RooBatchCompute Namespace Reference

Namespace for dispatching RooFit computations to various backends. More...

Namespaces
namespace	RF_ARCH

Classes
class	BracketAdapter
	Little adapter that gives a bracket operator to types that don't have one. More...
class	BracketAdapterWithMask
class	RooBatchComputeInterface
	The interface which should be implemented to provide optimised computation functions for implementations of RooAbsReal::evaluateSpan(). More...
struct	RunContext
	This struct enables passing computation data around between elements of a computation graph. More...

Typedefs
typedef std::vector< double >	ArgVector
typedef const double *__restrict	InputArr
typedef double *__restrict	RestrictArr
typedef std::vector< RooSpan< const double > >	VarVector

Enumerations
enum class	Architecture {   AVX512 , AVX2 , AVX , SSE4 ,   GENERIC , CUDA }
enum	Computer {   AddPdf , ArgusBG , Bernstein , BifurGauss ,   BreitWigner , Bukin , CBShape , Chebychev ,   ChiSquare , DstD0BG , Exponential , Gamma ,   Gaussian , Johnson , Landau , Lognormal ,   NegativeLogarithms , Novosibirsk , Poisson , Polynomial ,   ProdPdf , Ratio , Voigtian }

Functions
__roodevice__ double	fast_exp (double x)
__roodevice__ double	fast_isqrt (double x)
__roodevice__ double	fast_log (double x)
void	init ()
	Inspect hardware capabilities, and load the optimal library for RooFit computations.

Variables
constexpr uint16_t	bufferSize = 64
R__EXTERN RooBatchComputeInterface *	dispatchCPU = nullptr
	This dispatch pointer points to an implementation of the compute library, provided one has been loaded.
R__EXTERN RooBatchComputeInterface *	dispatchCUDA = nullptr

Detailed Description

Namespace for dispatching RooFit computations to various backends.

This namespace contains an interface for providing high-performance computation functions for use in RooAbsReal::evaluateSpan(), see RooBatchComputeInterface.

Furthermore, several implementations of this interface can be created, which reside in RooBatchCompute::RF_ARCH, where RF_ARCH may be replaced by the architecture that this implementation targets, e.g. SSE, AVX, etc.

Using the pointer RooBatchCompute::dispatch, a computation request can be dispatched to the fastest backend that is available on a specific platform.

Typedef Documentation

ArgVector

typedef std::vector<double> RooBatchCompute::ArgVector

Definition at line 37 of file RooBatchComputeTypes.h.

InputArr

typedef const double* __restrict RooBatchCompute::InputArr

Definition at line 39 of file RooBatchComputeTypes.h.

RestrictArr

typedef double* __restrict RooBatchCompute::RestrictArr

Definition at line 38 of file RooBatchComputeTypes.h.

VarVector

typedef std::vector<RooSpan<const double> > RooBatchCompute::VarVector

Definition at line 36 of file RooBatchComputeTypes.h.

Enumeration Type Documentation

Architecture

enum class RooBatchCompute::Architecture

strong

Enumerator
AVX512
AVX2
AVX
SSE4
GENERIC
CUDA

Definition at line 37 of file RooBatchCompute.h.

Computer

enum RooBatchCompute::Computer

Enumerator
AddPdf
ArgusBG
Bernstein
BifurGauss
BreitWigner
Bukin
CBShape
Chebychev
ChiSquare
DstD0BG
Exponential
Gamma
Gaussian
Johnson
Landau
Lognormal
NegativeLogarithms
Novosibirsk
Poisson
Polynomial
ProdPdf
Ratio
Voigtian

Definition at line 39 of file RooBatchCompute.h.

Function Documentation

fast_exp()

__roodevice__ double RooBatchCompute::fast_exp ( double  x )

inline

Definition at line 53 of file RooVDTHeaders.h.

fast_isqrt()

__roodevice__ double RooBatchCompute::fast_isqrt ( double  x )

inline

Definition at line 61 of file RooVDTHeaders.h.

fast_log()

__roodevice__ double RooBatchCompute::fast_log ( double  x )

inline

Definition at line 57 of file RooVDTHeaders.h.

init()

void RooBatchCompute::init

(

)

Inspect hardware capabilities, and load the optimal library for RooFit computations.

Definition at line 43 of file Initialisation.cxx.

Variable Documentation

bufferSize

constexpr uint16_t RooBatchCompute::bufferSize = 64

constexpr

Definition at line 38 of file Batches.h.

dispatchCPU

RooBatchCompute::RooBatchComputeInterface * RooBatchCompute::dispatchCPU = nullptr

This dispatch pointer points to an implementation of the compute library, provided one has been loaded.

Using a virtual call, computation requests are dispatched to backends with architecture-specific functions such as SSE, AVX, AVX2, etc.

See also

RooBatchComputeInterface, RooBatchComputeClass, RF_ARCH

Definition at line 93 of file RooBatchCompute.h.

dispatchCUDA

RooBatchCompute::RooBatchComputeInterface * RooBatchCompute::dispatchCUDA = nullptr

Definition at line 93 of file RooBatchCompute.h.