Merge pull request #3396 from FernandoS27/prometheus-1

Implement SpinLocks, Fibers and a Host Timer

src/core/CMakeLists.txt

@@ -547,6 +547,8 @@ add_library(core STATIC
     hle/service/vi/vi_u.h
     hle/service/wlan/wlan.cpp
     hle/service/wlan/wlan.h
+    host_timing.cpp
+    host_timing.h
     loader/deconstructed_rom_directory.cpp
     loader/deconstructed_rom_directory.h
     loader/elf.cpp

src/core/core_timing_util.cpp

@@ -49,6 +49,21 @@ s64 nsToCycles(std::chrono::nanoseconds ns) {
     return (Hardware::BASE_CLOCK_RATE * ns.count()) / 1000000000;
 }
 
+u64 msToClockCycles(std::chrono::milliseconds ns) {
+    const u128 temp = Common::Multiply64Into128(ns.count(), Hardware::CNTFREQ);
+    return Common::Divide128On32(temp, 1000).first;
+}
+
+u64 usToClockCycles(std::chrono::microseconds ns) {
+    const u128 temp = Common::Multiply64Into128(ns.count(), Hardware::CNTFREQ);
+    return Common::Divide128On32(temp, 1000000).first;
+}
+
+u64 nsToClockCycles(std::chrono::nanoseconds ns) {
+    const u128 temp = Common::Multiply64Into128(ns.count(), Hardware::CNTFREQ);
+    return Common::Divide128On32(temp, 1000000000).first;
+}
+
 u64 CpuCyclesToClockCycles(u64 ticks) {
     const u128 temporal = Common::Multiply64Into128(ticks, Hardware::CNTFREQ);
     return Common::Divide128On32(temporal, static_cast<u32>(Hardware::BASE_CLOCK_RATE)).first;

src/core/core_timing_util.h

@@ -13,6 +13,9 @@ namespace Core::Timing {
 s64 msToCycles(std::chrono::milliseconds ms);
 s64 usToCycles(std::chrono::microseconds us);
 s64 nsToCycles(std::chrono::nanoseconds ns);
+u64 msToClockCycles(std::chrono::milliseconds ns);
+u64 usToClockCycles(std::chrono::microseconds ns);
+u64 nsToClockCycles(std::chrono::nanoseconds ns);
 
 inline std::chrono::milliseconds CyclesToMs(s64 cycles) {
     return std::chrono::milliseconds(cycles * 1000 / Hardware::BASE_CLOCK_RATE);

src/core/host_timing.cpp

@@ -0,0 +1,206 @@
+// Copyright 2020 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include "core/host_timing.h"
+
+#include <algorithm>
+#include <mutex>
+#include <string>
+#include <tuple>
+
+#include "common/assert.h"
+#include "core/core_timing_util.h"
+
+namespace Core::HostTiming {
+
+std::shared_ptr<EventType> CreateEvent(std::string name, TimedCallback&& callback) {
+    return std::make_shared<EventType>(std::move(callback), std::move(name));
+}
+
+struct CoreTiming::Event {
+    u64 time;
+    u64 fifo_order;
+    u64 userdata;
+    std::weak_ptr<EventType> type;
+
+    // Sort by time, unless the times are the same, in which case sort by
+    // the order added to the queue
+    friend bool operator>(const Event& left, const Event& right) {
+        return std::tie(left.time, left.fifo_order) > std::tie(right.time, right.fifo_order);
+    }
+
+    friend bool operator<(const Event& left, const Event& right) {
+        return std::tie(left.time, left.fifo_order) < std::tie(right.time, right.fifo_order);
+    }
+};
+
+CoreTiming::CoreTiming() {
+    clock =
+        Common::CreateBestMatchingClock(Core::Hardware::BASE_CLOCK_RATE, Core::Hardware::CNTFREQ);
+}
+
+CoreTiming::~CoreTiming() = default;
+
+void CoreTiming::ThreadEntry(CoreTiming& instance) {
+    instance.ThreadLoop();
+}
+
+void CoreTiming::Initialize() {
+    event_fifo_id = 0;
+    const auto empty_timed_callback = [](u64, s64) {};
+    ev_lost = CreateEvent("_lost_event", empty_timed_callback);
+    timer_thread = std::make_unique<std::thread>(ThreadEntry, std::ref(*this));
+}
+
+void CoreTiming::Shutdown() {
+    paused = true;
+    shutting_down = true;
+    event.Set();
+    timer_thread->join();
+    ClearPendingEvents();
+    timer_thread.reset();
+    has_started = false;
+}
+
+void CoreTiming::Pause(bool is_paused) {
+    paused = is_paused;
+}
+
+void CoreTiming::SyncPause(bool is_paused) {
+    if (is_paused == paused && paused_set == paused) {
+        return;
+    }
+    Pause(is_paused);
+    event.Set();
+    while (paused_set != is_paused)
+        ;
+}
+
+bool CoreTiming::IsRunning() const {
+    return !paused_set;
+}
+
+bool CoreTiming::HasPendingEvents() const {
+    return !(wait_set && event_queue.empty());
+}
+
+void CoreTiming::ScheduleEvent(s64 ns_into_future, const std::shared_ptr<EventType>& event_type,
+                               u64 userdata) {
+    basic_lock.lock();
+    const u64 timeout = static_cast<u64>(GetGlobalTimeNs().count() + ns_into_future);
+
+    event_queue.emplace_back(Event{timeout, event_fifo_id++, userdata, event_type});
+
+    std::push_heap(event_queue.begin(), event_queue.end(), std::greater<>());
+    basic_lock.unlock();
+    event.Set();
+}
+
+void CoreTiming::UnscheduleEvent(const std::shared_ptr<EventType>& event_type, u64 userdata) {
+    basic_lock.lock();
+    const auto itr = std::remove_if(event_queue.begin(), event_queue.end(), [&](const Event& e) {
+        return e.type.lock().get() == event_type.get() && e.userdata == userdata;
+    });
+
+    // Removing random items breaks the invariant so we have to re-establish it.
+    if (itr != event_queue.end()) {
+        event_queue.erase(itr, event_queue.end());
+        std::make_heap(event_queue.begin(), event_queue.end(), std::greater<>());
+    }
+    basic_lock.unlock();
+}
+
+void CoreTiming::AddTicks(std::size_t core_index, u64 ticks) {
+    ticks_count[core_index] += ticks;
+}
+
+void CoreTiming::ResetTicks(std::size_t core_index) {
+    ticks_count[core_index] = 0;
+}
+
+u64 CoreTiming::GetCPUTicks() const {
+    return clock->GetCPUCycles();
+}
+
+u64 CoreTiming::GetClockTicks() const {
+    return clock->GetClockCycles();
+}
+
+void CoreTiming::ClearPendingEvents() {
+    event_queue.clear();
+}
+
+void CoreTiming::RemoveEvent(const std::shared_ptr<EventType>& event_type) {
+    basic_lock.lock();
+
+    const auto itr = std::remove_if(event_queue.begin(), event_queue.end(), [&](const Event& e) {
+        return e.type.lock().get() == event_type.get();
+    });
+
+    // Removing random items breaks the invariant so we have to re-establish it.
+    if (itr != event_queue.end()) {
+        event_queue.erase(itr, event_queue.end());
+        std::make_heap(event_queue.begin(), event_queue.end(), std::greater<>());
+    }
+    basic_lock.unlock();
+}
+
+std::optional<u64> CoreTiming::Advance() {
+    advance_lock.lock();
+    basic_lock.lock();
+    global_timer = GetGlobalTimeNs().count();
+
+    while (!event_queue.empty() && event_queue.front().time <= global_timer) {
+        Event evt = std::move(event_queue.front());
+        std::pop_heap(event_queue.begin(), event_queue.end(), std::greater<>());
+        event_queue.pop_back();
+        basic_lock.unlock();
+
+        if (auto event_type{evt.type.lock()}) {
+            event_type->callback(evt.userdata, global_timer - evt.time);
+        }
+
+        basic_lock.lock();
+    }
+
+    if (!event_queue.empty()) {
+        const u64 next_time = event_queue.front().time - global_timer;
+        basic_lock.unlock();
+        advance_lock.unlock();
+        return next_time;
+    } else {
+        basic_lock.unlock();
+        advance_lock.unlock();
+        return std::nullopt;
+    }
+}
+
+void CoreTiming::ThreadLoop() {
+    has_started = true;
+    while (!shutting_down) {
+        while (!paused) {
+            paused_set = false;
+            const auto next_time = Advance();
+            if (next_time) {
+                std::chrono::nanoseconds next_time_ns = std::chrono::nanoseconds(*next_time);
+                event.WaitFor(next_time_ns);
+            } else {
+                wait_set = true;
+                event.Wait();
+            }
+            wait_set = false;
+        }
+        paused_set = true;
+    }
+}
+
+std::chrono::nanoseconds CoreTiming::GetGlobalTimeNs() const {
+    return clock->GetTimeNS();
+}
+
+std::chrono::microseconds CoreTiming::GetGlobalTimeUs() const {
+    return clock->GetTimeUS();
+}
+
+} // namespace Core::HostTiming

src/core/host_timing.h

@@ -0,0 +1,160 @@
+// Copyright 2020 yuzu Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <atomic>
+#include <chrono>
+#include <functional>
+#include <memory>
+#include <mutex>
+#include <optional>
+#include <string>
+#include <thread>
+#include <vector>
+
+#include "common/common_types.h"
+#include "common/spin_lock.h"
+#include "common/thread.h"
+#include "common/threadsafe_queue.h"
+#include "common/wall_clock.h"
+#include "core/hardware_properties.h"
+
+namespace Core::HostTiming {
+
+/// A callback that may be scheduled for a particular core timing event.
+using TimedCallback = std::function<void(u64 userdata, s64 cycles_late)>;
+
+/// Contains the characteristics of a particular event.
+struct EventType {
+    EventType(TimedCallback&& callback, std::string&& name)
+        : callback{std::move(callback)}, name{std::move(name)} {}
+
+    /// The event's callback function.
+    TimedCallback callback;
+    /// A pointer to the name of the event.
+    const std::string name;
+};
+
+/**
+ * This is a system to schedule events into the emulated machine's future. Time is measured
+ * in main CPU clock cycles.
+ *
+ * To schedule an event, you first have to register its type. This is where you pass in the
+ * callback. You then schedule events using the type id you get back.
+ *
+ * The int cyclesLate that the callbacks get is how many cycles late it was.
+ * So to schedule a new event on a regular basis:
+ * inside callback:
+ *   ScheduleEvent(periodInCycles - cyclesLate, callback, "whatever")
+ */
+class CoreTiming {
+public:
+    CoreTiming();
+    ~CoreTiming();
+
+    CoreTiming(const CoreTiming&) = delete;
+    CoreTiming(CoreTiming&&) = delete;
+
+    CoreTiming& operator=(const CoreTiming&) = delete;
+    CoreTiming& operator=(CoreTiming&&) = delete;
+
+    /// CoreTiming begins at the boundary of timing slice -1. An initial call to Advance() is
+    /// required to end slice - 1 and start slice 0 before the first cycle of code is executed.
+    void Initialize();
+
+    /// Tears down all timing related functionality.
+    void Shutdown();
+
+    /// Pauses/Unpauses the execution of the timer thread.
+    void Pause(bool is_paused);
+
+    /// Pauses/Unpauses the execution of the timer thread and waits until paused.
+    void SyncPause(bool is_paused);
+
+    /// Checks if core timing is running.
+    bool IsRunning() const;
+
+    /// Checks if the timer thread has started.
+    bool HasStarted() const {
+        return has_started;
+    }
+
+    /// Checks if there are any pending time events.
+    bool HasPendingEvents() const;
+
+    /// Schedules an event in core timing
+    void ScheduleEvent(s64 ns_into_future, const std::shared_ptr<EventType>& event_type,
+                       u64 userdata = 0);
+
+    void UnscheduleEvent(const std::shared_ptr<EventType>& event_type, u64 userdata);
+
+    /// We only permit one event of each type in the queue at a time.
+    void RemoveEvent(const std::shared_ptr<EventType>& event_type);
+
+    void AddTicks(std::size_t core_index, u64 ticks);
+
+    void ResetTicks(std::size_t core_index);
+
+    /// Returns current time in emulated CPU cycles
+    u64 GetCPUTicks() const;
+
+    /// Returns current time in emulated in Clock cycles
+    u64 GetClockTicks() const;
+
+    /// Returns current time in microseconds.
+    std::chrono::microseconds GetGlobalTimeUs() const;
+
+    /// Returns current time in nanoseconds.
+    std::chrono::nanoseconds GetGlobalTimeNs() const;
+
+    /// Checks for events manually and returns time in nanoseconds for next event, threadsafe.
+    std::optional<u64> Advance();
+
+private:
+    struct Event;
+
+    /// Clear all pending events. This should ONLY be done on exit.
+    void ClearPendingEvents();
+
+    static void ThreadEntry(CoreTiming& instance);
+    void ThreadLoop();
+
+    std::unique_ptr<Common::WallClock> clock;
+
+    u64 global_timer = 0;
+
+    std::chrono::nanoseconds start_point;
+
+    // The queue is a min-heap using std::make_heap/push_heap/pop_heap.
+    // We don't use std::priority_queue because we need to be able to serialize, unserialize and
+    // erase arbitrary events (RemoveEvent()) regardless of the queue order. These aren't
+    // accomodated by the standard adaptor class.
+    std::vector<Event> event_queue;
+    u64 event_fifo_id = 0;
+
+    std::shared_ptr<EventType> ev_lost;
+    Common::Event event{};
+    Common::SpinLock basic_lock{};
+    Common::SpinLock advance_lock{};
+    std::unique_ptr<std::thread> timer_thread;
+    std::atomic<bool> paused{};
+    std::atomic<bool> paused_set{};
+    std::atomic<bool> wait_set{};
+    std::atomic<bool> shutting_down{};
+    std::atomic<bool> has_started{};
+
+    std::array<std::atomic<u64>, Core::Hardware::NUM_CPU_CORES> ticks_count{};
+};
+
+/// Creates a core timing event with the given name and callback.
+///
+/// @param name     The name of the core timing event to create.
+/// @param callback The callback to execute for the event.
+///
+/// @returns An EventType instance representing the created event.
+///
+std::shared_ptr<EventType> CreateEvent(std::string name, TimedCallback&& callback);
+
+} // namespace Core::HostTiming