settings: Add option to change shader directory

core: hid: Fix state of capture and home buttons

.lgtm.yml

@@ -1,13 +0,0 @@
-# SPDX-FileCopyrightText: 2020 yuzu Emulator Project
-# SPDX-License-Identifier: GPL-2.0-or-later
-
-path_classifiers:
-  library: "externals"
-extraction:
-  cpp:
-    prepare:
-      packages:
-      - "libsdl2-dev"
-      - "qtmultimedia5-dev"
-      - "libtbb-dev"
-      - "libjack-jackd2-dev"

src/CMakeLists.txt

@@ -113,6 +113,9 @@ else()
 
         $<$<CXX_COMPILER_ID:Clang>:-Wno-braced-scalar-init>
         $<$<CXX_COMPILER_ID:Clang>:-Wno-unused-private-field>
+        $<$<CXX_COMPILER_ID:Clang>:-Werror=shadow-uncaptured-local>
+        $<$<CXX_COMPILER_ID:Clang>:-Werror=implicit-fallthrough>
+        $<$<CXX_COMPILER_ID:Clang>:-Werror=type-limits>
         $<$<CXX_COMPILER_ID:AppleClang>:-Wno-braced-scalar-init>
         $<$<CXX_COMPILER_ID:AppleClang>:-Wno-unused-private-field>
     )

src/audio_core/renderer/system.cpp

@@ -436,10 +436,7 @@ void System::Stop() {
     }
 
     if (execution_mode == ExecutionMode::Auto) {
-        // Should wait for the system to terminate here, but core timing (should have) already
-        // stopped, so this isn't needed. Find a way to make this definite.
-
-        // terminate_event.Wait();
+        terminate_event.Wait();
     }
 }
 

src/common/settings.cpp

@@ -45,6 +45,7 @@ void LogSettings() {
     log_setting("System_LanguageIndex", values.language_index.GetValue());
     log_setting("System_RegionIndex", values.region_index.GetValue());
     log_setting("System_TimeZoneIndex", values.time_zone_index.GetValue());
+    log_setting("System_UnsafeMemoryLayout", values.use_unsafe_extended_memory_layout.GetValue());
     log_setting("Core_UseMultiCore", values.use_multi_core.GetValue());
     log_setting("CPU_Accuracy", values.cpu_accuracy.GetValue());
     log_setting("Renderer_UseResolutionScaling", values.resolution_setup.GetValue());
@@ -73,6 +74,7 @@ void LogSettings() {
     log_path("DataStorage_LoadDir", Common::FS::GetYuzuPath(Common::FS::YuzuPath::LoadDir));
     log_path("DataStorage_NANDDir", Common::FS::GetYuzuPath(Common::FS::YuzuPath::NANDDir));
     log_path("DataStorage_SDMCDir", Common::FS::GetYuzuPath(Common::FS::YuzuPath::SDMCDir));
+    log_path("DataStorage_ShaderDir", Common::FS::GetYuzuPath(Common::FS::YuzuPath::ShaderDir));
     log_setting("Debugging_ProgramArgs", values.program_args.GetValue());
     log_setting("Debugging_GDBStub", values.use_gdbstub.GetValue());
     log_setting("Input_EnableMotion", values.motion_enabled.GetValue());
@@ -191,7 +193,7 @@ void RestoreGlobalState(bool is_powered_on) {
 
     // Core
     values.use_multi_core.SetGlobal(true);
-    values.use_extended_memory_layout.SetGlobal(true);
+    values.use_unsafe_extended_memory_layout.SetGlobal(true);
 
     // CPU
     values.cpu_accuracy.SetGlobal(true);
@@ -226,7 +228,6 @@ void RestoreGlobalState(bool is_powered_on) {
     values.shader_backend.SetGlobal(true);
     values.use_asynchronous_shaders.SetGlobal(true);
     values.use_fast_gpu_time.SetGlobal(true);
-    values.use_pessimistic_flushes.SetGlobal(true);
     values.use_vulkan_driver_pipeline_cache.SetGlobal(true);
     values.bg_red.SetGlobal(true);
     values.bg_green.SetGlobal(true);

src/common/settings.h

@@ -388,7 +388,8 @@ struct Values {
 
     // Core
     SwitchableSetting<bool> use_multi_core{true, "use_multi_core"};
-    SwitchableSetting<bool> use_extended_memory_layout{false, "use_extended_memory_layout"};
+    SwitchableSetting<bool> use_unsafe_extended_memory_layout{false,
+                                                              "use_unsafe_extended_memory_layout"};
 
     // Cpu
     SwitchableSetting<CPUAccuracy, true> cpu_accuracy{CPUAccuracy::Auto, CPUAccuracy::Auto,
@@ -460,7 +461,6 @@ struct Values {
                                                           ShaderBackend::SPIRV, "shader_backend"};
     SwitchableSetting<bool> use_asynchronous_shaders{false, "use_asynchronous_shaders"};
     SwitchableSetting<bool> use_fast_gpu_time{true, "use_fast_gpu_time"};
-    SwitchableSetting<bool> use_pessimistic_flushes{false, "use_pessimistic_flushes"};
     SwitchableSetting<bool> use_vulkan_driver_pipeline_cache{true,
                                                              "use_vulkan_driver_pipeline_cache"};
 

src/core/core.cpp

@@ -137,7 +137,7 @@ struct System::Impl {
         device_memory = std::make_unique<Core::DeviceMemory>();
 
         is_multicore = Settings::values.use_multi_core.GetValue();
-        extended_memory_layout = Settings::values.use_extended_memory_layout.GetValue();
+        extended_memory_layout = Settings::values.use_unsafe_extended_memory_layout.GetValue();
 
         core_timing.SetMulticore(is_multicore);
         core_timing.Initialize([&system]() { system.RegisterHostThread(); });
@@ -169,7 +169,7 @@ struct System::Impl {
     void ReinitializeIfNecessary(System& system) {
         const bool must_reinitialize =
             is_multicore != Settings::values.use_multi_core.GetValue() ||
-            extended_memory_layout != Settings::values.use_extended_memory_layout.GetValue();
+            extended_memory_layout != Settings::values.use_unsafe_extended_memory_layout.GetValue();
 
         if (!must_reinitialize) {
             return;
@@ -178,7 +178,7 @@ struct System::Impl {
         LOG_DEBUG(Kernel, "Re-initializing");
 
         is_multicore = Settings::values.use_multi_core.GetValue();
-        extended_memory_layout = Settings::values.use_extended_memory_layout.GetValue();
+        extended_memory_layout = Settings::values.use_unsafe_extended_memory_layout.GetValue();
 
         Initialize(system);
     }
@@ -293,6 +293,7 @@ struct System::Impl {
         ASSERT(Kernel::KProcess::Initialize(main_process, system, "main",
                                             Kernel::KProcess::ProcessType::Userland, resource_limit)
                    .IsSuccess());
+        Kernel::KProcess::Register(system.Kernel(), main_process);
         kernel.MakeApplicationProcess(main_process);
         const auto [load_result, load_parameters] = app_loader->Load(*main_process, system);
         if (load_result != Loader::ResultStatus::Success) {

src/core/hid/emulated_controller.cpp

@@ -551,6 +551,8 @@ void EmulatedController::EnableSystemButtons() {
 void EmulatedController::DisableSystemButtons() {
     std::scoped_lock lock{mutex};
     system_buttons_enabled = false;
+    controller.home_button_state.raw = 0;
+    controller.capture_button_state.raw = 0;
 }
 
 void EmulatedController::ResetSystemButtons() {
@@ -734,6 +736,8 @@ void EmulatedController::SetButton(const Common::Input::CallbackStatus& callback
     if (is_configuring) {
         controller.npad_button_state.raw = NpadButton::None;
         controller.debug_pad_button_state.raw = 0;
+        controller.home_button_state.raw = 0;
+        controller.capture_button_state.raw = 0;
         lock.unlock();
         TriggerOnChange(ControllerTriggerType::Button, false);
         return;

src/core/hle/kernel/board/nintendo/nx/k_system_control.cpp

@@ -35,12 +35,13 @@ namespace {
 using namespace Common::Literals;
 
 u32 GetMemorySizeForInit() {
-    return Settings::values.use_extended_memory_layout ? Smc::MemorySize_8GB : Smc::MemorySize_4GB;
+    return Settings::values.use_unsafe_extended_memory_layout ? Smc::MemorySize_8GB
+                                                              : Smc::MemorySize_4GB;
 }
 
 Smc::MemoryArrangement GetMemoryArrangeForInit() {
-    return Settings::values.use_extended_memory_layout ? Smc::MemoryArrangement_8GB
-                                                       : Smc::MemoryArrangement_4GB;
+    return Settings::values.use_unsafe_extended_memory_layout ? Smc::MemoryArrangement_8GB
+                                                              : Smc::MemoryArrangement_4GB;
 }
 } // namespace
 

src/core/hle/kernel/k_auto_object.h

@@ -182,8 +182,8 @@ public:
     explicit KAutoObjectWithList(KernelCore& kernel) : KAutoObject(kernel) {}
 
     static int Compare(const KAutoObjectWithList& lhs, const KAutoObjectWithList& rhs) {
-        const u64 lid = lhs.GetId();
-        const u64 rid = rhs.GetId();
+        const uintptr_t lid = reinterpret_cast<uintptr_t>(std::addressof(lhs));
+        const uintptr_t rid = reinterpret_cast<uintptr_t>(std::addressof(rhs));
 
         if (lid < rid) {
             return -1;

src/core/hle/kernel/kernel.cpp

@@ -95,7 +95,7 @@ struct KernelCore::Impl {
                                        pt_heap_region.GetSize());
         }
 
-        InitializeHackSharedMemory();
+        InitializeHackSharedMemory(kernel);
         RegisterHostThread(nullptr);
     }
 
@@ -216,10 +216,12 @@ struct KernelCore::Impl {
             auto* main_thread{Kernel::KThread::Create(system.Kernel())};
             main_thread->SetCurrentCore(core);
             ASSERT(Kernel::KThread::InitializeMainThread(system, main_thread, core).IsSuccess());
+            KThread::Register(system.Kernel(), main_thread);
 
             auto* idle_thread{Kernel::KThread::Create(system.Kernel())};
             idle_thread->SetCurrentCore(core);
             ASSERT(Kernel::KThread::InitializeIdleThread(system, idle_thread, core).IsSuccess());
+            KThread::Register(system.Kernel(), idle_thread);
 
             schedulers[i]->Initialize(main_thread, idle_thread, core);
         }
@@ -230,6 +232,7 @@ struct KernelCore::Impl {
                                        const Core::Timing::CoreTiming& core_timing) {
         system_resource_limit = KResourceLimit::Create(system.Kernel());
         system_resource_limit->Initialize(&core_timing);
+        KResourceLimit::Register(kernel, system_resource_limit);
 
         const auto sizes{memory_layout->GetTotalAndKernelMemorySizes()};
         const auto total_size{sizes.first};
@@ -355,6 +358,7 @@ struct KernelCore::Impl {
             ASSERT(KThread::InitializeHighPriorityThread(system, shutdown_threads[core_id], {}, {},
                                                          core_id)
                        .IsSuccess());
+            KThread::Register(system.Kernel(), shutdown_threads[core_id]);
         }
     }
 
@@ -729,7 +733,7 @@ struct KernelCore::Impl {
         memory_manager->Initialize(management_region.GetAddress(), management_region.GetSize());
     }
 
-    void InitializeHackSharedMemory() {
+    void InitializeHackSharedMemory(KernelCore& kernel) {
         // Setup memory regions for emulated processes
         // TODO(bunnei): These should not be hardcoded regions initialized within the kernel
         constexpr std::size_t hid_size{0x40000};
@@ -746,14 +750,23 @@ struct KernelCore::Impl {
 
         hid_shared_mem->Initialize(system.DeviceMemory(), nullptr, Svc::MemoryPermission::None,
                                    Svc::MemoryPermission::Read, hid_size);
+        KSharedMemory::Register(kernel, hid_shared_mem);
+
         font_shared_mem->Initialize(system.DeviceMemory(), nullptr, Svc::MemoryPermission::None,
                                     Svc::MemoryPermission::Read, font_size);
+        KSharedMemory::Register(kernel, font_shared_mem);
+
         irs_shared_mem->Initialize(system.DeviceMemory(), nullptr, Svc::MemoryPermission::None,
                                    Svc::MemoryPermission::Read, irs_size);
+        KSharedMemory::Register(kernel, irs_shared_mem);
+
         time_shared_mem->Initialize(system.DeviceMemory(), nullptr, Svc::MemoryPermission::None,
                                     Svc::MemoryPermission::Read, time_size);
+        KSharedMemory::Register(kernel, time_shared_mem);
+
         hidbus_shared_mem->Initialize(system.DeviceMemory(), nullptr, Svc::MemoryPermission::None,
                                       Svc::MemoryPermission::Read, hidbus_size);
+        KSharedMemory::Register(kernel, hidbus_shared_mem);
     }
 
     std::mutex registered_objects_lock;
@@ -1072,12 +1085,15 @@ static std::jthread RunHostThreadFunc(KernelCore& kernel, KProcess* process,
     // Commit the thread reservation.
     thread_reservation.Commit();
 
+    // Register the thread.
+    KThread::Register(kernel, thread);
+
     return std::jthread(
         [&kernel, thread, thread_name{std::move(thread_name)}, func{std::move(func)}] {
             // Set the thread name.
             Common::SetCurrentThreadName(thread_name.c_str());
 
-            // Register the thread.
+            // Set the thread as current.
             kernel.RegisterHostThread(thread);
 
             // Run the callback.
@@ -1099,6 +1115,9 @@ std::jthread KernelCore::RunOnHostCoreProcess(std::string&& process_name,
     // Ensure that we don't hold onto any extra references.
     SCOPE_EXIT({ process->Close(); });
 
+    // Register the new process.
+    KProcess::Register(*this, process);
+
     // Run the host thread.
     return RunHostThreadFunc(*this, process, std::move(process_name), std::move(func));
 }
@@ -1124,6 +1143,9 @@ void KernelCore::RunOnGuestCoreProcess(std::string&& process_name, std::function
     // Ensure that we don't hold onto any extra references.
     SCOPE_EXIT({ process->Close(); });
 
+    // Register the new process.
+    KProcess::Register(*this, process);
+
     // Reserve a new thread from the process resource limit.
     KScopedResourceReservation thread_reservation(process, LimitableResource::ThreadCountMax);
     ASSERT(thread_reservation.Succeeded());
@@ -1136,6 +1158,9 @@ void KernelCore::RunOnGuestCoreProcess(std::string&& process_name, std::function
     // Commit the thread reservation.
     thread_reservation.Commit();
 
+    // Register the new thread.
+    KThread::Register(*this, thread);
+
     // Begin running the thread.
     ASSERT(R_SUCCEEDED(thread->Run()));
 }

src/core/hle/service/ipc_helpers.h

@@ -156,6 +156,7 @@ public:
 
             auto* session = Kernel::KSession::Create(kernel);
             session->Initialize(nullptr, 0);
+            Kernel::KSession::Register(kernel, session);
 
             auto next_manager = std::make_shared<Service::SessionRequestManager>(
                 kernel, manager->GetServerManager());

src/core/hle/service/kernel_helpers.cpp

@@ -25,6 +25,9 @@ ServiceContext::ServiceContext(Core::System& system_, std::string name_)
                                         Kernel::KProcess::ProcessType::KernelInternal,
                                         kernel.GetSystemResourceLimit())
                .IsSuccess());
+
+    // Register the process.
+    Kernel::KProcess::Register(kernel, process);
     process_created = true;
 }
 

src/core/hle/service/mutex.cpp

@@ -12,6 +12,9 @@ Mutex::Mutex(Core::System& system) : m_system(system) {
     m_event = Kernel::KEvent::Create(system.Kernel());
     m_event->Initialize(nullptr);
 
+    // Register the event.
+    Kernel::KEvent::Register(system.Kernel(), m_event);
+
     ASSERT(R_SUCCEEDED(m_event->Signal()));
 }
 

src/core/hle/service/server_manager.cpp

@@ -33,6 +33,9 @@ ServerManager::ServerManager(Core::System& system) : m_system{system}, m_serve_m
     // Initialize event.
     m_event = Kernel::KEvent::Create(system.Kernel());
     m_event->Initialize(nullptr);
+
+    // Register event.
+    Kernel::KEvent::Register(system.Kernel(), m_event);
 }
 
 ServerManager::~ServerManager() {
@@ -160,6 +163,9 @@ Result ServerManager::ManageDeferral(Kernel::KEvent** out_event) {
     // Initialize the event.
     m_deferral_event->Initialize(nullptr);
 
+    // Register the event.
+    Kernel::KEvent::Register(m_system.Kernel(), m_deferral_event);
+
     // Set the output.
     *out_event = m_deferral_event;
 

src/core/hle/service/sm/sm.cpp

@@ -64,6 +64,9 @@ Result ServiceManager::RegisterService(std::string name, u32 max_sessions,
     auto* port = Kernel::KPort::Create(kernel);
     port->Initialize(ServerSessionCountMax, false, 0);
 
+    // Register the port.
+    Kernel::KPort::Register(kernel, port);
+
     service_ports.emplace(name, port);
     registered_services.emplace(name, handler);
     if (deferral_event) {

src/core/hle/service/sm/sm_controller.cpp

@@ -49,6 +49,9 @@ void Controller::CloneCurrentObject(HLERequestContext& ctx) {
     // Commit the session reservation.
     session_reservation.Commit();
 
+    // Register the session.
+    Kernel::KSession::Register(system.Kernel(), session);
+
     // Register with server manager.
     session_manager->GetServerManager().RegisterSession(&session->GetServerSession(),
                                                         session_manager);

src/core/memory.cpp

@@ -462,7 +462,7 @@ struct Memory::Impl {
         }
 
         if (Settings::IsFastmemEnabled()) {
-            const bool is_read_enable = Settings::IsGPULevelHigh() || !cached;
+            const bool is_read_enable = !Settings::IsGPULevelExtreme() || !cached;
             system.DeviceMemory().buffer.Protect(vaddr, size, is_read_enable, !cached);
         }
 

src/shader_recompiler/backend/spirv/emit_spirv_context_get_set.cpp

@@ -473,7 +473,8 @@ void EmitSetFragColor(EmitContext& ctx, u32 index, u32 component, Id value) {
 }
 
 void EmitSetSampleMask(EmitContext& ctx, Id value) {
-    ctx.OpStore(ctx.sample_mask, value);
+    const Id pointer{ctx.OpAccessChain(ctx.output_u32, ctx.sample_mask, ctx.u32_zero_value)};
+    ctx.OpStore(pointer, value);
 }
 
 void EmitSetFragDepth(EmitContext& ctx, Id value) {

src/shader_recompiler/backend/spirv/spirv_emit_context.cpp

@@ -1572,7 +1572,8 @@ void EmitContext::DefineOutputs(const IR::Program& program) {
             Decorate(frag_depth, spv::Decoration::BuiltIn, spv::BuiltIn::FragDepth);
         }
         if (info.stores_sample_mask) {
-            sample_mask = DefineOutput(*this, U32[1], std::nullopt);
+            const Id array_type{TypeArray(U32[1], Const(1U))};
+            sample_mask = DefineOutput(*this, array_type, std::nullopt);
             Decorate(sample_mask, spv::Decoration::BuiltIn, spv::BuiltIn::SampleMask);
         }
         break;

src/video_core/buffer_cache/buffer_cache.h

@@ -1426,7 +1426,7 @@ bool BufferCache<P>::SynchronizeBufferNoModified(Buffer& buffer, VAddr cpu_addr,
                 .size = sub_size,
             });
             total_size_bytes += sub_size;
-            largest_copy = std::max(largest_copy, sub_size);
+            largest_copy = std::max<u64>(largest_copy, sub_size);
         }
         const std::span<BufferCopy> copies_span(copies.data(), copies.size());
         UploadMemory(buffer, total_size_bytes, largest_copy, copies_span);

src/video_core/buffer_cache/memory_tracker_base.h

@@ -170,7 +170,8 @@ private:
         std::size_t page_index{cpu_address >> HIGHER_PAGE_BITS};
         u64 page_offset{cpu_address & HIGHER_PAGE_MASK};
         while (remaining_size > 0) {
-            const std::size_t copy_amount{std::min(HIGHER_PAGE_SIZE - page_offset, remaining_size)};
+            const std::size_t copy_amount{
+                std::min<std::size_t>(HIGHER_PAGE_SIZE - page_offset, remaining_size)};
             auto* manager{top_tier[page_index]};
             if (manager) {
                 if constexpr (BOOL_BREAK) {
@@ -206,7 +207,8 @@ private:
         u64 begin = std::numeric_limits<u64>::max();
         u64 end = 0;
         while (remaining_size > 0) {
-            const std::size_t copy_amount{std::min(HIGHER_PAGE_SIZE - page_offset, remaining_size)};
+            const std::size_t copy_amount{
+                std::min<std::size_t>(HIGHER_PAGE_SIZE - page_offset, remaining_size)};
             auto* manager{top_tier[page_index]};
             const auto execute = [&] {
                 auto [new_begin, new_end] = func(manager, page_offset, copy_amount);

src/video_core/compatible_formats.cpp

@@ -126,15 +126,14 @@ constexpr std::array VIEW_CLASS_ASTC_8x8_RGBA{
     PixelFormat::ASTC_2D_8X8_SRGB,
 };
 
-// Missing formats:
-// PixelFormat::ASTC_2D_10X5_UNORM
-// PixelFormat::ASTC_2D_10X5_SRGB
-
-// Missing formats:
-// PixelFormat::ASTC_2D_10X6_SRGB
+constexpr std::array VIEW_CLASS_ASTC_10x5_RGBA{
+    PixelFormat::ASTC_2D_10X5_UNORM,
+    PixelFormat::ASTC_2D_10X5_SRGB,
+};
 
 constexpr std::array VIEW_CLASS_ASTC_10x6_RGBA{
     PixelFormat::ASTC_2D_10X6_UNORM,
+    PixelFormat::ASTC_2D_10X6_SRGB,
 };
 
 constexpr std::array VIEW_CLASS_ASTC_10x8_RGBA{
@@ -147,9 +146,10 @@ constexpr std::array VIEW_CLASS_ASTC_10x10_RGBA{
     PixelFormat::ASTC_2D_10X10_SRGB,
 };
 
-// Missing formats
-// ASTC_2D_12X10_UNORM,
-// ASTC_2D_12X10_SRGB,
+constexpr std::array VIEW_CLASS_ASTC_12x10_RGBA{
+    PixelFormat::ASTC_2D_12X10_UNORM,
+    PixelFormat::ASTC_2D_12X10_SRGB,
+};
 
 constexpr std::array VIEW_CLASS_ASTC_12x12_RGBA{
     PixelFormat::ASTC_2D_12X12_UNORM,
@@ -229,9 +229,11 @@ constexpr Table MakeViewTable() {
     EnableRange(view, VIEW_CLASS_ASTC_6x6_RGBA);
     EnableRange(view, VIEW_CLASS_ASTC_8x5_RGBA);
     EnableRange(view, VIEW_CLASS_ASTC_8x8_RGBA);
+    EnableRange(view, VIEW_CLASS_ASTC_10x5_RGBA);
     EnableRange(view, VIEW_CLASS_ASTC_10x6_RGBA);
     EnableRange(view, VIEW_CLASS_ASTC_10x8_RGBA);
     EnableRange(view, VIEW_CLASS_ASTC_10x10_RGBA);
+    EnableRange(view, VIEW_CLASS_ASTC_12x10_RGBA);
     EnableRange(view, VIEW_CLASS_ASTC_12x12_RGBA);
     return view;
 }

src/video_core/fence_manager.h

@@ -4,13 +4,20 @@
 #pragma once
 
 #include <algorithm>
+#include <condition_variable>
 #include <cstring>
 #include <deque>
 #include <functional>
 #include <memory>
+#include <mutex>
+#include <thread>
 #include <queue>
 
 #include "common/common_types.h"
+#include "common/microprofile.h"
+#include "common/scope_exit.h"
+#include "common/settings.h"
+#include "common/thread.h"
 #include "video_core/delayed_destruction_ring.h"
 #include "video_core/gpu.h"
 #include "video_core/host1x/host1x.h"
@@ -23,15 +30,26 @@ class FenceBase {
 public:
     explicit FenceBase(bool is_stubbed_) : is_stubbed{is_stubbed_} {}
 
+    bool IsStubbed() const {
+        return is_stubbed;
+    }
+
 protected:
     bool is_stubbed;
 };
 
-template <typename TFence, typename TTextureCache, typename TTBufferCache, typename TQueryCache>
+template <typename Traits>
 class FenceManager {
+    using TFence = typename Traits::FenceType;
+    using TTextureCache = typename Traits::TextureCacheType;
+    using TBufferCache = typename Traits::BufferCacheType;
+    using TQueryCache = typename Traits::QueryCacheType;
+    static constexpr bool can_async_check = Traits::HAS_ASYNC_CHECK;
+
 public:
     /// Notify the fence manager about a new frame
     void TickFrame() {
+        std::unique_lock lock(ring_guard);
         delayed_destruction_ring.Tick();
     }
 
@@ -46,17 +64,33 @@ public:
     }
 
     void SignalFence(std::function<void()>&& func) {
-        TryReleasePendingFences();
+        rasterizer.InvalidateGPUCache();
+        bool delay_fence = Settings::IsGPULevelHigh();
+        if constexpr (!can_async_check) {
+            TryReleasePendingFences<false>();
+        }
         const bool should_flush = ShouldFlush();
         CommitAsyncFlushes();
-        uncommitted_operations.emplace_back(std::move(func));
-        CommitOperations();
         TFence new_fence = CreateFence(!should_flush);
-        fences.push(new_fence);
+        if constexpr (can_async_check) {
+            guard.lock();
+        }
+        if (delay_fence) {
+            uncommitted_operations.emplace_back(std::move(func));
+        }
+        pending_operations.emplace_back(std::move(uncommitted_operations));
         QueueFence(new_fence);
+        if (!delay_fence) {
+            func();
+        }
+        fences.push(std::move(new_fence));
         if (should_flush) {
             rasterizer.FlushCommands();
         }
+        if constexpr (can_async_check) {
+            guard.unlock();
+            cv.notify_all();
+        }
     }
 
     void SignalSyncPoint(u32 value) {
@@ -66,29 +100,30 @@ public:
     }
 
     void WaitPendingFences() {
-        while (!fences.empty()) {
-            TFence& current_fence = fences.front();
-            if (ShouldWait()) {
-                WaitFence(current_fence);
-            }
-            PopAsyncFlushes();
-            auto operations = std::move(pending_operations.front());
-            pending_operations.pop_front();
-            for (auto& operation : operations) {
-                operation();
-            }
-            PopFence();
+        if constexpr (!can_async_check) {
+            TryReleasePendingFences<true>();
         }
     }
 
 protected:
     explicit FenceManager(VideoCore::RasterizerInterface& rasterizer_, Tegra::GPU& gpu_,
-                          TTextureCache& texture_cache_, TTBufferCache& buffer_cache_,
+                          TTextureCache& texture_cache_, TBufferCache& buffer_cache_,
                           TQueryCache& query_cache_)
         : rasterizer{rasterizer_}, gpu{gpu_}, syncpoint_manager{gpu.Host1x().GetSyncpointManager()},
-          texture_cache{texture_cache_}, buffer_cache{buffer_cache_}, query_cache{query_cache_} {}
+          texture_cache{texture_cache_}, buffer_cache{buffer_cache_}, query_cache{query_cache_} {
+        if constexpr (can_async_check) {
+            fence_thread =
+                std::jthread([this](std::stop_token token) { ReleaseThreadFunc(token); });
+        }
+    }
 
-    virtual ~FenceManager() = default;
+    virtual ~FenceManager() {
+        if constexpr (can_async_check) {
+            fence_thread.request_stop();
+            cv.notify_all();
+            fence_thread.join();
+        }
+    }
 
     /// Creates a Fence Interface, does not create a backend fence if 'is_stubbed' is
     /// true
@@ -104,15 +139,20 @@ protected:
     Tegra::GPU& gpu;
     Tegra::Host1x::SyncpointManager& syncpoint_manager;
     TTextureCache& texture_cache;
-    TTBufferCache& buffer_cache;
+    TBufferCache& buffer_cache;
     TQueryCache& query_cache;
 
 private:
+    template <bool force_wait>
     void TryReleasePendingFences() {
         while (!fences.empty()) {
             TFence& current_fence = fences.front();
             if (ShouldWait() && !IsFenceSignaled(current_fence)) {
-                return;
+                if constexpr (force_wait) {
+                    WaitFence(current_fence);
+                } else {
+                    return;
+                }
             }
             PopAsyncFlushes();
             auto operations = std::move(pending_operations.front());
@@ -120,7 +160,49 @@ private:
             for (auto& operation : operations) {
                 operation();
             }
-            PopFence();
+            {
+                std::unique_lock lock(ring_guard);
+                delayed_destruction_ring.Push(std::move(current_fence));
+            }
+            fences.pop();
+        }
+    }
+
+    void ReleaseThreadFunc(std::stop_token stop_token) {
+        std::string name = "GPUFencingThread";
+        MicroProfileOnThreadCreate(name.c_str());
+
+        // Cleanup
+        SCOPE_EXIT({ MicroProfileOnThreadExit(); });
+
+        Common::SetCurrentThreadName(name.c_str());
+        Common::SetCurrentThreadPriority(Common::ThreadPriority::High);
+
+        TFence current_fence;
+        std::deque<std::function<void()>> current_operations;
+        while (!stop_token.stop_requested()) {
+            {
+                std::unique_lock lock(guard);
+                cv.wait(lock, [&] { return stop_token.stop_requested() || !fences.empty(); });
+                if (stop_token.stop_requested()) [[unlikely]] {
+                    return;
+                }
+                current_fence = std::move(fences.front());
+                current_operations = std::move(pending_operations.front());
+                fences.pop();
+                pending_operations.pop_front();
+            }
+            if (!current_fence->IsStubbed()) {
+                WaitFence(current_fence);
+            }
+            PopAsyncFlushes();
+            for (auto& operation : current_operations) {
+                operation();
+            }
+            {
+                std::unique_lock lock(ring_guard);
+                delayed_destruction_ring.Push(std::move(current_fence));
+            }
         }
     }
 
@@ -154,19 +236,16 @@ private:
         query_cache.CommitAsyncFlushes();
     }
 
-    void PopFence() {
-        delayed_destruction_ring.Push(std::move(fences.front()));
-        fences.pop();
-    }
-
-    void CommitOperations() {
-        pending_operations.emplace_back(std::move(uncommitted_operations));
-    }
-
     std::queue<TFence> fences;
     std::deque<std::function<void()>> uncommitted_operations;
     std::deque<std::deque<std::function<void()>>> pending_operations;
 
+    std::mutex guard;
+    std::mutex ring_guard;
+    std::condition_variable cv;
+
+    std::jthread fence_thread;
+
     DelayedDestructionRing<TFence, 6> delayed_destruction_ring;
 };
 

src/video_core/memory_manager.cpp

@@ -82,6 +82,7 @@ void MemoryManager::SetEntry(size_t position, MemoryManager::EntryType entry) {
 }
 
 PTEKind MemoryManager::GetPageKind(GPUVAddr gpu_addr) const {
+    std::unique_lock<std::mutex> lock(guard);
     return kind_map.GetValueAt(gpu_addr);
 }
 
@@ -160,7 +161,10 @@ GPUVAddr MemoryManager::BigPageTableOp(GPUVAddr gpu_addr, [[maybe_unused]] VAddr
         }
         remaining_size -= big_page_size;
     }
-    kind_map.Map(gpu_addr, gpu_addr + size, kind);
+    {
+        std::unique_lock<std::mutex> lock(guard);
+        kind_map.Map(gpu_addr, gpu_addr + size, kind);
+    }
     return gpu_addr;
 }
 
@@ -553,6 +557,7 @@ size_t MemoryManager::MaxContinuousRange(GPUVAddr gpu_addr, size_t size) const {
 }
 
 size_t MemoryManager::GetMemoryLayoutSize(GPUVAddr gpu_addr, size_t max_size) const {
+    std::unique_lock<std::mutex> lock(guard);
     return kind_map.GetContinuousSizeFrom(gpu_addr);
 }
 
@@ -745,10 +750,10 @@ void MemoryManager::FlushCaching() {
         return;
     }
     accumulator->Callback([this](GPUVAddr addr, size_t size) {
-        GetSubmappedRangeImpl<false>(addr, size, page_stash);
+        GetSubmappedRangeImpl<false>(addr, size, page_stash2);
     });
-    rasterizer->InnerInvalidation(page_stash);
-    page_stash.clear();
+    rasterizer->InnerInvalidation(page_stash2);
+    page_stash2.clear();
     accumulator->Clear();
 }
 

src/video_core/memory_manager.h

@@ -5,6 +5,7 @@
 
 #include <atomic>
 #include <map>
+#include <mutex>
 #include <optional>
 #include <vector>
 
@@ -215,6 +216,9 @@ private:
 
     std::vector<u64> big_page_continuous;
     std::vector<std::pair<VAddr, std::size_t>> page_stash{};
+    std::vector<std::pair<VAddr, std::size_t>> page_stash2{};
+
+    mutable std::mutex guard;
 
     static constexpr size_t continuous_bits = 64;
 

src/video_core/query_cache.h

@@ -6,6 +6,7 @@
 #include <algorithm>
 #include <array>
 #include <cstring>
+#include <functional>
 #include <iterator>
 #include <list>
 #include <memory>
@@ -17,13 +18,19 @@
 
 #include "common/assert.h"
 #include "common/settings.h"
+#include "core/memory.h"
 #include "video_core/control/channel_state_cache.h"
 #include "video_core/engines/maxwell_3d.h"
 #include "video_core/memory_manager.h"
 #include "video_core/rasterizer_interface.h"
+#include "video_core/texture_cache/slot_vector.h"
 
 namespace VideoCommon {
 
+using AsyncJobId = SlotId;
+
+static constexpr AsyncJobId NULL_ASYNC_JOB_ID{0};
+
 template <class QueryCache, class HostCounter>
 class CounterStreamBase {
 public:
@@ -93,9 +100,13 @@ private:
 template <class QueryCache, class CachedQuery, class CounterStream, class HostCounter>
 class QueryCacheBase : public VideoCommon::ChannelSetupCaches<VideoCommon::ChannelInfo> {
 public:
-    explicit QueryCacheBase(VideoCore::RasterizerInterface& rasterizer_)
-        : rasterizer{rasterizer_}, streams{{CounterStream{static_cast<QueryCache&>(*this),
-                                                          VideoCore::QueryType::SamplesPassed}}} {}
+    explicit QueryCacheBase(VideoCore::RasterizerInterface& rasterizer_,
+                            Core::Memory::Memory& cpu_memory_)
+        : rasterizer{rasterizer_},
+          cpu_memory{cpu_memory_}, streams{{CounterStream{static_cast<QueryCache&>(*this),
+                                                          VideoCore::QueryType::SamplesPassed}}} {
+        (void)slot_async_jobs.insert(); // Null value
+    }
 
     void InvalidateRegion(VAddr addr, std::size_t size) {
         std::unique_lock lock{mutex};
@@ -126,10 +137,15 @@ public:
             query = Register(type, *cpu_addr, host_ptr, timestamp.has_value());
         }
 
-        query->BindCounter(Stream(type).Current(), timestamp);
-        if (Settings::values.use_asynchronous_gpu_emulation.GetValue()) {
-            AsyncFlushQuery(*cpu_addr);
+        auto result = query->BindCounter(Stream(type).Current(), timestamp);
+        if (result) {
+            auto async_job_id = query->GetAsyncJob();
+            auto& async_job = slot_async_jobs[async_job_id];
+            async_job.collected = true;
+            async_job.value = *result;
+            query->SetAsyncJob(NULL_ASYNC_JOB_ID);
         }
+        AsyncFlushQuery(query, timestamp, lock);
     }
 
     /// Updates counters from GPU state. Expected to be called once per draw, clear or dispatch.
@@ -173,15 +189,18 @@ public:
     }
 
     void CommitAsyncFlushes() {
+        std::unique_lock lock{mutex};
         committed_flushes.push_back(uncommitted_flushes);
         uncommitted_flushes.reset();
     }
 
     bool HasUncommittedFlushes() const {
+        std::unique_lock lock{mutex};
         return uncommitted_flushes != nullptr;
     }
 
     bool ShouldWaitAsyncFlushes() const {
+        std::unique_lock lock{mutex};
         if (committed_flushes.empty()) {
             return false;
         }
@@ -189,6 +208,7 @@ public:
     }
 
     void PopAsyncFlushes() {
+        std::unique_lock lock{mutex};
         if (committed_flushes.empty()) {
             return;
         }
@@ -197,15 +217,25 @@ public:
             committed_flushes.pop_front();
             return;
         }
-        for (VAddr query_address : *flush_list) {
-            FlushAndRemoveRegion(query_address, 4);
+        for (AsyncJobId async_job_id : *flush_list) {
+            AsyncJob& async_job = slot_async_jobs[async_job_id];
+            if (!async_job.collected) {
+                FlushAndRemoveRegion(async_job.query_location, 2, true);
+            }
         }
         committed_flushes.pop_front();
     }
 
 private:
+    struct AsyncJob {
+        bool collected = false;
+        u64 value = 0;
+        VAddr query_location = 0;
+        std::optional<u64> timestamp{};
+    };
+
     /// Flushes a memory range to guest memory and removes it from the cache.
-    void FlushAndRemoveRegion(VAddr addr, std::size_t size) {
+    void FlushAndRemoveRegion(VAddr addr, std::size_t size, bool async = false) {
         const u64 addr_begin = addr;
         const u64 addr_end = addr_begin + size;
         const auto in_range = [addr_begin, addr_end](const CachedQuery& query) {
@@ -226,7 +256,16 @@ private:
                     continue;
                 }
                 rasterizer.UpdatePagesCachedCount(query.GetCpuAddr(), query.SizeInBytes(), -1);
-                query.Flush();
+                AsyncJobId async_job_id = query.GetAsyncJob();
+                auto flush_result = query.Flush(async);
+                if (async_job_id == NULL_ASYNC_JOB_ID) {
+                    ASSERT_MSG(false, "This should not be reachable at all");
+                    continue;
+                }
+                AsyncJob& async_job = slot_async_jobs[async_job_id];
+                async_job.collected = true;
+                async_job.value = flush_result;
+                query.SetAsyncJob(NULL_ASYNC_JOB_ID);
             }
             std::erase_if(contents, in_range);
         }
@@ -253,26 +292,60 @@ private:
         return found != std::end(contents) ? &*found : nullptr;
     }
 
-    void AsyncFlushQuery(VAddr addr) {
-        if (!uncommitted_flushes) {
-            uncommitted_flushes = std::make_shared<std::vector<VAddr>>();
+    void AsyncFlushQuery(CachedQuery* query, std::optional<u64> timestamp,
+                         std::unique_lock<std::recursive_mutex>& lock) {
+        const AsyncJobId new_async_job_id = slot_async_jobs.insert();
+        {
+            AsyncJob& async_job = slot_async_jobs[new_async_job_id];
+            query->SetAsyncJob(new_async_job_id);
+            async_job.query_location = query->GetCpuAddr();
+            async_job.collected = false;
+
+            if (!uncommitted_flushes) {
+                uncommitted_flushes = std::make_shared<std::vector<AsyncJobId>>();
+            }
+            uncommitted_flushes->push_back(new_async_job_id);
         }
-        uncommitted_flushes->push_back(addr);
+        lock.unlock();
+        std::function<void()> operation([this, new_async_job_id, timestamp] {
+            std::unique_lock local_lock{mutex};
+            AsyncJob& async_job = slot_async_jobs[new_async_job_id];
+            u64 value = async_job.value;
+            VAddr address = async_job.query_location;
+            slot_async_jobs.erase(new_async_job_id);
+            local_lock.unlock();
+            if (timestamp) {
+                u64 timestamp_value = *timestamp;
+                cpu_memory.WriteBlockUnsafe(address + sizeof(u64), &timestamp_value, sizeof(u64));
+                cpu_memory.WriteBlockUnsafe(address, &value, sizeof(u64));
+                rasterizer.InvalidateRegion(address, sizeof(u64) * 2,
+                                            VideoCommon::CacheType::NoQueryCache);
+            } else {
+                u32 small_value = static_cast<u32>(value);
+                cpu_memory.WriteBlockUnsafe(address, &small_value, sizeof(u32));
+                rasterizer.InvalidateRegion(address, sizeof(u32),
+                                            VideoCommon::CacheType::NoQueryCache);
+            }
+        });
+        rasterizer.SyncOperation(std::move(operation));
     }
 
     static constexpr std::uintptr_t YUZU_PAGESIZE = 4096;
     static constexpr unsigned YUZU_PAGEBITS = 12;
 
-    VideoCore::RasterizerInterface& rasterizer;
+    SlotVector<AsyncJob> slot_async_jobs;
 
-    std::recursive_mutex mutex;
+    VideoCore::RasterizerInterface& rasterizer;
+    Core::Memory::Memory& cpu_memory;
+
+    mutable std::recursive_mutex mutex;
 
     std::unordered_map<u64, std::vector<CachedQuery>> cached_queries;
 
     std::array<CounterStream, VideoCore::NumQueryTypes> streams;
 
-    std::shared_ptr<std::vector<VAddr>> uncommitted_flushes{};
-    std::list<std::shared_ptr<std::vector<VAddr>>> committed_flushes;
+    std::shared_ptr<std::vector<AsyncJobId>> uncommitted_flushes{};
+    std::list<std::shared_ptr<std::vector<AsyncJobId>>> committed_flushes;
 };
 
 template <class QueryCache, class HostCounter>
@@ -291,12 +364,12 @@ public:
     virtual ~HostCounterBase() = default;
 
     /// Returns the current value of the query.
-    u64 Query() {
+    u64 Query(bool async = false) {
         if (result) {
             return *result;
         }
 
-        u64 value = BlockingQuery() + base_result;
+        u64 value = BlockingQuery(async) + base_result;
         if (dependency) {
             value += dependency->Query();
             dependency = nullptr;
@@ -317,7 +390,7 @@ public:
 
 protected:
     /// Returns the value of query from the backend API blocking as needed.
-    virtual u64 BlockingQuery() const = 0;
+    virtual u64 BlockingQuery(bool async = false) const = 0;
 
 private:
     std::shared_ptr<HostCounter> dependency; ///< Counter to add to this value.
@@ -340,26 +413,33 @@ public:
     CachedQueryBase& operator=(const CachedQueryBase&) = delete;
 
     /// Flushes the query to guest memory.
-    virtual void Flush() {
+    virtual u64 Flush(bool async = false) {
         // When counter is nullptr it means that it's just been reset. We are supposed to write a
         // zero in these cases.
-        const u64 value = counter ? counter->Query() : 0;
+        const u64 value = counter ? counter->Query(async) : 0;
+        if (async) {
+            return value;
+        }
         std::memcpy(host_ptr, &value, sizeof(u64));
 
         if (timestamp) {
             std::memcpy(host_ptr + TIMESTAMP_OFFSET, &*timestamp, sizeof(u64));
         }
+        return value;
     }
 
     /// Binds a counter to this query.
-    void BindCounter(std::shared_ptr<HostCounter> counter_, std::optional<u64> timestamp_) {
+    std::optional<u64> BindCounter(std::shared_ptr<HostCounter> counter_,
+                                   std::optional<u64> timestamp_) {
+        std::optional<u64> result{};
         if (counter) {
             // If there's an old counter set it means the query is being rewritten by the game.
             // To avoid losing the data forever, flush here.
-            Flush();
+            result = std::make_optional(Flush());
         }
         counter = std::move(counter_);
         timestamp = timestamp_;
+        return result;
     }
 
     VAddr GetCpuAddr() const noexcept {
@@ -374,6 +454,14 @@ public:
         return with_timestamp ? LARGE_QUERY_SIZE : SMALL_QUERY_SIZE;
     }
 
+    void SetAsyncJob(AsyncJobId assigned_async_job_) {
+        assigned_async_job = assigned_async_job_;
+    }
+
+    AsyncJobId GetAsyncJob() const {
+        return assigned_async_job;
+    }
+
 protected:
     /// Returns true when querying the counter may potentially block.
     bool WaitPending() const noexcept {
@@ -389,6 +477,7 @@ private:
     u8* host_ptr;                         ///< Writable host pointer.
     std::shared_ptr<HostCounter> counter; ///< Host counter to query, owns the dependency tree.
     std::optional<u64> timestamp;         ///< Timestamp to flush to guest memory.
+    AsyncJobId assigned_async_job;
 };
 
 } // namespace VideoCommon

src/video_core/renderer_opengl/gl_fence_manager.h

@@ -30,7 +30,17 @@ private:
 };
 
 using Fence = std::shared_ptr<GLInnerFence>;
-using GenericFenceManager = VideoCommon::FenceManager<Fence, TextureCache, BufferCache, QueryCache>;
+
+struct FenceManagerParams {
+    using FenceType = Fence;
+    using BufferCacheType = BufferCache;
+    using TextureCacheType = TextureCache;
+    using QueryCacheType = QueryCache;
+
+    static constexpr bool HAS_ASYNC_CHECK = false;
+};
+
+using GenericFenceManager = VideoCommon::FenceManager<FenceManagerParams>;
 
 class FenceManagerOpenGL final : public GenericFenceManager {
 public:

src/video_core/renderer_opengl/gl_query_cache.cpp

@@ -26,8 +26,8 @@ constexpr GLenum GetTarget(VideoCore::QueryType type) {
 
 } // Anonymous namespace
 
-QueryCache::QueryCache(RasterizerOpenGL& rasterizer_)
-    : QueryCacheBase(rasterizer_), gl_rasterizer{rasterizer_} {}
+QueryCache::QueryCache(RasterizerOpenGL& rasterizer_, Core::Memory::Memory& cpu_memory_)
+    : QueryCacheBase(rasterizer_, cpu_memory_), gl_rasterizer{rasterizer_} {}
 
 QueryCache::~QueryCache() = default;
 
@@ -74,7 +74,7 @@ void HostCounter::EndQuery() {
     glEndQuery(GetTarget(type));
 }
 
-u64 HostCounter::BlockingQuery() const {
+u64 HostCounter::BlockingQuery([[maybe_unused]] bool async) const {
     GLint64 value;
     glGetQueryObjecti64v(query.handle, GL_QUERY_RESULT, &value);
     return static_cast<u64>(value);
@@ -96,7 +96,7 @@ CachedQuery& CachedQuery::operator=(CachedQuery&& rhs) noexcept {
     return *this;
 }
 
-void CachedQuery::Flush() {
+u64 CachedQuery::Flush([[maybe_unused]] bool async) {
     // Waiting for a query while another query of the same target is enabled locks Nvidia's driver.
     // To avoid this disable and re-enable keeping the dependency stream.
     // But we only have to do this if we have pending waits to be done.
@@ -106,11 +106,13 @@ void CachedQuery::Flush() {
         stream.Update(false);
     }
 
-    VideoCommon::CachedQueryBase<HostCounter>::Flush();
+    auto result = VideoCommon::CachedQueryBase<HostCounter>::Flush();
 
     if (slice_counter) {
         stream.Update(true);
     }
+
+    return result;
 }
 
 } // namespace OpenGL

src/video_core/renderer_opengl/gl_query_cache.h

@@ -28,7 +28,7 @@ using CounterStream = VideoCommon::CounterStreamBase<QueryCache, HostCounter>;
 class QueryCache final
     : public VideoCommon::QueryCacheBase<QueryCache, CachedQuery, CounterStream, HostCounter> {
 public:
-    explicit QueryCache(RasterizerOpenGL& rasterizer_);
+    explicit QueryCache(RasterizerOpenGL& rasterizer_, Core::Memory::Memory& cpu_memory_);
     ~QueryCache();
 
     OGLQuery AllocateQuery(VideoCore::QueryType type);
@@ -51,7 +51,7 @@ public:
     void EndQuery();
 
 private:
-    u64 BlockingQuery() const override;
+    u64 BlockingQuery(bool async = false) const override;
 
     QueryCache& cache;
     const VideoCore::QueryType type;
@@ -70,7 +70,7 @@ public:
     CachedQuery(const CachedQuery&) = delete;
     CachedQuery& operator=(const CachedQuery&) = delete;
 
-    void Flush() override;
+    u64 Flush(bool async = false) override;
 
 private:
     QueryCache* cache;

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -63,7 +63,7 @@ RasterizerOpenGL::RasterizerOpenGL(Core::Frontend::EmuWindow& emu_window_, Tegra
       buffer_cache(*this, cpu_memory_, buffer_cache_runtime),
       shader_cache(*this, emu_window_, device, texture_cache, buffer_cache, program_manager,
                    state_tracker, gpu.ShaderNotify()),
-      query_cache(*this), accelerate_dma(buffer_cache, texture_cache),
+      query_cache(*this, cpu_memory_), accelerate_dma(buffer_cache, texture_cache),
       fence_manager(*this, gpu, texture_cache, buffer_cache, query_cache),
       blit_image(program_manager_) {}
 

src/video_core/renderer_opengl/gl_texture_cache.cpp

@@ -861,9 +861,12 @@ GLuint Image::StorageHandle() noexcept {
     case PixelFormat::ASTC_2D_8X5_SRGB:
     case PixelFormat::ASTC_2D_5X4_SRGB:
     case PixelFormat::ASTC_2D_5X5_SRGB:
+    case PixelFormat::ASTC_2D_10X5_SRGB:
+    case PixelFormat::ASTC_2D_10X6_SRGB:
     case PixelFormat::ASTC_2D_10X8_SRGB:
     case PixelFormat::ASTC_2D_6X6_SRGB:
     case PixelFormat::ASTC_2D_10X10_SRGB:
+    case PixelFormat::ASTC_2D_12X10_SRGB:
     case PixelFormat::ASTC_2D_12X12_SRGB:
     case PixelFormat::ASTC_2D_8X6_SRGB:
     case PixelFormat::ASTC_2D_6X5_SRGB:

src/video_core/renderer_opengl/maxwell_to_gl.h

@@ -100,10 +100,13 @@ constexpr std::array<FormatTuple, VideoCore::Surface::MaxPixelFormat> FORMAT_TAB
     {GL_COMPRESSED_RGBA_ASTC_6x6_KHR},                                // ASTC_2D_6X6_UNORM
     {GL_COMPRESSED_SRGB8_ALPHA8_ASTC_6x6_KHR},                        // ASTC_2D_6X6_SRGB
     {GL_COMPRESSED_RGBA_ASTC_10x6_KHR},                               // ASTC_2D_10X6_UNORM
+    {GL_COMPRESSED_SRGB8_ALPHA8_ASTC_10x6_KHR},                       // ASTC_2D_10X6_SRGB
     {GL_COMPRESSED_RGBA_ASTC_10x5_KHR},                               // ASTC_2D_10X5_UNORM
     {GL_COMPRESSED_SRGB8_ALPHA8_ASTC_10x5_KHR},                       // ASTC_2D_10X5_SRGB
     {GL_COMPRESSED_RGBA_ASTC_10x10_KHR},                              // ASTC_2D_10X10_UNORM
     {GL_COMPRESSED_SRGB8_ALPHA8_ASTC_10x10_KHR},                      // ASTC_2D_10X10_SRGB
+    {GL_COMPRESSED_RGBA_ASTC_12x10_KHR},                              // ASTC_2D_12X10_UNORM
+    {GL_COMPRESSED_SRGB8_ALPHA8_ASTC_12x10_KHR},                      // ASTC_2D_12X10_SRGB
     {GL_COMPRESSED_RGBA_ASTC_12x12_KHR},                              // ASTC_2D_12X12_UNORM
     {GL_COMPRESSED_SRGB8_ALPHA8_ASTC_12x12_KHR},                      // ASTC_2D_12X12_SRGB
     {GL_COMPRESSED_RGBA_ASTC_8x6_KHR},                                // ASTC_2D_8X6_UNORM

src/video_core/renderer_vulkan/maxwell_to_vk.cpp

@@ -197,10 +197,13 @@ struct FormatTuple {
     {VK_FORMAT_ASTC_6x6_UNORM_BLOCK},                          // ASTC_2D_6X6_UNORM
     {VK_FORMAT_ASTC_6x6_SRGB_BLOCK},                           // ASTC_2D_6X6_SRGB
     {VK_FORMAT_ASTC_10x6_UNORM_BLOCK},                         // ASTC_2D_10X6_UNORM
+    {VK_FORMAT_ASTC_10x6_SRGB_BLOCK},                          // ASTC_2D_10X6_SRGB
     {VK_FORMAT_ASTC_10x5_UNORM_BLOCK},                         // ASTC_2D_10X5_UNORM
     {VK_FORMAT_ASTC_10x5_SRGB_BLOCK},                          // ASTC_2D_10X5_SRGB
     {VK_FORMAT_ASTC_10x10_UNORM_BLOCK},                        // ASTC_2D_10X10_UNORM
     {VK_FORMAT_ASTC_10x10_SRGB_BLOCK},                         // ASTC_2D_10X10_SRGB
+    {VK_FORMAT_ASTC_12x10_UNORM_BLOCK},                        // ASTC_2D_12X10_UNORM
+    {VK_FORMAT_ASTC_12x10_SRGB_BLOCK},                         // ASTC_2D_12X10_SRGB
     {VK_FORMAT_ASTC_12x12_UNORM_BLOCK},                        // ASTC_2D_12X12_UNORM
     {VK_FORMAT_ASTC_12x12_SRGB_BLOCK},                         // ASTC_2D_12X12_SRGB
     {VK_FORMAT_ASTC_8x6_UNORM_BLOCK},                          // ASTC_2D_8X6_UNORM

src/video_core/renderer_vulkan/renderer_vulkan.cpp

@@ -134,7 +134,7 @@ void RendererVulkan::SwapBuffers(const Tegra::FramebufferConfig* framebuffer) {
     Frame* frame = present_manager.GetRenderFrame();
     blit_screen.DrawToSwapchain(frame, *framebuffer, use_accelerated, is_srgb);
     scheduler.Flush(*frame->render_ready);
-    scheduler.Record([this, frame](vk::CommandBuffer) { present_manager.PushFrame(frame); });
+    present_manager.Present(frame);
 
     gpu.RendererFrameEndNotify();
     rasterizer.TickFrame();

src/video_core/renderer_vulkan/vk_fence_manager.cpp

@@ -5,6 +5,7 @@
 
 #include "video_core/renderer_vulkan/vk_buffer_cache.h"
 #include "video_core/renderer_vulkan/vk_fence_manager.h"
+#include "video_core/renderer_vulkan/vk_query_cache.h"
 #include "video_core/renderer_vulkan/vk_scheduler.h"
 #include "video_core/renderer_vulkan/vk_texture_cache.h"
 #include "video_core/vulkan_common/vulkan_device.h"

src/video_core/renderer_vulkan/vk_fence_manager.h

@@ -40,7 +40,16 @@ private:
 };
 using Fence = std::shared_ptr<InnerFence>;
 
-using GenericFenceManager = VideoCommon::FenceManager<Fence, TextureCache, BufferCache, QueryCache>;
+struct FenceManagerParams {
+    using FenceType = Fence;
+    using BufferCacheType = BufferCache;
+    using TextureCacheType = TextureCache;
+    using QueryCacheType = QueryCache;
+
+    static constexpr bool HAS_ASYNC_CHECK = true;
+};
+
+using GenericFenceManager = VideoCommon::FenceManager<FenceManagerParams>;
 
 class FenceManager final : public GenericFenceManager {
 public:

src/video_core/renderer_vulkan/vk_present_manager.cpp

@@ -153,16 +153,19 @@ Frame* PresentManager::GetRenderFrame() {
     return frame;
 }
 
-void PresentManager::PushFrame(Frame* frame) {
+void PresentManager::Present(Frame* frame) {
     if (!use_present_thread) {
+        scheduler.WaitWorker();
         CopyToSwapchain(frame);
         free_queue.push(frame);
         return;
     }
 
-    std::unique_lock lock{queue_mutex};
-    present_queue.push(frame);
-    frame_cv.notify_one();
+    scheduler.Record([this, frame](vk::CommandBuffer) {
+        std::unique_lock lock{queue_mutex};
+        present_queue.push(frame);
+        frame_cv.notify_one();
+    });
 }
 
 void PresentManager::RecreateFrame(Frame* frame, u32 width, u32 height, bool is_srgb,

src/video_core/renderer_vulkan/vk_present_manager.h

@@ -45,7 +45,7 @@ public:
     Frame* GetRenderFrame();
 
     /// Pushes a frame for presentation
-    void PushFrame(Frame* frame);
+    void Present(Frame* frame);
 
     /// Recreates the present frame to match the provided parameters
     void RecreateFrame(Frame* frame, u32 width, u32 height, bool is_srgb,

src/video_core/renderer_vulkan/vk_query_cache.cpp

@@ -66,9 +66,10 @@ void QueryPool::Reserve(std::pair<VkQueryPool, u32> query) {
     }
 }
 
-QueryCache::QueryCache(VideoCore::RasterizerInterface& rasterizer_, const Device& device_,
+QueryCache::QueryCache(VideoCore::RasterizerInterface& rasterizer_,
+                       Core::Memory::Memory& cpu_memory_, const Device& device_,
                        Scheduler& scheduler_)
-    : QueryCacheBase{rasterizer_}, device{device_}, scheduler{scheduler_},
+    : QueryCacheBase{rasterizer_, cpu_memory_}, device{device_}, scheduler{scheduler_},
       query_pools{
           QueryPool{device_, scheduler_, QueryType::SamplesPassed},
       } {}
@@ -98,8 +99,10 @@ HostCounter::HostCounter(QueryCache& cache_, std::shared_ptr<HostCounter> depend
       query{cache_.AllocateQuery(type_)}, tick{cache_.GetScheduler().CurrentTick()} {
     const vk::Device* logical = &cache.GetDevice().GetLogical();
     cache.GetScheduler().Record([logical, query = query](vk::CommandBuffer cmdbuf) {
+        const bool use_precise = Settings::IsGPULevelHigh();
         logical->ResetQueryPool(query.first, query.second, 1);
-        cmdbuf.BeginQuery(query.first, query.second, VK_QUERY_CONTROL_PRECISE_BIT);
+        cmdbuf.BeginQuery(query.first, query.second,
+                          use_precise ? VK_QUERY_CONTROL_PRECISE_BIT : 0);
     });
 }
 
@@ -112,8 +115,10 @@ void HostCounter::EndQuery() {
         [query = query](vk::CommandBuffer cmdbuf) { cmdbuf.EndQuery(query.first, query.second); });
 }
 
-u64 HostCounter::BlockingQuery() const {
-    cache.GetScheduler().Wait(tick);
+u64 HostCounter::BlockingQuery(bool async) const {
+    if (!async) {
+        cache.GetScheduler().Wait(tick);
+    }
     u64 data;
     const VkResult query_result = cache.GetDevice().GetLogical().GetQueryResults(
         query.first, query.second, 1, sizeof(data), &data, sizeof(data),

src/video_core/renderer_vulkan/vk_query_cache.h

@@ -52,7 +52,8 @@ private:
 class QueryCache final
     : public VideoCommon::QueryCacheBase<QueryCache, CachedQuery, CounterStream, HostCounter> {
 public:
-    explicit QueryCache(VideoCore::RasterizerInterface& rasterizer_, const Device& device_,
+    explicit QueryCache(VideoCore::RasterizerInterface& rasterizer_,
+                        Core::Memory::Memory& cpu_memory_, const Device& device_,
                         Scheduler& scheduler_);
     ~QueryCache();
 
@@ -83,7 +84,7 @@ public:
     void EndQuery();
 
 private:
-    u64 BlockingQuery() const override;
+    u64 BlockingQuery(bool async = false) const override;
 
     QueryCache& cache;
     const VideoCore::QueryType type;

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -172,7 +172,8 @@ RasterizerVulkan::RasterizerVulkan(Core::Frontend::EmuWindow& emu_window_, Tegra
       buffer_cache(*this, cpu_memory_, buffer_cache_runtime),
       pipeline_cache(*this, device, scheduler, descriptor_pool, update_descriptor_queue,
                      render_pass_cache, buffer_cache, texture_cache, gpu.ShaderNotify()),
-      query_cache{*this, device, scheduler}, accelerate_dma(buffer_cache, texture_cache, scheduler),
+      query_cache{*this, cpu_memory_, device, scheduler},
+      accelerate_dma(buffer_cache, texture_cache, scheduler),
       fence_manager(*this, gpu, texture_cache, buffer_cache, query_cache, device, scheduler),
       wfi_event(device.GetLogical().CreateEvent()) {
     scheduler.SetQueryCache(query_cache);
@@ -675,7 +676,8 @@ bool RasterizerVulkan::AccelerateConditionalRendering() {
     const GPUVAddr condition_address{maxwell3d->regs.render_enable.Address()};
     Maxwell::ReportSemaphore::Compare cmp;
     if (gpu_memory->IsMemoryDirty(condition_address, sizeof(cmp),
-                                  VideoCommon::CacheType::BufferCache)) {
+                                  VideoCommon::CacheType::BufferCache |
+                                      VideoCommon::CacheType::QueryCache)) {
         return true;
     }
     return false;

src/video_core/shader_cache.cpp

@@ -228,14 +228,14 @@ const ShaderInfo* ShaderCache::MakeShaderInfo(GenericEnvironment& env, VAddr cpu
     auto info = std::make_unique<ShaderInfo>();
     if (const std::optional<u64> cached_hash{env.Analyze()}) {
         info->unique_hash = *cached_hash;
-        info->size_bytes = env.CachedSize();
+        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.ReadSize();
+        info->size_bytes = env.ReadSizeBytes();
     }
     const size_t size_bytes{info->size_bytes};
     const ShaderInfo* const result{info.get()};

src/video_core/shader_environment.cpp

@@ -170,15 +170,19 @@ std::optional<u64> GenericEnvironment::Analyze() {
 void GenericEnvironment::SetCachedSize(size_t size_bytes) {
     cached_lowest = start_address;
     cached_highest = start_address + static_cast<u32>(size_bytes);
-    code.resize(CachedSize());
+    code.resize(CachedSizeWords());
     gpu_memory->ReadBlock(program_base + cached_lowest, code.data(), code.size() * sizeof(u64));
 }
 
-size_t GenericEnvironment::CachedSize() const noexcept {
-    return cached_highest - cached_lowest + INST_SIZE;
+size_t GenericEnvironment::CachedSizeWords() const noexcept {
+    return CachedSizeBytes() / INST_SIZE;
 }
 
-size_t GenericEnvironment::ReadSize() const noexcept {
+size_t GenericEnvironment::CachedSizeBytes() const noexcept {
+    return static_cast<size_t>(cached_highest) - cached_lowest + INST_SIZE;
+}
+
+size_t GenericEnvironment::ReadSizeBytes() const noexcept {
     return read_highest - read_lowest + INST_SIZE;
 }
 
@@ -187,7 +191,7 @@ bool GenericEnvironment::CanBeSerialized() const noexcept {
 }
 
 u64 GenericEnvironment::CalculateHash() const {
-    const size_t size{ReadSize()};
+    const size_t size{ReadSizeBytes()};
     const auto data{std::make_unique<char[]>(size)};
     gpu_memory->ReadBlock(program_base + read_lowest, data.get(), size);
     return Common::CityHash64(data.get(), size);
@@ -198,7 +202,7 @@ void GenericEnvironment::Dump(u64 hash) {
 }
 
 void GenericEnvironment::Serialize(std::ofstream& file) const {
-    const u64 code_size{static_cast<u64>(CachedSize())};
+    const u64 code_size{static_cast<u64>(CachedSizeBytes())};
     const u64 num_texture_types{static_cast<u64>(texture_types.size())};
     const u64 num_texture_pixel_formats{static_cast<u64>(texture_pixel_formats.size())};
     const u64 num_cbuf_values{static_cast<u64>(cbuf_values.size())};

src/video_core/shader_environment.h

@@ -48,9 +48,11 @@ public:
 
     void SetCachedSize(size_t size_bytes);
 
-    [[nodiscard]] size_t CachedSize() const noexcept;
+    [[nodiscard]] size_t CachedSizeWords() const noexcept;
 
-    [[nodiscard]] size_t ReadSize() const noexcept;
+    [[nodiscard]] size_t CachedSizeBytes() const noexcept;
+
+    [[nodiscard]] size_t ReadSizeBytes() const noexcept;
 
     [[nodiscard]] bool CanBeSerialized() const noexcept;
 

src/video_core/surface.cpp

@@ -250,10 +250,13 @@ bool IsPixelFormatASTC(PixelFormat format) {
     case PixelFormat::ASTC_2D_6X6_UNORM:
     case PixelFormat::ASTC_2D_6X6_SRGB:
     case PixelFormat::ASTC_2D_10X6_UNORM:
+    case PixelFormat::ASTC_2D_10X6_SRGB:
     case PixelFormat::ASTC_2D_10X5_UNORM:
     case PixelFormat::ASTC_2D_10X5_SRGB:
     case PixelFormat::ASTC_2D_10X10_UNORM:
     case PixelFormat::ASTC_2D_10X10_SRGB:
+    case PixelFormat::ASTC_2D_12X10_UNORM:
+    case PixelFormat::ASTC_2D_12X10_SRGB:
     case PixelFormat::ASTC_2D_12X12_UNORM:
     case PixelFormat::ASTC_2D_12X12_SRGB:
     case PixelFormat::ASTC_2D_8X6_UNORM:
@@ -279,11 +282,13 @@ bool IsPixelFormatSRGB(PixelFormat format) {
     case PixelFormat::ASTC_2D_8X5_SRGB:
     case PixelFormat::ASTC_2D_5X4_SRGB:
     case PixelFormat::ASTC_2D_5X5_SRGB:
+    case PixelFormat::ASTC_2D_10X6_SRGB:
     case PixelFormat::ASTC_2D_10X8_SRGB:
     case PixelFormat::ASTC_2D_6X6_SRGB:
     case PixelFormat::ASTC_2D_10X5_SRGB:
     case PixelFormat::ASTC_2D_10X10_SRGB:
     case PixelFormat::ASTC_2D_12X12_SRGB:
+    case PixelFormat::ASTC_2D_12X10_SRGB:
     case PixelFormat::ASTC_2D_8X6_SRGB:
     case PixelFormat::ASTC_2D_6X5_SRGB:
         return true;

src/video_core/surface.h

@@ -95,10 +95,13 @@ enum class PixelFormat {
     ASTC_2D_6X6_UNORM,
     ASTC_2D_6X6_SRGB,
     ASTC_2D_10X6_UNORM,
+    ASTC_2D_10X6_SRGB,
     ASTC_2D_10X5_UNORM,
     ASTC_2D_10X5_SRGB,
     ASTC_2D_10X10_UNORM,
     ASTC_2D_10X10_SRGB,
+    ASTC_2D_12X10_UNORM,
+    ASTC_2D_12X10_SRGB,
     ASTC_2D_12X12_UNORM,
     ASTC_2D_12X12_SRGB,
     ASTC_2D_8X6_UNORM,
@@ -232,10 +235,13 @@ constexpr std::array<u8, MaxPixelFormat> BLOCK_WIDTH_TABLE = {{
     6,  // ASTC_2D_6X6_UNORM
     6,  // ASTC_2D_6X6_SRGB
     10, // ASTC_2D_10X6_UNORM
+    10, // ASTC_2D_10X6_SRGB
     10, // ASTC_2D_10X5_UNORM
     10, // ASTC_2D_10X5_SRGB
     10, // ASTC_2D_10X10_UNORM
     10, // ASTC_2D_10X10_SRGB
+    12, // ASTC_2D_12X10_UNORM
+    12, // ASTC_2D_12X10_SRGB
     12, // ASTC_2D_12X12_UNORM
     12, // ASTC_2D_12X12_SRGB
     8,  // ASTC_2D_8X6_UNORM
@@ -338,10 +344,13 @@ constexpr std::array<u8, MaxPixelFormat> BLOCK_HEIGHT_TABLE = {{
     6,  // ASTC_2D_6X6_UNORM
     6,  // ASTC_2D_6X6_SRGB
     6,  // ASTC_2D_10X6_UNORM
+    6,  // ASTC_2D_10X6_SRGB
     5,  // ASTC_2D_10X5_UNORM
     5,  // ASTC_2D_10X5_SRGB
     10, // ASTC_2D_10X10_UNORM
     10, // ASTC_2D_10X10_SRGB
+    10, // ASTC_2D_12X10_UNORM
+    10, // ASTC_2D_12X10_SRGB
     12, // ASTC_2D_12X12_UNORM
     12, // ASTC_2D_12X12_SRGB
     6,  // ASTC_2D_8X6_UNORM
@@ -444,10 +453,13 @@ constexpr std::array<u8, MaxPixelFormat> BITS_PER_BLOCK_TABLE = {{
     128, // ASTC_2D_6X6_UNORM
     128, // ASTC_2D_6X6_SRGB
     128, // ASTC_2D_10X6_UNORM
+    128, // ASTC_2D_10X6_SRGB
     128, // ASTC_2D_10X5_UNORM
     128, // ASTC_2D_10X5_SRGB
     128, // ASTC_2D_10X10_UNORM
     128, // ASTC_2D_10X10_SRGB
+    128, // ASTC_2D_12X10_UNORM
+    128, // ASTC_2D_12X10_SRGB
     128, // ASTC_2D_12X12_UNORM
     128, // ASTC_2D_12X12_SRGB
     128, // ASTC_2D_8X6_UNORM

src/video_core/texture_cache/format_lookup_table.cpp

@@ -210,6 +210,8 @@ PixelFormat PixelFormatFromTextureInfo(TextureFormat format, ComponentType red,
         return PixelFormat::ASTC_2D_6X6_SRGB;
     case Hash(TextureFormat::ASTC_2D_10X6, UNORM, LINEAR):
         return PixelFormat::ASTC_2D_10X6_UNORM;
+    case Hash(TextureFormat::ASTC_2D_10X6, UNORM, SRGB):
+        return PixelFormat::ASTC_2D_10X6_SRGB;
     case Hash(TextureFormat::ASTC_2D_10X5, UNORM, LINEAR):
         return PixelFormat::ASTC_2D_10X5_UNORM;
     case Hash(TextureFormat::ASTC_2D_10X5, UNORM, SRGB):
@@ -218,6 +220,10 @@ PixelFormat PixelFormatFromTextureInfo(TextureFormat format, ComponentType red,
         return PixelFormat::ASTC_2D_10X10_UNORM;
     case Hash(TextureFormat::ASTC_2D_10X10, UNORM, SRGB):
         return PixelFormat::ASTC_2D_10X10_SRGB;
+    case Hash(TextureFormat::ASTC_2D_12X10, UNORM, LINEAR):
+        return PixelFormat::ASTC_2D_12X10_UNORM;
+    case Hash(TextureFormat::ASTC_2D_12X10, UNORM, SRGB):
+        return PixelFormat::ASTC_2D_12X10_SRGB;
     case Hash(TextureFormat::ASTC_2D_12X12, UNORM, LINEAR):
         return PixelFormat::ASTC_2D_12X12_UNORM;
     case Hash(TextureFormat::ASTC_2D_12X12, UNORM, SRGB):

src/video_core/texture_cache/formatter.h

@@ -179,6 +179,8 @@ struct fmt::formatter<VideoCore::Surface::PixelFormat> : fmt::formatter<fmt::str
                 return "ASTC_2D_6X6_SRGB";
             case PixelFormat::ASTC_2D_10X6_UNORM:
                 return "ASTC_2D_10X6_UNORM";
+            case PixelFormat::ASTC_2D_10X6_SRGB:
+                return "ASTC_2D_10X6_SRGB";
             case PixelFormat::ASTC_2D_10X5_UNORM:
                 return "ASTC_2D_10X5_UNORM";
             case PixelFormat::ASTC_2D_10X5_SRGB:
@@ -187,6 +189,10 @@ struct fmt::formatter<VideoCore::Surface::PixelFormat> : fmt::formatter<fmt::str
                 return "ASTC_2D_10X10_UNORM";
             case PixelFormat::ASTC_2D_10X10_SRGB:
                 return "ASTC_2D_10X10_SRGB";
+            case PixelFormat::ASTC_2D_12X10_UNORM:
+                return "ASTC_2D_12X10_UNORM";
+            case PixelFormat::ASTC_2D_12X10_SRGB:
+                return "ASTC_2D_12X10_SRGB";
             case PixelFormat::ASTC_2D_12X12_UNORM:
                 return "ASTC_2D_12X12_UNORM";
             case PixelFormat::ASTC_2D_12X12_SRGB:

src/video_core/texture_cache/texture_cache.h

@@ -888,7 +888,7 @@ void TextureCache<P>::DownloadImageIntoBuffer(typename TextureCache<P>::Image* i
             buffer,
             download_map.buffer,
         };
-        std::array buffer_offsets{
+        std::array<u64, 2> buffer_offsets{
             buffer_offset,
             download_map.offset,
         };

src/video_core/vulkan_common/vulkan_device.cpp

@@ -617,7 +617,9 @@ bool Device::ShouldBoostClocks() const {
 
     const bool is_steam_deck = vendor_id == 0x1002 && device_id == 0x163F;
 
-    return validated_driver && !is_steam_deck;
+    const bool is_debugging = this->HasDebuggingToolAttached();
+
+    return validated_driver && !is_steam_deck && !is_debugging;
 }
 
 bool Device::GetSuitability(bool requires_swapchain) {

src/yuzu/configuration/config.cpp

@@ -497,7 +497,7 @@ void Config::ReadCoreValues() {
     qt_config->beginGroup(QStringLiteral("Core"));
 
     ReadGlobalSetting(Settings::values.use_multi_core);
-    ReadGlobalSetting(Settings::values.use_extended_memory_layout);
+    ReadGlobalSetting(Settings::values.use_unsafe_extended_memory_layout);
 
     qt_config->endGroup();
 }
@@ -520,6 +520,13 @@ void Config::ReadDataStorageValues() {
                     QString::fromStdString(FS::GetYuzuPathString(FS::YuzuPath::SDMCDir)))
             .toString()
             .toStdString());
+    FS::SetYuzuPath(
+        FS::YuzuPath::ShaderDir,
+        qt_config
+            ->value(QStringLiteral("shader_directory"),
+                    QString::fromStdString(FS::GetYuzuPathString(FS::YuzuPath::ShaderDir)))
+            .toString()
+            .toStdString());
     FS::SetYuzuPath(
         FS::YuzuPath::LoadDir,
         qt_config
@@ -713,7 +720,6 @@ void Config::ReadRendererValues() {
     ReadGlobalSetting(Settings::values.shader_backend);
     ReadGlobalSetting(Settings::values.use_asynchronous_shaders);
     ReadGlobalSetting(Settings::values.use_fast_gpu_time);
-    ReadGlobalSetting(Settings::values.use_pessimistic_flushes);
     ReadGlobalSetting(Settings::values.use_vulkan_driver_pipeline_cache);
     ReadGlobalSetting(Settings::values.bg_red);
     ReadGlobalSetting(Settings::values.bg_green);
@@ -1162,7 +1168,7 @@ void Config::SaveCoreValues() {
     qt_config->beginGroup(QStringLiteral("Core"));
 
     WriteGlobalSetting(Settings::values.use_multi_core);
-    WriteGlobalSetting(Settings::values.use_extended_memory_layout);
+    WriteGlobalSetting(Settings::values.use_unsafe_extended_memory_layout);
 
     qt_config->endGroup();
 }
@@ -1177,6 +1183,9 @@ void Config::SaveDataStorageValues() {
     WriteSetting(QStringLiteral("sdmc_directory"),
                  QString::fromStdString(FS::GetYuzuPathString(FS::YuzuPath::SDMCDir)),
                  QString::fromStdString(FS::GetYuzuPathString(FS::YuzuPath::SDMCDir)));
+    WriteSetting(QStringLiteral("shader_directory"),
+                 QString::fromStdString(FS::GetYuzuPathString(FS::YuzuPath::ShaderDir)),
+                 QString::fromStdString(FS::GetYuzuPathString(FS::YuzuPath::ShaderDir)));
     WriteSetting(QStringLiteral("load_directory"),
                  QString::fromStdString(FS::GetYuzuPathString(FS::YuzuPath::LoadDir)),
                  QString::fromStdString(FS::GetYuzuPathString(FS::YuzuPath::LoadDir)));
@@ -1359,7 +1368,6 @@ void Config::SaveRendererValues() {
                  Settings::values.shader_backend.UsingGlobal());
     WriteGlobalSetting(Settings::values.use_asynchronous_shaders);
     WriteGlobalSetting(Settings::values.use_fast_gpu_time);
-    WriteGlobalSetting(Settings::values.use_pessimistic_flushes);
     WriteGlobalSetting(Settings::values.use_vulkan_driver_pipeline_cache);
     WriteGlobalSetting(Settings::values.bg_red);
     WriteGlobalSetting(Settings::values.bg_green);

src/yuzu/configuration/configure_filesystem.cpp

@@ -19,6 +19,8 @@ ConfigureFilesystem::ConfigureFilesystem(QWidget* parent)
             [this] { SetDirectory(DirectoryTarget::NAND, ui->nand_directory_edit); });
     connect(ui->sdmc_directory_button, &QToolButton::pressed, this,
             [this] { SetDirectory(DirectoryTarget::SD, ui->sdmc_directory_edit); });
+    connect(ui->shader_directory_button, &QToolButton::pressed, this,
+            [this] { SetDirectory(DirectoryTarget::Shader, ui->shader_directory_edit); });
     connect(ui->gamecard_path_button, &QToolButton::pressed, this,
             [this] { SetDirectory(DirectoryTarget::Gamecard, ui->gamecard_path_edit); });
     connect(ui->dump_path_button, &QToolButton::pressed, this,
@@ -50,6 +52,8 @@ void ConfigureFilesystem::SetConfiguration() {
         QString::fromStdString(Common::FS::GetYuzuPathString(Common::FS::YuzuPath::NANDDir)));
     ui->sdmc_directory_edit->setText(
         QString::fromStdString(Common::FS::GetYuzuPathString(Common::FS::YuzuPath::SDMCDir)));
+    ui->shader_directory_edit->setText(
+        QString::fromStdString(Common::FS::GetYuzuPathString(Common::FS::YuzuPath::ShaderDir)));
     ui->gamecard_path_edit->setText(
         QString::fromStdString(Settings::values.gamecard_path.GetValue()));
     ui->dump_path_edit->setText(
@@ -72,6 +76,8 @@ void ConfigureFilesystem::ApplyConfiguration() {
                             ui->nand_directory_edit->text().toStdString());
     Common::FS::SetYuzuPath(Common::FS::YuzuPath::SDMCDir,
                             ui->sdmc_directory_edit->text().toStdString());
+    Common::FS::SetYuzuPath(Common::FS::YuzuPath::ShaderDir,
+                            ui->shader_directory_edit->text().toStdString());
     Common::FS::SetYuzuPath(Common::FS::YuzuPath::DumpDir,
                             ui->dump_path_edit->text().toStdString());
     Common::FS::SetYuzuPath(Common::FS::YuzuPath::LoadDir,
@@ -95,6 +101,9 @@ void ConfigureFilesystem::SetDirectory(DirectoryTarget target, QLineEdit* edit)
     case DirectoryTarget::SD:
         caption = tr("Select Emulated SD Directory...");
         break;
+    case DirectoryTarget::Shader:
+        caption = tr("Select Shader Directory...");
+        break;
     case DirectoryTarget::Gamecard:
         caption = tr("Select Gamecard Path...");
         break;

src/yuzu/configuration/configure_filesystem.h

@@ -30,6 +30,7 @@ private:
     enum class DirectoryTarget {
         NAND,
         SD,
+        Shader,
         Gamecard,
         Dump,
         Load,

src/yuzu/configuration/configure_filesystem.ui

@@ -25,40 +25,6 @@
         <string>Storage Directories</string>
        </property>
        <layout class="QGridLayout" name="gridLayout">
-        <item row="0" column="0">
-         <widget class="QLabel" name="label">
-          <property name="text">
-           <string>NAND</string>
-          </property>
-         </widget>
-        </item>
-        <item row="0" column="3">
-         <widget class="QToolButton" name="nand_directory_button">
-          <property name="text">
-           <string>...</string>
-          </property>
-         </widget>
-        </item>
-        <item row="0" column="2">
-         <widget class="QLineEdit" name="nand_directory_edit"/>
-        </item>
-        <item row="1" column="2">
-         <widget class="QLineEdit" name="sdmc_directory_edit"/>
-        </item>
-        <item row="1" column="0">
-         <widget class="QLabel" name="label_2">
-          <property name="text">
-           <string>SD Card</string>
-          </property>
-         </widget>
-        </item>
-        <item row="1" column="3">
-         <widget class="QToolButton" name="sdmc_directory_button">
-          <property name="text">
-           <string>...</string>
-          </property>
-         </widget>
-        </item>
         <item row="0" column="1">
          <spacer name="horizontalSpacer">
           <property name="orientation">
@@ -75,6 +41,57 @@
           </property>
          </spacer>
         </item>
+        <item row="0" column="0">
+         <widget class="QLabel" name="label">
+          <property name="text">
+           <string>NAND</string>
+          </property>
+         </widget>
+        </item>
+        <item row="1" column="0">
+         <widget class="QLabel" name="label_2">
+          <property name="text">
+           <string>SD Card</string>
+          </property>
+         </widget>
+        </item>
+        <item row="1" column="2">
+         <widget class="QLineEdit" name="sdmc_directory_edit"/>
+        </item>
+        <item row="1" column="3">
+         <widget class="QToolButton" name="sdmc_directory_button">
+          <property name="text">
+           <string>...</string>
+          </property>
+         </widget>
+        </item>
+        <item row="0" column="3">
+         <widget class="QToolButton" name="nand_directory_button">
+          <property name="text">
+           <string>...</string>
+          </property>
+         </widget>
+        </item>
+        <item row="0" column="2">
+         <widget class="QLineEdit" name="nand_directory_edit"/>
+        </item>
+        <item row="2" column="0">
+         <widget class="QLabel" name="label_4">
+          <property name="text">
+           <string>Shaders</string>
+          </property>
+         </widget>
+        </item>
+        <item row="2" column="2">
+         <widget class="QLineEdit" name="shader_directory_edit"/>
+        </item>
+        <item row="2" column="3">
+         <widget class="QToolButton" name="shader_directory_button">
+          <property name="text">
+           <string>...</string>
+          </property>
+         </widget>
+        </item>
        </layout>
       </widget>
      </item>

src/yuzu/configuration/configure_general.cpp

@@ -35,9 +35,6 @@ void ConfigureGeneral::SetConfiguration() {
 
     ui->use_multi_core->setEnabled(runtime_lock);
     ui->use_multi_core->setChecked(Settings::values.use_multi_core.GetValue());
-    ui->use_extended_memory_layout->setEnabled(runtime_lock);
-    ui->use_extended_memory_layout->setChecked(
-        Settings::values.use_extended_memory_layout.GetValue());
 
     ui->toggle_check_exit->setChecked(UISettings::values.confirm_before_closing.GetValue());
     ui->toggle_user_on_boot->setChecked(UISettings::values.select_user_on_boot.GetValue());
@@ -79,9 +76,6 @@ void ConfigureGeneral::ResetDefaults() {
 void ConfigureGeneral::ApplyConfiguration() {
     ConfigurationShared::ApplyPerGameSetting(&Settings::values.use_multi_core, ui->use_multi_core,
                                              use_multi_core);
-    ConfigurationShared::ApplyPerGameSetting(&Settings::values.use_extended_memory_layout,
-                                             ui->use_extended_memory_layout,
-                                             use_extended_memory_layout);
 
     if (Settings::IsConfiguringGlobal()) {
         UISettings::values.confirm_before_closing = ui->toggle_check_exit->isChecked();
@@ -141,9 +135,6 @@ void ConfigureGeneral::SetupPerGameUI() {
                                             Settings::values.use_speed_limit, use_speed_limit);
     ConfigurationShared::SetColoredTristate(ui->use_multi_core, Settings::values.use_multi_core,
                                             use_multi_core);
-    ConfigurationShared::SetColoredTristate(ui->use_extended_memory_layout,
-                                            Settings::values.use_extended_memory_layout,
-                                            use_extended_memory_layout);
 
     connect(ui->toggle_speed_limit, &QCheckBox::clicked, ui->speed_limit, [this]() {
         ui->speed_limit->setEnabled(ui->toggle_speed_limit->isChecked() &&

src/yuzu/configuration/configure_general.h

@@ -47,7 +47,6 @@ private:
 
     ConfigurationShared::CheckState use_speed_limit;
     ConfigurationShared::CheckState use_multi_core;
-    ConfigurationShared::CheckState use_extended_memory_layout;
 
     const Core::System& system;
 };

src/yuzu/configuration/configure_general.ui

@@ -61,13 +61,6 @@
             </property>
            </widget>
           </item>
-          <item>
-           <widget class="QCheckBox" name="use_extended_memory_layout">
-            <property name="text">
-             <string>Extended memory layout (8GB DRAM)</string>
-            </property>
-           </widget>
-          </item>
           <item>
            <widget class="QCheckBox" name="toggle_check_exit">
             <property name="text">

src/yuzu/configuration/configure_graphics_advanced.cpp

@@ -34,7 +34,6 @@ void ConfigureGraphicsAdvanced::SetConfiguration() {
     ui->async_astc->setChecked(Settings::values.async_astc.GetValue());
     ui->use_asynchronous_shaders->setChecked(Settings::values.use_asynchronous_shaders.GetValue());
     ui->use_fast_gpu_time->setChecked(Settings::values.use_fast_gpu_time.GetValue());
-    ui->use_pessimistic_flushes->setChecked(Settings::values.use_pessimistic_flushes.GetValue());
     ui->use_vulkan_driver_pipeline_cache->setChecked(
         Settings::values.use_vulkan_driver_pipeline_cache.GetValue());
 
@@ -71,8 +70,6 @@ void ConfigureGraphicsAdvanced::ApplyConfiguration() {
                                              use_asynchronous_shaders);
     ConfigurationShared::ApplyPerGameSetting(&Settings::values.use_fast_gpu_time,
                                              ui->use_fast_gpu_time, use_fast_gpu_time);
-    ConfigurationShared::ApplyPerGameSetting(&Settings::values.use_pessimistic_flushes,
-                                             ui->use_pessimistic_flushes, use_pessimistic_flushes);
     ConfigurationShared::ApplyPerGameSetting(&Settings::values.use_vulkan_driver_pipeline_cache,
                                              ui->use_vulkan_driver_pipeline_cache,
                                              use_vulkan_driver_pipeline_cache);
@@ -102,8 +99,6 @@ void ConfigureGraphicsAdvanced::SetupPerGameUI() {
         ui->use_asynchronous_shaders->setEnabled(
             Settings::values.use_asynchronous_shaders.UsingGlobal());
         ui->use_fast_gpu_time->setEnabled(Settings::values.use_fast_gpu_time.UsingGlobal());
-        ui->use_pessimistic_flushes->setEnabled(
-            Settings::values.use_pessimistic_flushes.UsingGlobal());
         ui->use_vulkan_driver_pipeline_cache->setEnabled(
             Settings::values.use_vulkan_driver_pipeline_cache.UsingGlobal());
         ui->anisotropic_filtering_combobox->setEnabled(
@@ -125,9 +120,6 @@ void ConfigureGraphicsAdvanced::SetupPerGameUI() {
                                             use_asynchronous_shaders);
     ConfigurationShared::SetColoredTristate(ui->use_fast_gpu_time,
                                             Settings::values.use_fast_gpu_time, use_fast_gpu_time);
-    ConfigurationShared::SetColoredTristate(ui->use_pessimistic_flushes,
-                                            Settings::values.use_pessimistic_flushes,
-                                            use_pessimistic_flushes);
     ConfigurationShared::SetColoredTristate(ui->use_vulkan_driver_pipeline_cache,
                                             Settings::values.use_vulkan_driver_pipeline_cache,
                                             use_vulkan_driver_pipeline_cache);

src/yuzu/configuration/configure_graphics_advanced.h

@@ -42,7 +42,6 @@ private:
     ConfigurationShared::CheckState async_astc;
     ConfigurationShared::CheckState use_asynchronous_shaders;
     ConfigurationShared::CheckState use_fast_gpu_time;
-    ConfigurationShared::CheckState use_pessimistic_flushes;
     ConfigurationShared::CheckState use_vulkan_driver_pipeline_cache;
 
     const Core::System& system;

src/yuzu/configuration/configure_graphics_advanced.ui

@@ -126,16 +126,6 @@
           </property>
          </widget>
         </item>
-        <item>
-         <widget class="QCheckBox" name="use_pessimistic_flushes">
-          <property name="toolTip">
-           <string>Enables pessimistic buffer flushes. This option will force unmodified buffers to be flushed, which can cost performance.</string>
-          </property>
-          <property name="text">
-           <string>Use pessimistic buffer flushes (Hack)</string>
-          </property>
-         </widget>
-        </item>
         <item>
          <widget class="QCheckBox" name="use_vulkan_driver_pipeline_cache">
           <property name="toolTip">

src/yuzu/configuration/configure_system.cpp

@@ -111,6 +111,9 @@ void ConfigureSystem::SetConfiguration() {
     ui->custom_rtc_edit->setDateTime(QDateTime::fromSecsSinceEpoch(rtc_time));
     ui->device_name_edit->setText(
         QString::fromUtf8(Settings::values.device_name.GetValue().c_str()));
+    ui->use_unsafe_extended_memory_layout->setEnabled(enabled);
+    ui->use_unsafe_extended_memory_layout->setChecked(
+        Settings::values.use_unsafe_extended_memory_layout.GetValue());
 
     if (Settings::IsConfiguringGlobal()) {
         ui->combo_language->setCurrentIndex(Settings::values.language_index.GetValue());
@@ -160,6 +163,9 @@ void ConfigureSystem::ApplyConfiguration() {
     ConfigurationShared::ApplyPerGameSetting(&Settings::values.region_index, ui->combo_region);
     ConfigurationShared::ApplyPerGameSetting(&Settings::values.time_zone_index,
                                              ui->combo_time_zone);
+    ConfigurationShared::ApplyPerGameSetting(&Settings::values.use_unsafe_extended_memory_layout,
+                                             ui->use_unsafe_extended_memory_layout,
+                                             use_unsafe_extended_memory_layout);
 
     if (Settings::IsConfiguringGlobal()) {
         // Guard if during game and set to game-specific value
@@ -215,6 +221,10 @@ void ConfigureSystem::SetupPerGameUI() {
         Settings::values.rng_seed.GetValue().has_value(),
         Settings::values.rng_seed.GetValue(true).has_value(), use_rng_seed);
 
+    ConfigurationShared::SetColoredTristate(ui->use_unsafe_extended_memory_layout,
+                                            Settings::values.use_unsafe_extended_memory_layout,
+                                            use_unsafe_extended_memory_layout);
+
     ui->custom_rtc_checkbox->setVisible(false);
     ui->custom_rtc_edit->setVisible(false);
 }

src/yuzu/configuration/configure_system.h

@@ -41,6 +41,7 @@ private:
     bool enabled = false;
 
     ConfigurationShared::CheckState use_rng_seed;
+    ConfigurationShared::CheckState use_unsafe_extended_memory_layout;
 
     Core::System& system;
 };

src/yuzu/configuration/configure_system.ui

@@ -478,6 +478,13 @@
             </property>
            </widget>
           </item>
+          <item row="7" column="0">
+           <widget class="QCheckBox" name="use_unsafe_extended_memory_layout">
+            <property name="text">
+             <string>Unsafe extended memory layout (8GB DRAM)</string>
+            </property>
+           </widget>
+          </item>
          </layout>
         </item>
        </layout>

src/yuzu/main.cpp

@@ -27,6 +27,7 @@
 #include "configuration/configure_input.h"
 #include "configuration/configure_per_game.h"
 #include "configuration/configure_tas.h"
+#include "core/file_sys/romfs_factory.h"
 #include "core/file_sys/vfs.h"
 #include "core/file_sys/vfs_real.h"
 #include "core/frontend/applets/cabinet.h"
@@ -4171,6 +4172,8 @@ void GMainWindow::OnReinitializeKeys(ReinitializeKeyBehavior behavior) {
     }
 
     Core::Crypto::KeyManager& keys = Core::Crypto::KeyManager::Instance();
+    bool all_keys_present{true};
+
     if (keys.BaseDeriveNecessary()) {
         Core::Crypto::PartitionDataManager pdm{vfs->OpenDirectory("", FileSys::Mode::Read)};
 
@@ -4195,6 +4198,7 @@ void GMainWindow::OnReinitializeKeys(ReinitializeKeyBehavior behavior) {
             errors += tr(" - Missing PRODINFO");
         }
         if (!errors.isEmpty()) {
+            all_keys_present = false;
             QMessageBox::warning(
                 this, tr("Derivation Components Missing"),
                 tr("Encryption keys are missing. "
@@ -4222,11 +4226,40 @@ void GMainWindow::OnReinitializeKeys(ReinitializeKeyBehavior behavior) {
 
     system->GetFileSystemController().CreateFactories(*vfs);
 
+    if (all_keys_present && !this->CheckSystemArchiveDecryption()) {
+        LOG_WARNING(Frontend, "Mii model decryption failed");
+        QMessageBox::warning(
+            this, tr("System Archive Decryption Failed"),
+            tr("Encryption keys failed to decrypt firmware. "
+               "<br>Please follow <a href='https://yuzu-emu.org/help/quickstart/'>the yuzu "
+               "quickstart guide</a> to get all your keys, firmware and "
+               "games."));
+    }
+
     if (behavior == ReinitializeKeyBehavior::Warning) {
         game_list->PopulateAsync(UISettings::values.game_dirs);
     }
 }
 
+bool GMainWindow::CheckSystemArchiveDecryption() {
+    constexpr u64 MiiModelId = 0x0100000000000802;
+
+    auto bis_system = system->GetFileSystemController().GetSystemNANDContents();
+    if (!bis_system) {
+        // Not having system BIS files is not an error.
+        return true;
+    }
+
+    auto mii_nca = bis_system->GetEntry(MiiModelId, FileSys::ContentRecordType::Data);
+    if (!mii_nca) {
+        // Not having the Mii model is not an error.
+        return true;
+    }
+
+    // Return whether we are able to decrypt the RomFS of the Mii model.
+    return mii_nca->GetRomFS().get() != nullptr;
+}
+
 std::optional<u64> GMainWindow::SelectRomFSDumpTarget(const FileSys::ContentProvider& installed,
                                                       u64 program_id) {
     const auto dlc_entries =

src/yuzu/main.h

@@ -392,6 +392,7 @@ private:
     void LoadTranslation();
     void OpenPerGameConfiguration(u64 title_id, const std::string& file_name);
     bool CheckDarkMode();
+    bool CheckSystemArchiveDecryption();
 
     QString GetTasStateDescription() const;
     bool CreateShortcut(const std::string& shortcut_path, const std::string& title,

src/yuzu_cmd/config.cpp

@@ -274,7 +274,7 @@ void Config::ReadValues() {
 
     // Core
     ReadSetting("Core", Settings::values.use_multi_core);
-    ReadSetting("Core", Settings::values.use_extended_memory_layout);
+    ReadSetting("Core", Settings::values.use_unsafe_extended_memory_layout);
 
     // Cpu
     ReadSetting("Cpu", Settings::values.cpu_accuracy);
@@ -327,7 +327,6 @@ void Config::ReadValues() {
     ReadSetting("Renderer", Settings::values.accelerate_astc);
     ReadSetting("Renderer", Settings::values.async_astc);
     ReadSetting("Renderer", Settings::values.use_fast_gpu_time);
-    ReadSetting("Renderer", Settings::values.use_pessimistic_flushes);
     ReadSetting("Renderer", Settings::values.use_vulkan_driver_pipeline_cache);
 
     ReadSetting("Renderer", Settings::values.bg_red);

src/yuzu_cmd/default_ini.h

@@ -163,9 +163,9 @@ keyboard_enabled =
 # 0: Disabled, 1 (default): Enabled
 use_multi_core =
 
-# Enable extended guest system memory layout (8GB DRAM)
+# Enable unsafe extended guest system memory layout (8GB DRAM)
 # 0 (default): Disabled, 1: Enabled
-use_extended_memory_layout =
+use_unsafe_extended_memory_layout =
 
 [Cpu]
 # Adjusts various optimizations.
@@ -374,10 +374,6 @@ use_asynchronous_gpu_emulation =
 # 0: Off, 1 (default): On
 use_fast_gpu_time =
 
-# Force unmodified buffers to be flushed, which can cost performance.
-# 0: Off (default), 1: On
-use_pessimistic_flushes =
-
 # Whether to use garbage collection or not for GPU caches.
 # 0 (default): Off, 1: On
 use_caches_gc =

vcpkg.json

@@ -49,7 +49,7 @@
     "overrides": [
         {
             "name": "catch2",
-            "version": "3.0.1"
+            "version": "3.3.1"
         },
         {
             "name": "fmt",