gemm_kernels_q5_k.c File Reference

GEMM/GEMV kernels with Q5_K quantized weights. More...

#include <stdint.h>
#include <stddef.h>
#include <string.h>
#include <stdlib.h>
#include "ckernel_quant.h"

Go to the source code of this file.

Macros
#define	CK_Q5K_STACK_Q8_BLOCKS   128
#define	QK_K   256

Functions
void	ck_q5_k_prepare_weight (const void *src, void *dst, int N, int K)
size_t	ck_q5_k_prepared_block_size (void)
static int	ck_q5k_debug_fp32_fallback (void)
static int	ck_q5k_debug_generic_dot (void)
static float	dot_q5_k_q8_k_row (const block_q5_K *w, const block_q8_K *x, int nb)
void	gemm_nt_q5_k (const float *A, const void *B, const float *bias, float *C, int M, int N, int K)
void	gemm_nt_q5_k_prepared (const float *A, const void *B_prepared, const float *bias, float *C, int M, int N, int K)
void	gemm_nt_q5_k_prepared_m4 (const float *A, const void *B_prepared, const float *bias, float *C, int M, int N, int K)
void	gemm_nt_q5_k_prepared_q8_m4_nrange (const void *A_q8, const void *B_prepared, const float *bias, float *C, int M, int N, int K, int n_begin, int n_end)
void	gemm_nt_q5_k_q8_k (const void *A_q8, const void *B, const float *bias, float *C, int M, int N, int K)
void	gemm_nt_q5_k_q8_k_ref (const void *A_q8, const void *B, const float *bias, float *C, int M, int N, int K)
void	gemm_nt_q5_k_ref (const float *A, const void *B, const float *bias, float *C, int M, int N, int K)
static void	gemm_nt_q5_k_ref_fp32 (const float *A, const void *B, const float *bias, float *C, int M, int N, int K)
void	gemm_q5_k_q8_k_compact_rows4 (float *output, int output_stride, const void *weights, const void *const input_rows[4], int rows, int output_dim, int input_dim)
void	gemv_q5_k (float *y, const void *W, const float *x, int M, int K)
void	gemv_q5_k_q8_k (float *y, const void *W, const void *x_q8, int M, int K)
void	gemv_q5_k_q8_k_ref (float *y, const void *W, const void *x_q8, int M, int K)
void	gemv_q5_k_ref (float *y, const void *W, const float *x, int M, int K)
static void	gemv_q5_k_ref_fp32 (float *y, const void *W, const float *x, int M, int K)
static uint8_t	q5_k_quant_value (const block_q5_K *block, int subblock, int i)
void	quantize_row_q8_k (const float *x, void *vy, int k)
static void	unpack_q5_k_scales (const uint8_t *scales, uint8_t *sc, uint8_t *m)

Detailed Description

GEMM/GEMV kernels with Q5_K quantized weights.

CK-ENGINE KERNEL RULES:

1. NO malloc/free - memory via bump allocator, pointers passed in
2. NO OpenMP - parallelization at orchestrator/codegen layer
3. API must define: inputs, outputs, workspace, and memory layouts
4. Pure computation - deterministic, no side effects

After changes: make test && make llamacpp-parity-full

Implements matrix multiplication where:

• Activations (input): FP32 (quantized internally to Q8_K for dot path)
• Weights: Q5_K (5-bit super-block quant)
• Output: FP32

Q5_K Format (256 weights per super-block):

• d: FP16 super-block scale
• dmin: FP16 super-block minimum
• scales[12]: 8 sub-block scales + 8 sub-block mins (6 bits each, packed)
• qh[32]: high bits for 256 weights (1 bit each)
• qs[128]: low 4 bits for 256 weights (4 bits each)

Total: 2 + 2 + 12 + 32 + 128 = 176 bytes per 256 weights = 5.5 bits/weight

Dequantization formula (matches llama.cpp): w = d * scale * q - dmin * mins where q = qs_val | (qh_bit << 4) = 5-bit value [0, 31]

Definition in file gemm_kernels_q5_k.c.

Macro Definition Documentation

CK_Q5K_STACK_Q8_BLOCKS

#define CK_Q5K_STACK_Q8_BLOCKS   128

Definition at line 46 of file gemm_kernels_q5_k.c.

QK_K

#define QK_K   256

Definition at line 45 of file gemm_kernels_q5_k.c.

Function Documentation

ck_q5_k_prepare_weight()

void ck_q5_k_prepare_weight	(	const void *	src,
		void *	dst,
		int	N,
		int	K
	)

Definition at line 131 of file gemm_kernels_q5_k.c.

{
    if (!src || !dst || N <= 0 || K <= 0 || (K % QK_K) != 0) return;
    const block_q5_K *input = (const block_q5_K *)src;
    block_q5_K_prepared *output = (block_q5_K_prepared *)dst;
    const size_t blocks = (size_t)N * (size_t)(K / QK_K);
    for (size_t b = 0; b < blocks; ++b) {
        output[b].d = input[b].d;
        output[b].dmin = input[b].dmin;
        unpack_q5_k_scales(input[b].scales, output[b].scales, output[b].mins);
        for (int sb = 0; sb < 8; ++sb) {
            for (int i = 0; i < 32; ++i) {
                output[b].qs[sb * 32 + i] = q5_k_quant_value(&input[b], sb, i);
            }
        }
    }
}

References q5_k_quant_value(), QK_K, and unpack_q5_k_scales().

ck_q5_k_prepared_block_size()

size_t ck_q5_k_prepared_block_size

(

void

)

Definition at line 126 of file gemm_kernels_q5_k.c.

{
    return sizeof(block_q5_K_prepared);
}

ck_q5k_debug_fp32_fallback()

static int ck_q5k_debug_fp32_fallback ( void  )

static

Definition at line 48 of file gemm_kernels_q5_k.c.

{
    static int cached = -1;
    if (cached < 0) {
        const char *env = getenv("CK_DEBUG_Q5K_FP32_FALLBACK");
        cached = (env && env[0] && env[0] != '0') ? 1 : 0;
    }
    return cached;
}

Referenced by gemm_nt_q5_k_ref(), and gemv_q5_k_ref().

ck_q5k_debug_generic_dot()

static int ck_q5k_debug_generic_dot ( void  )

static

Definition at line 58 of file gemm_kernels_q5_k.c.

{
    static int cached = -1;
    if (cached < 0) {
        const char *env = getenv("CK_DEBUG_Q5K_GENERIC_DOT");
        cached = (env && env[0] && env[0] != '0') ? 1 : 0;
    }
    return cached;
}

Referenced by dot_q5_k_q8_k_row().

dot_q5_k_q8_k_row()

static float dot_q5_k_q8_k_row ( const block_q5_K *  w, const block_q8_K *  x, int  nb  )

static

Definition at line 478 of file gemm_kernels_q5_k.c.

                                                                                 {
#if defined(__AVX2__)
    if (!ck_q5k_debug_generic_dot()) {
        return dot_q5_k_q8_k_row_avx2(w, x, nb);
    }
#endif
 
    static const uint32_t kmask1 = 0x3f3f3f3fU;
    static const uint32_t kmask2 = 0x0f0f0f0fU;
    static const uint32_t kmask3 = 0x03030303U;
 
    uint32_t utmp[4] = {0, 0, 0, 0};
    const uint8_t *scales = (const uint8_t *)&utmp[0];
    const uint8_t *mins = (const uint8_t *)&utmp[2];
 
    int8_t aux8[QK_K];
    int16_t aux16[8];
    float sums[8] = {0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f};
    int32_t aux32[8];
 
    float sumf = 0.0f;
    for (int b = 0; b < nb; ++b) {
        const block_q5_K *wb = &w[b];
        const block_q8_K *xb = &x[b];
        const uint8_t *q4 = wb->qs;
        const uint8_t *hm = wb->qh;
        int8_t *a = aux8;
        uint8_t m = 1;
        memset(aux32, 0, sizeof(aux32));
 
        for (int j = 0; j < QK_K / 64; ++j) {
            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
            a += 32;
            m <<= 1;
 
            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4);
            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
            a += 32;
            m <<= 1;
 
            q4 += 32;
        }
 
        memcpy(utmp, wb->scales, 12);
        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
        const uint32_t uaux = utmp[1] & kmask1;
        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
        utmp[2] = uaux;
        utmp[0] &= kmask1;
 
        int sumi = 0;
        for (int j = 0; j < QK_K / 16; ++j) {
            sumi += (int)xb->bsums[j] * (int)mins[j / 2];
        }
 
        a = aux8;
        const int8_t *q8 = xb->qs;
        int is = 0;
        for (int j = 0; j < QK_K / 32; ++j) {
            const int32_t scale = (int32_t)scales[is++];
 
            for (int l = 0; l < 8; ++l) aux16[l] = (int16_t)(q8[l] * a[l]);
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
 
            for (int l = 0; l < 8; ++l) aux16[l] = (int16_t)(q8[l] * a[l]);
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
 
            for (int l = 0; l < 8; ++l) aux16[l] = (int16_t)(q8[l] * a[l]);
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
 
            for (int l = 0; l < 8; ++l) aux16[l] = (int16_t)(q8[l] * a[l]);
            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
            q8 += 8; a += 8;
        }
 
        const float d = CK_FP16_TO_FP32(wb->d) * xb->d;
        for (int l = 0; l < 8; ++l) {
            sums[l] += d * (float)aux32[l];
        }
        const float dmin = CK_FP16_TO_FP32(wb->dmin) * xb->d;
        sumf -= dmin * (float)sumi;
    }
 
    for (int l = 0; l < 8; ++l) {
        sumf += sums[l];
    }
    return sumf;
}

References block_q8_K::bsums, CK_FP16_TO_FP32, ck_q5k_debug_generic_dot(), block_q8_K::d, QK_K, and block_q8_K::qs.

Referenced by gemm_nt_q5_k_q8_k_ref(), and gemv_q5_k_q8_k_ref().

gemm_nt_q5_k()

void gemm_nt_q5_k	(	const float *	A,
		const void *	B,
		const float *	bias,
		float *	C,
		int	M,
		int	N,
		int	K
	)

Definition at line 1001 of file gemm_kernels_q5_k.c.

{
#if defined(__AVX512F__)
    /* TODO: AVX-512 implementation */
    gemm_nt_q5_k_ref(A, B, bias, C, M, N, K);
#elif defined(__AVX2__)
    /* TODO: AVX-2 implementation */
    gemm_nt_q5_k_ref(A, B, bias, C, M, N, K);
#elif defined(__AVX__)
    /* TODO: AVX implementation */
    gemm_nt_q5_k_ref(A, B, bias, C, M, N, K);
#elif defined(__SSE4_1__)
    /* TODO: SSE4.1 implementation */
    gemm_nt_q5_k_ref(A, B, bias, C, M, N, K);
#else
    gemm_nt_q5_k_ref(A, B, bias, C, M, N, K);
#endif
}

References C, and gemm_nt_q5_k_ref().

gemm_nt_q5_k_prepared()

void gemm_nt_q5_k_prepared	(	const float *	A,
		const void *	B_prepared,
		const float *	bias,
		float *	C,
		int	M,
		int	N,
		int	K
	)

Definition at line 705 of file gemm_kernels_q5_k.c.

{
#if !defined(__AVX2__)
    (void)A; (void)B_prepared; (void)bias; (void)C;
    (void)M; (void)N; (void)K;
#else
    if (!A || !B_prepared || !C || M <= 0 || N <= 0 || K <= 0 ||
            (K % QK_K) != 0) return;
    const int blocks_per_row = K / QK_K;
    if (blocks_per_row > CK_Q5K_STACK_Q8_BLOCKS) return;
    const block_q5_K_prepared *W = (const block_q5_K_prepared *)B_prepared;
    for (int m = 0; m < M; ++m) {
        block_q8_K a_q8[blocks_per_row];
        quantize_row_q8_k(A + (size_t)m * K, a_q8, K);
        for (int n = 0; n < N; ++n) {
            const float sum = dot_q5_k_prepared_q8_k_row_avx2(
                    W + (size_t)n * blocks_per_row, a_q8, blocks_per_row);
            C[(size_t)m * N + n] = sum + (bias ? bias[n] : 0.0f);
        }
    }
#endif
}

References C, CK_Q5K_STACK_Q8_BLOCKS, QK_K, and quantize_row_q8_k().

gemm_nt_q5_k_prepared_m4()

void gemm_nt_q5_k_prepared_m4	(	const float *	A,
		const void *	B_prepared,
		const float *	bias,
		float *	C,
		int	M,
		int	N,
		int	K
	)

Definition at line 732 of file gemm_kernels_q5_k.c.

{
#if !defined(__AVX2__)
    (void)A; (void)B_prepared; (void)bias; (void)C;
    (void)M; (void)N; (void)K;
#else
    if (!A || !B_prepared || !C || M <= 0 || N <= 0 || K <= 0 ||
            (K % QK_K) != 0) return;
    const int blocks_per_row = K / QK_K;
    if (blocks_per_row > CK_Q5K_STACK_Q8_BLOCKS) return;
    const block_q5_K_prepared *W = (const block_q5_K_prepared *)B_prepared;
    for (int m = 0; m < M; m += 4) {
        const int rows = M - m < 4 ? M - m : 4;
        block_q8_K a_q8[4][blocks_per_row];
        const block_q8_K *row_ptrs[4] = {
            a_q8[0], a_q8[1], a_q8[2], a_q8[3],
        };
        for (int r = 0; r < rows; ++r) {
            quantize_row_q8_k(A + (size_t)(m + r) * K, a_q8[r], K);
        }
        for (int n = 0; n < N; ++n) {
            float sums[4];
            dot_q5_k_prepared_q8_k_m4_avx2(
                    W + (size_t)n * blocks_per_row,
                    row_ptrs, rows, blocks_per_row, sums);
            for (int r = 0; r < rows; ++r) {
                C[(size_t)(m + r) * N + n] = sums[r] + (bias ? bias[n] : 0.0f);
            }
        }
    }
#endif
}

References C, CK_Q5K_STACK_Q8_BLOCKS, QK_K, and quantize_row_q8_k().

gemm_nt_q5_k_prepared_q8_m4_nrange()

void gemm_nt_q5_k_prepared_q8_m4_nrange	(	const void *	A_q8,
		const void *	B_prepared,
		const float *	bias,
		float *	C,
		int	M,
		int	N,
		int	K,
		int	n_begin,
		int	n_end
	)

Definition at line 769 of file gemm_kernels_q5_k.c.

{
#if !defined(__AVX2__)
    (void)A_q8; (void)B_prepared; (void)bias; (void)C;
    (void)M; (void)N; (void)K; (void)n_begin; (void)n_end;
#else
    if (!A_q8 || !B_prepared || !C || M <= 0 || N <= 0 || K <= 0 ||
            (K % QK_K) != 0 || n_begin < 0 || n_end > N ||
            n_begin >= n_end) return;
    const int blocks_per_row = K / QK_K;
    if (blocks_per_row > CK_Q5K_STACK_Q8_BLOCKS) return;
    const block_q8_K *A = (const block_q8_K *)A_q8;
    const block_q5_K_prepared *W =
        (const block_q5_K_prepared *)B_prepared;
 
    for (int m = 0; m < M; m += 4) {
        const int rows = M - m < 4 ? M - m : 4;
        const block_q8_K *row_ptrs[4] = {
            A + (size_t)(m + 0) * blocks_per_row,
            A + (size_t)(m + (rows > 1 ? 1 : 0)) * blocks_per_row,
            A + (size_t)(m + (rows > 2 ? 2 : 0)) * blocks_per_row,
            A + (size_t)(m + (rows > 3 ? 3 : 0)) * blocks_per_row,
        };
        for (int n = n_begin; n < n_end; ++n) {
            float sums[4];
            dot_q5_k_prepared_q8_k_m4_avx2(
                W + (size_t)n * blocks_per_row,
                row_ptrs, rows, blocks_per_row, sums);
            for (int r = 0; r < rows; ++r) {
                C[(size_t)(m + r) * N + n] =
                    sums[r] + (bias ? bias[n] : 0.0f);
            }
        }
    }
#endif
}

References C, CK_Q5K_STACK_Q8_BLOCKS, and QK_K.

gemm_nt_q5_k_q8_k()

void gemm_nt_q5_k_q8_k	(	const void *	A_q8,
		const void *	B,
		const float *	bias,
		float *	C,
		int	M,
		int	N,
		int	K
	)

Definition at line 914 of file gemm_kernels_q5_k.c.

{
#if defined(__AVX512F__)
    /* TODO: AVX-512 implementation */
    gemm_nt_q5_k_q8_k_ref(A_q8, B, bias, C, M, N, K);
#elif defined(__AVX2__)
    /* TODO: AVX-2 implementation */
    gemm_nt_q5_k_q8_k_ref(A_q8, B, bias, C, M, N, K);
#elif defined(__AVX__)
    /* TODO: AVX implementation */
    gemm_nt_q5_k_q8_k_ref(A_q8, B, bias, C, M, N, K);
#elif defined(__SSE4_1__)
    /* TODO: SSE4.1 implementation */
    gemm_nt_q5_k_q8_k_ref(A_q8, B, bias, C, M, N, K);
#else
    gemm_nt_q5_k_q8_k_ref(A_q8, B, bias, C, M, N, K);
#endif
}

References C, and gemm_nt_q5_k_q8_k_ref().

gemm_nt_q5_k_q8_k_ref()

void gemm_nt_q5_k_q8_k_ref	(	const void *	A_q8,
		const void *	B,
		const float *	bias,
		float *	C,
		int	M,
		int	N,
		int	K
	)

Definition at line 678 of file gemm_kernels_q5_k.c.

{
    if (!A_q8 || !B || !C || M <= 0 || N <= 0 || K <= 0) {
        return;
    }
    if (K % QK_K != 0) {
        return;
    }
 
    const block_q8_K *A = (const block_q8_K *)A_q8;
    const block_q5_K *W = (const block_q5_K *)B;
    const int blocks_per_row = K / QK_K;
 
    for (int m = 0; m < M; ++m) {
        const block_q8_K *a_row = &A[m * blocks_per_row];
        for (int n = 0; n < N; ++n) {
            const block_q5_K *w_row = &W[n * blocks_per_row];
            const float sum = dot_q5_k_q8_k_row(w_row, a_row, blocks_per_row);
            C[m * N + n] = sum + (bias ? bias[n] : 0.0f);
        }
    }
}

References C, dot_q5_k_q8_k_row(), and QK_K.

Referenced by gemm_nt_q5_k_q8_k(), and gemm_nt_q5_k_ref().

gemm_nt_q5_k_ref()

void gemm_nt_q5_k_ref	(	const float *	A,
		const void *	B,
		const float *	bias,
		float *	C,
		int	M,
		int	N,
		int	K
	)

Definition at line 855 of file gemm_kernels_q5_k.c.

{
    if (!A || !B || !C || M <= 0 || N <= 0 || K <= 0) {
        return;
    }
    if (ck_q5k_debug_fp32_fallback()) {
        gemm_nt_q5_k_ref_fp32(A, B, bias, C, M, N, K);
        return;
    }
    if (K % QK_K != 0) {
        gemm_nt_q5_k_ref_fp32(A, B, bias, C, M, N, K);
        return;
    }
 
    const block_q5_K *blocks = (const block_q5_K *)B;
    const int blocks_per_col = K / QK_K;
    if (blocks_per_col > CK_Q5K_STACK_Q8_BLOCKS) {
        gemm_nt_q5_k_ref_fp32(A, B, bias, C, M, N, K);
        return;
    }
 
    for (int m = 0; m < M; ++m) {
        const float *a_row = &A[m * K];
        block_q8_K a_q8[CK_Q5K_STACK_Q8_BLOCKS];
        quantize_row_q8_k(a_row, a_q8, K);
        gemm_nt_q5_k_q8_k_ref(a_q8, blocks, bias, &C[m * N], 1, N, K);
    }
}

References C, ck_q5k_debug_fp32_fallback(), CK_Q5K_STACK_Q8_BLOCKS, gemm_nt_q5_k_q8_k_ref(), gemm_nt_q5_k_ref_fp32(), QK_K, and quantize_row_q8_k().

Referenced by gemm_nt_q5_k().

gemm_nt_q5_k_ref_fp32()

static void gemm_nt_q5_k_ref_fp32 ( const float *  A, const void *  B, const float *  bias, float *  C, int  M, int  N, int  K  )

static

Definition at line 603 of file gemm_kernels_q5_k.c.

{
    const block_q5_K *blocks = (const block_q5_K *)B;
    const int blocks_per_col = K / QK_K;
 
    for (int m = 0; m < M; m++) {
        const float *a_row = &A[m * K];
 
        for (int n = 0; n < N; n++) {
            float sum = 0.0f;
            const block_q5_K *w_row = &blocks[n * blocks_per_col];
            for (int b = 0; b < blocks_per_col; b++) {
                const block_q5_K *block = &w_row[b];
                const float d = CK_FP16_TO_FP32(block->d);
                const float dmin = CK_FP16_TO_FP32(block->dmin);
                uint8_t sc_arr[8], m_arr[8];
                unpack_q5_k_scales(block->scales, sc_arr, m_arr);
 
                for (int sb = 0; sb < 8; sb++) {
                    const float d_sub = d * (float)sc_arr[sb];
                    const float m_sub = dmin * (float)m_arr[sb];
 
                    for (int i = 0; i < 32; i++) {
                        const uint8_t q = q5_k_quant_value(block, sb, i);
                        sum += (d_sub * (float)q - m_sub) * a_row[b * QK_K + sb * 32 + i];
                    }
                }
            }
 
            C[m * N + n] = sum + (bias ? bias[n] : 0.0f);
        }
    }
}

References C, CK_FP16_TO_FP32, q5_k_quant_value(), QK_K, and unpack_q5_k_scales().

Referenced by gemm_nt_q5_k_ref().

gemm_q5_k_q8_k_compact_rows4()

void gemm_q5_k_q8_k_compact_rows4	(	float *	output,
		int	output_stride,
		const void *	weights,
		const void *const	input_rows[4],
		int	rows,
		int	output_dim,
		int	input_dim
	)

Definition at line 937 of file gemm_kernels_q5_k.c.

{
    if (!output || !weights || !input_rows || rows <= 0 || rows > 4 ||
        output_stride < output_dim || output_dim <= 0 || input_dim <= 0 ||
        (input_dim % QK_K) != 0) {
        return;
    }
    for (int row = 0; row < rows; ++row) {
        if (!input_rows[row]) return;
    }
 
#if defined(__AVX2__)
    const block_q5_K *blocks = (const block_q5_K *)weights;
    const int blocks_per_row = input_dim / QK_K;
    const block_q8_K *inputs[4] = {
        (const block_q8_K *)input_rows[0],
        (const block_q8_K *)input_rows[rows > 1 ? 1 : 0],
        (const block_q8_K *)input_rows[rows > 2 ? 2 : 0],
        (const block_q8_K *)input_rows[rows > 3 ? 3 : 0],
    };
    for (int n = 0; n < output_dim; ++n) {
        float values[4];
        dot_q5_k_q8_k_rows4_avx2(
            blocks + (size_t)n * (size_t)blocks_per_row,
            inputs, rows, blocks_per_row, values);
        for (int row = 0; row < rows; ++row) {
            output[(size_t)row * (size_t)output_stride + (size_t)n] =
                values[row];
        }
    }
#else
    for (int row = 0; row < rows; ++row) {
        gemv_q5_k_q8_k(
            output + (size_t)row * (size_t)output_stride,
            weights, input_rows[row], output_dim, input_dim);
    }
#endif
}

References gemv_q5_k_q8_k(), and QK_K.

Referenced by ck_moe_q4k_q5k_bucket_work().

gemv_q5_k()

void gemv_q5_k	(	float *	y,
		const void *	W,
		const float *	x,
		int	M,
		int	K
	)

Definition at line 982 of file gemm_kernels_q5_k.c.

{
#if defined(__AVX512F__)
    /* TODO: AVX-512 implementation */
    gemv_q5_k_ref(y, W, x, M, K);
#elif defined(__AVX2__)
    /* TODO: AVX-2 implementation */
    gemv_q5_k_ref(y, W, x, M, K);
#elif defined(__AVX__)
    /* TODO: AVX implementation */
    gemv_q5_k_ref(y, W, x, M, K);
#elif defined(__SSE4_1__)
    /* TODO: SSE4.1 implementation */
    gemv_q5_k_ref(y, W, x, M, K);
#else
    gemv_q5_k_ref(y, W, x, M, K);
#endif
}

References gemv_q5_k_ref().

gemv_q5_k_q8_k()

void gemv_q5_k_q8_k	(	float *	y,
		const void *	W,
		const void *	x_q8,
		int	M,
		int	K
	)

Definition at line 892 of file gemm_kernels_q5_k.c.

{
#if defined(__AVX512F__)
    /* TODO: AVX-512 implementation */
    gemv_q5_k_q8_k_ref(y, W, x_q8, M, K);
#elif defined(__AVX2__)
    /* TODO: AVX-2 implementation */
    gemv_q5_k_q8_k_ref(y, W, x_q8, M, K);
#elif defined(__AVX__)
    /* TODO: AVX implementation */
    gemv_q5_k_q8_k_ref(y, W, x_q8, M, K);
#elif defined(__SSE4_1__)
    /* TODO: SSE4.1 implementation */
    gemv_q5_k_q8_k_ref(y, W, x_q8, M, K);
#else
    gemv_q5_k_q8_k_ref(y, W, x_q8, M, K);
#endif
}

References gemv_q5_k_q8_k_ref().

Referenced by ck_moe_q4k_q5k_route_work(), gemm_q5_k_q8_k_compact_rows4(), and moe_swiglu_expert_forward_q4k_q5k_workspace().

gemv_q5_k_q8_k_ref()

void gemv_q5_k_q8_k_ref	(	float *	y,
		const void *	W,
		const void *	x_q8,
		int	M,
		int	K
	)

Definition at line 656 of file gemm_kernels_q5_k.c.

{
    if (!y || !W || !x_q8 || M <= 0 || K <= 0) {
        return;
    }
    if (K % QK_K != 0) {
        return;
    }
 
    const block_q5_K *blocks = (const block_q5_K *)W;
    const block_q8_K *x = (const block_q8_K *)x_q8;
    const int blocks_per_row = K / QK_K;
 
    for (int m = 0; m < M; ++m) {
        const block_q5_K *w_row = &blocks[m * blocks_per_row];
        y[m] = dot_q5_k_q8_k_row(w_row, x, blocks_per_row);
    }
}

References dot_q5_k_q8_k_row(), and QK_K.

Referenced by gemv_q5_k_q8_k(), and gemv_q5_k_ref().

gemv_q5_k_ref()

void gemv_q5_k_ref	(	float *	y,
		const void *	W,
		const float *	x,
		int	M,
		int	K
	)

Definition at line 819 of file gemm_kernels_q5_k.c.

{
    if (!y || !W || !x || M <= 0 || K <= 0) {
        return;
    }
    if (ck_q5k_debug_fp32_fallback()) {
        gemv_q5_k_ref_fp32(y, W, x, M, K);
        return;
    }
    if (K % QK_K != 0) {
        gemv_q5_k_ref_fp32(y, W, x, M, K);
        return;
    }
 
    const block_q5_K *blocks = (const block_q5_K *)W;
    const int blocks_per_row = K / QK_K;
    if (blocks_per_row > CK_Q5K_STACK_Q8_BLOCKS) {
        gemv_q5_k_ref_fp32(y, W, x, M, K);
        return;
    }
 
    block_q8_K x_q8[CK_Q5K_STACK_Q8_BLOCKS];
    /* Q8_K bytes are part of the numerical ABI. Use the shared provider,
     * whose FP-contraction policy is validated against llama.cpp. */
    quantize_row_q8_k(x, x_q8, K);
    gemv_q5_k_q8_k_ref(y, blocks, x_q8, M, K);
}

References ck_q5k_debug_fp32_fallback(), CK_Q5K_STACK_Q8_BLOCKS, gemv_q5_k_q8_k_ref(), gemv_q5_k_ref_fp32(), QK_K, and quantize_row_q8_k().

Referenced by gemv_q5_k().

gemv_q5_k_ref_fp32()

static void gemv_q5_k_ref_fp32 ( float *  y, const void *  W, const float *  x, int  M, int  K  )

static

Definition at line 572 of file gemm_kernels_q5_k.c.

{
    const block_q5_K *blocks = (const block_q5_K *)W;
    const int blocks_per_row = K / QK_K;
 
    for (int m = 0; m < M; m++) {
        const float *x_row = x;
        float sum = 0.0f;
 
        for (int b = 0; b < blocks_per_row; b++) {
            const block_q5_K *block = &blocks[m * blocks_per_row + b];
            const float d = CK_FP16_TO_FP32(block->d);
            const float dmin = CK_FP16_TO_FP32(block->dmin);
            uint8_t sc_arr[8], m_arr[8];
            unpack_q5_k_scales(block->scales, sc_arr, m_arr);
 
            for (int sb = 0; sb < 8; sb++) {
                const float d_sub = d * (float)sc_arr[sb];
                const float m_sub = dmin * (float)m_arr[sb];
 
                for (int i = 0; i < 32; i++) {
                    const uint8_t q = q5_k_quant_value(block, sb, i);
                    sum += (d_sub * (float)q - m_sub) * x_row[b * QK_K + sb * 32 + i];
                }
            }
        }
 
        y[m] = sum;
    }
}

References CK_FP16_TO_FP32, q5_k_quant_value(), QK_K, and unpack_q5_k_scales().

Referenced by gemv_q5_k_ref().

q5_k_quant_value()

static uint8_t q5_k_quant_value ( const block_q5_K *  block, int  subblock, int  i  )

inlinestatic

Definition at line 119 of file gemm_kernels_q5_k.c.

                                                                                     {
    const uint8_t *ql = block->qs + (subblock / 2) * 32;
    const uint8_t low = (subblock & 1) ? (uint8_t)(ql[i] >> 4) : (uint8_t)(ql[i] & 0x0F);
    const uint8_t high = (block->qh[i] & (uint8_t)(1u << subblock)) ? 16u : 0u;
    return (uint8_t)(low | high);
}

Referenced by ck_q5_k_prepare_weight(), gemm_nt_q5_k_ref_fp32(), and gemv_q5_k_ref_fp32().

quantize_row_q8_k()

void quantize_row_q8_k	(	const float *	x,
		void *	vy,
		int	k
	)

Definition at line 121 of file gemm_kernels_q4k_q8k.c.

                                                        {
    const char *ref_env = getenv("CK_DEBUG_Q8K_REF");
    if (ref_env && atoi(ref_env) != 0) {
        quantize_row_q8_k_ref(x, vy, k);
        return;
    }
#if defined(__AVX512F__) && defined(__AVX512BW__)
    quantize_row_q8_k_avx512(x, vy, k);
#elif defined(__AVX2__)
    quantize_row_q8_k_avx2(x, vy, k);
#elif defined(__AVX__)
    quantize_row_q8_k_avx(x, vy, k);
#elif defined(__SSE4_1__)
    quantize_row_q8_k_sse(x, vy, k);
#else
    quantize_row_q8_k_ref(x, vy, k);
#endif
}

Referenced by gemm_nt_q5_k_prepared(), gemm_nt_q5_k_prepared_m4(), gemm_nt_q5_k_ref(), and gemv_q5_k_ref().

unpack_q5_k_scales()

static void unpack_q5_k_scales ( const uint8_t *  scales, uint8_t *  sc, uint8_t *  m  )

inlinestatic

Definition at line 95 of file gemm_kernels_q5_k.c.

                                                  {
    sc[0] = scales[0] & 0x3F;
    sc[1] = scales[1] & 0x3F;
    sc[2] = scales[2] & 0x3F;
    sc[3] = scales[3] & 0x3F;
 
    m[0] = scales[4] & 0x3F;
    m[1] = scales[5] & 0x3F;
    m[2] = scales[6] & 0x3F;
    m[3] = scales[7] & 0x3F;
 
    sc[4] = (scales[8]  & 0x0F) | ((scales[0] >> 6) << 4);
    sc[5] = (scales[9]  & 0x0F) | ((scales[1] >> 6) << 4);
    sc[6] = (scales[10] & 0x0F) | ((scales[2] >> 6) << 4);
    sc[7] = (scales[11] & 0x0F) | ((scales[3] >> 6) << 4);
 
    m[4] = (scales[8]  >> 4) | ((scales[4] >> 6) << 4);
    m[5] = (scales[9]  >> 4) | ((scales[5] >> 6) << 4);
    m[6] = (scales[10] >> 4) | ((scales[6] >> 6) << 4);
    m[7] = (scales[11] >> 4) | ((scales[7] >> 6) << 4);
}

Referenced by ck_q5_k_prepare_weight(), gemm_nt_q5_k_ref_fp32(), and gemv_q5_k_ref_fp32().