gemv_omp.h File Reference

Go to the source code of this file.

Functions
void	gemv_fused_q5_0_bias_parallel_omp (float *y, const void *W, const float *x, const float *bias, int M, int K)
void	gemv_q5_0_q8_0_parallel_omp (float *y, const void *W, const void *x_q8, int M, int K)
void	gemv_q8_0_q8_0_parallel_omp (float *y, const void *W, const void *x_q8, int M, int K)

Function Documentation

gemv_fused_q5_0_bias_parallel_omp()

void gemv_fused_q5_0_bias_parallel_omp	(	float *	y,
		const void *	W,
		const float *	x,
		const float *	bias,
		int	M,
		int	K
	)

Definition at line 98 of file gemv_omp.c.

{
    const block_q5_0 *w_blocks = (const block_q5_0 *)W;
    const int blocks_per_row = K / QK5_0;
 
    /* Quantize input ONCE (serial, fast — K=896 → 28 blocks = 952 bytes) */
    block_q8_0 x_q8[K / QK8_0];
    quantize_row_q8_0(x, (void *)x_q8, K);
 
    /* Parallel GEMV over output rows */
    #pragma omp parallel for schedule(static)
    for (int row = 0; row < M; row++) {
        vec_dot_q5_0_q8_0(K, &y[row],
                          &w_blocks[row * blocks_per_row],
                          x_q8);
        if (bias) y[row] += bias[row];
    }
}

References QK5_0, QK8_0, quantize_row_q8_0(), and vec_dot_q5_0_q8_0().

gemv_q5_0_q8_0_parallel_omp()

void gemv_q5_0_q8_0_parallel_omp	(	float *	y,
		const void *	W,
		const void *	x_q8,
		int	M,
		int	K
	)

Definition at line 74 of file gemv_omp.c.

{
    const block_q5_0 *w_blocks = (const block_q5_0 *)W;
    const block_q8_0 *x_blocks = (const block_q8_0 *)x_q8;
    const int blocks_per_row = K / QK5_0;
 
    #pragma omp parallel for schedule(static)
    for (int row = 0; row < M; row++) {
        vec_dot_q5_0_q8_0(K, &y[row],
                          &w_blocks[row * blocks_per_row],
                          x_blocks);
    }
}

References QK5_0, and vec_dot_q5_0_q8_0().

gemv_q8_0_q8_0_parallel_omp()

void gemv_q8_0_q8_0_parallel_omp	(	float *	y,
		const void *	W,
		const void *	x_q8,
		int	M,
		int	K
	)

Definition at line 52 of file gemv_omp.c.

{
    const block_q8_0 *w_blocks = (const block_q8_0 *)W;
    const block_q8_0 *x_blocks = (const block_q8_0 *)x_q8;
    const int blocks_per_row = K / QK8_0;
 
    #pragma omp parallel for schedule(static)
    for (int row = 0; row < M; row++) {
        vec_dot_q8_0_q8_0(K, &y[row],
                          &w_blocks[row * blocks_per_row],
                          x_blocks);
    }
}

References QK8_0, and vec_dot_q8_0_q8_0().