CPU feature detection and dispatch macros. More...

#include <stdint.h>

Data Structures
struct	ck_capability_t
	CPU capability information structure. More...

Macros
#define	CK_GEMM_DISPATCH(...) gemm_ref(__VA_ARGS__)
	Dispatch to best available GEMM kernel.

#define	CK_GEMV_DISPATCH(...) gemv_ref(__VA_ARGS__)
	Dispatch to best available GEMV kernel.

#define	CK_HAS_AI_ACCEL 0

#define	CK_HAS_BEST_VECTOR 0

#define	CK_QGEMM_DISPATCH(...) qgemm_ref(__VA_ARGS__)
	Dispatch to best available quantized GEMV kernel For INT8/INT4 quantization with VNNI/AMX acceleration.

#define	CK_VECTOR_WIDTH 32 /* Scalar fallback */

Functions
static ck_capability_t	ck_get_capabilities (void)
	Get current platform capabilities.

Detailed Description

CPU feature detection and dispatch macros.

Defines standardized macros for SIMD instruction set detection and kernel dispatch. Use these instead of CPU model checks.

Feature Priority (best available): AMX (512-bit tile ops, Intel Sapphire Rapids+) AVX-512 (512-bit vector, Intel Skylake-X+) AVX2 (256-bit with FMA, Intel Haswell+) AVX (256-bit, Intel Sandy Bridge+) NEON/SVE2 (ARM) DSA (PowerPC) Reference (fallback)

Definition in file ck_features.h.

Macro Definition Documentation

◆ CK_GEMM_DISPATCH

#define CK_GEMM_DISPATCH ( ... ) gemm_ref(__VA_ARGS__)

Dispatch to best available GEMM kernel.

Usage: CK_GEMM_DISPATCH(y, W, x, M, K); expands to appropriate kernel call

Definition at line 174 of file ck_features.h.

◆ CK_GEMV_DISPATCH

#define CK_GEMV_DISPATCH ( ... ) gemv_ref(__VA_ARGS__)

Dispatch to best available GEMV kernel.

Definition at line 189 of file ck_features.h.

◆ CK_HAS_AI_ACCEL

#define CK_HAS_AI_ACCEL 0

Definition at line 150 of file ck_features.h.

◆ CK_HAS_BEST_VECTOR

#define CK_HAS_BEST_VECTOR 0

Definition at line 139 of file ck_features.h.

◆ CK_QGEMM_DISPATCH

#define CK_QGEMM_DISPATCH ( ... ) qgemm_ref(__VA_ARGS__)

Dispatch to best available quantized GEMV kernel For INT8/INT4 quantization with VNNI/AMX acceleration.

Definition at line 205 of file ck_features.h.

◆ CK_VECTOR_WIDTH

#define CK_VECTOR_WIDTH 32 /* Scalar fallback */

Definition at line 138 of file ck_features.h.

Function Documentation

◆ ck_get_capabilities()

static ck_capability_t ck_get_capabilities ( void )

inlinestatic

Get current platform capabilities.

Definition at line 226 of file ck_features.h.

                                                        {
    ck_capability_t cap = {
        .name = "unknown",
        .width = 32,
        .has_fma = 0,
        .has_ai_accel = 0,
        .best_kernel = "gemm_ref"
    };
 
#if defined(CK_HAS_AMX)
    cap.name = "AMX (Intel Sapphire Rapids+)";
    cap.width = 512;
    cap.has_fma = 1;
    cap.has_ai_accel = 1;
    cap.best_kernel = "gemm_amx";
#elif defined(CK_HAS_AVX512)
    cap.name = "AVX-512 (Intel Skylake-X+)";
    cap.width = 512;
    cap.has_fma = 1;
    cap.has_ai_accel = 1;
    cap.best_kernel = "gemm_avx512";
#elif defined(CK_HAS_AVX2_FMA)
    cap.name = "AVX2+FMA (Intel Haswell+)";
    cap.width = 256;
    cap.has_fma = 1;
    cap.has_ai_accel = 0;
    cap.best_kernel = "gemm_avx2";
#elif defined(CK_HAS_AVX)
    cap.name = "AVX (Intel Sandy Bridge+)";
    cap.width = 256;
    cap.has_fma = 0;
    cap.has_ai_accel = 0;
    cap.best_kernel = "gemm_avx";
#elif defined(CK_HAS_NEON)
    cap.name = "NEON (ARM)";
    cap.width = 128;
    cap.has_fma = 1;
    cap.has_ai_accel = 0;
    cap.best_kernel = "gemm_neon";
#elif defined(CK_HAS_ALTIVEC)
    cap.name = "AltiVec (PowerPC)";
    cap.width = 128;
    cap.has_fma = 1;
    cap.has_ai_accel = 0;
    cap.best_kernel = "gemm_altivec";
#endif
 
    return cap;
}

References ck_capability_t::best_kernel, ck_capability_t::has_ai_accel, ck_capability_t::has_fma, ck_capability_t::name, and ck_capability_t::width.

Referenced by main().

Data Structures

Macros

Functions