diff --git a/core/sys/info/cpu_intel.odin b/core/sys/info/cpu_intel.odin index d6fa9850728..95b53dda0d6 100644 --- a/core/sys/info/cpu_intel.odin +++ b/core/sys/info/cpu_intel.odin @@ -28,6 +28,23 @@ CPU_Feature :: enum u64 { ssse3, // Supplemental streaming SIMD extension 3 sse41, // Streaming SIMD extension 4 and 4.1 sse42, // Streaming SIMD extension 4 and 4.2 + + avx512bf16, // Vector Neural Network Instructions supporting bfloat16 + avx512bitalg, // Bit Algorithms + avx512bw, // Byte and Word instructions + avx512cd, // Conflict Detection instructions + avx512dq, // Doubleword and Quadword instructions + avx512er, // Exponential and Reciprocal instructions + avx512f, // Foundation + avx512fp16, // Vector 16-bit float instructions + avx512ifma, // Integer Fused Multiply Add + avx512pf, // Prefetch instructions + avx512vbmi, // Vector Byte Manipulation Instructions + avx512vbmi2, // Vector Byte Manipulation Instructions 2 + avx512vl, // Vector Length extensions + avx512vnni, // Vector Neural Network Instructions + avx512vp2intersect, // Vector Pair Intersection to a Pair of Mask Registers + avx512vpopcntdq, // Vector Population Count for Doubleword and Quadword } CPU_Features :: distinct bit_set[CPU_Feature; u64] @@ -82,9 +99,11 @@ init_cpu_features :: proc "c" () { // // See: crbug.com/375968 os_supports_avx := false + os_supports_avx512 := false if .os_xsave in set && is_set(26, ecx1) { eax, _ := xgetbv(0) os_supports_avx = is_set(1, eax) && is_set(2, eax) + os_supports_avx512 = is_set(5, eax) && is_set(6, eax) && is_set(7, eax) } if os_supports_avx { try_set(&set, .avx, 28, ecx1) @@ -94,11 +113,37 @@ init_cpu_features :: proc "c" () { return } - _, ebx7, _, _ := cpuid(7, 0) + _, ebx7, ecx7, edx7 := cpuid(7, 0) try_set(&set, .bmi1, 3, ebx7) if os_supports_avx { try_set(&set, .avx2, 5, ebx7) } + if os_supports_avx512 { + try_set(&set, .avx512f, 16, ebx7) + try_set(&set, .avx512dq, 17, ebx7) + try_set(&set, .avx512ifma, 21, ebx7) + try_set(&set, .avx512pf, 26, ebx7) + try_set(&set, .avx512er, 27, ebx7) + try_set(&set, .avx512cd, 28, ebx7) + try_set(&set, .avx512bw, 30, ebx7) + + // XMM/YMM are also required for 128/256-bit instructions + if os_supports_avx { + try_set(&set, .avx512vl, 31, ebx7) + } + + try_set(&set, .avx512vbmi, 1, ecx7) + try_set(&set, .avx512vbmi2, 6, ecx7) + try_set(&set, .avx512vnni, 11, ecx7) + try_set(&set, .avx512bitalg, 12, ecx7) + try_set(&set, .avx512vpopcntdq, 14, ecx7) + + try_set(&set, .avx512vp2intersect, 8, edx7) + try_set(&set, .avx512fp16, 23, edx7) + + eax7_1, _, _, _ := cpuid(7, 1) + try_set(&set, .avx512bf16, 5, eax7_1) + } try_set(&set, .bmi2, 8, ebx7) try_set(&set, .erms, 9, ebx7) try_set(&set, .rdseed, 18, ebx7)