graphics: initialize BDA tables before first buffer registration

Co-authored-by: Leandro Bras <shadowgaming.7@gmail.com>

.vscode/launch.json

@@ -27,7 +27,7 @@
             "args": [
                 "--game", "D:/gamesps5/PPSA07429",
                 "--printf-direction", "File",
-                "--graphics-debug-dump", "True"
+                "--graphics-debug-dump"
             ],
             "stopAtEntry": false,
             "cwd": "${workspaceFolder}\\_Build\\vscode-clang",

3rdparty/CMakeLists.txt

@@ -25,6 +25,23 @@ FetchContent_MakeAvailable(opus)
 set(BUILD_TESTING "${kyty_build_testing_saved}")
 install(FILES "${opus_SOURCE_DIR}/COPYING" DESTINATION "licenses/opus")
 
+# Headsets are driven through OpenXR. Only the static loader is built; the runtime that serves it
+# (for example SteamVR, Monado or Virtual Desktop) comes from the user's system.
+set(DYNAMIC_LOADER OFF CACHE BOOL "" FORCE)
+set(BUILD_WITH_SYSTEM_JSONCPP OFF CACHE BOOL "" FORCE)
+FetchContent_Declare(openxr
+	GIT_REPOSITORY https://github.com/KhronosGroup/OpenXR-SDK.git
+	GIT_TAG f2448a8797c85814aa892efc1ab8707900fbcc78 # release-1.1.63
+	GIT_SHALLOW FALSE)
+FetchContent_MakeAvailable(openxr)
+# Keep the loader's headers and library out of the installed package.
+set_property(DIRECTORY "${openxr_SOURCE_DIR}" PROPERTY EXCLUDE_FROM_ALL TRUE)
+# The loader uses exceptions internally.
+if(NOT KYTY_CLANG_CL)
+	target_compile_options(openxr_loader PRIVATE -fexceptions)
+endif()
+install(FILES "${openxr_SOURCE_DIR}/LICENSE" DESTINATION "licenses/openxr")
+
 # The guest instruction patcher needs a decoder/encoder and a small
 # trampoline assembler. Pin the same revisions used by shadPS4's implementation.
 set(XBYAK_BUILD_TESTS OFF CACHE BOOL "" FORCE)

CMakeLists.txt

@@ -336,7 +336,7 @@ if (KYTY_CLANG_CL)
 	set_source_files_properties(${kyty_emulator_src} PROPERTIES COMPILE_FLAGS "-Wno-pragma-pack -Wno-deprecated-declarations -D_TIMESPEC_DEFINED")
 endif()
 
-set(kyty_emulator_link_libraries common Vulkan::Headers spirv-tools-opt spirv-tools SDL3::SDL3-static imgui xxhash FFmpeg::ffmpeg fmt::fmt nlohmann_json::nlohmann_json LibAtrac9 kyty_git_version Zydis::Zydis xbyak::xbyak opus)
+set(kyty_emulator_link_libraries common Vulkan::Headers spirv-tools-opt spirv-tools SDL3::SDL3-static imgui xxhash FFmpeg::ffmpeg fmt::fmt nlohmann_json::nlohmann_json LibAtrac9 kyty_git_version Zydis::Zydis xbyak::xbyak opus OpenXR::openxr_loader)
 
 # Linux system libraries.
 if(LINUX)
@@ -655,6 +655,8 @@ if(BUILD_TESTING)
 	add_test(NAME save_data_dialog COMMAND $<TARGET_FILE:save_data_dialog_tests>)
 	add_test(NAME http_uri_parse COMMAND $<TARGET_FILE:http_uri_parse_tests>)
 	add_test(NAME shader_recompiler_compute COMMAND $<TARGET_FILE:shader_recompiler_compute_tests>)
+	add_test(NAME shader_recompiler_tessellation
+		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --tessellation-only)
 	add_test(NAME shader_recompiler_alignbyte
 		COMMAND $<TARGET_FILE:shader_recompiler_compute_tests> --alignbyte-only)
 	add_test(NAME graphics_pipeline_rasterization

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/logging/log.h

@@ -51,20 +51,26 @@ inline constexpr auto BrightWhite   = fmt::fg(fmt::terminal_color::bright_white)
 
 } // namespace Color
 
+// Out of line, so callers on small guest stacks (fibers) do not reserve the formatting frame.
+template <typename... Args>
+[[gnu::noinline]] void Printf(fmt::text_style style, const char* format, const Args&... args) {
+	Write(style, fmt::sprintf(format, args...));
+}
+
 } // namespace Log
 
 // NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
 #define LOGF(...)                                                                                  \
 	do {                                                                                           \
 		if (!::Log::IsSilent()) {                                                                  \
-			::Log::Write(::fmt::sprintf(__VA_ARGS__));                                             \
+			::Log::Printf({}, __VA_ARGS__);                                                        \
 		}                                                                                          \
 	} while (false)
 // NOLINTNEXTLINE(cppcoreguidelines-macro-usage)
 #define LOGF_COLOR(style, ...)                                                                     \
 	do {                                                                                           \
 		if (!::Log::IsSilent()) {                                                                  \
-			::Log::Write((style), ::fmt::sprintf(__VA_ARGS__));                                    \
+			::Log::Printf((style), __VA_ARGS__);                                                   \
 		}                                                                                          \
 	} while (false)
 

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/command_processor/commandProcessor.h

@@ -104,7 +104,6 @@ public:
 	                       uint32_t value);
 	void PrepareCpuFlip(uint64_t request_id);
 	void SynchronizeGpu();
-	void EmitGlobalBarrier();
 	void TriggerEopEventAtEndOfPipe(uint32_t interrupt_context_id);
 	void DispatchDirect(uint32_t thread_group_x, uint32_t thread_group_y, uint32_t thread_group_z,
 	                    uint32_t mode);

src/graphics/guest_gpu/command_processor/pm4Handlers.cpp

@@ -71,12 +71,6 @@ constexpr uint32_t NormalizeRegisterOffset(uint32_t raw_offset) {
 	return raw_offset & ~RegisterSelectorMask;
 }
 
-bool ReleaseMemGcrNeedsBarrier(uint32_t eop_event_type, uint32_t gcr_cntl) {
-	return eop_event_type != 0x28u ||
-	       (gcr_cntl & (GcrGl2MetadataInvalidate | GcrGl0VectorInvalidate | GcrGl1Invalidate |
-	                    GcrGl2Invalidate | GcrGl2Writeback)) != 0;
-}
-
 uint32_t ReleaseMemCacheActionFromGcr(uint32_t gcr_cntl) {
 	return ((gcr_cntl & GcrGl2Writeback) != 0 ? 0x38u : 0x00u);
 }
@@ -291,10 +285,6 @@ static void HwCtxIgnoreAaMaskRegister([[maybe_unused]] uint32_t cmd_offset,
 
 static void HwCtxIgnoreAlphaToMaskRegister([[maybe_unused]] uint32_t value) {}
 
-static void HwCtxIgnoreDisabledUserClipPlane(CommandProcessor& cp, uint32_t value) {
-	EXIT_NOT_IMPLEMENTED(value != 0 || cp.GetCtx().GetClipControl().user_clip_planes != 0);
-}
-
 static void HwCtxIgnorePrimitiveIdReset([[maybe_unused]] uint32_t value) {}
 
 static void HwCtxIgnoreVtxControl([[maybe_unused]] uint32_t value) {}
@@ -2263,30 +2253,14 @@ KYTY_CP_OP_PARSER(CpOpReleaseMem) {
 		}
 	};
 
-	if (data_sel == 0 || interrupt_selector == 4) {
-		if (eop_event_type != 0x28 || gcr_cntl != 0) {
-			cp.EmitGlobalBarrier();
-		}
-
+	// Cache actions only order GPU work, which the renderer's resource barriers already order.
+	if (data_sel == 0 || interrupt_selector == 4 ||
+	    (release_dst == ReleaseMemDstMemory && dst_gpu_addr == nullptr)) {
 		trigger_interrupt();
 
 		return 7;
 	}
 
-	if (release_dst == ReleaseMemDstMemory && dst_gpu_addr == nullptr) {
-		if (eop_event_type != 0x28 || gcr_cntl != 0) {
-			cp.EmitGlobalBarrier();
-		}
-
-		trigger_interrupt();
-
-		return 7;
-	}
-
-	if (ReleaseMemGcrNeedsBarrier(eop_event_type, gcr_cntl)) {
-		cp.EmitGlobalBarrier();
-	}
-
 	auto cache_action = ReleaseMemCacheActionFromGcr(gcr_cntl);
 	cache_policy      = 0;
 
@@ -2899,7 +2873,7 @@ void GraphicsInitJmpTablesCxIndirect() {
 
 	for (auto cmd_offset = Pm4::PA_CL_UCP_0_X; cmd_offset <= Pm4::PA_CL_UCP_5_W; cmd_offset++) {
 		g_hw_ctx_indirect_func[cmd_offset] = [](KYTY_HW_CTX_INDIRECT_ARGS) {
-			HwCtxIgnoreDisabledUserClipPlane(cp, value);
+			cp.GetCtx().SetUserClipPlane(cmd_offset - Pm4::PA_CL_UCP_0_X, value);
 		};
 	}
 

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;
 	}
 }
 
@@ -607,7 +607,6 @@ bool GuestGpu::Process(Submission& submission) {
 		case SubmissionType::SuspendPoint:
 			// Age caches at the frame boundary, not once per command list.
 			m_renderer.RunGarbageCollector();
-			cp.EmitGlobalBarrier();
 			m_renderer.GetCommandScheduler().DeferPriorityOperation(
 			    [ready = m_suspend_point_ready] { ready->release(); });
 			cp.BufferFlush();
@@ -906,6 +905,7 @@ 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);
@@ -1221,22 +1221,6 @@ void CommandProcessor::WriteAtEndOfPipe64(uint32_t cache_policy, uint32_t event_
 	                 interrupt_context_id);
 }
 
-void CommandProcessor::EmitGlobalBarrier() {
-	Common::LockGuard lock(m_renderer.GetMutex());
-
-	vk::MemoryBarrier2 barrier {};
-	barrier.srcStageMask  = vk::PipelineStageFlagBits2::eAllCommands;
-	barrier.srcAccessMask = vk::AccessFlagBits2::eMemoryWrite;
-	barrier.dstStageMask  = vk::PipelineStageFlagBits2::eAllCommands;
-	barrier.dstAccessMask = vk::AccessFlagBits2::eMemoryRead | vk::AccessFlagBits2::eMemoryWrite;
-
-	vk::DependencyInfo dependency {};
-	dependency.memoryBarrierCount = 1;
-	dependency.pMemoryBarriers    = &barrier;
-	GetScheduler().EndRendering();
-	CurrentBuffer().Handle().pipelineBarrier2(dependency);
-}
-
 void CommandProcessor::TriggerEopEventAtEndOfPipe(uint32_t interrupt_context_id) {
 	(void)CurrentBuffer().Handle();
 	GetScheduler().DeferPriorityOperation(
@@ -1256,21 +1240,17 @@ void CommandProcessor::TriggerEvent(uint32_t event_type, uint32_t event_index,
 	}
 
 	const auto valid_cache_event_index = event_index == 0x00000000 || event_index == 0x00000007;
+	// Partial flushes and cache writebacks only order GPU work, which the renderer's resource
+	// barriers already order.
 	switch (event_type) {
 		// CsPartialFlush, GsPartialFlush, PsPartialFlush.
 		case 0x00000007:
 		case 0x0000000f:
-		case 0x00000010: EmitGlobalBarrier(); break;
-		// CbDbDataWritebackInvalidate, CbDataWritebackInvalidate.
+		case 0x00000010: break;
+		// CbDbDataWritebackInvalidate, CbDataWritebackInvalidate, DbDataWritebackInvalidate,
+		// DbMetadataWritebackInvalidate, CbMetadataWritebackInvalidate.
 		case 0x00000016:
 		case 0x00000031:
-			if (!valid_cache_event_index) {
-				EXIT("unknown event type: 0x%08" PRIx32 ", 0x%08" PRIx32 "\n", event_type,
-				     event_index);
-			}
-			EmitGlobalBarrier();
-			break;
-		// DbDataWritebackInvalidate, DbMetadataWritebackInvalidate, CbMetadataWritebackInvalidate.
 		case 0x0000002a:
 		case 0x0000002c:
 		case 0x0000002e:
@@ -1278,7 +1258,6 @@ void CommandProcessor::TriggerEvent(uint32_t event_type, uint32_t event_index,
 				EXIT("unknown event type: 0x%08" PRIx32 ", 0x%08" PRIx32 "\n", event_type,
 				     event_index);
 			}
-			EmitGlobalBarrier();
 			break;
 		case 0x0000000d:
 		case 0x0000000e:

src/graphics/guest_gpu/hardwareContext.h

@@ -5,6 +5,8 @@
 #include "common/common.h"
 #include "graphics/guest_gpu/gpu_defs.h"
 
+#include <array>
+
 namespace Libs::Graphics::HW {
 
 struct ColorBase {
@@ -824,6 +826,11 @@ public:
 	void SetBlendColor(const BlendColor& color) { m_blend_color = color; }
 	[[nodiscard]] const ClipControl& GetClipControl() const { return m_clip_control; }
 	void SetClipControl(const ClipControl& control) { m_clip_control = control; }
+	// PA_CL_UCP_0_X..PA_CL_UCP_5_W as raw float bits, four per plane.
+	[[nodiscard]] const std::array<uint32_t, 24>& GetUserClipPlanes() const {
+		return m_user_clip_planes;
+	}
+	void SetUserClipPlane(uint32_t index, uint32_t value) { m_user_clip_planes[index] = value; }
 	[[nodiscard]] const RenderControl& GetRenderControl() const { return m_render_control; }
 	void SetRenderControl(const RenderControl& control) { m_render_control = control; }
 	[[nodiscard]] const DepthRenderOverride& GetDepthRenderOverride() const {
@@ -924,6 +931,8 @@ private:
 	ColorControl    m_color_control;
 	ScanModeControl m_scan_mode_control;
 
+	std::array<uint32_t, 24> m_user_clip_planes {};
+
 	AaSampleControl m_aa_sample_control;
 	AaConfig        m_aa_config;
 

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;
@@ -254,6 +256,10 @@ BufferCache::BufferCache(GraphicContext& graphics, CommandScheduler& scheduler,
 	m_gds_buffer.Flush(0, m_gds_buffer.Size());
 	SetVulkanObjectNameF(m_graphics.device, m_bda_pagetable_buffer.Handle(),
 	                     "BDA Page Table Buffer");
+	m_scheduler.BeginCommand();
+	for (auto* buffer: {&m_bda_pagetable_buffer, m_fault_manager.GetFaultBuffer()}) {
+		buffer->Fill(0, buffer->Size(), 0);
+	}
 	const auto null_id =
 	    m_slot_buffers.insert(m_graphics, m_scheduler, MemoryUsage::DeviceLocal, 0, AllFlags, 16);
 	EXIT_IF(null_id != NULL_BUFFER_ID);
@@ -467,6 +473,7 @@ 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 +665,7 @@ 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"
 
@@ -326,7 +327,7 @@ void CommandScheduler::CheckActive() const {
 }
 
 CommandBuffer& CommandScheduler::Current() {
-	CheckActive();
+	EXIT_IF(m_command.IsInvalid());
 	return m_command;
 }
 
@@ -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/context.cpp

@@ -104,8 +104,14 @@ void CommandBuffer::BeginRendering(const RenderState& state) const {
 	rendering.pDepthAttachment     = depth_stencil.has_depth ? &depth : nullptr;
 	rendering.pStencilAttachment   = depth_stencil.has_stencil ? &stencil : nullptr;
 	Handle().beginRendering(rendering);
+	// The open pass holds the cleared contents, so a later clear begins a new pass.
 	m_render_state = state;
-	m_rendering    = true;
+	for (auto& color: m_render_state.color_attachments) {
+		color.is_clear = false;
+	}
+	m_render_state.depth_stencil_attachment.depth_clear   = false;
+	m_render_state.depth_stencil_attachment.stencil_clear = false;
+	m_rendering                                           = true;
 }
 
 void CommandBuffer::EndRendering() const {

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;
-			}
-		}
 	}
 }
 
@@ -383,10 +373,9 @@ static void ClipPrint(const char* func, const HW::ClipControl& c) {
 static void ClipCheck(const HW::ClipControl& c) {
 	// dx_linear_attr_clip_enable preserves linear (noperspective) attributes at clip-generated
 	// vertices, which Vulkan provides as part of clipping and interpolation.
-	EXIT_NOT_IMPLEMENTED(c.user_clip_planes != 0 || c.user_clip_plane_mode != 0 ||
-	                     c.vertex_kill_any || c.user_clip_plane_negate_y ||
-	                     c.user_clip_plane_cull_only || c.cull_on_clipping_error_disable ||
-	                     c.force_viewport_index_from_vs_enable);
+	EXIT_NOT_IMPLEMENTED(c.user_clip_plane_mode != 0 || c.vertex_kill_any ||
+	                     c.user_clip_plane_negate_y || c.user_clip_plane_cull_only ||
+	                     c.cull_on_clipping_error_disable || c.force_viewport_index_from_vs_enable);
 }
 
 static void RcPrint(const char* func, const HW::RenderControl& c) {

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

@@ -143,7 +143,8 @@ NativeStorageBuffer(RenderContext& context, const PreparedBindings::BufferSource
 	return result;
 }
 
-bool IsSupportedDepthTextureEncoding(const ShaderTextureResource& descriptor, bool r128) {
+// Reserved and unsupported T# fields must be zero; perf_mod only scales sampler state.
+static bool IsCanonicalTextureEncoding(const ShaderTextureResource& descriptor, bool r128) {
 	constexpr uint32_t field1_reserved_mask = 0x200fff00u;
 	constexpr uint32_t field2_reserved_mask = 0xf0003000u;
 	const uint32_t     field3_expected = descriptor.DstSelXYZW() |
@@ -154,17 +155,25 @@ bool IsSupportedDepthTextureEncoding(const ShaderTextureResource& descriptor, bo
 	const uint32_t     field4_expected = descriptor.Depth() | (descriptor.BaseArray5() << 16u);
 	const uint32_t     field5_expected = (static_cast<uint32_t>(descriptor.PerfMod5()) << 20u) |
 	                                     (static_cast<uint32_t>(descriptor.MaxMip()) << 4u);
-	const bool         common          = (descriptor.fields[1] & field1_reserved_mask) == 0 &&
-	                                     (descriptor.fields[2] & field2_reserved_mask) == 0 &&
-	                                     descriptor.fields[3] == field3_expected;
-	if (r128) {
-		return common && descriptor.fields[4] == 0 && descriptor.fields[5] == 0 &&
-		       descriptor.fields[6] == 0 && descriptor.fields[7] == 0;
+	if ((descriptor.fields[1] & field1_reserved_mask) != 0 ||
+	    (descriptor.fields[2] & field2_reserved_mask) != 0 || descriptor.fields[3] != field3_expected) {
+		return false;
 	}
-	const bool full = common && descriptor.fields[4] == field4_expected &&
-	                  descriptor.fields[5] == field5_expected;
-	if (!full ||
-	    (descriptor.MsaaDepth() && !IsMultisampledTexture(descriptor.Type()))) {
+	if (r128) {
+		return descriptor.fields[4] == 0 && descriptor.fields[5] == 0 && descriptor.fields[6] == 0 &&
+		       descriptor.fields[7] == 0;
+	}
+	return descriptor.fields[4] == field4_expected && descriptor.fields[5] == field5_expected;
+}
+
+bool IsSupportedDepthTextureEncoding(const ShaderTextureResource& descriptor, bool r128) {
+	if (!IsCanonicalTextureEncoding(descriptor, r128)) {
+		return false;
+	}
+	if (r128) {
+		return true;
+	}
+	if (descriptor.MsaaDepth() && !IsMultisampledTexture(descriptor.Type())) {
 		return false;
 	}
 	const auto metadata_control = descriptor.fields[6] & 0x00ffffffu;
@@ -281,36 +290,12 @@ static bool IsSupportedStorageTextureDescriptor(const ShaderRecompiler::IR::Imag
 	       !descriptor.MsaaDepth();
 }
 
-static bool IsSupportedStorageTextureEncoding(const ShaderRecompiler::IR::ImageResource& resource,
-                                              const ShaderTextureResource& descriptor) {
-	constexpr uint32_t field1_reserved_mask = 0x200fff00u;
-	constexpr uint32_t field2_reserved_mask = 0xf0003000u;
-	constexpr uint32_t field5_expected      = 0x00700000u;
-	constexpr uint32_t field5_max_mip_mask  = 0x000000f0u;
-	const uint32_t     expected_field3 = descriptor.DstSelXYZW() |
-	                                     (static_cast<uint32_t>(descriptor.BaseLevel()) << 12u) |
-	                                     (static_cast<uint32_t>(descriptor.LastLevel()) << 16u) |
-	                                     (static_cast<uint32_t>(descriptor.TileMode()) << 20u) |
-	                                     (static_cast<uint32_t>(descriptor.Type()) << 28u);
-	const uint32_t     expected_field4 =
-	    descriptor.Depth() | (static_cast<uint32_t>(descriptor.BaseArray5()) << 16u);
-	const bool common = (descriptor.fields[1] & field1_reserved_mask) == 0 &&
-	                    (descriptor.fields[2] & field2_reserved_mask) == 0 &&
-	                    descriptor.fields[3] == expected_field3;
-	if (resource.r128) {
-		return common && descriptor.fields[4] == 0 && descriptor.fields[5] == 0 &&
-		       descriptor.fields[6] == 0 && descriptor.fields[7] == 0;
-	}
-	return common && descriptor.fields[4] == expected_field4 &&
-	       (descriptor.fields[5] & ~field5_max_mip_mask) == field5_expected;
-}
-
 void ValidateStorageTexture(const ShaderRecompiler::IR::ImageResource& resource,
                             const ShaderTextureResource& descriptor, uint64_t size) {
 	const auto format        = descriptor.Format();
 	const bool resource_ok   = IsSupportedStorageImageResource(resource);
 	const bool descriptor_ok = IsSupportedStorageTextureDescriptor(resource, descriptor);
-	const bool encoding_ok   = IsSupportedStorageTextureEncoding(resource, descriptor);
+	const bool encoding_ok   = IsCanonicalTextureEncoding(descriptor, resource.r128);
 	const bool uint_resource    = resource.numeric_class == Prospero::TextureNumericClass::Uint;
 	const bool raw_sint_storage = uint_resource &&
 	    ShaderRecompiler::IR::RawSignedStorageFormat(resource, format) != Prospero::BufferFormat::kInvalid;
@@ -740,6 +725,12 @@ void RenderExecutor::FindBuffers(std::span<PreparedBindings* const> stages) {
 			const auto descriptor = DecodeNativeDescriptor<ShaderBufferResource>(snapshot.buffers[resource]);
 			const auto address = descriptor.Base48();
 			auto size = descriptor.GetSize();
+			if (size != 0 && descriptor.Stride() == 0 && descriptor.OutOfBounds() == 2 &&
+			    !metadata.scalar) {
+				// OOB_SELECT=2 only tests NumRecords != 0; a zero-stride fetch
+				// still needs its full payload when NumRecords is smaller than it.
+				size = std::max(size, uint64_t {metadata.max_byte_extent});
+			}
 			if (address == 0 || size == 0) {
 				source = {};
 				continue;
@@ -757,14 +748,18 @@ void RenderExecutor::FindBuffers(std::span<PreparedBindings* const> stages) {
 							}
 						}
 					}
-				} else {
-					const auto& graphics = m_context.GetGraphics();
-					const auto limit = uint64_t {graphics.GetPhysicalDeviceProperties().limits.maxStorageBufferRange} -
-					                   (graphics.StorageMinAlignment() - 1);
-					size = std::min({size, uint64_t {256} * 1024 * 1024, limit});
 				}
+				//  PPSA15246
+				const auto& graphics = m_context.GetGraphics();
+				const auto  limit =
+				    uint64_t {graphics.GetPhysicalDeviceProperties().limits.maxStorageBufferRange} -
+				    (graphics.StorageMinAlignment() - 1);
+				size = std::min({size, uint64_t {256} * 1024 * 1024, limit});
+			}
+			// Finite GPU descriptors can span sparse PRT holes; clamp only oversized/sentinel ranges.
+			if (size >= (uint64_t {1} << 30u) || descriptor.NumRecords() == UINT32_MAX) {
+				size = Libs::LibKernel::Memory::ClampRangeSize(address, size);
 			}
-			size = Libs::LibKernel::Memory::ClampRangeSize(address, size);
 			source = {address, size, cache.FindBuffer(address, size)};
 		}
 	}
@@ -946,8 +941,12 @@ void RenderExecutor::CommitBindings(CommandBuffer&                     buffer,
 				for (const auto& resource: program.info.buffers) {
 					if (!resource.written || resource.descriptor_index == UINT32_MAX) continue;
 					const auto& written = writer->buffer_sources[resource.descriptor_index];
-					if (written.size != 0 && ImageRangeOverlaps(address, size,
-					                                          written.address, written.size)) {
+					// ADD_TID and swizzle move fixed-offset writes past write_extent.
+					const bool linear = (resource.packed_stride & ((1u << 14u) | (1u << 20u))) == 0u;
+					const auto written_size =
+					    linear ? std::min<uint64_t>(written.size, resource.write_extent) : written.size;
+					if (written_size != 0 && ImageRangeOverlaps(address, size,
+					                                          written.address, written_size)) {
 						EXIT("scalar resource reads overlap a shader buffer write\n");
 					}
 				}

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;
 }
@@ -764,10 +780,12 @@ PipelineCache::GraphicsPrograms PipelineCache::GetGraphicsPrograms(
 			}
 		}
 	}
-	if (context.GetClipControl().clip_disable) {
+	auto&       last_stage   = vertex_info[tess_active ? 2u : 0u];
+	const auto& clip_control = context.GetClipControl();
+	if (clip_control.clip_disable) {
 		const auto& viewport = context.GetScreenViewport().viewports[0];
 		const auto& limits   = m_graphics.GetPhysicalDeviceProperties().limits;
-		auto&       clip     = vertex_info[tess_active ? 2u : 0u].clip_space;
+		auto&       clip     = last_stage.clip_space;
 		clip.scale[0]        = viewport.xscale;
 		clip.scale[1]        = viewport.yscale;
 		clip.offset[0]       = viewport.xoffset;
@@ -777,9 +795,16 @@ PipelineCache::GraphicsPrograms PipelineCache::GetGraphicsPrograms(
 		clip.half_extent[1] =
 		    static_cast<float>(std::min(limits.maxViewportDimensions[1], 16384u)) * 0.5f;
 		clip.enabled = true;
+	} else if (clip_control.user_clip_planes != 0) {
+		// CLIP_DISABLE also disables user planes. How they combine with shader clip and cull
+		// distances is unverified; draw push constants only reach VS/mesh and PS, and the
+		// generated RECTLIST stages do not forward clip distances.
+		EXIT_NOT_IMPLEMENTED((last_stage.pa_cl_vs_out_cntl & 0xffffu) != 0 || tess_active ||
+		                     Prospero::IsRectList(user_config.GetPrimType()));
+		last_stage.user_clip_planes = clip_control.user_clip_planes;
 	}
 	uint32_t push_data_cursor = ShaderRecompiler::IR::PushData::GraphicsDrawDwords(
-	    mesh_active, object_primitive_id);
+	    mesh_active, object_primitive_id, last_stage.user_clip_planes);
 	GraphicsPrograms  result;
 	if (pixel_active) {
 		result.pixel = m_program_cache->Get(pixel_params, pixel_info, push_data_cursor);
@@ -993,8 +1018,13 @@ 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 +1054,13 @@ 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/shaderResourceBarrier.cpp

@@ -92,7 +92,8 @@ bool HasShaderBufferWrites(const ShaderStageRuntime& runtime) {
 	const auto& program   = *runtime.program;
 	const auto& resources = *runtime.resources;
 	EXIT_IF(resources.buffers.size() != program.info.buffers.size());
-	bool has_writes = false;
+	// Address writes reach memory outside the bound buffer descriptors.
+	bool has_writes = program.has_address_writes;
 	for (uint32_t i = 0; i < program.info.buffers.size(); i++) {
 		if (!program.info.buffers[i].written) {
 			continue;

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 {};
@@ -626,6 +636,9 @@ void CreatePipelineInternal(GraphicContext& graphics, PipelineCache::Pipeline& p
 	EXIT_NOT_IMPLEMENTED(pipeline.pipeline_layout == nullptr);
 
 	vk::ComputePipelineCreateInfo info {};
+	if (input_info.stage.program->info.uses_gds) {
+		info.flags = vk::PipelineCreateFlagBits::eDispatchBase;
+	}
 	info.stage             = comp_shader_stage_info;
 	info.layout            = pipeline.pipeline_layout;
 	info.basePipelineIndex = -1;

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,
@@ -127,9 +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;
@@ -185,14 +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;
 }
@@ -296,6 +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;
@@ -304,14 +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) {
@@ -431,7 +447,28 @@ void RenderExecutor::DispatchDirect(uint64_t submit_id, CommandBuffer& buffer,
 	}
 
 	vk_buffer.bindPipeline(vk::PipelineBindPoint::eCompute, pipeline.pipeline);
-	vk_buffer.dispatch(thread_group_x, thread_group_y, thread_group_z);
+	constexpr uint32_t DISPATCH_INITIATOR_ORDERED_APPEND_ENABLE = 1u << 3u;
+	const uint64_t     group_threads = uint64_t {std::max(input_info.threads_num[0], 1u)} *
+	                                   std::max(input_info.threads_num[1], 1u) *
+	                                   std::max(input_info.threads_num[2], 1u);
+	if ((mode & DISPATCH_INITIATOR_ORDERED_APPEND_ENABLE) != 0 && program.info.uses_gds &&
+	    group_threads <= input_info.wave_size && !input_info.lds_storage) {
+		// A single guest wave per group can run in launch order without a device-wide
+		// spinlock. Keep the guest workgroup IDs when splitting the dispatch.
+		for (uint32_t z = 0; z < thread_group_z; ++z) {
+			for (uint32_t y = 0; y < thread_group_y; ++y) {
+				for (uint32_t x = 0; x < thread_group_x; ++x) {
+					if ((x | y | z) != 0u) {
+						ShaderAccessBarrier(vk_buffer, vk::PipelineStageFlagBits::eComputeShader);
+					}
+					vk_buffer.dispatchBase(x, y, z, 1, 1, 1);
+				}
+			}
+		}
+	} else {
+		vk_buffer.dispatch(thread_group_x, thread_group_y, thread_group_z);
+	}
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::Dispatches);
 
 	// The removed host fence also ordered read-only dispatches before later writers.
 	ShaderAccessBarrier(vk_buffer, vk::PipelineStageFlagBits::eComputeShader);
@@ -462,6 +499,7 @@ void RenderExecutor::DispatchIndirect(uint64_t submit_id, CommandBuffer& buffer,
 	input_info.indirect_args_address = args_addr;
 	const auto compute_program = m_context.GetPipelineCache().GetComputeProgram(
 	    cs_regs, buffer.GetRegisters().GetShaderRegisters(), input_info);
+	if (!compute_program) return;
 	buffer.EndRendering();
 	auto& pipeline = m_context.GetPipelineCache().GetComputePipeline(input_info, compute_program);
 	auto& bindings = m_compute_bindings;
@@ -501,6 +539,7 @@ void RenderExecutor::DispatchIndirect(uint64_t submit_id, CommandBuffer& buffer,
 	                          1, &barrier, 0, nullptr, 0, nullptr);
 	vk_buffer.bindPipeline(vk::PipelineBindPoint::eCompute, pipeline.pipeline);
 	vk_buffer.dispatchIndirect(args_buffer->Handle(), args_offset);
+	Common::DebugCounters::Add(Common::DebugCounters::Counter::Dispatches);
 	ShaderAccessBarrier(vk_buffer, vk::PipelineStageFlagBits::eComputeShader);
 	ResetBindings();
 }

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"
@@ -1192,13 +1193,14 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 	if (!mesh_active) {
 		CommitIndexBuffer(vk_buffer, index_binding);
 	}
-	const auto draw_dwords = ShaderRecompiler::IR::PushData::GraphicsDrawDwords(
-	    mesh_active, state.ps_input_info.object_primitive_id);
+	using ShaderRecompiler::IR::PushData;
+	const auto draw_stages =
+	    vk::ShaderStageFlagBits::eFragment |
+	    (mesh_active ? vk::ShaderStageFlagBits::eMeshEXT : vk::ShaderStageFlagBits::eVertex);
+	const auto draw_dwords =
+	    PushData::GraphicsDrawDwords(mesh_active, state.ps_input_info.object_primitive_id);
 	if (draw_dwords != 0) {
-		std::array<uint32_t, ShaderRecompiler::IR::PushData::MeshDrawDwordCount> draw_data {};
-		const auto draw_stages = vk::ShaderStageFlagBits::eFragment |
-		                         (mesh_active ? vk::ShaderStageFlagBits::eMeshEXT
-		                                      : vk::ShaderStageFlagBits::eVertex);
+		std::array<uint32_t, PushData::MeshDrawDwordCount> draw_data {};
 		if (mesh_active) {
 			draw_data = {
 			    draw.index_count,
@@ -1212,6 +1214,12 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 		vk_buffer.pushConstants(pipeline.pipeline_layout, draw_stages, 0,
 		                        draw_dwords * sizeof(uint32_t), draw_data.data());
 	}
+	if (const auto planes = vertex_stages.back().user_clip_planes; planes != 0) {
+		vk_buffer.pushConstants(pipeline.pipeline_layout, draw_stages,
+		                        PushData::UserClipPlaneDword * sizeof(uint32_t),
+		                        std::bit_width(planes) * 4u * sizeof(uint32_t),
+		                        buffer.GetRegisters().GetUserClipPlanes().data());
+	}
 
 	auto& dynamic_state = m_context.GetCommandScheduler().GetDynamicState();
 	UpdateGraphicsDynamicState(buffer, dynamic_state, vertex_stages.back(), state.depth_info,
@@ -1219,6 +1227,20 @@ void RenderExecutor::ExecutePreparedDraw(uint64_t submit_id, CommandBuffer& buff
 	dynamic_state.SetAttachmentFeedbackLoopAspects(feedback_aspects);
 	dynamic_state.Commit(m_context.GetGraphics(), vk_buffer);
 
+	vk::PipelineStageFlags shader_write_stages = {};
+	for (const auto& stage: vertex_stages) {
+		if (HasShaderBufferWrites(stage.stage)) {
+			shader_write_stages |= ShaderPipelineStages(NativeShaderStage(stage.logical_stage));
+		}
+	}
+	if (state.ps_active && HasShaderBufferWrites(state.ps_input_info.stage)) {
+		shader_write_stages |= vk::PipelineStageFlagBits::eFragmentShader;
+	}
+	if (shader_write_stages) {
+		// As for dispatches, buffer writes wait for earlier accesses to the memory.
+		m_context.GetCommandScheduler().EndRendering();
+		ShaderWriteHazardBarrier(vk_buffer, shader_write_stages);
+	}
 	LogDrawPhase(draw.Name(), "BeginRendering");
 	if (!draw.IsIndexed()) {
 		SetDrawDebugPhase(buffer, submit_id, draw, 0x400u);
@@ -1233,19 +1255,12 @@ 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);
 	}
-	vk::PipelineStageFlags shader_write_stages = {};
-	for (const auto& stage: vertex_stages) {
-		if (HasShaderBufferWrites(stage.stage)) {
-			shader_write_stages |= ShaderPipelineStages(NativeShaderStage(stage.logical_stage));
-		}
-	}
-	if (state.ps_active && HasShaderBufferWrites(state.ps_input_info.stage)) {
-		shader_write_stages |= vk::PipelineStageFlagBits::eFragmentShader;
-	}
 	if (shader_write_stages) {
 		m_context.GetCommandScheduler().EndRendering();
 		ShaderWriteBarrier(vk_buffer, shader_write_stages);

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/openxrHeadset.cpp

@@ -0,0 +1,762 @@
+#include "graphics/presentation/openxrHeadset.h"
+
+#include "common/emulatorConfig.h"
+#include "common/logging/log.h"
+#include "common/stringUtils.h"
+#include "common/threads.h"
+#include "graphics/host_gpu/graphicContext.h"
+#include "graphics/host_gpu/renderer/render.h"
+#include "graphics/host_gpu/renderer/renderContext.h"
+#include "graphics/presentation/vrInput.h"
+#include "libs/hmd2.h"
+
+#define XR_USE_GRAPHICS_API_VULKAN
+#include <algorithm>
+#include <atomic>
+#include <chrono>
+#include <cmath>
+#include <cstdlib>
+#include <cstring>
+#include <fmt/format.h>
+#include <mutex>
+#include <openxr/openxr.h>
+#include <openxr/openxr_platform.h>
+#include <string>
+#include <thread>
+
+namespace Libs::Vr {
+
+namespace {
+
+using namespace Libs::Graphics;
+
+enum Action { TRIGGER, SQUEEZE, STICK, STICK_CLICK, PRIMARY, SECONDARY, MENU, GRIP, ACTION_COUNT };
+
+// Touch controller bindings: the path under /user/hand/<side>/input/ per hand, or null.
+constexpr struct {
+	const char*  name;
+	XrActionType type;
+	const char*  path[2];
+} ACTIONS[ACTION_COUNT] = {
+    {"trigger", XR_ACTION_TYPE_FLOAT_INPUT, {"trigger/value", "trigger/value"}},
+    {"squeeze", XR_ACTION_TYPE_FLOAT_INPUT, {"squeeze/value", "squeeze/value"}},
+    {"thumbstick", XR_ACTION_TYPE_VECTOR2F_INPUT, {"thumbstick", "thumbstick"}},
+    {"thumbstick_click", XR_ACTION_TYPE_BOOLEAN_INPUT, {"thumbstick/click", "thumbstick/click"}},
+    {"primary", XR_ACTION_TYPE_BOOLEAN_INPUT, {"x/click", "a/click"}},
+    {"secondary", XR_ACTION_TYPE_BOOLEAN_INPUT, {"y/click", "b/click"}},
+    {"menu", XR_ACTION_TYPE_BOOLEAN_INPUT, {"menu/click", nullptr}},
+    {"grip_pose", XR_ACTION_TYPE_POSE_INPUT, {"grip/pose", "grip/pose"}},
+};
+
+constexpr const char* HANDS[2] = {"left", "right"};
+
+// Eye images move from the render thread to the headset thread through these. One is being written,
+// one is the latest recorded frame and one may be held by the headset thread until the GPU is done.
+constexpr int STAGING_COUNT = 3;
+
+struct Staging {
+	VulkanImage image;
+	uint64_t    tick = 0;
+	VrPose head; // The head pose the game last read, so the picture can be placed in the world.
+};
+
+struct State {
+	XrInstance  instance    = XR_NULL_HANDLE;
+	XrSystemId  system      = XR_NULL_SYSTEM_ID;
+	XrSession   session     = XR_NULL_HANDLE;
+	XrSpace     view_space  = XR_NULL_HANDLE;
+	XrSpace     local_space = XR_NULL_HANDLE;
+	XrSwapchain swapchain   = XR_NULL_HANDLE; // Two array layers, one per eye.
+	std::vector<XrSwapchainImageVulkanKHR> swapchain_images;
+	vk::Extent2D                           eye {};
+	std::vector<std::string>               extensions[2]; // Vulkan instance and device extensions.
+	PFN_xrGetVulkanGraphicsDeviceKHR       get_graphics_device = nullptr;
+
+	GraphicContext*  graphics = nullptr;
+	std::thread      thread;
+	std::atomic_bool running {false};
+
+	// Headset thread only.
+	XrSessionState    session_state   = XR_SESSION_STATE_UNKNOWN;
+	bool              session_running = false;
+	uint64_t          shown_tick      = 0;
+	VrPose            shown_head;
+	vk::CommandPool   pool;
+	vk::CommandBuffer command;
+	vk::Fence         fence;
+
+	std::mutex       mutex;
+	Staging          staging[STAGING_COUNT];
+	int              latest    = -1;
+	int              held      = -1;
+	MasterSemaphore* semaphore = nullptr;
+
+	XrActionSet action_set = XR_NULL_HANDLE;
+	XrAction    actions[ACTION_COUNT] {};
+	XrPath      hand_path[2] {};
+	XrSpace     grip_space[2] {};
+
+	std::mutex        input_mutex;
+	VrControllerInput input[2];
+	VrPose            head;
+	VrPose            last_head_read;
+};
+
+State g;
+
+bool Ok(XrResult result, const char* what) {
+	if (XR_SUCCEEDED(result)) {
+		return true;
+	}
+	char text[XR_MAX_RESULT_STRING_SIZE] = {};
+	xrResultToString(g.instance, result, text);
+	Log::WriteToConsoleAndLog(fmt::format("OpenXR: {} failed: {}\n", what, text));
+	return false;
+}
+
+// Destroying the instance also destroys the session, spaces, swapchain and actions.
+void DestroyInstance(const char* reason) {
+	if (reason != nullptr) {
+		Log::WriteToConsoleAndLog(
+		    fmt::format("OpenXR: {}; showing the eye image in the window only\n", reason));
+	}
+	if (g.instance != XR_NULL_HANDLE) {
+		xrDestroyInstance(g.instance);
+	}
+	g.instance = XR_NULL_HANDLE;
+	g.session  = XR_NULL_HANDLE;
+}
+
+struct Access {
+	vk::PipelineStageFlags2 stage;
+	vk::AccessFlags2        access;
+};
+
+constexpr Access NO_ACCESS {vk::PipelineStageFlagBits2::eNone, vk::AccessFlagBits2::eNone};
+constexpr Access TRANSFER {vk::PipelineStageFlagBits2::eTransfer, vk::AccessFlagBits2::eNone};
+constexpr Access TRANSFER_READ {vk::PipelineStageFlagBits2::eTransfer,
+                                vk::AccessFlagBits2::eTransferRead};
+constexpr Access TRANSFER_WRITE {vk::PipelineStageFlagBits2::eTransfer,
+                                 vk::AccessFlagBits2::eTransferWrite};
+constexpr Access COLOR_ATTACHMENT {vk::PipelineStageFlagBits2::eColorAttachmentOutput,
+                                   vk::AccessFlagBits2::eColorAttachmentRead |
+                                       vk::AccessFlagBits2::eColorAttachmentWrite};
+
+void Barrier(vk::CommandBuffer command, vk::Image image, vk::ImageLayout from, vk::ImageLayout to,
+             uint32_t layers, Access src, Access dst) {
+	vk::ImageMemoryBarrier2 barrier {};
+	barrier.srcStageMask        = src.stage;
+	barrier.srcAccessMask       = src.access;
+	barrier.dstStageMask        = dst.stage;
+	barrier.dstAccessMask       = dst.access;
+	barrier.oldLayout           = from;
+	barrier.newLayout           = to;
+	barrier.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
+	barrier.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
+	barrier.image               = image;
+	barrier.subresourceRange    = {vk::ImageAspectFlagBits::eColor, 0, 1, 0, layers};
+	vk::DependencyInfo dependency {};
+	dependency.imageMemoryBarrierCount = 1;
+	dependency.pImageMemoryBarriers    = &barrier;
+	command.pipelineBarrier2(dependency);
+}
+
+XrQuaternionf Multiply(const XrQuaternionf& a, const XrQuaternionf& b) {
+	return {a.w * b.x + a.x * b.w + a.y * b.z - a.z * b.y,
+	        a.w * b.y - a.x * b.z + a.y * b.w + a.z * b.x,
+	        a.w * b.z + a.x * b.y - a.y * b.x + a.z * b.w,
+	        a.w * b.w - a.x * b.x - a.y * b.y - a.z * b.z};
+}
+
+XrVector3f Rotate(const XrQuaternionf& q, const XrVector3f& v) {
+	const XrVector3f t {2.0f * (q.y * v.z - q.z * v.y), 2.0f * (q.z * v.x - q.x * v.z),
+	                    2.0f * (q.x * v.y - q.y * v.x)};
+	return {v.x + q.w * t.x + (q.y * t.z - q.z * t.y), v.y + q.w * t.y + (q.z * t.x - q.x * t.z),
+	        v.z + q.w * t.z + (q.x * t.y - q.y * t.x)};
+}
+
+VrPose Locate(XrSpace space, XrTime time) {
+	VrPose          pose;
+	XrSpaceLocation location {XR_TYPE_SPACE_LOCATION};
+	if (XR_SUCCEEDED(xrLocateSpace(space, g.local_space, time, &location)) &&
+	    (location.locationFlags & XR_SPACE_LOCATION_ORIENTATION_VALID_BIT) != 0) {
+		const auto& [q, p] = location.pose;
+		pose               = {true, {q.x, q.y, q.z, q.w}, {p.x, p.y, p.z}};
+	}
+	return pose;
+}
+
+// The swapchain stores the eyes in a gamma-encoded format, so a raw copy of the display-encoded
+// game image comes out right. Staging uses the same layout with UNORM, which a blit writes from the
+// raw UNORM or float presentation frame without conversion.
+bool CreateSwapchain() {
+	uint32_t count = 0;
+	xrEnumerateSwapchainFormats(g.session, 0, &count, nullptr);
+	std::vector<int64_t> formats(count);
+	xrEnumerateSwapchainFormats(g.session, count, &count, formats.data());
+	constexpr std::pair<vk::Format, vk::Format> candidates[] = {
+	    {vk::Format::eR8G8B8A8Srgb, vk::Format::eR8G8B8A8Unorm},
+	    {vk::Format::eB8G8R8A8Srgb, vk::Format::eB8G8R8A8Unorm},
+	    {vk::Format::eR8G8B8A8Unorm, vk::Format::eR8G8B8A8Unorm},
+	    {vk::Format::eB8G8R8A8Unorm, vk::Format::eB8G8R8A8Unorm},
+	};
+	const auto* format =
+	    std::find_if(std::begin(candidates), std::end(candidates), [&](const auto& c) {
+		    return std::find(formats.begin(), formats.end(), static_cast<int64_t>(c.first)) !=
+			       formats.end();
+	    });
+	if (format == std::end(candidates)) {
+		return false;
+	}
+	XrSwapchainCreateInfo info {XR_TYPE_SWAPCHAIN_CREATE_INFO};
+	info.usageFlags = XR_SWAPCHAIN_USAGE_COLOR_ATTACHMENT_BIT | XR_SWAPCHAIN_USAGE_TRANSFER_DST_BIT;
+	info.format     = static_cast<int64_t>(format->first);
+	info.sampleCount = 1;
+	info.width       = g.eye.width;
+	info.height      = g.eye.height;
+	info.faceCount   = 1;
+	info.arraySize   = 2;
+	info.mipCount    = 1;
+	if (!Ok(xrCreateSwapchain(g.session, &info, &g.swapchain), "xrCreateSwapchain")) {
+		return false;
+	}
+	xrEnumerateSwapchainImages(g.swapchain, 0, &count, nullptr);
+	g.swapchain_images.assign(count, {XR_TYPE_SWAPCHAIN_IMAGE_VULKAN_KHR});
+	xrEnumerateSwapchainImages(
+	    g.swapchain, count, &count,
+	    reinterpret_cast<XrSwapchainImageBaseHeader*>(g.swapchain_images.data()));
+
+	vk::ImageCreateInfo staging {};
+	staging.imageType   = vk::ImageType::e2D;
+	staging.format      = format->second;
+	staging.extent      = vk::Extent3D {g.eye.width * 2, g.eye.height, 1};
+	staging.mipLevels   = 1;
+	staging.arrayLayers = 1;
+	staging.samples     = vk::SampleCountFlagBits::e1;
+	staging.tiling      = vk::ImageTiling::eOptimal;
+	staging.usage = vk::ImageUsageFlagBits::eTransferSrc | vk::ImageUsageFlagBits::eTransferDst;
+	for (auto& slot: g.staging) {
+		if (!g.graphics->CreateImage(staging, slot.image)) {
+			return false;
+		}
+	}
+	Log::WriteToConsoleAndLog(fmt::format("OpenXR: eye swapchain {}x{} {}\n", g.eye.width,
+	                                      g.eye.height, vk::to_string(format->first)));
+	return true;
+}
+
+bool CreateInput() {
+	XrActionSetCreateInfo set {XR_TYPE_ACTION_SET_CREATE_INFO};
+	std::strcpy(set.actionSetName, "gameplay");
+	std::strcpy(set.localizedActionSetName, "Gameplay");
+	if (!Ok(xrCreateActionSet(g.instance, &set, &g.action_set), "xrCreateActionSet")) {
+		return false;
+	}
+	for (int hand = 0; hand < 2; hand++) {
+		xrStringToPath(g.instance, fmt::format("/user/hand/{}", HANDS[hand]).c_str(),
+		               &g.hand_path[hand]);
+	}
+	std::vector<XrActionSuggestedBinding> bindings;
+	for (int i = 0; i < ACTION_COUNT; i++) {
+		XrActionCreateInfo info {XR_TYPE_ACTION_CREATE_INFO};
+		info.actionType          = ACTIONS[i].type;
+		info.countSubactionPaths = 2;
+		info.subactionPaths      = g.hand_path;
+		std::strcpy(info.actionName, ACTIONS[i].name);
+		std::strcpy(info.localizedActionName, ACTIONS[i].name);
+		if (!Ok(xrCreateAction(g.action_set, &info, &g.actions[i]), "xrCreateAction")) {
+			return false;
+		}
+		for (int hand = 0; hand < 2; hand++) {
+			if (ACTIONS[i].path[hand] != nullptr) {
+				XrPath path = XR_NULL_PATH;
+				xrStringToPath(
+				    g.instance,
+				    fmt::format("/user/hand/{}/input/{}", HANDS[hand], ACTIONS[i].path[hand])
+				        .c_str(),
+				    &path);
+				bindings.push_back({g.actions[i], path});
+			}
+		}
+	}
+	XrInteractionProfileSuggestedBinding suggested {XR_TYPE_INTERACTION_PROFILE_SUGGESTED_BINDING};
+	xrStringToPath(g.instance, "/interaction_profiles/oculus/touch_controller",
+	               &suggested.interactionProfile);
+	suggested.countSuggestedBindings = static_cast<uint32_t>(bindings.size());
+	suggested.suggestedBindings      = bindings.data();
+	XrSessionActionSetsAttachInfo attach {XR_TYPE_SESSION_ACTION_SETS_ATTACH_INFO};
+	attach.countActionSets = 1;
+	attach.actionSets      = &g.action_set;
+	if (!Ok(xrSuggestInteractionProfileBindings(g.instance, &suggested), "suggest bindings") ||
+	    !Ok(xrAttachSessionActionSets(g.session, &attach), "attach action sets")) {
+		return false;
+	}
+	for (int hand = 0; hand < 2; hand++) {
+		XrActionSpaceCreateInfo space {XR_TYPE_ACTION_SPACE_CREATE_INFO};
+		space.action            = g.actions[GRIP];
+		space.subactionPath     = g.hand_path[hand];
+		space.poseInActionSpace = {{0, 0, 0, 1}, {0, 0, 0}};
+		if (!Ok(xrCreateActionSpace(g.session, &space, &g.grip_space[hand]), "grip space")) {
+			return false;
+		}
+	}
+	return true;
+}
+
+// The head is tracked while the frame loop runs; the controllers only while the session has focus.
+void PollInput(XrTime time) {
+	const auto        head = Locate(g.view_space, time);
+	VrControllerInput input[2];
+	XrActiveActionSet active {g.action_set, XR_NULL_PATH};
+	XrActionsSyncInfo sync {XR_TYPE_ACTIONS_SYNC_INFO};
+	sync.countActiveActionSets = 1;
+	sync.activeActionSets      = &active;
+	if (g.action_set != XR_NULL_HANDLE && g.session_state == XR_SESSION_STATE_FOCUSED &&
+	    XR_SUCCEEDED(xrSyncActions(g.session, &sync))) {
+		for (int hand = 0; hand < 2; hand++) {
+			const auto get = [&](Action action, auto read, auto state) {
+				XrActionStateGetInfo info {XR_TYPE_ACTION_STATE_GET_INFO};
+				info.action        = g.actions[action];
+				info.subactionPath = g.hand_path[hand];
+				read(g.session, &info, &state);
+				return state;
+			};
+			const auto value = [&](Action action) {
+				return get(action, xrGetActionStateFloat,
+				           XrActionStateFloat {XR_TYPE_ACTION_STATE_FLOAT})
+				    .currentState;
+			};
+			const auto flag = [&](Action action) {
+				return get(action, xrGetActionStateBoolean,
+				           XrActionStateBoolean {XR_TYPE_ACTION_STATE_BOOLEAN})
+				           .currentState == XR_TRUE;
+			};
+			const auto stick = get(STICK, xrGetActionStateVector2f,
+			                       XrActionStateVector2f {XR_TYPE_ACTION_STATE_VECTOR2F});
+			auto&      out   = input[hand];
+			out.trigger      = value(TRIGGER);
+			out.squeeze      = value(SQUEEZE);
+			out.stick_x      = stick.currentState.x;
+			out.stick_y      = stick.currentState.y;
+			out.stick_click  = flag(STICK_CLICK);
+			out.primary      = flag(PRIMARY);
+			out.secondary    = flag(SECONDARY);
+			out.menu         = flag(MENU);
+			out.pose         = Locate(g.grip_space[hand], time);
+			out.connected =
+			    get(GRIP, xrGetActionStatePose, XrActionStatePose {XR_TYPE_ACTION_STATE_POSE})
+			        .isActive == XR_TRUE;
+		}
+	}
+	std::lock_guard lock(g.input_mutex);
+	g.head     = head;
+	g.input[0] = input[0];
+	g.input[1] = input[1];
+}
+
+void HandleEvents() {
+	for (XrEventDataBuffer event {XR_TYPE_EVENT_DATA_BUFFER};
+	     xrPollEvent(g.instance, &event) == XR_SUCCESS; event = {XR_TYPE_EVENT_DATA_BUFFER}) {
+		if (event.type == XR_TYPE_EVENT_DATA_INSTANCE_LOSS_PENDING) {
+			g.running = false;
+			continue;
+		}
+		if (event.type != XR_TYPE_EVENT_DATA_SESSION_STATE_CHANGED) {
+			continue;
+		}
+		g.session_state = reinterpret_cast<const XrEventDataSessionStateChanged&>(event).state;
+		LOGF("OpenXR: session state %d\n", static_cast<int>(g.session_state));
+		switch (g.session_state) {
+			case XR_SESSION_STATE_READY: {
+				XrSessionBeginInfo begin {XR_TYPE_SESSION_BEGIN_INFO};
+				begin.primaryViewConfigurationType = XR_VIEW_CONFIGURATION_TYPE_PRIMARY_STEREO;
+				g.session_running = Ok(xrBeginSession(g.session, &begin), "xrBeginSession");
+				break;
+			}
+			case XR_SESSION_STATE_STOPPING: {
+				xrEndSession(g.session);
+				g.session_running = false;
+				std::lock_guard lock(g.input_mutex);
+				g.head     = {};
+				g.input[0] = {};
+				g.input[1] = {};
+				break;
+			}
+			case XR_SESSION_STATE_EXITING:
+			case XR_SESSION_STATE_LOSS_PENDING: g.running = false; break;
+			default: break;
+		}
+	}
+}
+
+bool CopyToSwapchain(const VulkanImage& staging) {
+	uint32_t                    image = 0;
+	XrSwapchainImageAcquireInfo acquire {XR_TYPE_SWAPCHAIN_IMAGE_ACQUIRE_INFO};
+	{
+		Common::LockGuard queue_lock(g.graphics->queue_mutex);
+		if (!Ok(xrAcquireSwapchainImage(g.swapchain, &acquire, &image), "acquire")) {
+			return false;
+		}
+	}
+	XrSwapchainImageWaitInfo wait {XR_TYPE_SWAPCHAIN_IMAGE_WAIT_INFO};
+	wait.timeout = XR_INFINITE_DURATION;
+	bool copied  = Ok(xrWaitSwapchainImage(g.swapchain, &wait), "wait swapchain image");
+	if (copied) {
+		const auto target = g.swapchain_images[image].image;
+		(void)g.command.reset();
+		vk::CommandBufferBeginInfo begin {};
+		begin.flags = vk::CommandBufferUsageFlagBits::eOneTimeSubmit;
+		(void)g.command.begin(&begin);
+		Barrier(g.command, target, vk::ImageLayout::eUndefined,
+		        vk::ImageLayout::eTransferDstOptimal, 2, NO_ACCESS, TRANSFER_WRITE);
+		vk::ImageCopy copies[2] {};
+		for (uint32_t eye = 0; eye < 2; eye++) {
+			copies[eye].srcSubresource = {vk::ImageAspectFlagBits::eColor, 0, 0, 1};
+			copies[eye].srcOffset = vk::Offset3D {static_cast<int32_t>(eye * g.eye.width), 0, 0};
+			copies[eye].dstSubresource = {vk::ImageAspectFlagBits::eColor, 0, eye, 1};
+			copies[eye].extent         = vk::Extent3D {g.eye.width, g.eye.height, 1};
+		}
+		g.command.copyImage(staging.image, vk::ImageLayout::eTransferSrcOptimal, target,
+		                    vk::ImageLayout::eTransferDstOptimal, 2, copies);
+		Barrier(g.command, target, vk::ImageLayout::eTransferDstOptimal,
+		        vk::ImageLayout::eColorAttachmentOptimal, 2, TRANSFER_WRITE, COLOR_ATTACHMENT);
+		(void)g.command.end();
+		vk::SubmitInfo submit {};
+		submit.commandBufferCount = 1;
+		submit.pCommandBuffers    = &g.command;
+		(void)g.graphics->device.resetFences(1, &g.fence);
+		vk::Result result {};
+		{
+			Common::LockGuard queue_lock(g.graphics->queue_mutex);
+			result = g.graphics->queue.submit(1, &submit, g.fence);
+		}
+		copied = result == vk::Result::eSuccess &&
+		         g.graphics->device.waitForFences(1, &g.fence, VK_TRUE, UINT64_MAX) ==
+		             vk::Result::eSuccess;
+	}
+	XrSwapchainImageReleaseInfo release {XR_TYPE_SWAPCHAIN_IMAGE_RELEASE_INFO};
+	Common::LockGuard           queue_lock(g.graphics->queue_mutex);
+	return Ok(xrReleaseSwapchainImage(g.swapchain, &release), "release") && copied;
+}
+
+// Holds the newest recorded eye image until the GPU has finished it, without waiting, then copies
+// it into the swapchain. The swapchain keeps the last copy, so frames in between show it again.
+bool UpdateSwapchain() {
+	int              index     = -1;
+	MasterSemaphore* semaphore = nullptr;
+	{
+		std::lock_guard lock(g.mutex);
+		if (g.held < 0 && g.latest >= 0 && g.staging[g.latest].tick != g.shown_tick) {
+			g.held = g.latest;
+		}
+		index     = g.held;
+		semaphore = g.semaphore;
+	}
+	if (index >= 0) {
+		const auto& slot = g.staging[index];
+		semaphore->Refresh();
+		if (semaphore->IsFree(slot.tick)) {
+			if (CopyToSwapchain(slot.image)) {
+				g.shown_tick = slot.tick;
+				g.shown_head = slot.head;
+			}
+			std::lock_guard lock(g.mutex);
+			g.held = -1;
+		}
+	}
+	return g.shown_tick != 0;
+}
+
+// The game drew the picture for the head pose it read, so it is placed in the world at that pose
+// and the runtime reprojects it as the head moves. Eye offsets come from the runtime.
+void PlaceViews(XrTime time, XrCompositionLayerProjection& layer,
+                XrCompositionLayerProjectionView (&views)[2]) {
+	XrViewLocateInfo locate {XR_TYPE_VIEW_LOCATE_INFO};
+	locate.viewConfigurationType = XR_VIEW_CONFIGURATION_TYPE_PRIMARY_STEREO;
+	locate.displayTime           = time;
+	locate.space                 = g.view_space;
+	XrViewState state {XR_TYPE_VIEW_STATE};
+	XrView      located[2] {{XR_TYPE_VIEW}, {XR_TYPE_VIEW}};
+	uint32_t    count = 0;
+	const bool  valid =
+	    XR_SUCCEEDED(xrLocateViews(g.session, &locate, &state, 2, &count, located)) &&
+	    (state.viewStateFlags & XR_VIEW_STATE_POSITION_VALID_BIT) != 0;
+	const auto&         head = g.shown_head;
+	const XrQuaternionf head_q {head.orientation[0], head.orientation[1], head.orientation[2],
+	                            head.orientation[3]};
+	const XrVector3f    head_p {head.position[0], head.position[1], head.position[2]};
+	// The outer side of each eye is its temporal side: left for the left eye.
+	const float out    = std::atan(Hmd2::FOV_TAN_OUT);
+	const float in     = std::atan(Hmd2::FOV_TAN_IN);
+	const float top    = std::atan(Hmd2::FOV_TAN_TOP);
+	const float bottom = std::atan(Hmd2::FOV_TAN_BOTTOM);
+	for (int eye = 0; eye < 2; eye++) {
+		const float side = eye == 0 ? -Hmd2::EYE_OFFSET_M : Hmd2::EYE_OFFSET_M;
+		auto&       view = views[eye];
+		view.pose        = valid ? located[eye].pose : XrPosef {{0, 0, 0, 1}, {side, 0, 0}};
+		if (head.valid) {
+			const auto offset     = Rotate(head_q, view.pose.position);
+			view.pose.position    = {head_p.x + offset.x, head_p.y + offset.y, head_p.z + offset.z};
+			view.pose.orientation = Multiply(head_q, view.pose.orientation);
+		}
+		view.fov                = {eye == 0 ? -out : -in, eye == 0 ? in : out, top, -bottom};
+		view.subImage.swapchain = g.swapchain;
+		view.subImage.imageRect = {
+		    {0, 0}, {static_cast<int32_t>(g.eye.width), static_cast<int32_t>(g.eye.height)}};
+		view.subImage.imageArrayIndex = static_cast<uint32_t>(eye);
+	}
+	layer.space     = head.valid ? g.local_space : g.view_space;
+	layer.viewCount = 2;
+	layer.views     = views;
+}
+
+void ThreadMain() {
+	while (g.running) {
+		HandleEvents();
+		XrFrameWaitInfo wait {XR_TYPE_FRAME_WAIT_INFO};
+		XrFrameState    frame {XR_TYPE_FRAME_STATE};
+		if (!g.session_running || !Ok(xrWaitFrame(g.session, &wait, &frame), "xrWaitFrame")) {
+			std::this_thread::sleep_for(std::chrono::milliseconds(50));
+			continue;
+		}
+		PollInput(frame.predictedDisplayTime);
+		XrFrameBeginInfo begin {XR_TYPE_FRAME_BEGIN_INFO};
+		{
+			Common::LockGuard queue_lock(g.graphics->queue_mutex);
+			if (!Ok(xrBeginFrame(g.session, &begin), "xrBeginFrame")) {
+				continue;
+			}
+		}
+		XrCompositionLayerProjection     layer {XR_TYPE_COMPOSITION_LAYER_PROJECTION};
+		XrCompositionLayerProjectionView views[2] {{XR_TYPE_COMPOSITION_LAYER_PROJECTION_VIEW},
+		                                           {XR_TYPE_COMPOSITION_LAYER_PROJECTION_VIEW}};
+		const auto*    layers = reinterpret_cast<const XrCompositionLayerBaseHeader*>(&layer);
+		XrFrameEndInfo end {XR_TYPE_FRAME_END_INFO};
+		end.displayTime          = frame.predictedDisplayTime;
+		end.environmentBlendMode = XR_ENVIRONMENT_BLEND_MODE_OPAQUE;
+		if (frame.shouldRender == XR_TRUE && UpdateSwapchain()) {
+			PlaceViews(frame.predictedDisplayTime, layer, views);
+			end.layerCount = 1;
+			end.layers     = &layers;
+		}
+		Common::LockGuard queue_lock(g.graphics->queue_mutex);
+		(void)Ok(xrEndFrame(g.session, &end), "xrEndFrame");
+	}
+}
+
+} // namespace
+
+void OpenXrInit() {
+	if (!Config::VrEnabled()) {
+		return;
+	}
+	if (const char* env = std::getenv("KYTY_OPENXR");
+	    env != nullptr && std::strcmp(env, "0") == 0) {
+		return;
+	}
+	const char*          extensions[] = {XR_KHR_VULKAN_ENABLE_EXTENSION_NAME};
+	XrInstanceCreateInfo info {XR_TYPE_INSTANCE_CREATE_INFO};
+	std::strcpy(info.applicationInfo.applicationName, "KytyPS5");
+	std::strcpy(info.applicationInfo.engineName, "Kyty");
+	info.applicationInfo.apiVersion = XR_API_VERSION_1_0;
+	info.enabledExtensionCount      = 1;
+	info.enabledExtensionNames      = extensions;
+	if (XR_FAILED(xrCreateInstance(&info, &g.instance))) {
+		DestroyInstance("no runtime available");
+		return;
+	}
+	XrSystemGetInfo system {XR_TYPE_SYSTEM_GET_INFO};
+	system.formFactor = XR_FORM_FACTOR_HEAD_MOUNTED_DISPLAY;
+	if (XR_FAILED(xrGetSystem(g.instance, &system, &g.system))) {
+		DestroyInstance("no headset available");
+		return;
+	}
+	PFN_xrGetVulkanGraphicsRequirementsKHR get_requirements = nullptr;
+	PFN_xrGetVulkanInstanceExtensionsKHR   get_extensions[2] {};
+	xrGetInstanceProcAddr(g.instance, "xrGetVulkanGraphicsRequirementsKHR",
+	                      reinterpret_cast<PFN_xrVoidFunction*>(&get_requirements));
+	xrGetInstanceProcAddr(g.instance, "xrGetVulkanInstanceExtensionsKHR",
+	                      reinterpret_cast<PFN_xrVoidFunction*>(&get_extensions[0]));
+	// Both queries share one signature.
+	xrGetInstanceProcAddr(g.instance, "xrGetVulkanDeviceExtensionsKHR",
+	                      reinterpret_cast<PFN_xrVoidFunction*>(&get_extensions[1]));
+	xrGetInstanceProcAddr(g.instance, "xrGetVulkanGraphicsDeviceKHR",
+	                      reinterpret_cast<PFN_xrVoidFunction*>(&g.get_graphics_device));
+	XrGraphicsRequirementsVulkanKHR requirements {XR_TYPE_GRAPHICS_REQUIREMENTS_VULKAN_KHR};
+	if (get_requirements == nullptr || get_extensions[0] == nullptr ||
+	    get_extensions[1] == nullptr || g.get_graphics_device == nullptr ||
+	    XR_FAILED(get_requirements(g.instance, g.system, &requirements))) {
+		DestroyInstance("the runtime does not support XR_KHR_vulkan_enable");
+		return;
+	}
+	for (int i = 0; i < 2; i++) {
+		uint32_t size = 0;
+		get_extensions[i](g.instance, g.system, 0, &size, nullptr);
+		std::string names(size, '\0');
+		get_extensions[i](g.instance, g.system, size, &size, names.data());
+		g.extensions[i] = Common::Split(names.c_str(), ' ');
+	}
+	XrViewConfigurationView views[2] {{XR_TYPE_VIEW_CONFIGURATION_VIEW},
+	                                  {XR_TYPE_VIEW_CONFIGURATION_VIEW}};
+	uint32_t                count = 0;
+	xrEnumerateViewConfigurationViews(g.instance, g.system,
+	                                  XR_VIEW_CONFIGURATION_TYPE_PRIMARY_STEREO, 2, &count, views);
+	g.eye = {views[0].recommendedImageRectWidth, views[0].recommendedImageRectHeight};
+	XrSystemProperties properties {XR_TYPE_SYSTEM_PROPERTIES};
+	xrGetSystemProperties(g.instance, g.system, &properties);
+	Log::WriteToConsoleAndLog(fmt::format("OpenXR: headset '{}' ready\n", properties.systemName));
+}
+
+void OpenXrAppendExtensions(std::vector<const char*>&                   out,
+                            const std::vector<vk::ExtensionProperties>& available, bool device) {
+	if (g.instance == XR_NULL_HANDLE) {
+		return;
+	}
+	for (const auto& name: g.extensions[device ? 1 : 0]) {
+		const auto same = [&](const char* other) { return name == other; };
+		if (std::any_of(available.begin(), available.end(),
+		                [&](const auto& extension) { return same(extension.extensionName); }) &&
+		    std::none_of(out.begin(), out.end(), same)) {
+			out.push_back(name.c_str());
+		}
+	}
+}
+
+bool OpenXrPhysicalDevice(vk::Instance instance, vk::PhysicalDevice& out) {
+	VkPhysicalDevice device = VK_NULL_HANDLE;
+	if (g.instance == XR_NULL_HANDLE ||
+	    XR_FAILED(g.get_graphics_device(g.instance, g.system, static_cast<VkInstance>(instance),
+	                                    &device))) {
+		return false;
+	}
+	out = vk::PhysicalDevice(device);
+	return true;
+}
+
+void OpenXrStartSession(GraphicContext& graphics) {
+	if (g.instance == XR_NULL_HANDLE) {
+		return;
+	}
+	// The runtime must be asked for its GPU before a session is created.
+	vk::PhysicalDevice required;
+	if (!OpenXrPhysicalDevice(graphics.instance, required) ||
+	    required != graphics.physical_device) {
+		DestroyInstance("the headset runs on another GPU; select it with --gpu-index");
+		return;
+	}
+	g.graphics = &graphics;
+	XrGraphicsBindingVulkanKHR binding {XR_TYPE_GRAPHICS_BINDING_VULKAN_KHR};
+	binding.instance         = static_cast<VkInstance>(graphics.instance);
+	binding.physicalDevice   = static_cast<VkPhysicalDevice>(graphics.physical_device);
+	binding.device           = static_cast<VkDevice>(graphics.device);
+	binding.queueFamilyIndex = graphics.queue_family;
+	binding.queueIndex       = 0;
+	XrSessionCreateInfo info {XR_TYPE_SESSION_CREATE_INFO};
+	info.next     = &binding;
+	info.systemId = g.system;
+	XrReferenceSpaceCreateInfo space {XR_TYPE_REFERENCE_SPACE_CREATE_INFO};
+	space.poseInReferenceSpace = {{0, 0, 0, 1}, {0, 0, 0}};
+	const auto reference       = [&](XrReferenceSpaceType type, XrSpace& out) {
+		space.referenceSpaceType = type;
+		return Ok(xrCreateReferenceSpace(g.session, &space, &out), "xrCreateReferenceSpace");
+	};
+	if (!Ok(xrCreateSession(g.instance, &info, &g.session), "xrCreateSession") ||
+	    !reference(XR_REFERENCE_SPACE_TYPE_VIEW, g.view_space) ||
+	    !reference(XR_REFERENCE_SPACE_TYPE_LOCAL, g.local_space) || !CreateSwapchain()) {
+		DestroyInstance("the headset session could not be created");
+		return;
+	}
+	if (!CreateInput()) {
+		g.action_set = XR_NULL_HANDLE;
+	}
+	vk::CommandPoolCreateInfo pool {};
+	pool.queueFamilyIndex = graphics.queue_family;
+	pool.flags            = vk::CommandPoolCreateFlagBits::eResetCommandBuffer;
+	(void)graphics.device.createCommandPool(&pool, nullptr, &g.pool);
+	vk::CommandBufferAllocateInfo allocate {};
+	allocate.commandPool        = g.pool;
+	allocate.level              = vk::CommandBufferLevel::ePrimary;
+	allocate.commandBufferCount = 1;
+	(void)graphics.device.allocateCommandBuffers(&allocate, &g.command);
+	vk::FenceCreateInfo fence {};
+	(void)graphics.device.createFence(&fence, nullptr, &g.fence);
+
+	g.running = true;
+	g.thread  = std::thread(ThreadMain);
+}
+
+void OpenXrShutdown() {
+	g.running = false;
+	if (g.thread.joinable()) {
+		g.thread.join();
+	}
+	DestroyInstance(nullptr);
+	if (g.graphics == nullptr) {
+		return;
+	}
+	{
+		Common::LockGuard queue_lock(g.graphics->queue_mutex);
+		(void)g.graphics->device.waitIdle();
+	}
+	for (auto& slot: g.staging) {
+		if (slot.image.image != nullptr) {
+			g.graphics->DeleteImage(slot.image);
+		}
+	}
+	g.graphics->device.destroyFence(g.fence, nullptr);
+	g.graphics->device.destroyCommandPool(g.pool, nullptr);
+	g.graphics = nullptr;
+}
+
+VrPose OpenXrGetHeadPose() {
+	if (!g.running) {
+		return {};
+	}
+	std::lock_guard lock(g.input_mutex);
+	if (g.head.valid) {
+		g.last_head_read = g.head;
+	}
+	return g.head;
+}
+
+VrControllerInput OpenXrGetController(int hand) {
+	if (!g.running) {
+		return {};
+	}
+	std::lock_guard lock(g.input_mutex);
+	return g.input[hand];
+}
+
+void OpenXrQueueEyeSource(CommandBuffer& command, const VulkanImage& source) {
+	if (!g.running) {
+		return;
+	}
+	std::lock_guard lock(g.mutex);
+	int             index = 0;
+	while (index == g.held || index == g.latest) {
+		index++;
+	}
+	auto&      slot       = g.staging[index];
+	const auto vk_command = command.Handle();
+	Barrier(vk_command, slot.image.image, vk::ImageLayout::eUndefined,
+	        vk::ImageLayout::eTransferDstOptimal, 1, TRANSFER, TRANSFER_WRITE);
+	vk::ImageBlit blit {};
+	blit.srcSubresource = {vk::ImageAspectFlagBits::eColor, 0, 0, 1};
+	blit.srcOffsets[1]  = vk::Offset3D {static_cast<int32_t>(source.extent.width),
+	                                    static_cast<int32_t>(source.extent.height), 1};
+	blit.dstSubresource = {vk::ImageAspectFlagBits::eColor, 0, 0, 1};
+	blit.dstOffsets[1]  = vk::Offset3D {static_cast<int32_t>(slot.image.extent.width),
+	                                    static_cast<int32_t>(slot.image.extent.height), 1};
+	vk_command.blitImage(source.image, vk::ImageLayout::eTransferSrcOptimal, slot.image.image,
+	                     vk::ImageLayout::eTransferDstOptimal, 1, &blit, vk::Filter::eLinear);
+	Barrier(vk_command, slot.image.image, vk::ImageLayout::eTransferDstOptimal,
+	        vk::ImageLayout::eTransferSrcOptimal, 1, TRANSFER_WRITE, TRANSFER_READ);
+	auto& scheduler = command.GetContext().GetCommandScheduler();
+	slot.tick       = scheduler.CurrentTick();
+	g.semaphore     = &scheduler.GetMasterSemaphore();
+	g.latest        = index;
+	std::lock_guard input_lock(g.input_mutex);
+	slot.head = g.last_head_read;
+}
+
+} // namespace Libs::Vr

src/graphics/presentation/openxrHeadset.h

@@ -0,0 +1,41 @@
+#ifndef EMULATOR_SRC_GRAPHICS_PRESENTATION_OPENXRHEADSET_H_
+#define EMULATOR_SRC_GRAPHICS_PRESENTATION_OPENXRHEADSET_H_
+
+#include "graphics/host_gpu/vulkanCommon.h"
+
+#include <vector>
+
+namespace Libs::Graphics {
+struct GraphicContext;
+struct VulkanImage;
+class CommandBuffer;
+} // namespace Libs::Graphics
+
+// Output to a real headset through OpenXR. Everything here is inert unless the virtual headset is
+// enabled (--vr) and an OpenXR runtime with a headset is available; KYTY_OPENXR=0 turns it off.
+namespace Libs::Vr {
+
+// Creates the OpenXR instance. Call before the Vulkan instance exists, because the runtime names
+// the Vulkan extensions and the GPU it needs.
+void OpenXrInit();
+
+// Appends the Vulkan instance or device extensions the runtime needs, when available.
+void OpenXrAppendExtensions(std::vector<const char*>&                   out,
+                            const std::vector<vk::ExtensionProperties>& available, bool device);
+
+// The physical device the runtime renders with. False when OpenXR is not in use.
+bool OpenXrPhysicalDevice(vk::Instance instance, vk::PhysicalDevice& out);
+
+// Creates the session and starts the headset thread. Call once the device and queue exist.
+void OpenXrStartSession(Graphics::GraphicContext& graphics);
+
+// Stops the headset thread and destroys the session. Call before the device is destroyed.
+void OpenXrShutdown();
+
+// Render thread, headset flips only: scales the side-by-side eye image into a staging image for the
+// headset thread. The source is in transfer-source layout.
+void OpenXrQueueEyeSource(Graphics::CommandBuffer& command, const Graphics::VulkanImage& source);
+
+} // namespace Libs::Vr
+
+#endif // EMULATOR_SRC_GRAPHICS_PRESENTATION_OPENXRHEADSET_H_

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/presenter.h

@@ -27,7 +27,7 @@ public:
 	KYTY_CLASS_NO_COPY(Presenter);
 
 	[[nodiscard]] Frame&         PrepareFrame(CommandBuffer& command, const ImageInfo& info,
-	                                         bool hdr = false);
+	                                          bool hdr = false, bool headset = false);
 	[[nodiscard]] Frame&         PrepareBlankFrame(uint32_t width, uint32_t height, bool opaque,
 	                                               CommandBuffer* producer = nullptr);
 	[[nodiscard]] bool           PresentLastFrame();

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"
@@ -15,10 +16,12 @@
 #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"
 #include "libs/errno.h"
+#include "libs/hmd2.h"
 #include "libs/libs.h"
 #include "loader/systemContent.h"
 
@@ -49,6 +52,7 @@ constexpr int      VIDEO_OUT_FALSE                                      = 0;
 constexpr int      VIDEO_OUT_BUS_TYPE_MAIN                              = 0;
 constexpr int      VIDEO_OUT_BUS_TYPE_OVERLAY                           = 1;
 constexpr int      VIDEO_OUT_BUS_TYPE_SUB                               = 2;
+constexpr int      VIDEO_OUT_BUS_TYPE_HMD2_MAIN                          = 32;
 constexpr int      VIDEO_OUT_FLIP_MODE_VSYNC                            = 1;
 constexpr int      VIDEO_OUT_FLIP_MODE_VSYNC_MULTI                      = 4;
 constexpr int      VIDEO_OUT_FLIP_MODE_SLAVE                            = 8;
@@ -64,6 +68,7 @@ constexpr uint64_t VIDEO_OUT_REFRESH_RATE_59_94HZ                       = 3;
 constexpr uint64_t VIDEO_OUT_REFRESH_RATE_119_88HZ                      = 13;
 constexpr int      VIDEO_OUT_BUFFER_ATTRIBUTE_CATEGORY_UNCOMPRESSED     = 0;
 constexpr int      VIDEO_OUT_BUFFER_ATTRIBUTE_CATEGORY_COMPRESSED       = 1;
+constexpr int      VIDEO_OUT_BUFFER_ATTRIBUTE_CATEGORY_HMD2              = 0x10;
 constexpr uint64_t VIDEO_OUT_BUFFER_ATTRIBUTE_OPTION_STRICT_COLORIMETRY = 8;
 constexpr uint64_t VIDEO_OUT_BUFFER_ATTRIBUTE_OPTION_ALPHA_PREMULTIPLIED = 32;
 
@@ -250,8 +255,9 @@ private:
 
 struct VideoOutDriver::Impl {
 public:
-	static constexpr std::array VIDEO_OUT_BUSES {
-	    VIDEO_OUT_BUS_TYPE_MAIN, VIDEO_OUT_BUS_TYPE_OVERLAY, VIDEO_OUT_BUS_TYPE_SUB};
+	static constexpr std::array VIDEO_OUT_BUSES {VIDEO_OUT_BUS_TYPE_MAIN,
+	                                             VIDEO_OUT_BUS_TYPE_OVERLAY, VIDEO_OUT_BUS_TYPE_SUB,
+	                                             VIDEO_OUT_BUS_TYPE_HMD2_MAIN};
 	static constexpr int VIDEO_OUT_NUM_MAX = VIDEO_OUT_BUSES.size() + 1;
 
 	Impl(uint32_t width, uint32_t height, Graphics::Presenter& presenter)
@@ -514,6 +520,10 @@ static bool IsValidFlipMode(int mode) {
 	       mode == VIDEO_OUT_FLIP_MODE_SLAVE || mode == VIDEO_OUT_FLIP_MODE_MASTER;
 }
 
+// Set once the headset port shows eye images. The window then shows them instead of the main port,
+// so the picture is visible without a headset.
+static std::atomic<bool> g_headset_mirror {false};
+
 static int ReserveFlipRequest(VideoOutDriver::Impl& driver, int handle, int index, int flip_mode,
                               int64_t flip_arg, FlipRequestSource source, uint64_t& request_id) {
 	auto* video_out = driver.Get(handle);
@@ -532,8 +542,10 @@ static int ReserveFlipRequest(VideoOutDriver::Impl& driver, int handle, int inde
 	    (video_out->slave_count != 0) != (flip_mode == VIDEO_OUT_FLIP_MODE_MASTER)) {
 		return VIDEO_OUT_ERROR_INVALID_FLIP_MODE;
 	}
+	// The headset port can flip before its reprojection buffers are registered; those show blank.
 	if (video_out->closing ||
-	    (!IsSpecialBufferIndex(index) && !video_out->buffers[index].Occupied())) {
+	    (!IsSpecialBufferIndex(index) && !video_out->buffers[index].Occupied() &&
+	     video_out->bus != VIDEO_OUT_BUS_TYPE_HMD2_MAIN)) {
 		return VIDEO_OUT_ERROR_INVALID_INDEX;
 	}
 	if (!driver.GetFlipQueue().Reserve(*video_out, index, flip_arg, source, request_id)) {
@@ -738,6 +750,9 @@ bool VideoOutDriver::Impl::Close(int handle) {
 		master->slave_count--;
 	}
 	m_flip_queue.Cancel(config);
+	if (config.bus == VIDEO_OUT_BUS_TYPE_HMD2_MAIN) {
+		g_headset_mirror = false;
+	}
 	m_presenter.ClearLayer(config.bus);
 	DeleteVideoOutEvents(flip_events, VideoOutEventKind::Flip);
 	DeleteVideoOutEvents(pre_vblank_events, VideoOutEventKind::PreVblankStart);
@@ -909,6 +924,9 @@ 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 +956,7 @@ 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 +984,7 @@ 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;
@@ -1059,16 +1079,19 @@ void FlipQueue::Prepare(uint64_t request_id, Graphics::CommandBuffer& buffer) {
 		index          = request->index;
 	}
 
-	const bool          special = IsSpecialBufferIndex(index);
+	bool                special = IsSpecialBufferIndex(index);
 	Graphics::ImageInfo source_info;
 	uint32_t            width   = 0;
 	uint32_t            height  = 0;
 	bool                current = false;
 	bool                premultiplied_alpha = false;
 	bool                hdr                 = false;
+	bool                headset             = false;
 	{
 		Common::LockGuard lock(cfg->mutex);
 		current = cfg->opened && !cfg->closing && cfg->generation == generation;
+		headset = cfg->bus == VIDEO_OUT_BUS_TYPE_HMD2_MAIN;
+		special = special || (headset && !cfg->buffers[index].Occupied());
 		if (current && special) {
 			width  = cfg->width;
 			height = cfg->height;
@@ -1113,7 +1136,10 @@ void FlipQueue::Prepare(uint64_t request_id, Graphics::CommandBuffer& buffer) {
 		frame = &m_presenter.PrepareBlankFrame(width, height, index == VIDEO_OUT_BUFFER_INDEX_BLACK,
 		                                       &buffer);
 	} else {
-		frame = &m_presenter.PrepareFrame(buffer, source_info, hdr);
+		frame = &m_presenter.PrepareFrame(buffer, source_info, hdr, headset);
+		if (headset) {
+			g_headset_mirror = true;
+		}
 	}
 
 	Common::LockGuard lock(m_mutex);
@@ -1232,6 +1258,8 @@ bool FlipQueue::Flip(uint32_t micros) {
 	          [](const auto& a, const auto& b) { return a.cfg->bus < b.cfg->bus; });
 	bool due = true;
 	std::array<Graphics::Presenter::Layer, VideoOutDriver::Impl::VIDEO_OUT_BUSES.size()> layers {};
+	size_t                      layer_count = 0;
+	Graphics::Presenter::Frame* hidden      = nullptr; // The main port behind the headset image.
 	for (size_t i = 0; i < count; i++) {
 		auto& r = requests[i];
 		r.cfg->mutex.Lock();
@@ -1239,10 +1267,25 @@ bool FlipQueue::Flip(uint32_t micros) {
 		if (r.id == group) {
 			due &= IsFlipDueLocked(*r.cfg, r.generation);
 		}
-		layers[i] = {r.frame, r.cfg->bus, r.premultiplied_alpha};
+		if (r.cfg->bus == VIDEO_OUT_BUS_TYPE_MAIN &&
+		    g_headset_mirror.load(std::memory_order_relaxed)) {
+			hidden = r.frame;
+		} else {
+			const bool headset    = r.cfg->bus == VIDEO_OUT_BUS_TYPE_HMD2_MAIN;
+			layers[layer_count++] = {r.frame, headset ? VIDEO_OUT_BUS_TYPE_MAIN : r.cfg->bus,
+			                         r.premultiplied_alpha};
+		}
 	}
 	if (due) {
-		m_presenter.Present(std::span(layers.data(), count));
+		if (hidden != nullptr) {
+			m_presenter.Discard(*hidden);
+		}
+		if (layer_count != 0) {
+			m_presenter.Present(std::span(layers.data(), layer_count));
+		}
+		if (requests[0].cfg->bus == 0) {
+			Graphics::PerformanceOverlayRecordFlip(requests[0].cfg->width, requests[0].cfg->height);
+		}
 	}
 
 	m_mutex.Lock();
@@ -1451,6 +1494,33 @@ KYTY_SYSV_ABI int VideoOutRegisterBuffers2(int handle, int set_index, int buffer
 	if (option != nullptr) {
 		return VIDEO_OUT_ERROR_INVALID_OPTION;
 	}
+	// The headset port takes reprojection buffers with a zeroed attribute. Each names its eye
+	// texture, which is registered and flipped like a regular surface.
+	const bool headset = ctx->bus == VIDEO_OUT_BUS_TYPE_HMD2_MAIN;
+	if (headset != (category == VIDEO_OUT_BUFFER_ATTRIBUTE_CATEGORY_HMD2)) {
+		return VIDEO_OUT_ERROR_INVALID_CATEGORY;
+	}
+	VideoOutBufferAttribute2                              headset_attribute {};
+	std::array<VideoOutBuffers, VIDEO_OUT_BUFFER_NUM_MAX> headset_buffers {};
+	if (headset) {
+		for (int i = 0; i < buffer_num; i++) {
+			Hmd2::ReprojectionSurface surface {};
+			if (buffers[i].metadata != nullptr ||
+			    !Hmd2::GetReprojectionSurface(buffers[i].data, surface) ||
+			    (i != 0 && (surface.width != headset_attribute.width ||
+			                surface.height != headset_attribute.height ||
+			                surface.pixel_format != headset_attribute.pixel_format))) {
+				return VIDEO_OUT_ERROR_INVALID_VALUE;
+			}
+			headset_attribute.pixel_format = surface.pixel_format;
+			headset_attribute.width        = surface.width;
+			headset_attribute.height       = surface.height;
+			headset_buffers[i].data        = reinterpret_cast<const void*>(surface.address);
+		}
+		attribute = &headset_attribute;
+		buffers   = headset_buffers.data();
+		category  = VIDEO_OUT_BUFFER_ATTRIBUTE_CATEGORY_UNCOMPRESSED;
+	}
 	if (category != VIDEO_OUT_BUFFER_ATTRIBUTE_CATEGORY_UNCOMPRESSED &&
 	    category != VIDEO_OUT_BUFFER_ATTRIBUTE_CATEGORY_COMPRESSED) {
 		return VIDEO_OUT_ERROR_INVALID_CATEGORY;
@@ -1861,7 +1931,8 @@ static int ValidateOutputConfig(int handle, uint64_t mode, const VideoOutOutputO
 		}
 	}
 
-	if (mode != VIDEO_OUT_OUTPUT_MODE_DEFAULT && mode != VIDEO_OUT_OUTPUT_MODE_119_88HZ) {
+	if (mode != VIDEO_OUT_OUTPUT_MODE_DEFAULT && mode != VIDEO_OUT_OUTPUT_MODE_119_88HZ &&
+	    mode != Hmd2::OUTPUT_MODE_89_91HZ && mode != Hmd2::OUTPUT_MODE_119_88HZ) {
 		return VIDEO_OUT_ERROR_UNSUPPORTED_OUTPUT_MODE;
 	}
 

src/graphics/presentation/vrInput.h

@@ -0,0 +1,36 @@
+#ifndef EMULATOR_SRC_GRAPHICS_PRESENTATION_VRINPUT_H_
+#define EMULATOR_SRC_GRAPHICS_PRESENTATION_VRINPUT_H_
+
+namespace Libs::Vr {
+
+// A pose in the tracking space: metres, right-handed, +y up, -z forward (the OpenXR convention).
+struct VrPose {
+	bool  valid          = false;
+	float orientation[4] = {0.0f, 0.0f, 0.0f, 1.0f}; // x, y, z, w
+	float position[3]    = {0.0f, 0.0f, 0.0f};
+};
+
+// One headset controller. Sticks are -1..1 with +y up, the trigger and squeeze 0..1.
+struct VrControllerInput {
+	bool   connected = false;
+	VrPose pose;
+	float  trigger     = 0.0f;
+	float  squeeze     = 0.0f;
+	float  stick_x     = 0.0f;
+	float  stick_y     = 0.0f;
+	bool   primary     = false; // X on the left controller, A on the right one.
+	bool   secondary   = false; // Y on the left controller, B on the right one.
+	bool   stick_click = false;
+	bool   menu        = false;
+};
+
+// Latest head pose; not valid, at the origin, while no headset session tracks the head.
+[[nodiscard]] VrPose OpenXrGetHeadPose();
+
+// Latest state of the left (0) or right (1) controller. While it is not connected the caller keeps
+// using the keyboard or gamepad.
+[[nodiscard]] VrControllerInput OpenXrGetController(int hand);
+
+} // namespace Libs::Vr
+
+#endif // EMULATOR_SRC_GRAPHICS_PRESENTATION_VRINPUT_H_

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 +11,8 @@
 #include "graphics/host_gpu/renderer/render.h"
 #include "graphics/host_gpu/renderer/renderContext.h"
 #include "graphics/host_gpu/vulkanCommon.h"
+#include "graphics/presentation/openxrHeadset.h"
+#include "graphics/presentation/performanceOverlay.h"
 #include "graphics/presentation/presenter.h"
 #include "graphics/presentation/systemOverlay.h"
 #include "graphics/presentation/videoOut.h"
@@ -324,6 +327,7 @@ 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());
 	}
@@ -915,7 +919,8 @@ Presenter::Presenter(WindowContext& window): m_impl(std::make_unique<Impl>(windo
 
 Presenter::~Presenter() = default;
 
-Presenter::Frame& Presenter::PrepareFrame(CommandBuffer& buffer, const ImageInfo& info, bool hdr) {
+Presenter::Frame& Presenter::PrepareFrame(CommandBuffer& buffer, const ImageInfo& info, bool hdr,
+                                          bool headset) {
 	KYTY_PROFILER_FUNCTION();
 	EXIT_IF(buffer.IsInvalid());
 	auto frame_format = info.pixel_format;
@@ -935,6 +940,12 @@ Presenter::Frame& Presenter::PrepareFrame(CommandBuffer& buffer, const ImageInfo
 	frame->Configure(m_impl->window.graphic_ctx,
 	                 {image.backing.extent.width, image.backing.extent.height}, frame_format);
 	frame->CopyFrom(buffer, image);
+	if (headset) {
+		// The frame holds the raw display-encoded pixels the headset needs.
+		frame->Transit(buffer.Handle(), vk::ImageLayout::eTransferSrcOptimal,
+		               vk::AccessFlagBits2::eTransferRead);
+		Vr::OpenXrQueueEyeSource(buffer, frame->image);
+	}
 	return *frame;
 }
 
@@ -1039,8 +1050,8 @@ 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 +1071,7 @@ 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

@@ -1,6 +1,3 @@
-#include <SDL3/SDL.h>
-#include <SDL3/SDL_vulkan.h>
-
 #include "common/assert.h"
 #include "common/common.h"
 #include "common/emulatorConfig.h"
@@ -14,6 +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/openxrHeadset.h"
 #include "graphics/presentation/presenter.h"
 #include "graphics/presentation/systemOverlay.h"
 #include "graphics/presentation/videoOut.h"
@@ -23,6 +21,8 @@
 #include "libs/controller.h"
 #include "loader/systemContent.h"
 
+#include <SDL3/SDL.h>
+#include <SDL3/SDL_vulkan.h>
 #include <algorithm>
 #include <cstdio>
 #include <cstring>
@@ -170,6 +170,8 @@ static void VulkanFindPhysicalDevice(vk::Instance instance, vk::SurfaceKHR surfa
 			LOGF("Vulkan GPU index %d is unavailable; selecting automatically\n",
 			     Config::GetGpuIndex());
 		}
+	} else if (vk::PhysicalDevice headset; Vr::OpenXrPhysicalDevice(instance, headset)) {
+		devices = {headset}; // The headset runtime renders on this GPU.
 	}
 
 	vk::PhysicalDevice  best_device       = nullptr;
@@ -266,7 +268,6 @@ 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 +499,21 @@ 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 +625,13 @@ 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;
@@ -859,8 +878,11 @@ void WindowContext::CreateVulkan() {
 	}
 	VULKAN_HPP_DEFAULT_DISPATCHER.init(get_instance_proc_addr);
 
+	Vr::OpenXrInit();
+
 	VulkanExtensions r;
 	VulkanGetExtensions(r);
+	Vr::OpenXrAppendExtensions(r.required_extensions, r.available_extensions, false);
 	VulkanCheckInstanceVersion();
 
 	vk::ApplicationInfo app_info {};
@@ -966,7 +988,6 @@ 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 +1045,8 @@ 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);
 			}
@@ -1034,6 +1056,7 @@ void WindowContext::CreateVulkan() {
 			device_extensions.push_back(VK_EXT_ATTACHMENT_FEEDBACK_LOOP_LAYOUT_EXTENSION_NAME);
 			device_extensions.push_back(VK_EXT_ATTACHMENT_FEEDBACK_LOOP_DYNAMIC_STATE_EXTENSION_NAME);
 		}
+		Vr::OpenXrAppendExtensions(device_extensions, available_extensions, true);
 	}
 
 	graphic_ctx.device = VulkanCreateDevice(graphic_ctx, device_extensions);
@@ -1051,6 +1074,7 @@ void WindowContext::CreateVulkan() {
 	render_context = std::make_unique<RenderContext>(graphic_ctx);
 	LibKernel::Memory::InstallGpuResources(render_context.get());
 	presenter = std::make_unique<Presenter>(*this);
+	Vr::OpenXrStartSession(graphic_ctx);
 }
 
 void WindowContext::RefreshSurfaceCapabilities() {
@@ -1073,6 +1097,7 @@ void WindowContext::RecreateSurface() {
 }
 
 WindowContext::~WindowContext() {
+	Vr::OpenXrShutdown();
 	ShutdownSystemOverlayInput();
 	presenter.reset();
 	LibKernel::Memory::InstallGpuResources(nullptr);

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

@@ -209,20 +209,20 @@ int BufferTableAttribFromOffset(uint32_t raw_offset, int dword) {
 }
 
 Frontend::EmbeddedFetchPlan DetectEmbeddedVertexFetch(
-    const Decoder::Program& decoded, const ShaderVertexInputInfo* input_info,
+    const Decoder::Program& decoded, const ShaderVertexInputInfo& input_info,
     uint32_t user_data_base, uint32_t user_data_count, uint32_t wave_size) {
-	const uint32_t    vertex_index_reg   = input_info->logical_stage == ShaderType::Local ? 2u : 5u;
-	const uint32_t    instance_index_reg = input_info->logical_stage == ShaderType::Local ? 5u : 8u;
+	const uint32_t    vertex_index_reg   = input_info.logical_stage == ShaderType::Local ? 2u : 5u;
+	const uint32_t    instance_index_reg = input_info.logical_stage == ShaderType::Local ? 5u : 8u;
 	Frontend::EmbeddedFetchPlan data;
-	data.loads.reserve(input_info->resources_num);
+	data.loads.reserve(input_info.resources_num);
 	int32_t vertex_offset_candidate   = -1;
 	int32_t instance_offset_candidate = -1;
 	bool    vertex_offset_conflict    = false;
 	bool    instance_offset_conflict  = false;
 
 	const int shift_regs = 8;
-	const int attrib_reg = input_info->fetch_attrib_reg + shift_regs;
-	const int buffer_reg = input_info->fetch_buffer_reg + shift_regs;
+	const int attrib_reg = input_info.fetch_attrib_reg + shift_regs;
+	const int buffer_reg = input_info.fetch_buffer_reg + shift_regs;
 
 	std::array<EmbeddedFetchSgprInfo, 108> sgprs {};
 	std::array<bool, 256>                 vgpr_is_index {};
@@ -481,18 +481,13 @@ 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 +519,11 @@ 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 +556,30 @@ 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 +606,7 @@ 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 +614,122 @@ 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;
@@ -763,9 +808,9 @@ TranslateResult TranslateProgram(const Decoder::Program& decoded, const CompileO
 	const auto& cfg = *selected_cfg;
 	Frontend::EmbeddedFetchPlan embedded_fetch;
 	if ((options.stage == ShaderType::Vertex || options.stage == ShaderType::Local) &&
-	    options.input_info.vertex != nullptr && options.input_info.vertex->fetch_embedded) {
+	    options.input_info.vertex->fetch_embedded) {
 		embedded_fetch = DetectEmbeddedVertexFetch(
-		    decoded, options.input_info.vertex, options.user_data_base,
+		    decoded, *options.input_info.vertex, options.user_data_base,
 		    static_cast<uint32_t>(options.user_data.size()), options.wave_size);
 		if (!embedded_fetch.loads.empty()) {
 			LOGF("%s embedded vertex fetch plan: detected=%" PRIu64 "\n", GetDumpLabel(options),

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/Tessellation.cpp

@@ -21,15 +21,21 @@ struct TessellationAddress {
 	Kind     kind        = Kind::Unknown;
 	uint32_t coefficient = 0;
 	uint32_t constant    = 0;
+
+	bool operator==(const TessellationAddress&) const = default;
 };
 
-uint32_t ReflectTessellationStride(const Decoder::Program& program, bool local,
-                                   uint32_t control_points, uint32_t input_stride) {
+// Reflects the LS output stride, or the HS output stride and patch constant range.
+void ReflectTessellationLayout(const Decoder::Program& program, bool local,
+                               ShaderTessellationInputInfo& info) {
 	using namespace Decoder;
-	using Address = TessellationAddress;
-	using Kind    = Address::Kind;
-	std::array<Address, IR::NumVectorRegs> registers {};
-	registers[local ? 3u : 1u] =
+	using Address   = TessellationAddress;
+	using Kind      = Address::Kind;
+	using Registers = std::array<Address, IR::NumVectorRegs>;
+
+	const auto control_points = local ? info.input_control_points : info.output_control_points;
+	Registers  entry {}, registers;
+	entry[local ? 3u : 1u] =
 	    local ? Address {Kind::Affine, 1u, 0u} : Address {Kind::PackedControlPoint};
 	const auto constant = [](uint32_t value) { return Address {Kind::Affine, 0u, value}; };
 	const auto read     = [&](const Operand& operand) -> Address {
@@ -79,78 +85,122 @@ uint32_t ReflectTessellationStride(const Decoder::Program& program, bool local,
 		           ? value
 		           : Address {};
 	};
-	uint32_t stride = 0;
-	for (const auto& inst: program.instructions) {
-		// Stage exits preserve the active path's definitions. An internal join
-		// would require merging register definitions.
-		EXIT_NOT_IMPLEMENTED(IsDirectBranch(inst.opcode) &&
-		                     inst.branch_target != program.instructions.back().pc);
-		const bool local_store =
-		    inst.opcode == Opcode::DS_WRITE_B32 || inst.opcode == Opcode::DS_WRITE2_B32;
-		const bool buffer_store = inst.opcode == Opcode::BUFFER_STORE_DWORD ||
-		                          inst.opcode == Opcode::BUFFER_STORE_DWORDX2 ||
-		                          inst.opcode == Opcode::BUFFER_STORE_DWORDX3 ||
-		                          inst.opcode == Opcode::BUFFER_STORE_DWORDX4;
-		const bool control_store =
-		    !local && buffer_store && inst.src2.kind == OperandKind::Sgpr && inst.src2.reg == 2u;
-		const bool control_read =
-		    !local && (inst.opcode == Opcode::DS_READ_B32 || inst.opcode == Opcode::DS_READ2_B32);
-		if ((local && local_store) || control_store || control_read) {
-			const auto address = read(inst.src0);
-			if (address.kind != Kind::Affine) {
-				EXIT("%s tessellation address is not affine at pc 0x%08x\n", local ? "LS" : "HS",
-				     inst.pc);
-			}
-			if (address.coefficient != 0u) {
-				const auto expected = control_read ? input_stride : stride;
-				EXIT_NOT_IMPLEMENTED((address.coefficient & 3u) != 0u ||
-				                     (expected != 0u && expected != address.coefficient));
-				if (!control_read) stride = address.coefficient;
-			}
+	// A branch joins its target with the fallthrough path. Definitions that differ between
+	// the paths are unknown there. Loop headers repeat the walk until back edges are stable.
+	std::unordered_map<uint32_t, Registers> joins, loops;
+	const auto merge = [](Registers& into, const Registers& from) {
+		bool changed = false;
+		for (uint32_t index = 0; index < into.size(); index++) {
+			if (into[index] == from[index]) continue;
+			changed |= into[index] != Address {};
+			into[index] = {};
 		}
-		// Store vdata is a source, despite occupying the decoder's dst field.
-		if (local_store || buffer_store || inst.dst.kind != OperandKind::Vgpr) continue;
-		const auto lhs   = read(inst.src0);
-		const auto rhs   = read(inst.src1);
-		const auto third = read(inst.src2);
-		Address    value;
-		switch (inst.opcode) {
-			case Opcode::V_BFE_U32:
-				if (lhs.kind == Kind::PackedControlPoint && rhs.kind == Kind::Affine &&
-				    rhs.coefficient == 0u && third.kind == Kind::Affine &&
-				    third.coefficient == 0u) {
-					if (rhs.constant == 8u && third.constant == 5u) value = {Kind::Affine, 1u, 0u};
-					if (rhs.constant == 0u && third.constant == 8u) value = constant(0u);
+		return changed;
+	};
+
+	uint32_t stride = 0, patch_begin = UINT32_MAX, patch_end = 0;
+	for (bool repeat = true; repeat;) {
+		repeat    = false;
+		registers = entry;
+		for (const auto& inst: program.instructions) {
+			if (const auto join = joins.find(inst.pc); join != joins.end()) {
+				merge(registers, join->second);
+				joins.erase(join);
+			}
+			if (const auto loop = loops.find(inst.pc); loop != loops.end())
+				merge(registers, loop->second);
+			if (IsDirectBranch(inst.opcode)) {
+				const bool back = inst.branch_target <= inst.pc;
+				const auto [join, inserted] =
+				    (back ? loops : joins).try_emplace(inst.branch_target, registers);
+				const bool changed = !inserted && merge(join->second, registers);
+				repeat |= back && (inserted || changed);
+			}
+			const bool local_store  = IsDsWriteOpcode(inst.opcode);
+			const bool buffer_store = inst.opcode == Opcode::BUFFER_STORE_DWORD ||
+			                          inst.opcode == Opcode::BUFFER_STORE_DWORDX2 ||
+			                          inst.opcode == Opcode::BUFFER_STORE_DWORDX3 ||
+			                          inst.opcode == Opcode::BUFFER_STORE_DWORDX4;
+			const bool control_store = !local && buffer_store &&
+			                           inst.src2.kind == OperandKind::Sgpr && inst.src2.reg == 2u;
+			if ((local && local_store) || control_store) {
+				const auto address = read(inst.src0);
+				if (address.kind != Kind::Affine) {
+					EXIT("%s tessellation address is not affine at pc 0x%08x\n",
+					     local ? "LS" : "HS", inst.pc);
 				}
-				break;
-			case Opcode::V_MUL_U32_U24: value = multiply(low24(lhs), low24(rhs)); break;
-			case Opcode::V_MAD_U32_U24: value = add(multiply(low24(lhs), low24(rhs)), third); break;
-			case Opcode::V_LSHL_ADD_U32:
-				if (rhs.kind == Kind::Affine && rhs.coefficient == 0u && rhs.constant < 32u)
-					value = add(multiply(lhs, constant(1u << rhs.constant)), third);
-				break;
-			case Opcode::V_LSHLREV_B32:
-				if (lhs.kind == Kind::Affine && lhs.coefficient == 0u && lhs.constant < 32u)
-					value = multiply(rhs, constant(1u << lhs.constant));
-				break;
-			case Opcode::V_SUB_NC_U32:
-				if (lhs.kind == Kind::Affine && rhs.kind == Kind::Affine &&
-				    lhs.coefficient >= rhs.coefficient && lhs.constant >= rhs.constant)
-					value = affine(lhs.coefficient - rhs.coefficient, lhs.constant - rhs.constant);
-				break;
-			default: break;
+				if (address.coefficient != 0u) {
+					EXIT_NOT_IMPLEMENTED((address.coefficient & 3u) != 0u ||
+					                     (stride != 0u && stride != address.coefficient));
+					stride = address.coefficient;
+				} else if (control_store) {
+					// Control-point-independent HS outputs are patch constants.
+					const auto begin = address.constant + inst.offset;
+					patch_begin      = std::min(patch_begin, begin);
+					patch_end        = std::max(patch_end, begin + 4u * inst.data_dwords);
+				}
+			}
+			// Store vdata is a source, despite occupying the decoder's dst field.
+			if (local_store || buffer_store || inst.dst.kind != OperandKind::Vgpr) continue;
+			const auto lhs   = read(inst.src0);
+			const auto rhs   = read(inst.src1);
+			const auto third = read(inst.src2);
+			Address    value;
+			switch (inst.opcode) {
+				case Opcode::V_BFE_U32:
+					if (lhs.kind == Kind::PackedControlPoint && rhs.kind == Kind::Affine &&
+					    rhs.coefficient == 0u && third.kind == Kind::Affine &&
+					    third.coefficient == 0u) {
+						if (rhs.constant == 8u && third.constant == 5u)
+							value = {Kind::Affine, 1u, 0u};
+						if (rhs.constant == 0u && third.constant == 8u) value = constant(0u);
+					}
+					break;
+				case Opcode::V_AND_B32:
+					if ((lhs.kind == Kind::PackedControlPoint && rhs == constant(0xffu)) ||
+					    (rhs.kind == Kind::PackedControlPoint && lhs == constant(0xffu)))
+						value = constant(0u);
+					break;
+				case Opcode::V_ADD_NC_U32: value = add(lhs, rhs); break;
+				case Opcode::V_MUL_LO_U32: value = multiply(lhs, rhs); break;
+				case Opcode::V_MUL_U32_U24: value = multiply(low24(lhs), low24(rhs)); break;
+				case Opcode::V_MAD_U32_U24:
+					value = add(multiply(low24(lhs), low24(rhs)), third);
+					break;
+				case Opcode::V_LSHL_ADD_U32:
+					if (rhs.kind == Kind::Affine && rhs.coefficient == 0u && rhs.constant < 32u)
+						value = add(multiply(lhs, constant(1u << rhs.constant)), third);
+					break;
+				case Opcode::V_LSHLREV_B32:
+					if (lhs.kind == Kind::Affine && lhs.coefficient == 0u && lhs.constant < 32u)
+						value = multiply(rhs, constant(1u << lhs.constant));
+					break;
+				case Opcode::V_SUB_NC_U32:
+					if (lhs.kind == Kind::Affine && rhs.kind == Kind::Affine &&
+					    lhs.coefficient >= rhs.coefficient && lhs.constant >= rhs.constant)
+						value =
+						    affine(lhs.coefficient - rhs.coefficient, lhs.constant - rhs.constant);
+					break;
+				default: break;
+			}
+			if (inst.dst.sdwa_sel != 6u || inst.dst.op_sel || inst.dst.omod || inst.dst.clamp ||
+			    inst.dst.dpp)
+				value = {};
+			for (uint32_t index = 0;
+			     index < std::max(inst.data_dwords, 1u) && inst.dst.reg + index < registers.size();
+			     index++) {
+				registers[inst.dst.reg + index] = {};
+			}
+			registers.at(inst.dst.reg) = value;
 		}
-		if (inst.dst.sdwa_sel != 6u || inst.dst.op_sel || inst.dst.omod || inst.dst.clamp ||
-		    inst.dst.dpp)
-			value = {};
-		for (uint32_t index = 0;
-		     index < std::max(inst.data_dwords, 1u) && inst.dst.reg + index < registers.size(); index++) {
-			registers[inst.dst.reg + index] = {};
-		}
-		registers.at(inst.dst.reg) = value;
 	}
-	EXIT_NOT_IMPLEMENTED(stride == 0u);
-	return stride;
+	EXIT_NOT_IMPLEMENTED(!joins.empty() || stride == 0u);
+	(local ? info.ls_stride : info.hs_stride) = stride;
+	if (patch_end != 0u) {
+		EXIT_NOT_IMPLEMENTED(patch_begin < stride * control_points);
+		info.patch_base = patch_begin;
+		info.patch_size = patch_end - patch_begin;
+	}
 }
 
 const IR::Inst* TessellationBufferBase(const IR::Inst& inst) {
@@ -161,6 +211,15 @@ const IR::Inst* TessellationBufferBase(const IR::Inst& inst) {
 	                                                                                 : nullptr;
 }
 
+// Whether an access predicate, possibly a conjunction of EXEC masks, implies a condition.
+bool Implies(IR::Value predicate, IR::Value condition) {
+	predicate = predicate.Resolve();
+	if (predicate == condition.Resolve()) return true;
+	const auto* inst = predicate.TryInstruction();
+	return inst != nullptr && inst->GetOpcode() == IR::ValueOpcode::LogicalAnd &&
+	       (Implies(inst->Arg(0), condition) || Implies(inst->Arg(1), condition));
+}
+
 IR::Value ActiveAddress(IR::Block& block, IR::Block::iterator before, IR::Value value,
                         IR::Value predicate, std::unordered_map<IR::Inst*, IR::Value>& resolved) {
 	using namespace IR;
@@ -169,18 +228,28 @@ IR::Value ActiveAddress(IR::Block& block, IR::Block::iterator before, IR::Value
 	if (source == nullptr) return value;
 	if (const auto found = resolved.find(source); found != resolved.end()) return found->second;
 	if (source->GetOpcode() == ValueOpcode::SelectU32) {
-		return source->Arg(0).Resolve() == predicate.Resolve()
+		return Implies(predicate, source->Arg(0))
 		           ? ActiveAddress(block, before, source->Arg(1), predicate, resolved)
 		           : value;
 	}
-	if (source->GetOpcode() != ValueOpcode::IAdd32 && source->GetOpcode() != ValueOpcode::ISub32)
+	// Pure address arithmetic that stays behind EXEC selects; mul24 lowers to IMul32 of BFEs.
+	const auto opcode = source->GetOpcode();
+	if (opcode != ValueOpcode::IAdd32 && opcode != ValueOpcode::ISub32 &&
+	    opcode != ValueOpcode::IMul32 && opcode != ValueOpcode::BitFieldUExtract)
 		return value;
-	const auto lhs = ActiveAddress(block, before, source->Arg(0), predicate, resolved);
-	const auto rhs = ActiveAddress(block, before, source->Arg(1), predicate, resolved);
-	if (lhs != source->Arg(0).Resolve() || rhs != source->Arg(1).Resolve()) {
-		auto copy = block.PrependNewInst(before, source->GetOpcode(), {lhs, rhs},
-		                                 source->Flags<uint64_t>());
-		value     = Value(&*copy);
+	std::array<Value, 3> args;
+	bool                 changed = false;
+	for (size_t index = 0; index < source->NumArgs(); index++) {
+		args[index] = ActiveAddress(block, before, source->Arg(index), predicate, resolved);
+		changed |= args[index] != source->Arg(index).Resolve();
+	}
+	if (changed) {
+		const auto flags = source->Flags<uint64_t>();
+		const auto copy =
+		    source->NumArgs() == 2u
+		        ? block.PrependNewInst(before, opcode, {args[0], args[1]}, flags)
+		        : block.PrependNewInst(before, opcode, {args[0], args[1], args[2]}, flags);
+		value = Value(&*copy);
 	}
 	resolved.emplace(source, value);
 	return value;
@@ -193,11 +262,12 @@ void AnalyzeTessellationPrograms(std::span<const uint32_t> local, std::span<cons
 	const auto       local_program = Decoder::DecodeFrontProgram(local);
 	Decoder::Program control_program;
 	Decoder::DecodeProgram(control, control_program);
-	info.ls_stride = ReflectTessellationStride(local_program, true, info.input_control_points, 0u);
-	info.hs_stride = ReflectTessellationStride(control_program, false, info.output_control_points,
-	                                           info.ls_stride);
-	LOGF("Tessellation interface: input_cp=%u output_cp=%u ls_stride=%u hs_stride=%u\n",
-	     info.input_control_points, info.output_control_points, info.ls_stride, info.hs_stride);
+	ReflectTessellationLayout(local_program, true, info);
+	ReflectTessellationLayout(control_program, false, info);
+	LOGF("Tessellation interface: input_cp=%u output_cp=%u ls_stride=%u hs_stride=%u "
+	     "patch_base=0x%x patch_size=%u\n",
+	     info.input_control_points, info.output_control_points, info.ls_stride, info.hs_stride,
+	     info.patch_base, info.patch_size);
 }
 
 void LowerTessellationMemory(IR::Program& program, const CompileOptions& options) {
@@ -218,7 +288,7 @@ void LowerTessellationMemory(IR::Program& program, const CompileOptions& options
 		}
 	}
 	ConstantPropagationPass(program.blocks, program.wave_size);
-	uint32_t reads = 0, writes = 0, factors = 0;
+	uint32_t reads = 0, writes = 0, factors = 0, patches = 0;
 	for (auto* block: program.blocks) {
 		for (auto it = block->begin(); it != block->end(); ++it) {
 			auto&      inst   = *it;
@@ -261,15 +331,22 @@ void LowerTessellationMemory(IR::Program& program, const CompileOptions& options
 					kind = TessellationAttribute::Factor;
 					factors += components;
 				} else {
-					kind = write ? TessellationAttribute::ControlOutput
-					             : TessellationAttribute::EvaluationInput;
-					EXIT_NOT_IMPLEMENTED(write
-					                         ? options.stage != ShaderType::TessellationControl
-					                         : options.stage != ShaderType::TessellationEvaluation);
+					// After its barrier, the HS may read back outputs of other control points.
+					const bool output = options.stage == ShaderType::TessellationControl;
+					EXIT_NOT_IMPLEMENTED(
+					    !output && (write || options.stage != ShaderType::TessellationEvaluation));
+					kind = output ? TessellationAttribute::ControlOutput
+					              : TessellationAttribute::EvaluationInput;
 					if (address.IsImmediate() &&
 					    address.U32() >= tess.hs_stride * tess.output_control_points) {
-						EXIT_NOT_IMPLEMENTED(!write);
-						kind = TessellationAttribute::PatchOutput;
+						const auto begin = address.U32() + memory.offset;
+						EXIT_NOT_IMPLEMENTED(begin < tess.patch_base ||
+						                     begin + 4u * components >
+						                         tess.patch_base + tess.patch_size);
+						kind = output ? TessellationAttribute::PatchOutput
+						              : TessellationAttribute::PatchInput;
+						// HS read-backs must not hide misplaced patch stores from the check below.
+						if (write || !output) patches += components;
 					}
 				}
 			} else {
@@ -320,11 +397,14 @@ void LowerTessellationMemory(IR::Program& program, const CompileOptions& options
 	ConstantPropagationPass(program.blocks, program.wave_size);
 	RemoveIdentities(program);
 	EliminateDeadCode(program.blocks);
-	LOGF("%s tessellation lowering: reads=%u writes=%u factors=%u\n",
+	// Patch constant stores whose EXEC selects were not stripped would land in control points.
+	EXIT_NOT_IMPLEMENTED(options.stage == ShaderType::TessellationControl &&
+	                     tess.patch_size != 0u && patches == 0u);
+	LOGF("%s tessellation lowering: reads=%u writes=%u factors=%u patches=%u\n",
 	     options.stage == ShaderType::Local                 ? "LS"
 	     : options.stage == ShaderType::TessellationControl ? "HS"
 	                                                       : "TES",
-	     reads, writes, factors);
+	     reads, writes, factors, patches);
 }
 
 } // namespace Libs::Graphics::ShaderRecompiler

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

@@ -233,8 +233,8 @@ std::vector<uint32_t> EmitProgram(IR::Program& program, ShaderStageInputInfo inp
 	        ? 2u
 	        : 1u;
 	DefineModule(state);
-	ValidateNativeProgram(program, program.stage == ShaderType::Compute &&
-	                                   input_info.compute != nullptr && input_info.compute->lds_storage);
+	ValidateNativeProgram(program,
+	                      program.stage == ShaderType::Compute && input_info.compute->lds_storage);
 	EmitProgram(state);
 	state.builder.AddEntryPoint(ExecutionModelForStage(state.program.stage), state.main_func,
 	                            "main", state.interface_variables);

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();
@@ -77,7 +115,7 @@ uint32_t EmitF32ToU32(EmitterState& state, uint32_t src, bool signed_value) {
 	} else {
 		state.builder.AddFunction(spv::OpConvertFToU, TypeU32(state), converted_raw, trunc);
 	}
-	const auto nan = EmitClassifyF32(state, src).nan;
+	const auto nan = Unary(state, spv::OpIsNan, TypeBool(state), src);
 	if (signed_value) {
 		const auto below = Binary(state, spv::OpFOrdLessThanEqual, TypeBool(state), src,
 		                          ConstantF32(state, 0xcf000000u));
@@ -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);
@@ -210,31 +235,16 @@ 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) {
-	// 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);
+	// The clamp modifier runs in DX10 clamp mode: NaN becomes 0, as NClamp's NMax step does.
+	return EmitExt(state, TypeF32(state), GLSLstd450NClamp,
+	               {arg0, ConstantF32(state, 0), ConstantF32(state, 0x3f800000u)});
 }
 
 uint32_t EmitSMulHi(EmitterState& state, uint32_t arg0, uint32_t arg1) {

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

@@ -17,16 +17,9 @@ uint32_t EmitCompareU32Constant(EmitterState& state, spv::Op opcode, uint32_t va
 }
 
 F32Class EmitClassifyF32Bits(EmitterState& state, uint32_t bits) {
-	F32Class cls;
-	cls.bits            = bits;
-	const auto abs_bits = EmitAndConstant(state, cls.bits, 0x7fffffffu);
-	cls.nan  = EmitCompareU32Constant(state, spv::OpUGreaterThan, abs_bits, 0x7f800000u);
-	cls.zero = EmitCompareU32Constant(state, spv::OpIEqual, abs_bits, 0);
-	return cls;
-}
-
-F32Class EmitClassifyF32(EmitterState& state, uint32_t value) {
-	return EmitClassifyF32Bits(state, EmitBitcastF32ToU32(state, value));
+	const auto abs_bits = EmitAndConstant(state, bits, 0x7fffffffu);
+	return {EmitCompareU32Constant(state, spv::OpUGreaterThan, abs_bits, 0x7f800000u),
+	        EmitCompareU32Constant(state, spv::OpIEqual, abs_bits, 0)};
 }
 
 uint32_t EmitClassMaskBitMatch(EmitterState& state, uint32_t mask, uint32_t bit,
@@ -100,25 +93,54 @@ uint32_t EmitClassMaskF16(EmitterState& state, uint32_t bits, uint32_t mask) {
 	return EmitClassMaskBits(state, bits, mask, true);
 }
 
+static uint32_t EmitMinMaxF32(EmitterState& state, uint32_t lhs, uint32_t rhs, uint32_t lhs_bits,
+                              uint32_t rhs_bits, bool max_value) {
+	// RDNA2 V_MIN/V_MAX_F32 return the other source for a NaN source (S1 when both are NaN)
+	// and order -0 below +0. An unordered compare keeps S0 for a NaN S1 and, unlike NMin/NMax,
+	// always returns the exact bits of one source.
+	const auto either    = EmitOrU32(state, lhs_bits, rhs_bits);
+	const auto both_zero = EmitCompareU32Constant(state, spv::OpIEqual,
+	                                              EmitAndConstant(state, either, 0x7fffffffu), 0);
+	const auto zero =
+	    EmitBitcastU32ToF32(state, max_value ? EmitAndU32(state, lhs_bits, rhs_bits) : either);
+	const auto keep_lhs =
+	    Binary(state, max_value ? spv::OpFUnordGreaterThanEqual : spv::OpFUnordLessThan,
+		       TypeBool(state), lhs, rhs);
+	const auto ordered = Select(state, TypeF32(state), keep_lhs, lhs, rhs);
+	const auto numeric = Select(state, TypeF32(state), both_zero, zero, ordered);
+	return Select(state, TypeF32(state), Unary(state, spv::OpIsNan, TypeBool(state), lhs), rhs,
+	              numeric);
+}
+
 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, EmitBitcastF32ToU32(state, lhs),
+	                     EmitBitcastF32ToU32(state, rhs), max_value);
+}
 
-	const auto numeric_cond = state.builder.AllocateId();
-	state.builder.AddFunction(max_value ? spv::OpFOrdGreaterThanEqual : spv::OpFOrdLessThan,
-	                          TypeBool(state), numeric_cond, lhs, rhs);
-	const auto ordered_bits =
-	    EmitSelectValueU32(state, numeric_cond, lhs_class.bits, rhs_class.bits);
-
-	const auto both_zero    = EmitLogicalAndBool(state, lhs_class.zero, rhs_class.zero);
-	const auto zero_bits    = max_value ? EmitAndU32(state, lhs_class.bits, rhs_class.bits)
-	                                    : EmitOrU32(state, lhs_class.bits, rhs_class.bits);
-	const auto numeric_bits = EmitSelectValueU32(state, both_zero, zero_bits, ordered_bits);
-
-	const auto rhs_nan_bits =
-	    EmitSelectValueU32(state, rhs_class.nan, lhs_class.bits, numeric_bits);
-	const auto result_bits = EmitSelectValueU32(state, lhs_class.nan, rhs_class.bits, rhs_nan_bits);
-	return EmitBitcastU32ToF32(state, result_bits);
+uint32_t EmitFPMedTri32(EmitterState& state, uint32_t a, uint32_t b, uint32_t c) {
+	const auto a_bits = EmitBitcastF32ToU32(state, a);
+	const auto b_bits = EmitBitcastF32ToU32(state, b);
+	const auto c_bits = EmitBitcastF32ToU32(state, c);
+	const auto min_ab = EmitMinMaxF32(state, a, b, a_bits, b_bits, false);
+	const auto min3 =
+	    EmitMinMaxF32(state, min_ab, c, EmitBitcastF32ToU32(state, min_ab), c_bits, false);
+	const auto max_ab = EmitMinMaxF32(state, a, b, a_bits, b_bits, true);
+	const auto max3 =
+	    EmitMinMaxF32(state, max_ab, c, EmitBitcastF32ToU32(state, max_ab), c_bits, true);
+	// RDNA2 compares numerically against MAX3, including ties between signed zeros, then takes
+	// V_MAX_F32 of the other two sources in their original order.
+	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 x        = Select(state, TypeF32(state), a_is_max, b, a);
+	const auto y =
+	    Select(state, TypeF32(state), EmitLogicalOrBool(state, a_is_max, b_is_max), c, b);
+	const auto median = EmitMinMaxF32Value(state, x, y, true);
+	const auto is_nan = [&](uint32_t value) {
+		return Unary(state, spv::OpIsNan, TypeBool(state), value);
+	};
+	const auto any_nan =
+	    EmitLogicalOrBool(state, EmitLogicalOrBool(state, is_nan(a), is_nan(b)), is_nan(c));
+	return Select(state, TypeF32(state), any_nan, min3, median);
 }
 
 uint32_t EmitFlushF32DenormToSignedZero(EmitterState& state, uint32_t value) {

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

@@ -53,8 +53,7 @@ uint32_t EmitBuiltinU32(EmitterState& state, IR::StageInputKind kind, uint32_t c
 		const auto bits  = state.builder.AllocateId();
 		state.builder.AddFunction(spv::OpLoad, TypeBool(state), value, variable);
 		// SPI_BARYC_CNTL.FRONT_FACE_ALL_BITS selects uint masks or float +/-1.
-		const bool all_bits =
-		    state.input_info.pixel != nullptr && state.input_info.pixel->ps_front_face_all_bits;
+		const bool all_bits = state.input_info.pixel->ps_front_face_all_bits;
 		const auto front = ConstantU32(state, all_bits ? 0xffffffffu : 0x3f800000u);
 		const auto back  = ConstantU32(state, all_bits ? 0u : 0xbf800000u);
 		state.builder.AddFunction(spv::OpSelect, TypeU32(state), bits, value, front, back);
@@ -350,6 +349,38 @@ uint32_t ExportVector(ValueEmitContext& ctx, uint32_t data, const IR::ExportInfo
 	return vector;
 }
 
+uint32_t DistancePointer(EmitterState& state, IR::StageOutputKind kind, uint32_t index) {
+	if (state.program.stage == ShaderType::Mesh) {
+		return MeshOutputPointer(state, kind, index);
+	}
+	const auto variable = kind == IR::StageOutputKind::ClipDistance ? state.clip_distance_variable
+	                                                                : state.cull_distance_variable;
+	const auto pointer  = state.builder.AllocateId();
+	state.builder.AddFunction(spv::OpAccessChain,
+	                          TypePointer(state, spv::StorageClassOutput, TypeF32(state)), pointer,
+	                          variable, ConstantU32(state, index));
+	return pointer;
+}
+
+// The clipper keeps the half space where dot(plane, position) >= 0 for each UCP_ENA plane.
+void EmitUserClipDistances(EmitterState& state, uint32_t position) {
+	uint32_t index = 0;
+	for (auto mask = state.user_clip_planes; mask != 0; mask &= mask - 1u) {
+		const auto dword = IR::PushData::UserClipPlaneDword + 4u * std::countr_zero(mask);
+		uint32_t   components[4] {};
+		for (uint32_t i = 0; i < 4; i++) {
+			components[i] = Unary(state, spv::OpBitcast, TypeF32(state),
+			                      EmitPushDataLoad(state, ConstantU32(state, dword + i)));
+		}
+		const auto plane = state.builder.AllocateId();
+		state.builder.AddFunction(spv::OpCompositeConstruct, TypeF32Vector(state, 4), plane,
+		                          components[0], components[1], components[2], components[3]);
+		state.builder.AddFunction(
+		    spv::OpStore, DistancePointer(state, IR::StageOutputKind::ClipDistance, index++),
+		    Binary(state, spv::OpDot, TypeF32(state), plane, position));
+	}
+}
+
 void EmitAuxPositionExport(ValueEmitContext& ctx, uint32_t data, const IR::ExportInfo& exp) {
 	auto& state = ctx.state;
 	for (uint32_t component = 0; component < 4; component++) {
@@ -390,23 +421,13 @@ void EmitAuxPositionExport(ValueEmitContext& ctx, uint32_t data, const IR::Expor
 			state.builder.AddFunction(spv::OpStore, state.point_size_variable, f32);
 			continue;
 		}
-		auto StoreDistance = [&](IR::StageOutputKind kind, uint32_t variable, uint32_t index) {
-			if (index == UINT32_MAX) {
-				return;
+		auto StoreDistance = [&](IR::StageOutputKind kind, uint32_t index) {
+			if (index != UINT32_MAX) {
+				state.builder.AddFunction(spv::OpStore, DistancePointer(state, kind, index), f32);
 			}
-			uint32_t pointer;
-			if (state.program.stage == ShaderType::Mesh) {
-				pointer = MeshOutputPointer(state, kind, index);
-			} else {
-				pointer = state.builder.AllocateId();
-				state.builder.AddFunction(spv::OpAccessChain,
-				                          TypePointer(state, spv::StorageClassOutput, TypeF32(state)),
-				                          pointer, variable, ConstantU32(state, index));
-			}
-			state.builder.AddFunction(spv::OpStore, pointer, f32);
 		};
-		StoreDistance(IR::StageOutputKind::ClipDistance, state.clip_distance_variable, output.clip_distance);
-		StoreDistance(IR::StageOutputKind::CullDistance, state.cull_distance_variable, output.cull_distance);
+		StoreDistance(IR::StageOutputKind::ClipDistance, output.clip_distance);
+		StoreDistance(IR::StageOutputKind::CullDistance, output.cull_distance);
 	}
 }
 
@@ -597,6 +618,9 @@ void EmitSetAttribute(ValueEmitContext& ctx, const IR::Inst& inst) {
 		    state.input_info.vertex->clip_space.enabled) {
 			value = ConvertPositionToClipSpace(state, value);
 		}
+		if (exp.kind == IR::ExportTargetKind::Position && state.user_clip_planes != 0) {
+			EmitUserClipDistances(state, value);
+		}
 		if (state.program.stage == ShaderType::Mesh) {
 			const auto kind = exp.kind == IR::ExportTargetKind::Position
 			                      ? IR::StageOutputKind::Position
@@ -609,7 +633,6 @@ void EmitSetAttribute(ValueEmitContext& ctx, const IR::Inst& inst) {
 				const auto equal = state.builder.AllocateId();
 				const auto invalid = state.builder.AllocateId();
 				const auto distance = state.builder.AllocateId();
-				const auto distance_pointer = state.builder.AllocateId();
 				state.builder.AddFunction(spv::OpFOrdEqual, TypeBoolVector(state, 4), equal,
 				                          value, zero);
 				state.builder.AddFunction(spv::OpAll, TypeBool(state), invalid, equal);
@@ -618,11 +641,10 @@ void EmitSetAttribute(ValueEmitContext& ctx, const IR::Inst& inst) {
 				state.builder.AddFunction(spv::OpSelect, TypeF32(state), distance, invalid,
 				                          ConstantF32Value(state, -1.0f),
 				                          ConstantF32Value(state, 0.0f));
-				state.builder.AddFunction(
-				    spv::OpAccessChain, TypePointer(state, spv::StorageClassOutput, TypeF32(state)),
-				    distance_pointer, state.clip_distance_variable,
-				    ConstantU32(state, state.invalid_position_clip_distance));
-				state.builder.AddFunction(spv::OpStore, distance_pointer, distance);
+				state.builder.AddFunction(spv::OpStore,
+				                          DistancePointer(state, IR::StageOutputKind::ClipDistance,
+				                                          state.invalid_position_clip_distance),
+				                          distance);
 				static std::atomic_bool logged = false;
 				if (!logged.exchange(true, std::memory_order_relaxed)) {
 					Log::WriteToConsoleAndLog(
@@ -655,6 +677,14 @@ 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);

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

@@ -80,7 +80,7 @@ struct EmitterState {
 			lds_storage_class = spv::StorageClassWorkgroup;
 		}
 		if (program.info.uses_lds && program.stage == ShaderType::Compute &&
-		    input_info.compute != nullptr && input_info.compute->lds_storage) {
+		    input_info.compute->lds_storage) {
 			lds_storage_class = spv::StorageClassStorageBuffer;
 		}
 	}
@@ -104,9 +104,8 @@ struct EmitterState {
 	std::array<uint32_t, 3>                          u32_vector_types {};
 	std::array<uint32_t, 3>                          i32_vector_types {};
 	std::array<uint32_t, 3>                          f32_vector_types {};
-	std::array<uint32_t, 6>                          tess_variables {};
+	std::array<uint32_t, IR::NumTessellationAttributes> tess_variables {};
 	uint32_t                                         tess_inner_variable = 0;
-	uint32_t                                         tess_patch_base     = 0;
 
 	uint32_t                                         lane_count              = 1;
 	uint32_t                                         lane_half               = 0;
@@ -148,6 +147,7 @@ struct EmitterState {
 	uint32_t                   point_size_variable                   = 0;
 	uint32_t                   clip_distance_variable                = 0;
 	uint32_t                   invalid_position_clip_distance        = UINT32_MAX;
+	uint32_t                   user_clip_planes                      = 0;
 	uint32_t                   cull_distance_variable                = 0;
 	uint32_t                   layer_variable                        = 0;
 	uint32_t                   viewport_index_variable               = 0;
@@ -285,7 +285,6 @@ struct ImageSampleLayout {
 };
 
 struct F32Class {
-	uint32_t bits = 0;
 	uint32_t nan  = 0;
 	uint32_t zero = 0;
 };
@@ -510,8 +509,6 @@ inline constexpr auto EmitLogicalNotBool = EmitNative<spv::OpLogicalNot, IR::Typ
 
 F32Class EmitClassifyF32Bits(EmitterState& state, uint32_t bits);
 
-F32Class EmitClassifyF32(EmitterState& state, uint32_t value);
-
 uint32_t EmitClassMaskBitMatch(EmitterState& state, uint32_t mask, uint32_t bit,
                                uint32_t class_match);
 

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

@@ -157,9 +157,25 @@ void EmitMemoryOffsets(EmitterState& state) {
 	}
 }
 
+// Stages without a workgroup emulate LDS with one private array per invocation. A GPU keeps
+// hundreds of thousands of invocations in flight, so this size multiplies into driver local memory:
+// a fixed 8192 dwords (32 KB) asked for gigabytes and killed the device (a write fault in the
+// first heavy pixel shader that touched LDS). A pixel shader can only address the LDS that
+// EXTRA_LDS_SIZE reserves, a vertex shader the LDS_SIZE of its GS stage; hardware reads zero and
+// drops writes beyond it, as the bounds checks do.
+constexpr uint32_t FUNCTION_LDS_DWORDS = 1024u;
+
 uint32_t LdsDwordCount(const EmitterState& state) {
-	const auto* workgroup = ShaderWorkgroupInput(state.program.stage, state.input_info);
-	return workgroup != nullptr ? workgroup->lds_size_dwords : 8192u;
+	if (const auto* workgroup = ShaderWorkgroupInput(state.program.stage, state.input_info)) {
+		return workgroup->lds_size_dwords;
+	}
+	if (state.program.stage == ShaderType::Pixel) {
+		return std::max(state.input_info.pixel->lds_size_dwords, 1u);
+	}
+	if (state.program.stage == ShaderType::Vertex) {
+		return std::max(state.input_info.vertex->lds_size_dwords, 1u);
+	}
+	return FUNCTION_LDS_DWORDS;
 }
 
 void EnsureLdsStorage(EmitterState& state) {

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

@@ -244,7 +244,7 @@ void DefineDescriptors(EmitterState& state) {
 	counts[static_cast<size_t>(Kind::ShaderData)] =
 	    bindings.ShaderDataDwords() != 0 && !bindings.UsesPushData();
 	if (state.program.bindings.UsesPushData() || state.program.stage == ShaderType::Mesh ||
-	    state.object_primitive_id_variable != 0) {
+	    state.object_primitive_id_variable != 0 || state.user_clip_planes != 0) {
 		const auto type              = PushConstantBlockType(state);
 		state.push_constant_variable = state.builder.DefineGlobalVariable(
 		    TypePointer(state, spv::StorageClassPushConstant, type), spv::StorageClassPushConstant);
@@ -600,14 +600,24 @@ void DefineOutputs(EmitterState& state) {
 			cull_distance_count = std::max(cull_distance_count, output.index + 1);
 		}
 	}
+	const bool position = std::ranges::any_of(state.outputs, [](const OutputBinding& output) {
+		return output.kind == IR::StageOutputKind::Position;
+	});
+	if (position && state.input_info.vertex->user_clip_planes != 0) {
+		// PipelineCache rejects UCP_ENA with shader clip or cull distances.
+		EXIT_IF(clip_distance_count + cull_distance_count != 0);
+		state.user_clip_planes = state.input_info.vertex->user_clip_planes;
+		for (auto mask = state.user_clip_planes; mask != 0; mask &= mask - 1u) {
+			state.outputs.push_back(
+			    {{IR::StageOutputKind::ClipDistance, clip_distance_count++, 0, "gl_ClipDistance"}});
+		}
+	}
 	if (state.program.stage == ShaderType::Mesh) {
 		DefineMeshOutputs(state, clip_distance_count, cull_distance_count);
 		return;
 	}
-	if (state.program.stage == ShaderType::Vertex && clip_distance_count + cull_distance_count < 8u &&
-	    std::ranges::any_of(state.outputs, [](const OutputBinding& output) {
-		    return output.kind == IR::StageOutputKind::Position;
-	    })) {
+	if (state.program.stage == ShaderType::Vertex &&
+	    clip_distance_count + cull_distance_count < 8u && position) {
 		// Reserve one plane for the enabled PA_CL_CLIP_CNTL clipping-error cull.
 		state.invalid_position_clip_distance = clip_distance_count++;
 		state.outputs.push_back({{IR::StageOutputKind::ClipDistance,

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

@@ -233,14 +233,17 @@ void Invoke(Return (*emit)(Context&, Args...), ValueEmitContext& ctx, const IR::
 }
 
 void EmitDirectInstruction(ValueEmitContext& ctx, const IR::Inst& inst) {
-	if (ctx.half != 0 && (inst.GetOpcode() == IR::ValueOpcode::Ballot ||
-	                     inst.GetOpcode() == IR::ValueOpcode::ReadFirstLane ||
-	                     inst.GetOpcode() == IR::ValueOpcode::SubgroupUMinU32)) {
-		// These operations combine both emulated halves into one whole-wave result.
+	using Op = IR::ValueOpcode;
+	const auto op = inst.GetOpcode();
+	if (ctx.half != 0 &&
+	    (op == Op::Ballot || op == Op::ReadFirstLane || op == Op::SubgroupUMinU32 ||
+	     op == Op::ReadConstBuffer ||
+	     (op == Op::LoadAddressU32 && ctx.Memory(inst).kind == IR::ResourceKind::ScalarAddress))) {
+		// Whole-wave results and SMEM loads (SGPR operands, no EXEC) are identical in both halves.
 		ctx.Define(inst, ctx.other_half->Result(inst));
 		return;
 	}
-	switch (inst.GetOpcode()) {
+	switch (op) {
 #define VALUE_OPCODE(name, ...)                                                                    \
 	case IR::ValueOpcode::name: return Invoke(Emit##name, ctx, inst);
 #include "graphics/shader/recompiler/ir/opcodes/ValueOpcodes.inc"
@@ -307,6 +310,7 @@ 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);
 			}
@@ -622,6 +626,12 @@ void EmitProgram(EmitterState& state) {
 	}
 	if (state.program.dispatcher_fallback) {
 		auto& dispatch = dispatcher.emplace();
+		// Allocate both halves in program order so the SPIR-V is identical across runs.
+		const auto add_spill = [&](const IR::Inst* inst) {
+			for (uint32_t half = 0; half < state.lane_count; half++) {
+				dispatch.spills[half].emplace(inst, state.builder.AllocateId());
+			}
+		};
 		for (const auto* block: program.blocks) {
 			for (const auto& inst: *block) {
 				if (inst.GetOpcode() != IR::ValueOpcode::Phi) {
@@ -631,7 +641,7 @@ void EmitProgram(EmitterState& state) {
 					ctx.Fail(inst, "cannot be stored by the dispatcher");
 					break;
 				}
-				dispatch.spills[0].emplace(&inst, state.builder.AllocateId());
+				add_spill(&inst);
 			}
 		}
 		const auto mark_cross_block = [&](IR::Value value, const IR::Block* consumer) {
@@ -646,7 +656,7 @@ void EmitProgram(EmitterState& state) {
 				return;
 			}
 			if (!dispatch.spills[0].contains(definition)) {
-				dispatch.spills[0].emplace(definition, state.builder.AllocateId());
+				add_spill(definition);
 			}
 		};
 		for (const auto* block: program.blocks) {
@@ -668,12 +678,7 @@ void EmitProgram(EmitterState& state) {
 		dispatch.continue_label     = state.builder.AllocateId();
 		dispatch.merge_label        = state.builder.AllocateId();
 		ctx.dispatcher_spills       = &dispatch.spills[0];
-		if (state.lane_count == 2) {
-			for (const auto& [inst, id]: dispatch.spills[0]) {
-				dispatch.spills[1].emplace(inst, state.builder.AllocateId());
-			}
-			high.dispatcher_spills = &dispatch.spills[1];
-		}
+		high.dispatcher_spills      = &dispatch.spills[1];
 	}
 	DefineGetBdaPointer(state);
 	DefineIndirectBufferLoads(ctx);

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

@@ -5,6 +5,17 @@
 namespace Libs::Graphics::ShaderRecompiler::Spirv::Emitter {
 namespace {
 
+bool IsTessellationInput(IR::TessellationAttribute kind) {
+	using Attribute = IR::TessellationAttribute;
+	return kind == Attribute::ControlInput || kind == Attribute::EvaluationInput ||
+	       kind == Attribute::PatchInput;
+}
+
+bool IsTessellationPatch(IR::TessellationAttribute kind) {
+	using Attribute = IR::TessellationAttribute;
+	return kind == Attribute::PatchOutput || kind == Attribute::PatchInput;
+}
+
 uint32_t TessellationPointer(ValueEmitContext& ctx, const IR::Inst& inst) {
 	auto& state          = ctx.state;
 	using Attribute      = IR::TessellationAttribute;
@@ -12,8 +23,9 @@ uint32_t TessellationPointer(ValueEmitContext& ctx, const IR::Inst& inst) {
 	const auto& tess     = state.input_info.vertex->tess;
 	const auto  variable = state.tess_variables.at(static_cast<uint32_t>(kind));
 	EXIT_IF(variable == 0);
-	const auto input   = kind == Attribute::ControlInput || kind == Attribute::EvaluationInput;
-	const auto storage = input ? spv::StorageClassInput : spv::StorageClassOutput;
+	const bool patch = IsTessellationPatch(kind);
+	const auto storage =
+	    IsTessellationInput(kind) ? spv::StorageClassInput : spv::StorageClassOutput;
 	const auto pointer = state.builder.AllocateId();
 	if (kind == Attribute::Factor) {
 		EXIT_NOT_IMPLEMENTED(!inst.Arg(1).IsImmediate());
@@ -26,15 +38,13 @@ uint32_t TessellationPointer(ValueEmitContext& ctx, const IR::Inst& inst) {
 		return pointer;
 	}
 	auto address = ctx.Arg(inst, 1);
-	if (kind == Attribute::PatchOutput) {
-		address =
-		    EmitBinaryU32(state, spv::OpISub, address, ConstantU32(state, state.tess_patch_base));
+	if (patch) {
+		address = EmitBinaryU32(state, spv::OpISub, address, ConstantU32(state, tess.patch_base));
 	}
 	const bool local  = kind == Attribute::LocalOutput || kind == Attribute::ControlInput;
 	const auto stride = local ? tess.ls_stride : tess.hs_stride;
-	const auto offset = kind == Attribute::PatchOutput ? address
-	                                                   : EmitBinaryU32(state, spv::OpUMod, address,
-	                                                                   ConstantU32(state, stride));
+	const auto offset =
+	    patch ? address : EmitBinaryU32(state, spv::OpUMod, address, ConstantU32(state, stride));
 	const auto attribute =
 	    EmitBinaryU32(state, spv::OpShiftRightLogical, offset, ConstantU32(state, 4));
 	const auto component =
@@ -42,12 +52,14 @@ uint32_t TessellationPointer(ValueEmitContext& ctx, const IR::Inst& inst) {
 	                  EmitBinaryU32(state, spv::OpShiftRightLogical, offset, ConstantU32(state, 2)),
 	                  ConstantU32(state, 3));
 	const auto type = TypePointer(state, storage, TypeU32(state));
-	if (kind == Attribute::LocalOutput || kind == Attribute::PatchOutput) {
+	if (kind == Attribute::LocalOutput || patch) {
 		state.builder.AddFunction(spv::OpAccessChain, type, pointer, variable, attribute,
 		                          component);
 	} else {
+		// An HS writes only its own control point but may read any other.
 		const auto vertex =
-		    kind == Attribute::ControlOutput
+		    kind == Attribute::ControlOutput &&
+		            inst.GetOpcode() == IR::ValueOpcode::SetTessellationAttribute
 		        ? EmitLaneId(state)
 		        : EmitBinaryU32(state, spv::OpUDiv, address, ConstantU32(state, stride));
 		state.builder.AddFunction(spv::OpAccessChain, type, pointer, variable, vertex, attribute,
@@ -60,20 +72,12 @@ uint32_t TessellationPointer(ValueEmitContext& ctx, const IR::Inst& inst) {
 
 void DefineTessellationInterfaces(EmitterState& state) {
 	using Attribute = IR::TessellationAttribute;
-	std::array<bool, 6> used {};
-	uint32_t            patch_begin = UINT32_MAX, patch_end = 0;
+	std::array<bool, IR::NumTessellationAttributes> used {};
 	for (const auto* block: state.program.blocks) {
 		for (const auto& inst: *block) {
-			if (inst.GetOpcode() != IR::ValueOpcode::GetTessellationAttribute &&
-			    inst.GetOpcode() != IR::ValueOpcode::SetTessellationAttribute)
-				continue;
-			const auto kind = inst.Arg(0).U32();
-			used.at(kind)   = true;
-			if (kind == static_cast<uint32_t>(Attribute::PatchOutput)) {
-				EXIT_NOT_IMPLEMENTED(!inst.Arg(1).IsImmediate());
-				patch_begin = std::min(patch_begin, inst.Arg(1).U32());
-				patch_end   = std::max(patch_end, inst.Arg(1).U32() + 4u);
-			}
+			if (inst.GetOpcode() == IR::ValueOpcode::GetTessellationAttribute ||
+			    inst.GetOpcode() == IR::ValueOpcode::SetTessellationAttribute)
+				used.at(inst.Arg(0).U32()) = true;
 		}
 	}
 	if (std::ranges::none_of(used, [](bool value) { return value; })) return;
@@ -83,15 +87,16 @@ void DefineTessellationInterfaces(EmitterState& state) {
 	};
 	for (uint32_t index = 0; index < used.size(); index++) {
 		if (!used[index]) continue;
-		const auto kind    = static_cast<Attribute>(index);
-		const bool input   = kind == Attribute::ControlInput || kind == Attribute::EvaluationInput;
-		const auto storage = input ? spv::StorageClassInput : spv::StorageClassOutput;
-		uint32_t   type;
+		const auto kind  = static_cast<Attribute>(index);
+		const bool patch = IsTessellationPatch(kind);
+		const auto storage =
+		    IsTessellationInput(kind) ? spv::StorageClassInput : spv::StorageClassOutput;
+		uint32_t type;
 		if (kind == Attribute::Factor) {
 			type = array(TypeF32(state), 4u);
-		} else if (kind == Attribute::PatchOutput) {
-			state.tess_patch_base = patch_begin;
-			type = array(TypeU32Vector(state, 4), (patch_end - patch_begin + 15u) / 16u);
+		} else if (patch) {
+			// HS and TES share the reflected patch layout so their interfaces match.
+			type = array(TypeU32Vector(state, 4), (tess.patch_size + 15u) / 16u);
 		} else {
 			const bool local  = kind == Attribute::LocalOutput || kind == Attribute::ControlInput;
 			const auto stride = local ? tess.ls_stride : tess.hs_stride;
@@ -112,11 +117,10 @@ void DefineTessellationInterfaces(EmitterState& state) {
 			state.builder.AddAnnotation(spv::OpDecorate, state.tess_inner_variable,
 			                            spv::DecorationPatch);
 		} else {
-			state.builder.AddAnnotation(
-			    spv::OpDecorate, variable, spv::DecorationLocation,
-			    kind == Attribute::PatchOutput ? (tess.hs_stride + 15u) / 16u : 0u);
+			state.builder.AddAnnotation(spv::OpDecorate, variable, spv::DecorationLocation,
+			                            patch ? (tess.hs_stride + 15u) / 16u : 0u);
 		}
-		if (kind == Attribute::Factor || kind == Attribute::PatchOutput) {
+		if (kind == Attribute::Factor || patch) {
 			state.builder.AddAnnotation(spv::OpDecorate, variable, spv::DecorationPatch);
 		}
 	}
@@ -124,7 +128,9 @@ void DefineTessellationInterfaces(EmitterState& state) {
 
 void DefineTessellationExecutionModes(EmitterState& state) {
 	const auto& tess = state.input_info.vertex->tess;
-	EXIT_NOT_IMPLEMENTED(tess.domain != 1u || tess.partitioning != 2u ||
+	// CxVgtTfParam: triangles, integer or fractional-odd partitioning, clockwise output.
+	EXIT_NOT_IMPLEMENTED(tess.domain != 1u ||
+	                     (tess.partitioning != 0u && tess.partitioning != 2u) ||
 	                     tess.output_topology != 2u);
 	state.builder.RequireCapability(spv::CapabilityTessellation);
 	if (state.program.stage == ShaderType::TessellationControl) {
@@ -132,7 +138,9 @@ void DefineTessellationExecutionModes(EmitterState& state) {
 		                               tess.output_control_points);
 	} else {
 		state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeTriangles);
-		state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeSpacingFractionalOdd);
+		state.builder.AddExecutionMode(
+		    state.main_func, tess.partitioning == 0u ? spv::ExecutionModeSpacingEqual
+		                                             : spv::ExecutionModeSpacingFractionalOdd);
 		state.builder.AddExecutionMode(state.main_func, spv::ExecutionModeVertexOrderCw);
 	}
 }

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 {
@@ -468,8 +470,10 @@ bool ResolveSetpcDwordJumpTable(const Decoder::Program& program, uint32_t setpc_
 	if (!ResolveJumpTableEntryCount(program, load_index, load.src1, 2u, &entry_count)) {
 		return false;
 	}
+	// Code-table loads are not translated, so a selector that is also the load destination
+	// keeps its scaled index at the branch.
 	uint32_t selector_code = UINT32_MAX;
-	if (!ScalarOperandCode(load.src1, selector_code) || selector_code == offset_code ||
+	if (!ScalarOperandCode(load.src1, selector_code) ||
 	    ScalarCodeWrittenInRange(program, load_index + 1u, setpc_index, selector_code)) {
 		return false;
 	}
@@ -1650,6 +1654,32 @@ 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 +1711,56 @@ 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/ImageOps.cpp

@@ -195,6 +195,9 @@ constexpr MimgGatherInfo MIMG_GATHER_OPCODE_LIST[] = {
     {0x5fu, Opcode::IMAGE_GATHER4_C_LZ_O,
      ImageSampleFlagCompare | ImageSampleFlagLevelZero | ImageSampleFlagOffset},
     {0x61u, Opcode::IMAGE_GATHER4H, ImageSampleFlagGatherHorizontal},
+    {0xc8u, Opcode::IMAGE_GATHER4_C, ImageSampleFlagCompare | ImageSampleFlagAdjust},
+    {0xd8u, Opcode::IMAGE_GATHER4_C_O,
+     ImageSampleFlagCompare | ImageSampleFlagOffset | ImageSampleFlagAdjust},
 };
 
 constexpr MimgAtomicInfo MIMG_ATOMIC_OPCODE_LIST[] = {

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

@@ -70,6 +70,8 @@ constexpr MemoryOpcodeInfo MUBUF_OPCODE_LIST[] = {
     {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},
@@ -164,6 +166,8 @@ void ApplyMemoryInfo(Instruction& inst, const MemoryOpcodeInfo* info) {
 	inst.formatted   = info->formatted;
 }
 
+} // namespace
+
 bool IsDsWriteOpcode(Opcode opcode) {
 	switch (opcode) {
 		case Opcode::DS_WRITE_B8:
@@ -182,6 +186,8 @@ bool IsDsWriteOpcode(Opcode opcode) {
 	}
 }
 
+namespace {
+
 bool IsDsAtomicOpcode(Opcode opcode) {
 	switch (opcode) {
 		case Opcode::DS_ADD_U32:
@@ -312,6 +318,9 @@ 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 ||

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:

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

@@ -599,6 +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,
@@ -880,6 +882,7 @@ Program DecodeFrontProgram(std::span<const uint32_t> front);
 void DecodeProgram(std::span<const uint32_t> code, Program& program);
 bool IsConditionalBranch(Opcode opcode);
 bool IsDirectBranch(Opcode opcode);
+bool IsDsWriteOpcode(Opcode opcode);
 
 void DecodeScalarSource(uint32_t code, uint32_t pc, Operand& operand);
 void DecodeScalarDestination(uint32_t code, uint32_t pc, Operand& operand);

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()},

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

@@ -987,30 +987,6 @@ void ValidateTranslateOptions(const TranslateOptions& options) {
 	    options.stage != ShaderType::Local) {
 		EXIT("embedded fetch requires a vertex or local shader");
 	}
-	switch (options.stage) {
-		case ShaderType::Vertex:
-		case ShaderType::Local:
-		case ShaderType::TessellationControl:
-		case ShaderType::TessellationEvaluation:
-		case ShaderType::Mesh:
-			if (options.input_info.vertex == nullptr) {
-				EXIT("vertex shader translation has no vertex input metadata");
-			}
-			return;
-		case ShaderType::Pixel:
-			if (options.input_info.pixel == nullptr) {
-				EXIT("pixel shader translation has no pixel input metadata");
-			}
-			return;
-		case ShaderType::Compute:
-			if (options.input_info.compute == nullptr) {
-				EXIT("compute shader translation has no compute input metadata");
-			}
-			return;
-		default:
-			EXIT("shader translation has unsupported stage %u",
-			     static_cast<uint32_t>(options.stage));
-	}
 }
 
 } // namespace
@@ -1044,7 +1020,9 @@ IR::Program TranslateProgram(const Decoder::Program& decoded, const CFG::Graph&
 		case ShaderType::Compute:
 			result.scratch_dwords = options.input_info.compute->scratch_size_dwords;
 			break;
-		default: break; // ValidateTranslateOptions rejects unsupported stages.
+		default:
+			EXIT("shader translation has unsupported stage %u",
+			     static_cast<uint32_t>(options.stage));
 	}
 	result.dispatcher_fallback = cfg.irreducible || cfg.unsupported;
 	result.cfg_failure_kind    = cfg.failure_kind;

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

@@ -125,6 +125,7 @@ struct BufferResource {
 	uint32_t               descriptor_index   = UINT32_MAX;
 	uint32_t               first_use_pc       = 0;
 	uint32_t               max_byte_extent    = 0;
+	uint32_t               write_extent       = 0; // UINT32_MAX once a write is runtime-addressed or formatted
 	uint32_t               packed_stride      = 0;
 	Prospero::BufferFormat descriptor_format  = Prospero::BufferFormat::kInvalid;
 	uint32_t               descriptor_swizzle = DstSel(4, 5, 6, 7);
@@ -220,8 +221,11 @@ enum class TessellationAttribute {
 	ControlOutput,
 	EvaluationInput,
 	PatchOutput,
+	PatchInput,
 	Factor
 };
+constexpr uint32_t NumTessellationAttributes =
+    static_cast<uint32_t>(TessellationAttribute::Factor) + 1u;
 
 enum class StageInputKind {
 	VertexIndex,
@@ -356,7 +360,14 @@ struct PushData {
 	static constexpr uint32_t NoStart    = UINT32_MAX;
 	std::array<uint32_t, DwordCount> dwords {};
 
-	[[nodiscard]] static constexpr uint32_t GraphicsDrawDwords(bool mesh, bool object_id) {
+	// User clip planes 0..N follow the mesh draw parameters or object ID, four dwords each.
+	static constexpr uint32_t UserClipPlaneDword = MeshDrawDwordCount;
+
+	[[nodiscard]] static constexpr uint32_t GraphicsDrawDwords(bool mesh, bool object_id,
+	                                                           uint32_t user_clip_planes = 0) {
+		if (user_clip_planes != 0) {
+			return UserClipPlaneDword + 4u * std::bit_width(user_clip_planes);
+		}
 		return mesh ? MeshDrawDwordCount : static_cast<uint32_t>(object_id);
 	}
 
@@ -495,7 +506,6 @@ 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;