gpu: preserve finite buffer descriptors across sparse PRT holes

Avoid readbacks for compute metadata fills

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;
 }

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;
@@ -80,6 +89,7 @@ struct ConfigOptions {
 	bool                   tessellation_enabled        = false;
 	bool                   trophy_enabled              = true;
 	uint32_t               trophy_duration_seconds     = 6;
+	bool                   perf_overlay_enabled            = false;
 	bool                   playgo_hack_enabled         = false;
 	bool                   skip_notice_screen          = false;
 #if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
@@ -100,6 +110,13 @@ const std::string& GetAudioInputDevice();
 const std::optional<ControllerColor>& GetControllerColor();
 uint32_t GetControllerSpeakerVolume();
 uint32_t GetControllerVibrationIntensity();
+uint32_t GetKeyboardPlayer();
+// Player 0 is the configured user; later players derive their IDs from it.
+int32_t  GetPlayerUserId(int player);
+// Returns -1 when the user ID does not belong to any player.
+int      GetPlayerFromUserId(int32_t user_id);
+// Players after the first are guests: they have no profile, trophies or NP account.
+bool     IsGuestUserId(int32_t user_id);
 PresentMode GetPresentMode();
 int32_t GetGpuIndex();
 bool     FullscreenEnabled();

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/common/platform/sysLinuxFileIO.cpp

@@ -367,7 +367,7 @@ bool SysFileDeleteFile(const std::filesystem::path& name) {
 
 bool SysFileFlush(sys_file_t& f) {
 	if (f.type == SYS_FILE_FILE && f.f != nullptr) {
-		return (fflush(f.f) == 0);
+		return fflush(f.f) == 0 && fsync(fileno(f.f)) == 0;
 	}
 
 	return false;

src/emulator.cpp

@@ -89,42 +89,6 @@ static void MountSandboxDirs() {
 	MountOrCreateDir("_TempData/" + title_id, "/temp");
 }
 
-static bool ClearDirectoryContents(const std::filesystem::path& dir) {
-	bool ok = true;
-
-	for (const auto& entry: Common::File::GetDirEntries(dir)) {
-		if (entry.name == "." || entry.name == "..") {
-			continue;
-		}
-
-		auto path = dir / entry.name;
-
-		if (entry.is_file) {
-			Common::File::RemoveReadonly(path);
-			ok = Common::File::DeleteFile(path) && ok;
-		} else {
-			ok = ClearDirectoryContents(path) && ok;
-			ok = Common::File::DeleteDirectory(path) && ok;
-		}
-	}
-
-	return ok;
-}
-
-static void ClearDebugTextureFolder() {
-	const std::string debug_texture_folder = "_Textures";
-
-	if (!Common::File::IsDirectoryExisting(debug_texture_folder)) {
-		Common::File::CreateDirectories(debug_texture_folder);
-		return;
-	}
-
-	if (!ClearDirectoryContents(debug_texture_folder)) {
-		LOGF_COLOR(Log::Color::BrightYellow, "TextureDump: failed to completely clear %s\n",
-		           debug_texture_folder.c_str());
-	}
-}
-
 static void Init(const Config::ConfigOptions& cfg, const std::filesystem::path& param_json,
                  Common::Subsystems& subsystems) {
 	EXIT_IF(!Common::Thread::IsMainThread());
@@ -198,8 +162,6 @@ void Run(const RunOptions& options) {
 	Common::Subsystems subsystems(true);
 	Init(options.config, param_json, subsystems);
 
-	ClearDebugTextureFolder();
-
 	PrintSystemInfo();
 	std::string title_id;
 	Loader::SystemContentParamSfoGetString("TITLE_ID", &title_id);

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"
@@ -342,12 +343,11 @@ void CommandProcessor::WriteData(uint32_t* dst, const uint32_t* src, uint32_t dw
 		return;
 	}
 
-	if (write_one_address) {
-		for (uint32_t i = 0; i < dw_num; i++) {
-			dst[0] = src[i];
-		}
-	} else {
-		memcpy(dst, src, static_cast<size_t>(dw_num) * sizeof(uint32_t));
+	// Completion labels can be reused immediately; memcpy may write a destination twice.
+	auto* output = static_cast<volatile uint32_t*>(dst);
+	for (uint32_t i = 0; i < dw_num; i++) {
+		*output = src[i];
+		output += !write_one_address;
 	}
 }
 
@@ -906,6 +906,7 @@ void CommandProcessor::DrawIndirectMulti(uint32_t data_offset, uint32_t max_coun
 	if (draw_count == 0) {
 		return;
 	}
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::IndirectDraws, draw_count);
 
 	const auto args_size = indexed ? sizeof(DrawIndexedIndirectArgs) : sizeof(DrawIndirectArgs);
 	EXIT_NOT_IMPLEMENTED(stride_in_bytes < args_size);

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"
@@ -262,14 +263,8 @@ void TextureCache::DeleteImage(ImageId id) {
 		EXIT("TextureCache: deleting a GPU-modified image without resolving its contents\n");
 	}
 	m_download_images.erase(id);
-	if (image->info.HasMetadata()) {
-		const auto metadata = m_surface_metas.find(image->info.metadata.range.address);
-		if (metadata != m_surface_metas.end() &&
-		    image->info.metadata.kind == ImageMetadataKind::Htile &&
-		    metadata->second.type == MetaDataInfo::Type::HTile) {
-			// A later binding may have reused this address for another metadata type.
-			m_surface_metas.erase(metadata);
-		}
+	if (image->info.metadata.kind == ImageMetadataKind::Htile) {
+		m_surface_metas.erase(image->info.metadata.range.address);
 	}
 	UnregisterImage(id);
 	if (m_scheduler.Active()) {
@@ -954,6 +949,7 @@ void TextureCache::UploadImage(Image& image, Buffer& source, uint64_t source_off
 		const auto linear = m_tiler.Detile(source.Handle(), source_offset, info);
 		const auto copies = info.BufferCopies(linear.offset);
 		destination.Upload(copies, linear.buffer, linear.offset, linear.size);
+		Common::DebugCounters::Add(Common::DebugCounters::Counter::TextureUploadBytes, linear.size);
 		return;
 	}
 
@@ -999,6 +995,7 @@ void TextureCache::UploadImage(Image& image, Buffer& source, uint64_t source_off
 		copies[0].imageExtent       = info.extent;
 	}
 	destination.Upload(copies, linear.buffer, linear.offset, linear.size);
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::TextureUploadBytes, linear.size);
 }
 
 void TextureCache::InitializeImage(ImageId id) {
@@ -1039,7 +1036,7 @@ void TextureCache::MaterializeColorClear(ImageId id, const ImageDesc& desc,
 		std::scoped_lock lock {m_lock};
 		auto& image         = m_slot_images[id];
 		image.info.metadata = desc.info.metadata;
-		// Native color metadata must not retain a reused HTile/CMask/FMask clear flag.
+		// Native color metadata must not retain a reused HTile clear flag.
 		m_surface_metas.erase(range.address);
 		if (range.size == 0 || desc.info.resources.levels != 1) {
 			return;
@@ -1417,11 +1414,8 @@ vk::ImageView TextureCache::FindDepthTarget(ImageId id, const ImageDesc& desc) {
 	image.info.metadata = desc.info.metadata;
 	if (desc.info.HasMetadata()) {
 		const auto [metadata, inserted] = m_surface_metas.try_emplace(desc.info.metadata.range.address);
-		if (inserted) {
-			metadata->second.type = MetaDataInfo::Type::HTile;
-			if (image.info.htile_initial_clear) {
-				metadata->second.clear_mask.set();
-			}
+		if (inserted && image.info.htile_initial_clear) {
+			metadata->second.clear_mask.set();
 		}
 	}
 	RefreshImage(id);
@@ -1774,6 +1768,7 @@ bool TextureCache::DownloadImageMemory(ImageId id) {
 	download.Flush(offset, range.size);
 
 	DownloadImage(image, download, offset, range.size, image.info.resources.levels);
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::TextureDownloadBytes, range.size);
 	vk::BufferMemoryBarrier barrier {};
 	barrier.srcAccessMask = vk::AccessFlagBits::eMemoryWrite | vk::AccessFlagBits::eTransferWrite |
 	                        vk::AccessFlagBits::eShaderWrite;
@@ -1943,6 +1938,7 @@ void TextureCache::RunGarbageCollector() {
 				}
 			}
 			FreeImage(id);
+			Common::DebugCounters::Add(Common::DebugCounters::Counter::TextureEvictions);
 			if (m_total_used_memory < m_critical_gc_memory && aggressive) {
 				deletions >>= 2;
 				aggressive = false;

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

@@ -72,7 +72,16 @@ void RenderExecutor::ResolveRenderColorTarget(CommandBuffer& buffer, RenderColor
 		return;
 	}
 	const auto samples = render_sample_count(rt.attrib.num_fragments);
-	if (samples == 0 || rt.attrib.num_samples != rt.attrib.num_fragments) {
+	// TODO(PPSA24156): Replace this 2x MSAA approximation with proper 4-coverage/2-fragment EQAA.
+	const bool approximate_eqaa = rt.attrib.num_samples == 2 && rt.attrib.num_fragments == 1;
+	if (approximate_eqaa) {
+		static std::atomic_bool logged = false;
+		if (!logged.exchange(true, std::memory_order_relaxed)) {
+			Log::WriteToConsoleAndLog(
+			    "Warning: approximating 4-coverage/2-fragment EQAA with 2x MSAA.\n");
+		}
+	}
+	if (samples == 0 || (rt.attrib.num_samples != rt.attrib.num_fragments && !approximate_eqaa)) {
 		EXIT("unsupported render-target sample configuration: samples=%u fragments=%u\n",
 		     rt.attrib.num_samples, rt.attrib.num_fragments);
 	}
@@ -220,8 +229,7 @@ void RenderExecutor::ResolveRenderColorTarget(CommandBuffer& buffer, RenderColor
 	}
 	const bool has_dcc        = rt.info.dcc_compression_enable && rt.dcc_addr.addr != 0;
 	const bool has_cmask      = !rt.info.dcc_compression_enable && rt.info.cmask_fast_clear_enable &&
-	                            rt.cmask.addr != 0 && samples == 1 &&
-	                            !rt.info.fmask_compression_enable &&
+	                            rt.cmask.addr != 0 &&
 	                            !rt.attrib3.write_vrs_rate_hint_to_cmask;
 	if (has_dcc || has_cmask) {
 		TileSizeAlign metadata_size {};
@@ -233,7 +241,7 @@ void RenderExecutor::ResolveRenderColorTarget(CommandBuffer& buffer, RenderColor
 		} else {
 			(void)TileGetCmaskSize(width, height, layers, levels, metadata_size);
 		}
-		// DCC owns the clear when both planes are enabled; single-sample CMASK stays expanded.
+		// DCC owns the clear when both planes are enabled.
 		desc.info.metadata.kind = has_dcc ? ImageMetadataKind::Dcc : ImageMetadataKind::Cmask;
 		desc.info.metadata.range = {has_dcc ? rt.dcc_addr.addr : rt.cmask.addr, metadata_size.size};
 		desc.info.metadata.clear_word =

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/image/image.cpp

@@ -353,8 +353,12 @@ void Image::CopyImage(Image& source) {
 	std::vector<vk::ImageCopy> copies;
 	copies.reserve(levels);
 	for (uint32_t level = 0; level < levels; level++) {
-		const auto width  = std::max(source.backing.extent.width >> level, 1u);
-		const auto height = std::max(source.backing.extent.height >> level, 1u);
+		// Overlapping cache images can differ in size; copying the source extent into a smaller
+		// destination writes out of bounds and loses the device.
+		const auto width  = std::min(std::max(source.backing.extent.width >> level, 1u),
+		                             std::max(backing.extent.width >> level, 1u));
+		const auto height = std::min(std::max(source.backing.extent.height >> level, 1u),
+		                             std::max(backing.extent.height >> level, 1u));
 		const auto depth  = std::max(base_depth >> level, 1u);
 		const auto [source_layers, destination_layers] = SanitizeCopyLayers(source, *this, depth);
 		vk::ImageCopy copy {};

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

@@ -764,7 +764,10 @@ void RenderExecutor::FindBuffers(std::span<PreparedBindings* const> stages) {
 					size = std::min({size, uint64_t {256} * 1024 * 1024, limit});
 				}
 			}
-			size = Libs::LibKernel::Memory::ClampRangeSize(address, size);
+			// Finite GPU descriptors can span sparse PRT holes; clamp only oversized/sentinel ranges.
+			if (size >= (uint64_t {1} << 30u) || descriptor.NumRecords() == UINT32_MAX) {
+				size = Libs::LibKernel::Memory::ClampRangeSize(address, size);
+			}
 			source = {address, size, cache.FindBuffer(address, size)};
 		}
 	}

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>
@@ -273,6 +275,7 @@ struct PipelineCache::ProgramCache {
 		ShaderRecompiler::IR::ResourceSnapshot       resources;
 		ShaderRecompiler::IR::ResourceSpecialization specialization;
 		std::vector<Permutation>                    permutations;
+		bool                                       skip_dispatch = false;
 	};
 
 	struct ProgramKeyHash {
@@ -299,6 +302,9 @@ struct PipelineCache::ProgramCache {
 	                               ShaderRecompiler::IR::ResourceSpecialization specialization,
 	                               uint32_t push_data_start_dword,
 	                               std::span<const uint32_t> function_code) {
+		Common::DebugCounters::ScopedTimer timer(Common::DebugCounters::Counter::ShadersCompiled,
+		                                         Common::DebugCounters::Counter::ShaderCompileNs,
+		                                         Common::DebugCounters::Peak::ShaderCompileNs);
 		auto result = ShaderRecompiler::CompileProgram(std::move(translated), options,
 		                                               specialization, push_data_start_dword);
 		if (!ValidateShaderSpirv(options.dump_label, options.shader_hash, result.spirv,
@@ -355,6 +361,7 @@ struct PipelineCache::ProgramCache {
 		bool source_changed = false;
 		if (entry != programs.end()) {
 			auto& cached = entry->second;
+			if (cached.skip_dispatch) return {};
 			Permutation* permutation = nullptr;
 			if (cached.call_source) {
 				source_changed = cached.RefreshCallSource(runtime);
@@ -414,10 +421,18 @@ struct PipelineCache::ProgramCache {
 		DumpShaderOriginal(stage_name, options.shader_hash, params.code);
 		ShaderRecompiler::ShaderSource source;
 		if (entry == programs.end()) source = ShaderRecompiler::PrepareShaderSource(params.code, options);
+		if (source.unsupported_call_pc != UINT32_MAX) {
+			PipelineCacheLog("!!! TEMPORARY COMPUTE DISPATCH SKIP !!! shader=0x{:016x} pc=0x{:08x} "
+			                 "reason=unsupported GPU-selected callable target; rendering is incomplete",
+			                 params.hash, source.unsupported_call_pc);
+			programs.try_emplace(lookup_key, ShaderRecompiler::IR::ResourcePlan{}, nullptr)
+			    .first->second.skip_dispatch = true;
+			return {};
+		}
 		auto* current = entry == programs.end() ? &source : entry->second.call_source.get();
 		ShaderRecompiler::TranslateResult translated;
 		if (current != nullptr) {
-			if (current->call) {
+			if (!current->calls.empty()) {
 				if (entry == programs.end()) ShaderRecompiler::RefreshShaderSource(*current, runtime);
 				translated = ShaderRecompiler::TranslateProgram(current->linked->decoded, options);
 				translated.program.source_reads = current->reads;
@@ -428,7 +443,7 @@ struct PipelineCache::ProgramCache {
 			translated = ShaderRecompiler::TranslateProgram(params.code, options);
 		}
 		if (entry == programs.end()) {
-			auto retained = source.call ? std::make_unique<ShaderRecompiler::ShaderSource>(std::move(source)) : nullptr;
+			auto retained = !source.calls.empty() ? std::make_unique<ShaderRecompiler::ShaderSource>(std::move(source)) : nullptr;
 			entry = programs.try_emplace(lookup_key,
 			    ShaderRecompiler::IR::ExtractResourcePlan(translated.program), std::move(retained)).first;
 			EXIT_IF(!ShaderRecompiler::IR::MaterializeResources(
@@ -575,8 +590,8 @@ void PipelineCache::InitializeDriverCache() {
 		return;
 	}
 	if (!initial_data.empty()) {
-		PipelineCacheLog("Vulkan pipeline cache: loaded {} bytes from {}", initial_data.size(),
-		                 path);
+		PipelineCacheLog("Vulkan pipeline cache: loaded {:.2f} MB from {}",
+		                 static_cast<double>(initial_data.size()) / 1'000'000.0, path);
 	} else {
 		PipelineCacheLog("Vulkan pipeline cache: initialized empty");
 	}
@@ -616,8 +631,8 @@ void PipelineCache::Save() {
 	}
 	if (result != vk::Result::eSuccess || size == 0 ||
 	    size > std::numeric_limits<uint32_t>::max()) {
-		PipelineCacheLog("Vulkan pipeline cache: save failed ({}, {} bytes)",
-		                 vk::to_string(result), size);
+		PipelineCacheLog("Vulkan pipeline cache: save failed ({}, {:.2f} MB)",
+		                 vk::to_string(result), static_cast<double>(size) / 1'000'000.0);
 		return;
 	}
 	payload.resize(size);
@@ -649,7 +664,8 @@ void PipelineCache::Save() {
 		                 Common::PathToString(m_driver_cache_path));
 		return;
 	}
-	PipelineCacheLog("Vulkan pipeline cache: saved {} bytes to {}", payload.size(),
+	PipelineCacheLog("Vulkan pipeline cache: saved {:.2f} MB to {}",
+	                 static_cast<double>(payload.size()) / 1'000'000.0,
 	                 Common::PathToString(m_driver_cache_path));
 	m_driver_cache_dirty = false;
 }
@@ -993,8 +1009,13 @@ PipelineCache::Pipeline& PipelineCache::GetGraphicsPipeline(
 
 	auto cached = std::make_unique<Pipeline>();
 	LogPipelineTrace("CreatePipelineInternal begin", vs_id, ps_id);
-	CreatePipelineInternal(m_graphics, *cached, rendering, key.vertex_input, vertex_info,
-	                       ps_input_info, programs, static_params, m_driver_cache);
+	{
+		Common::DebugCounters::ScopedTimer timer(Common::DebugCounters::Counter::PipelinesCreated,
+		                                         Common::DebugCounters::Counter::PipelineCreateNs,
+		                                         Common::DebugCounters::Peak::PipelineCreateNs);
+		CreatePipelineInternal(m_graphics, *cached, rendering, key.vertex_input, vertex_info,
+		                       ps_input_info, programs, static_params, m_driver_cache);
+	}
 	LogPipelineTrace("CreatePipelineInternal done", vs_id, ps_id);
 
 	EXIT_NOT_IMPLEMENTED(cached->pipeline == nullptr);
@@ -1024,7 +1045,13 @@ PipelineCache::GetComputePipeline(const ShaderComputeInputInfo& input_info,
 	}
 
 	auto cached = std::make_unique<Pipeline>();
-	CreatePipelineInternal(m_graphics, *cached, input_info, compute_program.module, m_driver_cache);
+	{
+		Common::DebugCounters::ScopedTimer timer(Common::DebugCounters::Counter::PipelinesCreated,
+		                                         Common::DebugCounters::Counter::PipelineCreateNs,
+		                                         Common::DebugCounters::Peak::PipelineCreateNs);
+		CreatePipelineInternal(m_graphics, *cached, input_info, compute_program.module,
+		                       m_driver_cache);
+	}
 
 	EXIT_NOT_IMPLEMENTED(cached->pipeline == nullptr);
 	EXIT_NOT_IMPLEMENTED(cached->pipeline_layout == nullptr);

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

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

src/graphics/host_gpu/renderer/render.h

@@ -201,11 +201,11 @@ private:
 	void                      BindImage(ImageId id, bool storage);
 	void                      BindRenderTarget(ImageId id);
 	void                      ResetBindings();
-	void                     UpdateComputeMetaClears(const ShaderComputeInputInfo& input,
-	                                                 const CommandBuffer&          buffer);
-	[[nodiscard]] bool TryConsumeComputeImageClear(const ShaderComputeInputInfo& input,
-	                                              CommandBuffer& command, uint32_t group_x,
-	                                              uint32_t group_y, uint32_t group_z, uint32_t mode);
+	[[nodiscard]] bool        TryConsumeComputeMetaClear(const ShaderComputeInputInfo& input,
+	                                                     const CommandBuffer&          buffer);
+	[[nodiscard]] bool TryConsumeComputeFill(const ShaderComputeInputInfo& input,
+	                                        CommandBuffer& command, uint32_t group_x,
+	                                        uint32_t group_y, uint32_t group_z, uint32_t mode);
 
 	RenderContext&                        m_context;
 	GraphicsBindings                     m_graphics_bindings;

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"
@@ -50,8 +51,8 @@ static bool FillSourcesDisjoint(std::span<const ShaderRecompiler::IR::Descriptor
 	return true;
 }
 
-void RenderExecutor::UpdateComputeMetaClears(const ShaderComputeInputInfo& input,
-                                            const CommandBuffer&          buffer) {
+bool RenderExecutor::TryConsumeComputeMetaClear(const ShaderComputeInputInfo& input,
+                                                const CommandBuffer&          buffer) {
 	const auto& program   = *input.stage.program;
 	const auto& resources = *input.stage.resources;
 	if (resources.buffers.size() != program.info.buffers.size()) {
@@ -61,9 +62,10 @@ void RenderExecutor::UpdateComputeMetaClears(const ShaderComputeInputInfo& input
 	for (uint32_t i = 0; i < program.info.buffers.size(); i++) {
 		const auto& resource   = program.info.buffers[i];
 		const auto  descriptor = DecodeNativeDescriptor<ShaderBufferResource>(resources.buffers[i]);
-		// Reads and transforms of metadata do not establish a clear state.
+		// A metadata resource that is also read is not proven to be a full overwrite. Execute it
+		// conservatively instead of replacing the dispatch with a coarse full-surface clear.
 		if ((!resource.written || resource.read) && cache.IsMeta(descriptor.Base48())) {
-			return;
+			return false;
 		}
 	}
 
@@ -73,10 +75,13 @@ void RenderExecutor::UpdateComputeMetaClears(const ShaderComputeInputInfo& input
 			if (resource.written) {
 				const auto descriptor =
 				    DecodeNativeDescriptor<ShaderBufferResource>(resources.buffers[i]);
-				(void)cache.ClearMeta(descriptor.Base48());
+				if (cache.ClearMeta(descriptor.Base48())) {
+					return true;
+				}
 			}
 		}
 	}
+	return false;
 }
 
 bool ResolveComputeBufferFill(const ShaderComputeInputInfo& input, uint32_t group_x,
@@ -94,9 +99,10 @@ bool ResolveComputeBufferFill(const ShaderComputeInputInfo& input, uint32_t grou
 	constexpr std::array formats {
 	    Prospero::BufferFormat::k32UInt, Prospero::BufferFormat::k32_32UInt,
 	    Prospero::BufferFormat::k32_32_32UInt, Prospero::BufferFormat::k32_32_32_32UInt};
+	// Offset-or-swizzle bounds do not describe the element range proved by the shader.
 	if (descriptor.Stride() != element_size || descriptor.Format() != formats[fill.words - 1] ||
 	    descriptor.SwizzleEnabled() || descriptor.IndexStride() != 0 || descriptor.AddTid() ||
-	    descriptor.Base48() == 0) {
+	    descriptor.OutOfBounds() == 3 || descriptor.Base48() == 0) {
 		return false;
 	}
 	if (input.threads_num[0] == 0 || input.threads_num[0] != fill.group_stride[0] ||
@@ -108,7 +114,11 @@ bool ResolveComputeBufferFill(const ShaderComputeInputInfo& input, uint32_t grou
 	                                 ? group_x
 	                                 : static_cast<uint64_t>(group_x) * input.threads_num[0];
 	const auto     size        = descriptor.GetSize();
-	if (invocations != descriptor.NumRecords() || size == 0 || size > UINT32_MAX ||
+	const auto elements = descriptor.NumRecords();
+	const bool covered = fill.bounded_elements != 0
+	                         ? fill.bounded_elements == elements && invocations >= elements
+	                         : invocations == elements;
+	if (!covered || invocations > uint64_t {UINT32_MAX} + 1 || size == 0 || size > UINT32_MAX ||
 	    (input.dispatch_thread_dimensions &&
 	     (group_x % input.threads_num[0] != 0 || input.dispatch_threads_num[0] != group_x ||
 	      input.dispatch_threads_num[1] != 1 || input.dispatch_threads_num[2] != 1))) {
@@ -122,9 +132,9 @@ bool ResolveComputeBufferFill(const ShaderComputeInputInfo& input, uint32_t grou
 	return true;
 }
 
-bool RenderExecutor::TryConsumeComputeImageClear(const ShaderComputeInputInfo& input,
-                                                CommandBuffer& command, uint32_t group_x,
-                                                uint32_t group_y, uint32_t group_z, uint32_t mode) {
+bool RenderExecutor::TryConsumeComputeFill(const ShaderComputeInputInfo& input,
+                                          CommandBuffer& command, uint32_t group_x,
+                                          uint32_t group_y, uint32_t group_z, uint32_t mode) {
 	const auto& program   = *input.stage.program;
 	const auto& resources = *input.stage.resources;
 	const auto& fill      = resources.uniform_fill;
@@ -180,14 +190,26 @@ bool RenderExecutor::TryConsumeComputeImageClear(const ShaderComputeInputInfo& i
 	                              size)) {
 		return false;
 	}
-	if (!cache.ClearImageFromBuffer(command, descriptor.Base48(), size, packed_clear)) {
-		return false;
+	const bool image_clear =
+	    cache.ClearImageFromBuffer(command, descriptor.Base48(), size, packed_clear);
+	if (!image_clear) {
+		// CPU fills invalidate overlapping cached images. Preserve the native dispatch for
+		// partial or ambiguous image backings, including pooled metadata aliases.
+		if ((descriptor.Base48() & 3u) != 0 ||
+		    !m_context.IsMapped(descriptor.Base48(), size) ||
+		    cache.IsRegionRegistered(descriptor.Base48(), size)) {
+			return false;
+		}
+		// Metadata need not have an image backing yet. Keep proven fills in guest memory
+		// when CPU-owned, so color-clear discovery can read them without a GPU download.
+		m_context.GetBufferCache().FillBuffer(descriptor.Base48(), size, packed_clear, false);
 	}
 	static std::atomic<uint32_t> logged_clears {0};
 	if (logged_clears.fetch_add(1, std::memory_order_relaxed) < 32) {
-		LOGF("GraphicsRenderDispatchDirect: compute image clear shader=0x%016" PRIx64
+		LOGF("GraphicsRenderDispatchDirect: compute %s shader=0x%016" PRIx64
 		     " addr=0x%016" PRIx64 " size=0x%016" PRIx64 " value=0x%08" PRIx32 "\n",
-		     input.stage.program->shader_hash, descriptor.Base48(), size, packed_clear);
+		     image_clear ? "image clear" : "buffer fill", input.stage.program->shader_hash,
+		     descriptor.Base48(), size, packed_clear);
 	}
 	return true;
 }
@@ -252,10 +274,6 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 		return;
 	}
 
-	if (sh_ctx.GetCs().cs_regs.data_addr == 0) {
-		return;
-	}
-
 	constexpr uint32_t DISPATCH_INITIATOR_USE_THREAD_DIMENSIONS = 1u << 5u;
 	constexpr uint32_t DISPATCH_INITIATOR_BASE_BITS             = 0x41u;
 	constexpr uint32_t DISPATCH_INITIATOR_MODIFIER_BITS         = 0xa038u;
@@ -295,6 +313,7 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 	}
 	const auto compute_program =
 	    m_context.GetPipelineCache().GetComputeProgram(cs_regs, sh_regs, input_info);
+	if (!compute_program) return;
 	if (use_thread_dimensions) {
 		input_info.dispatch_threads_num[0]    = thread_group_x;
 		input_info.dispatch_threads_num[1]    = thread_group_y;
@@ -303,14 +322,12 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 
 	const auto& program   = *input_info.stage.program;
 	const auto& resources = *input_info.stage.resources;
-	if (resources.specialization_reads.empty()) {
-		// Cached metadata addresses can be reused as ordinary buffers; preserve their guest writes.
-		UpdateComputeMetaClears(input_info, buffer);
-		if (TryConsumeComputeImageClear(input_info, buffer, thread_group_x, thread_group_y,
-		                               thread_group_z, mode)) {
-			ResetBindings();
-			return;
-		}
+	if (resources.specialization_reads.empty() &&
+	    (TryConsumeComputeMetaClear(input_info, buffer) ||
+	     TryConsumeComputeFill(input_info, buffer, thread_group_x, thread_group_y,
+	                           thread_group_z, mode))) {
+		ResetBindings();
+		return;
 	}
 
 	if (use_thread_dimensions) {
@@ -431,6 +448,7 @@ 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);
@@ -461,6 +479,7 @@ void RenderExecutor::DispatchIndirect(uint64_t submit_id, CommandBuffer& buffer,
 	input_info.indirect_args_address = args_addr;
 	const auto compute_program = m_context.GetPipelineCache().GetComputeProgram(
 	    cs_regs, buffer.GetRegisters().GetShaderRegisters(), input_info);
+	if (!compute_program) return;
 	buffer.EndRendering();
 	auto& pipeline = m_context.GetPipelineCache().GetComputePipeline(input_info, compute_program);
 	auto& bindings = m_compute_bindings;
@@ -500,6 +519,7 @@ 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");
@@ -567,9 +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;
@@ -1060,7 +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();
@@ -1230,6 +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/ShaderRecompiler.cpp

@@ -481,18 +481,14 @@ Decoder::Program DecodeFusedProgram(std::span<const uint32_t> front, std::span<c
 	return result;
 }
 
-void PrepareCallTarget(ShaderSource& source, const CompileOptions& options) {
+void PrepareCallTargets(ShaderSource& source, const CompileOptions& options) {
 	using namespace IR;
+	EXIT_NOT_IMPLEMENTED(options.stage != ShaderType::Compute);
 	for (uint32_t i = 0; i < source.decoded.instructions.size(); ++i) {
 		if (source.decoded.instructions[i].opcode != Decoder::Opcode::S_SWAPPC_B64) continue;
-		EXIT_NOT_IMPLEMENTED(source.call.has_value() || options.stage != ShaderType::Compute);
-		source.call = ShaderSource::Call {.instruction = i};
+		source.calls.push_back({.instruction = i});
 	}
 	EXIT_IF(options.input_info.compute == nullptr);
-	const auto& call = source.decoded.instructions[source.call->instruction];
-	EXIT_NOT_IMPLEMENTED(call.src0.kind != Decoder::OperandKind::Sgpr ||
-	                     call.dst.kind != Decoder::OperandKind::Sgpr ||
-	                     call.src0.reg != call.dst.reg || call.dst.reg >= 105u);
 	const auto graph = CFG::BuildGraph(source.decoded);
 	Program query;
 	query.stage = options.stage;
@@ -524,7 +520,11 @@ void PrepareCallTarget(ShaderSource& source, const CompileOptions& options) {
 		return value;
 	};
 	Register = [&](uint32_t before, uint32_t code) -> Value {
-		const auto index = CFG::FindScalarDefinition(source.decoded, graph, before, code);
+		std::vector<uint32_t> crossed_calls;
+		const auto index = CFG::FindScalarDefinition(source.decoded, graph, before, code, crossed_calls);
+		for (const auto crossed: crossed_calls) {
+			std::ranges::find(source.calls, crossed, &ShaderSource::Call::instruction)->preserved_registers.set(code);
+		}
 		if (index == UINT32_MAX) {
 			EXIT_IF(code < query.user_data_base || code - query.user_data_base >= query.user_data_count);
 			return IREmitter(&block).GetUserData(static_cast<ScalarReg>(code));
@@ -557,18 +557,30 @@ void PrepareCallTarget(ShaderSource& source, const CompileOptions& options) {
 		EXIT_IF(!producer.ready || !producer.values.contains(code));
 		return producer.values.at(code).Resolve();
 	};
-	for (uint32_t i = 0; i < 2u; ++i) {
-		const auto value = Register(source.call->instruction, call.src0.reg + i);
-		EXIT_IF(!ValidateRuntimeValue(query, value, RuntimeValueType::Integer));
-		auto& root = query.value_storage.emplace_back(ValueOpcode::ReferenceU32);
-		root.SetArg(0, value);
-		source.call->target[i] = Value(&root);
+	for (auto& site: source.calls) {
+		const auto& call = source.decoded.instructions[site.instruction];
+		EXIT_NOT_IMPLEMENTED(call.src0.kind != Decoder::OperandKind::Sgpr ||
+		                     call.dst.kind != Decoder::OperandKind::Sgpr ||
+		                     call.src0.reg != call.dst.reg || call.dst.reg >= 105u);
+		for (uint32_t i = 0; i < 2u; ++i) {
+			const auto value = Register(site.instruction, call.src0.reg + i);
+			if (!ValidateRuntimeValue(query, value, RuntimeValueType::Integer)) {
+				EXIT_IF(value.Instruction()->GetOpcode() != ValueOpcode::ReadConstBuffer);
+				source.unsupported_call_pc = call.pc;
+				source.calls.clear();
+				return;
+			}
+			auto& root = query.value_storage.emplace_back(ValueOpcode::ReferenceU32);
+			root.SetArg(0, value);
+			site.target[i] = Value(&root);
+		}
 	}
 	RewriteToSsa(query.blocks);
 	ConstantPropagationPass(query.blocks, query.wave_size);
 	RemoveIdentities(query);
 	EliminateDeadCode(query.blocks);
-	for (auto& value: source.call->target) value = value.Instruction()->Arg(0).Resolve();
+	for (auto& site: source.calls)
+		for (auto& value: site.target) value = value.Instruction()->Arg(0).Resolve();
 	query.value_storage.clear();
 	for (auto& inst: block) {
 		inst.SetParent(nullptr);
@@ -595,7 +607,7 @@ ShaderSource PrepareShaderSource(std::span<const uint32_t> code, const CompileOp
 		Decoder::DecodeProgram(code, source.decoded);
 	}
 	if (source.decoded.has_swap_pc) {
-		PrepareCallTarget(source, options);
+		PrepareCallTargets(source, options);
 		source.code.assign(code.begin(), code.end());
 		source.decoded.code = source.code;
 	}
@@ -603,88 +615,122 @@ ShaderSource PrepareShaderSource(std::span<const uint32_t> code, const CompileOp
 }
 
 const Decoder::Program& RefreshShaderSource(ShaderSource& source, const IR::SrtRuntime& runtime) {
+	EXIT_IF(source.unsupported_call_pc != UINT32_MAX);
 	auto& reads = source.reads;
 	reads.clear();
 	IR::SrtReadCapture capture(runtime, reads);
 	const auto clean = IR::CleanRuntime(capture.ObservedRuntime());
 	IR::SrtWalker walker(source.call_targets, clean);
-	uint32_t low = 0, high = 0;
-	const auto& call = source.decoded.instructions[source.call->instruction];
-	if (!walker.Evaluate(source.call->target[0], low) || !walker.Evaluate(source.call->target[1], high))
-		EXIT("shader call at pc 0x%08x has an unavailable scalar target", call.pc);
-	// SWAPPC reads the old aliased pair and ignores the target's low two bits.
-	const auto address = ((uint64_t {high} << 32u) | low) & ~uint64_t {3};
-	EXIT_IF(address == 0u);
+	for (auto& site: source.calls) {
+		uint32_t low = 0, high = 0;
+		if (!walker.Evaluate(site.target[0], low) || !walker.Evaluate(site.target[1], high))
+			EXIT("shader call at pc 0x%08x has an unavailable scalar target",
+			     source.decoded.instructions[site.instruction].pc);
+		// SWAPPC reads the old aliased pair and ignores the target's low two bits.
+		const auto address = ((uint64_t {high} << 32u) | low) & ~uint64_t {3};
+		EXIT_IF(address == 0u);
+		site.address = address;
+	}
 	const auto query_reads = reads.size();
 	if (source.linked) {
 		auto& linked = *source.linked;
-		const auto function = std::span(linked.code).subspan(source.code.size());
-		if (clean.read_memory(clean.userdata, address, linked.observed_function) &&
-		    std::ranges::equal(function, linked.observed_function)) {
-			return linked.decoded;
+		bool unchanged = true;
+		for (size_t i = 0; i < source.calls.size(); ++i) {
+			const auto& function = linked.functions[i];
+			const auto expected = std::span(linked.code).subspan(function.offset, function.size);
+			const auto preceding = std::span(source.calls).first(i);
+			const auto previous = std::ranges::find(preceding, source.calls[i].address, &ShaderSource::Call::address);
+			if (previous != preceding.end()) {
+				const auto& observed = linked.functions[previous - preceding.begin()];
+				if (std::ranges::equal(expected, std::span(linked.code).subspan(observed.offset, observed.size))) continue;
+				unchanged = false;
+				break;
+			}
+			linked.observed_function.resize(function.size);
+			if (!clean.read_memory(clean.userdata, source.calls[i].address, linked.observed_function) ||
+			    !std::ranges::equal(expected, linked.observed_function)) {
+				unchanged = false;
+				break;
+			}
 		}
+		if (unchanged) return linked.decoded;
 		reads.resize(query_reads);
 	}
-	ShaderSource::Linked linked {.code = source.code,
-	                             .decoded = source.decoded};
+	ShaderSource::Linked linked {.code = source.code, .decoded = source.decoded};
 	const auto& last = source.decoded.instructions.back();
 	EXIT_NOT_IMPLEMENTED(last.opcode != Decoder::Opcode::S_ENDPGM);
-	// Keep the caller's data/footer bytes at their native offsets. The leaf's virtual CFG
-	// PCs follow decoded instructions; its physical code is independently read and owned.
-	const auto function_pc = last.pc + last.word_count * sizeof(uint32_t);
-	for (uint32_t index = 0;;) {
-		// The observed queue helper is a scalar leaf. Scalar encodings need at most one literal.
-		EXIT_IF(index >= 64u || address > UINT64_MAX - (uint64_t {index} + 2u) * sizeof(uint32_t));
-		std::array<uint32_t, 2> words {};
-		if (!clean.read_memory(clean.userdata, address + uint64_t {index} * sizeof(uint32_t),
-		                       std::span(words).first(1)))
-			EXIT("shader call at pc 0x%08x cannot read leaf code", call.pc);
-		const auto family = Decoder::GetInstructionFamily(words[0]);
-		EXIT_NOT_IMPLEMENTED(family != Decoder::Family::SOP1 && family != Decoder::Family::SOP2 &&
-		                     family != Decoder::Family::SOPC && family != Decoder::Family::SOPK &&
-		                     family != Decoder::Family::SOPP);
-		Decoder::Instruction inst;
-		Decoder::DecodeInstruction(words, 0u, inst);
-		if (inst.word_count == 2u) {
-			if (!clean.read_memory(clean.userdata, address + uint64_t {index + 1u} * sizeof(uint32_t),
-			                       std::span(words).last(1)))
-				EXIT("shader call at pc 0x%08x cannot read leaf literal", call.pc);
-			inst = {};
-			Decoder::DecodeInstruction(words, 0u, inst);
+	// Keep the caller's data/footer at native offsets. Expand leaf CFGs per call site
+	// before SSA, so return edges neither require a dispatcher nor merge resource roots.
+	uint32_t function_pc = last.pc + last.word_count * sizeof(uint32_t);
+	std::map<uint64_t, std::vector<Decoder::Instruction>> functions;
+	for (size_t i = 0; i < source.calls.size(); ++i) {
+		const auto& site = source.calls[i];
+		const auto& call = source.decoded.instructions[site.instruction];
+		const auto address = source.calls[i].address;
+		auto [entry, inserted] = functions.try_emplace(address);
+		auto& function = entry->second;
+		if (inserted) {
+			for (uint32_t index = 0;;) {
+				std::array<uint32_t, Decoder::MaxInstructionRawWords> words {};
+				EXIT_IF(index > UINT32_MAX / sizeof(uint32_t) - words.size() ||
+				        address > UINT64_MAX - (uint64_t {index} + words.size()) * sizeof(uint32_t));
+				uint32_t count = 1;
+				Decoder::Instruction inst;
+				for (;;) {
+					if (!clean.read_memory(clean.userdata, address + uint64_t {index} * sizeof(uint32_t),
+					                       std::span(words).first(count)))
+						EXIT("shader call at pc 0x%08x cannot read leaf code", call.pc);
+					inst = {};
+					Decoder::DecodeInstruction(words, 0u, inst);
+					if (inst.word_count <= count) break;
+					count = inst.word_count;
+				}
+				inst.pc = index * sizeof(uint32_t);
+				if (Decoder::IsDirectBranch(inst.opcode)) inst.branch_target += inst.pc;
+				index += inst.word_count;
+				EXIT_NOT_IMPLEMENTED(inst.opcode == Decoder::Opcode::S_SWAPPC_B64 ||
+				                     inst.opcode == Decoder::Opcode::S_GETPC_B64 ||
+				                     inst.opcode == Decoder::Opcode::S_ENDPGM);
+				function.push_back(inst);
+				if (inst.opcode == Decoder::Opcode::S_SETPC_B64) break;
+			}
 		}
-		inst.pc = index * sizeof(uint32_t);
-		if (Decoder::IsDirectBranch(inst.opcode)) inst.branch_target += inst.pc;
-		linked.code.insert(linked.code.end(), inst.raw, inst.raw + inst.word_count);
-		index += inst.word_count;
-		EXIT_NOT_IMPLEMENTED(inst.opcode == Decoder::Opcode::S_SWAPPC_B64 ||
-		                     inst.opcode == Decoder::Opcode::S_GETPC_B64 ||
-		                     inst.opcode == Decoder::Opcode::S_ENDPGM);
-		EXIT_IF(CFG::MayWriteScalarRegister(inst, call.dst.reg) || CFG::MayWriteScalarRegister(inst, call.dst.reg + 1u));
-		const bool returns = inst.opcode == Decoder::Opcode::S_SETPC_B64;
-		if (returns) {
-			EXIT_NOT_IMPLEMENTED(inst.src0.kind != Decoder::OperandKind::Sgpr || inst.src0.reg != call.dst.reg);
-			inst.opcode = Decoder::Opcode::S_BRANCH;
-			inst.branch_target = call.pc + call.word_count * sizeof(uint32_t);
-		} else if (Decoder::IsDirectBranch(inst.opcode)) {
-			EXIT_IF(inst.branch_target > UINT32_MAX - function_pc);
-			inst.branch_target += function_pc;
+		const uint32_t function_size = function.back().pc + function.back().word_count * sizeof(uint32_t);
+		EXIT_IF(function_pc > UINT32_MAX - function_size);
+		const auto function_offset = static_cast<uint32_t>(linked.code.size());
+		for (auto inst: function) {
+			EXIT_IF(CFG::MayWriteScalarRegister(inst, call.dst.reg) ||
+			        CFG::MayWriteScalarRegister(inst, call.dst.reg + 1u));
+			for (uint32_t code = 0; code < site.preserved_registers.size(); ++code) {
+				if (site.preserved_registers[code] && CFG::MayWriteScalarRegister(inst, code))
+					EXIT("shader call at pc 0x%08x clobbers scalar target dependency s%u", call.pc, code);
+			}
+			linked.code.insert(linked.code.end(), inst.raw, inst.raw + inst.word_count);
+			if (inst.opcode == Decoder::Opcode::S_SETPC_B64) {
+				EXIT_NOT_IMPLEMENTED(inst.src0.kind != Decoder::OperandKind::Sgpr || inst.src0.reg != call.dst.reg);
+				inst.opcode = Decoder::Opcode::S_BRANCH;
+				inst.branch_target = call.pc + call.word_count * sizeof(uint32_t);
+			} else if (Decoder::IsDirectBranch(inst.opcode)) {
+				EXIT_IF(inst.branch_target >= function_size);
+				inst.branch_target += function_pc;
+			}
+			inst.pc += function_pc;
+			linked.decoded.instructions.push_back(std::move(inst));
 		}
-		inst.pc += function_pc;
-		linked.decoded.instructions.push_back(std::move(inst));
-		if (returns) break;
-	}
-	const auto function_end = linked.decoded.instructions.back().pc +
-	                          linked.decoded.instructions.back().word_count * sizeof(uint32_t);
-	for (size_t i = source.decoded.instructions.size(); i + 1u < linked.decoded.instructions.size(); ++i) {
-		const auto& inst = linked.decoded.instructions[i];
-		if (Decoder::IsDirectBranch(inst.opcode))
-			EXIT_IF(inst.branch_target < function_pc || inst.branch_target >= function_end);
+		linked.functions.push_back({function_offset, function_size / 4u});
+		linked.decoded.instructions[site.instruction].branch_target = function_pc;
+		function_pc += function_size;
 	}
 	reads.resize(query_reads);
-	reads.emplace_back(address, (linked.code.size() - source.code.size()) * sizeof(uint32_t));
-	linked.decoded.instructions[source.call->instruction].branch_target = function_pc;
+	for (size_t i = 0; i < source.calls.size(); ++i) {
+		const auto preceding = std::span(source.calls).first(i);
+		if (std::ranges::find(preceding, source.calls[i].address, &ShaderSource::Call::address) == preceding.end())
+			reads.emplace_back(source.calls[i].address, linked.functions[i].size * sizeof(uint32_t));
+	}
 	linked.decoded.code = linked.code;
-	linked.observed_function.resize(linked.code.size() - source.code.size());
+	linked.observed_function.resize(std::ranges::max_element(linked.functions, {},
+	    &ShaderSource::Linked::Function::size)->size);
+
 	source.linked = std::move(linked);
 	++source.revision;
 	return source.linked->decoded;

src/graphics/shader/recompiler/ShaderRecompiler.h

@@ -8,6 +8,7 @@
 #include "graphics/shader/shader.h"
 
 #include <array>
+#include <bitset>
 #include <optional>
 #include <span>
 #include <vector>
@@ -47,16 +48,21 @@ struct ShaderSource {
 	struct Call {
 		uint32_t instruction;
 		std::array<IR::Value, 2> target;
+		std::bitset<128> preserved_registers;
+		uint64_t address = 0;
 	};
 	struct Linked {
+		struct Function { uint32_t offset; uint32_t size; };
 		std::vector<uint32_t> code;
 		Decoder::Program decoded;
+		std::vector<Function> functions;
 		std::vector<uint32_t> observed_function;
 	};
 	std::vector<uint32_t> code;
 	Decoder::Program decoded;
 	IR::ResourcePlan call_targets;
-	std::optional<Call> call;
+	std::vector<Call> calls;
+	uint32_t unsupported_call_pc = UINT32_MAX;
 	std::optional<Linked> linked;
 	uint64_t revision = 0;
 	std::vector<std::pair<uint64_t, uint64_t>> reads;

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

@@ -81,7 +81,8 @@ void ValidateNativeProgram(const IR::Program& program, bool lds_storage) {
 						}
 						uses_gds |= memory.kind == IR::ResourceKind::Gds;
 						uses_lds |= memory.kind == IR::ResourceKind::Lds;
-					} else if (memory.kind == IR::ResourceKind::Buffer) {
+					} else if (memory.kind == IR::ResourceKind::Buffer ||
+					           memory.kind == IR::ResourceKind::ScalarBuffer) {
 						if (memory.resource >= program.info.buffers.size()) {
 							Fail(program, "typed shader contains an invalid buffer resource");
 						}

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();
@@ -122,19 +160,6 @@ uint32_t EmitFPMad32(EmitterState& state, uint32_t a, uint32_t b, uint32_t c) {
 	return sum;
 }
 
-uint32_t EmitFPMedTri32(EmitterState& state, uint32_t a, uint32_t b, uint32_t c) {
-	const auto min_ab   = EmitMinMaxF32Value(state, a, b, false);
-	const auto min3     = EmitMinMaxF32Value(state, min_ab, c, false);
-	const auto max_ab   = EmitMinMaxF32Value(state, a, b, true);
-	const auto high_min = EmitMinMaxF32Value(state, max_ab, c, false);
-	const auto median   = EmitMinMaxF32Value(state, min_ab, high_min, true);
-	const auto nan_ab   = Binary(state, spv::OpLogicalOr, TypeBool(state),
-	                             EmitClassifyF32(state, a).nan, EmitClassifyF32(state, b).nan);
-	const auto any_nan =
-	    Binary(state, spv::OpLogicalOr, TypeBool(state), nan_ab, EmitClassifyF32(state, c).nan);
-	return Select(state, TypeF32(state), any_nan, min3, median);
-}
-
 uint32_t EmitFindUMsb64(EmitterState& state, uint32_t value) {
 	// GLSL FindUMsb is restricted to 32-bit components.
 	const auto unpacked = Unary(state, spv::OpBitcast, TypeU32Vector(state, 2), value);
@@ -210,28 +235,18 @@ uint32_t EmitCompositeExtractU32x2(EmitterState& state, uint32_t arg0, IR::Value
 }
 
 uint32_t EmitPackFloat2x16Rtz(EmitterState& state, uint32_t arg0, uint32_t arg1) {
-	const auto pair = state.builder.AllocateId();
-	state.builder.AddFunction(spv::OpCompositeConstruct, TypeF32Vector(state, 2), pair, arg0, arg1);
-	const auto packed = EmitPackHalf2x16(state, pair);
-	const auto back   = EmitExt(state, TypeF32Vector(state, 2), GLSLstd450UnpackHalf2x16, {packed});
-	auto       result = packed;
-	for (uint32_t i = 0; i < 2; i++) {
-		const auto rounded = state.builder.AllocateId();
-		state.builder.AddFunction(spv::OpCompositeExtract, TypeF32(state), rounded, back, i);
-		const auto above =
-		    Binary(state, spv::OpFOrdGreaterThan, TypeBool(state),
-		           EmitExt(state, TypeF32(state), GLSLstd450FAbs, {rounded}),
-		           EmitExt(state, TypeF32(state), GLSLstd450FAbs, {i == 0 ? arg0 : arg1}));
-		result = Binary(state, spv::OpISub, TypeU32(state), result,
-		                Select(state, TypeU32(state), above,
-		                       ConstantU32(state, i == 0 ? 1u : 0x10000u), ConstantU32(state, 0)));
-	}
-	return result;
+	const auto low  = EmitF32ToF16RtzBits(state, arg0);
+	const auto high = Binary(state, spv::OpShiftLeftLogical, TypeU32(state),
+	                         EmitF32ToF16RtzBits(state, arg1), ConstantU32(state, 16));
+	return EmitOrU32(state, low, high);
 }
 
 uint32_t EmitFPSaturate32(EmitterState& state, uint32_t arg0) {
-	return EmitExt(state, TypeF32(state), GLSLstd450FClamp,
-	               {arg0, ConstantF32(state, 0), ConstantF32(state, 0x3f800000u)});
+	// The clamp modifier runs in DX10 clamp mode: NaN becomes 0. FClamp leaves NaN undefined.
+	const auto clamped = EmitExt(state, TypeF32(state), GLSLstd450FClamp,
+	                             {arg0, ConstantF32(state, 0), ConstantF32(state, 0x3f800000u)});
+	return Select(state, TypeF32(state), EmitClassifyF32(state, arg0).nan, ConstantF32(state, 0),
+	              clamped);
 }
 
 uint32_t EmitSMulHi(EmitterState& state, uint32_t arg0, uint32_t arg1) {

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

@@ -100,10 +100,9 @@ uint32_t EmitClassMaskF16(EmitterState& state, uint32_t bits, uint32_t mask) {
 	return EmitClassMaskBits(state, bits, mask, true);
 }
 
-uint32_t EmitMinMaxF32Value(EmitterState& state, uint32_t lhs, uint32_t rhs, bool max_value) {
-	const auto lhs_class = EmitClassifyF32(state, lhs);
-	const auto rhs_class = EmitClassifyF32(state, rhs);
-
+static uint32_t EmitMinMaxF32(EmitterState& state, uint32_t lhs, uint32_t rhs,
+                             const F32Class& lhs_class, const F32Class& rhs_class,
+                             bool max_value) {
 	const auto numeric_cond = state.builder.AllocateId();
 	state.builder.AddFunction(max_value ? spv::OpFOrdGreaterThanEqual : spv::OpFOrdLessThan,
 	                          TypeBool(state), numeric_cond, lhs, rhs);
@@ -121,6 +120,34 @@ uint32_t EmitMinMaxF32Value(EmitterState& state, uint32_t lhs, uint32_t rhs, boo
 	return EmitBitcastU32ToF32(state, result_bits);
 }
 
+uint32_t EmitMinMaxF32Value(EmitterState& state, uint32_t lhs, uint32_t rhs, bool max_value) {
+	const auto lhs_class = EmitClassifyF32(state, lhs);
+	const auto rhs_class = EmitClassifyF32(state, rhs);
+	return EmitMinMaxF32(state, lhs, rhs, lhs_class, rhs_class, max_value);
+}
+
+uint32_t EmitFPMedTri32(EmitterState& state, uint32_t a, uint32_t b, uint32_t c) {
+	const auto a_class = EmitClassifyF32(state, a);
+	const auto b_class = EmitClassifyF32(state, b);
+	const auto c_class = EmitClassifyF32(state, c);
+	const auto min_ab = EmitMinMaxF32(state, a, b, a_class, b_class, false);
+	const auto min_ab_class = EmitClassifyF32(state, min_ab);
+	const auto min3 = EmitMinMaxF32(state, min_ab, c, min_ab_class, c_class, false);
+	const auto max_ab = EmitMinMaxF32(state, a, b, a_class, b_class, true);
+	const auto max_ab_class = EmitClassifyF32(state, max_ab);
+	const auto max3 = EmitMinMaxF32(state, max_ab, c, max_ab_class, c_class, true);
+	const auto max_bc = EmitMinMaxF32(state, b, c, b_class, c_class, true);
+	const auto max_ac = EmitMinMaxF32(state, a, c, a_class, c_class, true);
+	// RDNA2 compares numerically against MAX3, including ties between signed zeros.
+	const auto a_is_max = Binary(state, spv::OpFOrdEqual, TypeBool(state), max3, a);
+	const auto b_is_max = Binary(state, spv::OpFOrdEqual, TypeBool(state), max3, b);
+	const auto median = Select(state, TypeF32(state), a_is_max, max_bc,
+	                           Select(state, TypeF32(state), b_is_max, max_ac, max_ab));
+	const auto nan_ab = EmitLogicalOrBool(state, a_class.nan, b_class.nan);
+	const auto any_nan = EmitLogicalOrBool(state, nan_ab, c_class.nan);
+	return Select(state, TypeF32(state), any_nan, min3, median);
+}
+
 uint32_t EmitFlushF32DenormToSignedZero(EmitterState& state, uint32_t value) {
 	const auto bits      = state.builder.AllocateId();
 	const auto abs_bits  = state.builder.AllocateId();

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

@@ -655,6 +655,14 @@ void EmitStoreCompletion(EmitterState& state) {
 	                          spv::MemorySemanticsImageMemoryMask));
 }
 
+void EmitDataCacheInvalidate(EmitterState& state) {
+	// RDNA2 GL0/GL1 invalidation makes vector memory visible without a workgroup rendezvous.
+	state.builder.AddFunction(
+	    spv::OpMemoryBarrier, ConstantU32(state, spv::ScopeDevice),
+	    ConstantU32(state, spv::MemorySemanticsAcquireMask | spv::MemorySemanticsUniformMemoryMask |
+	                          spv::MemorySemanticsImageMemoryMask));
+}
+
 void EmitBarrier(EmitterState& state) {
 	const auto tessellation = state.program.stage == ShaderType::TessellationControl;
 	if (!tessellation && ShaderWorkgroupInput(state.program.stage, state.input_info) == nullptr) {

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

@@ -97,10 +97,14 @@ uint32_t GatherMip(ValueEmitContext& ctx, const IR::Inst& inst, const IR::Memory
 		              ConstantF32Value(state, 1.0f / 256.0f));
 	};
 	const auto sampler_control = ctx.Arg(inst, 6);
+	const auto mip_filter = Extract(sampler_control, 26, 2);
 	const auto preclamp = Binary(state, spv::OpINotEqual, TypeBool(state),
 	                            Extract(sampler_control, 28, 1), ConstantU32(state, 0));
-	const auto half = ConstantF32Value(state, 0.5f);
 	const auto zero = ZeroF32(state);
+	// Point selects the nearest mip; linear gathers return the lower, primary mip only.
+	const auto rounding = Select(state, TypeF32(state),
+	    Binary(state, spv::OpIEqual, TypeBool(state), mip_filter, ConstantU32(state, 1)),
+	    ConstantF32Value(state, 0.5f), zero);
 	const auto base = Unary(state, spv::OpConvertUToF, TypeF32(state),
 	                        Extract(ctx.Arg(inst, 4), 12, 4));
 	const auto minimum = Binary(state, spv::OpFSub, TypeF32(state),
@@ -108,19 +112,19 @@ uint32_t GatherMip(ValueEmitContext& ctx, const IR::Inst& inst, const IR::Memory
 	const auto last = ConstantF32Value(state, static_cast<float>(mip_count - 1u));
 	auto lod = Binary(state, spv::OpFAdd, TypeF32(state),
 	                  AddressF32(ctx, mem, address, layout.lod),
-	                  Select(state, TypeF32(state), preclamp, half, zero));
+	                  Select(state, TypeF32(state), preclamp, rounding, zero));
 	lod = EmitGlsl<GLSLstd450FClamp, IR::Type::F32>(
 	    state, lod, LodFixed(ctx.Arg(inst, 5), 0), LodFixed(ctx.Arg(inst, 5), 12));
 	// T# minimum LOD is in physical mip space and applies after the S# view-relative clamp.
 	lod = EmitGlsl<GLSLstd450FMax, IR::Type::F32>(state, lod, minimum);
 	lod = EmitGlsl<GLSLstd450FClamp, IR::Type::F32>(state, lod, zero, last);
 	lod = Binary(state, spv::OpFAdd, TypeF32(state), lod,
-	             Select(state, TypeF32(state), preclamp, zero, half));
+	             Select(state, TypeF32(state), preclamp, zero, rounding));
 	// The clamped value is nonnegative, so conversion performs the integer mip rounding.
 	const auto mip = EmitUMin32(state, Unary(state, spv::OpConvertFToU, TypeU32(state), lod),
 	                           ConstantU32(state, mip_count - 1u));
 	const auto mip_none = Binary(state, spv::OpIEqual, TypeBool(state),
-	                            Extract(sampler_control, 26, 2), ConstantU32(state, 0));
+	                            mip_filter, ConstantU32(state, 0));
 	return Select(state, TypeU32(state), mip_none, ConstantU32(state, 0), mip);
 }
 
@@ -284,10 +288,10 @@ uint32_t ResultVector(ValueEmitContext& ctx, uint32_t value,
 }
 
 uint32_t QueryDimensions(ValueEmitContext& ctx, const IR::MemoryInfo& mem,
-                         const IR::Inst& address) {
-	const auto  dimension = ctx.state.program.info.images.at(mem.resource).dimension;
+                         const IR::Inst& address, uint32_t resource, uint32_t array_index) {
+	const auto  dimension = ctx.state.program.info.images.at(resource).dimension;
 	const auto& info      = ImageDimensionInfoFor(dimension);
-	const auto  image     = LoadImageDescriptor(ctx.state, mem.resource);
+	const auto  image     = LoadImageDescriptor(ctx.state, resource, 0u, array_index);
 	const auto  size      = ctx.state.builder.AllocateId();
 	if (info.multisampled != 0u) {
 		ctx.state.builder.AddFunction(spv::OpImageQuerySize,
@@ -689,68 +693,58 @@ auto EmitImageAccess(ValueEmitContext& ctx, const IR::Inst& inst, uint32_t resul
 	}
 	const auto selected = EmitIndirectResourceIndex(
 	    state, key, image.indirect_mapping_offset, image.indirect_search_iterations, 0u);
-	struct ImageRun {
-		uint32_t first;
-		uint32_t count;
-		uint32_t resource;
-		uint32_t slot_bias;
+	const auto SameType = [&](uint32_t lhs, uint32_t rhs) {
+		const auto& first = state.program.info.images[lhs];
+		const auto& second = state.program.info.images[rhs];
+		return first.dimension == second.dimension && first.cube == second.cube &&
+		       IR::DescriptorBindingForImage(first) == IR::DescriptorBindingForImage(second) &&
+		       (inst.GetOpcode() != IR::ValueOpcode::ImageWrite ||
+		        first.shader_swizzle == second.shader_swizzle) &&
+		       first.mip_count == 1u && second.mip_count == 1u;
 	};
-	std::vector<ImageRun> runs;
-	for (uint32_t ordinal = 0; ordinal < image.indirect_resources.size(); ++ordinal) {
-		const auto resource = image.indirect_resources[ordinal];
-		const auto& candidate = state.program.info.images[resource];
-		const auto kind = *IR::DescriptorBindingForImage(candidate);
-		if (!runs.empty()) {
-			auto& run = runs.back();
-			const auto& first = state.program.info.images[run.resource];
-			const auto next_slot = run.slot_bias + ordinal;
-			if (candidate.dimension == first.dimension && candidate.cube == first.cube &&
-			    IR::DescriptorBindingForImage(first) == kind &&
-			    (inst.GetOpcode() != IR::ValueOpcode::ImageWrite ||
-			     candidate.shader_swizzle == first.shader_swizzle) &&
-			    candidate.mip_count == 1u && first.mip_count == 1u &&
-			    (ordinal == 1u || candidate.descriptor_index == next_slot)) {
-				// Native roots precede appended children; only the root may have a slot gap.
-				if (ordinal == 1u)
-					run.slot_bias = candidate.descriptor_index - ordinal;
-				++run.count;
-				continue;
-			}
-		}
-		runs.push_back({ordinal, 1u, resource,
-		                candidate.descriptor_index - ordinal});
+	// Descriptor order is unrelated to image type. Share each typed access and
+	// select its descriptor slot independently, including gaps between native roots.
+	struct ImageGroup { uint32_t resource; uint32_t count; };
+	std::vector<ImageGroup> groups;
+	std::vector<uint32_t> ordinal_groups;
+	for (const auto resource: image.indirect_resources) {
+		const auto found = std::ranges::find_if(groups, [&](const ImageGroup& group) {
+			return SameType(resource, group.resource);
+		});
+		ordinal_groups.push_back(static_cast<uint32_t>(found - groups.begin()));
+		if (found == groups.end()) groups.push_back({resource, 1u});
+		else ++found->count;
 	}
-	const auto EmitRun = [&](uint32_t index) {
-		const auto& run = runs[index];
-		if (run.count == 1u) return emit(run.resource, 0u);
-		auto slot = Binary(state, spv::OpIAdd, TypeU32(state), selected,
-		                   ConstantU32(state, run.slot_bias));
-		if (run.first == 0u) {
-			const auto root_slot = image.descriptor_index;
-			if (root_slot != run.slot_bias) {
-				const auto is_root = Binary(state, spv::OpIEqual, TypeBool(state), selected,
-				                            ConstantU32(state, 0u));
-				const auto root_or_child = state.builder.AllocateId();
-				state.builder.AddFunction(spv::OpSelect, TypeU32(state), root_or_child,
-				                          is_root, ConstantU32(state, root_slot), slot);
-				slot = root_or_child;
-			}
-		}
-		return emit(run.resource, slot);
-	};
-	auto run_index = ConstantU32(state, 0u);
-	for (uint32_t index = 1; index < runs.size(); ++index) {
+	const bool dynamic_slot = std::ranges::any_of(groups, [](const ImageGroup& group) {
+		return group.count > 1u;
+	});
+	auto group_index = ConstantU32(state, 0u);
+	auto slot_bias = ConstantU32(state, image.descriptor_index);
+	auto previous_bias = image.descriptor_index;
+	for (uint32_t ordinal = 1; ordinal < image.indirect_resources.size(); ++ordinal) {
+		const auto& candidate = state.program.info.images[image.indirect_resources[ordinal]];
+		const auto bias = candidate.descriptor_index - ordinal;
+		const bool select_group = ordinal_groups[ordinal] != ordinal_groups[ordinal - 1u];
+		const bool select_slot = dynamic_slot && bias != previous_bias;
+		previous_bias = bias;
+		if (!select_group && !select_slot) continue;
 		const auto at_or_after = Binary(state, spv::OpUGreaterThanEqual, TypeBool(state),
-		                               selected, ConstantU32(state, runs[index].first));
-		const auto next = state.builder.AllocateId();
-		state.builder.AddFunction(spv::OpSelect, TypeU32(state), next, at_or_after,
-		                          ConstantU32(state, index), run_index);
-		run_index = next;
+		                               selected, ConstantU32(state, ordinal));
+		if (select_group) group_index = Select(state, TypeU32(state), at_or_after,
+		                                      ConstantU32(state, ordinal_groups[ordinal]), group_index);
+		if (select_slot) slot_bias = Select(state, TypeU32(state), at_or_after,
+		                                   ConstantU32(state, bias), slot_bias);
 	}
-	return runs.size() == 1u
-	           ? EmitRun(0u)
-	           : EmitIndexSwitch(state, run_index, static_cast<uint32_t>(runs.size()),
-	                             result_type, EmitRun);
+	const auto slot = dynamic_slot
+	    ? Binary(state, spv::OpIAdd, TypeU32(state), selected, slot_bias) : 0u;
+	const auto EmitGroup = [&](uint32_t index) {
+		const auto& group = groups[index];
+		return emit(group.resource, group.count == 1u ? 0u : slot);
+	};
+	return groups.size() == 1u
+	           ? EmitGroup(0u)
+	           : EmitIndexSwitch(state, group_index, static_cast<uint32_t>(groups.size()),
+	                             result_type, EmitGroup);
 }
 
 template <typename Emit>
@@ -852,7 +846,10 @@ void EmitImage(ValueEmitContext& ctx, const IR::Inst& inst) {
 	const auto* address = ctx.ImageAddress(inst.Arg(image_info.needs_sampler ? 2 : 1));
 	if (op == IR::ValueOpcode::ImageQueryDimensions) {
 		state.builder.RequireCapability(spv::CapabilityImageQuery);
-		ctx.Define(inst, QueryDimensions(ctx, mem, *address));
+		ctx.Define(inst, EmitImageAccess(ctx, inst, TypeU32Vector(state, 4),
+		    [&](uint32_t resource, uint32_t array_index) {
+			    return QueryDimensions(ctx, mem, *address, resource, array_index);
+		    }));
 		return;
 	}
 	if (image_info.needs_sampler &&

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

@@ -218,6 +218,7 @@ inline constexpr auto EmitTtraceData   = EmitVoid;
 inline constexpr auto EmitInstPrefetch = EmitVoid;
 void                  EmitBarrier(EmitterState& state);
 void                  EmitStoreCompletion(EmitterState& state);
+void                  EmitDataCacheInvalidate(EmitterState& state);
 void                  EmitMeshAllocate(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitMeshDrawParameter(ValueEmitContext& ctx, const IR::Inst& inst);
 uint32_t              EmitGetTessellationAttribute(ValueEmitContext& ctx, const IR::Inst& inst);
@@ -296,6 +297,7 @@ inline constexpr auto EmitBufferAtomicSMax32    = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicUMax32    = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicAnd32     = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicAnd64     = EmitBufferAtomic64;
+inline constexpr auto EmitBufferAtomicIAdd64    = EmitBufferAtomic64;
 inline constexpr auto EmitBufferAtomicOr32      = EmitAtomic32;
 inline constexpr auto EmitBufferAtomicOr64      = EmitBufferAtomic64;
 inline constexpr auto EmitBufferAtomicXor32     = EmitAtomic32;
@@ -325,6 +327,7 @@ inline constexpr auto EmitSharedAtomicSMin32    = EmitAtomic32;
 inline constexpr auto EmitSharedAtomicUMin32    = EmitAtomic32;
 inline constexpr auto EmitSharedAtomicSMax32    = EmitAtomic32;
 inline constexpr auto EmitSharedAtomicUMax64    = EmitSharedAtomic64;
+inline constexpr auto EmitSharedAtomicCmpSwap32 = EmitAtomic32;
 inline constexpr auto EmitSharedAtomicCmpSwap64 = EmitSharedAtomic64;
 inline constexpr auto EmitSharedAtomicUMax32    = EmitAtomic32;
 inline constexpr auto EmitSharedAtomicAnd32     = EmitAtomic32;

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

@@ -867,7 +867,7 @@ void StoreFormattedPrepared(ValueEmitContext& ctx, const IR::Inst& inst,
 	});
 }
 
-uint32_t LoadScalarBuffer(ValueEmitContext& ctx, const IR::Inst& inst) {
+uint32_t LoadIndirectScalarBuffer(ValueEmitContext& ctx, const IR::Inst& inst) {
 	auto&       state  = ctx.state;
 	const auto& handle = *inst.Arg(0).ResolveInstruction();
 	const auto  word1  = ctx.Arg(handle, 1);
@@ -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);
 		    });
 	    });
 }
@@ -1545,8 +1597,29 @@ uint32_t EmitReadConst(ValueEmitContext& ctx, const IR::Inst& inst) {
 }
 
 void EmitReadConstBuffer(ValueEmitContext& ctx, const IR::Inst& inst) {
-	if (ctx.Memory(inst).planning_only) return;
-	ctx.Define(inst, LoadScalarBuffer(ctx, inst));
+	auto mem = ctx.Memory(inst);
+	if (mem.planning_only) return;
+	if (mem.kind == IR::ResourceKind::IndirectBuffer) {
+		ctx.Define(inst, LoadIndirectScalarBuffer(ctx, inst));
+		return;
+	}
+	auto& state        = ctx.state;
+	mem.kind           = IR::ResourceKind::ScalarBuffer;
+	auto index = Binary(state, spv::OpShiftRightLogical, TypeU32(state), ctx.Arg(inst, 1),
+	                    ConstantU32(state, 2));
+	if (mem.offset >= 4u) {
+		index = Binary(state, spv::OpIAdd, TypeU32(state), index, ConstantU32(state, mem.offset >> 2u));
+	}
+	const auto access    = PrepareMemoryResourceAccess(state, mem);
+	// Scalar buffer addressing aligns the base and each offset independently.
+	const auto element = Binary(
+	    state, spv::OpIAdd, TypeU32(state), index,
+	    Binary(state, spv::OpShiftRightLogical, TypeU32(state), access.byte_offset, ConstantU32(state, 2)));
+	const auto condition = EmitMemoryElementInBounds(state, access, element);
+	ctx.Define(inst, EmitValueOrZeroIfCondition(state, condition, [&]() {
+		           return EmitNative<spv::OpLoad, IR::Type::U32>(
+		               state, EmitMemoryElementPointer(state, access, element));
+	           }));
 }
 
 uint32_t EmitGetBufferResourceIndex(ValueEmitContext& ctx, const IR::Inst& inst) {

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);
@@ -269,8 +270,10 @@ MemoryResourceAccess PrepareMemoryResourceAccess(EmitterState& state, const IR::
 		case IR::ResourceKind::Flat:
 		case IR::ResourceKind::Global:
 			EXIT("physical address memory must use the BDA emitter\n");
+		case IR::ResourceKind::ScalarBuffer:
 		case IR::ResourceKind::Buffer: {
-			const auto bits = IR::StorageBufferElementBits(state.program, mem);
+			const auto bits = mem.kind == IR::ResourceKind::Buffer
+			                      ? IR::StorageBufferElementBits(state.program, mem) : 32u;
 			auto* buffer = &state.storage_buffers[2];
 			if (bits == 8u) {
 				buffer = &state.storage_buffers[0];
@@ -502,6 +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;
@@ -514,6 +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;
@@ -558,8 +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/backend/spirv/spirvEmitterProgram.cpp

@@ -307,6 +307,7 @@ void EmitBlock(ValueEmitContext& ctx, const IR::Block* block, EmitInstruction&&
 			ctx.state.lane_half = half;
 			if (half == 0 || (inst.GetOpcode() != IR::ValueOpcode::Barrier &&
 			                  inst.GetOpcode() != IR::ValueOpcode::StoreCompletion &&
+			                  inst.GetOpcode() != IR::ValueOpcode::DataCacheInvalidate &&
 			                  inst.GetOpcode() != IR::ValueOpcode::MeshAllocate)) {
 				emit_instruction(lane, inst);
 			}

src/graphics/shader/recompiler/frontend/cfg/ShaderCFG.cpp

@@ -8,6 +8,8 @@
 #include <deque>
 #include <list>
 #include <map>
+#include <optional>
+#include <set>
 #include <span>
 
 namespace Libs::Graphics::ShaderRecompiler::CFG {
@@ -1650,6 +1652,32 @@ Graph BuildGraph(const Decoder::Program& program) {
 	}
 	SortUnique(graph.code_table_load_pcs);
 
+	if (!indirect_setpc && std::ranges::any_of(program.instructions, [](const auto& inst) {
+		    return inst.opcode == Opcode::S_SWAPPC_B64 && inst.branch_target != UINT32_MAX;
+		})) {
+		// Linked leaf bodies follow the caller in native PC order. Visit them in
+		// execution order so returning to the caller does not manufacture goto loops.
+		std::vector<bool> visited(graph.blocks.size(), false);
+		std::vector<std::pair<uint32_t, bool>> pending {{graph.entry_block, false}};
+		std::vector<BasicBlock> ordered;
+		ordered.reserve(graph.blocks.size());
+		while (!pending.empty()) {
+			const auto [id, finished] = pending.back();
+			pending.pop_back();
+			if (finished) {
+				ordered.push_back(std::move(graph.blocks[id]));
+				continue;
+			}
+			if (visited[id]) continue;
+			visited[id] = true;
+			pending.emplace_back(id, true);
+			for (const auto successor: graph.blocks[id].successors)
+				pending.emplace_back(successor, false);
+		}
+		std::ranges::reverse(ordered);
+		ApplyBlockOrder(graph, std::move(ordered));
+	}
+
 	RecomputeAnalyses(graph);
 
 	if (indirect_setpc) {
@@ -1681,33 +1709,56 @@ Graph Structurize(const Graph& graph) {
 }
 
 bool MayWriteScalarRegister(const Decoder::Instruction& inst, uint32_t code) {
-	uint32_t destination = UINT32_MAX;
-	// Scalar ALU pair widths are not all decoded. The typed producer must confirm this word.
-	return InstructionWritesScalarCode(inst, code) ||
-	       (ScalarOperandCode(inst.dst, destination) && code == destination + 1u);
+	// Vector comparisons and carry results can write a scalar pair in wave64.
+	if (inst.family == Decoder::Family::VOP1 || inst.family == Decoder::Family::VOP2 ||
+	    inst.family == Decoder::Family::VOPC || inst.family == Decoder::Family::VOP3 ||
+	    inst.family == Decoder::Family::VOP3P) {
+		uint32_t destination = 0;
+		for (const auto* operand: {&inst.dst, &inst.dst2})
+			if (ScalarOperandCode(*operand, destination) && code >= destination && code <= destination + 1u)
+				return true;
+	}
+	return InstructionWritesScalarCode(inst, code);
 }
 
 uint32_t FindScalarDefinition(const Decoder::Program& program, const Graph& graph,
-                              uint32_t before, uint32_t code) {
+                              uint32_t before, uint32_t code, std::vector<uint32_t>& crossed_calls) {
 	EXIT_IF(before >= program.instructions.size());
 	const auto found = std::ranges::find_if(graph.blocks, [before](const auto& block) {
 		return block.inst_begin <= before && before < block.inst_end;
 	});
 	EXIT_IF(found == graph.blocks.end());
-	const auto* block = &*found;
-	for (size_t depth = 0; depth < graph.blocks.size(); ++depth) {
-		for (uint32_t i = before; i > block->inst_begin;) {
-			const auto& inst = program.instructions[--i];
-			EXIT_IF(inst.opcode == Opcode::S_SWAPPC_B64);
-			if (MayWriteScalarRegister(inst, code)) return i;
-		}
-		if (block->id == graph.entry_block) return UINT32_MAX;
-		EXIT_IF(block->predecessors.size() != 1u);
-		block = graph.FindBlock(block->predecessors[0]);
+	std::vector<std::pair<uint32_t, uint32_t>> pending {{found->id, before}};
+	std::set<std::pair<uint32_t, uint32_t>> visited;
+	std::optional<uint32_t> definition;
+	while (!pending.empty()) {
+		const auto point = pending.back();
+		pending.pop_back();
+		if (!visited.insert(point).second) continue;
+		const auto* block = graph.FindBlock(point.first);
 		EXIT_IF(block == nullptr);
-		before = block->inst_end;
+		uint32_t index = UINT32_MAX;
+		for (uint32_t i = point.second; i > block->inst_begin;) {
+			const auto& inst = program.instructions[--i];
+			if (inst.opcode == Opcode::S_SWAPPC_B64) {
+				EXIT_NOT_IMPLEMENTED(MayWriteScalarRegister(inst, code));
+				crossed_calls.push_back(i);
+				continue;
+			}
+			if (MayWriteScalarRegister(inst, code)) { index = i; break; }
+		}
+		if (index != UINT32_MAX || block->id == graph.entry_block) {
+			EXIT_IF(definition && *definition != index);
+			definition = index;
+			continue;
+		}
+		EXIT_IF(block->predecessors.empty());
+		for (const auto predecessor: block->predecessors) {
+			pending.emplace_back(predecessor, graph.blocks[predecessor].inst_end);
+		}
 	}
-	EXIT("scalar shader call source has cyclic reaching definitions");
+	EXIT_IF(!definition);
+	return *definition;
 }
 
 std::string BranchConditionToString(BranchCondition condition) {

src/graphics/shader/recompiler/frontend/cfg/ShaderCFG.h

@@ -115,7 +115,7 @@ struct Graph {
 Graph       BuildGraph(const Decoder::Program& program);
 bool MayWriteScalarRegister(const Decoder::Instruction& inst, uint32_t code);
 uint32_t FindScalarDefinition(const Decoder::Program& program, const Graph& graph,
-                              uint32_t before, uint32_t code);
+                              uint32_t before, uint32_t code, std::vector<uint32_t>& crossed_calls);
 // Returns structured control flow or failure diagnostics without changing the native graph.
 // On failure, failure_block is an original block ID or UINT32_MAX.
 Graph       Structurize(const Graph& graph);

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

@@ -67,8 +67,11 @@ constexpr MemoryOpcodeInfo MUBUF_OPCODE_LIST[] = {
     {0x3fu, Opcode::BUFFER_ATOMIC_FMIN, 1, 32},
     {0x40u, Opcode::BUFFER_ATOMIC_FMAX, 1, 32},
     {0x50u, Opcode::BUFFER_ATOMIC_SWAP_X2, 2, 32},
+    {0x52u, Opcode::BUFFER_ATOMIC_ADD_X2, 2, 32},
     {0x59u, Opcode::BUFFER_ATOMIC_AND_X2, 2, 32},
     {0x5au, Opcode::BUFFER_ATOMIC_OR_X2, 2, 32},
+    {0x71u, Opcode::BUFFER_GL0_INV, 0, 0},
+    {0x72u, Opcode::BUFFER_GL1_INV, 0, 0},
     {0x80u, Opcode::BUFFER_LOAD_FORMAT_D16_X, 1, 16, false, false, true},
     {0x83u, Opcode::BUFFER_LOAD_FORMAT_D16_XYZW, 2, 16, false, false, true},
     {0x84u, Opcode::BUFFER_STORE_FORMAT_D16_X, 1, 16, false, false, true},
@@ -114,6 +117,7 @@ constexpr MemoryOpcodeInfo DS_OPCODE_LIST[] = {
     {0x27u, Opcode::DS_MIN_RTN_U32, 1, 32},      {0x28u, Opcode::DS_MAX_RTN_U32, 1, 32},
     {0x29u, Opcode::DS_AND_RTN_B32, 1, 32},      {0x2au, Opcode::DS_OR_RTN_B32, 1, 32},
     {0x2bu, Opcode::DS_XOR_RTN_B32, 1, 32},      {0x2du, Opcode::DS_WRXCHG_RTN_B32, 1, 32},
+    {0x30u, Opcode::DS_CMPST_RTN_B32, 1, 32},
     {0x35u, Opcode::DS_SWIZZLE_B32, 1, 32},      {0x36u, Opcode::DS_READ_B32, 1, 32},
     {0x37u, Opcode::DS_READ2_B32, 2, 32},        {0x38u, Opcode::DS_READ2ST64_B32, 2, 32},
     {0x39u, Opcode::DS_READ_I8, 1, 8, true},     {0x3au, Opcode::DS_READ_U8, 1, 8},
@@ -200,6 +204,7 @@ bool IsDsAtomicOpcode(Opcode opcode) {
 		case Opcode::DS_MAX_U32:
 		case Opcode::DS_MAX_RTN_U32:
 		case Opcode::DS_MAX_RTN_U64:
+		case Opcode::DS_CMPST_RTN_B32:
 		case Opcode::DS_CMPST_RTN_B64:
 		case Opcode::DS_AND_B32:
 		case Opcode::DS_AND_RTN_B32:
@@ -220,6 +225,7 @@ uint32_t DsSourceCount(Opcode opcode) {
 		case Opcode::DS_WRITE2ST64_B32:
 		case Opcode::DS_WRITE2_B64:
 		case Opcode::DS_WRITE2ST64_B64:
+		case Opcode::DS_CMPST_RTN_B32:
 		case Opcode::DS_CMPST_RTN_B64:
 		case Opcode::DS_MSKOR_B32: return 3u;
 		case Opcode::DS_MIN_F32:
@@ -308,6 +314,9 @@ void DecodeMubuf(uint32_t pc, std::span<const uint32_t> code, uint32_t word_inde
 	if (inst.opcode == Opcode::UNSUPPORTED) {
 		SetUnsupported(inst, Family::MUBUF, opcode, "MUBUF opcode is not implemented");
 	}
+	if (inst.opcode == Opcode::BUFFER_GL0_INV || inst.opcode == Opcode::BUFFER_GL1_INV) {
+		return;
+	}
 
 	DecodeVectorGpr(vdata, inst.dst);
 	if (inst.opcode == Opcode::BUFFER_LOAD_FORMAT_D16_XYZW ||
@@ -454,6 +463,9 @@ void DecodeDs(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index,
 	if (inst.opcode == Opcode::DS_SWIZZLE_B32 && inst.offset >= 0xe000u) {
 		SetUnsupported(inst, Family::DS, opcode, "DS swizzle FFT mode is not implemented");
 	}
+	if (inst.gds && inst.opcode == Opcode::DS_CMPST_RTN_B32) {
+		SetUnsupported(inst, Family::DS, opcode, "DS compare-store return B32 GDS variant is not implemented");
+	}
 	if (inst.gds &&
 	    (inst.opcode == Opcode::DS_SWIZZLE_B32 || inst.opcode == Opcode::DS_PERMUTE_B32 ||
 	     inst.opcode == Opcode::DS_BPERMUTE_B32 ||

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

@@ -164,9 +164,23 @@ void DecodeSop1(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 	switch (inst.opcode) {
 		case Opcode::S_SWAPPC_B64:
 			inst.branch_target = UINT32_MAX;
-			inst.data_dwords = 2;
+			[[fallthrough]];
+		case Opcode::S_MOV_B64:
+		case Opcode::S_CMOV_B64:
+		case Opcode::S_NOT_B64:
+		case Opcode::S_WQM_B64:
+		case Opcode::S_BREV_B64:
+		case Opcode::S_BITSET0_B64:
+		case Opcode::S_BITSET1_B64:
+		case Opcode::S_AND_SAVEEXEC_B64:
+		case Opcode::S_ORN2_SAVEEXEC_B64:
+		case Opcode::S_QUADMASK_B64:
+		case Opcode::S_ANDN1_SAVEEXEC_B64:
+		case Opcode::S_BITREPLICATE_B64_B32:
+			inst.data_dwords = 2u;
 			break;
 		case Opcode::S_GETPC_B64:
+			inst.data_dwords = 2u;
 			inst.src_count = 0;
 			DecodeScalarDestination(sdst, pc, inst.dst);
 			return;
@@ -206,6 +220,8 @@ void DecodeSop2(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index
 
 	DecodeScalarDestination(sdst, pc, inst.dst);
 	DecodeBinarySources(pc, code, word_index, inst, ssrc0, ssrc1);
+	if (opcode == 0x0bu || (opcode >= 0x0fu && opcode <= 0x25u && (opcode & 1u)) || opcode == 0x29u)
+		inst.data_dwords = 2u;
 }
 
 void DecodeSopk(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index,

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

@@ -566,6 +566,9 @@ std::string InstructionToString(const Instruction& inst) {
 			                                               OperandToString(inst.src0).c_str()));
 		case Opcode::S_BARRIER:
 			return WithUnsupportedReason(inst, fmt::format("0x{:08x}: s_barrier", inst.pc));
+		case Opcode::BUFFER_GL0_INV:
+		case Opcode::BUFFER_GL1_INV:
+			return fmt::format("0x{:08x}: {}", inst.pc, magic_enum::enum_name(inst.opcode));
 		case Opcode::V_NOP:
 			return WithUnsupportedReason(inst, fmt::format("0x{:08x}: v_nop", inst.pc));
 		case Opcode::S_ENDPGM:
@@ -670,6 +673,7 @@ 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 +721,7 @@ 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,
@@ -598,6 +599,8 @@ enum class Opcode {
 	BUFFER_ATOMIC_XOR,
 	BUFFER_ATOMIC_FMIN,
 	BUFFER_ATOMIC_FMAX,
+	BUFFER_GL0_INV,
+	BUFFER_GL1_INV,
 	FLAT_LOAD_UBYTE,
 	FLAT_LOAD_SBYTE,
 	FLAT_LOAD_USHORT,
@@ -631,6 +634,7 @@ 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

@@ -940,6 +940,10 @@ void Translator::DS_PERMUTE(const Decoder::Instruction& inst, bool backward) {
 
 void Translator::EmitMemory(const Decoder::Instruction& inst) {
 	switch (inst.opcode) {
+		case Decoder::Opcode::BUFFER_GL0_INV:
+		case Decoder::Opcode::BUFFER_GL1_INV:
+			ir.Emit(IR::ValueOpcode::DataCacheInvalidate);
+			return;
 		case Decoder::Opcode::IMAGE_BVH_INTERSECT_RAY: {
 			const auto result = ir.Emit(IR::ValueOpcode::BvhIntersect,
 			    {ConstructU32x4(inst.src1, 4), MakeImageAddress(inst, inst.src0), ir.GetExec()},
@@ -1046,6 +1050,8 @@ 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 +1089,8 @@ 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;
@@ -128,6 +133,7 @@ struct BufferResource {
 	bool                   written            = false;
 	bool                   atomic             = false;
 	bool                   formatted          = false;
+	bool                   scalar             = false;
 	bool                   live               = false;
 	uint32_t               indirect_root              = NoIndirectBuffer;
 	uint32_t               indirect_mapping_offset    = 0;
@@ -489,7 +495,6 @@ struct ShaderInfo {
 	static constexpr uint32_t MaxBufferRoots  = 64;
 	static constexpr uint32_t MaxImages       = 64;
 	static constexpr uint32_t MaxSamplers     = 32;
-	static constexpr uint32_t MaxSampledPairs = 64;
 
 	std::vector<BufferResource>      buffers;
 	std::vector<ImageResource>       images;
@@ -621,6 +626,8 @@ 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

@@ -94,6 +94,7 @@ bool HasSideEffects(ValueOpcode opcode) {
 		case ValueOpcode::SetTessellationAttribute:
 		case ValueOpcode::MeshAllocate:
 		case ValueOpcode::StoreCompletion:
+		case ValueOpcode::DataCacheInvalidate:
 		case ValueOpcode::Barrier: return true;
 		default: return false;
 	}
@@ -118,6 +119,7 @@ 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 +139,7 @@ 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 +179,7 @@ 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)
@@ -348,6 +350,7 @@ VALUE_OPCODE(SetAttribute, Void, U32x4, U1)
 
 VALUE_OPCODE(ControlNop, Void)
 VALUE_OPCODE(StoreCompletion, Void)
+VALUE_OPCODE(DataCacheInvalidate, Void)
 VALUE_OPCODE(Barrier, Void)
 VALUE_OPCODE(Sendmsg, Void)
 VALUE_OPCODE(TtraceData, Void)

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
@@ -1023,6 +1029,19 @@ void ConstantPropagationPass(const BlockList& blocks, uint32_t wave_size) {
 		}
 		drain();
 	} while (normalized);
+	// Repeated register reads share a bitcast only within the same SSA block.
+	std::unordered_map<const Inst*, Inst*> bitcasts;
+	for (auto* block: blocks) {
+		bitcasts.clear();
+		for (auto& inst: *block) {
+			if (inst.GetOpcode() != ValueOpcode::BitCastF32U32 &&
+			    inst.GetOpcode() != ValueOpcode::BitCastU32F32) continue;
+			const auto* source = Arg(inst, 0).TryInstruction();
+			if (source == nullptr) continue;
+			const auto [entry, inserted] = bitcasts.try_emplace(source, &inst);
+			if (!inserted) Replace(inst, Value(entry->second));
+		}
+	}
 }
 
 } // namespace Libs::Graphics::ShaderRecompiler::IR

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

@@ -30,32 +30,62 @@ void RemoveIdentities(Program& program) {
 }
 
 void EliminateDeadCode(const BlockList& blocks) {
-	std::vector<Inst*> pending;
-	for (auto* block: blocks) {
-		for (auto& inst: *block) {
-			// Retained resource-planning instructions live outside the block list.
-			inst.live = inst.MayHaveSideEffects() || std::ranges::any_of(inst.Uses(),
-			    [](const Use& use) { return use.user->Parent() == nullptr; });
-			if (inst.live) pending.push_back(&inst);
-		}
-	}
-	while (!pending.empty()) {
-		const auto* inst = pending.back();
-		pending.pop_back();
-		for (size_t arg = 0; arg < inst->NumArgs(); ++arg) {
-			auto* input = inst->Arg(arg).TryInstruction();
-			if (input != nullptr && input->Parent() != nullptr && !input->live) {
-				input->live = true;
-				pending.push_back(input);
+	bool changed;
+	do {
+		std::vector<Inst*> pending;
+		for (auto* block: blocks) {
+			for (auto& inst: *block) {
+				// Retained resource-planning instructions live outside the block list.
+				inst.live = inst.MayHaveSideEffects() || std::ranges::any_of(inst.Uses(),
+				    [](const Use& use) { return use.user->Parent() == nullptr; });
+				if (inst.live) pending.push_back(&inst);
 			}
 		}
-	}
-	// Unlink every dead operand before freeing nodes in a dead Phi cycle.
-	for (auto* block: blocks)
-		for (auto& inst: *block)
-			if (!inst.live) inst.Invalidate();
-	for (auto* block: blocks)
-		std::erase_if(block->Instructions(), [](const Inst& inst) { return !inst.live; });
+		while (!pending.empty()) {
+			const auto* inst = pending.back();
+			pending.pop_back();
+			for (size_t arg = 0; arg < inst->NumArgs(); ++arg) {
+				auto* input = inst->Arg(arg).TryInstruction();
+				if (input != nullptr && input->Parent() != nullptr && !input->live) {
+					input->live = true;
+					pending.push_back(input);
+				}
+			}
+		}
+		// Unlink every dead operand before freeing nodes in a dead Phi cycle.
+		for (auto* block: blocks)
+			for (auto& inst: *block)
+				if (!inst.live) inst.Invalidate();
+		for (auto* block: blocks)
+			std::erase_if(block->Instructions(), [](const Inst& inst) { return !inst.live; });
+
+		changed = false;
+		for (auto* block: blocks) {
+			const auto& term = block->terminator;
+			if (term.kind != CFG::TerminatorKind::ConditionalBranch || term.loop_header ||
+			    block->condition.IsImmediate()) continue;
+			for (const bool take_true: {false, true}) {
+				const auto* empty = take_true ? term.false_block : term.true_block;
+				const auto* merge = take_true ? term.true_block : term.false_block;
+				if (empty == nullptr || merge == nullptr || !empty->empty() ||
+				    empty->terminator.loop_header || merge->terminator.loop_header ||
+				    empty->terminator.kind != CFG::TerminatorKind::Branch ||
+				    empty->terminator.true_block != merge ||
+				    std::ranges::any_of(*merge, [](const Inst& inst) {
+					    return inst.GetOpcode() == ValueOpcode::Phi;
+				    })) continue;
+				// Keep the CFG and SSA edges intact; only the empty detour's predicate is dead.
+				for (auto& inst: *block) {
+					if (inst.GetOpcode() == ValueOpcode::Reference &&
+					    inst.Arg(0).Resolve() == block->condition.Resolve())
+						inst.SetArg(0, Value(take_true));
+				}
+				block->condition = Value(take_true);
+				changed = true;
+				break;
+			}
+		}
+	} while (changed);
 }
 
 } // namespace Libs::Graphics::ShaderRecompiler::IR

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

@@ -292,11 +292,23 @@ bool NormalizeIndirectBuffer(DescriptorValue& value, bool indexed, const BufferR
 bool IsBoundedDescriptorTable(const ResourcePlan& program,
                               const DescriptorSource::IndirectDescriptor& indirect) {
 	const auto* table = Source(program, indirect.table_source);
-	return indirect.sources.empty() && !indirect.selector &&
+	return indirect.sources.empty() && (!indirect.selector || indirect.table_scalar) &&
 	       (indirect.key_count.IsEmpty() || indirect.table_maximum) && indirect.workgroup_axis == UINT32_MAX &&
 	       table != nullptr && table->dword_count == 4u;
 }
 
+bool CanReadIndexedSelector(const ShaderBufferResource& material, uint32_t offset) {
+	if (material.Type() != 0u || material.SwizzleEnabled() || material.AddTid() ||
+	    material.OutOfBounds() != 0u) return false;
+	if (material.Format() == Prospero::BufferFormat::kInvalid || material.NumRecords() == 0u ||
+	    offset >= material.Stride()) return true;
+	return material.Base48() != 0u &&
+	       ((material.Base48() | material.Stride() | offset) & 3u) == 0u &&
+	       material.NumRecords() <= MaxIndirectDescriptorProbes &&
+	       uint64_t {material.NumRecords() - 1u} * material.Stride() + offset + 4u <=
+	           uint64_t {UINT32_MAX} + 1u;
+}
+
 bool EvaluateIndirectCount(const DescriptorSource::IndirectDescriptor& indirect,
                            const SrtRuntime& runtime, SrtWalker& clean, uint32_t& count) {
 	if (indirect.workgroup_axis != UINT32_MAX) {
@@ -490,6 +502,15 @@ bool ResolveSelectorKeys(const DescriptorSource::IndirectDescriptor& indirect,
 			keys.push_back(key);
 			mask &= mask - 1u;
 		}
+	} else if (selector->stride == 0u) {
+		ShaderBufferResource material;
+		if (!DecodeBufferDescriptor(material_value, material) ||
+		    !CanReadIndexedSelector(material, selector->offset)) return false;
+		if (material.Format() != Prospero::BufferFormat::kInvalid && material.NumRecords() != 0u &&
+		    selector->offset < material.Stride() &&
+		    !read_keys(material.Base48(), material.GetSize(), selector->offset,
+		               material.Stride(), material.NumRecords())) return false;
+		keys.push_back(0u); // Every out-of-bounds indexed load returns zero.
 	} else {
 		ShaderBufferResource material;
 		if (!DecodeBufferDescriptor(material_value, material) || selector->stride < 4u ||
@@ -524,29 +545,29 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
                                    ResourceSnapshot&             snapshot,
                                    std::vector<DescriptorValue>& descriptors,
                                    std::vector<Specialization>&  specializations,
-                                   uint32_t maximum_resources, Normalize&& normalize) {
+                                   Normalize&& normalize) {
 	const auto& indirect = *program.descriptor_sources[source_index].indirect_descriptor;
 	const auto  descriptor_bytes = read_dwords * sizeof(uint32_t);
 	const auto& sources = indirect.sources;
 	const auto* selector = indirect.selector ? &*indirect.selector : nullptr;
+	const bool byte_keys = IsBoundedDescriptorTable(program, indirect);
 	auto* selector_keys = selector != nullptr
 	    ? &program.selector_keys[program.selector_groups[source_index]] : nullptr;
 	auto& keys = selector_keys != nullptr ? selector_keys->values : program.material_keys;
+	auto& byte_offsets = program.material_keys;
 	const auto children_begin = descriptors.size();
 	const auto root_resource = specializations[resource_index];
-	const auto intern_candidate = [&](const DescriptorValue& candidate, uint32_t& ordinal) {
-		ordinal = 0u;
-		if (candidate == descriptors[resource_index]) return true;
+	const auto intern_candidate = [&](const DescriptorValue& candidate) {
+		if (candidate == descriptors[resource_index]) return 0u;
 		const auto found = std::find(descriptors.begin() + children_begin, descriptors.end(), candidate);
-		ordinal = static_cast<uint32_t>(found - descriptors.begin() - children_begin + 1u);
+		const auto ordinal = static_cast<uint32_t>(found - descriptors.begin() - children_begin + 1u);
 		if (found == descriptors.end()) {
-			if (descriptors.size() >= maximum_resources) return false;
 			descriptors.push_back(candidate);
 			auto child = root_resource;
 			child.indirect_root = resource_index;
 			specializations.push_back(child);
 		}
-		return true;
+		return ordinal;
 	};
 	uint64_t table_base = 0;
 	uint64_t table_size = UINT64_MAX; // Scalar addresses have no buffer descriptor bounds.
@@ -558,7 +579,7 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 		    "indirect {} resource={} source={} table_source={} words={:08x},{:08x},{:08x},{:08x} "
 		    "base={:#x} size={} type={} runtime_stride={} records={} swizzle={} addtid={} oob={} "
 		    "maximum={} offset={} immediate={} record_bytes={} child_count={} child_stride={} "
-		    "child_offset={} indexed_pointer={} candidates={} capacity={} {}",
+		    "child_offset={} indexed_pointer={} candidates={} {}",
 		    reason, resource_index, source_index, indirect.table_source,
 		    table_value.dwords[0], table_value.dwords[1], table_value.dwords[2], table_value.dwords[3],
 		    table_base, table_size, table.Type(), table.Stride(), table.NumRecords(),
@@ -567,7 +588,7 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 		    indirect.table_offset, indirect.table_immediate, indirect.table_record_bytes,
 		    pointer != nullptr ? pointer->count : 1u, pointer != nullptr ? pointer->stride : 0u,
 		    pointer != nullptr ? pointer->offset : 0u, pointer != nullptr && pointer->indexed,
-		    descriptors.size(), maximum_resources, details));
+		    descriptors.size(), details));
 	};
 	const auto read_failed = [&](std::string_view reason, uint32_t row, uint32_t child,
 	                             uint64_t base, uint64_t offset, size_t bytes) {
@@ -583,7 +604,7 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 	if (sources.empty()) {
 		if (selector == nullptr) keys.clear();
 		DescriptorValue material_value;
-		if ((selector != nullptr && !selector_keys->resolved &&
+		if ((selector != nullptr && (byte_keys || !selector_keys->resolved) &&
 		     !clean.EvaluateDescriptor(selector->source, material_value)) ||
 		    !clean.EvaluateDescriptor(indirect.table_source, table_value)) {
 			return false;
@@ -596,16 +617,40 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 		} else {
 			return false;
 		}
-		if (IsBoundedDescriptorTable(program, indirect)) {
+		if (byte_keys && table_size > UINT32_MAX) return failed("table geometry");
+		bool use_selector = selector != nullptr;
+		if (byte_keys && selector != nullptr) {
+			ShaderBufferResource material;
+			if (!DecodeBufferDescriptor(material_value, material)) return false;
+			use_selector = CanReadIndexedSelector(material, selector->offset);
+			if (use_selector && material.Format() != Prospero::BufferFormat::kInvalid &&
+			    material.NumRecords() != 0u && selector->offset < material.Stride()) {
+				const auto begin = material.Base48() + selector->offset;
+				const auto end = begin + uint64_t {material.NumRecords() - 1u} * material.Stride() + 4u;
+				for (uint32_t i = 0; i < program.info.buffers.size(); ++i) {
+					const auto& buffer = program.info.buffers[i];
+					if (!buffer.written) continue;
+					ShaderBufferResource written;
+					if (program.descriptor_sources[buffer.source].indirect_descriptor ||
+					    !DecodeBufferDescriptor(snapshot.buffers[i], written) ||
+					    (written.Base48() != 0u && written.GetSize() != 0u &&
+					     begin < written.Base48() + written.GetSize() &&
+					     written.Base48() < end)) {
+						use_selector = false;
+						break;
+					}
+				}
+			}
+		}
+		if (byte_keys && !use_selector) {
 			if (!indirect.table_scalar && table.Format() == Prospero::BufferFormat::kInvalid) {
 				descriptors[resource_index] = {.dword_count = dword_count};
 				return true;
 			}
-			if (table_size > UINT32_MAX ||
-			    (indirect.table_record_bytes != 0u &&
+			if (indirect.table_record_bytes != 0u &&
 			     (table.Type() != 0u || table.SwizzleEnabled() || table.AddTid() ||
 			      table.OutOfBounds() != 0u || (table_base & 3u) != 0u ||
-			      (table.Stride() & 3u) != 0u || table.Stride() < indirect.table_record_bytes)))
+			      (table.Stride() & 3u) != 0u || table.Stride() < indirect.table_record_bytes))
 				return failed("table geometry");
 			const auto* pointer = indirect.pointer_table ? &*indirect.pointer_table : nullptr;
 			const bool indexed_pointer = pointer != nullptr && pointer->indexed;
@@ -640,10 +685,11 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 				return failed("candidate count", fmt::format("rows={} limit={}", count, MaxIndirectDescriptorProbes));
 			// Affine SMEM offsets wrap in U32 before adding the aligned immediate.
 			// Image heaps retain one bulk snapshot and their absolute byte-offset mapping.
+			auto& heap_words = program.material_keys;
 			if (image_heap) {
-				keys.resize(table_size / 4u);
-				if (!ReadScalarTable(table_base, table_size, 0u, runtime, keys))
-					return read_failed("image heap read", 0u, 0u, table_base, 0u, keys.size() * sizeof(uint32_t));
+				heap_words.resize(table_size / 4u);
+				if (!ReadScalarTable(table_base, table_size, 0u, runtime, heap_words))
+					return read_failed("image heap read", 0u, 0u, table_base, 0u, heap_words.size() * sizeof(uint32_t));
 			}
 			const auto mapping_offset = snapshot.flattened_srt.size();
 			descriptors[resource_index] = {.dword_count = dword_count};
@@ -670,15 +716,14 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 						                     std::span(candidate.dwords).first(read_dwords))) return false;
 					} else {
 						const auto table_word = table_offset / 4u;
-						std::copy_n(keys.begin() + table_word,
-						            std::min<size_t>(read_dwords, keys.size() - table_word),
+						std::copy_n(heap_words.begin() + table_word,
+						            std::min<size_t>(read_dwords, heap_words.size() - table_word),
 						            candidate.dwords.begin());
 					}
 					if (!normalize(candidate))
 						return normalization_failed(table_offset, word, step,
 						                            indirect.table_maximum ? "key" : "probe");
-					uint32_t ordinal;
-					if (!intern_candidate(candidate, ordinal)) return false;
+					const auto ordinal = intern_candidate(candidate);
 					if (ordinal == 0u) continue;
 					const auto mapping_word = image_heap ? table_offset / 4u : word * child_count + child;
 					snapshot.flattened_srt.resize(mapping_offset + 2u + mapping_word, 0u);
@@ -722,21 +767,38 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 				selector_keys->resolved = true;
 			}
 		}
+		if (byte_keys) {
+			// Keep cached selector values independent of each destination table's
+			// U32 arithmetic, and use the same byte keys as the full-heap GPU path.
+			byte_offsets.clear();
+			byte_offsets.reserve(keys.size());
+			for (const auto key: keys) {
+				const auto offset = uint64_t {(key * indirect.table_stride + indirect.table_offset) & ~3u} +
+				                    (indirect.table_immediate & ~3u);
+				byte_offsets.push_back(offset <= UINT32_MAX ? static_cast<uint32_t>(offset) : UINT32_MAX);
+			}
+			std::ranges::sort(byte_offsets);
+			byte_offsets.erase(std::unique(byte_offsets.begin(), byte_offsets.end()), byte_offsets.end());
+		}
 	}
 
 	const auto mapping_offset = snapshot.flattened_srt.size();
-	const auto key_count = sources.empty() ? keys.size() : sources.size();
+	const auto selected_keys = byte_keys ? std::span<const uint32_t>(byte_offsets)
+	                                    : std::span<const uint32_t>(keys);
+	const auto key_count = sources.empty() ? selected_keys.size() : sources.size();
 	snapshot.flattened_srt.resize(mapping_offset + 1u + key_count * 2u);
 	snapshot.flattened_srt[mapping_offset] = static_cast<uint32_t>(key_count);
 	for (uint32_t entry = 0; entry < key_count; ++entry) {
-		const auto key = sources.empty() ? keys[entry] : entry;
+		const auto key = sources.empty() ? selected_keys[entry] : entry;
 		uint64_t table_offset = 0;
 		DescriptorValue candidate;
 		candidate.dword_count = dword_count;
 		if (sources.empty()) {
-			table_offset = uint64_t {(key * indirect.table_stride + indirect.table_offset) & ~3u} +
-			               (indirect.table_immediate & ~3u);
-			if (!ReadScalarTable(table_base, table_size, table_offset, runtime,
+			table_offset = byte_keys ? key
+			    : uint64_t {(key * indirect.table_stride + indirect.table_offset) & ~3u} +
+			          (indirect.table_immediate & ~3u);
+			if (!(byte_keys && key == UINT32_MAX) &&
+			    !ReadScalarTable(table_base, table_size, table_offset, runtime,
 			                     std::span(candidate.dwords).first(read_dwords))) {
 				return read_failed("table descriptor read", key, 0u, table_base, table_offset, descriptor_bytes);
 			}
@@ -746,10 +808,10 @@ bool MaterializeIndirectDescriptor(const ResourcePlan&                         p
 		if (!normalize(candidate))
 			return normalization_failed(table_offset, key, indirect.table_stride, "key");
 		uint32_t ordinal = 0;
-		if (entry == 0) {
+		if (entry == 0 && !byte_keys) {
 			descriptors[resource_index] = candidate;
-		} else if (!intern_candidate(candidate, ordinal)) {
-			return false;
+		} else {
+			ordinal = intern_candidate(candidate);
 		}
 		snapshot.flattened_srt[mapping_offset + 1u + entry * 2u] = key;
 		snapshot.flattened_srt[mapping_offset + 2u + entry * 2u] = ordinal;
@@ -772,9 +834,8 @@ struct SamplerPlan {
 		uint32_t     source;
 		SamplerClass type;
 	};
-	std::array<std::array<uint32_t, 2>, ShaderInfo::MaxSamplers> mapping;
-	std::array<Binding, ShaderInfo::MaxSamplers>                bindings;
-	uint32_t                                                  sampler_count = 0;
+	std::vector<std::array<uint32_t, 2>> mapping;
+	std::vector<Binding>                bindings;
 };
 
 template <typename T, typename Keep>
@@ -913,6 +974,7 @@ static bool BuildResourceSpecialization(const ResourcePlan& program, ResourceSna
 			if (i >= program.info.images.size() && source != nullptr &&
 			    source->indirect_descriptor &&
 			    IsBoundedDescriptorTable(program, *source->indirect_descriptor) &&
+			    specialization.images[base_index].indirect_search_iterations == 0u &&
 			    !source->indirect_descriptor->pointer_table) {
 				const auto& indirect = *source->indirect_descriptor;
 				const size_t mapping = specialization.images[base_index].indirect_mapping_offset;
@@ -997,7 +1059,6 @@ static bool BuildResourceSpecialization(const ResourcePlan& program, ResourceSna
 			    is_1d_or_2d(image_class.dimension);
 			if ((!same_coordinates && !(is_2d(image.dimension) && is_2d(image_class.dimension)) &&
 			     !unfiltered_coordinates) ||
-			    image.mip_count != image_class.mip_count ||
 			    image.conversion_format != image_class.conversion_format ||
 			    (!storage_swizzle && image.shader_swizzle != image_class.shader_swizzle)) {
 				return SpecializationFail(
@@ -1011,11 +1072,9 @@ static bool BuildResourceSpecialization(const ResourcePlan& program, ResourceSna
 }
 
 bool BuildSamplerPlan(const ShaderInfo& base, SamplerPlan& plan) {
-	if (base.samplers.size() > plan.mapping.size()) {
-		return false;
-	}
-	std::array<uint8_t, ShaderInfo::MaxSamplers> usage {};
-	plan.sampler_count = static_cast<uint32_t>(base.samplers.size());
+	plan.mapping.resize(base.samplers.size());
+	plan.bindings.resize(base.samplers.size());
+	std::vector<uint8_t> usage(base.samplers.size());
 	for (const auto& pair: base.sampled_pairs) {
 		if (pair.image >= base.images.size() || pair.sampler >= base.samplers.size()) {
 			return false;
@@ -1037,8 +1096,8 @@ bool BuildSamplerPlan(const ShaderInfo& base, SamplerPlan& plan) {
 		bool       first   = true;
 		for (uint32_t type = 0; type < mapping.size(); type++) {
 			if ((classes & (1u << type)) == 0u) continue;
-			const auto target = first ? index : plan.sampler_count++;
-			if (target >= ShaderInfo::MaxSamplers) return false;
+			const auto target = first ? index : static_cast<uint32_t>(plan.bindings.size());
+			if (!first) plan.bindings.emplace_back();
 			mapping[type]         = target;
 			plan.bindings[target] = {index, static_cast<SamplerClass>(type)};
 			first                = false;
@@ -1178,7 +1237,7 @@ static UniformFillPlan AnalyzeUniformFill(const Program& program) {
 	}
 	if (store == nullptr || visited.size() != program.blocks.size()) return {};
 	for (const auto& buffer: program.info.buffers) {
-		if (buffer.read) return {};
+		if (buffer.read && (!buffer.scalar || buffer.written)) return {};
 	}
 	const auto& memory = program.memory_info.at(store->Flags<MemoryFlags>().index);
 	UniformFillPlan result;
@@ -1409,7 +1468,8 @@ 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 +1484,11 @@ 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 +1574,21 @@ 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()) {
@@ -1536,9 +1603,9 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 		return true;
 	};
 	const auto resource_failed = [&](std::string_view kind, uint32_t resource, uint32_t source,
-	                                 size_t candidates, uint32_t capacity) {
-		return SpecializationFail(fmt::format("indirect {} resource={} source={} candidates={} capacity={}",
-		    kind, resource, source, candidates, capacity));
+	                                 size_t candidates) {
+		return SpecializationFail(fmt::format("indirect {} resource={} source={} candidates={}",
+		    kind, resource, source, candidates));
 	};
 	snapshot.buffers.resize(program.info.buffers.size());
 	specialization.buffers.assign(program.info.buffers.size(), {});
@@ -1555,12 +1622,12 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 				if (active.empty() || active[base.source]) {
 					if (!MaterializeIndirectDescriptor(
 					        program, base.source, i, 4u, 4u, observed, clean, snapshot,
-					        snapshot.buffers, specialization.buffers, UINT32_MAX,
+					        snapshot.buffers, specialization.buffers,
 					        [&](DescriptorValue& value) {
 						        return NormalizeIndirectBuffer(
 						            value, source->indirect_descriptor->indexed, base);
 					        }))
-						return resource_failed("buffer", i, base.source, snapshot.buffers.size(), UINT32_MAX);
+						return resource_failed("buffer", i, base.source, snapshot.buffers.size());
 				}
 			} else if (!evaluate(base.source, snapshot.buffers[i], base.written)) {
 				return false;
@@ -1589,6 +1656,7 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 		    base.formatted ? descriptor.Format() : Prospero::BufferFormat::kInvalid;
 		buffer.descriptor_swizzle = base.formatted ? descriptor.DstSelXYZW() : DstSel(4, 5, 6, 7);
 		buffer.zero_stride_oob    = descriptor.OutOfBounds() == 0u && stride == 0u;
+		buffer.scalar_empty       = base.scalar && descriptor.GetSize() == 0u;
 	}
 	snapshot.images.resize(program.info.images.size());
 	specialization.images.resize(program.info.images.size());
@@ -1618,14 +1686,14 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 			const bool bounded_table = IsBoundedDescriptorTable(program, indirect);
 			if (!MaterializeIndirectDescriptor(
 			        program, image.source, i, 8u, image.r128 ? 4u : 8u, observed, clean, snapshot, snapshot.images,
-			        specialization.images, UINT32_MAX, [&](DescriptorValue& value) {
+			        specialization.images, [&](DescriptorValue& value) {
 				        // A broad heap also contains resources for other typed image operations.
 				        if (NullImageDescriptor(value) || !ValidImageDescriptor(value, image.r128) ||
 				            (bounded_table && DescriptorDimension(value, image.dimension) != image.dimension))
 					        value.dwords.fill(0u);
 				        return true;
 			        })) {
-				return resource_failed("image", i, image.source, snapshot.images.size(), UINT32_MAX);
+				return resource_failed("image", i, image.source, snapshot.images.size());
 			}
 		} else {
 			if (!evaluate(image.source, snapshot.images[i], image.written)) {
@@ -1651,9 +1719,9 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 				snapshot.samplers[i] = {.dword_count = 4u};
 				if ((active.empty() || active[source]) && !MaterializeIndirectDescriptor(
 				        program, source, i, 4u, 4u, observed, clean, snapshot, snapshot.samplers,
-				        specialization.samplers, ShaderInfo::MaxSamplers,
+				        specialization.samplers,
 				        [](DescriptorValue&) { return true; }))
-					return resource_failed("sampler", i, source, snapshot.samplers.size(), ShaderInfo::MaxSamplers);
+					return resource_failed("sampler", i, source, snapshot.samplers.size());
 			} else {
 				if (indirect != nullptr) source = indirect->sources[0];
 				if (!evaluate(source, snapshot.samplers[i])) return false;
@@ -1662,11 +1730,13 @@ bool MaterializeResources(const ResourcePlan& program, const SrtRuntime& runtime
 		if (program.info.samplers[base_index].gather_lod) {
 			const auto control = snapshot.samplers[i].dwords[2];
 			const auto filter = (control >> 26u) & 3u;
-			// MipNone always selects the base level. Explicit point gathers currently require
-			// encoded-zero primary and secondary bias; linear primary-mip selection is unsupported.
-			if (filter > 1u || (filter == 1u && (control & 0xfffffu) != 0u)) {
+			const auto perf_mip = (snapshot.samplers[i].dwords[1] >> 24u) & 15u;
+			// MipNone always selects the base level. Point/linear selection requires zero
+			// primary/secondary bias; linear mip snapping is not implemented.
+			if (filter > 2u || (filter != 0u && (control & 0xfffffu) != 0u) ||
+			    (filter == 2u && perf_mip != 0u)) {
 				return SpecializationFail(
-				    "explicit-LOD gather requires mip filtering None or Point with zero LOD biases");
+				    "explicit-LOD gather requires None, unbiased Point, or unbiased Linear without mip snapping");
 			}
 		}
 	}
@@ -1759,8 +1829,8 @@ void ApplyResourceSpecialization(Program& program, const ResourceSpecialization&
 	SamplerPlan sampler_plan;
 	EXIT_IF(!BuildSamplerPlan(program.info, sampler_plan));
 	const auto original_sampler_count = samplers.size();
-	samplers.reserve(sampler_plan.sampler_count);
-	for (uint32_t index = 0; index < sampler_plan.sampler_count; index++) {
+	samplers.reserve(sampler_plan.bindings.size());
+	for (uint32_t index = 0; index < sampler_plan.bindings.size(); index++) {
 		const auto& binding = sampler_plan.bindings[index];
 		if (index >= original_sampler_count) {
 			samplers.push_back(samplers[binding.source]);
@@ -1769,7 +1839,7 @@ void ApplyResourceSpecialization(Program& program, const ResourceSpecialization&
 		samplers[index].force_point_filtering = binding.type == SamplerClass::Integer;
 		samplers[index].integer_border        = binding.type != SamplerClass::Float;
 	}
-	for (uint32_t index = 0; index < sampler_plan.sampler_count; index++) {
+	for (uint32_t index = 0; index < sampler_plan.bindings.size(); index++) {
 		const auto& binding = sampler_plan.bindings[index];
 		for (auto& candidate: samplers[index].indirect_resources) {
 			candidate = sampler_plan.mapping[candidate][static_cast<uint32_t>(binding.type)];
@@ -1786,6 +1856,17 @@ void ApplyResourceSpecialization(Program& program, const ResourceSpecialization&
 	}
 
 	auto& memory_info = program.memory_info;
+	// Resolve empty scalar-table words before folding the indirect descriptors that use them.
+	for (auto* block: program.blocks) {
+		for (auto& inst: *block) {
+			if (inst.GetOpcode() != ValueOpcode::ReadConstBuffer) continue;
+			const auto& memory = memory_info[inst.Flags<MemoryFlags>().index];
+			if (memory.kind == ResourceKind::ScalarBuffer &&
+			    specialization.buffers[memory.resource].scalar_empty &&
+			    buffers[memory.resource].indirect_root == BufferResource::NoIndirectBuffer)
+				inst.ReplaceUsesWith(Value(0u));
+		}
+	}
 	const ImageRemap image_remap(specialization);
 	for (auto* block: program.blocks) {
 		for (auto it = block->begin(); it != block->end(); ++it) {
@@ -1885,6 +1966,7 @@ void ApplyResourceSpecialization(Program& program, const ResourceSpecialization&
 			EXIT_IF(image.indirect_root == memory.resource &&
 			        inst.GetOpcode() != ValueOpcode::ImageSampleRaw &&
 			        inst.GetOpcode() != ValueOpcode::ImageGatherRaw &&
+			        inst.GetOpcode() != ValueOpcode::ImageQueryDimensions &&
 			        inst.GetOpcode() != ValueOpcode::ImageRead &&
 			        inst.GetOpcode() != ValueOpcode::ImageWrite);
 		}

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

@@ -13,6 +13,7 @@ struct ResourceSpecialization {
 		Prospero::BufferFormat descriptor_format               = Prospero::BufferFormat::kInvalid;
 		uint32_t               descriptor_swizzle              = DstSel(4, 5, 6, 7);
 		bool                   zero_stride_oob                 = false;
+		bool                   scalar_empty                    = false;
 		uint32_t               indirect_root                   = BufferResource::NoIndirectBuffer;
 		uint32_t               indirect_mapping_offset         = 0;
 		uint32_t               indirect_search_iterations      = 0;

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 {
@@ -1556,14 +1558,49 @@ private:
 		return read;
 	}
 
+	const Inst* IndexedMaterialRead(const Inst* read, Value active,
+	                               DescriptorSource::IndirectDescriptor::SelectorRead& selector) const {
+		uint32_t lane = 0;
+		if (read != nullptr && read->GetOpcode() == ValueOpcode::CompositeExtractU32x2) {
+			if (!ImmediateU32(read->Arg(1), lane) || lane >= 2u) return nullptr;
+			read = read->Arg(0).Resolve().TryInstruction();
+			if (read == nullptr || read->GetOpcode() != ValueOpcode::LoadBufferU32x2) return nullptr;
+		} else if (read == nullptr || read->GetOpcode() != ValueOpcode::LoadBufferU32) {
+			return nullptr;
+		}
+		const auto& memory = m_program.memory_info[read->Flags<MemoryFlags>().index];
+		if (memory.kind != ResourceKind::Buffer || memory.typed || memory.formatted ||
+		    !memory.idxen || memory.offen || memory.offset > UINT32_MAX - lane * 4u ||
+		    read->Arg(2).Resolve() != Value(0u) || read->Arg(3).Resolve() != Value(0u) ||
+		    !EquivalentValue(m_program, active, read->Arg(4))) return nullptr;
+		selector.offset = memory.offset + lane * 4u;
+		return read;
+	}
+
 	bool MatchUniformizedMaterialKey(Value key, const Inst& image,
 	                                DescriptorSource::IndirectDescriptor& indirect,
-	                                DescriptorSource& material_source) {
+	                                DescriptorSource& material_source, bool indexed) {
 		const auto* selected = UniformizedMaterialValue(key, image);
 		if (selected == nullptr) return false;
 		const auto active = selected->Arg(0).Resolve();
 		DescriptorSource::IndirectDescriptor::SelectorRead selector;
 		const auto* read = SelectorReadValue(selected->Arg(1), active, selector);
+		if (indexed) {
+			read = IndexedMaterialRead(read, active, selector);
+			// This optional domain is checked against the dispatch's writable buffers.
+			// Other stages and unbounded/image writes need the original heap domain.
+			if (read == nullptr || m_program.stage != ShaderType::Compute ||
+			    std::ranges::any_of(m_program.blocks, [](const Block* block) {
+				    return std::ranges::any_of(*block, [](const Inst& inst) {
+					    const auto op = inst.GetOpcode();
+					    return ImageOpcodeInfoOf(op).resource_class == ImageResourceClass::Storage ||
+					           AddressOpcodeInfoOf(op).access == AddressAccess::Write;
+				    });
+			    }) || !MakeRuntimeTableSource(*read, material_source)) return false;
+			selector.source = InternSource(material_source);
+			indirect.selector = std::move(selector);
+			return true;
+		}
 		if (read == nullptr || read->GetOpcode() != ValueOpcode::LoadAddressU32 ||
 		    read->NumArgs() != 4u || !Implies(active, read->Arg(3)))
 			return false;
@@ -2813,7 +2850,7 @@ private:
 				} else result.maximum = std::max(result.maximum, SelectorMaximum(key, {}, 0u, active));
 			}
 		}
-		if (lane != nullptr) result = SelectedFiniteDomain(*lane, handle != nullptr ? lane->Arg(1) : Value {});
+		if (lane != nullptr) result = SelectedSelectorDomain(*lane, handle != nullptr ? lane->Arg(1) : Value {});
 		else if (handle == nullptr) result.maximum = SelectorMaximum(key);
 		if (result.maximum > UINT32_MAX / 4u ||
 		    uint64_t {result.maximum} * indirect.table_stride + indirect.table_offset > UINT32_MAX)
@@ -2969,6 +3006,13 @@ private:
 			// Keep live descriptor payloads, but project descriptor-only SMEM onto
 			// the GPU key using the same plan as other indirect image reads.
 			plan.retain_reads |= table_source.dword_count == 4u;
+		} else if (!sampler && !plan.table_indexed && table_source.dword_count == 4u &&
+		           MatchUniformizedMaterialKey(key, handle, indirect, material_source, true)) {
+			// Preserve the heap's byte-offset key so runtime geometry can choose the
+			// tighter selector domain without changing the GPU's descriptor reads.
+			indirect.table_scalar = true;
+			plan.table_read = plan.reads[0];
+			plan.retain_reads = true;
 		} else if (sampler && table_source.dword_count == 4u && !plan.table_indexed &&
 		           SetExactTableMaximum(key, indirect, &handle)) {
 			indirect.table_scalar = true;
@@ -3018,7 +3062,7 @@ private:
 				}
 			}
 			if (indirect.key_count.IsEmpty() &&
-			    !MatchUniformizedMaterialKey(key, handle, indirect, material_source) &&
+			    !MatchUniformizedMaterialKey(key, handle, indirect, material_source, false) &&
 			    !MatchReversedImageRange(key, indirect)) {
 				const auto maximum = SelectorMaximum(key);
 				if (maximum >= MaxIndirectDescriptorProbes) return false;
@@ -3062,6 +3106,10 @@ private:
 					indirect.selector_mask = selected.selector_mask;
 					indirect.signed_count = selected.signed_count;
 					indirect.workgroup_axis = selected.workgroup_axis;
+					if (selected.table_scalar) {
+						indirect.table_scalar = true;
+						plan.table_read = plan.reads[0];
+					}
 					material_source = m_sources[selected.selector->source];
 					shared_selector = true;
 					break;
@@ -3262,26 +3310,13 @@ private:
 			if (value == nullptr || (selected != nullptr && selected != value)) return false;
 			selected = value;
 		}
-		const auto* read = selected->Arg(1).Resolve().TryInstruction();
-		uint32_t lane = 0;
-		if (read != nullptr && read->GetOpcode() == ValueOpcode::CompositeExtractU32x2) {
-			if (!ImmediateU32(read->Arg(1), lane) || lane >= 2u) return false;
-			read = read->Arg(0).Resolve().TryInstruction();
-			if (read == nullptr || read->GetOpcode() != ValueOpcode::LoadBufferU32x2) return false;
-		} else if (read == nullptr || read->GetOpcode() != ValueOpcode::LoadBufferU32) {
-			return false;
-		}
-		const auto& memory = m_program.memory_info[read->Flags<MemoryFlags>().index];
-		uint32_t offset = 1u, scalar = 1u;
-		if (memory.kind != ResourceKind::Buffer || memory.typed || memory.formatted ||
-		    !memory.idxen || memory.offen || memory.offset > UINT32_MAX - lane * 4u ||
-		    !ImmediateU32(read->Arg(2), offset) || offset != 0u ||
-		    !ImmediateU32(read->Arg(3), scalar) || scalar != 0u ||
-		    !EquivalentValue(m_program, selected->Arg(0), read->Arg(4))) return false;
-		if (!MakeRuntimeTableSource(*read, material_source) ||
+		DescriptorSource::IndirectDescriptor::SelectorRead selector;
+		const auto* read = IndexedMaterialRead(selected->Arg(1).Resolve().TryInstruction(),
+		                                       selected->Arg(0), selector);
+		if (read == nullptr || !MakeRuntimeTableSource(*read, material_source) ||
 		    !BoundedSelectedIndex(read->Arg(1), read->Arg(4), indirect)) return false;
-		indirect.selector.emplace(DescriptorSource::IndirectDescriptor::SelectorRead{.source = InternSource(material_source),
-		                              .offset = memory.offset + lane * 4u});
+		selector.source = InternSource(material_source);
+		indirect.selector = std::move(selector);
 		return finish();
 	}
 
@@ -3370,10 +3405,12 @@ private:
 		return found ? SelectorDomain {maximum, counted ? count : Value {}} : SelectorDomain {};
 	}
 
-	SelectorDomain SelectedFiniteDomain(const Inst& lane, Value guard) const {
+	SelectorDomain SelectedSelectorDomain(const Inst& lane, Value guard) const {
 		if (guard.IsEmpty() ? !HasActiveLane(lane.Arg(1), lane.Parent())
 		                    : !Implies(guard, lane.Arg(1))) return {};
-		return FiniteSelectorDomain(lane.Arg(0), lane.Arg(1));
+		auto result = FiniteSelectorDomain(lane.Arg(0), lane.Arg(1));
+		if (result.maximum == UINT32_MAX) result.maximum = PossibleU32Bits(lane.Arg(0));
+		return result;
 	}
 
 	uint32_t SelectorMaximum(Value value, SelectorBounds bounds = {}, uint32_t depth = 0,
@@ -3390,7 +3427,7 @@ private:
 		};
 		switch (inst->GetOpcode()) {
 			case ValueOpcode::ReadFirstLane:
-				return SelectedFiniteDomain(*inst, guard).maximum;
+				return SelectedSelectorDomain(*inst, guard).maximum;
 			case ValueOpcode::ReadLane:
 				// ReadLane ignores EXEC; only context-free bounds on every source lane apply.
 				return PossibleU32Bits(inst->Arg(0));
@@ -3735,6 +3772,8 @@ private:
 		resource.written         = resource.written || write;
 		resource.atomic          = resource.atomic || atomic;
 		resource.formatted       = resource.formatted || memory.formatted;
+		resource.scalar          = resource.scalar || op == ValueOpcode::ReadConstBuffer ||
+		                           memory.kind == ResourceKind::ScalarBuffer;
 	}
 
 	uint32_t AddImage(uint32_t source, const MemoryInfo& memory, ValueOpcode op, uint32_t pc) {
@@ -3819,9 +3858,6 @@ private:
 				return;
 			}
 		}
-		if (m_info.sampled_pairs.size() >= ShaderInfo::MaxSampledPairs) {
-			Fail(pc, "sampled image/sampler pair limit exceeded");
-		}
 		m_info.sampled_pairs.push_back({image, sampler, pc});
 	}
 
@@ -3881,24 +3917,6 @@ private:
 		if (memory.planning_only || IsIndirectPlanningMemory(flags.index)) {
 			return;
 		}
-		if (op == ValueOpcode::ReadConstBuffer) {
-			const auto* handle = inst.Arg(0).Resolve().TryInstruction();
-			if (memory.kind != ResourceKind::ScalarBuffer || handle == nullptr ||
-			    handle->GetOpcode() != ValueOpcode::GetBufferResource || handle->NumArgs() != 4u) {
-				Fail(flags.pc, "scalar buffer read requires a four-DWORD buffer descriptor");
-			}
-			// Scalar descriptors can span disjoint guest mappings. Keep their raw GPU operands
-			// separate from handles that vector-resource planning may rewrite.
-			auto* block = inst.Parent();
-			const auto where = std::ranges::find_if(block->Instructions(),
-			    [&](const Inst& candidate) { return &candidate == &inst; });
-			auto& scalar = *block->PrependNewInst(where, ValueOpcode::GetBufferResource,
-			    {handle->Arg(0), handle->Arg(1), handle->Arg(2), handle->Arg(3)});
-			inst.SetArg(0, Value(&scalar));
-			m_program.memory_info[flags.index].kind = ResourceKind::IndirectBuffer;
-			m_info.uses_dma = true;
-			return;
-		}
 		Inst*    handle   = nullptr;
 		uint32_t source   = 0;
 		uint32_t resource = 0;
@@ -3910,11 +3928,12 @@ private:
 				source = indirect->source;
 			} else if (!GetHandle(inst.Arg(0), ValueOpcode::GetBufferResource, 4, flags.pc,
 			               memory.resource * 4u, handle, source)) {
-				if (memory.kind != ResourceKind::Buffer ||
+				if (memory.kind != (op == ValueOpcode::ReadConstBuffer ? ResourceKind::ScalarBuffer
+				                                                        : ResourceKind::Buffer) ||
 				    !memory.SupportsIndirectBufferLoad(op)) {
 					Fail(flags.pc,
 					     "buffer descriptor is not a valid runtime value; GPU-selected access "
-					     "requires a raw 8/16-bit, raw DWORD x1/x2/x3/x4, or formatted X load");
+					     "requires a scalar, raw 8/16-bit, raw DWORD x1/x2/x3/x4, or formatted X load");
 				}
 				m_program.memory_info[flags.index].kind = ResourceKind::IndirectBuffer;
 				m_info.uses_dma                         = true;
@@ -4055,11 +4074,71 @@ 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

@@ -307,7 +307,7 @@ void Visit(Program& program, ShaderStageInputInfo input_info, InputUsage& inputs
 			if (shared_access != SharedAccess::None) {
 				info.uses_lds |= kind == ResourceKind::Lds;
 				info.uses_gds |= kind == ResourceKind::Gds;
-			} else if (kind == ResourceKind::Buffer) {
+			} else if (kind == ResourceKind::Buffer || kind == ResourceKind::ScalarBuffer) {
 				if (memory.resource >= info.buffers.size()) {
 					return Fail("buffer operation has invalid resource metadata");
 				}
@@ -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");
 				}
@@ -508,6 +509,7 @@ void Visit(Program& program, ShaderStageInputInfo input_info, InputUsage& inputs
 		case ValueOpcode::ReadConst: info.uses_flattened_srt = true; break;
 		case ValueOpcode::BitwiseXor32: info.has_bitwise_xor = true; break;
 		case ValueOpcode::StoreCompletion: info.subgroup_barrier = true; break;
+		case ValueOpcode::DataCacheInvalidate: info.coherent_buffers = true; break;
 		case ValueOpcode::BvhIntersect: info.bvh = true; break;
 		case ValueOpcode::ConditionRef:
 			info.subgroup_vote |=

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

@@ -89,8 +89,7 @@ bool IsRawRead(const ResourcePlan& values, const Inst& inst) {
 	const auto& memory = values.memory_info[index];
 	if (op != ValueOpcode::LoadBufferU32) {
 		return (op == ValueOpcode::LoadAddressU32 && memory.kind == ResourceKind::ScalarAddress) ||
-		       (op == ValueOpcode::ReadConstBuffer &&
-		        (memory.kind == ResourceKind::ScalarBuffer || memory.kind == ResourceKind::IndirectBuffer));
+		       (op == ValueOpcode::ReadConstBuffer && memory.kind == ResourceKind::ScalarBuffer);
 	}
 	if (inst.NumArgs() != 5u || memory.kind != ResourceKind::Buffer || memory.typed ||
 	    memory.formatted || memory.coherent || memory.data_bits != 32u ||
@@ -160,9 +159,11 @@ bool IsRuntimeUniformOp(ValueOpcode op) {
 		case ValueOpcode::ULessThan32:
 		case ValueOpcode::ULessThanEqual32:
 		case ValueOpcode::IEqual32:
+		case ValueOpcode::IEqual64:
 		case ValueOpcode::UGreaterThan32:
 		case ValueOpcode::SGreaterThanEqual32:
 		case ValueOpcode::INotEqual32:
+		case ValueOpcode::INotEqual64:
 		case ValueOpcode::LogicalOr:
 		case ValueOpcode::LogicalAnd:
 		case ValueOpcode::LogicalXor:
@@ -1054,8 +1055,19 @@ bool SrtWalker::EvaluateInst(const Inst& inst, uint64_t& result) {
 				return true;
 			}
 			return false;
+		case ValueOpcode::IEqual64:
+		case ValueOpcode::INotEqual64:
+			if (!binary()) return false;
+			result = inst.GetOpcode() == ValueOpcode::IEqual64 ? a == b : a != b;
+			return true;
 		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;