Skip to content

Commit

Permalink
Fix detection of AVX2 support
Browse files Browse the repository at this point in the history 
Fixes NLnetLabs#222.
  • Loading branch information
@k0ekk0ek
k0ekk0ek committed Aug 1, 2024
1 parent 380abdb commit 00cb72e
Showing 1 changed file with 74 additions and 29 deletions.
103 changes: 74 additions & 29 deletions src/isadetection.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -3,6 +3,7 @@
*
* Slightly modified version of isadetection.h in simdjson.
*
* Copyright (c) 2024 NLnet Labs (Jeroen Koekkoek)
* Copyright (c) 2020- simdjson (Daniel Lemire,
* Geoff Langdale,
* John Keiser)
Expand Down Expand Up @@ -107,20 +108,22 @@ static inline uint32_t detect_supported_architectures() {
#elif defined(__x86_64__) || defined(_M_AMD64) // x64

// Can be found on Intel ISA Reference for CPUID
static const uint32_t cpuid_avx2_bit = 1 << 5; ///< @private Bit 5 of EBX for EAX=0x7
static const uint32_t cpuid_bmi1_bit = 1 << 3; ///< @private bit 3 of EBX for EAX=0x7
static const uint32_t cpuid_bmi2_bit = 1 << 8; ///< @private bit 8 of EBX for EAX=0x7
static const uint32_t cpuid_avx512f_bit = 1 << 16; ///< @private bit 16 of EBX for EAX=0x7
static const uint32_t cpuid_avx512dq_bit = 1 << 17; ///< @private bit 17 of EBX for EAX=0x7
static const uint32_t cpuid_avx512ifma_bit = 1 << 21; ///< @private bit 21 of EBX for EAX=0x7
static const uint32_t cpuid_avx512pf_bit = 1 << 26; ///< @private bit 26 of EBX for EAX=0x7
static const uint32_t cpuid_avx512er_bit = 1 << 27; ///< @private bit 27 of EBX for EAX=0x7
static const uint32_t cpuid_avx512cd_bit = 1 << 28; ///< @private bit 28 of EBX for EAX=0x7
static const uint32_t cpuid_avx512bw_bit = 1 << 30; ///< @private bit 30 of EBX for EAX=0x7
static const uint32_t cpuid_avx512vl_bit = 1U << 31; ///< @private bit 31 of EBX for EAX=0x7
static const uint32_t cpuid_avx512vbmi2_bit = 1 << 6; ///< @private bit 6 of ECX for EAX=0x7
static const uint32_t cpuid_sse42_bit = 1 << 20; ///< @private bit 20 of ECX for EAX=0x1
static const uint32_t cpuid_pclmulqdq_bit = 1 << 1; ///< @private bit 1 of ECX for EAX=0x1
static const uint32_t cpuid_avx2_bit = 1 << 5; ///< @private Bit 5 of EBX for EAX=0x7
static const uint32_t cpuid_bmi1_bit = 1 << 3; ///< @private bit 3 of EBX for EAX=0x7
static const uint32_t cpuid_bmi2_bit = 1 << 8; ///< @private bit 8 of EBX for EAX=0x7
static const uint32_t cpuid_avx512f_bit = 1 << 16; ///< @private bit 16 of EBX for EAX=0x7
static const uint32_t cpuid_avx512dq_bit = 1 << 17; ///< @private bit 17 of EBX for EAX=0x7
static const uint32_t cpuid_avx512ifma_bit = 1 << 21; ///< @private bit 21 of EBX for EAX=0x7
static const uint32_t cpuid_avx512pf_bit = 1 << 26; ///< @private bit 26 of EBX for EAX=0x7
static const uint32_t cpuid_avx512er_bit = 1 << 27; ///< @private bit 27 of EBX for EAX=0x7
static const uint32_t cpuid_avx512cd_bit = 1 << 28; ///< @private bit 28 of EBX for EAX=0x7
static const uint32_t cpuid_avx512bw_bit = 1 << 30; ///< @private bit 30 of EBX for EAX=0x7
static const uint32_t cpuid_avx512vl_bit = 1U << 31; ///< @private bit 31 of EBX for EAX=0x7
static const uint32_t cpuid_avx512vbmi2_bit = 1 << 6; ///< @private bit 6 of ECX for EAX=0x7
static const uint32_t cpuid_sse42_bit = 1 << 20; ///< @private bit 20 of ECX for EAX=0x1
static const uint32_t cpuid_pclmulqdq_bit = 1 << 1; ///< @private bit 1 of ECX for EAX=0x1
static const uint32_t cpuid_have_xgetbv_bit = 1 << 27; ///< @private bit 27 of ECX for EAX=0x1
static const uint32_t cpuid_have_avx_bit = 1 << 28; ///< @private bit 28 of ECX for EAX=0x1

static inline void cpuid(
uint32_t *eax, uint32_t *ebx, uint32_t *ecx, uint32_t *edx)
Expand All @@ -145,17 +148,26 @@ static inline void cpuid(
#endif
}

static inline uint32_t detect_supported_architectures(void) {
static inline void xgetbv(
uint32_t *edx, uint32_t *eax, uint32_t *ecx)
{
// FIXME: Implement support for _xgetbv intrinsic?
uint32_t a, c = *ecx, d;
asm volatile("xgetbv\n\t" : "=d"(d), "=a"(a), "+c"(c));
*edx = d;
*eax = a;
*ecx = c;
}

static inline uint32_t detect_supported_architectures(void)
{
uint32_t eax, ebx, ecx, edx;
uint32_t host_isa = 0x0;
uint32_t host_isa = 0x0, host_avx_isa = 0x0;

// ECX for EAX=0x7
eax = 0x7;
ecx = 0x0;
cpuid(&eax, &ebx, &ecx, &edx);
if (ebx & cpuid_avx2_bit) {
host_isa |= AVX2;
}
if (ebx & cpuid_bmi1_bit) {
host_isa |= BMI1;
}
Expand All @@ -164,46 +176,51 @@ static inline uint32_t detect_supported_architectures(void) {
host_isa |= BMI2;
}

if (ebx & cpuid_avx2_bit) {
host_avx_isa |= AVX2;
}

if (ebx & cpuid_avx512f_bit) {
host_isa |= AVX512F;
host_avx_isa |= AVX512F;
}

if (ebx & cpuid_avx512dq_bit) {
host_isa |= AVX512DQ;
host_avx_isa |= AVX512DQ;
}

if (ebx & cpuid_avx512ifma_bit) {
host_isa |= AVX512IFMA;
host_avx_isa |= AVX512IFMA;
}

if (ebx & cpuid_avx512pf_bit) {
host_isa |= AVX512PF;
host_avx_isa |= AVX512PF;
}

if (ebx & cpuid_avx512er_bit) {
host_isa |= AVX512ER;
host_avx_isa |= AVX512ER;
}

if (ebx & cpuid_avx512cd_bit) {
host_isa |= AVX512CD;
host_avx_isa |= AVX512CD;
}

if (ebx & cpuid_avx512bw_bit) {
host_isa |= AVX512BW;
host_avx_isa |= AVX512BW;
}

if (ebx & cpuid_avx512vl_bit) {
host_isa |= AVX512VL;
host_avx_isa |= AVX512VL;
}

if (ecx & cpuid_avx512vbmi2_bit) {
host_isa |= AVX512VBMI2;
host_avx_isa |= AVX512VBMI2;
}

bool have_avx = false, have_xgetbv = false;

// EBX for EAX=0x1
eax = 0x1;
cpuid(&eax, &ebx, &ecx, &edx);

if (ecx & cpuid_sse42_bit) {
host_isa |= SSE42;
}
Expand All @@ -212,6 +229,34 @@ static inline uint32_t detect_supported_architectures(void) {
host_isa |= PCLMULQDQ;
}

// Correct detection of AVX2 support requires more than checking the CPUID
// bit. Peter Cordes provides an excellent answer on Stack Overflow
// (https://stackoverflow.com/a/34071400) quoting the article Introduction
// to Intel Advanced Vector Extensions (search Wayback Machine).
//
// 1. Verify that the operating system supports XGETBV using
// CPUID.1:ECX.OSXSAVE bit 27 = 1.
// 2. Verify the processor supports the AVX instruction extensions using:
// CPUID.1:ECX bit 28 = 1.
// 3. Issue XGETBV, and verify that the feature-enabled mask at bits 1 and 2
// are 11b (XMM state and YMM state enabled by the operating system).


// Determine if the CPU supports AVX
have_avx = (ecx & cpuid_have_avx_bit) != 0;
// Determine if the Operating System supports XGETBV
have_xgetbv = (ecx & cpuid_have_xgetbv_bit) != 0;

if (have_avx && have_xgetbv) {
uint64_t xcr0;
ecx = 0x0;
xgetbv(&edx, &eax, &ecx);

xcr0 = ((uint64_t)edx << 32) | (uint64_t)eax;
if ((xcr0 & 0x6) == 0x6)
host_isa |= host_avx_isa;
}

return host_isa;
}
#else // fallback
Expand Down

0 comments on commit 00cb72e

Please sign in to comment.