void ROITracker::updateDelay( const PixelViewports& pvps,
const uint8_t* ticket )
{
LBASSERT( _needsUpdate );
LBASSERTINFO( ticket == _ticket, "Wrong ticket" );
if( ticket != _ticket )
{
LBERROR << "Wrong ticket" << std::endl;
return;
}
uint32_t totalAreaFound = 0;
for( uint32_t i = 0; i < pvps.size(); i++ )
totalAreaFound += pvps[ i ].getArea();
Area& area = (*_curFrame)[ _lastStage ].areas.back();
if( totalAreaFound < area.pvp.getArea()*4/5 )
{
// ROI cutted enough, reset failure statistics
area.lastSkip = 0;
}else
{
// disable ROI for next frames, if it was failing before,
// increase number of frames to skip
area.lastSkip = LB_MIN( area.lastSkip*2 + 1, 64 );
area.skip = area.lastSkip;
}
_needsUpdate = false;
}
//----------------------------------------------------------------------
// read
//----------------------------------------------------------------------
void SocketConnection::readNB( void* buffer, const uint64_t bytes )
{
if( isClosed() )
return;
WSABUF wsaBuffer = { LB_MIN( bytes, 65535 ),
reinterpret_cast< char* >( buffer ) };
DWORD flags = 0;
ResetEvent( _overlappedRead.hEvent );
_overlappedDone = 0;
const int result = WSARecv( _readFD, &wsaBuffer, 1, &_overlappedDone,
&flags, &_overlappedRead, 0 );
if( result == 0 ) // got data already
{
if( _overlappedDone == 0 ) // socket closed
{
LBDEBUG << "Got EOF, closing connection" << std::endl;
close();
}
SetEvent( _overlappedRead.hEvent );
}
else if( GetLastError() != WSA_IO_PENDING )
{
LBWARN << "Could not start overlapped receive: " << lunchbox::sysError
<< ", closing connection" << std::endl;
close();
}
}
void FramerateEqualizer::_init()
{
const Compound* compound = getCompound();
if( _nSamples > 0 || !compound )
return;
_nSamples = 1;
// Subscribe to child channel load events
const Compounds& children = compound->getChildren();
LBASSERT( _loadListeners.empty( ));
_loadListeners.resize( children.size( ));
for( size_t i = 0; i < children.size(); ++i )
{
Compound* child = children[i];
const uint32_t period = child->getInheritPeriod();
LoadListener& loadListener = _loadListeners[i];
loadListener.parent = this;
loadListener.period = period;
LoadSubscriber subscriber( &loadListener );
child->accept( subscriber );
_nSamples = LB_MAX( _nSamples, period );
}
_nSamples = LB_MIN( _nSamples, 100 );
}
void FrameData::adjustQuality( const float delta )
{
_quality += delta;
_quality = LB_MAX( _quality, 0.1f );
_quality = LB_MIN( _quality, 1.0f );
setDirty( DIRTY_FLAGS );
LBINFO << "Set non-idle image quality to " << _quality << std::endl;
}
void TreeEqualizer::_notifyLoadData( Node* node, Channel* channel,
const uint32_t nStatistics,
const Statistic* statistics )
{
if( !node )
return;
_notifyLoadData( node->left, channel, nStatistics, statistics );
_notifyLoadData( node->right, channel, nStatistics, statistics );
if( !node->compound || node->compound->getChannel() != channel )
return;
// gather relevant load data
const uint32_t taskID = node->compound->getTaskID();
int64_t startTime = std::numeric_limits< int64_t >::max();
int64_t endTime = 0;
bool loadSet = false;
int64_t timeTransmit = 0;
for( uint32_t i = 0; i < nStatistics && !loadSet; ++i )
{
const Statistic& stat = statistics[ i ];
if( stat.task != taskID ) // from different compound
continue;
switch( stat.type )
{
case Statistic::CHANNEL_CLEAR:
case Statistic::CHANNEL_DRAW:
case Statistic::CHANNEL_READBACK:
startTime = LB_MIN( startTime, stat.startTime );
endTime = LB_MAX( endTime, stat.endTime );
break;
case Statistic::CHANNEL_ASYNC_READBACK:
case Statistic::CHANNEL_FRAME_TRANSMIT:
timeTransmit += stat.endTime - stat.startTime;
break;
// assemble blocks on input frames, stop using subsequent data
case Statistic::CHANNEL_ASSEMBLE:
loadSet = true;
break;
default:
break;
}
}
if( startTime == std::numeric_limits< int64_t >::max( ))
return;
node->time = endTime - startTime;
node->time = LB_MAX( node->time, 1 );
node->time = LB_MAX( node->time, timeTransmit );
}
uint128_t computeMinMax() const
{
uint64_t xMax = 0;
uint64_t xMin = std::numeric_limits<uint64_t>::max();
for (ItemsCIter i = items.begin(); i != items.end(); ++i)
{
const Item& item = *i;
xMin = LB_MIN(xMin, item.start);
xMax = LB_MAX(xMax, item.end);
}
return uint128_t(xMax, xMin);
}
int64_t NamedPipeConnection::write( const void* buffer, const uint64_t bytes )
{
if( !isConnected() || _fd == INVALID_HANDLE_VALUE )
return -1;
DWORD wrote;
const DWORD use = LB_MIN( bytes, CO_WRITE_BUFFER_SIZE );
ResetEvent( _write.hEvent );
if( WriteFile( _fd, buffer, use, &wrote, &_write ))
return wrote;
if( GetLastError() != ERROR_IO_PENDING )
{
LBWARN << "Could not start write: " << lunchbox::sysError << std::endl;
return -1;
}
DWORD got = 0;
if( GetOverlappedResult( _fd, &_write, &got, false ))
return got;
switch( GetLastError( ))
{
case ERROR_PIPE_CONNECTED:
return 0;
case ERROR_IO_PENDING:
case ERROR_IO_INCOMPLETE:
{
if( WAIT_OBJECT_0 != WaitForSingleObject( _write.hEvent, INFINITE ))
throw Exception( Exception::TIMEOUT_WRITE );
break;
}
default:
LBWARN << "Write complete failed: " << lunchbox::sysError << std::endl;
}
if( GetOverlappedResult( _fd, &_write, &got, false ))
return got;
if( GetLastError() == ERROR_PIPE_CONNECTED )
return 0;
LBWARN << "Write complete failed: " << lunchbox::sysError << std::endl;
return -1;
}
void Channel::_updateNearFar( const mesh::BoundingSphere& boundingSphere )
{
// compute dynamic near/far plane of whole model
const FrameData& frameData = _getFrameData();
const eq::Matrix4f& rotation = frameData.getCameraRotation();
const eq::Matrix4f headTransform = getHeadTransform() * rotation;
eq::Matrix4f modelInv;
compute_inverse( headTransform, modelInv );
const eq::Vector3f zero = modelInv * eq::Vector3f::ZERO;
eq::Vector3f front = modelInv * eq::Vector3f( 0.0f, 0.0f, -1.0f );
front -= zero;
front.normalize();
front *= boundingSphere.w();
const eq::Vector3f center =
frameData.getCameraPosition().get_sub_vector< 3 >() -
boundingSphere.get_sub_vector< 3 >();
const eq::Vector3f nearPoint = headTransform * ( center - front );
const eq::Vector3f farPoint = headTransform * ( center + front );
if( useOrtho( ))
{
LBASSERTINFO( fabs( farPoint.z() - nearPoint.z() ) >
std::numeric_limits< float >::epsilon(),
nearPoint << " == " << farPoint );
setNearFar( -nearPoint.z(), -farPoint.z() );
}
else
{
// estimate minimal value of near plane based on frustum size
const eq::Frustumf& frustum = getFrustum();
const float width = fabs( frustum.right() - frustum.left() );
const float height = fabs( frustum.top() - frustum.bottom() );
const float size = LB_MIN( width, height );
const float minNear = frustum.near_plane() / size * .001f;
const float zNear = LB_MAX( minNear, -nearPoint.z() );
const float zFar = LB_MAX( zNear * 2.f, -farPoint.z() );
setNearFar( zNear, zFar );
}
}
template< class T > void _test()
{
T* lock = new T;
lock->set();
#ifdef LUNCHBOX_USE_OPENMP
const size_t nThreads = LB_MIN( lunchbox::OMP::getNThreads() * 3,
MAXTHREADS );
#else
const size_t nThreads = 16;
#endif
Thread< T > threads[MAXTHREADS];
for( size_t i = 1; i <= nThreads; i = i << 1 )
{
_running = true;
for( size_t j = 0; j < i; ++j )
{
threads[j].lock = lock;
TEST( threads[j].start( ));
}
lunchbox::sleep( 10 ); // let threads initialize
_clock.reset();
lock->unset();
lunchbox::sleep( TIME ); // let threads run
_running = false;
for( size_t j = 0; j < i; ++j )
TEST( threads[j].join( ));
const float time = _clock.getTimef();
TEST( !lock->isSet( ));
lock->set();
size_t ops = 0;
for( size_t j = 0; j < nThreads; ++j )
ops += threads[j].ops;
std::cout << std::setw(20) << lunchbox::className( lock ) << ", "
<< std::setw(12) << /*set, test, unset*/ 3 * ops / time
<< ", " << std::setw(3) << i << std::endl;
}
delete lock;
}
void FramerateEqualizer::LoadListener::notifyLoadData(
Channel* channel, const uint32_t frameNumber, const Statistics& statistics,
const Viewport& region )
{
// gather required load data
int64_t startTime = std::numeric_limits< int64_t >::max();
int64_t endTime = 0;
for( size_t i = 0; i < statistics.size(); ++i )
{
const eq::Statistic& data = statistics[i];
switch( data.type )
{
case eq::Statistic::CHANNEL_CLEAR:
case eq::Statistic::CHANNEL_DRAW:
case eq::Statistic::CHANNEL_ASSEMBLE:
case eq::Statistic::CHANNEL_READBACK:
startTime = LB_MIN( startTime, data.startTime );
endTime = LB_MAX( endTime, data.endTime );
break;
default:
break;
}
}
if( startTime == std::numeric_limits< int64_t >::max( ))
return;
if( startTime == endTime ) // very fast draws might report 0 times
++endTime;
for( std::deque< FrameTime >::iterator i = parent->_times.begin();
i != parent->_times.end(); ++i )
{
FrameTime& frameTime = *i;
if( frameTime.first != frameNumber )
continue;
const float time = static_cast< float >( endTime - startTime ) / period;
frameTime.second = LB_MAX( frameTime.second, time );
LBLOG( LOG_LB2 ) << "Frame " << frameNumber << " channel "
<< channel->getName() << " time " << time
<< " period " << period << std::endl;
}
}
void LoadEqualizer::_updateLeaf( Node* node )
{
const Compound* compound = node->compound;
const Channel* channel = compound->getChannel();
LBASSERT( channel );
const PixelViewport& pvp = channel->getPixelViewport();
node->resources = compound->isActive() ? compound->getUsage() : 0.f;
LBLOG( LOG_LB2 ) << channel->getName() << " active " << compound->isActive()
<< " using " << node->resources << std::endl;
LBASSERT( node->resources >= 0.f );
node->maxSize.x() = pvp.w;
node->maxSize.y() = pvp.h;
node->boundaryf = getBoundaryf();
node->boundary2i = getBoundary2i();
node->resistancef = getResistancef();
node->resistance2i = getResistance2i();
if( !compound->hasDestinationChannel( ))
return;
const float nResources = _getTotalResources();
if( getAssembleOnlyLimit() <= nResources - node->resources )
{
node->resources = 0.f;
return; // OPT
}
const float time = float( _getTotalTime( ));
const float assembleTime = float( _getAssembleTime( ));
if( assembleTime == 0.f || node->resources == 0.f )
return;
const float timePerResource = time / ( nResources - node->resources );
const float renderTime = timePerResource * node->resources ;
const float clampedAssembleTime = LB_MIN( assembleTime, renderTime );
const float newTimePerResource = (time + clampedAssembleTime) / nResources;
node->resources -= ( clampedAssembleTime / newTimePerResource );
if( node->resources < 0.f ) // may happen due to fp rounding
node->resources = 0.f;
}
//----------------------------------------------------------------------
// read
//----------------------------------------------------------------------
void NamedPipeConnection::readNB( void* buffer, const uint64_t bytes )
{
if( isClosed( ))
return;
ResetEvent( _read.hEvent );
DWORD use = LB_MIN( bytes, CO_READ_BUFFER_SIZE );
if( ReadFile( _fd, buffer, use, &_readDone, &_read ) )
{
LBASSERT( _readDone > 0 );
SetEvent( _read.hEvent );
}
else if( GetLastError() != ERROR_IO_PENDING )
{
LBWARN << "Could not start overlapped receive: " << lunchbox::sysError
<< ", closing connection" << std::endl;
close();
}
}
void Renderer::updateNearFar( const Vector4f& boundingSphere )
{
const Matrix4f& view = getViewMatrix();
Matrix4f viewInv;
compute_inverse( view, viewInv );
const Vector3f& zero = viewInv * Vector3f::ZERO;
Vector3f front = viewInv * Vector3f( 0.0f, 0.0f, -1.0f );
front -= zero;
front.normalize();
front *= boundingSphere.w();
const Vector3f& translation = getModelMatrix().get_translation();
const Vector3f& center = translation - boundingSphere.get_sub_vector< 3 >();
const Vector3f& nearPoint = view * ( center - front );
const Vector3f& farPoint = view * ( center + front );
if( _impl->useOrtho( ))
{
LBASSERTINFO( fabs( farPoint.z() - nearPoint.z() ) >
std::numeric_limits< float >::epsilon(),
nearPoint << " == " << farPoint );
setNearFar( -nearPoint.z(), -farPoint.z() );
}
else
{
// estimate minimal value of near plane based on frustum size
const eq::Frustumf& frustum = _impl->getFrustum();
const float width = fabs( frustum.right() - frustum.left() );
const float height = fabs( frustum.top() - frustum.bottom() );
const float size = LB_MIN( width, height );
const float minNear = frustum.near_plane() / size * .001f;
const float zNear = LB_MAX( minNear, -nearPoint.z() );
const float zFar = LB_MAX( zNear * 2.f, -farPoint.z() );
setNearFar( zNear, zFar );
}
}
void ROIFinder::_readbackInfo( util::ObjectManager& glObjects )
{
LBASSERT( glObjects.supportsEqTexture( ));
LBASSERT( glObjects.supportsEqFrameBufferObject( ));
PixelViewport pvp = _pvp;
pvp.apply( Zoom( GRID_SIZE, GRID_SIZE ));
pvp.w = LB_MIN( pvp.w+pvp.x, _pvpOriginal.w+_pvpOriginal.x ) - pvp.x;
pvp.h = LB_MIN( pvp.h+pvp.y, _pvpOriginal.h+_pvpOriginal.y ) - pvp.y;
LBASSERT( pvp.isValid());
// copy frame buffer to texture
const void* bufferKey = _getInfoKey( );
util::Texture* texture =
glObjects.obtainEqTexture( bufferKey, GL_TEXTURE_RECTANGLE_ARB );
#ifdef EQ_ROI_USE_DEPTH_TEXTURE
texture->copyFromFrameBuffer( GL_DEPTH_COMPONENT, pvp );
#else
texture->copyFromFrameBuffer( GL_RGBA, pvp );
#endif
// draw zoomed quad into FBO
const void* fboKey = _getInfoKey( );
util::FrameBufferObject* fbo = glObjects.getEqFrameBufferObject( fboKey );
if( fbo )
{
LBCHECK( fbo->resize( _pvp.w, _pvp.h ));
}
else
{
fbo = glObjects.newEqFrameBufferObject( fboKey );
LBCHECK( fbo->init( _pvp.w, _pvp.h, GL_RGBA32F, 0, 0 ));
}
fbo->bind();
texture->bind();
// Enable & download depth texture
glEnable( GL_TEXTURE_RECTANGLE_ARB );
texture->applyWrap();
texture->applyZoomFilter( FILTER_LINEAR );
// Enable shaders
GLuint program = glObjects.getProgram( shaderRBInfo );
if( program == util::ObjectManager::INVALID )
{
// Create fragment shader which reads depth values from
// rectangular textures
const GLuint shader = glObjects.newShader( shaderRBInfo,
GL_FRAGMENT_SHADER );
LBASSERT( shader != util::ObjectManager::INVALID );
#ifdef EQ_ROI_USE_DEPTH_TEXTURE
const GLchar* fShaderPtr = roiFragmentShader_glsl.c_str();
#else
const GLchar* fShaderPtr = roiFragmentShaderRGB_glsl.c_str();
#endif
EQ_GL_CALL( glShaderSource( shader, 1, &fShaderPtr, 0 ));
EQ_GL_CALL( glCompileShader( shader ));
GLint status;
glGetShaderiv( shader, GL_COMPILE_STATUS, &status );
if( !status )
LBERROR << "Failed to compile fragment shader for ROI finder"
<< std::endl;
program = glObjects.newProgram( shaderRBInfo );
EQ_GL_CALL( glAttachShader( program, shader ));
EQ_GL_CALL( glLinkProgram( program ));
glGetProgramiv( program, GL_LINK_STATUS, &status );
if( !status )
{
LBWARN << "Failed to link shader program for ROI finder"
<< std::endl;
return;
}
// use fragment shader and setup uniforms
EQ_GL_CALL( glUseProgram( program ));
GLint param = glGetUniformLocation( program, "texture" );
glUniform1i( param, 0 );
}
else
{
// use fragment shader
EQ_GL_CALL( glUseProgram( program ));
}
// Draw Quad
glDisable( GL_LIGHTING );
glColor3f( 1.0f, 1.0f, 1.0f );
glBegin( GL_QUADS );
glVertex3i( 0, 0, 0 );
glVertex3i( _pvp.w, 0, 0 );
glVertex3i( _pvp.w, _pvp.h, 0 );
glVertex3i( 0, _pvp.h, 0 );
glEnd();
// restore state
glDisable( GL_TEXTURE_RECTANGLE_ARB );
EQ_GL_CALL( glUseProgram( 0 ));
fbo->unbind();
// finish readback of info
LBASSERT( static_cast<int32_t>(_perBlockInfo.size()) >= _pvp.w*_pvp.h*4 );
texture = fbo->getColorTextures()[0];
LBASSERT( texture->getFormat() == GL_RGBA );
LBASSERT( texture->getType() == GL_FLOAT );
texture->download( &_perBlockInfo[0] );
}
template< class T, uint32_t hold > void _test()
{
T* lock = new T;
lock->set();
#ifdef LUNCHBOX_USE_OPENMP
const size_t nThreads = LB_MIN( lunchbox::OMP::getNThreads()*3, MAXTHREADS );
#else
const size_t nThreads = 16;
#endif
WriteThread< T, hold > writers[MAXTHREADS];
ReadThread< T, hold > readers[MAXTHREADS];
std::cout << " Class, write ops/ms, read ops/ms, w threads, "
<< "r threads" << std::endl;
for( size_t nWrite = 0; nWrite <= nThreads;
nWrite = (nWrite == 0) ? 1 : nWrite << 1 )
{
for( size_t i = 1; i <= nThreads; i = i << 1 )
{
if( i < nWrite )
continue;
const size_t nRead = i - nWrite;
_running = true;
for( size_t j = 0; j < nWrite; ++j )
{
writers[j].lock = lock;
TEST( writers[j].start( ));
}
for( size_t j = 0; j < nRead; ++j )
{
readers[j].lock = lock;
TESTINFO( readers[j].start(), j );
}
lunchbox::sleep( 10 ); // let threads initialize
_clock.reset();
lock->unset();
lunchbox::sleep( TIME ); // let threads run
_running = false;
for( size_t j = 0; j < nWrite; ++j )
TEST( writers[j].join( ));
for( size_t j = 0; j < nRead; ++j )
TEST( readers[j].join( ));
const double time = _clock.getTimed();
TEST( !lock->isSet( ));
lock->set();
size_t nWriteOps = 0;
double wTime = time * double( nWrite );
for( size_t j = 0; j < nWrite; ++j )
{
nWriteOps += writers[j].ops;
wTime -= writers[j].sTime;
}
if( nWrite > 0 )
wTime /= double( nWrite );
if( wTime == 0.f )
wTime = std::numeric_limits< double >::epsilon();
size_t nReadOps = 0;
double rTime = time * double( nRead );
for( size_t j = 0; j < nRead; ++j )
{
nReadOps += readers[j].ops;
rTime -= readers[j].sTime;
}
if( nRead > 0 )
rTime /= double( nRead );
if( rTime == 0.f )
rTime = std::numeric_limits< double >::epsilon();
std::cout << std::setw(20)<< lunchbox::className( lock ) << ", "
<< std::setw(12) << 3 * nWriteOps / wTime << ", "
<< std::setw(12) << 3 * nReadOps / rTime << ", "
<< std::setw(9) << nWrite << ", " << std::setw(9) << nRead
<< std::endl;
}
}
delete lock;
}
void ViewEqualizer::Listener::notifyLoadData(Channel* channel,
const uint32_t frameNumber,
const Statistics& statistics,
const Viewport& /*region*/)
{
Load& load = _getLoad(frameNumber);
if (load == Load::NONE)
return;
LBASSERT(_taskIDs.find(channel) != _taskIDs.end());
const uint32_t taskID = _taskIDs[channel];
// gather relevant load data
int64_t startTime = std::numeric_limits<int64_t>::max();
int64_t endTime = 0;
bool loadSet = false;
int64_t transmitTime = 0;
for (size_t i = 0; i < statistics.size() && !loadSet; ++i)
{
const Statistic& data = statistics[i];
if (data.task != taskID) // data from another compound
continue;
switch (data.type)
{
case Statistic::CHANNEL_CLEAR:
case Statistic::CHANNEL_DRAW:
case Statistic::CHANNEL_READBACK:
startTime = LB_MIN(startTime, data.startTime);
endTime = LB_MAX(endTime, data.endTime);
break;
case Statistic::CHANNEL_ASYNC_READBACK:
case Statistic::CHANNEL_FRAME_TRANSMIT:
transmitTime += data.startTime - data.endTime;
break;
case Statistic::CHANNEL_FRAME_WAIT_SENDTOKEN:
transmitTime -= data.endTime - data.startTime;
break;
// assemble blocks on input frames, stop using subsequent data
case Statistic::CHANNEL_ASSEMBLE:
loadSet = true;
break;
default:
break;
}
}
if (startTime == std::numeric_limits<int64_t>::max())
return;
LBASSERTINFO(load.missing > 0, load << " for " << channel->getName() << " "
<< channel->getSerial());
const int64_t time = LB_MAX(endTime - startTime, transmitTime);
load.time += time;
--load.missing;
if (load.missing == 0)
{
const float rTime = float(load.time) / float(load.nResources);
load.time = int64_t(rTime * sqrtf(float(load.nResources)));
}
LBLOG(LOG_LB1) << "Task " << taskID << ", added time " << time << " to "
<< load << " from " << channel->getName() << " "
<< channel->getSerial() << std::endl;
}
void LoadEqualizer::_computeSplit( Node* node, const float time,
LBDatas* datas, const Viewport& vp,
const Range& range )
{
LBLOG( LOG_LB2 ) << "_computeSplit " << vp << ", " << range << " time "
<< time << std::endl;
LBASSERTINFO( vp.isValid(), vp );
LBASSERTINFO( range.isValid(), range );
LBASSERTINFO( node->resources > 0.f || !vp.hasArea() || !range.hasData(),
"Assigning " << node->resources <<
" work to viewport " << vp << ", " << range );
Compound* compound = node->compound;
if( compound )
{
_assign( compound, vp, range );
return;
}
LBASSERT( node->left && node->right );
LBDatas workingSet = datas[ node->mode ];
const float leftTime = node->resources > 0 ?
time * node->left->resources / node->resources : 0.f;
float timeLeft = LB_MIN( leftTime, time ); // correct for fp rounding error
switch( node->mode )
{
case MODE_VERTICAL:
{
LBASSERT( range == Range::ALL );
float splitPos = vp.x;
const float end = vp.getXEnd();
while( timeLeft > std::numeric_limits< float >::epsilon() &&
splitPos < end )
{
LBLOG( LOG_LB2 ) << timeLeft << "ms left using "
<< workingSet.size() << " tiles" << std::endl;
// remove all irrelevant items from working set
for( LBDatas::iterator i = workingSet.begin();
i != workingSet.end(); )
{
const Data& data = *i;
if( data.vp.getXEnd() > splitPos )
++i;
else
i = workingSet.erase( i );
}
if( workingSet.empty( ))
break;
// find next 'discontinouity' in loads
float currentPos = 1.0f;
for( LBDatas::const_iterator i = workingSet.begin();
i != workingSet.end(); ++i )
{
const Data& data = *i;
if( data.vp.x > splitPos && data.vp.x < currentPos )
currentPos = data.vp.x;
const float xEnd = data.vp.getXEnd();
if( xEnd > splitPos && xEnd < currentPos )
currentPos = xEnd;
}
const float width = currentPos - splitPos;
LBASSERTINFO( width > 0.f, currentPos << "<=" << splitPos );
LBASSERT( currentPos <= 1.0f );
// accumulate normalized load in splitPos...currentPos
LBLOG( LOG_LB2 ) << "Computing load in X " << splitPos << "..."
<< currentPos << std::endl;
float currentTime = 0.f;
for( LBDatas::const_iterator i = workingSet.begin();
i != workingSet.end(); ++i )
{
const Data& data = *i;
if( data.vp.x >= currentPos ) // not yet needed data sets
break;
float yContrib = data.vp.h;
if( data.vp.y < vp.y )
yContrib -= (vp.y - data.vp.y);
const float dataEnd = data.vp.getYEnd();
const float vpEnd = vp.getYEnd();
if( dataEnd > vpEnd )
yContrib -= (dataEnd - vpEnd);
if( yContrib > 0.f )
{
const float percentage = ( width / data.vp.w ) *
( yContrib / data.vp.h );
currentTime += ( data.time * percentage );
LBLOG( LOG_LB2 ) << data.vp << " contributes "
<< yContrib << " in " << vp.h << " ("
<< percentage << ") with " << data.time
<< ": " << ( data.time * percentage )
<< " vp.y " << vp.y << " dataEnd "
<< dataEnd << " vpEnd " << vpEnd
<< std::endl;
LBASSERT( percentage < 1.01f )
}
}
LBLOG( LOG_LB2 ) << splitPos << "..." << currentPos << ": t="
<< currentTime << " of " << timeLeft
<< std::endl;
if( currentTime >= timeLeft ) // found last region
{
splitPos += ( width * timeLeft / currentTime );
timeLeft = 0.0f;
}
else
{
timeLeft -= currentTime;
splitPos = currentPos;
}
}
LBLOG( LOG_LB2 ) << "Should split at X " << splitPos << std::endl;
if( getDamping() < 1.f )
splitPos = (1.f - getDamping()) * splitPos +
getDamping() * node->split;
LBLOG( LOG_LB2 ) << "Dampened split at X " << splitPos << std::endl;
// There might be more time left due to MIN_PIXEL rounding by parent
// LBASSERTINFO( timeLeft <= .001f, timeLeft );
// Ensure minimum size
const Compound* root = getCompound();
const float pvpW = static_cast< float >(
root->getInheritPixelViewport().w );
const float boundary = static_cast< float >( node->boundary2i.x()) /
pvpW;
if( node->left->resources == 0.f )
splitPos = vp.x;
else if( node->right->resources == 0.f )
splitPos = end;
else if( boundary > 0 )
{
const float lengthRight = vp.getXEnd() - splitPos;
const float lengthLeft = splitPos - vp.x;
const float maxRight =
static_cast< float >( node->right->maxSize.x( )) / pvpW;
const float maxLeft =
static_cast< float >( node->left->maxSize.x( )) / pvpW;
if( lengthRight > maxRight )
splitPos = end - maxRight;
else if( lengthLeft > maxLeft )
splitPos = vp.x + maxLeft;
if( (splitPos - vp.x) < boundary )
splitPos = vp.x + boundary;
if( (end - splitPos) < boundary )
splitPos = end - boundary;
const uint32_t ratio =
static_cast< uint32_t >( splitPos / boundary + .5f );
splitPos = ratio * boundary;
}
splitPos = LB_MAX( splitPos, vp.x );
splitPos = LB_MIN( splitPos, end);
const float newPixelW = pvpW * splitPos;
const float oldPixelW = pvpW * node->split;
if( int( fabs(newPixelW - oldPixelW) ) < node->resistance2i.x( ))
splitPos = node->split;
else
node->split = splitPos;
LBLOG( LOG_LB2 ) << "Constrained split " << vp << " at X "
<< splitPos << std::endl;
// balance children
Viewport childVP = vp;
childVP.w = (splitPos - vp.x);
_computeSplit( node->left, leftTime, datas, childVP, range );
childVP.x = childVP.getXEnd();
childVP.w = end - childVP.x;
// Fix 2994111: Rounding errors with 2D LB and 16 sources
// Floating point rounding may create a width for the 'right'
// child which is slightly below the parent width. Correct it.
while( childVP.getXEnd() < end )
childVP.w += std::numeric_limits< float >::epsilon();
_computeSplit( node->right, time-leftTime, datas, childVP, range );
break;
}
case MODE_HORIZONTAL:
{
LBASSERT( range == Range::ALL );
float splitPos = vp.y;
const float end = vp.getYEnd();
while( timeLeft > std::numeric_limits< float >::epsilon() &&
splitPos < end )
{
LBLOG( LOG_LB2 ) << timeLeft << "ms left using "
<< workingSet.size() << " tiles" << std::endl;
// remove all unrelevant items from working set
for( LBDatas::iterator i = workingSet.begin();
i != workingSet.end(); )
{
const Data& data = *i;
if( data.vp.getYEnd() > splitPos )
++i;
else
i = workingSet.erase( i );
}
if( workingSet.empty( ))
break;
// find next 'discontinuouity' in loads
float currentPos = 1.0f;
for( LBDatas::const_iterator i = workingSet.begin();
i != workingSet.end(); ++i )
{
const Data& data = *i;
if( data.vp.y > splitPos && data.vp.y < currentPos )
currentPos = data.vp.y;
const float yEnd = data.vp.getYEnd();
if( yEnd > splitPos && yEnd < currentPos )
currentPos = yEnd;
}
const float height = currentPos - splitPos;
LBASSERTINFO( height > 0.f, currentPos << "<=" << splitPos );
LBASSERT( currentPos <= 1.0f );
// accumulate normalized load in splitPos...currentPos
LBLOG( LOG_LB2 ) << "Computing load in Y " << splitPos << "..."
<< currentPos << std::endl;
float currentTime = 0.f;
for( LBDatas::const_iterator i = workingSet.begin();
i != workingSet.end(); ++i )
{
const Data& data = *i;
if( data.vp.y >= currentPos ) // not yet needed data sets
break;
float xContrib = data.vp.w;
if( data.vp.x < vp.x )
xContrib -= (vp.x - data.vp.x);
const float dataEnd = data.vp.getXEnd();
const float vpEnd = vp.getXEnd();
if( dataEnd > vpEnd )
xContrib -= (dataEnd - vpEnd);
if( xContrib > 0.f )
{
const float percentage = ( height / data.vp.h ) *
( xContrib / data.vp.w );
currentTime += ( data.time * percentage );
LBLOG( LOG_LB2 ) << data.vp << " contributes "
<< xContrib << " in " << vp.w << " ("
<< percentage << ") with " << data.time
<< ": " << ( data.time * percentage )
<< " total " << currentTime << " vp.x "
<< vp.x << " dataEnd " << dataEnd
<< " vpEnd " << vpEnd << std::endl;
LBASSERT( percentage < 1.01f )
}
}
LBLOG( LOG_LB2 ) << splitPos << "..." << currentPos << ": t="
<< currentTime << " of " << timeLeft
<< std::endl;
if( currentTime >= timeLeft ) // found last region
{
splitPos += (height * timeLeft / currentTime );
timeLeft = 0.0f;
}
else
{
timeLeft -= currentTime;
splitPos = currentPos;
}
}
LBLOG( LOG_LB2 ) << "Should split at Y " << splitPos << std::endl;
if( getDamping() < 1.f )
splitPos = (1.f - getDamping( )) * splitPos +
getDamping() * node->split;
LBLOG( LOG_LB2 ) << "Dampened split at Y " << splitPos << std::endl;
const Compound* root = getCompound();
const float pvpH = static_cast< float >(
root->getInheritPixelViewport().h );
const float boundary = static_cast< float >(node->boundary2i.y( )) /
pvpH;
if( node->left->resources == 0.f )
splitPos = vp.y;
else if( node->right->resources == 0.f )
splitPos = end;
else if ( boundary > 0 )
{
const float lengthRight = vp.getYEnd() - splitPos;
const float lengthLeft = splitPos - vp.y;
const float maxRight =
static_cast< float >( node->right->maxSize.y( )) / pvpH;
const float maxLeft =
static_cast< float >( node->left->maxSize.y( )) / pvpH;
if( lengthRight > maxRight )
splitPos = end - maxRight;
else if( lengthLeft > maxLeft )
splitPos = vp.y + maxLeft;
if( (splitPos - vp.y) < boundary )
splitPos = vp.y + boundary;
if( (end - splitPos) < boundary )
splitPos = end - boundary;
const uint32_t ratio =
static_cast< uint32_t >( splitPos / boundary + .5f );
splitPos = ratio * boundary;
}
splitPos = LB_MAX( splitPos, vp.y );
splitPos = LB_MIN( splitPos, end );
const float newPixelH = pvpH * splitPos;
const float oldPixelH = pvpH * node->split;
if( int( fabs(newPixelH - oldPixelH) ) < node->resistance2i.y( ))
splitPos = node->split;
else
node->split = splitPos;
LBLOG( LOG_LB2 ) << "Constrained split " << vp << " at Y "
<< splitPos << std::endl;
Viewport childVP = vp;
childVP.h = (splitPos - vp.y);
_computeSplit( node->left, leftTime, datas, childVP, range );
childVP.y = childVP.getYEnd();
childVP.h = end - childVP.y;
while( childVP.getYEnd() < end )
childVP.h += std::numeric_limits< float >::epsilon();
_computeSplit( node->right, time - leftTime, datas, childVP, range);
break;
}
case MODE_DB:
{
LBASSERT( vp == Viewport::FULL );
float splitPos = range.start;
const float end = range.end;
while( timeLeft > std::numeric_limits< float >::epsilon() &&
splitPos < end )
{
LBLOG( LOG_LB2 ) << timeLeft << "ms left using "
<< workingSet.size() << " tiles" << std::endl;
// remove all irrelevant items from working set
for( LBDatas::iterator i = workingSet.begin();
i != workingSet.end(); )
{
const Data& data = *i;
if( data.range.end > splitPos )
++i;
else
i = workingSet.erase( i );
}
if( workingSet.empty( ))
break;
// find next 'discontinouity' in loads
float currentPos = 1.0f;
for( LBDatas::const_iterator i = workingSet.begin();
i != workingSet.end(); ++i )
{
const Data& data = *i;
currentPos = LB_MIN( currentPos, data.range.end );
}
const float size = currentPos - splitPos;
LBASSERTINFO( size > 0.f, currentPos << "<=" << splitPos );
LBASSERT( currentPos <= 1.0f );
// accumulate normalized load in splitPos...currentPos
LBLOG( LOG_LB2 ) << "Computing load in range " << splitPos
<< "..." << currentPos << std::endl;
float currentTime = 0.f;
for( LBDatas::const_iterator i = workingSet.begin();
i != workingSet.end(); ++i )
{
const Data& data = *i;
if( data.range.start >= currentPos ) // not yet needed data
break;
#if 0
// make sure we cover full area
LBASSERTINFO( data.range.start <= splitPos,
data.range.start << " > " << splitPos );
LBASSERTINFO( data.range.end >= currentPos,
data.range.end << " < " << currentPos);
#endif
currentTime += data.time * size / data.range.getSize();
}
LBLOG( LOG_LB2 ) << splitPos << "..." << currentPos << ": t="
<< currentTime << " of " << timeLeft
<< std::endl;
if( currentTime >= timeLeft ) // found last region
{
const float width = currentPos - splitPos;
splitPos += (width * timeLeft / currentTime );
timeLeft = 0.0f;
}
else
{
timeLeft -= currentTime;
splitPos = currentPos;
}
}
LBLOG( LOG_LB2 ) << "Should split at " << splitPos << std::endl;
if( getDamping() < 1.f )
splitPos = (1.f - getDamping( )) * splitPos +
getDamping() * node->split;
LBLOG( LOG_LB2 ) << "Dampened split at " << splitPos << std::endl;
const float boundary( node->boundaryf );
if( node->left->resources == 0.f )
splitPos = range.start;
else if( node->right->resources == 0.f )
splitPos = end;
const uint32_t ratio = static_cast< uint32_t >
( splitPos / boundary + .5f );
splitPos = ratio * boundary;
if( (splitPos - range.start) < boundary )
splitPos = range.start;
if( (end - splitPos) < boundary )
splitPos = end;
if( fabs( splitPos - node->split ) < node->resistancef )
splitPos = node->split;
else
node->split = splitPos;
LBLOG( LOG_LB2 ) << "Constrained split " << range << " at pos "
<< splitPos << std::endl;
Range childRange = range;
childRange.end = splitPos;
_computeSplit( node->left, leftTime, datas, vp, childRange );
childRange.start = childRange.end;
childRange.end = range.end;
_computeSplit( node->right, time - leftTime, datas, vp, childRange);
break;
}
default:
LBUNIMPLEMENTED;
}
}
void LoadEqualizer::_updateNode( Node* node, const Viewport& vp,
const Range& range )
{
Node* left = node->left;
Node* right = node->right;
LBASSERT( left );
LBASSERT( right );
Viewport leftVP = vp;
Viewport rightVP = vp;
Range leftRange = range;
Range rightRange = range;
switch( node->mode )
{
default:
LBUNIMPLEMENTED;
case MODE_VERTICAL:
leftVP.w = vp.w * .5f;
rightVP.x = leftVP.getXEnd();
rightVP.w = vp.getXEnd() - rightVP.x;
node->split = leftVP.getXEnd();
break;
case MODE_HORIZONTAL:
leftVP.h = vp.h * .5f;
rightVP.y = leftVP.getYEnd();
rightVP.h = vp.getYEnd() - rightVP.y;
node->split = leftVP.getYEnd();
break;
case MODE_DB:
leftRange.end = range.start + ( range.end - range.start ) * .5f;
rightRange.start = leftRange.end;
node->split = leftRange.end;
break;
}
_update( left, leftVP, leftRange );
_update( right, rightVP, rightRange );
node->resources = left->resources + right->resources;
if( left->resources == 0.f )
{
node->maxSize = right->maxSize;
node->boundary2i = right->boundary2i;
node->boundaryf = right->boundaryf;
node->resistance2i = right->resistance2i;
node->resistancef = right->resistancef;
}
else if( right->resources == 0.f )
{
node->maxSize = left->maxSize;
node->boundary2i = left->boundary2i;
node->boundaryf = left->boundaryf;
node->resistance2i = left->resistance2i;
node->resistancef = left->resistancef;
}
else
{
switch( node->mode )
{
case MODE_VERTICAL:
node->maxSize.x() = left->maxSize.x() + right->maxSize.x();
node->maxSize.y() = LB_MIN( left->maxSize.y(), right->maxSize.y());
node->boundary2i.x() = left->boundary2i.x()+ right->boundary2i.x();
node->boundary2i.y() = LB_MAX( left->boundary2i.y(),
right->boundary2i.y());
node->boundaryf = LB_MAX( left->boundaryf, right->boundaryf );
node->resistance2i.x() = LB_MAX( left->resistance2i.x(),
right->resistance2i.x( ));
node->resistance2i.y() = LB_MAX( left->resistance2i.y(),
right->resistance2i.y());
node->resistancef = LB_MAX( left->resistancef, right->resistancef );
break;
case MODE_HORIZONTAL:
node->maxSize.x() = LB_MIN( left->maxSize.x(), right->maxSize.x());
node->maxSize.y() = left->maxSize.y() + right->maxSize.y();
node->boundary2i.x() = LB_MAX( left->boundary2i.x(),
right->boundary2i.x() );
node->boundary2i.y() = left->boundary2i.y()+ right->boundary2i.y();
node->boundaryf = LB_MAX( left->boundaryf, right->boundaryf );
node->resistance2i.x() = LB_MAX( left->resistance2i.x(),
right->resistance2i.x() );
node->resistance2i.y() = LB_MAX( left->resistance2i.y(),
right->resistance2i.y( ));
node->resistancef = LB_MAX( left->resistancef, right->resistancef );
break;
case MODE_DB:
node->boundary2i.x() = LB_MAX( left->boundary2i.x(),
right->boundary2i.x() );
node->boundary2i.y() = LB_MAX( left->boundary2i.y(),
right->boundary2i.y() );
node->boundaryf = left->boundaryf + right->boundaryf;
node->resistance2i.x() = LB_MAX( left->resistance2i.x(),
right->resistance2i.x() );
node->resistance2i.y() = LB_MAX( left->resistance2i.y(),
right->resistance2i.y() );
node->resistancef = LB_MAX( left->resistancef, right->resistancef );
break;
default:
LBUNIMPLEMENTED;
}
}
}
void LoadEqualizer::notifyLoadData( Channel* channel,
const uint32_t frameNumber,
const Statistics& statistics,
const Viewport& region )
{
LBLOG( LOG_LB2 ) << statistics.size()
<< " samples from "<< channel->getName()
<< " @ " << frameNumber << std::endl;
for( std::deque< LBFrameData >::iterator i = _history.begin();
i != _history.end(); ++i )
{
LBFrameData& frameData = *i;
if( frameData.first != frameNumber )
continue;
// Found corresponding historical data set
LBDatas& items = frameData.second;
for( LBDatas::iterator j = items.begin(); j != items.end(); ++j )
{
Data& data = *j;
if( data.channel != channel )
continue;
// Found corresponding historical data item
const uint32_t taskID = data.taskID;
LBASSERTINFO( taskID > 0, channel->getName( ));
// gather relevant load data
int64_t startTime = std::numeric_limits< int64_t >::max();
int64_t endTime = 0;
bool loadSet = false;
int64_t transmitTime = 0;
for( size_t k = 0; k < statistics.size(); ++k )
{
const Statistic& stat = statistics[k];
if( stat.task == data.destTaskID )
_updateAssembleTime( data, stat );
// from different compound
if( stat.task != taskID || loadSet )
continue;
switch( stat.type )
{
case Statistic::CHANNEL_CLEAR:
case Statistic::CHANNEL_DRAW:
case Statistic::CHANNEL_READBACK:
startTime = LB_MIN( startTime, stat.startTime );
endTime = LB_MAX( endTime, stat.endTime );
break;
case Statistic::CHANNEL_ASYNC_READBACK:
case Statistic::CHANNEL_FRAME_TRANSMIT:
transmitTime += stat.endTime - stat.startTime;
break;
case Statistic::CHANNEL_FRAME_WAIT_SENDTOKEN:
transmitTime -= stat.endTime - stat.startTime;
break;
// assemble blocks on input frames, stop using subsequent data
case Statistic::CHANNEL_ASSEMBLE:
loadSet = true;
break;
default:
break;
}
}
if( startTime == std::numeric_limits< int64_t >::max( ))
return;
data.vp.apply( region ); // Update ROI
data.time = endTime - startTime;
data.time = LB_MAX( data.time, 1 );
data.time = LB_MAX( data.time, transmitTime );
data.assembleTime = LB_MAX( data.assembleTime, 0 );
LBLOG( LOG_LB2 ) << "Added time " << data.time << " (+"
<< data.assembleTime << ") for "
<< channel->getName() << " " << data.vp << ", "
<< data.range << " @ " << frameNumber << std::endl;
return;
// Note: if the same channel is used twice as a child, the
// load-compound association does not work.
}
}
}
void TreeEqualizer::_assign( Node* node, const Viewport& vp,
const Range& range )
{
LBLOG( LOG_LB2 ) << "assign " << vp << ", " << range << " time "
<< node->time << " split " << node->split << std::endl;
LBASSERTINFO( vp.isValid(), vp );
LBASSERTINFO( range.isValid(), range );
LBASSERTINFO( node->resources > 0.f || !vp.hasArea() || !range.hasData(),
"Assigning work to unused compound: " << vp << ", " << range);
Compound* compound = node->compound;
if( compound )
{
LBASSERTINFO( vp == Viewport::FULL || range == Range::ALL,
"Mixed 2D/DB load-balancing not implemented" );
compound->setViewport( vp );
compound->setRange( range );
LBLOG( LOG_LB2 ) << compound->getChannel()->getName() << " set " << vp
<< ", " << range << std::endl;
return;
}
switch( node->mode )
{
case MODE_VERTICAL:
{
// Ensure minimum size
const Compound* root = getCompound();
const float pvpW = float( root->getInheritPixelViewport().w );
const float end = vp.getXEnd();
const float boundary = float( node->boundary2i.x( )) / pvpW;
float absoluteSplit = vp.x + vp.w * node->split;
if( node->left->resources == 0.f )
absoluteSplit = vp.x;
else if( node->right->resources == 0.f )
absoluteSplit = end;
else if( boundary > 0 )
{
const float right = vp.getXEnd() - absoluteSplit;
const float left = absoluteSplit - vp.x;
const float maxRight = float( node->right->maxSize.x( )) / pvpW;
const float maxLeft = float( node->left->maxSize.x( )) / pvpW;
if( right > maxRight )
absoluteSplit = end - maxRight;
else if( left > maxLeft )
absoluteSplit = vp.x + maxLeft;
if( (absoluteSplit - vp.x) < boundary )
absoluteSplit = vp.x + boundary;
if( (end - absoluteSplit) < boundary )
absoluteSplit = end - boundary;
const uint32_t ratio = uint32_t( absoluteSplit / boundary + .5f );
absoluteSplit = ratio * boundary;
}
absoluteSplit = LB_MAX( absoluteSplit, vp.x );
absoluteSplit = LB_MIN( absoluteSplit, end);
node->split = (absoluteSplit - vp.x ) / vp.w;
LBLOG( LOG_LB2 ) << "Constrained split " << vp << " at X "
<< node->split << std::endl;
// traverse children
Viewport childVP = vp;
childVP.w = (absoluteSplit - vp.x);
_assign( node->left, childVP, range );
childVP.x = childVP.getXEnd();
childVP.w = end - childVP.x;
// Fix 2994111: Rounding errors with 2D LB and 16 sources
// Floating point rounding may create a width for the 'right'
// child which is slightly below the parent width. Correct it.
while( childVP.getXEnd() < end )
childVP.w += std::numeric_limits< float >::epsilon();
_assign( node->right, childVP, range );
break;
}
case MODE_HORIZONTAL:
{
// Ensure minimum size
const Compound* root = getCompound();
const float pvpH = float( root->getInheritPixelViewport().h );
const float end = vp.getYEnd();
const float boundary = float( node->boundary2i.y( )) / pvpH;
float absoluteSplit = vp.y + vp.h * node->split;
if( node->left->resources == 0.f )
absoluteSplit = vp.y;
else if( node->right->resources == 0.f )
absoluteSplit = end;
else if( boundary > 0 )
{
const float right = vp.getYEnd() - absoluteSplit;
const float left = absoluteSplit - vp.y;
const float maxRight = float( node->right->maxSize.y( )) / pvpH;
const float maxLeft = float( node->left->maxSize.y( )) / pvpH;
if( right > maxRight )
absoluteSplit = end - maxRight;
else if( left > maxLeft )
absoluteSplit = vp.y + maxLeft;
if( (absoluteSplit - vp.y) < boundary )
absoluteSplit = vp.y + boundary;
if( (end - absoluteSplit) < boundary )
absoluteSplit = end - boundary;
const uint32_t ratio = uint32_t( absoluteSplit / boundary + .5f );
absoluteSplit = ratio * boundary;
}
absoluteSplit = LB_MAX( absoluteSplit, vp.y );
absoluteSplit = LB_MIN( absoluteSplit, end);
node->split = (absoluteSplit - vp.y ) / vp.h;
LBLOG( LOG_LB2 ) << "Constrained split " << vp << " at X "
<< node->split << std::endl;
// traverse children
Viewport childVP = vp;
childVP.h = (absoluteSplit - vp.y);
_assign( node->left, childVP, range );
childVP.y = childVP.getYEnd();
childVP.h = end - childVP.y;
// Fix 2994111: Rounding errors with 2D LB and 16 sources
// Floating point rounding may create a width for the 'right'
// child which is slightly below the parent width. Correct it.
while( childVP.getYEnd() < end )
childVP.h += std::numeric_limits< float >::epsilon();
_assign( node->right, childVP, range );
break;
}
case MODE_DB:
{
LBASSERT( vp == Viewport::FULL );
const float end = range.end;
float absoluteSplit = range.start + (range.end-range.start)*node->split;
const float boundary( node->boundaryf );
if( node->left->resources == 0.f )
absoluteSplit = range.start;
else if( node->right->resources == 0.f )
absoluteSplit = end;
const uint32_t ratio = uint32_t( absoluteSplit / boundary + .5f );
absoluteSplit = ratio * boundary;
if( (absoluteSplit - range.start) < boundary )
absoluteSplit = range.start;
if( (end - absoluteSplit) < boundary )
absoluteSplit = end;
node->split = (absoluteSplit-range.start) / (range.end-range.start);
LBLOG( LOG_LB2 ) << "Constrained split " << range << " at pos "
<< node->split << std::endl;
Range childRange = range;
childRange.end = absoluteSplit;
_assign( node->left, vp, childRange );
childRange.start = childRange.end;
childRange.end = range.end;
_assign( node->right, vp, childRange);
break;
}
default:
LBUNIMPLEMENTED;
}
}
void TreeEqualizer::_update( Node* node )
{
if( !node )
return;
const Compound* compound = node->compound;
if( compound )
{
const Channel* channel = compound->getChannel();
const PixelViewport& pvp = channel->getPixelViewport();
LBASSERT( channel );
node->resources = compound->isActive() ? compound->getUsage() : 0.f;
node->maxSize.x() = pvp.w;
node->maxSize.y() = pvp.h;
node->boundaryf = _boundaryf;
node->boundary2i = _boundary2i;
return;
}
// else
LBASSERT( node->left );
LBASSERT( node->right );
_update( node->left );
_update( node->right );
node->resources = node->left->resources + node->right->resources;
if( node->left->resources == 0.f )
{
node->maxSize = node->right->maxSize;
node->boundary2i = node->right->boundary2i;
node->boundaryf = node->right->boundaryf;
node->time = node->right->time;
}
else if( node->right->resources == 0.f )
{
node->maxSize = node->left->maxSize;
node->boundary2i = node->left->boundary2i;
node->boundaryf = node->left->boundaryf;
node->time = node->left->time;
}
else
{
switch( node->mode )
{
case MODE_VERTICAL:
node->maxSize.x() = node->left->maxSize.x() +
node->right->maxSize.x();
node->maxSize.y() = LB_MIN( node->left->maxSize.y(),
node->right->maxSize.y() );
node->boundary2i.x() = node->left->boundary2i.x() +
node->right->boundary2i.x();
node->boundary2i.y() = LB_MAX( node->left->boundary2i.y(),
node->right->boundary2i.y());
node->boundaryf = LB_MAX( node->left->boundaryf,
node->right->boundaryf );
break;
case MODE_HORIZONTAL:
node->maxSize.x() = LB_MIN( node->left->maxSize.x(),
node->right->maxSize.x() );
node->maxSize.y() = node->left->maxSize.y() +
node->right->maxSize.y();
node->boundary2i.x() = LB_MAX( node->left->boundary2i.x(),
node->right->boundary2i.x() );
node->boundary2i.y() = node->left->boundary2i.y() +
node->right->boundary2i.y();
node->boundaryf = LB_MAX( node->left->boundaryf,
node->right->boundaryf );
break;
case MODE_DB:
node->boundary2i.x() = LB_MAX( node->left->boundary2i.x(),
node->right->boundary2i.x() );
node->boundary2i.y() = LB_MAX( node->left->boundary2i.y(),
node->right->boundary2i.y() );
node->boundaryf = node->left->boundaryf +node->right->boundaryf;
break;
default:
LBUNIMPLEMENTED;
}
node->time = node->left->time + node->right->time;
}
}
int64_t SocketConnection::write( const void* buffer, const uint64_t bytes )
{
if( !isConnected() || _writeFD == INVALID_SOCKET )
return -1;
DWORD wrote;
WSABUF wsaBuffer =
{
LB_MIN( bytes, 65535 ),
const_cast<char*>( static_cast< const char* >( buffer ))
};
ResetEvent( _overlappedWrite.hEvent );
if( WSASend(_writeFD, &wsaBuffer, 1, &wrote, 0, &_overlappedWrite, 0 ) == 0 )
// ok
return wrote;
if( WSAGetLastError() != WSA_IO_PENDING )
return -1;
const DWORD err = WaitForSingleObject( _overlappedWrite.hEvent, INFINITE );
switch( err )
{
case WAIT_FAILED:
case WAIT_ABANDONED:
{
LBWARN << "Write error" << lunchbox::sysError << std::endl;
return -1;
}
default:
LBWARN << "Unhandled write error " << err << ": " << lunchbox::sysError
<< std::endl;
// no break;
case WAIT_OBJECT_0:
break;
}
DWORD got = 0;
DWORD flags = 0;
if( WSAGetOverlappedResult( _writeFD, &_overlappedWrite, &got, false,
&flags ))
{
return got;
}
switch( WSAGetLastError() )
{
case WSA_IO_INCOMPLETE:
throw Exception( Exception::TIMEOUT_WRITE );
default:
{
LBWARN << "Write error : " << lunchbox::sysError << std::endl;
return -1;
}
}
LBUNREACHABLE;
return -1;
}
int64_t IBInterface::postRdmaWrite( const void* buffer, uint32_t numBytes )
{
#ifdef EQ_MEASURE_TIME
eq::lunchbox::Clock clock;
clock.reset();
#endif
ib_api_status_t ibStatus;
ib_wc_t wc;
ib_wc_t *wcDone,*wcFree;
wcFree = &wc;
wcFree->p_next = 0;
wcDone = 0;
#ifdef EQ_MEASURE_TIME
eq::lunchbox::Clock clockWait;
clockWait.reset();
#endif
// validation of the send job
do {
ibStatus = ib_poll_cq( _completionQueue->getWriteHandle(),
&wcFree,
&wcDone );
if ( ibStatus == IB_SUCCESS )
{
if ( wcDone->status != IB_WCS_SUCCESS )
{
LBWARN << "ERROR IN POLL WRITE"<< std::endl;
return -1;
}
_writePoll.getData()[wcDone->wr_id] = true;
wcFree = wcDone;
wcFree->p_next = 0;
wcDone = 0;
}
else if ( !_writePoll.getData()[ _numBufWrite ] )
{
ibStatus = IB_SUCCESS;
}
} while ( ibStatus == IB_SUCCESS );
#ifdef EQ_MEASURE_TIME
_timeTotalWriteWait += clockWait.getTimef();
#endif
uint32_t incBytes = 0;
uint32_t compt = 0;
uint32_t size;
size = LB_MIN( numBytes, EQ_MAXBLOCKBUFFER );
ib_local_ds_t list;
#ifdef EQ_MEASURE_TIME
eq::lunchbox::Clock clockCopy;
clockCopy.reset();
#endif
//memcpy( _writeBlocks[ _numBufWrite ]->buf.getData(),
// buffer , size );
eq::lunchbox::fastCopy( _writeBlocks[ _numBufWrite ]->buf.getData(),
buffer , size );
list.vaddr = _writeBlocks[ _numBufWrite ]->getVaddr();
#ifdef EQ_MEASURE_TIME
_timeCopyBufferWrite += clockCopy.getTimef();
#endif
list.lkey = _writeBlocks[_numBufWrite ]->getLocalKey();
list.length = size;
// A 64-bit work request identifier that is returned to the consumer
// as part of the work completion.
_wr.wr_id = _numBufWrite;
// A reference to an array of local data segments used by the send
// operation.
_wr.ds_array = &list;
// Number of local data segments specified by this work request.
_wr.num_ds = 1;
// The type of work request being submitted to the send queue.
_wr.wr_type = WR_SEND;
// A pointer used to chain work requests together. This permits multiple
// work requests to be posted to a queue pair through a single function
// call. This value is set to NULL to mark the end of the chain.
_wr.p_next = 0;
// This routine posts a work request to the send queue of a queue pair
ibStatus = ib_post_send( _queuePair, &_wr, 0 );
if ( ibStatus != IB_SUCCESS )
{
LBWARN << "ERROR IN POST SEND DATA"<< std::endl;
return -1;
}
_writePoll.getData()[ _numBufWrite ] = false;
if ( _numBufWrite == EQ_NUMBLOCKMEMORY -1 )
_ibConnection->incWriteInterface();
_numBufWrite = ( _numBufWrite + 1 ) % EQ_NUMBLOCKMEMORY;
#ifdef EQ_MEASURE_TIME
_timeTotalWrite += clock.getTimef();
#endif
return size;
}
int64_t IBInterface::readSync( void* buffer, uint32_t bytes )
{
#ifdef EQ_MEASURE_TIME
eq::lunchbox::Clock clock;
clock.reset();
#endif
uint32_t comptRead = 0;
int64_t sizebuf = _readPoll.getData()[ _numBufRead ];
// if no data in buffer, we ask for a receive operation
while ( sizebuf < 1 )
#ifdef EQ_MEASURE_TIME
eq::lunchbox::Clock clockWait;
clockWait.reset();
#endif
sizebuf = _waitPollCQ( _numBufRead );
#ifdef EQ_MEASURE_TIME
_timeTotalWaitPoll += clockWait.getTimef();
#endif
if ( sizebuf > _posReadInBuffer )
{
#ifdef EQ_MEASURE_TIME
eq::lunchbox::Clock clockCopy;
clockCopy.reset();
#endif
// find a better memcpy or a system that we don't need to use memcpy
// copy buffer
uint32_t nbRead = LB_MIN( bytes, sizebuf - _posReadInBuffer);
/* memcpy(reinterpret_cast<char*>( buffer ) + comptRead,
reinterpret_cast<char*>
( _readBlocks[_numBufRead]->buf.getData()) +
_posReadInBuffer, nbRead );*/
eq::lunchbox::fastCopy( reinterpret_cast<char*>( buffer ) + comptRead,
reinterpret_cast<char*>
( _readBlocks[_numBufRead]->buf.getData() ) +
_posReadInBuffer, nbRead );
_posReadInBuffer += nbRead;
#ifdef EQ_MEASURE_TIME
_timeCopyBufferRead += clockCopy.getTimef();
#endif
// all buffer has been taken
if ( _posReadInBuffer == sizebuf )
{
_completionQueue->removeEventRead();
// init var for next read
_posReadInBuffer = 0;
_readPoll.getData()[ _numBufRead ] = 0;
// next Read in on the next CQ
if (_numBufRead == EQ_NUMBLOCKMEMORY-1 )
_ibConnection->incReadInterface();
// notify that read is finnish
_ibPostRecv( _numBufRead );
// To do work with more buffer in a CQ
_numBufRead = ( _numBufRead + 1 ) % EQ_NUMBLOCKMEMORY;
}
return nbRead;
}
LBWARN << "ERROR IN READ SYNC"<< std::endl;
return -1;
#ifdef EQ_MEASURE_TIME
_floatTimeReadSync += clock.getTimef();
#endif
}