Make the ExeFS dumper take updates into account

maxwell_dma: Match official doc and support pitch->voxel copies

.github/ISSUE_TEMPLATE/bug-report-feature-request.md

@@ -1,4 +1,13 @@
-<!--
+---
+name: Bug Report / Feature Request
+about: Tech support does not belong here. You should only file an issue here if you think you have experienced an actual bug with yuzu or you are requesting a feature you believe would make yuzu better.
+title: ''
+labels: ''
+assignees: ''
+
+---
+
+<!---
 Please keep in mind yuzu is EXPERIMENTAL SOFTWARE.
 
 Please read the FAQ:

.github/ISSUE_TEMPLATE/config.yml

@@ -0,0 +1,8 @@
+blank_issues_enabled: false
+contact_links:
+  - name: yuzu Discord
+    url: https://discord.com/invite/u77vRWY
+    about: If you are experiencing an issue with yuzu, and you need tech support, or if you have a general question, try asking in the official yuzu Discord linked here. Piracy is not allowed.
+  - name: Community forums
+    url: https://community.citra-emu.org
+    about: This is an alternative place for tech support, however helpers there are not as active.

CMakeLists.txt

@@ -118,8 +118,17 @@ message(STATUS "Target architecture: ${ARCHITECTURE}")
 # Configure C++ standard
 # ===========================
 
-set(CMAKE_CXX_STANDARD 17)
-set(CMAKE_CXX_STANDARD_REQUIRED ON)
+# boost asio's concept usage doesn't play nicely with some compilers yet.
+add_definitions(-DBOOST_ASIO_DISABLE_CONCEPTS)
+if (MSVC)
+    add_compile_options(/std:c++latest)
+
+    # cubeb and boost still make use of deprecated result_of.
+    add_definitions(-D_HAS_DEPRECATED_RESULT_OF)
+else()
+    set(CMAKE_CXX_STANDARD 20)
+    set(CMAKE_CXX_STANDARD_REQUIRED ON)
+endif()
 
 # Output binaries to bin/
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${PROJECT_BINARY_DIR}/bin)
@@ -151,7 +160,7 @@ macro(yuzu_find_packages)
     #    Cmake Pkg Prefix  Version     Conan Pkg
         "Boost             1.71        boost/1.72.0"
         "Catch2            2.11        catch2/2.11.0"
-        "fmt               6.2         fmt/6.2.0"
+        "fmt               7.0         fmt/7.0.1"
     # can't use until https://github.com/bincrafters/community/issues/1173
         #"libzip            1.5         libzip/1.5.2@bincrafters/stable"
         "lz4               1.8         lz4/1.9.2"
@@ -211,7 +220,7 @@ if(ENABLE_QT)
 
         set(QT_PREFIX_HINT HINTS "${QT_PREFIX}")
     endif()
-    find_package(Qt5 5.9 COMPONENTS Widgets OpenGL ${QT_PREFIX_HINT})
+    find_package(Qt5 5.9 COMPONENTS Widgets ${QT_PREFIX_HINT})
     if (YUZU_USE_QT_WEB_ENGINE)
         find_package(Qt5 COMPONENTS WebEngineCore WebEngineWidgets)
     endif()
@@ -287,7 +296,7 @@ if (CONAN_REQUIRED_LIBS)
     if(ENABLE_QT)
         list(APPEND CMAKE_MODULE_PATH "${CONAN_QT_ROOT_RELEASE}")
         list(APPEND CMAKE_PREFIX_PATH "${CONAN_QT_ROOT_RELEASE}")
-        find_package(Qt5 5.9 REQUIRED COMPONENTS Widgets OpenGL)
+        find_package(Qt5 5.9 REQUIRED COMPONENTS Widgets)
         if (YUZU_USE_QT_WEB_ENGINE)
             find_package(Qt5 REQUIRED COMPONENTS WebEngineCore WebEngineWidgets)
         endif()
@@ -330,10 +339,15 @@ elseif(SDL2_FOUND)
 endif()
 
 # Ensure libusb is properly configured (based on dolphin libusb include)
-find_package(LibUSB)
-add_subdirectory(externals/libusb)
-set(LIBUSB_LIBRARIES usb)
-
+if(NOT APPLE)
+    include(FindPkgConfig)
+    find_package(LibUSB)
+endif()
+if (NOT LIBUSB_FOUND)
+    add_subdirectory(externals/libusb)
+    set(LIBUSB_INCLUDE_DIR "")
+    set(LIBUSB_LIBRARIES usb)
+endif()
 
 # Prefer the -pthread flag on Linux.
 set(THREADS_PREFER_PTHREAD_FLAG ON)

CMakeModules/CopyYuzuQt5Deps.cmake

@@ -15,7 +15,6 @@ function(copy_yuzu_Qt5_deps target_dir)
         icuuc*.dll
         Qt5Core$<$<CONFIG:Debug>:d>.*
         Qt5Gui$<$<CONFIG:Debug>:d>.*
-        Qt5OpenGL$<$<CONFIG:Debug>:d>.*
         Qt5Widgets$<$<CONFIG:Debug>:d>.*
     )
 

README.md

@@ -2,7 +2,7 @@ yuzu emulator
 =============
 [![Travis CI Build Status](https://travis-ci.com/yuzu-emu/yuzu.svg?branch=master)](https://travis-ci.com/yuzu-emu/yuzu)
 [![Azure Mainline CI Build Status](https://dev.azure.com/yuzu-emu/yuzu/_apis/build/status/yuzu%20mainline?branchName=master)](https://dev.azure.com/yuzu-emu/yuzu/)
-[![Discord](https://img.shields.io/discord/398318088170242053?color=%237289DA&label=yuzu&logo=discord&logoColor=white)](https://discord.gg/XQV6dn9)
+[![Discord](https://img.shields.io/discord/398318088170242053?color=%237289DA&label=yuzu&logo=discord&logoColor=white)](https://discord.com/invite/u77vRWY)
 
 yuzu is an experimental open-source emulator for the Nintendo Switch from the creators of [Citra](https://citra-emu.org/).
 
@@ -16,7 +16,7 @@ yuzu is licensed under the GPLv2 (or any later version). Refer to the license.tx
 
 Check out our [website](https://yuzu-emu.org/)!
 
-For development discussion, please join us on [Discord](https://discord.gg/XQV6dn9).
+For development discussion, please join us on [Discord](https://discord.com/invite/u77vRWY).
 
 ### Development
 

externals/find-modules/FindLibUSB.cmake

@@ -0,0 +1,43 @@
+# - Find libusb-1.0 library
+# This module defines
+#  LIBUSB_INCLUDE_DIR, where to find bluetooth.h
+#  LIBUSB_LIBRARIES, the libraries needed to use libusb-1.0.
+#  LIBUSB_FOUND, If false, do not try to use libusb-1.0.
+#
+# Copyright (c) 2009, Michal Cihar, <michal@cihar.com>
+#
+# vim: expandtab sw=4 ts=4 sts=4:
+
+if(ANDROID)
+       set(LIBUSB_FOUND FALSE CACHE INTERNAL "libusb-1.0 found")
+       message(STATUS "libusb-1.0 not found.")
+elseif (NOT LIBUSB_FOUND)
+    pkg_check_modules (LIBUSB_PKG libusb-1.0)
+
+    find_path(LIBUSB_INCLUDE_DIR NAMES libusb.h
+       PATHS
+       ${LIBUSB_PKG_INCLUDE_DIRS}
+       /usr/include/libusb-1.0
+       /usr/include
+       /usr/local/include/libusb-1.0
+       /usr/local/include
+    )
+
+    find_library(LIBUSB_LIBRARIES NAMES usb-1.0 usb
+       PATHS
+       ${LIBUSB_PKG_LIBRARY_DIRS}
+       /usr/lib
+       /usr/local/lib
+    )
+
+    if(LIBUSB_INCLUDE_DIR AND LIBUSB_LIBRARIES)
+       set(LIBUSB_FOUND TRUE CACHE INTERNAL "libusb-1.0 found")
+       message(STATUS "Found libusb-1.0: ${LIBUSB_INCLUDE_DIR}, ${LIBUSB_LIBRARIES}")
+    else(LIBUSB_INCLUDE_DIR AND LIBUSB_LIBRARIES)
+       set(LIBUSB_FOUND FALSE CACHE INTERNAL "libusb-1.0 found")
+       message(STATUS "libusb-1.0 not found.")
+    endif(LIBUSB_INCLUDE_DIR AND LIBUSB_LIBRARIES)
+
+    mark_as_advanced(LIBUSB_INCLUDE_DIR LIBUSB_LIBRARIES)
+endif ()
+

src/audio_core/cubeb_sink.cpp

@@ -193,7 +193,7 @@ long CubebSinkStream::DataCallback(cubeb_stream* stream, void* user_data, const
     const std::size_t samples_to_write = num_channels * num_frames;
     std::size_t samples_written;
 
-    if (Settings::values.enable_audio_stretching) {
+    if (Settings::values.enable_audio_stretching.GetValue()) {
         const std::vector<s16> in{impl->queue.Pop()};
         const std::size_t num_in{in.size() / num_channels};
         s16* const out{reinterpret_cast<s16*>(buffer)};

src/audio_core/stream.cpp

@@ -38,7 +38,7 @@ Stream::Stream(Core::Timing::CoreTiming& core_timing, u32 sample_rate, Format fo
       sink_stream{sink_stream}, core_timing{core_timing}, name{std::move(name_)} {
 
     release_event = Core::Timing::CreateEvent(
-        name, [this](u64 userdata, s64 cycles_late) { ReleaseActiveBuffer(); });
+        name, [this](u64 userdata, s64 cycles_late) { ReleaseActiveBuffer(cycles_late); });
 }
 
 void Stream::Play() {
@@ -66,15 +66,6 @@ s64 Stream::GetBufferReleaseNS(const Buffer& buffer) const {
     return ns.count();
 }
 
-s64 Stream::GetBufferReleaseNSHostTiming(const Buffer& buffer) const {
-    const std::size_t num_samples{buffer.GetSamples().size() / GetNumChannels()};
-    /// DSP signals before playing the last sample, in HLE we emulate this in this way
-    s64 base_samples = std::max<s64>(static_cast<s64>(num_samples) - 1, 0);
-    const auto ns =
-        std::chrono::nanoseconds((static_cast<u64>(base_samples) * 1000000000ULL) / sample_rate);
-    return ns.count();
-}
-
 static void VolumeAdjustSamples(std::vector<s16>& samples, float game_volume) {
     const float volume{std::clamp(Settings::Volume() - (1.0f - game_volume), 0.0f, 1.0f)};
 
@@ -89,7 +80,7 @@ static void VolumeAdjustSamples(std::vector<s16>& samples, float game_volume) {
     }
 }
 
-void Stream::PlayNextBuffer() {
+void Stream::PlayNextBuffer(s64 cycles_late) {
     if (!IsPlaying()) {
         // Ensure we are in playing state before playing the next buffer
         sink_stream.Flush();
@@ -114,18 +105,17 @@ void Stream::PlayNextBuffer() {
 
     sink_stream.EnqueueSamples(GetNumChannels(), active_buffer->GetSamples());
 
-    if (core_timing.IsHostTiming()) {
-        core_timing.ScheduleEvent(GetBufferReleaseNSHostTiming(*active_buffer), release_event, {});
-    } else {
-        core_timing.ScheduleEvent(GetBufferReleaseNS(*active_buffer), release_event, {});
-    }
+    core_timing.ScheduleEvent(
+        GetBufferReleaseNS(*active_buffer) -
+            (Settings::values.enable_audio_stretching.GetValue() ? 0 : cycles_late),
+        release_event, {});
 }
 
-void Stream::ReleaseActiveBuffer() {
+void Stream::ReleaseActiveBuffer(s64 cycles_late) {
     ASSERT(active_buffer);
     released_buffers.push(std::move(active_buffer));
     release_callback();
-    PlayNextBuffer();
+    PlayNextBuffer(cycles_late);
 }
 
 bool Stream::QueueBuffer(BufferPtr&& buffer) {

src/audio_core/stream.h

@@ -90,10 +90,10 @@ public:
 
 private:
     /// Plays the next queued buffer in the audio stream, starting playback if necessary
-    void PlayNextBuffer();
+    void PlayNextBuffer(s64 cycles_late = 0);
 
     /// Releases the actively playing buffer, signalling that it has been completed
-    void ReleaseActiveBuffer();
+    void ReleaseActiveBuffer(s64 cycles_late = 0);
 
     /// Gets the number of core cycles when the specified buffer will be released
     s64 GetBufferReleaseNS(const Buffer& buffer) const;

src/common/alignment.h

@@ -11,7 +11,9 @@ namespace Common {
 template <typename T>
 constexpr T AlignUp(T value, std::size_t size) {
     static_assert(std::is_unsigned_v<T>, "T must be an unsigned value.");
-    return static_cast<T>(value + (size - value % size) % size);
+    auto mod{static_cast<T>(value % size)};
+    value -= mod;
+    return static_cast<T>(mod == T{0} ? value : value + size);
 }
 
 template <typename T>

src/core/arm/cpu_interrupt_handler.cpp

@@ -2,8 +2,6 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
-#pragma once
-
 #include "common/thread.h"
 #include "core/arm/cpu_interrupt_handler.h"
 

src/core/arm/dynarmic/arm_dynarmic_32.cpp

@@ -142,10 +142,32 @@ std::shared_ptr<Dynarmic::A32::Jit> ARM_Dynarmic_32::MakeJit(Common::PageTable&
     // Timing
     config.wall_clock_cntpct = uses_wall_clock;
 
-    // Optimizations
-    if (Settings::values.disable_cpu_opt) {
-        config.enable_optimizations = false;
-        config.enable_fast_dispatch = false;
+    // Safe optimizations
+    if (Settings::values.cpu_accuracy != Settings::CPUAccuracy::Accurate) {
+        if (!Settings::values.cpuopt_page_tables) {
+            config.page_table = nullptr;
+        }
+        if (!Settings::values.cpuopt_block_linking) {
+            config.optimizations &= ~Dynarmic::OptimizationFlag::BlockLinking;
+        }
+        if (!Settings::values.cpuopt_return_stack_buffer) {
+            config.optimizations &= ~Dynarmic::OptimizationFlag::ReturnStackBuffer;
+        }
+        if (!Settings::values.cpuopt_fast_dispatcher) {
+            config.optimizations &= ~Dynarmic::OptimizationFlag::FastDispatch;
+        }
+        if (!Settings::values.cpuopt_context_elimination) {
+            config.optimizations &= ~Dynarmic::OptimizationFlag::GetSetElimination;
+        }
+        if (!Settings::values.cpuopt_const_prop) {
+            config.optimizations &= ~Dynarmic::OptimizationFlag::ConstProp;
+        }
+        if (!Settings::values.cpuopt_misc_ir) {
+            config.optimizations &= ~Dynarmic::OptimizationFlag::MiscIROpt;
+        }
+        if (!Settings::values.cpuopt_reduce_misalign_checks) {
+            config.only_detect_misalignment_via_page_table_on_page_boundary = false;
+        }
     }
 
     return std::make_unique<Dynarmic::A32::Jit>(config);

src/core/arm/dynarmic/arm_dynarmic_64.cpp

@@ -191,15 +191,37 @@ std::shared_ptr<Dynarmic::A64::Jit> ARM_Dynarmic_64::MakeJit(Common::PageTable&
     // Unpredictable instructions
     config.define_unpredictable_behaviour = true;
 
-    // Optimizations
-    if (Settings::values.disable_cpu_opt) {
-        config.enable_optimizations = false;
-        config.enable_fast_dispatch = false;
-    }
-
     // Timing
     config.wall_clock_cntpct = uses_wall_clock;
 
+    // Safe optimizations
+    if (Settings::values.cpu_accuracy != Settings::CPUAccuracy::Accurate) {
+        if (!Settings::values.cpuopt_page_tables) {
+            config.page_table = nullptr;
+        }
+        if (!Settings::values.cpuopt_block_linking) {
+            config.optimizations &= ~Dynarmic::OptimizationFlag::BlockLinking;
+        }
+        if (!Settings::values.cpuopt_return_stack_buffer) {
+            config.optimizations &= ~Dynarmic::OptimizationFlag::ReturnStackBuffer;
+        }
+        if (!Settings::values.cpuopt_fast_dispatcher) {
+            config.optimizations &= ~Dynarmic::OptimizationFlag::FastDispatch;
+        }
+        if (!Settings::values.cpuopt_context_elimination) {
+            config.optimizations &= ~Dynarmic::OptimizationFlag::GetSetElimination;
+        }
+        if (!Settings::values.cpuopt_const_prop) {
+            config.optimizations &= ~Dynarmic::OptimizationFlag::ConstProp;
+        }
+        if (!Settings::values.cpuopt_misc_ir) {
+            config.optimizations &= ~Dynarmic::OptimizationFlag::MiscIROpt;
+        }
+        if (!Settings::values.cpuopt_reduce_misalign_checks) {
+            config.only_detect_misalignment_via_page_table_on_page_boundary = false;
+        }
+    }
+
     return std::make_shared<Dynarmic::A64::Jit>(config);
 }
 

src/core/core.cpp

@@ -147,8 +147,8 @@ struct System::Impl {
 
         device_memory = std::make_unique<Core::DeviceMemory>(system);
 
-        is_multicore = Settings::values.use_multi_core;
-        is_async_gpu = is_multicore || Settings::values.use_asynchronous_gpu_emulation;
+        is_multicore = Settings::values.use_multi_core.GetValue();
+        is_async_gpu = is_multicore || Settings::values.use_asynchronous_gpu_emulation.GetValue();
 
         kernel.SetMulticore(is_multicore);
         cpu_manager.SetMulticore(is_multicore);
@@ -162,7 +162,7 @@ struct System::Impl {
         const auto current_time = std::chrono::duration_cast<std::chrono::seconds>(
             std::chrono::system_clock::now().time_since_epoch());
         Settings::values.custom_rtc_differential =
-            Settings::values.custom_rtc.value_or(current_time) - current_time;
+            Settings::values.custom_rtc.GetValue().value_or(current_time) - current_time;
 
         // Create a default fs if one doesn't already exist.
         if (virtual_filesystem == nullptr)

src/core/core_timing.cpp

@@ -172,7 +172,7 @@ void CoreTiming::ClearPendingEvents() {
 }
 
 void CoreTiming::RemoveEvent(const std::shared_ptr<EventType>& event_type) {
-    basic_lock.lock();
+    std::scoped_lock lock{basic_lock};
 
     const auto itr = std::remove_if(event_queue.begin(), event_queue.end(), [&](const Event& e) {
         return e.type.lock().get() == event_type.get();
@@ -183,12 +183,10 @@ void CoreTiming::RemoveEvent(const std::shared_ptr<EventType>& event_type) {
         event_queue.erase(itr, event_queue.end());
         std::make_heap(event_queue.begin(), event_queue.end(), std::greater<>());
     }
-    basic_lock.unlock();
 }
 
 std::optional<s64> CoreTiming::Advance() {
-    std::scoped_lock advance_scope{advance_lock};
-    std::scoped_lock basic_scope{basic_lock};
+    std::scoped_lock lock{advance_lock, basic_lock};
     global_timer = GetGlobalTimeNs().count();
 
     while (!event_queue.empty() && event_queue.front().time <= global_timer) {

src/core/crypto/key_manager.cpp

@@ -695,8 +695,9 @@ void KeyManager::WriteKeyToFile(KeyCategory category, std::string_view keyname,
 }
 
 void KeyManager::SetKey(S128KeyType id, Key128 key, u64 field1, u64 field2) {
-    if (s128_keys.find({id, field1, field2}) != s128_keys.end())
+    if (s128_keys.find({id, field1, field2}) != s128_keys.end() || key == Key128{}) {
         return;
+    }
     if (id == S128KeyType::Titlekey) {
         Key128 rights_id;
         std::memcpy(rights_id.data(), &field2, sizeof(u64));
@@ -716,8 +717,9 @@ void KeyManager::SetKey(S128KeyType id, Key128 key, u64 field1, u64 field2) {
             return std::tie(elem.second.type, elem.second.field1, elem.second.field2) ==
                    std::tie(id, field1, field2);
         });
-    if (iter2 != s128_file_id.end())
+    if (iter2 != s128_file_id.end()) {
         WriteKeyToFile(category, iter2->first, key);
+    }
 
     // Variable cases
     if (id == S128KeyType::KeyArea) {
@@ -745,16 +747,18 @@ void KeyManager::SetKey(S128KeyType id, Key128 key, u64 field1, u64 field2) {
 }
 
 void KeyManager::SetKey(S256KeyType id, Key256 key, u64 field1, u64 field2) {
-    if (s256_keys.find({id, field1, field2}) != s256_keys.end())
+    if (s256_keys.find({id, field1, field2}) != s256_keys.end() || key == Key256{}) {
         return;
+    }
     const auto iter = std::find_if(
         s256_file_id.begin(), s256_file_id.end(),
         [&id, &field1, &field2](const std::pair<std::string, KeyIndex<S256KeyType>> elem) {
             return std::tie(elem.second.type, elem.second.field1, elem.second.field2) ==
                    std::tie(id, field1, field2);
         });
-    if (iter != s256_file_id.end())
+    if (iter != s256_file_id.end()) {
         WriteKeyToFile(KeyCategory::Standard, iter->first, key);
+    }
     s256_keys[{id, field1, field2}] = key;
 }
 

src/core/file_sys/bis_factory.cpp

@@ -12,6 +12,10 @@
 
 namespace FileSys {
 
+constexpr u64 NAND_USER_SIZE = 0x680000000;  // 26624 MiB
+constexpr u64 NAND_SYSTEM_SIZE = 0xA0000000; // 2560 MiB
+constexpr u64 NAND_TOTAL_SIZE = 0x747C00000; // 29820 MiB
+
 BISFactory::BISFactory(VirtualDir nand_root_, VirtualDir load_root_, VirtualDir dump_root_)
     : nand_root(std::move(nand_root_)), load_root(std::move(load_root_)),
       dump_root(std::move(dump_root_)),
@@ -110,30 +114,29 @@ VirtualDir BISFactory::GetImageDirectory() const {
 
 u64 BISFactory::GetSystemNANDFreeSpace() const {
     const auto sys_dir = GetOrCreateDirectoryRelative(nand_root, "/system");
-    if (sys_dir == nullptr)
-        return 0;
+    if (sys_dir == nullptr) {
+        return GetSystemNANDTotalSpace();
+    }
 
     return GetSystemNANDTotalSpace() - sys_dir->GetSize();
 }
 
 u64 BISFactory::GetSystemNANDTotalSpace() const {
-    return static_cast<u64>(Settings::values.nand_system_size);
+    return NAND_SYSTEM_SIZE;
 }
 
 u64 BISFactory::GetUserNANDFreeSpace() const {
-    const auto usr_dir = GetOrCreateDirectoryRelative(nand_root, "/user");
-    if (usr_dir == nullptr)
-        return 0;
-
-    return GetUserNANDTotalSpace() - usr_dir->GetSize();
+    // For some reason games such as BioShock 1 checks whether this is exactly 0x680000000 bytes.
+    // Set the free space to be 1 MiB less than the total as a workaround to this issue.
+    return GetUserNANDTotalSpace() - 0x100000;
 }
 
 u64 BISFactory::GetUserNANDTotalSpace() const {
-    return static_cast<u64>(Settings::values.nand_user_size);
+    return NAND_USER_SIZE;
 }
 
 u64 BISFactory::GetFullNANDTotalSpace() const {
-    return static_cast<u64>(Settings::values.nand_total_size);
+    return NAND_TOTAL_SIZE;
 }
 
 VirtualDir BISFactory::GetBCATDirectory(u64 title_id) const {

src/core/file_sys/patch_manager.cpp

@@ -80,16 +80,6 @@ VirtualDir PatchManager::PatchExeFS(VirtualDir exefs) const {
     if (exefs == nullptr)
         return exefs;
 
-    if (Settings::values.dump_exefs) {
-        LOG_INFO(Loader, "Dumping ExeFS for title_id={:016X}", title_id);
-        const auto dump_dir =
-            Core::System::GetInstance().GetFileSystemController().GetModificationDumpRoot(title_id);
-        if (dump_dir != nullptr) {
-            const auto exefs_dir = GetOrCreateDirectoryRelative(dump_dir, "/exefs");
-            VfsRawCopyD(exefs, exefs_dir);
-        }
-    }
-
     const auto& installed = Core::System::GetInstance().GetContentProvider();
 
     const auto& disabled = Settings::values.disabled_addons[title_id];
@@ -135,6 +125,16 @@ VirtualDir PatchManager::PatchExeFS(VirtualDir exefs) const {
         }
     }
 
+    if (Settings::values.dump_exefs) {
+        LOG_INFO(Loader, "Dumping ExeFS for title_id={:016X}", title_id);
+        const auto dump_dir =
+                Core::System::GetInstance().GetFileSystemController().GetModificationDumpRoot(title_id);
+        if (dump_dir != nullptr) {
+            const auto exefs_dir = GetOrCreateDirectoryRelative(dump_dir, "/exefs");
+            VfsRawCopyD(exefs, exefs_dir);
+        }
+    }
+
     return exefs;
 }
 

src/core/file_sys/sdmc_factory.cpp

@@ -10,6 +10,8 @@
 
 namespace FileSys {
 
+constexpr u64 SDMC_TOTAL_SIZE = 0x10000000000; // 1 TiB
+
 SDMCFactory::SDMCFactory(VirtualDir dir_)
     : dir(std::move(dir_)), contents(std::make_unique<RegisteredCache>(
                                 GetOrCreateDirectoryRelative(dir, "/Nintendo/Contents/registered"),
@@ -46,7 +48,7 @@ u64 SDMCFactory::GetSDMCFreeSpace() const {
 }
 
 u64 SDMCFactory::GetSDMCTotalSpace() const {
-    return static_cast<u64>(Settings::values.sdmc_size);
+    return SDMC_TOTAL_SIZE;
 }
 
 } // namespace FileSys

src/core/file_sys/vfs_real.cpp

@@ -112,19 +112,26 @@ VirtualFile RealVfsFilesystem::MoveFile(std::string_view old_path_, std::string_
     const auto new_path =
         FileUtil::SanitizePath(new_path_, FileUtil::DirectorySeparator::PlatformDefault);
 
-    if (!FileUtil::Exists(old_path) || FileUtil::Exists(new_path) ||
-        FileUtil::IsDirectory(old_path) || !FileUtil::Rename(old_path, new_path))
-        return nullptr;
-
     if (cache.find(old_path) != cache.end()) {
-        auto cached = cache[old_path];
-        if (!cached.expired()) {
-            auto file = cached.lock();
-            file->Open(new_path, "r+b");
-            cache.erase(old_path);
-            cache[new_path] = file;
+        auto file = cache[old_path].lock();
+
+        if (!cache[old_path].expired()) {
+            file->Close();
         }
+
+        if (!FileUtil::Exists(old_path) || FileUtil::Exists(new_path) ||
+            FileUtil::IsDirectory(old_path) || !FileUtil::Rename(old_path, new_path)) {
+            return nullptr;
+        }
+
+        cache.erase(old_path);
+        file->Open(new_path, "r+b");
+        cache[new_path] = file;
+    } else {
+        UNREACHABLE();
+        return nullptr;
     }
+
     return OpenFile(new_path, Mode::ReadWrite);
 }
 

src/core/frontend/framebuffer_layout.cpp

@@ -29,7 +29,7 @@ FramebufferLayout DefaultFrameLayout(u32 width, u32 height) {
 
     const float window_aspect_ratio = static_cast<float>(height) / width;
     const float emulation_aspect_ratio = EmulationAspectRatio(
-        static_cast<AspectRatio>(Settings::values.aspect_ratio), window_aspect_ratio);
+        static_cast<AspectRatio>(Settings::values.aspect_ratio.GetValue()), window_aspect_ratio);
 
     const Common::Rectangle<u32> screen_window_area{0, 0, width, height};
     Common::Rectangle<u32> screen = MaxRectangle(screen_window_area, emulation_aspect_ratio);

src/core/hle/kernel/process.cpp

@@ -123,7 +123,7 @@ std::shared_ptr<Process> Process::Create(Core::System& system, std::string name,
                                                               : kernel.CreateNewUserProcessID();
     process->capabilities.InitializeForMetadatalessProcess();
 
-    std::mt19937 rng(Settings::values.rng_seed.value_or(0));
+    std::mt19937 rng(Settings::values.rng_seed.GetValue().value_or(0));
     std::uniform_int_distribution<u64> distribution;
     std::generate(process->random_entropy.begin(), process->random_entropy.end(),
                   [&] { return distribution(rng); });

src/core/hle/kernel/scheduler.cpp

@@ -6,6 +6,7 @@
 // licensed under GPLv2 or later under exception provided by the author.
 
 #include <algorithm>
+#include <mutex>
 #include <set>
 #include <unordered_set>
 #include <utility>
@@ -31,22 +32,20 @@ GlobalScheduler::GlobalScheduler(KernelCore& kernel) : kernel{kernel} {}
 GlobalScheduler::~GlobalScheduler() = default;
 
 void GlobalScheduler::AddThread(std::shared_ptr<Thread> thread) {
-    global_list_guard.lock();
+    std::scoped_lock lock{global_list_guard};
     thread_list.push_back(std::move(thread));
-    global_list_guard.unlock();
 }
 
 void GlobalScheduler::RemoveThread(std::shared_ptr<Thread> thread) {
-    global_list_guard.lock();
+    std::scoped_lock lock{global_list_guard};
     thread_list.erase(std::remove(thread_list.begin(), thread_list.end(), thread),
                       thread_list.end());
-    global_list_guard.unlock();
 }
 
 u32 GlobalScheduler::SelectThreads() {
     ASSERT(is_locked);
     const auto update_thread = [](Thread* thread, Scheduler& sched) {
-        sched.guard.lock();
+        std::scoped_lock lock{sched.guard};
         if (thread != sched.selected_thread_set.get()) {
             if (thread == nullptr) {
                 ++sched.idle_selection_count;
@@ -57,7 +56,6 @@ u32 GlobalScheduler::SelectThreads() {
             sched.is_context_switch_pending || (sched.selected_thread_set != sched.current_thread);
         sched.is_context_switch_pending = reschedule_pending;
         std::atomic_thread_fence(std::memory_order_seq_cst);
-        sched.guard.unlock();
         return reschedule_pending;
     };
     if (!is_reselection_pending.load()) {
@@ -757,11 +755,12 @@ void Scheduler::OnSwitch(void* this_scheduler) {
 
 void Scheduler::SwitchToCurrent() {
     while (true) {
-        guard.lock();
-        selected_thread = selected_thread_set;
-        current_thread = selected_thread;
-        is_context_switch_pending = false;
-        guard.unlock();
+        {
+            std::scoped_lock lock{guard};
+            selected_thread = selected_thread_set;
+            current_thread = selected_thread;
+            is_context_switch_pending = false;
+        }
         while (!is_context_switch_pending) {
             if (current_thread != nullptr && !current_thread->IsHLEThread()) {
                 current_thread->context_guard.lock();

src/core/hle/service/am/am.cpp

@@ -10,6 +10,7 @@
 #include "core/core.h"
 #include "core/file_sys/control_metadata.h"
 #include "core/file_sys/patch_manager.h"
+#include "core/file_sys/registered_cache.h"
 #include "core/file_sys/savedata_factory.h"
 #include "core/hle/ipc_helpers.h"
 #include "core/hle/kernel/kernel.h"
@@ -271,7 +272,7 @@ ISelfController::ISelfController(Core::System& system,
         {41, nullptr, "IsSystemBufferSharingEnabled"},
         {42, nullptr, "GetSystemSharedLayerHandle"},
         {43, nullptr, "GetSystemSharedBufferHandle"},
-        {44, nullptr, "CreateManagedDisplaySeparableLayer"},
+        {44, &ISelfController::CreateManagedDisplaySeparableLayer, "CreateManagedDisplaySeparableLayer"},
         {45, nullptr, "SetManagedDisplayLayerSeparationMode"},
         {50, &ISelfController::SetHandlesRequestToDisplay, "SetHandlesRequestToDisplay"},
         {51, nullptr, "ApproveToDisplay"},
@@ -461,6 +462,24 @@ void ISelfController::CreateManagedDisplayLayer(Kernel::HLERequestContext& ctx)
     rb.Push(*layer_id);
 }
 
+void ISelfController::CreateManagedDisplaySeparableLayer(Kernel::HLERequestContext& ctx) {
+    LOG_WARNING(Service_AM, "(STUBBED) called");
+
+    // TODO(Subv): Find out how AM determines the display to use, for now just
+    // create the layer in the Default display.
+    // This calls nn::vi::CreateRecordingLayer() which creates another layer.
+    // Currently we do not support more than 1 layer per display, output 1 layer id for now.
+    // Outputting 1 layer id instead of the expected 2 has not been observed to cause any adverse
+    // side effects.
+    // TODO: Support multiple layers
+    const auto display_id = nvflinger->OpenDisplay("Default");
+    const auto layer_id = nvflinger->CreateLayer(*display_id);
+
+    IPC::ResponseBuilder rb{ctx, 4};
+    rb.Push(RESULT_SUCCESS);
+    rb.Push(*layer_id);
+}
+
 void ISelfController::SetHandlesRequestToDisplay(Kernel::HLERequestContext& ctx) {
     LOG_WARNING(Service_AM, "(STUBBED) called");
 
@@ -730,14 +749,14 @@ void ICommonStateGetter::GetDefaultDisplayResolution(Kernel::HLERequestContext&
 
     if (Settings::values.use_docked_mode) {
         rb.Push(static_cast<u32>(Service::VI::DisplayResolution::DockedWidth) *
-                static_cast<u32>(Settings::values.resolution_factor));
+                static_cast<u32>(Settings::values.resolution_factor.GetValue()));
         rb.Push(static_cast<u32>(Service::VI::DisplayResolution::DockedHeight) *
-                static_cast<u32>(Settings::values.resolution_factor));
+                static_cast<u32>(Settings::values.resolution_factor.GetValue()));
     } else {
         rb.Push(static_cast<u32>(Service::VI::DisplayResolution::UndockedWidth) *
-                static_cast<u32>(Settings::values.resolution_factor));
+                static_cast<u32>(Settings::values.resolution_factor.GetValue()));
         rb.Push(static_cast<u32>(Service::VI::DisplayResolution::UndockedHeight) *
-                static_cast<u32>(Settings::values.resolution_factor));
+                static_cast<u32>(Settings::values.resolution_factor.GetValue()));
     }
 }
 
@@ -1353,14 +1372,25 @@ void IApplicationFunctions::GetDisplayVersion(Kernel::HLERequestContext& ctx) {
 
     std::array<u8, 0x10> version_string{};
 
-    FileSys::PatchManager pm{system.CurrentProcess()->GetTitleID()};
-    const auto res = pm.GetControlMetadata();
+    const auto res = [this] {
+        const auto title_id = system.CurrentProcess()->GetTitleID();
+
+        FileSys::PatchManager pm{title_id};
+        auto res = pm.GetControlMetadata();
+        if (res.first != nullptr) {
+            return res;
+        }
+
+        FileSys::PatchManager pm_update{FileSys::GetUpdateTitleID(title_id)};
+        return pm_update.GetControlMetadata();
+    }();
+
     if (res.first != nullptr) {
         const auto& version = res.first->GetVersionString();
         std::copy(version.begin(), version.end(), version_string.begin());
     } else {
-        constexpr u128 default_version = {1, 0};
-        std::memcpy(version_string.data(), default_version.data(), sizeof(u128));
+        constexpr char default_version[]{"1.0.0"};
+        std::memcpy(version_string.data(), default_version, sizeof(default_version));
     }
 
     IPC::ResponseBuilder rb{ctx, 6};
@@ -1377,7 +1407,19 @@ void IApplicationFunctions::GetDesiredLanguage(Kernel::HLERequestContext& ctx) {
     u32 supported_languages = 0;
     FileSys::PatchManager pm{system.CurrentProcess()->GetTitleID()};
 
-    const auto res = pm.GetControlMetadata();
+    const auto res = [this] {
+        const auto title_id = system.CurrentProcess()->GetTitleID();
+
+        FileSys::PatchManager pm{title_id};
+        auto res = pm.GetControlMetadata();
+        if (res.first != nullptr) {
+            return res;
+        }
+
+        FileSys::PatchManager pm_update{FileSys::GetUpdateTitleID(title_id)};
+        return pm_update.GetControlMetadata();
+    }();
+
     if (res.first != nullptr) {
         supported_languages = res.first->GetSupportedLanguages();
     }

src/core/hle/service/am/am.h

@@ -140,6 +140,7 @@ private:
     void SetOutOfFocusSuspendingEnabled(Kernel::HLERequestContext& ctx);
     void SetAlbumImageOrientation(Kernel::HLERequestContext& ctx);
     void CreateManagedDisplayLayer(Kernel::HLERequestContext& ctx);
+    void CreateManagedDisplaySeparableLayer(Kernel::HLERequestContext& ctx);
     void SetHandlesRequestToDisplay(Kernel::HLERequestContext& ctx);
     void SetIdleTimeDetectionExtension(Kernel::HLERequestContext& ctx);
     void GetIdleTimeDetectionExtension(Kernel::HLERequestContext& ctx);

src/core/hle/service/nim/nim.cpp

@@ -121,11 +121,83 @@ public:
             {39, nullptr, "PrepareShutdown"},
             {40, nullptr, "ListApplyDeltaTask"},
             {41, nullptr, "ClearNotEnoughSpaceStateOfApplyDeltaTask"},
-            {42, nullptr, "Unknown1"},
-            {43, nullptr, "Unknown2"},
-            {44, nullptr, "Unknown3"},
-            {45, nullptr, "Unknown4"},
-            {46, nullptr, "Unknown5"},
+            {42, nullptr, "Unknown42"},
+            {43, nullptr, "Unknown43"},
+            {44, nullptr, "Unknown44"},
+            {45, nullptr, "Unknown45"},
+            {46, nullptr, "Unknown46"},
+            {47, nullptr, "Unknown47"},
+            {48, nullptr, "Unknown48"},
+            {49, nullptr, "Unknown49"},
+            {50, nullptr, "Unknown50"},
+            {51, nullptr, "Unknown51"},
+            {52, nullptr, "Unknown52"},
+            {53, nullptr, "Unknown53"},
+            {54, nullptr, "Unknown54"},
+            {55, nullptr, "Unknown55"},
+            {56, nullptr, "Unknown56"},
+            {57, nullptr, "Unknown57"},
+            {58, nullptr, "Unknown58"},
+            {59, nullptr, "Unknown59"},
+            {60, nullptr, "Unknown60"},
+            {61, nullptr, "Unknown61"},
+            {62, nullptr, "Unknown62"},
+            {63, nullptr, "Unknown63"},
+            {64, nullptr, "Unknown64"},
+            {65, nullptr, "Unknown65"},
+            {66, nullptr, "Unknown66"},
+            {67, nullptr, "Unknown67"},
+            {68, nullptr, "Unknown68"},
+            {69, nullptr, "Unknown69"},
+            {70, nullptr, "Unknown70"},
+            {71, nullptr, "Unknown71"},
+            {72, nullptr, "Unknown72"},
+            {73, nullptr, "Unknown73"},
+            {74, nullptr, "Unknown74"},
+            {75, nullptr, "Unknown75"},
+            {76, nullptr, "Unknown76"},
+            {77, nullptr, "Unknown77"},
+            {78, nullptr, "Unknown78"},
+            {79, nullptr, "Unknown79"},
+            {80, nullptr, "Unknown80"},
+            {81, nullptr, "Unknown81"},
+            {82, nullptr, "Unknown82"},
+            {83, nullptr, "Unknown83"},
+            {84, nullptr, "Unknown84"},
+            {85, nullptr, "Unknown85"},
+            {86, nullptr, "Unknown86"},
+            {87, nullptr, "Unknown87"},
+            {88, nullptr, "Unknown88"},
+            {89, nullptr, "Unknown89"},
+            {90, nullptr, "Unknown90"},
+            {91, nullptr, "Unknown91"},
+            {92, nullptr, "Unknown92"},
+            {93, nullptr, "Unknown93"},
+            {94, nullptr, "Unknown94"},
+            {95, nullptr, "Unknown95"},
+            {96, nullptr, "Unknown96"},
+            {97, nullptr, "Unknown97"},
+            {98, nullptr, "Unknown98"},
+            {99, nullptr, "Unknown99"},
+            {100, nullptr, "Unknown100"},
+            {101, nullptr, "Unknown101"},
+            {102, nullptr, "Unknown102"},
+            {103, nullptr, "Unknown103"},
+            {104, nullptr, "Unknown104"},
+            {105, nullptr, "Unknown105"},
+            {106, nullptr, "Unknown106"},
+            {107, nullptr, "Unknown107"},
+            {108, nullptr, "Unknown108"},
+            {109, nullptr, "Unknown109"},
+            {110, nullptr, "Unknown110"},
+            {111, nullptr, "Unknown111"},
+            {112, nullptr, "Unknown112"},
+            {113, nullptr, "Unknown113"},
+            {114, nullptr, "Unknown114"},
+            {115, nullptr, "Unknown115"},
+            {116, nullptr, "Unknown116"},
+            {117, nullptr, "Unknown117"},
+            {118, nullptr, "Unknown118"},
         };
         // clang-format on
 
@@ -142,6 +214,7 @@ public:
             {1, nullptr, "RefreshDebugAvailability"},
             {2, nullptr, "ClearDebugResponse"},
             {3, nullptr, "RegisterDebugResponse"},
+            {4, nullptr, "IsLargeResourceAvailable"},
         };
         // clang-format on
 
@@ -164,6 +237,8 @@ public:
         static const FunctionInfo functions[] = {
             {0, nullptr, "RequestDeviceAuthenticationToken"},
             {1, nullptr, "RequestCachedDeviceAuthenticationToken"},
+            {2, nullptr, "RequestEdgeToken"},
+            {3, nullptr, "RequestCachedEdgeToken"},
             {100, nullptr, "RequestRegisterDeviceAccount"},
             {101, nullptr, "RequestUnregisterDeviceAccount"},
             {102, nullptr, "RequestDeviceAccountStatus"},
@@ -181,7 +256,8 @@ public:
             {305, nullptr, "RequestCreateVirtualAccount"},
             {306, nullptr, "RequestDeviceLinkStatus"},
             {400, nullptr, "GetAccountByVirtualAccount"},
-            {500, nullptr, "RequestSyncTicket"},
+            {401, nullptr, "GetVirtualAccount"},
+            {500, nullptr, "RequestSyncTicketLegacy"},
             {501, nullptr, "RequestDownloadTicket"},
             {502, nullptr, "RequestDownloadTicketForPrepurchasedContents"},
             {503, nullptr, "RequestSyncTicket"},

src/core/hle/service/npns/npns.cpp

@@ -30,6 +30,7 @@ public:
             {23, nullptr, "DestroyToken"},
             {24, nullptr, "DestroyTokenWithApplicationId"},
             {25, nullptr, "QueryIsTokenValid"},
+            {26, nullptr, "ListenToMyApplicationId"},
             {31, nullptr, "UploadTokenToBaaS"},
             {32, nullptr, "DestroyTokenForBaaS"},
             {33, nullptr, "CreateTokenForBaaS"},

src/core/hle/service/ns/ns.cpp

@@ -104,7 +104,7 @@ IApplicationManagerInterface::IApplicationManagerInterface()
         {94, nullptr, "LaunchApplication"},
         {95, nullptr, "GetApplicationLaunchInfo"},
         {96, nullptr, "AcquireApplicationLaunchInfo"},
-        {97, nullptr, "GetMainApplicationProgramIndex2"},
+        {97, nullptr, "GetMainApplicationProgramIndexByApplicationLaunchInfo"},
         {98, nullptr, "EnableApplicationAllThreadDumpOnCrash"},
         {99, nullptr, "LaunchDevMenu"},
         {100, nullptr, "ResetToFactorySettings"},
@@ -254,7 +254,7 @@ IApplicationManagerInterface::IApplicationManagerInterface()
         {2170, nullptr, "GetRightsEnvironmentStatus"},
         {2171, nullptr, "GetRightsEnvironmentStatusChangedEvent"},
         {2180, nullptr, "RequestExtendRightsInRightsEnvironment"},
-        {2181, nullptr, "GetLastResultOfExtendRightsInRightsEnvironment"},
+        {2181, nullptr, "GetResultOfExtendRightsInRightsEnvironment"},
         {2182, nullptr, "SetActiveRightsContextUsingStateToRightsEnvironment"},
         {2190, nullptr, "GetRightsEnvironmentHandleForApplication"},
         {2199, nullptr, "GetRightsEnvironmentCountForDebug"},
@@ -366,7 +366,8 @@ ResultVal<u8> IApplicationManagerInterface::GetApplicationDesiredLanguage(
     LOG_DEBUG(Service_NS, "called with supported_languages={:08X}", supported_languages);
 
     // Get language code from settings
-    const auto language_code = Set::GetLanguageCodeFromIndex(Settings::values.language_index);
+    const auto language_code =
+        Set::GetLanguageCodeFromIndex(Settings::values.language_index.GetValue());
 
     // Convert to application language, get priority list
     const auto application_language = ConvertToApplicationLanguage(language_code);
@@ -445,8 +446,8 @@ IApplicationVersionInterface::IApplicationVersionInterface()
 
 IApplicationVersionInterface::~IApplicationVersionInterface() = default;
 
-IContentManagerInterface::IContentManagerInterface()
-    : ServiceFramework{"IContentManagerInterface"} {
+IContentManagementInterface::IContentManagementInterface()
+    : ServiceFramework{"IContentManagementInterface"} {
     // clang-format off
     static const FunctionInfo functions[] = {
         {11, nullptr, "CalculateApplicationOccupiedSize"},
@@ -463,7 +464,7 @@ IContentManagerInterface::IContentManagerInterface()
     RegisterHandlers(functions);
 }
 
-IContentManagerInterface::~IContentManagerInterface() = default;
+IContentManagementInterface::~IContentManagementInterface() = default;
 
 IDocumentInterface::IDocumentInterface() : ServiceFramework{"IDocumentInterface"} {
     // clang-format off
@@ -545,7 +546,7 @@ NS::NS(const char* name) : ServiceFramework{name} {
         {7995, &NS::PushInterface<IAccountProxyInterface>, "GetAccountProxyInterface"},
         {7996, &NS::PushInterface<IApplicationManagerInterface>, "GetApplicationManagerInterface"},
         {7997, &NS::PushInterface<IDownloadTaskInterface>, "GetDownloadTaskInterface"},
-        {7998, &NS::PushInterface<IContentManagerInterface>, "GetContentManagementInterface"},
+        {7998, &NS::PushInterface<IContentManagementInterface>, "GetContentManagementInterface"},
         {7999, &NS::PushInterface<IDocumentInterface>, "GetDocumentInterface"},
     };
     // clang-format on
@@ -572,9 +573,9 @@ public:
             {6, nullptr, "TerminateApplication"},
             {7, nullptr, "PrepareLaunchProgramFromHost"},
             {8, nullptr, "LaunchApplication"},
-            {9, nullptr, "LaunchApplicationWithStorageId"},
-            {10, nullptr, "TerminateApplication2"},
-            {11, nullptr, "GetRunningApplicationProcessId"},
+            {9, nullptr, "LaunchApplicationWithStorageIdForDevelop"},
+            {10, nullptr, "IsSystemMemoryResourceLimitBoosted"},
+            {11, nullptr, "GetRunningApplicationProcessIdForDevelop"},
             {12, nullptr, "SetCurrentApplicationRightsEnvironmentCanBeActive"},
             {13, nullptr, "CreateApplicationResourceForDevelop"},
             {14, nullptr, "IsPreomiaForDevelop"},
@@ -636,6 +637,10 @@ public:
             {9, nullptr, "GetSystemUpdateNotificationEventForContentDelivery"},
             {10, nullptr, "NotifySystemUpdateForContentDelivery"},
             {11, nullptr, "PrepareShutdown"},
+            {12, nullptr, "Unknown12"},
+            {13, nullptr, "Unknown13"},
+            {14, nullptr, "Unknown14"},
+            {15, nullptr, "Unknown15"},
             {16, nullptr, "DestroySystemUpdateTask"},
             {17, nullptr, "RequestSendSystemUpdate"},
             {18, nullptr, "GetSendSystemUpdateProgress"},

src/core/hle/service/ns/ns.h

@@ -40,10 +40,10 @@ public:
     ~IApplicationVersionInterface() override;
 };
 
-class IContentManagerInterface final : public ServiceFramework<IContentManagerInterface> {
+class IContentManagementInterface final : public ServiceFramework<IContentManagementInterface> {
 public:
-    explicit IContentManagerInterface();
-    ~IContentManagerInterface() override;
+    explicit IContentManagementInterface();
+    ~IContentManagementInterface() override;
 };
 
 class IDocumentInterface final : public ServiceFramework<IDocumentInterface> {

src/core/hle/service/ns/pl_u.cpp

@@ -163,7 +163,7 @@ PL_U::PL_U(Core::System& system)
         {5, &PL_U::GetSharedFontInOrderOfPriority, "GetSharedFontInOrderOfPriority"},
         {6, nullptr, "GetSharedFontInOrderOfPriorityForSystem"},
         {100, nullptr, "RequestApplicationFunctionAuthorization"},
-        {101, nullptr, "RequestApplicationFunctionAuthorizationForSystem"},
+        {101, nullptr, "RequestApplicationFunctionAuthorizationByProcessId"},
         {102, nullptr, "RequestApplicationFunctionAuthorizationByApplicationId"},
         {1000, nullptr, "LoadNgWordDataForPlatformRegionChina"},
         {1001, nullptr, "GetNgWordDataSizeForPlatformRegionChina"},

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

@@ -144,7 +144,7 @@ void NVDRV::QueryEvent(Kernel::HLERequestContext& ctx) {
     }
 }
 
-void NVDRV::SetClientPID(Kernel::HLERequestContext& ctx) {
+void NVDRV::SetAruid(Kernel::HLERequestContext& ctx) {
     IPC::RequestParser rp{ctx};
     pid = rp.Pop<u64>();
     LOG_WARNING(Service_NVDRV, "(STUBBED) called, pid=0x{:X}", pid);
@@ -154,7 +154,7 @@ void NVDRV::SetClientPID(Kernel::HLERequestContext& ctx) {
     rb.Push<u32>(0);
 }
 
-void NVDRV::FinishInitialize(Kernel::HLERequestContext& ctx) {
+void NVDRV::SetGraphicsFirmwareMemoryMarginEnabled(Kernel::HLERequestContext& ctx) {
     LOG_WARNING(Service_NVDRV, "(STUBBED) called");
 
     IPC::ResponseBuilder rb{ctx, 2};
@@ -187,13 +187,14 @@ NVDRV::NVDRV(std::shared_ptr<Module> nvdrv, const char* name)
         {4, &NVDRV::QueryEvent, "QueryEvent"},
         {5, nullptr, "MapSharedMem"},
         {6, &NVDRV::GetStatus, "GetStatus"},
-        {7, nullptr, "ForceSetClientPID"},
-        {8, &NVDRV::SetClientPID, "SetClientPID"},
+        {7, nullptr, "SetAruidForTest"},
+        {8, &NVDRV::SetAruid, "SetAruid"},
         {9, &NVDRV::DumpGraphicsMemoryInfo, "DumpGraphicsMemoryInfo"},
         {10, nullptr, "InitializeDevtools"},
         {11, &NVDRV::Ioctl2, "Ioctl2"},
         {12, &NVDRV::Ioctl3, "Ioctl3"},
-        {13, &NVDRV::FinishInitialize, "FinishInitialize"},
+        {13, &NVDRV::SetGraphicsFirmwareMemoryMarginEnabled,
+         "SetGraphicsFirmwareMemoryMarginEnabled"},
     };
     RegisterHandlers(functions);
 }

src/core/hle/service/nvdrv/interface.h

@@ -29,8 +29,8 @@ private:
     void Close(Kernel::HLERequestContext& ctx);
     void Initialize(Kernel::HLERequestContext& ctx);
     void QueryEvent(Kernel::HLERequestContext& ctx);
-    void SetClientPID(Kernel::HLERequestContext& ctx);
-    void FinishInitialize(Kernel::HLERequestContext& ctx);
+    void SetAruid(Kernel::HLERequestContext& ctx);
+    void SetGraphicsFirmwareMemoryMarginEnabled(Kernel::HLERequestContext& ctx);
     void GetStatus(Kernel::HLERequestContext& ctx);
     void DumpGraphicsMemoryInfo(Kernel::HLERequestContext& ctx);
     void IoctlBase(Kernel::HLERequestContext& ctx, IoctlVersion version);

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

@@ -10,19 +10,19 @@ namespace Service::Nvidia {
 
 NVMEMP::NVMEMP() : ServiceFramework("nvmemp") {
     static const FunctionInfo functions[] = {
-        {0, &NVMEMP::Cmd0, "Cmd0"},
-        {1, &NVMEMP::Cmd1, "Cmd1"},
+        {0, &NVMEMP::Open, "Open"},
+        {1, &NVMEMP::GetAruid, "GetAruid"},
     };
     RegisterHandlers(functions);
 }
 
 NVMEMP::~NVMEMP() = default;
 
-void NVMEMP::Cmd0(Kernel::HLERequestContext& ctx) {
+void NVMEMP::Open(Kernel::HLERequestContext& ctx) {
     UNIMPLEMENTED();
 }
 
-void NVMEMP::Cmd1(Kernel::HLERequestContext& ctx) {
+void NVMEMP::GetAruid(Kernel::HLERequestContext& ctx) {
     UNIMPLEMENTED();
 }
 

src/core/hle/service/nvdrv/nvmemp.h

@@ -14,8 +14,8 @@ public:
     ~NVMEMP() override;
 
 private:
-    void Cmd0(Kernel::HLERequestContext& ctx);
-    void Cmd1(Kernel::HLERequestContext& ctx);
+    void Open(Kernel::HLERequestContext& ctx);
+    void GetAruid(Kernel::HLERequestContext& ctx);
 };
 
 } // namespace Service::Nvidia

src/core/hle/service/pcie/pcie.cpp

@@ -36,6 +36,9 @@ public:
             {18, nullptr, "ReleaseIrq"},
             {19, nullptr, "SetIrqEnable"},
             {20, nullptr, "SetAspmEnable"},
+            {21, nullptr, "SetResetUponResumeEnable"},
+            {22, nullptr, "Unknown22"},
+            {23, nullptr, "Unknown23"},
         };
         // clang-format on
 

src/core/hle/service/pcv/pcv.cpp

@@ -42,6 +42,9 @@ public:
             {24, nullptr, "GetModuleStateTable"},
             {25, nullptr, "GetPowerDomainStateTable"},
             {26, nullptr, "GetFuseInfo"},
+            {27, nullptr, "GetDramId"},
+            {28, nullptr, "IsPoweredOn"},
+            {29, nullptr, "GetVoltage"},
         };
         // clang-format on
 

src/core/hle/service/pm/pm.cpp

@@ -78,13 +78,13 @@ public:
         : ServiceFramework{"pm:dmnt"}, kernel(kernel) {
         // clang-format off
         static const FunctionInfo functions[] = {
-            {0, nullptr, "GetDebugProcesses"},
-            {1, nullptr, "StartDebugProcess"},
-            {2, &DebugMonitor::GetTitlePid, "GetTitlePid"},
-            {3, nullptr, "EnableDebugForTitleId"},
-            {4, &DebugMonitor::GetApplicationPid, "GetApplicationPid"},
-            {5, nullptr, "EnableDebugForApplication"},
-            {6, nullptr, "DisableDebug"},
+            {0, nullptr, "GetJitDebugProcessIdList"},
+            {1, nullptr, "StartProcess"},
+            {2, &DebugMonitor::GetProcessId, "GetProcessId"},
+            {3, nullptr, "HookToCreateProcess"},
+            {4, &DebugMonitor::GetApplicationProcessId, "GetApplicationProcessId"},
+            {5, nullptr, "HookToCreateApplicationProgress"},
+            {6, nullptr, "ClearHook"},
         };
         // clang-format on
 
@@ -92,7 +92,7 @@ public:
     }
 
 private:
-    void GetTitlePid(Kernel::HLERequestContext& ctx) {
+    void GetProcessId(Kernel::HLERequestContext& ctx) {
         IPC::RequestParser rp{ctx};
         const auto title_id = rp.PopRaw<u64>();
 
@@ -114,7 +114,7 @@ private:
         rb.Push((*process)->GetProcessID());
     }
 
-    void GetApplicationPid(Kernel::HLERequestContext& ctx) {
+    void GetApplicationProcessId(Kernel::HLERequestContext& ctx) {
         LOG_DEBUG(Service_PM, "called");
         GetApplicationPidGeneric(ctx, kernel.GetProcessList());
     }
@@ -163,15 +163,15 @@ public:
         : ServiceFramework{"pm:shell"}, kernel(kernel) {
         // clang-format off
         static const FunctionInfo functions[] = {
-            {0, nullptr, "LaunchProcess"},
-            {1, nullptr, "TerminateProcessByPid"},
-            {2, nullptr, "TerminateProcessByTitleId"},
-            {3, nullptr, "GetProcessEventWaiter"},
-            {4, nullptr, "GetProcessEventType"},
+            {0, nullptr, "LaunchProgram"},
+            {1, nullptr, "TerminateProcess"},
+            {2, nullptr, "TerminateProgram"},
+            {3, nullptr, "GetProcessEventHandle"},
+            {4, nullptr, "GetProcessEventInfo"},
             {5, nullptr, "NotifyBootFinished"},
-            {6, &Shell::GetApplicationPid, "GetApplicationPid"},
+            {6, &Shell::GetApplicationProcessIdForShell, "GetApplicationProcessIdForShell"},
             {7, nullptr, "BoostSystemMemoryResourceLimit"},
-            {8, nullptr, "EnableAdditionalSystemThreads"},
+            {8, nullptr, "BoostApplicationThreadResourceLimit"},
             {9, nullptr, "GetBootFinishedEventHandle"},
         };
         // clang-format on
@@ -180,7 +180,7 @@ public:
     }
 
 private:
-    void GetApplicationPid(Kernel::HLERequestContext& ctx) {
+    void GetApplicationProcessIdForShell(Kernel::HLERequestContext& ctx) {
         LOG_DEBUG(Service_PM, "called");
         GetApplicationPidGeneric(ctx, kernel.GetProcessList());
     }

src/core/hle/service/prepo/prepo.cpp

@@ -42,6 +42,11 @@ public:
             {40101, nullptr, "SetUserAgreementCheckEnabled"},
             {50100, nullptr, "ReadAllApplicationReportFiles"},
             {90100, nullptr, "ReadAllReportFiles"},
+            {90101, nullptr, "Unknown90101"},
+            {90102, nullptr, "Unknown90102"},
+            {90200, nullptr, "GetStatistics"},
+            {90201, nullptr, "GetThroughputHistory"},
+            {90300, nullptr, "GetLastUploadError"},
         };
         // clang-format on
 

src/core/hle/service/psc/psc.cpp

@@ -24,6 +24,8 @@ public:
             {4, nullptr, "Cancel"},
             {5, nullptr, "PrintModuleInformation"},
             {6, nullptr, "GetModuleInformation"},
+            {10, nullptr, "Unknown10"},
+            {11, nullptr, "Unknown11"},
         };
         // clang-format on
 

src/core/hle/service/ptm/psm.cpp

@@ -35,6 +35,7 @@ public:
             {15, nullptr, "GetBatteryAgePercentage"},
             {16, nullptr, "GetBatteryChargeInfoEvent"},
             {17, nullptr, "GetBatteryChargeInfoFields"},
+            {18, nullptr, "GetBatteryChargeCalibratedEvent"},
         };
         // clang-format on
 

src/core/hle/service/set/set.cpp

@@ -91,7 +91,7 @@ void GetAvailableLanguageCodesImpl(Kernel::HLERequestContext& ctx, std::size_t m
 }
 
 void GetKeyCodeMapImpl(Kernel::HLERequestContext& ctx) {
-    const auto language_code = available_language_codes[Settings::values.language_index];
+    const auto language_code = available_language_codes[Settings::values.language_index.GetValue()];
     const auto key_code =
         std::find_if(language_to_layout.cbegin(), language_to_layout.cend(),
                      [=](const auto& element) { return element.first == language_code; });
@@ -99,7 +99,7 @@ void GetKeyCodeMapImpl(Kernel::HLERequestContext& ctx) {
     if (key_code == language_to_layout.cend()) {
         LOG_ERROR(Service_SET,
                   "Could not find keyboard layout for language index {}, defaulting to English us",
-                  Settings::values.language_index);
+                  Settings::values.language_index.GetValue());
     } else {
         layout = key_code->second;
     }
@@ -163,11 +163,11 @@ void SET::GetQuestFlag(Kernel::HLERequestContext& ctx) {
 }
 
 void SET::GetLanguageCode(Kernel::HLERequestContext& ctx) {
-    LOG_DEBUG(Service_SET, "called {}", Settings::values.language_index);
+    LOG_DEBUG(Service_SET, "called {}", Settings::values.language_index.GetValue());
 
     IPC::ResponseBuilder rb{ctx, 4};
     rb.Push(RESULT_SUCCESS);
-    rb.PushEnum(available_language_codes[Settings::values.language_index]);
+    rb.PushEnum(available_language_codes[Settings::values.language_index.GetValue()]);
 }
 
 void SET::GetRegionCode(Kernel::HLERequestContext& ctx) {
@@ -175,7 +175,7 @@ void SET::GetRegionCode(Kernel::HLERequestContext& ctx) {
 
     IPC::ResponseBuilder rb{ctx, 3};
     rb.Push(RESULT_SUCCESS);
-    rb.Push(Settings::values.region_index);
+    rb.Push(Settings::values.region_index.GetValue());
 }
 
 void SET::GetKeyCodeMap(Kernel::HLERequestContext& ctx) {

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

@@ -12,7 +12,7 @@
 
 namespace Service::SM {
 
-void Controller::ConvertSessionToDomain(Kernel::HLERequestContext& ctx) {
+void Controller::ConvertCurrentObjectToDomain(Kernel::HLERequestContext& ctx) {
     ASSERT_MSG(ctx.Session()->IsSession(), "Session is already a domain");
     LOG_DEBUG(Service, "called, server_session={}", ctx.Session()->GetObjectId());
     ctx.Session()->ConvertToDomain();
@@ -22,7 +22,7 @@ void Controller::ConvertSessionToDomain(Kernel::HLERequestContext& ctx) {
     rb.Push<u32>(1); // Converted sessions start with 1 request handler
 }
 
-void Controller::DuplicateSession(Kernel::HLERequestContext& ctx) {
+void Controller::CloneCurrentObject(Kernel::HLERequestContext& ctx) {
     // TODO(bunnei): This is just creating a new handle to the same Session. I assume this is wrong
     // and that we probably want to actually make an entirely new Session, but we still need to
     // verify this on hardware.
@@ -33,10 +33,10 @@ void Controller::DuplicateSession(Kernel::HLERequestContext& ctx) {
     rb.PushMoveObjects(ctx.Session()->GetParent()->Client());
 }
 
-void Controller::DuplicateSessionEx(Kernel::HLERequestContext& ctx) {
-    LOG_WARNING(Service, "(STUBBED) called, using DuplicateSession");
+void Controller::CloneCurrentObjectEx(Kernel::HLERequestContext& ctx) {
+    LOG_WARNING(Service, "(STUBBED) called, using CloneCurrentObject");
 
-    DuplicateSession(ctx);
+    CloneCurrentObject(ctx);
 }
 
 void Controller::QueryPointerBufferSize(Kernel::HLERequestContext& ctx) {
@@ -47,13 +47,14 @@ void Controller::QueryPointerBufferSize(Kernel::HLERequestContext& ctx) {
     rb.Push<u16>(0x1000);
 }
 
+// https://switchbrew.org/wiki/IPC_Marshalling
 Controller::Controller() : ServiceFramework("IpcController") {
     static const FunctionInfo functions[] = {
-        {0x00000000, &Controller::ConvertSessionToDomain, "ConvertSessionToDomain"},
-        {0x00000001, nullptr, "ConvertDomainToSession"},
-        {0x00000002, &Controller::DuplicateSession, "DuplicateSession"},
-        {0x00000003, &Controller::QueryPointerBufferSize, "QueryPointerBufferSize"},
-        {0x00000004, &Controller::DuplicateSessionEx, "DuplicateSessionEx"},
+        {0, &Controller::ConvertCurrentObjectToDomain, "ConvertCurrentObjectToDomain"},
+        {1, nullptr, "CopyFromCurrentDomain"},
+        {2, &Controller::CloneCurrentObject, "CloneCurrentObject"},
+        {3, &Controller::QueryPointerBufferSize, "QueryPointerBufferSize"},
+        {4, &Controller::CloneCurrentObjectEx, "CloneCurrentObjectEx"},
     };
     RegisterHandlers(functions);
 }

src/core/hle/service/sm/controller.h

@@ -14,9 +14,9 @@ public:
     ~Controller() override;
 
 private:
-    void ConvertSessionToDomain(Kernel::HLERequestContext& ctx);
-    void DuplicateSession(Kernel::HLERequestContext& ctx);
-    void DuplicateSessionEx(Kernel::HLERequestContext& ctx);
+    void ConvertCurrentObjectToDomain(Kernel::HLERequestContext& ctx);
+    void CloneCurrentObject(Kernel::HLERequestContext& ctx);
+    void CloneCurrentObjectEx(Kernel::HLERequestContext& ctx);
     void QueryPointerBufferSize(Kernel::HLERequestContext& ctx);
 };
 

src/core/hle/service/sockets/nsd.cpp

@@ -14,6 +14,7 @@ NSD::NSD(const char* name) : ServiceFramework(name) {
         {12, nullptr, "GetDeviceId"},
         {13, nullptr, "DeleteSettings"},
         {14, nullptr, "ImportSettings"},
+        {15, nullptr, "SetChangeEnvironmentIdentifierDisabled"},
         {20, nullptr, "Resolve"},
         {21, nullptr, "ResolveEx"},
         {30, nullptr, "GetNasServiceSetting"},
@@ -28,6 +29,11 @@ NSD::NSD(const char* name) : ServiceFramework(name) {
         {60, nullptr, "ReadSaveDataFromFsForTest"},
         {61, nullptr, "WriteSaveDataToFsForTest"},
         {62, nullptr, "DeleteSaveDataOfFsForTest"},
+        {63, nullptr, "IsChangeEnvironmentIdentifierDisabled"},
+        {64, nullptr, "SetWithoutDomainExchangeFqdns"},
+        {100, nullptr, "GetApplicationServerEnvironmentType"},
+        {101, nullptr, "SetApplicationServerEnvironmentType"},
+        {102, nullptr, "DeleteApplicationServerEnvironmentType"},
     };
     // clang-format on
 

src/core/hle/service/sockets/sfdnsres.cpp

@@ -7,7 +7,7 @@
 
 namespace Service::Sockets {
 
-void SFDNSRES::GetAddrInfo(Kernel::HLERequestContext& ctx) {
+void SFDNSRES::GetAddrInfoRequest(Kernel::HLERequestContext& ctx) {
     struct Parameters {
         u8 use_nsd_resolve;
         u32 unknown;
@@ -29,15 +29,20 @@ SFDNSRES::SFDNSRES() : ServiceFramework("sfdnsres") {
     static const FunctionInfo functions[] = {
         {0, nullptr, "SetDnsAddressesPrivate"},
         {1, nullptr, "GetDnsAddressPrivate"},
-        {2, nullptr, "GetHostByName"},
-        {3, nullptr, "GetHostByAddr"},
-        {4, nullptr, "GetHostStringError"},
-        {5, nullptr, "GetGaiStringError"},
-        {6, &SFDNSRES::GetAddrInfo, "GetAddrInfo"},
-        {7, nullptr, "GetNameInfo"},
-        {8, nullptr, "RequestCancelHandle"},
-        {9, nullptr, "CancelSocketCall"},
-        {11, nullptr, "ClearDnsIpServerAddressArray"},
+        {2, nullptr, "GetHostByNameRequest"},
+        {3, nullptr, "GetHostByAddrRequest"},
+        {4, nullptr, "GetHostStringErrorRequest"},
+        {5, nullptr, "GetGaiStringErrorRequest"},
+        {6, &SFDNSRES::GetAddrInfoRequest, "GetAddrInfoRequest"},
+        {7, nullptr, "GetNameInfoRequest"},
+        {8, nullptr, "RequestCancelHandleRequest"},
+        {9, nullptr, "CancelRequest"},
+        {10, nullptr, "GetHostByNameRequestWithOptions"},
+        {11, nullptr, "GetHostByAddrRequestWithOptions"},
+        {12, nullptr, "GetAddrInfoRequestWithOptions"},
+        {13, nullptr, "GetNameInfoRequestWithOptions"},
+        {14, nullptr, "ResolverSetOptionRequest"},
+        {15, nullptr, "ResolverGetOptionRequest"},
     };
     RegisterHandlers(functions);
 }

src/core/hle/service/sockets/sfdnsres.h

@@ -15,7 +15,7 @@ public:
     ~SFDNSRES() override;
 
 private:
-    void GetAddrInfo(Kernel::HLERequestContext& ctx);
+    void GetAddrInfoRequest(Kernel::HLERequestContext& ctx);
 };
 
 } // namespace Service::Sockets

src/core/hle/service/spl/module.cpp

@@ -19,7 +19,7 @@ namespace Service::SPL {
 
 Module::Interface::Interface(std::shared_ptr<Module> module, const char* name)
     : ServiceFramework(name), module(std::move(module)),
-      rng(Settings::values.rng_seed.value_or(std::time(nullptr))) {}
+      rng(Settings::values.rng_seed.GetValue().value_or(std::time(nullptr))) {}
 
 Module::Interface::~Interface() = default;
 

src/core/hle/service/spl/spl.cpp

@@ -9,35 +9,36 @@ namespace Service::SPL {
 SPL::SPL(std::shared_ptr<Module> module) : Module::Interface(std::move(module), "spl:") {
     static const FunctionInfo functions[] = {
         {0, nullptr, "GetConfig"},
-        {1, nullptr, "UserExpMod"},
+        {1, nullptr, "ModularExponentiate"},
         {2, nullptr, "GenerateAesKek"},
         {3, nullptr, "LoadAesKey"},
         {4, nullptr, "GenerateAesKey"},
         {5, nullptr, "SetConfig"},
         {7, &SPL::GetRandomBytes, "GetRandomBytes"},
-        {9, nullptr, "LoadSecureExpModKey"},
-        {10, nullptr, "SecureExpMod"},
+        {9, nullptr, "ImportLotusKey"},
+        {10, nullptr, "DecryptLotusMessage"},
         {11, nullptr, "IsDevelopment"},
         {12, nullptr, "GenerateSpecificAesKey"},
-        {13, nullptr, "DecryptPrivk"},
+        {13, nullptr, "DecryptDeviceUniqueData"},
         {14, nullptr, "DecryptAesKey"},
-        {15, nullptr, "DecryptAesCtr"},
+        {15, nullptr, "CryptAesCtr"},
         {16, nullptr, "ComputeCmac"},
-        {17, nullptr, "LoadRsaOaepKey"},
-        {18, nullptr, "UnwrapRsaOaepWrappedTitleKey"},
+        {17, nullptr, "ImportEsKey"},
+        {18, nullptr, "UnwrapTitleKey"},
         {19, nullptr, "LoadTitleKey"},
-        {20, nullptr, "UnwrapAesWrappedTitleKey"},
-        {21, nullptr, "LockAesEngine"},
-        {22, nullptr, "UnlockAesEngine"},
-        {23, nullptr, "GetSplWaitEvent"},
-        {24, nullptr, "SetSharedData"},
-        {25, nullptr, "GetSharedData"},
-        {26, nullptr, "ImportSslRsaKey"},
-        {27, nullptr, "SecureExpModWithSslKey"},
-        {28, nullptr, "ImportEsRsaKey"},
-        {29, nullptr, "SecureExpModWithEsKey"},
-        {30, nullptr, "EncryptManuRsaKeyForImport"},
-        {31, nullptr, "GetPackage2Hash"},
+        {20, nullptr, "PrepareEsCommonKey"},
+        {21, nullptr, "AllocateAesKeyslot"},
+        {22, nullptr, "DeallocateAesKeySlot"},
+        {23, nullptr, "GetAesKeyslotAvailableEvent"},
+        {24, nullptr, "SetBootReason"},
+        {25, nullptr, "GetBootReason"},
+        {26, nullptr, "DecryptAndStoreSslClientCertKey"},
+        {27, nullptr, "ModularExponentiateWithSslClientCertKey"},
+        {28, nullptr, "DecryptAndStoreDrmDeviceCertKey"},
+        {29, nullptr, "ModularExponentiateWithDrmDeviceCertKey"},
+        {30, nullptr, "ReencryptDeviceUniqueData "},
+        {31, nullptr, "PrepareEsArchiveKey"}, // This is also GetPackage2Hash?
+        {32, nullptr, "LoadPreparedAesKey"},
     };
     RegisterHandlers(functions);
 }

src/core/hle/service/time/time.cpp

@@ -90,6 +90,13 @@ public:
         : ServiceFramework("ISteadyClock"), clock_core{clock_core}, system{system} {
         static const FunctionInfo functions[] = {
             {0, &ISteadyClock::GetCurrentTimePoint, "GetCurrentTimePoint"},
+            {2, nullptr, "GetTestOffset"},
+            {3, nullptr, "SetTestOffset"},
+            {100, nullptr, "GetRtcValue"},
+            {101, nullptr, "IsRtcResetDetected"},
+            {102, nullptr, "GetSetupResultValue"},
+            {200, nullptr, "GetInternalOffset"},
+            {201, nullptr, "SetInternalOffset"},
         };
         RegisterHandlers(functions);
     }

src/core/hle/service/usb/usb.cpp

@@ -20,7 +20,7 @@ public:
         static const FunctionInfo functions[] = {
             {0, nullptr, "GetDsEndpoint"},
             {1, nullptr, "GetSetupEvent"},
-            {2, nullptr, "Unknown"},
+            {2, nullptr, "Unknown2"},
             {3, nullptr, "EnableInterface"},
             {4, nullptr, "DisableInterface"},
             {5, nullptr, "CtrlInPostBufferAsync"},
@@ -55,6 +55,7 @@ public:
             {9, nullptr, "SetBinaryObjectStore"},
             {10, nullptr, "Enable"},
             {11, nullptr, "Disable"},
+            {12, nullptr, "Unknown12"},
         };
         // clang-format on
 
@@ -69,13 +70,13 @@ public:
         static const FunctionInfo functions[] = {
             {0, nullptr, "Open"},
             {1, nullptr, "Close"},
-            {2, nullptr, "Unknown1"},
+            {2, nullptr, "Unknown2"},
             {3, nullptr, "Populate"},
             {4, nullptr, "PostBufferAsync"},
             {5, nullptr, "GetXferReport"},
             {6, nullptr, "PostBufferMultiAsync"},
-            {7, nullptr, "Unknown3"},
-            {8, nullptr, "Unknown4"},
+            {7, nullptr, "Unknown7"},
+            {8, nullptr, "Unknown8"},
         };
         // clang-format on
 
@@ -88,13 +89,13 @@ public:
     explicit IClientIfSession() : ServiceFramework{"IClientIfSession"} {
         // clang-format off
         static const FunctionInfo functions[] = {
-            {0, nullptr, "Unknown1"},
+            {0, nullptr, "Unknown0"},
             {1, nullptr, "SetInterface"},
             {2, nullptr, "GetInterface"},
             {3, nullptr, "GetAlternateInterface"},
             {4, nullptr, "GetCurrentFrame"},
             {5, nullptr, "CtrlXferAsync"},
-            {6, nullptr, "Unknown2"},
+            {6, nullptr, "Unknown6"},
             {7, nullptr, "GetCtrlXferReport"},
             {8, nullptr, "ResetDevice"},
             {9, nullptr, "OpenUsbEp"},
@@ -118,7 +119,7 @@ public:
             {5, nullptr, "DestroyInterfaceAvailableEvent"},
             {6, nullptr, "GetInterfaceStateChangeEvent"},
             {7, nullptr, "AcquireUsbIf"},
-            {8, nullptr, "Unknown1"},
+            {8, nullptr, "Unknown8"},
         };
         // clang-format on
 
@@ -179,8 +180,8 @@ public:
             {4, nullptr, "GetFwRevision"},
             {5, nullptr, "GetManufacturerId"},
             {6, nullptr, "GetDeviceId"},
-            {7, nullptr, "Unknown1"},
-            {8, nullptr, "Unknown2"},
+            {7, nullptr, "Unknown7"},
+            {8, nullptr, "Unknown8"},
         };
         // clang-format on
 
@@ -215,12 +216,12 @@ public:
     explicit USB_PM() : ServiceFramework{"usb:pm"} {
         // clang-format off
         static const FunctionInfo functions[] = {
-            {0, nullptr, "Unknown1"},
-            {1, nullptr, "Unknown2"},
-            {2, nullptr, "Unknown3"},
-            {3, nullptr, "Unknown4"},
-            {4, nullptr, "Unknown5"},
-            {5, nullptr, "Unknown6"},
+            {0, nullptr, "Unknown0"},
+            {1, nullptr, "Unknown1"},
+            {2, nullptr, "Unknown2"},
+            {3, nullptr, "Unknown3"},
+            {4, nullptr, "Unknown4"},
+            {5, nullptr, "Unknown5"},
         };
         // clang-format on
 

src/core/hle/service/vi/vi.cpp

@@ -519,9 +519,9 @@ private:
             IGBPConnectRequestParcel request{ctx.ReadBuffer()};
             IGBPConnectResponseParcel response{
                 static_cast<u32>(static_cast<u32>(DisplayResolution::UndockedWidth) *
-                                 Settings::values.resolution_factor),
+                                 Settings::values.resolution_factor.GetValue()),
                 static_cast<u32>(static_cast<u32>(DisplayResolution::UndockedHeight) *
-                                 Settings::values.resolution_factor)};
+                                 Settings::values.resolution_factor.GetValue())};
             ctx.WriteBuffer(response.Serialize());
             break;
         }
@@ -700,6 +700,7 @@ public:
             {3215, nullptr, "SetDisplayGamma"},
             {3216, nullptr, "GetDisplayCmuLuma"},
             {3217, nullptr, "SetDisplayCmuLuma"},
+            {6013, nullptr, "GetLayerPresentationSubmissionTimestamps"},
             {8225, nullptr, "GetSharedBufferMemoryHandleId"},
             {8250, nullptr, "OpenSharedLayer"},
             {8251, nullptr, "CloseSharedLayer"},
@@ -748,14 +749,14 @@ private:
 
         if (Settings::values.use_docked_mode) {
             rb.Push(static_cast<u32>(Service::VI::DisplayResolution::DockedWidth) *
-                    static_cast<u32>(Settings::values.resolution_factor));
+                    static_cast<u32>(Settings::values.resolution_factor.GetValue()));
             rb.Push(static_cast<u32>(Service::VI::DisplayResolution::DockedHeight) *
-                    static_cast<u32>(Settings::values.resolution_factor));
+                    static_cast<u32>(Settings::values.resolution_factor.GetValue()));
         } else {
             rb.Push(static_cast<u32>(Service::VI::DisplayResolution::UndockedWidth) *
-                    static_cast<u32>(Settings::values.resolution_factor));
+                    static_cast<u32>(Settings::values.resolution_factor.GetValue()));
             rb.Push(static_cast<u32>(Service::VI::DisplayResolution::UndockedHeight) *
-                    static_cast<u32>(Settings::values.resolution_factor));
+                    static_cast<u32>(Settings::values.resolution_factor.GetValue()));
         }
 
         rb.PushRaw<float>(60.0f); // This wouldn't seem to be correct for 30 fps games.
@@ -785,6 +786,7 @@ public:
             {2300, nullptr, "AcquireLayerTexturePresentingEvent"},
             {2301, nullptr, "ReleaseLayerTexturePresentingEvent"},
             {2302, nullptr, "GetDisplayHotplugEvent"},
+            {2303, nullptr, "GetDisplayModeChangedEvent"},
             {2402, nullptr, "GetDisplayHotplugState"},
             {2501, nullptr, "GetCompositorErrorInfo"},
             {2601, nullptr, "GetDisplayErrorEvent"},
@@ -1029,9 +1031,9 @@ private:
         // between docked and undocked dimensions. We take the liberty of applying
         // the resolution scaling factor here.
         rb.Push(static_cast<u64>(DisplayResolution::UndockedWidth) *
-                static_cast<u32>(Settings::values.resolution_factor));
+                static_cast<u32>(Settings::values.resolution_factor.GetValue()));
         rb.Push(static_cast<u64>(DisplayResolution::UndockedHeight) *
-                static_cast<u32>(Settings::values.resolution_factor));
+                static_cast<u32>(Settings::values.resolution_factor.GetValue()));
     }
 
     void SetLayerScalingMode(Kernel::HLERequestContext& ctx) {
@@ -1064,8 +1066,8 @@ private:
         LOG_WARNING(Service_VI, "(STUBBED) called");
 
         DisplayInfo display_info;
-        display_info.width *= static_cast<u64>(Settings::values.resolution_factor);
-        display_info.height *= static_cast<u64>(Settings::values.resolution_factor);
+        display_info.width *= static_cast<u64>(Settings::values.resolution_factor.GetValue());
+        display_info.height *= static_cast<u64>(Settings::values.resolution_factor.GetValue());
         ctx.WriteBuffer(&display_info, sizeof(DisplayInfo));
         IPC::ResponseBuilder rb{ctx, 4};
         rb.Push(RESULT_SUCCESS);

src/core/hle/service/vi/vi_u.cpp

@@ -12,6 +12,7 @@ VI_U::VI_U(std::shared_ptr<NVFlinger::NVFlinger> nv_flinger)
     : ServiceFramework{"vi:u"}, nv_flinger{std::move(nv_flinger)} {
     static const FunctionInfo functions[] = {
         {0, &VI_U::GetDisplayService, "GetDisplayService"},
+        {1, nullptr, "GetDisplayServiceWithProxyNameExchange"},
     };
     RegisterHandlers(functions);
 }

src/core/hle/service/wlan/wlan.cpp

@@ -15,34 +15,37 @@ public:
     explicit WLANInfra() : ServiceFramework{"wlan:inf"} {
         // clang-format off
         static const FunctionInfo functions[] = {
-            {0, nullptr, "Unknown1"},
-            {1, nullptr, "Unknown2"},
+            {0, nullptr, "OpenMode"},
+            {1, nullptr, "CloseMode"},
             {2, nullptr, "GetMacAddress"},
             {3, nullptr, "StartScan"},
             {4, nullptr, "StopScan"},
             {5, nullptr, "Connect"},
             {6, nullptr, "CancelConnect"},
             {7, nullptr, "Disconnect"},
-            {8, nullptr, "Unknown3"},
-            {9, nullptr, "Unknown4"},
+            {8, nullptr, "GetConnectionEvent"},
+            {9, nullptr, "GetConnectionStatus"},
             {10, nullptr, "GetState"},
             {11, nullptr, "GetScanResult"},
             {12, nullptr, "GetRssi"},
             {13, nullptr, "ChangeRxAntenna"},
-            {14, nullptr, "Unknown5"},
-            {15, nullptr, "Unknown6"},
+            {14, nullptr, "GetFwVersion"},
+            {15, nullptr, "RequestSleep"},
             {16, nullptr, "RequestWakeUp"},
             {17, nullptr, "RequestIfUpDown"},
-            {18, nullptr, "Unknown7"},
-            {19, nullptr, "Unknown8"},
-            {20, nullptr, "Unknown9"},
-            {21, nullptr, "Unknown10"},
-            {22, nullptr, "Unknown11"},
-            {23, nullptr, "Unknown12"},
-            {24, nullptr, "Unknown13"},
-            {25, nullptr, "Unknown14"},
-            {26, nullptr, "Unknown15"},
-            {27, nullptr, "Unknown16"},
+            {18, nullptr, "Unknown18"},
+            {19, nullptr, "Unknown19"},
+            {20, nullptr, "Unknown20"},
+            {21, nullptr, "Unknown21"},
+            {22, nullptr, "Unknown22"},
+            {23, nullptr, "Unknown23"},
+            {24, nullptr, "Unknown24"},
+            {25, nullptr, "Unknown25"},
+            {26, nullptr, "Unknown26"},
+            {27, nullptr, "Unknown27"},
+            {28, nullptr, "Unknown28"},
+            {29, nullptr, "Unknown29"},
+            {30, nullptr, "Unknown30"},
         };
         // clang-format on
 
@@ -55,12 +58,12 @@ public:
     explicit WLANLocal() : ServiceFramework{"wlan:lcl"} {
         // clang-format off
         static const FunctionInfo functions[] = {
-            {0, nullptr, "Unknown1"},
-            {1, nullptr, "Unknown2"},
-            {2, nullptr, "Unknown3"},
-            {3, nullptr, "Unknown4"},
-            {4, nullptr, "Unknown5"},
-            {5, nullptr, "Unknown6"},
+            {0, nullptr, "Unknown0"},
+            {1, nullptr, "Unknown1"},
+            {2, nullptr, "Unknown2"},
+            {3, nullptr, "Unknown3"},
+            {4, nullptr, "Unknown4"},
+            {5, nullptr, "Unknown5"},
             {6, nullptr, "GetMacAddress"},
             {7, nullptr, "CreateBss"},
             {8, nullptr, "DestroyBss"},
@@ -72,38 +75,42 @@ public:
             {14, nullptr, "CancelJoin"},
             {15, nullptr, "Disconnect"},
             {16, nullptr, "SetBeaconLostCount"},
-            {17, nullptr, "Unknown7"},
-            {18, nullptr, "Unknown8"},
-            {19, nullptr, "Unknown9"},
+            {17, nullptr, "Unknown17"},
+            {18, nullptr, "Unknown18"},
+            {19, nullptr, "Unknown19"},
             {20, nullptr, "GetBssIndicationEvent"},
             {21, nullptr, "GetBssIndicationInfo"},
             {22, nullptr, "GetState"},
             {23, nullptr, "GetAllowedChannels"},
             {24, nullptr, "AddIe"},
             {25, nullptr, "DeleteIe"},
-            {26, nullptr, "Unknown10"},
-            {27, nullptr, "Unknown11"},
+            {26, nullptr, "Unknown26"},
+            {27, nullptr, "Unknown27"},
             {28, nullptr, "CreateRxEntry"},
             {29, nullptr, "DeleteRxEntry"},
-            {30, nullptr, "Unknown12"},
-            {31, nullptr, "Unknown13"},
+            {30, nullptr, "Unknown30"},
+            {31, nullptr, "Unknown31"},
             {32, nullptr, "AddMatchingDataToRxEntry"},
             {33, nullptr, "RemoveMatchingDataFromRxEntry"},
             {34, nullptr, "GetScanResult"},
-            {35, nullptr, "Unknown14"},
+            {35, nullptr, "Unknown35"},
             {36, nullptr, "SetActionFrameWithBeacon"},
             {37, nullptr, "CancelActionFrameWithBeacon"},
             {38, nullptr, "CreateRxEntryForActionFrame"},
             {39, nullptr, "DeleteRxEntryForActionFrame"},
-            {40, nullptr, "Unknown15"},
-            {41, nullptr, "Unknown16"},
+            {40, nullptr, "Unknown40"},
+            {41, nullptr, "Unknown41"},
             {42, nullptr, "CancelGetActionFrame"},
             {43, nullptr, "GetRssi"},
-            {44, nullptr, "Unknown17"},
-            {45, nullptr, "Unknown18"},
-            {46, nullptr, "Unknown19"},
-            {47, nullptr, "Unknown20"},
-            {48, nullptr, "Unknown21"},
+            {44, nullptr, "Unknown44"},
+            {45, nullptr, "Unknown45"},
+            {46, nullptr, "Unknown46"},
+            {47, nullptr, "Unknown47"},
+            {48, nullptr, "Unknown48"},
+            {49, nullptr, "Unknown49"},
+            {50, nullptr, "Unknown50"},
+            {51, nullptr, "Unknown51"},
+            {52, nullptr, "Unknown52"},
         };
         // clang-format on
 
@@ -142,18 +149,19 @@ public:
     explicit WLANSocketManager() : ServiceFramework{"wlan:soc"} {
         // clang-format off
         static const FunctionInfo functions[] = {
-            {0, nullptr, "Unknown1"},
-            {1, nullptr, "Unknown2"},
-            {2, nullptr, "Unknown3"},
-            {3, nullptr, "Unknown4"},
-            {4, nullptr, "Unknown5"},
-            {5, nullptr, "Unknown6"},
+            {0, nullptr, "Unknown0"},
+            {1, nullptr, "Unknown1"},
+            {2, nullptr, "Unknown2"},
+            {3, nullptr, "Unknown3"},
+            {4, nullptr, "Unknown4"},
+            {5, nullptr, "Unknown5"},
             {6, nullptr, "GetMacAddress"},
             {7, nullptr, "SwitchTsfTimerFunction"},
-            {8, nullptr, "Unknown7"},
-            {9, nullptr, "Unknown8"},
-            {10, nullptr, "Unknown9"},
-            {11, nullptr, "Unknown10"},
+            {8, nullptr, "Unknown8"},
+            {9, nullptr, "Unknown9"},
+            {10, nullptr, "Unknown10"},
+            {11, nullptr, "Unknown11"},
+            {12, nullptr, "Unknown12"},
         };
         // clang-format on
 

src/core/perf_stats.cpp

@@ -119,13 +119,14 @@ double PerfStats::GetLastFrameTimeScale() {
 }
 
 void FrameLimiter::DoFrameLimiting(microseconds current_system_time_us) {
-    if (!Settings::values.use_frame_limit || Settings::values.use_multi_core) {
+    if (!Settings::values.use_frame_limit.GetValue() ||
+        Settings::values.use_multi_core.GetValue()) {
         return;
     }
 
     auto now = Clock::now();
 
-    const double sleep_scale = Settings::values.frame_limit / 100.0;
+    const double sleep_scale = Settings::values.frame_limit.GetValue() / 100.0;
 
     // Max lag caused by slow frames. Shouldn't be more than the length of a frame at the current
     // speed percent or it will clamp too much and prevent this from properly limiting to that

src/core/settings.cpp

@@ -62,6 +62,7 @@ const std::array<const char*, NumMouseButtons> mapping = {{
 }
 
 Values values = {};
+bool configuring_global = true;
 
 std::string GetTimeZoneString() {
     static constexpr std::array<const char*, 46> timezones{{
@@ -73,9 +74,9 @@ std::string GetTimeZoneString() {
         "UCT",       "Universal", "UTC", "W-SU",    "WET",     "Zulu",
     }};
 
-    ASSERT(Settings::values.time_zone_index < timezones.size());
+    ASSERT(Settings::values.time_zone_index.GetValue() < timezones.size());
 
-    return timezones[Settings::values.time_zone_index];
+    return timezones[Settings::values.time_zone_index.GetValue()];
 }
 
 void Apply() {
@@ -97,25 +98,25 @@ void LogSetting(const std::string& name, const T& value) {
 
 void LogSettings() {
     LOG_INFO(Config, "yuzu Configuration:");
-    LogSetting("System_UseDockedMode", Settings::values.use_docked_mode);
-    LogSetting("System_RngSeed", Settings::values.rng_seed.value_or(0));
+    LogSetting("Controls_UseDockedMode", Settings::values.use_docked_mode);
+    LogSetting("System_RngSeed", Settings::values.rng_seed.GetValue().value_or(0));
     LogSetting("System_CurrentUser", Settings::values.current_user);
-    LogSetting("System_LanguageIndex", Settings::values.language_index);
-    LogSetting("System_RegionIndex", Settings::values.region_index);
-    LogSetting("System_TimeZoneIndex", Settings::values.time_zone_index);
-    LogSetting("Core_UseMultiCore", Settings::values.use_multi_core);
-    LogSetting("Renderer_UseResolutionFactor", Settings::values.resolution_factor);
-    LogSetting("Renderer_UseFrameLimit", Settings::values.use_frame_limit);
-    LogSetting("Renderer_FrameLimit", Settings::values.frame_limit);
-    LogSetting("Renderer_UseDiskShaderCache", Settings::values.use_disk_shader_cache);
-    LogSetting("Renderer_GPUAccuracyLevel", Settings::values.gpu_accuracy);
+    LogSetting("System_LanguageIndex", Settings::values.language_index.GetValue());
+    LogSetting("System_RegionIndex", Settings::values.region_index.GetValue());
+    LogSetting("System_TimeZoneIndex", Settings::values.time_zone_index.GetValue());
+    LogSetting("Core_UseMultiCore", Settings::values.use_multi_core.GetValue());
+    LogSetting("Renderer_UseResolutionFactor", Settings::values.resolution_factor.GetValue());
+    LogSetting("Renderer_UseFrameLimit", Settings::values.use_frame_limit.GetValue());
+    LogSetting("Renderer_FrameLimit", Settings::values.frame_limit.GetValue());
+    LogSetting("Renderer_UseDiskShaderCache", Settings::values.use_disk_shader_cache.GetValue());
+    LogSetting("Renderer_GPUAccuracyLevel", Settings::values.gpu_accuracy.GetValue());
     LogSetting("Renderer_UseAsynchronousGpuEmulation",
-               Settings::values.use_asynchronous_gpu_emulation);
-    LogSetting("Renderer_UseVsync", Settings::values.use_vsync);
-    LogSetting("Renderer_UseAssemblyShaders", Settings::values.use_assembly_shaders);
-    LogSetting("Renderer_AnisotropicFilteringLevel", Settings::values.max_anisotropy);
+               Settings::values.use_asynchronous_gpu_emulation.GetValue());
+    LogSetting("Renderer_UseVsync", Settings::values.use_vsync.GetValue());
+    LogSetting("Renderer_UseAssemblyShaders", Settings::values.use_assembly_shaders.GetValue());
+    LogSetting("Renderer_AnisotropicFilteringLevel", Settings::values.max_anisotropy.GetValue());
     LogSetting("Audio_OutputEngine", Settings::values.sink_id);
-    LogSetting("Audio_EnableAudioStretching", Settings::values.enable_audio_stretching);
+    LogSetting("Audio_EnableAudioStretching", Settings::values.enable_audio_stretching.GetValue());
     LogSetting("Audio_OutputDevice", Settings::values.audio_device_id);
     LogSetting("DataStorage_UseVirtualSd", Settings::values.use_virtual_sd);
     LogSetting("DataStorage_NandDir", FileUtil::GetUserPath(FileUtil::UserPath::NANDDir));
@@ -131,15 +132,56 @@ float Volume() {
     if (values.audio_muted) {
         return 0.0f;
     }
-    return values.volume;
+    return values.volume.GetValue();
 }
 
 bool IsGPULevelExtreme() {
-    return values.gpu_accuracy == GPUAccuracy::Extreme;
+    return values.gpu_accuracy.GetValue() == GPUAccuracy::Extreme;
 }
 
 bool IsGPULevelHigh() {
-    return values.gpu_accuracy == GPUAccuracy::Extreme || values.gpu_accuracy == GPUAccuracy::High;
+    return values.gpu_accuracy.GetValue() == GPUAccuracy::Extreme ||
+           values.gpu_accuracy.GetValue() == GPUAccuracy::High;
+}
+
+void RestoreGlobalState() {
+    // If a game is running, DO NOT restore the global settings state
+    if (Core::System::GetInstance().IsPoweredOn()) {
+        return;
+    }
+
+    // Audio
+    values.enable_audio_stretching.SetGlobal(true);
+    values.volume.SetGlobal(true);
+
+    // Core
+    values.use_multi_core.SetGlobal(true);
+
+    // Renderer
+    values.renderer_backend.SetGlobal(true);
+    values.vulkan_device.SetGlobal(true);
+    values.aspect_ratio.SetGlobal(true);
+    values.max_anisotropy.SetGlobal(true);
+    values.use_frame_limit.SetGlobal(true);
+    values.frame_limit.SetGlobal(true);
+    values.use_disk_shader_cache.SetGlobal(true);
+    values.gpu_accuracy.SetGlobal(true);
+    values.use_asynchronous_gpu_emulation.SetGlobal(true);
+    values.use_vsync.SetGlobal(true);
+    values.use_assembly_shaders.SetGlobal(true);
+    values.use_fast_gpu_time.SetGlobal(true);
+    values.force_30fps_mode.SetGlobal(true);
+    values.bg_red.SetGlobal(true);
+    values.bg_green.SetGlobal(true);
+    values.bg_blue.SetGlobal(true);
+
+    // System
+    values.language_index.SetGlobal(true);
+    values.region_index.SetGlobal(true);
+    values.time_zone_index.SetGlobal(true);
+    values.rng_seed.SetGlobal(true);
+    values.custom_rtc.SetGlobal(true);
+    values.sound_index.SetGlobal(true);
 }
 
 } // namespace Settings

src/core/settings.h

@@ -346,31 +346,6 @@ struct TouchscreenInput {
     u32 rotation_angle;
 };
 
-enum class NANDTotalSize : u64 {
-    S29_1GB = 0x747C00000ULL,
-};
-
-enum class NANDUserSize : u64 {
-    S26GB = 0x680000000ULL,
-};
-
-enum class NANDSystemSize : u64 {
-    S2_5GB = 0xA0000000,
-};
-
-enum class SDMCSize : u64 {
-    S1GB = 0x40000000,
-    S2GB = 0x80000000,
-    S4GB = 0x100000000ULL,
-    S8GB = 0x200000000ULL,
-    S16GB = 0x400000000ULL,
-    S32GB = 0x800000000ULL,
-    S64GB = 0x1000000000ULL,
-    S128GB = 0x2000000000ULL,
-    S256GB = 0x4000000000ULL,
-    S1TB = 0x10000000000ULL,
-};
-
 enum class RendererBackend {
     OpenGL = 0,
     Vulkan = 1,
@@ -382,20 +357,102 @@ enum class GPUAccuracy : u32 {
     Extreme = 2,
 };
 
+enum class CPUAccuracy {
+    Accurate = 0,
+    DebugMode = 1,
+};
+
+extern bool configuring_global;
+
+template <typename Type>
+class Setting final {
+public:
+    Setting() = default;
+    explicit Setting(Type val) : global{val} {}
+    ~Setting() = default;
+    void SetGlobal(bool to_global) {
+        use_global = to_global;
+    }
+    bool UsingGlobal() const {
+        return use_global;
+    }
+    Type GetValue(bool need_global = false) const {
+        if (use_global || need_global) {
+            return global;
+        }
+        return local;
+    }
+    void SetValue(const Type& value) {
+        if (use_global) {
+            global = value;
+        } else {
+            local = value;
+        }
+    }
+
+private:
+    bool use_global = true;
+    Type global{};
+    Type local{};
+};
+
 struct Values {
+    // Audio
+    std::string audio_device_id;
+    std::string sink_id;
+    bool audio_muted;
+    Setting<bool> enable_audio_stretching;
+    Setting<float> volume;
+
+    // Core
+    Setting<bool> use_multi_core;
+
+    // Cpu
+    CPUAccuracy cpu_accuracy;
+
+    bool cpuopt_page_tables;
+    bool cpuopt_block_linking;
+    bool cpuopt_return_stack_buffer;
+    bool cpuopt_fast_dispatcher;
+    bool cpuopt_context_elimination;
+    bool cpuopt_const_prop;
+    bool cpuopt_misc_ir;
+    bool cpuopt_reduce_misalign_checks;
+
+    // Renderer
+    Setting<RendererBackend> renderer_backend;
+    bool renderer_debug;
+    Setting<int> vulkan_device;
+
+    Setting<u16> resolution_factor = Setting(static_cast<u16>(1));
+    Setting<int> aspect_ratio;
+    Setting<int> max_anisotropy;
+    Setting<bool> use_frame_limit;
+    Setting<u16> frame_limit;
+    Setting<bool> use_disk_shader_cache;
+    Setting<GPUAccuracy> gpu_accuracy;
+    Setting<bool> use_asynchronous_gpu_emulation;
+    Setting<bool> use_vsync;
+    Setting<bool> use_assembly_shaders;
+    Setting<bool> force_30fps_mode;
+    Setting<bool> use_fast_gpu_time;
+
+    Setting<float> bg_red;
+    Setting<float> bg_green;
+    Setting<float> bg_blue;
+
     // System
-    bool use_docked_mode;
-    std::optional<u32> rng_seed;
+    Setting<std::optional<u32>> rng_seed;
     // Measured in seconds since epoch
-    std::optional<std::chrono::seconds> custom_rtc;
+    Setting<std::optional<std::chrono::seconds>> custom_rtc;
     // Set on game boot, reset on stop. Seconds difference between current time and `custom_rtc`
     std::chrono::seconds custom_rtc_differential;
 
     s32 current_user;
-    s32 language_index;
-    s32 region_index;
-    s32 time_zone_index;
-    s32 sound_index;
+    Setting<s32> language_index;
+    Setting<s32> region_index;
+    Setting<s32> time_zone_index;
+    Setting<s32> sound_index;
 
     // Controls
     std::array<PlayerInput, 10> players;
@@ -419,51 +476,13 @@ struct Values {
     u16 udp_input_port;
     u8 udp_pad_index;
 
-    // Core
-    bool use_multi_core;
+    bool use_docked_mode;
 
     // Data Storage
     bool use_virtual_sd;
     bool gamecard_inserted;
     bool gamecard_current_game;
     std::string gamecard_path;
-    NANDTotalSize nand_total_size;
-    NANDSystemSize nand_system_size;
-    NANDUserSize nand_user_size;
-    SDMCSize sdmc_size;
-
-    // Renderer
-    RendererBackend renderer_backend;
-    bool renderer_debug;
-    int vulkan_device;
-
-    u16 resolution_factor{1};
-    int aspect_ratio;
-    int max_anisotropy;
-    bool use_frame_limit;
-    u16 frame_limit;
-    bool use_disk_shader_cache;
-    GPUAccuracy gpu_accuracy;
-    bool use_asynchronous_gpu_emulation;
-    bool use_vsync;
-    bool use_assembly_shaders;
-    bool force_30fps_mode;
-    bool use_fast_gpu_time;
-
-    float bg_red;
-    float bg_green;
-    float bg_blue;
-
-    std::string log_filter;
-
-    bool use_dev_keys;
-
-    // Audio
-    bool audio_muted;
-    std::string sink_id;
-    bool enable_audio_stretching;
-    std::string audio_device_id;
-    float volume;
 
     // Debugging
     bool record_frame_times;
@@ -474,10 +493,13 @@ struct Values {
     bool dump_nso;
     bool reporting_services;
     bool quest_flag;
-    bool disable_cpu_opt;
     bool disable_macro_jit;
 
-    // BCAT
+    // Misceallaneous
+    std::string log_filter;
+    bool use_dev_keys;
+
+    // Services
     std::string bcat_backend;
     bool bcat_boxcat_local;
 
@@ -501,4 +523,7 @@ std::string GetTimeZoneString();
 void Apply();
 void LogSettings();
 
+// Restore the global state of all applicable settings in the Values struct
+void RestoreGlobalState();
+
 } // namespace Settings

src/core/telemetry_session.cpp

@@ -189,19 +189,24 @@ void TelemetrySession::AddInitialInfo(Loader::AppLoader& app_loader) {
     // Log user configuration information
     constexpr auto field_type = Telemetry::FieldType::UserConfig;
     AddField(field_type, "Audio_SinkId", Settings::values.sink_id);
-    AddField(field_type, "Audio_EnableAudioStretching", Settings::values.enable_audio_stretching);
-    AddField(field_type, "Core_UseMultiCore", Settings::values.use_multi_core);
-    AddField(field_type, "Renderer_Backend", TranslateRenderer(Settings::values.renderer_backend));
-    AddField(field_type, "Renderer_ResolutionFactor", Settings::values.resolution_factor);
-    AddField(field_type, "Renderer_UseFrameLimit", Settings::values.use_frame_limit);
-    AddField(field_type, "Renderer_FrameLimit", Settings::values.frame_limit);
-    AddField(field_type, "Renderer_UseDiskShaderCache", Settings::values.use_disk_shader_cache);
+    AddField(field_type, "Audio_EnableAudioStretching",
+             Settings::values.enable_audio_stretching.GetValue());
+    AddField(field_type, "Core_UseMultiCore", Settings::values.use_multi_core.GetValue());
+    AddField(field_type, "Renderer_Backend",
+             TranslateRenderer(Settings::values.renderer_backend.GetValue()));
+    AddField(field_type, "Renderer_ResolutionFactor",
+             Settings::values.resolution_factor.GetValue());
+    AddField(field_type, "Renderer_UseFrameLimit", Settings::values.use_frame_limit.GetValue());
+    AddField(field_type, "Renderer_FrameLimit", Settings::values.frame_limit.GetValue());
+    AddField(field_type, "Renderer_UseDiskShaderCache",
+             Settings::values.use_disk_shader_cache.GetValue());
     AddField(field_type, "Renderer_GPUAccuracyLevel",
-             TranslateGPUAccuracyLevel(Settings::values.gpu_accuracy));
+             TranslateGPUAccuracyLevel(Settings::values.gpu_accuracy.GetValue()));
     AddField(field_type, "Renderer_UseAsynchronousGpuEmulation",
-             Settings::values.use_asynchronous_gpu_emulation);
-    AddField(field_type, "Renderer_UseVsync", Settings::values.use_vsync);
-    AddField(field_type, "Renderer_UseAssemblyShaders", Settings::values.use_assembly_shaders);
+             Settings::values.use_asynchronous_gpu_emulation.GetValue());
+    AddField(field_type, "Renderer_UseVsync", Settings::values.use_vsync.GetValue());
+    AddField(field_type, "Renderer_UseAssemblyShaders",
+             Settings::values.use_assembly_shaders.GetValue());
     AddField(field_type, "System_UseDockedMode", Settings::values.use_docked_mode);
 }
 

src/input_common/CMakeLists.txt

@@ -30,7 +30,8 @@ if(SDL2_FOUND)
     target_compile_definitions(input_common PRIVATE HAVE_SDL2)
 endif()
 
-target_link_libraries(input_common PUBLIC ${LIBUSB_LIBRARIES})
+target_include_directories(input_common SYSTEM PRIVATE ${LIBUSB_INCLUDE_DIR})
+target_link_libraries(input_common PRIVATE ${LIBUSB_LIBRARIES})
 
 create_target_directory_groups(input_common)
 target_link_libraries(input_common PUBLIC core PRIVATE common Boost::boost)

src/input_common/gcadapter/gc_adapter.cpp

@@ -4,6 +4,7 @@
 
 #include <chrono>
 #include <thread>
+#include <libusb.h>
 #include "common/logging/log.h"
 #include "input_common/gcadapter/gc_adapter.h"
 
@@ -24,12 +25,16 @@ Adapter::Adapter() {
     LOG_INFO(Input, "GC Adapter Initialization started");
 
     current_status = NO_ADAPTER_DETECTED;
-    libusb_init(&libusb_ctx);
 
-    StartScanThread();
+    const int init_res = libusb_init(&libusb_ctx);
+    if (init_res == LIBUSB_SUCCESS) {
+        StartScanThread();
+    } else {
+        LOG_ERROR(Input, "libusb could not be initialized. failed with error = {}", init_res);
+    }
 }
 
-GCPadStatus Adapter::GetPadStatus(int port, const std::array<u8, 37>& adapter_payload) {
+GCPadStatus Adapter::GetPadStatus(std::size_t port, const std::array<u8, 37>& adapter_payload) {
     GCPadStatus pad = {};
     bool get_origin = false;
 
@@ -194,7 +199,7 @@ void Adapter::StartScanThread() {
     }
 
     detect_thread_running = true;
-    detect_thread = std::thread([=] { ScanThreadFunc(); });
+    detect_thread = std::thread(&Adapter::ScanThreadFunc, this);
 }
 
 void Adapter::StopScanThread() {
@@ -211,17 +216,26 @@ void Adapter::Setup() {
     adapter_controllers_status.fill(ControllerTypes::None);
 
     // pointer to list of connected usb devices
-    libusb_device** devices;
+    libusb_device** devices{};
 
     // populate the list of devices, get the count
-    const std::size_t device_count = libusb_get_device_list(libusb_ctx, &devices);
+    const ssize_t device_count = libusb_get_device_list(libusb_ctx, &devices);
+    if (device_count < 0) {
+        LOG_ERROR(Input, "libusb_get_device_list failed with error: {}", device_count);
+        detect_thread_running = false; // Stop the loop constantly checking for gc adapter
+        // TODO: For hotplug+gc adapter checkbox implementation, revert this.
+        return;
+    }
 
-    for (std::size_t index = 0; index < device_count; ++index) {
-        if (CheckDeviceAccess(devices[index])) {
-            // GC Adapter found and accessible, registering it
-            GetGCEndpoint(devices[index]);
-            break;
+    if (devices != nullptr) {
+        for (std::size_t index = 0; index < static_cast<std::size_t>(device_count); ++index) {
+            if (CheckDeviceAccess(devices[index])) {
+                // GC Adapter found and accessible, registering it
+                GetGCEndpoint(devices[index]);
+                break;
+            }
         }
+        libusb_free_device_list(devices, 1);
     }
 }
 
@@ -279,7 +293,13 @@ bool Adapter::CheckDeviceAccess(libusb_device* device) {
 
 void Adapter::GetGCEndpoint(libusb_device* device) {
     libusb_config_descriptor* config = nullptr;
-    libusb_get_config_descriptor(device, 0, &config);
+    const int config_descriptor_return = libusb_get_config_descriptor(device, 0, &config);
+    if (config_descriptor_return != LIBUSB_SUCCESS) {
+        LOG_ERROR(Input, "libusb_get_config_descriptor failed with error = {}",
+                  config_descriptor_return);
+        return;
+    }
+
     for (u8 ic = 0; ic < config->bNumInterfaces; ic++) {
         const libusb_interface* interfaceContainer = &config->interface[ic];
         for (int i = 0; i < interfaceContainer->num_altsetting; i++) {
@@ -338,11 +358,11 @@ void Adapter::Reset() {
     }
 }
 
-bool Adapter::DeviceConnected(int port) {
+bool Adapter::DeviceConnected(std::size_t port) {
     return adapter_controllers_status[port] != ControllerTypes::None;
 }
 
-void Adapter::ResetDeviceType(int port) {
+void Adapter::ResetDeviceType(std::size_t port) {
     adapter_controllers_status[port] = ControllerTypes::None;
 }
 

src/input_common/gcadapter/gc_adapter.h

@@ -7,10 +7,14 @@
 #include <functional>
 #include <mutex>
 #include <thread>
-#include <libusb.h>
+#include <unordered_map>
 #include "common/common_types.h"
 #include "common/threadsafe_queue.h"
 
+struct libusb_context;
+struct libusb_device;
+struct libusb_device_handle;
+
 namespace GCAdapter {
 
 enum {
@@ -96,6 +100,9 @@ public:
     void BeginConfiguration();
     void EndConfiguration();
 
+    /// Returns true if there is a device connected to port
+    bool DeviceConnected(std::size_t port);
+
     std::array<Common::SPSCQueue<GCPadStatus>, 4>& GetPadQueue();
     const std::array<Common::SPSCQueue<GCPadStatus>, 4>& GetPadQueue() const;
 
@@ -103,7 +110,7 @@ public:
     const std::array<GCState, 4>& GetPadState() const;
 
 private:
-    GCPadStatus GetPadStatus(int port, const std::array<u8, 37>& adapter_payload);
+    GCPadStatus GetPadStatus(std::size_t port, const std::array<u8, 37>& adapter_payload);
 
     void PadToState(const GCPadStatus& pad, GCState& state);
 
@@ -115,11 +122,8 @@ private:
     /// Stop scanning for the adapter
     void StopScanThread();
 
-    /// Returns true if there is a device connected to port
-    bool DeviceConnected(int port);
-
     /// Resets status of device connected to port
-    void ResetDeviceType(int port);
+    void ResetDeviceType(std::size_t port);
 
     /// Returns true if we successfully gain access to GC Adapter
     bool CheckDeviceAccess(libusb_device* device);

src/input_common/gcadapter/gc_poller.cpp

@@ -6,6 +6,7 @@
 #include <list>
 #include <mutex>
 #include <utility>
+#include "common/assert.h"
 #include "common/threadsafe_queue.h"
 #include "input_common/gcadapter/gc_adapter.h"
 #include "input_common/gcadapter/gc_poller.h"
@@ -20,7 +21,10 @@ public:
     ~GCButton() override;
 
     bool GetStatus() const override {
-        return gcadapter->GetPadState()[port].buttons.at(button);
+        if (gcadapter->DeviceConnected(port)) {
+            return gcadapter->GetPadState()[port].buttons.at(button);
+        }
+        return false;
     }
 
 private:
@@ -43,13 +47,17 @@ public:
     }
 
     bool GetStatus() const override {
-        const float axis_value = (gcadapter->GetPadState()[port].axes.at(axis) - 128.0f) / 128.0f;
-        if (trigger_if_greater) {
-            // TODO: Might be worthwile to set a slider for the trigger threshold. It is currently
-            // always set to 0.5 in configure_input_player.cpp ZL/ZR HandleClick
-            return axis_value > threshold;
+        if (gcadapter->DeviceConnected(port)) {
+            const float axis_value =
+                (gcadapter->GetPadState()[port].axes.at(axis) - 128.0f) / 128.0f;
+            if (trigger_if_greater) {
+                // TODO: Might be worthwile to set a slider for the trigger threshold. It is
+                // currently always set to 0.5 in configure_input_player.cpp ZL/ZR HandleClick
+                return axis_value > threshold;
+            }
+            return axis_value < -threshold;
         }
-        return axis_value < -threshold;
+        return false;
     }
 
 private:
@@ -94,9 +102,12 @@ std::unique_ptr<Input::ButtonDevice> GCButtonFactory::Create(const Common::Param
         return std::make_unique<GCAxisButton>(port, axis, threshold, trigger_if_greater,
                                               adapter.get());
     }
+
+    UNREACHABLE();
+    return nullptr;
 }
 
-Common::ParamPackage GCButtonFactory::GetNextInput() {
+Common::ParamPackage GCButtonFactory::GetNextInput() const {
     Common::ParamPackage params;
     GCAdapter::GCPadStatus pad;
     auto& queue = adapter->GetPadQueue();
@@ -147,11 +158,14 @@ public:
         : port(port_), axis_x(axis_x_), axis_y(axis_y_), deadzone(deadzone_), gcadapter(adapter) {}
 
     float GetAxis(int axis) const {
-        std::lock_guard lock{mutex};
-        // division is not by a perfect 128 to account for some variance in center location
-        // e.g. my device idled at 131 in X, 120 in Y, and full range of motion was in range
-        // [20-230]
-        return (gcadapter->GetPadState()[port].axes.at(axis) - 128.0f) / 95.0f;
+        if (gcadapter->DeviceConnected(port)) {
+            std::lock_guard lock{mutex};
+            // division is not by a perfect 128 to account for some variance in center location
+            // e.g. my device idled at 131 in X, 120 in Y, and full range of motion was in range
+            // [20-230]
+            return (gcadapter->GetPadState()[port].axes.at(axis) - 128.0f) / 95.0f;
+        }
+        return 0.0f;
     }
 
     std::pair<float, float> GetAnalog(int axis_x, int axis_y) const {
@@ -249,7 +263,7 @@ Common::ParamPackage GCAnalogFactory::GetNextInput() {
             const u8 axis = static_cast<u8>(pad.axis);
             if (analog_x_axis == -1) {
                 analog_x_axis = axis;
-                controller_number = port;
+                controller_number = static_cast<int>(port);
             } else if (analog_y_axis == -1 && analog_x_axis != axis && controller_number == port) {
                 analog_y_axis = axis;
             }

src/input_common/gcadapter/gc_poller.h

@@ -25,7 +25,7 @@ public:
      */
     std::unique_ptr<Input::ButtonDevice> Create(const Common::ParamPackage& params) override;
 
-    Common::ParamPackage GetNextInput();
+    Common::ParamPackage GetNextInput() const;
 
     /// For device input configuration/polling
     void BeginConfiguration();

src/input_common/main.cpp

@@ -4,7 +4,6 @@
 
 #include <memory>
 #include <thread>
-#include <libusb.h>
 #include "common/param_package.h"
 #include "input_common/analog_from_button.h"
 #include "input_common/gcadapter/gc_adapter.h"

src/input_common/udp/client.cpp

@@ -234,7 +234,7 @@ CalibrationConfigurationJob::CalibrationConfigurationJob(
     std::function<void(Status)> status_callback,
     std::function<void(u16, u16, u16, u16)> data_callback) {
 
-    std::thread([=] {
+    std::thread([=, this] {
         constexpr u16 CALIBRATION_THRESHOLD = 100;
 
         u16 min_x{UINT16_MAX};

src/video_core/buffer_cache/buffer_cache.h

@@ -96,7 +96,8 @@ public:
         }
         if (is_written) {
             map->MarkAsModified(true, GetModifiedTicks());
-            if (Settings::IsGPULevelHigh() && Settings::values.use_asynchronous_gpu_emulation) {
+            if (Settings::IsGPULevelHigh() &&
+                Settings::values.use_asynchronous_gpu_emulation.GetValue()) {
                 MarkForAsyncFlush(map);
             }
             if (!map->is_written) {
@@ -369,7 +370,8 @@ private:
         }
         if (modified_inheritance) {
             map->MarkAsModified(true, GetModifiedTicks());
-            if (Settings::IsGPULevelHigh() && Settings::values.use_asynchronous_gpu_emulation) {
+            if (Settings::IsGPULevelHigh() &&
+                Settings::values.use_asynchronous_gpu_emulation.GetValue()) {
                 MarkForAsyncFlush(map);
             }
         }

src/video_core/engines/maxwell_dma.cpp

@@ -14,50 +14,45 @@
 
 namespace Tegra::Engines {
 
+using namespace Texture;
+
 MaxwellDMA::MaxwellDMA(Core::System& system, MemoryManager& memory_manager)
     : system{system}, memory_manager{memory_manager} {}
 
 void MaxwellDMA::CallMethod(u32 method, u32 method_argument, bool is_last_call) {
-    ASSERT_MSG(method < Regs::NUM_REGS,
-               "Invalid MaxwellDMA register, increase the size of the Regs structure");
+    ASSERT_MSG(method < NUM_REGS, "Invalid MaxwellDMA register");
 
     regs.reg_array[method] = method_argument;
 
-#define MAXWELLDMA_REG_INDEX(field_name)                                                           \
-    (offsetof(Tegra::Engines::MaxwellDMA::Regs, field_name) / sizeof(u32))
-
-    switch (method) {
-    case MAXWELLDMA_REG_INDEX(exec): {
-        HandleCopy();
-        break;
+    if (method == offsetof(Regs, launch_dma) / sizeof(u32)) {
+        Launch();
     }
-    }
-
-#undef MAXWELLDMA_REG_INDEX
 }
 
 void MaxwellDMA::CallMultiMethod(u32 method, const u32* base_start, u32 amount,
                                  u32 methods_pending) {
-    for (std::size_t i = 0; i < amount; i++) {
+    for (size_t i = 0; i < amount; ++i) {
         CallMethod(method, base_start[i], methods_pending - static_cast<u32>(i) <= 1);
     }
 }
 
-void MaxwellDMA::HandleCopy() {
-    LOG_TRACE(HW_GPU, "Requested a DMA copy");
-
-    const GPUVAddr source = regs.src_address.Address();
-    const GPUVAddr dest = regs.dst_address.Address();
+void MaxwellDMA::Launch() {
+    LOG_TRACE(Render_OpenGL, "DMA copy 0x{:x} -> 0x{:x}", static_cast<GPUVAddr>(regs.offset_in),
+              static_cast<GPUVAddr>(regs.offset_out));
 
     // TODO(Subv): Perform more research and implement all features of this engine.
-    ASSERT(regs.exec.enable_swizzle == 0);
-    ASSERT(regs.exec.query_mode == Regs::QueryMode::None);
-    ASSERT(regs.exec.query_intr == Regs::QueryIntr::None);
-    ASSERT(regs.exec.copy_mode == Regs::CopyMode::Unk2);
-    ASSERT(regs.dst_params.pos_x == 0);
-    ASSERT(regs.dst_params.pos_y == 0);
+    const LaunchDMA& launch = regs.launch_dma;
+    ASSERT(launch.remap_enable == 0);
+    ASSERT(launch.semaphore_type == LaunchDMA::SemaphoreType::NONE);
+    ASSERT(launch.interrupt_type == LaunchDMA::InterruptType::NONE);
+    ASSERT(launch.data_transfer_type == LaunchDMA::DataTransferType::NON_PIPELINED);
+    ASSERT(regs.dst_params.origin.x == 0);
+    ASSERT(regs.dst_params.origin.y == 0);
 
-    if (!regs.exec.is_dst_linear && !regs.exec.is_src_linear) {
+    const bool is_src_pitch = launch.src_memory_layout == LaunchDMA::MemoryLayout::PITCH;
+    const bool is_dst_pitch = launch.dst_memory_layout == LaunchDMA::MemoryLayout::PITCH;
+
+    if (!is_src_pitch && !is_dst_pitch) {
         // If both the source and the destination are in block layout, assert.
         UNREACHABLE_MSG("Tiled->Tiled DMA transfers are not yet implemented");
         return;
@@ -66,144 +61,161 @@ void MaxwellDMA::HandleCopy() {
     // All copies here update the main memory, so mark all rasterizer states as invalid.
     system.GPU().Maxwell3D().OnMemoryWrite();
 
-    if (regs.exec.is_dst_linear && regs.exec.is_src_linear) {
-        // When the enable_2d bit is disabled, the copy is performed as if we were copying a 1D
-        // buffer of length `x_count`, otherwise we copy a 2D image of dimensions (x_count,
-        // y_count).
-        if (!regs.exec.enable_2d) {
-            memory_manager.CopyBlock(dest, source, regs.x_count);
-            return;
-        }
-
-        // If both the source and the destination are in linear layout, perform a line-by-line
-        // copy. We're going to take a subrect of size (x_count, y_count) from the source
-        // rectangle. There is no need to manually flush/invalidate the regions because
-        // CopyBlock does that for us.
-        for (u32 line = 0; line < regs.y_count; ++line) {
-            const GPUVAddr source_line = source + line * regs.src_pitch;
-            const GPUVAddr dest_line = dest + line * regs.dst_pitch;
-            memory_manager.CopyBlock(dest_line, source_line, regs.x_count);
-        }
-        return;
-    }
-
-    ASSERT(regs.exec.enable_2d == 1);
-
-    if (regs.exec.is_dst_linear && !regs.exec.is_src_linear) {
-
-        ASSERT(regs.src_params.BlockDepth() == 0);
-        // Optimized path for micro copies.
-        if (regs.dst_pitch * regs.y_count < Texture::GetGOBSize() && regs.dst_pitch <= 64) {
-            const u32 bytes_per_pixel = regs.dst_pitch / regs.x_count;
-            const std::size_t src_size = Texture::GetGOBSize();
-            const std::size_t dst_size = regs.dst_pitch * regs.y_count;
-            u32 pos_x = regs.src_params.pos_x;
-            u32 pos_y = regs.src_params.pos_y;
-            const u64 offset =
-                Texture::GetGOBOffset(regs.src_params.size_x, regs.src_params.size_y, pos_x, pos_y,
-                                      regs.src_params.BlockDepth(), bytes_per_pixel);
-            const u32 x_in_gob = 64 / bytes_per_pixel;
-            pos_x = pos_x % x_in_gob;
-            pos_y = pos_y % 8;
-
-            if (read_buffer.size() < src_size) {
-                read_buffer.resize(src_size);
-            }
-
-            if (write_buffer.size() < dst_size) {
-                write_buffer.resize(dst_size);
-            }
-
-            if (Settings::IsGPULevelExtreme()) {
-                memory_manager.ReadBlock(source + offset, read_buffer.data(), src_size);
-                memory_manager.ReadBlock(dest, write_buffer.data(), dst_size);
-            } else {
-                memory_manager.ReadBlockUnsafe(source + offset, read_buffer.data(), src_size);
-                memory_manager.ReadBlockUnsafe(dest, write_buffer.data(), dst_size);
-            }
-
-            Texture::UnswizzleSubrect(regs.x_count, regs.y_count, regs.dst_pitch,
-                                      regs.src_params.size_x, bytes_per_pixel, read_buffer.data(),
-                                      write_buffer.data(), regs.src_params.BlockHeight(), pos_x,
-                                      pos_y);
-
-            memory_manager.WriteBlock(dest, write_buffer.data(), dst_size);
-
-            return;
-        }
-        // If the input is tiled and the output is linear, deswizzle the input and copy it over.
-        const u32 bytes_per_pixel = regs.dst_pitch / regs.x_count;
-        const std::size_t src_size = Texture::CalculateSize(
-            true, bytes_per_pixel, regs.src_params.size_x, regs.src_params.size_y,
-            regs.src_params.size_z, regs.src_params.BlockHeight(), regs.src_params.BlockDepth());
-
-        const std::size_t src_layer_size = Texture::CalculateSize(
-            true, bytes_per_pixel, regs.src_params.size_x, regs.src_params.size_y, 1,
-            regs.src_params.BlockHeight(), regs.src_params.BlockDepth());
-
-        const std::size_t dst_size = regs.dst_pitch * regs.y_count;
-
-        if (read_buffer.size() < src_size) {
-            read_buffer.resize(src_size);
-        }
-
-        if (write_buffer.size() < dst_size) {
-            write_buffer.resize(dst_size);
-        }
-
-        if (Settings::IsGPULevelExtreme()) {
-            memory_manager.ReadBlock(source, read_buffer.data(), src_size);
-            memory_manager.ReadBlock(dest, write_buffer.data(), dst_size);
-        } else {
-            memory_manager.ReadBlockUnsafe(source, read_buffer.data(), src_size);
-            memory_manager.ReadBlockUnsafe(dest, write_buffer.data(), dst_size);
-        }
-
-        Texture::UnswizzleSubrect(
-            regs.x_count, regs.y_count, regs.dst_pitch, regs.src_params.size_x, bytes_per_pixel,
-            read_buffer.data() + src_layer_size * regs.src_params.pos_z, write_buffer.data(),
-            regs.src_params.BlockHeight(), regs.src_params.pos_x, regs.src_params.pos_y);
-
-        memory_manager.WriteBlock(dest, write_buffer.data(), dst_size);
+    if (is_src_pitch && is_dst_pitch) {
+        CopyPitchToPitch();
     } else {
-        ASSERT(regs.dst_params.BlockDepth() == 0);
+        ASSERT(launch.multi_line_enable == 1);
 
-        const u32 bytes_per_pixel = regs.src_pitch / regs.x_count;
-
-        const std::size_t dst_size = Texture::CalculateSize(
-            true, bytes_per_pixel, regs.dst_params.size_x, regs.dst_params.size_y,
-            regs.dst_params.size_z, regs.dst_params.BlockHeight(), regs.dst_params.BlockDepth());
-
-        const std::size_t dst_layer_size = Texture::CalculateSize(
-            true, bytes_per_pixel, regs.dst_params.size_x, regs.dst_params.size_y, 1,
-            regs.dst_params.BlockHeight(), regs.dst_params.BlockDepth());
-
-        const std::size_t src_size = regs.src_pitch * regs.y_count;
-
-        if (read_buffer.size() < src_size) {
-            read_buffer.resize(src_size);
-        }
-
-        if (write_buffer.size() < dst_size) {
-            write_buffer.resize(dst_size);
-        }
-
-        if (Settings::IsGPULevelExtreme()) {
-            memory_manager.ReadBlock(source, read_buffer.data(), src_size);
-            memory_manager.ReadBlock(dest, write_buffer.data(), dst_size);
+        if (!is_src_pitch && is_dst_pitch) {
+            CopyBlockLinearToPitch();
         } else {
-            memory_manager.ReadBlockUnsafe(source, read_buffer.data(), src_size);
-            memory_manager.ReadBlockUnsafe(dest, write_buffer.data(), dst_size);
+            CopyPitchToBlockLinear();
         }
-
-        // If the input is linear and the output is tiled, swizzle the input and copy it over.
-        Texture::SwizzleSubrect(
-            regs.x_count, regs.y_count, regs.src_pitch, regs.dst_params.size_x, bytes_per_pixel,
-            write_buffer.data() + dst_layer_size * regs.dst_params.pos_z, read_buffer.data(),
-            regs.dst_params.BlockHeight(), regs.dst_params.pos_x, regs.dst_params.pos_y);
-
-        memory_manager.WriteBlock(dest, write_buffer.data(), dst_size);
     }
 }
 
+void MaxwellDMA::CopyPitchToPitch() {
+    // When `multi_line_enable` bit is disabled the copy is performed as if we were copying a 1D
+    // buffer of length `line_length_in`.
+    // Otherwise we copy a 2D image of dimensions (line_length_in, line_count).
+    if (!regs.launch_dma.multi_line_enable) {
+        memory_manager.CopyBlock(regs.offset_out, regs.offset_in, regs.line_length_in);
+        return;
+    }
+
+    // Perform a line-by-line copy.
+    // We're going to take a subrect of size (line_length_in, line_count) from the source rectangle.
+    // There is no need to manually flush/invalidate the regions because CopyBlock does that for us.
+    for (u32 line = 0; line < regs.line_count; ++line) {
+        const GPUVAddr source_line = regs.offset_in + static_cast<size_t>(line) * regs.pitch_in;
+        const GPUVAddr dest_line = regs.offset_out + static_cast<size_t>(line) * regs.pitch_out;
+        memory_manager.CopyBlock(dest_line, source_line, regs.line_length_in);
+    }
+}
+
+void MaxwellDMA::CopyBlockLinearToPitch() {
+    ASSERT(regs.src_params.block_size.depth == 0);
+
+    // Optimized path for micro copies.
+    const size_t dst_size = static_cast<size_t>(regs.pitch_out) * regs.line_count;
+    if (dst_size < GOB_SIZE && regs.pitch_out <= GOB_SIZE_X) {
+        FastCopyBlockLinearToPitch();
+        return;
+    }
+
+    // Deswizzle the input and copy it over.
+    const u32 bytes_per_pixel = regs.pitch_out / regs.line_length_in;
+    const Parameters& src_params = regs.src_params;
+    const u32 width = src_params.width;
+    const u32 height = src_params.height;
+    const u32 depth = src_params.depth;
+    const u32 block_height = src_params.block_size.height;
+    const u32 block_depth = src_params.block_size.depth;
+    const size_t src_size =
+        CalculateSize(true, bytes_per_pixel, width, height, depth, block_height, block_depth);
+    const size_t src_layer_size =
+        CalculateSize(true, bytes_per_pixel, width, height, 1, block_height, block_depth);
+
+    if (read_buffer.size() < src_size) {
+        read_buffer.resize(src_size);
+    }
+    if (write_buffer.size() < dst_size) {
+        write_buffer.resize(dst_size);
+    }
+
+    if (Settings::IsGPULevelExtreme()) {
+        memory_manager.ReadBlock(regs.offset_in, read_buffer.data(), src_size);
+        memory_manager.ReadBlock(regs.offset_out, write_buffer.data(), dst_size);
+    } else {
+        memory_manager.ReadBlockUnsafe(regs.offset_in, read_buffer.data(), src_size);
+        memory_manager.ReadBlockUnsafe(regs.offset_out, write_buffer.data(), dst_size);
+    }
+
+    UnswizzleSubrect(regs.line_length_in, regs.line_count, regs.pitch_out, width, bytes_per_pixel,
+                     read_buffer.data() + src_layer_size * src_params.layer, write_buffer.data(),
+                     block_height, src_params.origin.x, src_params.origin.y);
+
+    memory_manager.WriteBlock(regs.offset_out, write_buffer.data(), dst_size);
+}
+
+void MaxwellDMA::CopyPitchToBlockLinear() {
+    const auto& dst_params = regs.dst_params;
+    const u32 bytes_per_pixel = regs.pitch_in / regs.line_length_in;
+    const u32 width = dst_params.width;
+    const u32 height = dst_params.height;
+    const u32 depth = dst_params.depth;
+    const u32 block_height = dst_params.block_size.height;
+    const u32 block_depth = dst_params.block_size.depth;
+    const size_t dst_size =
+        CalculateSize(true, bytes_per_pixel, width, height, depth, block_height, block_depth);
+    const size_t dst_layer_size =
+        CalculateSize(true, bytes_per_pixel, width, height, 1, block_height, block_depth);
+
+    const size_t src_size = static_cast<size_t>(regs.pitch_in) * regs.line_count;
+
+    if (read_buffer.size() < src_size) {
+        read_buffer.resize(src_size);
+    }
+    if (write_buffer.size() < dst_size) {
+        write_buffer.resize(dst_size);
+    }
+
+    if (Settings::IsGPULevelExtreme()) {
+        memory_manager.ReadBlock(regs.offset_in, read_buffer.data(), src_size);
+        memory_manager.ReadBlock(regs.offset_out, write_buffer.data(), dst_size);
+    } else {
+        memory_manager.ReadBlockUnsafe(regs.offset_in, read_buffer.data(), src_size);
+        memory_manager.ReadBlockUnsafe(regs.offset_out, write_buffer.data(), dst_size);
+    }
+
+    // If the input is linear and the output is tiled, swizzle the input and copy it over.
+    if (regs.dst_params.block_size.depth > 0) {
+        ASSERT(dst_params.layer == 0);
+        SwizzleSliceToVoxel(regs.line_length_in, regs.line_count, regs.pitch_in, width, height,
+                            bytes_per_pixel, block_height, block_depth, dst_params.origin.x,
+                            dst_params.origin.y, write_buffer.data(), read_buffer.data());
+    } else {
+        SwizzleSubrect(regs.line_length_in, regs.line_count, regs.pitch_in, width, bytes_per_pixel,
+                       write_buffer.data() + dst_layer_size * dst_params.layer, read_buffer.data(),
+                       block_height, dst_params.origin.x, dst_params.origin.y);
+    }
+
+    memory_manager.WriteBlock(regs.offset_out, write_buffer.data(), dst_size);
+}
+
+void MaxwellDMA::FastCopyBlockLinearToPitch() {
+    const u32 bytes_per_pixel = regs.pitch_out / regs.line_length_in;
+    const size_t src_size = GOB_SIZE;
+    const size_t dst_size = static_cast<size_t>(regs.pitch_out) * regs.line_count;
+    u32 pos_x = regs.src_params.origin.x;
+    u32 pos_y = regs.src_params.origin.y;
+    const u64 offset = GetGOBOffset(regs.src_params.width, regs.src_params.height, pos_x, pos_y,
+                                    regs.src_params.block_size.height, bytes_per_pixel);
+    const u32 x_in_gob = 64 / bytes_per_pixel;
+    pos_x = pos_x % x_in_gob;
+    pos_y = pos_y % 8;
+
+    if (read_buffer.size() < src_size) {
+        read_buffer.resize(src_size);
+    }
+
+    if (write_buffer.size() < dst_size) {
+        write_buffer.resize(dst_size);
+    }
+
+    if (Settings::IsGPULevelExtreme()) {
+        memory_manager.ReadBlock(regs.offset_in + offset, read_buffer.data(), src_size);
+        memory_manager.ReadBlock(regs.offset_out, write_buffer.data(), dst_size);
+    } else {
+        memory_manager.ReadBlockUnsafe(regs.offset_in + offset, read_buffer.data(), src_size);
+        memory_manager.ReadBlockUnsafe(regs.offset_out, write_buffer.data(), dst_size);
+    }
+
+    UnswizzleSubrect(regs.line_length_in, regs.line_count, regs.pitch_out, regs.src_params.width,
+                     bytes_per_pixel, read_buffer.data(), write_buffer.data(),
+                     regs.src_params.block_size.height, pos_x, pos_y);
+
+    memory_manager.WriteBlock(regs.offset_out, write_buffer.data(), dst_size);
+}
+
 } // namespace Tegra::Engines

src/video_core/engines/maxwell_dma.h

@@ -24,12 +24,167 @@ class MemoryManager;
 namespace Tegra::Engines {
 
 /**
- * This Engine is known as GK104_Copy. Documentation can be found in:
+ * This engine is known as gk104_copy. Documentation can be found in:
+ * https://github.com/NVIDIA/open-gpu-doc/blob/master/classes/dma-copy/clb0b5.h
  * https://github.com/envytools/envytools/blob/master/rnndb/fifo/gk104_copy.xml
  */
 
 class MaxwellDMA final : public EngineInterface {
 public:
+    struct PackedGPUVAddr {
+        u32 upper;
+        u32 lower;
+
+        constexpr operator GPUVAddr() const noexcept {
+            return (static_cast<GPUVAddr>(upper & 0xff) << 32) | lower;
+        }
+    };
+
+    union BlockSize {
+        BitField<0, 4, u32> width;
+        BitField<4, 4, u32> height;
+        BitField<8, 4, u32> depth;
+        BitField<12, 4, u32> gob_height;
+    };
+    static_assert(sizeof(BlockSize) == 4);
+
+    union Origin {
+        BitField<0, 16, u32> x;
+        BitField<16, 16, u32> y;
+    };
+    static_assert(sizeof(Origin) == 4);
+
+    struct Parameters {
+        BlockSize block_size;
+        u32 width;
+        u32 height;
+        u32 depth;
+        u32 layer;
+        Origin origin;
+    };
+    static_assert(sizeof(Parameters) == 24);
+
+    struct Semaphore {
+        PackedGPUVAddr address;
+        u32 payload;
+    };
+    static_assert(sizeof(Semaphore) == 12);
+
+    struct RenderEnable {
+        enum class Mode : u32 {
+            FALSE = 0,
+            TRUE = 1,
+            CONDITIONAL = 2,
+            RENDER_IF_EQUAL = 3,
+            RENDER_IF_NOT_EQUAL = 4,
+        };
+
+        PackedGPUVAddr address;
+        BitField<0, 3, Mode> mode;
+    };
+    static_assert(sizeof(RenderEnable) == 12);
+
+    enum class PhysModeTarget : u32 {
+        LOCAL_FB = 0,
+        COHERENT_SYSMEM = 1,
+        NONCOHERENT_SYSMEM = 2,
+    };
+    using PhysMode = BitField<0, 2, PhysModeTarget>;
+
+    union LaunchDMA {
+        enum class DataTransferType : u32 {
+            NONE = 0,
+            PIPELINED = 1,
+            NON_PIPELINED = 2,
+        };
+
+        enum class SemaphoreType : u32 {
+            NONE = 0,
+            RELEASE_ONE_WORD_SEMAPHORE = 1,
+            RELEASE_FOUR_WORD_SEMAPHORE = 2,
+        };
+
+        enum class InterruptType : u32 {
+            NONE = 0,
+            BLOCKING = 1,
+            NON_BLOCKING = 2,
+        };
+
+        enum class MemoryLayout : u32 {
+            BLOCKLINEAR = 0,
+            PITCH = 1,
+        };
+
+        enum class Type : u32 {
+            VIRTUAL = 0,
+            PHYSICAL = 1,
+        };
+
+        enum class SemaphoreReduction : u32 {
+            IMIN = 0,
+            IMAX = 1,
+            IXOR = 2,
+            IAND = 3,
+            IOR = 4,
+            IADD = 5,
+            INC = 6,
+            DEC = 7,
+            FADD = 0xA,
+        };
+
+        enum class SemaphoreReductionSign : u32 {
+            SIGNED = 0,
+            UNSIGNED = 1,
+        };
+
+        enum class BypassL2 : u32 {
+            USE_PTE_SETTING = 0,
+            FORCE_VOLATILE = 1,
+        };
+
+        BitField<0, 2, DataTransferType> data_transfer_type;
+        BitField<2, 1, u32> flush_enable;
+        BitField<3, 2, SemaphoreType> semaphore_type;
+        BitField<5, 2, InterruptType> interrupt_type;
+        BitField<7, 1, MemoryLayout> src_memory_layout;
+        BitField<8, 1, MemoryLayout> dst_memory_layout;
+        BitField<9, 1, u32> multi_line_enable;
+        BitField<10, 1, u32> remap_enable;
+        BitField<11, 1, u32> rmwdisable;
+        BitField<12, 1, Type> src_type;
+        BitField<13, 1, Type> dst_type;
+        BitField<14, 4, SemaphoreReduction> semaphore_reduction;
+        BitField<18, 1, SemaphoreReductionSign> semaphore_reduction_sign;
+        BitField<19, 1, u32> reduction_enable;
+        BitField<20, 1, BypassL2> bypass_l2;
+    };
+    static_assert(sizeof(LaunchDMA) == 4);
+
+    struct RemapConst {
+        enum Swizzle : u32 {
+            SRC_X = 0,
+            SRC_Y = 1,
+            SRC_Z = 2,
+            SRC_W = 3,
+            CONST_A = 4,
+            CONST_B = 5,
+            NO_WRITE = 6,
+        };
+
+        PackedGPUVAddr address;
+
+        union {
+            BitField<0, 3, Swizzle> dst_x;
+            BitField<4, 3, Swizzle> dst_y;
+            BitField<8, 3, Swizzle> dst_z;
+            BitField<12, 3, Swizzle> dst_w;
+            BitField<16, 2, u32> component_size_minus_one;
+            BitField<20, 2, u32> num_src_components_minus_one;
+            BitField<24, 2, u32> num_dst_components_minus_one;
+        };
+    };
+    static_assert(sizeof(RemapConst) == 12);
+
     explicit MaxwellDMA(Core::System& system, MemoryManager& memory_manager);
     ~MaxwellDMA() = default;
 
@@ -40,144 +195,19 @@ public:
     void CallMultiMethod(u32 method, const u32* base_start, u32 amount,
                          u32 methods_pending) override;
 
-    struct Regs {
-        static constexpr std::size_t NUM_REGS = 0x1D6;
-
-        struct Parameters {
-            union {
-                BitField<0, 4, u32> block_depth;
-                BitField<4, 4, u32> block_height;
-                BitField<8, 4, u32> block_width;
-            };
-            u32 size_x;
-            u32 size_y;
-            u32 size_z;
-            u32 pos_z;
-            union {
-                BitField<0, 16, u32> pos_x;
-                BitField<16, 16, u32> pos_y;
-            };
-
-            u32 BlockHeight() const {
-                return block_height.Value();
-            }
-
-            u32 BlockDepth() const {
-                return block_depth.Value();
-            }
-        };
-
-        static_assert(sizeof(Parameters) == 24, "Parameters has wrong size");
-
-        enum class ComponentMode : u32 {
-            Src0 = 0,
-            Src1 = 1,
-            Src2 = 2,
-            Src3 = 3,
-            Const0 = 4,
-            Const1 = 5,
-            Zero = 6,
-        };
-
-        enum class CopyMode : u32 {
-            None = 0,
-            Unk1 = 1,
-            Unk2 = 2,
-        };
-
-        enum class QueryMode : u32 {
-            None = 0,
-            Short = 1,
-            Long = 2,
-        };
-
-        enum class QueryIntr : u32 {
-            None = 0,
-            Block = 1,
-            NonBlock = 2,
-        };
-
-        union {
-            struct {
-                INSERT_UNION_PADDING_WORDS(0xC0);
-
-                struct {
-                    union {
-                        BitField<0, 2, CopyMode> copy_mode;
-                        BitField<2, 1, u32> flush;
-
-                        BitField<3, 2, QueryMode> query_mode;
-                        BitField<5, 2, QueryIntr> query_intr;
-
-                        BitField<7, 1, u32> is_src_linear;
-                        BitField<8, 1, u32> is_dst_linear;
-
-                        BitField<9, 1, u32> enable_2d;
-                        BitField<10, 1, u32> enable_swizzle;
-                    };
-                } exec;
-
-                INSERT_UNION_PADDING_WORDS(0x3F);
-
-                struct {
-                    u32 address_high;
-                    u32 address_low;
-
-                    GPUVAddr Address() const {
-                        return static_cast<GPUVAddr>((static_cast<GPUVAddr>(address_high) << 32) |
-                                                     address_low);
-                    }
-                } src_address;
-
-                struct {
-                    u32 address_high;
-                    u32 address_low;
-
-                    GPUVAddr Address() const {
-                        return static_cast<GPUVAddr>((static_cast<GPUVAddr>(address_high) << 32) |
-                                                     address_low);
-                    }
-                } dst_address;
-
-                u32 src_pitch;
-                u32 dst_pitch;
-                u32 x_count;
-                u32 y_count;
-
-                INSERT_UNION_PADDING_WORDS(0xB8);
-
-                u32 const0;
-                u32 const1;
-                union {
-                    BitField<0, 4, ComponentMode> component0;
-                    BitField<4, 4, ComponentMode> component1;
-                    BitField<8, 4, ComponentMode> component2;
-                    BitField<12, 4, ComponentMode> component3;
-                    BitField<16, 2, u32> component_size;
-                    BitField<20, 3, u32> src_num_components;
-                    BitField<24, 3, u32> dst_num_components;
-
-                    u32 SrcBytePerPixel() const {
-                        return src_num_components.Value() * component_size.Value();
-                    }
-                    u32 DstBytePerPixel() const {
-                        return dst_num_components.Value() * component_size.Value();
-                    }
-                } swizzle_config;
-
-                Parameters dst_params;
-
-                INSERT_UNION_PADDING_WORDS(1);
-
-                Parameters src_params;
-
-                INSERT_UNION_PADDING_WORDS(0x13);
-            };
-            std::array<u32, NUM_REGS> reg_array;
-        };
-    } regs{};
-
 private:
+    /// Performs the copy from the source buffer to the destination buffer as configured in the
+    /// registers.
+    void Launch();
+
+    void CopyPitchToPitch();
+
+    void CopyBlockLinearToPitch();
+
+    void CopyPitchToBlockLinear();
+
+    void FastCopyBlockLinearToPitch();
+
     Core::System& system;
 
     MemoryManager& memory_manager;
@@ -185,28 +215,58 @@ private:
     std::vector<u8> read_buffer;
     std::vector<u8> write_buffer;
 
-    /// Performs the copy from the source buffer to the destination buffer as configured in the
-    /// registers.
-    void HandleCopy();
-};
+    static constexpr std::size_t NUM_REGS = 0x800;
+    struct Regs {
+        union {
+            struct {
+                u32 reserved[0x40];
+                u32 nop;
+                u32 reserved01[0xf];
+                u32 pm_trigger;
+                u32 reserved02[0x3f];
+                Semaphore semaphore;
+                u32 reserved03[0x2];
+                RenderEnable render_enable;
+                PhysMode src_phys_mode;
+                PhysMode dst_phys_mode;
+                u32 reserved04[0x26];
+                LaunchDMA launch_dma;
+                u32 reserved05[0x3f];
+                PackedGPUVAddr offset_in;
+                PackedGPUVAddr offset_out;
+                u32 pitch_in;
+                u32 pitch_out;
+                u32 line_length_in;
+                u32 line_count;
+                u32 reserved06[0xb8];
+                RemapConst remap_const;
+                Parameters dst_params;
+                u32 reserved07[0x1];
+                Parameters src_params;
+                u32 reserved08[0x275];
+                u32 pm_trigger_end;
+                u32 reserved09[0x3ba];
+            };
+            std::array<u32, NUM_REGS> reg_array;
+        };
+    } regs{};
 
 #define ASSERT_REG_POSITION(field_name, position)                                                  \
     static_assert(offsetof(MaxwellDMA::Regs, field_name) == position * 4,                          \
                   "Field " #field_name " has invalid position")
 
-ASSERT_REG_POSITION(exec, 0xC0);
-ASSERT_REG_POSITION(src_address, 0x100);
-ASSERT_REG_POSITION(dst_address, 0x102);
-ASSERT_REG_POSITION(src_pitch, 0x104);
-ASSERT_REG_POSITION(dst_pitch, 0x105);
-ASSERT_REG_POSITION(x_count, 0x106);
-ASSERT_REG_POSITION(y_count, 0x107);
-ASSERT_REG_POSITION(const0, 0x1C0);
-ASSERT_REG_POSITION(const1, 0x1C1);
-ASSERT_REG_POSITION(swizzle_config, 0x1C2);
-ASSERT_REG_POSITION(dst_params, 0x1C3);
-ASSERT_REG_POSITION(src_params, 0x1CA);
+    ASSERT_REG_POSITION(launch_dma, 0xC0);
+    ASSERT_REG_POSITION(offset_in, 0x100);
+    ASSERT_REG_POSITION(offset_out, 0x102);
+    ASSERT_REG_POSITION(pitch_in, 0x104);
+    ASSERT_REG_POSITION(pitch_out, 0x105);
+    ASSERT_REG_POSITION(line_length_in, 0x106);
+    ASSERT_REG_POSITION(line_count, 0x107);
+    ASSERT_REG_POSITION(remap_const, 0x1C0);
+    ASSERT_REG_POSITION(dst_params, 0x1C3);
+    ASSERT_REG_POSITION(src_params, 0x1CA);
 
 #undef ASSERT_REG_POSITION
+};
 
 } // namespace Tegra::Engines

src/video_core/gpu.cpp

@@ -157,7 +157,7 @@ u64 GPU::GetTicks() const {
     constexpr u64 gpu_ticks_den = 625;
 
     u64 nanoseconds = system.CoreTiming().GetGlobalTimeNs().count();
-    if (Settings::values.use_fast_gpu_time) {
+    if (Settings::values.use_fast_gpu_time.GetValue()) {
         nanoseconds /= 256;
     }
     const u64 nanoseconds_num = nanoseconds / gpu_ticks_den;

src/video_core/query_cache.h

@@ -132,7 +132,7 @@ public:
         }
 
         query->BindCounter(Stream(type).Current(), timestamp);
-        if (Settings::values.use_asynchronous_gpu_emulation) {
+        if (Settings::values.use_asynchronous_gpu_emulation.GetValue()) {
             AsyncFlushQuery(cpu_addr);
         }
     }

src/video_core/renderer_base.cpp

@@ -18,7 +18,7 @@ RendererBase::~RendererBase() = default;
 void RendererBase::RefreshBaseSettings() {
     UpdateCurrentFramebufferLayout();
 
-    renderer_settings.use_framelimiter = Settings::values.use_frame_limit;
+    renderer_settings.use_framelimiter = Settings::values.use_frame_limit.GetValue();
     renderer_settings.set_background_color = true;
 }
 

src/video_core/renderer_opengl/gl_device.cpp

@@ -229,15 +229,15 @@ Device::Device()
     // uniform buffers as "push constants"
     has_fast_buffer_sub_data = is_nvidia && !disable_fast_buffer_sub_data;
 
-    use_assembly_shaders = Settings::values.use_assembly_shaders && GLAD_GL_NV_gpu_program5 &&
-                           GLAD_GL_NV_compute_program5 && GLAD_GL_NV_transform_feedback &&
-                           GLAD_GL_NV_transform_feedback2;
+    use_assembly_shaders = Settings::values.use_assembly_shaders.GetValue() &&
+                           GLAD_GL_NV_gpu_program5 && GLAD_GL_NV_compute_program5 &&
+                           GLAD_GL_NV_transform_feedback && GLAD_GL_NV_transform_feedback2;
 
     LOG_INFO(Render_OpenGL, "Renderer_VariableAOFFI: {}", has_variable_aoffi);
     LOG_INFO(Render_OpenGL, "Renderer_ComponentIndexingBug: {}", has_component_indexing_bug);
     LOG_INFO(Render_OpenGL, "Renderer_PreciseBug: {}", has_precise_bug);
 
-    if (Settings::values.use_assembly_shaders && !use_assembly_shaders) {
+    if (Settings::values.use_assembly_shaders.GetValue() && !use_assembly_shaders) {
         LOG_ERROR(Render_OpenGL, "Assembly shaders enabled but not supported");
     }
 }

src/video_core/renderer_opengl/gl_shader_disk_cache.cpp

@@ -213,7 +213,7 @@ ShaderDiskCacheOpenGL::~ShaderDiskCacheOpenGL() = default;
 std::optional<std::vector<ShaderDiskCacheEntry>> ShaderDiskCacheOpenGL::LoadTransferable() {
     // Skip games without title id
     const bool has_title_id = system.CurrentProcess()->GetTitleID() != 0;
-    if (!Settings::values.use_disk_shader_cache || !has_title_id) {
+    if (!Settings::values.use_disk_shader_cache.GetValue() || !has_title_id) {
         return {};
     }
 

src/video_core/renderer_opengl/renderer_opengl.cpp

@@ -455,8 +455,8 @@ void RendererOpenGL::LoadColorToActiveGLTexture(u8 color_r, u8 color_g, u8 color
 void RendererOpenGL::InitOpenGLObjects() {
     frame_mailbox = std::make_unique<FrameMailbox>();
 
-    glClearColor(Settings::values.bg_red, Settings::values.bg_green, Settings::values.bg_blue,
-                 0.0f);
+    glClearColor(Settings::values.bg_red.GetValue(), Settings::values.bg_green.GetValue(),
+                 Settings::values.bg_blue.GetValue(), 0.0f);
 
     // Create shader programs
     OGLShader vertex_shader;
@@ -561,8 +561,8 @@ void RendererOpenGL::ConfigureFramebufferTexture(TextureInfo& texture,
 void RendererOpenGL::DrawScreen(const Layout::FramebufferLayout& layout) {
     if (renderer_settings.set_background_color) {
         // Update background color before drawing
-        glClearColor(Settings::values.bg_red, Settings::values.bg_green, Settings::values.bg_blue,
-                     0.0f);
+        glClearColor(Settings::values.bg_red.GetValue(), Settings::values.bg_green.GetValue(),
+                     Settings::values.bg_blue.GetValue(), 0.0f);
     }
 
     // Set projection matrix

src/video_core/renderer_vulkan/fixed_pipeline_state.cpp

@@ -39,52 +39,18 @@ constexpr std::array POLYGON_OFFSET_ENABLE_LUT = {
 
 } // Anonymous namespace
 
-void FixedPipelineState::DepthStencil::Fill(const Maxwell& regs) noexcept {
-    raw = 0;
-    front.action_stencil_fail.Assign(PackStencilOp(regs.stencil_front_op_fail));
-    front.action_depth_fail.Assign(PackStencilOp(regs.stencil_front_op_zfail));
-    front.action_depth_pass.Assign(PackStencilOp(regs.stencil_front_op_zpass));
-    front.test_func.Assign(PackComparisonOp(regs.stencil_front_func_func));
-    if (regs.stencil_two_side_enable) {
-        back.action_stencil_fail.Assign(PackStencilOp(regs.stencil_back_op_fail));
-        back.action_depth_fail.Assign(PackStencilOp(regs.stencil_back_op_zfail));
-        back.action_depth_pass.Assign(PackStencilOp(regs.stencil_back_op_zpass));
-        back.test_func.Assign(PackComparisonOp(regs.stencil_back_func_func));
-    } else {
-        back.action_stencil_fail.Assign(front.action_stencil_fail);
-        back.action_depth_fail.Assign(front.action_depth_fail);
-        back.action_depth_pass.Assign(front.action_depth_pass);
-        back.test_func.Assign(front.test_func);
-    }
-    depth_test_enable.Assign(regs.depth_test_enable);
-    depth_write_enable.Assign(regs.depth_write_enabled);
-    depth_bounds_enable.Assign(regs.depth_bounds_enable);
-    stencil_enable.Assign(regs.stencil_enable);
-    depth_test_func.Assign(PackComparisonOp(regs.depth_test_func));
-}
-
-void FixedPipelineState::Rasterizer::Fill(const Maxwell& regs) noexcept {
+void FixedPipelineState::Fill(const Maxwell& regs, bool has_extended_dynamic_state) {
     const auto& clip = regs.view_volume_clip_control;
     const std::array enabled_lut = {regs.polygon_offset_point_enable,
                                     regs.polygon_offset_line_enable,
                                     regs.polygon_offset_fill_enable};
     const u32 topology_index = static_cast<u32>(regs.draw.topology.Value());
 
-    u32 packed_front_face = PackFrontFace(regs.front_face);
-    if (regs.screen_y_control.triangle_rast_flip != 0) {
-        // Flip front face
-        packed_front_face = 1 - packed_front_face;
-    }
-
     raw = 0;
-    topology.Assign(topology_index);
     primitive_restart_enable.Assign(regs.primitive_restart.enabled != 0 ? 1 : 0);
-    cull_enable.Assign(regs.cull_test_enabled != 0 ? 1 : 0);
     depth_bias_enable.Assign(enabled_lut[POLYGON_OFFSET_ENABLE_LUT[topology_index]] != 0 ? 1 : 0);
     depth_clamp_disabled.Assign(regs.view_volume_clip_control.depth_clamp_disabled.Value());
     ndc_minus_one_to_one.Assign(regs.depth_mode == Maxwell::DepthMode::MinusOneToOne ? 1 : 0);
-    cull_face.Assign(PackCullFace(regs.cull_face));
-    front_face.Assign(packed_front_face);
     polygon_mode.Assign(PackPolygonMode(regs.polygon_mode_front));
     patch_control_points_minus_one.Assign(regs.patch_vertices - 1);
     tessellation_primitive.Assign(static_cast<u32>(regs.tess_mode.prim.Value()));
@@ -93,19 +59,37 @@ void FixedPipelineState::Rasterizer::Fill(const Maxwell& regs) noexcept {
     logic_op_enable.Assign(regs.logic_op.enable != 0 ? 1 : 0);
     logic_op.Assign(PackLogicOp(regs.logic_op.operation));
     rasterize_enable.Assign(regs.rasterize_enable != 0 ? 1 : 0);
-    std::memcpy(&point_size, &regs.point_size, sizeof(point_size)); // TODO: C++20 std::bit_cast
-}
 
-void FixedPipelineState::ColorBlending::Fill(const Maxwell& regs) noexcept {
+    std::memcpy(&point_size, &regs.point_size, sizeof(point_size)); // TODO: C++20 std::bit_cast
+
+    for (std::size_t index = 0; index < Maxwell::NumVertexArrays; ++index) {
+        binding_divisors[index] =
+            regs.instanced_arrays.IsInstancingEnabled(index) ? regs.vertex_array[index].divisor : 0;
+    }
+
+    for (std::size_t index = 0; index < Maxwell::NumVertexAttributes; ++index) {
+        const auto& input = regs.vertex_attrib_format[index];
+        auto& attribute = attributes[index];
+        attribute.raw = 0;
+        attribute.enabled.Assign(input.IsConstant() ? 0 : 1);
+        attribute.buffer.Assign(input.buffer);
+        attribute.offset.Assign(input.offset);
+        attribute.type.Assign(static_cast<u32>(input.type.Value()));
+        attribute.size.Assign(static_cast<u32>(input.size.Value()));
+    }
+
     for (std::size_t index = 0; index < std::size(attachments); ++index) {
         attachments[index].Fill(regs, index);
     }
-}
 
-void FixedPipelineState::ViewportSwizzles::Fill(const Maxwell& regs) noexcept {
     const auto& transform = regs.viewport_transform;
-    std::transform(transform.begin(), transform.end(), swizzles.begin(),
+    std::transform(transform.begin(), transform.end(), viewport_swizzles.begin(),
                    [](const auto& viewport) { return static_cast<u16>(viewport.swizzle.raw); });
+
+    if (!has_extended_dynamic_state) {
+        no_extended_dynamic_state.Assign(1);
+        dynamic_state.Fill(regs);
+    }
 }
 
 void FixedPipelineState::BlendingAttachment::Fill(const Maxwell& regs, std::size_t index) {
@@ -147,20 +131,57 @@ void FixedPipelineState::BlendingAttachment::Fill(const Maxwell& regs, std::size
     enable.Assign(1);
 }
 
-void FixedPipelineState::Fill(const Maxwell& regs) {
-    rasterizer.Fill(regs);
-    depth_stencil.Fill(regs);
-    color_blending.Fill(regs);
-    viewport_swizzles.Fill(regs);
+void FixedPipelineState::DynamicState::Fill(const Maxwell& regs) {
+    const u32 topology_index = static_cast<u32>(regs.draw.topology.Value());
+    u32 packed_front_face = PackFrontFace(regs.front_face);
+    if (regs.screen_y_control.triangle_rast_flip != 0) {
+        // Flip front face
+        packed_front_face = 1 - packed_front_face;
+    }
+
+    raw1 = 0;
+    raw2 = 0;
+    front.action_stencil_fail.Assign(PackStencilOp(regs.stencil_front_op_fail));
+    front.action_depth_fail.Assign(PackStencilOp(regs.stencil_front_op_zfail));
+    front.action_depth_pass.Assign(PackStencilOp(regs.stencil_front_op_zpass));
+    front.test_func.Assign(PackComparisonOp(regs.stencil_front_func_func));
+    if (regs.stencil_two_side_enable) {
+        back.action_stencil_fail.Assign(PackStencilOp(regs.stencil_back_op_fail));
+        back.action_depth_fail.Assign(PackStencilOp(regs.stencil_back_op_zfail));
+        back.action_depth_pass.Assign(PackStencilOp(regs.stencil_back_op_zpass));
+        back.test_func.Assign(PackComparisonOp(regs.stencil_back_func_func));
+    } else {
+        back.action_stencil_fail.Assign(front.action_stencil_fail);
+        back.action_depth_fail.Assign(front.action_depth_fail);
+        back.action_depth_pass.Assign(front.action_depth_pass);
+        back.test_func.Assign(front.test_func);
+    }
+    stencil_enable.Assign(regs.stencil_enable);
+    depth_write_enable.Assign(regs.depth_write_enabled);
+    depth_bounds_enable.Assign(regs.depth_bounds_enable);
+    depth_test_enable.Assign(regs.depth_test_enable);
+    front_face.Assign(packed_front_face);
+    depth_test_func.Assign(PackComparisonOp(regs.depth_test_func));
+    topology.Assign(topology_index);
+    cull_face.Assign(PackCullFace(regs.cull_face));
+    cull_enable.Assign(regs.cull_test_enabled != 0 ? 1 : 0);
+
+    for (std::size_t index = 0; index < Maxwell::NumVertexArrays; ++index) {
+        const auto& input = regs.vertex_array[index];
+        VertexBinding& binding = vertex_bindings[index];
+        binding.raw = 0;
+        binding.enabled.Assign(input.IsEnabled() ? 1 : 0);
+        binding.stride.Assign(static_cast<u16>(input.stride.Value()));
+    }
 }
 
 std::size_t FixedPipelineState::Hash() const noexcept {
-    const u64 hash = Common::CityHash64(reinterpret_cast<const char*>(this), sizeof *this);
+    const u64 hash = Common::CityHash64(reinterpret_cast<const char*>(this), Size());
     return static_cast<std::size_t>(hash);
 }
 
 bool FixedPipelineState::operator==(const FixedPipelineState& rhs) const noexcept {
-    return std::memcmp(this, &rhs, sizeof *this) == 0;
+    return std::memcmp(this, &rhs, Size()) == 0;
 }
 
 u32 FixedPipelineState::PackComparisonOp(Maxwell::ComparisonOp op) noexcept {

src/video_core/renderer_vulkan/fixed_pipeline_state.h

@@ -60,14 +60,6 @@ struct FixedPipelineState {
 
         void Fill(const Maxwell& regs, std::size_t index);
 
-        std::size_t Hash() const noexcept;
-
-        bool operator==(const BlendingAttachment& rhs) const noexcept;
-
-        bool operator!=(const BlendingAttachment& rhs) const noexcept {
-            return !operator==(rhs);
-        }
-
         constexpr std::array<bool, 4> Mask() const noexcept {
             return {mask_r != 0, mask_g != 0, mask_b != 0, mask_a != 0};
         }
@@ -97,83 +89,77 @@ struct FixedPipelineState {
         }
     };
 
-    struct VertexInput {
-        union Binding {
-            u16 raw;
-            BitField<0, 1, u16> enabled;
-            BitField<1, 12, u16> stride;
-        };
+    union VertexAttribute {
+        u32 raw;
+        BitField<0, 1, u32> enabled;
+        BitField<1, 5, u32> buffer;
+        BitField<6, 14, u32> offset;
+        BitField<20, 3, u32> type;
+        BitField<23, 6, u32> size;
 
-        union Attribute {
-            u32 raw;
-            BitField<0, 1, u32> enabled;
-            BitField<1, 5, u32> buffer;
-            BitField<6, 14, u32> offset;
-            BitField<20, 3, u32> type;
-            BitField<23, 6, u32> size;
-
-            constexpr Maxwell::VertexAttribute::Type Type() const noexcept {
-                return static_cast<Maxwell::VertexAttribute::Type>(type.Value());
-            }
-
-            constexpr Maxwell::VertexAttribute::Size Size() const noexcept {
-                return static_cast<Maxwell::VertexAttribute::Size>(size.Value());
-            }
-        };
-
-        std::array<Binding, Maxwell::NumVertexArrays> bindings;
-        std::array<u32, Maxwell::NumVertexArrays> binding_divisors;
-        std::array<Attribute, Maxwell::NumVertexAttributes> attributes;
-
-        void SetBinding(std::size_t index, bool enabled, u32 stride, u32 divisor) noexcept {
-            auto& binding = bindings[index];
-            binding.raw = 0;
-            binding.enabled.Assign(enabled ? 1 : 0);
-            binding.stride.Assign(static_cast<u16>(stride));
-            binding_divisors[index] = divisor;
+        constexpr Maxwell::VertexAttribute::Type Type() const noexcept {
+            return static_cast<Maxwell::VertexAttribute::Type>(type.Value());
         }
 
-        void SetAttribute(std::size_t index, bool enabled, u32 buffer, u32 offset,
-                          Maxwell::VertexAttribute::Type type,
-                          Maxwell::VertexAttribute::Size size) noexcept {
-            auto& attribute = attributes[index];
-            attribute.raw = 0;
-            attribute.enabled.Assign(enabled ? 1 : 0);
-            attribute.buffer.Assign(buffer);
-            attribute.offset.Assign(offset);
-            attribute.type.Assign(static_cast<u32>(type));
-            attribute.size.Assign(static_cast<u32>(size));
+        constexpr Maxwell::VertexAttribute::Size Size() const noexcept {
+            return static_cast<Maxwell::VertexAttribute::Size>(size.Value());
         }
     };
 
-    struct Rasterizer {
+    template <std::size_t Position>
+    union StencilFace {
+        BitField<Position + 0, 3, u32> action_stencil_fail;
+        BitField<Position + 3, 3, u32> action_depth_fail;
+        BitField<Position + 6, 3, u32> action_depth_pass;
+        BitField<Position + 9, 3, u32> test_func;
+
+        Maxwell::StencilOp ActionStencilFail() const noexcept {
+            return UnpackStencilOp(action_stencil_fail);
+        }
+
+        Maxwell::StencilOp ActionDepthFail() const noexcept {
+            return UnpackStencilOp(action_depth_fail);
+        }
+
+        Maxwell::StencilOp ActionDepthPass() const noexcept {
+            return UnpackStencilOp(action_depth_pass);
+        }
+
+        Maxwell::ComparisonOp TestFunc() const noexcept {
+            return UnpackComparisonOp(test_func);
+        }
+    };
+
+    union VertexBinding {
+        u16 raw;
+        BitField<0, 12, u16> stride;
+        BitField<12, 1, u16> enabled;
+    };
+
+    struct DynamicState {
         union {
-            u32 raw;
-            BitField<0, 4, u32> topology;
-            BitField<4, 1, u32> primitive_restart_enable;
-            BitField<5, 1, u32> cull_enable;
-            BitField<6, 1, u32> depth_bias_enable;
-            BitField<7, 1, u32> depth_clamp_disabled;
-            BitField<8, 1, u32> ndc_minus_one_to_one;
-            BitField<9, 2, u32> cull_face;
-            BitField<11, 1, u32> front_face;
-            BitField<12, 2, u32> polygon_mode;
-            BitField<14, 5, u32> patch_control_points_minus_one;
-            BitField<19, 2, u32> tessellation_primitive;
-            BitField<21, 2, u32> tessellation_spacing;
-            BitField<23, 1, u32> tessellation_clockwise;
-            BitField<24, 1, u32> logic_op_enable;
-            BitField<25, 4, u32> logic_op;
-            BitField<29, 1, u32> rasterize_enable;
+            u32 raw1;
+            StencilFace<0> front;
+            StencilFace<12> back;
+            BitField<24, 1, u32> stencil_enable;
+            BitField<25, 1, u32> depth_write_enable;
+            BitField<26, 1, u32> depth_bounds_enable;
+            BitField<27, 1, u32> depth_test_enable;
+            BitField<28, 1, u32> front_face;
+            BitField<29, 3, u32> depth_test_func;
         };
+        union {
+            u32 raw2;
+            BitField<0, 4, u32> topology;
+            BitField<4, 2, u32> cull_face;
+            BitField<6, 1, u32> cull_enable;
+        };
+        std::array<VertexBinding, Maxwell::NumVertexArrays> vertex_bindings;
 
-        // TODO(Rodrigo): Move this to push constants
-        u32 point_size;
+        void Fill(const Maxwell& regs);
 
-        void Fill(const Maxwell& regs) noexcept;
-
-        constexpr Maxwell::PrimitiveTopology Topology() const noexcept {
-            return static_cast<Maxwell::PrimitiveTopology>(topology.Value());
+        Maxwell::ComparisonOp DepthTestFunc() const noexcept {
+            return UnpackComparisonOp(depth_test_func);
         }
 
         Maxwell::CullFace CullFace() const noexcept {
@@ -183,70 +169,36 @@ struct FixedPipelineState {
         Maxwell::FrontFace FrontFace() const noexcept {
             return UnpackFrontFace(front_face.Value());
         }
-    };
 
-    struct DepthStencil {
-        template <std::size_t Position>
-        union StencilFace {
-            BitField<Position + 0, 3, u32> action_stencil_fail;
-            BitField<Position + 3, 3, u32> action_depth_fail;
-            BitField<Position + 6, 3, u32> action_depth_pass;
-            BitField<Position + 9, 3, u32> test_func;
-
-            Maxwell::StencilOp ActionStencilFail() const noexcept {
-                return UnpackStencilOp(action_stencil_fail);
-            }
-
-            Maxwell::StencilOp ActionDepthFail() const noexcept {
-                return UnpackStencilOp(action_depth_fail);
-            }
-
-            Maxwell::StencilOp ActionDepthPass() const noexcept {
-                return UnpackStencilOp(action_depth_pass);
-            }
-
-            Maxwell::ComparisonOp TestFunc() const noexcept {
-                return UnpackComparisonOp(test_func);
-            }
-        };
-
-        union {
-            u32 raw;
-            StencilFace<0> front;
-            StencilFace<12> back;
-            BitField<24, 1, u32> depth_test_enable;
-            BitField<25, 1, u32> depth_write_enable;
-            BitField<26, 1, u32> depth_bounds_enable;
-            BitField<27, 1, u32> stencil_enable;
-            BitField<28, 3, u32> depth_test_func;
-        };
-
-        void Fill(const Maxwell& regs) noexcept;
-
-        Maxwell::ComparisonOp DepthTestFunc() const noexcept {
-            return UnpackComparisonOp(depth_test_func);
+        constexpr Maxwell::PrimitiveTopology Topology() const noexcept {
+            return static_cast<Maxwell::PrimitiveTopology>(topology.Value());
         }
     };
 
-    struct ColorBlending {
-        std::array<BlendingAttachment, Maxwell::NumRenderTargets> attachments;
-
-        void Fill(const Maxwell& regs) noexcept;
+    union {
+        u32 raw;
+        BitField<0, 1, u32> no_extended_dynamic_state;
+        BitField<2, 1, u32> primitive_restart_enable;
+        BitField<3, 1, u32> depth_bias_enable;
+        BitField<4, 1, u32> depth_clamp_disabled;
+        BitField<5, 1, u32> ndc_minus_one_to_one;
+        BitField<6, 2, u32> polygon_mode;
+        BitField<8, 5, u32> patch_control_points_minus_one;
+        BitField<13, 2, u32> tessellation_primitive;
+        BitField<15, 2, u32> tessellation_spacing;
+        BitField<17, 1, u32> tessellation_clockwise;
+        BitField<18, 1, u32> logic_op_enable;
+        BitField<19, 4, u32> logic_op;
+        BitField<23, 1, u32> rasterize_enable;
     };
+    u32 point_size;
+    std::array<u32, Maxwell::NumVertexArrays> binding_divisors;
+    std::array<VertexAttribute, Maxwell::NumVertexAttributes> attributes;
+    std::array<BlendingAttachment, Maxwell::NumRenderTargets> attachments;
+    std::array<u16, Maxwell::NumViewports> viewport_swizzles;
+    DynamicState dynamic_state;
 
-    struct ViewportSwizzles {
-        std::array<u16, Maxwell::NumViewports> swizzles;
-
-        void Fill(const Maxwell& regs) noexcept;
-    };
-
-    VertexInput vertex_input;
-    Rasterizer rasterizer;
-    DepthStencil depth_stencil;
-    ColorBlending color_blending;
-    ViewportSwizzles viewport_swizzles;
-
-    void Fill(const Maxwell& regs);
+    void Fill(const Maxwell& regs, bool has_extended_dynamic_state);
 
     std::size_t Hash() const noexcept;
 
@@ -255,6 +207,11 @@ struct FixedPipelineState {
     bool operator!=(const FixedPipelineState& rhs) const noexcept {
         return !operator==(rhs);
     }
+
+    std::size_t Size() const noexcept {
+        const std::size_t total_size = sizeof *this;
+        return total_size - (no_extended_dynamic_state != 0 ? 0 : sizeof(DynamicState));
+    }
 };
 static_assert(std::has_unique_object_representations_v<FixedPipelineState>);
 static_assert(std::is_trivially_copyable_v<FixedPipelineState>);

src/video_core/renderer_vulkan/renderer_vulkan.cpp

@@ -409,7 +409,7 @@ bool RendererVulkan::PickDevices() {
         return false;
     }
 
-    const s32 device_index = Settings::values.vulkan_device;
+    const s32 device_index = Settings::values.vulkan_device.GetValue();
     if (device_index < 0 || device_index >= static_cast<s32>(devices->size())) {
         LOG_ERROR(Render_Vulkan, "Invalid device index {}!", device_index);
         return false;

src/video_core/renderer_vulkan/vk_device.cpp

@@ -313,6 +313,16 @@ bool VKDevice::Create() {
         LOG_INFO(Render_Vulkan, "Device doesn't support custom border colors");
     }
 
+    VkPhysicalDeviceExtendedDynamicStateFeaturesEXT dynamic_state;
+    if (ext_extended_dynamic_state) {
+        dynamic_state.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTENDED_DYNAMIC_STATE_FEATURES_EXT;
+        dynamic_state.pNext = nullptr;
+        dynamic_state.extendedDynamicState = VK_TRUE;
+        SetNext(next, dynamic_state);
+    } else {
+        LOG_INFO(Render_Vulkan, "Device doesn't support extended dynamic state");
+    }
+
     if (!ext_depth_range_unrestricted) {
         LOG_INFO(Render_Vulkan, "Device doesn't support depth range unrestricted");
     }
@@ -541,6 +551,7 @@ std::vector<const char*> VKDevice::LoadExtensions() {
     bool has_ext_subgroup_size_control{};
     bool has_ext_transform_feedback{};
     bool has_ext_custom_border_color{};
+    bool has_ext_extended_dynamic_state{};
     for (const auto& extension : physical.EnumerateDeviceExtensionProperties()) {
         Test(extension, nv_viewport_swizzle, VK_NV_VIEWPORT_SWIZZLE_EXTENSION_NAME, true);
         Test(extension, khr_uniform_buffer_standard_layout,
@@ -558,6 +569,8 @@ std::vector<const char*> VKDevice::LoadExtensions() {
              false);
         Test(extension, has_ext_custom_border_color, VK_EXT_CUSTOM_BORDER_COLOR_EXTENSION_NAME,
              false);
+        Test(extension, has_ext_extended_dynamic_state,
+             VK_EXT_EXTENDED_DYNAMIC_STATE_EXTENSION_NAME, false);
         if (Settings::values.renderer_debug) {
             Test(extension, nv_device_diagnostics_config,
                  VK_NV_DEVICE_DIAGNOSTICS_CONFIG_EXTENSION_NAME, true);
@@ -643,6 +656,19 @@ std::vector<const char*> VKDevice::LoadExtensions() {
         }
     }
 
+    if (has_ext_extended_dynamic_state) {
+        VkPhysicalDeviceExtendedDynamicStateFeaturesEXT dynamic_state;
+        dynamic_state.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_EXTENDED_DYNAMIC_STATE_FEATURES_EXT;
+        dynamic_state.pNext = nullptr;
+        features.pNext = &dynamic_state;
+        physical.GetFeatures2KHR(features);
+
+        if (dynamic_state.extendedDynamicState) {
+            extensions.push_back(VK_EXT_EXTENDED_DYNAMIC_STATE_EXTENSION_NAME);
+            ext_extended_dynamic_state = true;
+        }
+    }
+
     return extensions;
 }
 

src/video_core/renderer_vulkan/vk_device.h

@@ -182,6 +182,11 @@ public:
         return ext_custom_border_color;
     }
 
+    /// Returns true if the device supports VK_EXT_extended_dynamic_state.
+    bool IsExtExtendedDynamicStateSupported() const {
+        return ext_extended_dynamic_state;
+    }
+
     /// Returns the vendor name reported from Vulkan.
     std::string_view GetVendorName() const {
         return vendor_name;
@@ -239,6 +244,7 @@ private:
     bool ext_shader_viewport_index_layer{};    ///< Support for VK_EXT_shader_viewport_index_layer.
     bool ext_transform_feedback{};             ///< Support for VK_EXT_transform_feedback.
     bool ext_custom_border_color{};            ///< Support for VK_EXT_custom_border_color.
+    bool ext_extended_dynamic_state{};         ///< Support for VK_EXT_extended_dynamic_state.
     bool nv_device_diagnostics_config{};       ///< Support for VK_NV_device_diagnostics_config.
 
     // Telemetry parameters

src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp

@@ -176,20 +176,32 @@ std::vector<vk::ShaderModule> VKGraphicsPipeline::CreateShaderModules(
 
 vk::Pipeline VKGraphicsPipeline::CreatePipeline(const RenderPassParams& renderpass_params,
                                                 const SPIRVProgram& program) const {
-    const auto& vi = fixed_state.vertex_input;
-    const auto& ds = fixed_state.depth_stencil;
-    const auto& cd = fixed_state.color_blending;
-    const auto& rs = fixed_state.rasterizer;
-    const auto& viewport_swizzles = fixed_state.viewport_swizzles.swizzles;
+    const auto& state = fixed_state;
+    const auto& viewport_swizzles = state.viewport_swizzles;
+
+    FixedPipelineState::DynamicState dynamic;
+    if (device.IsExtExtendedDynamicStateSupported()) {
+        // Insert dummy values, as long as they are valid they don't matter as extended dynamic
+        // state is ignored
+        dynamic.raw1 = 0;
+        dynamic.raw2 = 0;
+        for (FixedPipelineState::VertexBinding& binding : dynamic.vertex_bindings) {
+            // Enable all vertex bindings
+            binding.raw = 0;
+            binding.enabled.Assign(1);
+        }
+    } else {
+        dynamic = state.dynamic_state;
+    }
 
     std::vector<VkVertexInputBindingDescription> vertex_bindings;
     std::vector<VkVertexInputBindingDivisorDescriptionEXT> vertex_binding_divisors;
-    for (std::size_t index = 0; index < std::size(vi.bindings); ++index) {
-        const auto& binding = vi.bindings[index];
+    for (std::size_t index = 0; index < Maxwell::NumVertexArrays; ++index) {
+        const auto& binding = dynamic.vertex_bindings[index];
         if (!binding.enabled) {
             continue;
         }
-        const bool instanced = vi.binding_divisors[index] != 0;
+        const bool instanced = state.binding_divisors[index] != 0;
         const auto rate = instanced ? VK_VERTEX_INPUT_RATE_INSTANCE : VK_VERTEX_INPUT_RATE_VERTEX;
 
         auto& vertex_binding = vertex_bindings.emplace_back();
@@ -200,14 +212,14 @@ vk::Pipeline VKGraphicsPipeline::CreatePipeline(const RenderPassParams& renderpa
         if (instanced) {
             auto& binding_divisor = vertex_binding_divisors.emplace_back();
             binding_divisor.binding = static_cast<u32>(index);
-            binding_divisor.divisor = vi.binding_divisors[index];
+            binding_divisor.divisor = state.binding_divisors[index];
         }
     }
 
     std::vector<VkVertexInputAttributeDescription> vertex_attributes;
     const auto& input_attributes = program[0]->entries.attributes;
-    for (std::size_t index = 0; index < std::size(vi.attributes); ++index) {
-        const auto& attribute = vi.attributes[index];
+    for (std::size_t index = 0; index < state.attributes.size(); ++index) {
+        const auto& attribute = state.attributes[index];
         if (!attribute.enabled) {
             continue;
         }
@@ -244,15 +256,15 @@ vk::Pipeline VKGraphicsPipeline::CreatePipeline(const RenderPassParams& renderpa
     input_assembly_ci.sType = VK_STRUCTURE_TYPE_PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO;
     input_assembly_ci.pNext = nullptr;
     input_assembly_ci.flags = 0;
-    input_assembly_ci.topology = MaxwellToVK::PrimitiveTopology(device, rs.Topology());
+    input_assembly_ci.topology = MaxwellToVK::PrimitiveTopology(device, dynamic.Topology());
     input_assembly_ci.primitiveRestartEnable =
-        rs.primitive_restart_enable != 0 && SupportsPrimitiveRestart(input_assembly_ci.topology);
+        state.primitive_restart_enable != 0 && SupportsPrimitiveRestart(input_assembly_ci.topology);
 
     VkPipelineTessellationStateCreateInfo tessellation_ci;
     tessellation_ci.sType = VK_STRUCTURE_TYPE_PIPELINE_TESSELLATION_STATE_CREATE_INFO;
     tessellation_ci.pNext = nullptr;
     tessellation_ci.flags = 0;
-    tessellation_ci.patchControlPoints = rs.patch_control_points_minus_one.Value() + 1;
+    tessellation_ci.patchControlPoints = state.patch_control_points_minus_one.Value() + 1;
 
     VkPipelineViewportStateCreateInfo viewport_ci;
     viewport_ci.sType = VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO;
@@ -280,13 +292,13 @@ vk::Pipeline VKGraphicsPipeline::CreatePipeline(const RenderPassParams& renderpa
     rasterization_ci.sType = VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO;
     rasterization_ci.pNext = nullptr;
     rasterization_ci.flags = 0;
-    rasterization_ci.depthClampEnable = rs.depth_clamp_disabled == 0 ? VK_TRUE : VK_FALSE;
-    rasterization_ci.rasterizerDiscardEnable = rs.rasterize_enable == 0 ? VK_TRUE : VK_FALSE;
+    rasterization_ci.depthClampEnable = state.depth_clamp_disabled == 0 ? VK_TRUE : VK_FALSE;
+    rasterization_ci.rasterizerDiscardEnable = state.rasterize_enable == 0 ? VK_TRUE : VK_FALSE;
     rasterization_ci.polygonMode = VK_POLYGON_MODE_FILL;
     rasterization_ci.cullMode =
-        rs.cull_enable ? MaxwellToVK::CullFace(rs.CullFace()) : VK_CULL_MODE_NONE;
-    rasterization_ci.frontFace = MaxwellToVK::FrontFace(rs.FrontFace());
-    rasterization_ci.depthBiasEnable = rs.depth_bias_enable;
+        dynamic.cull_enable ? MaxwellToVK::CullFace(dynamic.CullFace()) : VK_CULL_MODE_NONE;
+    rasterization_ci.frontFace = MaxwellToVK::FrontFace(dynamic.FrontFace());
+    rasterization_ci.depthBiasEnable = state.depth_bias_enable;
     rasterization_ci.depthBiasConstantFactor = 0.0f;
     rasterization_ci.depthBiasClamp = 0.0f;
     rasterization_ci.depthBiasSlopeFactor = 0.0f;
@@ -307,14 +319,15 @@ vk::Pipeline VKGraphicsPipeline::CreatePipeline(const RenderPassParams& renderpa
     depth_stencil_ci.sType = VK_STRUCTURE_TYPE_PIPELINE_DEPTH_STENCIL_STATE_CREATE_INFO;
     depth_stencil_ci.pNext = nullptr;
     depth_stencil_ci.flags = 0;
-    depth_stencil_ci.depthTestEnable = ds.depth_test_enable;
-    depth_stencil_ci.depthWriteEnable = ds.depth_write_enable;
-    depth_stencil_ci.depthCompareOp =
-        ds.depth_test_enable ? MaxwellToVK::ComparisonOp(ds.DepthTestFunc()) : VK_COMPARE_OP_ALWAYS;
-    depth_stencil_ci.depthBoundsTestEnable = ds.depth_bounds_enable;
-    depth_stencil_ci.stencilTestEnable = ds.stencil_enable;
-    depth_stencil_ci.front = GetStencilFaceState(ds.front);
-    depth_stencil_ci.back = GetStencilFaceState(ds.back);
+    depth_stencil_ci.depthTestEnable = dynamic.depth_test_enable;
+    depth_stencil_ci.depthWriteEnable = dynamic.depth_write_enable;
+    depth_stencil_ci.depthCompareOp = dynamic.depth_test_enable
+                                          ? MaxwellToVK::ComparisonOp(dynamic.DepthTestFunc())
+                                          : VK_COMPARE_OP_ALWAYS;
+    depth_stencil_ci.depthBoundsTestEnable = dynamic.depth_bounds_enable;
+    depth_stencil_ci.stencilTestEnable = dynamic.stencil_enable;
+    depth_stencil_ci.front = GetStencilFaceState(dynamic.front);
+    depth_stencil_ci.back = GetStencilFaceState(dynamic.back);
     depth_stencil_ci.minDepthBounds = 0.0f;
     depth_stencil_ci.maxDepthBounds = 0.0f;
 
@@ -324,7 +337,7 @@ vk::Pipeline VKGraphicsPipeline::CreatePipeline(const RenderPassParams& renderpa
         static constexpr std::array COMPONENT_TABLE = {
             VK_COLOR_COMPONENT_R_BIT, VK_COLOR_COMPONENT_G_BIT, VK_COLOR_COMPONENT_B_BIT,
             VK_COLOR_COMPONENT_A_BIT};
-        const auto& blend = cd.attachments[index];
+        const auto& blend = state.attachments[index];
 
         VkColorComponentFlags color_components = 0;
         for (std::size_t i = 0; i < COMPONENT_TABLE.size(); ++i) {
@@ -354,11 +367,27 @@ vk::Pipeline VKGraphicsPipeline::CreatePipeline(const RenderPassParams& renderpa
     color_blend_ci.pAttachments = cb_attachments.data();
     std::memset(color_blend_ci.blendConstants, 0, sizeof(color_blend_ci.blendConstants));
 
-    static constexpr std::array dynamic_states = {
+    std::vector dynamic_states = {
         VK_DYNAMIC_STATE_VIEWPORT,           VK_DYNAMIC_STATE_SCISSOR,
         VK_DYNAMIC_STATE_DEPTH_BIAS,         VK_DYNAMIC_STATE_BLEND_CONSTANTS,
         VK_DYNAMIC_STATE_DEPTH_BOUNDS,       VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK,
-        VK_DYNAMIC_STATE_STENCIL_WRITE_MASK, VK_DYNAMIC_STATE_STENCIL_REFERENCE};
+        VK_DYNAMIC_STATE_STENCIL_WRITE_MASK, VK_DYNAMIC_STATE_STENCIL_REFERENCE,
+    };
+    if (device.IsExtExtendedDynamicStateSupported()) {
+        static constexpr std::array extended = {
+            VK_DYNAMIC_STATE_CULL_MODE_EXT,
+            VK_DYNAMIC_STATE_FRONT_FACE_EXT,
+            VK_DYNAMIC_STATE_PRIMITIVE_TOPOLOGY_EXT,
+            VK_DYNAMIC_STATE_VERTEX_INPUT_BINDING_STRIDE_EXT,
+            VK_DYNAMIC_STATE_DEPTH_TEST_ENABLE_EXT,
+            VK_DYNAMIC_STATE_DEPTH_WRITE_ENABLE_EXT,
+            VK_DYNAMIC_STATE_DEPTH_COMPARE_OP_EXT,
+            VK_DYNAMIC_STATE_DEPTH_BOUNDS_TEST_ENABLE_EXT,
+            VK_DYNAMIC_STATE_STENCIL_TEST_ENABLE_EXT,
+            VK_DYNAMIC_STATE_STENCIL_OP_EXT,
+        };
+        dynamic_states.insert(dynamic_states.end(), extended.begin(), extended.end());
+    }
 
     VkPipelineDynamicStateCreateInfo dynamic_state_ci;
     dynamic_state_ci.sType = VK_STRUCTURE_TYPE_PIPELINE_DYNAMIC_STATE_CREATE_INFO;

src/video_core/renderer_vulkan/vk_pipeline_cache.cpp

@@ -116,12 +116,12 @@ u32 FillDescriptorLayout(const ShaderEntries& entries,
 } // Anonymous namespace
 
 std::size_t GraphicsPipelineCacheKey::Hash() const noexcept {
-    const u64 hash = Common::CityHash64(reinterpret_cast<const char*>(this), sizeof *this);
+    const u64 hash = Common::CityHash64(reinterpret_cast<const char*>(this), Size());
     return static_cast<std::size_t>(hash);
 }
 
 bool GraphicsPipelineCacheKey::operator==(const GraphicsPipelineCacheKey& rhs) const noexcept {
-    return std::memcmp(&rhs, this, sizeof *this) == 0;
+    return std::memcmp(&rhs, this, Size()) == 0;
 }
 
 std::size_t ComputePipelineCacheKey::Hash() const noexcept {
@@ -312,18 +312,19 @@ VKPipelineCache::DecompileShaders(const GraphicsPipelineCacheKey& key) {
     const auto& gpu = system.GPU().Maxwell3D();
 
     Specialization specialization;
-    if (fixed_state.rasterizer.Topology() == Maxwell::PrimitiveTopology::Points) {
+    if (fixed_state.dynamic_state.Topology() == Maxwell::PrimitiveTopology::Points ||
+        device.IsExtExtendedDynamicStateSupported()) {
         float point_size;
-        std::memcpy(&point_size, &fixed_state.rasterizer.point_size, sizeof(float));
+        std::memcpy(&point_size, &fixed_state.point_size, sizeof(float));
         specialization.point_size = point_size;
         ASSERT(point_size != 0.0f);
     }
     for (std::size_t i = 0; i < Maxwell::NumVertexAttributes; ++i) {
-        const auto& attribute = fixed_state.vertex_input.attributes[i];
+        const auto& attribute = fixed_state.attributes[i];
         specialization.enabled_attributes[i] = attribute.enabled.Value() != 0;
         specialization.attribute_types[i] = attribute.Type();
     }
-    specialization.ndc_minus_one_to_one = fixed_state.rasterizer.ndc_minus_one_to_one;
+    specialization.ndc_minus_one_to_one = fixed_state.ndc_minus_one_to_one;
 
     SPIRVProgram program;
     std::vector<VkDescriptorSetLayoutBinding> bindings;

src/video_core/renderer_vulkan/vk_pipeline_cache.h

@@ -44,10 +44,10 @@ class VKUpdateDescriptorQueue;
 using Maxwell = Tegra::Engines::Maxwell3D::Regs;
 
 struct GraphicsPipelineCacheKey {
-    FixedPipelineState fixed_state;
     RenderPassParams renderpass_params;
+    u32 padding;
     std::array<GPUVAddr, Maxwell::MaxShaderProgram> shaders;
-    u64 padding; // This is necessary for unique object representations
+    FixedPipelineState fixed_state;
 
     std::size_t Hash() const noexcept;
 
@@ -56,6 +56,10 @@ struct GraphicsPipelineCacheKey {
     bool operator!=(const GraphicsPipelineCacheKey& rhs) const noexcept {
         return !operator==(rhs);
     }
+
+    std::size_t Size() const noexcept {
+        return sizeof(renderpass_params) + sizeof(padding) + sizeof(shaders) + fixed_state.Size();
+    }
 };
 static_assert(std::has_unique_object_representations_v<GraphicsPipelineCacheKey>);
 static_assert(std::is_trivially_copyable_v<GraphicsPipelineCacheKey>);

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -186,13 +186,22 @@ bool HasToPreserveDepthContents(bool is_clear, const Maxwell& regs) {
            scissor.max_y < regs.zeta_height;
 }
 
+template <std::size_t N>
+std::array<VkDeviceSize, N> ExpandStrides(const std::array<u16, N>& strides) {
+    std::array<VkDeviceSize, N> expanded;
+    std::copy(strides.begin(), strides.end(), expanded.begin());
+    return expanded;
+}
+
 } // Anonymous namespace
 
 class BufferBindings final {
 public:
-    void AddVertexBinding(VkBuffer buffer, VkDeviceSize offset) {
+    void AddVertexBinding(VkBuffer buffer, VkDeviceSize offset, VkDeviceSize size, u32 stride) {
         vertex.buffers[vertex.num_buffers] = buffer;
         vertex.offsets[vertex.num_buffers] = offset;
+        vertex.sizes[vertex.num_buffers] = size;
+        vertex.strides[vertex.num_buffers] = static_cast<u16>(stride);
         ++vertex.num_buffers;
     }
 
@@ -202,76 +211,76 @@ public:
         index.type = type;
     }
 
-    void Bind(VKScheduler& scheduler) const {
+    void Bind(const VKDevice& device, VKScheduler& scheduler) const {
         // Use this large switch case to avoid dispatching more memory in the record lambda than
         // what we need. It looks horrible, but it's the best we can do on standard C++.
         switch (vertex.num_buffers) {
         case 0:
-            return BindStatic<0>(scheduler);
+            return BindStatic<0>(device, scheduler);
         case 1:
-            return BindStatic<1>(scheduler);
+            return BindStatic<1>(device, scheduler);
         case 2:
-            return BindStatic<2>(scheduler);
+            return BindStatic<2>(device, scheduler);
         case 3:
-            return BindStatic<3>(scheduler);
+            return BindStatic<3>(device, scheduler);
         case 4:
-            return BindStatic<4>(scheduler);
+            return BindStatic<4>(device, scheduler);
         case 5:
-            return BindStatic<5>(scheduler);
+            return BindStatic<5>(device, scheduler);
         case 6:
-            return BindStatic<6>(scheduler);
+            return BindStatic<6>(device, scheduler);
         case 7:
-            return BindStatic<7>(scheduler);
+            return BindStatic<7>(device, scheduler);
         case 8:
-            return BindStatic<8>(scheduler);
+            return BindStatic<8>(device, scheduler);
         case 9:
-            return BindStatic<9>(scheduler);
+            return BindStatic<9>(device, scheduler);
         case 10:
-            return BindStatic<10>(scheduler);
+            return BindStatic<10>(device, scheduler);
         case 11:
-            return BindStatic<11>(scheduler);
+            return BindStatic<11>(device, scheduler);
         case 12:
-            return BindStatic<12>(scheduler);
+            return BindStatic<12>(device, scheduler);
         case 13:
-            return BindStatic<13>(scheduler);
+            return BindStatic<13>(device, scheduler);
         case 14:
-            return BindStatic<14>(scheduler);
+            return BindStatic<14>(device, scheduler);
         case 15:
-            return BindStatic<15>(scheduler);
+            return BindStatic<15>(device, scheduler);
         case 16:
-            return BindStatic<16>(scheduler);
+            return BindStatic<16>(device, scheduler);
         case 17:
-            return BindStatic<17>(scheduler);
+            return BindStatic<17>(device, scheduler);
         case 18:
-            return BindStatic<18>(scheduler);
+            return BindStatic<18>(device, scheduler);
         case 19:
-            return BindStatic<19>(scheduler);
+            return BindStatic<19>(device, scheduler);
         case 20:
-            return BindStatic<20>(scheduler);
+            return BindStatic<20>(device, scheduler);
         case 21:
-            return BindStatic<21>(scheduler);
+            return BindStatic<21>(device, scheduler);
         case 22:
-            return BindStatic<22>(scheduler);
+            return BindStatic<22>(device, scheduler);
         case 23:
-            return BindStatic<23>(scheduler);
+            return BindStatic<23>(device, scheduler);
         case 24:
-            return BindStatic<24>(scheduler);
+            return BindStatic<24>(device, scheduler);
         case 25:
-            return BindStatic<25>(scheduler);
+            return BindStatic<25>(device, scheduler);
         case 26:
-            return BindStatic<26>(scheduler);
+            return BindStatic<26>(device, scheduler);
         case 27:
-            return BindStatic<27>(scheduler);
+            return BindStatic<27>(device, scheduler);
         case 28:
-            return BindStatic<28>(scheduler);
+            return BindStatic<28>(device, scheduler);
         case 29:
-            return BindStatic<29>(scheduler);
+            return BindStatic<29>(device, scheduler);
         case 30:
-            return BindStatic<30>(scheduler);
+            return BindStatic<30>(device, scheduler);
         case 31:
-            return BindStatic<31>(scheduler);
+            return BindStatic<31>(device, scheduler);
         case 32:
-            return BindStatic<32>(scheduler);
+            return BindStatic<32>(device, scheduler);
         }
         UNREACHABLE();
     }
@@ -282,6 +291,8 @@ private:
         std::size_t num_buffers = 0;
         std::array<VkBuffer, Maxwell::NumVertexArrays> buffers;
         std::array<VkDeviceSize, Maxwell::NumVertexArrays> offsets;
+        std::array<VkDeviceSize, Maxwell::NumVertexArrays> sizes;
+        std::array<u16, Maxwell::NumVertexArrays> strides;
     } vertex;
 
     struct {
@@ -291,15 +302,23 @@ private:
     } index;
 
     template <std::size_t N>
-    void BindStatic(VKScheduler& scheduler) const {
-        if (index.buffer) {
-            BindStatic<N, true>(scheduler);
+    void BindStatic(const VKDevice& device, VKScheduler& scheduler) const {
+        if (device.IsExtExtendedDynamicStateSupported()) {
+            if (index.buffer) {
+                BindStatic<N, true, true>(scheduler);
+            } else {
+                BindStatic<N, false, true>(scheduler);
+            }
         } else {
-            BindStatic<N, false>(scheduler);
+            if (index.buffer) {
+                BindStatic<N, true, false>(scheduler);
+            } else {
+                BindStatic<N, false, false>(scheduler);
+            }
         }
     }
 
-    template <std::size_t N, bool is_indexed>
+    template <std::size_t N, bool is_indexed, bool has_extended_dynamic_state>
     void BindStatic(VKScheduler& scheduler) const {
         static_assert(N <= Maxwell::NumVertexArrays);
         if constexpr (N == 0) {
@@ -311,6 +330,31 @@ private:
         std::copy(vertex.buffers.begin(), vertex.buffers.begin() + N, buffers.begin());
         std::copy(vertex.offsets.begin(), vertex.offsets.begin() + N, offsets.begin());
 
+        if constexpr (has_extended_dynamic_state) {
+            // With extended dynamic states we can specify the length and stride of a vertex buffer
+            std::array<VkDeviceSize, N> sizes;
+            std::array<u16, N> strides;
+            std::copy(vertex.sizes.begin(), vertex.sizes.begin() + N, sizes.begin());
+            std::copy(vertex.strides.begin(), vertex.strides.begin() + N, strides.begin());
+
+            if constexpr (is_indexed) {
+                scheduler.Record(
+                    [buffers, offsets, sizes, strides, index = index](vk::CommandBuffer cmdbuf) {
+                        cmdbuf.BindIndexBuffer(index.buffer, index.offset, index.type);
+                        cmdbuf.BindVertexBuffers2EXT(0, static_cast<u32>(N), buffers.data(),
+                                                     offsets.data(), sizes.data(),
+                                                     ExpandStrides(strides).data());
+                    });
+            } else {
+                scheduler.Record([buffers, offsets, sizes, strides](vk::CommandBuffer cmdbuf) {
+                    cmdbuf.BindVertexBuffers2EXT(0, static_cast<u32>(N), buffers.data(),
+                                                 offsets.data(), sizes.data(),
+                                                 ExpandStrides(strides).data());
+                });
+            }
+            return;
+        }
+
         if constexpr (is_indexed) {
             // Indexed draw
             scheduler.Record([buffers, offsets, index = index](vk::CommandBuffer cmdbuf) {
@@ -369,7 +413,7 @@ void RasterizerVulkan::Draw(bool is_indexed, bool is_instanced) {
 
     const auto& gpu = system.GPU().Maxwell3D();
     GraphicsPipelineCacheKey key;
-    key.fixed_state.Fill(gpu.regs);
+    key.fixed_state.Fill(gpu.regs, device.IsExtExtendedDynamicStateSupported());
 
     buffer_cache.Map(CalculateGraphicsStreamBufferSize(is_indexed));
 
@@ -402,7 +446,7 @@ void RasterizerVulkan::Draw(bool is_indexed, bool is_instanced) {
 
     UpdateDynamicStates();
 
-    buffer_bindings.Bind(scheduler);
+    buffer_bindings.Bind(device, scheduler);
 
     BeginTransformFeedback();
 
@@ -822,7 +866,7 @@ RasterizerVulkan::DrawParameters RasterizerVulkan::SetupGeometry(FixedPipelineSt
     const auto& gpu = system.GPU().Maxwell3D();
     const auto& regs = gpu.regs;
 
-    SetupVertexArrays(fixed_state.vertex_input, buffer_bindings);
+    SetupVertexArrays(buffer_bindings);
 
     const u32 base_instance = regs.vb_base_instance;
     const u32 num_instances = is_instanced ? gpu.mme_draw.instance_count : 1;
@@ -893,6 +937,17 @@ void RasterizerVulkan::UpdateDynamicStates() {
     UpdateBlendConstants(regs);
     UpdateDepthBounds(regs);
     UpdateStencilFaces(regs);
+    if (device.IsExtExtendedDynamicStateSupported()) {
+        UpdateCullMode(regs);
+        UpdateDepthBoundsTestEnable(regs);
+        UpdateDepthTestEnable(regs);
+        UpdateDepthWriteEnable(regs);
+        UpdateDepthCompareOp(regs);
+        UpdateFrontFace(regs);
+        UpdatePrimitiveTopology(regs);
+        UpdateStencilOp(regs);
+        UpdateStencilTestEnable(regs);
+    }
 }
 
 void RasterizerVulkan::BeginTransformFeedback() {
@@ -940,41 +995,25 @@ void RasterizerVulkan::EndTransformFeedback() {
         [](vk::CommandBuffer cmdbuf) { cmdbuf.EndTransformFeedbackEXT(0, 0, nullptr, nullptr); });
 }
 
-void RasterizerVulkan::SetupVertexArrays(FixedPipelineState::VertexInput& vertex_input,
-                                         BufferBindings& buffer_bindings) {
+void RasterizerVulkan::SetupVertexArrays(BufferBindings& buffer_bindings) {
     const auto& regs = system.GPU().Maxwell3D().regs;
 
-    for (std::size_t index = 0; index < Maxwell::NumVertexAttributes; ++index) {
-        const auto& attrib = regs.vertex_attrib_format[index];
-        if (attrib.IsConstant()) {
-            vertex_input.SetAttribute(index, false, 0, 0, {}, {});
-            continue;
-        }
-        vertex_input.SetAttribute(index, true, attrib.buffer, attrib.offset, attrib.type.Value(),
-                                  attrib.size.Value());
-    }
-
     for (std::size_t index = 0; index < Maxwell::NumVertexArrays; ++index) {
         const auto& vertex_array = regs.vertex_array[index];
         if (!vertex_array.IsEnabled()) {
-            vertex_input.SetBinding(index, false, 0, 0);
             continue;
         }
-        vertex_input.SetBinding(
-            index, true, vertex_array.stride,
-            regs.instanced_arrays.IsInstancingEnabled(index) ? vertex_array.divisor : 0);
-
         const GPUVAddr start{vertex_array.StartAddress()};
         const GPUVAddr end{regs.vertex_array_limit[index].LimitAddress()};
 
         ASSERT(end >= start);
-        const std::size_t size{end - start};
+        const std::size_t size = end - start;
         if (size == 0) {
-            buffer_bindings.AddVertexBinding(DefaultBuffer(), 0);
+            buffer_bindings.AddVertexBinding(DefaultBuffer(), 0, DEFAULT_BUFFER_SIZE, 0);
             continue;
         }
         const auto info = buffer_cache.UploadMemory(start, size);
-        buffer_bindings.AddVertexBinding(info.handle, info.offset);
+        buffer_bindings.AddVertexBinding(info.handle, info.offset, size, vertex_array.stride);
     }
 }
 
@@ -1326,6 +1365,117 @@ void RasterizerVulkan::UpdateStencilFaces(Tegra::Engines::Maxwell3D::Regs& regs)
     }
 }
 
+void RasterizerVulkan::UpdateCullMode(Tegra::Engines::Maxwell3D::Regs& regs) {
+    if (!state_tracker.TouchCullMode()) {
+        return;
+    }
+    scheduler.Record(
+        [enabled = regs.cull_test_enabled, cull_face = regs.cull_face](vk::CommandBuffer cmdbuf) {
+            cmdbuf.SetCullModeEXT(enabled ? MaxwellToVK::CullFace(cull_face) : VK_CULL_MODE_NONE);
+        });
+}
+
+void RasterizerVulkan::UpdateDepthBoundsTestEnable(Tegra::Engines::Maxwell3D::Regs& regs) {
+    if (!state_tracker.TouchDepthBoundsTestEnable()) {
+        return;
+    }
+    scheduler.Record([enable = regs.depth_bounds_enable](vk::CommandBuffer cmdbuf) {
+        cmdbuf.SetDepthBoundsTestEnableEXT(enable);
+    });
+}
+
+void RasterizerVulkan::UpdateDepthTestEnable(Tegra::Engines::Maxwell3D::Regs& regs) {
+    if (!state_tracker.TouchDepthTestEnable()) {
+        return;
+    }
+    scheduler.Record([enable = regs.depth_test_enable](vk::CommandBuffer cmdbuf) {
+        cmdbuf.SetDepthTestEnableEXT(enable);
+    });
+}
+
+void RasterizerVulkan::UpdateDepthWriteEnable(Tegra::Engines::Maxwell3D::Regs& regs) {
+    if (!state_tracker.TouchDepthWriteEnable()) {
+        return;
+    }
+    scheduler.Record([enable = regs.depth_write_enabled](vk::CommandBuffer cmdbuf) {
+        cmdbuf.SetDepthWriteEnableEXT(enable);
+    });
+}
+
+void RasterizerVulkan::UpdateDepthCompareOp(Tegra::Engines::Maxwell3D::Regs& regs) {
+    if (!state_tracker.TouchDepthCompareOp()) {
+        return;
+    }
+    scheduler.Record([func = regs.depth_test_func](vk::CommandBuffer cmdbuf) {
+        cmdbuf.SetDepthCompareOpEXT(MaxwellToVK::ComparisonOp(func));
+    });
+}
+
+void RasterizerVulkan::UpdateFrontFace(Tegra::Engines::Maxwell3D::Regs& regs) {
+    if (!state_tracker.TouchFrontFace()) {
+        return;
+    }
+
+    VkFrontFace front_face = MaxwellToVK::FrontFace(regs.front_face);
+    if (regs.screen_y_control.triangle_rast_flip != 0) {
+        front_face = front_face == VK_FRONT_FACE_CLOCKWISE ? VK_FRONT_FACE_COUNTER_CLOCKWISE
+                                                           : VK_FRONT_FACE_CLOCKWISE;
+    }
+    scheduler.Record(
+        [front_face](vk::CommandBuffer cmdbuf) { cmdbuf.SetFrontFaceEXT(front_face); });
+}
+
+void RasterizerVulkan::UpdatePrimitiveTopology(Tegra::Engines::Maxwell3D::Regs& regs) {
+    if (!state_tracker.TouchPrimitiveTopology()) {
+        return;
+    }
+    const Maxwell::PrimitiveTopology primitive_topology = regs.draw.topology.Value();
+    scheduler.Record([this, primitive_topology](vk::CommandBuffer cmdbuf) {
+        cmdbuf.SetPrimitiveTopologyEXT(MaxwellToVK::PrimitiveTopology(device, primitive_topology));
+    });
+}
+
+void RasterizerVulkan::UpdateStencilOp(Tegra::Engines::Maxwell3D::Regs& regs) {
+    if (!state_tracker.TouchStencilOp()) {
+        return;
+    }
+    const Maxwell::StencilOp fail = regs.stencil_front_op_fail;
+    const Maxwell::StencilOp zfail = regs.stencil_front_op_zfail;
+    const Maxwell::StencilOp zpass = regs.stencil_front_op_zpass;
+    const Maxwell::ComparisonOp compare = regs.stencil_front_func_func;
+    if (regs.stencil_two_side_enable) {
+        scheduler.Record([fail, zfail, zpass, compare](vk::CommandBuffer cmdbuf) {
+            cmdbuf.SetStencilOpEXT(VK_STENCIL_FACE_FRONT_AND_BACK, MaxwellToVK::StencilOp(fail),
+                                   MaxwellToVK::StencilOp(zpass), MaxwellToVK::StencilOp(zfail),
+                                   MaxwellToVK::ComparisonOp(compare));
+        });
+    } else {
+        const Maxwell::StencilOp back_fail = regs.stencil_back_op_fail;
+        const Maxwell::StencilOp back_zfail = regs.stencil_back_op_zfail;
+        const Maxwell::StencilOp back_zpass = regs.stencil_back_op_zpass;
+        const Maxwell::ComparisonOp back_compare = regs.stencil_back_func_func;
+        scheduler.Record([fail, zfail, zpass, compare, back_fail, back_zfail, back_zpass,
+                          back_compare](vk::CommandBuffer cmdbuf) {
+            cmdbuf.SetStencilOpEXT(VK_STENCIL_FACE_FRONT_BIT, MaxwellToVK::StencilOp(fail),
+                                   MaxwellToVK::StencilOp(zpass), MaxwellToVK::StencilOp(zfail),
+                                   MaxwellToVK::ComparisonOp(compare));
+            cmdbuf.SetStencilOpEXT(VK_STENCIL_FACE_BACK_BIT, MaxwellToVK::StencilOp(back_fail),
+                                   MaxwellToVK::StencilOp(back_zpass),
+                                   MaxwellToVK::StencilOp(back_zfail),
+                                   MaxwellToVK::ComparisonOp(back_compare));
+        });
+    }
+}
+
+void RasterizerVulkan::UpdateStencilTestEnable(Tegra::Engines::Maxwell3D::Regs& regs) {
+    if (!state_tracker.TouchStencilTestEnable()) {
+        return;
+    }
+    scheduler.Record([enable = regs.stencil_enable](vk::CommandBuffer cmdbuf) {
+        cmdbuf.SetStencilTestEnableEXT(enable);
+    });
+}
+
 std::size_t RasterizerVulkan::CalculateGraphicsStreamBufferSize(bool is_indexed) const {
     std::size_t size = CalculateVertexArraysSize();
     if (is_indexed) {

src/video_core/renderer_vulkan/vk_rasterizer.h

@@ -185,8 +185,7 @@ private:
 
     bool WalkAttachmentOverlaps(const CachedSurfaceView& attachment);
 
-    void SetupVertexArrays(FixedPipelineState::VertexInput& vertex_input,
-                           BufferBindings& buffer_bindings);
+    void SetupVertexArrays(BufferBindings& buffer_bindings);
 
     void SetupIndexBuffer(BufferBindings& buffer_bindings, DrawParameters& params, bool is_indexed);
 
@@ -246,6 +245,16 @@ private:
     void UpdateDepthBounds(Tegra::Engines::Maxwell3D::Regs& regs);
     void UpdateStencilFaces(Tegra::Engines::Maxwell3D::Regs& regs);
 
+    void UpdateCullMode(Tegra::Engines::Maxwell3D::Regs& regs);
+    void UpdateDepthBoundsTestEnable(Tegra::Engines::Maxwell3D::Regs& regs);
+    void UpdateDepthTestEnable(Tegra::Engines::Maxwell3D::Regs& regs);
+    void UpdateDepthWriteEnable(Tegra::Engines::Maxwell3D::Regs& regs);
+    void UpdateDepthCompareOp(Tegra::Engines::Maxwell3D::Regs& regs);
+    void UpdateFrontFace(Tegra::Engines::Maxwell3D::Regs& regs);
+    void UpdatePrimitiveTopology(Tegra::Engines::Maxwell3D::Regs& regs);
+    void UpdateStencilOp(Tegra::Engines::Maxwell3D::Regs& regs);
+    void UpdateStencilTestEnable(Tegra::Engines::Maxwell3D::Regs& regs);
+
     std::size_t CalculateGraphicsStreamBufferSize(bool is_indexed) const;
 
     std::size_t CalculateComputeStreamBufferSize() const;

src/video_core/renderer_vulkan/vk_state_tracker.cpp

@@ -36,6 +36,15 @@ Flags MakeInvalidationFlags() {
     flags[BlendConstants] = true;
     flags[DepthBounds] = true;
     flags[StencilProperties] = true;
+    flags[CullMode] = true;
+    flags[DepthBoundsEnable] = true;
+    flags[DepthTestEnable] = true;
+    flags[DepthWriteEnable] = true;
+    flags[DepthCompareOp] = true;
+    flags[FrontFace] = true;
+    flags[PrimitiveTopology] = true;
+    flags[StencilOp] = true;
+    flags[StencilTestEnable] = true;
     return flags;
 }
 
@@ -75,6 +84,57 @@ void SetupDirtyStencilProperties(Tables& tables) {
     table[OFF(stencil_back_func_mask)] = StencilProperties;
 }
 
+void SetupDirtyCullMode(Tables& tables) {
+    auto& table = tables[0];
+    table[OFF(cull_face)] = CullMode;
+    table[OFF(cull_test_enabled)] = CullMode;
+}
+
+void SetupDirtyDepthBoundsEnable(Tables& tables) {
+    tables[0][OFF(depth_bounds_enable)] = DepthBoundsEnable;
+}
+
+void SetupDirtyDepthTestEnable(Tables& tables) {
+    tables[0][OFF(depth_test_enable)] = DepthTestEnable;
+}
+
+void SetupDirtyDepthWriteEnable(Tables& tables) {
+    tables[0][OFF(depth_write_enabled)] = DepthWriteEnable;
+}
+
+void SetupDirtyDepthCompareOp(Tables& tables) {
+    tables[0][OFF(depth_test_func)] = DepthCompareOp;
+}
+
+void SetupDirtyFrontFace(Tables& tables) {
+    auto& table = tables[0];
+    table[OFF(front_face)] = FrontFace;
+    table[OFF(screen_y_control)] = FrontFace;
+}
+
+void SetupDirtyPrimitiveTopology(Tables& tables) {
+    tables[0][OFF(draw.topology)] = PrimitiveTopology;
+}
+
+void SetupDirtyStencilOp(Tables& tables) {
+    auto& table = tables[0];
+    table[OFF(stencil_front_op_fail)] = StencilOp;
+    table[OFF(stencil_front_op_zfail)] = StencilOp;
+    table[OFF(stencil_front_op_zpass)] = StencilOp;
+    table[OFF(stencil_front_func_func)] = StencilOp;
+    table[OFF(stencil_back_op_fail)] = StencilOp;
+    table[OFF(stencil_back_op_zfail)] = StencilOp;
+    table[OFF(stencil_back_op_zpass)] = StencilOp;
+    table[OFF(stencil_back_func_func)] = StencilOp;
+
+    // Table 0 is used by StencilProperties
+    tables[1][OFF(stencil_two_side_enable)] = StencilOp;
+}
+
+void SetupDirtyStencilTestEnable(Tables& tables) {
+    tables[0][OFF(stencil_enable)] = StencilTestEnable;
+}
+
 } // Anonymous namespace
 
 StateTracker::StateTracker(Core::System& system)
@@ -90,6 +150,14 @@ void StateTracker::Initialize() {
     SetupDirtyBlendConstants(tables);
     SetupDirtyDepthBounds(tables);
     SetupDirtyStencilProperties(tables);
+    SetupDirtyCullMode(tables);
+    SetupDirtyDepthBoundsEnable(tables);
+    SetupDirtyDepthTestEnable(tables);
+    SetupDirtyDepthWriteEnable(tables);
+    SetupDirtyDepthCompareOp(tables);
+    SetupDirtyFrontFace(tables);
+    SetupDirtyPrimitiveTopology(tables);
+    SetupDirtyStencilOp(tables);
 }
 
 void StateTracker::InvalidateCommandBufferState() {

src/video_core/renderer_vulkan/vk_state_tracker.h

@@ -26,6 +26,16 @@ enum : u8 {
     DepthBounds,
     StencilProperties,
 
+    CullMode,
+    DepthBoundsEnable,
+    DepthTestEnable,
+    DepthWriteEnable,
+    DepthCompareOp,
+    FrontFace,
+    PrimitiveTopology,
+    StencilOp,
+    StencilTestEnable,
+
     Last
 };
 static_assert(Last <= std::numeric_limits<u8>::max());
@@ -64,6 +74,46 @@ public:
         return Exchange(Dirty::StencilProperties, false);
     }
 
+    bool TouchCullMode() {
+        return Exchange(Dirty::CullMode, false);
+    }
+
+    bool TouchDepthBoundsTestEnable() {
+        return Exchange(Dirty::DepthBoundsEnable, false);
+    }
+
+    bool TouchDepthTestEnable() {
+        return Exchange(Dirty::DepthTestEnable, false);
+    }
+
+    bool TouchDepthBoundsEnable() {
+        return Exchange(Dirty::DepthBoundsEnable, false);
+    }
+
+    bool TouchDepthWriteEnable() {
+        return Exchange(Dirty::DepthWriteEnable, false);
+    }
+
+    bool TouchDepthCompareOp() {
+        return Exchange(Dirty::DepthCompareOp, false);
+    }
+
+    bool TouchFrontFace() {
+        return Exchange(Dirty::FrontFace, false);
+    }
+
+    bool TouchPrimitiveTopology() {
+        return Exchange(Dirty::PrimitiveTopology, false);
+    }
+
+    bool TouchStencilOp() {
+        return Exchange(Dirty::StencilOp, false);
+    }
+
+    bool TouchStencilTestEnable() {
+        return Exchange(Dirty::StencilTestEnable, false);
+    }
+
 private:
     bool Exchange(std::size_t id, bool new_value) const noexcept {
         auto& flags = system.GPU().Maxwell3D().dirty.flags;

src/video_core/renderer_vulkan/vk_stream_buffer.cpp

@@ -121,7 +121,7 @@ void VKStreamBuffer::CreateBuffers(VkBufferUsageFlags usage) {
 
     // Substract from the preferred heap size some bytes to avoid getting out of memory.
     const VkDeviceSize heap_size = memory_properties.memoryHeaps[preferred_heap].size;
-    const VkDeviceSize allocable_size = heap_size - 4 * 1024 * 1024;
+    const VkDeviceSize allocable_size = heap_size - 9 * 1024 * 1024;
 
     VkBufferCreateInfo buffer_ci;
     buffer_ci.sType = VK_STRUCTURE_TYPE_BUFFER_CREATE_INFO;

src/video_core/renderer_vulkan/wrapper.cpp

@@ -88,6 +88,16 @@ void Load(VkDevice device, DeviceDispatch& dld) noexcept {
     X(vkCmdSetStencilWriteMask);
     X(vkCmdSetViewport);
     X(vkCmdWaitEvents);
+    X(vkCmdBindVertexBuffers2EXT);
+    X(vkCmdSetCullModeEXT);
+    X(vkCmdSetDepthBoundsTestEnableEXT);
+    X(vkCmdSetDepthCompareOpEXT);
+    X(vkCmdSetDepthTestEnableEXT);
+    X(vkCmdSetDepthWriteEnableEXT);
+    X(vkCmdSetFrontFaceEXT);
+    X(vkCmdSetPrimitiveTopologyEXT);
+    X(vkCmdSetStencilOpEXT);
+    X(vkCmdSetStencilTestEnableEXT);
     X(vkCreateBuffer);
     X(vkCreateBufferView);
     X(vkCreateCommandPool);

src/video_core/renderer_vulkan/wrapper.h

@@ -207,6 +207,16 @@ struct DeviceDispatch : public InstanceDispatch {
     PFN_vkCmdSetStencilWriteMask vkCmdSetStencilWriteMask;
     PFN_vkCmdSetViewport vkCmdSetViewport;
     PFN_vkCmdWaitEvents vkCmdWaitEvents;
+    PFN_vkCmdBindVertexBuffers2EXT vkCmdBindVertexBuffers2EXT;
+    PFN_vkCmdSetCullModeEXT vkCmdSetCullModeEXT;
+    PFN_vkCmdSetDepthBoundsTestEnableEXT vkCmdSetDepthBoundsTestEnableEXT;
+    PFN_vkCmdSetDepthCompareOpEXT vkCmdSetDepthCompareOpEXT;
+    PFN_vkCmdSetDepthTestEnableEXT vkCmdSetDepthTestEnableEXT;
+    PFN_vkCmdSetDepthWriteEnableEXT vkCmdSetDepthWriteEnableEXT;
+    PFN_vkCmdSetFrontFaceEXT vkCmdSetFrontFaceEXT;
+    PFN_vkCmdSetPrimitiveTopologyEXT vkCmdSetPrimitiveTopologyEXT;
+    PFN_vkCmdSetStencilOpEXT vkCmdSetStencilOpEXT;
+    PFN_vkCmdSetStencilTestEnableEXT vkCmdSetStencilTestEnableEXT;
     PFN_vkCreateBuffer vkCreateBuffer;
     PFN_vkCreateBufferView vkCreateBufferView;
     PFN_vkCreateCommandPool vkCreateCommandPool;
@@ -969,6 +979,50 @@ public:
                              buffer_barriers.data(), image_barriers.size(), image_barriers.data());
     }
 
+    void BindVertexBuffers2EXT(u32 first_binding, u32 binding_count, const VkBuffer* buffers,
+                               const VkDeviceSize* offsets, const VkDeviceSize* sizes,
+                               const VkDeviceSize* strides) const noexcept {
+        dld->vkCmdBindVertexBuffers2EXT(handle, first_binding, binding_count, buffers, offsets,
+                                        sizes, strides);
+    }
+
+    void SetCullModeEXT(VkCullModeFlags cull_mode) const noexcept {
+        dld->vkCmdSetCullModeEXT(handle, cull_mode);
+    }
+
+    void SetDepthBoundsTestEnableEXT(bool enable) const noexcept {
+        dld->vkCmdSetDepthBoundsTestEnableEXT(handle, enable ? VK_TRUE : VK_FALSE);
+    }
+
+    void SetDepthCompareOpEXT(VkCompareOp compare_op) const noexcept {
+        dld->vkCmdSetDepthCompareOpEXT(handle, compare_op);
+    }
+
+    void SetDepthTestEnableEXT(bool enable) const noexcept {
+        dld->vkCmdSetDepthTestEnableEXT(handle, enable ? VK_TRUE : VK_FALSE);
+    }
+
+    void SetDepthWriteEnableEXT(bool enable) const noexcept {
+        dld->vkCmdSetDepthWriteEnableEXT(handle, enable ? VK_TRUE : VK_FALSE);
+    }
+
+    void SetFrontFaceEXT(VkFrontFace front_face) const noexcept {
+        dld->vkCmdSetFrontFaceEXT(handle, front_face);
+    }
+
+    void SetPrimitiveTopologyEXT(VkPrimitiveTopology primitive_topology) const noexcept {
+        dld->vkCmdSetPrimitiveTopologyEXT(handle, primitive_topology);
+    }
+
+    void SetStencilOpEXT(VkStencilFaceFlags face_mask, VkStencilOp fail_op, VkStencilOp pass_op,
+                         VkStencilOp depth_fail_op, VkCompareOp compare_op) const noexcept {
+        dld->vkCmdSetStencilOpEXT(handle, face_mask, fail_op, pass_op, depth_fail_op, compare_op);
+    }
+
+    void SetStencilTestEnableEXT(bool enable) const noexcept {
+        dld->vkCmdSetStencilTestEnableEXT(handle, enable ? VK_TRUE : VK_FALSE);
+    }
+
     void BindTransformFeedbackBuffersEXT(u32 first, u32 count, const VkBuffer* buffers,
                                          const VkDeviceSize* offsets,
                                          const VkDeviceSize* sizes) const noexcept {

src/video_core/texture_cache/surface_params.cpp

@@ -343,8 +343,7 @@ std::size_t SurfaceParams::GetLayerSize(bool as_host_size, bool uncompressed) co
         size += GetInnerMipmapMemorySize(level, as_host_size, uncompressed);
     }
     if (is_tiled && is_layered) {
-        return Common::AlignBits(size,
-                                 Tegra::Texture::GetGOBSizeShift() + block_height + block_depth);
+        return Common::AlignBits(size, Tegra::Texture::GOB_SIZE_SHIFT + block_height + block_depth);
     }
     return size;
 }
@@ -418,7 +417,7 @@ std::tuple<u32, u32, u32> SurfaceParams::GetBlockOffsetXYZ(u32 offset) const {
     const u32 block_size = GetBlockSize();
     const u32 block_index = offset / block_size;
     const u32 gob_offset = offset % block_size;
-    const u32 gob_index = gob_offset / static_cast<u32>(Tegra::Texture::GetGOBSize());
+    const u32 gob_index = gob_offset / static_cast<u32>(Tegra::Texture::GOB_SIZE);
     const u32 x_gob_pixels = 64U / GetBytesPerPixel();
     const u32 x_block_pixels = x_gob_pixels << block_width;
     const u32 y_block_pixels = 8U << block_height;

src/video_core/texture_cache/surface_params.h

@@ -204,7 +204,7 @@ public:
     static std::size_t AlignLayered(const std::size_t out_size, const u32 block_height,
                                     const u32 block_depth) {
         return Common::AlignBits(out_size,
-                                 Tegra::Texture::GetGOBSizeShift() + block_height + block_depth);
+                                 Tegra::Texture::GOB_SIZE_SHIFT + block_height + block_depth);
     }
 
     /// Converts a width from a type of surface into another. This helps represent the

src/video_core/texture_cache/texture_cache.h

@@ -249,7 +249,7 @@ public:
             auto& surface = render_targets[index].target;
             surface->MarkAsRenderTarget(false, NO_RT);
             const auto& cr_params = surface->GetSurfaceParams();
-            if (!cr_params.is_tiled && Settings::values.use_asynchronous_gpu_emulation) {
+            if (!cr_params.is_tiled && Settings::values.use_asynchronous_gpu_emulation.GetValue()) {
                 AsyncFlushSurface(surface);
             }
         }

src/video_core/textures/decoders.cpp

@@ -6,6 +6,7 @@
 #include <cstring>
 #include "common/alignment.h"
 #include "common/assert.h"
+#include "common/bit_util.h"
 #include "video_core/gpu.h"
 #include "video_core/textures/decoders.h"
 #include "video_core/textures/texture.h"
@@ -37,20 +38,10 @@ struct alignas(64) SwizzleTable {
     std::array<std::array<u16, M>, N> values{};
 };
 
-constexpr u32 gob_size_x_shift = 6;
-constexpr u32 gob_size_y_shift = 3;
-constexpr u32 gob_size_z_shift = 0;
-constexpr u32 gob_size_shift = gob_size_x_shift + gob_size_y_shift + gob_size_z_shift;
+constexpr u32 FAST_SWIZZLE_ALIGN = 16;
 
-constexpr u32 gob_size_x = 1U << gob_size_x_shift;
-constexpr u32 gob_size_y = 1U << gob_size_y_shift;
-constexpr u32 gob_size_z = 1U << gob_size_z_shift;
-constexpr u32 gob_size = 1U << gob_size_shift;
-
-constexpr u32 fast_swizzle_align = 16;
-
-constexpr auto legacy_swizzle_table = SwizzleTable<gob_size_y, gob_size_x, gob_size_z>();
-constexpr auto fast_swizzle_table = SwizzleTable<gob_size_y, 4, fast_swizzle_align>();
+constexpr auto LEGACY_SWIZZLE_TABLE = SwizzleTable<GOB_SIZE_X, GOB_SIZE_X, GOB_SIZE_Z>();
+constexpr auto FAST_SWIZZLE_TABLE = SwizzleTable<GOB_SIZE_Y, 4, FAST_SWIZZLE_ALIGN>();
 
 /**
  * This function manages ALL the GOBs(Group of Bytes) Inside a single block.
@@ -69,17 +60,17 @@ void PreciseProcessBlock(u8* const swizzled_data, u8* const unswizzled_data, con
         u32 y_address = z_address;
         u32 pixel_base = layer_z * z + y_start * stride_x;
         for (u32 y = y_start; y < y_end; y++) {
-            const auto& table = legacy_swizzle_table[y % gob_size_y];
+            const auto& table = LEGACY_SWIZZLE_TABLE[y % GOB_SIZE_Y];
             for (u32 x = x_start; x < x_end; x++) {
-                const u32 swizzle_offset{y_address + table[x * bytes_per_pixel % gob_size_x]};
+                const u32 swizzle_offset{y_address + table[x * bytes_per_pixel % GOB_SIZE_X]};
                 const u32 pixel_index{x * out_bytes_per_pixel + pixel_base};
                 data_ptrs[unswizzle] = swizzled_data + swizzle_offset;
                 data_ptrs[!unswizzle] = unswizzled_data + pixel_index;
                 std::memcpy(data_ptrs[0], data_ptrs[1], bytes_per_pixel);
             }
             pixel_base += stride_x;
-            if ((y + 1) % gob_size_y == 0)
-                y_address += gob_size;
+            if ((y + 1) % GOB_SIZE_Y == 0)
+                y_address += GOB_SIZE;
         }
         z_address += xy_block_size;
     }
@@ -104,18 +95,18 @@ void FastProcessBlock(u8* const swizzled_data, u8* const unswizzled_data, const
         u32 y_address = z_address;
         u32 pixel_base = layer_z * z + y_start * stride_x;
         for (u32 y = y_start; y < y_end; y++) {
-            const auto& table = fast_swizzle_table[y % gob_size_y];
-            for (u32 xb = x_startb; xb < x_endb; xb += fast_swizzle_align) {
-                const u32 swizzle_offset{y_address + table[(xb / fast_swizzle_align) % 4]};
+            const auto& table = FAST_SWIZZLE_TABLE[y % GOB_SIZE_Y];
+            for (u32 xb = x_startb; xb < x_endb; xb += FAST_SWIZZLE_ALIGN) {
+                const u32 swizzle_offset{y_address + table[(xb / FAST_SWIZZLE_ALIGN) % 4]};
                 const u32 out_x = xb * out_bytes_per_pixel / bytes_per_pixel;
                 const u32 pixel_index{out_x + pixel_base};
                 data_ptrs[unswizzle ? 1 : 0] = swizzled_data + swizzle_offset;
                 data_ptrs[unswizzle ? 0 : 1] = unswizzled_data + pixel_index;
-                std::memcpy(data_ptrs[0], data_ptrs[1], fast_swizzle_align);
+                std::memcpy(data_ptrs[0], data_ptrs[1], FAST_SWIZZLE_ALIGN);
             }
             pixel_base += stride_x;
-            if ((y + 1) % gob_size_y == 0)
-                y_address += gob_size;
+            if ((y + 1) % GOB_SIZE_Y == 0)
+                y_address += GOB_SIZE;
         }
         z_address += xy_block_size;
     }
@@ -138,9 +129,9 @@ void SwizzledData(u8* const swizzled_data, u8* const unswizzled_data, const bool
     auto div_ceil = [](const u32 x, const u32 y) { return ((x + y - 1) / y); };
     const u32 stride_x = width * out_bytes_per_pixel;
     const u32 layer_z = height * stride_x;
-    const u32 gob_elements_x = gob_size_x / bytes_per_pixel;
-    constexpr u32 gob_elements_y = gob_size_y;
-    constexpr u32 gob_elements_z = gob_size_z;
+    const u32 gob_elements_x = GOB_SIZE_X / bytes_per_pixel;
+    constexpr u32 gob_elements_y = GOB_SIZE_Y;
+    constexpr u32 gob_elements_z = GOB_SIZE_Z;
     const u32 block_x_elements = gob_elements_x;
     const u32 block_y_elements = gob_elements_y * block_height;
     const u32 block_z_elements = gob_elements_z * block_depth;
@@ -148,7 +139,7 @@ void SwizzledData(u8* const swizzled_data, u8* const unswizzled_data, const bool
     const u32 blocks_on_x = div_ceil(aligned_width, block_x_elements);
     const u32 blocks_on_y = div_ceil(height, block_y_elements);
     const u32 blocks_on_z = div_ceil(depth, block_z_elements);
-    const u32 xy_block_size = gob_size * block_height;
+    const u32 xy_block_size = GOB_SIZE * block_height;
     const u32 block_size = xy_block_size * block_depth;
     u32 tile_offset = 0;
     for (u32 zb = 0; zb < blocks_on_z; zb++) {
@@ -182,7 +173,7 @@ void CopySwizzledData(u32 width, u32 height, u32 depth, u32 bytes_per_pixel,
                       bool unswizzle, u32 block_height, u32 block_depth, u32 width_spacing) {
     const u32 block_height_size{1U << block_height};
     const u32 block_depth_size{1U << block_depth};
-    if (bytes_per_pixel % 3 != 0 && (width * bytes_per_pixel) % fast_swizzle_align == 0) {
+    if (bytes_per_pixel % 3 != 0 && (width * bytes_per_pixel) % FAST_SWIZZLE_ALIGN == 0) {
         SwizzledData<true>(swizzled_data, unswizzled_data, unswizzle, width, height, depth,
                            bytes_per_pixel, out_bytes_per_pixel, block_height_size,
                            block_depth_size, width_spacing);
@@ -259,25 +250,26 @@ std::vector<u8> UnswizzleTexture(u8* address, u32 tile_size_x, u32 tile_size_y,
 }
 
 void SwizzleSubrect(u32 subrect_width, u32 subrect_height, u32 source_pitch, u32 swizzled_width,
-                    u32 bytes_per_pixel, u8* swizzled_data, u8* unswizzled_data,
+                    u32 bytes_per_pixel, u8* swizzled_data, const u8* unswizzled_data,
                     u32 block_height_bit, u32 offset_x, u32 offset_y) {
     const u32 block_height = 1U << block_height_bit;
-    const u32 image_width_in_gobs{(swizzled_width * bytes_per_pixel + (gob_size_x - 1)) /
-                                  gob_size_x};
+    const u32 image_width_in_gobs =
+        (swizzled_width * bytes_per_pixel + (GOB_SIZE_X - 1)) / GOB_SIZE_X;
     for (u32 line = 0; line < subrect_height; ++line) {
         const u32 dst_y = line + offset_y;
         const u32 gob_address_y =
-            (dst_y / (gob_size_y * block_height)) * gob_size * block_height * image_width_in_gobs +
-            ((dst_y % (gob_size_y * block_height)) / gob_size_y) * gob_size;
-        const auto& table = legacy_swizzle_table[dst_y % gob_size_y];
+            (dst_y / (GOB_SIZE_Y * block_height)) * GOB_SIZE * block_height * image_width_in_gobs +
+            ((dst_y % (GOB_SIZE_Y * block_height)) / GOB_SIZE_Y) * GOB_SIZE;
+        const auto& table = LEGACY_SWIZZLE_TABLE[dst_y % GOB_SIZE_Y];
         for (u32 x = 0; x < subrect_width; ++x) {
             const u32 dst_x = x + offset_x;
             const u32 gob_address =
-                gob_address_y + (dst_x * bytes_per_pixel / gob_size_x) * gob_size * block_height;
-            const u32 swizzled_offset = gob_address + table[(dst_x * bytes_per_pixel) % gob_size_x];
-            u8* source_line = unswizzled_data + line * source_pitch + x * bytes_per_pixel;
-            u8* dest_addr = swizzled_data + swizzled_offset;
+                gob_address_y + (dst_x * bytes_per_pixel / GOB_SIZE_X) * GOB_SIZE * block_height;
+            const u32 swizzled_offset = gob_address + table[(dst_x * bytes_per_pixel) % GOB_SIZE_X];
+            const u32 unswizzled_offset = line * source_pitch + x * bytes_per_pixel;
 
+            const u8* const source_line = unswizzled_data + unswizzled_offset;
+            u8* const dest_addr = swizzled_data + swizzled_offset;
             std::memcpy(dest_addr, source_line, bytes_per_pixel);
         }
     }
@@ -289,14 +281,15 @@ void UnswizzleSubrect(u32 subrect_width, u32 subrect_height, u32 dest_pitch, u32
     const u32 block_height = 1U << block_height_bit;
     for (u32 line = 0; line < subrect_height; ++line) {
         const u32 y2 = line + offset_y;
-        const u32 gob_address_y = (y2 / (gob_size_y * block_height)) * gob_size * block_height +
-                                  ((y2 % (gob_size_y * block_height)) / gob_size_y) * gob_size;
-        const auto& table = legacy_swizzle_table[y2 % gob_size_y];
+        const u32 gob_address_y = (y2 / (GOB_SIZE_Y * block_height)) * GOB_SIZE * block_height +
+                                  ((y2 % (GOB_SIZE_Y * block_height)) / GOB_SIZE_Y) * GOB_SIZE;
+        const auto& table = LEGACY_SWIZZLE_TABLE[y2 % GOB_SIZE_Y];
         for (u32 x = 0; x < subrect_width; ++x) {
             const u32 x2 = (x + offset_x) * bytes_per_pixel;
-            const u32 gob_address = gob_address_y + (x2 / gob_size_x) * gob_size * block_height;
-            const u32 swizzled_offset = gob_address + table[x2 % gob_size_x];
-            u8* dest_line = unswizzled_data + line * dest_pitch + x * bytes_per_pixel;
+            const u32 gob_address = gob_address_y + (x2 / GOB_SIZE_X) * GOB_SIZE * block_height;
+            const u32 swizzled_offset = gob_address + table[x2 % GOB_SIZE_X];
+            const u32 unswizzled_offset = line * dest_pitch + x * bytes_per_pixel;
+            u8* dest_line = unswizzled_data + unswizzled_offset;
             u8* source_addr = swizzled_data + swizzled_offset;
 
             std::memcpy(dest_line, source_addr, bytes_per_pixel);
@@ -304,21 +297,48 @@ void UnswizzleSubrect(u32 subrect_width, u32 subrect_height, u32 dest_pitch, u32
     }
 }
 
+void SwizzleSliceToVoxel(u32 line_length_in, u32 line_count, u32 pitch, u32 width, u32 height,
+                         u32 bytes_per_pixel, u32 block_height, u32 block_depth, u32 origin_x,
+                         u32 origin_y, u8* output, const u8* input) {
+    UNIMPLEMENTED_IF(origin_x > 0);
+    UNIMPLEMENTED_IF(origin_y > 0);
+
+    const u32 stride = width * bytes_per_pixel;
+    const u32 gobs_in_x = (stride + GOB_SIZE_X - 1) / GOB_SIZE_X;
+    const u32 block_size = gobs_in_x << (GOB_SIZE_SHIFT + block_height + block_depth);
+
+    const u32 block_height_mask = (1U << block_height) - 1;
+    const u32 x_shift = Common::CountTrailingZeroes32(GOB_SIZE << (block_height + block_depth));
+
+    for (u32 line = 0; line < line_count; ++line) {
+        const auto& table = LEGACY_SWIZZLE_TABLE[line % GOB_SIZE_Y];
+        const u32 block_y = line / GOB_SIZE_Y;
+        const u32 dst_offset_y =
+            (block_y >> block_height) * block_size + (block_y & block_height_mask) * GOB_SIZE;
+        for (u32 x = 0; x < line_length_in; ++x) {
+            const u32 dst_offset =
+                ((x / GOB_SIZE_X) << x_shift) + dst_offset_y + table[x % GOB_SIZE_X];
+            const u32 src_offset = x * bytes_per_pixel + line * pitch;
+            std::memcpy(output + dst_offset, input + src_offset, bytes_per_pixel);
+        }
+    }
+}
+
 void SwizzleKepler(const u32 width, const u32 height, const u32 dst_x, const u32 dst_y,
                    const u32 block_height_bit, const std::size_t copy_size, const u8* source_data,
                    u8* swizzle_data) {
     const u32 block_height = 1U << block_height_bit;
-    const u32 image_width_in_gobs{(width + gob_size_x - 1) / gob_size_x};
+    const u32 image_width_in_gobs{(width + GOB_SIZE_X - 1) / GOB_SIZE_X};
     std::size_t count = 0;
     for (std::size_t y = dst_y; y < height && count < copy_size; ++y) {
         const std::size_t gob_address_y =
-            (y / (gob_size_y * block_height)) * gob_size * block_height * image_width_in_gobs +
-            ((y % (gob_size_y * block_height)) / gob_size_y) * gob_size;
-        const auto& table = legacy_swizzle_table[y % gob_size_y];
+            (y / (GOB_SIZE_Y * block_height)) * GOB_SIZE * block_height * image_width_in_gobs +
+            ((y % (GOB_SIZE_Y * block_height)) / GOB_SIZE_Y) * GOB_SIZE;
+        const auto& table = LEGACY_SWIZZLE_TABLE[y % GOB_SIZE_Y];
         for (std::size_t x = dst_x; x < width && count < copy_size; ++x) {
             const std::size_t gob_address =
-                gob_address_y + (x / gob_size_x) * gob_size * block_height;
-            const std::size_t swizzled_offset = gob_address + table[x % gob_size_x];
+                gob_address_y + (x / GOB_SIZE_X) * GOB_SIZE * block_height;
+            const std::size_t swizzled_offset = gob_address + table[x % GOB_SIZE_X];
             const u8* source_line = source_data + count;
             u8* dest_addr = swizzle_data + swizzled_offset;
             count++;
@@ -373,9 +393,9 @@ std::vector<u8> DecodeTexture(const std::vector<u8>& texture_data, TextureFormat
 std::size_t CalculateSize(bool tiled, u32 bytes_per_pixel, u32 width, u32 height, u32 depth,
                           u32 block_height, u32 block_depth) {
     if (tiled) {
-        const u32 aligned_width = Common::AlignBits(width * bytes_per_pixel, gob_size_x_shift);
-        const u32 aligned_height = Common::AlignBits(height, gob_size_y_shift + block_height);
-        const u32 aligned_depth = Common::AlignBits(depth, gob_size_z_shift + block_depth);
+        const u32 aligned_width = Common::AlignBits(width * bytes_per_pixel, GOB_SIZE_X_SHIFT);
+        const u32 aligned_height = Common::AlignBits(height, GOB_SIZE_Y_SHIFT + block_height);
+        const u32 aligned_depth = Common::AlignBits(depth, GOB_SIZE_Z_SHIFT + block_depth);
         return aligned_width * aligned_height * aligned_depth;
     } else {
         return width * height * depth * bytes_per_pixel;
@@ -386,14 +406,14 @@ u64 GetGOBOffset(u32 width, u32 height, u32 dst_x, u32 dst_y, u32 block_height,
                  u32 bytes_per_pixel) {
     auto div_ceil = [](const u32 x, const u32 y) { return ((x + y - 1) / y); };
     const u32 gobs_in_block = 1 << block_height;
-    const u32 y_blocks = gob_size_y << block_height;
-    const u32 x_per_gob = gob_size_x / bytes_per_pixel;
+    const u32 y_blocks = GOB_SIZE_Y << block_height;
+    const u32 x_per_gob = GOB_SIZE_X / bytes_per_pixel;
     const u32 x_blocks = div_ceil(width, x_per_gob);
-    const u32 block_size = gob_size * gobs_in_block;
+    const u32 block_size = GOB_SIZE * gobs_in_block;
     const u32 stride = block_size * x_blocks;
     const u32 base = (dst_y / y_blocks) * stride + (dst_x / x_per_gob) * block_size;
     const u32 relative_y = dst_y % y_blocks;
-    return base + (relative_y / gob_size_y) * gob_size;
+    return base + (relative_y / GOB_SIZE_Y) * GOB_SIZE;
 }
 
 } // namespace Tegra::Texture

src/video_core/textures/decoders.h

@@ -10,15 +10,15 @@
 
 namespace Tegra::Texture {
 
-// GOBSize constant. Calculated by 64 bytes in x multiplied by 8 y coords, represents
-// an small rect of (64/bytes_per_pixel)X8.
-inline std::size_t GetGOBSize() {
-    return 512;
-}
+constexpr u32 GOB_SIZE_X = 64;
+constexpr u32 GOB_SIZE_Y = 8;
+constexpr u32 GOB_SIZE_Z = 1;
+constexpr u32 GOB_SIZE = GOB_SIZE_X * GOB_SIZE_Y * GOB_SIZE_Z;
 
-inline std::size_t GetGOBSizeShift() {
-    return 9;
-}
+constexpr std::size_t GOB_SIZE_X_SHIFT = 6;
+constexpr std::size_t GOB_SIZE_Y_SHIFT = 3;
+constexpr std::size_t GOB_SIZE_Z_SHIFT = 0;
+constexpr std::size_t GOB_SIZE_SHIFT = GOB_SIZE_X_SHIFT + GOB_SIZE_Y_SHIFT + GOB_SIZE_Z_SHIFT;
 
 /// Unswizzles a swizzled texture without changing its format.
 void UnswizzleTexture(u8* unswizzled_data, u8* address, u32 tile_size_x, u32 tile_size_y,
@@ -48,14 +48,32 @@ std::size_t CalculateSize(bool tiled, u32 bytes_per_pixel, u32 width, u32 height
 
 /// Copies an untiled subrectangle into a tiled surface.
 void SwizzleSubrect(u32 subrect_width, u32 subrect_height, u32 source_pitch, u32 swizzled_width,
-                    u32 bytes_per_pixel, u8* swizzled_data, u8* unswizzled_data, u32 block_height,
-                    u32 offset_x, u32 offset_y);
+                    u32 bytes_per_pixel, u8* swizzled_data, const u8* unswizzled_data,
+                    u32 block_height_bit, u32 offset_x, u32 offset_y);
 
 /// Copies a tiled subrectangle into a linear surface.
 void UnswizzleSubrect(u32 subrect_width, u32 subrect_height, u32 dest_pitch, u32 swizzled_width,
                       u32 bytes_per_pixel, u8* swizzled_data, u8* unswizzled_data, u32 block_height,
                       u32 offset_x, u32 offset_y);
 
+/// @brief Swizzles a 2D array of pixels into a 3D texture
+/// @param line_length_in  Number of pixels per line
+/// @param line_count      Number of lines
+/// @param pitch           Number of bytes per line
+/// @param width           Width of the swizzled texture
+/// @param height          Height of the swizzled texture
+/// @param bytes_per_pixel Number of bytes used per pixel
+/// @param block_height    Block height shift
+/// @param block_depth     Block depth shift
+/// @param origin_x        Column offset in pixels of the swizzled texture
+/// @param origin_y        Row offset in pixels of the swizzled texture
+/// @param output          Pointer to the pixels of the swizzled texture
+/// @param input           Pointer to the 2D array of pixels used as input
+/// @pre input and output points to an array large enough to hold the number of bytes used
+void SwizzleSliceToVoxel(u32 line_length_in, u32 line_count, u32 pitch, u32 width, u32 height,
+                         u32 bytes_per_pixel, u32 block_height, u32 block_depth, u32 origin_x,
+                         u32 origin_y, u8* output, const u8* input);
+
 void SwizzleKepler(u32 width, u32 height, u32 dst_x, u32 dst_y, u32 block_height,
                    std::size_t copy_size, const u8* source_data, u8* swizzle_data);
 

src/video_core/textures/texture.cpp

@@ -48,7 +48,7 @@ constexpr std::array<float, 256> SRGB_CONVERSION_LUT = {
 };
 
 unsigned SettingsMinimumAnisotropy() noexcept {
-    switch (static_cast<Anisotropy>(Settings::values.max_anisotropy)) {
+    switch (static_cast<Anisotropy>(Settings::values.max_anisotropy.GetValue())) {
     default:
     case Anisotropy::Default:
         return 1U;

src/video_core/video_core.cpp

@@ -19,7 +19,7 @@ namespace {
 std::unique_ptr<VideoCore::RendererBase> CreateRenderer(Core::Frontend::EmuWindow& emu_window,
                                                         Core::System& system,
                                                         Core::Frontend::GraphicsContext& context) {
-    switch (Settings::values.renderer_backend) {
+    switch (Settings::values.renderer_backend.GetValue()) {
     case Settings::RendererBackend::OpenGL:
         return std::make_unique<OpenGL::RendererOpenGL>(emu_window, system, context);
 #ifdef HAS_VULKAN
@@ -42,7 +42,7 @@ std::unique_ptr<Tegra::GPU> CreateGPU(Core::Frontend::EmuWindow& emu_window, Cor
         return nullptr;
     }
 
-    if (Settings::values.use_asynchronous_gpu_emulation) {
+    if (Settings::values.use_asynchronous_gpu_emulation.GetValue()) {
         return std::make_unique<VideoCommon::GPUAsynch>(system, std::move(renderer),
                                                         std::move(context));
     }
@@ -51,8 +51,8 @@ std::unique_ptr<Tegra::GPU> CreateGPU(Core::Frontend::EmuWindow& emu_window, Cor
 
 u16 GetResolutionScaleFactor(const RendererBase& renderer) {
     return static_cast<u16>(
-        Settings::values.resolution_factor != 0
-            ? Settings::values.resolution_factor
+        Settings::values.resolution_factor.GetValue() != 0
+            ? Settings::values.resolution_factor.GetValue()
             : renderer.GetRenderWindow().GetFramebufferLayout().GetScalingRatio());
 }
 

src/yuzu/CMakeLists.txt

@@ -24,10 +24,18 @@ add_executable(yuzu
     compatibility_list.h
     configuration/config.cpp
     configuration/config.h
+    configuration/configuration_shared.cpp
+    configuration/configuration_shared.h
     configuration/configure.ui
     configuration/configure_audio.cpp
     configuration/configure_audio.h
     configuration/configure_audio.ui
+    configuration/configure_cpu.cpp
+    configuration/configure_cpu.h
+    configuration/configure_cpu.ui
+    configuration/configure_cpu_debug.cpp
+    configuration/configure_cpu_debug.h
+    configuration/configure_cpu_debug.ui
     configuration/configure_debug.cpp
     configuration/configure_debug.h
     configuration/configure_debug.ui
@@ -60,9 +68,12 @@ add_executable(yuzu
     configuration/configure_mouse_advanced.cpp
     configuration/configure_mouse_advanced.h
     configuration/configure_mouse_advanced.ui
-    configuration/configure_per_general.cpp
-    configuration/configure_per_general.h
-    configuration/configure_per_general.ui
+    configuration/configure_per_game.cpp
+    configuration/configure_per_game.h
+    configuration/configure_per_game.ui
+    configuration/configure_per_game_addons.cpp
+    configuration/configure_per_game_addons.h
+    configuration/configure_per_game_addons.ui
     configuration/configure_profile_manager.cpp
     configuration/configure_profile_manager.h
     configuration/configure_profile_manager.ui
@@ -93,11 +104,13 @@ add_executable(yuzu
     game_list_p.h
     game_list_worker.cpp
     game_list_worker.h
+    hotkeys.cpp
+    hotkeys.h
+    install_dialog.cpp
+    install_dialog.h
     loading_screen.cpp
     loading_screen.h
     loading_screen.ui
-    hotkeys.cpp
-    hotkeys.h
     main.cpp
     main.h
     main.ui
@@ -147,7 +160,7 @@ endif()
 create_target_directory_groups(yuzu)
 
 target_link_libraries(yuzu PRIVATE common core input_common video_core)
-target_link_libraries(yuzu PRIVATE Boost::boost glad Qt5::OpenGL Qt5::Widgets)
+target_link_libraries(yuzu PRIVATE Boost::boost glad Qt5::Widgets)
 target_link_libraries(yuzu PRIVATE ${PLATFORM_LIBRARIES} Threads::Threads)
 
 if (ENABLE_VULKAN AND NOT WIN32)

src/yuzu/bootmanager.cpp

@@ -145,7 +145,7 @@ public:
 
         // disable vsync for any shared contexts
         auto format = share_context->format();
-        format.setSwapInterval(main_surface ? Settings::values.use_vsync : 0);
+        format.setSwapInterval(main_surface ? Settings::values.use_vsync.GetValue() : 0);
 
         context = std::make_unique<QOpenGLContext>();
         context->setShareContext(share_context);
@@ -495,7 +495,7 @@ void GRenderWindow::resizeEvent(QResizeEvent* event) {
 
 std::unique_ptr<Core::Frontend::GraphicsContext> GRenderWindow::CreateSharedContext() const {
 #ifdef HAS_OPENGL
-    if (Settings::values.renderer_backend == Settings::RendererBackend::OpenGL) {
+    if (Settings::values.renderer_backend.GetValue() == Settings::RendererBackend::OpenGL) {
         auto c = static_cast<OpenGLSharedContext*>(main_context.get());
         // Bind the shared contexts to the main surface in case the backend wants to take over
         // presentation
@@ -511,7 +511,7 @@ bool GRenderWindow::InitRenderTarget() {
 
     first_frame = false;
 
-    switch (Settings::values.renderer_backend) {
+    switch (Settings::values.renderer_backend.GetValue()) {
     case Settings::RendererBackend::OpenGL:
         if (!InitializeOpenGL()) {
             return false;
@@ -538,7 +538,7 @@ bool GRenderWindow::InitRenderTarget() {
     OnFramebufferSizeChanged();
     BackupGeometry();
 
-    if (Settings::values.renderer_backend == Settings::RendererBackend::OpenGL) {
+    if (Settings::values.renderer_backend.GetValue() == Settings::RendererBackend::OpenGL) {
         if (!LoadOpenGL()) {
             return false;
         }

src/yuzu/configuration/config.cpp

@@ -13,17 +13,20 @@
 #include "input_common/udp/client.h"
 #include "yuzu/configuration/config.h"
 
-Config::Config() {
+Config::Config(const std::string& config_file, bool is_global) {
     // TODO: Don't hardcode the path; let the frontend decide where to put the config files.
-    qt_config_loc = FileUtil::GetUserPath(FileUtil::UserPath::ConfigDir) + "qt-config.ini";
+    qt_config_loc = FileUtil::GetUserPath(FileUtil::UserPath::ConfigDir) + config_file;
     FileUtil::CreateFullPath(qt_config_loc);
     qt_config =
         std::make_unique<QSettings>(QString::fromStdString(qt_config_loc), QSettings::IniFormat);
+    global = is_global;
     Reload();
 }
 
 Config::~Config() {
-    Save();
+    if (global) {
+        Save();
+    }
 }
 
 const std::array<int, Settings::NativeButton::NumButtons> Config::default_buttons = {
@@ -212,7 +215,7 @@ const std::array<int, Settings::NativeKeyboard::NumKeyboardMods> Config::default
 // UISetting::values.shortcuts, which is alphabetically ordered.
 // clang-format off
 const std::array<UISettings::Shortcut, 16> Config::default_hotkeys{{
-    {QStringLiteral("Capture Screenshot"),       QStringLiteral("Main Window"), {QStringLiteral("Ctrl+P"), Qt::ApplicationShortcut}},
+    {QStringLiteral("Capture Screenshot"),       QStringLiteral("Main Window"), {QStringLiteral("Ctrl+P"), Qt::WidgetWithChildrenShortcut}},
     {QStringLiteral("Change Docked Mode"),       QStringLiteral("Main Window"), {QStringLiteral("F10"), Qt::ApplicationShortcut}},
     {QStringLiteral("Continue/Pause Emulation"), QStringLiteral("Main Window"), {QStringLiteral("F4"), Qt::WindowShortcut}},
     {QStringLiteral("Decrease Speed Limit"),     QStringLiteral("Main Window"), {QStringLiteral("-"), Qt::ApplicationShortcut}},
@@ -220,8 +223,8 @@ const std::array<UISettings::Shortcut, 16> Config::default_hotkeys{{
     {QStringLiteral("Exit yuzu"),                QStringLiteral("Main Window"), {QStringLiteral("Ctrl+Q"), Qt::WindowShortcut}},
     {QStringLiteral("Fullscreen"),               QStringLiteral("Main Window"), {QStringLiteral("F11"), Qt::WindowShortcut}},
     {QStringLiteral("Increase Speed Limit"),     QStringLiteral("Main Window"), {QStringLiteral("+"), Qt::ApplicationShortcut}},
-    {QStringLiteral("Load Amiibo"),              QStringLiteral("Main Window"), {QStringLiteral("F2"), Qt::ApplicationShortcut}},
-    {QStringLiteral("Load File"),                QStringLiteral("Main Window"), {QStringLiteral("Ctrl+O"), Qt::WindowShortcut}},
+    {QStringLiteral("Load Amiibo"),              QStringLiteral("Main Window"), {QStringLiteral("F2"), Qt::WidgetWithChildrenShortcut}},
+    {QStringLiteral("Load File"),                QStringLiteral("Main Window"), {QStringLiteral("Ctrl+O"), Qt::WidgetWithChildrenShortcut}},
     {QStringLiteral("Mute Audio"),               QStringLiteral("Main Window"), {QStringLiteral("Ctrl+M"), Qt::WindowShortcut}},
     {QStringLiteral("Restart Emulation"),        QStringLiteral("Main Window"), {QStringLiteral("F6"), Qt::WindowShortcut}},
     {QStringLiteral("Stop Emulation"),           QStringLiteral("Main Window"), {QStringLiteral("F5"), Qt::WindowShortcut}},
@@ -402,16 +405,19 @@ void Config::ApplyDefaultProfileIfInputInvalid() {
 void Config::ReadAudioValues() {
     qt_config->beginGroup(QStringLiteral("Audio"));
 
-    Settings::values.sink_id = ReadSetting(QStringLiteral("output_engine"), QStringLiteral("auto"))
-                                   .toString()
-                                   .toStdString();
-    Settings::values.enable_audio_stretching =
-        ReadSetting(QStringLiteral("enable_audio_stretching"), true).toBool();
-    Settings::values.audio_device_id =
-        ReadSetting(QStringLiteral("output_device"), QStringLiteral("auto"))
-            .toString()
-            .toStdString();
-    Settings::values.volume = ReadSetting(QStringLiteral("volume"), 1).toFloat();
+    if (global) {
+        Settings::values.sink_id =
+            ReadSetting(QStringLiteral("output_engine"), QStringLiteral("auto"))
+                .toString()
+                .toStdString();
+        Settings::values.audio_device_id =
+            ReadSetting(QStringLiteral("output_device"), QStringLiteral("auto"))
+                .toString()
+                .toStdString();
+    }
+    ReadSettingGlobal(Settings::values.enable_audio_stretching,
+                      QStringLiteral("enable_audio_stretching"), true);
+    ReadSettingGlobal(Settings::values.volume, QStringLiteral("volume"), 1);
 
     qt_config->endGroup();
 }
@@ -440,6 +446,8 @@ void Config::ReadControlValues() {
             .toInt());
     Settings::values.udp_pad_index =
         static_cast<u8>(ReadSetting(QStringLiteral("udp_pad_index"), 0).toUInt());
+    Settings::values.use_docked_mode =
+        ReadSetting(QStringLiteral("use_docked_mode"), false).toBool();
 
     qt_config->endGroup();
 }
@@ -447,7 +455,7 @@ void Config::ReadControlValues() {
 void Config::ReadCoreValues() {
     qt_config->beginGroup(QStringLiteral("Core"));
 
-    Settings::values.use_multi_core = ReadSetting(QStringLiteral("use_multi_core"), false).toBool();
+    ReadSettingGlobal(Settings::values.use_multi_core, QStringLiteral("use_multi_core"), false);
 
     qt_config->endGroup();
 }
@@ -497,22 +505,6 @@ void Config::ReadDataStorageValues() {
         ReadSetting(QStringLiteral("gamecard_current_game"), false).toBool();
     Settings::values.gamecard_path =
         ReadSetting(QStringLiteral("gamecard_path"), QStringLiteral("")).toString().toStdString();
-    Settings::values.nand_total_size = static_cast<Settings::NANDTotalSize>(
-        ReadSetting(QStringLiteral("nand_total_size"),
-                    QVariant::fromValue<u64>(static_cast<u64>(Settings::NANDTotalSize::S29_1GB)))
-            .toULongLong());
-    Settings::values.nand_user_size = static_cast<Settings::NANDUserSize>(
-        ReadSetting(QStringLiteral("nand_user_size"),
-                    QVariant::fromValue<u64>(static_cast<u64>(Settings::NANDUserSize::S26GB)))
-            .toULongLong());
-    Settings::values.nand_system_size = static_cast<Settings::NANDSystemSize>(
-        ReadSetting(QStringLiteral("nand_system_size"),
-                    QVariant::fromValue<u64>(static_cast<u64>(Settings::NANDSystemSize::S2_5GB)))
-            .toULongLong());
-    Settings::values.sdmc_size = static_cast<Settings::SDMCSize>(
-        ReadSetting(QStringLiteral("sdmc_size"),
-                    QVariant::fromValue<u64>(static_cast<u64>(Settings::SDMCSize::S16GB)))
-            .toULongLong());
 
     qt_config->endGroup();
 }
@@ -532,8 +524,6 @@ void Config::ReadDebuggingValues() {
     Settings::values.reporting_services =
         ReadSetting(QStringLiteral("reporting_services"), false).toBool();
     Settings::values.quest_flag = ReadSetting(QStringLiteral("quest_flag"), false).toBool();
-    Settings::values.disable_cpu_opt =
-        ReadSetting(QStringLiteral("disable_cpu_opt"), false).toBool();
     Settings::values.disable_macro_jit =
         ReadSetting(QStringLiteral("disable_macro_jit"), false).toBool();
 
@@ -625,35 +615,59 @@ void Config::ReadPathValues() {
     qt_config->endGroup();
 }
 
+void Config::ReadCpuValues() {
+    qt_config->beginGroup(QStringLiteral("Cpu"));
+
+    if (global) {
+        Settings::values.cpu_accuracy = static_cast<Settings::CPUAccuracy>(
+            ReadSetting(QStringLiteral("cpu_accuracy"), 0).toInt());
+
+        Settings::values.cpuopt_page_tables =
+            ReadSetting(QStringLiteral("cpuopt_page_tables"), true).toBool();
+        Settings::values.cpuopt_block_linking =
+            ReadSetting(QStringLiteral("cpuopt_block_linking"), true).toBool();
+        Settings::values.cpuopt_return_stack_buffer =
+            ReadSetting(QStringLiteral("cpuopt_return_stack_buffer"), true).toBool();
+        Settings::values.cpuopt_fast_dispatcher =
+            ReadSetting(QStringLiteral("cpuopt_fast_dispatcher"), true).toBool();
+        Settings::values.cpuopt_context_elimination =
+            ReadSetting(QStringLiteral("cpuopt_context_elimination"), true).toBool();
+        Settings::values.cpuopt_const_prop =
+            ReadSetting(QStringLiteral("cpuopt_const_prop"), true).toBool();
+        Settings::values.cpuopt_misc_ir =
+            ReadSetting(QStringLiteral("cpuopt_misc_ir"), true).toBool();
+        Settings::values.cpuopt_reduce_misalign_checks =
+            ReadSetting(QStringLiteral("cpuopt_reduce_misalign_checks"), true).toBool();
+    }
+
+    qt_config->endGroup();
+}
+
 void Config::ReadRendererValues() {
     qt_config->beginGroup(QStringLiteral("Renderer"));
 
-    Settings::values.renderer_backend =
-        static_cast<Settings::RendererBackend>(ReadSetting(QStringLiteral("backend"), 0).toInt());
-    Settings::values.renderer_debug = ReadSetting(QStringLiteral("debug"), false).toBool();
-    Settings::values.vulkan_device = ReadSetting(QStringLiteral("vulkan_device"), 0).toInt();
-    Settings::values.aspect_ratio = ReadSetting(QStringLiteral("aspect_ratio"), 0).toInt();
-    Settings::values.max_anisotropy = ReadSetting(QStringLiteral("max_anisotropy"), 0).toInt();
-    Settings::values.use_frame_limit =
-        ReadSetting(QStringLiteral("use_frame_limit"), true).toBool();
-    Settings::values.frame_limit = ReadSetting(QStringLiteral("frame_limit"), 100).toUInt();
-    Settings::values.use_disk_shader_cache =
-        ReadSetting(QStringLiteral("use_disk_shader_cache"), true).toBool();
-    const int gpu_accuracy_level = ReadSetting(QStringLiteral("gpu_accuracy"), 0).toInt();
-    Settings::values.gpu_accuracy = static_cast<Settings::GPUAccuracy>(gpu_accuracy_level);
-    Settings::values.use_asynchronous_gpu_emulation =
-        ReadSetting(QStringLiteral("use_asynchronous_gpu_emulation"), false).toBool();
-    Settings::values.use_vsync = ReadSetting(QStringLiteral("use_vsync"), true).toBool();
-    Settings::values.use_assembly_shaders =
-        ReadSetting(QStringLiteral("use_assembly_shaders"), false).toBool();
-    Settings::values.use_fast_gpu_time =
-        ReadSetting(QStringLiteral("use_fast_gpu_time"), true).toBool();
-    Settings::values.force_30fps_mode =
-        ReadSetting(QStringLiteral("force_30fps_mode"), false).toBool();
+    ReadSettingGlobal(Settings::values.renderer_backend, QStringLiteral("backend"), 0);
+    ReadSettingGlobal(Settings::values.renderer_debug, QStringLiteral("debug"), false);
+    ReadSettingGlobal(Settings::values.vulkan_device, QStringLiteral("vulkan_device"), 0);
+    ReadSettingGlobal(Settings::values.aspect_ratio, QStringLiteral("aspect_ratio"), 0);
+    ReadSettingGlobal(Settings::values.max_anisotropy, QStringLiteral("max_anisotropy"), 0);
+    ReadSettingGlobal(Settings::values.use_frame_limit, QStringLiteral("use_frame_limit"), true);
+    ReadSettingGlobal(Settings::values.frame_limit, QStringLiteral("frame_limit"), 100);
+    ReadSettingGlobal(Settings::values.use_disk_shader_cache,
+                      QStringLiteral("use_disk_shader_cache"), true);
+    ReadSettingGlobal(Settings::values.gpu_accuracy, QStringLiteral("gpu_accuracy"), 0);
+    ReadSettingGlobal(Settings::values.use_asynchronous_gpu_emulation,
+                      QStringLiteral("use_asynchronous_gpu_emulation"), false);
+    ReadSettingGlobal(Settings::values.use_vsync, QStringLiteral("use_vsync"), true);
+    ReadSettingGlobal(Settings::values.use_assembly_shaders, QStringLiteral("use_assembly_shaders"),
+                      false);
+    ReadSettingGlobal(Settings::values.use_fast_gpu_time, QStringLiteral("use_fast_gpu_time"),
+                      true);
+    ReadSettingGlobal(Settings::values.force_30fps_mode, QStringLiteral("force_30fps_mode"), false);
 
-    Settings::values.bg_red = ReadSetting(QStringLiteral("bg_red"), 0.0).toFloat();
-    Settings::values.bg_green = ReadSetting(QStringLiteral("bg_green"), 0.0).toFloat();
-    Settings::values.bg_blue = ReadSetting(QStringLiteral("bg_blue"), 0.0).toFloat();
+    ReadSettingGlobal(Settings::values.bg_red, QStringLiteral("bg_red"), 0.0);
+    ReadSettingGlobal(Settings::values.bg_green, QStringLiteral("bg_green"), 0.0);
+    ReadSettingGlobal(Settings::values.bg_blue, QStringLiteral("bg_blue"), 0.0);
 
     qt_config->endGroup();
 }
@@ -665,11 +679,13 @@ void Config::ReadShortcutValues() {
         const auto& [keyseq, context] = shortcut;
         qt_config->beginGroup(group);
         qt_config->beginGroup(name);
+        // No longer using ReadSetting for shortcut.second as it innacurately returns a value of 1
+        // for WidgetWithChildrenShortcut which is a value of 3. Needed to fix shortcuts the open
+        // a file dialog in windowed mode
         UISettings::values.shortcuts.push_back(
             {name,
              group,
-             {ReadSetting(QStringLiteral("KeySeq"), keyseq).toString(),
-              ReadSetting(QStringLiteral("Context"), context).toInt()}});
+             {ReadSetting(QStringLiteral("KeySeq"), keyseq).toString(), shortcut.second}});
         qt_config->endGroup();
         qt_config->endGroup();
     }
@@ -680,35 +696,45 @@ void Config::ReadShortcutValues() {
 void Config::ReadSystemValues() {
     qt_config->beginGroup(QStringLiteral("System"));
 
-    Settings::values.use_docked_mode =
-        ReadSetting(QStringLiteral("use_docked_mode"), false).toBool();
+    ReadSettingGlobal(Settings::values.current_user, QStringLiteral("current_user"), 0);
+    Settings::values.current_user =
+        std::clamp<int>(Settings::values.current_user, 0, Service::Account::MAX_USERS - 1);
 
-    Settings::values.current_user = std::clamp<int>(
-        ReadSetting(QStringLiteral("current_user"), 0).toInt(), 0, Service::Account::MAX_USERS - 1);
+    ReadSettingGlobal(Settings::values.language_index, QStringLiteral("language_index"), 1);
 
-    Settings::values.language_index = ReadSetting(QStringLiteral("language_index"), 1).toInt();
+    ReadSettingGlobal(Settings::values.region_index, QStringLiteral("region_index"), 1);
 
-    Settings::values.region_index = ReadSetting(QStringLiteral("region_index"), 1).toInt();
+    ReadSettingGlobal(Settings::values.time_zone_index, QStringLiteral("time_zone_index"), 0);
 
-    Settings::values.time_zone_index = ReadSetting(QStringLiteral("time_zone_index"), 0).toInt();
-
-    const auto rng_seed_enabled = ReadSetting(QStringLiteral("rng_seed_enabled"), false).toBool();
-    if (rng_seed_enabled) {
-        Settings::values.rng_seed = ReadSetting(QStringLiteral("rng_seed"), 0).toULongLong();
-    } else {
-        Settings::values.rng_seed = std::nullopt;
+    bool rng_seed_enabled;
+    ReadSettingGlobal(rng_seed_enabled, QStringLiteral("rng_seed_enabled"), false);
+    bool rng_seed_global =
+        global || qt_config->value(QStringLiteral("rng_seed/use_global"), true).toBool();
+    Settings::values.rng_seed.SetGlobal(rng_seed_global);
+    if (global || !rng_seed_global) {
+        if (rng_seed_enabled) {
+            Settings::values.rng_seed.SetValue(
+                ReadSetting(QStringLiteral("rng_seed"), 0).toULongLong());
+        } else {
+            Settings::values.rng_seed.SetValue(std::nullopt);
+        }
     }
 
-    const auto custom_rtc_enabled =
-        ReadSetting(QStringLiteral("custom_rtc_enabled"), false).toBool();
-    if (custom_rtc_enabled) {
-        Settings::values.custom_rtc =
-            std::chrono::seconds(ReadSetting(QStringLiteral("custom_rtc"), 0).toULongLong());
-    } else {
-        Settings::values.custom_rtc = std::nullopt;
+    bool custom_rtc_enabled;
+    ReadSettingGlobal(custom_rtc_enabled, QStringLiteral("custom_rtc_enabled"), false);
+    bool custom_rtc_global =
+        global || qt_config->value(QStringLiteral("custom_rtc/use_global"), true).toBool();
+    Settings::values.custom_rtc.SetGlobal(custom_rtc_global);
+    if (global || !custom_rtc_global) {
+        if (custom_rtc_enabled) {
+            Settings::values.custom_rtc.SetValue(
+                std::chrono::seconds(ReadSetting(QStringLiteral("custom_rtc"), 0).toULongLong()));
+        } else {
+            Settings::values.custom_rtc.SetValue(std::nullopt);
+        }
     }
 
-    Settings::values.sound_index = ReadSetting(QStringLiteral("sound_index"), 1).toInt();
+    ReadSettingGlobal(Settings::values.sound_index, QStringLiteral("sound_index"), 1);
 
     qt_config->endGroup();
 }
@@ -802,18 +828,21 @@ void Config::ReadWebServiceValues() {
 }
 
 void Config::ReadValues() {
-    ReadControlValues();
+    if (global) {
+        ReadControlValues();
+        ReadDataStorageValues();
+        ReadDebuggingValues();
+        ReadDisabledAddOnValues();
+        ReadServiceValues();
+        ReadUIValues();
+        ReadWebServiceValues();
+        ReadMiscellaneousValues();
+    }
     ReadCoreValues();
+    ReadCpuValues();
     ReadRendererValues();
     ReadAudioValues();
-    ReadDataStorageValues();
     ReadSystemValues();
-    ReadMiscellaneousValues();
-    ReadDebuggingValues();
-    ReadWebServiceValues();
-    ReadServiceValues();
-    ReadDisabledAddOnValues();
-    ReadUIValues();
 }
 
 void Config::SavePlayerValues() {
@@ -900,30 +929,36 @@ void Config::SaveTouchscreenValues() {
 }
 
 void Config::SaveValues() {
-    SaveControlValues();
+    if (global) {
+        SaveControlValues();
+        SaveDataStorageValues();
+        SaveDebuggingValues();
+        SaveDisabledAddOnValues();
+        SaveServiceValues();
+        SaveUIValues();
+        SaveWebServiceValues();
+        SaveMiscellaneousValues();
+    }
     SaveCoreValues();
+    SaveCpuValues();
     SaveRendererValues();
     SaveAudioValues();
-    SaveDataStorageValues();
     SaveSystemValues();
-    SaveMiscellaneousValues();
-    SaveDebuggingValues();
-    SaveWebServiceValues();
-    SaveServiceValues();
-    SaveDisabledAddOnValues();
-    SaveUIValues();
 }
 
 void Config::SaveAudioValues() {
     qt_config->beginGroup(QStringLiteral("Audio"));
 
-    WriteSetting(QStringLiteral("output_engine"), QString::fromStdString(Settings::values.sink_id),
-                 QStringLiteral("auto"));
-    WriteSetting(QStringLiteral("enable_audio_stretching"),
-                 Settings::values.enable_audio_stretching, true);
-    WriteSetting(QStringLiteral("output_device"),
-                 QString::fromStdString(Settings::values.audio_device_id), QStringLiteral("auto"));
-    WriteSetting(QStringLiteral("volume"), Settings::values.volume, 1.0f);
+    if (global) {
+        WriteSetting(QStringLiteral("output_engine"),
+                     QString::fromStdString(Settings::values.sink_id), QStringLiteral("auto"));
+        WriteSetting(QStringLiteral("output_device"),
+                     QString::fromStdString(Settings::values.audio_device_id),
+                     QStringLiteral("auto"));
+    }
+    WriteSettingGlobal(QStringLiteral("enable_audio_stretching"),
+                       Settings::values.enable_audio_stretching, true);
+    WriteSettingGlobal(QStringLiteral("volume"), Settings::values.volume, 1.0f);
 
     qt_config->endGroup();
 }
@@ -946,6 +981,7 @@ void Config::SaveControlValues() {
     WriteSetting(QStringLiteral("udp_input_port"), Settings::values.udp_input_port,
                  InputCommon::CemuhookUDP::DEFAULT_PORT);
     WriteSetting(QStringLiteral("udp_pad_index"), Settings::values.udp_pad_index, 0);
+    WriteSetting(QStringLiteral("use_docked_mode"), Settings::values.use_docked_mode, false);
 
     qt_config->endGroup();
 }
@@ -953,7 +989,7 @@ void Config::SaveControlValues() {
 void Config::SaveCoreValues() {
     qt_config->beginGroup(QStringLiteral("Core"));
 
-    WriteSetting(QStringLiteral("use_multi_core"), Settings::values.use_multi_core, false);
+    WriteSettingGlobal(QStringLiteral("use_multi_core"), Settings::values.use_multi_core, false);
 
     qt_config->endGroup();
 }
@@ -982,18 +1018,7 @@ void Config::SaveDataStorageValues() {
                  false);
     WriteSetting(QStringLiteral("gamecard_path"),
                  QString::fromStdString(Settings::values.gamecard_path), QStringLiteral(""));
-    WriteSetting(QStringLiteral("nand_total_size"),
-                 QVariant::fromValue<u64>(static_cast<u64>(Settings::values.nand_total_size)),
-                 QVariant::fromValue<u64>(static_cast<u64>(Settings::NANDTotalSize::S29_1GB)));
-    WriteSetting(QStringLiteral("nand_user_size"),
-                 QVariant::fromValue<u64>(static_cast<u64>(Settings::values.nand_user_size)),
-                 QVariant::fromValue<u64>(static_cast<u64>(Settings::NANDUserSize::S26GB)));
-    WriteSetting(QStringLiteral("nand_system_size"),
-                 QVariant::fromValue<u64>(static_cast<u64>(Settings::values.nand_system_size)),
-                 QVariant::fromValue<u64>(static_cast<u64>(Settings::NANDSystemSize::S2_5GB)));
-    WriteSetting(QStringLiteral("sdmc_size"),
-                 QVariant::fromValue<u64>(static_cast<u64>(Settings::values.sdmc_size)),
-                 QVariant::fromValue<u64>(static_cast<u64>(Settings::SDMCSize::S16GB)));
+
     qt_config->endGroup();
 }
 
@@ -1009,7 +1034,6 @@ void Config::SaveDebuggingValues() {
     WriteSetting(QStringLiteral("dump_exefs"), Settings::values.dump_exefs, false);
     WriteSetting(QStringLiteral("dump_nso"), Settings::values.dump_nso, false);
     WriteSetting(QStringLiteral("quest_flag"), Settings::values.quest_flag, false);
-    WriteSetting(QStringLiteral("disable_cpu_opt"), Settings::values.disable_cpu_opt, false);
     WriteSetting(QStringLiteral("disable_macro_jit"), Settings::values.disable_macro_jit, false);
 
     qt_config->endGroup();
@@ -1073,32 +1097,63 @@ void Config::SavePathValues() {
     qt_config->endGroup();
 }
 
+void Config::SaveCpuValues() {
+    qt_config->beginGroup(QStringLiteral("Cpu"));
+
+    if (global) {
+        WriteSetting(QStringLiteral("cpu_accuracy"),
+                     static_cast<int>(Settings::values.cpu_accuracy), 0);
+
+        WriteSetting(QStringLiteral("cpuopt_page_tables"), Settings::values.cpuopt_page_tables,
+                     true);
+        WriteSetting(QStringLiteral("cpuopt_block_linking"), Settings::values.cpuopt_block_linking,
+                     true);
+        WriteSetting(QStringLiteral("cpuopt_return_stack_buffer"),
+                     Settings::values.cpuopt_return_stack_buffer, true);
+        WriteSetting(QStringLiteral("cpuopt_fast_dispatcher"),
+                     Settings::values.cpuopt_fast_dispatcher, true);
+        WriteSetting(QStringLiteral("cpuopt_context_elimination"),
+                     Settings::values.cpuopt_context_elimination, true);
+        WriteSetting(QStringLiteral("cpuopt_const_prop"), Settings::values.cpuopt_const_prop, true);
+        WriteSetting(QStringLiteral("cpuopt_misc_ir"), Settings::values.cpuopt_misc_ir, true);
+        WriteSetting(QStringLiteral("cpuopt_reduce_misalign_checks"),
+                     Settings::values.cpuopt_reduce_misalign_checks, true);
+    }
+
+    qt_config->endGroup();
+}
+
 void Config::SaveRendererValues() {
     qt_config->beginGroup(QStringLiteral("Renderer"));
 
-    WriteSetting(QStringLiteral("backend"), static_cast<int>(Settings::values.renderer_backend), 0);
+    WriteSettingGlobal(QStringLiteral("backend"),
+                       static_cast<int>(Settings::values.renderer_backend.GetValue(global)),
+                       Settings::values.renderer_backend.UsingGlobal(), 0);
     WriteSetting(QStringLiteral("debug"), Settings::values.renderer_debug, false);
-    WriteSetting(QStringLiteral("vulkan_device"), Settings::values.vulkan_device, 0);
-    WriteSetting(QStringLiteral("aspect_ratio"), Settings::values.aspect_ratio, 0);
-    WriteSetting(QStringLiteral("max_anisotropy"), Settings::values.max_anisotropy, 0);
-    WriteSetting(QStringLiteral("use_frame_limit"), Settings::values.use_frame_limit, true);
-    WriteSetting(QStringLiteral("frame_limit"), Settings::values.frame_limit, 100);
-    WriteSetting(QStringLiteral("use_disk_shader_cache"), Settings::values.use_disk_shader_cache,
-                 true);
-    WriteSetting(QStringLiteral("gpu_accuracy"), static_cast<int>(Settings::values.gpu_accuracy),
-                 0);
-    WriteSetting(QStringLiteral("use_asynchronous_gpu_emulation"),
-                 Settings::values.use_asynchronous_gpu_emulation, false);
-    WriteSetting(QStringLiteral("use_vsync"), Settings::values.use_vsync, true);
-    WriteSetting(QStringLiteral("use_assembly_shaders"), Settings::values.use_assembly_shaders,
-                 false);
-    WriteSetting(QStringLiteral("use_fast_gpu_time"), Settings::values.use_fast_gpu_time, true);
-    WriteSetting(QStringLiteral("force_30fps_mode"), Settings::values.force_30fps_mode, false);
+    WriteSettingGlobal(QStringLiteral("vulkan_device"), Settings::values.vulkan_device, 0);
+    WriteSettingGlobal(QStringLiteral("aspect_ratio"), Settings::values.aspect_ratio, 0);
+    WriteSettingGlobal(QStringLiteral("max_anisotropy"), Settings::values.max_anisotropy, 0);
+    WriteSettingGlobal(QStringLiteral("use_frame_limit"), Settings::values.use_frame_limit, true);
+    WriteSettingGlobal(QStringLiteral("frame_limit"), Settings::values.frame_limit, 100);
+    WriteSettingGlobal(QStringLiteral("use_disk_shader_cache"),
+                       Settings::values.use_disk_shader_cache, true);
+    WriteSettingGlobal(QStringLiteral("gpu_accuracy"),
+                       static_cast<int>(Settings::values.gpu_accuracy.GetValue(global)),
+                       Settings::values.gpu_accuracy.UsingGlobal(), 0);
+    WriteSettingGlobal(QStringLiteral("use_asynchronous_gpu_emulation"),
+                       Settings::values.use_asynchronous_gpu_emulation, false);
+    WriteSettingGlobal(QStringLiteral("use_vsync"), Settings::values.use_vsync, true);
+    WriteSettingGlobal(QStringLiteral("use_assembly_shaders"),
+                       Settings::values.use_assembly_shaders, false);
+    WriteSettingGlobal(QStringLiteral("use_fast_gpu_time"), Settings::values.use_fast_gpu_time,
+                       true);
+    WriteSettingGlobal(QStringLiteral("force_30fps_mode"), Settings::values.force_30fps_mode,
+                       false);
 
     // Cast to double because Qt's written float values are not human-readable
-    WriteSetting(QStringLiteral("bg_red"), static_cast<double>(Settings::values.bg_red), 0.0);
-    WriteSetting(QStringLiteral("bg_green"), static_cast<double>(Settings::values.bg_green), 0.0);
-    WriteSetting(QStringLiteral("bg_blue"), static_cast<double>(Settings::values.bg_blue), 0.0);
+    WriteSettingGlobal(QStringLiteral("bg_red"), Settings::values.bg_red, 0.0);
+    WriteSettingGlobal(QStringLiteral("bg_green"), Settings::values.bg_green, 0.0);
+    WriteSettingGlobal(QStringLiteral("bg_blue"), Settings::values.bg_blue, 0.0);
 
     qt_config->endGroup();
 }
@@ -1126,23 +1181,28 @@ void Config::SaveShortcutValues() {
 void Config::SaveSystemValues() {
     qt_config->beginGroup(QStringLiteral("System"));
 
-    WriteSetting(QStringLiteral("use_docked_mode"), Settings::values.use_docked_mode, false);
     WriteSetting(QStringLiteral("current_user"), Settings::values.current_user, 0);
-    WriteSetting(QStringLiteral("language_index"), Settings::values.language_index, 1);
-    WriteSetting(QStringLiteral("region_index"), Settings::values.region_index, 1);
-    WriteSetting(QStringLiteral("time_zone_index"), Settings::values.time_zone_index, 0);
+    WriteSettingGlobal(QStringLiteral("language_index"), Settings::values.language_index, 1);
+    WriteSettingGlobal(QStringLiteral("region_index"), Settings::values.region_index, 1);
+    WriteSettingGlobal(QStringLiteral("time_zone_index"), Settings::values.time_zone_index, 0);
 
-    WriteSetting(QStringLiteral("rng_seed_enabled"), Settings::values.rng_seed.has_value(), false);
-    WriteSetting(QStringLiteral("rng_seed"), Settings::values.rng_seed.value_or(0), 0);
+    WriteSettingGlobal(QStringLiteral("rng_seed_enabled"),
+                       Settings::values.rng_seed.GetValue(global).has_value(),
+                       Settings::values.rng_seed.UsingGlobal(), false);
+    WriteSettingGlobal(QStringLiteral("rng_seed"),
+                       Settings::values.rng_seed.GetValue(global).value_or(0),
+                       Settings::values.rng_seed.UsingGlobal(), 0);
 
-    WriteSetting(QStringLiteral("custom_rtc_enabled"), Settings::values.custom_rtc.has_value(),
-                 false);
-    WriteSetting(QStringLiteral("custom_rtc"),
-                 QVariant::fromValue<long long>(
-                     Settings::values.custom_rtc.value_or(std::chrono::seconds{}).count()),
-                 0);
+    WriteSettingGlobal(QStringLiteral("custom_rtc_enabled"),
+                       Settings::values.custom_rtc.GetValue(global).has_value(),
+                       Settings::values.custom_rtc.UsingGlobal(), false);
+    WriteSettingGlobal(
+        QStringLiteral("custom_rtc"),
+        QVariant::fromValue<long long>(
+            Settings::values.custom_rtc.GetValue(global).value_or(std::chrono::seconds{}).count()),
+        Settings::values.custom_rtc.UsingGlobal(), 0);
 
-    WriteSetting(QStringLiteral("sound_index"), Settings::values.sound_index, 1);
+    WriteSettingGlobal(QStringLiteral("sound_index"), Settings::values.sound_index, 1);
 
     qt_config->endGroup();
 }
@@ -1234,6 +1294,34 @@ QVariant Config::ReadSetting(const QString& name, const QVariant& default_value)
     return result;
 }
 
+template <typename Type>
+void Config::ReadSettingGlobal(Settings::Setting<Type>& setting, const QString& name) {
+    const bool use_global = qt_config->value(name + QStringLiteral("/use_global"), true).toBool();
+    setting.SetGlobal(use_global);
+    if (global || !use_global) {
+        setting.SetValue(ReadSetting(name).value<Type>());
+    }
+}
+
+template <typename Type>
+void Config::ReadSettingGlobal(Settings::Setting<Type>& setting, const QString& name,
+                               const QVariant& default_value) {
+    const bool use_global = qt_config->value(name + QStringLiteral("/use_global"), true).toBool();
+    setting.SetGlobal(use_global);
+    if (global || !use_global) {
+        setting.SetValue(ReadSetting(name, default_value).value<Type>());
+    }
+}
+
+template <typename Type>
+void Config::ReadSettingGlobal(Type& setting, const QString& name,
+                               const QVariant& default_value) const {
+    const bool use_global = qt_config->value(name + QStringLiteral("/use_global"), true).toBool();
+    if (global || !use_global) {
+        setting = ReadSetting(name, default_value).value<Type>();
+    }
+}
+
 void Config::WriteSetting(const QString& name, const QVariant& value) {
     qt_config->setValue(name, value);
 }
@@ -1244,6 +1332,40 @@ void Config::WriteSetting(const QString& name, const QVariant& value,
     qt_config->setValue(name, value);
 }
 
+template <typename Type>
+void Config::WriteSettingGlobal(const QString& name, const Settings::Setting<Type>& setting) {
+    if (!global) {
+        qt_config->setValue(name + QStringLiteral("/use_global"), setting.UsingGlobal());
+    }
+    if (global || !setting.UsingGlobal()) {
+        qt_config->setValue(name, setting.GetValue(global));
+    }
+}
+
+template <typename Type>
+void Config::WriteSettingGlobal(const QString& name, const Settings::Setting<Type>& setting,
+                                const QVariant& default_value) {
+    if (!global) {
+        qt_config->setValue(name + QStringLiteral("/use_global"), setting.UsingGlobal());
+    }
+    if (global || !setting.UsingGlobal()) {
+        qt_config->setValue(name + QStringLiteral("/default"),
+                            setting.GetValue(global) == default_value.value<Type>());
+        qt_config->setValue(name, setting.GetValue(global));
+    }
+}
+
+void Config::WriteSettingGlobal(const QString& name, const QVariant& value, bool use_global,
+                                const QVariant& default_value) {
+    if (!global) {
+        qt_config->setValue(name + QStringLiteral("/use_global"), use_global);
+    }
+    if (global || !use_global) {
+        qt_config->setValue(name + QStringLiteral("/default"), value == default_value);
+        qt_config->setValue(name, value);
+    }
+}
+
 void Config::Reload() {
     ReadValues();
     // To apply default value changes