1  #include <stdlib.h>
       2  #include <cpuid.h>
       3  
       4  /* Check if the OS supports executing AVX512FP16 instructions.  */
       5  
       6  #define XCR_XFEATURE_ENABLED_MASK	0x0
       7  
       8  #define XSTATE_FP	0x1
       9  #define XSTATE_SSE	0x2
      10  #define XSTATE_YMM	0x4
      11  #define XSTATE_OPMASK	0x20
      12  #define XSTATE_ZMM	0x40
      13  #define XSTATE_HI_ZMM	0x80
      14  
      15  static int
      16  check_osxsave (void)
      17  {
      18    unsigned int eax, ebx, ecx, edx;
      19  
      20    if (!__get_cpuid (1, &eax, &ebx, &ecx, &edx))
      21      return 0;
      22  
      23    return (ecx & bit_OSXSAVE) != 0;
      24  }
      25  
      26  static int
      27  avx512fp16_os_support (void)
      28  {
      29    unsigned int eax, edx;
      30    unsigned int ecx = XCR_XFEATURE_ENABLED_MASK;
      31    unsigned int mask = XSTATE_MASK;
      32  
      33    if (!check_osxsave ())
      34      return 0;
      35  
      36    __asm__ ("xgetbv" : "=a" (eax), "=d" (edx) : "c" (ecx));
      37  
      38    return ((eax & mask) == mask);
      39  }
      40  
      41  static void do_test (void);
      42  
      43  int
      44  main ()
      45  {
      46    unsigned int eax, ebx, ecx, edx;
      47  
      48    if (!avx512fp16_os_support ())
      49      return 0;
      50  
      51    if (__get_cpuid_max (0, NULL) < 7)
      52      return 0;
      53  
      54    __cpuid_count (7, 0, eax, ebx, ecx, edx);
      55  
      56      /* Run AVX512FP16 test only if host has ISA support.  */
      57    if (((ebx & (bit_AVX512F | bit_AVX512BW))
      58         == (bit_AVX512F | bit_AVX512BW))
      59        && (edx & bit_AVX512FP16)
      60        && AVX512VL (ebx))
      61      {
      62        do_test ();
      63  #ifdef DEBUG
      64        printf ("PASSED\n");
      65  #endif
      66        return 0;
      67      }
      68  
      69  #ifdef DEBUG
      70    printf ("SKIPPED\n");
      71  #endif
      72  
      73    return 0;
      74  }