ckernel_dtype.h File Reference

#include <stdint.h>
#include <stddef.h>

Go to the source code of this file.

Macros
#define	CK_DT_MASK(dt)   (1u << (uint32_t)(dt))

Typedefs
typedef uint32_t	CKDataTypeMask

Enumerations
enum	CKDataType {   CK_DT_FP32 = 0 , CK_DT_BF16 , CK_DT_FP16 , CK_DT_INT8 ,   CK_DT_INT4 , CK_DT_Q4_0 , CK_DT_Q4_1 , CK_DT_Q4_K ,   CK_DT_Q6_K , CK_DT_Q8_0 , CK_DT_Q8_K , CK_DT_Q5_0 ,   CK_DT_Q5_1 , CK_DT_Q5_K , CK_DT_COUNT }
	Supported data types in C-Kernel-Engine. More...

Functions
static size_t	ck_dtype_block_bytes (CKDataType dt)
	Get bytes per block for quantized types. More...
static size_t	ck_dtype_block_size (CKDataType dt)
	Get the number of elements per quantization block. More...
static size_t	ck_dtype_bytes (CKDataType dt)
	Get bytes per element for non-quantized types. More...
static int	ck_dtype_is_quantized (CKDataType dt)
	Check if a data type is block-quantized (GGML-style) More...
static size_t	ck_dtype_row_bytes (CKDataType dt, size_t n_elements)
	Calculate total bytes for n_elements of given dtype. More...
static int	ck_dtype_supported (CKDataTypeMask mask, CKDataType dt)

Macro Definition Documentation

CK_DT_MASK

#define CK_DT_MASK

(

dt

)

(1u << (uint32_t)(dt))

Definition at line 53 of file ckernel_dtype.h.

Typedef Documentation

CKDataTypeMask

typedef uint32_t CKDataTypeMask

Definition at line 51 of file ckernel_dtype.h.

Enumeration Type Documentation

CKDataType

enum CKDataType

Supported data types in C-Kernel-Engine.

Standard types:

• CK_DT_FP32: 32-bit float (baseline, full precision)
• CK_DT_BF16: Brain Float 16 (1+8+7 bits)
• CK_DT_FP16: IEEE Half Precision (1+5+10 bits)

Simple quantization:

• CK_DT_INT8: 8-bit signed integer
• CK_DT_INT4: 4-bit signed integer (simple, no scales)

GGML-compatible quantization (block-based with scales):

• CK_DT_Q4_0: 4-bit, 32 weights/block, 1 FP16 scale
• CK_DT_Q4_K: 4-bit k-quant, 256 weights/block, nested scales (Q4_K_M)
• CK_DT_Q6_K: 6-bit k-quant, 256 weights/block, per-16 scales
• CK_DT_Q8_0: 8-bit, 32 weights/block, 1 FP16 scale
• CK_DT_Q8_K: 8-bit k-quant, 256 weights/block, FP32 scale + bsums
• CK_DT_Q5_0: 5-bit, 32 weights/block, 1 FP16 scale

Enumerator
CK_DT_FP32
CK_DT_BF16
CK_DT_FP16
CK_DT_INT8
CK_DT_INT4
CK_DT_Q4_0
CK_DT_Q4_1
CK_DT_Q4_K
CK_DT_Q6_K
CK_DT_Q8_0
CK_DT_Q8_K
CK_DT_Q5_0
CK_DT_Q5_1
CK_DT_Q5_K
CK_DT_COUNT

Definition at line 27 of file ckernel_dtype.h.

             {
    /* Standard floating-point types */
    CK_DT_FP32 = 0,      /* 4 bytes per element */
    CK_DT_BF16,          /* 2 bytes per element */
    CK_DT_FP16,          /* 2 bytes per element */
 
    /* Simple integer types (legacy) */
    CK_DT_INT8,          /* 1 byte per element */
    CK_DT_INT4,          /* 0.5 bytes per element (packed) */
 
    /* GGML-compatible block quantization */
    CK_DT_Q4_0,          /* 4.5 bits/weight (18 bytes per 32 weights) */
    CK_DT_Q4_1,          /* 5.0 bits/weight (20 bytes per 32 weights) */
    CK_DT_Q4_K,          /* 4.5 bits/weight (144 bytes per 256 weights) - Q4_K_M */
    CK_DT_Q6_K,          /* 6.5 bits/weight (210 bytes per 256 weights) */
    CK_DT_Q8_0,          /* 8.5 bits/weight (34 bytes per 32 weights) */
    CK_DT_Q8_K,          /* 9.125 bits/weight (292 bytes per 256 weights) */
    CK_DT_Q5_0,          /* 5.5 bits/weight (22 bytes per 32 weights) */
    CK_DT_Q5_1,          /* 6.0 bits/weight (24 bytes per 32 weights) */
    CK_DT_Q5_K,          /* 5.5 bits/weight (176 bytes per 256 weights) - Q5_K super-block */
 
    CK_DT_COUNT
} CKDataType;

Function Documentation

ck_dtype_block_bytes()

static size_t ck_dtype_block_bytes ( CKDataType  dt )

inlinestatic

Get bytes per block for quantized types.

Definition at line 109 of file ckernel_dtype.h.

{
    switch (dt) {
    case CK_DT_Q4_0:
        return 18;   /* 2 (scale) + 16 (32 x 4-bit) */
    case CK_DT_Q4_1:
        return 20;   /* 2 (scale) + 2 (min) + 16 (32 x 4-bit) */
    case CK_DT_Q5_0:
        return 22;   /* 2 (scale) + 4 (high bit) + 16 (low 4-bit) */
    case CK_DT_Q5_1:
        return 24;   /* 2 (scale) + 2 (min) + 4 (high bit) + 16 (low 4-bit) */
    case CK_DT_Q4_K:
        return 144;  /* 2 + 2 + 12 + 128 */
    case CK_DT_Q5_K:
        return 176;  /* 2 + 2 + 12 + 32 + 128 */
    case CK_DT_Q6_K:
        return 210;  /* 2 + 16 + 128 + 64 */
    case CK_DT_Q8_0:
        return 34;   /* 2 (scale) + 32 (32 x 8-bit) */
    case CK_DT_Q8_K:
        return 292;  /* 4 (scale) + 256 (int8) + 32 (bsums) */
    default:
        return ck_dtype_bytes(dt);
    }
}

References CK_DT_Q4_0, CK_DT_Q4_1, CK_DT_Q4_K, CK_DT_Q5_0, CK_DT_Q5_1, CK_DT_Q5_K, CK_DT_Q6_K, CK_DT_Q8_0, CK_DT_Q8_K, and ck_dtype_bytes().

Referenced by ck_dtype_row_bytes().

ck_dtype_block_size()

static size_t ck_dtype_block_size ( CKDataType  dt )

inlinestatic

Get the number of elements per quantization block.

Definition at line 87 of file ckernel_dtype.h.

{
    switch (dt) {
    case CK_DT_Q4_0:
    case CK_DT_Q4_1:
    case CK_DT_Q5_0:
    case CK_DT_Q5_1:
    case CK_DT_Q8_0:
        return 32;
    case CK_DT_Q4_K:
    case CK_DT_Q5_K:
    case CK_DT_Q6_K:
    case CK_DT_Q8_K:
        return 256;
    default:
        return 1; /* Non-quantized types: 1 element per "block" */
    }
}

References CK_DT_Q4_0, CK_DT_Q4_1, CK_DT_Q4_K, CK_DT_Q5_0, CK_DT_Q5_1, CK_DT_Q5_K, CK_DT_Q6_K, CK_DT_Q8_0, and CK_DT_Q8_K.

Referenced by ck_dtype_row_bytes().

ck_dtype_bytes()

static size_t ck_dtype_bytes ( CKDataType  dt )

inlinestatic

Get bytes per element for non-quantized types.

Note

For quantized types, use ck_dtype_block_bytes() and ck_dtype_block_size()

Definition at line 68 of file ckernel_dtype.h.

{
    switch (dt) {
    case CK_DT_BF16:
    case CK_DT_FP16:
        return 2;
    case CK_DT_INT8:
        return 1;
    case CK_DT_INT4:
        return 1; /* Note: actually 0.5, but stored as pairs */
    case CK_DT_FP32:
    default:
        return 4;
    }
}

References CK_DT_BF16, CK_DT_FP16, CK_DT_FP32, CK_DT_INT4, and CK_DT_INT8.

Referenced by ck_dtype_block_bytes(), and ck_dtype_row_bytes().

ck_dtype_is_quantized()

static int ck_dtype_is_quantized ( CKDataType  dt )

inlinestatic

Check if a data type is block-quantized (GGML-style)

Definition at line 58 of file ckernel_dtype.h.

{
    return dt == CK_DT_Q4_0 || dt == CK_DT_Q4_1 || dt == CK_DT_Q5_0 || dt == CK_DT_Q5_1 ||
           dt == CK_DT_Q5_K || dt == CK_DT_Q4_K || dt == CK_DT_Q6_K || dt == CK_DT_Q8_0 || dt == CK_DT_Q8_K;
}

References CK_DT_Q4_0, CK_DT_Q4_1, CK_DT_Q4_K, CK_DT_Q5_0, CK_DT_Q5_1, CK_DT_Q5_K, CK_DT_Q6_K, CK_DT_Q8_0, and CK_DT_Q8_K.

Referenced by ck_dtype_row_bytes().

ck_dtype_row_bytes()

static size_t ck_dtype_row_bytes ( CKDataType  dt, size_t  n_elements  )

inlinestatic

Calculate total bytes for n_elements of given dtype.

Definition at line 138 of file ckernel_dtype.h.

{
    if (ck_dtype_is_quantized(dt)) {
        size_t block_size = ck_dtype_block_size(dt);
        size_t n_blocks = (n_elements + block_size - 1) / block_size;
        return n_blocks * ck_dtype_block_bytes(dt);
    }
    return n_elements * ck_dtype_bytes(dt);
}

References ck_dtype_block_bytes(), ck_dtype_block_size(), ck_dtype_bytes(), and ck_dtype_is_quantized().

Referenced by buffer_bytes(), build_plan(), ck_qkv_project_head_major_q4_k(), ck_qkv_project_head_major_quant(), ck_qkv_project_head_major_token_q4_k(), ck_qkv_project_head_major_token_q4_k_q8_k(), ck_qkv_project_head_major_token_quant(), embedding_forward_q4_k(), embedding_forward_q6_k(), embedding_forward_q8_0(), fused_mlp_swiglu_prefill_w1w2_quant(), fused_mlp_swiglu_prefill_w1w2_quant_scratch_size(), fused_rmsnorm_qkv_prefill_head_major_quant(), fused_rmsnorm_qkv_prefill_head_major_quant_scratch_size(), mega_fused_outproj_mlp_prefill(), mega_fused_outproj_mlp_prefill_scratch_size(), model_forward_prefill_impl(), model_layer_0_decode(), model_layer_0_prefill(), model_layer_10_decode(), model_layer_10_prefill(), model_layer_11_decode(), model_layer_11_prefill(), model_layer_12_decode(), model_layer_12_prefill(), model_layer_13_decode(), model_layer_13_prefill(), model_layer_14_decode(), model_layer_14_prefill(), model_layer_15_decode(), model_layer_15_prefill(), model_layer_16_decode(), model_layer_16_prefill(), model_layer_17_decode(), model_layer_17_prefill(), model_layer_18_decode(), model_layer_18_prefill(), model_layer_19_decode(), model_layer_19_prefill(), model_layer_1_decode(), model_layer_1_prefill(), model_layer_20_decode(), model_layer_20_prefill(), model_layer_21_decode(), model_layer_21_prefill(), model_layer_22_decode(), model_layer_22_prefill(), model_layer_23_decode(), model_layer_23_prefill(), model_layer_2_decode(), model_layer_2_prefill(), model_layer_3_decode(), model_layer_3_prefill(), model_layer_4_decode(), model_layer_4_prefill(), model_layer_5_decode(), model_layer_5_prefill(), model_layer_6_decode(), model_layer_6_prefill(), model_layer_7_decode(), model_layer_7_prefill(), model_layer_8_decode(), model_layer_8_prefill(), model_layer_9_decode(), model_layer_9_prefill(), out_proj_head_major_q5_0_q8_0(), out_proj_head_major_q8_0_q8_0(), quantize_attn_out_head_major_q8_0(), qwen2_0_5b_decode_forward_prefill_impl(), qwen2_0_5b_decode_layer_0_decode(), qwen2_0_5b_decode_layer_0_prefill(), qwen2_0_5b_decode_layer_10_decode(), qwen2_0_5b_decode_layer_10_prefill(), qwen2_0_5b_decode_layer_11_decode(), qwen2_0_5b_decode_layer_11_prefill(), qwen2_0_5b_decode_layer_12_decode(), qwen2_0_5b_decode_layer_12_prefill(), qwen2_0_5b_decode_layer_13_decode(), qwen2_0_5b_decode_layer_13_prefill(), qwen2_0_5b_decode_layer_14_decode(), qwen2_0_5b_decode_layer_14_prefill(), qwen2_0_5b_decode_layer_15_decode(), qwen2_0_5b_decode_layer_15_prefill(), qwen2_0_5b_decode_layer_16_decode(), qwen2_0_5b_decode_layer_16_prefill(), qwen2_0_5b_decode_layer_17_decode(), qwen2_0_5b_decode_layer_17_prefill(), qwen2_0_5b_decode_layer_18_decode(), qwen2_0_5b_decode_layer_18_prefill(), qwen2_0_5b_decode_layer_19_decode(), qwen2_0_5b_decode_layer_19_prefill(), qwen2_0_5b_decode_layer_1_decode(), qwen2_0_5b_decode_layer_1_prefill(), qwen2_0_5b_decode_layer_20_decode(), qwen2_0_5b_decode_layer_20_prefill(), qwen2_0_5b_decode_layer_21_decode(), qwen2_0_5b_decode_layer_21_prefill(), qwen2_0_5b_decode_layer_22_decode(), qwen2_0_5b_decode_layer_22_prefill(), qwen2_0_5b_decode_layer_23_decode(), qwen2_0_5b_decode_layer_23_prefill(), qwen2_0_5b_decode_layer_2_decode(), qwen2_0_5b_decode_layer_2_prefill(), qwen2_0_5b_decode_layer_3_decode(), qwen2_0_5b_decode_layer_3_prefill(), qwen2_0_5b_decode_layer_4_decode(), qwen2_0_5b_decode_layer_4_prefill(), qwen2_0_5b_decode_layer_5_decode(), qwen2_0_5b_decode_layer_5_prefill(), qwen2_0_5b_decode_layer_6_decode(), qwen2_0_5b_decode_layer_6_prefill(), qwen2_0_5b_decode_layer_7_decode(), qwen2_0_5b_decode_layer_7_prefill(), qwen2_0_5b_decode_layer_8_decode(), qwen2_0_5b_decode_layer_8_prefill(), qwen2_0_5b_decode_layer_9_decode(), and qwen2_0_5b_decode_layer_9_prefill().

ck_dtype_supported()

static int ck_dtype_supported ( CKDataTypeMask  mask, CKDataType  dt  )

inlinestatic

Definition at line 148 of file ckernel_dtype.h.

{
    return (mask & CK_DT_MASK(dt)) != 0;
}

References CK_DT_MASK, and mask.