"Merge Tagged PR 6598"

.ci/templates/build-msvc.yml

@@ -8,7 +8,7 @@ steps:
   displayName: 'Install vulkan-sdk'
 - script: python -m pip install --upgrade pip conan
   displayName: 'Install conan'
-- script: refreshenv && mkdir build && cd build && cmake -G "Visual Studio 17 2022" -A x64 -DYUZU_USE_BUNDLED_QT=1 -DYUZU_USE_BUNDLED_SDL2=1 -DYUZU_USE_QT_WEB_ENGINE=ON -DENABLE_COMPATIBILITY_LIST_DOWNLOAD=ON -DYUZU_ENABLE_COMPATIBILITY_REPORTING=${COMPAT} -DYUZU_TESTS=OFF -DUSE_DISCORD_PRESENCE=ON -DENABLE_QT_TRANSLATION=ON -DDISPLAY_VERSION=${{ parameters['version'] }} -DCMAKE_BUILD_TYPE=Release .. && cd ..
+- script: refreshenv && mkdir build && cd build && cmake -G "Visual Studio 16 2019" -A x64 -DYUZU_USE_BUNDLED_QT=1 -DYUZU_USE_BUNDLED_SDL2=1 -DYUZU_USE_QT_WEB_ENGINE=ON -DENABLE_COMPATIBILITY_LIST_DOWNLOAD=ON -DYUZU_ENABLE_COMPATIBILITY_REPORTING=${COMPAT} -DYUZU_TESTS=OFF -DUSE_DISCORD_PRESENCE=ON -DENABLE_QT_TRANSLATION=ON -DDISPLAY_VERSION=${{ parameters['version'] }} -DCMAKE_BUILD_TYPE=Release .. && cd ..
   displayName: 'Configure CMake'
 - task: MSBuild@1
   displayName: 'Build'

.ci/yuzu-mainline-step2.yml

@@ -47,7 +47,7 @@ stages:
     timeoutInMinutes: 120
     displayName: 'msvc'
     pool:
-      vmImage: windows-2022
+      vmImage: windows-2019
     steps:
     - template: ./templates/sync-source.yml
       parameters:

.ci/yuzu-patreon-step2.yml

@@ -12,7 +12,7 @@ stages:
     timeoutInMinutes: 120
     displayName: 'windows-msvc'
     pool:
-      vmImage: windows-2022
+      vmImage: windows-2019
     steps:
     - template: ./templates/sync-source.yml
       parameters:

src/audio_core/stream.cpp

@@ -86,6 +86,16 @@ static void VolumeAdjustSamples(std::vector<s16>& samples, float game_volume) {
 }
 
 void Stream::PlayNextBuffer(std::chrono::nanoseconds ns_late) {
+#ifndef _WIN32
+    auto now = std::chrono::steady_clock::now();
+    auto duration = now.time_since_epoch();
+    auto nanoseconds = std::chrono::duration_cast<std::chrono::nanoseconds>(duration);
+
+    if (nanoseconds > expected_cb_time) {
+        ns_late = nanoseconds - expected_cb_time;
+    }
+#endif
+
     if (!IsPlaying()) {
         // Ensure we are in playing state before playing the next buffer
         sink_stream.Flush();
@@ -120,6 +130,9 @@ void Stream::PlayNextBuffer(std::chrono::nanoseconds ns_late) {
         ns_late = {};
     }
 
+#ifndef _WIN32
+    expected_cb_time = nanoseconds + (buffer_release_ns - ns_late);
+#endif
     core_timing.ScheduleEvent(buffer_release_ns - ns_late, release_event, {});
 }
 

src/audio_core/stream.h

@@ -123,6 +123,9 @@ private:
     SinkStream& sink_stream;                ///< Output sink for the stream
     Core::Timing::CoreTiming& core_timing;  ///< Core timing instance.
     std::string name;                       ///< Name of the stream, must be unique
+#ifndef _WIN32
+    std::chrono::nanoseconds expected_cb_time = {}; ///< Estimated time of next callback
+#endif
 };
 
 using StreamPtr = std::shared_ptr<Stream>;

src/core/arm/dynarmic/arm_dynarmic_32.cpp

@@ -125,7 +125,9 @@ public:
     }
 
     void AddTicks(u64 ticks) override {
-        ASSERT_MSG(!parent.uses_wall_clock, "This should never happen - dynarmic ticking disabled");
+        if (parent.uses_wall_clock) {
+            return;
+        }
 
         // Divide the number of ticks by the amount of CPU cores. TODO(Subv): This yields only a
         // rough approximation of the amount of executed ticks in the system, it may be thrown off
@@ -142,7 +144,12 @@ public:
     }
 
     u64 GetTicksRemaining() override {
-        ASSERT_MSG(!parent.uses_wall_clock, "This should never happen - dynarmic ticking disabled");
+        if (parent.uses_wall_clock) {
+            if (!parent.interrupt_handlers[parent.core_index].IsInterrupted()) {
+                return minimum_run_cycles;
+            }
+            return 0U;
+        }
 
         return std::max<s64>(parent.system.CoreTiming().GetDowncount(), 0);
     }
@@ -172,7 +179,7 @@ public:
     Core::Memory::Memory& memory;
     std::size_t num_interpreted_instructions{};
     bool debugger_enabled{};
-    static constexpr u64 minimum_run_cycles = 1000U;
+    static constexpr u64 minimum_run_cycles = 10000U;
 };
 
 std::shared_ptr<Dynarmic::A32::Jit> ARM_Dynarmic_32::MakeJit(Common::PageTable* page_table) const {
@@ -200,7 +207,7 @@ std::shared_ptr<Dynarmic::A32::Jit> ARM_Dynarmic_32::MakeJit(Common::PageTable*
 
     // Timing
     config.wall_clock_cntpct = uses_wall_clock;
-    config.enable_cycle_counting = !uses_wall_clock;
+    config.enable_cycle_counting = true;
 
     // Code cache size
     config.code_cache_size = 512_MiB;

src/core/arm/dynarmic/arm_dynarmic_64.cpp

@@ -166,7 +166,9 @@ public:
     }
 
     void AddTicks(u64 ticks) override {
-        ASSERT_MSG(!parent.uses_wall_clock, "This should never happen - dynarmic ticking disabled");
+        if (parent.uses_wall_clock) {
+            return;
+        }
 
         // Divide the number of ticks by the amount of CPU cores. TODO(Subv): This yields only a
         // rough approximation of the amount of executed ticks in the system, it may be thrown off
@@ -181,7 +183,12 @@ public:
     }
 
     u64 GetTicksRemaining() override {
-        ASSERT_MSG(!parent.uses_wall_clock, "This should never happen - dynarmic ticking disabled");
+        if (parent.uses_wall_clock) {
+            if (!parent.interrupt_handlers[parent.core_index].IsInterrupted()) {
+                return minimum_run_cycles;
+            }
+            return 0U;
+        }
 
         return std::max<s64>(parent.system.CoreTiming().GetDowncount(), 0);
     }
@@ -216,7 +223,7 @@ public:
     u64 tpidrro_el0 = 0;
     u64 tpidr_el0 = 0;
     bool debugger_enabled{};
-    static constexpr u64 minimum_run_cycles = 1000U;
+    static constexpr u64 minimum_run_cycles = 10000U;
 };
 
 std::shared_ptr<Dynarmic::A64::Jit> ARM_Dynarmic_64::MakeJit(Common::PageTable* page_table,
@@ -260,7 +267,7 @@ std::shared_ptr<Dynarmic::A64::Jit> ARM_Dynarmic_64::MakeJit(Common::PageTable*
 
     // Timing
     config.wall_clock_cntpct = uses_wall_clock;
-    config.enable_cycle_counting = !uses_wall_clock;
+    config.enable_cycle_counting = true;
 
     // Code cache size
     config.code_cache_size = 512_MiB;

src/core/core.cpp

@@ -138,6 +138,7 @@ struct System::Impl {
 
         kernel.Suspend(false);
         core_timing.SyncPause(false);
+        cpu_manager.Pause(false);
         is_paused = false;
 
         return status;
@@ -149,6 +150,7 @@ struct System::Impl {
 
         core_timing.SyncPause(true);
         kernel.Suspend(true);
+        cpu_manager.Pause(true);
         is_paused = true;
 
         return status;
@@ -158,6 +160,7 @@ struct System::Impl {
         std::unique_lock<std::mutex> lk(suspend_guard);
         kernel.Suspend(true);
         core_timing.SyncPause(true);
+        cpu_manager.Pause(true);
         return lk;
     }
 
@@ -165,6 +168,7 @@ struct System::Impl {
         if (!is_paused) {
             core_timing.SyncPause(false);
             kernel.Suspend(false);
+            cpu_manager.Pause(false);
         }
     }
 
@@ -330,8 +334,6 @@ struct System::Impl {
             gpu_core->NotifyShutdown();
         }
 
-        kernel.ShutdownCores();
-        cpu_manager.Shutdown();
         debugger.reset();
         services.reset();
         service_manager.reset();

src/core/core_timing.cpp

@@ -61,12 +61,7 @@ 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) {
-        const auto hardware_concurrency = std::thread::hardware_concurrency();
-        size_t id = 0;
-        worker_threads.emplace_back(ThreadEntry, std::ref(*this), id++);
-        if (hardware_concurrency > 8) {
-            worker_threads.emplace_back(ThreadEntry, std::ref(*this), id++);
-        }
+        worker_threads.emplace_back(ThreadEntry, std::ref(*this), 0);
     }
 }
 
@@ -228,14 +223,11 @@ std::optional<s64> CoreTiming::Advance() {
         event_queue.pop_back();
 
         if (const auto event_type{evt.type.lock()}) {
-            sequence_mutex.lock();
+
             event_mutex.unlock();
 
-            event_type->guard.lock();
-            sequence_mutex.unlock();
             const s64 delay = static_cast<s64>(GetGlobalTimeNs().count() - evt.time);
             event_type->callback(evt.user_data, std::chrono::nanoseconds{delay});
-            event_type->guard.unlock();
 
             event_mutex.lock();
             pending_events.fetch_sub(1, std::memory_order_relaxed);

src/core/core_timing.h

@@ -32,7 +32,6 @@ struct EventType {
     TimedCallback callback;
     /// A pointer to the name of the event.
     const std::string name;
-    mutable std::mutex guard;
 };
 
 /**
@@ -157,7 +156,6 @@ private:
     std::condition_variable wait_pause_cv;
     std::condition_variable wait_signal_cv;
     mutable std::mutex event_mutex;
-    mutable std::mutex sequence_mutex;
 
     std::atomic<bool> paused_state{};
     bool is_paused{};

src/core/cpu_manager.cpp

@@ -25,8 +25,10 @@ void CpuManager::ThreadStart(std::stop_token stop_token, CpuManager& cpu_manager
 }
 
 void CpuManager::Initialize() {
+    running_mode = true;
     num_cores = is_multicore ? Core::Hardware::NUM_CPU_CORES : 1;
     gpu_barrier = std::make_unique<Common::Barrier>(num_cores + 1);
+    pause_barrier = std::make_unique<Common::Barrier>(num_cores + 1);
 
     for (std::size_t core = 0; core < num_cores; core++) {
         core_data[core].host_thread = std::jthread(ThreadStart, std::ref(*this), core);
@@ -34,11 +36,8 @@ void CpuManager::Initialize() {
 }
 
 void CpuManager::Shutdown() {
-    for (std::size_t core = 0; core < num_cores; core++) {
-        if (core_data[core].host_thread.joinable()) {
-            core_data[core].host_thread.join();
-        }
-    }
+    running_mode = false;
+    Pause(false);
 }
 
 void CpuManager::GuestThreadFunction() {
@@ -65,10 +64,6 @@ void CpuManager::IdleThreadFunction() {
     }
 }
 
-void CpuManager::ShutdownThreadFunction() {
-    ShutdownThread();
-}
-
 ///////////////////////////////////////////////////////////////////////////////
 ///                             MultiCore                                   ///
 ///////////////////////////////////////////////////////////////////////////////
@@ -181,13 +176,41 @@ void CpuManager::PreemptSingleCore(bool from_running_enviroment) {
     }
 }
 
-void CpuManager::ShutdownThread() {
+void CpuManager::SuspendThread() {
     auto& kernel = system.Kernel();
-    auto core = is_multicore ? kernel.CurrentPhysicalCoreIndex() : 0;
-    auto* current_thread = kernel.GetCurrentEmuThread();
+    kernel.CurrentScheduler()->OnThreadStart();
 
-    Common::Fiber::YieldTo(current_thread->GetHostContext(), *core_data[core].host_context);
-    UNREACHABLE();
+    while (true) {
+        auto core = is_multicore ? kernel.CurrentPhysicalCoreIndex() : 0;
+        auto& scheduler = *kernel.CurrentScheduler();
+        Kernel::KThread* current_thread = scheduler.GetSchedulerCurrentThread();
+        Common::Fiber::YieldTo(current_thread->GetHostContext(), *core_data[core].host_context);
+
+        // This shouldn't be here. This is here because the scheduler needs the current
+        // thread to have dispatch disabled before explicitly rescheduling. Ideally in the
+        // future this will be called by RequestScheduleOnInterrupt and explicitly disabling
+        // dispatch outside the scheduler will not be necessary.
+        current_thread->DisableDispatch();
+
+        scheduler.RescheduleCurrentCore();
+    }
+}
+
+void CpuManager::Pause(bool paused) {
+    std::scoped_lock lk{pause_lock};
+
+    if (pause_state == paused) {
+        return;
+    }
+
+    // Set the new state
+    pause_state.store(paused);
+
+    // Wake up any waiting threads
+    pause_state.notify_all();
+
+    // Wait for all threads to successfully change state before returning
+    pause_barrier->Sync();
 }
 
 void CpuManager::RunThread(std::size_t core) {
@@ -218,9 +241,27 @@ void CpuManager::RunThread(std::size_t core) {
         system.GPU().ObtainContext();
     }
 
-    auto* current_thread = system.Kernel().CurrentScheduler()->GetIdleThread();
-    Kernel::SetCurrentThread(system.Kernel(), current_thread);
-    Common::Fiber::YieldTo(data.host_context, *current_thread->GetHostContext());
+    {
+        // Set the current thread on entry
+        auto* current_thread = system.Kernel().CurrentScheduler()->GetIdleThread();
+        Kernel::SetCurrentThread(system.Kernel(), current_thread);
+    }
+
+    while (running_mode) {
+        if (pause_state.load(std::memory_order_relaxed)) {
+            // Wait for caller to acknowledge pausing
+            pause_barrier->Sync();
+
+            // Wait until unpaused
+            pause_state.wait(true, std::memory_order_relaxed);
+
+            // Wait for caller to acknowledge unpausing
+            pause_barrier->Sync();
+        }
+
+        auto current_thread = system.Kernel().CurrentScheduler()->GetSchedulerCurrentThread();
+        Common::Fiber::YieldTo(data.host_context, *current_thread->GetHostContext());
+    }
 }
 
 } // namespace Core

src/core/cpu_manager.h

@@ -50,14 +50,16 @@ public:
     void Initialize();
     void Shutdown();
 
+    void Pause(bool paused);
+
     std::function<void()> GetGuestThreadStartFunc() {
         return [this] { GuestThreadFunction(); };
     }
     std::function<void()> GetIdleThreadStartFunc() {
         return [this] { IdleThreadFunction(); };
     }
-    std::function<void()> GetShutdownThreadStartFunc() {
-        return [this] { ShutdownThreadFunction(); };
+    std::function<void()> GetSuspendThreadStartFunc() {
+        return [this] { SuspendThread(); };
     }
 
     void PreemptSingleCore(bool from_running_enviroment = true);
@@ -70,7 +72,6 @@ private:
     void GuestThreadFunction();
     void GuestRewindFunction();
     void IdleThreadFunction();
-    void ShutdownThreadFunction();
 
     void MultiCoreRunGuestThread();
     void MultiCoreRunGuestLoop();
@@ -82,7 +83,7 @@ private:
 
     static void ThreadStart(std::stop_token stop_token, CpuManager& cpu_manager, std::size_t core);
 
-    void ShutdownThread();
+    void SuspendThread();
     void RunThread(std::size_t core);
 
     struct CoreData {
@@ -90,7 +91,12 @@ private:
         std::jthread host_thread;
     };
 
+    std::atomic<bool> running_mode{};
+    std::atomic<bool> pause_state{};
+    std::unique_ptr<Common::Barrier> pause_barrier{};
     std::unique_ptr<Common::Barrier> gpu_barrier{};
+    std::mutex pause_lock{};
+
     std::array<CoreData, Core::Hardware::NUM_CPU_CORES> core_data{};
 
     bool is_async_gpu{};

src/core/hle/kernel/k_thread.cpp

@@ -269,7 +269,7 @@ Result KThread::InitializeIdleThread(Core::System& system, KThread* thread, s32
 Result KThread::InitializeHighPriorityThread(Core::System& system, KThread* thread,
                                              KThreadFunction func, uintptr_t arg, s32 virt_core) {
     return InitializeThread(thread, func, arg, {}, {}, virt_core, nullptr, ThreadType::HighPriority,
-                            system.GetCpuManager().GetShutdownThreadStartFunc());
+                            system.GetCpuManager().GetSuspendThreadStartFunc());
 }
 
 Result KThread::InitializeUserThread(Core::System& system, KThread* thread, KThreadFunction func,
@@ -741,19 +741,6 @@ void KThread::Continue() {
     KScheduler::OnThreadStateChanged(kernel, this, old_state);
 }
 
-void KThread::WaitUntilSuspended() {
-    // Make sure we have a suspend requested.
-    ASSERT(IsSuspendRequested());
-
-    // Loop until the thread is not executing on any core.
-    for (std::size_t i = 0; i < static_cast<std::size_t>(Core::Hardware::NUM_CPU_CORES); ++i) {
-        KThread* core_thread{};
-        do {
-            core_thread = kernel.Scheduler(i).GetSchedulerCurrentThread();
-        } while (core_thread == this);
-    }
-}
-
 Result KThread::SetActivity(Svc::ThreadActivity activity) {
     // Lock ourselves.
     KScopedLightLock lk(activity_pause_lock);

src/core/hle/kernel/k_thread.h

@@ -208,8 +208,6 @@ public:
 
     void Continue();
 
-    void WaitUntilSuspended();
-
     constexpr void SetSyncedIndex(s32 index) {
         synced_index = index;
     }

src/core/hle/kernel/kernel.cpp

@@ -76,7 +76,7 @@ struct KernelCore::Impl {
         InitializeMemoryLayout();
         Init::InitializeKPageBufferSlabHeap(system);
         InitializeSchedulers();
-        InitializeShutdownThreads();
+        InitializeSuspendThreads();
         InitializePreemption(kernel);
 
         RegisterHostThread();
@@ -143,9 +143,9 @@ struct KernelCore::Impl {
         CleanupObject(system_resource_limit);
 
         for (u32 core_id = 0; core_id < Core::Hardware::NUM_CPU_CORES; core_id++) {
-            if (shutdown_threads[core_id]) {
-                shutdown_threads[core_id]->Close();
-                shutdown_threads[core_id] = nullptr;
+            if (suspend_threads[core_id]) {
+                suspend_threads[core_id]->Close();
+                suspend_threads[core_id] = nullptr;
             }
 
             schedulers[core_id]->Finalize();
@@ -247,13 +247,13 @@ struct KernelCore::Impl {
         system.CoreTiming().ScheduleEvent(time_interval, preemption_event);
     }
 
-    void InitializeShutdownThreads() {
+    void InitializeSuspendThreads() {
         for (u32 core_id = 0; core_id < Core::Hardware::NUM_CPU_CORES; core_id++) {
-            shutdown_threads[core_id] = KThread::Create(system.Kernel());
-            ASSERT(KThread::InitializeHighPriorityThread(system, shutdown_threads[core_id], {}, {},
+            suspend_threads[core_id] = KThread::Create(system.Kernel());
+            ASSERT(KThread::InitializeHighPriorityThread(system, suspend_threads[core_id], {}, {},
                                                          core_id)
                        .IsSuccess());
-            shutdown_threads[core_id]->SetName(fmt::format("SuspendThread:{}", core_id));
+            suspend_threads[core_id]->SetName(fmt::format("SuspendThread:{}", core_id));
         }
     }
 
@@ -775,7 +775,7 @@ struct KernelCore::Impl {
     std::weak_ptr<ServiceThread> default_service_thread;
     Common::ThreadWorker service_threads_manager;
 
-    std::array<KThread*, Core::Hardware::NUM_CPU_CORES> shutdown_threads;
+    std::array<KThread*, Core::Hardware::NUM_CPU_CORES> suspend_threads;
     std::array<Core::CPUInterruptHandler, Core::Hardware::NUM_CPU_CORES> interrupts{};
     std::array<std::unique_ptr<Kernel::KScheduler>, Core::Hardware::NUM_CPU_CORES> schedulers{};
 
@@ -1085,27 +1085,16 @@ const Kernel::KSharedMemory& KernelCore::GetHidBusSharedMem() const {
 
 void KernelCore::Suspend(bool suspended) {
     const bool should_suspend{exception_exited || suspended};
-    const auto activity = should_suspend ? ProcessActivity::Paused : ProcessActivity::Runnable;
-
-    for (auto* process : GetProcessList()) {
-        process->SetActivity(activity);
-
-        if (should_suspend) {
-            // Wait for execution to stop
-            for (auto* thread : process->GetThreadList()) {
-                thread->WaitUntilSuspended();
-            }
+    const auto state{should_suspend ? ThreadState::Runnable : ThreadState::Waiting};
+    {
+        KScopedSchedulerLock lk{*this};
+        for (auto* thread : impl->suspend_threads) {
+            thread->SetState(state);
+            thread->SetWaitReasonForDebugging(ThreadWaitReasonForDebugging::Suspended);
         }
     }
 }
 
-void KernelCore::ShutdownCores() {
-    for (auto* thread : impl->shutdown_threads) {
-        void(thread->Run());
-    }
-    InterruptAllPhysicalCores();
-}
-
 bool KernelCore::IsMulticore() const {
     return impl->is_multicore;
 }

src/core/hle/kernel/kernel.h

@@ -280,9 +280,6 @@ public:
     /// Exceptional exit all processes.
     void ExceptionalExit();
 
-    /// Notify emulated CPU cores to shut down.
-    void ShutdownCores();
-
     bool IsMulticore() const;
 
     bool IsShuttingDown() const;

src/video_core/gpu_thread.cpp

@@ -31,8 +31,7 @@ static void RunThread(std::stop_token stop_token, Core::System& system,
     VideoCore::RasterizerInterface* const rasterizer = renderer.ReadRasterizer();
 
     while (!stop_token.stop_requested()) {
-        CommandDataContainer next;
-        state.queue.Pop(next, stop_token);
+        CommandDataContainer next = state.queue.PopWait(stop_token);
         if (stop_token.stop_requested()) {
             break;
         }

src/video_core/gpu_thread.h

@@ -10,7 +10,7 @@
 #include <thread>
 #include <variant>
 
-#include "common/bounded_threadsafe_queue.h"
+#include "common/threadsafe_queue.h"
 #include "video_core/framebuffer_config.h"
 
 namespace Tegra {
@@ -96,7 +96,7 @@ struct CommandDataContainer {
 
 /// Struct used to synchronize the GPU thread
 struct SynchState final {
-    using CommandQueue = Common::MPSCQueue<CommandDataContainer>;
+    using CommandQueue = Common::MPSCQueue<CommandDataContainer, true>;
     std::mutex write_lock;
     CommandQueue queue;
     u64 last_fence{};

src/video_core/macro/macro_hle.cpp

@@ -3,6 +3,8 @@
 
 #include <array>
 #include <vector>
+#include "common/scope_exit.h"
+#include "video_core/dirty_flags.h"
 #include "video_core/engines/maxwell_3d.h"
 #include "video_core/macro/macro.h"
 #include "video_core/macro/macro_hle.h"
@@ -58,6 +60,7 @@ void HLE_0217920100488FF7(Engines::Maxwell3D& maxwell3d, const std::vector<u32>&
     maxwell3d.regs.index_array.first = parameters[3];
     maxwell3d.regs.reg_array[0x446] = element_base; // vertex id base?
     maxwell3d.regs.index_array.count = parameters[1];
+    maxwell3d.dirty.flags[VideoCommon::Dirty::IndexBuffer] = true;
     maxwell3d.regs.vb_element_base = element_base;
     maxwell3d.regs.vb_base_instance = base_instance;
     maxwell3d.mme_draw.instance_count = instance_count;
@@ -80,10 +83,67 @@ void HLE_0217920100488FF7(Engines::Maxwell3D& maxwell3d, const std::vector<u32>&
     maxwell3d.mme_draw.current_mode = Engines::Maxwell3D::MMEDrawMode::Undefined;
 }
 
-constexpr std::array<std::pair<u64, HLEFunction>, 3> hle_funcs{{
+// Multidraw Indirect
+void HLE_3f5e74b9c9a50164(Engines::Maxwell3D& maxwell3d, const std::vector<u32>& parameters) {
+    SCOPE_EXIT({
+        // Clean everything.
+        maxwell3d.regs.reg_array[0x446] = 0x0; // vertex id base?
+        maxwell3d.regs.index_array.count = 0;
+        maxwell3d.regs.vb_element_base = 0x0;
+        maxwell3d.regs.vb_base_instance = 0x0;
+        maxwell3d.mme_draw.instance_count = 0;
+        maxwell3d.CallMethodFromMME(0x8e3, 0x640);
+        maxwell3d.CallMethodFromMME(0x8e4, 0x0);
+        maxwell3d.CallMethodFromMME(0x8e5, 0x0);
+        maxwell3d.mme_draw.current_mode = Engines::Maxwell3D::MMEDrawMode::Undefined;
+        maxwell3d.dirty.flags[VideoCommon::Dirty::IndexBuffer] = true;
+    });
+    const u32 start_indirect = parameters[0];
+    const u32 end_indirect = parameters[1];
+    if (start_indirect >= end_indirect) {
+        // Nothing to do.
+        return;
+    }
+    const auto topology =
+        static_cast<Tegra::Engines::Maxwell3D::Regs::PrimitiveTopology>(parameters[2]);
+    maxwell3d.regs.draw.topology.Assign(topology);
+    const u32 padding = parameters[3];
+    const std::size_t max_draws = parameters[4];
+
+    const u32 indirect_words = 5 + padding;
+    const std::size_t first_draw = start_indirect;
+    const std::size_t effective_draws = end_indirect - start_indirect;
+    const std::size_t last_draw = start_indirect + std::min(effective_draws, max_draws);
+
+    for (std::size_t index = first_draw; index < last_draw; index++) {
+        const std::size_t base = index * indirect_words + 5;
+        const u32 num_vertices = parameters[base];
+        const u32 instance_count = parameters[base + 1];
+        const u32 first_index = parameters[base + 2];
+        const u32 base_vertex = parameters[base + 3];
+        const u32 base_instance = parameters[base + 4];
+        maxwell3d.regs.index_array.first = first_index;
+        maxwell3d.regs.reg_array[0x446] = base_vertex;
+        maxwell3d.regs.index_array.count = num_vertices;
+        maxwell3d.regs.vb_element_base = base_vertex;
+        maxwell3d.regs.vb_base_instance = base_instance;
+        maxwell3d.mme_draw.instance_count = instance_count;
+        maxwell3d.CallMethodFromMME(0x8e3, 0x640);
+        maxwell3d.CallMethodFromMME(0x8e4, base_vertex);
+        maxwell3d.CallMethodFromMME(0x8e5, base_instance);
+        maxwell3d.dirty.flags[VideoCommon::Dirty::IndexBuffer] = true;
+        if (maxwell3d.ShouldExecute()) {
+            maxwell3d.Rasterizer().Draw(true, true);
+        }
+        maxwell3d.mme_draw.current_mode = Engines::Maxwell3D::MMEDrawMode::Undefined;
+    }
+}
+
+constexpr std::array<std::pair<u64, HLEFunction>, 4> hle_funcs{{
     {0x771BB18C62444DA0, &HLE_771BB18C62444DA0},
     {0x0D61FC9FAAC9FCAD, &HLE_0D61FC9FAAC9FCAD},
     {0x0217920100488FF7, &HLE_0217920100488FF7},
+    {0x3f5e74b9c9a50164, &HLE_3f5e74b9c9a50164},
 }};
 
 class HLEMacroImpl final : public CachedMacro {
@@ -99,6 +159,7 @@ private:
     Engines::Maxwell3D& maxwell3d;
     HLEFunction func;
 };
+
 } // Anonymous namespace
 
 HLEMacro::HLEMacro(Engines::Maxwell3D& maxwell3d_) : maxwell3d{maxwell3d_} {}