				blockDesc.hasForceThresholds = !!(unit.mFlags & PxcNpWorkUnitFlag::eFORCE_THRESHOLD);
				blockDesc.disableStrongFriction = !!(unit.mFlags & PxcNpWorkUnitFlag::eDISABLE_STRONG_FRICTION);
				blockDesc.bodyState0 = (unit.mFlags & PxcNpWorkUnitFlag::eARTICULATION_BODY0) ? PxSolverContactDesc::eARTICULATION : PxSolverContactDesc::eDYNAMIC_BODY;
				//second body is articulation
				if (unit.mFlags & PxcNpWorkUnitFlag::eARTICULATION_BODY1)
				{
					//kinematic link
					if (desc.linkIndexB == 0xff)
					{
						blockDesc.bodyState1 = PxSolverContactDesc::eSTATIC_BODY;
					}
					else
					{
						blockDesc.bodyState1 = PxSolverContactDesc::eARTICULATION;
					}
				}
				else
				{
					blockDesc.bodyState1 = (unit.mFlags & PxcNpWorkUnitFlag::eHAS_KINEMATIC_ACTOR) ? PxSolverContactDesc::eKINEMATIC_BODY :
						((unit.mFlags & PxcNpWorkUnitFlag::eDYNAMIC_BODY1) ? PxSolverContactDesc::eDYNAMIC_BODY : PxSolverContactDesc::eSTATIC_BODY);
				}
				//blockDesc.flags = unit.flags;

				unit.setInvMassScaleFromDominance(blockDesc.invMassScales);

				blockDesc.restDistance = unit.mRestDistance;
				blockDesc.frictionPtr = unit.mFrictionDataPtr;
				blockDesc.frictionCount = unit.mFrictionPatchCount;
				blockDesc.maxCCDSeparation = (flags & PxRigidBodyFlag::eENABLE_SPECULATIVE_CCD) ? ccdMaxSeparation : PX_MAX_F32;
				blockDesc.offsetSlop = unit.mOffsetSlop;
			}

#if DY_BATCH_CONSTRAINTS
			SolverConstraintPrepState::Enum state = SolverConstraintPrepState::eUNBATCHABLE;
			if(header.stride == 4)
			{
				//KS - todo - plumb in axisConstraintCount into this method to keep track of the number of axes
				state = createFinalizeSolverContacts4(cmOutputs, *threadContextLocal,
					 blockDescs,
					 invDt,
					 dt,
					 bounceThreshold,
					 frictionOffsetThreshold,
					 correlationDist,
					 rigidContactBiasCoefficient,
					 blockAllocator);
			}
			if(SolverConstraintPrepState::eSUCCESS != state)
#endif
			{
				for(PxU32 i = 0; i < header.stride; ++i)
				{
					PxSolverConstraintDesc& desc = contactDescPtr[header.startIndex+i];
					PxsContactManager* cm = reinterpret_cast<PxsContactManager*>(desc.constraint);
					PxsContactManagerOutput& output = *cmOutputs[i];

					createFinalizeSolverContacts(blockDescs[i], output, *threadContextLocal,
						invDt, dt, bounceThreshold, frictionOffsetThreshold, correlationDist, rigidContactBiasCoefficient,
						blockAllocator, Z);
			
					getContactManagerConstraintDesc(output,*cm,desc);
				}
			}

			for (PxU32 i = 0; i < header.stride; ++i)
			{
				updateFrictionAnchorCountAndPosition(contactDescPtr[header.startIndex + i], *cmOutputs[i], blockDescs[i]);
			}

			for (PxU32 i = 0; i < header.stride; ++i)
			{
				PxsContactManager* cm = cms[i];

				PxcNpWorkUnit& unit = cm->getWorkUnit();
				unit.mFrictionDataPtr = blockDescs[i].frictionPtr;
				unit.mFrictionPatchCount = blockDescs[i].frictionCount;
				axisConstraintCount += blockDescs[i].axisConstraintCount;
			}
		}
		else if(contactDescPtr[header.startIndex].constraintType == DY_SC_TYPE_RB_1D)
		{
			SolverConstraintShaderPrepDesc shaderDescs[4];
			PxSolverConstraintPrepDesc descs[4];

			for (PxU32 i = 0; i < header.stride; ++i)
			{
				PxSolverConstraintDesc& desc = contactDescPtr[header.startIndex + i];
				const Constraint* constraint = reinterpret_cast<const Constraint*>(desc.constraint);

				SolverConstraintShaderPrepDesc& shaderPrepDesc = shaderDescs[i];
				PxSolverConstraintPrepDesc& prepDesc = descs[i];

				const PxConstraintSolverPrep solverPrep = constraint->solverPrep;
				const void* constantBlock = constraint->constantBlock;
				const PxU32 constantBlockByteSize = constraint->constantBlockSize;
				const PxTransform& pose0 = (constraint->body0 ? constraint->body0->getPose() : idt);
				const PxTransform& pose1 = (constraint->body1 ? constraint->body1->getPose() : idt);
				const PxSolverBody* sbody0 = desc.bodyA;
				const PxSolverBody* sbody1 = desc.bodyB;
				PxSolverBodyData* sbodyData0 = &solverBodyData[desc.linkIndexA == PxSolverConstraintDesc::RIGID_BODY ? desc.bodyADataIndex : 0];
				PxSolverBodyData* sbodyData1 = &solverBodyData[desc.linkIndexB == PxSolverConstraintDesc::RIGID_BODY ? desc.bodyBDataIndex : 0];

				shaderPrepDesc.constantBlock = constantBlock;
				shaderPrepDesc.constantBlockByteSize = constantBlockByteSize;
				shaderPrepDesc.constraint = constraint;
				shaderPrepDesc.solverPrep = solverPrep;

				prepDesc.desc = &desc;
				prepDesc.bodyFrame0 = pose0;
				prepDesc.bodyFrame1 = pose1;
				prepDesc.data0 = sbodyData0;
				prepDesc.data1 = sbodyData1;
				prepDesc.body0 = sbody0;
				prepDesc.body1 = sbody1;
				prepDesc.linBreakForce = constraint->linBreakForce;
				prepDesc.angBreakForce = constraint->angBreakForce;
				prepDesc.writeback = &context.getConstraintWriteBackPool()[constraint->index];
				setupConstraintFlags(prepDesc, constraint->flags);
				prepDesc.minResponseThreshold = constraint->minResponseThreshold;
			}

#if DY_BATCH_CONSTRAINTS && DY_BATCH_1D
			SolverConstraintPrepState::Enum state = SolverConstraintPrepState::eUNBATCHABLE;
			if(header.stride == 4)
			{
				PxU32 totalRows;
				state = setupSolverConstraint4
					(shaderDescs, descs, dt, invDt, jointBiasCoefficient, totalRows,
					blockAllocator);

				axisConstraintCount += totalRows;
			}
			if(state != SolverConstraintPrepState::eSUCCESS)
#endif
			{
				for(PxU32 i = 0; i < header.stride; ++i)
				{
					axisConstraintCount += SetupSolverConstraint(shaderDescs[i], descs[i], blockAllocator, dt, invDt, jointBiasCoefficient);
				}
			}
		}
	}

#if PX_ENABLE_SIM_STATS
	threadContextLocal->getSimStats().numAxisSolverConstraints += axisConstraintCount;
#else
	PX_CATCH_UNDEFINED_ENABLE_SIM_STATS
#endif

	context.putThreadContext(threadContextLocal);
	return PxU32(axisConstraintCount); //Can't write to threadContex as it's shared!!!!
}

class PxsCreateFinalizeContactsTask : public Cm::Task
{
	PxsCreateFinalizeContactsTask& operator=(const PxsCreateFinalizeContactsTask&);
public:
	PxsCreateFinalizeContactsTask( const PxU32 numConstraints, PxSolverConstraintDesc* descArray, PxSolverBodyData* solverBodyData,
		ThreadContext& threadContext, DynamicsContext& context, PxU32 startIndex, PxU32 endIndex, PxsContactManagerOutputIterator& outputs) :
			Cm::Task(context.getContextId()),
			mNumConstraints(numConstraints), mDescArray(descArray), mSolverBodyData(solverBodyData),
			mThreadContext(threadContext), mDynamicsContext(context),
			mOutputs(outputs),
			mStartIndex(startIndex), mEndIndex(endIndex)
	{}

	virtual void runInternal() PX_OVERRIDE
	{
		createFinalizeContacts_Parallel(mSolverBodyData, mThreadContext, mDynamicsContext, mStartIndex, mEndIndex, mOutputs);
	}

	virtual const char* getName() const PX_OVERRIDE
	{
		return "PxsDynamics.createFinalizeContacts";
	}

public:
	const PxU32 mNumConstraints;
	PxSolverConstraintDesc* mDescArray;
	PxSolverBodyData* mSolverBodyData;
	ThreadContext& mThreadContext;
	DynamicsContext& mDynamicsContext;
	PxsContactManagerOutputIterator& mOutputs;
	PxU32 mStartIndex;
	PxU32 mEndIndex;
};

PxU8* BlockAllocator::reserveConstraintData(const PxU32 size)
{
	mTotalConstraintByteSize += size;
	return mConstraintBlockStream.reserve(size, mConstraintBlockManager);
}

PxU8* BlockAllocator::reserveFrictionData(const PxU32 size)
{
	return mFrictionPatchStreamPair.reserve<PxU8>(size);
}

class PxsCreateArticConstraintsTask : public Cm::Task
{
	PxsCreateArticConstraintsTask& operator=(const PxsCreateArticConstraintsTask&);	

public:

	static const PxU32 NbArticsPerTask = 32;

	PxsCreateArticConstraintsTask(Dy::FeatherstoneArticulation** articulations, const PxU32 nbArticulations, PxSolverBodyData* solverBodyData, ThreadContext& threadContext, DynamicsContext& context,
		PxsContactManagerOutputIterator& outputs) :
		Cm::Task(context.getContextId()),
		mArticulations(articulations),
		mNbArticulations(nbArticulations),
		mSolverBodyData(solverBodyData),
		mThreadContext(threadContext), mDynamicsContext(context),
		mOutputs(outputs)
	{}

	virtual void runInternal() PX_OVERRIDE
	{
		const PxReal correlationDist = mDynamicsContext.getCorrelationDistance();
		const PxReal bounceThreshold = mDynamicsContext.getBounceThreshold();
		const PxReal frictionOffsetThreshold = mDynamicsContext.getFrictionOffsetThreshold();
		const PxReal dt = mDynamicsContext.getDt();
		const PxReal invDt = PxMin(mDynamicsContext.getMaxBiasCoefficient(), mDynamicsContext.getInvDt());
		const PxReal ccdMaxSeparation = mDynamicsContext.getCCDSeparationThreshold();
		const PxReal rigidContactBiasCoefficient = mDynamicsContext.getBiasCoefficients().rigidContact;
		const PxReal jointBiasCoefficient = mDynamicsContext.getBiasCoefficients().joint;

		ThreadContext* threadContext = mDynamicsContext.getThreadContext();
		threadContext->mConstraintBlockStream.reset(); //ensure there's no left-over memory that belonged to another island

		threadContext->mZVector.forceSize_Unsafe(0);
		threadContext->mZVector.reserve(mThreadContext.mMaxArticulationLinks);
		threadContext->mZVector.forceSize_Unsafe(mThreadContext.mMaxArticulationLinks);

		for (PxU32 i = 0; i < mNbArticulations; ++i)
		{
			mArticulations[i]->prepareStaticConstraints(dt, invDt, mOutputs, *threadContext, correlationDist, bounceThreshold, frictionOffsetThreshold,
				ccdMaxSeparation, rigidContactBiasCoefficient, jointBiasCoefficient,
				mSolverBodyData, mThreadContext.mConstraintBlockManager, mDynamicsContext.getConstraintWriteBackPool().begin());
		}

		mDynamicsContext.putThreadContext(threadContext);
	}

	virtual const char* getName() const PX_OVERRIDE
	{
		return "PxsDynamics.createFinalizeContacts";
	}

public:

	Dy::FeatherstoneArticulation** mArticulations;
	const PxU32 mNbArticulations;
	PxSolverBodyData* mSolverBodyData;
	ThreadContext& mThreadContext;
	DynamicsContext& mDynamicsContext;
	PxsContactManagerOutputIterator& mOutputs;
};

static void createFinalizeConstraints(DynamicsContext& context, IslandContext& islandContext, PxSolverBodyData* solverBodyData, PxsContactManagerOutputIterator& outputs, PxBaseTask* continuation, PxU32 numThreads, PxU64 contextID, bool enhancedDeterminism)
{
	PX_PROFILE_ZONE("CreateConstraints", contextID);
	PX_UNUSED(contextID);

	ThreadContext& threadContext = *islandContext.mThreadContext;
	PxU32 descCount = threadContext.mNumDifferentBodyConstraints;
	PxU32 selfConstraintDescCount = threadContext.contactDescArraySize - (threadContext.mNumDifferentBodyConstraints + threadContext.mNumStaticConstraints);

	PxArray<PxU32>& accumulatedConstraintsPerPartition = threadContext.mConstraintsPerPartition;

	PxU32 numHeaders = 0;
	PxU32 currentPartition = 0;
	PxU32 maxJ = descCount == 0 ? 0 : accumulatedConstraintsPerPartition[0];

	const PxU32 maxBatchPartition = 0xFFFFFFFF;

	const PxU32 maxBatchSize = enhancedDeterminism ? 1u : 4u;

	PxU32 headersPerPartition = 0;
	for(PxU32 a = 0; a < descCount;)
	{
		PxU32 loopMax = PxMin(maxJ - a, maxBatchSize);
		PxU16 j = 0;
		if(loopMax > 0)
		{
			PxConstraintBatchHeader& header = threadContext.contactConstraintBatchHeaders[numHeaders++];
			
			j=1;
			PxSolverConstraintDesc& desc = threadContext.orderedContactConstraints[a];
			if(!isArticulationConstraint(desc) && (desc.constraintType == DY_SC_TYPE_RB_CONTACT || 
				desc.constraintType == DY_SC_TYPE_RB_1D) && currentPartition < maxBatchPartition)
			{
				for(; j < loopMax && desc.constraintType == threadContext.orderedContactConstraints[a+j].constraintType && 
					!isArticulationConstraint(threadContext.orderedContactConstraints[a+j]); ++j);
			}
			header.startIndex = a;
			header.stride = j;
			headersPerPartition++;
		}
		if(maxJ == (a + j) && maxJ != descCount)
		{
			//Go to next partition!
			accumulatedConstraintsPerPartition[currentPartition] = headersPerPartition;
			headersPerPartition = 0;
			currentPartition++;
			maxJ = accumulatedConstraintsPerPartition[currentPartition];
		}
		a+= j;
	}
	if(descCount)
		accumulatedConstraintsPerPartition[currentPartition] = headersPerPartition;

	accumulatedConstraintsPerPartition.forceSize_Unsafe(threadContext.mMaxPartitions);

	for(PxU32 a = 0; a < selfConstraintDescCount; ++a)
	{
		PxConstraintBatchHeader& header = threadContext.contactConstraintBatchHeaders[numHeaders++];
		header.startIndex = a + descCount;
		header.stride = 1;
	}

	threadContext.numContactConstraintBatches = numHeaders;

	Cm::FlushPool& flushPool = context.getTaskPool();

	{
		PxSolverConstraintDesc* descBegin = threadContext.orderedContactConstraints;
			
		//Choose an appropriate number of constraint prep tasks. This must be proportionate to the number of constraints to prep and the number
		//of worker threads available.
		const PxU32 TaskBlockSize = 16;
		const PxU32 TaskBlockLargeSize = 64;
		const PxU32 BlockAllocationSize = 64;

		PxU32 numTasks = (numHeaders+TaskBlockLargeSize-1)/TaskBlockLargeSize;

		if(numTasks)
		{
			if(numTasks < numThreads)
				numTasks = PxMax(1u, (numHeaders+TaskBlockSize-1)/TaskBlockSize);

			const PxU32 constraintsPerTask = (numHeaders + numTasks-1)/numTasks;

			for(PxU32 i = 0; i < numTasks; i+=BlockAllocationSize)
			{
				PxU32 blockSize = PxMin(numTasks - i, BlockAllocationSize);

				PxsCreateFinalizeContactsTask* tasks = reinterpret_cast<PxsCreateFinalizeContactsTask*>(flushPool.allocate(sizeof(PxsCreateFinalizeContactsTask)*blockSize));

				for(PxU32 a = 0; a < blockSize; ++a)
				{
					PxU32 startIndex = (a + i) * constraintsPerTask;
					PxU32 endIndex = PxMin(startIndex + constraintsPerTask, numHeaders);
					PxsCreateFinalizeContactsTask* pTask = PX_PLACEMENT_NEW(&tasks[a], PxsCreateFinalizeContactsTask(descCount, descBegin, solverBodyData, threadContext, context, startIndex, endIndex, outputs));

					pTask->setContinuation(continuation);
					pTask->removeReference();
				}
			}
		}
	}

	const PxU32 articCount = islandContext.mCounts.articulations;

	for (PxU32 i = 0; i < articCount; i += PxsCreateArticConstraintsTask::NbArticsPerTask)
	{
		const PxU32 nbToProcess = PxMin(articCount - i, PxsCreateArticConstraintsTask::NbArticsPerTask);

		PxsCreateArticConstraintsTask* task = PX_PLACEMENT_NEW(flushPool.allocate(sizeof(PxsCreateArticConstraintsTask)), PxsCreateArticConstraintsTask)
			(threadContext.mArticulationArray + i, nbToProcess, solverBodyData, threadContext, context, outputs);

		task->setContinuation(continuation);
		task->removeReference();
	}
}

void PxsSolverCreateFinalizeConstraintsTask::runInternal()
{
	const PxU32 numThreads = getTaskManager()->getCpuDispatcher()->getWorkerCount();
	createFinalizeConstraints(mContext, mIslandContext, mContext.mSolverBodyDataPool.begin(), mOutputs, mCont, numThreads, mContextID, mEnhancedDeterminism);
}

}
}