CPU tensor kernels

Always-available CPU reference tensor kernels. More...

Collaboration diagram for CPU tensor kernels:

Classes
class	uni20::kernel::cpu::GemmLoop< T, MR, NR, KR >
	Generic M×N×K contraction engine with full K-dimension recursion. More...

Functions
template<typename T , std::size_t MR, std::size_t NR, std::size_t KR>
void	uni20::kernel::contract_strided (static_vector< extent_strides< 2 >, MR > const &Mgrp, static_vector< extent_strides< 2 >, NR > const &Ngrp, static_vector< extent_strides< 2 >, KR > const &Kgrp, T alpha, T const *A, T const *B, T beta, T *C)
	Execute the CPU tensor contraction using precomputed stride groupings.
	uni20::kernel::cpu::GemmLoop< T, MR, NR, KR >::GemmLoop (static_vector< extent_strides< 2 >, MR > const &Mgrp, static_vector< extent_strides< 2 >, NR > const &Ngrp, static_vector< extent_strides< 2 >, KR > const &Kgrp, T alpha, T beta) noexcept
	Build the loop engine for a fused contraction.
void	uni20::kernel::cpu::GemmLoop< T, MR, NR, KR >::run (T const *A0, T const *B0, T *C0) noexcept
	Perform C = β·C + α·(A ⋅ B) over all fused M, N, and K dimensions.

Detailed Description

Always-available CPU reference tensor kernels.

Function Documentation

contract_strided()

template<typename T , std::size_t MR, std::size_t NR, std::size_t KR>

void uni20::kernel::contract_strided	(	static_vector< extent_strides< 2 >, MR > const &	Mgrp,
		static_vector< extent_strides< 2 >, NR > const &	Ngrp,
		static_vector< extent_strides< 2 >, KR > const &	Kgrp,
		T	alpha,
		T const *	A,
		T const *	B,
		T	beta,
		T *	C
	)

Execute the CPU tensor contraction using precomputed stride groupings.

Template Parameters

T	Scalar type stored in the tensors.
MR	Number of fused M dimensions.
NR	Number of fused N dimensions.
KR	Number of fused K dimensions.

Parameters

Mgrp	Metadata describing extents and strides for the fused M dimensions.
Ngrp	Metadata describing extents and strides for the fused N dimensions.
Kgrp	Metadata describing extents and strides for the fused K dimensions.
alpha	Scaling factor applied to the contraction output.
A	Pointer to the base of the left-hand operand.
B	Pointer to the base of the right-hand operand.
beta	Scaling factor applied to the pre-existing contents of the destination tensor.
C	Pointer to the base of the destination tensor.

GemmLoop()

template<typename T , std::size_t MR, std::size_t NR, std::size_t KR>

uni20::kernel::cpu::GemmLoop< T, MR, NR, KR >::GemmLoop ( static_vector< extent_strides< 2 >, MR > const &  Mgrp, static_vector< extent_strides< 2 >, NR > const &  Ngrp, static_vector< extent_strides< 2 >, KR > const &  Kgrp, T  alpha, T  beta  )

inlinenoexcept

Build the loop engine for a fused contraction.

Parameters

Mgrp	Metadata describing extents and strides for the fused M dimensions.
Ngrp	Metadata describing extents and strides for the fused N dimensions.
Kgrp	Metadata describing extents and strides for the fused K dimensions.
alpha	Scaling factor applied to the contraction output.
beta	Scaling factor applied to the pre-existing contents of the destination tensor.

run()

template<typename T , std::size_t MR, std::size_t NR, std::size_t KR>

void uni20::kernel::cpu::GemmLoop< T, MR, NR, KR >::run ( T const *  A0, T const *  B0, T *  C0  )

inlinenoexcept

Perform C = β·C + α·(A ⋅ B) over all fused M, N, and K dimensions.

Parameters

A0	Pointer to the base of the left-hand operand.
B0	Pointer to the base of the right-hand operand.
C0	Pointer to the base of the destination tensor.