detail::KernelGemm< T, Bits, Arch > Struct Template Reference

#include <kernel_gemm.h>

Public Types
using	K = Microkernel< T, Bits, Arch >
using	Helpers = KernelHelpers< T, Bits, Arch >

Static Public Member Functions
static void	gemm (const T *A, my_size_t M, my_size_t K_len, my_size_t strideA, const T *B, my_size_t N, my_size_t strideB, T *C, my_size_t strideC) noexcept
	Register-blocked GEMM: C[M,N] = A[M,K] × B[K,N].

Static Public Attributes
static constexpr my_size_t	simdWidth = K::simdWidth
static constexpr my_size_t	MR = K::MR
	Tile height: rows of C computed per micro-kernel invocation.
static constexpr my_size_t	NR_VECS = K::NR_VECS
	Number of SIMD vectors per tile column. The tile width is NR = NR_VECS × simdWidth.
static constexpr my_size_t	NR = K::NR
	Tile width: columns of C computed per wide micro-kernel invocation.

Member Typedef Documentation

Helpers

template<typename T , my_size_t Bits, typename Arch >

using detail::KernelGemm< T, Bits, Arch >::Helpers = KernelHelpers<T, Bits, Arch>

K

template<typename T , my_size_t Bits, typename Arch >

using detail::KernelGemm< T, Bits, Arch >::K = Microkernel<T, Bits, Arch>

Member Function Documentation

gemm()

template<typename T , my_size_t Bits, typename Arch >

static void detail::KernelGemm< T, Bits, Arch >::gemm ( const T *  A, my_size_t  M, my_size_t  K_len, my_size_t  strideA, const T *  B, my_size_t  N, my_size_t  strideB, T *  C, my_size_t  strideC  )

inlinestaticnoexcept

Register-blocked GEMM: C[M,N] = A[M,K] × B[K,N].

Top-level dispatcher that tiles the output matrix and routes each tile to the appropriate micro-kernel based on its position.

All pointers address raw physical memory with padded row strides. The caller must ensure the favorable layout (see Memory Layout Requirements).

Parameters

A	Pointer to first element of A
M	Number of rows of A (and C)
K_len	Contraction length (columns of A, rows of B)
strideA	Physical row stride of A (≥ K_len, includes padding)
B	Pointer to first element of B
N	Number of columns of B (and C)
strideB	Physical row stride of B (≥ N, includes padding)
C	Pointer to first element of C (output, zero-initialized not required)
strideC	Physical row stride of C (≥ N, includes padding)

Here is the call graph for this function:

Member Data Documentation

MR

template<typename T , my_size_t Bits, typename Arch >

constexpr my_size_t detail::KernelGemm< T, Bits, Arch >::MR = K::MR

staticconstexpr

Tile height: rows of C computed per micro-kernel invocation.

NR

template<typename T , my_size_t Bits, typename Arch >

constexpr my_size_t detail::KernelGemm< T, Bits, Arch >::NR = K::NR

staticconstexpr

Tile width: columns of C computed per wide micro-kernel invocation.

NR_VECS

template<typename T , my_size_t Bits, typename Arch >

constexpr my_size_t detail::KernelGemm< T, Bits, Arch >::NR_VECS = K::NR_VECS

staticconstexpr

Number of SIMD vectors per tile column. The tile width is NR = NR_VECS × simdWidth.

simdWidth

template<typename T , my_size_t Bits, typename Arch >

constexpr my_size_t detail::KernelGemm< T, Bits, Arch >::simdWidth = K::simdWidth

staticconstexpr

---

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

• core/include/fused/kernel_ops/kernel_gemm.h