merge 'dev' into 'metal-dev'

src/core/core_timing.cpp

@@ -26,6 +26,24 @@ std::shared_ptr<EventType> CreateEvent(std::string name, TimedCallback&& callbac
     return std::make_shared<EventType>(std::move(callback), std::move(name));
 }
 
+struct CoreTiming::Event {
+    s64 time;
+    u64 fifo_order;
+    std::weak_ptr<EventType> type;
+    s64 reschedule_time;
+    heap_t::handle_type handle{};
+
+    // 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::CreateOptimalClock()} {}
 
 CoreTiming::~CoreTiming() {
@@ -69,7 +87,7 @@ void CoreTiming::Pause(bool is_paused) {
 }
 
 void CoreTiming::SyncPause(bool is_paused) {
-    if (is_paused == paused && paused_set == is_paused) {
+    if (is_paused == paused && paused_set == paused) {
         return;
     }
 
@@ -94,7 +112,7 @@ bool CoreTiming::IsRunning() const {
 
 bool CoreTiming::HasPendingEvents() const {
     std::scoped_lock lock{basic_lock};
-    return !event_queue.empty();
+    return !(wait_set && event_queue.empty());
 }
 
 void CoreTiming::ScheduleEvent(std::chrono::nanoseconds ns_into_future,
@@ -103,8 +121,8 @@ void CoreTiming::ScheduleEvent(std::chrono::nanoseconds ns_into_future,
         std::scoped_lock scope{basic_lock};
         const auto next_time{absolute_time ? ns_into_future : GetGlobalTimeNs() + ns_into_future};
 
-        event_queue.emplace_back(Event{next_time.count(), event_fifo_id++, event_type});
+        auto h{event_queue.emplace(Event{next_time.count(), event_fifo_id++, event_type, 0})};
-        std::push_heap(event_queue.begin(), event_queue.end(), std::greater<>());
+        (*h).handle = h;
     }
 
     event.Set();
@@ -118,9 +136,9 @@ void CoreTiming::ScheduleLoopingEvent(std::chrono::nanoseconds start_time,
         std::scoped_lock scope{basic_lock};
         const auto next_time{absolute_time ? start_time : GetGlobalTimeNs() + start_time};
 
-        event_queue.emplace_back(
+        auto h{event_queue.emplace(
-            Event{next_time.count(), event_fifo_id++, event_type, resched_time.count()});
+            Event{next_time.count(), event_fifo_id++, event_type, resched_time.count()})};
-        std::push_heap(event_queue.begin(), event_queue.end(), std::greater<>());
+        (*h).handle = h;
     }
 
     event.Set();
@@ -131,11 +149,17 @@ void CoreTiming::UnscheduleEvent(const std::shared_ptr<EventType>& event_type,
     {
         std::scoped_lock lk{basic_lock};
 
-        event_queue.erase(
+        std::vector<heap_t::handle_type> to_remove;
-            std::remove_if(event_queue.begin(), event_queue.end(),
+        for (auto itr = event_queue.begin(); itr != event_queue.end(); itr++) {
-                           [&](const Event& e) { return e.type.lock().get() == event_type.get(); }),
+            const Event& e = *itr;
-            event_queue.end());
+            if (e.type.lock().get() == event_type.get()) {
-        std::make_heap(event_queue.begin(), event_queue.end(), std::greater<>());
+                to_remove.push_back(itr->handle);
+            }
+        }
+
+        for (auto& h : to_remove) {
+            event_queue.erase(h);
+        }
 
         event_type->sequence_number++;
     }
@@ -148,7 +172,7 @@ void CoreTiming::UnscheduleEvent(const std::shared_ptr<EventType>& event_type,
 
 void CoreTiming::AddTicks(u64 ticks_to_add) {
     cpu_ticks += ticks_to_add;
-    downcount -= static_cast<s64>(ticks_to_add);
+    downcount -= static_cast<s64>(cpu_ticks);
 }
 
 void CoreTiming::Idle() {
@@ -156,7 +180,7 @@ void CoreTiming::Idle() {
 }
 
 void CoreTiming::ResetTicks() {
-    downcount.store(MAX_SLICE_LENGTH, std::memory_order_release);
+    downcount = MAX_SLICE_LENGTH;
 }
 
 u64 CoreTiming::GetClockTicks() const {
@@ -177,38 +201,48 @@ std::optional<s64> CoreTiming::Advance() {
     std::scoped_lock lock{advance_lock, basic_lock};
     global_timer = GetGlobalTimeNs().count();
 
-    while (!event_queue.empty() && event_queue.front().time <= global_timer) {
+    while (!event_queue.empty() && event_queue.top().time <= global_timer) {
-        Event evt = std::move(event_queue.front());
+        const Event& evt = event_queue.top();
-        std::pop_heap(event_queue.begin(), event_queue.end(), std::greater<>());
-        event_queue.pop_back();
 
-        if (const auto event_type = evt.type.lock()) {
+        if (const auto event_type{evt.type.lock()}) {
             const auto evt_time = evt.time;
             const auto evt_sequence_num = event_type->sequence_number;
 
-            basic_lock.unlock();
+            if (evt.reschedule_time == 0) {
+                event_queue.pop();
 
-            const auto new_schedule_time = event_type->callback(
+                basic_lock.unlock();
-                evt_time, std::chrono::nanoseconds{GetGlobalTimeNs().count() - evt_time});
 
-            basic_lock.lock();
+                event_type->callback(
+                    evt_time, std::chrono::nanoseconds{GetGlobalTimeNs().count() - evt_time});
 
-            if (evt_sequence_num != event_type->sequence_number) {
+                basic_lock.lock();
-                continue;
+            } else {
-            }
+                basic_lock.unlock();
 
-            if (new_schedule_time.has_value() || evt.reschedule_time != 0) {
+                const auto new_schedule_time{event_type->callback(
-                const auto next_schedule_time = new_schedule_time.value_or(
+                    evt_time, std::chrono::nanoseconds{GetGlobalTimeNs().count() - evt_time})};
-                    std::chrono::nanoseconds{evt.reschedule_time});
 
-                auto next_time = evt.time + next_schedule_time.count();
+                basic_lock.lock();
-                if (evt.time < pause_end_time) {
+
-                    next_time = pause_end_time + next_schedule_time.count();
+                if (evt_sequence_num != event_type->sequence_number) {
+                    // Heap handle is invalidated after external modification.
+                    continue;
                 }
 
-                event_queue.emplace_back(Event{next_time, event_fifo_id++, evt.type,
+                const auto next_schedule_time{new_schedule_time.has_value()
-                                               next_schedule_time.count()});
+                                                  ? new_schedule_time.value().count()
-                std::push_heap(event_queue.begin(), event_queue.end(), std::greater<>());
+                                                  : evt.reschedule_time};
+
+                // If this event was scheduled into a pause, its time now is going to be way
+                // behind. Re-set this event to continue from the end of the pause.
+                auto next_time{evt.time + next_schedule_time};
+                if (evt.time < pause_end_time) {
+                    next_time = pause_end_time + next_schedule_time;
+                }
+
+                event_queue.update(evt.handle, Event{next_time, event_fifo_id++, evt.type,
+                                                     next_schedule_time, evt.handle});
             }
         }
 
@@ -216,7 +250,7 @@ std::optional<s64> CoreTiming::Advance() {
     }
 
     if (!event_queue.empty()) {
-        return event_queue.front().time;
+        return event_queue.top().time;
     } else {
         return std::nullopt;
     }
@@ -235,7 +269,7 @@ void CoreTiming::ThreadLoop() {
 #ifdef _WIN32
                     while (!paused && !event.IsSet() && wait_time > 0) {
                         wait_time = *next_time - GetGlobalTimeNs().count();
-                        if (wait_time >= 1'000'000) { // 1ms
+                        if (wait_time >= timer_resolution_ns) {
                             Common::Windows::SleepForOneTick();
                         } else {
 #ifdef ARCHITECTURE_x86_64
@@ -256,8 +290,10 @@ void CoreTiming::ThreadLoop() {
             } else {
                 // Queue is empty, wait until another event is scheduled and signals us to
                 // continue.
+                wait_set = true;
                 event.Wait();
             }
+            wait_set = false;
         }
 
         paused_set = true;
@@ -291,4 +327,10 @@ std::chrono::microseconds CoreTiming::GetGlobalTimeUs() const {
     return std::chrono::microseconds{Common::WallClock::CPUTickToUS(cpu_ticks)};
 }
 
+#ifdef _WIN32
+void CoreTiming::SetTimerResolutionNs(std::chrono::nanoseconds ns) {
+    timer_resolution_ns = ns.count();
+}
+#endif
+
 } // namespace Core::Timing

src/core/core_timing.h

@@ -11,7 +11,8 @@
 #include <optional>
 #include <string>
 #include <thread>
-#include <vector>
+
+#include <boost/heap/fibonacci_heap.hpp>
 
 #include "common/common_types.h"
 #include "common/thread.h"
@@ -42,6 +43,18 @@ enum class UnscheduleEventType {
     NoWait,
 };
 
+/**
+ * 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 s64 ns_late that the callbacks get is how many ns late it was.
+ * So to schedule a new event on a regular basis:
+ * inside callback:
+ *   ScheduleEvent(period_in_ns - ns_late, callback, "whatever")
+ */
 class CoreTiming {
 public:
     CoreTiming();
@@ -53,56 +66,99 @@ public:
     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(std::function<void()>&& on_thread_init_);
+
+    /// Clear all pending events. This should ONLY be done on exit.
     void ClearPendingEvents();
+
+    /// Sets if emulation is multicore or single core, must be set before Initialize
     void SetMulticore(bool is_multicore_) {
         is_multicore = is_multicore_;
     }
+
+    /// 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(std::chrono::nanoseconds ns_into_future,
                        const std::shared_ptr<EventType>& event_type, bool absolute_time = false);
+
+    /// Schedules an event which will automatically re-schedule itself with the given time, until
+    /// unscheduled
     void ScheduleLoopingEvent(std::chrono::nanoseconds start_time,
                               std::chrono::nanoseconds resched_time,
                               const std::shared_ptr<EventType>& event_type,
                               bool absolute_time = false);
+
     void UnscheduleEvent(const std::shared_ptr<EventType>& event_type,
                          UnscheduleEventType type = UnscheduleEventType::Wait);
+
     void AddTicks(u64 ticks_to_add);
+
     void ResetTicks();
+
     void Idle();
+
     s64 GetDowncount() const {
-        return downcount.load(std::memory_order_relaxed);
+        return downcount;
     }
+
+    /// 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;
+
+    /// 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<s64> Advance();
 
+#ifdef _WIN32
+    void SetTimerResolutionNs(std::chrono::nanoseconds ns);
+#endif
+
 private:
-    struct Event {
+    struct Event;
-        s64 time;
-        u64 fifo_order;
-        std::shared_ptr<EventType> type;
-        bool operator>(const Event& other) const {
-            return std::tie(time, fifo_order) > std::tie(other.time, other.fifo_order);
-        }
-    };
 
     static void ThreadEntry(CoreTiming& instance);
     void ThreadLoop();
+
     void Reset();
 
     std::unique_ptr<Common::WallClock> clock;
-    std::atomic<s64> global_timer{0};
+
-    std::vector<Event> event_queue;
+    s64 global_timer = 0;
-    std::atomic<u64> event_fifo_id{0};
+
+#ifdef _WIN32
+    s64 timer_resolution_ns;
+#endif
+
+    using heap_t =
+        boost::heap::fibonacci_heap<CoreTiming::Event, boost::heap::compare<std::greater<>>>;
+
+    heap_t event_queue;
+    u64 event_fifo_id = 0;
 
     Common::Event event{};
     Common::Event pause_event{};
@@ -117,12 +173,20 @@ private:
     std::function<void()> on_thread_init{};
 
     bool is_multicore{};
-    std::atomic<s64> pause_end_time{};
+    s64 pause_end_time{};
 
-    std::atomic<u64> cpu_ticks{};
+    /// Cycle timing
-    std::atomic<s64> downcount{};
+    u64 cpu_ticks{};
+    s64 downcount{};
 };
 
+/// 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::Timing

src/core/cpu_manager.cpp

@@ -1,12 +1,6 @@
 // SPDX-FileCopyrightText: Copyright 2018 yuzu Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
 
-#include <algorithm>
-#include <atomic>
-#include <memory>
-#include <thread>
-#include <vector>
-
 #include "common/fiber.h"
 #include "common/microprofile.h"
 #include "common/scope_exit.h"
@@ -30,7 +24,6 @@ void CpuManager::Initialize() {
     num_cores = is_multicore ? Core::Hardware::NUM_CPU_CORES : 1;
     gpu_barrier = std::make_unique<Common::Barrier>(num_cores + 1);
 
-    core_data.resize(num_cores);
     for (std::size_t core = 0; core < num_cores; core++) {
         core_data[core].host_thread =
             std::jthread([this, core](std::stop_token token) { RunThread(token, core); });
@@ -38,10 +31,10 @@ void CpuManager::Initialize() {
 }
 
 void CpuManager::Shutdown() {
-    for (auto& data : core_data) {
+    for (std::size_t core = 0; core < num_cores; core++) {
-        if (data.host_thread.joinable()) {
+        if (core_data[core].host_thread.joinable()) {
-            data.host_thread.request_stop();
+            core_data[core].host_thread.request_stop();
-            data.host_thread.join();
+            core_data[core].host_thread.join();
         }
     }
 }
@@ -73,7 +66,12 @@ void CpuManager::HandleInterrupt() {
     Kernel::KInterruptManager::HandleInterrupt(kernel, static_cast<s32>(core_index));
 }
 
+///////////////////////////////////////////////////////////////////////////////
+///                             MultiCore                                   ///
+///////////////////////////////////////////////////////////////////////////////
+
 void CpuManager::MultiCoreRunGuestThread() {
+    // Similar to UserModeThreadStarter in HOS
     auto& kernel = system.Kernel();
     auto* thread = Kernel::GetCurrentThreadPointer(kernel);
     kernel.CurrentScheduler()->OnThreadStart();
@@ -90,6 +88,10 @@ void CpuManager::MultiCoreRunGuestThread() {
 }
 
 void CpuManager::MultiCoreRunIdleThread() {
+    // Not accurate to HOS. Remove this entire method when singlecore is removed.
+    // See notes in KScheduler::ScheduleImpl for more information about why this
+    // is inaccurate.
+
     auto& kernel = system.Kernel();
     kernel.CurrentScheduler()->OnThreadStart();
 
@@ -103,6 +105,10 @@ void CpuManager::MultiCoreRunIdleThread() {
     }
 }
 
+///////////////////////////////////////////////////////////////////////////////
+///                             SingleCore                                   ///
+///////////////////////////////////////////////////////////////////////////////
+
 void CpuManager::SingleCoreRunGuestThread() {
     auto& kernel = system.Kernel();
     auto* thread = Kernel::GetCurrentThreadPointer(kernel);
@@ -148,16 +154,19 @@ void CpuManager::PreemptSingleCore(bool from_running_environment) {
         system.CoreTiming().Advance();
         kernel.SetIsPhantomModeForSingleCore(false);
     }
-    current_core.store((current_core + 1) % Core::Hardware::NUM_CPU_CORES, std::memory_order_release);
+    current_core.store((current_core + 1) % Core::Hardware::NUM_CPU_CORES);
     system.CoreTiming().ResetTicks();
     kernel.Scheduler(current_core).PreemptSingleCore();
 
+    // We've now been scheduled again, and we may have exchanged schedulers.
+    // Reload the scheduler in case it's different.
     if (!kernel.Scheduler(current_core).IsIdle()) {
         idle_count = 0;
     }
 }
 
 void CpuManager::GuestActivate() {
+    // Similar to the HorizonKernelMain callback in HOS
     auto& kernel = system.Kernel();
     auto* scheduler = kernel.CurrentScheduler();
 
@@ -175,19 +184,27 @@ void CpuManager::ShutdownThread() {
 }
 
 void CpuManager::RunThread(std::stop_token token, std::size_t core) {
+    /// Initialization
     system.RegisterCoreThread(core);
-    std::string name = is_multicore ? "CPUCore_" + std::to_string(core) : "CPUThread";
+    std::string name;
+    if (is_multicore) {
+        name = "CPUCore_" + std::to_string(core);
+    } else {
+        name = "CPUThread";
+    }
     MicroProfileOnThreadCreate(name.c_str());
     Common::SetCurrentThreadName(name.c_str());
     Common::SetCurrentThreadPriority(Common::ThreadPriority::Critical);
     auto& data = core_data[core];
     data.host_context = Common::Fiber::ThreadToFiber();
 
+    // Cleanup
     SCOPE_EXIT {
         data.host_context->Exit();
         MicroProfileOnThreadExit();
     };
 
+    // Running
     if (!gpu_barrier->Sync(token)) {
         return;
     }

src/video_core/gpu.cpp

@@ -40,23 +40,10 @@ struct GPU::Impl {
     explicit Impl(GPU& gpu_, Core::System& system_, bool is_async_, bool use_nvdec_)
         : gpu{gpu_}, system{system_}, host1x{system.Host1x()}, use_nvdec{use_nvdec_},
           shader_notify{std::make_unique<VideoCore::ShaderNotify>()}, is_async{is_async_},
-          gpu_thread{system_, is_async_}, scheduler{std::make_unique<Control::Scheduler>(gpu)} {
+          gpu_thread{system_, is_async_}, scheduler{std::make_unique<Control::Scheduler>(gpu)} {}
-        Initialize();
-    }
 
     ~Impl() = default;
 
-    void Initialize() {
-        // Initialize the GPU memory manager
-        memory_manager = std::make_unique<Tegra::MemoryManager>(system);
-
-        // Initialize the command buffer
-        command_buffer.reserve(COMMAND_BUFFER_SIZE);
-
-        // Initialize the fence manager
-        fence_manager = std::make_unique<FenceManager>();
-    }
-
     std::shared_ptr<Control::ChannelState> CreateChannel(s32 channel_id) {
         auto channel_state = std::make_shared<Tegra::Control::ChannelState>(channel_id);
         channels.emplace(channel_id, channel_state);
@@ -104,15 +91,14 @@ struct GPU::Impl {
 
     /// Flush all current written commands into the host GPU for execution.
     void FlushCommands() {
-        if (!command_buffer.empty()) {
+        rasterizer->FlushCommands();
-            rasterizer->ExecuteCommands(command_buffer);
-            command_buffer.clear();
-        }
     }
 
     /// Synchronizes CPU writes with Host GPU memory.
     void InvalidateGPUCache() {
-        rasterizer->InvalidateGPUCache();
+        std::function<void(PAddr, size_t)> callback_writes(
+            [this](PAddr address, size_t size) { rasterizer->OnCacheInvalidation(address, size); });
+        system.GatherGPUDirtyMemory(callback_writes);
     }
 
     /// Signal the ending of command list.
@@ -122,10 +108,11 @@ struct GPU::Impl {
     }
 
     /// Request a host GPU memory flush from the CPU.
-    u64 RequestSyncOperation(std::function<void()>&& action) {
+    template <typename Func>
+    [[nodiscard]] u64 RequestSyncOperation(Func&& action) {
         std::unique_lock lck{sync_request_mutex};
         const u64 fence = ++last_sync_fence;
-        sync_requests.emplace_back(std::move(action), fence);
+        sync_requests.emplace_back(action);
         return fence;
     }
 
@@ -143,12 +130,12 @@ struct GPU::Impl {
     void TickWork() {
         std::unique_lock lck{sync_request_mutex};
         while (!sync_requests.empty()) {
-            auto& request = sync_requests.front();
+            auto request = std::move(sync_requests.front());
+            sync_requests.pop_front();
             sync_request_mutex.unlock();
-            request.first();
+            request();
             current_sync_fence.fetch_add(1, std::memory_order_release);
             sync_request_mutex.lock();
-            sync_requests.pop_front();
             sync_request_cv.notify_all();
         }
     }
@@ -235,6 +222,7 @@ struct GPU::Impl {
     /// This can be used to launch any necessary threads and register any necessary
     /// core timing events.
     void Start() {
+        Settings::UpdateGPUAccuracy();
         gpu_thread.StartThread(*renderer, renderer->Context(), *scheduler);
     }
 
@@ -264,7 +252,7 @@ struct GPU::Impl {
 
     /// Notify rasterizer that any caches of the specified region should be flushed to Switch memory
     void FlushRegion(DAddr addr, u64 size) {
-        rasterizer->FlushRegion(addr, size);
+        gpu_thread.FlushRegion(addr, size);
     }
 
     VideoCore::RasterizerDownloadArea OnCPURead(DAddr addr, u64 size) {
@@ -284,7 +272,7 @@ struct GPU::Impl {
 
     /// Notify rasterizer that any caches of the specified region should be invalidated
     void InvalidateRegion(DAddr addr, u64 size) {
-        rasterizer->InvalidateRegion(addr, size);
+        gpu_thread.InvalidateRegion(addr, size);
     }
 
     bool OnCPUWrite(DAddr addr, u64 size) {
@@ -293,7 +281,57 @@ struct GPU::Impl {
 
     /// Notify rasterizer that any caches of the specified region should be flushed and invalidated
     void FlushAndInvalidateRegion(DAddr addr, u64 size) {
-        rasterizer->FlushAndInvalidateRegion(addr, size);
+        gpu_thread.FlushAndInvalidateRegion(addr, size);
+    }
+
+    void RequestComposite(std::vector<Tegra::FramebufferConfig>&& layers,
+                          std::vector<Service::Nvidia::NvFence>&& fences) {
+        size_t num_fences{fences.size()};
+        size_t current_request_counter{};
+        {
+            std::unique_lock<std::mutex> lk(request_swap_mutex);
+            if (free_swap_counters.empty()) {
+                current_request_counter = request_swap_counters.size();
+                request_swap_counters.emplace_back(num_fences);
+            } else {
+                current_request_counter = free_swap_counters.front();
+                request_swap_counters[current_request_counter] = num_fences;
+                free_swap_counters.pop_front();
+            }
+        }
+        const auto wait_fence =
+            RequestSyncOperation([this, current_request_counter, &layers, &fences, num_fences] {
+                auto& syncpoint_manager = host1x.GetSyncpointManager();
+                if (num_fences == 0) {
+                    renderer->Composite(layers);
+                }
+                const auto executer = [this, current_request_counter, layers_copy = layers]() {
+                    {
+                        std::unique_lock<std::mutex> lk(request_swap_mutex);
+                        if (--request_swap_counters[current_request_counter] != 0) {
+                            return;
+                        }
+                        free_swap_counters.push_back(current_request_counter);
+                    }
+                    renderer->Composite(layers_copy);
+                };
+                for (size_t i = 0; i < num_fences; i++) {
+                    syncpoint_manager.RegisterGuestAction(fences[i].id, fences[i].value, executer);
+                }
+            });
+        gpu_thread.TickGPU();
+        WaitForSyncOperation(wait_fence);
+    }
+
+    std::vector<u8> GetAppletCaptureBuffer() {
+        std::vector<u8> out;
+
+        const auto wait_fence =
+            RequestSyncOperation([&] { out = renderer->GetAppletCaptureBuffer(); });
+        gpu_thread.TickGPU();
+        WaitForSyncOperation(wait_fence);
+
+        return out;
     }
 
     GPU& gpu;
@@ -310,12 +348,16 @@ struct GPU::Impl {
     /// When true, we are about to shut down emulation session, so terminate outstanding tasks
     std::atomic_bool shutting_down{};
 
+    std::array<std::atomic<u32>, Service::Nvidia::MaxSyncPoints> syncpoints{};
+
+    std::array<std::list<u32>, Service::Nvidia::MaxSyncPoints> syncpt_interrupts;
+
     std::mutex sync_mutex;
     std::mutex device_mutex;
 
     std::condition_variable sync_cv;
 
-    std::list<std::pair<std::function<void()>, u64>> sync_requests;
+    std::list<std::function<void()>> sync_requests;
     std::atomic<u64> current_sync_fence{};
     u64 last_sync_fence{};
     std::mutex sync_request_mutex;
@@ -331,13 +373,182 @@ struct GPU::Impl {
     Tegra::Control::ChannelState* current_channel;
     s32 bound_channel{-1};
 
-    std::unique_ptr<Tegra::MemoryManager> memory_manager;
+    std::deque<size_t> free_swap_counters;
-    std::vector<u32> command_buffer;
+    std::deque<size_t> request_swap_counters;
-    std::unique_ptr<FenceManager> fence_manager;
+    std::mutex request_swap_mutex;
-
-    static constexpr size_t COMMAND_BUFFER_SIZE = 4 * 1024 * 1024;
 };
 
-// ... (rest of the implementation remains the same)
+GPU::GPU(Core::System& system, bool is_async, bool use_nvdec)
+    : impl{std::make_unique<Impl>(*this, system, is_async, use_nvdec)} {}
+
+GPU::~GPU() = default;
+
+std::shared_ptr<Control::ChannelState> GPU::AllocateChannel() {
+    return impl->AllocateChannel();
+}
+
+void GPU::InitChannel(Control::ChannelState& to_init, u64 program_id) {
+    impl->InitChannel(to_init, program_id);
+}
+
+void GPU::BindChannel(s32 channel_id) {
+    impl->BindChannel(channel_id);
+}
+
+void GPU::ReleaseChannel(Control::ChannelState& to_release) {
+    impl->ReleaseChannel(to_release);
+}
+
+void GPU::InitAddressSpace(Tegra::MemoryManager& memory_manager) {
+    impl->InitAddressSpace(memory_manager);
+}
+
+void GPU::BindRenderer(std::unique_ptr<VideoCore::RendererBase> renderer) {
+    impl->BindRenderer(std::move(renderer));
+}
+
+void GPU::FlushCommands() {
+    impl->FlushCommands();
+}
+
+void GPU::InvalidateGPUCache() {
+    impl->InvalidateGPUCache();
+}
+
+void GPU::OnCommandListEnd() {
+    impl->OnCommandListEnd();
+}
+
+u64 GPU::RequestFlush(DAddr addr, std::size_t size) {
+    return impl->RequestSyncOperation(
+        [this, addr, size]() { impl->rasterizer->FlushRegion(addr, size); });
+}
+
+u64 GPU::CurrentSyncRequestFence() const {
+    return impl->CurrentSyncRequestFence();
+}
+
+void GPU::WaitForSyncOperation(u64 fence) {
+    return impl->WaitForSyncOperation(fence);
+}
+
+void GPU::TickWork() {
+    impl->TickWork();
+}
+
+/// Gets a mutable reference to the Host1x interface
+Host1x::Host1x& GPU::Host1x() {
+    return impl->host1x;
+}
+
+/// Gets an immutable reference to the Host1x interface.
+const Host1x::Host1x& GPU::Host1x() const {
+    return impl->host1x;
+}
+
+Engines::Maxwell3D& GPU::Maxwell3D() {
+    return impl->Maxwell3D();
+}
+
+const Engines::Maxwell3D& GPU::Maxwell3D() const {
+    return impl->Maxwell3D();
+}
+
+Engines::KeplerCompute& GPU::KeplerCompute() {
+    return impl->KeplerCompute();
+}
+
+const Engines::KeplerCompute& GPU::KeplerCompute() const {
+    return impl->KeplerCompute();
+}
+
+Tegra::DmaPusher& GPU::DmaPusher() {
+    return impl->DmaPusher();
+}
+
+const Tegra::DmaPusher& GPU::DmaPusher() const {
+    return impl->DmaPusher();
+}
+
+VideoCore::RendererBase& GPU::Renderer() {
+    return impl->Renderer();
+}
+
+const VideoCore::RendererBase& GPU::Renderer() const {
+    return impl->Renderer();
+}
+
+VideoCore::ShaderNotify& GPU::ShaderNotify() {
+    return impl->ShaderNotify();
+}
+
+const VideoCore::ShaderNotify& GPU::ShaderNotify() const {
+    return impl->ShaderNotify();
+}
+
+void GPU::RequestComposite(std::vector<Tegra::FramebufferConfig>&& layers,
+                           std::vector<Service::Nvidia::NvFence>&& fences) {
+    impl->RequestComposite(std::move(layers), std::move(fences));
+}
+
+std::vector<u8> GPU::GetAppletCaptureBuffer() {
+    return impl->GetAppletCaptureBuffer();
+}
+
+u64 GPU::GetTicks() const {
+    return impl->GetTicks();
+}
+
+bool GPU::IsAsync() const {
+    return impl->IsAsync();
+}
+
+bool GPU::UseNvdec() const {
+    return impl->UseNvdec();
+}
+
+void GPU::RendererFrameEndNotify() {
+    impl->RendererFrameEndNotify();
+}
+
+void GPU::Start() {
+    impl->Start();
+}
+
+void GPU::NotifyShutdown() {
+    impl->NotifyShutdown();
+}
+
+void GPU::ObtainContext() {
+    impl->ObtainContext();
+}
+
+void GPU::ReleaseContext() {
+    impl->ReleaseContext();
+}
+
+void GPU::PushGPUEntries(s32 channel, Tegra::CommandList&& entries) {
+    impl->PushGPUEntries(channel, std::move(entries));
+}
+
+VideoCore::RasterizerDownloadArea GPU::OnCPURead(PAddr addr, u64 size) {
+    return impl->OnCPURead(addr, size);
+}
+
+void GPU::FlushRegion(DAddr addr, u64 size) {
+    impl->FlushRegion(addr, size);
+}
+
+void GPU::InvalidateRegion(DAddr addr, u64 size) {
+    impl->InvalidateRegion(addr, size);
+}
+
+bool GPU::OnCPUWrite(DAddr addr, u64 size) {
+    return impl->OnCPUWrite(addr, size);
+}
+
+void GPU::FlushAndInvalidateRegion(DAddr addr, u64 size) {
+    impl->FlushAndInvalidateRegion(addr, size);
+}
 
 } // namespace Tegra

src/video_core/optimized_rasterizer.cpp

@@ -1,221 +0,0 @@
-#include "video_core/optimized_rasterizer.h"
-#include "common/settings.h"
-#include "video_core/gpu.h"
-#include "video_core/memory_manager.h"
-#include "video_core/engines/maxwell_3d.h"
-
-namespace VideoCore {
-
-OptimizedRasterizer::OptimizedRasterizer(Core::System& system, Tegra::GPU& gpu)
-    : system{system}, gpu{gpu}, memory_manager{gpu.MemoryManager()} {
-    InitializeShaderCache();
-}
-
-OptimizedRasterizer::~OptimizedRasterizer() = default;
-
-void OptimizedRasterizer::Draw(bool is_indexed, u32 instance_count) {
-    MICROPROFILE_SCOPE(GPU_Rasterization);
-
-    PrepareRendertarget();
-    UpdateDynamicState();
-
-    if (is_indexed) {
-        DrawIndexed(instance_count);
-    } else {
-        DrawArrays(instance_count);
-    }
-}
-
-void OptimizedRasterizer::Clear(u32 layer_count) {
-    MICROPROFILE_SCOPE(GPU_Rasterization);
-
-    PrepareRendertarget();
-    ClearFramebuffer(layer_count);
-}
-
-void OptimizedRasterizer::DispatchCompute() {
-    MICROPROFILE_SCOPE(GPU_Compute);
-
-    PrepareCompute();
-    LaunchComputeShader();
-}
-
-void OptimizedRasterizer::ResetCounter(VideoCommon::QueryType type) {
-    query_cache.ResetCounter(type);
-}
-
-void OptimizedRasterizer::Query(GPUVAddr gpu_addr, VideoCommon::QueryType type,
-                                VideoCommon::QueryPropertiesFlags flags, u32 payload, u32 subreport) {
-    query_cache.Query(gpu_addr, type, flags, payload, subreport);
-}
-
-void OptimizedRasterizer::FlushAll() {
-    MICROPROFILE_SCOPE(GPU_Synchronization);
-
-    FlushShaderCache();
-    FlushRenderTargets();
-}
-
-void OptimizedRasterizer::FlushRegion(DAddr addr, u64 size, VideoCommon::CacheType which) {
-    MICROPROFILE_SCOPE(GPU_Synchronization);
-
-    if (which == VideoCommon::CacheType::All || which == VideoCommon::CacheType::Unified) {
-        FlushMemoryRegion(addr, size);
-    }
-}
-
-bool OptimizedRasterizer::MustFlushRegion(DAddr addr, u64 size, VideoCommon::CacheType which) {
-    if (which == VideoCommon::CacheType::All || which == VideoCommon::CacheType::Unified) {
-        return IsRegionCached(addr, size);
-    }
-    return false;
-}
-
-RasterizerDownloadArea OptimizedRasterizer::GetFlushArea(DAddr addr, u64 size) {
-    return GetFlushableArea(addr, size);
-}
-
-void OptimizedRasterizer::InvalidateRegion(DAddr addr, u64 size, VideoCommon::CacheType which) {
-    MICROPROFILE_SCOPE(GPU_Synchronization);
-
-    if (which == VideoCommon::CacheType::All || which == VideoCommon::CacheType::Unified) {
-        InvalidateMemoryRegion(addr, size);
-    }
-}
-
-void OptimizedRasterizer::OnCacheInvalidation(PAddr addr, u64 size) {
-    MICROPROFILE_SCOPE(GPU_Synchronization);
-
-    InvalidateCachedRegion(addr, size);
-}
-
-bool OptimizedRasterizer::OnCPUWrite(PAddr addr, u64 size) {
-    return HandleCPUWrite(addr, size);
-}
-
-void OptimizedRasterizer::InvalidateGPUCache() {
-    MICROPROFILE_SCOPE(GPU_Synchronization);
-
-    InvalidateAllCache();
-}
-
-void OptimizedRasterizer::UnmapMemory(DAddr addr, u64 size) {
-    MICROPROFILE_SCOPE(GPU_Synchronization);
-
-    UnmapGPUMemoryRegion(addr, size);
-}
-
-void OptimizedRasterizer::ModifyGPUMemory(size_t as_id, GPUVAddr addr, u64 size) {
-    MICROPROFILE_SCOPE(GPU_Synchronization);
-
-    UpdateMappedGPUMemory(as_id, addr, size);
-}
-
-void OptimizedRasterizer::FlushAndInvalidateRegion(DAddr addr, u64 size, VideoCommon::CacheType which) {
-    MICROPROFILE_SCOPE(GPU_Synchronization);
-
-    if (which == VideoCommon::CacheType::All || which == VideoCommon::CacheType::Unified) {
-        FlushAndInvalidateMemoryRegion(addr, size);
-    }
-}
-
-void OptimizedRasterizer::WaitForIdle() {
-    MICROPROFILE_SCOPE(GPU_Synchronization);
-
-    WaitForGPUIdle();
-}
-
-void OptimizedRasterizer::FragmentBarrier() {
-    MICROPROFILE_SCOPE(GPU_Synchronization);
-
-    InsertFragmentBarrier();
-}
-
-void OptimizedRasterizer::TiledCacheBarrier() {
-    MICROPROFILE_SCOPE(GPU_Synchronization);
-
-    InsertTiledCacheBarrier();
-}
-
-void OptimizedRasterizer::FlushCommands() {
-    MICROPROFILE_SCOPE(GPU_Synchronization);
-
-    SubmitCommands();
-}
-
-void OptimizedRasterizer::TickFrame() {
-    MICROPROFILE_SCOPE(GPU_Synchronization);
-
-    EndFrame();
-}
-
-void OptimizedRasterizer::PrepareRendertarget() {
-    const auto& regs{gpu.Maxwell3D().regs};
-    const auto& framebuffer{regs.framebuffer};
-
-    render_targets.resize(framebuffer.num_color_buffers);
-    for (std::size_t index = 0; index < framebuffer.num_color_buffers; ++index) {
-        render_targets[index] = GetColorBuffer(index);
-    }
-
-    depth_stencil = GetDepthBuffer();
-}
-
-void OptimizedRasterizer::UpdateDynamicState() {
-    const auto& regs{gpu.Maxwell3D().regs};
-
-    UpdateViewport(regs.viewport_transform);
-    UpdateScissor(regs.scissor_test);
-    UpdateDepthBias(regs.polygon_offset_units, regs.polygon_offset_clamp, regs.polygon_offset_factor);
-    UpdateBlendConstants(regs.blend_color);
-    UpdateStencilFaceMask(regs.stencil_front_func_mask, regs.stencil_back_func_mask);
-}
-
-void OptimizedRasterizer::DrawIndexed(u32 instance_count) {
-    const auto& draw_state{gpu.Maxwell3D().draw_manager->GetDrawState()};
-    const auto& index_buffer{memory_manager.ReadBlockUnsafe(draw_state.index_buffer.Address(),
-                                                            draw_state.index_buffer.size)};
-
-    shader_cache.BindComputeShader();
-    shader_cache.BindGraphicsShader();
-
-    DrawElementsInstanced(draw_state.topology, draw_state.index_buffer.count,
-                          draw_state.index_buffer.format, index_buffer.data(), instance_count);
-}
-
-void OptimizedRasterizer::DrawArrays(u32 instance_count) {
-    const auto& draw_state{gpu.Maxwell3D().draw_manager->GetDrawState()};
-
-    shader_cache.BindComputeShader();
-    shader_cache.BindGraphicsShader();
-
-    DrawArraysInstanced(draw_state.topology, draw_state.vertex_buffer.first,
-                        draw_state.vertex_buffer.count, instance_count);
-}
-
-void OptimizedRasterizer::ClearFramebuffer(u32 layer_count) {
-    const auto& regs{gpu.Maxwell3D().regs};
-    const auto& clear_state{regs.clear_buffers};
-
-    if (clear_state.R || clear_state.G || clear_state.B || clear_state.A) {
-        ClearColorBuffers(clear_state.R, clear_state.G, clear_state.B, clear_state.A,
-                          regs.clear_color[0], regs.clear_color[1], regs.clear_color[2],
-                          regs.clear_color[3], layer_count);
-    }
-
-    if (clear_state.Z || clear_state.S) {
-        ClearDepthStencilBuffer(clear_state.Z, clear_state.S, regs.clear_depth, regs.clear_stencil,
-                                layer_count);
-    }
-}
-
-void OptimizedRasterizer::PrepareCompute() {
-    shader_cache.BindComputeShader();
-}
-
-void OptimizedRasterizer::LaunchComputeShader() {
-    const auto& launch_desc{gpu.KeplerCompute().launch_description};
-    DispatchCompute(launch_desc.grid_dim_x, launch_desc.grid_dim_y, launch_desc.grid_dim_z);
-}
-
-} // namespace VideoCore

src/video_core/optimized_rasterizer.h

@@ -1,73 +0,0 @@
-#pragma once
-
-#include <memory>
-#include <vector>
-#include "common/common_types.h"
-#include "video_core/rasterizer_interface.h"
-#include "video_core/engines/maxwell_3d.h"
-
-namespace Core {
-class System;
-}
-
-namespace Tegra {
-class GPU;
-class MemoryManager;
-}
-
-namespace VideoCore {
-
-class ShaderCache;
-class QueryCache;
-
-class OptimizedRasterizer final : public RasterizerInterface {
-public:
-    explicit OptimizedRasterizer(Core::System& system, Tegra::GPU& gpu);
-    ~OptimizedRasterizer() override;
-
-    void Draw(bool is_indexed, u32 instance_count) override;
-    void Clear(u32 layer_count) override;
-    void DispatchCompute() override;
-    void ResetCounter(VideoCommon::QueryType type) override;
-    void Query(GPUVAddr gpu_addr, VideoCommon::QueryType type,
-               VideoCommon::QueryPropertiesFlags flags, u32 payload, u32 subreport) override;
-    void FlushAll() override;
-    void FlushRegion(DAddr addr, u64 size, VideoCommon::CacheType which) override;
-    bool MustFlushRegion(DAddr addr, u64 size, VideoCommon::CacheType which) override;
-    RasterizerDownloadArea GetFlushArea(DAddr addr, u64 size) override;
-    void InvalidateRegion(DAddr addr, u64 size, VideoCommon::CacheType which) override;
-    void OnCacheInvalidation(PAddr addr, u64 size) override;
-    bool OnCPUWrite(PAddr addr, u64 size) override;
-    void InvalidateGPUCache() override;
-    void UnmapMemory(DAddr addr, u64 size) override;
-    void ModifyGPUMemory(size_t as_id, GPUVAddr addr, u64 size) override;
-    void FlushAndInvalidateRegion(DAddr addr, u64 size, VideoCommon::CacheType which) override;
-    void WaitForIdle() override;
-    void FragmentBarrier() override;
-    void TiledCacheBarrier() override;
-    void FlushCommands() override;
-    void TickFrame() override;
-
-private:
-    void PrepareRendertarget();
-    void UpdateDynamicState();
-    void DrawIndexed(u32 instance_count);
-    void DrawArrays(u32 instance_count);
-    void ClearFramebuffer(u32 layer_count);
-    void PrepareCompute();
-    void LaunchComputeShader();
-
-    Core::System& system;
-    Tegra::GPU& gpu;
-    Tegra::MemoryManager& memory_manager;
-
-    std::unique_ptr<ShaderCache> shader_cache;
-    std::unique_ptr<QueryCache> query_cache;
-
-    std::vector<RenderTargetConfig> render_targets;
-    DepthStencilConfig depth_stencil;
-
-    // Add any additional member variables needed for the optimized rasterizer
-};
-
-} // namespace VideoCore

src/video_core/shader_cache.cpp

@@ -3,18 +3,9 @@
 
 #include <algorithm>
 #include <array>
-#include <atomic>
-#include <filesystem>
-#include <fstream>
-#include <mutex>
-#include <thread>
 #include <vector>
 
 #include "common/assert.h"
-#include "common/fs/file.h"
-#include "common/fs/path_util.h"
-#include "common/logging/log.h"
-#include "common/thread_worker.h"
 #include "shader_recompiler/frontend/maxwell/control_flow.h"
 #include "shader_recompiler/object_pool.h"
 #include "video_core/control/channel_state.h"
@@ -28,288 +19,233 @@
 
 namespace VideoCommon {
 
-constexpr size_t MAX_SHADER_CACHE_SIZE = 1024 * 1024 * 1024; // 1GB
-
-class ShaderCacheWorker : public Common::ThreadWorker {
-public:
-    explicit ShaderCacheWorker(const std::string& name) : ThreadWorker(name) {}
-    ~ShaderCacheWorker() = default;
-
-    void CompileShader(ShaderInfo* shader) {
-        Push([shader]() {
-            // Compile shader here
-            // This is a placeholder for the actual compilation process
-            std::this_thread::sleep_for(std::chrono::milliseconds(10));
-            shader->is_compiled.store(true, std::memory_order_release);
-        });
-    }
-};
-
-class ShaderCache::Impl {
-public:
-    explicit Impl(Tegra::MaxwellDeviceMemoryManager& device_memory_)
-        : device_memory{device_memory_}, workers{CreateWorkers()} {
-        LoadCache();
-    }
-
-    ~Impl() {
-        SaveCache();
-    }
-
-    void InvalidateRegion(VAddr addr, size_t size) {
-        std::scoped_lock lock{invalidation_mutex};
-        InvalidatePagesInRegion(addr, size);
-        RemovePendingShaders();
-    }
-
-    void OnCacheInvalidation(VAddr addr, size_t size) {
-        std::scoped_lock lock{invalidation_mutex};
-        InvalidatePagesInRegion(addr, size);
-    }
-
-    void SyncGuestHost() {
-        std::scoped_lock lock{invalidation_mutex};
-        RemovePendingShaders();
-    }
-
-    bool RefreshStages(std::array<u64, 6>& unique_hashes);
-    const ShaderInfo* ComputeShader();
-    void GetGraphicsEnvironments(GraphicsEnvironments& result, const std::array<u64, NUM_PROGRAMS>& unique_hashes);
-
-    ShaderInfo* TryGet(VAddr addr) const {
-        std::scoped_lock lock{lookup_mutex};
-
-        const auto it = lookup_cache.find(addr);
-        if (it == lookup_cache.end()) {
-            return nullptr;
-        }
-        return it->second->data;
-    }
-
-    void Register(std::unique_ptr<ShaderInfo> data, VAddr addr, size_t size) {
-        std::scoped_lock lock{invalidation_mutex, lookup_mutex};
-
-        const VAddr addr_end = addr + size;
-        Entry* const entry = NewEntry(addr, addr_end, data.get());
-
-        const u64 page_end = (addr_end + SUYU_PAGESIZE - 1) >> SUYU_PAGEBITS;
-        for (u64 page = addr >> SUYU_PAGEBITS; page < page_end; ++page) {
-            invalidation_cache[page].push_back(entry);
-        }
-
-        storage.push_back(std::move(data));
-
-        device_memory.UpdatePagesCachedCount(addr, size, 1);
-    }
-
-private:
-    std::vector<std::unique_ptr<ShaderCacheWorker>> CreateWorkers() {
-        const size_t num_workers = std::thread::hardware_concurrency();
-        std::vector<std::unique_ptr<ShaderCacheWorker>> workers;
-        workers.reserve(num_workers);
-        for (size_t i = 0; i < num_workers; ++i) {
-            workers.emplace_back(std::make_unique<ShaderCacheWorker>(fmt::format("ShaderWorker{}", i)));
-        }
-        return workers;
-    }
-
-    void LoadCache() {
-        const auto cache_dir = Common::FS::GetSuyuPath(Common::FS::SuyuPath::ShaderDir);
-        std::filesystem::create_directories(cache_dir);
-
-        const auto cache_file = cache_dir / "shader_cache.bin";
-        if (!std::filesystem::exists(cache_file)) {
-            return;
-        }
-
-        std::ifstream file(cache_file, std::ios::binary);
-        if (!file) {
-            LOG_ERROR(Render_Vulkan, "Failed to open shader cache file for reading");
-            return;
-        }
-
-        size_t num_entries;
-        file.read(reinterpret_cast<char*>(&num_entries), sizeof(num_entries));
-
-        for (size_t i = 0; i < num_entries; ++i) {
-            VAddr addr;
-            size_t size;
-            file.read(reinterpret_cast<char*>(&addr), sizeof(addr));
-            file.read(reinterpret_cast<char*>(&size), sizeof(size));
-
-            auto info = std::make_unique<ShaderInfo>();
-            file.read(reinterpret_cast<char*>(info.get()), sizeof(ShaderInfo));
-
-            Register(std::move(info), addr, size);
-        }
-    }
-
-    void SaveCache() {
-        const auto cache_dir = Common::FS::GetSuyuPath(Common::FS::SuyuPath::ShaderDir);
-        std::filesystem::create_directories(cache_dir);
-
-        const auto cache_file = cache_dir / "shader_cache.bin";
-        std::ofstream file(cache_file, std::ios::binary | std::ios::trunc);
-        if (!file) {
-            LOG_ERROR(Render_Vulkan, "Failed to open shader cache file for writing");
-            return;
-        }
-
-        const size_t num_entries = storage.size();
-        file.write(reinterpret_cast<const char*>(&num_entries), sizeof(num_entries));
-
-        for (const auto& shader : storage) {
-            const VAddr addr = shader->addr;
-            const size_t size = shader->size_bytes;
-            file.write(reinterpret_cast<const char*>(&addr), sizeof(addr));
-            file.write(reinterpret_cast<const char*>(&size), sizeof(size));
-            file.write(reinterpret_cast<const char*>(shader.get()), sizeof(ShaderInfo));
-        }
-    }
-
-    void InvalidatePagesInRegion(VAddr addr, size_t size) {
-        const VAddr addr_end = addr + size;
-        const u64 page_end = (addr_end + SUYU_PAGESIZE - 1) >> SUYU_PAGEBITS;
-        for (u64 page = addr >> SUYU_PAGEBITS; page < page_end; ++page) {
-            auto it = invalidation_cache.find(page);
-            if (it == invalidation_cache.end()) {
-                continue;
-            }
-            InvalidatePageEntries(it->second, addr, addr_end);
-        }
-    }
-
-    void RemovePendingShaders() {
-        if (marked_for_removal.empty()) {
-            return;
-        }
-        // Remove duplicates
-        std::sort(marked_for_removal.begin(), marked_for_removal.end());
-        marked_for_removal.erase(std::unique(marked_for_removal.begin(), marked_for_removal.end()),
-                                 marked_for_removal.end());
-
-        std::vector<ShaderInfo*> removed_shaders;
-
-        std::scoped_lock lock{lookup_mutex};
-        for (Entry* const entry : marked_for_removal) {
-            removed_shaders.push_back(entry->data);
-
-            const auto it = lookup_cache.find(entry->addr_start);
-            ASSERT(it != lookup_cache.end());
-            lookup_cache.erase(it);
-        }
-        marked_for_removal.clear();
-
-        if (!removed_shaders.empty()) {
-            RemoveShadersFromStorage(removed_shaders);
-        }
-    }
-
-    void InvalidatePageEntries(std::vector<Entry*>& entries, VAddr addr, VAddr addr_end) {
-        size_t index = 0;
-        while (index < entries.size()) {
-            Entry* const entry = entries[index];
-            if (!entry->Overlaps(addr, addr_end)) {
-                ++index;
-                continue;
-            }
-
-            UnmarkMemory(entry);
-            RemoveEntryFromInvalidationCache(entry);
-            marked_for_removal.push_back(entry);
-        }
-    }
-
-    void RemoveEntryFromInvalidationCache(const Entry* entry) {
-        const u64 page_end = (entry->addr_end + SUYU_PAGESIZE - 1) >> SUYU_PAGEBITS;
-        for (u64 page = entry->addr_start >> SUYU_PAGEBITS; page < page_end; ++page) {
-            const auto entries_it = invalidation_cache.find(page);
-            ASSERT(entries_it != invalidation_cache.end());
-            std::vector<Entry*>& entries = entries_it->second;
-
-            const auto entry_it = std::find(entries.begin(), entries.end(), entry);
-            ASSERT(entry_it != entries.end());
-            entries.erase(entry_it);
-        }
-    }
-
-    void UnmarkMemory(Entry* entry) {
-        if (!entry->is_memory_marked) {
-            return;
-        }
-        entry->is_memory_marked = false;
-
-        const VAddr addr = entry->addr_start;
-        const size_t size = entry->addr_end - addr;
-        device_memory.UpdatePagesCachedCount(addr, size, -1);
-    }
-
-    void RemoveShadersFromStorage(const std::vector<ShaderInfo*>& removed_shaders) {
-        storage.erase(
-            std::remove_if(storage.begin(), storage.end(),
-                           [&removed_shaders](const std::unique_ptr<ShaderInfo>& shader) {
-                               return std::find(removed_shaders.begin(), removed_shaders.end(),
-                                                shader.get()) != removed_shaders.end();
-                           }),
-            storage.end());
-    }
-
-    Entry* NewEntry(VAddr addr, VAddr addr_end, ShaderInfo* data) {
-        auto entry = std::make_unique<Entry>(Entry{addr, addr_end, data});
-        Entry* const entry_pointer = entry.get();
-
-        lookup_cache.emplace(addr, std::move(entry));
-        return entry_pointer;
-    }
-
-    Tegra::MaxwellDeviceMemoryManager& device_memory;
-    std::vector<std::unique_ptr<ShaderCacheWorker>> workers;
-
-    mutable std::mutex lookup_mutex;
-    std::mutex invalidation_mutex;
-
-    std::unordered_map<VAddr, std::unique_ptr<Entry>> lookup_cache;
-    std::unordered_map<u64, std::vector<Entry*>> invalidation_cache;
-    std::vector<std::unique_ptr<ShaderInfo>> storage;
-    std::vector<Entry*> marked_for_removal;
-};
-
-ShaderCache::ShaderCache(Tegra::MaxwellDeviceMemoryManager& device_memory_)
-    : impl{std::make_unique<Impl>(device_memory_)} {}
-
-ShaderCache::~ShaderCache() = default;
-
 void ShaderCache::InvalidateRegion(VAddr addr, size_t size) {
-    impl->InvalidateRegion(addr, size);
+    std::scoped_lock lock{invalidation_mutex};
+    InvalidatePagesInRegion(addr, size);
+    RemovePendingShaders();
 }
 
 void ShaderCache::OnCacheInvalidation(VAddr addr, size_t size) {
-    impl->OnCacheInvalidation(addr, size);
+    std::scoped_lock lock{invalidation_mutex};
+    InvalidatePagesInRegion(addr, size);
 }
 
 void ShaderCache::SyncGuestHost() {
-    impl->SyncGuestHost();
+    std::scoped_lock lock{invalidation_mutex};
+    RemovePendingShaders();
 }
 
+ShaderCache::ShaderCache(Tegra::MaxwellDeviceMemoryManager& device_memory_)
+    : device_memory{device_memory_} {}
+
 bool ShaderCache::RefreshStages(std::array<u64, 6>& unique_hashes) {
-    return impl->RefreshStages(unique_hashes);
+    auto& dirty{maxwell3d->dirty.flags};
+    if (!dirty[VideoCommon::Dirty::Shaders]) {
+        return last_shaders_valid;
+    }
+    dirty[VideoCommon::Dirty::Shaders] = false;
+
+    const GPUVAddr base_addr{maxwell3d->regs.program_region.Address()};
+    for (size_t index = 0; index < Tegra::Engines::Maxwell3D::Regs::MaxShaderProgram; ++index) {
+        if (!maxwell3d->regs.IsShaderConfigEnabled(index)) {
+            unique_hashes[index] = 0;
+            continue;
+        }
+        const auto& shader_config{maxwell3d->regs.pipelines[index]};
+        const auto program{static_cast<Tegra::Engines::Maxwell3D::Regs::ShaderType>(index)};
+        if (program == Tegra::Engines::Maxwell3D::Regs::ShaderType::Pixel &&
+            !maxwell3d->regs.rasterize_enable) {
+            unique_hashes[index] = 0;
+            continue;
+        }
+        const GPUVAddr shader_addr{base_addr + shader_config.offset};
+        const std::optional<VAddr> cpu_shader_addr{gpu_memory->GpuToCpuAddress(shader_addr)};
+        if (!cpu_shader_addr) {
+            LOG_ERROR(HW_GPU, "Invalid GPU address for shader 0x{:016x}", shader_addr);
+            last_shaders_valid = false;
+            return false;
+        }
+        const ShaderInfo* shader_info{TryGet(*cpu_shader_addr)};
+        if (!shader_info) {
+            const u32 start_address{shader_config.offset};
+            GraphicsEnvironment env{*maxwell3d, *gpu_memory, program, base_addr, start_address};
+            shader_info = MakeShaderInfo(env, *cpu_shader_addr);
+        }
+        shader_infos[index] = shader_info;
+        unique_hashes[index] = shader_info->unique_hash;
+    }
+    last_shaders_valid = true;
+    return true;
 }
 
 const ShaderInfo* ShaderCache::ComputeShader() {
-    return impl->ComputeShader();
+    const GPUVAddr program_base{kepler_compute->regs.code_loc.Address()};
+    const auto& qmd{kepler_compute->launch_description};
+    const GPUVAddr shader_addr{program_base + qmd.program_start};
+    const std::optional<VAddr> cpu_shader_addr{gpu_memory->GpuToCpuAddress(shader_addr)};
+    if (!cpu_shader_addr) {
+        LOG_ERROR(HW_GPU, "Invalid GPU address for shader 0x{:016x}", shader_addr);
+        return nullptr;
+    }
+    if (const ShaderInfo* const shader = TryGet(*cpu_shader_addr)) {
+        return shader;
+    }
+    ComputeEnvironment env{*kepler_compute, *gpu_memory, program_base, qmd.program_start};
+    return MakeShaderInfo(env, *cpu_shader_addr);
 }
 
 void ShaderCache::GetGraphicsEnvironments(GraphicsEnvironments& result,
                                           const std::array<u64, NUM_PROGRAMS>& unique_hashes) {
-    impl->GetGraphicsEnvironments(result, unique_hashes);
+    size_t env_index{};
+    const GPUVAddr base_addr{maxwell3d->regs.program_region.Address()};
+    for (size_t index = 0; index < NUM_PROGRAMS; ++index) {
+        if (unique_hashes[index] == 0) {
+            continue;
+        }
+        const auto program{static_cast<Tegra::Engines::Maxwell3D::Regs::ShaderType>(index)};
+        auto& env{result.envs[index]};
+        const u32 start_address{maxwell3d->regs.pipelines[index].offset};
+        env = GraphicsEnvironment{*maxwell3d, *gpu_memory, program, base_addr, start_address};
+        env.SetCachedSize(shader_infos[index]->size_bytes);
+        result.env_ptrs[env_index++] = &env;
+    }
 }
 
 ShaderInfo* ShaderCache::TryGet(VAddr addr) const {
-    return impl->TryGet(addr);
+    std::scoped_lock lock{lookup_mutex};
+
+    const auto it = lookup_cache.find(addr);
+    if (it == lookup_cache.end()) {
+        return nullptr;
+    }
+    return it->second->data;
 }
 
 void ShaderCache::Register(std::unique_ptr<ShaderInfo> data, VAddr addr, size_t size) {
-    impl->Register(std::move(data), addr, size);
+    std::scoped_lock lock{invalidation_mutex, lookup_mutex};
+
+    const VAddr addr_end = addr + size;
+    Entry* const entry = NewEntry(addr, addr_end, data.get());
+
+    const u64 page_end = (addr_end + SUYU_PAGESIZE - 1) >> SUYU_PAGEBITS;
+    for (u64 page = addr >> SUYU_PAGEBITS; page < page_end; ++page) {
+        invalidation_cache[page].push_back(entry);
+    }
+
+    storage.push_back(std::move(data));
+
+    device_memory.UpdatePagesCachedCount(addr, size, 1);
+}
+
+void ShaderCache::InvalidatePagesInRegion(VAddr addr, size_t size) {
+    const VAddr addr_end = addr + size;
+    const u64 page_end = (addr_end + SUYU_PAGESIZE - 1) >> SUYU_PAGEBITS;
+    for (u64 page = addr >> SUYU_PAGEBITS; page < page_end; ++page) {
+        auto it = invalidation_cache.find(page);
+        if (it == invalidation_cache.end()) {
+            continue;
+        }
+        InvalidatePageEntries(it->second, addr, addr_end);
+    }
+}
+
+void ShaderCache::RemovePendingShaders() {
+    if (marked_for_removal.empty()) {
+        return;
+    }
+    // Remove duplicates
+    std::ranges::sort(marked_for_removal);
+    marked_for_removal.erase(std::unique(marked_for_removal.begin(), marked_for_removal.end()),
+                             marked_for_removal.end());
+
+    boost::container::small_vector<ShaderInfo*, 16> removed_shaders;
+
+    std::scoped_lock lock{lookup_mutex};
+    for (Entry* const entry : marked_for_removal) {
+        removed_shaders.push_back(entry->data);
+
+        const auto it = lookup_cache.find(entry->addr_start);
+        ASSERT(it != lookup_cache.end());
+        lookup_cache.erase(it);
+    }
+    marked_for_removal.clear();
+
+    if (!removed_shaders.empty()) {
+        RemoveShadersFromStorage(removed_shaders);
+    }
+}
+
+void ShaderCache::InvalidatePageEntries(std::vector<Entry*>& entries, VAddr addr, VAddr addr_end) {
+    size_t index = 0;
+    while (index < entries.size()) {
+        Entry* const entry = entries[index];
+        if (!entry->Overlaps(addr, addr_end)) {
+            ++index;
+            continue;
+        }
+
+        UnmarkMemory(entry);
+        RemoveEntryFromInvalidationCache(entry);
+        marked_for_removal.push_back(entry);
+    }
+}
+
+void ShaderCache::RemoveEntryFromInvalidationCache(const Entry* entry) {
+    const u64 page_end = (entry->addr_end + SUYU_PAGESIZE - 1) >> SUYU_PAGEBITS;
+    for (u64 page = entry->addr_start >> SUYU_PAGEBITS; page < page_end; ++page) {
+        const auto entries_it = invalidation_cache.find(page);
+        ASSERT(entries_it != invalidation_cache.end());
+        std::vector<Entry*>& entries = entries_it->second;
+
+        const auto entry_it = std::ranges::find(entries, entry);
+        ASSERT(entry_it != entries.end());
+        entries.erase(entry_it);
+    }
+}
+
+void ShaderCache::UnmarkMemory(Entry* entry) {
+    if (!entry->is_memory_marked) {
+        return;
+    }
+    entry->is_memory_marked = false;
+
+    const VAddr addr = entry->addr_start;
+    const size_t size = entry->addr_end - addr;
+    device_memory.UpdatePagesCachedCount(addr, size, -1);
+}
+
+void ShaderCache::RemoveShadersFromStorage(std::span<ShaderInfo*> removed_shaders) {
+    // Remove them from the cache
+    std::erase_if(storage, [&removed_shaders](const std::unique_ptr<ShaderInfo>& shader) {
+        return std::ranges::find(removed_shaders, shader.get()) != removed_shaders.end();
+    });
+}
+
+ShaderCache::Entry* ShaderCache::NewEntry(VAddr addr, VAddr addr_end, ShaderInfo* data) {
+    auto entry = std::make_unique<Entry>(Entry{addr, addr_end, data});
+    Entry* const entry_pointer = entry.get();
+
+    lookup_cache.emplace(addr, std::move(entry));
+    return entry_pointer;
+}
+
+const ShaderInfo* ShaderCache::MakeShaderInfo(GenericEnvironment& env, VAddr cpu_addr) {
+    auto info = std::make_unique<ShaderInfo>();
+    if (const std::optional<u64> cached_hash{env.Analyze()}) {
+        info->unique_hash = *cached_hash;
+        info->size_bytes = env.CachedSizeBytes();
+    } else {
+        // Slow path, not really hit on commercial games
+        // Build a control flow graph to get the real shader size
+        Shader::ObjectPool<Shader::Maxwell::Flow::Block> flow_block;
+        Shader::Maxwell::Flow::CFG cfg{env, flow_block, env.StartAddress()};
+        info->unique_hash = env.CalculateHash();
+        info->size_bytes = env.ReadSizeBytes();
+    }
+    const size_t size_bytes{info->size_bytes};
+    const ShaderInfo* const result{info.get()};
+    Register(std::move(info), cpu_addr, size_bytes);
+    return result;
 }
 
 } // namespace VideoCommon