early-access version 2805

README.md

@@ -1,7 +1,7 @@
 yuzu emulator early access
 =============
 
-This is the source code for early-access 2804.
+This is the source code for early-access 2805.
 
 ## Legal Notice
 

src/common/thread.cpp

@@ -47,6 +47,9 @@ void SetCurrentThreadPriority(ThreadPriority new_priority) {
     case ThreadPriority::VeryHigh:
         windows_priority = THREAD_PRIORITY_HIGHEST;
         break;
+    case ThreadPriority::Critical:
+        windows_priority = THREAD_PRIORITY_TIME_CRITICAL;
+        break;
     default:
         windows_priority = THREAD_PRIORITY_NORMAL;
         break;
@@ -59,9 +62,10 @@ void SetCurrentThreadPriority(ThreadPriority new_priority) {
 void SetCurrentThreadPriority(ThreadPriority new_priority) {
     pthread_t this_thread = pthread_self();
 
-    s32 max_prio = sched_get_priority_max(SCHED_OTHER);
+    const auto scheduling_type = SCHED_OTHER;
-    s32 min_prio = sched_get_priority_min(SCHED_OTHER);
+    s32 max_prio = sched_get_priority_max(scheduling_type);
-    u32 level = static_cast<u32>(new_priority) + 1;
+    s32 min_prio = sched_get_priority_min(scheduling_type);
+    u32 level = std::max(static_cast<u32>(new_priority) + 1, 4U);
 
     struct sched_param params;
     if (max_prio > min_prio) {
@@ -70,7 +74,7 @@ void SetCurrentThreadPriority(ThreadPriority new_priority) {
         params.sched_priority = min_prio - ((min_prio - max_prio) * level) / 4;
     }
 
-    pthread_setschedparam(this_thread, SCHED_OTHER, &params);
+    pthread_setschedparam(this_thread, scheduling_type, &params);
 }
 
 #endif

src/common/thread.h

@@ -92,6 +92,7 @@ enum class ThreadPriority : u32 {
     Normal = 1,
     High = 2,
     VeryHigh = 3,
+    Critical = 4,
 };
 
 void SetCurrentThreadPriority(ThreadPriority new_priority);

src/common/uint128.h

@@ -30,6 +30,10 @@ namespace Common {
 #else
     return _udiv128(r[1], r[0], d, &remainder);
 #endif
+#else
+#ifdef __SIZEOF_INT128__
+    const auto product = static_cast<unsigned __int128>(a) * static_cast<unsigned __int128>(b);
+    return static_cast<u64>(product / d);
 #else
     const u64 diva = a / d;
     const u64 moda = a % d;
@@ -37,6 +41,7 @@ namespace Common {
     const u64 modb = b % d;
     return diva * b + moda * divb + moda * modb / d;
 #endif
+#endif
 }
 
 // This function multiplies 2 u64 values and produces a u128 value;

src/common/x64/native_clock.cpp

@@ -65,8 +65,10 @@ 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_} {
-    time_point.inner.last_measure = FencedRDTSC();
+    TimePoint new_time_point{};
-    time_point.inner.accumulated_ticks = 0U;
+    new_time_point.last_measure = FencedRDTSC();
+    new_time_point.accumulated_ticks = 0U;
+    time_point.store(new_time_point);
     ns_rtsc_factor = GetFixedPoint64Factor(NS_RATIO, rtsc_frequency);
     us_rtsc_factor = GetFixedPoint64Factor(US_RATIO, rtsc_frequency);
     ms_rtsc_factor = GetFixedPoint64Factor(MS_RATIO, rtsc_frequency);
@@ -76,34 +78,31 @@ NativeClock::NativeClock(u64 emulated_cpu_frequency_, u64 emulated_clock_frequen
 
 u64 NativeClock::GetRTSC() {
     TimePoint new_time_point{};
-    TimePoint current_time_point{};
+    TimePoint current_time_point = time_point.load(std::memory_order_acquire);
-
-    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;
+        u64 diff = current_measure - current_time_point.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
+        new_time_point.last_measure = current_measure > current_time_point.last_measure
-                                                ? current_measure
+                                          ? current_measure
-                                                : current_time_point.inner.last_measure;
+                                          : current_time_point.last_measure;
-        new_time_point.inner.accumulated_ticks = current_time_point.inner.accumulated_ticks + diff;
+        new_time_point.accumulated_ticks = current_time_point.accumulated_ticks + diff;
-    } while (!Common::AtomicCompareAndSwap(time_point.pack.data(), new_time_point.pack,
+    } while (!time_point.compare_exchange_weak(
-                                           current_time_point.pack, current_time_point.pack));
+        current_time_point, new_time_point, std::memory_order_release, std::memory_order_acquire));
     /// The clock cannot be more precise than the guest timer, remove the lower bits
-    return new_time_point.inner.accumulated_ticks & inaccuracy_mask;
+    return new_time_point.accumulated_ticks;
 }
 
 void NativeClock::Pause(bool is_paused) {
     if (!is_paused) {
-        TimePoint current_time_point{};
         TimePoint new_time_point{};
-
+        TimePoint current_time_point = time_point.load(std::memory_order_acquire);
-        current_time_point.pack = Common::AtomicLoad128(time_point.pack.data());
         do {
-            new_time_point.pack = current_time_point.pack;
+            new_time_point = current_time_point;
-            new_time_point.inner.last_measure = FencedRDTSC();
+            new_time_point.last_measure = FencedRDTSC();
-        } while (!Common::AtomicCompareAndSwap(time_point.pack.data(), new_time_point.pack,
+        } while (!time_point.compare_exchange_weak(current_time_point, new_time_point,
-                                               current_time_point.pack, current_time_point.pack));
+                                                   std::memory_order_release,
+                                                   std::memory_order_acquire));
     }
 }
 

src/common/x64/native_clock.h

@@ -3,6 +3,7 @@
 
 #pragma once
 
+#include <atomic>
 #include "common/wall_clock.h"
 
 namespace Common {
@@ -28,21 +29,12 @@ public:
 private:
     u64 GetRTSC();
 
-    union alignas(16) TimePoint {
+    struct alignas(16) TimePoint {
-        TimePoint() : pack{} {}
+        u64 last_measure{};
-        u128 pack{};
+        u64 accumulated_ticks{};
-        struct Inner {
-            u64 last_measure{};
-            u64 accumulated_ticks{};
-        } inner;
     };
 
-    /// value used to reduce the native clocks accuracy as some apss rely on
+    std::atomic<TimePoint> time_point;
-    /// undefined behavior where the level of accuracy in the clock shouldn't
-    /// be higher.
-    static constexpr u64 inaccuracy_mask = ~(UINT64_C(0x400) - 1);
-
-    TimePoint time_point;
     // factors
     u64 clock_rtsc_factor{};
     u64 cpu_rtsc_factor{};

src/core/core_timing.cpp

@@ -7,6 +7,7 @@
 #include <tuple>
 
 #include "common/microprofile.h"
+#include "common/thread.h"
 #include "core/core_timing.h"
 #include "core/core_timing_util.h"
 #include "core/hardware_properties.h"
@@ -45,7 +46,7 @@ void CoreTiming::ThreadEntry(CoreTiming& instance) {
     constexpr char name[] = "yuzu:HostTiming";
     MicroProfileOnThreadCreate(name);
     Common::SetCurrentThreadName(name);
-    Common::SetCurrentThreadPriority(Common::ThreadPriority::VeryHigh);
+    Common::SetCurrentThreadPriority(Common::ThreadPriority::Critical);
     instance.on_thread_init();
     instance.ThreadLoop();
     MicroProfileOnThreadExit();
@@ -59,68 +60,96 @@ void CoreTiming::Initialize(std::function<void()>&& on_thread_init_) {
     const auto empty_timed_callback = [](std::uintptr_t, std::chrono::nanoseconds) {};
     ev_lost = CreateEvent("_lost_event", empty_timed_callback);
     if (is_multicore) {
-        timer_thread = std::make_unique<std::thread>(ThreadEntry, std::ref(*this));
+        const auto hardware_concurrency = std::thread::hardware_concurrency();
+        worker_threads.emplace_back(ThreadEntry, std::ref(*this));
+        if (hardware_concurrency > 8) {
+            worker_threads.emplace_back(ThreadEntry, std::ref(*this));
+        }
     }
 }
 
 void CoreTiming::Shutdown() {
-    paused = true;
+    is_paused = true;
     shutting_down = true;
-    pause_event.Set();
+    {
-    event.Set();
+        std::unique_lock main_lock(event_mutex);
-    if (timer_thread) {
+        event_cv.notify_all();
-        timer_thread->join();
+        wait_pause_cv.notify_all();
     }
+    for (auto& thread : worker_threads) {
+        thread.join();
+    }
+    worker_threads.clear();
     ClearPendingEvents();
-    timer_thread.reset();
     has_started = false;
 }
 
-void CoreTiming::Pause(bool is_paused) {
+void CoreTiming::Pause(bool is_paused_) {
-    paused = is_paused;
+    std::unique_lock main_lock(event_mutex);
-    pause_event.Set();
+    if (is_paused_ == paused_state.load(std::memory_order_relaxed)) {
-}
-
-void CoreTiming::SyncPause(bool is_paused) {
-    if (is_paused == paused && paused_set == paused) {
         return;
     }
-    Pause(is_paused);
+    if (is_multicore) {
-    if (timer_thread) {
+        is_paused = is_paused_;
-        if (!is_paused) {
+        event_cv.notify_all();
-            pause_event.Set();
+        if (!is_paused_) {
+            wait_pause_cv.notify_all();
+        }
+    }
+    paused_state.store(is_paused_, std::memory_order_relaxed);
+}
+
+void CoreTiming::SyncPause(bool is_paused_) {
+    std::unique_lock main_lock(event_mutex);
+    if (is_paused_ == paused_state.load(std::memory_order_relaxed)) {
+        return;
+    }
+
+    if (is_multicore) {
+        is_paused = is_paused_;
+        event_cv.notify_all();
+        if (!is_paused_) {
+            wait_pause_cv.notify_all();
+        }
+    }
+    paused_state.store(is_paused_, std::memory_order_relaxed);
+    if (is_multicore) {
+        if (is_paused_) {
+            wait_signal_cv.wait(main_lock, [this] { return pause_count == worker_threads.size(); });
+        } else {
+            wait_signal_cv.wait(main_lock, [this] { return pause_count == 0; });
         }
-        event.Set();
-        while (paused_set != is_paused)
-            ;
     }
 }
 
 bool CoreTiming::IsRunning() const {
-    return !paused_set;
+    return !paused_state.load(std::memory_order_acquire);
 }
 
 bool CoreTiming::HasPendingEvents() const {
-    return !(wait_set && event_queue.empty());
+    std::unique_lock main_lock(event_mutex);
+    return !event_queue.empty();
 }
 
 void CoreTiming::ScheduleEvent(std::chrono::nanoseconds ns_into_future,
                                const std::shared_ptr<EventType>& event_type,
                                std::uintptr_t user_data) {
-    {
-        std::scoped_lock scope{basic_lock};
-        const u64 timeout = static_cast<u64>((GetGlobalTimeNs() + ns_into_future).count());
 
-        event_queue.emplace_back(Event{timeout, event_fifo_id++, user_data, event_type});
+    std::unique_lock main_lock(event_mutex);
+    const u64 timeout = static_cast<u64>((GetGlobalTimeNs() + ns_into_future).count());
 
-        std::push_heap(event_queue.begin(), event_queue.end(), std::greater<>());
+    event_queue.emplace_back(Event{timeout, event_fifo_id++, user_data, event_type});
+
+    std::push_heap(event_queue.begin(), event_queue.end(), std::greater<>());
+
+    if (is_multicore) {
+        event_cv.notify_one();
     }
-    event.Set();
 }
 
 void CoreTiming::UnscheduleEvent(const std::shared_ptr<EventType>& event_type,
                                  std::uintptr_t user_data) {
-    std::scoped_lock scope{basic_lock};
+    std::unique_lock main_lock(event_mutex);
     const auto itr = std::remove_if(event_queue.begin(), event_queue.end(), [&](const Event& e) {
         return e.type.lock().get() == event_type.get() && e.user_data == user_data;
     });
@@ -168,11 +197,12 @@ u64 CoreTiming::GetClockTicks() const {
 }
 
 void CoreTiming::ClearPendingEvents() {
+    std::unique_lock main_lock(event_mutex);
     event_queue.clear();
 }
 
 void CoreTiming::RemoveEvent(const std::shared_ptr<EventType>& event_type) {
-    std::scoped_lock lock{basic_lock};
+    std::unique_lock main_lock(event_mutex);
 
     const auto itr = std::remove_if(event_queue.begin(), event_queue.end(), [&](const Event& e) {
         return e.type.lock().get() == event_type.get();
@@ -186,21 +216,22 @@ void CoreTiming::RemoveEvent(const std::shared_ptr<EventType>& event_type) {
 }
 
 std::optional<s64> CoreTiming::Advance() {
-    std::scoped_lock lock{advance_lock, basic_lock};
     global_timer = GetGlobalTimeNs().count();
 
+    std::unique_lock main_lock(event_mutex);
     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();
+        event_mutex.unlock();
 
         if (const auto event_type{evt.type.lock()}) {
-            event_type->callback(
+            std::unique_lock lk(event_type->guard);
-                evt.user_data, std::chrono::nanoseconds{static_cast<s64>(global_timer - evt.time)});
+            event_type->callback(evt.user_data, std::chrono::nanoseconds{static_cast<s64>(
+                                                    GetGlobalTimeNs().count() - evt.time)});
         }
 
-        basic_lock.lock();
+        event_mutex.lock();
         global_timer = GetGlobalTimeNs().count();
     }
 
@@ -213,26 +244,34 @@ std::optional<s64> CoreTiming::Advance() {
 }
 
 void CoreTiming::ThreadLoop() {
+    const auto predicate = [this] { return !event_queue.empty() || is_paused; };
     has_started = true;
     while (!shutting_down) {
-        while (!paused) {
+        while (!is_paused && !shutting_down) {
-            paused_set = false;
             const auto next_time = Advance();
             if (next_time) {
                 if (*next_time > 0) {
                     std::chrono::nanoseconds next_time_ns = std::chrono::nanoseconds(*next_time);
-                    event.WaitFor(next_time_ns);
+                    std::unique_lock main_lock(event_mutex);
+                    event_cv.wait_for(main_lock, next_time_ns, predicate);
                 }
             } else {
-                wait_set = true;
+                std::unique_lock main_lock(event_mutex);
-                event.Wait();
+                event_cv.wait(main_lock, predicate);
             }
-            wait_set = false;
         }
-        paused_set = true;
+        std::unique_lock main_lock(event_mutex);
-        clock->Pause(true);
+        pause_count++;
-        pause_event.Wait();
+        if (pause_count == worker_threads.size()) {
-        clock->Pause(false);
+            clock->Pause(true);
+            wait_signal_cv.notify_all();
+        }
+        wait_pause_cv.wait(main_lock, [this] { return !is_paused || shutting_down; });
+        pause_count--;
+        if (pause_count == 0) {
+            clock->Pause(false);
+            wait_signal_cv.notify_all();
+        }
     }
 }
 

src/core/core_timing.h

@@ -5,6 +5,7 @@
 
 #include <atomic>
 #include <chrono>
+#include <condition_variable>
 #include <functional>
 #include <memory>
 #include <mutex>
@@ -14,7 +15,6 @@
 #include <vector>
 
 #include "common/common_types.h"
-#include "common/thread.h"
 #include "common/wall_clock.h"
 
 namespace Core::Timing {
@@ -32,6 +32,7 @@ struct EventType {
     TimedCallback callback;
     /// A pointer to the name of the event.
     const std::string name;
+    mutable std::mutex guard;
 };
 
 /**
@@ -146,19 +147,21 @@ private:
     u64 event_fifo_id = 0;
 
     std::shared_ptr<EventType> ev_lost;
-    Common::Event event{};
-    Common::Event pause_event{};
-    std::mutex basic_lock;
-    std::mutex 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::function<void()> on_thread_init{};
 
+    std::vector<std::thread> worker_threads;
+
+    std::condition_variable event_cv;
+    std::condition_variable wait_pause_cv;
+    std::condition_variable wait_signal_cv;
+    mutable std::mutex event_mutex;
+
+    std::atomic<bool> paused_state{};
+    bool is_paused{};
+    bool shutting_down{};
     bool is_multicore{};
+    size_t pause_count{};
 
     /// Cycle timing
     u64 ticks{};

src/core/hle/service/nvdrv/core/nvmap.cpp

@@ -41,22 +41,18 @@ NvResult NvMap::Handle::Alloc(Flags pFlags, u32 pAlign, u8 pKind, u64 pAddress)
     size = Common::AlignUp(size, PAGE_SIZE);
     aligned_size = Common::AlignUp(size, align);
     address = pAddress;
-
-    // TODO: pin init
-
     allocated = true;
 
     return NvResult::Success;
 }
 
 NvResult NvMap::Handle::Duplicate(bool internal_session) {
+    std::scoped_lock lock(mutex);
     // Unallocated handles cannot be duplicated as duplication requires memory accounting (in HOS)
     if (!allocated) [[unlikely]] {
         return NvResult::BadValue;
     }
 
-    std::scoped_lock lock(mutex);
-
     // If we internally use FromId the duplication tracking of handles won't work accurately due to
     // us not implementing per-process handle refs.
     if (internal_session) {

src/core/hle/service/nvdrv/devices/nvhost_gpu.cpp

@@ -270,12 +270,12 @@ NvResult nvhost_gpu::SubmitGPFIFOImpl(IoctlSubmitGpfifo& params, std::vector<u8>
         }
     }
 
-    gpu.PushGPUEntries(bind_id, std::move(entries));
     params.fence.id = channel_syncpoint;
 
     u32 increment{(flags.fence_increment.Value() != 0 ? 2 : 0) +
                   (flags.increment_value.Value() != 0 ? params.fence.value : 0)};
     params.fence.value = syncpoint_manager.IncrementSyncpointMaxExt(channel_syncpoint, increment);
+    gpu.PushGPUEntries(bind_id, std::move(entries));
 
     if (flags.fence_increment.Value()) {
         if (flags.suppress_wfi.Value()) {

src/tests/core/core_timing.cpp

@@ -8,6 +8,7 @@
 #include <chrono>
 #include <cstdlib>
 #include <memory>
+#include <mutex>
 #include <string>
 
 #include "core/core.h"
@@ -21,13 +22,14 @@ std::array<s64, 5> delays{};
 
 std::bitset<CB_IDS.size()> callbacks_ran_flags;
 u64 expected_callback = 0;
+std::mutex control_mutex;
 
 template <unsigned int IDX>
 void HostCallbackTemplate(std::uintptr_t user_data, std::chrono::nanoseconds ns_late) {
+    std::unique_lock<std::mutex> lk(control_mutex);
     static_assert(IDX < CB_IDS.size(), "IDX out of range");
     callbacks_ran_flags.set(IDX);
     REQUIRE(CB_IDS[IDX] == user_data);
-    REQUIRE(CB_IDS[IDX] == CB_IDS[calls_order[expected_callback]]);
     delays[IDX] = ns_late.count();
     ++expected_callback;
 }