hkDemo::Result BroadphaseAddRemoveDemo::stepDemo()
{
#if defined(SELF_BENCHMARK)
LARGE_INTEGER counter0;
printf("Benchmarking...\r\n");
::QueryPerformanceCounter(&counter0);
for(int i=0;i<512;++i)
{
spawnFirework();
}
LARGE_INTEGER counter1;
LARGE_INTEGER frequency;
::QueryPerformanceCounter(&counter1);
::QueryPerformanceFrequency(&frequency);
printf("Time: %.I64d ms\r\n",((counter1.QuadPart-counter0.QuadPart)*1000)/frequency.QuadPart);
exit(1);
#endif
if(m_env->m_window->getKeyboard().wasKeyPressed('B'))
{
m_useBatch=!m_useBatch;
}
if(m_env->m_window->getKeyboard().wasKeyPressed('D'))
{
m_draw=!m_draw;
}
if(m_env->m_gamePad->wasButtonPressed(HKG_PAD_BUTTON_2)) m_useBatch=!m_useBatch;
if(m_env->m_gamePad->wasButtonPressed(HKG_PAD_BUTTON_3)) m_draw=!m_draw;
m_world->lock();
m_timer-=m_world->getSolverInfo()->m_deltaTime;
if(m_timer<0)
{
HK_TIMER_BEGIN_LIST( "DemoAddBodies", "Add" );
spawnFirework();
if(m_bodies.getSize()>=m_maxbodies) m_timer+=m_rate;
HK_TIMER_SPLIT_LIST( "Remove" );
flushFirework();
HK_TIMER_END_LIST();
}
m_world->unlock();
char text[256];
hkString::sprintf(text,"%d bodies [%d bodies per second]\r\n[B]atchs: %s\r\n[D]raw: %s",m_bodies.getSize(),(int)(m_count/m_rate),
m_useBatch?"ON":"OFF",
m_draw?"ON":"OFF");
m_env->m_textDisplay->outputText(text,8,(int)m_env->m_window->getHeight()-96);
return hkDefaultPhysicsDemo::stepDemo();
}
hkDemo::Result OptimizedWorldRaycastDemo::stepDemo()
{
m_world->lock();
m_time += m_timestep;
//
// Create a number of ray directions
//
const int NUM_DIRECTIONS = 8; const hkReal increment = 2.0f;
//const int NUM_DIRECTIONS = 125; const hkReal increment = 0.5f;
hkVector4 rayDirections[NUM_DIRECTIONS];
{
int i = 0;
for (hkReal x = -1.0f; x <= 1.0f; x += increment)
{
for (hkReal y = -1.0f; y <= 1.0f; y += increment)
{
for (hkReal z = -1.0f; z <= 1.0f; z+= increment)
{
rayDirections[i++].set(x,y,z);
}
}
}
}
hkReal angle = 0.5f * m_time;
// In these first chapter we simply call the hkpWorld castRay version with no further optimizations
{
angle += HK_REAL_PI * 0.3f;
hkReal xPos = m_rayLength * 3.0f * hkMath::sin(angle);
hkReal zPos = m_rayLength * 3.0f * hkMath::cos(angle);
hkpWorldRayCastInput inputs[NUM_DIRECTIONS];
hkpClosestRayHitCollector collectors[NUM_DIRECTIONS];
HK_TIMER_BEGIN("NotOptimized", HK_NULL);
{
for (int i = 0; i < NUM_DIRECTIONS; i++)
{
inputs[i].m_from.set( xPos, 0, zPos);
inputs[i].m_to.setAddMul4(inputs[i].m_from, rayDirections[i], 1.f*m_rayLength);
m_world->castRay(inputs[i], collectors[i]);
}
}
HK_TIMER_END();
displayHits( inputs, NUM_DIRECTIONS, collectors, hkColor::RED);
}
// This second set of raycasts (yellow) is using a the group functionality of the hkpWorldRayCaster.
// Internally it builds an the aabb cache. This is an optimization for broadphase raycasts.
// Note that we need to pass in array of collectors to the hkpWorldRayCaster.
// As the hkpWorldRayCaster has no clue about the size of each collector, we have to pass in the size.
// Interesting if we use a size of 0, than all raycast will report to the same collector.
{
angle += HK_REAL_PI * 0.3f;
hkReal xPos = m_rayLength * 3.0f * hkMath::sin(angle);
hkReal zPos = m_rayLength * 3.0f * hkMath::cos(angle);
hkpWorldRayCastInput inputs[NUM_DIRECTIONS];
hkpClosestRayHitCollector collectors[NUM_DIRECTIONS];
//
// Cast Rays
//
HK_TIMER_BEGIN("CastRayGroup", HK_NULL);
{
for (int i = 0; i < NUM_DIRECTIONS; i++)
{
inputs[i].m_from.set( xPos, 0, zPos);
inputs[i].m_to.setAddMul4(inputs[i].m_from, rayDirections[i], 1.f*m_rayLength);
}
hkpWorldRayCaster rayCaster;
rayCaster.castRayGroup( *m_world->getBroadPhase(), inputs, NUM_DIRECTIONS, m_world->getCollisionFilter(), collectors, sizeof(collectors[0]) );
}
HK_TIMER_END();
displayHits( inputs, NUM_DIRECTIONS, collectors, hkColor::CYAN);
}
// This set of raycasts (yellow) is using a the hkpBroadPhaseAabbCache.
// So by building the aabb cache. This is an optimization for broadphase raycasts.
// The idea is that the aabb cache is actually a reduced broadphase, just storing
// objects inside the aabb. Therefore using this cache can speed up rayCasts significantly.
// Unfortunately we cannot use our simple hkpWorld::castRay() function, but have to use
// a small helper class hkpWorldRayCaster (which is actually used by the hkpWorld::castRay()).
{
angle += HK_REAL_PI * 0.3f;
hkReal xPos = m_rayLength * 3.0f * hkMath::sin(angle);
hkReal zPos = m_rayLength * 3.0f * hkMath::cos(angle);
hkpWorldRayCastInput inputs[NUM_DIRECTIONS];
hkpClosestRayHitCollector collectors[NUM_DIRECTIONS];
//
// Calc Cache
//
HK_TIMER_BEGIN_LIST("UseAabbCache", "CalcCache");
hkpBroadPhaseAabbCache* cache;
int cacheSize;
{
hkAabb aabb;
aabb.m_min.set( xPos - m_rayLength, -m_rayLength, zPos-m_rayLength);
aabb.m_max.set( xPos + m_rayLength, +m_rayLength, zPos+m_rayLength);
cacheSize = m_world->getBroadPhase()->getAabbCacheSize();
cache = reinterpret_cast<hkpBroadPhaseAabbCache*>(hkAllocateStack<char>(cacheSize));
m_world->getBroadPhase()->calcAabbCache( aabb, cache );
}
//
// Cast Rays
//
HK_TIMER_SPLIT_LIST("CastRays");
{
for (int i = 0; i < NUM_DIRECTIONS; i++)
{
inputs[i].m_from.set( xPos, 0, zPos);
inputs[i].m_to.setAddMul4(inputs[i].m_from, rayDirections[i], 1.f*m_rayLength);
hkpWorldRayCaster rayCaster;
rayCaster.castRay( *m_world->getBroadPhase(), inputs[i], m_world->getCollisionFilter(), cache, collectors[i] );
}
}
HK_TIMER_END_LIST();
hkDeallocateStack<char>( (char*)cache);
displayHits( inputs, NUM_DIRECTIONS, collectors, hkColor::YELLOW);
}
// This third set of raycasts (blue) is using a the hkpBroadPhaseAabbCache and
// also makes use of the fact that many rays starting at the same position can
// be handled specially.
// Unfortunately we cannot use our simple hkpWorld::castRay() function, but have to use
// a small helper class hkpWorldRayCaster (which is actually used by the hkpWorld::castRay()).
{
angle += HK_REAL_PI * 0.3f;
hkReal xPos = m_rayLength * 3.0f * hkMath::sin(angle);
hkReal zPos = m_rayLength * 3.0f * hkMath::cos(angle);
hkpWorldRayCastInput inputs[NUM_DIRECTIONS];
hkpClosestRayHitCollector collectors[NUM_DIRECTIONS];
//
// Calc Cache
//
HK_TIMER_BEGIN_LIST("CacheSameStart", "CalcCache");
hkpBroadPhaseAabbCache* cache;
int cacheSize;
{
hkAabb aabb;
aabb.m_min.set( xPos - m_rayLength, -m_rayLength, zPos-m_rayLength);
aabb.m_max.set( xPos + m_rayLength, +m_rayLength, zPos+m_rayLength);
cacheSize = m_world->getBroadPhase()->getAabbCacheSize();
cache = reinterpret_cast<hkpBroadPhaseAabbCache*>(hkAllocateStack<char>(cacheSize));
m_world->getBroadPhase()->calcAabbCache( aabb, cache );
}
//
// Cast Rays
//
HK_TIMER_SPLIT_LIST("CastRays");
{
for (int i = 0; i < NUM_DIRECTIONS; i++)
{
inputs[i].m_from.set( xPos, 0, zPos);
inputs[i].m_to.setAddMul4(inputs[i].m_from, rayDirections[i], 1.f*m_rayLength);
}
hkpWorldRayCaster rayCaster;
rayCaster.castRaysFromSinglePoint( *m_world->getBroadPhase(), inputs, NUM_DIRECTIONS, m_world->getCollisionFilter(), cache, collectors, hkSizeOf( collectors[0] ) );
}
HK_TIMER_END_LIST();
hkDeallocateStack<char>( (char*)cache);
displayHits( inputs, NUM_DIRECTIONS, collectors, hkColor::BLUE);
}
// This 4th set of raycasts (green) is using a the AabbPhantom
// A phantom is like a persistent aabb cache, so it only makes sense if our aabb shows some
// framecoherency.
// Unfortunetaly the current phantom implementations simply casts the ray against all objects
// overlapping with the phantom. Therefor, if too many objects are intersecting the phantoms
// aabb, performance could get bad.
{
angle += HK_REAL_PI * 0.3f;
hkReal xPos = m_rayLength * 3.0f * hkMath::sin(angle);
hkReal zPos = m_rayLength * 3.0f * hkMath::cos(angle);
hkpWorldRayCastInput inputs[NUM_DIRECTIONS];
hkpClosestRayHitCollector collectors[NUM_DIRECTIONS];
HK_TIMER_BEGIN_LIST("AabbPhantom", "MovePhantom");
{
hkAabb aabb;
aabb.m_min.set( xPos - m_rayLength, -m_rayLength, zPos-m_rayLength);
aabb.m_max.set( xPos + m_rayLength, +m_rayLength, zPos+m_rayLength);
m_phantom->setAabb( aabb );
}
HK_TIMER_SPLIT_LIST("CastRays");
{
for (int i = 0; i < NUM_DIRECTIONS; i++)
{
inputs[i].m_from.set( xPos, 0, zPos);
inputs[i].m_to.setAddMul4(inputs[i].m_from, rayDirections[i], 1.f*m_rayLength);
m_phantom->castRay(inputs[i], collectors[i]);
}
}
HK_TIMER_END_LIST();
displayHits( inputs, NUM_DIRECTIONS, collectors, hkColor::GREEN);
}
// fastest (purple), combine all optimizations above
// We are using a phantom to set up the broadphase cache. This avoids the problem of the hkpPhantom::castRay()
{
angle += HK_REAL_PI * 0.3f;
hkReal xPos = m_rayLength * 3.0f * hkMath::sin(angle);
hkReal zPos = m_rayLength * 3.0f * hkMath::cos(angle);
hkpWorldRayCastInput inputs[NUM_DIRECTIONS];
hkpClosestRayHitCollector collectors[NUM_DIRECTIONS];
hkpBroadPhaseAabbCache* cache;
int cacheSize;
HK_TIMER_BEGIN_LIST("AllOpt.", "MovePhantomAndDoCache");
{
hkAabb aabb;
aabb.m_min.set( xPos - m_rayLength, -m_rayLength, zPos-m_rayLength);
aabb.m_max.set( xPos + m_rayLength, +m_rayLength, zPos+m_rayLength);
m_phantomUseCache->setAabb( aabb );
cacheSize = m_world->getBroadPhase()->getAabbCacheSize();
cache = reinterpret_cast<hkpBroadPhaseAabbCache*>(hkAllocateStack<char>(cacheSize));
m_world->getBroadPhase()->calcAabbCache( m_phantomUseCache->getOverlappingCollidables(), cache );
}
HK_TIMER_SPLIT_LIST("CastRays");
{
for (int i = 0; i < NUM_DIRECTIONS; i++)
{
inputs[i].m_from.set( xPos, 0, zPos);
inputs[i].m_to.setAddMul4(inputs[i].m_from, rayDirections[i], 1.f*m_rayLength);
}
hkpWorldRayCaster rayCaster;
rayCaster.castRaysFromSinglePoint( *m_world->getBroadPhase(), inputs, NUM_DIRECTIONS, m_world->getCollisionFilter(), cache, collectors, hkSizeOf( collectors[0] ) );
}
HK_TIMER_END_LIST();
hkDeallocateStack<char>( (char*)cache);
displayHits( inputs, NUM_DIRECTIONS, collectors, hkColor::PURPLE);
}
m_world->unlock();
return hkDefaultPhysicsDemo::stepDemo();
}
void AabbQueryDemo::queryAabbMT()
{
HK_TIMER_BEGIN_LIST("queryAabbMT", "setup");
hkAabb aabb;
// Grab a body from the world, and compute it's AABB + some padding
const hkpCollidable* queryCollidable;
{
queryCollidable = m_collidables[m_collIdx];
hkpRigidBody* rb = hkGetRigidBody(queryCollidable);
queryCollidable->getShape()->getAabb(rb->getTransform(), 20.0f, aabb);
}
hkArray<hkpKdTreeAabbCommand> commands;
commands.setSize(numQueries);
//
// For performance timings, we query the same AABB multiple times and overwrite the results.
// When using this, make sure to use different output arrays!
//
hkArray<hkPrimitiveId> output;
output.setSize(50);
//
// Set up the commands for the job
//
for (int i=0; i<commands.getSize(); i++)
{
hkpKdTreeAabbCommand& command = commands[i];
command.m_aabb = aabb;
command.m_results = output.begin();
command.m_resultsCapacity = output.getSize();
command.m_numResultsOut = 0;
}
//
// Setup the job
//
hkArray<hkpRayCastQueryJobHeader> header(1);
hkpKdTreeAabbJob aabbJob(header.begin(), commands.begin(), commands.getSize(), &m_semaphore);
aabbJob.m_numTrees = 1;
aabbJob.m_trees[0] = m_world->m_kdTreeManager->getTree();
m_jobQueue->addJob( *reinterpret_cast<hkJobQueue::JobQueueEntry*>(&aabbJob), hkJobQueue::JOB_HIGH_PRIORITY );
HK_TIMER_SPLIT_LIST("process");
m_jobThreadPool->processAllJobs( m_jobQueue );
m_jobThreadPool->waitForCompletion();
m_semaphore.acquire();
HK_TIMER_END_LIST();
//
// Run the same query on the broadphase for comparison purposes
//
hkArray<hkpBroadPhaseHandlePair> sapHits;
{
HK_TIME_CODE_BLOCK("BroadphaseQueryAabb", HK_NULL);
for (int i=0; i<numQueries; i++)
{
sapHits.clear();
m_world->getBroadPhase()->querySingleAabb( aabb, sapHits );
}
}
//
// Check results and draw
//
for (int i=0; i<commands.getSize(); i++)
{
hkpKdTreeAabbCommand& command = commands[i];
hkBool jobOk = compareHitArrays(command.m_aabb, command.m_results, command.m_numResultsOut, sapHits);
for (int j=0; j<command.m_numResultsOut; j++)
{
HK_SET_OBJECT_COLOR(command.m_results[j], hkColor::YELLOW);
}
HK_SET_OBJECT_COLOR((hkUlong)queryCollidable, hkColor::LIME);
if( !jobOk )
{
m_env->m_textDisplay->outputText("MT Hit lits differed!", 20, 250, (hkUint32) hkColor::RED);
}
}
}