Merge pull request #5275 from FernandoS27/fast-native-clock

X86/NativeClock: Improve performance of clock calculations on hot path.

5 files changed, 174 insertions, 104 deletions

diff --git a/src/common/CMakeLists.txt b/src/common/CMakeLists.txt
index 9824c5564..f77575a00 100644
--- a/src/common/CMakeLists.txt
+++ b/src/common/CMakeLists.txt
@@ -98,7 +98,6 @@ add_library(common STATIC
     algorithm.h
     alignment.h
     assert.h
-    atomic_ops.cpp
     atomic_ops.h
     detached_tasks.cpp
     detached_tasks.h
diff --git a/src/common/atomic_ops.cpp b/src/common/atomic_ops.cpp
deleted file mode 100644
index 1612d0e67..000000000
--- a/src/common/atomic_ops.cpp
+++ /dev/null
@@ -1,75 +0,0 @@
-// Copyright 2020 yuzu Emulator Project
-// Licensed under GPLv2 or any later version
-// Refer to the license.txt file included.
-
-#include <cstring>
-
-#include "common/atomic_ops.h"
-
-#if _MSC_VER
-#include <intrin.h>
-#endif
-
-namespace Common {
-
-#if _MSC_VER
-
-bool AtomicCompareAndSwap(volatile u8* pointer, u8 value, u8 expected) {
-    const u8 result =
-        _InterlockedCompareExchange8(reinterpret_cast<volatile char*>(pointer), value, expected);
-    return result == expected;
-}
-
-bool AtomicCompareAndSwap(volatile u16* pointer, u16 value, u16 expected) {
-    const u16 result =
-        _InterlockedCompareExchange16(reinterpret_cast<volatile short*>(pointer), value, expected);
-    return result == expected;
-}
-
-bool AtomicCompareAndSwap(volatile u32* pointer, u32 value, u32 expected) {
-    const u32 result =
-        _InterlockedCompareExchange(reinterpret_cast<volatile long*>(pointer), value, expected);
-    return result == expected;
-}
-
-bool AtomicCompareAndSwap(volatile u64* pointer, u64 value, u64 expected) {
-    const u64 result = _InterlockedCompareExchange64(reinterpret_cast<volatile __int64*>(pointer),
-                                                     value, expected);
-    return result == expected;
-}
-
-bool AtomicCompareAndSwap(volatile u64* pointer, u128 value, u128 expected) {
-    return _InterlockedCompareExchange128(reinterpret_cast<volatile __int64*>(pointer), value[1],
-                                          value[0],
-                                          reinterpret_cast<__int64*>(expected.data())) != 0;
-}
-
-#else
-
-bool AtomicCompareAndSwap(volatile u8* pointer, u8 value, u8 expected) {
-    return __sync_bool_compare_and_swap(pointer, expected, value);
-}
-
-bool AtomicCompareAndSwap(volatile u16* pointer, u16 value, u16 expected) {
-    return __sync_bool_compare_and_swap(pointer, expected, value);
-}
-
-bool AtomicCompareAndSwap(volatile u32* pointer, u32 value, u32 expected) {
-    return __sync_bool_compare_and_swap(pointer, expected, value);
-}
-
-bool AtomicCompareAndSwap(volatile u64* pointer, u64 value, u64 expected) {
-    return __sync_bool_compare_and_swap(pointer, expected, value);
-}
-
-bool AtomicCompareAndSwap(volatile u64* pointer, u128 value, u128 expected) {
-    unsigned __int128 value_a;
-    unsigned __int128 expected_a;
-    std::memcpy(&value_a, value.data(), sizeof(u128));
-    std::memcpy(&expected_a, expected.data(), sizeof(u128));
-    return __sync_bool_compare_and_swap((unsigned __int128*)pointer, expected_a, value_a);
-}
-
-#endif
-
-} // namespace Common
diff --git a/src/common/atomic_ops.h b/src/common/atomic_ops.h
index b46888589..2b1f515e8 100644
--- a/src/common/atomic_ops.h
+++ b/src/common/atomic_ops.h
@@ -4,14 +4,75 @@
 
 #pragma once
 
+#include <cstring>
+#include <memory>
+
 #include "common/common_types.h"
 
+#if _MSC_VER
+#include <intrin.h>
+#endif
+
 namespace Common {
 
-[[nodiscard]] bool AtomicCompareAndSwap(volatile u8* pointer, u8 value, u8 expected);
-[[nodiscard]] bool AtomicCompareAndSwap(volatile u16* pointer, u16 value, u16 expected);
-[[nodiscard]] bool AtomicCompareAndSwap(volatile u32* pointer, u32 value, u32 expected);
-[[nodiscard]] bool AtomicCompareAndSwap(volatile u64* pointer, u64 value, u64 expected);
-[[nodiscard]] bool AtomicCompareAndSwap(volatile u64* pointer, u128 value, u128 expected);
+#if _MSC_VER
+
+[[nodiscard]] inline bool AtomicCompareAndSwap(volatile u8* pointer, u8 value, u8 expected) {
+    const u8 result =
+        _InterlockedCompareExchange8(reinterpret_cast<volatile char*>(pointer), value, expected);
+    return result == expected;
+}
+
+[[nodiscard]] inline bool AtomicCompareAndSwap(volatile u16* pointer, u16 value, u16 expected) {
+    const u16 result =
+        _InterlockedCompareExchange16(reinterpret_cast<volatile short*>(pointer), value, expected);
+    return result == expected;
+}
+
+[[nodiscard]] inline bool AtomicCompareAndSwap(volatile u32* pointer, u32 value, u32 expected) {
+    const u32 result =
+        _InterlockedCompareExchange(reinterpret_cast<volatile long*>(pointer), value, expected);
+    return result == expected;
+}
+
+[[nodiscard]] inline bool AtomicCompareAndSwap(volatile u64* pointer, u64 value, u64 expected) {
+    const u64 result = _InterlockedCompareExchange64(reinterpret_cast<volatile __int64*>(pointer),
+                                                     value, expected);
+    return result == expected;
+}
+
+[[nodiscard]] inline bool AtomicCompareAndSwap(volatile u64* pointer, u128 value, u128 expected) {
+    return _InterlockedCompareExchange128(reinterpret_cast<volatile __int64*>(pointer), value[1],
+                                          value[0],
+                                          reinterpret_cast<__int64*>(expected.data())) != 0;
+}
+
+#else
+
+[[nodiscard]] inline bool AtomicCompareAndSwap(volatile u8* pointer, u8 value, u8 expected) {
+    return __sync_bool_compare_and_swap(pointer, expected, value);
+}
+
+[[nodiscard]] inline bool AtomicCompareAndSwap(volatile u16* pointer, u16 value, u16 expected) {
+    return __sync_bool_compare_and_swap(pointer, expected, value);
+}
+
+[[nodiscard]] inline bool AtomicCompareAndSwap(volatile u32* pointer, u32 value, u32 expected) {
+    return __sync_bool_compare_and_swap(pointer, expected, value);
+}
+
+[[nodiscard]] inline bool AtomicCompareAndSwap(volatile u64* pointer, u64 value, u64 expected) {
+    return __sync_bool_compare_and_swap(pointer, expected, value);
+}
+
+[[nodiscard]] inline bool AtomicCompareAndSwap(volatile u64* pointer, u128 value, u128 expected) {
+    unsigned __int128 value_a;
+    unsigned __int128 expected_a;
+    std::memcpy(&value_a, value.data(), sizeof(u128));
+    std::memcpy(&expected_a, expected.data(), sizeof(u128));
+    return __sync_bool_compare_and_swap((unsigned __int128*)pointer, expected_a, value_a);
+}
+
+#endif
 
 } // namespace Common
diff --git a/src/common/x64/native_clock.cpp b/src/common/x64/native_clock.cpp
index eb8a7782f..a65f6b832 100644
--- a/src/common/x64/native_clock.cpp
+++ b/src/common/x64/native_clock.cpp
@@ -2,19 +2,74 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
+#include <array>
 #include <chrono>
+#include <limits>
 #include <mutex>
 #include <thread>
 
 #ifdef _MSC_VER
 #include <intrin.h>
+
+#pragma intrinsic(__umulh)
+#pragma intrinsic(_udiv128)
 #else
 #include <x86intrin.h>
 #endif
 
+#include "common/atomic_ops.h"
 #include "common/uint128.h"
 #include "common/x64/native_clock.h"
 
+namespace {
+
+[[nodiscard]] u64 GetFixedPoint64Factor(u64 numerator, u64 divisor) {
+#ifdef __SIZEOF_INT128__
+    const auto base = static_cast<unsigned __int128>(numerator) << 64ULL;
+    return static_cast<u64>(base / divisor);
+#elif defined(_M_X64) || defined(_M_ARM64)
+    std::array<u64, 2> r = {0, numerator};
+    u64 remainder;
+#if _MSC_VER < 1923
+    return udiv128(r[1], r[0], divisor, &remainder);
+#else
+    return _udiv128(r[1], r[0], divisor, &remainder);
+#endif
+#else
+    // This one is bit more inaccurate.
+    return MultiplyAndDivide64(std::numeric_limits<u64>::max(), numerator, divisor);
+#endif
+}
+
+[[nodiscard]] u64 MultiplyHigh(u64 a, u64 b) {
+#ifdef __SIZEOF_INT128__
+    return (static_cast<unsigned __int128>(a) * static_cast<unsigned __int128>(b)) >> 64;
+#elif defined(_M_X64) || defined(_M_ARM64)
+    return __umulh(a, b); // MSVC
+#else
+    // Generic fallback
+    const u64 a_lo = u32(a);
+    const u64 a_hi = a >> 32;
+    const u64 b_lo = u32(b);
+    const u64 b_hi = b >> 32;
+
+    const u64 a_x_b_hi = a_hi * b_hi;
+    const u64 a_x_b_mid = a_hi * b_lo;
+    const u64 b_x_a_mid = b_hi * a_lo;
+    const u64 a_x_b_lo = a_lo * b_lo;
+
+    const u64 carry_bit = (static_cast<u64>(static_cast<u32>(a_x_b_mid)) +
+                           static_cast<u64>(static_cast<u32>(b_x_a_mid)) + (a_x_b_lo >> 32)) >>
+                          32;
+
+    const u64 multhi = a_x_b_hi + (a_x_b_mid >> 32) + (b_x_a_mid >> 32) + carry_bit;
+
+    return multhi;
+#endif
+}
+
+} // namespace
+
 namespace Common {
 
 u64 EstimateRDTSCFrequency() {
@@ -48,54 +103,71 @@ NativeClock::NativeClock(u64 emulated_cpu_frequency_, u64 emulated_clock_frequen
     : WallClock(emulated_cpu_frequency_, emulated_clock_frequency_, true), rtsc_frequency{
                                                                                rtsc_frequency_} {
     _mm_mfence();
-    last_measure = __rdtsc();
-    accumulated_ticks = 0U;
+    time_point.inner.last_measure = __rdtsc();
+    time_point.inner.accumulated_ticks = 0U;
+    ns_rtsc_factor = GetFixedPoint64Factor(1000000000, rtsc_frequency);
+    us_rtsc_factor = GetFixedPoint64Factor(1000000, rtsc_frequency);
+    ms_rtsc_factor = GetFixedPoint64Factor(1000, rtsc_frequency);
+    clock_rtsc_factor = GetFixedPoint64Factor(emulated_clock_frequency, rtsc_frequency);
+    cpu_rtsc_factor = GetFixedPoint64Factor(emulated_cpu_frequency, rtsc_frequency);
 }
 
 u64 NativeClock::GetRTSC() {
-    std::scoped_lock scope{rtsc_serialize};
-    _mm_mfence();
-    const u64 current_measure = __rdtsc();
-    u64 diff = current_measure - last_measure;
-    diff = diff & ~static_cast<u64>(static_cast<s64>(diff) >> 63); // max(diff, 0)
-    if (current_measure > last_measure) {
-        last_measure = current_measure;
-    }
-    accumulated_ticks += diff;
+    TimePoint new_time_point{};
+    TimePoint current_time_point{};
+    do {
+        current_time_point.pack = time_point.pack;
+        _mm_mfence();
+        const u64 current_measure = __rdtsc();
+        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
+                                                ? current_measure
+                                                : 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));
     /// The clock cannot be more precise than the guest timer, remove the lower bits
-    return accumulated_ticks & inaccuracy_mask;
+    return new_time_point.inner.accumulated_ticks & inaccuracy_mask;
 }
 
 void NativeClock::Pause(bool is_paused) {
     if (!is_paused) {
-        _mm_mfence();
-        last_measure = __rdtsc();
+        TimePoint current_time_point{};
+        TimePoint new_time_point{};
+        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();
+        } while (!Common::AtomicCompareAndSwap(time_point.pack.data(), new_time_point.pack,
+                                               current_time_point.pack));
     }
 }
 
 std::chrono::nanoseconds NativeClock::GetTimeNS() {
     const u64 rtsc_value = GetRTSC();
-    return std::chrono::nanoseconds{MultiplyAndDivide64(rtsc_value, 1000000000, rtsc_frequency)};
+    return std::chrono::nanoseconds{MultiplyHigh(rtsc_value, ns_rtsc_factor)};
 }
 
 std::chrono::microseconds NativeClock::GetTimeUS() {
     const u64 rtsc_value = GetRTSC();
-    return std::chrono::microseconds{MultiplyAndDivide64(rtsc_value, 1000000, rtsc_frequency)};
+    return std::chrono::microseconds{MultiplyHigh(rtsc_value, us_rtsc_factor)};
 }
 
 std::chrono::milliseconds NativeClock::GetTimeMS() {
     const u64 rtsc_value = GetRTSC();
-    return std::chrono::milliseconds{MultiplyAndDivide64(rtsc_value, 1000, rtsc_frequency)};
+    return std::chrono::milliseconds{MultiplyHigh(rtsc_value, ms_rtsc_factor)};
 }
 
 u64 NativeClock::GetClockCycles() {
     const u64 rtsc_value = GetRTSC();
-    return MultiplyAndDivide64(rtsc_value, emulated_clock_frequency, rtsc_frequency);
+    return MultiplyHigh(rtsc_value, clock_rtsc_factor);
 }
 
 u64 NativeClock::GetCPUCycles() {
     const u64 rtsc_value = GetRTSC();
-    return MultiplyAndDivide64(rtsc_value, emulated_cpu_frequency, rtsc_frequency);
+    return MultiplyHigh(rtsc_value, cpu_rtsc_factor);
 }
 
 } // namespace X64
diff --git a/src/common/x64/native_clock.h b/src/common/x64/native_clock.h
index 6d1e32ac8..7cbd400d2 100644
--- a/src/common/x64/native_clock.h
+++ b/src/common/x64/native_clock.h
@@ -6,7 +6,6 @@
 
 #include <optional>
 
-#include "common/spin_lock.h"
 #include "common/wall_clock.h"
 
 namespace Common {
@@ -32,14 +31,28 @@ public:
 private:
     u64 GetRTSC();
 
+    union alignas(16) TimePoint {
+        TimePoint() : pack{} {}
+        u128 pack{};
+        struct Inner {
+            u64 last_measure{};
+            u64 accumulated_ticks{};
+        } inner;
+    };
+
     /// value used to reduce the native clocks accuracy as some apss rely on
     /// undefined behavior where the level of accuracy in the clock shouldn't
     /// be higher.
     static constexpr u64 inaccuracy_mask = ~(UINT64_C(0x400) - 1);
 
-    SpinLock rtsc_serialize{};
-    u64 last_measure{};
-    u64 accumulated_ticks{};
+    TimePoint time_point;
+    // factors
+    u64 clock_rtsc_factor{};
+    u64 cpu_rtsc_factor{};
+    u64 ns_rtsc_factor{};
+    u64 us_rtsc_factor{};
+    u64 ms_rtsc_factor{};
+
     u64 rtsc_frequency;
 };
 } // namespace X64