diff options
Diffstat (limited to '')
-rw-r--r-- | src/common/x64/cpu_detect.cpp | 123 | ||||
-rw-r--r-- | src/common/x64/cpu_detect.h | 79 | ||||
-rw-r--r-- | src/common/x64/native_clock.cpp | 61 | ||||
-rw-r--r-- | src/common/x64/native_clock.h | 2 |
4 files changed, 159 insertions, 106 deletions
diff --git a/src/common/x64/cpu_detect.cpp b/src/common/x64/cpu_detect.cpp index fbeacc7e2..f5296b32a 100644 --- a/src/common/x64/cpu_detect.cpp +++ b/src/common/x64/cpu_detect.cpp @@ -1,8 +1,11 @@ -// Copyright 2013 Dolphin Emulator Project / 2015 Citra Emulator Project -// Licensed under GPLv2 or any later version -// Refer to the license.txt file included. +// Copyright 2013 Dolphin Emulator Project / 2015 Citra Emulator Project / 2022 Yuzu Emulator +// Project Licensed under GPLv2 or any later version Refer to the license.txt file included. +#include <array> #include <cstring> +#include <iterator> +#include <string_view> +#include "common/bit_util.h" #include "common/common_types.h" #include "common/x64/cpu_detect.h" @@ -17,7 +20,7 @@ // clang-format on #endif -static inline void __cpuidex(int info[4], int function_id, int subfunction_id) { +static inline void __cpuidex(int info[4], u32 function_id, u32 subfunction_id) { #if defined(__DragonFly__) || defined(__FreeBSD__) // Despite the name, this is just do_cpuid() with ECX as second input. cpuid_count((u_int)function_id, (u_int)subfunction_id, (u_int*)info); @@ -30,7 +33,7 @@ static inline void __cpuidex(int info[4], int function_id, int subfunction_id) { #endif } -static inline void __cpuid(int info[4], int function_id) { +static inline void __cpuid(int info[4], u32 function_id) { return __cpuidex(info, function_id, 0); } @@ -45,6 +48,17 @@ static inline u64 _xgetbv(u32 index) { namespace Common { +CPUCaps::Manufacturer CPUCaps::ParseManufacturer(std::string_view brand_string) { + if (brand_string == "GenuineIntel") { + return Manufacturer::Intel; + } else if (brand_string == "AuthenticAMD") { + return Manufacturer::AMD; + } else if (brand_string == "HygonGenuine") { + return Manufacturer::Hygon; + } + return Manufacturer::Unknown; +} + // Detects the various CPU features static CPUCaps Detect() { CPUCaps caps = {}; @@ -53,75 +67,74 @@ static CPUCaps Detect() { // yuzu at all anyway int cpu_id[4]; - memset(caps.brand_string, 0, sizeof(caps.brand_string)); - // Detect CPU's CPUID capabilities and grab CPU string + // Detect CPU's CPUID capabilities and grab manufacturer string __cpuid(cpu_id, 0x00000000); - u32 max_std_fn = cpu_id[0]; // EAX - - std::memcpy(&caps.brand_string[0], &cpu_id[1], sizeof(int)); - std::memcpy(&caps.brand_string[4], &cpu_id[3], sizeof(int)); - std::memcpy(&caps.brand_string[8], &cpu_id[2], sizeof(int)); - if (cpu_id[1] == 0x756e6547 && cpu_id[2] == 0x6c65746e && cpu_id[3] == 0x49656e69) - caps.manufacturer = Manufacturer::Intel; - else if (cpu_id[1] == 0x68747541 && cpu_id[2] == 0x444d4163 && cpu_id[3] == 0x69746e65) - caps.manufacturer = Manufacturer::AMD; - else if (cpu_id[1] == 0x6f677948 && cpu_id[2] == 0x656e6975 && cpu_id[3] == 0x6e65476e) - caps.manufacturer = Manufacturer::Hygon; - else - caps.manufacturer = Manufacturer::Unknown; + const u32 max_std_fn = cpu_id[0]; // EAX - __cpuid(cpu_id, 0x80000000); + std::memset(caps.brand_string, 0, std::size(caps.brand_string)); + std::memcpy(&caps.brand_string[0], &cpu_id[1], sizeof(u32)); + std::memcpy(&caps.brand_string[4], &cpu_id[3], sizeof(u32)); + std::memcpy(&caps.brand_string[8], &cpu_id[2], sizeof(u32)); + + caps.manufacturer = CPUCaps::ParseManufacturer(caps.brand_string); - u32 max_ex_fn = cpu_id[0]; + // Set reasonable default cpu string even if brand string not available + std::strncpy(caps.cpu_string, caps.brand_string, std::size(caps.brand_string)); - // Set reasonable default brand string even if brand string not available - strcpy(caps.cpu_string, caps.brand_string); + __cpuid(cpu_id, 0x80000000); + + const u32 max_ex_fn = cpu_id[0]; // Detect family and other miscellaneous features if (max_std_fn >= 1) { __cpuid(cpu_id, 0x00000001); - if ((cpu_id[3] >> 25) & 1) - caps.sse = true; - if ((cpu_id[3] >> 26) & 1) - caps.sse2 = true; - if ((cpu_id[2]) & 1) - caps.sse3 = true; - if ((cpu_id[2] >> 9) & 1) - caps.ssse3 = true; - if ((cpu_id[2] >> 19) & 1) - caps.sse4_1 = true; - if ((cpu_id[2] >> 20) & 1) - caps.sse4_2 = true; - if ((cpu_id[2] >> 25) & 1) - caps.aes = true; + caps.sse = Common::Bit<25>(cpu_id[3]); + caps.sse2 = Common::Bit<26>(cpu_id[3]); + caps.sse3 = Common::Bit<0>(cpu_id[2]); + caps.pclmulqdq = Common::Bit<1>(cpu_id[2]); + caps.ssse3 = Common::Bit<9>(cpu_id[2]); + caps.sse4_1 = Common::Bit<19>(cpu_id[2]); + caps.sse4_2 = Common::Bit<20>(cpu_id[2]); + caps.movbe = Common::Bit<22>(cpu_id[2]); + caps.popcnt = Common::Bit<23>(cpu_id[2]); + caps.aes = Common::Bit<25>(cpu_id[2]); + caps.f16c = Common::Bit<29>(cpu_id[2]); // AVX support requires 3 separate checks: // - Is the AVX bit set in CPUID? // - Is the XSAVE bit set in CPUID? // - XGETBV result has the XCR bit set. - if (((cpu_id[2] >> 28) & 1) && ((cpu_id[2] >> 27) & 1)) { + if (Common::Bit<28>(cpu_id[2]) && Common::Bit<27>(cpu_id[2])) { if ((_xgetbv(_XCR_XFEATURE_ENABLED_MASK) & 0x6) == 0x6) { caps.avx = true; - if ((cpu_id[2] >> 12) & 1) + if (Common::Bit<12>(cpu_id[2])) caps.fma = true; } } if (max_std_fn >= 7) { __cpuidex(cpu_id, 0x00000007, 0x00000000); - // Can't enable AVX2 unless the XSAVE/XGETBV checks above passed - if ((cpu_id[1] >> 5) & 1) - caps.avx2 = caps.avx; - if ((cpu_id[1] >> 3) & 1) - caps.bmi1 = true; - if ((cpu_id[1] >> 8) & 1) - caps.bmi2 = true; - // Checks for AVX512F, AVX512CD, AVX512VL, AVX512DQ, AVX512BW (Intel Skylake-X/SP) - if ((cpu_id[1] >> 16) & 1 && (cpu_id[1] >> 28) & 1 && (cpu_id[1] >> 31) & 1 && - (cpu_id[1] >> 17) & 1 && (cpu_id[1] >> 30) & 1) { - caps.avx512 = caps.avx2; + // Can't enable AVX{2,512} unless the XSAVE/XGETBV checks above passed + if (caps.avx) { + caps.avx2 = Common::Bit<5>(cpu_id[1]); + caps.avx512f = Common::Bit<16>(cpu_id[1]); + caps.avx512dq = Common::Bit<17>(cpu_id[1]); + caps.avx512cd = Common::Bit<28>(cpu_id[1]); + caps.avx512bw = Common::Bit<30>(cpu_id[1]); + caps.avx512vl = Common::Bit<31>(cpu_id[1]); + caps.avx512vbmi = Common::Bit<1>(cpu_id[2]); + caps.avx512bitalg = Common::Bit<12>(cpu_id[2]); } + + caps.bmi1 = Common::Bit<3>(cpu_id[1]); + caps.bmi2 = Common::Bit<8>(cpu_id[1]); + caps.sha = Common::Bit<29>(cpu_id[1]); + + caps.gfni = Common::Bit<8>(cpu_id[2]); + + __cpuidex(cpu_id, 0x00000007, 0x00000001); + caps.avx_vnni = caps.avx && Common::Bit<4>(cpu_id[0]); } } @@ -138,15 +151,13 @@ static CPUCaps Detect() { if (max_ex_fn >= 0x80000001) { // Check for more features __cpuid(cpu_id, 0x80000001); - if ((cpu_id[2] >> 16) & 1) - caps.fma4 = true; + caps.lzcnt = Common::Bit<5>(cpu_id[2]); + caps.fma4 = Common::Bit<16>(cpu_id[2]); } if (max_ex_fn >= 0x80000007) { __cpuid(cpu_id, 0x80000007); - if (cpu_id[3] & (1 << 8)) { - caps.invariant_tsc = true; - } + caps.invariant_tsc = Common::Bit<8>(cpu_id[3]); } if (max_std_fn >= 0x16) { diff --git a/src/common/x64/cpu_detect.h b/src/common/x64/cpu_detect.h index e3b63302e..40c48b132 100644 --- a/src/common/x64/cpu_detect.h +++ b/src/common/x64/cpu_detect.h @@ -1,42 +1,65 @@ -// Copyright 2013 Dolphin Emulator Project / 2015 Citra Emulator Project -// Licensed under GPLv2 or any later version -// Refer to the license.txt file included. +// Copyright 2013 Dolphin Emulator Project / 2015 Citra Emulator Project / 2022 Yuzu Emulator +// Project Project Licensed under GPLv2 or any later version Refer to the license.txt file included. #pragma once -namespace Common { +#include <string_view> +#include "common/common_types.h" -enum class Manufacturer : u32 { - Intel = 0, - AMD = 1, - Hygon = 2, - Unknown = 3, -}; +namespace Common { /// x86/x64 CPU capabilities that may be detected by this module struct CPUCaps { + + enum class Manufacturer : u8 { + Unknown = 0, + Intel = 1, + AMD = 2, + Hygon = 3, + }; + + static Manufacturer ParseManufacturer(std::string_view brand_string); + Manufacturer manufacturer; - char cpu_string[0x21]; - char brand_string[0x41]; - bool sse; - bool sse2; - bool sse3; - bool ssse3; - bool sse4_1; - bool sse4_2; - bool lzcnt; - bool avx; - bool avx2; - bool avx512; - bool bmi1; - bool bmi2; - bool fma; - bool fma4; - bool aes; - bool invariant_tsc; + char brand_string[13]; + + char cpu_string[48]; + u32 base_frequency; u32 max_frequency; u32 bus_frequency; + + bool sse : 1; + bool sse2 : 1; + bool sse3 : 1; + bool ssse3 : 1; + bool sse4_1 : 1; + bool sse4_2 : 1; + + bool avx : 1; + bool avx_vnni : 1; + bool avx2 : 1; + bool avx512f : 1; + bool avx512dq : 1; + bool avx512cd : 1; + bool avx512bw : 1; + bool avx512vl : 1; + bool avx512vbmi : 1; + bool avx512bitalg : 1; + + bool aes : 1; + bool bmi1 : 1; + bool bmi2 : 1; + bool f16c : 1; + bool fma : 1; + bool fma4 : 1; + bool gfni : 1; + bool invariant_tsc : 1; + bool lzcnt : 1; + bool movbe : 1; + bool pclmulqdq : 1; + bool popcnt : 1; + bool sha : 1; }; /** diff --git a/src/common/x64/native_clock.cpp b/src/common/x64/native_clock.cpp index 91b842829..7fd9d22f8 100644 --- a/src/common/x64/native_clock.cpp +++ b/src/common/x64/native_clock.cpp @@ -4,33 +4,55 @@ #include <array> #include <chrono> -#include <limits> -#include <mutex> #include <thread> #include "common/atomic_ops.h" #include "common/uint128.h" #include "common/x64/native_clock.h" +#ifdef _MSC_VER +#include <intrin.h> +#endif + namespace Common { +#ifdef _MSC_VER +__forceinline static u64 FencedRDTSC() { + _mm_lfence(); + _ReadWriteBarrier(); + const u64 result = __rdtsc(); + _mm_lfence(); + _ReadWriteBarrier(); + return result; +} +#else +static u64 FencedRDTSC() { + u64 result; + asm volatile("lfence\n\t" + "rdtsc\n\t" + "shl $32, %%rdx\n\t" + "or %%rdx, %0\n\t" + "lfence" + : "=a"(result) + : + : "rdx", "memory", "cc"); + return result; +} +#endif + u64 EstimateRDTSCFrequency() { // Discard the first result measuring the rdtsc. - _mm_mfence(); - __rdtsc(); + FencedRDTSC(); std::this_thread::sleep_for(std::chrono::milliseconds{1}); - _mm_mfence(); - __rdtsc(); + FencedRDTSC(); // Get the current time. const auto start_time = std::chrono::steady_clock::now(); - _mm_mfence(); - const u64 tsc_start = __rdtsc(); + const u64 tsc_start = FencedRDTSC(); // Wait for 200 milliseconds. std::this_thread::sleep_for(std::chrono::milliseconds{200}); const auto end_time = std::chrono::steady_clock::now(); - _mm_mfence(); - const u64 tsc_end = __rdtsc(); + const u64 tsc_end = FencedRDTSC(); // Calculate differences. const u64 timer_diff = static_cast<u64>( std::chrono::duration_cast<std::chrono::nanoseconds>(end_time - start_time).count()); @@ -44,8 +66,7 @@ NativeClock::NativeClock(u64 emulated_cpu_frequency_, u64 emulated_clock_frequen u64 rtsc_frequency_) : WallClock(emulated_cpu_frequency_, emulated_clock_frequency_, true), rtsc_frequency{ rtsc_frequency_} { - _mm_mfence(); - time_point.inner.last_measure = __rdtsc(); + time_point.inner.last_measure = FencedRDTSC(); time_point.inner.accumulated_ticks = 0U; ns_rtsc_factor = GetFixedPoint64Factor(NS_RATIO, rtsc_frequency); us_rtsc_factor = GetFixedPoint64Factor(US_RATIO, rtsc_frequency); @@ -57,10 +78,10 @@ NativeClock::NativeClock(u64 emulated_cpu_frequency_, u64 emulated_clock_frequen u64 NativeClock::GetRTSC() { TimePoint new_time_point{}; TimePoint current_time_point{}; + + current_time_point.pack = Common::AtomicLoad128(time_point.pack.data()); do { - current_time_point.pack = time_point.pack; - _mm_mfence(); - const u64 current_measure = __rdtsc(); + const u64 current_measure = FencedRDTSC(); u64 diff = current_measure - current_time_point.inner.last_measure; diff = diff & ~static_cast<u64>(static_cast<s64>(diff) >> 63); // max(diff, 0) new_time_point.inner.last_measure = current_measure > current_time_point.inner.last_measure @@ -68,7 +89,7 @@ u64 NativeClock::GetRTSC() { : current_time_point.inner.last_measure; new_time_point.inner.accumulated_ticks = current_time_point.inner.accumulated_ticks + diff; } while (!Common::AtomicCompareAndSwap(time_point.pack.data(), new_time_point.pack, - current_time_point.pack)); + current_time_point.pack, current_time_point.pack)); /// The clock cannot be more precise than the guest timer, remove the lower bits return new_time_point.inner.accumulated_ticks & inaccuracy_mask; } @@ -77,13 +98,13 @@ void NativeClock::Pause(bool is_paused) { if (!is_paused) { TimePoint current_time_point{}; TimePoint new_time_point{}; + + current_time_point.pack = Common::AtomicLoad128(time_point.pack.data()); do { - current_time_point.pack = time_point.pack; new_time_point.pack = current_time_point.pack; - _mm_mfence(); - new_time_point.inner.last_measure = __rdtsc(); + new_time_point.inner.last_measure = FencedRDTSC(); } while (!Common::AtomicCompareAndSwap(time_point.pack.data(), new_time_point.pack, - current_time_point.pack)); + current_time_point.pack, current_time_point.pack)); } } diff --git a/src/common/x64/native_clock.h b/src/common/x64/native_clock.h index 7cbd400d2..2c3082ea0 100644 --- a/src/common/x64/native_clock.h +++ b/src/common/x64/native_clock.h @@ -4,8 +4,6 @@ #pragma once -#include <optional> - #include "common/wall_clock.h" namespace Common { |