Evaluates a RooAbsReal object in other ways than recursive graph traversal.

Currently, it is being used for evaluating a RooAbsReal object and supplying the value to the minimizer, during a fit. The class scans the dependencies and schedules the computations in a secure and efficient way. The computations take place in the RooBatchCompute library and can be carried off by either the CPU or a CUDA-supporting GPU. The Evaluator class takes care of data transfers. An instance of this class is created every time RooAbsPdf::fitTo() is called and gets destroyed when the fitting ends.

Definition at line 36 of file Evaluator.h.

Public Member Functions
	Evaluator (const RooAbsReal &absReal, bool useGPU=false)
	Construct a new Evaluator.

	~Evaluator ()

RooArgSet	getParameters () const
	Gets all the parameters of the RooAbsReal.

void	print (std::ostream &os)

std::span< const double >	run ()
	Returns the value of the top node in the computation graph.

void	setInput (std::string const &name, std::span< const double > inputArray, bool isOnDevice)

void	setOffsetMode (RooFit::EvalContext::OffsetMode)
	Sets the offset mode for evaluation.

Private Member Functions
void	assignToGPU (NodeInfo &info)
	Assign a node to be computed in the GPU.

void	computeCPUNode (const RooAbsArg *node, NodeInfo &info)

std::span< const double >	getValHeterogeneous ()
	Returns the value of the top node in the computation graph.

void	markGPUNodes ()
	Decides which nodes are assigned to the GPU in a CUDA fit.

void	processVariable (NodeInfo &nodeInfo)
	Process a variable in the computation graph.

void	setClientsDirty (NodeInfo &nodeInfo)
	Flags all the clients of a given node dirty.

void	setOperMode (RooAbsArg *arg, RooAbsArg::OperMode opMode)
	Temporarily change the operation mode of a RooAbsArg until the Evaluator gets deleted.

void	syncDataTokens ()
	If there are servers with the same name that got de-duplicated in the `_nodes` list, we need to set their data tokens too.

void	updateOutputSizes ()

Private Attributes
std::unique_ptr< RooBatchCompute::AbsBufferManager >	_bufferManager

std::stack< std::unique_ptr< ChangeOperModeRAII > >	_changeOperModeRAIIs

RooFit::EvalContext	_evalContextCPU

RooFit::EvalContext	_evalContextCUDA

bool	_needToUpdateOutputSizes = false

int	_nEvaluations = 0

std::vector< NodeInfo >	_nodes

RooAbsReal &	_topNode

const bool	_useGPU = false

#include <RooFit/Evaluator.h>

Constructor & Destructor Documentation

◆ Evaluator()

RooFit::Evaluator::Evaluator	(	const RooAbsReal &	absReal,
		bool	useGPU = `false`
	)

Construct a new Evaluator.

The constructor analyzes and saves metadata about the graph, useful for the evaluation of it that will be done later. In case the CUDA mode is selected, there's also some CUDA-related initialization.

Parameters

[in]	absReal	The RooAbsReal object that sits on top of the computation graph that we want to evaluate.
[in]	useGPU	Whether the evaluation should be preferably done on the GPU.

Definition at line 149 of file Evaluator.cxx.

◆ ~Evaluator()

RooFit::Evaluator::~Evaluator ( )

Definition at line 331 of file Evaluator.cxx.

Member Function Documentation

◆ assignToGPU()

void RooFit::Evaluator::assignToGPU ( NodeInfo & info )

private

Assign a node to be computed in the GPU.

Scan it's clients and also assign them in case they only depend on GPU nodes.

Definition at line 524 of file Evaluator.cxx.

◆ computeCPUNode()

void RooFit::Evaluator::computeCPUNode	(	const RooAbsArg *	node,
		NodeInfo &	info
	)

private

Definition at line 340 of file Evaluator.cxx.

◆ getParameters()

RooArgSet RooFit::Evaluator::getParameters ( ) const

Gets all the parameters of the RooAbsReal.

This is in principle not necessary, because we can always ask the RooAbsReal itself, but the Evaluator has the cached information to get the answer quicker. Therefore, this is not meant to be used in general, just where it matters.

Warning: If we find another solution to get the parameters efficiently, this function might be removed without notice.

Definition at line 662 of file Evaluator.cxx.

◆ getValHeterogeneous()

std::span< const double > RooFit::Evaluator::getValHeterogeneous ( )

private

Returns the value of the top node in the computation graph.

Definition at line 449 of file Evaluator.cxx.

◆ markGPUNodes()

void RooFit::Evaluator::markGPUNodes ( )

private

Decides which nodes are assigned to the GPU in a CUDA fit.

Definition at line 559 of file Evaluator.cxx.

◆ print()

void RooFit::Evaluator::print ( std::ostream & os )

Definition at line 600 of file Evaluator.cxx.

◆ processVariable()

void RooFit::Evaluator::processVariable ( NodeInfo & nodeInfo )

private

Process a variable in the computation graph.

This is a separate non-inlined function such that we can see in performance profiles how long this takes.

Definition at line 395 of file Evaluator.cxx.

◆ run()

std::span< const double > RooFit::Evaluator::run ( )

Returns the value of the top node in the computation graph.

Definition at line 419 of file Evaluator.cxx.

◆ setClientsDirty()

void RooFit::Evaluator::setClientsDirty ( NodeInfo & nodeInfo )

private

Flags all the clients of a given node dirty.

This is a separate non-inlined function such that we can see in performance profiles how long this takes.

Definition at line 411 of file Evaluator.cxx.

◆ setInput()

void RooFit::Evaluator::setInput	(	std::string const &	name,
		std::span< const double >	inputArray,
		bool	isOnDevice
	)

Definition at line 240 of file Evaluator.cxx.

◆ setOffsetMode()

void RooFit::Evaluator::setOffsetMode ( RooFit::EvalContext::OffsetMode mode )

Sets the offset mode for evaluation.

This function sets the offset mode for evaluation to the specified mode. It updates the offset mode for both CPU and CUDA evaluation contexts.

Parameters

mode	The offset mode to be set.

Note: This function marks reducer nodes as dirty if the offset mode is changed, because only reducer nodes can use offsetting.

Definition at line 684 of file Evaluator.cxx.

◆ setOperMode()

void RooFit::Evaluator::setOperMode	(	RooAbsArg *	arg,
		RooAbsArg::OperMode	opMode
	)

private

Temporarily change the operation mode of a RooAbsArg until the Evaluator gets deleted.

Definition at line 593 of file Evaluator.cxx.

◆ syncDataTokens()

void RooFit::Evaluator::syncDataTokens ( )

private

If there are servers with the same name that got de-duplicated in the _nodes list, we need to set their data tokens too.

We find such nodes by visiting the servers of every known node.

Definition at line 224 of file Evaluator.cxx.

◆ updateOutputSizes()

void RooFit::Evaluator::updateOutputSizes ( )

private

Definition at line 291 of file Evaluator.cxx.

Member Data Documentation

◆ _bufferManager

std::unique_ptr<RooBatchCompute::AbsBufferManager> RooFit::Evaluator::_bufferManager

private

Definition at line 59 of file Evaluator.h.

◆ _changeOperModeRAIIs

std::stack<std::unique_ptr<ChangeOperModeRAII> > RooFit::Evaluator::_changeOperModeRAIIs

private

Definition at line 67 of file Evaluator.h.

◆ _evalContextCPU

RooFit::EvalContext RooFit::Evaluator::_evalContextCPU

private

Definition at line 64 of file Evaluator.h.

◆ _evalContextCUDA

RooFit::EvalContext RooFit::Evaluator::_evalContextCUDA

private

Definition at line 65 of file Evaluator.h.

◆ _needToUpdateOutputSizes

bool RooFit::Evaluator::_needToUpdateOutputSizes = false

private

Definition at line 63 of file Evaluator.h.

◆ _nEvaluations

int RooFit::Evaluator::_nEvaluations = 0

private

Definition at line 62 of file Evaluator.h.

◆ _nodes

std::vector<NodeInfo> RooFit::Evaluator::_nodes

private

Definition at line 66 of file Evaluator.h.

◆ _topNode

RooAbsReal& RooFit::Evaluator::_topNode

private

Definition at line 60 of file Evaluator.h.

◆ _useGPU

const bool RooFit::Evaluator::_useGPU = false

private

Definition at line 61 of file Evaluator.h.

Libraries for RooFit::Evaluator:

[legend]

The documentation for this class was generated from the following files:

roofit/roofitcore/inc/RooFit/Evaluator.h
roofit/roofitcore/src/RooFit/Evaluator.cxx

Public Member Functions

Private Member Functions

Private Attributes

Constructor & Destructor Documentation

◆ Evaluator()

◆ ~Evaluator()

Member Function Documentation

◆ assignToGPU()

◆ computeCPUNode()

◆ getParameters()

◆ getValHeterogeneous()

◆ markGPUNodes()

◆ print()

◆ processVariable()

◆ run()

◆ setClientsDirty()

◆ setInput()

◆ setOffsetMode()

◆ setOperMode()

◆ syncDataTokens()

◆ updateOutputSizes()

Member Data Documentation

◆ _bufferManager

◆ _changeOperModeRAIIs

◆ _evalContextCPU

◆ _evalContextCUDA

◆ _needToUpdateOutputSizes

◆ _nEvaluations

◆ _nodes

◆ _topNode

◆ _useGPU