graphics: materialize CMASK clears for multisampled targets (PPSA13794)

Add in-game performance panel

Co-authored-by: nmzik <Nmzik@mail.ru>

CMakeLists.txt

@@ -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

assets/localization/ar-AE.json

@@ -1,3 +1,4 @@
 {
-  "trophy.toast.earned": "تم كسب الكأس!"
+  "trophy.toast.earned": "تم كسب الكأس!",
+  "user.guest": "ضيف"
 }

assets/localization/cs-CZ.json

@@ -1,3 +1,4 @@
 {
-  "trophy.toast.earned": "Trofej získána!"
+  "trophy.toast.earned": "Trofej získána!",
+  "user.guest": "Host"
 }

assets/localization/da-DK.json

@@ -1,3 +1,4 @@
 {
-  "trophy.toast.earned": "Trofæ optjent!"
+  "trophy.toast.earned": "Trofæ optjent!",
+  "user.guest": "Gæst"
 }

assets/localization/de-DE.json

@@ -1,3 +1,4 @@
 {
-  "trophy.toast.earned": "Trophäe erhalten!"
+  "trophy.toast.earned": "Trophäe erhalten!",
+  "user.guest": "Gast"
 }

assets/localization/el-GR.json

@@ -1,3 +1,4 @@
 {
-  "trophy.toast.earned": "Κερδίσατε ένα τρόπαιο!"
+  "trophy.toast.earned": "Κερδίσατε ένα τρόπαιο!",
+  "user.guest": "Επισκ."
 }

assets/localization/en-GB.json

@@ -1,3 +1,4 @@
 {
-  "trophy.toast.earned": "Trophy earned!"
+  "trophy.toast.earned": "Trophy earned!",
+  "user.guest": "Guest"
 }

assets/localization/en-US.json

@@ -1,3 +1,4 @@
 {
-  "trophy.toast.earned": "Trophy earned!"
+  "trophy.toast.earned": "Trophy earned!",
+  "user.guest": "Guest"
 }

assets/localization/es-419.json

@@ -1,3 +1,4 @@
 {
-  "trophy.toast.earned": "¡Ganaste un trofeo!"
+  "trophy.toast.earned": "¡Ganaste un trofeo!",
+  "user.guest": "Invitado"
 }

assets/localization/es-ES.json

@@ -1,3 +1,4 @@
 {
-  "trophy.toast.earned": "¡Ganaste un trofeo!"
+  "trophy.toast.earned": "¡Ganaste un trofeo!",
+  "user.guest": "Invitado"
 }

assets/localization/fi-FI.json

@@ -1,3 +1,4 @@
 {
-  "trophy.toast.earned": "Pokaali ansaittu!"
+  "trophy.toast.earned": "Pokaali ansaittu!",
+  "user.guest": "Vieras"
 }

assets/localization/fr-CA.json

@@ -1,3 +1,4 @@
 {
-  "trophy.toast.earned": "Trophée obtenu !"
+  "trophy.toast.earned": "Trophée obtenu !",
+  "user.guest": "Invité"
 }

assets/localization/fr-FR.json

@@ -1,3 +1,4 @@
 {
-  "trophy.toast.earned": "Trophée obtenu !"
+  "trophy.toast.earned": "Trophée obtenu !",
+  "user.guest": "Invité"
 }

assets/localization/hu-HU.json

@@ -1,3 +1,4 @@
 {
-  "trophy.toast.earned": "Trófea megszerezve!"
+  "trophy.toast.earned": "Trófea megszerezve!",
+  "user.guest": "Vendég"
 }

assets/localization/id-ID.json

@@ -1,3 +1,4 @@
 {
-  "trophy.toast.earned": "Trofi diperoleh!"
+  "trophy.toast.earned": "Trofi diperoleh!",
+  "user.guest": "Tamu"
 }

assets/localization/it-IT.json

@@ -1,3 +1,4 @@
 {
-  "trophy.toast.earned": "Trofeo ottenuto!"
+  "trophy.toast.earned": "Trofeo ottenuto!",
+  "user.guest": "Ospite"
 }

assets/localization/ja-JP.json

@@ -1,3 +1,4 @@
 {
-  "trophy.toast.earned": "トロフィーを獲得しました！"
+  "trophy.toast.earned": "トロフィーを獲得しました！",
+  "user.guest": "ゲスト"
 }

assets/localization/ko-KR.json

@@ -1,3 +1,4 @@
 {
-  "trophy.toast.earned": "트로피를 획득했습니다!"
+  "trophy.toast.earned": "트로피를 획득했습니다!",
+  "user.guest": "게스트"
 }

assets/localization/nl-NL.json

@@ -1,3 +1,4 @@
 {
-  "trophy.toast.earned": "Trofee behaald!"
+  "trophy.toast.earned": "Trofee behaald!",
+  "user.guest": "Gast"
 }

assets/localization/no-NO.json

@@ -1,3 +1,4 @@
 {
-  "trophy.toast.earned": "Trofé opptjent!"
+  "trophy.toast.earned": "Trofé opptjent!",
+  "user.guest": "Gjest"
 }

assets/localization/pl-PL.json

@@ -1,3 +1,4 @@
 {
-  "trophy.toast.earned": "Zdobyto trofeum!"
+  "trophy.toast.earned": "Zdobyto trofeum!",
+  "user.guest": "Gość"
 }

assets/localization/pt-BR.json

@@ -1,3 +1,4 @@
 {
-  "trophy.toast.earned": "Troféu conquistado!"
+  "trophy.toast.earned": "Troféu conquistado!",
+  "user.guest": "Convidado"
 }

assets/localization/pt-PT.json

@@ -1,3 +1,4 @@
 {
-  "trophy.toast.earned": "Troféu conquistado!"
+  "trophy.toast.earned": "Troféu conquistado!",
+  "user.guest": "Convidado"
 }

assets/localization/ro-RO.json

@@ -1,3 +1,4 @@
 {
-  "trophy.toast.earned": "Trofeu câștigat!"
+  "trophy.toast.earned": "Trofeu câștigat!",
+  "user.guest": "Invitat"
 }

assets/localization/ru-RU.json

@@ -1,3 +1,4 @@
 {
-  "trophy.toast.earned": "Трофей получен!"
+  "trophy.toast.earned": "Трофей получен!",
+  "user.guest": "Гость"
 }

assets/localization/sv-SE.json

@@ -1,3 +1,4 @@
 {
-  "trophy.toast.earned": "Trofé intjänad!"
+  "trophy.toast.earned": "Trofé intjänad!",
+  "user.guest": "Gäst"
 }

assets/localization/th-TH.json

@@ -1,3 +1,4 @@
 {
-  "trophy.toast.earned": "ได้รับถ้วยรางวัลแล้ว!"
+  "trophy.toast.earned": "ได้รับถ้วยรางวัลแล้ว!",
+  "user.guest": "แขก"
 }

assets/localization/tr-TR.json

@@ -1,3 +1,4 @@
 {
-  "trophy.toast.earned": "Kupa kazanıldı!"
+  "trophy.toast.earned": "Kupa kazanıldı!",
+  "user.guest": "Misafir"
 }

assets/localization/vi-VN.json

@@ -1,3 +1,4 @@
 {
-  "trophy.toast.earned": "Đã nhận cúp!"
+  "trophy.toast.earned": "Đã nhận cúp!",
+  "user.guest": "Khách"
 }

assets/localization/zh-Hans.json

@@ -1,3 +1,4 @@
 {
-  "trophy.toast.earned": "已获得奖杯！"
+  "trophy.toast.earned": "已获得奖杯！",
+  "user.guest": "访客"
 }

assets/localization/zh-Hant.json

@@ -1,3 +1,4 @@
 {
-  "trophy.toast.earned": "已獲得獎盃！"
+  "trophy.toast.earned": "已獲得獎盃！",
+  "user.guest": "訪客"
 }

src/common/CMakeLists.txt

@@ -43,6 +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})

src/common/debugCounters.h

@@ -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_ */

src/common/emulatorConfig.cpp

@@ -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;
 }
@@ -164,6 +189,10 @@ bool SyncRawImageBuffersEnabled() {
 	return g_config->sync_raw_image_buffers;
 }
 
+bool SupportZmaskEnabled() {
+	return g_config->support_zmask;
+}
+
 bool TessellationEnabled() {
 	return g_config->tessellation_enabled;
 }

src/common/emulatorConfig.h

@@ -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;
@@ -77,9 +86,11 @@ struct ConfigOptions {
 	bool                   renderdoc_enabled           = false;
 	bool                   readback_linear_images      = false;
 	bool                   sync_raw_image_buffers      = false;
+	bool                   support_zmask               = false;
 	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 +111,13 @@ 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();
@@ -134,6 +152,7 @@ bool GpuAssistedValidationEnabled();
 bool RenderDocEnabled();
 bool ReadbackLinearImagesEnabled();
 bool SyncRawImageBuffersEnabled();
+bool SupportZmaskEnabled();
 bool TessellationEnabled();
 bool TrophyEnabled();
 uint32_t GetTrophyDurationSeconds();

src/common/hostStats.cpp

@@ -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

src/common/hostStats.h

@@ -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_ */

src/graphics/guest_gpu/graphicsRun.cpp

@@ -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"
@@ -906,6 +907,7 @@ 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);

src/graphics/host_gpu/graphicContext.h

@@ -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);
 

src/graphics/host_gpu/renderer/cache/bufferCache.cpp

@@ -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"
@@ -177,6 +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;
@@ -467,6 +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) {
@@ -658,6 +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;
 	});

src/graphics/host_gpu/renderer/cache/textureCache.cpp

@@ -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"
@@ -262,15 +263,6 @@ 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);
-		}
-	}
 	UnregisterImage(id);
 	if (m_scheduler.Active()) {
 		m_scheduler.DeferOperation([this, id] { m_slot_images.erase(id); });
@@ -724,6 +716,13 @@ ImageId TextureCache::ResolveDepthOverlap(const ImageInfo& requested, BindingTyp
 		           "TextureCache: unsupported unequal-sample depth overlap copy (%u -> %u)\n",
 		           cached.backing.samples, replacement.backing.samples);
 	}
+	if (info.IsDepth() && info.metadata.kind == ImageMetadataKind::Htile) {
+		// Copied depth supersedes pending clears from the previous native image.
+		const auto metadata = m_surface_metas.find(info.metadata.range.address);
+		if (metadata != m_surface_metas.end()) {
+			metadata->second.clear_mask.reset();
+		}
+	}
 	FreeImage(cached_id);
 	return replacement_id;
 }
@@ -954,6 +953,7 @@ 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 +999,7 @@ 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 +1040,7 @@ 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 +1418,8 @@ 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 +1772,7 @@ 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;
@@ -1881,6 +1880,7 @@ void TextureCache::UnmapMemory(uint64_t address, uint64_t size) {
 		EXIT("TextureCache: invalid unmap range\n");
 	}
 	std::scoped_lock lock {m_lock};
+	// HTILE belongs to guest memory, independently of native image replacements.
 	for (auto metadata = m_surface_metas.begin(); metadata != m_surface_metas.end();) {
 		const auto base = metadata->first;
 		if (base >= address && base < address + size) {
@@ -1943,6 +1943,7 @@ 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;

src/graphics/host_gpu/renderer/cache/textureCache.h

@@ -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;
 	};

src/graphics/host_gpu/renderer/colorRenderTarget.cpp

@@ -220,8 +220,7 @@ 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 +232,7 @@ 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 =

src/graphics/host_gpu/renderer/commandScheduler.cpp

@@ -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,

src/graphics/host_gpu/renderer/debug.cpp

@@ -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;
-			}
-		}
 	}
 }
 

src/graphics/host_gpu/renderer/depthRenderTarget.cpp

@@ -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;

src/graphics/host_gpu/renderer/depthRenderTarget.h

@@ -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.

src/graphics/host_gpu/renderer/masterSemaphore.cpp

@@ -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();

src/graphics/host_gpu/renderer/pipeline/descriptors.cpp

@@ -126,9 +126,13 @@ NativeStorageBuffer(RenderContext& context, const PreparedBindings::BufferSource
 	if (size > graphics.GetPhysicalDeviceProperties().limits.maxStorageBufferRange) {
 		EXIT("storage buffer range is unsupported\n");
 	}
-	auto [buffer, offset] = context.GetBufferCache().ObtainBuffer(
-	    address, size, resource.written,
+	const bool metadata_read = resource.read && Config::SupportZmaskEnabled() &&
+	                           (address & 0x7fffu) == 0 &&
+	                           context.GetTextureCache().IsMeta(address);
+	auto [buffer, offset]    = context.GetBufferCache().ObtainBuffer(
+	    address, size, resource.written || metadata_read,
 	    resource.formatted || (resource.read && Config::SyncRawImageBuffersEnabled()), id);
+	if (metadata_read) buffer->Fill(offset, size, 0xf); // HTILE ZMask::kUncompressed.
 	const auto aligned_offset = Common::AlignDown(offset, alignment);
 	const auto adjustment     = offset - aligned_offset;
 	const auto max_range      = graphics.GetPhysicalDeviceProperties().limits.maxStorageBufferRange;

src/graphics/host_gpu/renderer/pipeline/pipelineCache.cpp

@@ -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"
@@ -24,6 +25,7 @@
 #include <array>
 #include <atomic>
 #include <cctype>
+#include <chrono>
 #include <cstdio>
 #include <cstring>
 #include <fmt/format.h>
@@ -299,6 +301,9 @@ 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,
@@ -993,8 +998,13 @@ 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);
@@ -1024,7 +1034,13 @@ 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);

src/graphics/host_gpu/renderer/renderCompute.cpp

@@ -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"
@@ -94,9 +95,10 @@ 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 +110,11 @@ 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))) {
@@ -252,10 +258,6 @@ 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;
@@ -431,6 +433,7 @@ 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);
+	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);
@@ -500,6 +503,7 @@ 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();
 }

src/graphics/host_gpu/renderer/renderDraw.cpp

@@ -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);

src/graphics/host_gpu/vma.cpp

@@ -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;

src/graphics/presentation/performanceOverlay.cpp

@@ -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

src/graphics/presentation/performanceOverlay.h

@@ -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_

src/graphics/presentation/renderDoc.cpp

@@ -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;
 		}

src/graphics/presentation/systemOverlay.cpp

@@ -1,7 +1,5 @@
 #include "graphics/presentation/systemOverlay.h"
 
-#include <SDL3/SDL.h>
-
 #include "common/assert.h"
 #include "common/emulatorConfig.h"
 #include "common/file.h"
@@ -11,10 +9,11 @@
 #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"
@@ -22,6 +21,7 @@
 #include "loader/systemContent.h"
 #include "stb_image.h"
 
+#include <SDL3/SDL.h>
 #include <algorithm>
 #include <array>
 #include <atomic>
@@ -712,7 +712,8 @@ SystemOverlayVisualState GetSystemOverlayVisualState() noexcept {
 	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) + static_cast<decltype(core.revision)>(cursor.revision)};
+	            (splash ? 2u : 0u) + static_cast<decltype(core.revision)>(cursor.revision) +
+	            PerformanceOverlayRevision()};
 }
 
 bool ProcessSystemOverlayInput(const SDL_Event& event) {
@@ -1662,8 +1663,7 @@ struct SystemOverlay::Impl {
 		}
 	}
 
-
-	void DrawTrophyNotification(vk::Extent2D frame_extent, Clock::time_point now) {
+	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;
@@ -1731,7 +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) {
@@ -1799,7 +1799,7 @@ struct SystemOverlay::Impl {
 		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) &&
-		    !GetHostInputTouchPadVisualState().enabled) {
+		    !GetHostInputTouchPadVisualState().enabled && !PerformanceOverlayEnabled()) {
 			return false;
 		}
 		const auto prepared_session = snapshot.session;
@@ -1830,6 +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) {
@@ -1838,7 +1839,7 @@ 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;

src/graphics/presentation/videoOut.cpp

@@ -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();

src/graphics/presentation/window/hostInput.cpp

@@ -254,7 +254,8 @@ 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());

src/graphics/presentation/window/swapchain.cpp

@@ -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());
 	}
@@ -1039,8 +1042,8 @@ void Presenter::Impl::Present() {
 		// 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 && swapchain.PrepareSystemOverlay();
+		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);
@@ -1060,6 +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) {

src/graphics/presentation/window/vulkanWindow.cpp

@@ -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);
 			}

src/graphics/presentation/window/window.cpp

@@ -1,21 +1,20 @@
 #include "graphics/presentation/window.h"
 
-#include <SDL3/SDL.h>
-
 #include "common/assert.h"
 #include "common/common.h"
 #include "common/emulatorConfig.h"
 #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
@@ -459,12 +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;

src/graphics/shader/recompiler/backend/spirv/spirvEmitterAlu.cpp

@@ -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();
@@ -210,28 +248,18 @@ 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) {

src/graphics/shader/recompiler/backend/spirv/spirvEmitterInstructions.h

@@ -296,6 +296,7 @@ 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 +326,7 @@ 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;

src/graphics/shader/recompiler/backend/spirv/spirvEmitterInternal.h

@@ -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;

src/graphics/shader/recompiler/backend/spirv/spirvEmitterMemory.cpp

@@ -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);
 		    });
 	    });
 }

src/graphics/shader/recompiler/backend/spirv/spirvEmitterMemoryHelpers.cpp

@@ -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),

src/graphics/shader/recompiler/frontend/decode/MemoryOps.cpp

@@ -67,6 +67,7 @@ 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},
     {0x80u, Opcode::BUFFER_LOAD_FORMAT_D16_X, 1, 16, false, false, true},
@@ -114,6 +115,7 @@ 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 +202,7 @@ 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 +223,7 @@ 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:
@@ -454,6 +458,9 @@ 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 ||

src/graphics/shader/recompiler/frontend/decode/ShaderDecoder.cpp

@@ -670,6 +670,7 @@ std::string InstructionToString(const Instruction& inst) {
 		case Opcode::BUFFER_ATOMIC_CMPSWAP:
 		case Opcode::BUFFER_ATOMIC_SWAP_X2:
 		case Opcode::BUFFER_ATOMIC_ADD:
+		case Opcode::BUFFER_ATOMIC_ADD_X2:
 		case Opcode::BUFFER_ATOMIC_SUB:
 		case Opcode::BUFFER_ATOMIC_SMIN:
 		case Opcode::BUFFER_ATOMIC_UMIN:
@@ -717,6 +718,7 @@ std::string InstructionToString(const Instruction& inst) {
 		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:

src/graphics/shader/recompiler/frontend/decode/ShaderDecoder.h

@@ -586,6 +586,7 @@ enum class Opcode {
 	BUFFER_ATOMIC_CMPSWAP,
 	BUFFER_ATOMIC_SWAP_X2,
 	BUFFER_ATOMIC_ADD,
+	BUFFER_ATOMIC_ADD_X2,
 	BUFFER_ATOMIC_SUB,
 	BUFFER_ATOMIC_SMIN,
 	BUFFER_ATOMIC_UMIN,
@@ -631,6 +632,7 @@ enum class Opcode {
 	DS_MAX_U32,
 	DS_MAX_RTN_U32,
 	DS_MAX_RTN_U64,
+	DS_CMPST_RTN_B32,
 	DS_CMPST_RTN_B64,
 	DS_AND_B32,
 	DS_AND_RTN_B32,

src/graphics/shader/recompiler/frontend/translate/Float.cpp

@@ -27,7 +27,7 @@ void Translator::PackedFloat16(const Decoder::Instruction& inst, IR::ValueOpcode
 	if (quiet_snan) {
 		const auto quiet_snan_lane = [&](const Decoder::Operand& operand, bool high) {
 			const auto bits     = ReadU16LaneAsU32(operand, high, false);
-			const auto exponent = ir.BitwiseAnd(bits, IR::U32(IR::Value(0x7c00u)));
+			const auto exponent = ir.BitwiseAnd(bits, IR::U32(IR::Value(0x7e00u)));
 			const auto payload  = ir.BitwiseAnd(bits, IR::U32(IR::Value(0x01ffu)));
 			const auto snan     = ir.LogicalAnd(ir.IEqual(exponent, IR::U32(IR::Value(0x7c00u))),
 			                                    ir.INotEqual(payload, IR::U32(IR::Value(0u))));

src/graphics/shader/recompiler/frontend/translate/Memory.cpp

@@ -1046,6 +1046,8 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 			return BUFFER_ATOMIC(inst, IR::ValueOpcode::BufferAtomicFMin32);
 		case Decoder::Opcode::BUFFER_ATOMIC_FMAX:
 			return BUFFER_ATOMIC(inst, IR::ValueOpcode::BufferAtomicFMax32);
+		case Decoder::Opcode::BUFFER_ATOMIC_ADD_X2:
+			return BUFFER_ATOMIC(inst, IR::ValueOpcode::BufferAtomicIAdd64);
 
 		case Decoder::Opcode::DS_ADD_U32:
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicIAdd32, false);
@@ -1083,6 +1085,8 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicUMax32, true);
 		case Decoder::Opcode::DS_MAX_RTN_U64:
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicUMax64, true);
+		case Decoder::Opcode::DS_CMPST_RTN_B32:
+			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicCmpSwap32, true);
 		case Decoder::Opcode::DS_CMPST_RTN_B64:
 			return DS_ATOMIC(inst, IR::ValueOpcode::SharedAtomicCmpSwap64, true);
 		case Decoder::Opcode::DS_AND_B32:

src/graphics/shader/recompiler/frontend/translate/Translate.cpp

@@ -233,13 +233,6 @@ IR::Value Translator::ReadOperand(const Decoder::Operand& operand, IR::Type type
 		if (type == IR::Type::F64) {
 			if (operand.kind == Decoder::OperandKind::LiteralConstant) {
 				pair = {IR::U32(IR::Value(0u)), IR::U32(IR::Value(operand.value))};
-			} else if (operand.kind == Decoder::OperandKind::FloatInlineConstant) {
-				const auto bits = operand.value == 0x3e22f983u
-				                      ? 0x3fc45f306dc9c882ull
-				                      : std::bit_cast<uint64_t>(static_cast<double>(
-				                            std::bit_cast<float>(operand.value)));
-				pair            = {IR::U32(IR::Value(static_cast<uint32_t>(bits))),
-				                   IR::U32(IR::Value(static_cast<uint32_t>(bits >> 32u)))};
 			}
 			if (operand.absolute) {
 				pair[1] = ir.BitwiseAnd(pair[1], IR::U32(IR::Value(0x7fffffffu)));
@@ -488,6 +481,14 @@ std::array<IR::U32, 2> Translator::ReadU32Pair(const Decoder::Operand& operand)
 	if (operand.kind == Decoder::OperandKind::VccLo) {
 		return {ir.GetVccLo(), ir.GetVccHi()};
 	}
+	if (operand.kind == Decoder::OperandKind::FloatInlineConstant) {
+		const auto bits =
+		    operand.value == 0x3e22f983u
+		        ? 0x3fc45f306dc9c882ull
+		        : std::bit_cast<uint64_t>(static_cast<double>(std::bit_cast<float>(operand.value)));
+		return {IR::U32(IR::Value(static_cast<uint32_t>(bits))),
+		        IR::U32(IR::Value(static_cast<uint32_t>(bits >> 32u)))};
+	}
 	const auto low = ApplyBitSourceModifiers(operand, ReadRawU32(operand));
 	IR::U32    high(IR::Value(0u));
 	if (operand.kind == Decoder::OperandKind::Sgpr || operand.kind == Decoder::OperandKind::Vgpr) {
@@ -546,9 +547,11 @@ IR::U32 Translator::Read16LaneBits(const Decoder::Operand& operand, bool high_la
 	auto bits             = ReadRawU32(operand);
 	if (operand.kind == Decoder::OperandKind::FloatInlineConstant) {
 		// Inline floats encode the operand's width before SDWA or OP_SEL selects bits.
-		const auto half      = IR::F16(ir.Emit(IR::ValueOpcode::ConvertF16F32, {ir.BitCastF32(bits)}));
-		const auto half_bits = IR::U16(ir.Emit(IR::ValueOpcode::BitCastU16F16, {half}));
-		bits                = IR::U32(ir.Emit(IR::ValueOpcode::ConvertU32U16, {half_bits}));
+		const auto value = operand.value;
+		bits             = IR::U32(IR::Value(value == 0x3e22f983u
+		                                         ? 0x3118u
+		                                         : ((value >> 16u) & 0x8000u) |
+		                                               ((((value >> 23u) & 0xffu) - 112u) << 10u)));
 	}
 	bits = ApplyBitSourceModifiers(raw_operand, bits);
 	uint32_t offset = (high_lane ? operand.op_sel_hi : operand.op_sel) ? 16u : 0u;

src/graphics/shader/recompiler/ir/ResourceSnapshot.h

@@ -25,6 +25,7 @@ struct UniformFill {
 	std::array<uint32_t, 3> group_stride {};
 	uint32_t                 words        = 0;
 	uint32_t                 value        = 0;
+	uint32_t                 bounded_elements = 0;
 
 	bool operator==(const UniformFill&) const = default;
 };

src/graphics/shader/recompiler/ir/ShaderIR.h

@@ -82,6 +82,11 @@ struct MemoryInfo {
 	bool                    coherent                                              = false;
 	bool                    planning_only                                         = false;
 
+	[[nodiscard]] bool IsCoherentBufferPairLoad(ValueOpcode opcode) const {
+		return kind == ResourceKind::Buffer && opcode == ValueOpcode::LoadBufferU32x2 && coherent &&
+		       !typed && !formatted && data_bits == 32u;
+	}
+
 	[[nodiscard]] bool SupportsIndirectBufferLoad(ValueOpcode opcode) const {
 		if (typed) return false;
 		if (data_bits == 8u) return !formatted && opcode == ValueOpcode::LoadBufferU8;
@@ -622,6 +627,8 @@ struct CompiledShaderInfo {
 struct UniformFillPlan {
 	UniformFill          fill;
 	std::array<Value, 4> values;
+	Value                pattern_size;
+	Value                element_count;
 };
 
 // Resource analysis retained by the shader cache. It owns immutable descriptor/SRT,

src/graphics/shader/recompiler/ir/opcodes/ValueOpcodes.cpp

@@ -118,6 +118,7 @@ BufferAccess BufferAccessOf(ValueOpcode opcode) {
 		case ValueOpcode::BufferAtomicCmpSwap32:
 		case ValueOpcode::BufferAtomicSwap64:
 		case ValueOpcode::BufferAtomicIAdd32:
+		case ValueOpcode::BufferAtomicIAdd64:
 		case ValueOpcode::BufferAtomicISub32:
 		case ValueOpcode::BufferAtomicSMin32:
 		case ValueOpcode::BufferAtomicUMin32:
@@ -137,6 +138,7 @@ BufferAccess BufferAccessOf(ValueOpcode opcode) {
 uint32_t BufferComponentCount(ValueOpcode opcode) {
 	switch (opcode) {
 		case ValueOpcode::BufferAtomicSwap64:
+		case ValueOpcode::BufferAtomicIAdd64:
 		case ValueOpcode::BufferAtomicAnd64:
 		case ValueOpcode::BufferAtomicOr64:
 		case ValueOpcode::LoadBufferU32x2:
@@ -176,6 +178,7 @@ SharedAccess SharedAccessOf(ValueOpcode opcode) {
 		case ValueOpcode::SharedAtomicSMax32:
 		case ValueOpcode::SharedAtomicUMax32:
 		case ValueOpcode::SharedAtomicUMax64:
+		case ValueOpcode::SharedAtomicCmpSwap32:
 		case ValueOpcode::SharedAtomicCmpSwap64:
 		case ValueOpcode::SharedAtomicAnd32:
 		case ValueOpcode::SharedAtomicOr32:

src/graphics/shader/recompiler/ir/opcodes/ValueOpcodes.inc

@@ -268,6 +268,7 @@ VALUE_OPCODE(BufferAtomicSwap32, U32, BufferResource, U32, U32, U32, U32, U1)
 VALUE_OPCODE(BufferAtomicCmpSwap32, U32, BufferResource, U32, U32, U32, U32, U32, U1)
 VALUE_OPCODE(BufferAtomicSwap64, U64, BufferResource, U32, U32, U32, U64, U1)
 VALUE_OPCODE(BufferAtomicIAdd32, U32, BufferResource, U32, U32, U32, U32, U1)
+VALUE_OPCODE(BufferAtomicIAdd64, U64, BufferResource, U32, U32, U32, U64, U1)
 VALUE_OPCODE(BufferAtomicISub32, U32, BufferResource, U32, U32, U32, U32, U1)
 VALUE_OPCODE(BufferAtomicSMin32, U32, BufferResource, U32, U32, U32, U32, U1)
 VALUE_OPCODE(BufferAtomicUMin32, U32, BufferResource, U32, U32, U32, U32, U1)
@@ -306,6 +307,7 @@ VALUE_OPCODE(SharedAtomicUMin32, U32, U32, U32, U1)
 VALUE_OPCODE(SharedAtomicSMax32, U32, U32, U32, U1)
 VALUE_OPCODE(SharedAtomicUMax32, U32, U32, U32, U1)
 VALUE_OPCODE(SharedAtomicUMax64, U64, U32, U64, U1)
+VALUE_OPCODE(SharedAtomicCmpSwap32, U32, U32, U32, U32, U1)
 VALUE_OPCODE(SharedAtomicCmpSwap64, U64, U32, U64, U64, U1)
 VALUE_OPCODE(SharedAtomicAnd32, U32, U32, U32, U1)
 VALUE_OPCODE(SharedAtomicOr32, U32, U32, U32, U1)

src/graphics/shader/recompiler/ir/passes/ConstantPropagation.cpp

@@ -51,11 +51,13 @@ public:
 	}
 
 	void QueueUsers(const std::vector<Use>& users) {
-		++generation;
-		affected.clear();
 		for (const auto& use: users) affected.push_back(use.user);
+	}
+
+	void FlushUsers() {
+		++generation;
 		// Pattern folds inspect nested operands even when intermediate operations
-		// stay dynamic. Visit consumers once per change, including cyclic Phi edges.
+		// stay dynamic. Visit their union once per drain, including cyclic Phi edges.
 		for (size_t i = 0; i < affected.size(); ++i) {
 			auto* inst = affected[i];
 			const auto found = locations.find(inst);
@@ -64,6 +66,7 @@ public:
 			Queue(inst);
 			for (const auto& use: inst->Uses()) affected.push_back(use.user);
 		}
+		affected.clear();
 	}
 
 	Block::iterator Prepend(Block& block, Block::iterator before, ValueOpcode opcode,
@@ -977,27 +980,30 @@ void ConstantPropagationPass(const BlockList& blocks, uint32_t wave_size) {
 	std::vector<Use> users;
 	std::vector<Value> operands;
 	const auto drain = [&] {
-		for (size_t i = 0; i < worklist.pending.size(); ++i) {
-			auto* inst = worklist.pending[i];
-			auto& location = worklist.locations.at(inst);
-			location.queued = false;
-			const auto opcode = inst->GetOpcode();
-			if (opcode == ValueOpcode::Identity || opcode == ValueOpcode::Void) continue;
-			const auto where = location.instruction;
-			users.assign(inst->Uses().begin(), inst->Uses().end());
-			operands.clear();
-			for (size_t arg = 0; arg < inst->NumArgs(); ++arg) operands.push_back(inst->Arg(arg));
-			mask_projection.Fold(*inst, wave_size);
-			FoldInstruction(*inst->Parent(), where, worklist);
-			bool changed = opcode != inst->GetOpcode() || operands.size() != inst->NumArgs();
-			for (size_t arg = 0; !changed && arg < operands.size(); ++arg)
-				changed = operands[arg] != inst->Arg(arg);
-			if (changed) {
-				worklist.QueueUsers(users);
-				worklist.Queue(inst);
+		do {
+			for (size_t i = 0; i < worklist.pending.size(); ++i) {
+				auto* inst = worklist.pending[i];
+				auto& location = worklist.locations.at(inst);
+				location.queued = false;
+				const auto opcode = inst->GetOpcode();
+				if (opcode == ValueOpcode::Identity || opcode == ValueOpcode::Void) continue;
+				const auto where = location.instruction;
+				users.assign(inst->Uses().begin(), inst->Uses().end());
+				operands.clear();
+				for (size_t arg = 0; arg < inst->NumArgs(); ++arg) operands.push_back(inst->Arg(arg));
+				mask_projection.Fold(*inst, wave_size);
+				FoldInstruction(*inst->Parent(), where, worklist);
+				bool changed = opcode != inst->GetOpcode() || operands.size() != inst->NumArgs();
+				for (size_t arg = 0; !changed && arg < operands.size(); ++arg)
+					changed = operands[arg] != inst->Arg(arg);
+				if (changed) {
+					worklist.QueueUsers(users);
+					worklist.Queue(inst);
+				}
 			}
-		}
-		worklist.pending.clear();
+			worklist.pending.clear();
+			worklist.FlushUsers();
+		} while (!worklist.pending.empty());
 	};
 	drain();
 	// Normalize retained PHI/select values only after every supported field read has

src/graphics/shader/recompiler/ir/passes/ResourceMaterialization.cpp

@@ -1409,7 +1409,8 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 		}
 		if (op != ValueOpcode::LogicalAnd && op != ValueOpcode::LogicalOr &&
 		    op != ValueOpcode::LogicalNot && op != ValueOpcode::IEqual32 &&
-		    op != ValueOpcode::INotEqual32 && op != ValueOpcode::ULessThanEqual32 &&
+		    op != ValueOpcode::INotEqual32 && op != ValueOpcode::ULessThan32 &&
+		    op != ValueOpcode::ULessThanEqual32 &&
 		    op != ValueOpcode::SelectU1 && op != ValueOpcode::SelectU32) return finish(opaque());
 		auto& node = plan.value_storage.emplace_back(op);
 		for (size_t index = 0; index < inst->NumArgs(); ++index)
@@ -1424,10 +1425,11 @@ ResourcePlan ExtractResourcePlan(const Program& program) {
 		for (auto it = first; it != plan.value_storage.end(); ++it) it->Invalidate();
 		plan.value_storage.erase(first, plan.value_storage.end());
 	}
-	plan.uniform_fill = AnalyzeUniformFill(program);
-	for (uint32_t i = 0; i < plan.uniform_fill.fill.words; ++i) {
-		plan.uniform_fill.values[i] = Clone(plan.uniform_fill.values[i]);
-	}
+	plan.uniform_fill = program.uniform_fill.fill.kind != UniformFillKind::None
+	    ? program.uniform_fill : AnalyzeUniformFill(program);
+	for (auto& value: plan.uniform_fill.values) value = Clone(value);
+	plan.uniform_fill.pattern_size = Clone(plan.uniform_fill.pattern_size);
+	plan.uniform_fill.element_count = Clone(plan.uniform_fill.element_count);
 	plan.capture_specialization_reads |=
 	    capture_indirect_reads || capture_control_reads || !plan.source_reads.empty();
 	if (plan.capture_specialization_reads) {
@@ -1513,15 +1515,21 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 	}
 	snapshot.uniform_fill = {};
 	const auto& fill = program.uniform_fill;
-	const auto words = fill.fill.words;
+	auto words = fill.fill.words;
 	std::array<uint32_t, 4> stored {};
 	bool uniform_fill = words != 0;
+	if (uniform_fill && !fill.pattern_size.IsEmpty())
+		uniform_fill = clean.Evaluate(fill.pattern_size, words) && words >= 1 && words <= 4;
+	uint32_t elements = 0;
+	if (uniform_fill && !fill.element_count.IsEmpty())
+		uniform_fill = clean.Evaluate(fill.element_count, elements) && elements != 0;
 	for (uint32_t i = 0; i < words && uniform_fill; ++i) {
 		uniform_fill = clean.Evaluate(fill.values[i], stored[i]) && stored[i] == stored[0];
 	}
 	if (uniform_fill) {
 		snapshot.uniform_fill = fill.fill;
 		snapshot.uniform_fill.value = stored[0];
+		snapshot.uniform_fill.bounded_elements = elements;
 	}
 	const auto evaluate = [&](uint32_t source, DescriptorValue& value, bool written = false) {
 		if (source >= program.descriptor_sources.size()) {

src/graphics/shader/recompiler/ir/passes/ResourceTracking.cpp

@@ -15,6 +15,8 @@
 #include <tuple>
 #include <unordered_map>
 #include <utility>
+#define XXH_INLINE_ALL
+#include <xxhash.h>
 
 namespace Libs::Graphics::ShaderRecompiler::IR {
 namespace {
@@ -4040,11 +4042,71 @@ private:
 	bool                                       m_shader_writes = false;
 };
 
+UniformFillPlan RecognizePatternFill(Program& program, const Decoder::Program& decoded,
+                                   ShaderStageInputInfo input) {
+	struct Pattern {
+		size_t code_words;
+		uint64_t hash;
+		uint32_t wave_size_mask;
+		bool bounded;
+	};
+	constexpr std::array patterns {
+		// Bounded 64-lane variant.
+		Pattern {.code_words = 69, .hash = 0x0704b9818e3cba06ull, .wave_size_mask = 64, .bounded = true},
+		// Dispatch thread dimensions, 64-lane variants.
+		Pattern {.code_words = 42, .hash = 0x92ce57f21310ef02ull, .wave_size_mask = 64, .bounded = false},
+		Pattern {.code_words = 40, .hash = 0x374161027f58bee7ull, .wave_size_mask = 64, .bounded = false},
+		Pattern {.code_words = 40, .hash = 0x7d7172c56c2e296dull, .wave_size_mask = 64, .bounded = false},
+		// Dispatch thread dimensions, valid in both wave sizes.
+		Pattern {.code_words = 38, .hash = 0xacbda1a1ccdd4198ull, .wave_size_mask = 32 | 64, .bounded = false},
+	};
+	if (program.stage != ShaderType::Compute || input.compute == nullptr ||
+	    decoded.instructions.empty() || program.user_data_base != 0 || program.user_data_count < 10 ||
+	    program.info.buffers.size() != 1 || !program.info.images.empty()) return {};
+	const auto& cs = *input.compute;
+	const bool matching_inputs = cs.workgroup_register == 10 && cs.thread_ids_num == 1 &&
+	    cs.group_id[0] && !cs.group_id[1] && !cs.group_id[2] && cs.threads_num[0] == 64 &&
+	    cs.threads_num[1] == 1 && cs.threads_num[2] == 1 && cs.float_mode == 0xc0;
+	if (!matching_inputs) return {};
+	const auto& last = decoded.instructions.back();
+	const size_t words = last.pc / sizeof(uint32_t) + last.word_count;
+	if (words > decoded.code.size() ||
+	    std::ranges::find(patterns, words, &Pattern::code_words) == patterns.end()) return {};
+	// Hash executable instructions through S_ENDPGM, excluding binary metadata.
+	// Verified variants implement Destination[id] = value[id % pattern_size].
+	const auto hash = XXH3_64bits(decoded.code.data(), words * sizeof(uint32_t));
+	const auto pattern = std::ranges::find_if(patterns, [&](const Pattern& candidate) {
+		return candidate.code_words == words && candidate.hash == hash;
+	});
+	if (pattern == patterns.end() || (program.wave_size & pattern->wave_size_mask) == 0 ||
+	    cs.dispatch_thread_dimensions != !pattern->bounded) return {};
+	std::array<Value, 10> userdata;
+	for (auto* block: program.blocks) {
+		for (auto& inst: *block) {
+			if (inst.GetOpcode() != ValueOpcode::GetUserData) continue;
+			const auto reg = RegIndex(inst.Arg(0).ScalarRegister());
+			if (reg < userdata.size()) userdata[reg] = Value(&inst);
+		}
+	}
+	for (uint32_t reg = 4; reg < 10; ++reg)
+		if ((reg != 8 || pattern->bounded) && userdata[reg].IsEmpty()) return {};
+	UniformFillPlan result;
+	result.fill.kind = UniformFillKind::Buffer;
+	result.fill.words = 1;
+	result.fill.group_stride[0] = 64;
+	std::copy_n(userdata.begin() + 4, 4, result.values.begin());
+	result.pattern_size = userdata[9];
+	// Only the bounded variant reads an explicit count; the other uses thread dimensions.
+	if (pattern->bounded) result.element_count = userdata[8];
+	return result;
+}
+
 } // namespace
 
 void TrackResources(Program& program, const Decoder::Program& decoded, const CFG::Graph& native_cfg,
                     ShaderStageInputInfo input) {
 	Tracker(program, decoded, native_cfg, input).Run();
+	program.uniform_fill = RecognizePatternFill(program, decoded, input);
 }
 
 } // namespace Libs::Graphics::ShaderRecompiler::IR

src/graphics/shader/recompiler/ir/passes/ShaderInfoCollection.cpp

@@ -340,6 +340,7 @@ void Visit(Program& program, ShaderStageInputInfo input_info, InputUsage& inputs
 			}
 			if (kind == ResourceKind::Buffer) {
 				info.coherent_buffers |= memory.coherent;
+				info.buffer_int64_atomics |= memory.IsCoherentBufferPairLoad(op);
 				if (memory.resource >= info.buffers.size()) {
 					return Fail("buffer operation has invalid resource metadata");
 				}

src/graphics/shader/recompiler/ir/passes/SrtWalker.cpp

@@ -1054,7 +1054,13 @@ bool SrtWalker::EvaluateInst(const Inst& inst, uint64_t& result) {
 			}
 			return false;
 		case ValueOpcode::ULessThan32:
-			if (binary()) {
+			// No U32 is below zero, including an unknown loop-carried value.
+			if (!Arg(inst, 1, b)) return false;
+			if (static_cast<uint32_t>(b) == 0u) {
+				result = 0u;
+				return true;
+			}
+			if (Arg(inst, 0, a)) {
 				result = static_cast<uint32_t>(a) < static_cast<uint32_t>(b);
 				return true;
 			}

src/graphics/shader/shader.cpp

@@ -583,16 +583,16 @@ ShaderParams PrepareProgram(const HW::VertexShaderInfo& regs, const HW::Context&
 		     mesh.input_primitive, sh.m_vgtGsOutPrimType, sh.m_vgtGsMaxVertOut,
 		     group.primitive_group_size, group.vertex_group_size, mesh.max_vertices);
 	}
-	mesh.max_primitives = mesh.fast_launch ? mesh.max_vertices :
-	                      native_ngg ? mesh.InputPrimitiveCount(mesh.max_vertices) :
-	                      group.primitive_group_size * (sh.m_vgtGsMaxVertOut - 2u);
 	mesh.primitives_per_group = std::min({static_cast<uint32_t>(group.primitive_group_size),
 	                                      mesh.InputPrimitiveCount(group.vertex_group_size),
-	                                      native_ngg ? mesh.max_primitives :
+	                                      native_ngg ? mesh.InputPrimitiveCount(mesh.max_vertices) :
 	                                          mesh.max_vertices / sh.m_vgtGsMaxVertOut});
 	EXIT_IF(mesh.primitives_per_group == 0u);
 	mesh.vertices_per_group = mesh.InputVertexCount(mesh.primitives_per_group);
-	const auto threads = native_ngg && !mesh.fast_launch ? mesh.vertices_per_group : mesh.max_vertices;
+	mesh.max_primitives = mesh.primitives_per_group * (sh.m_geNggSubgrpCntl & 0x1ffu);
+	const auto threads = std::max(native_ngg && !mesh.fast_launch ? mesh.vertices_per_group
+	                                                            : mesh.max_vertices,
+	                              mesh.max_primitives);
 	mesh.threads_num[0] = ((threads + mesh.wave_size - 1u) / mesh.wave_size) * mesh.wave_size;
 	mesh.threads_num[1] = mesh.threads_num[2] = 1u;
 	return params;

src/kernel/fileSystem.cpp

@@ -4,6 +4,7 @@
 #include "common/assert.h"
 #include "common/common.h"
 #include "common/dateTime.h"
+#include "common/debugCounters.h"
 #include "common/emulatorConfig.h"
 #include "common/file.h"
 #include "common/hash.h"
@@ -338,14 +339,20 @@ static bool HasWindowsForbiddenFilenameCharacter(const std::string& relative_pat
 std::filesystem::path MountPoints::ResolvePath(const std::string& mounted_name) {
 	Common::LockGuard lock(m_mutex);
 
+	// Guests join mount points and names with extra slashes (e.g. "/addcont0//file").
+	std::string normalized = Common::FixFilenameSlash(mounted_name);
+	normalized.erase(std::unique(normalized.begin(), normalized.end(),
+	                             [](char a, char b) { return a == '/' && b == '/'; }),
+	                 normalized.end());
+
 	// Match the entire guest path so a mount root works with or without a trailing slash.
-	const auto mounted_path = Common::FixDirectorySlash(mounted_name);
+	const auto mounted_path = Common::FixDirectorySlash(normalized);
 	const auto it           = std::find_if(
 	    m_mount_pairs.begin(), m_mount_pairs.end(),
 	    [&mounted_path](const MountPair& p) { return mounted_path.starts_with(p.point); });
 	if (it != m_mount_pairs.end()) {
 		const auto& p = *it;
-		auto rel_path = Common::RemoveFirst(Common::FixFilenameSlash(mounted_name), p.point.size());
+		auto rel_path = Common::RemoveFirst(normalized, p.point.size());
 		while (rel_path.starts_with('/')) {
 			rel_path = Common::RemoveFirst(rel_path, 1);
 		}
@@ -663,6 +670,7 @@ int64_t KYTY_SYSV_ABI KernelRead(int d, void* buf, size_t nbytes) {
 	                         std::min<uint64_t>(nbytes, remaining));
 	uint32_t bytes_read = 0;
 	file->f.Read(buf, static_cast<uint32_t>(nbytes), &bytes_read);
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::GuestFileReadBytes, bytes_read);
 
 	file->mutex.Unlock();
 
@@ -725,6 +733,7 @@ int64_t KYTY_SYSV_ABI KernelWrite(int d, const void* buf, size_t nbytes) {
 		file->f.Seek(file->f.Size());
 	}
 	file->f.Write(buf, static_cast<uint32_t>(nbytes), &bytes_written);
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::GuestFileWriteBytes, bytes_written);
 	if (file->sync_writes) {
 		file->f.Flush();
 	}
@@ -789,6 +798,7 @@ int64_t KYTY_SYSV_ABI KernelPread(int d, void* buf, size_t nbytes, int64_t offse
 	uint32_t bytes_read = 0;
 	file->f.Seek(offset);
 	file->f.Read(buf, static_cast<uint32_t>(nbytes), &bytes_read);
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::GuestFileReadBytes, bytes_read);
 	file->f.Seek(pos);
 
 	file->mutex.Unlock();
@@ -890,6 +900,7 @@ int64_t KYTY_SYSV_ABI KernelPreadv(int d, const KernelIovec* iov, int iovcnt, in
 		Memory::InvalidateMemory(reinterpret_cast<uint64_t>(buffer.iov_base), count);
 		uint32_t bytes = 0;
 		file->f.Read(buffer.iov_base, count, &bytes);
+		Common::DebugCounters::Add(Common::DebugCounters::Counter::GuestFileReadBytes, bytes);
 		bytes_read += bytes;
 		remaining -= bytes;
 		if (bytes < count) {
@@ -939,6 +950,7 @@ int64_t KYTY_SYSV_ABI KernelPwrite(int d, const void* buf, size_t nbytes, int64_
 	uint32_t bytes_written = 0;
 	file->f.Seek(file->append ? file->f.Size() : static_cast<uint64_t>(offset));
 	file->f.Write(buf, static_cast<uint32_t>(nbytes), &bytes_written);
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::GuestFileWriteBytes, bytes_written);
 	if (file->sync_writes) {
 		file->f.Flush();
 	}
@@ -1010,6 +1022,7 @@ static int64_t Writev(int d, const KernelIovec* iov, int iovcnt, std::optional<i
 		}
 		uint32_t bytes = 0;
 		file->f.Write(buffer.iov_base, static_cast<uint32_t>(buffer.iov_len), &bytes);
+		Common::DebugCounters::Add(Common::DebugCounters::Counter::GuestFileWriteBytes, bytes);
 		bytes_written += bytes;
 		if (bytes < buffer.iov_len) {
 			break;

src/kernel/memory.cpp

@@ -4414,4 +4414,39 @@ int KYTY_SYSV_ABI KernelMemoryPoolGetBlockStats(KernelMemoryPoolBlockStats* outp
 	return OK;
 }
 
+DebugStats GetDebugStats() {
+	DebugStats stats;
+	stats.direct_total       = PhysicalMemory::Size();
+	stats.flexible_total     = FlexibleMemory::Size();
+	stats.pool_committed     = g_memory_pool_committed.load(std::memory_order_relaxed);
+	stats.page_entries_total = PAGE_TABLE_POOL_ENTRIES;
+	if (g_physical_memory != nullptr) {
+		Common::LockGuard lock(g_physical_memory->GetMutex());
+		for (const auto& [start, block]: g_physical_memory->GetPhysicalBlocks()) {
+			switch (block.kind) {
+				case PhysicalMemory::AllocationKind::Direct:
+					stats.direct_allocated += block.size;
+					break;
+				case PhysicalMemory::AllocationKind::Pooled:
+					stats.pooled_allocated += block.size;
+					break;
+				case PhysicalMemory::AllocationKind::Automatic:
+					stats.automatic_allocated += block.size;
+					break;
+			}
+		}
+		for (const auto& mapping: g_physical_memory->GetMappings()) {
+			stats.direct_mapped += mapping.map_size;
+		}
+	}
+	if (g_flexible_memory != nullptr) {
+		stats.flexible_used = stats.flexible_total - g_flexible_memory->Available();
+	}
+	if (g_virtual_ranges != nullptr) {
+		stats.cpu_page_entries = g_virtual_ranges->CountPageTableEntries(false);
+		stats.gpu_page_entries = g_virtual_ranges->CountPageTableEntries(true);
+	}
+	return stats;
+}
+
 } // namespace Libs::LibKernel::Memory

src/kernel/memory.h

@@ -15,6 +15,23 @@ namespace Libs::LibKernel::Memory {
 void Initialize();
 void Shutdown();
 
+// Guest memory usage for the performance panel. Takes only bookkeeping locks, so it never
+// waits on the GPU; call it only while the memory subsystem is initialized.
+struct DebugStats {
+	uint64_t direct_total        = 0;
+	uint64_t direct_allocated    = 0; // sceKernelAllocateDirectMemory.
+	uint64_t pooled_allocated    = 0; // Memory pool expansions.
+	uint64_t automatic_allocated = 0; // Extended memory.
+	uint64_t direct_mapped       = 0;
+	uint64_t flexible_total      = 0;
+	uint64_t flexible_used       = 0;
+	uint64_t pool_committed      = 0;
+	uint64_t cpu_page_entries    = 0; // 2 MiB page-table entries in use.
+	uint64_t gpu_page_entries    = 0;
+	uint64_t page_entries_total  = 0; // Per side.
+};
+[[nodiscard]] DebugStats GetDebugStats();
+
 struct Lifecycle {
 	static constexpr const char* name       = "Memory";
 	static constexpr auto        initialize = Libs::LibKernel::Memory::Initialize;

src/kernel/module.h

@@ -0,0 +1,24 @@
+#ifndef EMULATOR_SRC_KERNEL_MODULE_H_
+#define EMULATOR_SRC_KERNEL_MODULE_H_
+
+#include "common/abi.h"
+
+#include <cstddef>
+#include <cstdint>
+
+namespace Libs::LibKernel {
+
+using KernelModule = int32_t;
+struct KernelLoadModuleOpt;
+struct KernelUnloadModuleOpt;
+
+KernelModule KYTY_SYSV_ABI KernelLoadStartModule(const char* module_file_name, size_t args,
+                                                 const void* argp, uint32_t flags,
+                                                 const KernelLoadModuleOpt* opt, int* res);
+int KYTY_SYSV_ABI KernelStopUnloadModule(KernelModule handle, size_t args, const void* argp,
+                                        uint32_t flags, const KernelUnloadModuleOpt* opt,
+                                        int* res);
+
+} // namespace Libs::LibKernel
+
+#endif // EMULATOR_SRC_KERNEL_MODULE_H_

src/launcher/forms/configuration_edit_dialog.ui

@@ -66,7 +66,6 @@
             <item row="1" column="1">
              <widget class="QSpinBox" name="spinBox_user_id">
               <property name="toolTip"><string>Local user ID exposed to games</string></property>
-              <property name="maximum"><number>2147483647</number></property>
              </widget>
             </item>
             <item row="2" column="0">
@@ -175,12 +174,6 @@
              </widget>
             </item>
             <item row="7" column="1">
-             <widget class="QCheckBox" name="checkBox_sync_raw_image_buffers">
-              <property name="toolTip"><string>Can fix missing textures and incorrect transparency in some games. May reduce performance.</string></property>
-              <property name="text"><string>Accurate texture synchronization</string></property>
-             </widget>
-            </item>
-            <item row="8" column="1">
              <widget class="QCheckBox" name="checkBox_hdr">
               <property name="toolTip"><string>Enable HDR10 output for supported games. Requires HDR enabled on the host display.</string></property>
               <property name="text"><string>Enable HDR display support</string></property>
@@ -243,6 +236,12 @@
               <property name="text"><string>Auto-hide cursor</string></property>
              </widget>
             </item>
+            <item>
+             <widget class="QCheckBox" name="checkBox_perf_overlay">
+              <property name="toolTip"><string>Show FPS, frame times, CPU/GPU usage, RAM and memory pressure in the game window. Press F2 (Cmd+P on macOS) to toggle it, and add Shift for the detailed view.</string></property>
+              <property name="text"><string>Show performance panel</string></property>
+             </widget>
+            </item>
             <item>
              <widget class="QCheckBox" name="checkBox_skip_notice_screen">
               <property name="toolTip"><string>Skip startup logos and notices in supported games.</string></property>
@@ -347,6 +346,16 @@
             </item>
            </layout>
           </item>
+          <item row="3" column="0">
+           <widget class="QLabel" name="label_keyboard_player">
+            <property name="text"><string>Keyboard player:</string></property>
+           </widget>
+          </item>
+          <item row="3" column="1">
+           <widget class="QComboBox" name="comboBox_keyboard_player">
+            <property name="toolTip"><string>Gamepads fill the remaining players in connection order. Every player uses the same input mapping.</string></property>
+           </widget>
+          </item>
          </layout>
         </widget>
        </item>
@@ -468,6 +477,31 @@
          </layout>
         </widget>
        </item>
+       <item>
+        <widget class="QGroupBox" name="rendering_compatibility_group">
+         <property name="title"><string>Rendering compatibility</string></property>
+         <layout class="QVBoxLayout" name="renderingCompatibilityLayout">
+          <item>
+           <widget class="QLabel" name="label_rendering_compatibility">
+            <property name="text"><string>These settings may help with rendering bugs at a slight performance cost.</string></property>
+            <property name="wordWrap"><bool>true</bool></property>
+           </widget>
+          </item>
+          <item>
+           <widget class="QCheckBox" name="checkBox_sync_raw_image_buffers">
+            <property name="toolTip"><string>Can fix missing textures and incorrect transparency (Resident Evil 2).</string></property>
+            <property name="text"><string>Accurate texture synchronization</string></property>
+           </widget>
+          </item>
+          <item>
+           <widget class="QCheckBox" name="checkBox_support_zmask">
+            <property name="toolTip"><string>Can restore missing geometry (High on Life 2).</string></property>
+            <property name="text"><string>Support ZMask</string></property>
+           </widget>
+          </item>
+         </layout>
+        </widget>
+       </item>
        <item>
         <spacer name="debugSpacer">
          <property name="orientation"><enum>Qt::Vertical</enum></property>

src/launcher/include/configuration.h

@@ -48,11 +48,14 @@ struct ControllerSettings {
 	QString color;
 	int     speaker_volume      = 50;
 	int     vibration_intensity = 100;
+	// 0 shares player 1 with the first gamepad; 1..MAX_PLAYERS dedicates that player to the keyboard.
+	int     keyboard_player     = 0;
 
 	void WriteSettings(QSettings* s) const {
 		s->setValue("controller_color", color);
 		s->setValue("controller_speaker_volume", speaker_volume);
 		s->setValue("controller_vibration_intensity", vibration_intensity);
+		s->setValue("controller_keyboard_player", keyboard_player);
 	}
 
 	void ReadSettings(QSettings* s) {
@@ -65,6 +68,8 @@ struct ControllerSettings {
 		};
 		speaker_volume      = read_percent("controller_speaker_volume", 50);
 		vibration_intensity = read_percent("controller_vibration_intensity", 100);
+		keyboard_player = qBound(0, s->value("controller_keyboard_player", 0).toInt(),
+		                         Config::MAX_PLAYERS);
 	}
 };
 
@@ -122,9 +127,11 @@ public:
 	bool                   hide_cursor_enabled         = false;
 	bool                   readback_linear_images      = false;
 	bool                   sync_raw_image_buffers      = false;
+	bool                   support_zmask               = false;
 	bool                   tessellation_enabled        = false;
 	bool                   trophy_enabled              = true;
 	int                    trophy_duration_seconds     = 6;
+	bool                   perf_overlay_enabled        = false;
 	bool                   skip_notice_screen          = false;
 	bool                   network_connected           = false;
 	int                    vblank_frequency            = 60;
@@ -160,9 +167,11 @@ public:
 		hide_cursor_enabled         = other.hide_cursor_enabled;
 		readback_linear_images      = other.readback_linear_images;
 		sync_raw_image_buffers      = other.sync_raw_image_buffers;
+		support_zmask               = other.support_zmask;
 		tessellation_enabled        = other.tessellation_enabled;
 		trophy_enabled              = other.trophy_enabled;
 		trophy_duration_seconds     = other.trophy_duration_seconds;
+		perf_overlay_enabled        = other.perf_overlay_enabled;
 		skip_notice_screen          = other.skip_notice_screen;
 		network_connected           = other.network_connected;
 		vblank_frequency            = other.vblank_frequency;

src/launcher/src/configuration.cpp

@@ -14,9 +14,11 @@ QVariantMap Configuration::GameSettings() const {
 	KYTY_CFG_SET(hide_cursor_enabled);
 	KYTY_CFG_SET(readback_linear_images);
 	KYTY_CFG_SET(sync_raw_image_buffers);
+	KYTY_CFG_SET(support_zmask);
 	KYTY_CFG_SET(tessellation_enabled);
 	KYTY_CFG_SET(trophy_enabled);
 	KYTY_CFG_SET(trophy_duration_seconds);
+	KYTY_CFG_SET(perf_overlay_enabled);
 	KYTY_CFG_SET(skip_notice_screen);
 	KYTY_CFG_SET(network_connected);
 	KYTY_CFG_SET(vblank_frequency);
@@ -62,9 +64,11 @@ void Configuration::ReadGameSettingsValues(const Settings& s) {
 	KYTY_CFG_GET(hide_cursor_enabled);
 	KYTY_CFG_GET(readback_linear_images);
 	KYTY_CFG_GET(sync_raw_image_buffers);
+	KYTY_CFG_GET(support_zmask);
 	KYTY_CFG_GET(tessellation_enabled);
 	trophy_enabled   = s.value("trophy_enabled", trophy_enabled).toBool();
 	KYTY_CFG_GET(trophy_duration_seconds);
+	KYTY_CFG_GET(perf_overlay_enabled);
 	KYTY_CFG_GET(skip_notice_screen);
 	KYTY_CFG_GET(network_connected);
 	vblank_frequency = s.value("vblank_frequency", vblank_frequency).toInt();

src/launcher/src/configurationEditDialog.cpp

@@ -209,10 +209,19 @@ static void ListInit(QComboBox* combo, T value) {
 void ConfigurationEditDialog::Init(const Configuration& info) {
 	m_ui->lineEdit_user_name->setMaxLength(static_cast<int>(Config::MAX_USER_NAME_LENGTH));
 	m_ui->lineEdit_user_name->setText(info.user_name);
+	m_ui->spinBox_user_id->setMaximum(Config::MAX_USER_ID);
 	m_ui->spinBox_user_id->setValue(info.user_id);
 	UpdateControllerColorButton(m_ui->button_controller_color, info.controller.color);
 	m_ui->slider_controller_vibration->setValue(info.controller.vibration_intensity);
 	m_ui->slider_controller_volume->setValue(info.controller.speaker_volume);
+	auto* keyboard_player = m_ui->comboBox_keyboard_player;
+	keyboard_player->clear();
+	keyboard_player->addItem(tr("Shared with player 1 (default)"), 0);
+	for (int player = 1; player <= Config::MAX_PLAYERS; player++) {
+		keyboard_player->addItem(tr("Player %1").arg(player), player);
+	}
+	keyboard_player->setCurrentIndex(
+	    qMax(0, keyboard_player->findData(info.controller.keyboard_player)));
 	auto* microphone = m_ui->comboBox_audio_input_device;
 	microphone->clear();
 	microphone->addItem(tr("None"), QString {});
@@ -278,11 +287,13 @@ void ConfigurationEditDialog::Init(const Configuration& info) {
 	m_ui->checkBox_hide_cursor->setChecked(info.hide_cursor_enabled);
 	m_ui->checkBox_readback->setChecked(info.readback_linear_images);
 	m_ui->checkBox_sync_raw_image_buffers->setChecked(info.sync_raw_image_buffers);
+	m_ui->checkBox_support_zmask->setChecked(info.support_zmask);
 	m_ui->checkBox_tessellation->setChecked(info.tessellation_enabled);
 	m_ui->checkBox_skip_notice_screen->setChecked(info.skip_notice_screen);
 	m_ui->checkBox_network_connected->setChecked(info.network_connected);
 	m_ui->checkBox_trophy_notifications->setChecked(info.trophy_enabled);
 	m_ui->spinBox_trophy_duration->setValue(info.trophy_duration_seconds);
+	m_ui->checkBox_perf_overlay->setChecked(info.perf_overlay_enabled);
 	m_ui->spinBox_vblank_frequency->setValue(info.vblank_frequency);
 	m_ui->comboBox_console_language->clear();
 	m_ui->comboBox_console_language->addItems(CONSOLE_LANGUAGE_NAMES);
@@ -458,6 +469,7 @@ static void UpdateInfo(Configuration& info, Ui::ConfigurationEditDialog& ui, boo
 		info.controller.color = ui.button_controller_color->property("controllerColor").toString();
 		info.controller.vibration_intensity = ui.slider_controller_vibration->value();
 		info.controller.speaker_volume      = ui.slider_controller_volume->value();
+		info.controller.keyboard_player     = ui.comboBox_keyboard_player->currentData().toInt();
 	}
 	info.screen_resolution =
 	    TextToEnum<Configuration::Resolution>(ui.comboBox_screen_resolution->currentText());
@@ -469,11 +481,13 @@ static void UpdateInfo(Configuration& info, Ui::ConfigurationEditDialog& ui, boo
 	info.hide_cursor_enabled       = ui.checkBox_hide_cursor->isChecked();
 	info.readback_linear_images    = ui.checkBox_readback->isChecked();
 	info.sync_raw_image_buffers    = ui.checkBox_sync_raw_image_buffers->isChecked();
+	info.support_zmask             = ui.checkBox_support_zmask->isChecked();
 	info.tessellation_enabled      = ui.checkBox_tessellation->isChecked();
 	info.skip_notice_screen        = ui.checkBox_skip_notice_screen->isChecked();
 	info.network_connected         = ui.checkBox_network_connected->isChecked();
 	info.trophy_enabled            = ui.checkBox_trophy_notifications->isChecked();
 	info.trophy_duration_seconds   = ui.spinBox_trophy_duration->value();
+	info.perf_overlay_enabled      = ui.checkBox_perf_overlay->isChecked();
 	info.vblank_frequency          = ui.spinBox_vblank_frequency->value();
 	info.console_language          = ui.comboBox_console_language->currentIndex();
 	info.vulkan_validation_enabled = ui.checkBox_vulkan_validation->isChecked();

src/launcher/src/inputMappingDialog.cpp

@@ -203,7 +203,8 @@ public:
 
 		auto* layout = new QVBoxLayout(this);
 		m_label      = new QLabel(
-		    tr("Press a key or mouse button.\nF1, F7, and F11 are reserved; Esc cancels."), this);
+		    tr("Press a key or mouse button.\nF1, F2, F7, F8, and F11 are reserved; Esc cancels."),
+		    this);
 		m_label->setAlignment(Qt::AlignCenter);
 		layout->addWidget(m_label);
 	}
@@ -219,7 +220,8 @@ protected:
 			reject();
 			return;
 		}
-		if (event->key() == Qt::Key_F1 || event->key() == Qt::Key_F7 ||
+		if (event->key() == Qt::Key_F1 || event->key() == Qt::Key_F2 ||
+		    event->key() == Qt::Key_F7 || event->key() == Qt::Key_F8 ||
 		    event->key() == Qt::Key_F11) {
 			m_label->setText(tr("That key is reserved by the emulator."));
 			return;

src/launcher/src/mainDialog.cpp

@@ -246,6 +246,7 @@ static QStringList CreateEmulatorArgs(const Configuration& info) {
 	}
 	args << "--controller-volume" << QString::number(info.controller.speaker_volume);
 	args << "--controller-vibration" << QString::number(info.controller.vibration_intensity);
+	args << "--keyboard-player" << QString::number(info.controller.keyboard_player);
 	args << "--present-mode" << EnumToText(info.present_mode);
 	if (info.gpu_index >= 0) {
 		args << "--gpu" << QString::number(info.gpu_index);
@@ -261,8 +262,10 @@ static QStringList CreateEmulatorArgs(const Configuration& info) {
 	}
 	args << "--readback-linear-images" << BoolArg(info.readback_linear_images);
 	args << "--sync-raw-image-buffers" << BoolArg(info.sync_raw_image_buffers);
+	args << "--support-zmask" << BoolArg(info.support_zmask);
 	args << "--trophy-notifications" << BoolArg(info.trophy_enabled);
 	args << "--trophy-duration" << QString::number(info.trophy_duration_seconds);
+	args << "--perf-overlay" << BoolArg(info.perf_overlay_enabled);
 	args << "--skip-notice-screen" << BoolArg(info.skip_notice_screen);
 	args << "--network-connected" << BoolArg(info.network_connected);
 	if (info.tessellation_enabled) {

src/libs/audio.cpp

@@ -1,6 +1,7 @@
 #include "libs/audio.h"
 
 #include "common/assert.h"
+#include "common/debugCounters.h"
 #include "common/emulatorConfig.h"
 #include "common/logging/log.h"
 #include "common/threads.h"
@@ -403,6 +404,9 @@ bool Audio::QueueSdlAudio(PortOut* port, const void* data, bool blocking, float
 			queued = SDL_GetAudioStreamQueued(port->stream);
 		}
 		if (queued < static_cast<int>(prepared_size)) {
+			if (port->queue_primed) {
+				Common::DebugCounters::Add(Common::DebugCounters::Counter::AudioUnderruns);
+			}
 			port->queue_primed = false;
 		}
 	}

src/libs/controller.cpp

@@ -18,6 +18,7 @@
 #include <atomic>
 #include <cmath>
 #include <cstring>
+#include <memory>
 #include <vector>
 
 namespace Libs::Controller {
@@ -107,18 +108,19 @@ struct ControllerState {
 
 class GameController {
 public:
-	GameController() = default;
+	explicit GameController(int player): m_player(player) {}
 
 	KYTY_CLASS_NO_COPY(GameController);
 
+	[[nodiscard]] int Handle() const { return PAD_HANDLE + m_player; }
+
 	int InitPort();
 	int OpenPort(int user_id, int type, int index);
 	int GetPortHandle(int user_id, int type, int index);
 	int ClosePort(int handle);
 	int CheckPortHandle(int handle);
 
-	void Connect(int id);
-	void Disconnect(int id);
+	void SetDevices(int active_id);
 	void Button(int id, uint32_t button, bool down);
 	void Axis(int id, Axis axis, int value);
 	void RightStick(int id, int x, int y);
@@ -128,7 +130,6 @@ public:
 	void ResetOrientation();
 	void ToggleGyroEmulation(int id);
 	void ResetInputState();
-	void ReleaseHostPads();
 	void GetConnectionInfo(bool* flag, int* count);
 	void SetVibration(uint8_t large_motor, uint8_t small_motor);
 	int  GetActiveControllerId();
@@ -145,8 +146,10 @@ private:
 	enum class PortState { Uninitialized, Closed, Open };
 
 	int CheckPortHandleLocked(int handle) const;
+	[[nodiscard]] bool PadOpenArgsAreValid(int user_id, int type, int index) const;
+
+	const int m_player;
 
-	void CheckActive();
 	void AddState();
 	// Emulated tilt: the right stick drives pitch and roll instead of the stick axes.
 	void UpdateEmulatedTilt();
@@ -157,7 +160,6 @@ private:
 
 	Common::Mutex    m_mutex;
 	PortState        m_port_state      = PortState::Uninitialized;
-	std::vector<int> m_connected_ids;
 	int              m_active_id       = -1;
 	bool             m_connected       = false;
 	int              m_connected_count = 0;
@@ -184,7 +186,39 @@ private:
 	std::array<std::atomic<uint32_t>, 3> m_setting_steps {};
 };
 
-static GameController* g_controller = nullptr;
+static std::array<std::unique_ptr<GameController>, Config::MAX_PLAYERS> g_players;
+
+// Maps host gamepads to players; the keyboard follows the configured keyboard player.
+static Common::Mutex g_router_mutex;
+static std::vector<std::pair<int, int>> g_pad_players;
+
+static uint32_t g_keyboard_player = 0;
+
+static void RefreshPlayers();
+static void StopHostPad(SDL_Gamepad* pad);
+
+static int KeyboardPlayer() {
+	return g_keyboard_player == 0 ? 0 : static_cast<int>(g_keyboard_player) - 1;
+}
+
+// Caller holds g_router_mutex.
+static GameController* PlayerForDevice(int id) {
+	if (id == HOST_INPUT_CONTROLLER_ID) {
+		return g_players[KeyboardPlayer()].get();
+	}
+	for (const auto& [pad, player]: g_pad_players) {
+		if (pad == id) {
+			return player >= 0 ? g_players[player].get() : nullptr;
+		}
+	}
+	return nullptr;
+}
+
+static GameController* PlayerForHandle(int handle) {
+	return handle >= PAD_HANDLE && handle < PAD_HANDLE + Config::MAX_PLAYERS
+	           ? g_players[handle - PAD_HANDLE].get()
+	           : nullptr;
+}
 
 static void pad_fill_data(PadData* data, const ControllerState& state, bool connected,
                           int connected_count) {
@@ -228,13 +262,15 @@ constexpr std::array SPEAKER_VOLUME = {1.0f, 0.0f, 0.02f, 0.12f, 0.42f};
 constexpr std::array INTENSITY      = {1.0f, 0.0f, 0.33f, 0.66f};
 
 void CycleSetting(Setting setting) {
-	if (g_controller != nullptr) {
-		g_controller->CycleSetting(setting);
+	for (const auto& player: g_players) {
+		if (player != nullptr) {
+			player->CycleSetting(setting);
+		}
 	}
 }
 
 float GetSettingScale(Setting setting) {
-	return g_controller != nullptr ? g_controller->GetSettingScale(setting) : 1.0f;
+	return g_players[0] != nullptr ? g_players[0]->GetSettingScale(setting) : 1.0f;
 }
 
 void GameController::CycleSetting(Setting setting) {
@@ -404,35 +440,53 @@ static int32_t TriggerEffectState(const PadTriggerEffectCommand& command, int va
 }
 
 void Initialize() {
-	EXIT_IF(g_controller != nullptr);
+	EXIT_IF(g_players[0] != nullptr);
 
-	g_controller = new GameController;
-	g_controller->Connect(HOST_INPUT_CONTROLLER_ID);
+	for (int i = 0; i < Config::MAX_PLAYERS; i++) {
+		g_players[i] = std::make_unique<GameController>(i);
+	}
+	g_keyboard_player = Config::GetKeyboardPlayer();
+	RefreshPlayers();
 }
 
 void Shutdown() {
 	EmergencyShutdown();
-	delete g_controller;
-	g_controller = nullptr;
+	for (auto& player: g_players) {
+		player.reset();
+	}
 }
 
 void EmergencyShutdown() {
-	if (g_controller != nullptr) {
-		g_controller->ReleaseHostPads();
+	Common::LockGuard lock(g_router_mutex);
+	DualSenseHaptics::Shutdown();
+	std::vector<SDL_Gamepad*> pads;
+	for (const auto& [id, player]: g_pad_players) {
+		if (auto* pad = SDL_GetGamepadFromID(static_cast<SDL_JoystickID>(id))) {
+			StopHostPad(pad);
+			pads.push_back(pad);
+		}
+	}
+	if (!pads.empty()) {
+		SDL_Delay(RELEASE_FLUSH_MS);
+		for (auto* pad: pads) {
+			SDL_CloseGamepad(pad);
+		}
+	}
+	g_pad_players.clear();
+	for (const auto& player: g_players) {
+		if (player != nullptr) {
+			player->SetDevices(-1);
+		}
 	}
 }
 
-void GameController::Connect(int id) {
+void GameController::SetDevices(int active_id) {
 	Common::LockGuard lock(m_mutex);
-
-	if (std::find(m_connected_ids.begin(), m_connected_ids.end(), id) != m_connected_ids.end()) {
+	if (m_active_id == active_id) {
 		return;
 	}
-
-	m_connected_ids.push_back(id);
-
-	if (id != HOST_INPUT_CONTROLLER_ID) {
-		if (auto* pad = SDL_GetGamepadFromID(static_cast<SDL_JoystickID>(id));
+	if (active_id != HOST_INPUT_CONTROLLER_ID && active_id != -1) {
+		if (auto* pad = SDL_GetGamepadFromID(static_cast<SDL_JoystickID>(active_id));
 		    pad != nullptr) {
 			if (const auto& color = Config::GetControllerColor()) {
 				(void)SDL_SetGamepadLED(pad, (*color)[0], (*color)[1], (*color)[2]);
@@ -446,42 +500,11 @@ void GameController::Connect(int id) {
 		}
 	}
 
-	CheckActive();
-}
-
-void GameController::Disconnect(int id) {
-	Common::LockGuard lock(m_mutex);
-
-	const auto it = std::find(m_connected_ids.begin(), m_connected_ids.end(), id);
-	EXIT_IF(it == m_connected_ids.end());
-
-	m_connected_ids.erase(it);
-
-	CheckActive();
-}
-
-void GameController::CheckActive() {
-	int  new_active_id = -1;
-	bool new_connected = false;
-
-	if (!m_connected_ids.empty()) {
-		new_active_id = m_connected_ids[0];
-		for (const auto id: m_connected_ids) {
-			if (id != HOST_INPUT_CONTROLLER_ID) {
-				new_active_id = id;
-				break;
-			}
-		}
-		new_connected = true;
-	}
-
-	if (m_connected == new_connected && m_active_id == new_active_id) {
-		return;
-	}
+	const bool new_connected = active_id != -1;
 	if (!m_connected && new_connected) {
 		m_connected_count++;
 	}
-	m_active_id     = new_active_id;
+	m_active_id     = active_id;
 	m_connected     = new_connected;
 	m_state         = {};
 	m_gyro_time     = 0;
@@ -513,7 +536,7 @@ void GameController::AddState() {
 void GameController::Button(int id, uint32_t button, bool down) {
 	Common::LockGuard lock(m_mutex);
 
-	// The keyboard shares the player-1 pad with the active gamepad.
+	// Keyboard events are routed to the configured player's port.
 	if (m_active_id == id || id == HOST_INPUT_CONTROLLER_ID) {
 		m_state.time = LibKernel::KernelGetProcessTime();
 
@@ -798,39 +821,16 @@ void GameController::ResetInputState() {
 	AddState();
 }
 
-void GameController::ReleaseHostPads() {
-	Common::LockGuard lock(m_mutex);
-	DualSenseHaptics::Shutdown();
-
-	std::vector<SDL_Gamepad*> pads;
-	for (const auto id: m_connected_ids) {
-		if (id == HOST_INPUT_CONTROLLER_ID) {
-			continue;
-		}
-		if (auto* pad = SDL_GetGamepadFromID(static_cast<SDL_JoystickID>(id));
-		    pad != nullptr) {
-			if (SDL_GetGamepadType(pad) == SDL_GAMEPAD_TYPE_PS5) {
-				DualSenseEffects effect {};
-				effect.enable_bits     = 0x0c;
-				effect.right_trigger[0] = 0x05;
-				effect.left_trigger[0]  = 0x05;
-				(void)SDL_SendGamepadEffect(pad, &effect, sizeof(effect));
-			}
-			(void)SDL_RumbleGamepad(pad, 0, 0, 0);
-			(void)SDL_SetGamepadLED(pad, 0, 0, 0);
-			pads.push_back(pad);
-		}
+static void StopHostPad(SDL_Gamepad* pad) {
+	if (SDL_GetGamepadType(pad) == SDL_GAMEPAD_TYPE_PS5) {
+		DualSenseEffects effect {};
+		effect.enable_bits     = 0x0c;
+		effect.right_trigger[0] = 0x05;
+		effect.left_trigger[0]  = 0x05;
+		(void)SDL_SendGamepadEffect(pad, &effect, sizeof(effect));
 	}
-
-	if (!pads.empty()) {
-		SDL_Delay(RELEASE_FLUSH_MS);
-		for (auto* pad: pads) {
-			SDL_CloseGamepad(pad);
-		}
-	}
-
-	m_connected_ids.clear();
-	CheckActive();
+	(void)SDL_RumbleGamepad(pad, 0, 0, 0);
+	(void)SDL_SetGamepadLED(pad, 0, 0, 0);
 }
 
 void GameController::SetVibration(uint8_t large_motor, uint8_t small_motor) {
@@ -847,7 +847,9 @@ void GameController::ApplyVibration() {
 	const auto scale       = duration != 0 ? GetSettingScale(Setting::VibrationIntensity) : 0.0f;
 	const auto large_motor = Scale(m_vibration[0], scale);
 	const auto small_motor = Scale(m_vibration[1], scale);
-	if (DualSenseHaptics::SetVibration(m_active_id, large_motor, small_motor, duration)) {
+	// The DualSense haptics backend drives a single pad, so only player 1 uses it.
+	if (m_player == 0 &&
+	    DualSenseHaptics::SetVibration(m_active_id, large_motor, small_motor, duration)) {
 		return;
 	}
 
@@ -999,49 +1001,149 @@ int GameController::ReadStates(ControllerState* states, int states_num, bool* fl
 	return ret_num;
 }
 
+// Keep every opened pad, including those waiting for a free player slot.
+// Caller holds g_router_mutex.
+static void RefreshPlayers() {
+	std::array<bool, Config::MAX_PLAYERS> taken {};
+	const int keyboard = KeyboardPlayer();
+	if (g_keyboard_player != 0) {
+		taken[keyboard] = true;
+	}
+	for (auto& [id, player]: g_pad_players) {
+		if (g_keyboard_player != 0 && player == keyboard) {
+			player = -1;
+		}
+		if (player >= 0) {
+			taken[player] = true;
+		}
+	}
+	for (auto& [id, player]: g_pad_players) {
+		if (player >= 0) {
+			continue;
+		}
+		const auto free = std::find(taken.begin(), taken.end(), false);
+		if (free != taken.end()) {
+			player = static_cast<int>(free - taken.begin());
+			*free  = true;
+			LOGF("Controller %d assigned to player %d\n", id, player + 1);
+		}
+	}
+	std::array<int, Config::MAX_PLAYERS> active;
+	active.fill(-1);
+	active[keyboard] = HOST_INPUT_CONTROLLER_ID;
+	for (const auto& [id, player]: g_pad_players) {
+		if (player >= 0) {
+			active[player] = id;
+		}
+	}
+	// Stop old owners before configuring new owners of the same physical pad.
+	for (int player = 0; player < Config::MAX_PLAYERS; player++) {
+		const int previous = g_players[player]->GetActiveControllerId();
+		if (previous != active[player]) {
+			if (player == 0) {
+				(void)DualSenseHaptics::SetVibration(previous, 0, 0, 0);
+			}
+			if (auto* pad = SDL_GetGamepadFromID(static_cast<SDL_JoystickID>(previous))) {
+				StopHostPad(pad);
+			}
+		}
+	}
+	for (int player = 0; player < Config::MAX_PLAYERS; player++) {
+		g_players[player]->SetDevices(active[player]);
+	}
+}
+
 void Connect(int id) {
-	g_controller->Connect(id);
+	Common::LockGuard lock(g_router_mutex);
+	if (std::any_of(g_pad_players.begin(), g_pad_players.end(),
+	                [id](const auto& pad) { return pad.first == id; })) {
+		return;
+	}
+	g_pad_players.emplace_back(id, -1);
+	RefreshPlayers();
 }
 
 void Disconnect(int id) {
-	g_controller->Disconnect(id);
+	Common::LockGuard lock(g_router_mutex);
+	std::erase_if(g_pad_players, [id](const auto& pad) { return pad.first == id; });
+	RefreshPlayers();
+}
+
+void CycleKeyboardPlayer() {
+	Common::LockGuard lock(g_router_mutex);
+	g_keyboard_player = (g_keyboard_player + 1) % (Config::MAX_PLAYERS + 1);
+	RefreshPlayers();
+	if (g_keyboard_player == 0) {
+		LOGF("Keyboard shared with player 1\n");
+	} else {
+		LOGF("Keyboard assigned to player %u\n", g_keyboard_player);
+	}
 }
 
 void SetButton(int id, uint32_t button, bool down) {
-	g_controller->Button(id, button, down);
+	Common::LockGuard lock(g_router_mutex);
+	if (auto* player = PlayerForDevice(id)) {
+		player->Button(id, button, down);
+	}
 }
 
 void SetAxis(int id, Axis axis, int value) {
-	g_controller->Axis(id, axis, value);
+	Common::LockGuard lock(g_router_mutex);
+	if (auto* player = PlayerForDevice(id)) {
+		player->Axis(id, axis, value);
+	}
 }
 
 void SetRightStick(int id, int x, int y) {
-	g_controller->RightStick(id, x, y);
+	Common::LockGuard lock(g_router_mutex);
+	if (auto* player = PlayerForDevice(id)) {
+		player->RightStick(id, x, y);
+	}
 }
 
 void SetTouchPad(int id, int finger, bool down, float x, float y) {
-	g_controller->TouchPad(id, finger, down, x, y);
+	Common::LockGuard lock(g_router_mutex);
+	if (auto* player = PlayerForDevice(id)) {
+		player->TouchPad(id, finger, down, x, y);
+	}
 }
 
 void SetSensor(int id, Sensor sensor, const float* data, uint64_t time_us) {
-	g_controller->Motion(id, sensor, data, time_us);
+	Common::LockGuard lock(g_router_mutex);
+	if (auto* player = PlayerForDevice(id)) {
+		player->Motion(id, sensor, data, time_us);
+	}
 }
 
 void ToggleGyroEmulation(int id) {
-	g_controller->ToggleGyroEmulation(id);
+	Common::LockGuard lock(g_router_mutex);
+	if (auto* player = PlayerForDevice(id)) {
+		player->ToggleGyroEmulation(id);
+	}
 }
 
 void ResetInputState() {
-	g_controller->ResetInputState();
+	for (const auto& player: g_players) {
+		if (player != nullptr) {
+			player->ResetInputState();
+		}
+	}
 }
 
 int GetActiveControllerId() {
-	return g_controller != nullptr ? g_controller->GetActiveControllerId() : -1;
+	return g_players[0] != nullptr ? g_players[0]->GetActiveControllerId() : -1;
 }
 
-static bool PadOpenArgsAreValid(int user_id, int type, int index) {
-	// The host keyboard and gamepad feed one standard port for the configured user.
-	return user_id == Config::GetUserId() && type == 0 && index == 0;
+bool IsPlayerConnected(int player) {
+	if (player < 0 || player >= Config::MAX_PLAYERS || g_players[player] == nullptr) {
+		return false;
+	}
+	return g_players[player]->GetActiveControllerId() != -1;
+}
+
+bool GameController::PadOpenArgsAreValid(int user_id, int type, int index) const {
+	// The host keyboard and gamepads of a player feed one standard port for that player's user.
+	return user_id == Config::GetPlayerUserId(m_player) && type == 0 && index == 0;
 }
 
 int GameController::InitPort() {
@@ -1064,7 +1166,7 @@ int GameController::OpenPort(int user_id, int type, int index) {
 		return PAD_ERROR_ALREADY_OPENED;
 	}
 	m_port_state = PortState::Open;
-	return PAD_HANDLE;
+	return Handle();
 }
 
 int GameController::GetPortHandle(int user_id, int type, int index) {
@@ -1072,15 +1174,16 @@ int GameController::GetPortHandle(int user_id, int type, int index) {
 	if (m_port_state == PortState::Uninitialized) {
 		return PAD_ERROR_NOT_INITIALIZED;
 	}
-	return m_port_state == PortState::Open && PadOpenArgsAreValid(user_id, type, index) ? PAD_HANDLE
-	                                                                               : PAD_ERROR_NO_HANDLE;
+	return m_port_state == PortState::Open && PadOpenArgsAreValid(user_id, type, index)
+	           ? Handle()
+	           : PAD_ERROR_NO_HANDLE;
 }
 
 int GameController::CheckPortHandleLocked(int handle) const {
 	if (m_port_state == PortState::Uninitialized) {
 		return PAD_ERROR_NOT_INITIALIZED;
 	}
-	return m_port_state == PortState::Open && handle == PAD_HANDLE ? OK : PAD_ERROR_INVALID_HANDLE;
+	return m_port_state == PortState::Open && handle == Handle() ? OK : PAD_ERROR_INVALID_HANDLE;
 }
 
 int GameController::CheckPortHandle(int handle) {
@@ -1098,12 +1201,19 @@ int GameController::ClosePort(int handle) {
 }
 
 int PadCheckHandle(int handle) {
-	return g_controller->CheckPortHandle(handle);
+	if (g_players[0] == nullptr) {
+		return PAD_ERROR_NOT_INITIALIZED;
+	}
+	auto* player = PlayerForHandle(handle);
+	return player != nullptr ? player->CheckPortHandle(handle) : PAD_ERROR_INVALID_HANDLE;
 }
 
 int KYTY_SYSV_ABI PadInit() {
 	PRINT_NAME();
-	return g_controller->InitPort();
+	for (const auto& player: g_players) {
+		(void)player->InitPort();
+	}
+	return OK;
 }
 
 int KYTY_SYSV_ABI PadOpen(int user_id, int type, int index, const void* param) {
@@ -1115,7 +1225,8 @@ int KYTY_SYSV_ABI PadOpen(int user_id, int type, int index, const void* param) {
 	     "\t param   = 0x%016" PRIx64 "\n",
 	     user_id, type, index, reinterpret_cast<uint64_t>(param));
 
-	return g_controller->OpenPort(user_id, type, index);
+	const int player = Config::GetPlayerFromUserId(user_id);
+	return player >= 0 ? g_players[player]->OpenPort(user_id, type, index) : PAD_ERROR_INVALID_ARG;
 }
 
 int KYTY_SYSV_ABI PadGetHandle(int user_id, int type, int index) {
@@ -1126,13 +1237,19 @@ int KYTY_SYSV_ABI PadGetHandle(int user_id, int type, int index) {
 	     "\t index   = %d\n",
 	     user_id, type, index);
 
-	return g_controller->GetPortHandle(user_id, type, index);
+	const int player = Config::GetPlayerFromUserId(user_id);
+	return player >= 0 ? g_players[player]->GetPortHandle(user_id, type, index)
+	                   : PAD_ERROR_NO_HANDLE;
 }
 
 int KYTY_SYSV_ABI PadClose(int handle) {
 	PRINT_NAME();
 	LOGF("\t handle = %d\n", handle);
-	return g_controller->ClosePort(handle);
+	auto* player = PlayerForHandle(handle);
+	if (player == nullptr) {
+		return PAD_ERROR_INVALID_HANDLE;
+	}
+	return player->ClosePort(handle);
 }
 
 int KYTY_SYSV_ABI PadSetMotionSensorState(int handle, bool enable) {
@@ -1143,7 +1260,7 @@ int KYTY_SYSV_ABI PadSetMotionSensorState(int handle, bool enable) {
 	}
 
 	LOGF("\t enable = %s\n", (enable ? "true" : "false"));
-	g_controller->SetMotionSensorState(enable);
+	PlayerForHandle(handle)->SetMotionSensorState(enable);
 
 	return OK;
 }
@@ -1167,7 +1284,7 @@ int KYTY_SYSV_ABI PadResetOrientation(int handle) {
 		return error;
 	}
 
-	g_controller->ResetOrientation();
+	PlayerForHandle(handle)->ResetOrientation();
 	return OK;
 }
 
@@ -1183,7 +1300,7 @@ int KYTY_SYSV_ABI PadGetControllerInformation(int handle, PadControllerInformati
 
 	int  connected_count = 0;
 	bool connected       = false;
-	g_controller->GetConnectionInfo(&connected, &connected_count);
+	PlayerForHandle(handle)->GetConnectionInfo(&connected, &connected_count);
 
 	std::memset(info, 0, sizeof(*info));
 
@@ -1228,7 +1345,7 @@ int KYTY_SYSV_ABI PadReadState(int handle, PadData* data) {
 	bool            connected       = false;
 	ControllerState state;
 
-	g_controller->ReadState(&state, &connected, &connected_count);
+	PlayerForHandle(handle)->ReadState(&state, &connected, &connected_count);
 
 	pad_fill_data(data, state, connected, connected_count);
 
@@ -1252,11 +1369,12 @@ int KYTY_SYSV_ABI PadRead(int handle, PadData* data, int num) {
 	bool            connected       = false;
 	ControllerState states[64]      = {};
 
-	int ret_num = g_controller->ReadStates(states, num, &connected, &connected_count);
+	auto* player  = PlayerForHandle(handle);
+	int   ret_num = player->ReadStates(states, num, &connected, &connected_count);
 
 	if (!connected || ret_num == 0) {
 		if (connected) {
-			g_controller->ReadState(&states[0], &connected, &connected_count);
+			player->ReadState(&states[0], &connected, &connected_count);
 		}
 		ret_num = 1;
 	}
@@ -1270,6 +1388,7 @@ int KYTY_SYSV_ABI PadRead(int handle, PadData* data, int num) {
 
 int KYTY_SYSV_ABI PadSetVibration(int handle, const PadVibrationParam* param) {
 	PRINT_NAME();
+	Common::LockGuard lock(g_router_mutex);
 
 	if (const auto error = PadCheckHandle(handle); error != OK) {
 		return error;
@@ -1282,7 +1401,7 @@ int KYTY_SYSV_ABI PadSetVibration(int handle, const PadVibrationParam* param) {
 	     "\t small_motor = %d\n",
 	     static_cast<int>(param->large_motor), static_cast<int>(param->small_motor));
 
-	g_controller->SetVibration(param->large_motor, param->small_motor);
+	PlayerForHandle(handle)->SetVibration(param->large_motor, param->small_motor);
 
 	return OK;
 }
@@ -1299,6 +1418,7 @@ int KYTY_SYSV_ABI PadResetLightBar(int handle) {
 
 int KYTY_SYSV_ABI PadSetLightBar(int handle, const PadLightBarParam* param) {
 	PRINT_NAME();
+	Common::LockGuard lock(g_router_mutex);
 
 	if (const auto error = PadCheckHandle(handle); error != OK) {
 		return error;
@@ -1307,13 +1427,14 @@ int KYTY_SYSV_ABI PadSetLightBar(int handle, const PadLightBarParam* param) {
 		return PAD_ERROR_INVALID_ARG;
 	}
 
-	g_controller->SetLightBar(param->r, param->g, param->b);
+	PlayerForHandle(handle)->SetLightBar(param->r, param->g, param->b);
 
 	return OK;
 }
 
 int KYTY_SYSV_ABI PadSetTriggerEffect(int handle, const PadTriggerEffectParam* param) {
 	PRINT_NAME();
+	Common::LockGuard lock(g_router_mutex);
 
 	if (const auto error = PadCheckHandle(handle); error != OK) {
 		return error;
@@ -1322,7 +1443,7 @@ int KYTY_SYSV_ABI PadSetTriggerEffect(int handle, const PadTriggerEffectParam* p
 		return PAD_ERROR_INVALID_ARG;
 	}
 
-	return g_controller->SetTriggerEffect(*param) ? OK : PAD_ERROR_INVALID_ARG;
+	return PlayerForHandle(handle)->SetTriggerEffect(*param) ? OK : PAD_ERROR_INVALID_ARG;
 }
 
 int KYTY_SYSV_ABI PadGetTriggerEffectState(int handle, PadTriggerEffectStateInformation* info) {
@@ -1337,7 +1458,7 @@ int KYTY_SYSV_ABI PadGetTriggerEffectState(int handle, PadTriggerEffectStateInfo
 		return PAD_ERROR_INVALID_ARG;
 	}
 
-	g_controller->GetTriggerEffectState(info);
+	PlayerForHandle(handle)->GetTriggerEffectState(info);
 
 	return OK;
 }

src/libs/controller.h

@@ -68,6 +68,8 @@ inline int controller_get_axis(int min, int max, int value) {
 
 void Connect(int id);
 void Disconnect(int id);
+// Moves the keyboard to the next player (shared with player 1, then 1-4) while the game runs.
+void CycleKeyboardPlayer();
 void SetButton(int id, uint32_t button, bool down);
 void SetAxis(int id, Axis axis, int value);
 void SetRightStick(int id, int x, int y);
@@ -76,6 +78,8 @@ void SetSensor(int id, Sensor sensor, const float* data, uint64_t time_us);
 void ToggleGyroEmulation(int id);
 void ResetInputState();
 int  GetActiveControllerId();
+// True when the player has a keyboard or gamepad assigned.
+bool IsPlayerConnected(int player);
 
 enum class Setting { SpeakerVolume, VibrationIntensity, TriggerEffectIntensity };
 

src/libs/libAmpr.cpp

@@ -25,6 +25,9 @@
 #include <string>
 #include <thread>
 #include <unordered_map>
+#include <atomic>
+#include <cinttypes>
+#include <iterator>
 #include <variant>
 #include <vector>
 
@@ -176,6 +179,14 @@ static int GetHostPathStat(const std::filesystem::path& host_path, LibKernel::Fi
 	return OK;
 }
 
+// Mounts can appear after a failed lookup (e.g. additional content), so forget negative results.
+void InvalidateFailedResolves() {
+	std::scoped_lock lock(g_mutex);
+	for (auto it = g_resolved_paths.begin(); it != g_resolved_paths.end();) {
+		it = it->second.result != OK ? g_resolved_paths.erase(it) : std::next(it);
+	}
+}
+
 static int ResolveOnePath(const char* guest_path, uint32_t* id, uint64_t* size) {
 	if (guest_path == nullptr || guest_path[0] == '\0') {
 		return LibKernel::KERNEL_ERROR_EINVAL;
@@ -225,6 +236,10 @@ static int ResolveOnePath(const char* guest_path, uint32_t* id, uint64_t* size)
 				log_missing = true;
 			}
 		}
+		if (!log_missing && info.result == OK && path.starts_with("/addcont")) {
+			LOGF("\tAPR resolve: %s -> %s (%" PRIu64 " bytes)\n", guest_path,
+			     Common::PathToString(info.host_path).c_str(), info.file_size);
+		}
 		if (log_missing) {
 			LOGF("\tAPR resolve missing path: %s -> %s\n", guest_path, Common::PathToString(info.host_path).c_str());
 		}
@@ -1181,6 +1196,13 @@ static int ExecuteCommand(const CommandBufferState::Command& entry) {
 		uint64_t bytes_read = 0;
 		auto     result = ReadHostFileToGuest(host_path, command.file_offset, command.destination,
 		                                      command.size, &bytes_read);
+		static std::atomic<uint32_t> s_logged_dlc_reads {0};
+		if (host_path.filename() == "re_dlc_000.pak" && s_logged_dlc_reads.fetch_add(1) < 32) {
+			LOGF("\tAPR read dlc: id=0x%08" PRIx32 ", offset=0x%016" PRIx64 ", size=0x%016" PRIx64
+			     ", read=0x%016" PRIx64 ", result=0x%08" PRIx32 "\n",
+			     command.file_id, command.file_offset, command.size, bytes_read,
+			     static_cast<uint32_t>(result));
+		}
 		if (result != OK) {
 			LOGF("\tAPR submit read failed: id=0x%08" PRIx32 ", result=0x%08" PRIx32 ", path=%s\n",
 			     command.file_id, static_cast<uint32_t>(result), Common::PathToString(host_path).c_str());

src/libs/libAppContent.cpp

@@ -22,13 +22,16 @@
 
 namespace Libs {
 
+namespace AprShared {
+void InvalidateFailedResolves();
+}
+
 LIB_VERSION("AppContent", 1, "AppContentUtil", 1, 1);
 
 namespace AppContent {
 
 constexpr uint32_t APP_CONTENT_APPPARAM_SKU_FLAG_FULL = 3;
 constexpr int      APP_CONTENT_ERROR_PARAMETER        = -2133262334; /* 0x80D90002 */
-constexpr int      APP_CONTENT_ERROR_BUSY             = -2133262333; /* 0x80D90003 */
 constexpr int      APP_CONTENT_ERROR_NOT_MOUNTED      = -2133262332; /* 0x80D90004 */
 constexpr int      APP_CONTENT_ERROR_NOT_FOUND        = -2133262331; /* 0x80D90005 */
 constexpr int      APP_CONTENT_ERROR_MOUNT_FULL       = -2133262330; /* 0x80D90006 */
@@ -207,7 +210,10 @@ int KYTY_SYSV_ABI AppContentAddcontMount(uint32_t                         servic
 
 	for (const auto& m: g_addcont_mounts) {
 		if (m.label == entry->entitlement_label) {
-			return APP_CONTENT_ERROR_BUSY;
+			const auto guest_point = m.point.substr(0, m.point.size() - 1);
+			std::strncpy(mount_point->data, guest_point.c_str(), sizeof(mount_point->data) - 1);
+			LOGF("\t already mounted %s\n", guest_point.c_str());
+			return OK;
 		}
 	}
 
@@ -232,7 +238,15 @@ int KYTY_SYSV_ABI AppContentAddcontMount(uint32_t                         servic
 	LOGF("\t mounted %s -> %s\n", mount.point.c_str(),
 	     Common::PathToString(entry->mount_dir).c_str());
 
-	std::strncpy(mount_point->data, mount.point.c_str(), sizeof(mount_point->data) - 1);
+	AprShared::InvalidateFailedResolves();
+
+	// Titles scan /download0/pdlc for downloaded content; it must exist even when empty.
+	if (const auto pdlc = LibKernel::FileSystem::GetRealFilename("/download0/pdlc"); !pdlc.empty()) {
+		std::filesystem::create_directories(pdlc, ec);
+	}
+
+	const auto guest_point = mount.point.substr(0, mount.point.size() - 1);
+	std::strncpy(mount_point->data, guest_point.c_str(), sizeof(mount_point->data) - 1);
 	g_addcont_mounts.push_back(std::move(mount));
 
 	return OK;
@@ -251,8 +265,8 @@ int KYTY_SYSV_ABI AppContentAddcontUnmount(const AppContentMountPoint* mount_poi
 		return APP_CONTENT_ERROR_NOT_MOUNTED;
 	}
 
-	const std::string point(mount_point->data,
-	                        strnlen(mount_point->data, sizeof(mount_point->data)));
+	const std::string point = Common::FixDirectorySlash(
+	    std::string_view(mount_point->data, strnlen(mount_point->data, sizeof(mount_point->data))));
 
 	std::lock_guard lock(g_addcont_mutex);
 
@@ -263,6 +277,7 @@ int KYTY_SYSV_ABI AppContentAddcontUnmount(const AppContentMountPoint* mount_poi
 	}
 	LibKernel::FileSystem::Umount(point);
 	g_addcont_mounts.erase(it);
+	AprShared::InvalidateFailedResolves();
 
 	return OK;
 }

src/libs/libKernel.cpp

@@ -11,6 +11,7 @@
 #include "kernel/eventQueue.h"
 #include "kernel/fileSystem.h"
 #include "kernel/memory.h"
+#include "kernel/module.h"
 #include "kernel/pthread.h"
 #include "kernel/semaphore.h"
 #include "kernel/syncOnAddress.h"
@@ -65,7 +66,6 @@ LIB_DEFINE(InitLibKernel_1_Apr);
 
 namespace LibKernel {
 
-using KernelModule                   = int32_t;
 using get_thread_atexit_count_func_t = KYTY_SYSV_ABI int (*)(KernelModule);
 using thread_atexit_report_func_t    = KYTY_SYSV_ABI void (*)(KernelModule);
 
@@ -1201,9 +1201,9 @@ static int KYTY_SYSV_ABI KernelRaiseException(Pthread thread, int signum) {
 	return OK;
 }
 
-static KYTY_SYSV_ABI KernelModule KernelLoadStartModule(const char* module_file_name, size_t args,
-                                                        const void* argp, uint32_t flags,
-                                                        const KernelLoadModuleOpt* opt, int* res) {
+KernelModule KYTY_SYSV_ABI KernelLoadStartModule(const char* module_file_name, size_t args,
+                                                 const void* argp, uint32_t flags,
+                                                 const KernelLoadModuleOpt* opt, int* res) {
 	PRINT_NAME();
 
 	// EXIT_NOT_IMPLEMENTED(!Common::Thread::IsMainThread());
@@ -1255,9 +1255,9 @@ static KYTY_SYSV_ABI KernelModule KernelLoadStartModule(const char* module_file_
 	return static_cast<KernelModule>(handle);
 }
 
-static int KYTY_SYSV_ABI KernelStopUnloadModule(KernelModule handle, size_t args, const void* argp,
-                                                uint32_t flags, const KernelUnloadModuleOpt* opt,
-                                                int* res) {
+int KYTY_SYSV_ABI KernelStopUnloadModule(KernelModule handle, size_t args, const void* argp,
+                                        uint32_t flags, const KernelUnloadModuleOpt* opt,
+                                        int* res) {
 	PRINT_NAME();
 
 	// EXIT_NOT_IMPLEMENTED(!Common::Thread::IsMainThread());
@@ -2006,6 +2006,11 @@ int KYTY_SYSV_ABI listen(int s, int backlog) {
 	return Network::Net::Listen(s, backlog);
 }
 
+int KYTY_SYSV_ABI shutdown(int s, int how) {
+	PRINT_NAME();
+	return Network::Net::Shutdown(s, how);
+}
+
 int KYTY_SYSV_ABI accept(int s, void* addr, uint32_t* addrlen) {
 	PRINT_NAME();
 	return Network::Net::Accept(s, addr, addrlen);
@@ -2242,6 +2247,7 @@ LIB_DEFINE(InitLibKernel_1_Posix) {
 	LIB_FUNC("KuOmgKoqCdY", Posix::bind);
 	LIB_FUNC("XVL8So3QJUk", Posix::connect);
 	LIB_FUNC("pxnCmagrtao", Posix::listen);
+	LIB_FUNC("TUuiYS2kE8s", Posix::shutdown);
 	LIB_FUNC("3e+4Iv7IJ8U", Posix::accept);
 	LIB_FUNC("RenI1lL1WFk", Posix::getsockname);
 	LIB_FUNC("6O8EwYOgH9Y", Posix::getsockopt);

src/libs/libPlayGo.cpp

@@ -221,11 +221,17 @@ int KYTY_SYSV_ABI PlayGoGetChunkId(int handle, uint16_t* out_chunk_id_list,
 	if (out_entries == nullptr) {
 		return PLAYGO_ERROR_BAD_POINTER;
 	}
-	if (number_of_entries != 0 && out_chunk_id_list == nullptr) {
-		return PLAYGO_ERROR_BAD_POINTER;
+	if (out_chunk_id_list != nullptr && number_of_entries == 0) {
+		return PLAYGO_ERROR_BAD_SIZE;
 	}
 	ensure_chunks_loaded();
 
+	// A null output buffer asks for the total chunk count.
+	if (out_chunk_id_list == nullptr) {
+		*out_entries = g_chunks_num;
+		return OK;
+	}
+
 	const uint32_t entries = std::min(number_of_entries, g_chunks_num);
 	for (uint32_t i = 0; i < entries; i++) {
 		out_chunk_id_list[i] = static_cast<uint16_t>(i);