#include "ckernel_engine.h"

Functions
CKMathBackend	ckernel_backend_native (void)

static void	ckernel_sgemm_native (int M, int N, int K, const float A, int lda, const float B, int ldb, const float bias, float C, int ldc)

Function Documentation

◆ ckernel_backend_native()

CKMathBackend ckernel_backend_native ( void )

Obtain the built-in native backend (single-node CPU, C + intrinsics).

Definition at line 39 of file backend_native.c.

{
    CKMathBackend b;
    b.sgemm = &ckernel_sgemm_native;
    return b;
}

References ckernel_sgemm_native(), and CKMathBackend::sgemm.

◆ ckernel_sgemm_native()

static void ckernel_sgemm_native	(	int	M,
		int	N,
		int	K,
		const float *	A,
		int	lda,
		const float *	B,
		int	ldb,
		const float *	bias,
		float *	C,
		int	ldc
	)

static

Native GEMM backend that directly reuses the C-Transformer GEMM kernel.

Layout assumptions (identical to C-Transformer/main.c):

A: [M x K], row-major, A(i,k) = A[i*K + k]
B: [N x K], row-major, B(j,k) = B[j*K + k]
C: [M x N], row-major, C(i,j) = C[i*N + j]

This is a straight copy of gemm_blocked_serial with a thin wrapper to match the CKMathBackend.sgemm signature. It is intentionally minimal.

Definition at line 18 of file backend_native.c.

{
    /* Honor caller-provided strides so padded matrices still compute correctly. */
    for (int i = 0; i < M; ++i) {
        const float *a_row = A + (size_t)i * lda;
        float *c_row = C + (size_t)i * ldc;
        for (int j = 0; j < N; ++j) {
            float sum = bias ? bias[j] : 0.0f;
            const float *b_row = B + (size_t)j * ldb;
            for (int k = 0; k < K; ++k) {
                sum += a_row[k] * b_row[k];
            }
            c_row[j] = sum;
        }
    }
}

References C.

Referenced by ckernel_backend_native().

Functions

Function Documentation

◆ ckernel_backend_native()

◆ ckernel_sgemm_native()