cpu_features.h File Reference

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

Go to the source code of this file.

Data Structures
struct	CPUInfo
struct	GEMMParams

Functions
void	cpu_features_init (void)
const CPUInfo *	get_cpu_info (void)
const GEMMParams *	get_gemm_params (void)
void	print_cpu_info (void)

Variables
CPUInfo	g_cpu_info
int	g_cpu_initialized
GEMMParams	g_gemm_params

Function Documentation

cpu_features_init()

void cpu_features_init

(

void

)

Definition at line 376 of file cpu_features.c.

                             {
    if (g_cpu_initialized) return;
 
    memset(&g_cpu_info, 0, sizeof(g_cpu_info));
    memset(&g_gemm_params, 0, sizeof(g_gemm_params));
 
    // Detect SIMD features
#ifdef X86_CPU
    detect_x86_features(&g_cpu_info);
    detect_x86_cache_sizes(&g_cpu_info);
#endif
 
    // Linux sysfs fallback for cache sizes
#if defined(__linux__)
    if (g_cpu_info.l1d_size == 0) {
        detect_linux_cache_sizes(&g_cpu_info);
    }
#endif
 
    // Detect physical cores
    g_cpu_info.num_cores = detect_physical_cores();
 
    // Compute GEMM parameters based on detected hardware
    compute_gemm_params(&g_cpu_info, &g_gemm_params);
 
    g_cpu_initialized = 1;
}

References compute_gemm_params(), detect_physical_cores(), g_cpu_info, g_cpu_initialized, g_gemm_params, CPUInfo::l1d_size, and CPUInfo::num_cores.

Referenced by get_cpu_info(), get_gemm_params(), and print_cpu_info().

get_cpu_info()

const CPUInfo * get_cpu_info

(

void

)

Definition at line 409 of file cpu_features.c.

                                  {
    if (!g_cpu_initialized) cpu_features_init();
    return &g_cpu_info;
}

References cpu_features_init(), g_cpu_info, and g_cpu_initialized.

Referenced by gemm_init_threads().

get_gemm_params()

const GEMMParams * get_gemm_params

(

void

)

Definition at line 404 of file cpu_features.c.

                                        {
    if (!g_cpu_initialized) cpu_features_init();
    return &g_gemm_params;
}

References cpu_features_init(), g_cpu_initialized, and g_gemm_params.

print_cpu_info()

void print_cpu_info

(

void

)

Definition at line 414 of file cpu_features.c.

                          {
    if (!g_cpu_initialized) cpu_features_init();
 
    printf("=== CPU Info ===\n");
    printf("Physical cores: %d\n", g_cpu_info.num_cores);
    printf("L1 Data Cache:  %zu KB\n", g_cpu_info.l1d_size / 1024);
    printf("L2 Cache:       %zu KB\n", g_cpu_info.l2_size / 1024);
    printf("L3 Cache:       %zu MB\n", g_cpu_info.l3_size / (1024 * 1024));
    printf("Cache line:     %zu bytes\n", g_cpu_info.l1_line_size);
    printf("AVX:            %s\n", g_cpu_info.has_avx ? "yes" : "no");
    printf("AVX2:           %s\n", g_cpu_info.has_avx2 ? "yes" : "no");
    printf("AVX-512F:       %s\n", g_cpu_info.has_avx512f ? "yes" : "no");
    printf("AVX-512BW/DQ/VL:%s/%s/%s\n",
           g_cpu_info.has_avx512bw ? "yes" : "no",
           g_cpu_info.has_avx512dq ? "yes" : "no",
           g_cpu_info.has_avx512vl ? "yes" : "no");
    printf("AVX-512 VNNI:   %s\n", g_cpu_info.has_avx512_vnni ? "yes" : "no");
    printf("AVX-512 BF16:   %s\n", g_cpu_info.has_avx512_bf16 ? "yes" : "no");
    printf("AMX tile/int8/bf16: %s/%s/%s\n",
           g_cpu_info.has_amx_tile ? "yes" : "no",
           g_cpu_info.has_amx_int8 ? "yes" : "no",
           g_cpu_info.has_amx_bf16 ? "yes" : "no");
    printf("FMA:            %s\n", g_cpu_info.has_fma ? "yes" : "no");
    printf("\n=== GEMM Blocking Parameters ===\n");
    printf("MR (microkernel rows): %d\n", g_gemm_params.MR);
    printf("NR (microkernel cols): %d\n", g_gemm_params.NR);
    printf("MC (M block):          %d\n", g_gemm_params.MC);
    printf("NC (N block):          %d\n", g_gemm_params.NC);
    printf("KC (K block):          %d\n", g_gemm_params.KC);
    printf("\n");
}

References cpu_features_init(), g_cpu_info, g_cpu_initialized, g_gemm_params, CPUInfo::has_amx_bf16, CPUInfo::has_amx_int8, CPUInfo::has_amx_tile, CPUInfo::has_avx, CPUInfo::has_avx2, CPUInfo::has_avx512_bf16, CPUInfo::has_avx512_vnni, CPUInfo::has_avx512bw, CPUInfo::has_avx512dq, CPUInfo::has_avx512f, CPUInfo::has_avx512vl, CPUInfo::has_fma, GEMMParams::KC, CPUInfo::l1_line_size, CPUInfo::l1d_size, CPUInfo::l2_size, CPUInfo::l3_size, GEMMParams::MC, GEMMParams::MR, GEMMParams::NC, GEMMParams::NR, and CPUInfo::num_cores.

Variable Documentation

g_cpu_info

CPUInfo g_cpu_info

extern

CPU Feature Detection and Cache-Aware Parameter Tuning

Detects CPU features, cache sizes, and core counts at runtime. Computes optimal GEMM blocking parameters based on actual hardware.

Definition at line 28 of file cpu_features.c.

{0};

Referenced by cpu_features_init(), get_cpu_info(), and print_cpu_info().

g_cpu_initialized

int g_cpu_initialized

extern

Definition at line 30 of file cpu_features.c.

Referenced by cpu_features_init(), get_cpu_info(), get_gemm_params(), and print_cpu_info().

g_gemm_params

GEMMParams g_gemm_params

extern

Definition at line 29 of file cpu_features.c.

{0};

Referenced by cpu_features_init(), get_gemm_params(), and print_cpu_info().