gpu: clamp resolved buffer ranges to storage buffer limits

gpu: preserve zero-stride buffer fetch payloads
gpu: preserve ordered-append command order
2026-10-22 11:56:58 +02:00 · 2026-10-19 13:53:18 +02:00 · 2026-10-19 12:23:43 +02:00 · 2026-10-19 11:46:39 +02:00 · 2026-10-19 05:26:41 +02:00 · 2026-10-19 05:17:32 +02:00
171 changed files with 9398 additions and 1443 deletions
@@ -113,7 +113,7 @@ jobs:
        shell: cmd
        run: |
          cmake --build _Build/windows --target launcher ^
-            audio_out2_port_tests pad_haptics_tests controller_settings_tests ime_dialog_tests lru_cache_tests virtual_memory_allocation_tests ^
+            audio_out2_port_tests pad_haptics_tests controller_settings_tests host_input_tests ime_dialog_tests lru_cache_tests virtual_memory_allocation_tests ^
            shader_recompiler_compute_tests ^
            resource_tracking_tests scalar_provenance_tests ^
            resource_materialization_tests shader_vertex_metadata_tests archive_file_tests avplayer_file_tests --parallel
@@ -121,7 +121,7 @@ jobs:
      - name: Test
        shell: cmd
        run: |
-          ctest --test-dir _Build/windows --output-on-failure -R "^(audio_out2_port|pad_haptics|controller_settings|ime_dialog|lru_cache|resource_tracking|scalar_provenance|resource_materialization|shader_vertex_metadata|archive_file|avplayer_file|virtual_memory_allocation)$"
+          ctest --test-dir _Build/windows --output-on-failure -R "^(audio_out2_port|pad_haptics|controller_settings|host_input|ime_dialog|lru_cache|resource_tracking|scalar_provenance|resource_materialization|shader_vertex_metadata|archive_file|avplayer_file|virtual_memory_allocation)$"

      - name: Install
        shell: cmd
@@ -202,7 +202,7 @@ jobs:
        shell: bash
        run: |
          cmake --build _Build/macos \
-            --target launcher audio_out2_port_tests pad_haptics_tests controller_settings_tests ime_dialog_tests lru_cache_tests virtual_memory_allocation_tests \
+            --target launcher audio_out2_port_tests pad_haptics_tests controller_settings_tests host_input_tests ime_dialog_tests lru_cache_tests virtual_memory_allocation_tests \
              memory_tracker_tests page_manager_tests shader_recompiler_compute_tests \
              resource_tracking_tests scalar_provenance_tests \
              resource_materialization_tests shader_vertex_metadata_tests archive_file_tests avplayer_file_tests \
@@ -212,7 +212,7 @@ jobs:
        shell: bash
        run: |
          ctest --test-dir _Build/macos --output-on-failure \
-            -R '^(audio_out2_port|pad_haptics|controller_settings|ime_dialog|lru_cache|memory_tracker|page_manager|resource_tracking|scalar_provenance|resource_materialization|shader_vertex_metadata|archive_file|avplayer_file|virtual_memory_allocation)$'
+            -R '^(audio_out2_port|pad_haptics|controller_settings|host_input|ime_dialog|lru_cache|memory_tracker|page_manager|resource_tracking|scalar_provenance|resource_materialization|shader_vertex_metadata|archive_file|avplayer_file|virtual_memory_allocation)$'

      - name: Install
        shell: bash
@@ -388,7 +388,7 @@ jobs:
        run: |
          cmake --build _Build/linux \
            --target launcher page_manager_tests memory_tracker_tests \
-              audio_out2_port_tests pad_haptics_tests controller_settings_tests ime_dialog_tests lru_cache_tests virtual_memory_allocation_tests \
+              audio_out2_port_tests pad_haptics_tests controller_settings_tests host_input_tests ime_dialog_tests lru_cache_tests virtual_memory_allocation_tests \
              resource_tracking_tests scalar_provenance_tests \
              resource_materialization_tests shader_vertex_metadata_tests archive_file_tests avplayer_file_tests \
            --parallel
@@ -397,7 +397,7 @@ jobs:
        shell: bash
        run: |
          ctest --test-dir _Build/linux --output-on-failure \
-            -R '^(audio_out2_port|pad_haptics|controller_settings|ime_dialog|lru_cache|memory_tracker|page_manager|resource_tracking|scalar_provenance|resource_materialization|shader_vertex_metadata|archive_file|avplayer_file|virtual_memory_allocation)$'
+            -R '^(audio_out2_port|pad_haptics|controller_settings|host_input|ime_dialog|lru_cache|memory_tracker|page_manager|resource_tracking|scalar_provenance|resource_materialization|shader_vertex_metadata|archive_file|avplayer_file|virtual_memory_allocation)$'

      - name: Install
        shell: bash
@@ -407,12 +407,12 @@ jobs:
      - name: Verify artifacts
        shell: bash
        run: |
-          file _Build/linux/install/launcher
+          file _Build/linux/install/kyty_launcher
          file _Build/linux/install/kyty_emulator
          test -x _Build/linux/install/kyty_updater
          file _Build/linux/install/kyty_emulator | grep -q "ELF 64-bit LSB .*x86-64"
          ldd _Build/linux/install/kyty_emulator > /dev/null
-          readelf -d _Build/linux/install/launcher | grep -q 'RPATH.*\$ORIGIN/lib'
+          readelf -d _Build/linux/install/kyty_launcher | grep -q 'RPATH.*\$ORIGIN/lib'
          while IFS= read -r -d '' binary; do
            while read -r dependency; do
              test -e "_Build/linux/install/lib/$dependency"
@@ -421,8 +421,8 @@ jobs:
                sed -n 's/.*Shared library: \[\(libQt6[^]]*\|libicu[^]]*\)\].*/\1/p'
            )
          done < <(
-            find _Build/linux/install/launcher _Build/linux/install/plugins \
-              -type f \( -name launcher -o -name '*.so' \) -print0
+            find _Build/linux/install/kyty_launcher _Build/linux/install/plugins \
+              -type f \( -name kyty_launcher -o -name '*.so' \) -print0
          )
          _Build/linux/install/kyty_emulator --help > /dev/null

@@ -497,7 +497,7 @@ jobs:
          linux_dir="artifacts/KytyPS5-Linux-x86_64"
          test -d "$linux_dir"
          chmod a+x \
-            "$linux_dir/launcher" \
+            "$linux_dir/kyty_launcher" \
            "$linux_dir/kyty_emulator" \
            "$linux_dir/kyty_updater"
          tar -C "$linux_dir" -czf \
@@ -5,7 +5,7 @@
            "name": "Debug launcher UI",
            "type": "cppvsdbg",
            "request": "launch",
-            "program": "${workspaceFolder}\\_Build\\vscode-clang\\launcher\\launcher.exe",
+            "program": "${workspaceFolder}\\_Build\\vscode-clang\\launcher\\kyty_launcher.exe",
            "args": [],
            "stopAtEntry": false,
            "cwd": "${workspaceFolder}\\_Build\\vscode-clang\\launcher",
@@ -13,7 +13,7 @@
            "externalConsole": false,
            "linux": {
                "type": "cppdbg",
-                "program": "${workspaceFolder}/_Build/vscode-clang/launcher/launcher",
+                "program": "${workspaceFolder}/_Build/vscode-clang/launcher/kyty_launcher",
                "cwd": "${workspaceFolder}/_Build/vscode-clang/launcher",
                "MIMode": "gdb",
                "miDebuggerPath": "/usr/bin/gdb"
@@ -350,7 +350,7 @@ set(inc_headers
 	${KYTY_THIRD_PARTY_DIR}/SPIRV-Tools/include
 	${KYTY_THIRD_PARTY_DIR}/SPIRV-Headers/include
 	${KYTY_THIRD_PARTY_DIR}/fmt/include
-	${KYTY_THIRD_PARTY_DIR}/xxhash
+	${KYTY_THIRD_PARTY_DIR}/xxHash
 	${KYTY_THIRD_PARTY_DIR}/cpuinfo/include
 	${KYTY_THIRD_PARTY_DIR}/renderdoc
 	${KYTY_THIRD_PARTY_DIR}/stb
@@ -526,6 +526,10 @@ add_executable(controller_settings_tests EXCLUDE_FROM_ALL
 target_link_libraries(controller_settings_tests common fmt::fmt SDL3::SDL3-static)
 target_include_directories(controller_settings_tests PRIVATE ${inc_headers})

+add_executable(host_input_tests EXCLUDE_FROM_ALL "${KYTY_TESTS_DIR}/HostInputTests.cpp")
+target_link_libraries(host_input_tests common fmt::fmt SDL3::SDL3-static)
+target_include_directories(host_input_tests PRIVATE ${inc_headers})
+
 add_executable(save_data_memory_tests EXCLUDE_FROM_ALL
 	"${KYTY_TESTS_DIR}/SaveDataMemoryTests.cpp"
 )
@@ -644,6 +648,7 @@ if(BUILD_TESTING)
 	add_test(NAME ngs2_sampler COMMAND $<TARGET_FILE:ngs2_sampler_tests>)
 	add_test(NAME pad_haptics COMMAND $<TARGET_FILE:pad_haptics_tests>)
 	add_test(NAME controller_settings COMMAND $<TARGET_FILE:controller_settings_tests>)
+	add_test(NAME host_input COMMAND $<TARGET_FILE:host_input_tests>)
 	add_test(NAME save_data_memory COMMAND $<TARGET_FILE:save_data_memory_tests>)
 	add_test(NAME png_enc COMMAND $<TARGET_FILE:png_enc_tests>)
 	add_test(NAME ces COMMAND $<TARGET_FILE:ces_tests>)
@@ -723,6 +728,7 @@ if(BUILD_TESTING)
 		ngs2_sampler_tests
 		pad_haptics_tests
 		controller_settings_tests
+		host_input_tests
 		save_data_memory_tests
 		png_enc_tests
 		ces_tests
@@ -760,6 +766,14 @@ add_custom_target(kyty_trophy_sound
 add_dependencies(kyty_emulator kyty_trophy_sound)
 install(FILES "${KYTY_TROPHY_UNLOCK_SOUND}" DESTINATION assets/sounds)

+add_custom_target(kyty_touch_cursors
+	COMMAND ${CMAKE_COMMAND} -E copy_directory
+		"${CMAKE_CURRENT_SOURCE_DIR}/assets/cursors"
+		"$<TARGET_FILE_DIR:kyty_emulator>/assets/cursors"
+	VERBATIM)
+add_dependencies(kyty_emulator kyty_touch_cursors)
+install(DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/assets/cursors" DESTINATION assets)
+
 target_link_libraries(kyty_emulator ${kyty_emulator_link_libraries})
 if (WIN32)
 	target_link_libraries(kyty_emulator opengl32)
@@ -275,11 +275,11 @@ From the repository root, follow the launcher configure, build and install steps
 [Building on Linux](#building-on-linux), omitting `-DCMAKE_PREFIX_PATH="$Qt6_DIR"`; CMake finds
 Fedora's packaged Qt 6 automatically.

-Run `./_Build/linux/install/launcher` on regular Fedora or inside the container. To launch a
+Run `./_Build/linux/install/kyty_launcher` on regular Fedora or inside the container. To launch a
 container build from the host, leave the container with `exit`, then run from the repository root:

 ```bash
-distrobox enter kyty-build -- ./_Build/linux/install/launcher
+distrobox enter kyty-build -- ./_Build/linux/install/kyty_launcher
 ```

 Qt libraries and plugins are bundled. You can also launch directly on a Fedora Atomic host
@@ -345,7 +345,7 @@ Use `_Build/linux` instead of `_Build/windows` for a Linux build.

 A ready-made Visual Studio Code setup is included in [`.vscode`](.vscode). It configures CMake
 Tools to build the project with Ninja and `clang-cl` and provides launch profiles for both
-`launcher.exe` and `kyty_emulator.exe`. It is Windows-only: VS Code settings cannot select a
+`kyty_launcher.exe` and `kyty_emulator.exe`. It is Windows-only: VS Code settings cannot select a
 compiler per platform, so on Linux configure from the command line as shown above.

 Before using it:
@@ -365,11 +365,11 @@ Update your graphics driver before reporting rendering problems.
 To use the graphical launcher:

 ```powershell
-.\_Build\windows\install\launcher.exe
+.\_Build\windows\install\kyty_launcher.exe
 ```

 ```bash
-./_Build/linux/install/launcher
+./_Build/linux/install/kyty_launcher
 ```

 ```bash
@@ -0,0 +1,4 @@
+{
+  "trophy.toast.earned": "تم كسب الكأس!",
+  "user.guest": "ضيف"
+}
@@ -0,0 +1,4 @@
+{
+  "trophy.toast.earned": "Trofej získána!",
+  "user.guest": "Host"
+}
@@ -0,0 +1,4 @@
+{
+  "trophy.toast.earned": "Trofæ optjent!",
+  "user.guest": "Gæst"
+}
@@ -0,0 +1,4 @@
+{
+  "trophy.toast.earned": "Trophäe erhalten!",
+  "user.guest": "Gast"
+}
@@ -0,0 +1,4 @@
+{
+  "trophy.toast.earned": "Κερδίσατε ένα τρόπαιο!",
+  "user.guest": "Επισκ."
+}
@@ -0,0 +1,4 @@
+{
+  "trophy.toast.earned": "Trophy earned!",
+  "user.guest": "Guest"
+}
@@ -0,0 +1,4 @@
+{
+  "trophy.toast.earned": "Trophy earned!",
+  "user.guest": "Guest"
+}
@@ -0,0 +1,4 @@
+{
+  "trophy.toast.earned": "¡Ganaste un trofeo!",
+  "user.guest": "Invitado"
+}
@@ -0,0 +1,4 @@
+{
+  "trophy.toast.earned": "¡Ganaste un trofeo!",
+  "user.guest": "Invitado"
+}
@@ -0,0 +1,4 @@
+{
+  "trophy.toast.earned": "Pokaali ansaittu!",
+  "user.guest": "Vieras"
+}
@@ -0,0 +1,4 @@
+{
+  "trophy.toast.earned": "Trophée obtenu !",
+  "user.guest": "Invité"
+}
@@ -0,0 +1,4 @@
+{
+  "trophy.toast.earned": "Trophée obtenu !",
+  "user.guest": "Invité"
+}
@@ -0,0 +1,4 @@
+{
+  "trophy.toast.earned": "Trófea megszerezve!",
+  "user.guest": "Vendég"
+}
@@ -0,0 +1,4 @@
+{
+  "trophy.toast.earned": "Trofi diperoleh!",
+  "user.guest": "Tamu"
+}
@@ -0,0 +1,4 @@
+{
+  "trophy.toast.earned": "Trofeo ottenuto!",
+  "user.guest": "Ospite"
+}
@@ -0,0 +1,4 @@
+{
+  "trophy.toast.earned": "トロフィーを獲得しました！",
+  "user.guest": "ゲスト"
+}
@@ -0,0 +1,4 @@
+{
+  "trophy.toast.earned": "트로피를 획득했습니다!",
+  "user.guest": "게스트"
+}
@@ -0,0 +1,4 @@
+{
+  "trophy.toast.earned": "Trofee behaald!",
+  "user.guest": "Gast"
+}
@@ -0,0 +1,4 @@
+{
+  "trophy.toast.earned": "Trofé opptjent!",
+  "user.guest": "Gjest"
+}
@@ -0,0 +1,4 @@
+{
+  "trophy.toast.earned": "Zdobyto trofeum!",
+  "user.guest": "Gość"
+}
@@ -0,0 +1,4 @@
+{
+  "trophy.toast.earned": "Troféu conquistado!",
+  "user.guest": "Convidado"
+}
@@ -0,0 +1,4 @@
+{
+  "trophy.toast.earned": "Troféu conquistado!",
+  "user.guest": "Convidado"
+}
@@ -0,0 +1,4 @@
+{
+  "trophy.toast.earned": "Trofeu câștigat!",
+  "user.guest": "Invitat"
+}
@@ -0,0 +1,4 @@
+{
+  "trophy.toast.earned": "Трофей получен!",
+  "user.guest": "Гость"
+}
@@ -0,0 +1,4 @@
+{
+  "trophy.toast.earned": "Trofé intjänad!",
+  "user.guest": "Gäst"
+}
@@ -0,0 +1,4 @@
+{
+  "trophy.toast.earned": "ได้รับถ้วยรางวัลแล้ว!",
+  "user.guest": "แขก"
+}
@@ -0,0 +1,4 @@
+{
+  "trophy.toast.earned": "Kupa kazanıldı!",
+  "user.guest": "Misafir"
+}
@@ -0,0 +1,4 @@
+{
+  "trophy.toast.earned": "Đã nhận cúp!",
+  "user.guest": "Khách"
+}
@@ -0,0 +1,4 @@
+{
+  "trophy.toast.earned": "已获得奖杯！",
+  "user.guest": "访客"
+}
@@ -0,0 +1,4 @@
+{
+  "trophy.toast.earned": "已獲得獎盃！",
+  "user.guest": "訪客"
+}
@@ -0,0 +1,20 @@
+# Notification localization
+
+Trophy notification subtitles use the configured console language. Trophy names still come
+from the game's metadata. The launcher trophy dialogs are unchanged.
+
+Edit `assets/localization/<locale>.json`, using flat UTF-8 string values:
+
+```json
+{
+  "trophy.toast.earned": "Trophy earned!"
+}
+```
+
+CMake converts the catalogs into C++ byte arrays in the build directory. The emulator parses
+the embedded catalogs once and falls back to `en-US.json` for missing translations. No language
+files or translation service are needed at runtime. Rebuild after editing a catalog.
+
+For local testing, rebuild `kyty_emulator`, choose a game's Console Language in the launcher,
+and unlock a trophy with notifications enabled. Installed system fonts provide non-Latin
+glyphs; the selected language needs an appropriate font on the host.
@@ -12,9 +12,23 @@ endif()

 add_library(common STATIC ${common_src} ${common_headers})

+set(localization_dir "${PROJECT_SOURCE_DIR}/assets/localization")
+file(GLOB localization_catalogs CONFIGURE_DEPENDS "${localization_dir}/*.json")
+set(localization_data "${CMAKE_CURRENT_BINARY_DIR}/localizationData.h")
+add_custom_command(
+	OUTPUT "${localization_data}"
+	COMMAND ${CMAKE_COMMAND} "-DINPUTS=${localization_catalogs}" "-DOUTPUT=${localization_data}"
+		-P "${KYTY_SOURCE_DIR}/embed_localization.cmake"
+	DEPENDS ${localization_catalogs} "${KYTY_SOURCE_DIR}/embed_localization.cmake"
+	COMMENT "Embed notification localization catalogs"
+	VERBATIM
+)
+target_sources(common PRIVATE "${localization_data}")
+
 target_include_directories(common
 	PUBLIC ${KYTY_SOURCE_DIR}
 	PRIVATE
+		${CMAKE_CURRENT_BINARY_DIR}
 		${KYTY_THIRD_PARTY_DIR}/cpuinfo/include
 )

@@ -29,6 +43,10 @@ target_link_libraries(common
 		ZArchive::zarchive
 		nlohmann_json::nlohmann_json
 )
+if(APPLE)
+	# hostStats.cpp reads GPU utilization from the IOAccelerator registry entries.
+	target_link_libraries(common PRIVATE "-framework CoreFoundation" "-framework IOKit")
+endif()

 get_property(inc_headers TARGET common PROPERTY INCLUDE_DIRECTORIES)
 set(check_headers ${CMAKE_CURRENT_SOURCE_DIR})
@@ -0,0 +1,117 @@
+#ifndef KYTY_COMMON_DEBUG_COUNTERS_H_
+#define KYTY_COMMON_DEBUG_COUNTERS_H_
+
+#include <atomic>
+#include <chrono>
+#include <cstdint>
+
+namespace Common::DebugCounters {
+
+static_assert(std::atomic<uint64_t>::is_always_lock_free && std::atomic<bool>::is_always_lock_free);
+
+// Collect events only while the panel is visible. Relaxed, lock-free atomics keep updates
+// cheap and safe in signal handlers. Rates come from comparing two snapshots.
+enum class Counter {
+	// GPU work.
+	Draws,
+	DrawInstances,
+	IndirectDraws,
+	Dispatches,
+	QueueSubmits,
+	GpuWaits, // CPU blocked on a GPU timeline semaphore.
+	GpuWaitNs,
+	GpuFaults, // Guest writes to GPU-tracked memory.
+	ShadersCompiled,
+	ShaderCompileNs,
+	PipelinesCreated,
+	PipelineCreateNs,
+	// Caches.
+	TextureUploadBytes,
+	TextureDownloadBytes,
+	TextureEvictions,
+	BufferUploadBytes,
+	BufferDownloadBytes,
+	BufferEvictions,
+	// Presentation.
+	FlipsRequested,
+	FlipsRejected, // Guest flip queue was full.
+	LateVblanks,   // The present thread fell behind the vblank rate.
+	SwapchainRecreations,
+	DroppedPresents,
+	// Emulator.
+	EmulatedInstructions, // Host CPU lacked an instruction the guest used.
+	UnresolvedImportCalls,
+	GuestFileReadBytes,
+	GuestFileWriteBytes,
+	AudioUnderruns,
+	Count
+};
+
+enum class Peak { ShaderCompileNs, PipelineCreateNs, Count };
+
+namespace Detail {
+inline std::atomic<bool>     g_enabled {false};
+inline std::atomic<uint64_t> g_counters[static_cast<int>(Counter::Count)] {};
+inline std::atomic<uint64_t> g_peaks[static_cast<int>(Peak::Count)] {};
+} // namespace Detail
+
+[[nodiscard]] inline bool Enabled() noexcept {
+	return Detail::g_enabled.load(std::memory_order_relaxed);
+}
+
+inline void SetEnabled(bool enabled) noexcept {
+	Detail::g_enabled.store(enabled, std::memory_order_relaxed);
+}
+
+inline void Add(Counter counter, uint64_t value = 1) noexcept {
+	if (!Enabled()) {
+		return;
+	}
+	Detail::g_counters[static_cast<int>(counter)].fetch_add(value, std::memory_order_relaxed);
+}
+
+class ScopedTimer final {
+public:
+	ScopedTimer(Counter count, Counter duration, Peak peak = Peak::Count)
+	    : m_count(count), m_duration(duration), m_peak(peak),
+	      m_start(Enabled() ? std::chrono::steady_clock::now()
+	                        : std::chrono::steady_clock::time_point {}) {}
+	~ScopedTimer() {
+		if (m_start == std::chrono::steady_clock::time_point {} || !Enabled()) {
+			return;
+		}
+		const auto elapsed =
+		    static_cast<uint64_t>(std::chrono::duration_cast<std::chrono::nanoseconds>(
+		                              std::chrono::steady_clock::now() - m_start)
+		                              .count());
+		Add(m_count);
+		Add(m_duration, elapsed);
+		if (m_peak != Peak::Count) {
+			auto& peak     = Detail::g_peaks[static_cast<int>(m_peak)];
+			auto  previous = peak.load(std::memory_order_relaxed);
+			while (previous < elapsed &&
+			       !peak.compare_exchange_weak(previous, elapsed, std::memory_order_relaxed)) {
+			}
+		}
+	}
+	ScopedTimer(const ScopedTimer&)            = delete;
+	ScopedTimer& operator=(const ScopedTimer&) = delete;
+
+private:
+	Counter                               m_count;
+	Counter                               m_duration;
+	Peak                                  m_peak;
+	std::chrono::steady_clock::time_point m_start;
+};
+
+[[nodiscard]] inline uint64_t Get(Counter counter) noexcept {
+	return Detail::g_counters[static_cast<int>(counter)].load(std::memory_order_relaxed);
+}
+
+[[nodiscard]] inline uint64_t Get(Peak peak) noexcept {
+	return Detail::g_peaks[static_cast<int>(peak)].load(std::memory_order_relaxed);
+}
+
+} // namespace Common::DebugCounters
+
+#endif /* KYTY_COMMON_DEBUG_COUNTERS_H_ */
@@ -1,6 +1,7 @@
 #include "common/emulatorConfig.h"

 #include "common/assert.h"
+#include "common/debugCounters.h"

 #include <algorithm>
 #include <memory>
@@ -16,6 +17,7 @@ void Initialize() {
 }

 void Shutdown() {
+	Common::DebugCounters::SetEnabled(false);
 	g_config.reset();
 }

@@ -24,8 +26,10 @@ void Load(const ConfigOptions& cfg) {
 	EXIT_IF(cfg.user_name.empty() || cfg.user_name.size() > MAX_USER_NAME_LENGTH);
 	EXIT_IF(!IsConfiguredUserIdValid(cfg.user_id));
 	EXIT_IF(cfg.controller_speaker_volume > 100 || cfg.controller_vibration_intensity > 100);
+	EXIT_IF(cfg.keyboard_player > static_cast<uint32_t>(MAX_PLAYERS));

 	*g_config = cfg;
+	Common::DebugCounters::SetEnabled(cfg.perf_overlay_enabled);
 }

 uint32_t GetScreenWidth() {
@@ -60,6 +64,27 @@ uint32_t GetControllerVibrationIntensity() {
 	return g_config->controller_vibration_intensity;
 }

+uint32_t GetKeyboardPlayer() {
+	return g_config->keyboard_player;
+}
+
+int32_t GetPlayerUserId(int player) {
+	return g_config->user_id + player * PLAYER_USER_ID_STEP;
+}
+
+int GetPlayerFromUserId(int32_t user_id) {
+	for (int player = 0; player < MAX_PLAYERS; player++) {
+		if (user_id == GetPlayerUserId(player)) {
+			return player;
+		}
+	}
+	return -1;
+}
+
+bool IsGuestUserId(int32_t user_id) {
+	return GetPlayerFromUserId(user_id) > 0;
+}
+
 PresentMode GetPresentMode() {
 	return g_config->present_mode;
 }
@@ -6,6 +6,7 @@
 #include <cstddef>
 #include <array>
 #include <filesystem>
+#include <limits>
 #include <optional>
 #include <string>
 #include <vector>
@@ -35,10 +36,16 @@ constexpr uint32_t MAX_CONSOLE_LANGUAGE     = 29;
 constexpr std::size_t MAX_USER_NAME_LENGTH = 16;
 constexpr int32_t DEFAULT_USER_ID           = 1000;

+constexpr int     MAX_PLAYERS         = 4;
+constexpr int32_t PLAYER_USER_ID_STEP = 0x100;
+constexpr int32_t MAX_USER_ID =
+    std::numeric_limits<int32_t>::max() - (MAX_PLAYERS - 1) * PLAYER_USER_ID_STEP;
+
 constexpr bool IsConfiguredUserIdValid(int32_t user_id) {
 	constexpr int32_t USER_ID_EVERYONE = 0xfe;
 	constexpr int32_t USER_ID_SYSTEM   = 0xff;
-	return user_id >= 0 && user_id != USER_ID_EVERYONE && user_id != USER_ID_SYSTEM;
+	return user_id >= 0 && user_id <= MAX_USER_ID && user_id != USER_ID_EVERYONE &&
+	       user_id != USER_ID_SYSTEM;
 }

 struct ConfigOptions {
@@ -50,6 +57,8 @@ struct ConfigOptions {
 	std::optional<ControllerColor> controller_color;
 	uint32_t               controller_speaker_volume      = 50;
 	uint32_t               controller_vibration_intensity = 100;
+	// 0 shares player 1 with the first gamepad; 1..MAX_PLAYERS dedicates that player to the keyboard.
+	uint32_t               keyboard_player                = 0;
 	PresentMode            present_mode                = PresentMode::Mailbox;
 	int32_t                gpu_index                   = -1;
 	bool                   fullscreen_enabled          = false;
@@ -80,6 +89,7 @@ struct ConfigOptions {
 	bool                   tessellation_enabled        = false;
 	bool                   trophy_enabled              = true;
 	uint32_t               trophy_duration_seconds     = 6;
+	bool                   perf_overlay_enabled            = false;
 	bool                   playgo_hack_enabled         = false;
 	bool                   skip_notice_screen          = false;
 #if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
@@ -100,6 +110,13 @@ const std::string& GetAudioInputDevice();
 const std::optional<ControllerColor>& GetControllerColor();
 uint32_t GetControllerSpeakerVolume();
 uint32_t GetControllerVibrationIntensity();
+uint32_t GetKeyboardPlayer();
+// Player 0 is the configured user; later players derive their IDs from it.
+int32_t  GetPlayerUserId(int player);
+// Returns -1 when the user ID does not belong to any player.
+int      GetPlayerFromUserId(int32_t user_id);
+// Players after the first are guests: they have no profile, trophies or NP account.
+bool     IsGuestUserId(int32_t user_id);
 PresentMode GetPresentMode();
 int32_t GetGpuIndex();
 bool     FullscreenEnabled();
@@ -0,0 +1,580 @@
+#include "common/hostStats.h"
+
+#include "common/config.h"
+
+#include <algorithm>
+#include <chrono>
+#include <thread>
+#include <unordered_map>
+
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
+#ifndef NOMINMAX
+#define NOMINMAX
+#endif
+#include <windows.h> // IWYU pragma: keep
+// clang-format off
+#include <psapi.h>
+// clang-format on
+#elif defined(__APPLE__)
+#include <CoreFoundation/CoreFoundation.h>
+#include <IOKit/IOKitLib.h>
+#include <cstring>
+#include <libproc.h>
+#include <mach/mach.h>
+#include <notify.h>
+#include <sys/resource.h>
+#include <sys/sysctl.h>
+#include <unistd.h>
+#else
+#include <filesystem>
+#include <fstream>
+#include <sstream>
+#include <sys/resource.h>
+#include <unistd.h>
+#endif
+
+namespace Common {
+
+namespace {
+
+// Cumulative counters; rates come from the difference between two samples.
+struct Counters {
+	struct Thread {
+		uint64_t    cpu_ns;
+		std::string name;
+	};
+	std::optional<uint64_t>              process_cpu_ns;
+	std::optional<uint64_t>              system_busy;
+	std::optional<uint64_t>              system_total;
+	std::optional<uint64_t>              context_switches;
+	std::optional<uint64_t>              page_faults;
+	std::optional<uint64_t>              page_ins;
+	std::optional<uint64_t>              disk_read;
+	std::optional<uint64_t>              disk_write;
+	std::unordered_map<uint64_t, Thread> threads;
+	uint64_t                             wall_ns = 0;
+};
+
+uint64_t WallNanoseconds() {
+	return static_cast<uint64_t>(std::chrono::duration_cast<std::chrono::nanoseconds>(
+	                                 std::chrono::steady_clock::now().time_since_epoch())
+	                                 .count());
+}
+
+#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
+
+uint64_t FileTimeTo100Ns(const FILETIME& time) {
+	return (static_cast<uint64_t>(time.dwHighDateTime) << 32u) | time.dwLowDateTime;
+}
+
+struct PlatformState {};
+
+void SamplePlatform(PlatformState& /*state*/, HostStats& stats, Counters& now) {
+	const auto process = GetCurrentProcess();
+	FILETIME   creation {};
+	FILETIME   exit {};
+	FILETIME   kernel {};
+	FILETIME   user {};
+	if (GetProcessTimes(process, &creation, &exit, &kernel, &user) != 0) {
+		now.process_cpu_ns = (FileTimeTo100Ns(kernel) + FileTimeTo100Ns(user)) * 100u;
+	}
+	FILETIME idle {};
+	if (GetSystemTimes(&idle, &kernel, &user) != 0) {
+		// Kernel time includes idle time.
+		now.system_total = FileTimeTo100Ns(kernel) + FileTimeTo100Ns(user);
+		now.system_busy  = *now.system_total - FileTimeTo100Ns(idle);
+	}
+
+	PROCESS_MEMORY_COUNTERS memory {};
+	if (K32GetProcessMemoryInfo(process, &memory, sizeof(memory)) != 0) {
+		stats.process_memory   = memory.WorkingSetSize;
+		stats.process_resident = memory.WorkingSetSize;
+		now.page_faults        = memory.PageFaultCount;
+	}
+	IO_COUNTERS io {};
+	if (GetProcessIoCounters(process, &io) != 0) {
+		now.disk_read  = io.ReadTransferCount;
+		now.disk_write = io.WriteTransferCount;
+	}
+
+	MEMORYSTATUSEX system {};
+	system.dwLength = sizeof(system);
+	if (GlobalMemoryStatusEx(&system) != 0) {
+		stats.system_memory_total = system.ullTotalPhys;
+		stats.system_memory_used  = system.ullTotalPhys - system.ullAvailPhys;
+		stats.system_free         = system.ullAvailPhys;
+		stats.memory_pressure     = system.dwMemoryLoad >= 90   ? MemoryPressure::Critical
+		                            : system.dwMemoryLoad >= 80 ? MemoryPressure::Warning
+		                                                        : MemoryPressure::Normal;
+	}
+}
+
+#else
+
+void SampleRusage(Counters& now) {
+	rusage usage {};
+	if (getrusage(RUSAGE_SELF, &usage) != 0) {
+		return;
+	}
+	const auto to_ns = [](const timeval& time) {
+		return static_cast<uint64_t>(time.tv_sec) * 1000000000u +
+		       static_cast<uint64_t>(time.tv_usec) * 1000u;
+	};
+	now.process_cpu_ns   = to_ns(usage.ru_utime) + to_ns(usage.ru_stime);
+	now.context_switches = static_cast<uint64_t>(usage.ru_nvcsw + usage.ru_nivcsw);
+	now.page_faults      = static_cast<uint64_t>(usage.ru_minflt + usage.ru_majflt);
+	now.page_ins         = static_cast<uint64_t>(usage.ru_majflt);
+}
+
+#if defined(__APPLE__)
+
+template <typename T>
+std::optional<T> SysctlValue(const char* name) {
+	T      value {};
+	size_t size = sizeof(value);
+	if (sysctlbyname(name, &value, &size, nullptr, 0) != 0 || size != sizeof(value)) {
+		return std::nullopt;
+	}
+	return value;
+}
+
+struct PlatformState {
+	mach_port_t             host          = mach_host_self();
+	int                     thermal_token = -1;
+	bool                    thermal_ready = false;
+	std::string             gpu_model;
+	std::optional<uint32_t> gpu_cores;
+
+	~PlatformState() {
+		if (thermal_ready) {
+			notify_cancel(thermal_token);
+		}
+		mach_port_deallocate(mach_task_self(), host);
+	}
+};
+
+std::optional<int64_t> DictionaryNumber(CFDictionaryRef dictionary, CFStringRef key) {
+	const auto* value  = CFDictionaryGetValue(dictionary, key);
+	int64_t     number = 0;
+	if (value == nullptr || CFGetTypeID(value) != CFNumberGetTypeID() ||
+	    !CFNumberGetValue(static_cast<CFNumberRef>(value), kCFNumberSInt64Type, &number) ||
+	    number < 0) {
+		return std::nullopt;
+	}
+	return number;
+}
+
+std::string RegistryString(io_object_t service, CFStringRef key) {
+	std::string result;
+	auto*       value = IORegistryEntryCreateCFProperty(service, key, kCFAllocatorDefault, 0);
+	if (value == nullptr) {
+		return result;
+	}
+	if (CFGetTypeID(value) == CFStringGetTypeID()) {
+		char buffer[128] {};
+		if (CFStringGetCString(static_cast<CFStringRef>(value), buffer, sizeof(buffer),
+		                       kCFStringEncodingUTF8)) {
+			result = buffer;
+		}
+	} else if (CFGetTypeID(value) == CFDataGetTypeID()) {
+		// Intel Macs store the model as NUL-terminated data.
+		const auto* data   = static_cast<CFDataRef>(value);
+		const auto* bytes  = reinterpret_cast<const char*>(CFDataGetBytePtr(data));
+		const auto  length = static_cast<size_t>(CFDataGetLength(data));
+		result.assign(bytes, strnlen(bytes, length));
+	}
+	CFRelease(value);
+	return result;
+}
+
+// IOAccelerator exposes whole-device statistics; there is no per-process figure.
+void SampleGpu(PlatformState& state, HostStats& stats) {
+	io_iterator_t iterator = 0;
+	if (IOServiceGetMatchingServices(MACH_PORT_NULL, IOServiceMatching("IOAccelerator"),
+	                                 &iterator) != KERN_SUCCESS) {
+		return;
+	}
+	const auto take_max = [](std::optional<double>& target, std::optional<int64_t> value) {
+		if (value) {
+			target = std::max(target.value_or(0.0), std::min(100.0, static_cast<double>(*value)));
+		}
+	};
+	const auto add = [](std::optional<uint64_t>& target, std::optional<int64_t> value) {
+		if (value) {
+			target = target.value_or(0) + static_cast<uint64_t>(*value);
+		}
+	};
+	for (io_object_t service = IOIteratorNext(iterator); service != 0;
+	     service             = IOIteratorNext(iterator)) {
+		if (state.gpu_model.empty()) {
+			state.gpu_model = RegistryString(service, CFSTR("model"));
+			auto*   cores   = IORegistryEntryCreateCFProperty(service, CFSTR("gpu-core-count"),
+			                                                  kCFAllocatorDefault, 0);
+			int64_t count   = 0;
+			if (cores != nullptr) {
+				if (CFGetTypeID(cores) == CFNumberGetTypeID() &&
+				    CFNumberGetValue(static_cast<CFNumberRef>(cores), kCFNumberSInt64Type,
+				                     &count) &&
+				    count > 0) {
+					state.gpu_cores = static_cast<uint32_t>(count);
+				}
+				CFRelease(cores);
+			}
+		}
+		auto* statistics = IORegistryEntryCreateCFProperty(service, CFSTR("PerformanceStatistics"),
+		                                                   kCFAllocatorDefault, 0);
+		if (statistics != nullptr) {
+			if (CFGetTypeID(statistics) == CFDictionaryGetTypeID()) {
+				const auto* values = static_cast<CFDictionaryRef>(statistics);
+				take_max(stats.gpu_percent,
+				         DictionaryNumber(values, CFSTR("Device Utilization %")));
+				take_max(stats.gpu_renderer_percent,
+				         DictionaryNumber(values, CFSTR("Renderer Utilization %")));
+				take_max(stats.gpu_tiler_percent,
+				         DictionaryNumber(values, CFSTR("Tiler Utilization %")));
+				add(stats.gpu_memory_in_use,
+				    DictionaryNumber(values, CFSTR("In use system memory")));
+				add(stats.gpu_memory_allocated,
+				    DictionaryNumber(values, CFSTR("Alloc system memory")));
+				add(stats.gpu_recoveries, DictionaryNumber(values, CFSTR("recoveryCount")));
+			}
+			CFRelease(statistics);
+		}
+		IOObjectRelease(service);
+	}
+	IOObjectRelease(iterator);
+	stats.gpu_model = state.gpu_model;
+	stats.gpu_cores = state.gpu_cores;
+}
+
+void SampleThreads(HostStats& stats, Counters& now) {
+	thread_act_array_t     threads = nullptr;
+	mach_msg_type_number_t count   = 0;
+	const auto             task    = mach_task_self();
+	if (task_threads(task, &threads, &count) != KERN_SUCCESS) {
+		return;
+	}
+	stats.thread_count = count;
+	for (mach_msg_type_number_t i = 0; i < count; i++) {
+		thread_identifier_info_data_t identifier {};
+		mach_msg_type_number_t        size = THREAD_IDENTIFIER_INFO_COUNT;
+		thread_basic_info_data_t      basic {};
+		mach_msg_type_number_t        basic_size = THREAD_BASIC_INFO_COUNT;
+		if (thread_info(threads[i], THREAD_IDENTIFIER_INFO,
+		                reinterpret_cast<thread_info_t>(&identifier), &size) == KERN_SUCCESS &&
+		    thread_info(threads[i], THREAD_BASIC_INFO, reinterpret_cast<thread_info_t>(&basic),
+		                &basic_size) == KERN_SUCCESS) {
+			const auto to_ns = [](const time_value_t& time) {
+				return static_cast<uint64_t>(time.seconds) * 1000000000u +
+				       static_cast<uint64_t>(time.microseconds) * 1000u;
+			};
+			auto& thread  = now.threads[identifier.thread_id];
+			thread.cpu_ns = to_ns(basic.user_time) + to_ns(basic.system_time);
+			thread_extended_info_data_t extended {};
+			mach_msg_type_number_t      extended_size = THREAD_EXTENDED_INFO_COUNT;
+			if (thread_info(threads[i], THREAD_EXTENDED_INFO,
+			                reinterpret_cast<thread_info_t>(&extended),
+			                &extended_size) == KERN_SUCCESS) {
+				thread.name = std::string(extended.pth_name,
+				                          strnlen(extended.pth_name, sizeof(extended.pth_name)));
+			}
+		}
+		mach_port_deallocate(task, threads[i]);
+	}
+	vm_deallocate(task, reinterpret_cast<vm_address_t>(threads), count * sizeof(thread_act_t));
+}
+
+void SamplePlatform(PlatformState& state, HostStats& stats, Counters& now) {
+	SampleRusage(now);
+
+	host_cpu_load_info_data_t load {};
+	mach_msg_type_number_t    load_count = HOST_CPU_LOAD_INFO_COUNT;
+	if (host_statistics(state.host, HOST_CPU_LOAD_INFO, reinterpret_cast<host_info_t>(&load),
+	                    &load_count) == KERN_SUCCESS) {
+		const uint64_t busy = static_cast<uint64_t>(load.cpu_ticks[CPU_STATE_USER]) +
+		                      load.cpu_ticks[CPU_STATE_SYSTEM] + load.cpu_ticks[CPU_STATE_NICE];
+		now.system_busy     = busy;
+		now.system_total    = busy + load.cpu_ticks[CPU_STATE_IDLE];
+	}
+
+	// Physical footprint is what Activity Monitor shows in its Memory column.
+	task_vm_info_data_t    vm_info {};
+	mach_msg_type_number_t vm_info_count = TASK_VM_INFO_COUNT;
+	if (task_info(mach_task_self(), TASK_VM_INFO, reinterpret_cast<task_info_t>(&vm_info),
+	              &vm_info_count) == KERN_SUCCESS) {
+		stats.process_memory      = vm_info.phys_footprint;
+		stats.process_resident    = vm_info.resident_size;
+		stats.process_compressed  = vm_info.compressed;
+		stats.process_file_backed = vm_info.external;
+		stats.process_virtual     = vm_info.virtual_size;
+	}
+	rusage_info_v2 usage {};
+	if (proc_pid_rusage(getpid(), RUSAGE_INFO_V2, reinterpret_cast<rusage_info_t*>(&usage)) == 0) {
+		now.disk_read  = usage.ri_diskio_bytesread;
+		now.disk_write = usage.ri_diskio_byteswritten;
+	}
+
+	stats.system_memory_total   = SysctlValue<uint64_t>("hw.memsize");
+	vm_size_t              page = 0;
+	vm_statistics64_data_t vm {};
+	mach_msg_type_number_t count = HOST_VM_INFO64_COUNT;
+	if (host_page_size(state.host, &page) == KERN_SUCCESS &&
+	    host_statistics64(state.host, HOST_VM_INFO64, reinterpret_cast<host_info64_t>(&vm),
+	                      &count) == KERN_SUCCESS) {
+		// Activity Monitor's "Memory Used": app memory + wired + compressed.
+		const uint64_t app       = vm.internal_page_count > vm.purgeable_count
+		                               ? vm.internal_page_count - vm.purgeable_count
+		                               : 0;
+		stats.system_memory_used = (app + vm.wire_count + vm.compressor_page_count) * page;
+		stats.system_wired       = static_cast<uint64_t>(vm.wire_count) * page;
+		stats.system_compressed  = static_cast<uint64_t>(vm.compressor_page_count) * page;
+		stats.system_cached =
+		    (static_cast<uint64_t>(vm.external_page_count) + vm.purgeable_count) * page;
+		stats.system_free = static_cast<uint64_t>(vm.free_count) * page;
+	}
+
+	if (const auto swap = SysctlValue<xsw_usage>("vm.swapusage")) {
+		stats.swap_used  = swap->xsu_used;
+		stats.swap_total = swap->xsu_total;
+	}
+
+	// 1 = normal, 2 = warning, 4 = critical, as reported to dispatch memory-pressure sources.
+	if (const auto level = SysctlValue<int>("kern.memorystatus_vm_pressure_level")) {
+		stats.memory_pressure = *level >= 4   ? MemoryPressure::Critical
+		                        : *level >= 2 ? MemoryPressure::Warning
+		                                      : MemoryPressure::Normal;
+	}
+
+	// 0 nominal, 1 moderate, 2 heavy, 3 trapping, 4 sleeping.
+	if (!state.thermal_ready) {
+		state.thermal_ready = notify_register_check("com.apple.system.thermalpressurelevel",
+		                                            &state.thermal_token) == NOTIFY_STATUS_OK;
+	}
+	uint64_t thermal = 0;
+	if (state.thermal_ready &&
+	    notify_get_state(state.thermal_token, &thermal) == NOTIFY_STATUS_OK) {
+		stats.thermal_pressure = thermal >= 3   ? ThermalPressure::Critical
+		                         : thermal == 2 ? ThermalPressure::Heavy
+		                         : thermal == 1 ? ThermalPressure::Moderate
+		                                        : ThermalPressure::Nominal;
+	}
+
+	SampleThreads(stats, now);
+	SampleGpu(state, stats);
+}
+
+#else
+
+using ProcValues = std::unordered_map<std::string, uint64_t>;
+
+// Reads "<key>: <number> [kB]" lines; kB values are returned in bytes.
+ProcValues ReadProcValues(const char* path) {
+	ProcValues    values;
+	std::ifstream file(path);
+	std::string   line;
+	while (std::getline(file, line)) {
+		const auto colon = line.find(':');
+		if (colon == std::string::npos) {
+			continue;
+		}
+		std::istringstream stream(line.substr(colon + 1));
+		uint64_t           number = 0;
+		std::string        unit;
+		if (stream >> number) {
+			stream >> unit;
+			values[line.substr(0, colon)] = unit == "kB" ? number * 1024u : number;
+		}
+	}
+	return values;
+}
+
+std::optional<uint64_t> Find(const ProcValues& values, const char* key) {
+	const auto it = values.find(key);
+	return it != values.end() ? std::optional(it->second) : std::nullopt;
+}
+
+struct PlatformState {};
+
+void SampleThreads(HostStats& stats, Counters& now) {
+	static const double ns_per_tick = 1e9 / static_cast<double>(std::max(sysconf(_SC_CLK_TCK), 1L));
+	std::error_code     error;
+	uint32_t            count = 0;
+	for (std::filesystem::directory_iterator it("/proc/self/task", error), end; !error && it != end;
+	     it.increment(error)) {
+		count++;
+		std::ifstream stat_file(it->path() / "stat");
+		std::string   stat;
+		uint64_t      id = 0;
+		if (!(stat_file >> id)) {
+			continue; // The thread may have exited during enumeration.
+		}
+		std::getline(stat_file, stat);
+		const auto close = stat.rfind(')');
+		if (close == std::string::npos) {
+			continue;
+		}
+		// Fields after the command name start at field 3 (state); utime and stime are 14 and 15.
+		std::istringstream fields(stat.substr(close + 2));
+		std::string        field;
+		for (int index = 3; index < 14 && fields >> field; index++) {
+		}
+		uint64_t user   = 0;
+		uint64_t system = 0;
+		if (!(fields >> user >> system)) {
+			continue;
+		}
+		auto& thread  = now.threads[id];
+		thread.cpu_ns = static_cast<uint64_t>(static_cast<double>(user + system) * ns_per_tick);
+		std::ifstream name_file(it->path() / "comm");
+		std::getline(name_file, thread.name);
+	}
+	if (!error) {
+		stats.thread_count = count;
+	}
+}
+
+void SamplePlatform(PlatformState& /*state*/, HostStats& stats, Counters& now) {
+	SampleRusage(now);
+
+	std::ifstream cpu_file("/proc/stat");
+	std::string   label;
+	if (cpu_file >> label && label == "cpu") {
+		uint64_t total = 0;
+		uint64_t idle  = 0;
+		uint64_t value = 0;
+		for (int index = 0; index < 8 && cpu_file >> value; index++) {
+			total += value;
+			if (index == 3 || index == 4) { // idle, iowait
+				idle += value;
+			}
+		}
+		now.system_total = total;
+		now.system_busy  = total - idle;
+	}
+
+	const auto status         = ReadProcValues("/proc/self/status");
+	stats.process_memory      = Find(status, "VmRSS");
+	stats.process_resident    = Find(status, "VmRSS");
+	stats.process_file_backed = Find(status, "RssFile");
+	stats.process_compressed  = Find(status, "VmSwap");
+	stats.process_virtual     = Find(status, "VmSize");
+
+	const auto io  = ReadProcValues("/proc/self/io");
+	now.disk_read  = Find(io, "read_bytes");
+	now.disk_write = Find(io, "write_bytes");
+
+	const auto memory    = ReadProcValues("/proc/meminfo");
+	const auto total     = Find(memory, "MemTotal");
+	const auto available = Find(memory, "MemAvailable");
+	if (total && available && *available <= *total) {
+		stats.system_memory_total = total;
+		stats.system_memory_used  = *total - *available;
+	}
+	stats.system_free = Find(memory, "MemFree");
+	if (const auto cached = Find(memory, "Cached")) {
+		stats.system_cached = *cached + Find(memory, "Buffers").value_or(0);
+	}
+	const auto swap_total = Find(memory, "SwapTotal");
+	const auto swap_free  = Find(memory, "SwapFree");
+	if (swap_total && swap_free && *swap_free <= *swap_total) {
+		stats.swap_total = swap_total;
+		stats.swap_used  = *swap_total - *swap_free;
+	}
+
+	// Pressure stall information: share of the last 10 s some task waited on memory.
+	std::ifstream pressure("/proc/pressure/memory");
+	std::string   kind;
+	std::string   average;
+	double        stalled = 0.0;
+	if (pressure >> kind >> average && kind == "some" && average.rfind("avg10=", 0) == 0 &&
+	    std::istringstream(average.substr(6)) >> stalled) {
+		stats.memory_pressure = stalled >= 20.0  ? MemoryPressure::Critical
+		                        : stalled >= 5.0 ? MemoryPressure::Warning
+		                                         : MemoryPressure::Normal;
+	} else if (total && available && *total != 0) {
+		const double free_share = static_cast<double>(*available) / static_cast<double>(*total);
+		stats.memory_pressure   = free_share < 0.10   ? MemoryPressure::Critical
+		                          : free_share < 0.25 ? MemoryPressure::Warning
+		                                              : MemoryPressure::Normal;
+	}
+
+	SampleThreads(stats, now);
+}
+
+#endif
+#endif
+
+constexpr size_t BUSIEST_THREADS = 6;
+
+} // namespace
+
+struct HostStatsSampler::Impl {
+	PlatformState state;
+	Counters      last;
+	uint32_t      logical_cores = std::max(std::thread::hardware_concurrency(), 1u);
+	bool          has_last      = false;
+};
+
+HostStatsSampler::HostStatsSampler(): m_impl(std::make_unique<Impl>()) {}
+
+HostStatsSampler::~HostStatsSampler() = default;
+
+HostStats HostStatsSampler::Sample() {
+	HostStats stats;
+	stats.logical_cores = m_impl->logical_cores;
+
+	Counters now;
+	now.wall_ns = WallNanoseconds();
+	SamplePlatform(m_impl->state, stats, now);
+
+	const auto& last    = m_impl->last;
+	const auto  elapsed = static_cast<double>(now.wall_ns - last.wall_ns) / 1e9;
+	if (m_impl->has_last && elapsed > 0.0) {
+		const auto rate = [&](std::optional<uint64_t> current,
+		                      std::optional<uint64_t> previous) -> std::optional<double> {
+			if (!current || !previous || *current < *previous) {
+				return std::nullopt;
+			}
+			return static_cast<double>(*current - *previous) / elapsed;
+		};
+		if (const auto cpu = rate(now.process_cpu_ns, last.process_cpu_ns)) {
+			stats.process_cpu_percent =
+			    std::clamp(*cpu / 1e9 * 100.0 / stats.logical_cores, 0.0, 100.0);
+		}
+		if (now.system_total && last.system_total && now.system_busy && last.system_busy &&
+		    *now.system_total > *last.system_total && *now.system_busy >= *last.system_busy) {
+			stats.system_cpu_percent =
+			    std::clamp(static_cast<double>(*now.system_busy - *last.system_busy) /
+			                   static_cast<double>(*now.system_total - *last.system_total) * 100.0,
+			               0.0, 100.0);
+		}
+		stats.context_switches_per_s = rate(now.context_switches, last.context_switches);
+		stats.page_faults_per_s      = rate(now.page_faults, last.page_faults);
+		stats.page_ins_per_s         = rate(now.page_ins, last.page_ins);
+		stats.disk_read_per_s        = rate(now.disk_read, last.disk_read);
+		stats.disk_write_per_s       = rate(now.disk_write, last.disk_write);
+
+		for (const auto& [id, thread]: now.threads) {
+			// A thread missing from the last sample started during this interval.
+			const auto     previous = last.threads.find(id);
+			const uint64_t before   = previous != last.threads.end() ? previous->second.cpu_ns : 0;
+			if (thread.cpu_ns < before) {
+				continue;
+			}
+			const double percent =
+			    static_cast<double>(thread.cpu_ns - before) / 1e9 / elapsed * 100.0;
+			stats.busiest_threads.push_back(
+			    {thread.name.empty() ? "thread " + std::to_string(id) : thread.name, percent});
+		}
+		std::sort(stats.busiest_threads.begin(), stats.busiest_threads.end(),
+		          [](const auto& a, const auto& b) { return a.cpu_percent > b.cpu_percent; });
+		if (stats.busiest_threads.size() > BUSIEST_THREADS) {
+			stats.busiest_threads.resize(BUSIEST_THREADS);
+		}
+	}
+	m_impl->last     = std::move(now);
+	m_impl->has_last = true;
+	return stats;
+}
+
+} // namespace Common
@@ -0,0 +1,86 @@
+#ifndef KYTY_COMMON_HOST_STATS_H_
+#define KYTY_COMMON_HOST_STATS_H_
+
+#include <cstdint>
+#include <memory>
+#include <optional>
+#include <string>
+#include <vector>
+
+namespace Common {
+
+enum class MemoryPressure { Unknown, Normal, Warning, Critical };
+enum class ThermalPressure { Unknown, Nominal, Moderate, Heavy, Critical };
+
+struct ThreadUsage {
+	std::string name;
+	double      cpu_percent = 0.0; // Of one core.
+};
+
+// Host-wide and process-wide resource usage. Missing values are std::nullopt; rates are per
+// second since the previous sample.
+struct HostStats {
+	uint32_t logical_cores = 0;
+
+	// Processing.
+	std::optional<double>    process_cpu_percent; // Share of all logical cores, 0-100.
+	std::optional<double>    system_cpu_percent;  // Whole machine, 0-100.
+	std::optional<uint32_t>  thread_count;
+	std::vector<ThreadUsage> busiest_threads; // Busiest first.
+	std::optional<double>    context_switches_per_s;
+	std::optional<double>    page_faults_per_s;
+	std::optional<double>    page_ins_per_s; // Faults that read from disk.
+	ThermalPressure          thermal_pressure = ThermalPressure::Unknown;
+
+	// Emulator process memory.
+	std::optional<uint64_t> process_memory; // Physical footprint / resident set.
+	std::optional<uint64_t> process_resident;
+	std::optional<uint64_t> process_compressed; // Compressed or swapped out.
+	std::optional<uint64_t> process_file_backed;
+	std::optional<uint64_t> process_virtual; // Reserved address space.
+
+	// System memory.
+	std::optional<uint64_t> system_memory_used;
+	std::optional<uint64_t> system_memory_total;
+	std::optional<uint64_t> system_wired;
+	std::optional<uint64_t> system_compressed;
+	std::optional<uint64_t> system_cached; // File cache, reclaimable.
+	std::optional<uint64_t> system_free;
+	std::optional<uint64_t> swap_used;
+	std::optional<uint64_t> swap_total;
+	MemoryPressure          memory_pressure = MemoryPressure::Unknown;
+
+	// Disk I/O of the emulator process.
+	std::optional<double> disk_read_per_s;
+	std::optional<double> disk_write_per_s;
+
+	// GPU, whole device (macOS IOAccelerator).
+	std::string             gpu_model;
+	std::optional<uint32_t> gpu_cores;
+	std::optional<double>   gpu_percent;
+	std::optional<double>   gpu_renderer_percent;
+	std::optional<double>   gpu_tiler_percent;
+	std::optional<uint64_t> gpu_memory_in_use; // System memory the GPU driver has in use.
+	std::optional<uint64_t> gpu_memory_allocated;
+	std::optional<uint64_t> gpu_recoveries; // GPU resets since boot.
+};
+
+// May block for a few milliseconds (thread enumeration, OS registry reads), so call it off the
+// render thread. The first sample has no rates.
+class HostStatsSampler final {
+public:
+	HostStatsSampler();
+	~HostStatsSampler();
+	HostStatsSampler(const HostStatsSampler&)            = delete;
+	HostStatsSampler& operator=(const HostStatsSampler&) = delete;
+
+	[[nodiscard]] HostStats Sample();
+
+private:
+	struct Impl;
+	std::unique_ptr<Impl> m_impl;
+};
+
+} // namespace Common
+
+#endif /* KYTY_COMMON_HOST_STATS_H_ */
@@ -0,0 +1,53 @@
+#include "common/localization.h"
+
+#include <array>
+#include <cstddef>
+#include <nlohmann/json.hpp>
+#include <string>
+
+#include "localizationData.h"
+
+namespace Common::Localization {
+namespace {
+
+constexpr std::array<std::string_view, 30> Locales = {
+    "ja-JP", "en-US", "fr-FR",   "es-ES",   "de-DE", "it-IT", "nl-NL", "pt-PT",
+    "ru-RU", "ko-KR", "zh-Hant", "zh-Hans", "fi-FI", "sv-SE", "da-DK", "no-NO",
+    "pl-PL", "pt-BR", "en-GB",   "tr-TR",   "es-419", "ar-AE", "fr-CA", "cs-CZ",
+    "hu-HU", "el-GR", "ro-RO",   "th-TH",   "vi-VN", "id-ID"};
+
+size_t LanguageIndex(int console_language) {
+	return console_language >= 0 && console_language < static_cast<int>(Locales.size())
+	           ? static_cast<size_t>(console_language)
+	           : 1;
+}
+
+} // namespace
+
+std::string_view LocaleName(int console_language) {
+	return Locales[LanguageIndex(console_language)];
+}
+
+std::string_view Get(int console_language, std::string_view key) {
+	static const auto catalogs = [] {
+		std::array<nlohmann::json, Locales.size()> result;
+		for (size_t i = 0; i < Locales.size(); ++i) {
+			const auto bytes = EmbeddedCatalog(Locales[i]);
+			result[i] = nlohmann::json::parse(bytes.begin(), bytes.end(), nullptr, false);
+		}
+		return result;
+	}();
+	for (const auto index: std::array {LanguageIndex(console_language), size_t {1}}) {
+		const auto& catalog = catalogs[index];
+		const auto  found   = catalog.find(key);
+		if (found != catalog.end() && found->is_string()) {
+			const auto& text = found->get_ref<const std::string&>();
+			if (!text.empty()) {
+				return text;
+			}
+		}
+	}
+	return key;
+}
+
+} // namespace Common::Localization
@@ -0,0 +1,13 @@
+#pragma once
+
+#include <string_view>
+
+namespace Common::Localization {
+
+// Console language indices follow the system configuration; invalid indices use English.
+[[nodiscard]] std::string_view LocaleName(int console_language);
+
+// Translations have static lifetime. Missing translations use English, then the supplied key.
+[[nodiscard]] std::string_view Get(int console_language, std::string_view key);
+
+} // namespace Common::Localization
@@ -5,6 +5,7 @@
 #include "common/emulatorConfig.h"

 #include <cstdio>
+#include <cstdlib>
 #include <filesystem>
 #include <fmt/format.h>
 #include <memory>
@@ -45,8 +46,8 @@ std::shared_ptr<spdlog::logger> MakeFileLogger(std::string                  name
 		std::filesystem::create_directories(parent);
 	}

-	auto sink =
-	    std::make_shared<spdlog::sinks::basic_file_sink_mt>(path.native(), true);
+	const bool truncate = std::getenv(Log::RelaunchedVariable) == nullptr;
+	auto       sink = std::make_shared<spdlog::sinks::basic_file_sink_mt>(path.native(), truncate);
 	return MakeLogger(std::move(name), std::move(sink));
 }

@@ -21,6 +21,10 @@ struct Lifecycle {

 enum class Direction { Silent, Console, File };

+// Set by LoadExec for the program it starts, which continues the log file of the one it replaces
+// instead of starting it over.
+inline constexpr const char* RelaunchedVariable = "KYTY_RELAUNCHED";
+
 Direction GetDirection();
 bool      IsSilent();
 void      Write(std::string_view text);
@@ -367,7 +367,7 @@ bool SysFileDeleteFile(const std::filesystem::path& name) {

 bool SysFileFlush(sys_file_t& f) {
 	if (f.type == SYS_FILE_FILE && f.f != nullptr) {
-		return (fflush(f.f) == 0);
+		return fflush(f.f) == 0 && fsync(fileno(f.f)) == 0;
 	}

 	return false;
@@ -1,9 +1,9 @@
 #include "common/trophies.h"

 #include "common/file.h"
+#include "common/localization.h"

 #include <algorithm>
-#include <array>
 #include <charconv>
 #include <fmt/format.h>
 #include <nlohmann/json.hpp>
@@ -105,13 +105,7 @@ Package ParsePackage(std::span<const std::byte> data, int console_language) {
 	    !conf["defaultLanguage"].is_string()) {
 		return {};
 	}
-	static constexpr std::array<std::string_view, 30> locales = {
-	    "ja-JP", "en-US", "fr-FR",   "es-ES",   "de-DE",  "it-IT", "nl-NL", "pt-PT",
-	    "ru-RU", "ko-KR", "zh-Hant", "zh-Hans", "fi-FI",  "sv-SE", "da-DK", "no-NO",
-	    "pl-PL", "pt-BR", "en-GB",   "tr-TR",   "es-419", "ar-AE", "fr-CA", "cs-CZ",
-	    "hu-HU", "el-GR", "ro-RO",   "th-TH",   "vi-VN",  "id-ID"};
-	const auto locale =
-	    locales[console_language >= 0 && console_language < locales.size() ? console_language : 1];
+	const auto locale = Localization::LocaleName(console_language);
 	auto meta = ReadJson(files, fmt::format("tropmeta_{}.json", locale));
 	if (meta.is_null()) {
 		meta = ReadJson(files, fmt::format("tropmeta_{}.json",
@@ -0,0 +1,62 @@
+if(NOT DEFINED INPUTS OR NOT DEFINED OUTPUT)
+	message(FATAL_ERROR "embed_localization.cmake requires INPUTS and OUTPUT")
+endif()
+
+set(contents "// Generated from assets/localization; do not edit.\n#pragma once\n#include <span>\n#include <string_view>\n\nnamespace {\n")
+set(lookup "std::span<const unsigned char> EmbeddedCatalog(std::string_view locale) {\n")
+set(continuation "[89ab][0-9a-f]")
+string(CONCAT utf8_character
+	"[0-7][0-9a-f]|(c[2-9a-f]|d[0-9a-f])${continuation}"
+	"|e0[ab][0-9a-f]${continuation}|e[1-9abcef]${continuation}${continuation}"
+	"|ed[89][0-9a-f]${continuation}|f0[9ab][0-9a-f]${continuation}${continuation}"
+	"|f[123]${continuation}${continuation}${continuation}|f48[0-9a-f]${continuation}${continuation}")
+foreach(catalog IN LISTS INPUTS)
+	file(READ "${catalog}" bytes HEX)
+	# Validate UTF-8, excluding overlong sequences, surrogates, and values above U+10FFFF.
+	if(NOT bytes MATCHES "^(${utf8_character})*$")
+		message(FATAL_ERROR "Localization must be UTF-8: ${catalog}")
+	endif()
+	file(READ "${catalog}" json)
+	string(JSON type ERROR_VARIABLE error TYPE "${json}")
+	if(error OR NOT type STREQUAL "OBJECT")
+		message(FATAL_ERROR "Invalid localization object in ${catalog}: ${error}")
+	endif()
+	string(JSON count LENGTH "${json}")
+	if(count GREATER 0)
+		math(EXPR last "${count} - 1")
+		foreach(index RANGE ${last})
+			string(JSON key MEMBER "${json}" ${index})
+			string(JSON type TYPE "${json}" "${key}")
+			if(NOT type STREQUAL "STRING")
+				message(FATAL_ERROR "Localization value '${key}' must be a string: ${catalog}")
+			endif()
+			string(JSON text GET "${json}" "${key}")
+			if(key STREQUAL "" OR text STREQUAL "")
+				message(FATAL_ERROR "Localization keys and values must not be empty: ${catalog}")
+			endif()
+		endforeach()
+	endif()
+	get_filename_component(locale "${catalog}" NAME_WE)
+	if(locale STREQUAL "en-US")
+		string(JSON earned ERROR_VARIABLE error GET "${json}" "trophy.toast.earned")
+		if(error OR earned STREQUAL "")
+			message(FATAL_ERROR "English localization requires trophy.toast.earned: ${catalog}")
+		endif()
+		set(english_found TRUE)
+	endif()
+	string(MAKE_C_IDENTIFIER "${locale}" symbol)
+	string(REGEX REPLACE "(..)" "0x\\1," bytes "${bytes}")
+	string(APPEND contents "constexpr unsigned char catalog_${symbol}[] = {${bytes}};\n")
+	string(APPEND lookup "\tif (locale == \"${locale}\") return catalog_${symbol};\n")
+endforeach()
+if(NOT english_found)
+	message(FATAL_ERROR "Localization requires an en-US.json fallback")
+endif()
+string(APPEND contents "\n${lookup}\treturn {};\n}\n} // namespace\n")
+
+get_filename_component(output_dir "${OUTPUT}" DIRECTORY)
+file(MAKE_DIRECTORY "${output_dir}")
+file(WRITE "${OUTPUT}.tmp" "${contents}")
+execute_process(COMMAND "${CMAKE_COMMAND}" -E copy_if_different "${OUTPUT}.tmp" "${OUTPUT}"
+	COMMAND_ERROR_IS_FATAL ANY)
+file(REMOVE "${OUTPUT}.tmp")
@@ -89,42 +89,6 @@ static void MountSandboxDirs() {
 	MountOrCreateDir("_TempData/" + title_id, "/temp");
 }

-static bool ClearDirectoryContents(const std::filesystem::path& dir) {
-	bool ok = true;
-
-	for (const auto& entry: Common::File::GetDirEntries(dir)) {
-		if (entry.name == "." || entry.name == "..") {
-			continue;
-		}
-
-		auto path = dir / entry.name;
-
-		if (entry.is_file) {
-			Common::File::RemoveReadonly(path);
-			ok = Common::File::DeleteFile(path) && ok;
-		} else {
-			ok = ClearDirectoryContents(path) && ok;
-			ok = Common::File::DeleteDirectory(path) && ok;
-		}
-	}
-
-	return ok;
-}
-
-static void ClearDebugTextureFolder() {
-	const std::string debug_texture_folder = "_Textures";
-
-	if (!Common::File::IsDirectoryExisting(debug_texture_folder)) {
-		Common::File::CreateDirectories(debug_texture_folder);
-		return;
-	}
-
-	if (!ClearDirectoryContents(debug_texture_folder)) {
-		LOGF_COLOR(Log::Color::BrightYellow, "TextureDump: failed to completely clear %s\n",
-		           debug_texture_folder.c_str());
-	}
-}
-
 static void Init(const Config::ConfigOptions& cfg, const std::filesystem::path& param_json,
                 Common::Subsystems& subsystems) {
 	EXIT_IF(!Common::Thread::IsMainThread());
@@ -198,8 +162,6 @@ void Run(const RunOptions& options) {
 	Common::Subsystems subsystems(true);
 	Init(options.config, param_json, subsystems);

-	ClearDebugTextureFolder();
-
 	PrintSystemInfo();
 	std::string title_id;
 	Loader::SystemContentParamSfoGetString("TITLE_ID", &title_id);
@@ -1,6 +1,7 @@
 #include "graphics/guest_gpu/graphicsRun.h"

 #include "common/assert.h"
+#include "common/debugCounters.h"
 #include "common/emulatorConfig.h"
 #include "common/logging/log.h"
 #include "common/profiler.h"
@@ -342,12 +343,11 @@ void CommandProcessor::WriteData(uint32_t* dst, const uint32_t* src, uint32_t dw
 		return;
 	}

-	if (write_one_address) {
-		for (uint32_t i = 0; i < dw_num; i++) {
-			dst[0] = src[i];
-		}
-	} else {
-		memcpy(dst, src, static_cast<size_t>(dw_num) * sizeof(uint32_t));
+	// Completion labels can be reused immediately; memcpy may write a destination twice.
+	auto* output = static_cast<volatile uint32_t*>(dst);
+	for (uint32_t i = 0; i < dw_num; i++) {
+		*output = src[i];
+		output += !write_one_address;
 	}
 }

@@ -906,6 +906,7 @@ void CommandProcessor::DrawIndirectMulti(uint32_t data_offset, uint32_t max_coun
 	if (draw_count == 0) {
 		return;
 	}
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::IndirectDraws, draw_count);

 	const auto args_size = indexed ? sizeof(DrawIndexedIndirectArgs) : sizeof(DrawIndirectArgs);
 	EXIT_NOT_IMPLEMENTED(stride_in_bytes < args_size);
@@ -8,6 +8,7 @@

 #include <map>
 #include <mutex>
+#include <span>
 #include <tuple>
 #include <vector>
 #include <vk_mem_alloc.h>
@@ -99,6 +100,14 @@ struct GraphicContext {
 	[[nodiscard]] bool CanReportMemoryUsage() const noexcept { return memory_budget_ext_enabled; }
 	[[nodiscard]] uint64_t GetDeviceMemoryUsage() const;
 	[[nodiscard]] uint64_t GetTotalMemoryBudget() const;
+
+	struct HeapUsage {
+		uint64_t usage        = 0; // Whole process, as reported by VK_EXT_memory_budget.
+		uint64_t budget       = 0;
+		uint32_t allocations  = 0;
+		bool     device_local = false;
+	};
+	[[nodiscard]] uint32_t GetHeapUsage(std::span<HeapUsage, VK_MAX_MEMORY_HEAPS> heaps) const;
 	[[nodiscard]] bool     CreateImage(const vk::ImageCreateInfo& info, VulkanImage& image);
 	void                   DeleteImage(VulkanImage& image);

@@ -66,6 +66,18 @@ RegionManager* MemoryTracker::GetOrCreateRegion(uint64_t index) {
 	return ptr;
 }

+void MemoryTracker::QueueBdaSync(uint64_t vaddr, uint64_t size) {
+	const auto begin = Common::AlignDown(vaddr, TRACKER_PAGE_SIZE);
+	const auto end   = Common::AlignUp(vaddr + size, TRACKER_PAGE_SIZE);
+	std::lock_guard lock(m_bda_sync_mutex);
+	m_bda_sync.Add(begin, end - begin);
+}
+
+RangeSet MemoryTracker::TakeBdaSync() {
+	std::lock_guard lock(m_bda_sync_mutex);
+	return std::exchange(m_bda_sync, RangeSet {});
+}
+
 bool MemoryTracker::IsRegionCpuModified(uint64_t vaddr, uint64_t size) {
 	CheckNotInUploadCallback();
 	return Iterate<true>(vaddr, size, [](RegionManager* manager, uint64_t offset, uint64_t bytes) {
@@ -88,6 +100,7 @@ void MemoryTracker::MarkRegionAsCpuModified(uint64_t vaddr, uint64_t size) {
 		std::scoped_lock lock(manager->lock);
 		manager->ChangeState<DirtySource::Cpu, true>(manager->GetCpuAddr() + offset, bytes);
 	});
+	QueueBdaSync(vaddr, size);
 }

 void MemoryTracker::MarkRegionAsGpuModified(uint64_t vaddr, uint64_t size) {
@@ -128,6 +141,8 @@ void MemoryTracker::UntrackMemory(uint64_t vaddr, uint64_t size) {
 	Iterate<false>(vaddr, size, [](RegionManager* manager, uint64_t offset, uint64_t bytes) {
 		manager->ChangeState<DirtySource::Cpu, true>(manager->GetCpuAddr() + offset, bytes);
 	});
+	locks.clear();
+	QueueBdaSync(vaddr, size);
 }

 } // namespace Libs::Graphics
@@ -29,6 +29,10 @@ public:
 	void               MarkRegionAsGpuModified(uint64_t vaddr, uint64_t size);
 	void               UnmarkRegionAsGpuModified(uint64_t vaddr, uint64_t size);
 	void               UntrackMemory(uint64_t vaddr, uint64_t size);
+	// Queue after changing ownership; a concurrent drain must not consume an event
+	// before the dirty pages it describes become visible.
+	void                   QueueBdaSync(uint64_t vaddr, uint64_t size);
+	[[nodiscard]] RangeSet TakeBdaSync();
 	// Removes protection from a range and flushes GPU-owned data when required.
 	template <typename Flush>
 	void InvalidateRegion(uint64_t vaddr, uint64_t size, Flush&& on_flush) noexcept {
@@ -51,6 +55,7 @@ public:
 				on_flush();
 			}
 		});
+		QueueBdaSync(vaddr, size);
 	}
 #if KYTY_BUILD == KYTY_BUILD_DEBUG
 	void ValidateGpuDirtyPages(const RangeSet& dirty, uint64_t vaddr, uint64_t size,
@@ -149,6 +154,8 @@ private:
 	std::unique_ptr<std::atomic<RegionManager*>[]> m_regions;
 	std::vector<std::unique_ptr<RegionManager>>    m_region_storage;
 	std::mutex                                     m_region_mutex;
+	std::mutex                                     m_bda_sync_mutex;
+	RangeSet                                       m_bda_sync;
 	PageManager&                                   m_page_manager;
 };

@@ -2,6 +2,7 @@

 #include "common/alignment.h"
 #include "common/assert.h"
+#include "common/debugCounters.h"
 #include "common/logging/log.h"
 #include "common/profiler.h"
 #include "graphics/guest_gpu/graphicsRun.h"
@@ -126,6 +127,7 @@ void BufferCache::ChangeRegister(BufferId id) {
 		}
 		WriteDataBuffer(m_bda_pagetable_buffer, table_offset,
 		                addresses.data(), addresses.size() * sizeof(vk::DeviceAddress));
+		m_memory_tracker.QueueBdaSync(buffer.CpuAddress(), buffer.Size());
 	} else {
 		const auto found = m_buffers.find(buffer.CpuAddress());
 		EXIT_IF(found == m_buffers.end() || found->second != id);
@@ -176,6 +178,7 @@ bool BufferCache::DownloadBufferMemory(Buffer& buffer, uint64_t vaddr, uint64_t
 	if (copies.empty()) {
 		return false;
 	}
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::BufferDownloadBytes, total_size);

 	auto [mapped, offset] = m_download_buffer.Map(total_size, 64);
 	std::unique_ptr<Buffer> temporary;
@@ -466,6 +469,7 @@ vk::Buffer BufferCache::UploadCopies(Buffer& buffer, std::span<vk::BufferCopy> c
 	if (copies.empty()) {
 		return nullptr;
 	}
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::BufferUploadBytes, total_size);

 	const auto source = UploadMemory(m_graphics, m_scheduler, m_staging_buffer, total_size, 4,
 	    vk::BufferUsageFlagBits::eTransferSrc, [&](uint8_t* mapped, uint64_t base_offset) {
@@ -657,6 +661,7 @@ void BufferCache::RunGarbageCollector() {
 		} else {
 			m_memory_tracker.UntrackMemory(buffer.CpuAddress(), buffer.Size());
 			DeleteBuffer(id);
+			Common::DebugCounters::Add(Common::DebugCounters::Counter::BufferEvictions);
 		}
 		return ++retire_count == limit;
 	});
@@ -685,6 +690,16 @@ void BufferCache::ProcessFaultBuffer() {
 	m_fault_manager.ProcessFaultBuffer();
 }

+void BufferCache::SynchronizeBda(const RangeSet& mapped) {
+	// Take before walking: writes queued during synchronization remain pending.
+	const auto pending = m_memory_tracker.TakeBdaSync();
+	pending.ForEach([&](uint64_t begin, uint64_t end) {
+		mapped.ForEachInRange(begin, end - begin, [this](uint64_t start, uint64_t finish) {
+			SynchronizeBuffersInRange(start, finish - start);
+		});
+	});
+}
+
 void BufferCache::SynchronizeBuffersInRange(uint64_t vaddr, uint64_t size) {
 	const auto end = vaddr + size;
 	auto       it  = m_buffers.upper_bound(vaddr);
@@ -80,7 +80,10 @@ public:
 	[[nodiscard]] bool IsRegionCpuModified(uint64_t vaddr, uint64_t size);
 	[[nodiscard]] bool IsRegionGpuModified(uint64_t vaddr, uint64_t size);
 	void               ProcessFaultBuffer();
-	void               SynchronizeBuffersInRange(uint64_t vaddr, uint64_t size);
+	void               SynchronizeBda(const RangeSet& mapped);
+	void               QueueBdaSync(uint64_t vaddr, uint64_t size) {
+		m_memory_tracker.QueueBdaSync(vaddr, size);
+	}
 	void               RunGarbageCollector();

 private:
@@ -114,6 +117,7 @@ private:
 	void DeleteBuffer(BufferId id);
 	[[nodiscard]] bool SynchronizeBuffer(Buffer& buffer, uint64_t vaddr, uint64_t size,
 	                                     bool is_written, bool is_texel_buffer);
+	void SynchronizeBuffersInRange(uint64_t vaddr, uint64_t size);
 	[[nodiscard]] vk::Buffer UploadCopies(Buffer& buffer, std::span<vk::BufferCopy> copies,
 	                                      uint64_t total_size);
 	[[nodiscard]] bool SynchronizeBufferFromImage(Buffer& buffer, uint64_t vaddr, uint64_t size);
@@ -67,11 +67,9 @@ Buffer::Buffer(GraphicContext& graphics, CommandScheduler& scheduler, MemoryUsag
 	buffer_info.usage       = flags;

 	const bool with_bda = bool(flags & vk::BufferUsageFlagBits::eShaderDeviceAddress);
-	const VmaAllocationCreateFlags bda_flag =
-	    with_bda ? VMA_ALLOCATION_CREATE_DEDICATED_MEMORY_BIT : 0;
 	VmaAllocationCreateInfo allocation_info {};
 	allocation_info.flags =
-	    VMA_ALLOCATION_CREATE_WITHIN_BUDGET_BIT | bda_flag | AllocationFlags(usage);
+	    VMA_ALLOCATION_CREATE_WITHIN_BUDGET_BIT | AllocationFlags(usage);
 	allocation_info.usage = AllocationUsage(usage);
 	allocation_info.preferredFlags = usage == MemoryUsage::DeviceLocal
 	                                     ? VkMemoryPropertyFlags {}
@@ -2,6 +2,7 @@

 #include "common/alignment.h"
 #include "common/assert.h"
+#include "common/debugCounters.h"
 #include "common/emulatorConfig.h"
 #include "common/logging/log.h"
 #include "common/profiler.h"
@@ -262,14 +263,8 @@ void TextureCache::DeleteImage(ImageId id) {
 		EXIT("TextureCache: deleting a GPU-modified image without resolving its contents\n");
 	}
 	m_download_images.erase(id);
-	if (image->info.HasMetadata()) {
-		const auto metadata = m_surface_metas.find(image->info.metadata.range.address);
-		if (metadata != m_surface_metas.end() &&
-		    image->info.metadata.kind == ImageMetadataKind::Htile &&
-		    metadata->second.type == MetaDataInfo::Type::HTile) {
-			// A later binding may have reused this address for another metadata type.
-			m_surface_metas.erase(metadata);
-		}
+	if (image->info.metadata.kind == ImageMetadataKind::Htile) {
+		m_surface_metas.erase(image->info.metadata.range.address);
 	}
 	UnregisterImage(id);
 	if (m_scheduler.Active()) {
@@ -954,6 +949,7 @@ void TextureCache::UploadImage(Image& image, Buffer& source, uint64_t source_off
 		const auto linear = m_tiler.Detile(source.Handle(), source_offset, info);
 		const auto copies = info.BufferCopies(linear.offset);
 		destination.Upload(copies, linear.buffer, linear.offset, linear.size);
+		Common::DebugCounters::Add(Common::DebugCounters::Counter::TextureUploadBytes, linear.size);
 		return;
 	}

@@ -999,6 +995,7 @@ void TextureCache::UploadImage(Image& image, Buffer& source, uint64_t source_off
 		copies[0].imageExtent       = info.extent;
 	}
 	destination.Upload(copies, linear.buffer, linear.offset, linear.size);
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::TextureUploadBytes, linear.size);
 }

 void TextureCache::InitializeImage(ImageId id) {
@@ -1039,7 +1036,7 @@ void TextureCache::MaterializeColorClear(ImageId id, const ImageDesc& desc,
 		std::scoped_lock lock {m_lock};
 		auto& image         = m_slot_images[id];
 		image.info.metadata = desc.info.metadata;
-		// Native color metadata must not retain a reused HTile/CMask/FMask clear flag.
+		// Native color metadata must not retain a reused HTile clear flag.
 		m_surface_metas.erase(range.address);
 		if (range.size == 0 || desc.info.resources.levels != 1) {
 			return;
@@ -1417,11 +1414,8 @@ vk::ImageView TextureCache::FindDepthTarget(ImageId id, const ImageDesc& desc) {
 	image.info.metadata = desc.info.metadata;
 	if (desc.info.HasMetadata()) {
 		const auto [metadata, inserted] = m_surface_metas.try_emplace(desc.info.metadata.range.address);
-		if (inserted) {
-			metadata->second.type = MetaDataInfo::Type::HTile;
-			if (image.info.htile_initial_clear) {
-				metadata->second.clear_mask.set();
-			}
+		if (inserted && image.info.htile_initial_clear) {
+			metadata->second.clear_mask.set();
 		}
 	}
 	RefreshImage(id);
@@ -1774,6 +1768,7 @@ bool TextureCache::DownloadImageMemory(ImageId id) {
 	download.Flush(offset, range.size);

 	DownloadImage(image, download, offset, range.size, image.info.resources.levels);
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::TextureDownloadBytes, range.size);
 	vk::BufferMemoryBarrier barrier {};
 	barrier.srcAccessMask = vk::AccessFlagBits::eMemoryWrite | vk::AccessFlagBits::eTransferWrite |
 	                        vk::AccessFlagBits::eShaderWrite;
@@ -1943,6 +1938,7 @@ void TextureCache::RunGarbageCollector() {
 				}
 			}
 			FreeImage(id);
+			Common::DebugCounters::Add(Common::DebugCounters::Counter::TextureEvictions);
 			if (m_total_used_memory < m_critical_gc_memory && aggressive) {
 				deletions >>= 2;
 				aggressive = false;
@@ -81,9 +81,6 @@ public:

 private:
 	struct MetaDataInfo {
-		enum class Type : uint8_t { CMask, FMask, HTile };
-
-		Type type;
 		// PS5 DB_DEPTH_VIEW encodes 13-bit base and last array-slice indices.
 		std::bitset<8192> clear_mask;
 	};
@@ -72,7 +72,16 @@ void RenderExecutor::ResolveRenderColorTarget(CommandBuffer& buffer, RenderColor
 		return;
 	}
 	const auto samples = render_sample_count(rt.attrib.num_fragments);
-	if (samples == 0 || rt.attrib.num_samples != rt.attrib.num_fragments) {
+	// TODO(PPSA24156): Replace this 2x MSAA approximation with proper 4-coverage/2-fragment EQAA.
+	const bool approximate_eqaa = rt.attrib.num_samples == 2 && rt.attrib.num_fragments == 1;
+	if (approximate_eqaa) {
+		static std::atomic_bool logged = false;
+		if (!logged.exchange(true, std::memory_order_relaxed)) {
+			Log::WriteToConsoleAndLog(
+			    "Warning: approximating 4-coverage/2-fragment EQAA with 2x MSAA.\n");
+		}
+	}
+	if (samples == 0 || (rt.attrib.num_samples != rt.attrib.num_fragments && !approximate_eqaa)) {
 		EXIT("unsupported render-target sample configuration: samples=%u fragments=%u\n",
 		     rt.attrib.num_samples, rt.attrib.num_fragments);
 	}
@@ -220,8 +229,7 @@ void RenderExecutor::ResolveRenderColorTarget(CommandBuffer& buffer, RenderColor
 	}
 	const bool has_dcc        = rt.info.dcc_compression_enable && rt.dcc_addr.addr != 0;
 	const bool has_cmask      = !rt.info.dcc_compression_enable && rt.info.cmask_fast_clear_enable &&
-	                            rt.cmask.addr != 0 && samples == 1 &&
-	                            !rt.info.fmask_compression_enable &&
+	                            rt.cmask.addr != 0 &&
 	                            !rt.attrib3.write_vrs_rate_hint_to_cmask;
 	if (has_dcc || has_cmask) {
 		TileSizeAlign metadata_size {};
@@ -233,7 +241,7 @@ void RenderExecutor::ResolveRenderColorTarget(CommandBuffer& buffer, RenderColor
 		} else {
 			(void)TileGetCmaskSize(width, height, layers, levels, metadata_size);
 		}
-		// DCC owns the clear when both planes are enabled; single-sample CMASK stays expanded.
+		// DCC owns the clear when both planes are enabled.
 		desc.info.metadata.kind = has_dcc ? ImageMetadataKind::Dcc : ImageMetadataKind::Cmask;
 		desc.info.metadata.range = {has_dcc ? rt.dcc_addr.addr : rt.cmask.addr, metadata_size.size};
 		desc.info.metadata.clear_word =
@@ -1,6 +1,7 @@
 #include "graphics/host_gpu/renderer/commandScheduler.h"

 #include "common/assert.h"
+#include "common/debugCounters.h"
 #include "common/logging/log.h"
 #include "graphics/host_gpu/graphicContext.h"

@@ -373,6 +374,7 @@ uint64_t CommandScheduler::Submit(SubmitInfo submit) {

 		result = graphics.queue.submit(1, &submit_info, nullptr);
 	}
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::QueueSubmits);

 	if (result != vk::Result::eSuccess) {
 		ReportVulkanFatal("vkQueueSubmit", result, tick, m_command.m_debug_op,
@@ -294,16 +294,6 @@ static void RtCheck(const HW::RenderTarget& rt) {
 		EXIT_NOT_IMPLEMENTED(rt.dcc.overwrite_combiner_disable != false);
 		// EXIT_NOT_IMPLEMENTED(rt.dcc.data_write_on_dcc_clear_to_reg != false);
 		EXIT_NOT_IMPLEMENTED(rt.dcc.dcc_clear_key_enable != false);
-		if (rt.cmask.addr != 0x0000000000000000 || rt.fmask.addr != 0x0000000000000000 ||
-		    rt.dcc_addr.addr != 0x0000000000000000) {
-			static bool logged = false;
-			if (!logged) {
-				LOGF("RenderTarget: temporary: ignoring PS5 metadata addresses cmask=0x%016" PRIx64
-				     " fmask=0x%016" PRIx64 " dcc=0x%016" PRIx64 "\n",
-				     rt.cmask.addr, rt.fmask.addr, rt.dcc_addr.addr);
-				logged = true;
-			}
-		}
 	}
 }

@@ -311,7 +311,6 @@ void RenderExecutor::ResolveRenderDepthTarget(CommandBuffer& buffer, RenderDepth
 	}
 	r.desc = MakeDepthTargetDesc(buffer, z);
 	r.depth_clear_enable      = rc.depth_clear_enable;
-	r.depth_load_clear_enable = r.depth_clear_enable;
 	r.depth_clear_value       = hw.GetDepthClearValue();
 	r.depth_test_enable       = dc.z_enable;
 	r.depth_write_enable      = r.depth_test_enable && dc.z_write_enable &&
@@ -421,9 +420,9 @@ vk::ImageAspectFlags RenderDepthInfo::AttachmentWriteAspects() const {
 	}

 	const auto           available = ImageViewOps::DepthAspectMask(format);
-	vk::ImageAspectFlags writes {};
+	vk::ImageAspectFlags writes = load_clear_aspects & available;
 	if ((available & vk::ImageAspectFlagBits::eDepth) &&
-	    (depth_load_clear_enable || depth_write_enable)) {
+	    depth_write_enable) {
 		writes |= vk::ImageAspectFlagBits::eDepth;
 	}
 	if (!(available & vk::ImageAspectFlagBits::eStencil)) {
@@ -461,8 +460,7 @@ vk::ImageAspectFlags RenderDepthInfo::AttachmentWriteAspects() const {
 		       (can_pass && depth_pass && state.passOp != vk::StencilOp::eKeep) ||
 		       (can_pass && depth_fail && state.depthFailOp != vk::StencilOp::eKeep);
 	};
-	if (stencil_clear_enable ||
-	    (stencil_test_enable && (face_writes(stencil_front) || face_writes(stencil_back)))) {
+	if (stencil_test_enable && (face_writes(stencil_front) || face_writes(stencil_back))) {
 		writes |= vk::ImageAspectFlagBits::eStencil;
 	}
 	return writes;
@@ -22,7 +22,7 @@ struct RenderDepthInfo {
 	// Discovery keeps guest image information but can remap the view into a larger cache image.
 	TextureCache::ImageDesc     desc;
 	bool                        depth_clear_enable       = false;
-	bool                        depth_load_clear_enable  = false;
+	vk::ImageAspectFlags        load_clear_aspects;
 	float                       depth_clear_value        = 0.0f;
 	bool                        depth_test_enable        = false;
 	// Effective draw writes; discovery applies test, target-write and clear controls.
@@ -353,8 +353,12 @@ void Image::CopyImage(Image& source) {
 	std::vector<vk::ImageCopy> copies;
 	copies.reserve(levels);
 	for (uint32_t level = 0; level < levels; level++) {
-		const auto width  = std::max(source.backing.extent.width >> level, 1u);
-		const auto height = std::max(source.backing.extent.height >> level, 1u);
+		// Overlapping cache images can differ in size; copying the source extent into a smaller
+		// destination writes out of bounds and loses the device.
+		const auto width  = std::min(std::max(source.backing.extent.width >> level, 1u),
+		                             std::max(backing.extent.width >> level, 1u));
+		const auto height = std::min(std::max(source.backing.extent.height >> level, 1u),
+		                             std::max(backing.extent.height >> level, 1u));
 		const auto depth  = std::max(base_depth >> level, 1u);
 		const auto [source_layers, destination_layers] = SanitizeCopyLayers(source, *this, depth);
 		vk::ImageCopy copy {};
@@ -1,6 +1,7 @@
 #include "graphics/host_gpu/renderer/masterSemaphore.h"

 #include "common/assert.h"
+#include "common/debugCounters.h"
 #include "graphics/host_gpu/graphicContext.h"

 namespace Libs::Graphics {
@@ -49,6 +50,8 @@ void MasterSemaphore::Wait(uint64_t tick) {
 	wait_info.pSemaphores    = &m_semaphore;
 	wait_info.pValues        = &tick;

+	Common::DebugCounters::ScopedTimer timer(Common::DebugCounters::Counter::GpuWaits,
+	                                         Common::DebugCounters::Counter::GpuWaitNs);
 	const auto result = m_graphics.device.waitSemaphores(&wait_info, UINT64_MAX);
 	EXIT_NOT_IMPLEMENTED(result != vk::Result::eSuccess);
 	Refresh();
@@ -740,6 +740,12 @@ void RenderExecutor::FindBuffers(std::span<PreparedBindings* const> stages) {
 			const auto descriptor = DecodeNativeDescriptor<ShaderBufferResource>(snapshot.buffers[resource]);
 			const auto address = descriptor.Base48();
 			auto size = descriptor.GetSize();
+			if (size != 0 && descriptor.Stride() == 0 && descriptor.OutOfBounds() == 2 &&
+			    !metadata.scalar) {
+				// OOB_SELECT=2 only tests NumRecords != 0; a zero-stride fetch
+				// still needs its full payload when NumRecords is smaller than it.
+				size = std::max(size, uint64_t {metadata.max_byte_extent});
+			}
 			if (address == 0 || size == 0) {
 				source = {};
 				continue;
@@ -757,14 +763,18 @@ void RenderExecutor::FindBuffers(std::span<PreparedBindings* const> stages) {
 							}
 						}
 					}
-				} else {
-					const auto& graphics = m_context.GetGraphics();
-					const auto limit = uint64_t {graphics.GetPhysicalDeviceProperties().limits.maxStorageBufferRange} -
-					                   (graphics.StorageMinAlignment() - 1);
-					size = std::min({size, uint64_t {256} * 1024 * 1024, limit});
 				}
+				//  PPSA15246
+				const auto& graphics = m_context.GetGraphics();
+				const auto  limit =
+				    uint64_t {graphics.GetPhysicalDeviceProperties().limits.maxStorageBufferRange} -
+				    (graphics.StorageMinAlignment() - 1);
+				size = std::min({size, uint64_t {256} * 1024 * 1024, limit});
+			}
+			// Finite GPU descriptors can span sparse PRT holes; clamp only oversized/sentinel ranges.
+			if (size >= (uint64_t {1} << 30u) || descriptor.NumRecords() == UINT32_MAX) {
+				size = Libs::LibKernel::Memory::ClampRangeSize(address, size);
 			}
-			size = Libs::LibKernel::Memory::ClampRangeSize(address, size);
 			source = {address, size, cache.FindBuffer(address, size)};
 		}
 	}
@@ -1,6 +1,7 @@
 #include "graphics/host_gpu/renderer/pipeline/pipelineCache.h"

 #include "common/assert.h"
+#include "common/debugCounters.h"
 #include "common/emulatorConfig.h"
 #include "common/file.h"
 #include "common/logging/log.h"
@@ -18,13 +19,13 @@
 #include "graphics/shader/recompiler/ShaderRecompiler.h"
 #include "graphics/shader/shaderCompiler.h"
 #include "kernel/memory.h"
-#include "kytyGitVersion.h"
 #include "loader/systemContent.h"

 #include <algorithm>
 #include <array>
 #include <atomic>
 #include <cctype>
+#include <chrono>
 #include <cstdio>
 #include <cstring>
 #include <fmt/format.h>
@@ -81,7 +82,8 @@ std::string DriverCacheSignature(const vk::PhysicalDeviceProperties& properties)
 		uuid[i * 2]     = hex[properties.pipelineCacheUUID[i] >> 4u];
 		uuid[i * 2 + 1] = hex[properties.pipelineCacheUUID[i] & 0xfu];
 	}
-	return fmt::format("KytyPC1:{}:{:08x}:{:08x}:{:08x}:{}\n", KYTY_GIT_REVISION,
+	// Driver cache entries already distinguish shader code and pipeline state across builds.
+	return fmt::format("KytyPC2:{:08x}:{:08x}:{:08x}:{}\n",
 	                   properties.vendorID, properties.deviceID, properties.driverVersion, uuid);
 }

@@ -273,6 +275,7 @@ struct PipelineCache::ProgramCache {
 		ShaderRecompiler::IR::ResourceSnapshot       resources;
 		ShaderRecompiler::IR::ResourceSpecialization specialization;
 		std::vector<Permutation>                    permutations;
+		bool                                       skip_dispatch = false;
 	};

 	struct ProgramKeyHash {
@@ -299,6 +302,9 @@ struct PipelineCache::ProgramCache {
 	                               ShaderRecompiler::IR::ResourceSpecialization specialization,
 	                               uint32_t push_data_start_dword,
 	                               std::span<const uint32_t> function_code) {
+		Common::DebugCounters::ScopedTimer timer(Common::DebugCounters::Counter::ShadersCompiled,
+		                                         Common::DebugCounters::Counter::ShaderCompileNs,
+		                                         Common::DebugCounters::Peak::ShaderCompileNs);
 		auto result = ShaderRecompiler::CompileProgram(std::move(translated), options,
 		                                               specialization, push_data_start_dword);
 		if (!ValidateShaderSpirv(options.dump_label, options.shader_hash, result.spirv,
@@ -355,6 +361,7 @@ struct PipelineCache::ProgramCache {
 		bool source_changed = false;
 		if (entry != programs.end()) {
 			auto& cached = entry->second;
+			if (cached.skip_dispatch) return {};
 			Permutation* permutation = nullptr;
 			if (cached.call_source) {
 				source_changed = cached.RefreshCallSource(runtime);
@@ -414,10 +421,18 @@ struct PipelineCache::ProgramCache {
 		DumpShaderOriginal(stage_name, options.shader_hash, params.code);
 		ShaderRecompiler::ShaderSource source;
 		if (entry == programs.end()) source = ShaderRecompiler::PrepareShaderSource(params.code, options);
+		if (source.unsupported_call_pc != UINT32_MAX) {
+			PipelineCacheLog("!!! TEMPORARY COMPUTE DISPATCH SKIP !!! shader=0x{:016x} pc=0x{:08x} "
+			                 "reason=unsupported GPU-selected callable target; rendering is incomplete",
+			                 params.hash, source.unsupported_call_pc);
+			programs.try_emplace(lookup_key, ShaderRecompiler::IR::ResourcePlan{}, nullptr)
+			    .first->second.skip_dispatch = true;
+			return {};
+		}
 		auto* current = entry == programs.end() ? &source : entry->second.call_source.get();
 		ShaderRecompiler::TranslateResult translated;
 		if (current != nullptr) {
-			if (current->call) {
+			if (!current->calls.empty()) {
 				if (entry == programs.end()) ShaderRecompiler::RefreshShaderSource(*current, runtime);
 				translated = ShaderRecompiler::TranslateProgram(current->linked->decoded, options);
 				translated.program.source_reads = current->reads;
@@ -428,7 +443,7 @@ struct PipelineCache::ProgramCache {
 			translated = ShaderRecompiler::TranslateProgram(params.code, options);
 		}
 		if (entry == programs.end()) {
-			auto retained = source.call ? std::make_unique<ShaderRecompiler::ShaderSource>(std::move(source)) : nullptr;
+			auto retained = !source.calls.empty() ? std::make_unique<ShaderRecompiler::ShaderSource>(std::move(source)) : nullptr;
 			entry = programs.try_emplace(lookup_key,
 			    ShaderRecompiler::IR::ExtractResourcePlan(translated.program), std::move(retained)).first;
 			EXIT_IF(!ShaderRecompiler::IR::MaterializeResources(
@@ -516,17 +531,6 @@ void PipelineCache::InitializeDriverCache() {
 		PipelineCacheLog("Vulkan pipeline cache: disabled (non-Release build)");
 		return;
 	}
-	const std::string_view git_hash     = KYTY_GIT_HASH;
-	const std::string_view git_revision = KYTY_GIT_REVISION;
-	if (git_hash == "unknown" || git_revision == "unknown") {
-		PipelineCacheLog("Vulkan pipeline cache: disabled (unknown git revision)");
-		return;
-	}
-	if (git_hash.ends_with("-dirty")) {
-		PipelineCacheLog("Vulkan pipeline cache: disabled (dirty build)");
-		return;
-	}
-
 	m_driver_cache_path     = std::filesystem::path("_PipelineCache") / (title_id + ".bin");
 	const auto path         = Common::PathToString(m_driver_cache_path);
 	const bool cache_exists = Common::File::IsFileExisting(m_driver_cache_path);
@@ -559,7 +563,7 @@ void PipelineCache::InitializeDriverCache() {
 			    XXH3_64bits(initial_data.data(), initial_data.size()) != payload_hash) {
 				initial_data.clear();
 				PipelineCacheLog(
-				    "Vulkan pipeline cache: invalidating {} (driver, emulator, or data mismatch)",
+				    "Vulkan pipeline cache: invalidating {} (driver or data mismatch)",
 				    path);
 			}
 		} else {
@@ -586,8 +590,8 @@ void PipelineCache::InitializeDriverCache() {
 		return;
 	}
 	if (!initial_data.empty()) {
-		PipelineCacheLog("Vulkan pipeline cache: loaded {} bytes from {}", initial_data.size(),
-		                 path);
+		PipelineCacheLog("Vulkan pipeline cache: loaded {:.2f} MB from {}",
+		                 static_cast<double>(initial_data.size()) / 1'000'000.0, path);
 	} else {
 		PipelineCacheLog("Vulkan pipeline cache: initialized empty");
 	}
@@ -627,8 +631,8 @@ void PipelineCache::Save() {
 	}
 	if (result != vk::Result::eSuccess || size == 0 ||
 	    size > std::numeric_limits<uint32_t>::max()) {
-		PipelineCacheLog("Vulkan pipeline cache: save failed ({}, {} bytes)",
-		                 vk::to_string(result), size);
+		PipelineCacheLog("Vulkan pipeline cache: save failed ({}, {:.2f} MB)",
+		                 vk::to_string(result), static_cast<double>(size) / 1'000'000.0);
 		return;
 	}
 	payload.resize(size);
@@ -660,7 +664,8 @@ void PipelineCache::Save() {
 		                 Common::PathToString(m_driver_cache_path));
 		return;
 	}
-	PipelineCacheLog("Vulkan pipeline cache: saved {} bytes to {}", payload.size(),
+	PipelineCacheLog("Vulkan pipeline cache: saved {:.2f} MB to {}",
+	                 static_cast<double>(payload.size()) / 1'000'000.0,
 	                 Common::PathToString(m_driver_cache_path));
 	m_driver_cache_dirty = false;
 }
@@ -1004,8 +1009,13 @@ PipelineCache::Pipeline& PipelineCache::GetGraphicsPipeline(

 	auto cached = std::make_unique<Pipeline>();
 	LogPipelineTrace("CreatePipelineInternal begin", vs_id, ps_id);
-	CreatePipelineInternal(m_graphics, *cached, rendering, key.vertex_input, vertex_info,
-	                       ps_input_info, programs, static_params, m_driver_cache);
+	{
+		Common::DebugCounters::ScopedTimer timer(Common::DebugCounters::Counter::PipelinesCreated,
+		                                         Common::DebugCounters::Counter::PipelineCreateNs,
+		                                         Common::DebugCounters::Peak::PipelineCreateNs);
+		CreatePipelineInternal(m_graphics, *cached, rendering, key.vertex_input, vertex_info,
+		                       ps_input_info, programs, static_params, m_driver_cache);
+	}
 	LogPipelineTrace("CreatePipelineInternal done", vs_id, ps_id);

 	EXIT_NOT_IMPLEMENTED(cached->pipeline == nullptr);
@@ -1035,7 +1045,13 @@ PipelineCache::GetComputePipeline(const ShaderComputeInputInfo& input_info,
 	}

 	auto cached = std::make_unique<Pipeline>();
-	CreatePipelineInternal(m_graphics, *cached, input_info, compute_program.module, m_driver_cache);
+	{
+		Common::DebugCounters::ScopedTimer timer(Common::DebugCounters::Counter::PipelinesCreated,
+		                                         Common::DebugCounters::Counter::PipelineCreateNs,
+		                                         Common::DebugCounters::Peak::PipelineCreateNs);
+		CreatePipelineInternal(m_graphics, *cached, input_info, compute_program.module,
+		                       m_driver_cache);
+	}

 	EXIT_NOT_IMPLEMENTED(cached->pipeline == nullptr);
 	EXIT_NOT_IMPLEMENTED(cached->pipeline_layout == nullptr);
@@ -196,21 +196,27 @@ static void CreateDescriptorLayout(GraphicContext& graphics, PipelineCache::Pipe
 	const auto& limits = graphics.GetPhysicalDeviceProperties().limits;
 	uint64_t descriptor_count = 0;
 	uint64_t storage_buffers = 0;
-	std::array<uint64_t, 32> stage_buffers {}, stage_resources {};
+	uint64_t samplers = 0;
+	std::array<uint64_t, 32> stage_buffers {}, stage_resources {}, stage_samplers {};
 	for (const auto& binding: bindings) {
 		descriptor_count += binding.descriptorCount;
 		const bool storage = binding.descriptorType == vk::DescriptorType::eStorageBuffer;
+		const bool sampler = binding.descriptorType == vk::DescriptorType::eSampler;
 		if (storage) storage_buffers += binding.descriptorCount;
+		if (sampler) samplers += binding.descriptorCount;
 		for (auto stages = static_cast<uint32_t>(binding.stageFlags); stages != 0; stages &= stages - 1u) {
 			const auto stage = std::countr_zero(stages);
 			if (storage) stage_buffers[stage] += binding.descriptorCount;
-			if (binding.descriptorType != vk::DescriptorType::eSampler)
+			if (sampler) stage_samplers[stage] += binding.descriptorCount;
+			else
 				stage_resources[stage] += binding.descriptorCount;
 			if (stage_buffers[stage] > limits.maxPerStageDescriptorStorageBuffers ||
-			    stage_resources[stage] > limits.maxPerStageResources) {
+			    stage_resources[stage] > limits.maxPerStageResources ||
+			    stage_samplers[stage] > limits.maxPerStageDescriptorSamplers) {
 				EXIT("native descriptor stage %u exceeds device limits: buffers=%" PRIu64
-				     "/%u resources=%" PRIu64 "/%u\n", stage, stage_buffers[stage],
-				     limits.maxPerStageDescriptorStorageBuffers, stage_resources[stage], limits.maxPerStageResources);
+				     "/%u resources=%" PRIu64 "/%u samplers=%" PRIu64 "/%u\n", stage, stage_buffers[stage],
+				     limits.maxPerStageDescriptorStorageBuffers, stage_resources[stage], limits.maxPerStageResources,
+				     stage_samplers[stage], limits.maxPerStageDescriptorSamplers);
 			}
 		}
 	}
@@ -218,6 +224,10 @@ static void CreateDescriptorLayout(GraphicContext& graphics, PipelineCache::Pipe
 		EXIT("native storage-buffer descriptors exceed device limit: %" PRIu64 "/%u\n",
 		     storage_buffers, limits.maxDescriptorSetStorageBuffers);
 	}
+	if (samplers > limits.maxDescriptorSetSamplers) {
+		EXIT("native sampler descriptors exceed device limit: %" PRIu64 "/%u\n",
+		     samplers, limits.maxDescriptorSetSamplers);
+	}
 	pipeline.uses_push_descriptors = descriptor_count <= graphics.max_push_descriptors;

 	vk::DescriptorSetLayoutCreateInfo create {};
@@ -626,6 +636,9 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 	EXIT_NOT_IMPLEMENTED(pipeline.pipeline_layout == nullptr);

 	vk::ComputePipelineCreateInfo info {};
+	if (input_info.stage.program->info.uses_gds) {
+		info.flags = vk::PipelineCreateFlagBits::eDispatchBase;
+	}
 	info.stage             = comp_shader_stage_info;
 	info.layout            = pipeline.pipeline_layout;
 	info.basePipelineIndex = -1;
@@ -201,11 +201,11 @@ private:
 	void                      BindImage(ImageId id, bool storage);
 	void                      BindRenderTarget(ImageId id);
 	void                      ResetBindings();
-	void                     UpdateComputeMetaClears(const ShaderComputeInputInfo& input,
-	                                                 const CommandBuffer&          buffer);
-	[[nodiscard]] bool TryConsumeComputeImageClear(const ShaderComputeInputInfo& input,
-	                                              CommandBuffer& command, uint32_t group_x,
-	                                              uint32_t group_y, uint32_t group_z, uint32_t mode);
+	[[nodiscard]] bool        TryConsumeComputeMetaClear(const ShaderComputeInputInfo& input,
+	                                                     const CommandBuffer&          buffer);
+	[[nodiscard]] bool TryConsumeComputeFill(const ShaderComputeInputInfo& input,
+	                                        CommandBuffer& command, uint32_t group_x,
+	                                        uint32_t group_y, uint32_t group_z, uint32_t mode);

 	RenderContext&                        m_context;
 	GraphicsBindings                     m_graphics_bindings;
@@ -1,5 +1,6 @@
 #include "common/assert.h"
 #include "common/common.h"
+#include "common/debugCounters.h"
 #include "common/emulatorConfig.h"
 #include "common/file.h"
 #include "common/logging/log.h"
@@ -50,8 +51,8 @@ static bool FillSourcesDisjoint(std::span<const ShaderRecompiler::IR::Descriptor
 	return true;
 }

-void RenderExecutor::UpdateComputeMetaClears(const ShaderComputeInputInfo& input,
-                                            const CommandBuffer&          buffer) {
+bool RenderExecutor::TryConsumeComputeMetaClear(const ShaderComputeInputInfo& input,
+                                                const CommandBuffer&          buffer) {
 	const auto& program   = *input.stage.program;
 	const auto& resources = *input.stage.resources;
 	if (resources.buffers.size() != program.info.buffers.size()) {
@@ -61,9 +62,10 @@ void RenderExecutor::UpdateComputeMetaClears(const ShaderComputeInputInfo& input
 	for (uint32_t i = 0; i < program.info.buffers.size(); i++) {
 		const auto& resource   = program.info.buffers[i];
 		const auto  descriptor = DecodeNativeDescriptor<ShaderBufferResource>(resources.buffers[i]);
-		// Reads and transforms of metadata do not establish a clear state.
+		// A metadata resource that is also read is not proven to be a full overwrite. Execute it
+		// conservatively instead of replacing the dispatch with a coarse full-surface clear.
 		if ((!resource.written || resource.read) && cache.IsMeta(descriptor.Base48())) {
-			return;
+			return false;
 		}
 	}

@@ -73,10 +75,13 @@ void RenderExecutor::UpdateComputeMetaClears(const ShaderComputeInputInfo& input
 			if (resource.written) {
 				const auto descriptor =
 				    DecodeNativeDescriptor<ShaderBufferResource>(resources.buffers[i]);
-				(void)cache.ClearMeta(descriptor.Base48());
+				if (cache.ClearMeta(descriptor.Base48())) {
+					return true;
+				}
 			}
 		}
 	}
+	return false;
 }

 bool ResolveComputeBufferFill(const ShaderComputeInputInfo& input, uint32_t group_x,
@@ -94,9 +99,10 @@ bool ResolveComputeBufferFill(const ShaderComputeInputInfo& input, uint32_t grou
 	constexpr std::array formats {
 	    Prospero::BufferFormat::k32UInt, Prospero::BufferFormat::k32_32UInt,
 	    Prospero::BufferFormat::k32_32_32UInt, Prospero::BufferFormat::k32_32_32_32UInt};
+	// Offset-or-swizzle bounds do not describe the element range proved by the shader.
 	if (descriptor.Stride() != element_size || descriptor.Format() != formats[fill.words - 1] ||
 	    descriptor.SwizzleEnabled() || descriptor.IndexStride() != 0 || descriptor.AddTid() ||
-	    descriptor.Base48() == 0) {
+	    descriptor.OutOfBounds() == 3 || descriptor.Base48() == 0) {
 		return false;
 	}
 	if (input.threads_num[0] == 0 || input.threads_num[0] != fill.group_stride[0] ||
@@ -108,7 +114,11 @@ bool ResolveComputeBufferFill(const ShaderComputeInputInfo& input, uint32_t grou
 	                                 ? group_x
 	                                 : static_cast<uint64_t>(group_x) * input.threads_num[0];
 	const auto     size        = descriptor.GetSize();
-	if (invocations != descriptor.NumRecords() || size == 0 || size > UINT32_MAX ||
+	const auto elements = descriptor.NumRecords();
+	const bool covered = fill.bounded_elements != 0
+	                         ? fill.bounded_elements == elements && invocations >= elements
+	                         : invocations == elements;
+	if (!covered || invocations > uint64_t {UINT32_MAX} + 1 || size == 0 || size > UINT32_MAX ||
 	    (input.dispatch_thread_dimensions &&
 	     (group_x % input.threads_num[0] != 0 || input.dispatch_threads_num[0] != group_x ||
 	      input.dispatch_threads_num[1] != 1 || input.dispatch_threads_num[2] != 1))) {
@@ -122,9 +132,9 @@ bool ResolveComputeBufferFill(const ShaderComputeInputInfo& input, uint32_t grou
 	return true;
 }

-bool RenderExecutor::TryConsumeComputeImageClear(const ShaderComputeInputInfo& input,
-                                                CommandBuffer& command, uint32_t group_x,
-                                                uint32_t group_y, uint32_t group_z, uint32_t mode) {
+bool RenderExecutor::TryConsumeComputeFill(const ShaderComputeInputInfo& input,
+                                          CommandBuffer& command, uint32_t group_x,
+                                          uint32_t group_y, uint32_t group_z, uint32_t mode) {
 	const auto& program   = *input.stage.program;
 	const auto& resources = *input.stage.resources;
 	const auto& fill      = resources.uniform_fill;
@@ -180,14 +190,26 @@ bool RenderExecutor::TryConsumeComputeImageClear(const ShaderComputeInputInfo& i
 	                              size)) {
 		return false;
 	}
-	if (!cache.ClearImageFromBuffer(command, descriptor.Base48(), size, packed_clear)) {
-		return false;
+	const bool image_clear =
+	    cache.ClearImageFromBuffer(command, descriptor.Base48(), size, packed_clear);
+	if (!image_clear) {
+		// CPU fills invalidate overlapping cached images. Preserve the native dispatch for
+		// partial or ambiguous image backings, including pooled metadata aliases.
+		if ((descriptor.Base48() & 3u) != 0 ||
+		    !m_context.IsMapped(descriptor.Base48(), size) ||
+		    cache.IsRegionRegistered(descriptor.Base48(), size)) {
+			return false;
+		}
+		// Metadata need not have an image backing yet. Keep proven fills in guest memory
+		// when CPU-owned, so color-clear discovery can read them without a GPU download.
+		m_context.GetBufferCache().FillBuffer(descriptor.Base48(), size, packed_clear, false);
 	}
 	static std::atomic<uint32_t> logged_clears {0};
 	if (logged_clears.fetch_add(1, std::memory_order_relaxed) < 32) {
-		LOGF("GraphicsRenderDispatchDirect: compute image clear shader=0x%016" PRIx64
+		LOGF("GraphicsRenderDispatchDirect: compute %s shader=0x%016" PRIx64
 		     " addr=0x%016" PRIx64 " size=0x%016" PRIx64 " value=0x%08" PRIx32 "\n",
-		     input.stage.program->shader_hash, descriptor.Base48(), size, packed_clear);
+		     image_clear ? "image clear" : "buffer fill", input.stage.program->shader_hash,
+		     descriptor.Base48(), size, packed_clear);
 	}
 	return true;
 }
@@ -252,10 +274,6 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 		return;
 	}

-	if (sh_ctx.GetCs().cs_regs.data_addr == 0) {
-		return;
-	}
-
 	constexpr uint32_t DISPATCH_INITIATOR_USE_THREAD_DIMENSIONS = 1u << 5u;
 	constexpr uint32_t DISPATCH_INITIATOR_BASE_BITS             = 0x41u;
 	constexpr uint32_t DISPATCH_INITIATOR_MODIFIER_BITS         = 0xa038u;
@@ -295,6 +313,7 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 	}
 	const auto compute_program =
 	    m_context.GetPipelineCache().GetComputeProgram(cs_regs, sh_regs, input_info);
+	if (!compute_program) return;
 	if (use_thread_dimensions) {
 		input_info.dispatch_threads_num[0]    = thread_group_x;
 		input_info.dispatch_threads_num[1]    = thread_group_y;
@@ -303,14 +322,12 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,

 	const auto& program   = *input_info.stage.program;
 	const auto& resources = *input_info.stage.resources;
-	if (resources.specialization_reads.empty()) {
-		// Cached metadata addresses can be reused as ordinary buffers; preserve their guest writes.
-		UpdateComputeMetaClears(input_info, buffer);
-		if (TryConsumeComputeImageClear(input_info, buffer, thread_group_x, thread_group_y,
-		                               thread_group_z, mode)) {
-			ResetBindings();
-			return;
-		}
+	if (resources.specialization_reads.empty() &&
+	    (TryConsumeComputeMetaClear(input_info, buffer) ||
+	     TryConsumeComputeFill(input_info, buffer, thread_group_x, thread_group_y,
+	                           thread_group_z, mode))) {
+		ResetBindings();
+		return;
 	}

 	if (use_thread_dimensions) {
@@ -430,7 +447,28 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 	}

 	vk_buffer.bindPipeline(vk::PipelineBindPoint::eCompute, pipeline.pipeline);
-	vk_buffer.dispatch(thread_group_x, thread_group_y, thread_group_z);
+	constexpr uint32_t DISPATCH_INITIATOR_ORDERED_APPEND_ENABLE = 1u << 3u;
+	const uint64_t     group_threads = uint64_t {std::max(input_info.threads_num[0], 1u)} *
+	                                   std::max(input_info.threads_num[1], 1u) *
+	                                   std::max(input_info.threads_num[2], 1u);
+	if ((mode & DISPATCH_INITIATOR_ORDERED_APPEND_ENABLE) != 0 && program.info.uses_gds &&
+	    group_threads <= input_info.wave_size && !input_info.lds_storage) {
+		// A single guest wave per group can run in launch order without a device-wide
+		// spinlock. Keep the guest workgroup IDs when splitting the dispatch.
+		for (uint32_t z = 0; z < thread_group_z; ++z) {
+			for (uint32_t y = 0; y < thread_group_y; ++y) {
+				for (uint32_t x = 0; x < thread_group_x; ++x) {
+					if ((x | y | z) != 0u) {
+						ShaderAccessBarrier(vk_buffer, vk::PipelineStageFlagBits::eComputeShader);
+					}
+					vk_buffer.dispatchBase(x, y, z, 1, 1, 1);
+				}
+			}
+		}
+	} else {
+		vk_buffer.dispatch(thread_group_x, thread_group_y, thread_group_z);
+	}
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::Dispatches);

 	// The removed host fence also ordered read-only dispatches before later writers.
 	ShaderAccessBarrier(vk_buffer, vk::PipelineStageFlagBits::eComputeShader);
@@ -461,6 +499,7 @@ void RenderExecutor::DispatchIndirect(uint64_t submit_id, CommandBuffer& buffer,
 	input_info.indirect_args_address = args_addr;
 	const auto compute_program = m_context.GetPipelineCache().GetComputeProgram(
 	    cs_regs, buffer.GetRegisters().GetShaderRegisters(), input_info);
+	if (!compute_program) return;
 	buffer.EndRendering();
 	auto& pipeline = m_context.GetPipelineCache().GetComputePipeline(input_info, compute_program);
 	auto& bindings = m_compute_bindings;
@@ -500,6 +539,7 @@ void RenderExecutor::DispatchIndirect(uint64_t submit_id, CommandBuffer& buffer,
 	                          1, &barrier, 0, nullptr, 0, nullptr);
 	vk_buffer.bindPipeline(vk::PipelineBindPoint::eCompute, pipeline.pipeline);
 	vk_buffer.dispatchIndirect(args_buffer->Handle(), args_offset);
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::Dispatches);
 	ShaderAccessBarrier(vk_buffer, vk::PipelineStageFlagBits::eComputeShader);
 	ResetBindings();
 }
@@ -90,6 +90,7 @@ bool RenderContext::IsMapped(uint64_t vaddr, uint64_t size) const noexcept {
 void RenderContext::MapMemory(uint64_t vaddr, uint64_t size) {
 	std::lock_guard lock(m_mapped_ranges_mutex);
 	m_mapped_ranges.Add(vaddr, size);
+	m_buffer_cache.QueueBdaSync(vaddr, size);
 }

 void RenderContext::UnmapMemory(uint64_t vaddr, uint64_t size) {
@@ -129,9 +130,7 @@ void RenderContext::PrepareBda() {
 		m_bda_logged = true;
 	}
 	std::shared_lock lock(m_mapped_ranges_mutex);
-	m_mapped_ranges.ForEach([this](uint64_t start, uint64_t end) {
-		m_buffer_cache.SynchronizeBuffersInRange(start, end - start);
-	});
+	m_buffer_cache.SynchronizeBda(m_mapped_ranges);
 	m_fault_process_pending = true;
 }

@@ -2,6 +2,7 @@

 #include "common/assert.h"
 #include "common/common.h"
+#include "common/debugCounters.h"
 #include "common/emulatorConfig.h"
 #include "common/file.h"
 #include "common/logging/log.h"
@@ -507,14 +508,15 @@ RenderState RenderExecutor::AcquireRenderTargets(CommandBuffer& buffer, RenderCo
 		}
 		const auto  image_view = cache.FindDepthTarget(depth.image_id, depth.desc);
 		const auto& metadata   = depth.desc.info.metadata;
-		if (metadata.kind == ImageMetadataKind::Htile && depth.depth_clear_enable &&
-		    !cache.ClearMeta(metadata.range.address)) {
-			EXIT("failed to acquire HTile metadata for a depth clear\n");
-		}
 		const bool meta_clear =
 		    metadata.kind == ImageMetadataKind::Htile &&
 		    cache.IsMetaCleared(metadata.range.address, depth.desc.view_info.base_layer);
-		depth.depth_load_clear_enable = depth.depth_clear_enable || meta_clear;
+		depth.load_clear_aspects = {};
+		if (depth.depth_clear_enable || meta_clear)
+			depth.load_clear_aspects |= vk::ImageAspectFlagBits::eDepth;
+		// Apply pending clears to compressed stencil as well as depth.
+		if (depth.stencil_clear_enable || (meta_clear && metadata.stencil_compressed))
+			depth.load_clear_aspects |= vk::ImageAspectFlagBits::eStencil;
 		if (meta_clear &&
 		    !cache.TouchMeta(metadata.range.address, depth.desc.view_info.base_layer, false)) {
 			EXIT("failed to consume HTile clear state\n");
@@ -536,7 +538,10 @@ RenderState RenderExecutor::AcquireRenderTargets(CommandBuffer& buffer, RenderCo
 			}
 		}
 		if (feedback_aspects && !m_context.GetGraphics().attachment_feedback_loop_enabled) {
-			EXIT("depth attachment feedback loop is not supported by the host\n");
+			static std::atomic_bool logged = false;
+			if (!logged.exchange(true, std::memory_order_relaxed)) {
+				std::printf("depth attachment feedback loop is not supported by the host\n");
+			}
 		}
 		auto layout = depth_attachment_layout(depth);
 		if (sampled_aspects & ~DepthReadableAspects(layout)) {
@@ -564,9 +569,11 @@ RenderState RenderExecutor::AcquireRenderTargets(CommandBuffer& buffer, RenderCo
 		attachment.clear_value[0] = std::bit_cast<uint32_t>(depth.depth_clear_value);
 		attachment.clear_value[1] = depth.stencil_clear_value;
 		attachment.has_depth      = static_cast<bool>(aspects & vk::ImageAspectFlagBits::eDepth);
-		attachment.depth_clear    = depth.depth_load_clear_enable;
+		attachment.depth_clear =
+		    static_cast<bool>(depth.load_clear_aspects & vk::ImageAspectFlagBits::eDepth);
 		attachment.has_stencil    = static_cast<bool>(aspects & vk::ImageAspectFlagBits::eStencil);
-		attachment.stencil_clear  = depth.stencil_clear_enable;
+		attachment.stencil_clear =
+		    static_cast<bool>(depth.load_clear_aspects & vk::ImageAspectFlagBits::eStencil);
 	}
 	if (color_count == 0 && !depth.image_id) {
 		const auto& limits = buffer.GetGraphics().GetPhysicalDeviceProperties().limits;
@@ -1057,7 +1064,7 @@ bool RenderExecutor::CopyDepthToColor(CommandBuffer& buffer, const DrawCallInfo&
 	                     state.ps_input_info.ps_sample_mask_export_enable);
 	vk::ImageAspectFlags feedback;
 	const auto rendering = AcquireRenderTargets(buffer, &color, 1, depth, feedback);
-	if (depth.depth_load_clear_enable) {
+	if (depth.load_clear_aspects) {
 		m_context.GetCommandScheduler().BeginRendering(rendering);
 	}
 	auto& cache = m_context.GetTextureCache();
@@ -1227,6 +1234,8 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 	} else {
 		EmitDrawPrimitives(ucfg, vk_buffer, draw, emit);
 	}
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::Draws);
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::DrawInstances, draw.instance_count);

 	if (!draw.IsIndexed()) {
 		SetDrawDebugPhase(buffer, submit_id, draw, 0x600u);
@@ -97,6 +97,24 @@ uint64_t GraphicContext::GetDeviceMemoryUsage() const {
 	return usage;
 }

+uint32_t GraphicContext::GetHeapUsage(std::span<HeapUsage, VK_MAX_MEMORY_HEAPS> heaps) const {
+	if (allocator == nullptr) {
+		return 0;
+	}
+	VmaBudget budgets[VK_MAX_MEMORY_HEAPS] {};
+	vmaGetHeapBudgets(allocator, budgets);
+	const auto count = physical_device_memory_properties.memoryHeapCount;
+	for (uint32_t heap = 0; heap < count; heap++) {
+		const auto& properties = physical_device_memory_properties.memoryHeaps[heap];
+		heaps[heap] = {.usage        = budgets[heap].usage,
+		               .budget       = budgets[heap].budget,
+		               .allocations  = budgets[heap].statistics.allocationCount,
+		               .device_local = static_cast<bool>(properties.flags &
+		                                                 vk::MemoryHeapFlagBits::eDeviceLocal)};
+	}
+	return count;
+}
+
 uint64_t GraphicContext::GetTotalMemoryBudget() const {
 	if (allocator == nullptr) {
 		return 0;
@@ -0,0 +1,747 @@
+#include "graphics/presentation/performanceOverlay.h"
+
+#include "common/debugCounters.h"
+#include "common/emulatorConfig.h"
+#include "common/hostStats.h"
+#include "graphics/host_gpu/graphicContext.h"
+#include "imgui.h"
+#include "kernel/memory.h"
+
+#include <algorithm>
+#include <array>
+#include <atomic>
+#include <chrono>
+#include <condition_variable>
+#include <cstdarg>
+#include <cstdio>
+#include <mutex>
+#include <optional>
+#include <stop_token>
+#include <thread>
+
+namespace Libs::Graphics {
+
+namespace {
+
+using Clock   = std::chrono::steady_clock;
+using Counter = Common::DebugCounters::Counter;
+using Peak    = Common::DebugCounters::Peak;
+
+constexpr auto   REFRESH_INTERVAL  = std::chrono::seconds(1);
+constexpr size_t FRAME_HISTORY     = 120;
+constexpr float  FRAME_GRAPH_MS    = 50.0f;
+constexpr float  FRAME_TIME_GOOD   = 17.5f;
+constexpr float  FRAME_TIME_MEDIUM = 34.0f;
+constexpr size_t COUNTER_COUNT     = static_cast<size_t>(Counter::Count);
+
+// Tables as wide as their columns; the default fills the window, which an auto-resizing
+// window then never shrinks.
+constexpr ImGuiTableFlags TABLE_FLAGS =
+    ImGuiTableFlags_SizingFixedFit | ImGuiTableFlags_NoHostExtendX;
+
+constexpr ImU32 GOOD    = IM_COL32(96, 220, 120, 255);
+constexpr ImU32 MEDIUM  = IM_COL32(230, 195, 60, 255);
+constexpr ImU32 BAD     = IM_COL32(235, 80, 70, 255);
+constexpr ImU32 NEUTRAL = IM_COL32(160, 160, 160, 255);
+
+struct FrameWindow {
+	std::array<float, FRAME_HISTORY> intervals_ms {};
+	size_t                           next         = 0;
+	size_t                           count        = 0;
+	uint32_t                         flips        = 0;
+	float                            max_ms       = 0.0f;
+	uint32_t                         guest_width  = 0;
+	uint32_t                         guest_height = 0;
+	Clock::time_point                last_flip    = {};
+	Clock::time_point                window_start = {};
+};
+
+// Shown values, recomputed once per refresh interval.
+struct Snapshot {
+	double                              fps          = 0.0;
+	double                              frame_ms     = 0.0;
+	double                              max_frame_ms = 0.0;
+	double                              low_fps      = 0.0; // From the 99th percentile.
+	double                              stalled_s    = 0.0;
+	uint32_t                            guest_width  = 0;
+	uint32_t                            guest_height = 0;
+	std::array<double, COUNTER_COUNT>   rates {};
+	std::array<uint64_t, COUNTER_COUNT> totals {};
+	uint64_t                            shader_compile_peak_ns  = 0;
+	uint64_t                            pipeline_create_peak_ns = 0;
+	Common::HostStats                   host;
+	std::optional<uint64_t>             device_memory;
+	uint64_t                            device_budget = 0;
+	std::array<GraphicContext::HeapUsage, VK_MAX_MEMORY_HEAPS> heaps {};
+	uint32_t                                                   heap_count = 0;
+	std::array<float, FRAME_HISTORY>                           history {};
+	size_t                                                     history_count = 0;
+	Clock::time_point                                          sample_time {};
+	LibKernel::Memory::DebugStats                              guest_memory;
+};
+
+std::mutex            g_frame_mutex;
+FrameWindow           g_frames;
+std::atomic<uint64_t> g_generation {0};
+std::atomic<bool>     g_details {false};
+std::atomic<uint64_t> g_revision {0};
+
+std::mutex                  g_host_mutex;
+Common::HostStats           g_host_stats;
+uint64_t                    g_host_generation = 0;
+std::condition_variable_any g_sampler_wait;
+std::jthread                g_sampler;
+
+void SetOverlayEnabled(bool enabled) {
+	{
+		std::scoped_lock lock(g_host_mutex);
+		if (enabled) {
+			std::scoped_lock frame_lock(g_frame_mutex);
+			g_frames = {};
+			g_generation.fetch_add(1, std::memory_order_release);
+		}
+		Common::DebugCounters::SetEnabled(enabled);
+	}
+	g_sampler_wait.notify_all();
+}
+
+bool GuestStalled(Clock::time_point now) {
+	std::scoped_lock lock(g_frame_mutex);
+	return now - g_frames.last_flip >= REFRESH_INTERVAL;
+}
+
+// Host statistics can block for milliseconds (IOKit, thread enumeration), so sample them off
+// the present path. While the guest stops flipping (loading screens), ask for one repaint per
+// interval.
+void SamplerThread(const std::stop_token& stop) {
+	std::optional<Common::HostStatsSampler> sampler;
+	uint64_t                                sampled_generation = 0;
+	std::unique_lock                        lock(g_host_mutex);
+	while (!stop.stop_requested()) {
+		g_sampler_wait.wait(lock, stop, [] { return Common::DebugCounters::Enabled(); });
+		if (stop.stop_requested()) {
+			break;
+		}
+		const auto generation = g_generation.load(std::memory_order_acquire);
+		lock.unlock();
+		if (!sampler || generation != sampled_generation) {
+			sampler.emplace();
+			sampled_generation = generation;
+		}
+		auto stats = sampler->Sample();
+		lock.lock();
+		g_host_stats      = std::move(stats);
+		g_host_generation = generation;
+		if (Common::DebugCounters::Enabled() && GuestStalled(Clock::now())) {
+			g_revision.fetch_add(1, std::memory_order_acq_rel);
+		}
+		g_sampler_wait.wait_for(lock, stop, REFRESH_INTERVAL, [generation] {
+			return !Common::DebugCounters::Enabled() ||
+			       generation != g_generation.load(std::memory_order_acquire);
+		});
+	}
+}
+
+void EnsureSampler() {
+	static std::once_flag once;
+	std::call_once(once, [] { g_sampler = std::jthread(SamplerThread); });
+}
+
+ImU32 FrameTimeColor(double ms) {
+	return ms <= FRAME_TIME_GOOD ? GOOD : ms <= FRAME_TIME_MEDIUM ? MEDIUM : BAD;
+}
+
+// Green below `warn`, yellow below `bad`, red above.
+ImU32 LevelColor(double value, double warn, double bad) {
+	return value < warn ? GOOD : value < bad ? MEDIUM : BAD;
+}
+
+ImVec4 Color(ImU32 color) {
+	return ImGui::ColorConvertU32ToFloat4(color);
+}
+
+std::string Bytes(uint64_t bytes) {
+	constexpr double GIB = 1024.0 * 1024.0 * 1024.0;
+	constexpr double MIB = 1024.0 * 1024.0;
+	char             buffer[32];
+	if (static_cast<double>(bytes) >= GIB) {
+		std::snprintf(buffer, sizeof(buffer), "%.2f GB", static_cast<double>(bytes) / GIB);
+	} else {
+		std::snprintf(buffer, sizeof(buffer), "%.0f MB", static_cast<double>(bytes) / MIB);
+	}
+	return buffer;
+}
+
+std::string Bytes(std::optional<uint64_t> bytes) {
+	return bytes ? Bytes(*bytes) : "n/a";
+}
+
+std::string Throughput(double bytes_per_second) {
+	char buffer[32];
+	std::snprintf(buffer, sizeof(buffer), "%.1f MB/s", bytes_per_second / (1024.0 * 1024.0));
+	return buffer;
+}
+
+std::string Percent(std::optional<double> value) {
+	if (!value) {
+		return "n/a";
+	}
+	char buffer[16];
+	std::snprintf(buffer, sizeof(buffer), "%.0f%%", *value);
+	return buffer;
+}
+
+void Row(const char* label, const std::string& value, std::optional<ImU32> color = std::nullopt) {
+	ImGui::TableNextRow();
+	ImGui::TableNextColumn();
+	ImGui::TextDisabled("%s", label);
+	ImGui::TableNextColumn();
+	if (color) {
+		ImGui::TextColored(Color(*color), "%s", value.c_str());
+	} else {
+		ImGui::TextUnformatted(value.c_str());
+	}
+}
+
+[[gnu::format(printf, 1, 2)]] std::string Fmt(const char* format, ...) {
+	char    buffer[128];
+	va_list args;
+	va_start(args, format);
+	std::vsnprintf(buffer, sizeof(buffer), format, args);
+	va_end(args);
+	return buffer;
+}
+
+bool BeginSection(const char* title) {
+	ImGui::SeparatorText(title);
+	return ImGui::BeginTable(title, 2, TABLE_FLAGS);
+}
+
+void DrawFrameGraph(const std::array<float, FRAME_HISTORY>& history, size_t count, float width,
+                    float height) {
+	const ImVec2 origin = ImGui::GetCursorScreenPos();
+	auto*        draw   = ImGui::GetWindowDrawList();
+	draw->AddRectFilled(origin, {origin.x + width, origin.y + height}, IM_COL32(0, 0, 0, 90));
+	const auto y_of = [&](float ms) {
+		return origin.y + height - std::min(ms, FRAME_GRAPH_MS) / FRAME_GRAPH_MS * height;
+	};
+	for (const float reference: {1000.0f / 60.0f, 1000.0f / 30.0f}) {
+		draw->AddLine({origin.x, y_of(reference)}, {origin.x + width, y_of(reference)},
+		              IM_COL32(255, 255, 255, 45));
+	}
+	const float bar = width / static_cast<float>(FRAME_HISTORY);
+	for (size_t i = 0; i < count; i++) {
+		const float ms = history[i];
+		const float x  = origin.x + width - static_cast<float>(count - i) * bar;
+		draw->AddRectFilled({x, y_of(ms)}, {x + std::max(bar - 1.0f, 1.0f), origin.y + height},
+		                    FrameTimeColor(ms));
+	}
+	ImGui::Dummy({width, height});
+}
+
+const char* PressureName(Common::MemoryPressure pressure) {
+	switch (pressure) {
+		case Common::MemoryPressure::Normal: return "Normal";
+		case Common::MemoryPressure::Warning: return "Warning";
+		case Common::MemoryPressure::Critical: return "Critical";
+		case Common::MemoryPressure::Unknown: break;
+	}
+	return "n/a";
+}
+
+ImU32 PressureColor(Common::MemoryPressure pressure) {
+	switch (pressure) {
+		case Common::MemoryPressure::Normal: return GOOD;
+		case Common::MemoryPressure::Warning: return MEDIUM;
+		case Common::MemoryPressure::Critical: return BAD;
+		case Common::MemoryPressure::Unknown: break;
+	}
+	return NEUTRAL;
+}
+
+const char* ThermalName(Common::ThermalPressure thermal) {
+	switch (thermal) {
+		case Common::ThermalPressure::Nominal: return "Nominal";
+		case Common::ThermalPressure::Moderate: return "Moderate";
+		case Common::ThermalPressure::Heavy: return "Heavy (throttling)";
+		case Common::ThermalPressure::Critical: return "Critical (throttling)";
+		case Common::ThermalPressure::Unknown: break;
+	}
+	return "n/a";
+}
+
+ImU32 ThermalColor(Common::ThermalPressure thermal) {
+	switch (thermal) {
+		case Common::ThermalPressure::Nominal: return GOOD;
+		case Common::ThermalPressure::Moderate: return MEDIUM;
+		case Common::ThermalPressure::Heavy:
+		case Common::ThermalPressure::Critical: return BAD;
+		case Common::ThermalPressure::Unknown: break;
+	}
+	return NEUTRAL;
+}
+
+const char* PresentModeName(Config::PresentMode mode) {
+	switch (mode) {
+		case Config::PresentMode::Fifo: return "FIFO";
+		case Config::PresentMode::Mailbox: return "Mailbox";
+		case Config::PresentMode::Immediate: return "Immediate";
+	}
+	return "?";
+}
+
+double Rate(const Snapshot& shown, Counter counter) {
+	return shown.rates[static_cast<size_t>(counter)];
+}
+
+uint64_t Total(const Snapshot& shown, Counter counter) {
+	return shown.totals[static_cast<size_t>(counter)];
+}
+
+void TransferRow(const Snapshot& shown, const char* label, Counter upload, Counter download) {
+	Row(label,
+	    Throughput(Rate(shown, upload)) + " up / " + Throughput(Rate(shown, download)) + " down");
+}
+
+void GpuActivityRows(const Snapshot& shown) {
+	const auto draws = Rate(shown, Counter::Draws);
+	Row("Draw calls", Fmt("%.0f/frame  %.0f/s", draws / std::max(shown.fps, 1.0), draws));
+	// Includes present pacing waits, so a high value alone is not a stall.
+	Row("CPU waiting on GPU",
+	    Fmt("%.0f ms/s  (%.0f waits/s)", Rate(shown, Counter::GpuWaitNs) / 1e6,
+	        Rate(shown, Counter::GpuWaits)));
+}
+
+void CompilationRows(const Snapshot& shown) {
+	const auto row = [&](const char* label, Counter count_counter, Counter duration_counter,
+	                     uint64_t peak_ns) {
+		const auto count = Total(shown, count_counter);
+		Row(label,
+		    Fmt("%llu  (avg %.1f ms, peak %.1f ms)", static_cast<unsigned long long>(count),
+		        count != 0 ? static_cast<double>(Total(shown, duration_counter)) / 1e6 /
+		                         static_cast<double>(count)
+		                   : 0.0,
+		        static_cast<double>(peak_ns) / 1e6),
+		    Rate(shown, count_counter) > 0.0 ? MEDIUM : GOOD);
+	};
+	row("Shader recompiler", Counter::ShadersCompiled, Counter::ShaderCompileNs,
+	    shown.shader_compile_peak_ns);
+	row("Pipeline compile", Counter::PipelinesCreated, Counter::PipelineCreateNs,
+	    shown.pipeline_create_peak_ns);
+}
+
+void Refresh(Snapshot& shown, const GraphicContext& graphics, Clock::time_point now,
+             uint64_t generation) {
+	const double seconds = shown.sample_time == Clock::time_point {}
+	                           ? 0.0
+	                           : std::chrono::duration<double>(now - shown.sample_time).count();
+	for (size_t i = 0; i < COUNTER_COUNT; i++) {
+		const auto value = Common::DebugCounters::Get(static_cast<Counter>(i));
+		shown.rates[i] =
+		    seconds > 0.0 ? static_cast<double>(value - shown.totals[i]) / seconds : 0.0;
+		shown.totals[i] = value;
+	}
+	shown.shader_compile_peak_ns  = Common::DebugCounters::Get(Peak::ShaderCompileNs);
+	shown.pipeline_create_peak_ns = Common::DebugCounters::Get(Peak::PipelineCreateNs);
+	shown.sample_time             = now;
+	{
+		std::scoped_lock lock(g_host_mutex);
+		if (g_host_generation == generation) {
+			shown.host = g_host_stats;
+		}
+	}
+	shown.heap_count = graphics.GetHeapUsage(shown.heaps);
+	shown.device_memory =
+	    graphics.CanReportMemoryUsage() ? std::optional<uint64_t>(0) : std::nullopt;
+	shown.device_budget = 0;
+	const bool discrete =
+	    graphics.GetPhysicalDeviceProperties().deviceType == vk::PhysicalDeviceType::eDiscreteGpu;
+	for (uint32_t i = 0; i < shown.heap_count; i++) {
+		const auto& heap = shown.heaps[i];
+		if (shown.device_memory && (!discrete || heap.device_local)) {
+			*shown.device_memory += heap.usage;
+			shown.device_budget += heap.budget;
+		}
+	}
+	if (g_details.load(std::memory_order_acquire)) {
+		shown.guest_memory = LibKernel::Memory::GetDebugStats();
+	}
+}
+
+void DrawCompact(const Snapshot& shown) {
+	if (!ImGui::BeginTable("##performance", 2, TABLE_FLAGS)) {
+		return;
+	}
+	const auto& host = shown.host;
+	if (shown.max_frame_ms > 0.0) {
+		Row("Worst frame", Fmt("%.1f ms", shown.max_frame_ms), FrameTimeColor(shown.max_frame_ms));
+	}
+	Row("CPU", host.process_cpu_percent
+	               ? Fmt("%.0f%%  of %u threads", *host.process_cpu_percent, host.logical_cores)
+	               : std::string("n/a"));
+	Row("GPU", Percent(host.gpu_percent));
+	GpuActivityRows(shown);
+	CompilationRows(shown);
+	if (shown.device_memory) {
+		Row("GPU memory", Bytes(*shown.device_memory) + " / " + Bytes(shown.device_budget));
+	}
+	TransferRow(shown, "Texture transfers", Counter::TextureUploadBytes,
+	            Counter::TextureDownloadBytes);
+	TransferRow(shown, "Buffer transfers", Counter::BufferUploadBytes,
+	            Counter::BufferDownloadBytes);
+	Row("Emulator RAM", Bytes(host.process_memory));
+	if (host.system_memory_used && host.system_memory_total) {
+		Row("System RAM",
+		    Bytes(*host.system_memory_used) + " / " + Bytes(*host.system_memory_total));
+	}
+	ImGui::EndTable();
+}
+
+void DrawFramesAndCpu(const Snapshot& shown, vk::Extent2D extent) {
+	const auto& host = shown.host;
+	if (BeginSection("Frames")) {
+		if (shown.low_fps > 0.0) {
+			Row("1% low", Fmt("%.1f FPS", shown.low_fps), FrameTimeColor(1000.0 / shown.low_fps));
+		}
+		Row("Worst frame", Fmt("%.1f ms", shown.max_frame_ms), FrameTimeColor(shown.max_frame_ms));
+		Row("Guest output", shown.guest_width != 0
+		                        ? Fmt("%ux%u", shown.guest_width, shown.guest_height)
+		                        : std::string("n/a"));
+		Row("Window", Fmt("%ux%u  %s  %u Hz", extent.width, extent.height,
+		                  PresentModeName(Config::GetPresentMode()), Config::GetVblankFrequency()));
+		Row("Flips", Fmt("%.0f/s requested", Rate(shown, Counter::FlipsRequested)));
+		const auto rejected = Rate(shown, Counter::FlipsRejected);
+		Row("Flip queue full", Fmt("%.0f/s", rejected), rejected > 0.0 ? BAD : GOOD);
+		const auto late = Rate(shown, Counter::LateVblanks);
+		Row("Late vblanks", Fmt("%.0f/s", late), LevelColor(late, 1.0, 10.0));
+		Row("Swapchain rebuilds",
+		    Fmt("%llu  (dropped %llu)",
+		        static_cast<unsigned long long>(Total(shown, Counter::SwapchainRecreations)),
+		        static_cast<unsigned long long>(Total(shown, Counter::DroppedPresents))));
+		ImGui::EndTable();
+	}
+	if (BeginSection("CPU")) {
+		Row("Emulator",
+		    host.process_cpu_percent
+		        ? Fmt("%.0f%% of %u threads  (%.1f cores)", *host.process_cpu_percent,
+		              host.logical_cores, *host.process_cpu_percent * host.logical_cores / 100.0)
+		        : std::string("n/a"));
+		Row("Machine", Percent(host.system_cpu_percent),
+		    host.system_cpu_percent ? std::optional(LevelColor(*host.system_cpu_percent, 70, 90))
+		                            : std::nullopt);
+		if (host.thread_count) {
+			Row("Host threads", Fmt("%u", *host.thread_count));
+		}
+		for (const auto& thread: host.busiest_threads) {
+			Row(Fmt("  %.24s", thread.name.c_str()).c_str(), Fmt("%.0f%%", thread.cpu_percent),
+			    LevelColor(thread.cpu_percent, 70, 95));
+		}
+		if (host.context_switches_per_s) {
+			Row("Context switches", Fmt("%.0f/s", *host.context_switches_per_s));
+		}
+		if (host.page_faults_per_s) {
+			Row("Page faults", Fmt("%.0f/s  (%.0f from disk)", *host.page_faults_per_s,
+			                       host.page_ins_per_s.value_or(0.0)));
+		}
+		Row("Thermal", ThermalName(host.thermal_pressure), ThermalColor(host.thermal_pressure));
+		const auto emulated = Rate(shown, Counter::EmulatedInstructions);
+		Row("Emulated instrs", Fmt("%.0f/s", emulated), LevelColor(emulated, 1000, 100000));
+		const auto unresolved = Rate(shown, Counter::UnresolvedImportCalls);
+		Row("Missing HLE calls",
+		    Fmt("%.0f/s  (%llu total)", unresolved,
+		        static_cast<unsigned long long>(Total(shown, Counter::UnresolvedImportCalls))),
+		    unresolved > 0.0 ? MEDIUM : GOOD);
+		ImGui::EndTable();
+	}
+}
+
+void DrawGpu(const Snapshot& shown, const GraphicContext& graphics) {
+	const auto& host = shown.host;
+	if (BeginSection("GPU")) {
+		Row("Device", host.gpu_model.empty()
+		                  ? std::string(graphics.GetPhysicalDeviceProperties().deviceName.data())
+		              : host.gpu_cores
+		                  ? Fmt("%s (%u cores)", host.gpu_model.c_str(), *host.gpu_cores)
+		                  : host.gpu_model);
+		Row("Utilization", Percent(host.gpu_percent),
+		    host.gpu_percent ? std::optional(LevelColor(*host.gpu_percent, 80, 95)) : std::nullopt);
+		if (host.gpu_renderer_percent || host.gpu_tiler_percent) {
+			Row("Renderer / tiler",
+			    Percent(host.gpu_renderer_percent) + " / " + Percent(host.gpu_tiler_percent));
+		}
+		if (host.gpu_memory_in_use) {
+			Row("Driver memory", Bytes(*host.gpu_memory_in_use) + " in use / " +
+			                         Bytes(host.gpu_memory_allocated) + " allocated");
+		}
+		if (host.gpu_recoveries) {
+			Row("GPU resets", Fmt("%llu", static_cast<unsigned long long>(*host.gpu_recoveries)),
+			    *host.gpu_recoveries > 0 ? BAD : GOOD);
+		}
+		const double fps = std::max(shown.fps, 1.0);
+		GpuActivityRows(shown);
+		Row("Instances", Fmt("%.0f/frame", Rate(shown, Counter::DrawInstances) / fps));
+		Row("Indirect draws", Fmt("%.0f/s", Rate(shown, Counter::IndirectDraws)));
+		Row("Dispatches", Fmt("%.0f/frame  %.0f/s", Rate(shown, Counter::Dispatches) / fps,
+		                      Rate(shown, Counter::Dispatches)));
+		Row("Queue submits", Fmt("%.0f/s", Rate(shown, Counter::QueueSubmits)));
+		const auto faults = Rate(shown, Counter::GpuFaults);
+		Row("Write-tracking faults", Fmt("%.0f/s", faults), LevelColor(faults, 20000, 100000));
+		CompilationRows(shown);
+		ImGui::EndTable();
+	}
+	if (BeginSection("GPU memory")) {
+		if (shown.device_memory) {
+			const double share = shown.device_budget != 0
+			                         ? static_cast<double>(*shown.device_memory) /
+			                               static_cast<double>(shown.device_budget) * 100.0
+			                         : 0.0;
+			Row("Vulkan budget", Bytes(*shown.device_memory) + " / " + Bytes(shown.device_budget),
+			    LevelColor(share, 80, 95));
+		}
+		for (size_t heap = 0; heap < shown.heap_count; heap++) {
+			const auto& usage = shown.heaps[heap];
+			Row(Fmt("  Heap %zu%s", heap, usage.device_local ? " (device)" : "").c_str(),
+			    Bytes(usage.usage) + " / " + Bytes(usage.budget) +
+			        Fmt("  %u allocs", usage.allocations));
+		}
+		TransferRow(shown, "Texture transfers", Counter::TextureUploadBytes,
+		            Counter::TextureDownloadBytes);
+		Row("Texture evictions", Fmt("%.0f/s", Rate(shown, Counter::TextureEvictions)));
+		TransferRow(shown, "Buffer transfers", Counter::BufferUploadBytes,
+		            Counter::BufferDownloadBytes);
+		Row("Buffer evictions", Fmt("%.0f/s", Rate(shown, Counter::BufferEvictions)));
+		ImGui::EndTable();
+	}
+}
+
+void DrawMemoryAndIo(const Snapshot& shown) {
+	const auto& host  = shown.host;
+	const auto& guest = shown.guest_memory;
+	if (BeginSection("PS5 memory")) {
+		const auto direct_used = guest.direct_allocated + guest.pooled_allocated;
+		Row("Direct", Bytes(direct_used) + " / " + Bytes(guest.direct_total),
+		    LevelColor(guest.direct_total != 0 ? static_cast<double>(direct_used) /
+		                                             static_cast<double>(guest.direct_total) * 100.0
+		                                       : 0.0,
+		               85, 97));
+		Row("  Mapped", Bytes(guest.direct_mapped));
+		Row("  Pool",
+		    Bytes(guest.pooled_allocated) + " (" + Bytes(guest.pool_committed) + " committed)");
+		Row("Flexible", Bytes(guest.flexible_used) + " / " + Bytes(guest.flexible_total));
+		Row("Extended", Bytes(guest.automatic_allocated));
+		// Past the pool size the guest sees no page-table entries available.
+		const auto entries = std::max(guest.cpu_page_entries, guest.gpu_page_entries);
+		Row("Page table (2 MB)",
+		    Fmt("CPU %llu  GPU %llu  / %llu",
+		        static_cast<unsigned long long>(guest.cpu_page_entries),
+		        static_cast<unsigned long long>(guest.gpu_page_entries),
+		        static_cast<unsigned long long>(guest.page_entries_total)),
+		    entries > guest.page_entries_total            ? BAD
+		    : entries * 10 > guest.page_entries_total * 9 ? MEDIUM
+		                                                  : GOOD);
+		ImGui::EndTable();
+	}
+	if (BeginSection("Emulator process")) {
+		Row("Footprint", Bytes(host.process_memory));
+		Row("Resident", Bytes(host.process_resident));
+		if (host.process_compressed) {
+			Row("Compressed / swapped", Bytes(*host.process_compressed),
+			    *host.process_compressed > (1ull << 30u) ? MEDIUM : GOOD);
+		}
+		Row("File-backed", Bytes(host.process_file_backed));
+		Row("Address space", Bytes(host.process_virtual));
+		ImGui::EndTable();
+	}
+	if (BeginSection("System memory")) {
+		if (host.system_memory_used && host.system_memory_total) {
+			Row("Used", Bytes(*host.system_memory_used) + " / " + Bytes(*host.system_memory_total));
+		}
+		if (host.system_wired) {
+			Row("Wired", Bytes(*host.system_wired));
+		}
+		if (host.system_compressed) {
+			Row("Compressed", Bytes(*host.system_compressed));
+		}
+		Row("Cached files", Bytes(host.system_cached));
+		Row("Free", Bytes(host.system_free));
+		if (host.swap_used) {
+			Row("Swap", Bytes(*host.swap_used) + " / " + Bytes(host.swap_total),
+			    *host.swap_used > (4ull << 30u) ? MEDIUM : GOOD);
+		}
+		Row("Pressure", PressureName(host.memory_pressure), PressureColor(host.memory_pressure));
+		ImGui::EndTable();
+	}
+	if (BeginSection("I/O & audio")) {
+		Row("Disk read / write", Throughput(host.disk_read_per_s.value_or(0.0)) + " / " +
+		                             Throughput(host.disk_write_per_s.value_or(0.0)));
+		Row("Guest file read / write", Throughput(Rate(shown, Counter::GuestFileReadBytes)) +
+		                                   " / " +
+		                                   Throughput(Rate(shown, Counter::GuestFileWriteBytes)));
+		const auto underruns = Rate(shown, Counter::AudioUnderruns);
+		Row("Audio underruns",
+		    Fmt("%.0f/s  (%llu total)", underruns,
+		        static_cast<unsigned long long>(Total(shown, Counter::AudioUnderruns))),
+		    underruns > 0.0 ? BAD : GOOD);
+		ImGui::EndTable();
+	}
+}
+
+} // namespace
+
+void PerformanceOverlayRecordFlip(uint32_t guest_width, uint32_t guest_height) noexcept {
+	if (!Common::DebugCounters::Enabled()) {
+		return;
+	}
+	std::scoped_lock lock(g_frame_mutex);
+	const auto       now = Clock::now();
+	if (g_frames.window_start == Clock::time_point {}) {
+		g_frames.window_start = now;
+	}
+	auto& frames = g_frames;
+	if (frames.last_flip != Clock::time_point {}) {
+		const float ms = std::chrono::duration<float, std::milli>(now - frames.last_flip).count();
+		frames.intervals_ms[frames.next] = ms;
+		frames.next                      = (frames.next + 1) % FRAME_HISTORY;
+		frames.count                     = std::min(frames.count + 1, FRAME_HISTORY);
+		frames.max_ms                    = std::max(frames.max_ms, ms);
+	}
+	frames.last_flip    = now;
+	frames.guest_width  = guest_width;
+	frames.guest_height = guest_height;
+	frames.flips++;
+}
+
+bool PerformanceOverlayEnabled() noexcept {
+	return Common::DebugCounters::Enabled();
+}
+
+void TogglePerformanceOverlay() noexcept {
+	SetOverlayEnabled(!Common::DebugCounters::Enabled());
+	g_revision.fetch_add(1, std::memory_order_acq_rel);
+}
+
+void TogglePerformanceOverlayDetails() noexcept {
+	if (Common::DebugCounters::Enabled()) {
+		g_details.store(!g_details.load(std::memory_order_acquire), std::memory_order_release);
+	} else {
+		g_details.store(true, std::memory_order_release);
+		SetOverlayEnabled(true);
+	}
+	g_revision.fetch_add(1, std::memory_order_acq_rel);
+}
+
+uint64_t PerformanceOverlayRevision() noexcept {
+	return g_revision.load(std::memory_order_acquire);
+}
+
+float DrawPerformanceOverlay(const GraphicContext& graphics, vk::Extent2D extent) {
+	if (!PerformanceOverlayEnabled()) {
+		return 0.0f;
+	}
+	EnsureSampler();
+
+	static Snapshot shown;
+	static uint64_t shown_generation = 0;
+	static bool     shown_details    = false;
+	const auto      generation       = g_generation.load(std::memory_order_acquire);
+	const bool      details          = g_details.load(std::memory_order_acquire);
+	const auto      now              = Clock::now();
+	if (generation != shown_generation) {
+		shown            = {};
+		shown_generation = generation;
+	}
+	if (shown.sample_time == Clock::time_point {} || now - shown.sample_time >= REFRESH_INTERVAL ||
+	    details != shown_details) {
+		{
+			std::scoped_lock lock(g_frame_mutex);
+			auto&            frames = g_frames;
+			shown.history_count     = frames.count;
+			for (size_t i = 0; i < frames.count; i++) {
+				shown.history[i] =
+				    frames.intervals_ms[(frames.next + FRAME_HISTORY - frames.count + i) %
+				                        FRAME_HISTORY];
+			}
+			if (frames.window_start == Clock::time_point {}) {
+				frames.window_start = now;
+			}
+			const double seconds = std::chrono::duration<double>(now - frames.window_start).count();
+			shown.fps            = seconds > 0.0 ? frames.flips / seconds : 0.0;
+			shown.frame_ms       = frames.flips > 0 ? seconds * 1000.0 / frames.flips : 0.0;
+			shown.max_frame_ms   = frames.max_ms;
+			shown.stalled_s   = frames.last_flip == Clock::time_point {}
+			                        ? 0.0
+			                        : std::chrono::duration<double>(now - frames.last_flip).count();
+			shown.guest_width = frames.guest_width;
+			shown.guest_height  = frames.guest_height;
+			frames.flips        = 0;
+			frames.max_ms       = 0.0f;
+			frames.window_start = now;
+		}
+		shown_details = details;
+		if (shown.history_count > 0) {
+			auto       sorted     = shown.history;
+			const auto end        = sorted.begin() + shown.history_count;
+			const auto percentile = sorted.begin() + shown.history_count * 99 / 100;
+			std::nth_element(sorted.begin(), percentile, end);
+			shown.low_fps = *percentile > 0.0f ? 1000.0 / *percentile : 0.0;
+		}
+		Refresh(shown, graphics, now, generation);
+	}
+
+	const float scale = std::clamp(std::min(static_cast<float>(extent.width) / 1920.0f,
+	                                        static_cast<float>(extent.height) / 1080.0f),
+	                               0.75f, 1.5f);
+	ImGui::SetNextWindowPos({static_cast<float>(extent.width) - 12.0f * scale, 12.0f * scale},
+	                        ImGuiCond_Always, {1.0f, 0.0f});
+	ImGui::SetNextWindowBgAlpha(details ? 0.82f : 0.72f);
+	constexpr ImGuiWindowFlags FLAGS =
+	    ImGuiWindowFlags_NoDecoration | ImGuiWindowFlags_NoInputs | ImGuiWindowFlags_NoNav |
+	    ImGuiWindowFlags_NoMove | ImGuiWindowFlags_NoSavedSettings |
+	    ImGuiWindowFlags_NoFocusOnAppearing | ImGuiWindowFlags_NoBringToFrontOnFocus |
+	    ImGuiWindowFlags_AlwaysAutoResize;
+	ImGui::PushFont(nullptr, (details ? 13.0f : 15.0f) * scale);
+	ImGui::PushStyleVar(ImGuiStyleVar_WindowRounding, 8.0f * scale);
+	ImGui::PushStyleVar(ImGuiStyleVar_WindowPadding, {10.0f * scale, 8.0f * scale});
+	ImGui::PushStyleVar(ImGuiStyleVar_ItemSpacing, {8.0f * scale, 2.0f * scale});
+	ImGui::PushStyleVar(ImGuiStyleVar_CellPadding, {4.0f * scale, 1.0f * scale});
+	if (ImGui::Begin("##kyty_performance_overlay", nullptr, FLAGS)) {
+		if (shown.fps > 0.0) {
+			ImGui::TextColored(Color(FrameTimeColor(shown.frame_ms)), "%.1f FPS   %.1f ms",
+			                   shown.fps, shown.frame_ms);
+		} else if (shown.stalled_s >= 1.0) {
+			ImGui::TextColored(Color(BAD), "No frames for %.0f s", shown.stalled_s);
+		} else {
+			ImGui::TextDisabled("Waiting for frames");
+		}
+		DrawFrameGraph(shown.history, shown.history_count, (details ? 330.0f : 260.0f) * scale,
+		               46.0f * scale);
+		if (details) {
+			// Three columns keep the detailed panel within a 720p window.
+			if (ImGui::BeginTable("##details", 3, TABLE_FLAGS)) {
+				ImGui::TableNextRow();
+				ImGui::TableNextColumn();
+				DrawFramesAndCpu(shown, extent);
+				ImGui::TableNextColumn();
+				DrawGpu(shown, graphics);
+				ImGui::TableNextColumn();
+				DrawMemoryAndIo(shown);
+				ImGui::EndTable();
+			}
+		} else {
+			DrawCompact(shown);
+		}
+#if defined(__APPLE__)
+		ImGui::TextDisabled(details ? "Cmd+Shift+P compact   Cmd+P hide"
+		                            : "Cmd+Shift+P details   Cmd+P hide");
+#else
+		ImGui::TextDisabled(details ? "Shift+F2 compact   F2 hide" : "Shift+F2 details   F2 hide");
+#endif
+	}
+	const float bottom = ImGui::GetWindowPos().y + ImGui::GetWindowSize().y;
+	ImGui::End();
+	ImGui::PopStyleVar(4);
+	ImGui::PopFont();
+	return bottom;
+}
+
+} // namespace Libs::Graphics
@@ -0,0 +1,30 @@
+#ifndef EMULATOR_SRC_GRAPHICS_PRESENTATION_PERFORMANCEOVERLAY_H_
+#define EMULATOR_SRC_GRAPHICS_PRESENTATION_PERFORMANCEOVERLAY_H_
+
+#include "graphics/host_gpu/vulkanCommon.h"
+
+#include <cstdint>
+
+namespace Libs::Graphics {
+
+struct GraphicContext;
+
+// Called for each main-bus guest flip, with the guest's video-out size.
+void PerformanceOverlayRecordFlip(uint32_t guest_width, uint32_t guest_height) noexcept;
+
+[[nodiscard]] bool PerformanceOverlayEnabled() noexcept;
+void               TogglePerformanceOverlay() noexcept;
+// Switches between the compact and the detailed panel, showing the panel if it was hidden.
+void TogglePerformanceOverlayDetails() noexcept;
+
+// Changes when the panel needs a repaint without a guest flip (toggled, or the guest stalled).
+[[nodiscard]] uint64_t PerformanceOverlayRevision() noexcept;
+
+// Must run between ImGui::NewFrame() and ImGui::Render(), on the present thread while the
+// graphics subsystem is alive. Draws in the top-right corner and returns the panel's bottom
+// edge, or 0 when hidden.
+float DrawPerformanceOverlay(const GraphicContext& graphics, vk::Extent2D extent);
+
+} // namespace Libs::Graphics
+
+#endif // EMULATOR_SRC_GRAPHICS_PRESENTATION_PERFORMANCEOVERLAY_H_
@@ -162,7 +162,8 @@ void RenderDocOnGuestFlip(RenderContext& renderer) {
 		return;
 	}
 	if (state == RenderDocState::Capturing) {
-		LOGF("RenderDoc: captured guest flip %u/2\n", ++g_captured_flips);
+		++g_captured_flips;
+		LOGF("RenderDoc: captured guest flip %u/2\n", g_captured_flips);
 		if (g_captured_flips < 2) {
 			return;
 		}
@@ -1,18 +1,19 @@
 #include "graphics/presentation/systemOverlay.h"

-#include <SDL3/SDL.h>
-
 #include "common/assert.h"
 #include "common/emulatorConfig.h"
 #include "common/file.h"
+#include "common/localization.h"
 #include "common/logging/log.h"
 #include "common/stringUtils.h"
 #include "gpu_blit_shaders/gpu_imgui_hdr_spv.h"
 #include "graphics/host_gpu/graphicContext.h"
 #include "graphics/presentation/ddsImage.h"
+#include "graphics/presentation/performanceOverlay.h"
+#include "graphics/presentation/window/hostInput.h"
 #include "imgui.h"
-#include "imgui_internal.h"
 #include "imgui_impl_vulkan.h"
+#include "imgui_internal.h"
 #include "libs/controller.h"
 #include "libs/dialog.h"
 #include "libs/ime.h"
@@ -20,6 +21,7 @@
 #include "loader/systemContent.h"
 #include "stb_image.h"

+#include <SDL3/SDL.h>
 #include <algorithm>
 #include <array>
 #include <atomic>
@@ -29,6 +31,7 @@
 #include <cstring>
 #include <deque>
 #include <fmt/format.h>
+#include <initializer_list>
 #include <limits>
 #include <memory>
 #include <mutex>
@@ -149,6 +152,47 @@ bool GetOverlaySnapshot(OverlaySnapshot* snapshot) {
 }

 constexpr size_t INPUT_QUEUE_CAPACITY = 128;
+constexpr int    MAX_TOUCH_CURSOR_IMAGE_DIMENSION = 2048;
+constexpr float  TOUCH_CURSOR_IMAGE_SIZE          = 48.0f;
+
+enum class TouchParticleSkin : uint8_t { Classic, BlueNeon, Monochrome, Crimson, RedBlue };
+
+TouchParticleSkin GetTouchParticleSkin() {
+	static const auto skin = [] {
+		for (const auto& value: Config::GetKeymap()) {
+			constexpr std::string_view key = "TouchParticleSkin=";
+			if (std::string_view(value).starts_with(key)) {
+				const auto name = std::string_view(value).substr(key.size());
+				if (name == "BlueNeon") {
+					return TouchParticleSkin::BlueNeon;
+				}
+				if (name == "Monochrome") {
+					return TouchParticleSkin::Monochrome;
+				}
+				if (name == "Crimson") {
+					return TouchParticleSkin::Crimson;
+				}
+				if (name == "RedBlue") {
+					return TouchParticleSkin::RedBlue;
+				}
+			}
+		}
+		return TouchParticleSkin::Classic;
+	}();
+	return skin;
+}
+
+bool TouchCursorImageEnabled() {
+	static const bool enabled = [] {
+		for (const auto& value: Config::GetKeymap()) {
+			if (std::string_view(value) == "TouchCursorImage=Off") {
+				return false;
+			}
+		}
+		return true;
+	}();
+	return enabled;
+}

 enum class InputKind : uint8_t {
 	Button,
@@ -238,6 +282,61 @@ std::vector<std::string_view> WrapTrophyTitle(std::string_view text, float font_
 	}
 }

+// ImGui's built-in font covers Latin only. Merge installed fonts for the selected console language.
+void AddSystemFontFallbacks(ImGuiIO& io) {
+	io.Fonts->AddFontDefault();
+	const auto add = [&](std::initializer_list<const char*> paths) {
+		for (const char* path: paths) {
+			std::error_code error;
+			if (std::filesystem::is_regular_file(path, error)) {
+				ImFontConfig config;
+				config.MergeMode = true;
+				config.Flags = ImFontFlags_NoLoadError;
+				if (io.Fonts->AddFontFromFileTTF(path, 0.0f, &config) != nullptr) {
+					return true;
+				}
+			}
+		}
+		return false;
+	};
+	bool found =
+	    add({"C:/Windows/Fonts/segoeui.ttf", "/System/Library/Fonts/Supplemental/Arial Unicode.ttf",
+	         "/Library/Fonts/Arial Unicode.ttf", "/usr/share/fonts/truetype/dejavu/DejaVuSans.ttf",
+	         "/usr/share/fonts/TTF/DejaVuSans.ttf", "/usr/share/fonts/noto/NotoSans-Regular.ttf"});
+	const auto  locale     = Common::Localization::LocaleName(Config::GetConsoleLanguage());
+	const char* noto_cjk[] = {"/usr/share/fonts/opentype/noto/NotoSansCJK-Regular.ttc",
+	                          "/usr/share/fonts/noto-cjk/NotoSansCJK-Regular.ttc",
+	                          "/usr/share/fonts/google-noto-cjk/NotoSansCJK-Regular.ttc"};
+	if (locale == "ja-JP") {
+		found = add({"C:/Windows/Fonts/meiryo.ttc", "C:/Windows/Fonts/YuGothM.ttc",
+		             "/System/Library/Fonts/ヒラギノ角ゴシック W3.ttc", noto_cjk[0], noto_cjk[1],
+		             noto_cjk[2]});
+	} else if (locale == "ko-KR") {
+		found = add({"C:/Windows/Fonts/malgun.ttf", "/System/Library/Fonts/AppleSDGothicNeo.ttc",
+		             noto_cjk[0], noto_cjk[1], noto_cjk[2]});
+	} else if (locale == "zh-Hans" || locale == "zh-Hant") {
+		found =
+		    add({locale == "zh-Hans" ? "C:/Windows/Fonts/msyh.ttc" : "C:/Windows/Fonts/msjh.ttc",
+		         "/System/Library/Fonts/PingFang.ttc", noto_cjk[0], noto_cjk[1], noto_cjk[2]});
+	} else if (locale == "th-TH") {
+		found = add({"C:/Windows/Fonts/leelawui.ttf", "C:/Windows/Fonts/tahoma.ttf",
+		             "/System/Library/Fonts/Supplemental/Ayuthaya.ttf",
+		             "/usr/share/fonts/truetype/tlwg/Garuda.ttf",
+		             "/usr/share/fonts/truetype/noto/NotoSansThai-Regular.ttf",
+		             "/usr/share/fonts/noto/NotoSansThai-Regular.ttf"});
+	} else if (locale == "ar-AE") {
+		found = add({"C:/Windows/Fonts/tahoma.ttf", "/System/Library/Fonts/Supplemental/Arial.ttf",
+		             "/usr/share/fonts/truetype/dejavu/DejaVuSans.ttf",
+		             "/usr/share/fonts/TTF/DejaVuSans.ttf",
+		             "/usr/share/fonts/truetype/noto/NotoSansArabic-Regular.ttf",
+		             "/usr/share/fonts/noto/NotoSansArabic-Regular.ttf"});
+	}
+	if (!found) {
+		LOGF("[Trophy] no installed fallback font for %.*s\n", static_cast<int>(locale.size()),
+		     locale.data());
+	}
+}
+
 void ClearInputEvents() {
 	std::scoped_lock lock(g_input_mutex);
 	g_input_events.clear();
@@ -399,6 +498,63 @@ std::u16string Utf8ToUtf16(std::string_view text) {
 	return result;
 }

+// The Arabic notification is a single RTL run. Use presentation forms because ImGui does not
+// shape Arabic. This intentionally applies only to the catalog subtitle, never game-provided names.
+std::string PrepareTrophySubtitle(int language) {
+	const auto text = Common::Localization::Get(language, "trophy.toast.earned");
+	if (Common::Localization::LocaleName(language) != "ar-AE") {
+		return std::string(text);
+	}
+	struct ArabicForm {
+		char16_t letter;
+		char16_t isolated;
+		bool     joins_next;
+	};
+	static constexpr ArabicForm FORMS[] = {
+	    {0x0621, 0xfe80, false}, {0x0622, 0xfe81, false}, {0x0623, 0xfe83, false},
+	    {0x0624, 0xfe85, false}, {0x0625, 0xfe87, false}, {0x0626, 0xfe89, true},
+	    {0x0627, 0xfe8d, false}, {0x0628, 0xfe8f, true},  {0x0629, 0xfe93, false},
+	    {0x062a, 0xfe95, true},  {0x062b, 0xfe99, true},  {0x062c, 0xfe9d, true},
+	    {0x062d, 0xfea1, true},  {0x062e, 0xfea5, true},  {0x062f, 0xfea9, false},
+	    {0x0630, 0xfeab, false}, {0x0631, 0xfead, false}, {0x0632, 0xfeaf, false},
+	    {0x0633, 0xfeb1, true},  {0x0634, 0xfeb5, true},  {0x0635, 0xfeb9, true},
+	    {0x0636, 0xfebd, true},  {0x0637, 0xfec1, true},  {0x0638, 0xfec5, true},
+	    {0x0639, 0xfec9, true},  {0x063a, 0xfecd, true},  {0x0641, 0xfed1, true},
+	    {0x0642, 0xfed5, true},  {0x0643, 0xfed9, true},  {0x0644, 0xfedd, true},
+	    {0x0645, 0xfee1, true},  {0x0646, 0xfee5, true},  {0x0647, 0xfee9, true},
+	    {0x0648, 0xfeed, false}, {0x0649, 0xfeef, false}, {0x064a, 0xfef1, true},
+	};
+	const auto form = [](char16_t c) -> const ArabicForm* {
+		for (const auto& value: FORMS) {
+			if (value.letter == c) {
+				return &value;
+			}
+		}
+		return nullptr;
+	};
+	const auto logical = Utf8ToUtf16(text);
+	if (std::none_of(logical.begin(), logical.end(),
+	                 [](char16_t c) { return c >= 0x0600 && c <= 0x06ff; })) {
+		return std::string(text);
+	}
+	auto visual = logical;
+	for (size_t i = 0; i < logical.size(); ++i) {
+		const auto* current = form(logical[i]);
+		if (current == nullptr) {
+			continue;
+		}
+		const auto* previous = i > 0 ? form(logical[i - 1]) : nullptr;
+		const auto* next     = i + 1 < logical.size() ? form(logical[i + 1]) : nullptr;
+		const bool  joins_previous =
+		    previous != nullptr && previous->joins_next && current->letter != 0x0621;
+		const bool joins_next = current->joins_next && next != nullptr && next->letter != 0x0621;
+		visual[i]             = static_cast<char16_t>(current->isolated + (joins_next ? 2 : 0) +
+		                                              (joins_previous ? 1 : 0));
+	}
+	std::reverse(visual.begin(), visual.end());
+	return Common::Utf16ToUtf8(visual);
+}
+
 std::string VisibleText(const Ime::HostSnapshot& snapshot, size_t max_units) {
 	const size_t cursor        = std::min<size_t>(snapshot.cursor, snapshot.text.size());
 	const size_t payload_units = max_units > 4 ? max_units - 4 : 1;
@@ -552,9 +708,12 @@ SystemOverlayVisualState GetSystemOverlayVisualState() noexcept {
 	const auto system = SystemDialog::GetVisualState();
 	const bool trophy_active = g_trophy_notification_active.load(std::memory_order_acquire);
 	const bool splash        = g_splash_active.load(std::memory_order_acquire);
-	return {core.active || dialog.active || system.active || trophy_active || splash,
+	const auto cursor        = GetHostInputTouchPadVisualState();
+	return {core.active || dialog.active || system.active || trophy_active || splash ||
+	            (cursor.enabled && cursor.cursor_visible),
 	        core.revision + dialog.revision + system.revision + (trophy_active ? 1u : 0u) +
-	            (splash ? 2u : 0u)};
+	            (splash ? 2u : 0u) + static_cast<decltype(core.revision)>(cursor.revision) +
+	            PerformanceOverlayRevision()};
 }

 bool ProcessSystemOverlayInput(const SDL_Event& event) {
@@ -694,6 +853,20 @@ void NotifyTrophyUnlocked(std::string_view name, int32_t grade,
 }

 struct SystemOverlay::Impl {
+	enum class TouchParticleIcon : uint8_t { Square, Triangle, Cross, Circle };
+
+	struct TouchParticle {
+		ImVec2            position;
+		Clock::time_point born;
+		TouchParticleIcon icon;
+	};
+
+	struct TouchTrailSegment {
+		ImVec2            from;
+		ImVec2            to;
+		Clock::time_point born;
+	};
+
 	explicit Impl(GraphicContext& context): graphics(context) {}

 	~Impl() {
@@ -705,6 +878,12 @@ struct SystemOverlay::Impl {
 		ReleaseVulkan();
 		if (imgui_context != nullptr) {
 			ImGui::SetCurrentContext(imgui_context);
+			if (hand_cursor_image) {
+				ImGui::UnregisterUserTexture(hand_cursor_image.get());
+			}
+			if (drawing_cursor_image) {
+				ImGui::UnregisterUserTexture(drawing_cursor_image.get());
+			}
 			if (trophy_notification_image) {
 				ImGui::UnregisterUserTexture(trophy_notification_image.get());
 			}
@@ -730,11 +909,14 @@ struct SystemOverlay::Impl {
 		io.ConfigNavCursorVisibleAlways = true;
 		io.BackendFlags |= ImGuiBackendFlags_HasGamepad;
 		io.BackendPlatformName = "Kyty system overlay input";
+		AddSystemFontFallbacks(io);
 		ImGui::StyleColorsDark();
 		auto& style          = ImGui::GetStyle();
 		style.WindowRounding = 10.0f;
 		style.FrameRounding  = 6.0f;
 		style.ItemSpacing    = {8.0f, 8.0f};
+		hand_cursor_image    = LoadTouchCursorImage("hand.png");
+		drawing_cursor_image = LoadTouchCursorImage("drawing.png");
 	}

 	void PlayTrophySound() {
@@ -775,6 +957,55 @@ struct SystemOverlay::Impl {
 		}
 	}

+	std::unique_ptr<ImTextureData> LoadTouchCursorImage(std::string_view filename) {
+		const char* base_path = SDL_GetBasePath();
+		if (base_path == nullptr) {
+			LOGF("Could not locate the emulator folder for touchpad cursor images: %s\n",
+			     SDL_GetError());
+			return nullptr;
+		}
+		const std::filesystem::path path =
+		    std::filesystem::path(base_path) / "assets" / "cursors" / std::string(filename);
+
+		Common::File file;
+		if (!file.Open(path, Common::File::Mode::Read)) {
+			LOGF("Touchpad cursor image not found: %s\n", Common::PathToString(path).c_str());
+			return nullptr;
+		}
+		const auto encoded = file.ReadWholeBuffer();
+		if (encoded.empty() ||
+		    encoded.size() > static_cast<std::size_t>(std::numeric_limits<int>::max())) {
+			LOGF("Touchpad cursor image is empty or too large: %s\n",
+			     Common::PathToString(path).c_str());
+			return nullptr;
+		}
+
+		int width = 0;
+		int height = 0;
+		auto* pixels = stbi_load_from_memory(
+		    reinterpret_cast<const stbi_uc*>(encoded.data()), static_cast<int>(encoded.size()),
+		    &width, &height, nullptr, 4);
+		if (pixels == nullptr || width <= 0 || height <= 0 ||
+		    width > MAX_TOUCH_CURSOR_IMAGE_DIMENSION ||
+		    height > MAX_TOUCH_CURSOR_IMAGE_DIMENSION) {
+			LOGF("Could not load touchpad cursor image %s: %s\n",
+			     Common::PathToString(path).c_str(),
+			     pixels == nullptr ? stbi_failure_reason() : "image dimensions exceed 2048x2048");
+			stbi_image_free(pixels);
+			return nullptr;
+		}
+
+		auto texture = std::make_unique<ImTextureData>();
+		texture->Create(ImTextureFormat_RGBA32, width, height);
+		std::memcpy(texture->Pixels, pixels,
+		            static_cast<std::size_t>(width) * static_cast<std::size_t>(height) * 4);
+		texture->UseColors = true;
+		stbi_image_free(pixels);
+		ImGui::RegisterUserTexture(texture.get());
+		LOGF("Loaded touchpad cursor image: %s\n", Common::PathToString(path).c_str());
+		return texture;
+	}
+
 	std::unique_ptr<ImTextureData> LoadTrophyIcon(std::span<const std::byte> encoded) {
 		if (encoded.empty() ||
 		    encoded.size() > static_cast<std::size_t>(std::numeric_limits<int>::max())) {
@@ -1191,7 +1422,248 @@ struct SystemOverlay::Impl {
 		}
 	}

-	void DrawTrophyNotification(vk::Extent2D frame_extent, Clock::time_point now) {
+	void EmitTouchParticles(ImVec2 from, ImVec2 to, Clock::time_point now) {
+		const float dx       = to.x - from.x;
+		const float dy       = to.y - from.y;
+		const float distance = std::hypot(dx, dy);
+		if (distance < 0.01f) {
+			return;
+		}
+		if (distance >= 1.0f) {
+			trail_segments.push_back({from, to, now});
+			if (trail_segments.size() > MAX_TOUCH_TRAIL_SEGMENTS) {
+				trail_segments.erase(trail_segments.begin());
+			}
+		}
+
+		constexpr float PARTICLE_SPACING = 24.0f;
+		float           traveled         = 0.0f;
+		while (particle_distance + distance - traveled >= PARTICLE_SPACING) {
+			traveled += PARTICLE_SPACING - particle_distance;
+			const float fraction = std::clamp(traveled / distance, 0.0f, 1.0f);
+			particles.push_back(
+			    {{from.x + dx * fraction, from.y + dy * fraction}, now, next_particle_icon});
+			next_particle_icon = static_cast<TouchParticleIcon>(
+			    (static_cast<uint8_t>(next_particle_icon) + 1) % 4);
+			if (particles.size() > MAX_TOUCH_PARTICLES) {
+				particles.erase(particles.begin());
+			}
+			particle_distance = 0.0f;
+		}
+		particle_distance += distance - traveled;
+	}
+
+	void DrawTouchParticles(ImDrawList* draw, Clock::time_point now) {
+		const auto skin = GetTouchParticleSkin();
+		constexpr float PARTICLE_LIFETIME = 1.0f;
+		constexpr float TRAIL_LIFETIME    = 0.35f;
+		std::erase_if(trail_segments, [now](const auto& segment) {
+			return std::chrono::duration<float>(now - segment.born).count() >= TRAIL_LIFETIME;
+		});
+		std::erase_if(particles, [now](const auto& particle) {
+			return std::chrono::duration<float>(now - particle.born).count() >=
+			       PARTICLE_LIFETIME;
+		});
+		for (const auto& segment: trail_segments) {
+			const float age =
+			    std::chrono::duration<float>(now - segment.born).count() / TRAIL_LIFETIME;
+			const float fade = 1.0f - age;
+			if (skin == TouchParticleSkin::RedBlue) {
+				draw->AddLine(segment.from, segment.to,
+				              IM_COL32(20, 75, 220, static_cast<int>(105.0f * fade)),
+				              9.0f * (1.0f - age * 0.5f));
+				draw->AddLine(segment.from, segment.to,
+				              IM_COL32(255, 45, 65, static_cast<int>(230.0f * fade)),
+				              4.0f * (1.0f - age * 0.5f));
+			} else if (skin == TouchParticleSkin::Crimson) {
+				draw->AddLine(segment.from, segment.to,
+				              IM_COL32(110, 0, 8, static_cast<int>(130.0f * fade)),
+				              9.0f * (1.0f - age * 0.5f));
+				draw->AddLine(segment.from, segment.to,
+				              IM_COL32(255, 35, 20, static_cast<int>(235.0f * fade)),
+				              4.0f * (1.0f - age * 0.5f));
+			} else {
+				draw->AddLine(segment.from, segment.to,
+				              IM_COL32(35, 145, 255, static_cast<int>(180.0f * fade)),
+				              5.0f * (1.0f - age * 0.5f));
+			}
+		}
+
+		for (const auto& particle: particles) {
+			const float age =
+			    std::chrono::duration<float>(now - particle.born).count() / PARTICLE_LIFETIME;
+			const float alpha = 1.0f - age;
+			const float size  = 9.0f * (1.0f - age * 0.35f);
+			const float drift = (static_cast<uint8_t>(particle.icon) % 2 == 0 ? -1.0f : 1.0f) *
+			                    8.0f * age;
+			const ImVec2 center {particle.position.x + drift, particle.position.y - 18.0f * age};
+			const auto faded = [alpha, skin, icon = particle.icon](int r, int g, int b) {
+				if (skin == TouchParticleSkin::BlueNeon) {
+					r = 75;
+					g = 220;
+					b = 255;
+				} else if (skin == TouchParticleSkin::RedBlue) {
+					const bool red = icon == TouchParticleIcon::Square ||
+					                 icon == TouchParticleIcon::Cross;
+					r = red ? 255 : 75;
+					g = red ? 65 : 175;
+					b = red ? 85 : 255;
+				} else if (skin == TouchParticleSkin::Monochrome) {
+					r = 235;
+					g = 245;
+					b = 255;
+				} else if (skin == TouchParticleSkin::Crimson) {
+					const bool ember = icon == TouchParticleIcon::Triangle ||
+					                   icon == TouchParticleIcon::Circle;
+					r = 255;
+					g = ember ? 205 : 115;
+					b = ember ? 70 : 15;
+				}
+				return IM_COL32(r, g, b, static_cast<int>(255.0f * alpha));
+			};
+
+			if (skin != TouchParticleSkin::Classic) {
+				const ImU32 glow =
+				    skin == TouchParticleSkin::BlueNeon
+				        ? IM_COL32(0, 115, 255, static_cast<int>(38.0f * alpha))
+				        : (skin == TouchParticleSkin::RedBlue
+				               ? IM_COL32(75, 100, 255, static_cast<int>(55.0f * alpha))
+				               : (skin == TouchParticleSkin::Crimson
+				               ? IM_COL32(255, 70, 0, static_cast<int>(70.0f * alpha))
+				               : IM_COL32(115, 165, 255, static_cast<int>(30.0f * alpha))));
+				draw->AddCircleFilled(center, size + 6.0f, glow);
+			}
+			const ImU32 particle_glow =
+			    skin == TouchParticleSkin::Crimson
+			        ? (particle.icon == TouchParticleIcon::Triangle ||
+			                   particle.icon == TouchParticleIcon::Circle
+			               ? IM_COL32(255, 195, 30, static_cast<int>(135.0f * alpha))
+			               : IM_COL32(255, 95, 0, static_cast<int>(135.0f * alpha)))
+			        : (skin == TouchParticleSkin::RedBlue
+			               ? (particle.icon == TouchParticleIcon::Square ||
+			                          particle.icon == TouchParticleIcon::Cross
+			                      ? IM_COL32(255, 40, 65, static_cast<int>(105.0f * alpha))
+			                      : IM_COL32(35, 125, 255, static_cast<int>(105.0f * alpha)))
+			               : IM_COL32(35, 145, 255,
+			                          static_cast<int>((skin == TouchParticleSkin::BlueNeon
+			                                                ? 100.0f
+			                                                : 65.0f) *
+			                                           alpha)));
+			draw->AddCircleFilled(center, size + 2.0f, particle_glow);
+			switch (particle.icon) {
+				case TouchParticleIcon::Square:
+					draw->AddRect({center.x - size * 0.65f, center.y - size * 0.65f},
+					              {center.x + size * 0.65f, center.y + size * 0.65f},
+					              faded(255, 105, 180), 2.0f, 0, 2.2f);
+					break;
+				case TouchParticleIcon::Triangle:
+					draw->AddLine({center.x, center.y - size}, {center.x - size, center.y + size},
+					              faded(80, 230, 130), 2.2f);
+					draw->AddLine({center.x - size, center.y + size},
+					              {center.x + size, center.y + size}, faded(80, 230, 130), 2.2f);
+					draw->AddLine({center.x + size, center.y + size},
+					              {center.x, center.y - size}, faded(80, 230, 130), 2.2f);
+					break;
+				case TouchParticleIcon::Cross:
+					draw->AddLine({center.x - size * 0.75f, center.y - size * 0.75f},
+					              {center.x + size * 0.75f, center.y + size * 0.75f},
+					              faded(90, 170, 255), 2.5f);
+					draw->AddLine({center.x + size * 0.75f, center.y - size * 0.75f},
+					              {center.x - size * 0.75f, center.y + size * 0.75f},
+					              faded(90, 170, 255), 2.5f);
+					break;
+				case TouchParticleIcon::Circle:
+					draw->AddCircle(center, size * 0.8f, faded(255, 105, 105), 0, 2.2f);
+					break;
+			}
+		}
+	}
+
+	void DrawTouchPadCursor(const HostInputTouchPadVisualState& state) {
+		if (!TouchCursorImageEnabled()) {
+			particles.clear();
+			trail_segments.clear();
+			return;
+		}
+		const auto now = Clock::now();
+		if (state.enabled && state.position_valid) {
+			const ImVec2 target {state.x * extent.width, state.y * extent.height};
+			if (!cursor_initialized || state.gesture != cursor_gesture) {
+				cursor_position    = target;
+				cursor_initialized = true;
+				cursor_gesture     = state.gesture;
+				particle_distance  = 0.0f;
+			} else {
+				const float dt = last_cursor_update == Clock::time_point {}
+				                     ? 1.0f / 60.0f
+				                     : std::clamp(std::chrono::duration<float>(now -
+				                                                             last_cursor_update)
+				                                      .count(),
+				                                  1.0f / 1000.0f, 0.05f);
+				const ImVec2 previous = cursor_position;
+				const float  follow   = 1.0f - std::exp(-18.0f * dt);
+				cursor_position.x += (target.x - cursor_position.x) * follow;
+				cursor_position.y += (target.y - cursor_position.y) * follow;
+				if (state.drawing) {
+					EmitTouchParticles(previous, cursor_position, now);
+				}
+			}
+			last_cursor_update = now;
+		}
+		auto* draw = ImGui::GetForegroundDrawList();
+		DrawTouchParticles(draw, now);
+		if (!state.enabled || !state.cursor_visible ||
+		    !state.position_valid || !cursor_initialized) {
+			return;
+		}
+
+		const ImVec2 position = cursor_position;
+		const auto& cursor_image =
+		    state.drawing ? drawing_cursor_image : hand_cursor_image;
+		if (cursor_image != nullptr) {
+			const float scale = TOUCH_CURSOR_IMAGE_SIZE /
+			                    static_cast<float>(std::max(cursor_image->Height, 1));
+			const ImVec2 size {static_cast<float>(cursor_image->Width) * scale,
+			                   static_cast<float>(cursor_image->Height) * scale};
+			draw->AddImage(cursor_image->GetTexRef(), position,
+			               {position.x + size.x, position.y + size.y});
+			return;
+		}
+
+		const ImU32  blue     = IM_COL32(35, 145, 255, 220);
+		const ImVec2 hand[] = {
+		    {position.x, position.y},         {position.x, position.y + 20.0f},
+		    {position.x + 4.0f, position.y + 16.0f},
+		    {position.x + 9.0f, position.y + 27.0f},
+		    {position.x + 13.0f, position.y + 25.0f},
+		    {position.x + 10.0f, position.y + 17.0f},
+		    {position.x + 18.0f, position.y + 18.0f},
+		    {position.x + 20.0f, position.y + 15.0f},
+		    {position.x + 19.0f, position.y + 12.0f},
+		    {position.x + 12.0f, position.y + 10.0f},
+		    {position.x + 12.0f, position.y + 4.0f},
+		    {position.x + 10.0f, position.y + 2.0f},
+		    {position.x + 8.0f, position.y + 3.0f},
+		    {position.x + 8.0f, position.y + 10.0f},
+		    {position.x + 6.0f, position.y + 2.0f},
+		    {position.x + 4.0f, position.y + 2.0f},
+		    {position.x + 3.0f, position.y + 10.0f},
+		    {position.x + 2.0f, position.y + 1.0f},
+		};
+		draw->AddConvexPolyFilled(hand, IM_ARRAYSIZE(hand), IM_COL32(250, 250, 250, 255));
+		for (std::size_t i = 0; i < IM_ARRAYSIZE(hand); i++) {
+			draw->AddLine(hand[i], hand[(i + 1) % IM_ARRAYSIZE(hand)],
+			              IM_COL32(20, 25, 35, 255), 1.5f);
+		}
+		if (state.drawing) {
+			const ImVec2 pen_start {position.x + 12.0f, position.y + 8.0f};
+			const ImVec2 pen_end {position.x + 23.0f, position.y - 3.0f};
+			draw->AddLine(pen_start, pen_end, IM_COL32(10, 45, 95, 255), 5.0f);
+			draw->AddLine(pen_start, pen_end, blue, 2.5f);
+		}
+	}
+
+	void DrawTrophyNotification(vk::Extent2D frame_extent, Clock::time_point now, float min_top) {
 		std::erase_if(retired_trophy_images, [](const auto& texture) {
 			if (texture->Status != ImTextureStatus_Destroyed) {
 				return false;
@@ -1221,8 +1693,9 @@ struct SystemOverlay::Impl {
 				trophy_notification = std::move(g_trophy_notifications.front());
 				g_trophy_notifications.pop_front();
 			}
-			trophy_notification_born  = now;
-			trophy_notification_image = LoadTrophyIcon(trophy_notification->icon_png);
+			trophy_notification_subtitle = PrepareTrophySubtitle(Config::GetConsoleLanguage());
+			trophy_notification_born     = now;
+			trophy_notification_image    = LoadTrophyIcon(trophy_notification->icon_png);
 			PlayTrophySound();
 		}
 		const float age = std::chrono::duration<float>(now - trophy_notification_born).count();
@@ -1258,7 +1731,7 @@ struct SystemOverlay::Impl {
 		const float  alpha    = std::min(fade_in, fade_out);
 		const float  slide    = (1.0f - fade_in) * (width + 24.0f);
 		const ImVec2 top_left {static_cast<float>(frame_extent.width) - 32.0f - width + slide,
-		                       32.0f};
+		                       std::max(32.0f, min_top + 12.0f)};
 		const ImVec2 bottom_right {top_left.x + width, top_left.y + height};
 		auto*        draw = ImGui::GetForegroundDrawList();
 		const auto   fade = [alpha](int r, int g, int b, int a) {
@@ -1306,8 +1779,17 @@ struct SystemOverlay::Impl {
 			              line.data(), line.data() + line.size());
 			line_y += line_height;
 		}
-		draw->AddText(nullptr, 20.0f * screen_scale, {text_left, line_y + 8.0f * screen_scale},
-		              fade(176, 180, 192, 255), "Trophy earned!");
+		const float subtitle_size = 20.0f * screen_scale;
+		const float subtitle_width =
+		    ImGui::GetFont()
+		        ->CalcTextSizeA(subtitle_size, std::numeric_limits<float>::max(), 0.0f,
+		                        trophy_notification_subtitle.c_str())
+		        .x;
+		const float available_width = std::max(1.0f, width - text_offset - text_padding);
+		const float fitted_size = std::max(
+		    1.0f, subtitle_size * std::min(1.0f, available_width / std::max(1.0f, subtitle_width)));
+		draw->AddText(nullptr, fitted_size, {text_left, line_y + 8.0f * screen_scale},
+		              fade(176, 180, 192, 255), trophy_notification_subtitle.c_str());
 		draw->PopClipRect();
 	}

@@ -1316,7 +1798,8 @@ struct SystemOverlay::Impl {
 		const bool      has_overlay = GetOverlaySnapshot(&snapshot);
 		const auto now = Clock::now();
 		if (!has_overlay && !g_trophy_notification_active.load(std::memory_order_acquire) &&
-		    !g_splash_active.load(std::memory_order_acquire)) {
+		    !g_splash_active.load(std::memory_order_acquire) &&
+		    !GetHostInputTouchPadVisualState().enabled && !PerformanceOverlayEnabled()) {
 			return false;
 		}
 		const auto prepared_session = snapshot.session;
@@ -1347,6 +1830,7 @@ struct SystemOverlay::Impl {
 		ImGui_ImplVulkan_NewFrame();
 		ImGui::NewFrame();
 		DrawSplash(frame_extent);
+		const float performance_bottom = DrawPerformanceOverlay(graphics, frame_extent);
 		GetOverlaySnapshot(&snapshot);
 		if (has_overlay && snapshot.session == prepared_session) {
 			if (snapshot.session.kind == OverlayKind::Dialog) {
@@ -1355,7 +1839,8 @@ struct SystemOverlay::Impl {
 				DrawIme(snapshot.ime, frame_extent);
 			}
 		}
-		DrawTrophyNotification(frame_extent, now);
+		DrawTrophyNotification(frame_extent, now, performance_bottom);
+		DrawTouchPadCursor(GetHostInputTouchPadVisualState());
 		ImGui::Render();
 		extent = frame_extent;
 		return true;
@@ -1410,7 +1895,20 @@ struct SystemOverlay::Impl {
 	ImVec2                                right_stick {};
 	OverlaySession                              session;
 	std::unique_ptr<ImTextureData>              trophy_notification_image;
+	std::unique_ptr<ImTextureData>              hand_cursor_image;
+	std::unique_ptr<ImTextureData>              drawing_cursor_image;
+	bool                                        cursor_initialized = false;
+	ImVec2                                      cursor_position {};
+	float                                       particle_distance  = 0.0f;
+	TouchParticleIcon                           next_particle_icon = TouchParticleIcon::Square;
+	Clock::time_point                           last_cursor_update;
+	uint64_t                                    cursor_gesture = 0;
+	std::vector<TouchParticle>                  particles;
+	std::vector<TouchTrailSegment>              trail_segments;
+	static constexpr std::size_t                MAX_TOUCH_PARTICLES      = 64;
+	static constexpr std::size_t                MAX_TOUCH_TRAIL_SEGMENTS = 128;
 	std::optional<TrophyNotification>           trophy_notification;
+	std::string                                 trophy_notification_subtitle;
 	Clock::time_point                           trophy_notification_born;
 	std::vector<std::unique_ptr<ImTextureData>> retired_trophy_images;
 	std::array<std::unique_ptr<DdsImage>, 2>      splash_dds;
@@ -3,6 +3,7 @@
 #include "common/abi.h"
 #include "common/assert.h"
 #include "common/common.h"
+#include "common/debugCounters.h"
 #include "common/emulatorConfig.h"
 #include "common/logging/log.h"
 #include "common/profiler.h"
@@ -15,6 +16,7 @@
 #include "graphics/host_gpu/renderer/image/imageInfo.h"
 #include "graphics/host_gpu/renderer/render.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
+#include "graphics/presentation/performanceOverlay.h"
 #include "graphics/presentation/presenter.h"
 #include "graphics/presentation/renderDoc.h"
 #include "kernel/pthread.h"
@@ -909,6 +911,9 @@ void VideoOutDriver::Impl::PresentThread(std::stop_token token) {

 		const auto frame_end = Common::Timer::QueryPerformanceCounter();
 		total_wait += static_cast<int64_t>(period) - static_cast<int64_t>(frame_end - frame_begin);
+		if (total_wait < 0) {
+			Common::DebugCounters::Add(Common::DebugCounters::Counter::LateVblanks);
+		}
 	}
 }

@@ -938,6 +943,7 @@ bool FlipQueue::Reserve(VideoOutConfig& cfg, int index, int64_t flip_arg, FlipRe
 		m_done_cond_var.SignalAll();
 	}
 	if (cfg.flip_status.flipPendingNum >= VIDEO_OUT_FLIP_QUEUE_CAPACITY) {
+		Common::DebugCounters::Add(Common::DebugCounters::Counter::FlipsRejected);
 		return false;
 	}
 	auto& pending = source == FlipRequestSource::GpuEop ? m_requests : m_cpu_requests;
@@ -965,6 +971,7 @@ bool FlipQueue::Reserve(VideoOutConfig& cfg, int index, int64_t flip_arg, FlipRe

 	pending.push_back(r);
 	request_id = r.id;
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::FlipsRequested);

 	cfg.flip_status.flipPendingNum++;
 	cfg.flip_status.submitProcessTimeCounter = r.submit_ptc;
@@ -1243,6 +1250,9 @@ bool FlipQueue::Flip(uint32_t micros) {
 	}
 	if (due) {
 		m_presenter.Present(std::span(layers.data(), count));
+		if (requests[0].cfg->bus == 0) {
+			Graphics::PerformanceOverlayRecordFlip(requests[0].cfg->width, requests[0].cfg->height);
+		}
 	}

 	m_mutex.Lock();
@@ -12,8 +12,10 @@
 #include <cmath>
 #include <cstddef>
 #include <cstdlib>
+#include <mutex>
 #include <string>
 #include <string_view>
+#include <unordered_map>

 namespace Libs::Graphics {

@@ -27,6 +29,8 @@ struct ControlInfo {
 	float            touch_x  = 0.0f;

 	Controller::Setting setting {};
+	bool                gyro          = false;
+	bool                touchpad_mode = false;
 };

 static constexpr std::array CONTROL_INFO = {
@@ -61,10 +65,66 @@ static constexpr std::array CONTROL_INFO = {
    ControlInfo {.name = "VibrationIntensity", .setting = Controller::Setting::VibrationIntensity},
    ControlInfo {.name    = "TriggerEffectIntensity",
                 .setting = Controller::Setting::TriggerEffectIntensity},
+    ControlInfo {.name = "GyroToggle", .gyro = true},
+    ControlInfo {.name = "TouchPadMode", .touchpad_mode = true},
 };

 constexpr std::size_t INVALID_CONTROL = CONTROL_INFO.size();

+struct PadSource {
+	std::string_view name;
+	int              button; // SDL_GamepadButton, or -1 for an analog trigger
+};
+
+constexpr std::array PAD_SOURCES = {
+    PadSource {"South", SDL_GAMEPAD_BUTTON_SOUTH},
+    PadSource {"East", SDL_GAMEPAD_BUTTON_EAST},
+    PadSource {"West", SDL_GAMEPAD_BUTTON_WEST},
+    PadSource {"North", SDL_GAMEPAD_BUTTON_NORTH},
+    PadSource {"Back", SDL_GAMEPAD_BUTTON_BACK},
+    PadSource {"Guide", SDL_GAMEPAD_BUTTON_GUIDE},
+    PadSource {"Start", SDL_GAMEPAD_BUTTON_START},
+    PadSource {"LeftStick", SDL_GAMEPAD_BUTTON_LEFT_STICK},
+    PadSource {"RightStick", SDL_GAMEPAD_BUTTON_RIGHT_STICK},
+    PadSource {"LeftShoulder", SDL_GAMEPAD_BUTTON_LEFT_SHOULDER},
+    PadSource {"RightShoulder", SDL_GAMEPAD_BUTTON_RIGHT_SHOULDER},
+    PadSource {"DpadUp", SDL_GAMEPAD_BUTTON_DPAD_UP},
+    PadSource {"DpadDown", SDL_GAMEPAD_BUTTON_DPAD_DOWN},
+    PadSource {"DpadLeft", SDL_GAMEPAD_BUTTON_DPAD_LEFT},
+    PadSource {"DpadRight", SDL_GAMEPAD_BUTTON_DPAD_RIGHT},
+    PadSource {"Misc1", SDL_GAMEPAD_BUTTON_MISC1},
+    PadSource {"RightPaddle1", SDL_GAMEPAD_BUTTON_RIGHT_PADDLE1},
+    PadSource {"LeftPaddle1", SDL_GAMEPAD_BUTTON_LEFT_PADDLE1},
+    PadSource {"RightPaddle2", SDL_GAMEPAD_BUTTON_RIGHT_PADDLE2},
+    PadSource {"LeftPaddle2", SDL_GAMEPAD_BUTTON_LEFT_PADDLE2},
+    PadSource {"Touchpad", SDL_GAMEPAD_BUTTON_TOUCHPAD},
+    PadSource {"LeftTrigger", -1},
+    PadSource {"RightTrigger", -1},
+};
+
+constexpr std::size_t LEFT_TRIGGER_SOURCE  = PAD_SOURCES.size() - 2;
+constexpr std::size_t RIGHT_TRIGGER_SOURCE = PAD_SOURCES.size() - 1;
+constexpr std::size_t INVALID_SOURCE       = PAD_SOURCES.size();
+
+struct PadDefault {
+	std::string_view control;
+	std::string_view source;
+};
+
+constexpr std::array PAD_DEFAULTS = {
+    PadDefault {"Cross", "South"},         PadDefault {"Circle", "East"},
+    PadDefault {"Square", "West"},         PadDefault {"Triangle", "North"},
+    PadDefault {"Options", "Start"},       PadDefault {"L3", "LeftStick"},
+    PadDefault {"R3", "RightStick"},       PadDefault {"L1", "LeftShoulder"},
+    PadDefault {"R1", "RightShoulder"},    PadDefault {"L2", "LeftTrigger"},
+    PadDefault {"R2", "RightTrigger"},     PadDefault {"Up", "DpadUp"},
+    PadDefault {"Down", "DpadDown"},       PadDefault {"Left", "DpadLeft"},
+    PadDefault {"Right", "DpadRight"},     PadDefault {"TouchPad", "Touchpad"},
+    PadDefault {"GyroToggle", "Back"},
+};
+
+constexpr std::string_view PAD_PREFIX = "Pad:";
+
 struct Binding {
 	SDL_Keycode key          = SDLK_UNKNOWN;
 	uint8_t     mouse_button = 0;
@@ -74,6 +134,7 @@ struct Binding {
 constexpr int              MOUSE_POLL_INTERVAL_MS = 33;
 constexpr uint64_t         CURSOR_IDLE_HIDE_MS    = 2000;
 constexpr std::string_view MOUSE_SENSITIVITY      = "MouseSensitivity=";
+constexpr std::string_view TOUCH_PARTICLE_SKIN    = "TouchParticleSkin=";

 struct MouseJoystickState {
 	bool     enabled   = false;
@@ -85,6 +146,38 @@ MouseJoystickState g_mouse;
 SDL_Window*        g_mouse_window   = nullptr;
 uint64_t           g_cursor_hide_at = 0;

+struct MouseTouchPadState {
+	bool     enabled        = false;
+	bool     cursor_visible = true;
+	bool     drawing        = false;
+	bool     position_valid = false;
+	float    x              = 0.5f;
+	float    y              = 0.5f;
+	uint64_t revision       = 0;
+	uint64_t gesture        = 0;
+};
+
+struct GamepadTouchPadState {
+	bool     enabled   = false;
+	bool     drawing   = false;
+	int      id        = -1;
+	int      x_axis    = 128;
+	int      y_axis    = 128;
+	float    x         = 0.5f;
+	float    y         = 0.5f;
+	uint64_t next_poll = 0;
+	uint64_t last_poll = 0;
+};
+
+constexpr int   GAMEPAD_TOUCHPAD_POLL_MS = 8;
+constexpr int   GAMEPAD_TOUCHPAD_DEADZONE = 18;
+constexpr float GAMEPAD_TOUCHPAD_SPEED    = 0.8f;
+
+MouseTouchPadState   g_mouse_touchpad;
+GamepadTouchPadState g_gamepad_touchpad;
+std::mutex           g_touchpad_mutex;
+std::unordered_map<int, std::array<bool, 2>> g_pad_trigger_pressed;
+
 std::size_t ControlFromName(std::string_view name) {
 	const auto info = std::find_if(CONTROL_INFO.begin(), CONTROL_INFO.end(),
 	                               [name](const auto& item) { return item.name == name; });
@@ -126,6 +219,14 @@ public:
 	InputMap() {
 		for (const auto& value: Config::GetKeymap()) {
 			const std::string_view entry = value;
+			if (entry == "Keyboard=Disabled") {
+				m_custom = true;
+				continue;
+			}
+			if (entry == "Pad=Disabled") {
+				m_custom_pad = true;
+				continue;
+			}
 			if (entry.starts_with(MOUSE_SENSITIVITY)) {
 				const float sensitivity =
 				    std::strtof(value.c_str() + MOUSE_SENSITIVITY.size(), nullptr);
@@ -133,6 +234,13 @@ public:
 				    std::clamp(std::isfinite(sensitivity) ? sensitivity : 1.0f, 0.1f, 5.0f);
 				continue;
 			}
+			if (entry.starts_with(TOUCH_PARTICLE_SKIN) || entry.starts_with("TouchCursorImage=")) {
+				continue;
+			}
+			if (entry.starts_with(PAD_PREFIX)) {
+				AddPadEntry(entry.substr(PAD_PREFIX.size()), value);
+				continue;
+			}
 			const auto split = entry.find('=');

 			Binding binding;
@@ -146,17 +254,33 @@ public:
 			}

 			const bool reserved = binding.key == SDLK_ESCAPE || binding.key == SDLK_F1 ||
-			                      binding.key == SDLK_F7 || binding.key == SDLK_F11;
+			                      binding.key == SDLK_F2 || binding.key == SDLK_F7 ||
+			                      binding.key == SDLK_F8 || binding.key == SDLK_F11;
 			if (binding.control == INVALID_CONTROL || reserved ||
 			    (binding.key == SDLK_UNKNOWN && binding.mouse_button == 0)) {
 				EXIT("Invalid input mapping: %s\n", value.c_str());
 			}
 			Add(binding);
 		}
+		if (!m_custom_pad) {
+			for (const auto& item: PAD_DEFAULTS) {
+				AssignPad(ControlFromName(item.control), SourceFromName(item.source));
+			}
+		}
 	}
-	[[nodiscard]] bool  Custom() const { return m_size != 0; }
+	[[nodiscard]] bool  Custom() const { return m_custom; }
 	[[nodiscard]] float MouseSensitivity() const { return m_mouse_sensitivity; }

+	[[nodiscard]] std::size_t PadControl(std::size_t source) const {
+		return source < m_pad_controls.size() ? m_pad_controls[source] : INVALID_CONTROL;
+	}
+
+	static std::size_t SourceFromButton(int button) {
+		const auto item = std::find_if(PAD_SOURCES.begin(), PAD_SOURCES.end(),
+		                               [button](const auto& s) { return s.button == button; });
+		return static_cast<std::size_t>(std::distance(PAD_SOURCES.begin(), item));
+	}
+
 	[[nodiscard]] std::size_t FindKey(int key_code) const {
 		key_code = NormalizeKey(static_cast<SDL_Keycode>(key_code));
 		const auto binding =
@@ -173,7 +297,39 @@ public:
 	}

 private:
+	static std::size_t SourceFromName(std::string_view name) {
+		const auto item = std::find_if(PAD_SOURCES.begin(), PAD_SOURCES.end(),
+		                               [name](const auto& s) { return s.name == name; });
+		return static_cast<std::size_t>(std::distance(PAD_SOURCES.begin(), item));
+	}
+
+	void AssignPad(std::size_t control, std::size_t source) {
+		for (auto& existing: m_pad_controls) {
+			if (existing == control) {
+				existing = INVALID_CONTROL;
+			}
+		}
+		m_pad_controls[source] = control;
+	}
+
+	void AddPadEntry(std::string_view entry, const std::string& original) {
+		const auto split   = entry.find('=');
+		const auto control = split != std::string_view::npos ? ControlFromName(entry.substr(0, split))
+		                                                     : INVALID_CONTROL;
+		const auto source  = split != std::string_view::npos ? SourceFromName(entry.substr(split + 1))
+		                                                     : INVALID_SOURCE;
+		if (control == INVALID_CONTROL || source == INVALID_SOURCE) {
+			EXIT("Invalid input mapping: %s\n", original.c_str());
+		}
+		const auto& info = CONTROL_INFO[control];
+		if (info.button == 0 && !info.gyro && info.axis != Controller::Axis::AxisMax) {
+			EXIT("Invalid input mapping: %s\n", original.c_str());
+		}
+		m_custom_pad = true;
+		AssignPad(control, source);
+	}
 	void Add(const Binding& binding) {
+		m_custom = true;
 		for (std::size_t index = 0; index < m_size;) {
 			if (Conflicts(m_bindings[index], binding)) {
 				m_bindings[index] = m_bindings[--m_size];
@@ -186,6 +342,13 @@ private:
 	}

 	std::array<Binding, CONTROL_INFO.size()> m_bindings {};
+	std::array<std::size_t, PAD_SOURCES.size()> m_pad_controls = [] {
+		std::array<std::size_t, PAD_SOURCES.size()> a {};
+		a.fill(INVALID_CONTROL);
+		return a;
+	}();
+	bool                                     m_custom = false;
+	bool                                     m_custom_pad = false;
 	std::size_t                              m_size              = 0;
 	float                                    m_mouse_sensitivity = 1.0f;
 };
@@ -211,6 +374,39 @@ void SetTouchPad(float x, bool down) {
 	Controller::SetTouchPad(Controller::HOST_INPUT_CONTROLLER_ID, 0, down, x, 0.5f);
 }

+void SetSwipeTouchPad(bool down, float x, float y) {
+	Controller::SetTouchPad(Controller::HOST_INPUT_CONTROLLER_ID, 1, down, x, y);
+}
+
+void RecordMouseTouchPadPosition(float x, float y) {
+	int width  = 0;
+	int height = 0;
+	if (g_mouse_window != nullptr) {
+		SDL_GetWindowSize(g_mouse_window, &width, &height);
+	}
+	if (width > 0 && height > 0) {
+		x = std::clamp(x / static_cast<float>(width), 0.0f, 1.0f);
+		y = std::clamp(y / static_cast<float>(height), 0.0f, 1.0f);
+	} else {
+		x = 0.5f;
+		y = 0.5f;
+	}
+
+	std::scoped_lock lock(g_touchpad_mutex);
+	if (!g_mouse_touchpad.enabled) {
+		return;
+	}
+	g_mouse_touchpad.x              = x;
+	g_mouse_touchpad.y              = y;
+	g_mouse_touchpad.position_valid = true;
+	g_mouse_touchpad.revision++;
+}
+
+bool TouchPadModeActive() {
+	std::scoped_lock lock(g_touchpad_mutex);
+	return g_mouse_touchpad.enabled || g_gamepad_touchpad.enabled;
+}
+
 uint32_t DefaultKeyboardButton(int key_code) {
 	switch (NormalizeKey(static_cast<SDL_Keycode>(key_code))) {
 		case SDLK_UP: return Controller::PAD_BUTTON_UP;
@@ -270,6 +466,18 @@ void SetControl(std::size_t control, bool down) {
 		SetButton(info.button, down);
 		return;
 	}
+	if (info.gyro) {
+		if (down) {
+			Controller::ToggleGyroEmulation(Controller::HOST_INPUT_CONTROLLER_ID);
+		}
+		return;
+	}
+	if (info.touchpad_mode) {
+		if (down) {
+			HostInputToggleTouchPadMode();
+		}
+		return;
+	}
 	if (info.axis == Controller::Axis::AxisMax) {
 		CycleSettingOnPress(info.setting, down);
 		return;
@@ -297,6 +505,11 @@ void DefaultKeyboardInput(int key_code, bool down) {

 	switch (NormalizeKey(static_cast<SDL_Keycode>(key_code))) {
 		case SDLK_BACKSPACE: SetTouchPad(0.25f, down); return;
+		case SDLK_F6:
+			if (down) {
+				HostInputToggleTouchPadMode();
+			}
+			return;
 		case SDLK_TAB: SetTouchPad(0.75f, down); return;
 		case SDLK_1: CycleSettingOnPress(Controller::Setting::SpeakerVolume, down); return;
 		case SDLK_2: CycleSettingOnPress(Controller::Setting::VibrationIntensity, down); return;
@@ -399,8 +612,134 @@ bool IsCursorActivity(const SDL_Event& event) {
 	return which != SDL_TOUCH_MOUSEID && which != SDL_PEN_MOUSEID;
 }

+std::array<int, 2> GamepadLeftStick(int id) {
+	auto* pad = SDL_GetGamepadFromID(id);
+	if (pad == nullptr) {
+		return {128, 128};
+	}
+	const auto value = [pad](SDL_GamepadAxis axis) {
+		return Controller::controller_get_axis(SDL_JOYSTICK_AXIS_MIN, SDL_JOYSTICK_AXIS_MAX,
+		                                       SDL_GetGamepadAxis(pad, axis));
+	};
+	return {value(SDL_GAMEPAD_AXIS_LEFTX), value(SDL_GAMEPAD_AXIS_LEFTY)};
+}
+
+void ApplyPadControl(int id, std::size_t control, std::size_t source, bool down) {
+	if (control == INVALID_CONTROL) {
+		return;
+	}
+	const auto& info = CONTROL_INFO[control];
+	if (info.gyro) {
+		if (down) {
+			Controller::ToggleGyroEmulation(id);
+		}
+	} else if (info.touchpad_mode) {
+		if (down) {
+			HostInputToggleGamepadTouchPad(id);
+		}
+	} else if (info.button == Controller::PAD_BUTTON_CROSS && HostInputGamepadDrawButton(id, down)) {
+		return;
+	} else if (info.button == Controller::PAD_BUTTON_TOUCH_PAD) {
+		// Real touchpads report their own finger position; other inputs emulate a click.
+		if (source != PAD_SOURCES.size() && PAD_SOURCES[source].button != SDL_GAMEPAD_BUTTON_TOUCHPAD) {
+			Controller::SetTouchPad(id, 0, down, info.touch_x != 0.0f ? info.touch_x : 0.5f, 0.5f);
+		}
+		Controller::SetButton(id, Controller::PAD_BUTTON_TOUCH_PAD, down);
+	} else if (info.button == Controller::PAD_BUTTON_L2) {
+		Controller::SetAxis(id, Controller::Axis::TriggerLeft, down ? 255 : 0);
+	} else if (info.button == Controller::PAD_BUTTON_R2) {
+		Controller::SetAxis(id, Controller::Axis::TriggerRight, down ? 255 : 0);
+	} else if (info.button != 0) {
+		Controller::SetButton(id, info.button, down);
+	} else if (info.axis == Controller::Axis::AxisMax) {
+		CycleSettingOnPress(info.setting, down);
+	}
+}
+
+int PollGamepadTouchPad(uint64_t now_ms) {
+	float x       = 0;
+	float y       = 0;
+	int   id      = -1;
+	bool  drawing = false;
+	{
+		std::scoped_lock lock(g_touchpad_mutex);
+		if (!g_gamepad_touchpad.enabled) {
+			return -1;
+		}
+		if (now_ms < g_gamepad_touchpad.next_poll) {
+			return static_cast<int>(g_gamepad_touchpad.next_poll - now_ms);
+		}
+
+		const float dt = g_gamepad_touchpad.last_poll == 0 || now_ms <= g_gamepad_touchpad.last_poll
+		                     ? 0.008f
+		                     : std::min(static_cast<float>(now_ms - g_gamepad_touchpad.last_poll) *
+		                                    0.001f,
+		                                0.05f);
+		g_gamepad_touchpad.last_poll = now_ms;
+		g_gamepad_touchpad.next_poll = now_ms + GAMEPAD_TOUCHPAD_POLL_MS;
+		const auto velocity          = [](int axis) {
+            const int offset = axis - 128;
+            if (std::abs(offset) <= GAMEPAD_TOUCHPAD_DEADZONE) {
+                return 0.0f;
+            }
+            const int adjusted =
+                offset > 0 ? offset - GAMEPAD_TOUCHPAD_DEADZONE : offset + GAMEPAD_TOUCHPAD_DEADZONE;
+            return static_cast<float>(adjusted) / static_cast<float>(127 - GAMEPAD_TOUCHPAD_DEADZONE);
+		};
+		const float old_x = g_gamepad_touchpad.x;
+		const float old_y = g_gamepad_touchpad.y;
+		g_gamepad_touchpad.x = std::clamp(
+		    old_x + velocity(g_gamepad_touchpad.x_axis) * GAMEPAD_TOUCHPAD_SPEED * dt, 0.0f, 1.0f);
+		g_gamepad_touchpad.y = std::clamp(
+		    old_y + velocity(g_gamepad_touchpad.y_axis) * GAMEPAD_TOUCHPAD_SPEED * dt, 0.0f, 1.0f);
+		if (old_x != g_gamepad_touchpad.x || old_y != g_gamepad_touchpad.y) {
+			g_mouse_touchpad.revision++;
+		}
+		id      = g_gamepad_touchpad.id;
+		x       = g_gamepad_touchpad.x;
+		y       = g_gamepad_touchpad.y;
+		drawing = g_gamepad_touchpad.drawing;
+	}
+	if (drawing) {
+		Controller::SetTouchPad(id, 1, true, x, y);
+	}
+	return GAMEPAD_TOUCHPAD_POLL_MS;
+}
+
 } // namespace

+void HostInputPadButton(int id, int button, bool down) {
+	const auto  source = InputMap::SourceFromButton(button);
+	const auto& map    = GetInputMap();
+	ApplyPadControl(id, map.PadControl(source), source, down);
+}
+
+void HostInputPadTrigger(int id, bool left, int value) {
+	const auto  source     = left ? LEFT_TRIGGER_SOURCE : RIGHT_TRIGGER_SOURCE;
+	const auto  control    = GetInputMap().PadControl(source);
+	if (control == INVALID_CONTROL) {
+		return;
+	}
+	const auto& info = CONTROL_INFO[control];
+	if (info.button == Controller::PAD_BUTTON_L2) {
+		Controller::SetAxis(id, Controller::Axis::TriggerLeft, value);
+	} else if (info.button == Controller::PAD_BUTTON_R2) {
+		Controller::SetAxis(id, Controller::Axis::TriggerRight, value);
+	} else {
+		auto& state = g_pad_trigger_pressed[id][left ? 0 : 1];
+		const bool down = state ? value > 32 : value > 96;
+		if (down != state) {
+			state = down;
+			ApplyPadControl(id, control, source, down);
+		}
+	}
+}
+
+void HostInputDisconnectGamepad(int id) {
+	g_pad_trigger_pressed.erase(id);
+	HostInputDisconnectGamepadTouchPad(id);
+}
+
 void HostInputInit(SDL_Window* window) {
 	GetInputMap();
 	g_mouse_window = window;
@@ -410,6 +749,16 @@ void HostInputInit(SDL_Window* window) {
 }

 void HostInputShutdown() {
+	g_pad_trigger_pressed.clear();
+	HostInputSetTouchPadMode(false);
+	int gamepad_touch_id = -1;
+	{
+		std::scoped_lock lock(g_touchpad_mutex);
+		gamepad_touch_id = g_gamepad_touchpad.enabled ? g_gamepad_touchpad.id : -1;
+	}
+	if (gamepad_touch_id >= 0) {
+		HostInputDisconnectGamepadTouchPad(gamepad_touch_id);
+	}
 	if (g_mouse.enabled) {
 		SetRelativeMouseMode(false);
 		CenterMouseStick();
@@ -438,6 +787,219 @@ void HostInputMouseButton(uint8_t mouse_button, bool down) {
 	}
 }

+void HostInputMouseTouchPad(bool down, float x, float y) {
+	RecordMouseTouchPadPosition(x, y);
+	{
+		std::scoped_lock lock(g_touchpad_mutex);
+		if (!g_mouse_touchpad.enabled) {
+			return;
+		}
+		if (down && !g_mouse_touchpad.drawing) {
+			g_mouse_touchpad.gesture++;
+		}
+		g_mouse_touchpad.drawing = down;
+		g_mouse_touchpad.revision++;
+		x = g_mouse_touchpad.x;
+		y = g_mouse_touchpad.y;
+	}
+	SetSwipeTouchPad(down, x, y);
+}
+
+void HostInputMouseMotion(float x, float y) {
+	RecordMouseTouchPadPosition(x, y);
+	std::scoped_lock lock(g_touchpad_mutex);
+	if (g_mouse_touchpad.enabled && g_mouse_touchpad.drawing) {
+		SetSwipeTouchPad(true, g_mouse_touchpad.x, g_mouse_touchpad.y);
+	}
+}
+
+void HostInputSetTouchPadMode(bool enabled) {
+	if (enabled) {
+		int gamepad_id = -1;
+		{
+			std::scoped_lock lock(g_touchpad_mutex);
+			if (g_gamepad_touchpad.enabled) {
+				gamepad_id = g_gamepad_touchpad.id;
+			}
+		}
+		if (gamepad_id >= 0) {
+			HostInputDisconnectGamepadTouchPad(gamepad_id);
+		}
+	}
+	{
+		std::scoped_lock lock(g_touchpad_mutex);
+		if (g_mouse_touchpad.enabled == enabled) {
+			return;
+		}
+	}
+	if (!enabled) {
+		float x = 0;
+		float y = 0;
+		(void) SDL_GetMouseState(&x, &y);
+		HostInputMouseTouchPad(false, x, y);
+	}
+	bool show_system_cursor = false;
+	{
+		std::scoped_lock lock(g_touchpad_mutex);
+		g_mouse_touchpad.enabled        = enabled;
+		g_mouse_touchpad.cursor_visible = true;
+		g_mouse_touchpad.drawing        = false;
+		g_mouse_touchpad.position_valid = false;
+		g_mouse_touchpad.gesture++;
+		g_mouse_touchpad.revision++;
+		show_system_cursor = !enabled && !g_gamepad_touchpad.enabled;
+	}
+	if (enabled) {
+		if (!SDL_HideCursor()) {
+			LOGF("Failed to hide mouse cursor for touchpad mode: %s\n", SDL_GetError());
+		}
+	} else if (show_system_cursor && !Config::HideCursorEnabled() && !SDL_ShowCursor()) {
+		LOGF("Failed to restore mouse cursor after touchpad mode: %s\n", SDL_GetError());
+	}
+	LOGF("Mouse touchpad mode: %s\n", enabled ? "enabled" : "disabled");
+}
+
+void HostInputToggleTouchPadMode() {
+	bool enabled = false;
+	{
+		std::scoped_lock lock(g_touchpad_mutex);
+		enabled = !g_mouse_touchpad.enabled;
+	}
+	HostInputSetTouchPadMode(enabled);
+}
+
+void HostInputToggleGamepadTouchPad(int id) {
+	bool mouse_mode = false;
+	{
+		std::scoped_lock lock(g_touchpad_mutex);
+		mouse_mode = g_mouse_touchpad.enabled;
+	}
+	if (mouse_mode) {
+		HostInputSetTouchPadMode(false);
+	}
+
+	const auto stick = GamepadLeftStick(id);
+	bool  enabled = false;
+	float x       = 0.5f;
+	float y       = 0.5f;
+	{
+		std::scoped_lock lock(g_touchpad_mutex);
+		if (g_gamepad_touchpad.enabled && g_gamepad_touchpad.id != id) {
+			return;
+		}
+		enabled = !g_gamepad_touchpad.enabled;
+		if (enabled) {
+			g_gamepad_touchpad           = {};
+			g_gamepad_touchpad.enabled   = true;
+			g_gamepad_touchpad.id        = id;
+			g_gamepad_touchpad.x_axis    = stick[0];
+			g_gamepad_touchpad.y_axis    = stick[1];
+			g_gamepad_touchpad.last_poll = SDL_GetTicks();
+			g_gamepad_touchpad.next_poll = g_gamepad_touchpad.last_poll + GAMEPAD_TOUCHPAD_POLL_MS;
+			g_mouse_touchpad.gesture++;
+		} else {
+			x                          = g_gamepad_touchpad.x;
+			y                          = g_gamepad_touchpad.y;
+			g_gamepad_touchpad         = {};
+		}
+		g_mouse_touchpad.revision++;
+	}
+	Controller::SetAxis(id, Controller::Axis::LeftX, enabled ? 128 : stick[0]);
+	Controller::SetAxis(id, Controller::Axis::LeftY, enabled ? 128 : stick[1]);
+	if (enabled) {
+		if (!SDL_HideCursor()) {
+			LOGF("Failed to hide mouse cursor for gamepad touchpad mode: %s\n", SDL_GetError());
+		}
+	} else {
+		Controller::SetTouchPad(id, 1, false, x, y);
+		if (!Config::HideCursorEnabled() && !SDL_ShowCursor()) {
+			LOGF("Failed to restore mouse cursor after gamepad touchpad mode: %s\n",
+			     SDL_GetError());
+		}
+	}
+	LOGF("Gamepad touchpad mode: %s\n", enabled ? "enabled" : "disabled");
+}
+
+bool HostInputGamepadDrawButton(int id, bool down) {
+	float x = 0;
+	float y = 0;
+	{
+		std::scoped_lock lock(g_touchpad_mutex);
+		if (!g_gamepad_touchpad.enabled || g_gamepad_touchpad.id != id) {
+			return false;
+		}
+		if (!down && !g_gamepad_touchpad.drawing) {
+			return false;
+		}
+		if (down && !g_gamepad_touchpad.drawing) {
+			g_mouse_touchpad.gesture++;
+		}
+		g_gamepad_touchpad.drawing = down;
+		x                          = g_gamepad_touchpad.x;
+		y                          = g_gamepad_touchpad.y;
+		g_mouse_touchpad.revision++;
+	}
+	Controller::SetTouchPad(id, 1, down, x, y);
+	return true;
+}
+
+bool HostInputGamepadAxis(int id, bool horizontal, int value) {
+	std::scoped_lock lock(g_touchpad_mutex);
+	if (!g_gamepad_touchpad.enabled || g_gamepad_touchpad.id != id) {
+		return false;
+	}
+	(horizontal ? g_gamepad_touchpad.x_axis : g_gamepad_touchpad.y_axis) = std::clamp(value, 0, 255);
+	g_gamepad_touchpad.next_poll = SDL_GetTicks();
+	return true;
+}
+
+void HostInputDisconnectGamepadTouchPad(int id) {
+	float x = 0;
+	float y = 0;
+	{
+		std::scoped_lock lock(g_touchpad_mutex);
+		if (!g_gamepad_touchpad.enabled || g_gamepad_touchpad.id != id) {
+			return;
+		}
+		x                  = g_gamepad_touchpad.x;
+		y                  = g_gamepad_touchpad.y;
+		g_gamepad_touchpad = {};
+		g_mouse_touchpad.revision++;
+	}
+	Controller::SetTouchPad(id, 1, false, x, y);
+	const auto stick = GamepadLeftStick(id);
+	Controller::SetAxis(id, Controller::Axis::LeftX, stick[0]);
+	Controller::SetAxis(id, Controller::Axis::LeftY, stick[1]);
+	if (!TouchPadModeActive() && !Config::HideCursorEnabled() && !SDL_ShowCursor()) {
+		LOGF("Failed to restore mouse cursor after gamepad disconnect: %s\n", SDL_GetError());
+	}
+}
+
+HostInputTouchPadVisualState GetHostInputTouchPadVisualState() {
+	std::scoped_lock             lock(g_touchpad_mutex);
+	HostInputTouchPadVisualState state {};
+	if (g_gamepad_touchpad.enabled) {
+		state.enabled        = true;
+		state.cursor_visible = true;
+		state.drawing        = g_gamepad_touchpad.drawing;
+		state.x              = g_gamepad_touchpad.x;
+		state.y              = g_gamepad_touchpad.y;
+		state.revision       = g_mouse_touchpad.revision;
+		state.gesture        = g_mouse_touchpad.gesture;
+		state.position_valid = true;
+		return state;
+	}
+	state.enabled        = g_mouse_touchpad.enabled;
+	state.cursor_visible = g_mouse_touchpad.cursor_visible;
+	state.drawing        = g_mouse_touchpad.drawing;
+	state.x              = g_mouse_touchpad.x;
+	state.y              = g_mouse_touchpad.y;
+	state.revision       = g_mouse_touchpad.revision;
+	state.gesture        = g_mouse_touchpad.gesture;
+	state.position_valid = g_mouse_touchpad.position_valid;
+	return state;
+}
+
 void HostInputToggleMouseToJoystick() {
 	if (g_mouse.enabled) {
 		SetRelativeMouseMode(false);
@@ -467,16 +1029,19 @@ bool HostInputWaitEvent(SDL_Event* event, int max_wait_ms) {
 		timeout = PollMouse(SDL_GetTicks());
 	}

-	if (g_cursor_hide_at != 0) {
+	if (g_cursor_hide_at != 0 && !TouchPadModeActive()) {
 		const auto now_ms = SDL_GetTicks();
 		const int  cursor_timeout =
 		    now_ms < g_cursor_hide_at ? static_cast<int>(g_cursor_hide_at - now_ms) : 0;
 		timeout = timeout < 0 ? cursor_timeout : std::min(timeout, cursor_timeout);
 	}
+	if (const int touch_timeout = PollGamepadTouchPad(SDL_GetTicks()); touch_timeout >= 0) {
+		timeout = timeout < 0 ? touch_timeout : std::min(timeout, touch_timeout);
+	}
 	timeout              = timeout < 0 ? max_wait_ms : std::min(timeout, max_wait_ms);
 	const bool has_event = SDL_WaitEventTimeout(event, timeout);

-	if (Config::HideCursorEnabled()) {
+	if (Config::HideCursorEnabled() && !TouchPadModeActive()) {
 		const auto now_ms = SDL_GetTicks();
 		if (has_event && !g_mouse.enabled && IsCursorActivity(*event) &&
 		    SDL_GetWindowFromEvent(event) == g_mouse_window) {
@@ -8,10 +8,33 @@ union SDL_Event;

 namespace Libs::Graphics {

+struct HostInputTouchPadVisualState {
+	bool     enabled        = false;
+	bool     cursor_visible = true;
+	bool     drawing        = false;
+	bool     position_valid = false;
+	float    x              = 0.5f;
+	float    y              = 0.5f;
+	uint64_t revision       = 0;
+	uint64_t gesture        = 0;
+};
+
 void               HostInputInit(SDL_Window* window);
 void               HostInputShutdown();
 void               HostInputKey(int key_code, bool down);
 void               HostInputMouseButton(uint8_t mouse_button, bool down);
+void               HostInputMouseTouchPad(bool down, float x, float y);
+void               HostInputMouseMotion(float x, float y);
+void               HostInputPadButton(int id, int sdl_button, bool down);
+void               HostInputPadTrigger(int id, bool left, int value);
+void               HostInputDisconnectGamepad(int id);
+void               HostInputSetTouchPadMode(bool enabled);
+void               HostInputToggleTouchPadMode();
+void               HostInputToggleGamepadTouchPad(int id);
+[[nodiscard]] bool HostInputGamepadDrawButton(int id, bool down);
+[[nodiscard]] bool HostInputGamepadAxis(int id, bool horizontal, int value);
+void               HostInputDisconnectGamepadTouchPad(int id);
+[[nodiscard]] HostInputTouchPadVisualState GetHostInputTouchPadVisualState();
 void               HostInputToggleMouseToJoystick();
 [[nodiscard]] bool HostInputWaitEvent(SDL_Event* event, int max_wait_ms);

@@ -1,5 +1,6 @@
 #include "common/assert.h"
 #include "common/common.h"
+#include "common/debugCounters.h"
 #include "common/emulatorConfig.h"
 #include "common/logging/log.h"
 #include "common/profiler.h"
@@ -10,6 +11,7 @@
 #include "graphics/host_gpu/renderer/render.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
 #include "graphics/host_gpu/vulkanCommon.h"
+#include "graphics/presentation/performanceOverlay.h"
 #include "graphics/presentation/presenter.h"
 #include "graphics/presentation/systemOverlay.h"
 #include "graphics/presentation/videoOut.h"
@@ -324,6 +326,7 @@ struct Presenter::Impl {
 	void RecoverSwapchain(Swapchain::Status status) {
 		LOGF("Recovering Vulkan swapchain%s\n",
 		     status == Swapchain::Status::SurfaceLost ? " and surface" : "");
+		Common::DebugCounters::Add(Common::DebugCounters::Counter::SwapchainRecreations);
 		swapchain.Recreate(status == Swapchain::Status::SurfaceLost);
 		frames.SetFormat(swapchain.Format());
 	}
@@ -945,7 +948,6 @@ Presenter::Frame& Presenter::PrepareBlankFrame(uint32_t width, uint32_t height,
 	auto*             frame  = m_impl->frames.Acquire({width, height}, format);
 	frame->hdr   = false;
 	frame->blank = true;
-	Common::LockGuard render_lock(m_impl->renderer.GetMutex());
 	frame->Configure(m_impl->window.graphic_ctx, {width, height}, format);
 	vk::ClearColorValue clear {};
 	clear.float32[3] = opaque ? 1.0f : 0.0f;
@@ -953,6 +955,7 @@ Presenter::Frame& Presenter::PrepareBlankFrame(uint32_t width, uint32_t height,
 		EXIT_IF(producer->IsInvalid());
 		frame->Clear(*producer, clear);
 	} else {
+		Common::LockGuard lock(m_impl->present_mutex);
 		auto& command = m_impl->present_scheduler.BeginCommand();
 		frame->Clear(command, clear);
 		frame->present_tick = m_impl->present_scheduler.Submit();
@@ -1036,18 +1039,17 @@ void Presenter::Impl::Present() {
 			RecoverSwapchain(status);
 			continue;
 		}
-		{
-			Common::LockGuard render_lock(renderer.GetMutex());
-			auto&             command = present_scheduler.BeginCommand();
-			const bool        draw_system_overlay =
-			    overlay_visual.active && swapchain.PrepareSystemOverlay();
-			swapchain.RecordPresentCommands(command, layers[0].frame, layers[1],
-			                                draw_system_overlay);
-			const auto tick = swapchain.Submit(present_scheduler);
-			for (const auto& layer: layers) {
-				if (layer.frame != nullptr) {
-					layer.frame->present_tick = tick;
-				}
+		// present_mutex owns the prepared frames and this scheduler. Only queue submission
+		// is shared with rendering, and it has its own queue_mutex.
+		auto&      command = present_scheduler.BeginCommand();
+		const bool draw_system_overlay = (overlay_visual.active || PerformanceOverlayEnabled()) &&
+		                                 swapchain.PrepareSystemOverlay();
+		swapchain.RecordPresentCommands(command, layers[0].frame, layers[1],
+		                                draw_system_overlay);
+		const auto tick = swapchain.Submit(present_scheduler);
+		for (const auto& layer: layers) {
+			if (layer.frame != nullptr) {
+				layer.frame->present_tick = tick;
 			}
 		}
 		status = swapchain.Present();
@@ -1061,6 +1063,7 @@ void Presenter::Impl::Present() {
 		return;
 	}
 	LOGF("Vulkan presentation retry exhausted; dropping frame\n");
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::DroppedPresents);
 }

 void Presenter::Discard(Frame& frame) {
@@ -266,7 +266,6 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 #endif
 #if !defined(__APPLE__)
 		check_feature(device_features2.features.depthClamp, "depthClamp");
-		check_feature(fragment_barycentric.fragmentShaderBarycentric, "fragmentShaderBarycentric");
 #endif

 		check_feature(features12.samplerMirrorClampToEdge, "samplerMirrorClampToEdge",
@@ -498,7 +497,21 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 		image_atomic_int64.pNext = supported_features2.pNext;
 		supported_features2.pNext = &image_atomic_int64;
 	}
+	const bool barycentric_extension =
+	    HasExtension(device_extensions, VK_KHR_FRAGMENT_SHADER_BARYCENTRIC_EXTENSION_NAME);
+	vk::PhysicalDeviceFragmentShaderBarycentricFeaturesKHR supported_barycentric {};
+	if (barycentric_extension) {
+		supported_barycentric.pNext = supported_features2.pNext;
+		supported_features2.pNext   = &supported_barycentric;
+	}
 	physical_device.getFeatures2(&supported_features2);
+	const bool barycentric_supported =
+	    barycentric_extension && supported_barycentric.fragmentShaderBarycentric == VK_TRUE;
+	if (!barycentric_supported) {
+		Log::WriteToConsoleAndLog(
+		    "WARNING: Your GPU is not fully supported (fragmentShaderBarycentric is unavailable). "
+		    "Some games may render incorrectly or crash.\n");
+	}
 	graphics.shader_image_int64_atomics_enabled = image_atomic_int64.shaderImageInt64Atomics;

 	auto features12 = WindowContext::RequiredVulkan12Features();
@@ -610,9 +623,13 @@ static vk::Device VulkanCreateDevice(GraphicContext& graphics,
 	robustness2.pNext = &features12;
 #else
 	vk::PhysicalDeviceFragmentShaderBarycentricFeaturesKHR fragment_barycentric {};
-	fragment_barycentric.pNext                     = &features12;
-	fragment_barycentric.fragmentShaderBarycentric = VK_TRUE;
-	robustness2.pNext                              = &fragment_barycentric;
+	if (barycentric_supported) {
+		fragment_barycentric.pNext                     = &features12;
+		fragment_barycentric.fragmentShaderBarycentric = VK_TRUE;
+		robustness2.pNext                              = &fragment_barycentric;
+	} else {
+		robustness2.pNext = &features12;
+	}
 #endif
 	robustness2.robustBufferAccess2 = supported_robustness2.robustBufferAccess2;
 	robustness2.robustImageAccess2  = supported_robustness2.robustImageAccess2;
@@ -966,7 +983,6 @@ void WindowContext::CreateVulkan() {
 #else
 	device_extensions.push_back(VK_EXT_DEPTH_CLIP_ENABLE_EXTENSION_NAME);
 	device_extensions.push_back(VK_EXT_COLOR_WRITE_ENABLE_EXTENSION_NAME);
-	device_extensions.push_back(VK_KHR_FRAGMENT_SHADER_BARYCENTRIC_EXTENSION_NAME);
 #endif

 #ifdef KYTY_ENABLE_DEBUG_PRINTF
@@ -1024,7 +1040,8 @@ void WindowContext::CreateVulkan() {
 		                             VK_EXT_PROVOKING_VERTEX_EXTENSION_NAME,
 		                             VK_EXT_MESH_SHADER_EXTENSION_NAME,
 		                             VK_KHR_WORKGROUP_MEMORY_EXPLICIT_LAYOUT_EXTENSION_NAME,
-		                             VK_EXT_DEPTH_RANGE_UNRESTRICTED_EXTENSION_NAME}) {
+		                             VK_EXT_DEPTH_RANGE_UNRESTRICTED_EXTENSION_NAME,
+		                             VK_KHR_FRAGMENT_SHADER_BARYCENTRIC_EXTENSION_NAME}) {
 			if (HasExtension(available_extensions, extension)) {
 				device_extensions.push_back(extension);
 			}
@@ -1,6 +1,4 @@
-#include "graphics/presentation/window.h"
-
-#include <SDL3/SDL.h>
+#include "graphics/presentation/window.h"

 #include "common/assert.h"
 #include "common/common.h"
@@ -8,14 +6,15 @@
 #include "common/file.h"
 #include "common/logging/log.h"
 #include "common/profiler.h"
+#include "common/stringUtils.h"
 #include "common/systemInfo.h"
 #include "common/threads.h"
 #include "common/timer.h"
-#include "common/stringUtils.h"
 #include "graphics/host_gpu/graphicContext.h"
 #include "graphics/host_gpu/renderer/render.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
 #include "graphics/host_gpu/vulkanCommon.h"
+#include "graphics/presentation/performanceOverlay.h"
 #include "graphics/presentation/renderDoc.h"
 #include "graphics/presentation/systemOverlay.h"
 #include "graphics/presentation/window/hostInput.h"
@@ -24,13 +23,14 @@
 #include "libs/controller.h"
 #include "loader/systemContent.h"

+#include <SDL3/SDL.h>
 #include <cstdlib>
+#include <filesystem>
 #include <fmt/format.h>
 #include <memory>
 #include <string>
 #include <vector>
 #include <vulkan/vk_platform.h>
-#include <filesystem>

 #define STB_IMAGE_IMPLEMENTATION
 #define STBI_NO_SIMD
@@ -54,25 +54,6 @@ struct EventKeyboard {
 	double   timestamp_seconds;
 };

-static uint32_t ControllerButtonToPadButton(int button) {
-	switch (button) {
-		case SDL_GAMEPAD_BUTTON_SOUTH: return Controller::PAD_BUTTON_CROSS;
-		case SDL_GAMEPAD_BUTTON_EAST: return Controller::PAD_BUTTON_CIRCLE;
-		case SDL_GAMEPAD_BUTTON_WEST: return Controller::PAD_BUTTON_SQUARE;
-		case SDL_GAMEPAD_BUTTON_NORTH: return Controller::PAD_BUTTON_TRIANGLE;
-		case SDL_GAMEPAD_BUTTON_START: return Controller::PAD_BUTTON_OPTIONS;
-		case SDL_GAMEPAD_BUTTON_LEFT_STICK: return Controller::PAD_BUTTON_L3;
-		case SDL_GAMEPAD_BUTTON_RIGHT_STICK: return Controller::PAD_BUTTON_R3;
-		case SDL_GAMEPAD_BUTTON_LEFT_SHOULDER: return Controller::PAD_BUTTON_L1;
-		case SDL_GAMEPAD_BUTTON_RIGHT_SHOULDER: return Controller::PAD_BUTTON_R1;
-		case SDL_GAMEPAD_BUTTON_DPAD_UP: return Controller::PAD_BUTTON_UP;
-		case SDL_GAMEPAD_BUTTON_DPAD_DOWN: return Controller::PAD_BUTTON_DOWN;
-		case SDL_GAMEPAD_BUTTON_DPAD_LEFT: return Controller::PAD_BUTTON_LEFT;
-		case SDL_GAMEPAD_BUTTON_DPAD_RIGHT: return Controller::PAD_BUTTON_RIGHT;
-		case SDL_GAMEPAD_BUTTON_TOUCHPAD: return Controller::PAD_BUTTON_TOUCH_PAD;
-		default: return 0;
-	}
-}

 static Controller::Axis ControllerAxisFromSdl(int axis_id) {
 	switch (axis_id) {
@@ -268,6 +249,15 @@ static void GameEventMouse([[maybe_unused]] const EventMouse& mb) {
 		}

 		HostInputMouseButton(mouse_button, mb.down);
+		if (!mb.touch && mb.left) {
+			HostInputMouseTouchPad(mb.down, static_cast<float>(mb.x), static_cast<float>(mb.y));
+		}
+	} else if (mb.motion && !mb.touch) {
+		if (mb.left) {
+			HostInputMouseTouchPad(true, static_cast<float>(mb.x), static_cast<float>(mb.y));
+		} else {
+			HostInputMouseMotion(static_cast<float>(mb.x), static_cast<float>(mb.y));
+		}
 	}
 }

@@ -316,6 +306,7 @@ static void GameEventController([[maybe_unused]] const EventController& f) {
 	}

 	if (f.removed) {
+		HostInputDisconnectGamepad(f.id);
 		if (auto* pad = SDL_GetGamepadFromID(f.id); pad != nullptr) {
 			Controller::Disconnect(f.id);
 			SDL_CloseGamepad(pad);
@@ -323,30 +314,22 @@ static void GameEventController([[maybe_unused]] const EventController& f) {
 	}

 	if (f.down || f.up) {
-		const auto button = ControllerButtonToPadButton(f.button);
-		if (button != 0) {
-			Controller::SetButton(f.id, button, f.down);
-		} else if (f.button == SDL_GAMEPAD_BUTTON_BACK) {
-			// Pads without a touchpad (Switch, Xbox, ...) have no way to click it, which
-			// blocks games that require it. Use their Back/Select button as a centered click.
-			auto* pad = SDL_GetGamepadFromID(f.id);
-			if (pad != nullptr && SDL_GetNumGamepadTouchpads(pad) == 0) {
-				if (f.down) {
-					Controller::SetTouchPad(f.id, 0, true, 0.5f, 0.5f);
-				}
-				Controller::SetButton(f.id, Controller::PAD_BUTTON_TOUCH_PAD, f.down);
-				if (f.up) {
-					Controller::SetTouchPad(f.id, 0, false, 0.5f, 0.5f);
-				}
-			}
-		}
+		HostInputPadButton(f.id, f.button, f.down);
 	}

 	if (f.axis) {
-		const auto axis = ControllerAxisFromSdl(f.axis_id);
-		if (axis != Controller::Axis::AxisMax) {
-			Controller::SetAxis(f.id, axis,
+		if (f.axis_id == SDL_GAMEPAD_AXIS_LEFT_TRIGGER ||
+		    f.axis_id == SDL_GAMEPAD_AXIS_RIGHT_TRIGGER) {
+			HostInputPadTrigger(f.id, f.axis_id == SDL_GAMEPAD_AXIS_LEFT_TRIGGER,
 			                    ControllerAxisValueFromSdl(f.axis_id, f.axis_value));
+		} else if (const auto axis = ControllerAxisFromSdl(f.axis_id);
+		           axis != Controller::Axis::AxisMax) {
+			const int value = ControllerAxisValueFromSdl(f.axis_id, f.axis_value);
+			if ((f.axis_id == SDL_GAMEPAD_AXIS_LEFTX || f.axis_id == SDL_GAMEPAD_AXIS_LEFTY) &&
+			    HostInputGamepadAxis(f.id, f.axis_id == SDL_GAMEPAD_AXIS_LEFTX, value)) {
+				return;
+			}
+			Controller::SetAxis(f.id, axis, value);
 		}
 	}
 }
@@ -431,6 +414,7 @@ void WindowContext::ProcessWindowEvent(const SDL_WindowEvent& event) {
 			break;
 		case SDL_EVENT_WINDOW_MOUSE_LEAVE:
 			LOGF("Mouse left window %" PRIu32 "\n", window_event.windowID);
+			HostInputSetTouchPadMode(false);
 			break;
 		case SDL_EVENT_WINDOW_FOCUS_GAINED:
 			LOGF("Window %" PRIu32 " gained keyboard focus\n", window_event.windowID);
@@ -475,12 +459,37 @@ void WindowContext::ProcessDisplayEvent(const SDL_DisplayEvent& display) {
 void WindowContext::ProcessEvent(double time_s) {
 	auto& game  = loop;
 	auto* event = &game.event;
-	if ((event->type == SDL_EVENT_KEY_DOWN || event->type == SDL_EVENT_KEY_UP) &&
-	    event->key.key == SDLK_F7) {
-		if (event->type == SDL_EVENT_KEY_DOWN && event->key.repeat == 0) {
-			HostInputToggleMouseToJoystick();
+	if (event->type == SDL_EVENT_KEY_DOWN || event->type == SDL_EVENT_KEY_UP) {
+		const bool pressed = event->type == SDL_EVENT_KEY_DOWN && event->key.repeat == 0;
+		switch (event->key.key) {
+			case SDLK_F7:
+				if (pressed) {
+					HostInputToggleMouseToJoystick();
+				}
+				return;
+			case SDLK_F8:
+				if (pressed) {
+					Controller::CycleKeyboardPlayer();
+				}
+				return;
+#if defined(__APPLE__)
+			case SDLK_P:
+				if ((event->key.mod & SDL_KMOD_GUI) == 0) {
+					break;
+				}
+				[[fallthrough]];
+#endif
+			case SDLK_F2:
+				if (pressed) {
+					if ((event->key.mod & SDL_KMOD_SHIFT) != 0) {
+						TogglePerformanceOverlayDetails();
+					} else {
+						TogglePerformanceOverlay();
+					}
+				}
+				return;
+			default: break;
 		}
-		return;
 	}
 	if (ProcessSystemOverlayInput(*event)) {
 		return;
@@ -481,18 +481,14 @@ Decoder::Program DecodeFusedProgram(std::span<const uint32_t> front, std::span<c
 	return result;
 }

-void PrepareCallTarget(ShaderSource& source, const CompileOptions& options) {
+void PrepareCallTargets(ShaderSource& source, const CompileOptions& options) {
 	using namespace IR;
+	EXIT_NOT_IMPLEMENTED(options.stage != ShaderType::Compute);
 	for (uint32_t i = 0; i < source.decoded.instructions.size(); ++i) {
 		if (source.decoded.instructions[i].opcode != Decoder::Opcode::S_SWAPPC_B64) continue;
-		EXIT_NOT_IMPLEMENTED(source.call.has_value() || options.stage != ShaderType::Compute);
-		source.call = ShaderSource::Call {.instruction = i};
+		source.calls.push_back({.instruction = i});
 	}
 	EXIT_IF(options.input_info.compute == nullptr);
-	const auto& call = source.decoded.instructions[source.call->instruction];
-	EXIT_NOT_IMPLEMENTED(call.src0.kind != Decoder::OperandKind::Sgpr ||
-	                     call.dst.kind != Decoder::OperandKind::Sgpr ||
-	                     call.src0.reg != call.dst.reg || call.dst.reg >= 105u);
 	const auto graph = CFG::BuildGraph(source.decoded);
 	Program query;
 	query.stage = options.stage;
@@ -524,7 +520,11 @@ void PrepareCallTarget(ShaderSource& source, const CompileOptions& options) {
 		return value;
 	};
 	Register = [&](uint32_t before, uint32_t code) -> Value {
-		const auto index = CFG::FindScalarDefinition(source.decoded, graph, before, code);
+		std::vector<uint32_t> crossed_calls;
+		const auto index = CFG::FindScalarDefinition(source.decoded, graph, before, code, crossed_calls);
+		for (const auto crossed: crossed_calls) {
+			std::ranges::find(source.calls, crossed, &ShaderSource::Call::instruction)->preserved_registers.set(code);
+		}
 		if (index == UINT32_MAX) {
 			EXIT_IF(code < query.user_data_base || code - query.user_data_base >= query.user_data_count);
 			return IREmitter(&block).GetUserData(static_cast<ScalarReg>(code));
@@ -557,18 +557,30 @@ void PrepareCallTarget(ShaderSource& source, const CompileOptions& options) {
 		EXIT_IF(!producer.ready || !producer.values.contains(code));
 		return producer.values.at(code).Resolve();
 	};
-	for (uint32_t i = 0; i < 2u; ++i) {
-		const auto value = Register(source.call->instruction, call.src0.reg + i);
-		EXIT_IF(!ValidateRuntimeValue(query, value, RuntimeValueType::Integer));
-		auto& root = query.value_storage.emplace_back(ValueOpcode::ReferenceU32);
-		root.SetArg(0, value);
-		source.call->target[i] = Value(&root);
+	for (auto& site: source.calls) {
+		const auto& call = source.decoded.instructions[site.instruction];
+		EXIT_NOT_IMPLEMENTED(call.src0.kind != Decoder::OperandKind::Sgpr ||
+		                     call.dst.kind != Decoder::OperandKind::Sgpr ||
+		                     call.src0.reg != call.dst.reg || call.dst.reg >= 105u);
+		for (uint32_t i = 0; i < 2u; ++i) {
+			const auto value = Register(site.instruction, call.src0.reg + i);
+			if (!ValidateRuntimeValue(query, value, RuntimeValueType::Integer)) {
+				EXIT_IF(value.Instruction()->GetOpcode() != ValueOpcode::ReadConstBuffer);
+				source.unsupported_call_pc = call.pc;
+				source.calls.clear();
+				return;
+			}
+			auto& root = query.value_storage.emplace_back(ValueOpcode::ReferenceU32);
+			root.SetArg(0, value);
+			site.target[i] = Value(&root);
+		}
 	}
 	RewriteToSsa(query.blocks);
 	ConstantPropagationPass(query.blocks, query.wave_size);
 	RemoveIdentities(query);
 	EliminateDeadCode(query.blocks);
-	for (auto& value: source.call->target) value = value.Instruction()->Arg(0).Resolve();
+	for (auto& site: source.calls)
+		for (auto& value: site.target) value = value.Instruction()->Arg(0).Resolve();
 	query.value_storage.clear();
 	for (auto& inst: block) {
 		inst.SetParent(nullptr);
@@ -595,7 +607,7 @@ ShaderSource PrepareShaderSource(std::span<const uint32_t> code, const CompileOp
 		Decoder::DecodeProgram(code, source.decoded);
 	}
 	if (source.decoded.has_swap_pc) {
-		PrepareCallTarget(source, options);
+		PrepareCallTargets(source, options);
 		source.code.assign(code.begin(), code.end());
 		source.decoded.code = source.code;
 	}
@@ -603,88 +615,122 @@ ShaderSource PrepareShaderSource(std::span<const uint32_t> code, const CompileOp
 }

 const Decoder::Program& RefreshShaderSource(ShaderSource& source, const IR::SrtRuntime& runtime) {
+	EXIT_IF(source.unsupported_call_pc != UINT32_MAX);
 	auto& reads = source.reads;
 	reads.clear();
 	IR::SrtReadCapture capture(runtime, reads);
 	const auto clean = IR::CleanRuntime(capture.ObservedRuntime());
 	IR::SrtWalker walker(source.call_targets, clean);
-	uint32_t low = 0, high = 0;
-	const auto& call = source.decoded.instructions[source.call->instruction];
-	if (!walker.Evaluate(source.call->target[0], low) || !walker.Evaluate(source.call->target[1], high))
-		EXIT("shader call at pc 0x%08x has an unavailable scalar target", call.pc);
-	// SWAPPC reads the old aliased pair and ignores the target's low two bits.
-	const auto address = ((uint64_t {high} << 32u) | low) & ~uint64_t {3};
-	EXIT_IF(address == 0u);
+	for (auto& site: source.calls) {
+		uint32_t low = 0, high = 0;
+		if (!walker.Evaluate(site.target[0], low) || !walker.Evaluate(site.target[1], high))
+			EXIT("shader call at pc 0x%08x has an unavailable scalar target",
+			     source.decoded.instructions[site.instruction].pc);
+		// SWAPPC reads the old aliased pair and ignores the target's low two bits.
+		const auto address = ((uint64_t {high} << 32u) | low) & ~uint64_t {3};
+		EXIT_IF(address == 0u);
+		site.address = address;
+	}
 	const auto query_reads = reads.size();
 	if (source.linked) {
 		auto& linked = *source.linked;
-		const auto function = std::span(linked.code).subspan(source.code.size());
-		if (clean.read_memory(clean.userdata, address, linked.observed_function) &&
-		    std::ranges::equal(function, linked.observed_function)) {
-			return linked.decoded;
+		bool unchanged = true;
+		for (size_t i = 0; i < source.calls.size(); ++i) {
+			const auto& function = linked.functions[i];
+			const auto expected = std::span(linked.code).subspan(function.offset, function.size);
+			const auto preceding = std::span(source.calls).first(i);
+			const auto previous = std::ranges::find(preceding, source.calls[i].address, &ShaderSource::Call::address);
+			if (previous != preceding.end()) {
+				const auto& observed = linked.functions[previous - preceding.begin()];
+				if (std::ranges::equal(expected, std::span(linked.code).subspan(observed.offset, observed.size))) continue;
+				unchanged = false;
+				break;
+			}
+			linked.observed_function.resize(function.size);
+			if (!clean.read_memory(clean.userdata, source.calls[i].address, linked.observed_function) ||
+			    !std::ranges::equal(expected, linked.observed_function)) {
+				unchanged = false;
+				break;
+			}
 		}
+		if (unchanged) return linked.decoded;
 		reads.resize(query_reads);
 	}
-	ShaderSource::Linked linked {.code = source.code,
-	                             .decoded = source.decoded};
+	ShaderSource::Linked linked {.code = source.code, .decoded = source.decoded};
 	const auto& last = source.decoded.instructions.back();
 	EXIT_NOT_IMPLEMENTED(last.opcode != Decoder::Opcode::S_ENDPGM);
-	// Keep the caller's data/footer bytes at their native offsets. The leaf's virtual CFG
-	// PCs follow decoded instructions; its physical code is independently read and owned.
-	const auto function_pc = last.pc + last.word_count * sizeof(uint32_t);
-	for (uint32_t index = 0;;) {
-		// The observed queue helper is a scalar leaf. Scalar encodings need at most one literal.
-		EXIT_IF(index >= 64u || address > UINT64_MAX - (uint64_t {index} + 2u) * sizeof(uint32_t));
-		std::array<uint32_t, 2> words {};
-		if (!clean.read_memory(clean.userdata, address + uint64_t {index} * sizeof(uint32_t),
-		                       std::span(words).first(1)))
-			EXIT("shader call at pc 0x%08x cannot read leaf code", call.pc);
-		const auto family = Decoder::GetInstructionFamily(words[0]);
-		EXIT_NOT_IMPLEMENTED(family != Decoder::Family::SOP1 && family != Decoder::Family::SOP2 &&
-		                     family != Decoder::Family::SOPC && family != Decoder::Family::SOPK &&
-		                     family != Decoder::Family::SOPP);
-		Decoder::Instruction inst;
-		Decoder::DecodeInstruction(words, 0u, inst);
-		if (inst.word_count == 2u) {
-			if (!clean.read_memory(clean.userdata, address + uint64_t {index + 1u} * sizeof(uint32_t),
-			                       std::span(words).last(1)))
-				EXIT("shader call at pc 0x%08x cannot read leaf literal", call.pc);
-			inst = {};
-			Decoder::DecodeInstruction(words, 0u, inst);
+	// Keep the caller's data/footer at native offsets. Expand leaf CFGs per call site
+	// before SSA, so return edges neither require a dispatcher nor merge resource roots.
+	uint32_t function_pc = last.pc + last.word_count * sizeof(uint32_t);
+	std::map<uint64_t, std::vector<Decoder::Instruction>> functions;
+	for (size_t i = 0; i < source.calls.size(); ++i) {
+		const auto& site = source.calls[i];
+		const auto& call = source.decoded.instructions[site.instruction];
+		const auto address = source.calls[i].address;
+		auto [entry, inserted] = functions.try_emplace(address);
+		auto& function = entry->second;
+		if (inserted) {
+			for (uint32_t index = 0;;) {
+				std::array<uint32_t, Decoder::MaxInstructionRawWords> words {};
+				EXIT_IF(index > UINT32_MAX / sizeof(uint32_t) - words.size() ||
+				        address > UINT64_MAX - (uint64_t {index} + words.size()) * sizeof(uint32_t));
+				uint32_t count = 1;
+				Decoder::Instruction inst;
+				for (;;) {
+					if (!clean.read_memory(clean.userdata, address + uint64_t {index} * sizeof(uint32_t),
+					                       std::span(words).first(count)))
+						EXIT("shader call at pc 0x%08x cannot read leaf code", call.pc);
+					inst = {};
+					Decoder::DecodeInstruction(words, 0u, inst);
+					if (inst.word_count <= count) break;
+					count = inst.word_count;
+				}
+				inst.pc = index * sizeof(uint32_t);
+				if (Decoder::IsDirectBranch(inst.opcode)) inst.branch_target += inst.pc;
+				index += inst.word_count;
+				EXIT_NOT_IMPLEMENTED(inst.opcode == Decoder::Opcode::S_SWAPPC_B64 ||
+				                     inst.opcode == Decoder::Opcode::S_GETPC_B64 ||
+				                     inst.opcode == Decoder::Opcode::S_ENDPGM);
+				function.push_back(inst);
+				if (inst.opcode == Decoder::Opcode::S_SETPC_B64) break;
+			}
 		}
-		inst.pc = index * sizeof(uint32_t);
-		if (Decoder::IsDirectBranch(inst.opcode)) inst.branch_target += inst.pc;
-		linked.code.insert(linked.code.end(), inst.raw, inst.raw + inst.word_count);
-		index += inst.word_count;
-		EXIT_NOT_IMPLEMENTED(inst.opcode == Decoder::Opcode::S_SWAPPC_B64 ||
-		                     inst.opcode == Decoder::Opcode::S_GETPC_B64 ||
-		                     inst.opcode == Decoder::Opcode::S_ENDPGM);
-		EXIT_IF(CFG::MayWriteScalarRegister(inst, call.dst.reg) || CFG::MayWriteScalarRegister(inst, call.dst.reg + 1u));
-		const bool returns = inst.opcode == Decoder::Opcode::S_SETPC_B64;
-		if (returns) {
-			EXIT_NOT_IMPLEMENTED(inst.src0.kind != Decoder::OperandKind::Sgpr || inst.src0.reg != call.dst.reg);
-			inst.opcode = Decoder::Opcode::S_BRANCH;
-			inst.branch_target = call.pc + call.word_count * sizeof(uint32_t);
-		} else if (Decoder::IsDirectBranch(inst.opcode)) {
-			EXIT_IF(inst.branch_target > UINT32_MAX - function_pc);
-			inst.branch_target += function_pc;
+		const uint32_t function_size = function.back().pc + function.back().word_count * sizeof(uint32_t);
+		EXIT_IF(function_pc > UINT32_MAX - function_size);
+		const auto function_offset = static_cast<uint32_t>(linked.code.size());
+		for (auto inst: function) {
+			EXIT_IF(CFG::MayWriteScalarRegister(inst, call.dst.reg) ||
+			        CFG::MayWriteScalarRegister(inst, call.dst.reg + 1u));
+			for (uint32_t code = 0; code < site.preserved_registers.size(); ++code) {
+				if (site.preserved_registers[code] && CFG::MayWriteScalarRegister(inst, code))
+					EXIT("shader call at pc 0x%08x clobbers scalar target dependency s%u", call.pc, code);
+			}
+			linked.code.insert(linked.code.end(), inst.raw, inst.raw + inst.word_count);
+			if (inst.opcode == Decoder::Opcode::S_SETPC_B64) {
+				EXIT_NOT_IMPLEMENTED(inst.src0.kind != Decoder::OperandKind::Sgpr || inst.src0.reg != call.dst.reg);
+				inst.opcode = Decoder::Opcode::S_BRANCH;
+				inst.branch_target = call.pc + call.word_count * sizeof(uint32_t);
+			} else if (Decoder::IsDirectBranch(inst.opcode)) {
+				EXIT_IF(inst.branch_target >= function_size);
+				inst.branch_target += function_pc;
+			}
+			inst.pc += function_pc;
+			linked.decoded.instructions.push_back(std::move(inst));
 		}
-		inst.pc += function_pc;
-		linked.decoded.instructions.push_back(std::move(inst));
-		if (returns) break;
-	}
-	const auto function_end = linked.decoded.instructions.back().pc +
-	                          linked.decoded.instructions.back().word_count * sizeof(uint32_t);
-	for (size_t i = source.decoded.instructions.size(); i + 1u < linked.decoded.instructions.size(); ++i) {
-		const auto& inst = linked.decoded.instructions[i];
-		if (Decoder::IsDirectBranch(inst.opcode))
-			EXIT_IF(inst.branch_target < function_pc || inst.branch_target >= function_end);
+		linked.functions.push_back({function_offset, function_size / 4u});
+		linked.decoded.instructions[site.instruction].branch_target = function_pc;
+		function_pc += function_size;
 	}
 	reads.resize(query_reads);
-	reads.emplace_back(address, (linked.code.size() - source.code.size()) * sizeof(uint32_t));
-	linked.decoded.instructions[source.call->instruction].branch_target = function_pc;
+	for (size_t i = 0; i < source.calls.size(); ++i) {
+		const auto preceding = std::span(source.calls).first(i);
+		if (std::ranges::find(preceding, source.calls[i].address, &ShaderSource::Call::address) == preceding.end())
+			reads.emplace_back(source.calls[i].address, linked.functions[i].size * sizeof(uint32_t));
+	}
 	linked.decoded.code = linked.code;
-	linked.observed_function.resize(linked.code.size() - source.code.size());
+	linked.observed_function.resize(std::ranges::max_element(linked.functions, {},
+	    &ShaderSource::Linked::Function::size)->size);
+
 	source.linked = std::move(linked);
 	++source.revision;
 	return source.linked->decoded;
@@ -8,6 +8,7 @@
 #include "graphics/shader/shader.h"

 #include <array>
+#include <bitset>
 #include <optional>
 #include <span>
 #include <vector>
@@ -47,16 +48,21 @@ struct ShaderSource {
 	struct Call {
 		uint32_t instruction;
 		std::array<IR::Value, 2> target;
+		std::bitset<128> preserved_registers;
+		uint64_t address = 0;
 	};
 	struct Linked {
+		struct Function { uint32_t offset; uint32_t size; };
 		std::vector<uint32_t> code;
 		Decoder::Program decoded;
+		std::vector<Function> functions;
 		std::vector<uint32_t> observed_function;
 	};
 	std::vector<uint32_t> code;
 	Decoder::Program decoded;
 	IR::ResourcePlan call_targets;
-	std::optional<Call> call;
+	std::vector<Call> calls;
+	uint32_t unsupported_call_pc = UINT32_MAX;
 	std::optional<Linked> linked;
 	uint64_t revision = 0;
 	std::vector<std::pair<uint64_t, uint64_t>> reads;
@@ -67,6 +67,44 @@ uint32_t EmitExt(EmitterState& state, uint32_t type, uint32_t opcode,
 	return result;
 }

+uint32_t EmitF32ToF16RtzBits(EmitterState& state, uint32_t value) {
+	// Native half conversion causes regressions; use compact integer RTZ conversion.
+	const auto bits = EmitBitcastF32ToU32(state, value);
+	const auto abs  = EmitAndConstant(state, bits, 0x7fffffffu);
+	const auto sign = EmitAndConstant(
+	    state, Binary(state, spv::OpShiftRightLogical, TypeU32(state), bits, ConstantU32(state, 16)),
+	    0x8000u);
+	const auto exp = Binary(state, spv::OpShiftRightLogical, TypeU32(state), abs,
+	                        ConstantU32(state, 23));
+	const auto mant = EmitAndConstant(state, abs, 0x007fffffu);
+
+	// Align the hidden bit to the half subnormal mantissa. Even unused shifts must be < 32.
+	const auto shift = EmitUMin32(
+	    state, Binary(state, spv::OpISub, TypeU32(state), ConstantU32(state, 126), exp),
+	    ConstantU32(state, 31));
+	const auto sub = Binary(state, spv::OpShiftRightLogical, TypeU32(state),
+	                        EmitOrU32(state, mant, ConstantU32(state, 0x00800000u)), shift);
+
+	// Drop 13 mantissa bits and rebias the exponent; finite overflow saturates to max half.
+	const auto truncated = Binary(state, spv::OpShiftRightLogical, TypeU32(state), abs,
+	                              ConstantU32(state, 13));
+	const auto normal = EmitUMin32(
+	    state, Binary(state, spv::OpISub, TypeU32(state), truncated, ConstantU32(state, 0x1c000u)),
+	    ConstantU32(state, 0x7bffu));
+	const auto finite = Select(state, TypeU32(state),
+	                           EmitCompareU32Constant(state, spv::OpULessThan, abs, 0x38800000u),
+	                           sub, normal);
+	const auto nan = EmitOrU32(state, EmitAndConstant(state, truncated, 0x03ffu),
+	                           ConstantU32(state, 0x7e00u));
+	const auto special = Select(state, TypeU32(state),
+	                            EmitCompareU32Constant(state, spv::OpIEqual, mant, 0),
+	                            ConstantU32(state, 0x7c00u), nan);
+	const auto magnitude = Select(
+	    state, TypeU32(state),
+	    EmitCompareU32Constant(state, spv::OpUGreaterThanEqual, abs, 0x7f800000u), special, finite);
+	return EmitOrU32(state, sign, magnitude);
+}
+
 uint32_t EmitF32ToU32(EmitterState& state, uint32_t src, bool signed_value) {
 	const auto trunc         = EmitTruncF32Value(state, src);
 	const auto converted_raw = state.builder.AllocateId();
@@ -122,19 +160,6 @@ uint32_t EmitFPMad32(EmitterState& state, uint32_t a, uint32_t b, uint32_t c) {
 	return sum;
 }

-uint32_t EmitFPMedTri32(EmitterState& state, uint32_t a, uint32_t b, uint32_t c) {
-	const auto min_ab   = EmitMinMaxF32Value(state, a, b, false);
-	const auto min3     = EmitMinMaxF32Value(state, min_ab, c, false);
-	const auto max_ab   = EmitMinMaxF32Value(state, a, b, true);
-	const auto high_min = EmitMinMaxF32Value(state, max_ab, c, false);
-	const auto median   = EmitMinMaxF32Value(state, min_ab, high_min, true);
-	const auto nan_ab   = Binary(state, spv::OpLogicalOr, TypeBool(state),
-	                             EmitClassifyF32(state, a).nan, EmitClassifyF32(state, b).nan);
-	const auto any_nan =
-	    Binary(state, spv::OpLogicalOr, TypeBool(state), nan_ab, EmitClassifyF32(state, c).nan);
-	return Select(state, TypeF32(state), any_nan, min3, median);
-}
-
 uint32_t EmitFindUMsb64(EmitterState& state, uint32_t value) {
 	// GLSL FindUMsb is restricted to 32-bit components.
 	const auto unpacked = Unary(state, spv::OpBitcast, TypeU32Vector(state, 2), value);
@@ -210,28 +235,18 @@ uint32_t EmitCompositeExtractU32x2(EmitterState& state, uint32_t arg0, IR::Value
 }

 uint32_t EmitPackFloat2x16Rtz(EmitterState& state, uint32_t arg0, uint32_t arg1) {
-	const auto pair = state.builder.AllocateId();
-	state.builder.AddFunction(spv::OpCompositeConstruct, TypeF32Vector(state, 2), pair, arg0, arg1);
-	const auto packed = EmitPackHalf2x16(state, pair);
-	const auto back   = EmitExt(state, TypeF32Vector(state, 2), GLSLstd450UnpackHalf2x16, {packed});
-	auto       result = packed;
-	for (uint32_t i = 0; i < 2; i++) {
-		const auto rounded = state.builder.AllocateId();
-		state.builder.AddFunction(spv::OpCompositeExtract, TypeF32(state), rounded, back, i);
-		const auto above =
-		    Binary(state, spv::OpFOrdGreaterThan, TypeBool(state),
-		           EmitExt(state, TypeF32(state), GLSLstd450FAbs, {rounded}),
-		           EmitExt(state, TypeF32(state), GLSLstd450FAbs, {i == 0 ? arg0 : arg1}));
-		result = Binary(state, spv::OpISub, TypeU32(state), result,
-		                Select(state, TypeU32(state), above,
-		                       ConstantU32(state, i == 0 ? 1u : 0x10000u), ConstantU32(state, 0)));
-	}
-	return result;
+	const auto low  = EmitF32ToF16RtzBits(state, arg0);
+	const auto high = Binary(state, spv::OpShiftLeftLogical, TypeU32(state),
+	                         EmitF32ToF16RtzBits(state, arg1), ConstantU32(state, 16));
+	return EmitOrU32(state, low, high);
 }

 uint32_t EmitFPSaturate32(EmitterState& state, uint32_t arg0) {
-	return EmitExt(state, TypeF32(state), GLSLstd450FClamp,
-	               {arg0, ConstantF32(state, 0), ConstantF32(state, 0x3f800000u)});
+	// The clamp modifier runs in DX10 clamp mode: NaN becomes 0. FClamp leaves NaN undefined.
+	const auto clamped = EmitExt(state, TypeF32(state), GLSLstd450FClamp,
+	                             {arg0, ConstantF32(state, 0), ConstantF32(state, 0x3f800000u)});
+	return Select(state, TypeF32(state), EmitClassifyF32(state, arg0).nan, ConstantF32(state, 0),
+	              clamped);
 }

 uint32_t EmitSMulHi(EmitterState& state, uint32_t arg0, uint32_t arg1) {
@@ -100,10 +100,9 @@ uint32_t EmitClassMaskF16(EmitterState& state, uint32_t bits, uint32_t mask) {
 	return EmitClassMaskBits(state, bits, mask, true);
 }

-uint32_t EmitMinMaxF32Value(EmitterState& state, uint32_t lhs, uint32_t rhs, bool max_value) {
-	const auto lhs_class = EmitClassifyF32(state, lhs);
-	const auto rhs_class = EmitClassifyF32(state, rhs);
-
+static uint32_t EmitMinMaxF32(EmitterState& state, uint32_t lhs, uint32_t rhs,
+                             const F32Class& lhs_class, const F32Class& rhs_class,
+                             bool max_value) {
 	const auto numeric_cond = state.builder.AllocateId();
 	state.builder.AddFunction(max_value ? spv::OpFOrdGreaterThanEqual : spv::OpFOrdLessThan,
 	                          TypeBool(state), numeric_cond, lhs, rhs);
@@ -121,6 +120,34 @@ uint32_t EmitMinMaxF32Value(EmitterState& state, uint32_t lhs, uint32_t rhs, boo
 	return EmitBitcastU32ToF32(state, result_bits);
 }

+uint32_t EmitMinMaxF32Value(EmitterState& state, uint32_t lhs, uint32_t rhs, bool max_value) {
+	const auto lhs_class = EmitClassifyF32(state, lhs);
+	const auto rhs_class = EmitClassifyF32(state, rhs);
+	return EmitMinMaxF32(state, lhs, rhs, lhs_class, rhs_class, max_value);
+}
+
+uint32_t EmitFPMedTri32(EmitterState& state, uint32_t a, uint32_t b, uint32_t c) {
+	const auto a_class = EmitClassifyF32(state, a);
+	const auto b_class = EmitClassifyF32(state, b);
+	const auto c_class = EmitClassifyF32(state, c);
+	const auto min_ab = EmitMinMaxF32(state, a, b, a_class, b_class, false);
+	const auto min_ab_class = EmitClassifyF32(state, min_ab);
+	const auto min3 = EmitMinMaxF32(state, min_ab, c, min_ab_class, c_class, false);
+	const auto max_ab = EmitMinMaxF32(state, a, b, a_class, b_class, true);
+	const auto max_ab_class = EmitClassifyF32(state, max_ab);
+	const auto max3 = EmitMinMaxF32(state, max_ab, c, max_ab_class, c_class, true);
+	const auto max_bc = EmitMinMaxF32(state, b, c, b_class, c_class, true);
+	const auto max_ac = EmitMinMaxF32(state, a, c, a_class, c_class, true);
+	// RDNA2 compares numerically against MAX3, including ties between signed zeros.
+	const auto a_is_max = Binary(state, spv::OpFOrdEqual, TypeBool(state), max3, a);
+	const auto b_is_max = Binary(state, spv::OpFOrdEqual, TypeBool(state), max3, b);
+	const auto median = Select(state, TypeF32(state), a_is_max, max_bc,
+	                           Select(state, TypeF32(state), b_is_max, max_ac, max_ab));
+	const auto nan_ab = EmitLogicalOrBool(state, a_class.nan, b_class.nan);
+	const auto any_nan = EmitLogicalOrBool(state, nan_ab, c_class.nan);
+	return Select(state, TypeF32(state), any_nan, min3, median);
+}
+
 uint32_t EmitFlushF32DenormToSignedZero(EmitterState& state, uint32_t value) {
 	const auto bits      = state.builder.AllocateId();
 	const auto abs_bits  = state.builder.AllocateId();
@@ -52,9 +52,12 @@ uint32_t EmitBuiltinU32(EmitterState& state, IR::StageInputKind kind, uint32_t c
 		const auto value = state.builder.AllocateId();
 		const auto bits  = state.builder.AllocateId();
 		state.builder.AddFunction(spv::OpLoad, TypeBool(state), value, variable);
-		// PS5 initializes v_front_face with float +1.0/-1.0 bits.
-		state.builder.AddFunction(spv::OpSelect, TypeU32(state), bits, value,
-		                          ConstantU32(state, 0x3f800000u), ConstantU32(state, 0xbf800000u));
+		// SPI_BARYC_CNTL.FRONT_FACE_ALL_BITS selects uint masks or float +/-1.
+		const bool all_bits =
+		    state.input_info.pixel != nullptr && state.input_info.pixel->ps_front_face_all_bits;
+		const auto front = ConstantU32(state, all_bits ? 0xffffffffu : 0x3f800000u);
+		const auto back  = ConstantU32(state, all_bits ? 0u : 0xbf800000u);
+		state.builder.AddFunction(spv::OpSelect, TypeU32(state), bits, value, front, back);
 		return bits;
 	}
 	if (kind == IR::StageInputKind::VertexIndex || kind == IR::StageInputKind::InstanceIndex ||
@@ -652,6 +655,14 @@ void EmitStoreCompletion(EmitterState& state) {
 	                          spv::MemorySemanticsImageMemoryMask));
 }

+void EmitDataCacheInvalidate(EmitterState& state) {
+	// RDNA2 GL0/GL1 invalidation makes vector memory visible without a workgroup rendezvous.
+	state.builder.AddFunction(
+	    spv::OpMemoryBarrier, ConstantU32(state, spv::ScopeDevice),
+	    ConstantU32(state, spv::MemorySemanticsAcquireMask | spv::MemorySemanticsUniformMemoryMask |
+	                          spv::MemorySemanticsImageMemoryMask));
+}
+
 void EmitBarrier(EmitterState& state) {
 	const auto tessellation = state.program.stage == ShaderType::TessellationControl;
 	if (!tessellation && ShaderWorkgroupInput(state.program.stage, state.input_info) == nullptr) {
@@ -97,10 +97,14 @@ uint32_t GatherMip(ValueEmitContext& ctx, const IR::Inst& inst, const IR::Memory
 		              ConstantF32Value(state, 1.0f / 256.0f));
 	};
 	const auto sampler_control = ctx.Arg(inst, 6);
+	const auto mip_filter = Extract(sampler_control, 26, 2);
 	const auto preclamp = Binary(state, spv::OpINotEqual, TypeBool(state),
 	                            Extract(sampler_control, 28, 1), ConstantU32(state, 0));
-	const auto half = ConstantF32Value(state, 0.5f);
 	const auto zero = ZeroF32(state);
+	// Point selects the nearest mip; linear gathers return the lower, primary mip only.
+	const auto rounding = Select(state, TypeF32(state),
+	    Binary(state, spv::OpIEqual, TypeBool(state), mip_filter, ConstantU32(state, 1)),
+	    ConstantF32Value(state, 0.5f), zero);
 	const auto base = Unary(state, spv::OpConvertUToF, TypeF32(state),
 	                        Extract(ctx.Arg(inst, 4), 12, 4));
 	const auto minimum = Binary(state, spv::OpFSub, TypeF32(state),
@@ -108,19 +112,19 @@ uint32_t GatherMip(ValueEmitContext& ctx, const IR::Inst& inst, const IR::Memory
 	const auto last = ConstantF32Value(state, static_cast<float>(mip_count - 1u));
 	auto lod = Binary(state, spv::OpFAdd, TypeF32(state),
 	                  AddressF32(ctx, mem, address, layout.lod),
-	                  Select(state, TypeF32(state), preclamp, half, zero));
+	                  Select(state, TypeF32(state), preclamp, rounding, zero));
 	lod = EmitGlsl<GLSLstd450FClamp, IR::Type::F32>(
 	    state, lod, LodFixed(ctx.Arg(inst, 5), 0), LodFixed(ctx.Arg(inst, 5), 12));
 	// T# minimum LOD is in physical mip space and applies after the S# view-relative clamp.
 	lod = EmitGlsl<GLSLstd450FMax, IR::Type::F32>(state, lod, minimum);
 	lod = EmitGlsl<GLSLstd450FClamp, IR::Type::F32>(state, lod, zero, last);
 	lod = Binary(state, spv::OpFAdd, TypeF32(state), lod,
-	             Select(state, TypeF32(state), preclamp, zero, half));
+	             Select(state, TypeF32(state), preclamp, zero, rounding));
 	// The clamped value is nonnegative, so conversion performs the integer mip rounding.
 	const auto mip = EmitUMin32(state, Unary(state, spv::OpConvertFToU, TypeU32(state), lod),
 	                           ConstantU32(state, mip_count - 1u));
 	const auto mip_none = Binary(state, spv::OpIEqual, TypeBool(state),
-	                            Extract(sampler_control, 26, 2), ConstantU32(state, 0));
+	                            mip_filter, ConstantU32(state, 0));
 	return Select(state, TypeU32(state), mip_none, ConstantU32(state, 0), mip);
 }

@@ -284,10 +288,10 @@ uint32_t ResultVector(ValueEmitContext& ctx, uint32_t value,
 }

 uint32_t QueryDimensions(ValueEmitContext& ctx, const IR::MemoryInfo& mem,
-                         const IR::Inst& address) {
-	const auto  dimension = ctx.state.program.info.images.at(mem.resource).dimension;
+                         const IR::Inst& address, uint32_t resource, uint32_t array_index) {
+	const auto  dimension = ctx.state.program.info.images.at(resource).dimension;
 	const auto& info      = ImageDimensionInfoFor(dimension);
-	const auto  image     = LoadImageDescriptor(ctx.state, mem.resource);
+	const auto  image     = LoadImageDescriptor(ctx.state, resource, 0u, array_index);
 	const auto  size      = ctx.state.builder.AllocateId();
 	if (info.multisampled != 0u) {
 		ctx.state.builder.AddFunction(spv::OpImageQuerySize,
@@ -689,68 +693,58 @@ auto EmitImageAccess(ValueEmitContext& ctx, const IR::Inst& inst, uint32_t resul
 	}
 	const auto selected = EmitIndirectResourceIndex(
 	    state, key, image.indirect_mapping_offset, image.indirect_search_iterations, 0u);
-	struct ImageRun {
-		uint32_t first;
-		uint32_t count;
-		uint32_t resource;
-		uint32_t slot_bias;
+	const auto SameType = [&](uint32_t lhs, uint32_t rhs) {
+		const auto& first = state.program.info.images[lhs];
+		const auto& second = state.program.info.images[rhs];
+		return first.dimension == second.dimension && first.cube == second.cube &&
+		       IR::DescriptorBindingForImage(first) == IR::DescriptorBindingForImage(second) &&
+		       (inst.GetOpcode() != IR::ValueOpcode::ImageWrite ||
+		        first.shader_swizzle == second.shader_swizzle) &&
+		       first.mip_count == 1u && second.mip_count == 1u;
 	};
-	std::vector<ImageRun> runs;
-	for (uint32_t ordinal = 0; ordinal < image.indirect_resources.size(); ++ordinal) {
-		const auto resource = image.indirect_resources[ordinal];
-		const auto& candidate = state.program.info.images[resource];
-		const auto kind = *IR::DescriptorBindingForImage(candidate);
-		if (!runs.empty()) {
-			auto& run = runs.back();
-			const auto& first = state.program.info.images[run.resource];
-			const auto next_slot = run.slot_bias + ordinal;
-			if (candidate.dimension == first.dimension && candidate.cube == first.cube &&
-			    IR::DescriptorBindingForImage(first) == kind &&
-			    (inst.GetOpcode() != IR::ValueOpcode::ImageWrite ||
-			     candidate.shader_swizzle == first.shader_swizzle) &&
-			    candidate.mip_count == 1u && first.mip_count == 1u &&
-			    (ordinal == 1u || candidate.descriptor_index == next_slot)) {
-				// Native roots precede appended children; only the root may have a slot gap.
-				if (ordinal == 1u)
-					run.slot_bias = candidate.descriptor_index - ordinal;
-				++run.count;
-				continue;
-			}
-		}
-		runs.push_back({ordinal, 1u, resource,
-		                candidate.descriptor_index - ordinal});
+	// Descriptor order is unrelated to image type. Share each typed access and
+	// select its descriptor slot independently, including gaps between native roots.
+	struct ImageGroup { uint32_t resource; uint32_t count; };
+	std::vector<ImageGroup> groups;
+	std::vector<uint32_t> ordinal_groups;
+	for (const auto resource: image.indirect_resources) {
+		const auto found = std::ranges::find_if(groups, [&](const ImageGroup& group) {
+			return SameType(resource, group.resource);
+		});
+		ordinal_groups.push_back(static_cast<uint32_t>(found - groups.begin()));
+		if (found == groups.end()) groups.push_back({resource, 1u});
+		else ++found->count;
 	}
-	const auto EmitRun = [&](uint32_t index) {
-		const auto& run = runs[index];
-		if (run.count == 1u) return emit(run.resource, 0u);
-		auto slot = Binary(state, spv::OpIAdd, TypeU32(state), selected,
-		                   ConstantU32(state, run.slot_bias));
-		if (run.first == 0u) {
-			const auto root_slot = image.descriptor_index;
-			if (root_slot != run.slot_bias) {
-				const auto is_root = Binary(state, spv::OpIEqual, TypeBool(state), selected,
-				                            ConstantU32(state, 0u));
-				const auto root_or_child = state.builder.AllocateId();
-				state.builder.AddFunction(spv::OpSelect, TypeU32(state), root_or_child,
-				                          is_root, ConstantU32(state, root_slot), slot);
-				slot = root_or_child;
-			}
-		}
-		return emit(run.resource, slot);
-	};
-	auto run_index = ConstantU32(state, 0u);
-	for (uint32_t index = 1; index < runs.size(); ++index) {
+	const bool dynamic_slot = std::ranges::any_of(groups, [](const ImageGroup& group) {
+		return group.count > 1u;
+	});
+	auto group_index = ConstantU32(state, 0u);
+	auto slot_bias = ConstantU32(state, image.descriptor_index);
+	auto previous_bias = image.descriptor_index;
+	for (uint32_t ordinal = 1; ordinal < image.indirect_resources.size(); ++ordinal) {
+		const auto& candidate = state.program.info.images[image.indirect_resources[ordinal]];
+		const auto bias = candidate.descriptor_index - ordinal;
+		const bool select_group = ordinal_groups[ordinal] != ordinal_groups[ordinal - 1u];
+		const bool select_slot = dynamic_slot && bias != previous_bias;
+		previous_bias = bias;
+		if (!select_group && !select_slot) continue;
 		const auto at_or_after = Binary(state, spv::OpUGreaterThanEqual, TypeBool(state),
-		                               selected, ConstantU32(state, runs[index].first));
-		const auto next = state.builder.AllocateId();
-		state.builder.AddFunction(spv::OpSelect, TypeU32(state), next, at_or_after,
-		                          ConstantU32(state, index), run_index);
-		run_index = next;
+		                               selected, ConstantU32(state, ordinal));
+		if (select_group) group_index = Select(state, TypeU32(state), at_or_after,
+		                                      ConstantU32(state, ordinal_groups[ordinal]), group_index);
+		if (select_slot) slot_bias = Select(state, TypeU32(state), at_or_after,
+		                                   ConstantU32(state, bias), slot_bias);
 	}
-	return runs.size() == 1u
-	           ? EmitRun(0u)
-	           : EmitIndexSwitch(state, run_index, static_cast<uint32_t>(runs.size()),
-	                             result_type, EmitRun);
+	const auto slot = dynamic_slot
+	    ? Binary(state, spv::OpIAdd, TypeU32(state), selected, slot_bias) : 0u;
+	const auto EmitGroup = [&](uint32_t index) {
+		const auto& group = groups[index];
+		return emit(group.resource, group.count == 1u ? 0u : slot);
+	};
+	return groups.size() == 1u
+	           ? EmitGroup(0u)
+	           : EmitIndexSwitch(state, group_index, static_cast<uint32_t>(groups.size()),
+	                             result_type, EmitGroup);
 }

 template <typename Emit>
@@ -852,7 +846,10 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 	const auto* address = ctx.ImageAddress(inst.Arg(image_info.needs_sampler ? 2 : 1));
 	if (op == IR::ValueOpcode::ImageQueryDimensions) {
 		state.builder.RequireCapability(spv::CapabilityImageQuery);
-		ctx.Define(inst, QueryDimensions(ctx, mem, *address));
+		ctx.Define(inst, EmitImageAccess(ctx, inst, TypeU32Vector(state, 4),
+		    [&](uint32_t resource, uint32_t array_index) {
+			    return QueryDimensions(ctx, mem, *address, resource, array_index);
+		    }));
 		return;
 	}
 	if (image_info.needs_sampler &&
@@ -218,6 +218,7 @@ inline constexpr auto EmitTtraceData   = EmitVoid;
 inline constexpr auto EmitInstPrefetch = EmitVoid;
 void                  EmitBarrier(EmitterState& state);
 void                  EmitStoreCompletion(EmitterState& state);
+void                  EmitDataCacheInvalidate(EmitterState& state);
 void                  EmitMeshAllocate(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitMeshDrawParameter(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitGetTessellationAttribute(ValueEmitContext& ctx, const IR::Inst& inst);
@@ -296,6 +297,7 @@ inline constexpr auto EmitBufferAtomicSMax32    = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicUMax32    = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicAnd32     = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicAnd64     = EmitBufferAtomic64;
+inline constexpr auto EmitBufferAtomicIAdd64    = EmitBufferAtomic64;
 inline constexpr auto EmitBufferAtomicOr32      = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicOr64      = EmitBufferAtomic64;
 inline constexpr auto EmitBufferAtomicXor32     = EmitAtomic32;
@@ -325,6 +327,7 @@ inline constexpr auto EmitSharedAtomicSMin32    = EmitAtomic32;
 inline constexpr auto EmitSharedAtomicUMin32    = EmitAtomic32;
 inline constexpr auto EmitSharedAtomicSMax32    = EmitAtomic32;
 inline constexpr auto EmitSharedAtomicUMax64    = EmitSharedAtomic64;
+inline constexpr auto EmitSharedAtomicCmpSwap32 = EmitAtomic32;
 inline constexpr auto EmitSharedAtomicCmpSwap64 = EmitSharedAtomic64;
 inline constexpr auto EmitSharedAtomicUMax32    = EmitAtomic32;
 inline constexpr auto EmitSharedAtomicAnd32     = EmitAtomic32;
@@ -420,6 +420,7 @@ void EnsureLdsStorage(EmitterState& state);
 struct MemoryResourceAccess {
 	IR::ResourceKind      kind             = IR::ResourceKind::None;
 	uint32_t              object_pointer   = 0;
+	uint32_t              descriptor_index = 0;
 	uint32_t              length           = 0;
 	uint32_t              byte_offset      = 0;
 	uint32_t              element_bits     = 32;
@@ -1075,6 +1075,63 @@ uint32_t LoadIndirectFormattedX(ValueEmitContext& ctx, const IR::Inst& inst) {
 	});
 }

+uint32_t LoadRawBuffer(ValueEmitContext& ctx, const IR::Inst& inst, const IR::MemoryInfo& mem,
+                       const MemoryResourceAccess& resource, uint32_t components) {
+	auto&      state  = ctx.state;
+	const auto scalar = [&] {
+		std::array<uint32_t, 4> values {};
+		for (uint32_t component = 0; component < components; component++) {
+			values[component] =
+			    LoadWordPrepared(ctx, inst, RebaseRawComponent(mem, component), resource);
+		}
+		return ConstructU32Composite(state, components, values);
+	};
+	if (!mem.IsCoherentBufferPairLoad(inst.GetOpcode())) return scalar();
+
+	const auto first  = ByteAddress(ctx, inst, mem, resource.byte_offset);
+	const auto second = ByteAddress(ctx, inst, RebaseRawComponent(mem, 1), resource.byte_offset);
+	const auto wide =
+	    PrepareStorageBufferResourceAccess(state, mem, state.storage_buffers[3],
+	                                       resource.non_uniform ? resource.descriptor_index : 0u);
+	const auto index =
+	    Binary(state, spv::OpShiftRightLogical, TypeU32(state), first, ConstantU32(state, 3));
+	auto atomic =
+	    Binary(state, spv::OpIEqual, TypeBool(state),
+	           Binary(state, spv::OpBitwiseAnd, TypeU32(state), first, ConstantU32(state, 7)),
+	           ConstantU32(state, 0));
+	atomic = AndCondition(state, atomic,
+	                      Binary(state, spv::OpIEqual, TypeBool(state), second,
+	                             EmitAddU32(state, first, ConstantU32(state, 4))));
+	// The 64-bit alias requires the entire pair; partial and swizzled reads retain DWORD bounds.
+	atomic = AndCondition(state, atomic, EmitMemoryElementInBounds(state, wide, index));
+	const auto atomic_label = state.builder.AllocateId(), scalar_label = state.builder.AllocateId();
+	const auto merge = state.builder.AllocateId();
+	state.builder.AddFunction(spv::OpSelectionMerge, merge, spv::SelectionControlMaskNone);
+	state.builder.AddFunction(spv::OpBranchConditional, atomic, atomic_label, scalar_label);
+	EmitLabel(state, atomic_label);
+	const auto pointer = EmitStorageBufferElementPointer(
+	    state, wide, index, TypeStorageBufferElementPointer(state, 64));
+	const auto snapshot = state.builder.AllocateId();
+	state.builder.AddFunction(spv::OpAtomicLoad, TypeU64(state), snapshot, pointer,
+	                          ConstantU32(state, spv::ScopeDevice),
+	                          ConstantU32(state, spv::MemorySemanticsAcquireMask |
+	                                                 spv::MemorySemanticsUniformMemoryMask));
+	const std::array<uint32_t, 4> words {
+	    EmitCompositeExtractU64(state, snapshot, IR::Value(0u)),
+	    EmitCompositeExtractU64(state, snapshot, IR::Value(1u))};
+	const auto pair = ConstructU32Composite(state, 2, words);
+	state.builder.AddFunction(spv::OpBranch, merge);
+	EmitLabel(state, scalar_label);
+	const auto fallback    = scalar();
+	const auto scalar_exit = state.current_label;
+	state.builder.AddFunction(spv::OpBranch, merge);
+	EmitLabel(state, merge);
+	const auto result = state.builder.AllocateId();
+	state.builder.AddFunction(spv::OpPhi, TypeU32Composite(state, 2), result, pair, atomic_label,
+	                          fallback, scalar_exit);
+	return result;
+}
+
 uint32_t LoadBuffer(ValueEmitContext& ctx, const IR::Inst& inst, uint32_t components) {
 	auto& state = ctx.state;
 	const auto mem = ctx.Memory(inst);
@@ -1100,12 +1157,7 @@ uint32_t LoadBuffer(ValueEmitContext& ctx, const IR::Inst& inst, uint32_t compon
 				    return LoadFormattedPrepared(ctx, inst, selected, resource, info, components);
 			    }
 			    if (resource.object_pointer == 0u) return ConstantU32CompositeZero(state, components);
-			    std::array<uint32_t, 4> values {};
-			    for (uint32_t component = 0; component < components; component++) {
-				    values[component] =
-				        LoadWordPrepared(ctx, inst, RebaseRawComponent(selected, component), resource);
-			    }
-			    return ConstructU32Composite(state, components, values);
+			    return LoadRawBuffer(ctx, inst, selected, resource, components);
 		    });
 	    });
 }
@@ -207,10 +207,11 @@ MemoryResourceAccess PrepareStorageBufferResourceAccess(EmitterState& state,
 	const bool non_uniform = array_index != 0u;
 	const auto static_index = state.program.info.buffers.at(mem.resource).descriptor_index;
 	if (!non_uniform) array_index = ConstantU32(state, static_index);
-	MemoryResourceAccess access {
-	    .kind = mem.kind,
-	    .non_uniform = non_uniform,
-	    .memory_access = mem.coherent ? spv::MemoryAccessVolatileMask : spv::MemoryAccessMaskNone};
+	MemoryResourceAccess access {.kind             = mem.kind,
+	                             .descriptor_index = array_index,
+	                             .non_uniform      = non_uniform,
+	                             .memory_access    = mem.coherent ? spv::MemoryAccessVolatileMask
+	                                                              : spv::MemoryAccessMaskNone};
 	access.object_pointer = state.builder.AllocateId();
 	state.builder.AddFunction(spv::OpAccessChain, buffer.pointer_type, access.object_pointer,
 	                          buffer.variable, array_index);
@@ -504,6 +505,7 @@ uint32_t EncodePackedColorComponent(EmitterState& state, const Format::BufferFor

 spv::Op SpirvAtomicOpcode(IR::ValueOpcode opcode) {
 	switch (opcode) {
+		case IR::ValueOpcode::SharedAtomicCmpSwap32:
 		case IR::ValueOpcode::SharedAtomicCmpSwap64:
 		case IR::ValueOpcode::ImageAtomicCompareSwap32:
 		case IR::ValueOpcode::BufferAtomicCmpSwap32: return spv::OpAtomicCompareExchange;
@@ -516,6 +518,7 @@ spv::Op SpirvAtomicOpcode(IR::ValueOpcode opcode) {
 		case IR::ValueOpcode::ImageAtomicIAdd64:
 		case IR::ValueOpcode::BufferAtomicIAdd32:
 		case IR::ValueOpcode::SharedAtomicIAdd64:
+		case IR::ValueOpcode::BufferAtomicIAdd64:
 		case IR::ValueOpcode::SharedAtomicIAdd32: return spv::OpAtomicIAdd;
 		case IR::ValueOpcode::BufferAtomicISub32:
 		case IR::ValueOpcode::SharedAtomicISub32: return spv::OpAtomicISub;
@@ -560,8 +563,12 @@ uint32_t EmitAtomicOperation(ValueEmitContext& ctx, const IR::Inst& inst, uint32
 	const bool wide   = inst.GetType() == IR::Type::U64;
 	const auto type   = wide ? TypeU64(ctx.state) : TypeU32(ctx.state);
 	if (opcode == spv::OpAtomicCompareExchange) {
-		const auto desired    = ctx.Arg(inst, inst.NumArgs() - 3);
-		const auto comparator = ctx.Arg(inst, inst.NumArgs() - 2);
+		auto desired    = ctx.Arg(inst, inst.NumArgs() - 3);
+		auto comparator = ctx.Arg(inst, inst.NumArgs() - 2);
+		if (inst.GetOpcode() == IR::ValueOpcode::SharedAtomicCmpSwap32) {
+			// RDNA2 DS compares DATA and stores DATA2, opposite to BUFFER atomics.
+			std::swap(desired, comparator);
+		}
 		ctx.state.builder.AddFunction(
 		    spv::OpAtomicCompareExchange, TypeU32(ctx.state), old, pointer,
 		    ConstantU32(ctx.state, scope), ConstantU32(ctx.state, spv::MemorySemanticsMaskNone),
@@ -307,6 +307,7 @@ void EmitBlock(ValueEmitContext& ctx, const IR::Block* block, EmitInstruction&&
 			ctx.state.lane_half = half;
 			if (half == 0 || (inst.GetOpcode() != IR::ValueOpcode::Barrier &&
 			                  inst.GetOpcode() != IR::ValueOpcode::StoreCompletion &&
+			                  inst.GetOpcode() != IR::ValueOpcode::DataCacheInvalidate &&
 			                  inst.GetOpcode() != IR::ValueOpcode::MeshAllocate)) {
 				emit_instruction(lane, inst);
 			}
@@ -8,6 +8,8 @@
 #include <deque>
 #include <list>
 #include <map>
+#include <optional>
+#include <set>
 #include <span>

 namespace Libs::Graphics::ShaderRecompiler::CFG {
@@ -1650,6 +1652,32 @@ Graph BuildGraph(const Decoder::Program& program) {
 	}
 	SortUnique(graph.code_table_load_pcs);

+	if (!indirect_setpc && std::ranges::any_of(program.instructions, [](const auto& inst) {
+		    return inst.opcode == Opcode::S_SWAPPC_B64 && inst.branch_target != UINT32_MAX;
+		})) {
+		// Linked leaf bodies follow the caller in native PC order. Visit them in
+		// execution order so returning to the caller does not manufacture goto loops.
+		std::vector<bool> visited(graph.blocks.size(), false);
+		std::vector<std::pair<uint32_t, bool>> pending {{graph.entry_block, false}};
+		std::vector<BasicBlock> ordered;
+		ordered.reserve(graph.blocks.size());
+		while (!pending.empty()) {
+			const auto [id, finished] = pending.back();
+			pending.pop_back();
+			if (finished) {
+				ordered.push_back(std::move(graph.blocks[id]));
+				continue;
+			}
+			if (visited[id]) continue;
+			visited[id] = true;
+			pending.emplace_back(id, true);
+			for (const auto successor: graph.blocks[id].successors)
+				pending.emplace_back(successor, false);
+		}
+		std::ranges::reverse(ordered);
+		ApplyBlockOrder(graph, std::move(ordered));
+	}
+
 	RecomputeAnalyses(graph);

 	if (indirect_setpc) {
@@ -1681,33 +1709,56 @@ Graph Structurize(const Graph& graph) {
 }

 bool MayWriteScalarRegister(const Decoder::Instruction& inst, uint32_t code) {
-	uint32_t destination = UINT32_MAX;
-	// Scalar ALU pair widths are not all decoded. The typed producer must confirm this word.
-	return InstructionWritesScalarCode(inst, code) ||
-	       (ScalarOperandCode(inst.dst, destination) && code == destination + 1u);
+	// Vector comparisons and carry results can write a scalar pair in wave64.
+	if (inst.family == Decoder::Family::VOP1 || inst.family == Decoder::Family::VOP2 ||
+	    inst.family == Decoder::Family::VOPC || inst.family == Decoder::Family::VOP3 ||
+	    inst.family == Decoder::Family::VOP3P) {
+		uint32_t destination = 0;
+		for (const auto* operand: {&inst.dst, &inst.dst2})
+			if (ScalarOperandCode(*operand, destination) && code >= destination && code <= destination + 1u)
+				return true;
+	}
+	return InstructionWritesScalarCode(inst, code);
 }

 uint32_t FindScalarDefinition(const Decoder::Program& program, const Graph& graph,
-                              uint32_t before, uint32_t code) {
+                              uint32_t before, uint32_t code, std::vector<uint32_t>& crossed_calls) {
 	EXIT_IF(before >= program.instructions.size());
 	const auto found = std::ranges::find_if(graph.blocks, [before](const auto& block) {
 		return block.inst_begin <= before && before < block.inst_end;
 	});
 	EXIT_IF(found == graph.blocks.end());
-	const auto* block = &*found;
-	for (size_t depth = 0; depth < graph.blocks.size(); ++depth) {
-		for (uint32_t i = before; i > block->inst_begin;) {
-			const auto& inst = program.instructions[--i];
-			EXIT_IF(inst.opcode == Opcode::S_SWAPPC_B64);
-			if (MayWriteScalarRegister(inst, code)) return i;
-		}
-		if (block->id == graph.entry_block) return UINT32_MAX;
-		EXIT_IF(block->predecessors.size() != 1u);
-		block = graph.FindBlock(block->predecessors[0]);
+	std::vector<std::pair<uint32_t, uint32_t>> pending {{found->id, before}};
+	std::set<std::pair<uint32_t, uint32_t>> visited;
+	std::optional<uint32_t> definition;
+	while (!pending.empty()) {
+		const auto point = pending.back();
+		pending.pop_back();
+		if (!visited.insert(point).second) continue;
+		const auto* block = graph.FindBlock(point.first);
 		EXIT_IF(block == nullptr);
-		before = block->inst_end;
+		uint32_t index = UINT32_MAX;
+		for (uint32_t i = point.second; i > block->inst_begin;) {
+			const auto& inst = program.instructions[--i];
+			if (inst.opcode == Opcode::S_SWAPPC_B64) {
+				EXIT_NOT_IMPLEMENTED(MayWriteScalarRegister(inst, code));
+				crossed_calls.push_back(i);
+				continue;
+			}
+			if (MayWriteScalarRegister(inst, code)) { index = i; break; }
+		}
+		if (index != UINT32_MAX || block->id == graph.entry_block) {
+			EXIT_IF(definition && *definition != index);
+			definition = index;
+			continue;
+		}
+		EXIT_IF(block->predecessors.empty());
+		for (const auto predecessor: block->predecessors) {
+			pending.emplace_back(predecessor, graph.blocks[predecessor].inst_end);
+		}
 	}
-	EXIT("scalar shader call source has cyclic reaching definitions");
+	EXIT_IF(!definition);
+	return *definition;
 }

 std::string BranchConditionToString(BranchCondition condition) {
@@ -115,7 +115,7 @@ struct Graph {
 Graph       BuildGraph(const Decoder::Program& program);
 bool MayWriteScalarRegister(const Decoder::Instruction& inst, uint32_t code);
 uint32_t FindScalarDefinition(const Decoder::Program& program, const Graph& graph,
-                              uint32_t before, uint32_t code);
+                              uint32_t before, uint32_t code, std::vector<uint32_t>& crossed_calls);
 // Returns structured control flow or failure diagnostics without changing the native graph.
 // On failure, failure_block is an original block ID or UINT32_MAX.
 Graph       Structurize(const Graph& graph);
@@ -67,8 +67,11 @@ constexpr MemoryOpcodeInfo MUBUF_OPCODE_LIST[] = {
    {0x3fu, Opcode::BUFFER_ATOMIC_FMIN, 1, 32},
    {0x40u, Opcode::BUFFER_ATOMIC_FMAX, 1, 32},
    {0x50u, Opcode::BUFFER_ATOMIC_SWAP_X2, 2, 32},
+    {0x52u, Opcode::BUFFER_ATOMIC_ADD_X2, 2, 32},
    {0x59u, Opcode::BUFFER_ATOMIC_AND_X2, 2, 32},
    {0x5au, Opcode::BUFFER_ATOMIC_OR_X2, 2, 32},
+    {0x71u, Opcode::BUFFER_GL0_INV, 0, 0},
+    {0x72u, Opcode::BUFFER_GL1_INV, 0, 0},
    {0x80u, Opcode::BUFFER_LOAD_FORMAT_D16_X, 1, 16, false, false, true},
    {0x83u, Opcode::BUFFER_LOAD_FORMAT_D16_XYZW, 2, 16, false, false, true},
    {0x84u, Opcode::BUFFER_STORE_FORMAT_D16_X, 1, 16, false, false, true},
@@ -114,6 +117,7 @@ constexpr MemoryOpcodeInfo DS_OPCODE_LIST[] = {
    {0x27u, Opcode::DS_MIN_RTN_U32, 1, 32},      {0x28u, Opcode::DS_MAX_RTN_U32, 1, 32},
    {0x29u, Opcode::DS_AND_RTN_B32, 1, 32},      {0x2au, Opcode::DS_OR_RTN_B32, 1, 32},
    {0x2bu, Opcode::DS_XOR_RTN_B32, 1, 32},      {0x2du, Opcode::DS_WRXCHG_RTN_B32, 1, 32},
+    {0x30u, Opcode::DS_CMPST_RTN_B32, 1, 32},
    {0x35u, Opcode::DS_SWIZZLE_B32, 1, 32},      {0x36u, Opcode::DS_READ_B32, 1, 32},
    {0x37u, Opcode::DS_READ2_B32, 2, 32},        {0x38u, Opcode::DS_READ2ST64_B32, 2, 32},
    {0x39u, Opcode::DS_READ_I8, 1, 8, true},     {0x3au, Opcode::DS_READ_U8, 1, 8},
@@ -200,6 +204,7 @@ bool IsDsAtomicOpcode(Opcode opcode) {
 		case Opcode::DS_MAX_U32:
 		case Opcode::DS_MAX_RTN_U32:
 		case Opcode::DS_MAX_RTN_U64:
+		case Opcode::DS_CMPST_RTN_B32:
 		case Opcode::DS_CMPST_RTN_B64:
 		case Opcode::DS_AND_B32:
 		case Opcode::DS_AND_RTN_B32:
@@ -220,6 +225,7 @@ uint32_t DsSourceCount(Opcode opcode) {
 		case Opcode::DS_WRITE2ST64_B32:
 		case Opcode::DS_WRITE2_B64:
 		case Opcode::DS_WRITE2ST64_B64:
+		case Opcode::DS_CMPST_RTN_B32:
 		case Opcode::DS_CMPST_RTN_B64:
 		case Opcode::DS_MSKOR_B32: return 3u;
 		case Opcode::DS_MIN_F32:
@@ -308,6 +314,9 @@ void DecodeMubuf(uint32_t pc, std::span<const uint32_t> code, uint32_t word_inde
 	if (inst.opcode == Opcode::UNSUPPORTED) {
 		SetUnsupported(inst, Family::MUBUF, opcode, "MUBUF opcode is not implemented");
 	}
+	if (inst.opcode == Opcode::BUFFER_GL0_INV || inst.opcode == Opcode::BUFFER_GL1_INV) {
+		return;
+	}

 	DecodeVectorGpr(vdata, inst.dst);
 	if (inst.opcode == Opcode::BUFFER_LOAD_FORMAT_D16_XYZW ||
@@ -454,6 +463,9 @@ void DecodeDs(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index,
 	if (inst.opcode == Opcode::DS_SWIZZLE_B32 && inst.offset >= 0xe000u) {
 		SetUnsupported(inst, Family::DS, opcode, "DS swizzle FFT mode is not implemented");
 	}
+	if (inst.gds && inst.opcode == Opcode::DS_CMPST_RTN_B32) {
+		SetUnsupported(inst, Family::DS, opcode, "DS compare-store return B32 GDS variant is not implemented");
+	}
 	if (inst.gds &&
 	    (inst.opcode == Opcode::DS_SWIZZLE_B32 || inst.opcode == Opcode::DS_PERMUTE_B32 ||
 	     inst.opcode == Opcode::DS_BPERMUTE_B32 ||
@@ -164,9 +164,23 @@ void DecodeSop1(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 	switch (inst.opcode) {
 		case Opcode::S_SWAPPC_B64:
 			inst.branch_target = UINT32_MAX;
-			inst.data_dwords = 2;
+			[[fallthrough]];
+		case Opcode::S_MOV_B64:
+		case Opcode::S_CMOV_B64:
+		case Opcode::S_NOT_B64:
+		case Opcode::S_WQM_B64:
+		case Opcode::S_BREV_B64:
+		case Opcode::S_BITSET0_B64:
+		case Opcode::S_BITSET1_B64:
+		case Opcode::S_AND_SAVEEXEC_B64:
+		case Opcode::S_ORN2_SAVEEXEC_B64:
+		case Opcode::S_QUADMASK_B64:
+		case Opcode::S_ANDN1_SAVEEXEC_B64:
+		case Opcode::S_BITREPLICATE_B64_B32:
+			inst.data_dwords = 2u;
 			break;
 		case Opcode::S_GETPC_B64:
+			inst.data_dwords = 2u;
 			inst.src_count = 0;
 			DecodeScalarDestination(sdst, pc, inst.dst);
 			return;
@@ -206,6 +220,8 @@ void DecodeSop2(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index

 	DecodeScalarDestination(sdst, pc, inst.dst);
 	DecodeBinarySources(pc, code, word_index, inst, ssrc0, ssrc1);
+	if (opcode == 0x0bu || (opcode >= 0x0fu && opcode <= 0x25u && (opcode & 1u)) || opcode == 0x29u)
+		inst.data_dwords = 2u;
 }

 void DecodeSopk(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index,
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Almo7aya	cd4eab9d65	gpu: clamp resolved buffer ranges to storage buffer limits	2026-10-19 13:53:18 +02:00
nmzik	fde0b98e69	gpu: preserve zero-stride buffer fetch payloads	2026-10-19 12:23:43 +02:00
shadowbeat070	d4465faa79	gpu: preserve ordered-append command order	2026-10-19 11:46:39 +02:00
nmzik	e5f0ee7fc2	gpu: preserve finite buffer descriptors across sparse PRT holes	2026-10-19 05:26:41 +02:00
nmzik	5f29374d5b	gpu: approximate PPSA24156 EQAA with 2x MSAA and warn once	2026-10-19 05:17:32 +02:00
nmzik	2c730e58ea	revert HTile lifetime changes and opt-in ZMask support	2026-10-19 05:04:56 +02:00
Pixeluted	e87fb8c859	kernel: hand guest mutex ownership directly to FIFO waiters	2026-10-19 05:04:56 +02:00
Pixeluted	440245f616	gpu: bound overlapping image copies by both native backings	2026-10-19 05:04:56 +02:00
nmzik	05d25ef28c	revert `01ce97388`: restore compute metadata clear consumption	2026-10-19 05:04:56 +02:00
Pixeluted	cad6aafdc5	shader: detect and loudly skip unsupported GPU-buffer callables	2026-10-19 05:04:56 +02:00
nmzik	defb2f11d3	Share median classifications and preserve RDNA2 signed-zero selection	2026-10-19 02:34:25 +02:00
nmzik	75d7a4311d	Share repeated SSA bitcasts and guard emitted shader size	2026-10-19 02:34:25 +02:00
nmzik	bb505a3ae8	Recognize packed descriptor keys in GPU sampler waterfalls	2026-10-19 02:34:25 +02:00
nmzik	ab8566591f	shader: link multiple compute helpers without synthetic return loops	2026-10-19 02:34:25 +02:00
nmzik	b2d3ce6dfc	shader: size expanded sampler bindings independently of native roots	2026-10-19 02:34:25 +02:00
nmzik	09c993dfe0	shader: query dimensions through shared indirect image selection	2026-10-19 02:34:25 +02:00
nmzik	de81300b28	shader: bound indexed material image selectors before heap expansion	2026-10-19 02:34:25 +02:00
nmzik	ac3711ce7c	shader: select the primary mip for linear explicit-LOD gathers	2026-10-19 02:34:24 +02:00
nmzik	8b51370ca1	shader: eliminate predicates for empty branch detours	2026-10-19 02:34:24 +02:00
nmzik	e9bcc77e0b	shader: retain resource guards for 64-bit scalar comparisons	2026-10-18 20:22:54 +02:00
nmzik	8639e0de81	shader: share indirect image accesses by type and guard emitted size	2026-10-18 20:16:49 +02:00
nmzik	7675579e79	kernel: implement POSIX fsync through the shared file API	2026-10-18 20:15:53 +02:00
nmzik	831862a50b	razor: stub	2026-10-18 19:29:54 +02:00
nmzik	099de8dea6	graphics: publish each PM4 WRITE_DATA word exactly once	2026-10-18 16:17:51 +02:00
opcod3	f4bdb52922	agc: restore verified fused shader size name	2026-10-18 16:12:28 +02:00
nmzik	22ef16a044	shader: implement PS5 vector data cache invalidation	2026-10-18 15:57:14 +02:00
nmzik	37df7f3278	Show pipeline cache sizes in MB and remove obsolete texture folder cleanup	2026-10-18 03:57:45 +02:00
nmzik	33208a545a	shader: remove artificial sampled image and sampler pair limit	2026-10-18 03:20:38 +02:00
Santiago Paquete	4059179e9d	graphics: write proven compute buffer fills to guest memory (#1628 ) Avoid readbacks for compute metadata fills Co-authored-by: nmzik <Nmzik@mail.ru>	2026-10-18 05:42:14 +02:00
nmzik	fe3857a171	graphics: materialize CMASK clears for multisampled targets (PPSA13794)	2026-10-18 02:37:56 +02:00
nmzik	c3e7ec8285	graphics: fix Jusant regression from `d8c4a9af3`	2026-10-18 01:03:37 +02:00
nmzik	763d3be3a9	graphics: add opt-in ZMask support and launcher rendering options	2026-10-17 17:50:02 +02:00
nmzik	d8c4a9af38	graphics: preserve HTILE metadata across native image replacement	2026-10-17 17:31:25 +02:00
Juan Pablo Farias	6aa3846fb2	feat: add an in-game performance panel (#1090 ) Add in-game performance panel Co-authored-by: nmzik <Nmzik@mail.ru>	2026-10-17 14:47:51 +02:00
Felipe	415387f00d	Support multiple controllers for local multiplayer (#1583 ) * Support multiple controllers for local co-op	2026-10-17 13:15:22 +02:00
Matthew Hurley	890393fbe6	vulkan: start without fragmentShaderBarycentric on GPUs that lack it (#1590 ) * vulkan: start without fragmentShaderBarycentric on GPUs that lack it Pascal (GTX 10xx), RDNA 1 and older drivers have no VK_KHR_fragment_shader_barycentric, so device selection rejected them with 'Could not find suitable device' (#1432, #1416, #762). The SPIR-V backend only emits the capability for pixel shaders with per-vertex/barycentric inputs, so make the extension and feature optional: warn at selection and creation, enable only when supported. Shaders that need custom barycentrics fail at pipeline creation instead of blocking startup for every game. * vulkan: fix macOS build and warn about unsupported GPUs --------- Co-authored-by: Matthew Hurley <171001439+theantipopau@users.noreply.github.com> Co-authored-by: nmzik <Nmzik@mail.ru>	2026-10-17 12:55:33 +02:00
nmzik	9f88043ebe	shader: use compact integer RTZ half conversion	2026-10-17 12:23:12 +02:00
nmzik	c2173801d4	shader: skip resources in proven empty unsigned loops	2026-10-17 10:38:14 +02:00
nmzik	56d03f16e9	network: UDPP2P	2026-10-17 10:24:45 +02:00
nmzik	3ecc6b7d1b	network: implement POSIX socket shutdown on Linux	2026-10-17 10:10:16 +02:00
nmzik	5bf37f52ad	renderer: clear compressed stencil with pending depth metadata	2026-10-17 09:50:11 +02:00
nmzik	57791645e4	renderer: track attachment load clears by image aspect	2026-10-17 09:34:51 +02:00
nmzik	d1fe1324c8	shader: keep coherent 64bit atomic	2026-10-17 07:52:20 +02:00
nmzik	b485dd24be	system service: stub	2026-10-17 07:30:21 +02:00
nmzik	ae25558321	shader: size NGG primitive outputs from amplification	2026-10-17 06:54:45 +02:00
nmzik	ec08ea197b	shader: new fill pattern	2026-10-17 06:24:29 +02:00
nmzik	f6f1ac8b40	shader: batch dependent folds to avoid repeated SSA graph walks	2026-10-17 06:20:02 +02:00
Nawid3333	b82512be06	shader: read float inline constants on 64-bit operands as f64 bits (#1576 ) A float inline constant (240-248) read by a 64-bit operand is the f64 bit pattern for integer B64 opcodes too: AMD's compilers emit v_lshrrev_b64 v[4:5], v0, 1.0 for 0x3ff0000000000000 >> lane. ReadU32Pair returned {f32 bits, 0}, so only F64-typed reads had the right value. Based on TheCruZ's 284deab28 (closed #1219). Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>	2026-10-17 05:21:06 +02:00
Nawid3333	ba8b0c85b8	shader: keep quiet NaNs with payload bits out of the packed f16 min/max sNaN path (#1579 ) V_PK_MIN_F16 and V_PK_MAX_F16 return the quieted NaN for a signaling NaN operand and the other operand for a quiet one (per half, as V_MIN_F16 and V_MAX_F16 do). The signaling test masked the half with 0x7c00 and checked for a nonzero payload, so a quiet NaN with payload bits such as 0x7e01 was taken as signaling and returned. Mask with 0x7e00 so the quiet bit must be clear. Two more vectors in PackedMinMaxF16NanAndSignedZeroEdges cover it. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>	2026-10-17 05:12:37 +02:00
Nawid3333	b6bd9ecae4	shader: read float inline constants on 16-bit operands as half bits (#1582 ) The 16-bit encodings of the float inline constants are fixed: the exact halves of 0.5, 1.0, 2.0 and 4.0 and their negations, and 0x3118 for 1/(2pi). Read16LaneBits converted the 32-bit value at run time instead, through PackHalf2x16, which the AMD driver rounds toward zero, so 1/(2pi) became 0x3117. Use the encodings directly. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>	2026-10-17 05:05:40 +02:00
Felipe	b95ea4bb5f	Support mountable-content DLC (addcont with real content) (#1581 ) Support mountable-content DLC (addcont) and fix PlayGo chunk count query	2026-10-17 04:44:28 +02:00
Nawid3333	e93b4b0378	shader: support returning 32-bit LDS compare-exchange (DS_CMPST_RTN_B32) (#1578 ) * shader: support DS_CMPST_RTN_B32 Assassin's Creed Valhalla (PPSA01490, #281) stops on DS opcode 0x30 in a compute shader. DS compares DATA and stores DATA2, the opposite of BUFFER_ATOMIC_CMPSWAP, so the 32-bit atomic path swaps the two for this opcode, as the 64-bit form does. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * shader: reject unsupported GDS compare-store variant --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> Co-authored-by: nmzik <Nmzik@mail.ru>	2026-10-17 04:37:16 +02:00
nmzik	0fd41ae206	Fix Linux CI xxHash include path (#1580 ) Fix xxHash include path capitalization for test builds	2026-10-17 04:37:00 +02:00
nmzik	9202bddad2	recognize verified pattern fills Co-authored-by: Juan Pablo Farias <juanpablo.profissional.a.f@gmail.com>	2026-10-17 04:09:54 +02:00
nmzik	32587337b5	renderdoc: advance capture flips independently of logging	2026-10-17 04:09:54 +02:00
Çağatay Güçlü	02ead0e2e1	shader: 64-bit buffer atomic add (BUFFER_ATOMIC_ADD_X2)	2026-10-17 04:09:54 +02:00
NeySlim	4f1f3b1bb3	shader: make the f32 clamp modifier turn NaN into 0	2026-10-17 04:09:53 +02:00
Nawid3333	42bb0da520	loader: keep the caller's flags in the TLS stubs (#1571 ) A patched TLS load calls TlsRegStub and SafeCall, which moved RSP past the red zone with sub/add rsp,0x80. ADD and SUB write the arithmetic flags, so the sub ran before SafeCall's pushfq saved them and the add ran after its popfq restored them: guest code that compares, loads fs:[0] and then branches on the flags could branch the wrong way. lea moves RSP without touching the flags. The entry lea plus a two-byte nop keeps every patched offset in both stubs; the exit lea is one byte longer and moves only ret. TlsStubsPreserveFlags calls the stubs with two flag patterns and checks the status flags, the output register and RAX; it fails with the old bytes. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>	2026-10-17 00:05:08 +02:00
nmzik	5dec523143	Stub RazorCpu profiling markers in a dedicated library	2026-10-16 19:16:42 +02:00
nmzik	d01633e18e	Revert "Fix scalar buffer reads across sparse guest mappings" This reverts commit `d7974dc763`.	2026-10-16 19:13:36 +02:00
nmzik	e7ac242039	Load packaged system modules through the native module lifecycle	2026-10-16 19:13:36 +02:00
nmzik	d7974dc763	Fix scalar buffer reads across sparse guest mappings	2026-10-16 18:20:31 +02:00
shadowbeat070	b899717900	renderer: log unsupported depth feedback once	2026-10-16 17:24:45 +02:00
shadowbeat070	d58aefde49	renderer: reduce synchronization and allocation overhead Performance changes from #1495: - Sync changed BDA ranges instead of scanning all mappings. - Let VMA suballocate BDA buffers. - Remove the renderer lock from presentation. - Keep compatible pipeline caches across builds. - Fold masked shifts that always produce zero.	2026-10-16 17:18:49 +02:00
Pipe530	4c6be6cd1d	Localize trophy notifications with embedded JSON catalogs	2026-10-16 16:19:43 +02:00
Felipe	584c2e8ca6	Play snd0.at9 during the startup splash screen (#1545 ) launcher: play snd0.at9 on title selection	2026-10-16 15:59:03 +02:00
nmzik	717b50155a	launcher: rename executable to kyty_launcher	2026-10-16 12:57:28 +02:00
shadowbeat070	57b161019d	shader: honor front-face input format	2026-10-16 12:57:27 +02:00
Felipe	3ff6606fcb	Input mapping tabs UI, emulated touchpad, and rumble for non-PS5 controllers (#1541 ) * Revamp input mapping dialog with tabs, gamepad remapping and gyro emulation toggle Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com> * Add emulated touchpad (mouse/gamepad) with particle skins; widen input dialog Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com> * Disable mouse touchpad mode when the cursor leaves the window Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com> * Add touchpad hand-cursor toggle; share touchpad options across both tabs Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com> * Hide touchpad trail and particles when the visual toggle is off Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com> * Rename input mapping tabs Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com> * Rumble non-DualSense gamepads from haptic audio Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com> * Fix input mapping, touchpad transitions, and rumble fallback --------- Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com> Co-authored-by: nmzik <Nmzik@mail.ru>	2026-10-16 04:04:37 +02:00
Ali Almohaya	6a0a666a41	loader: keep the output and log of a game LoadExec relaunches (#1480 ) * loader: keep the output of a process LoadExec relaunches LoadExec on Windows starts a new emulator process and exits the current one. The new process was created without a console and without the standard handles, so everything it printed was lost: a redirected log of a game that relaunches itself ended at the relaunch. Hand the new process the standard input, output and error handles, inheriting only those, and let it share the console when there is one. * loader: flush output before LoadExec starts the new program On Windows, LoadExec flushed the C streams only after the new process had been created, so output the old process still held could appear after the new one's first lines. On Linux they were not flushed at all before exec, which discards them. Flush the log and every stream before starting the new program. * loader: continue the log file across LoadExec With --printf-direction File, the program LoadExec starts is given the same --printf-output-file and opened it anew, emptying it. Everything the first program had logged, its LoadExec line included, was lost. LoadExec now sets KYTY_RELAUNCHED for the program it starts, which then appends to the log file instead of replacing it. On Linux the variable goes into the environment passed to exec rather than this process's own, which other threads may be reading. * loader: show LoadExec switches on the console LoadExec logged the program it switched to only when guest output was logged, and then only into the log: on a console the new program's startup output followed the old one's without anything marking where one program ended and the next began. Write the switch to the console as well as the log, with the program's arguments.	2026-10-16 03:01:17 +02:00