From dd12dd4c67dd4bc6e7a7d071b925afc38e687f8b Mon Sep 17 00:00:00 2001
From: Morph <39850852+Morph1984@users.noreply.github.com>
Date: Sat, 22 Apr 2023 22:55:03 -0400
Subject: x64: Deduplicate RDTSC usage

---
 src/common/x64/cpu_detect.cpp |  3 +++
 src/common/x64/cpu_wait.cpp   | 20 +-------------------
 src/common/x64/rdtsc.cpp      | 39 +++++++++++++++++++++++++++++++++++++++
 src/common/x64/rdtsc.h        | 37 +++++++++++++++++++++++++++++++++++++
 4 files changed, 80 insertions(+), 19 deletions(-)
 create mode 100644 src/common/x64/rdtsc.cpp
 create mode 100644 src/common/x64/rdtsc.h

(limited to 'src/common/x64')

diff --git a/src/common/x64/cpu_detect.cpp b/src/common/x64/cpu_detect.cpp
index 72ed6e96c..c998b1197 100644
--- a/src/common/x64/cpu_detect.cpp
+++ b/src/common/x64/cpu_detect.cpp
@@ -14,6 +14,7 @@
 #include "common/common_types.h"
 #include "common/logging/log.h"
 #include "common/x64/cpu_detect.h"
+#include "common/x64/rdtsc.h"
 
 #ifdef _WIN32
 #include <windows.h>
@@ -187,6 +188,8 @@ static CPUCaps Detect() {
             caps.tsc_frequency = static_cast<u64>(caps.crystal_frequency) *
                                  caps.tsc_crystal_ratio_numerator /
                                  caps.tsc_crystal_ratio_denominator;
+        } else {
+            caps.tsc_frequency = X64::EstimateRDTSCFrequency();
         }
     }
 
diff --git a/src/common/x64/cpu_wait.cpp b/src/common/x64/cpu_wait.cpp
index cfeef6a3d..c53dd4945 100644
--- a/src/common/x64/cpu_wait.cpp
+++ b/src/common/x64/cpu_wait.cpp
@@ -9,19 +9,11 @@
 
 #include "common/x64/cpu_detect.h"
 #include "common/x64/cpu_wait.h"
+#include "common/x64/rdtsc.h"
 
 namespace Common::X64 {
 
 #ifdef _MSC_VER
-__forceinline static u64 FencedRDTSC() {
-    _mm_lfence();
-    _ReadWriteBarrier();
-    const u64 result = __rdtsc();
-    _mm_lfence();
-    _ReadWriteBarrier();
-    return result;
-}
-
 __forceinline static void TPAUSE() {
     // 100,000 cycles is a reasonable amount of time to wait to save on CPU resources.
     // For reference:
@@ -32,16 +24,6 @@ __forceinline static void TPAUSE() {
     _tpause(0, FencedRDTSC() + PauseCycles);
 }
 #else
-static u64 FencedRDTSC() {
-    u64 eax;
-    u64 edx;
-    asm volatile("lfence\n\t"
-                 "rdtsc\n\t"
-                 "lfence\n\t"
-                 : "=a"(eax), "=d"(edx));
-    return (edx << 32) | eax;
-}
-
 static void TPAUSE() {
     // 100,000 cycles is a reasonable amount of time to wait to save on CPU resources.
     // For reference:
diff --git a/src/common/x64/rdtsc.cpp b/src/common/x64/rdtsc.cpp
new file mode 100644
index 000000000..9273274a3
--- /dev/null
+++ b/src/common/x64/rdtsc.cpp
@@ -0,0 +1,39 @@
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#include <thread>
+
+#include "common/steady_clock.h"
+#include "common/uint128.h"
+#include "common/x64/rdtsc.h"
+
+namespace Common::X64 {
+
+template <u64 Nearest>
+static u64 RoundToNearest(u64 value) {
+    const auto mod = value % Nearest;
+    return mod >= (Nearest / 2) ? (value - mod + Nearest) : (value - mod);
+}
+
+u64 EstimateRDTSCFrequency() {
+    // Discard the first result measuring the rdtsc.
+    FencedRDTSC();
+    std::this_thread::sleep_for(std::chrono::milliseconds{1});
+    FencedRDTSC();
+
+    // Get the current time.
+    const auto start_time = RealTimeClock::Now();
+    const u64 tsc_start = FencedRDTSC();
+    // Wait for 100 milliseconds.
+    std::this_thread::sleep_for(std::chrono::milliseconds{100});
+    const auto end_time = RealTimeClock::Now();
+    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());
+    const u64 tsc_diff = tsc_end - tsc_start;
+    const u64 tsc_freq = MultiplyAndDivide64(tsc_diff, 1000000000ULL, timer_diff);
+    return RoundToNearest<100'000>(tsc_freq);
+}
+
+} // namespace Common::X64
diff --git a/src/common/x64/rdtsc.h b/src/common/x64/rdtsc.h
new file mode 100644
index 000000000..0ec4f52f9
--- /dev/null
+++ b/src/common/x64/rdtsc.h
@@ -0,0 +1,37 @@
+// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
+// SPDX-License-Identifier: GPL-2.0-or-later
+
+#pragma once
+
+#ifdef _MSC_VER
+#include <intrin.h>
+#endif
+
+#include "common/common_types.h"
+
+namespace Common::X64 {
+
+#ifdef _MSC_VER
+__forceinline static u64 FencedRDTSC() {
+    _mm_lfence();
+    _ReadWriteBarrier();
+    const u64 result = __rdtsc();
+    _mm_lfence();
+    _ReadWriteBarrier();
+    return result;
+}
+#else
+static inline u64 FencedRDTSC() {
+    u64 eax;
+    u64 edx;
+    asm volatile("lfence\n\t"
+                 "rdtsc\n\t"
+                 "lfence\n\t"
+                 : "=a"(eax), "=d"(edx));
+    return (edx << 32) | eax;
+}
+#endif
+
+u64 EstimateRDTSCFrequency();
+
+} // namespace Common::X64
-- 
cgit v1.2.3


From 1492a65454d6a03f641b136cc61e68870be00218 Mon Sep 17 00:00:00 2001
From: Morph <39850852+Morph1984@users.noreply.github.com>
Date: Sat, 22 Apr 2023 23:08:28 -0400
Subject: (wall, native)_clock: Rework NativeClock

---
 src/common/steady_clock.cpp     |   5 +-
 src/common/wall_clock.cpp       |  73 +++++++-----------
 src/common/wall_clock.h         |  54 ++++++-------
 src/common/x64/native_clock.cpp | 165 ++++++----------------------------------
 src/common/x64/native_clock.h   |  56 +++++---------
 5 files changed, 94 insertions(+), 259 deletions(-)

(limited to 'src/common/x64')

diff --git a/src/common/steady_clock.cpp b/src/common/steady_clock.cpp
index 782859196..9415eed29 100644
--- a/src/common/steady_clock.cpp
+++ b/src/common/steady_clock.cpp
@@ -28,13 +28,12 @@ static s64 GetSystemTimeNS() {
     // GetSystemTimePreciseAsFileTime returns the file time in 100ns units.
     static constexpr s64 Multiplier = 100;
     // Convert Windows epoch to Unix epoch.
-    static constexpr s64 WindowsEpochToUnixEpochNS = 0x19DB1DED53E8000LL;
+    static constexpr s64 WindowsEpochToUnixEpoch = 0x19DB1DED53E8000LL;
 
     FILETIME filetime;
     GetSystemTimePreciseAsFileTime(&filetime);
     return Multiplier * ((static_cast<s64>(filetime.dwHighDateTime) << 32) +
-                         static_cast<s64>(filetime.dwLowDateTime)) -
-           WindowsEpochToUnixEpochNS;
+                         static_cast<s64>(filetime.dwLowDateTime) - WindowsEpochToUnixEpoch);
 }
 #endif
 
diff --git a/src/common/wall_clock.cpp b/src/common/wall_clock.cpp
index 817e71d52..ad8db06b0 100644
--- a/src/common/wall_clock.cpp
+++ b/src/common/wall_clock.cpp
@@ -2,88 +2,71 @@
 // SPDX-License-Identifier: GPL-2.0-or-later
 
 #include "common/steady_clock.h"
-#include "common/uint128.h"
 #include "common/wall_clock.h"
 
 #ifdef ARCHITECTURE_x86_64
 #include "common/x64/cpu_detect.h"
 #include "common/x64/native_clock.h"
+#include "common/x64/rdtsc.h"
 #endif
 
 namespace Common {
 
 class StandardWallClock final : public WallClock {
 public:
-    explicit StandardWallClock(u64 emulated_cpu_frequency_, u64 emulated_clock_frequency_)
-        : WallClock{emulated_cpu_frequency_, emulated_clock_frequency_, false},
-          start_time{SteadyClock::Now()} {}
+    explicit StandardWallClock() : start_time{SteadyClock::Now()} {}
 
-    std::chrono::nanoseconds GetTimeNS() override {
+    std::chrono::nanoseconds GetTimeNS() const override {
         return SteadyClock::Now() - start_time;
     }
 
-    std::chrono::microseconds GetTimeUS() override {
-        return std::chrono::duration_cast<std::chrono::microseconds>(GetTimeNS());
+    std::chrono::microseconds GetTimeUS() const override {
+        return static_cast<std::chrono::microseconds>(GetHostTicksElapsed() / NsToUsRatio::den);
     }
 
-    std::chrono::milliseconds GetTimeMS() override {
-        return std::chrono::duration_cast<std::chrono::milliseconds>(GetTimeNS());
+    std::chrono::milliseconds GetTimeMS() const override {
+        return static_cast<std::chrono::milliseconds>(GetHostTicksElapsed() / NsToMsRatio::den);
     }
 
-    u64 GetClockCycles() override {
-        const u128 temp = Common::Multiply64Into128(GetTimeNS().count(), emulated_clock_frequency);
-        return Common::Divide128On32(temp, NS_RATIO).first;
+    u64 GetCNTPCT() const override {
+        return GetHostTicksElapsed() * NsToCNTPCTRatio::num / NsToCNTPCTRatio::den;
     }
 
-    u64 GetCPUCycles() override {
-        const u128 temp = Common::Multiply64Into128(GetTimeNS().count(), emulated_cpu_frequency);
-        return Common::Divide128On32(temp, NS_RATIO).first;
+    u64 GetHostTicksNow() const override {
+        return static_cast<u64>(SteadyClock::Now().time_since_epoch().count());
     }
 
-    void Pause([[maybe_unused]] bool is_paused) override {
-        // Do nothing in this clock type.
+    u64 GetHostTicksElapsed() const override {
+        return static_cast<u64>(GetTimeNS().count());
+    }
+
+    bool IsNative() const override {
+        return false;
     }
 
 private:
     SteadyClock::time_point start_time;
 };
 
+std::unique_ptr<WallClock> CreateOptimalClock() {
 #ifdef ARCHITECTURE_x86_64
-
-std::unique_ptr<WallClock> CreateBestMatchingClock(u64 emulated_cpu_frequency,
-                                                   u64 emulated_clock_frequency) {
     const auto& caps = GetCPUCaps();
-    u64 rtsc_frequency = 0;
-    if (caps.invariant_tsc) {
-        rtsc_frequency = caps.tsc_frequency ? caps.tsc_frequency : EstimateRDTSCFrequency();
-    }
 
-    // Fallback to StandardWallClock if the hardware TSC does not have the precision greater than:
-    // - A nanosecond
-    // - The emulated CPU frequency
-    // - The emulated clock counter frequency (CNTFRQ)
-    if (rtsc_frequency <= WallClock::NS_RATIO || rtsc_frequency <= emulated_cpu_frequency ||
-        rtsc_frequency <= emulated_clock_frequency) {
-        return std::make_unique<StandardWallClock>(emulated_cpu_frequency,
-                                                   emulated_clock_frequency);
+    if (caps.invariant_tsc && caps.tsc_frequency >= WallClock::CNTFRQ) {
+        return std::make_unique<X64::NativeClock>(caps.tsc_frequency);
     } else {
-        return std::make_unique<X64::NativeClock>(emulated_cpu_frequency, emulated_clock_frequency,
-                                                  rtsc_frequency);
+        // Fallback to StandardWallClock if the hardware TSC
+        // - Is not invariant
+        // - Is not more precise than CNTFRQ
+        return std::make_unique<StandardWallClock>();
     }
-}
-
 #else
-
-std::unique_ptr<WallClock> CreateBestMatchingClock(u64 emulated_cpu_frequency,
-                                                   u64 emulated_clock_frequency) {
-    return std::make_unique<StandardWallClock>(emulated_cpu_frequency, emulated_clock_frequency);
-}
-
+    return std::make_unique<StandardWallClock>();
 #endif
+}
 
-std::unique_ptr<WallClock> CreateStandardWallClock(u64 emulated_cpu_frequency,
-                                                   u64 emulated_clock_frequency) {
-    return std::make_unique<StandardWallClock>(emulated_cpu_frequency, emulated_clock_frequency);
+std::unique_ptr<WallClock> CreateStandardWallClock() {
+    return std::make_unique<StandardWallClock>();
 }
 
 } // namespace Common
diff --git a/src/common/wall_clock.h b/src/common/wall_clock.h
index 157ec5eae..a73e6e644 100644
--- a/src/common/wall_clock.h
+++ b/src/common/wall_clock.h
@@ -5,6 +5,7 @@
 
 #include <chrono>
 #include <memory>
+#include <ratio>
 
 #include "common/common_types.h"
 
@@ -12,50 +13,43 @@ namespace Common {
 
 class WallClock {
 public:
-    static constexpr u64 NS_RATIO = 1'000'000'000;
-    static constexpr u64 US_RATIO = 1'000'000;
-    static constexpr u64 MS_RATIO = 1'000;
+    static constexpr u64 CNTFRQ = 19'200'000; // CNTPCT_EL0 Frequency = 19.2 MHz
 
     virtual ~WallClock() = default;
 
-    /// Returns current wall time in nanoseconds
-    [[nodiscard]] virtual std::chrono::nanoseconds GetTimeNS() = 0;
+    /// @returns The time in nanoseconds since the construction of this clock.
+    virtual std::chrono::nanoseconds GetTimeNS() const = 0;
 
-    /// Returns current wall time in microseconds
-    [[nodiscard]] virtual std::chrono::microseconds GetTimeUS() = 0;
+    /// @returns The time in microseconds since the construction of this clock.
+    virtual std::chrono::microseconds GetTimeUS() const = 0;
 
-    /// Returns current wall time in milliseconds
-    [[nodiscard]] virtual std::chrono::milliseconds GetTimeMS() = 0;
+    /// @returns The time in milliseconds since the construction of this clock.
+    virtual std::chrono::milliseconds GetTimeMS() const = 0;
 
-    /// Returns current wall time in emulated clock cycles
-    [[nodiscard]] virtual u64 GetClockCycles() = 0;
+    /// @returns The guest CNTPCT ticks since the construction of this clock.
+    virtual u64 GetCNTPCT() const = 0;
 
-    /// Returns current wall time in emulated cpu cycles
-    [[nodiscard]] virtual u64 GetCPUCycles() = 0;
+    /// @returns The raw host timer ticks since an indeterminate epoch.
+    virtual u64 GetHostTicksNow() const = 0;
 
-    virtual void Pause(bool is_paused) = 0;
+    /// @returns The raw host timer ticks since the construction of this clock.
+    virtual u64 GetHostTicksElapsed() const = 0;
 
-    /// Tells if the wall clock, uses the host CPU's hardware clock
-    [[nodiscard]] bool IsNative() const {
-        return is_native;
-    }
+    /// @returns Whether the clock directly uses the host's hardware clock.
+    virtual bool IsNative() const = 0;
 
 protected:
-    explicit WallClock(u64 emulated_cpu_frequency_, u64 emulated_clock_frequency_, bool is_native_)
-        : emulated_cpu_frequency{emulated_cpu_frequency_},
-          emulated_clock_frequency{emulated_clock_frequency_}, is_native{is_native_} {}
+    using NsRatio = std::nano;
+    using UsRatio = std::micro;
+    using MsRatio = std::milli;
 
-    u64 emulated_cpu_frequency;
-    u64 emulated_clock_frequency;
-
-private:
-    bool is_native;
+    using NsToUsRatio = std::ratio_divide<std::nano, std::micro>;
+    using NsToMsRatio = std::ratio_divide<std::nano, std::milli>;
+    using NsToCNTPCTRatio = std::ratio<CNTFRQ, std::nano::den>;
 };
 
-[[nodiscard]] std::unique_ptr<WallClock> CreateBestMatchingClock(u64 emulated_cpu_frequency,
-                                                                 u64 emulated_clock_frequency);
+std::unique_ptr<WallClock> CreateOptimalClock();
 
-[[nodiscard]] std::unique_ptr<WallClock> CreateStandardWallClock(u64 emulated_cpu_frequency,
-                                                                 u64 emulated_clock_frequency);
+std::unique_ptr<WallClock> CreateStandardWallClock();
 
 } // namespace Common
diff --git a/src/common/x64/native_clock.cpp b/src/common/x64/native_clock.cpp
index 277b00662..5d1eb0590 100644
--- a/src/common/x64/native_clock.cpp
+++ b/src/common/x64/native_clock.cpp
@@ -1,164 +1,45 @@
 // SPDX-FileCopyrightText: Copyright 2020 yuzu Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
-#include <array>
-#include <chrono>
-#include <thread>
-
-#include "common/atomic_ops.h"
-#include "common/steady_clock.h"
 #include "common/uint128.h"
 #include "common/x64/native_clock.h"
+#include "common/x64/rdtsc.h"
 
-#ifdef _MSC_VER
-#include <intrin.h>
-#endif
+namespace Common::X64 {
 
-namespace Common {
+NativeClock::NativeClock(u64 rdtsc_frequency_)
+    : start_ticks{FencedRDTSC()}, rdtsc_frequency{rdtsc_frequency_},
+      ns_rdtsc_factor{GetFixedPoint64Factor(NsRatio::den, rdtsc_frequency)},
+      us_rdtsc_factor{GetFixedPoint64Factor(UsRatio::den, rdtsc_frequency)},
+      ms_rdtsc_factor{GetFixedPoint64Factor(MsRatio::den, rdtsc_frequency)},
+      cntpct_rdtsc_factor{GetFixedPoint64Factor(CNTFRQ, rdtsc_frequency)} {}
 
-#ifdef _MSC_VER
-__forceinline static u64 FencedRDTSC() {
-    _mm_lfence();
-    _ReadWriteBarrier();
-    const u64 result = __rdtsc();
-    _mm_lfence();
-    _ReadWriteBarrier();
-    return result;
-}
-#else
-static u64 FencedRDTSC() {
-    u64 eax;
-    u64 edx;
-    asm volatile("lfence\n\t"
-                 "rdtsc\n\t"
-                 "lfence\n\t"
-                 : "=a"(eax), "=d"(edx));
-    return (edx << 32) | eax;
+std::chrono::nanoseconds NativeClock::GetTimeNS() const {
+    return std::chrono::nanoseconds{MultiplyHigh(GetHostTicksElapsed(), ns_rdtsc_factor)};
 }
-#endif
 
-template <u64 Nearest>
-static u64 RoundToNearest(u64 value) {
-    const auto mod = value % Nearest;
-    return mod >= (Nearest / 2) ? (value - mod + Nearest) : (value - mod);
+std::chrono::microseconds NativeClock::GetTimeUS() const {
+    return std::chrono::microseconds{MultiplyHigh(GetHostTicksElapsed(), us_rdtsc_factor)};
 }
 
-u64 EstimateRDTSCFrequency() {
-    // Discard the first result measuring the rdtsc.
-    FencedRDTSC();
-    std::this_thread::sleep_for(std::chrono::milliseconds{1});
-    FencedRDTSC();
-
-    // Get the current time.
-    const auto start_time = Common::RealTimeClock::Now();
-    const u64 tsc_start = FencedRDTSC();
-    // Wait for 250 milliseconds.
-    std::this_thread::sleep_for(std::chrono::milliseconds{250});
-    const auto end_time = Common::RealTimeClock::Now();
-    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());
-    const u64 tsc_diff = tsc_end - tsc_start;
-    const u64 tsc_freq = MultiplyAndDivide64(tsc_diff, 1000000000ULL, timer_diff);
-    return RoundToNearest<1000>(tsc_freq);
+std::chrono::milliseconds NativeClock::GetTimeMS() const {
+    return std::chrono::milliseconds{MultiplyHigh(GetHostTicksElapsed(), ms_rdtsc_factor)};
 }
 
-namespace X64 {
-NativeClock::NativeClock(u64 emulated_cpu_frequency_, u64 emulated_clock_frequency_,
-                         u64 rtsc_frequency_)
-    : WallClock(emulated_cpu_frequency_, emulated_clock_frequency_, true), rtsc_frequency{
-                                                                               rtsc_frequency_} {
-    // Thread to re-adjust the RDTSC frequency after 10 seconds has elapsed.
-    time_sync_thread = std::jthread{[this](std::stop_token token) {
-        // Get the current time.
-        const auto start_time = Common::RealTimeClock::Now();
-        const u64 tsc_start = FencedRDTSC();
-        // Wait for 10 seconds.
-        if (!Common::StoppableTimedWait(token, std::chrono::seconds{10})) {
-            return;
-        }
-        const auto end_time = Common::RealTimeClock::Now();
-        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());
-        const u64 tsc_diff = tsc_end - tsc_start;
-        const u64 tsc_freq = MultiplyAndDivide64(tsc_diff, 1000000000ULL, timer_diff);
-        rtsc_frequency = tsc_freq;
-        CalculateAndSetFactors();
-    }};
-
-    time_point.inner.last_measure = FencedRDTSC();
-    time_point.inner.accumulated_ticks = 0U;
-    CalculateAndSetFactors();
+u64 NativeClock::GetCNTPCT() const {
+    return MultiplyHigh(GetHostTicksElapsed(), cntpct_rdtsc_factor);
 }
 
-u64 NativeClock::GetRTSC() {
-    TimePoint new_time_point{};
-    TimePoint current_time_point{};
-
-    current_time_point.pack = Common::AtomicLoad128(time_point.pack.data());
-    do {
-        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
-                                                ? 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, current_time_point.pack));
-    return new_time_point.inner.accumulated_ticks;
+u64 NativeClock::GetHostTicksNow() const {
+    return FencedRDTSC();
 }
 
-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 {
-            new_time_point.pack = current_time_point.pack;
-            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));
-    }
-}
-
-std::chrono::nanoseconds NativeClock::GetTimeNS() {
-    const u64 rtsc_value = GetRTSC();
-    return std::chrono::nanoseconds{MultiplyHigh(rtsc_value, ns_rtsc_factor)};
+u64 NativeClock::GetHostTicksElapsed() const {
+    return FencedRDTSC() - start_ticks;
 }
 
-std::chrono::microseconds NativeClock::GetTimeUS() {
-    const u64 rtsc_value = GetRTSC();
-    return std::chrono::microseconds{MultiplyHigh(rtsc_value, us_rtsc_factor)};
+bool NativeClock::IsNative() const {
+    return true;
 }
 
-std::chrono::milliseconds NativeClock::GetTimeMS() {
-    const u64 rtsc_value = GetRTSC();
-    return std::chrono::milliseconds{MultiplyHigh(rtsc_value, ms_rtsc_factor)};
-}
-
-u64 NativeClock::GetClockCycles() {
-    const u64 rtsc_value = GetRTSC();
-    return MultiplyHigh(rtsc_value, clock_rtsc_factor);
-}
-
-u64 NativeClock::GetCPUCycles() {
-    const u64 rtsc_value = GetRTSC();
-    return MultiplyHigh(rtsc_value, cpu_rtsc_factor);
-}
-
-void NativeClock::CalculateAndSetFactors() {
-    ns_rtsc_factor = GetFixedPoint64Factor(NS_RATIO, rtsc_frequency);
-    us_rtsc_factor = GetFixedPoint64Factor(US_RATIO, rtsc_frequency);
-    ms_rtsc_factor = GetFixedPoint64Factor(MS_RATIO, rtsc_frequency);
-    clock_rtsc_factor = GetFixedPoint64Factor(emulated_clock_frequency, rtsc_frequency);
-    cpu_rtsc_factor = GetFixedPoint64Factor(emulated_cpu_frequency, rtsc_frequency);
-}
-
-} // namespace X64
-
-} // namespace Common
+} // namespace Common::X64
diff --git a/src/common/x64/native_clock.h b/src/common/x64/native_clock.h
index 03ca291d8..d6f8626c1 100644
--- a/src/common/x64/native_clock.h
+++ b/src/common/x64/native_clock.h
@@ -3,58 +3,36 @@
 
 #pragma once
 
-#include "common/polyfill_thread.h"
 #include "common/wall_clock.h"
 
-namespace Common {
+namespace Common::X64 {
 
-namespace X64 {
 class NativeClock final : public WallClock {
 public:
-    explicit NativeClock(u64 emulated_cpu_frequency_, u64 emulated_clock_frequency_,
-                         u64 rtsc_frequency_);
+    explicit NativeClock(u64 rdtsc_frequency_);
 
-    std::chrono::nanoseconds GetTimeNS() override;
+    std::chrono::nanoseconds GetTimeNS() const override;
 
-    std::chrono::microseconds GetTimeUS() override;
+    std::chrono::microseconds GetTimeUS() const override;
 
-    std::chrono::milliseconds GetTimeMS() override;
+    std::chrono::milliseconds GetTimeMS() const override;
 
-    u64 GetClockCycles() override;
+    u64 GetCNTPCT() const override;
 
-    u64 GetCPUCycles() override;
+    u64 GetHostTicksNow() const override;
 
-    void Pause(bool is_paused) override;
+    u64 GetHostTicksElapsed() const override;
 
-private:
-    u64 GetRTSC();
-
-    void CalculateAndSetFactors();
-
-    union alignas(16) TimePoint {
-        TimePoint() : pack{} {}
-        u128 pack{};
-        struct Inner {
-            u64 last_measure{};
-            u64 accumulated_ticks{};
-        } inner;
-    };
-
-    TimePoint time_point;
+    bool IsNative() const override;
 
-    // factors
-    u64 clock_rtsc_factor{};
-    u64 cpu_rtsc_factor{};
-    u64 ns_rtsc_factor{};
-    u64 us_rtsc_factor{};
-    u64 ms_rtsc_factor{};
-
-    u64 rtsc_frequency;
+private:
+    u64 start_ticks;
+    u64 rdtsc_frequency;
 
-    std::jthread time_sync_thread;
+    u64 ns_rdtsc_factor;
+    u64 us_rdtsc_factor;
+    u64 ms_rdtsc_factor;
+    u64 cntpct_rdtsc_factor;
 };
-} // namespace X64
-
-u64 EstimateRDTSCFrequency();
 
-} // namespace Common
+} // namespace Common::X64
-- 
cgit v1.2.3


From 907507886d755fa56099713c4b8f05bb640a8b7d Mon Sep 17 00:00:00 2001
From: Morph <39850852+Morph1984@users.noreply.github.com>
Date: Sun, 28 May 2023 17:45:47 -0400
Subject: (wall, native)_clock: Add GetGPUTick

Allows us to directly calculate the GPU tick without double conversion to and from the host clock tick.
---
 src/common/wall_clock.cpp       |  8 ++++++--
 src/common/wall_clock.h         | 20 +++++++++++++++++++-
 src/common/x64/native_clock.cpp |  7 ++++++-
 src/common/x64/native_clock.h   |  3 +++
 src/core/core_timing.cpp        |  7 +++++++
 src/core/core_timing.h          |  3 +++
 src/video_core/gpu.cpp          | 11 +++--------
 7 files changed, 47 insertions(+), 12 deletions(-)

(limited to 'src/common/x64')

diff --git a/src/common/wall_clock.cpp b/src/common/wall_clock.cpp
index ad8db06b0..dc0dcbd68 100644
--- a/src/common/wall_clock.cpp
+++ b/src/common/wall_clock.cpp
@@ -32,6 +32,10 @@ public:
         return GetHostTicksElapsed() * NsToCNTPCTRatio::num / NsToCNTPCTRatio::den;
     }
 
+    u64 GetGPUTick() const override {
+        return GetHostTicksElapsed() * NsToGPUTickRatio::num / NsToGPUTickRatio::den;
+    }
+
     u64 GetHostTicksNow() const override {
         return static_cast<u64>(SteadyClock::Now().time_since_epoch().count());
     }
@@ -52,12 +56,12 @@ std::unique_ptr<WallClock> CreateOptimalClock() {
 #ifdef ARCHITECTURE_x86_64
     const auto& caps = GetCPUCaps();
 
-    if (caps.invariant_tsc && caps.tsc_frequency >= WallClock::CNTFRQ) {
+    if (caps.invariant_tsc && caps.tsc_frequency >= WallClock::GPUTickFreq) {
         return std::make_unique<X64::NativeClock>(caps.tsc_frequency);
     } else {
         // Fallback to StandardWallClock if the hardware TSC
         // - Is not invariant
-        // - Is not more precise than CNTFRQ
+        // - Is not more precise than GPUTickFreq
         return std::make_unique<StandardWallClock>();
     }
 #else
diff --git a/src/common/wall_clock.h b/src/common/wall_clock.h
index 56c18ca25..fcfdd637c 100644
--- a/src/common/wall_clock.h
+++ b/src/common/wall_clock.h
@@ -13,7 +13,8 @@ namespace Common {
 
 class WallClock {
 public:
-    static constexpr u64 CNTFRQ = 19'200'000; // CNTPCT_EL0 Frequency = 19.2 MHz
+    static constexpr u64 CNTFRQ = 19'200'000;       // CNTPCT_EL0 Frequency = 19.2 MHz
+    static constexpr u64 GPUTickFreq = 614'400'000; // GM20B GPU Tick Frequency = 614.4 MHz
 
     virtual ~WallClock() = default;
 
@@ -29,6 +30,9 @@ public:
     /// @returns The guest CNTPCT ticks since the construction of this clock.
     virtual u64 GetCNTPCT() const = 0;
 
+    /// @returns The guest GPU ticks since the construction of this clock.
+    virtual u64 GetGPUTick() const = 0;
+
     /// @returns The raw host timer ticks since an indeterminate epoch.
     virtual u64 GetHostTicksNow() const = 0;
 
@@ -46,6 +50,10 @@ public:
         return us * UsToCNTPCTRatio::num / UsToCNTPCTRatio::den;
     }
 
+    static inline u64 NSToGPUTick(u64 ns) {
+        return ns * NsToGPUTickRatio::num / NsToGPUTickRatio::den;
+    }
+
     static inline u64 CNTPCTToNS(u64 cntpct) {
         return cntpct * NsToCNTPCTRatio::den / NsToCNTPCTRatio::num;
     }
@@ -54,6 +62,14 @@ public:
         return cntpct * UsToCNTPCTRatio::den / UsToCNTPCTRatio::num;
     }
 
+    static inline u64 GPUTickToNS(u64 gpu_tick) {
+        return gpu_tick * NsToGPUTickRatio::den / NsToGPUTickRatio::num;
+    }
+
+    static inline u64 CNTPCTToGPUTick(u64 cntpct) {
+        return cntpct * CNTPCTToGPUTickRatio::num / CNTPCTToGPUTickRatio::den;
+    }
+
 protected:
     using NsRatio = std::nano;
     using UsRatio = std::micro;
@@ -63,6 +79,8 @@ protected:
     using NsToMsRatio = std::ratio_divide<std::nano, std::milli>;
     using NsToCNTPCTRatio = std::ratio<CNTFRQ, std::nano::den>;
     using UsToCNTPCTRatio = std::ratio<CNTFRQ, std::micro::den>;
+    using NsToGPUTickRatio = std::ratio<GPUTickFreq, std::nano::den>;
+    using CNTPCTToGPUTickRatio = std::ratio<GPUTickFreq, CNTFRQ>;
 };
 
 std::unique_ptr<WallClock> CreateOptimalClock();
diff --git a/src/common/x64/native_clock.cpp b/src/common/x64/native_clock.cpp
index 5d1eb0590..7d2a26bd9 100644
--- a/src/common/x64/native_clock.cpp
+++ b/src/common/x64/native_clock.cpp
@@ -12,7 +12,8 @@ NativeClock::NativeClock(u64 rdtsc_frequency_)
       ns_rdtsc_factor{GetFixedPoint64Factor(NsRatio::den, rdtsc_frequency)},
       us_rdtsc_factor{GetFixedPoint64Factor(UsRatio::den, rdtsc_frequency)},
       ms_rdtsc_factor{GetFixedPoint64Factor(MsRatio::den, rdtsc_frequency)},
-      cntpct_rdtsc_factor{GetFixedPoint64Factor(CNTFRQ, rdtsc_frequency)} {}
+      cntpct_rdtsc_factor{GetFixedPoint64Factor(CNTFRQ, rdtsc_frequency)},
+      gputick_rdtsc_factor{GetFixedPoint64Factor(GPUTickFreq, rdtsc_frequency)} {}
 
 std::chrono::nanoseconds NativeClock::GetTimeNS() const {
     return std::chrono::nanoseconds{MultiplyHigh(GetHostTicksElapsed(), ns_rdtsc_factor)};
@@ -30,6 +31,10 @@ u64 NativeClock::GetCNTPCT() const {
     return MultiplyHigh(GetHostTicksElapsed(), cntpct_rdtsc_factor);
 }
 
+u64 NativeClock::GetGPUTick() const {
+    return MultiplyHigh(GetHostTicksElapsed(), gputick_rdtsc_factor);
+}
+
 u64 NativeClock::GetHostTicksNow() const {
     return FencedRDTSC();
 }
diff --git a/src/common/x64/native_clock.h b/src/common/x64/native_clock.h
index d6f8626c1..334415eff 100644
--- a/src/common/x64/native_clock.h
+++ b/src/common/x64/native_clock.h
@@ -19,6 +19,8 @@ public:
 
     u64 GetCNTPCT() const override;
 
+    u64 GetGPUTick() const override;
+
     u64 GetHostTicksNow() const override;
 
     u64 GetHostTicksElapsed() const override;
@@ -33,6 +35,7 @@ private:
     u64 us_rdtsc_factor;
     u64 ms_rdtsc_factor;
     u64 cntpct_rdtsc_factor;
+    u64 gputick_rdtsc_factor;
 };
 
 } // namespace Common::X64
diff --git a/src/core/core_timing.cpp b/src/core/core_timing.cpp
index 9a1d5a69a..e57bca70a 100644
--- a/src/core/core_timing.cpp
+++ b/src/core/core_timing.cpp
@@ -197,6 +197,13 @@ u64 CoreTiming::GetClockTicks() const {
     return ticks;
 }
 
+u64 CoreTiming::GetGPUTicks() const {
+    if (is_multicore) [[likely]] {
+        return clock->GetGPUTick();
+    }
+    return Common::WallClock::CNTPCTToGPUTick(ticks);
+}
+
 std::optional<s64> CoreTiming::Advance() {
     std::scoped_lock lock{advance_lock, basic_lock};
     global_timer = GetGlobalTimeNs().count();
diff --git a/src/core/core_timing.h b/src/core/core_timing.h
index fdacdd94a..1873852c4 100644
--- a/src/core/core_timing.h
+++ b/src/core/core_timing.h
@@ -119,6 +119,9 @@ public:
     /// Returns the current CNTPCT tick value.
     u64 GetClockTicks() const;
 
+    /// Returns the current GPU tick value.
+    u64 GetGPUTicks() const;
+
     /// Returns current time in microseconds.
     std::chrono::microseconds GetGlobalTimeUs() const;
 
diff --git a/src/video_core/gpu.cpp b/src/video_core/gpu.cpp
index 70762c51a..db385076d 100644
--- a/src/video_core/gpu.cpp
+++ b/src/video_core/gpu.cpp
@@ -193,18 +193,13 @@ struct GPU::Impl {
     }
 
     [[nodiscard]] u64 GetTicks() const {
-        // This values were reversed engineered by fincs from NVN
-        // The GPU clock is 614.4 MHz
-        using NsToGPUTickRatio = std::ratio<614'400'000, std::nano::den>;
-        static_assert(NsToGPUTickRatio::num == 384 && NsToGPUTickRatio::den == 625);
-
-        u64 nanoseconds = system.CoreTiming().GetGlobalTimeNs().count();
+        u64 gpu_tick = system.CoreTiming().GetGPUTicks();
 
         if (Settings::values.use_fast_gpu_time.GetValue()) {
-            nanoseconds /= 256;
+            gpu_tick /= 256;
         }
 
-        return nanoseconds * NsToGPUTickRatio::num / NsToGPUTickRatio::den;
+        return gpu_tick;
     }
 
     [[nodiscard]] bool IsAsync() const {
-- 
cgit v1.2.3


From 4303ed614d0d758d9e9bcdef8afee3274769d2fb Mon Sep 17 00:00:00 2001
From: Morph <39850852+Morph1984@users.noreply.github.com>
Date: Wed, 28 Jun 2023 01:07:10 -0400
Subject: x64: Add detection of monitorx instructions

monitorx introduces 2 instructions: MONITORX and MWAITX.
---
 src/common/telemetry.cpp      | 1 +
 src/common/x64/cpu_detect.cpp | 1 +
 src/common/x64/cpu_detect.h   | 1 +
 3 files changed, 3 insertions(+)

(limited to 'src/common/x64')

diff --git a/src/common/telemetry.cpp b/src/common/telemetry.cpp
index 91352912d..929ed67e4 100644
--- a/src/common/telemetry.cpp
+++ b/src/common/telemetry.cpp
@@ -93,6 +93,7 @@ void AppendCPUInfo(FieldCollection& fc) {
     add_field("CPU_Extension_x64_GFNI", caps.gfni);
     add_field("CPU_Extension_x64_INVARIANT_TSC", caps.invariant_tsc);
     add_field("CPU_Extension_x64_LZCNT", caps.lzcnt);
+    add_field("CPU_Extension_x64_MONITORX", caps.monitorx);
     add_field("CPU_Extension_x64_MOVBE", caps.movbe);
     add_field("CPU_Extension_x64_PCLMULQDQ", caps.pclmulqdq);
     add_field("CPU_Extension_x64_POPCNT", caps.popcnt);
diff --git a/src/common/x64/cpu_detect.cpp b/src/common/x64/cpu_detect.cpp
index c998b1197..780120a5b 100644
--- a/src/common/x64/cpu_detect.cpp
+++ b/src/common/x64/cpu_detect.cpp
@@ -168,6 +168,7 @@ static CPUCaps Detect() {
         __cpuid(cpu_id, 0x80000001);
         caps.lzcnt = Common::Bit<5>(cpu_id[2]);
         caps.fma4 = Common::Bit<16>(cpu_id[2]);
+        caps.monitorx = Common::Bit<29>(cpu_id[2]);
     }
 
     if (max_ex_fn >= 0x80000007) {
diff --git a/src/common/x64/cpu_detect.h b/src/common/x64/cpu_detect.h
index 8253944d6..756459417 100644
--- a/src/common/x64/cpu_detect.h
+++ b/src/common/x64/cpu_detect.h
@@ -63,6 +63,7 @@ struct CPUCaps {
     bool gfni : 1;
     bool invariant_tsc : 1;
     bool lzcnt : 1;
+    bool monitorx : 1;
     bool movbe : 1;
     bool pclmulqdq : 1;
     bool popcnt : 1;
-- 
cgit v1.2.3


From 3d868baaa44152e7a4bd8c64905443fd9a08adce Mon Sep 17 00:00:00 2001
From: Morph <39850852+Morph1984@users.noreply.github.com>
Date: Wed, 28 Jun 2023 01:07:59 -0400
Subject: x64: cpu_wait: Make use of MWAITX in MicroSleep

MWAITX is equivalent to UMWAIT on Intel's Alder Lake CPUs.
We can emulate TPAUSE by using MONITORX in conjunction with MWAITX to wait for 100K cycles.
---
 src/common/x64/cpu_wait.cpp | 33 +++++++++++++++++++++------------
 1 file changed, 21 insertions(+), 12 deletions(-)

(limited to 'src/common/x64')

diff --git a/src/common/x64/cpu_wait.cpp b/src/common/x64/cpu_wait.cpp
index c53dd4945..11b9c4d83 100644
--- a/src/common/x64/cpu_wait.cpp
+++ b/src/common/x64/cpu_wait.cpp
@@ -13,24 +13,30 @@
 
 namespace Common::X64 {
 
+namespace {
+
+// 100,000 cycles is a reasonable amount of time to wait to save on CPU resources.
+// For reference:
+// At 1 GHz, 100K cycles is 100us
+// At 2 GHz, 100K cycles is 50us
+// At 4 GHz, 100K cycles is 25us
+constexpr auto PauseCycles = 100'000U;
+
+} // Anonymous namespace
+
 #ifdef _MSC_VER
 __forceinline static void TPAUSE() {
-    // 100,000 cycles is a reasonable amount of time to wait to save on CPU resources.
-    // For reference:
-    // At 1 GHz, 100K cycles is 100us
-    // At 2 GHz, 100K cycles is 50us
-    // At 4 GHz, 100K cycles is 25us
-    static constexpr auto PauseCycles = 100'000;
     _tpause(0, FencedRDTSC() + PauseCycles);
 }
+
+__forceinline static void MWAITX() {
+    // monitor_var should be aligned to a cache line.
+    alignas(64) u64 monitor_var{};
+    _mm_monitorx(&monitor_var, 0, 0);
+    _mm_mwaitx(/* extensions*/ 2, /* hints */ 0, /* cycles */ PauseCycles);
+}
 #else
 static void TPAUSE() {
-    // 100,000 cycles is a reasonable amount of time to wait to save on CPU resources.
-    // For reference:
-    // At 1 GHz, 100K cycles is 100us
-    // At 2 GHz, 100K cycles is 50us
-    // At 4 GHz, 100K cycles is 25us
-    static constexpr auto PauseCycles = 100'000;
     const auto tsc = FencedRDTSC() + PauseCycles;
     const auto eax = static_cast<u32>(tsc & 0xFFFFFFFF);
     const auto edx = static_cast<u32>(tsc >> 32);
@@ -40,9 +46,12 @@ static void TPAUSE() {
 
 void MicroSleep() {
     static const bool has_waitpkg = GetCPUCaps().waitpkg;
+    static const bool has_monitorx = GetCPUCaps().monitorx;
 
     if (has_waitpkg) {
         TPAUSE();
+    } else if (has_monitorx) {
+        MWAITX();
     } else {
         std::this_thread::yield();
     }
-- 
cgit v1.2.3


From 2b68a3cbbf144b97aa524eb1dd17aad34cdf1a67 Mon Sep 17 00:00:00 2001
From: Morph <39850852+Morph1984@users.noreply.github.com>
Date: Wed, 28 Jun 2023 01:19:41 -0400
Subject: x64: cpu_wait: Remove magic values

---
 src/common/x64/cpu_wait.cpp | 11 ++++++++---
 1 file changed, 8 insertions(+), 3 deletions(-)

(limited to 'src/common/x64')

diff --git a/src/common/x64/cpu_wait.cpp b/src/common/x64/cpu_wait.cpp
index 11b9c4d83..ea16c8490 100644
--- a/src/common/x64/cpu_wait.cpp
+++ b/src/common/x64/cpu_wait.cpp
@@ -26,21 +26,26 @@ constexpr auto PauseCycles = 100'000U;
 
 #ifdef _MSC_VER
 __forceinline static void TPAUSE() {
-    _tpause(0, FencedRDTSC() + PauseCycles);
+    static constexpr auto RequestC02State = 0U;
+    _tpause(RequestC02State, FencedRDTSC() + PauseCycles);
 }
 
 __forceinline static void MWAITX() {
+    static constexpr auto EnableWaitTimeFlag = 1U << 1;
+    static constexpr auto RequestC1State = 0U;
+
     // monitor_var should be aligned to a cache line.
     alignas(64) u64 monitor_var{};
     _mm_monitorx(&monitor_var, 0, 0);
-    _mm_mwaitx(/* extensions*/ 2, /* hints */ 0, /* cycles */ PauseCycles);
+    _mm_mwaitx(EnableWaitTimeFlag, RequestC1State, PauseCycles);
 }
 #else
 static void TPAUSE() {
+    static constexpr auto RequestC02State = 0U;
     const auto tsc = FencedRDTSC() + PauseCycles;
     const auto eax = static_cast<u32>(tsc & 0xFFFFFFFF);
     const auto edx = static_cast<u32>(tsc >> 32);
-    asm volatile("tpause %0" : : "r"(0), "d"(edx), "a"(eax));
+    asm volatile("tpause %0" : : "r"(RequestC02State), "d"(edx), "a"(eax));
 }
 #endif
 
-- 
cgit v1.2.3


From 295fc7d0f8f0b6158307c5c9b11a60516f9eb221 Mon Sep 17 00:00:00 2001
From: Morph <39850852+Morph1984@users.noreply.github.com>
Date: Wed, 28 Jun 2023 01:35:16 -0400
Subject: x64: cpu_wait: Implement MWAITX for non-MSVC compilers

---
 src/common/x64/cpu_wait.cpp | 10 ++++++++++
 1 file changed, 10 insertions(+)

(limited to 'src/common/x64')

diff --git a/src/common/x64/cpu_wait.cpp b/src/common/x64/cpu_wait.cpp
index ea16c8490..41d385f59 100644
--- a/src/common/x64/cpu_wait.cpp
+++ b/src/common/x64/cpu_wait.cpp
@@ -47,6 +47,16 @@ static void TPAUSE() {
     const auto edx = static_cast<u32>(tsc >> 32);
     asm volatile("tpause %0" : : "r"(RequestC02State), "d"(edx), "a"(eax));
 }
+
+static void MWAITX() {
+    static constexpr auto EnableWaitTimeFlag = 1U << 1;
+    static constexpr auto RequestC1State = 0U;
+
+    // monitor_var should be aligned to a cache line.
+    alignas(64) u64 monitor_var{};
+    asm volatile("monitorx" : : "a"(&monitor_var), "c"(0), "d"(0));
+    asm volatile("mwaitx" : : "a"(RequestC1State), "b"(PauseCycles), "c"(EnableWaitTimeFlag));
+}
 #endif
 
 void MicroSleep() {
-- 
cgit v1.2.3