clustering::math::GemmPlan< T, Backend > Class Template Reference

Reusable GEMM plan: packs B once at construction, amortizes the packing cost across repeated execute calls with varying A. More...

#include <clustering/math/gemm_plan.h>

Public Member Functions
template<Layout LB>
	GemmPlan (const NDArray< T, 2, LB > &B, Pool pool)
	Construct the plan and fully pre-pack B into m_Bp.
template<Layout LA>
void	execute (const NDArray< T, 2, LA > &A, NDArray< T, 2 > &C, T alpha=T{1}, T beta=T{0}) const noexcept
	Execute the plan: compute C := alpha * A * B + beta * C against the pre-packed B captured at construction.
std::size_t	kDim () const noexcept
	Inner dimension captured at construction (B.rows).
std::size_t	nDim () const noexcept
	Column count captured at construction (B.cols).
const T *	debugBpData () const noexcept
	Debug accessor exposing the packed B pointer so tests can pin alignment.
std::size_t	debugScratchSize () const noexcept
	Debug accessor exposing the scratch capacity so tests can pin the sizing formula.
	GemmPlan (const GemmPlan &)=delete
GemmPlan &	operator= (const GemmPlan &)=delete
	GemmPlan (GemmPlan &&) noexcept=default
	Defaulted move constructor; transfers the packed B panel and scratch.
GemmPlan &	operator= (GemmPlan &&) noexcept=default
	Defaulted move assignment; transfers the packed B panel and scratch.
	~GemmPlan ()=default

Detailed Description

template<class T, class Backend = detail::ReferenceGemm>
class clustering::math::GemmPlan< T, Backend >

Reusable GEMM plan: packs B once at construction, amortizes the packing cost across repeated execute calls with varying A.

Ownership model:

• m_Bp holds a packB-shaped copy of B covering every (jcIdx, pcIdx) block; the source B may be destroyed after construction.
• m_scratch holds per-worker A-packing arenas sized workerCount * kMc * kKc.
• m_pool is captured by value at construction. execute uses it unchanged.

Concurrency: execute is const with respect to the plan's observable shape but mutates m_scratch via pointer aliasing into the per-worker slice. Two concurrent execute calls on the same plan alias the same scratch; callers that pipeline across application threads construct one plan per thread.

Lifetime: move-only. m_Bp and m_scratch are std::vector; move is noexcept.

Template Parameters

T	Element type (float or double).
Backend	Backend tag; defaulted to detail::ReferenceGemm. GemmPlan does NOT route through the backend – it calls gemmRunPrepacked directly with its owned m_Bp. The template parameter exists so GemmPlan is selected at the same <T, Backend> call site as the one-shot gemm entry.

Definition at line 40 of file gemm_plan.h.

Constructor & Destructor Documentation

GemmPlan() [1/3]

template<class T, class Backend = detail::ReferenceGemm>

template<Layout LB>

clustering::math::GemmPlan< T, Backend >::GemmPlan ( const NDArray< T, 2, LB > & B, Pool pool )

inline

Construct the plan and fully pre-pack B into m_Bp.

Template Parameters

LB	Layout tag of B (Contig or MaybeStrided).

Parameters

B	Source matrix (K x N); contents are copied into m_Bp, the source may outlive the plan or vice versa without coupling.
pool	Pool captured by value for use by execute; pool.workerCount() determines m_scratch size and is fixed for the plan's lifetime.

Definition at line 55 of file gemm_plan.h.

GemmPlan() [2/3]

template<class T, class Backend = detail::ReferenceGemm>

clustering::math::GemmPlan< T, Backend >::GemmPlan ( const GemmPlan< T, Backend > & )

delete

GemmPlan() [3/3]

template<class T, class Backend = detail::ReferenceGemm>

clustering::math::GemmPlan< T, Backend >::GemmPlan ( GemmPlan< T, Backend > && )

defaultnoexcept

Defaulted move constructor; transfers the packed B panel and scratch.

~GemmPlan()

template<class T, class Backend = detail::ReferenceGemm>

clustering::math::GemmPlan< T, Backend >::~GemmPlan ( )

default

Member Function Documentation

debugBpData()

template<class T, class Backend = detail::ReferenceGemm>

const T * clustering::math::GemmPlan< T, Backend >::debugBpData ( ) const

inlinenodiscardnoexcept

Debug accessor exposing the packed B pointer so tests can pin alignment.

Not part of the stable API.

Definition at line 131 of file gemm_plan.h.

debugScratchSize()

template<class T, class Backend = detail::ReferenceGemm>

std::size_t clustering::math::GemmPlan< T, Backend >::debugScratchSize ( ) const

inlinenodiscardnoexcept

Debug accessor exposing the scratch capacity so tests can pin the sizing formula.

Not part of the stable API.

Definition at line 135 of file gemm_plan.h.

execute()

template<class T, class Backend = detail::ReferenceGemm>

template<Layout LA>

void clustering::math::GemmPlan< T, Backend >::execute ( const NDArray< T, 2, LA > & A, NDArray< T, 2 > & C, T alpha = T{1}, T beta = T{0} ) const

inlinenoexcept

Execute the plan: compute C := alpha * A * B + beta * C against the pre-packed B captured at construction.

Template Parameters

LA	Layout tag of A.

Parameters

A	Input matrix (M x kDim()).
C	Output matrix (M x nDim()); must be mutable.
alpha	Scalar multiplier on A*B; defaults to 1.
beta	Scalar multiplier on the prior C; defaults to 0.

Definition at line 109 of file gemm_plan.h.

kDim()

template<class T, class Backend = detail::ReferenceGemm>

std::size_t clustering::math::GemmPlan< T, Backend >::kDim ( ) const

inlinenodiscardnoexcept

Inner dimension captured at construction (B.rows).

Definition at line 124 of file gemm_plan.h.

nDim()

template<class T, class Backend = detail::ReferenceGemm>

std::size_t clustering::math::GemmPlan< T, Backend >::nDim ( ) const

inlinenodiscardnoexcept

Column count captured at construction (B.cols).

Definition at line 127 of file gemm_plan.h.

operator=() [1/2]

template<class T, class Backend = detail::ReferenceGemm>

GemmPlan & clustering::math::GemmPlan< T, Backend >::operator= ( const GemmPlan< T, Backend > & )

delete

operator=() [2/2]

template<class T, class Backend = detail::ReferenceGemm>

GemmPlan & clustering::math::GemmPlan< T, Backend >::operator= ( GemmPlan< T, Backend > && )

defaultnoexcept

Defaulted move assignment; transfers the packed B panel and scratch.

---

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

• include/clustering/math/gemm_plan.h