TaskBatchPtr batchTasksWalk( const TaskNode::Task &task, const std::set<const TaskBatch *> &ancestors = std::set<const TaskBatch *>() )
{
// Acquire a batch with this task placed in it,
// and check that we haven't discovered a cyclic
// dependency.
TaskBatchPtr batch = acquireBatch( task );
if( ancestors.find( batch.get() ) != ancestors.end() )
{
throw IECore::Exception( ( boost::format( "Dispatched tasks cannot have cyclic dependencies but %s is involved in a cycle." ) % batch->plug()->relativeName( batch->plug()->ancestor<ScriptNode>() ) ).str() );
}
// Ask the task what preTasks and postTasks it would like.
TaskNode::Tasks preTasks;
TaskNode::Tasks postTasks;
{
Context::Scope scopedTaskContext( task.context() );
task.plug()->preTasks( preTasks );
task.plug()->postTasks( postTasks );
}
// Collect all the batches the postTasks belong in.
// We grab these first because they need to be included
// in the ancestors for cycle detection when getting
// the preTask batches.
TaskBatches postBatches;
for( TaskNode::Tasks::const_iterator it = postTasks.begin(); it != postTasks.end(); ++it )
{
postBatches.push_back( batchTasksWalk( *it ) );
}
// Collect all the batches the preTasks belong in,
// and add them as preTasks for our batch.
std::set<const TaskBatch *> preTaskAncestors( ancestors );
preTaskAncestors.insert( batch.get() );
for( TaskBatches::const_iterator it = postBatches.begin(), eIt = postBatches.end(); it != eIt; ++it )
{
preTaskAncestors.insert( it->get() );
}
for( TaskNode::Tasks::const_iterator it = preTasks.begin(); it != preTasks.end(); ++it )
{
addPreTask( batch.get(), batchTasksWalk( *it, preTaskAncestors ) );
}
// As far as TaskBatch and doDispatch() are concerned, there
// is no such thing as a postTask, so we emulate them by making
// this batch a preTask of each of the postTask batches. We also
// add the postTask batches as preTasks for the root, so that they
// are reachable from doDispatch().
for( TaskBatches::const_iterator it = postBatches.begin(), eIt = postBatches.end(); it != eIt; ++it )
{
addPreTask( it->get(), batch, /* forPostTask = */ true );
addPreTask( m_rootBatch.get(), *it );
}
return batch;
}
// Hash used to determine how to coalesce tasks into batches.
// If `batchHash( task1 ) == batchHash( task2 )` then the two
// tasks can be placed in the same batch.
IECore::MurmurHash batchHash( const TaskNode::Task &task )
{
MurmurHash result;
result.append( (uint64_t)task.node() );
// We ignore the frame because the whole point of batching
// is to allow multiple frames to be placed in the same
// batch if the context is otherwise identical.
result.append( contextHash( task.context(), /* ignoreFrame = */ true ) );
return result;
}
// Hash used to determine how to coalesce tasks into batches.
// If `batchHash( task1 ) == batchHash( task2 )` then the two
// tasks can be placed in the same batch.
IECore::MurmurHash batchHash( const TaskNode::Task &task )
{
MurmurHash result;
result.append( (uint64_t)task.node() );
const Context *context = task.context();
std::vector<IECore::InternedString> names;
context->names( names );
for( std::vector<IECore::InternedString>::const_iterator it = names.begin(); it != names.end(); ++it )
{
// Ignore the UI values since they should be irrelevant
// to execution.
if( it->string().compare( 0, 3, "ui:" ) == 0 )
{
continue;
}
// Ignore the frame, since the whole point of batching
// is to allow multiple frames to be placed in the same
// batch if the context is otherwise identical.
///
// There is one exception to this though - if the task is
// a no-op, then we don't want to coalesce, because then
// every single frame of the no-op would be placed in the
// same batch, and all downstream frames would then be forced
// to depend unnecessarily on all upstream frames.
if( *it == g_frame && task.hash() != MurmurHash() )
{
continue;
}
result.append( *it );
context->get<const IECore::Data>( *it )->hash( result );
}
return result;
}
TaskBatchPtr acquireBatch( const TaskNode::Task &task )
{
// See if we've previously visited this task, and therefore
// have placed it in a batch already, which we can return
// unchanged. The `taskToBatchMapHash` is used as the unique
// identity of a task.
MurmurHash taskToBatchMapHash = task.hash();
// Prevent identical tasks from different nodes from being
// coalesced.
taskToBatchMapHash.append( (uint64_t)task.node() );
if( task.hash() == IECore::MurmurHash() )
{
// Prevent no-ops from coalescing into a single batch, as this
// would break parallelism - see `DispatcherTest.testNoOpDoesntBreakFrameParallelism()`
taskToBatchMapHash.append( contextHash( task.context() ) );
}
const TaskToBatchMap::const_iterator it = m_tasksToBatches.find( taskToBatchMapHash );
if( it != m_tasksToBatches.end() )
{
return it->second;
}
// We haven't seen this task before, so we need to find
// an appropriate batch to put it in. This may be one of
// our current batches, or we may need to make a new one
// entirely if the current batch is full.
const bool requiresSequenceExecution = task.plug()->requiresSequenceExecution();
TaskBatchPtr batch = nullptr;
const MurmurHash batchMapHash = batchHash( task );
BatchMap::iterator bIt = m_currentBatches.find( batchMapHash );
if( bIt != m_currentBatches.end() )
{
TaskBatchPtr candidateBatch = bIt->second;
// Unfortunately we have to track batch size separately from `batch->frames().size()`,
// because no-ops don't update `frames()`, but _do_ count towards batch size.
IntDataPtr batchSizeData = candidateBatch->blindData()->member<IntData>( g_sizeBlindDataName );
const IntPlug *batchSizePlug = task.node()->dispatcherPlug()->getChild<const IntPlug>( g_batchSize );
const int batchSizeLimit = ( batchSizePlug ) ? batchSizePlug->getValue() : 1;
if( requiresSequenceExecution || ( batchSizeData->readable() < batchSizeLimit ) )
{
batch = candidateBatch;
batchSizeData->writable()++;
}
}
if( !batch )
{
batch = new TaskBatch( task.plug(), task.context() );
batch->blindData()->writable()[g_sizeBlindDataName] = new IntData( 1 );
m_currentBatches[batchMapHash] = batch;
}
// Now we have an appropriate batch, update it to include
// the frame for our task, and any other relevant information.
if( task.hash() != MurmurHash() )
{
float frame = task.context()->getFrame();
std::vector<float> &frames = batch->frames();
if( requiresSequenceExecution )
{
frames.insert( std::lower_bound( frames.begin(), frames.end(), frame ), frame );
}
else
{
frames.push_back( frame );
}
}
const BoolPlug *immediatePlug = task.node()->dispatcherPlug()->getChild<const BoolPlug>( g_immediatePlugName );
if( immediatePlug && immediatePlug->getValue() )
{
/// \todo Should we be scoping a context for this, to allow the plug to
/// have expressions on it? If so, should we be doing the same before
/// calling requiresSequenceExecution()? Or should we instead require that
/// they always be constant?
batch->blindData()->writable()[g_immediateBlindDataName] = g_trueBoolData;
}
// Remember which batch we stored this task in, for
// the next time someone asks for it.
m_tasksToBatches[taskToBatchMapHash] = batch;
return batch;
}
TaskBatchPtr acquireBatch( const TaskNode::Task &task )
{
// See if we've previously visited this task, and therefore
// have placed it in a batch already, which we can return
// unchanged.
MurmurHash taskToBatchMapHash = task.hash();
taskToBatchMapHash.append( (uint64_t)task.node() );
if( task.hash() == MurmurHash() )
{
// Make sure we don't coalesce all no-ops into a single
// batch. See comments in batchHash().
taskToBatchMapHash.append( task.context()->getFrame() );
}
const TaskToBatchMap::const_iterator it = m_tasksToBatches.find( taskToBatchMapHash );
if( it != m_tasksToBatches.end() )
{
return it->second;
}
// We haven't seen this task before, so we need to find
// an appropriate batch to put it in. This may be one of
// our current batches, or we may need to make a new one
// entirely.
TaskBatchPtr batch = NULL;
const MurmurHash batchMapHash = batchHash( task );
BatchMap::iterator bIt = m_currentBatches.find( batchMapHash );
if( bIt != m_currentBatches.end() )
{
TaskBatchPtr candidateBatch = bIt->second;
const IntPlug *batchSizePlug = task.node()->dispatcherPlug()->getChild<const IntPlug>( g_batchSize );
const size_t batchSize = ( batchSizePlug ) ? batchSizePlug->getValue() : 1;
if( task.plug()->requiresSequenceExecution() || ( candidateBatch->frames().size() < batchSize ) )
{
batch = candidateBatch;
}
}
if( !batch )
{
batch = new TaskBatch( task.plug(), task.context() );
m_currentBatches[batchMapHash] = batch;
}
// Now we have an appropriate batch, update it to include
// the frame for our task, and any other relevant information.
if( task.hash() != MurmurHash() )
{
float frame = task.context()->getFrame();
std::vector<float> &frames = batch->frames();
if( std::find( frames.begin(), frames.end(), frame ) == frames.end() )
{
if( task.plug()->requiresSequenceExecution() )
{
frames.insert( std::lower_bound( frames.begin(), frames.end(), frame ), frame );
}
else
{
frames.push_back( frame );
}
}
}
const BoolPlug *immediatePlug = task.node()->dispatcherPlug()->getChild<const BoolPlug>( g_immediatePlugName );
if( immediatePlug && immediatePlug->getValue() )
{
/// \todo Should we be scoping a context for this, to allow the plug to
/// have expressions on it? If so, should we be doing the same before
/// calling requiresSequenceExecution()? Or should we instead require that
/// they always be constant?
batch->blindData()->writable()[g_immediateBlindDataName] = g_trueBoolData;
}
// Remember which batch we stored this task in, for
// the next time someone asks for it.
m_tasksToBatches[taskToBatchMapHash] = batch;
return batch;
}