From 2d2f3d4eb20d04893c2f1a2af5a3c5b1179f005e Mon Sep 17 00:00:00 2001 From: SChernykh Date: Tue, 9 Jun 2020 17:47:23 +0200 Subject: [PATCH] Fixed detection of AVX2/AVX512 --- .../argon2/arch/x86_64/lib/argon2-avx512f.c | 5 +---- src/backend/cpu/platform/BasicCpuInfo.cpp | 18 ++++++++++++++++-- 2 files changed, 17 insertions(+), 6 deletions(-) diff --git a/src/3rdparty/argon2/arch/x86_64/lib/argon2-avx512f.c b/src/3rdparty/argon2/arch/x86_64/lib/argon2-avx512f.c index 282c2f419..f9df1b875 100644 --- a/src/3rdparty/argon2/arch/x86_64/lib/argon2-avx512f.c +++ b/src/3rdparty/argon2/arch/x86_64/lib/argon2-avx512f.c @@ -309,10 +309,7 @@ void xmrig_ar2_fill_segment_avx512f(const argon2_instance_t *instance, argon2_po } extern int cpu_flags_has_avx512f(void); - -// Argon2 AVX-512F implementation is broken -// TODO: enable it back when it's fixed -int xmrig_ar2_check_avx512f(void) { return 0 /*cpu_flags_has_avx512f()*/; } +int xmrig_ar2_check_avx512f(void) { return cpu_flags_has_avx512f(); } #else diff --git a/src/backend/cpu/platform/BasicCpuInfo.cpp b/src/backend/cpu/platform/BasicCpuInfo.cpp index eb0429805..bf152851f 100644 --- a/src/backend/cpu/platform/BasicCpuInfo.cpp +++ b/src/backend/cpu/platform/BasicCpuInfo.cpp @@ -119,10 +119,24 @@ static inline int32_t get_masked(int32_t val, int32_t h, int32_t l) } +static inline uint64_t xgetbv() +{ +#ifdef _MSC_VER + return _xgetbv(_XCR_XFEATURE_ENABLED_MASK); +#else + uint32_t eax_reg = 0; + uint32_t edx_reg = 0; + __asm__ __volatile__("xgetbv": "=a"(eax_reg), "=d"(edx_reg) : "c"(0) : "cc"); + return (static_cast(edx_reg) << 32) | eax_reg; +#endif +} + +static inline bool has_xcr_avx2() { return (xgetbv() & 0x06) == 0x06; } +static inline bool has_xcr_avx512() { return (xgetbv() & 0xE6) == 0xE6; } static inline bool has_osxsave() { return has_feature(PROCESSOR_INFO, ECX_Reg, 1 << 27); } static inline bool has_aes_ni() { return has_feature(PROCESSOR_INFO, ECX_Reg, 1 << 25); } -static inline bool has_avx2() { return has_feature(EXTENDED_FEATURES, EBX_Reg, 1 << 5) && has_osxsave(); } -static inline bool has_avx512f() { return has_feature(EXTENDED_FEATURES, EBX_Reg, 1 << 16) && has_osxsave(); } +static inline bool has_avx2() { return has_feature(EXTENDED_FEATURES, EBX_Reg, 1 << 5) && has_osxsave() && has_xcr_avx2(); } +static inline bool has_avx512f() { return has_feature(EXTENDED_FEATURES, EBX_Reg, 1 << 16) && has_osxsave() && has_xcr_avx512(); } static inline bool has_bmi2() { return has_feature(EXTENDED_FEATURES, EBX_Reg, 1 << 8); } static inline bool has_pdpe1gb() { return has_feature(PROCESSOR_EXT_INFO, EDX_Reg, 1 << 26); } static inline bool has_sse2() { return has_feature(PROCESSOR_INFO, EDX_Reg, 1 << 26); }