Merge pull request #1144 from lunkhound/pr-threading-refactor

Bullet 2 threading refactor: moved parallel-for calls into core libs

src/BulletDynamics/Dynamics/btDiscreteDynamicsWorldMt.cpp

@@ -108,8 +108,108 @@ struct InplaceSolverIslandCallbackMt : public btSimulationIslandManagerMt::Islan
 };
 
 
+///
+/// btConstraintSolverPoolMt
+///
 
-btDiscreteDynamicsWorldMt::btDiscreteDynamicsWorldMt(btDispatcher* dispatcher,btBroadphaseInterface* pairCache,btConstraintSolver* constraintSolver, btCollisionConfiguration* collisionConfiguration)
+btConstraintSolverPoolMt::ThreadSolver* btConstraintSolverPoolMt::getAndLockThreadSolver()
+{
+    int i = 0;
+#if BT_THREADSAFE
+    i = btGetCurrentThreadIndex() % m_solvers.size();
+#endif // #if BT_THREADSAFE
+    while ( true )
+    {
+        ThreadSolver& solver = m_solvers[ i ];
+        if ( solver.mutex.tryLock() )
+        {
+            return &solver;
+        }
+        // failed, try the next one
+        i = ( i + 1 ) % m_solvers.size();
+    }
+    return NULL;
+}
+
+void btConstraintSolverPoolMt::init( btConstraintSolver** solvers, int numSolvers )
+{
+    m_solverType = BT_SEQUENTIAL_IMPULSE_SOLVER;
+    m_solvers.resize( numSolvers );
+    for ( int i = 0; i < numSolvers; ++i )
+    {
+        m_solvers[ i ].solver = solvers[ i ];
+    }
+    if ( numSolvers > 0 )
+    {
+        m_solverType = solvers[ 0 ]->getSolverType();
+    }
+}
+
+// create the solvers for me
+btConstraintSolverPoolMt::btConstraintSolverPoolMt( int numSolvers )
+{
+    btAlignedObjectArray<btConstraintSolver*> solvers;
+    solvers.reserve( numSolvers );
+    for ( int i = 0; i < numSolvers; ++i )
+    {
+        btConstraintSolver* solver = new btSequentialImpulseConstraintSolver();
+        solvers.push_back( solver );
+    }
+    init( &solvers[ 0 ], numSolvers );
+}
+
+// pass in fully constructed solvers (destructor will delete them)
+btConstraintSolverPoolMt::btConstraintSolverPoolMt( btConstraintSolver** solvers, int numSolvers )
+{
+    init( solvers, numSolvers );
+}
+
+btConstraintSolverPoolMt::~btConstraintSolverPoolMt()
+{
+    // delete all solvers
+    for ( int i = 0; i < m_solvers.size(); ++i )
+    {
+        ThreadSolver& solver = m_solvers[ i ];
+        delete solver.solver;
+        solver.solver = NULL;
+    }
+}
+
+///solve a group of constraints
+btScalar btConstraintSolverPoolMt::solveGroup( btCollisionObject** bodies,
+    int numBodies,
+    btPersistentManifold** manifolds,
+    int numManifolds,
+    btTypedConstraint** constraints,
+    int numConstraints,
+    const btContactSolverInfo& info,
+    btIDebugDraw* debugDrawer,
+    btDispatcher* dispatcher
+)
+{
+    ThreadSolver* ts = getAndLockThreadSolver();
+    ts->solver->solveGroup( bodies, numBodies, manifolds, numManifolds, constraints, numConstraints, info, debugDrawer, dispatcher );
+    ts->mutex.unlock();
+    return 0.0f;
+}
+
+void btConstraintSolverPoolMt::reset()
+{
+    for ( int i = 0; i < m_solvers.size(); ++i )
+    {
+        ThreadSolver& solver = m_solvers[ i ];
+        solver.mutex.lock();
+        solver.solver->reset();
+        solver.mutex.unlock();
+    }
+}
+
+
+///
+/// btDiscreteDynamicsWorldMt
+///
+
+btDiscreteDynamicsWorldMt::btDiscreteDynamicsWorldMt(btDispatcher* dispatcher, btBroadphaseInterface* pairCache, btConstraintSolverPoolMt* constraintSolver, btCollisionConfiguration* collisionConfiguration)
 : btDiscreteDynamicsWorld(dispatcher,pairCache,constraintSolver,collisionConfiguration)
 {
 	if (m_ownsIslandManager)
@@ -124,8 +224,8 @@ btDiscreteDynamicsWorldMt::btDiscreteDynamicsWorldMt(btDispatcher* dispatcher,bt
 	{
 		void* mem = btAlignedAlloc(sizeof(btSimulationIslandManagerMt),16);
 		btSimulationIslandManagerMt* im = new (mem) btSimulationIslandManagerMt();
-        m_islandManager = im;
         im->setMinimumSolverBatchSize( m_solverInfo.m_minimumSolverBatchSize );
+        m_islandManager = im;
 	}
 }
 
@@ -145,7 +245,7 @@ btDiscreteDynamicsWorldMt::~btDiscreteDynamicsWorldMt()
 }
 
 
-void	btDiscreteDynamicsWorldMt::solveConstraints(btContactSolverInfo& solverInfo)
+void btDiscreteDynamicsWorldMt::solveConstraints(btContactSolverInfo& solverInfo)
 {
 	BT_PROFILE("solveConstraints");
 
@@ -160,3 +260,68 @@ void	btDiscreteDynamicsWorldMt::solveConstraints(btContactSolverInfo& solverInfo
 }
 
 
+struct UpdaterUnconstrainedMotion : public btIParallelForBody
+{
+    btScalar timeStep;
+    btRigidBody** rigidBodies;
+
+    void forLoop( int iBegin, int iEnd ) const BT_OVERRIDE
+    {
+        for ( int i = iBegin; i < iEnd; ++i )
+        {
+            btRigidBody* body = rigidBodies[ i ];
+            if ( !body->isStaticOrKinematicObject() )
+            {
+                //don't integrate/update velocities here, it happens in the constraint solver
+                body->applyDamping( timeStep );
+                body->predictIntegratedTransform( timeStep, body->getInterpolationWorldTransform() );
+            }
+        }
+    }
+};
+
+
+void btDiscreteDynamicsWorldMt::predictUnconstraintMotion( btScalar timeStep )
+{
+    BT_PROFILE( "predictUnconstraintMotion" );
+    if ( m_nonStaticRigidBodies.size() > 0 )
+    {
+        UpdaterUnconstrainedMotion update;
+        update.timeStep = timeStep;
+        update.rigidBodies = &m_nonStaticRigidBodies[ 0 ];
+        int grainSize = 50;  // num of iterations per task for task scheduler
+        btParallelFor( 0, m_nonStaticRigidBodies.size(), grainSize, update );
+    }
+}
+
+
+void btDiscreteDynamicsWorldMt::createPredictiveContacts( btScalar timeStep )
+{
+    BT_PROFILE( "createPredictiveContacts" );
+    releasePredictiveContacts();
+    if ( m_nonStaticRigidBodies.size() > 0 )
+    {
+        UpdaterCreatePredictiveContacts update;
+        update.world = this;
+        update.timeStep = timeStep;
+        update.rigidBodies = &m_nonStaticRigidBodies[ 0 ];
+        int grainSize = 50;  // num of iterations per task for task scheduler
+        btParallelFor( 0, m_nonStaticRigidBodies.size(), grainSize, update );
+    }
+}
+
+
+void btDiscreteDynamicsWorldMt::integrateTransforms( btScalar timeStep )
+{
+    BT_PROFILE( "integrateTransforms" );
+    if ( m_nonStaticRigidBodies.size() > 0 )
+    {
+        UpdaterIntegrateTransforms update;
+        update.world = this;
+        update.timeStep = timeStep;
+        update.rigidBodies = &m_nonStaticRigidBodies[ 0 ];
+        int grainSize = 50;  // num of iterations per task for task scheduler
+        btParallelFor( 0, m_nonStaticRigidBodies.size(), grainSize, update );
+    }
+}
+

src/BulletDynamics/Dynamics/btDiscreteDynamicsWorldMt.h

@@ -18,24 +18,116 @@ subject to the following restrictions:
 #define BT_DISCRETE_DYNAMICS_WORLD_MT_H
 
 #include "btDiscreteDynamicsWorld.h"
+#include "btSimulationIslandManagerMt.h"
+#include "BulletDynamics/ConstraintSolver/btConstraintSolver.h"
 
 struct InplaceSolverIslandCallbackMt;
 
+///
+/// btConstraintSolverPoolMt - masquerades as a constraint solver, but really it is a threadsafe pool of them.
+///
+///  Each solver in the pool is protected by a mutex.  When solveGroup is called from a thread,
+///  the pool looks for a solver that isn't being used by another thread, locks it, and dispatches the
+///  call to the solver.
+///  So long as there are at least as many solvers as there are hardware threads, it should never need to
+///  spin wait.
+///
+class btConstraintSolverPoolMt : public btConstraintSolver
+{
+public:
+    // create the solvers for me
+    explicit btConstraintSolverPoolMt( int numSolvers );
+
+    // pass in fully constructed solvers (destructor will delete them)
+    btConstraintSolverPoolMt( btConstraintSolver** solvers, int numSolvers );
+
+    virtual ~btConstraintSolverPoolMt();
+
+    ///solve a group of constraints
+    virtual btScalar solveGroup( btCollisionObject** bodies,
+        int numBodies,
+        btPersistentManifold** manifolds,
+        int numManifolds,
+        btTypedConstraint** constraints,
+        int numConstraints,
+        const btContactSolverInfo& info,
+        btIDebugDraw* debugDrawer,
+        btDispatcher* dispatcher
+    ) BT_OVERRIDE;
+
+    virtual	void reset() BT_OVERRIDE;
+    virtual btConstraintSolverType getSolverType() const BT_OVERRIDE { return m_solverType; }
+
+private:
+    const static size_t kCacheLineSize = 128;
+    struct ThreadSolver
+    {
+        btConstraintSolver* solver;
+        btSpinMutex mutex;
+        char _cachelinePadding[ kCacheLineSize - sizeof( btSpinMutex ) - sizeof( void* ) ];  // keep mutexes from sharing a cache line
+    };
+    btAlignedObjectArray<ThreadSolver> m_solvers;
+    btConstraintSolverType m_solverType;
+
+    ThreadSolver* getAndLockThreadSolver();
+    void init( btConstraintSolver** solvers, int numSolvers );
+};
+
+
+
 ///
 /// btDiscreteDynamicsWorldMt -- a version of DiscreteDynamicsWorld with some minor changes to support
 ///                              solving simulation islands on multiple threads.
 ///
+///  Should function exactly like btDiscreteDynamicsWorld.
+///  Also 3 methods that iterate over all of the rigidbodies can run in parallel:
+///     - predictUnconstraintMotion
+///     - integrateTransforms
+///     - createPredictiveContacts
+///
 ATTRIBUTE_ALIGNED16(class) btDiscreteDynamicsWorldMt : public btDiscreteDynamicsWorld
 {
 protected:
     InplaceSolverIslandCallbackMt* m_solverIslandCallbackMt;
 
-    virtual void	solveConstraints(btContactSolverInfo& solverInfo);
+    virtual void solveConstraints(btContactSolverInfo& solverInfo) BT_OVERRIDE;
+
+    virtual void predictUnconstraintMotion( btScalar timeStep ) BT_OVERRIDE;
+
+    struct UpdaterCreatePredictiveContacts : public btIParallelForBody
+    {
+        btScalar timeStep;
+        btRigidBody** rigidBodies;
+        btDiscreteDynamicsWorldMt* world;
+
+        void forLoop( int iBegin, int iEnd ) const BT_OVERRIDE
+        {
+            world->createPredictiveContactsInternal( &rigidBodies[ iBegin ], iEnd - iBegin, timeStep );
+        }
+    };
+    virtual void createPredictiveContacts( btScalar timeStep ) BT_OVERRIDE;
+
+    struct UpdaterIntegrateTransforms : public btIParallelForBody
+    {
+        btScalar timeStep;
+        btRigidBody** rigidBodies;
+        btDiscreteDynamicsWorldMt* world;
+
+        void forLoop( int iBegin, int iEnd ) const BT_OVERRIDE
+        {
+            world->integrateTransformsInternal( &rigidBodies[ iBegin ], iEnd - iBegin, timeStep );
+        }
+    };
+    virtual void integrateTransforms( btScalar timeStep ) BT_OVERRIDE;
 
 public:
 	BT_DECLARE_ALIGNED_ALLOCATOR();
 
-	btDiscreteDynamicsWorldMt(btDispatcher* dispatcher,btBroadphaseInterface* pairCache,btConstraintSolver* constraintSolver,btCollisionConfiguration* collisionConfiguration);
+	btDiscreteDynamicsWorldMt(btDispatcher* dispatcher,
+        btBroadphaseInterface* pairCache,
+        btConstraintSolverPoolMt* constraintSolver,   // Note this should be a solver-pool for multi-threading
+        btCollisionConfiguration* collisionConfiguration
+    );
 	virtual ~btDiscreteDynamicsWorldMt();
 };
 

src/BulletDynamics/Dynamics/btSimulationIslandManagerMt.cpp

@@ -15,6 +15,7 @@ subject to the following restrictions:
 
 
 #include "LinearMath/btScalar.h"
+#include "LinearMath/btThreads.h"
 #include "btSimulationIslandManagerMt.h"
 #include "BulletCollision/BroadphaseCollision/btDispatcher.h"
 #include "BulletCollision/NarrowPhaseCollision/btPersistentManifold.h"
@@ -44,7 +45,7 @@ btSimulationIslandManagerMt::btSimulationIslandManagerMt()
 {
     m_minimumSolverBatchSize = calcBatchCost(0, 128, 0);
     m_batchIslandMinBodyCount = 32;
-    m_islandDispatch = defaultIslandDispatch;
+    m_islandDispatch = parallelIslandDispatch;
     m_batchIsland = NULL;
 }
 
@@ -545,8 +546,9 @@ void btSimulationIslandManagerMt::mergeIslands()
 }
 
 
-void btSimulationIslandManagerMt::defaultIslandDispatch( btAlignedObjectArray<Island*>* islandsPtr, IslandCallback* callback )
+void btSimulationIslandManagerMt::serialIslandDispatch( btAlignedObjectArray<Island*>* islandsPtr, IslandCallback* callback )
 {
+    BT_PROFILE( "serialIslandDispatch" );
     // serial dispatch
     btAlignedObjectArray<Island*>& islands = *islandsPtr;
     for ( int i = 0; i < islands.size(); ++i )
@@ -565,6 +567,41 @@ void btSimulationIslandManagerMt::defaultIslandDispatch( btAlignedObjectArray<Is
     }
 }
 
+struct UpdateIslandDispatcher : public btIParallelForBody
+{
+    btAlignedObjectArray<btSimulationIslandManagerMt::Island*>* islandsPtr;
+    btSimulationIslandManagerMt::IslandCallback* callback;
+
+    void forLoop( int iBegin, int iEnd ) const BT_OVERRIDE
+    {
+        for ( int i = iBegin; i < iEnd; ++i )
+        {
+            btSimulationIslandManagerMt::Island* island = ( *islandsPtr )[ i ];
+            btPersistentManifold** manifolds = island->manifoldArray.size() ? &island->manifoldArray[ 0 ] : NULL;
+            btTypedConstraint** constraintsPtr = island->constraintArray.size() ? &island->constraintArray[ 0 ] : NULL;
+            callback->processIsland( &island->bodyArray[ 0 ],
+                island->bodyArray.size(),
+                manifolds,
+                island->manifoldArray.size(),
+                constraintsPtr,
+                island->constraintArray.size(),
+                island->id
+            );
+        }
+    }
+};
+
+void btSimulationIslandManagerMt::parallelIslandDispatch( btAlignedObjectArray<Island*>* islandsPtr, IslandCallback* callback )
+{
+    BT_PROFILE( "parallelIslandDispatch" );
+    int grainSize = 1;  // iterations per task
+    UpdateIslandDispatcher dispatcher;
+    dispatcher.islandsPtr = islandsPtr;
+    dispatcher.callback = callback;
+    btParallelFor( 0, islandsPtr->size(), grainSize, dispatcher );
+}
+
+
 ///@todo: this is random access, it can be walked 'cache friendly'!
 void btSimulationIslandManagerMt::buildAndProcessIslands( btDispatcher* dispatcher,
                                                         btCollisionWorld* collisionWorld,

src/BulletDynamics/Dynamics/btSimulationIslandManagerMt.h

@@ -59,7 +59,8 @@ public:
                                     ) = 0;
     };
     typedef void( *IslandDispatchFunc ) ( btAlignedObjectArray<Island*>* islands, IslandCallback* callback );
-    static void defaultIslandDispatch( btAlignedObjectArray<Island*>* islands, IslandCallback* callback );
+    static void serialIslandDispatch( btAlignedObjectArray<Island*>* islands, IslandCallback* callback );
+    static void parallelIslandDispatch( btAlignedObjectArray<Island*>* islandsPtr, IslandCallback* callback );
 protected:
     btAlignedObjectArray<Island*> m_allocatedIslands;  // owner of all Islands
     btAlignedObjectArray<Island*> m_activeIslands;  // islands actively in use