| 87 | } |
| 88 | |
| 89 | bool feature_detect_vnni() { |
| 90 | uint32_t eax, ebx, ecx, edx; |
| 91 | |
| 92 | // check cpu support |
| 93 | #if defined(_WIN32) |
| 94 | int cpuInfo[4]; |
| 95 | __cpuid(cpuInfo, 7); |
| 96 | eax = cpuInfo[0]; |
| 97 | ebx = cpuInfo[1]; |
| 98 | ecx = cpuInfo[2]; |
| 99 | edx = cpuInfo[3]; |
| 100 | #else |
| 101 | asm volatile("cpuid\n" |
| 102 | : "=a"(eax), "=b"(ebx), "=c"(ecx), "=d"(edx) |
| 103 | : "a"(7), "c"(0) |
| 104 | : "cc"); |
| 105 | #endif |
| 106 | // avx512f ---> 16 ebx |
| 107 | // avx512dq ---> 17 ebx |
| 108 | // avx512bw ---> 30 ebx |
| 109 | // avx512vl ---> 31 ebx |
| 110 | // avx512vnni --->11 ecx |
| 111 | if (!(bit(ebx, 16) && bit(ebx, 17) && bit(ebx, 30) && bit(ebx, 31) && bit(ecx, 11))) |
| 112 | return false; |
| 113 | |
| 114 | // check os support |
| 115 | asm volatile("xgetbv" : "=a"(eax), "=d"(edx) : "c"(0)); |
| 116 | |
| 117 | return (eax & 6) == 6; |
| 118 | } |
| 119 | |
| 120 | bool feature_detect_avx_fma(int ftr) { |
| 121 | // see Detecting Availability and Support in |