PerformanceEstimation.hpp File Reference

This file is for internal use only and should not be used. More...

#include "popops/Expr.hpp"
#include <poplar/PerfEstimateFunc.hpp>
#include <poplar/Target.hpp>
#include <poplar/VectorLayout.hpp>
#include <popops/OperationDef.hpp>

Go to the source code of this file.

Classes
struct	popops::internal::MultiSliceTargetParameters
	Cycle estimate for MultiSlice vertex. More...
struct	popops::internal::MultiUpdateOpTargetParameters
	Cycle estimate for MultiUpdateAdd vertex. More...
struct	popops::internal::CastTargetParameters
	Target parameters used in cast estimation. More...

Namespaces
namespace	popops
	Common functions, such as elementwise and reductions.

Functions
std::uint64_t	popops::internal::basicOpSupervisorOverhead (const bool isScaledPtr64Type=false)
	Supervisor context cycle estimate for BinaryOp and UnaryOp Supervisor/ MultiVertex codelets. More...
std::uint64_t	popops::internal::basicOpLoopCycles (const unsigned numElems, const unsigned vectorSize, const unsigned cyclesPerVector)
	Cycle cost for processing an arbitrary number of elements given the cycle cost for processing a vector of elements simultaneously. More...
std::uint64_t	popops::internal::binaryOpInnerLoopCycles (const poplar::Target &target, const poplar::Type &type, const unsigned cyclesPerVector, const bool vectorize, const unsigned numElems, const std::uint64_t overheadPerLoop)
	Cycle cost for processing an arbitrary number of elements given the cycle cost for processing a vector of elements simultaneously. More...
std::uint64_t	popops::internal::getDynamicSlice1DEstimate (const poplar::Target &target, const poplar::Type &type, const unsigned regionSize, const unsigned numSubElements)
	Cycle estimate for Dynamic Slice 1D vertex. More...
std::uint64_t	popops::internal::getBinaryOp1DInPlaceEstimate (const poplar::Target &target, const poplar::Type &type, const popops::expr::BinaryOpType op, const unsigned numElems)
	Cycle estimate for Binary-1D In-Place MultiVertex. More...

Detailed Description

This file is for internal use only and should not be used.

Function Documentation

basicOpLoopCycles()

std::uint64_t popops::internal::basicOpLoopCycles	(	const unsigned	numElems,
		const unsigned	vectorSize,
		const unsigned	cyclesPerVector
	)

Cycle cost for processing an arbitrary number of elements given the cycle cost for processing a vector of elements simultaneously.

Parameters

numElems	Number of elements in the operand.
vectorSize	Size of vector of elements that can be processed simultaneously.
cyclesPerVector	Cycle cost per vector of elements.

Returns

Estimated number of cycles.

basicOpSupervisorOverhead()

std::uint64_t popops::internal::basicOpSupervisorOverhead

(

const bool

isScaledPtr64Type = false

)

Supervisor context cycle estimate for BinaryOp and UnaryOp Supervisor/ MultiVertex codelets.

Parameters

isScaledPtr64Type

Is true if vector vertex field is ScaledPtr64 and false otherwise.

Returns

Estimated number of cycles.

binaryOpInnerLoopCycles()

std::uint64_t popops::internal::binaryOpInnerLoopCycles	(	const poplar::Target &	target,
		const poplar::Type &	type,
		const unsigned	cyclesPerVector,
		const bool	vectorize,
		const unsigned	numElems,
		const std::uint64_t	overheadPerLoop
	)

Cycle cost for processing an arbitrary number of elements given the cycle cost for processing a vector of elements simultaneously.

Parameters

target	The target on which the operation should be estimated
type	Data type of elements.
cyclesPerVector	Cycle cost per vector of elements.
vectorSize	Size of vector of elements that be processed simultaneously.
numElems	Number of elements in the operand.
overheadPerLoop	Cycles overhead per loop.

Returns

Estimated number of cycles.

getBinaryOp1DInPlaceEstimate()

std::uint64_t popops::internal::getBinaryOp1DInPlaceEstimate	(	const poplar::Target &	target,
		const poplar::Type &	type,
		const popops::expr::BinaryOpType	op,
		const unsigned	numElems
	)

Cycle estimate for Binary-1D In-Place MultiVertex.

Parameters

target	The target on which the operation should be estimated.
type	Data type of elements.
op	Binary operator type.
numElems	Total number of output elements.

Returns

Estimated number of cycles.

getDynamicSlice1DEstimate()

std::uint64_t popops::internal::getDynamicSlice1DEstimate	(	const poplar::Target &	target,
		const poplar::Type &	type,
		const unsigned	regionSize,
		const unsigned	numSubElements
	)

Cycle estimate for Dynamic Slice 1D vertex.

Parameters

target	The target on which the operation should be estimated.
type	Data type of elements.
regionSize	Number of elements in tensor to be sliced.
numSubElements	Number of elements per slice.

Returns

Estimated number of cycles.