libkernel: implement sceKernelAioWaitRequests (#1109)

kernel: implement batch AIO completion waits

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

src/graphics/presentation/window/hostInput.cpp

@@ -454,7 +454,7 @@ void HostInputToggleMouseToJoystick() {
 	LOGF("Mouse to right stick: enabled (F7 to release)\n");
 }
 
-bool HostInputWaitEvent(SDL_Event* event) {
+bool HostInputWaitEvent(SDL_Event* event, int max_wait_ms) {
 	int timeout = -1;
 	if (!g_mouse.enabled || SDL_GetKeyboardFocus() != g_mouse_window) {
 		g_mouse.next_poll = 0;
@@ -473,6 +473,7 @@ bool HostInputWaitEvent(SDL_Event* event) {
 		    now_ms < g_cursor_hide_at ? static_cast<int>(g_cursor_hide_at - now_ms) : 0;
 		timeout = timeout < 0 ? cursor_timeout : std::min(timeout, cursor_timeout);
 	}
+	timeout              = timeout < 0 ? max_wait_ms : std::min(timeout, max_wait_ms);
 	const bool has_event = SDL_WaitEventTimeout(event, timeout);
 
 	if (Config::HideCursorEnabled()) {

src/graphics/presentation/window/hostInput.h

@@ -13,7 +13,7 @@ void               HostInputShutdown();
 void               HostInputKey(int key_code, bool down);
 void               HostInputMouseButton(uint8_t mouse_button, bool down);
 void               HostInputToggleMouseToJoystick();
-[[nodiscard]] bool HostInputWaitEvent(SDL_Event* event);
+[[nodiscard]] bool HostInputWaitEvent(SDL_Event* event, int max_wait_ms);
 
 } // namespace Libs::Graphics
 

src/graphics/presentation/window/swapchain.cpp

@@ -998,7 +998,7 @@ void Presenter::Impl::Present() {
 		}
 
 		presented_overlay_revision.store(overlay_visual.revision, std::memory_order_release);
-		window.UpdateTitle();
+		window.loop.presented_frames.fetch_add(1, std::memory_order_relaxed);
 		return;
 	}
 	LOGF("Vulkan presentation retry exhausted; dropping frame\n");

src/graphics/presentation/window/window.cpp

@@ -722,6 +722,10 @@ void WindowContext::Run() {
 	loop.need_exit = false;
 	loop.paused.store(false, std::memory_order_release);
 
+	constexpr uint64_t title_interval_ms = 1000;
+	auto               title_time        = SDL_GetTicks();
+	auto               title_frames      = loop.presented_frames.load(std::memory_order_relaxed);
+
 	while (!loop.need_exit) {
 		if (loop.paused.load(std::memory_order_acquire)) {
 			if (!timer.IsPaused()) {
@@ -731,7 +735,20 @@ void WindowContext::Run() {
 			timer.Resume();
 		}
 
-		if (!HostInputWaitEvent(&loop.event)) {
+		// Refresh the title on the main thread without making presentation wait for it.
+		const auto now     = SDL_GetTicks();
+		const auto elapsed = now - title_time;
+		if (elapsed >= title_interval_ms) {
+			const auto frames = loop.presented_frames.load(std::memory_order_relaxed);
+			if (frames != 0) {
+				UpdateTitle(frames, static_cast<double>(frames - title_frames) * 1000.0 /
+				                        static_cast<double>(elapsed));
+			}
+			title_time   = now;
+			title_frames = frames;
+		}
+		const auto wait_ms = static_cast<int>(title_interval_ms - (now - title_time));
+		if (!HostInputWaitEvent(&loop.event, wait_ms)) {
 			continue;
 		}
 		ProcessEvent(timer.GetTimeS());
@@ -913,7 +930,7 @@ void WindowContext::UpdateIcon() {
 	}
 }
 
-void WindowContext::UpdateTitle() {
+void WindowContext::UpdateTitle(uint64_t frame_num, double current_fps) {
 	static char title[128];
 	static char title_id[12];
 	static char app_ver[12];
@@ -923,10 +940,6 @@ void WindowContext::UpdateTitle() {
 	static bool has_app_ver =
 	    Loader::SystemContentParamSfoGetString("APP_VER", app_ver, sizeof(app_ver));
 	static const std::string processor_name = Common::GetSystemInfo().ProcessorName;
-	static uint64_t fps_start   = Common::Timer::QueryPerformanceCounter();
-	static uint64_t frame_num   = 0;
-	static uint64_t fps_frames  = 0;
-	static double   current_fps = 0.0;
 
 #if KYTY_BUILD == KYTY_BUILD_DEBUG
 	static constexpr auto build_type = "Debug";
@@ -936,17 +949,6 @@ void WindowContext::UpdateTitle() {
 	static constexpr auto build_type = "Unknown";
 #endif
 
-	const auto now       = Common::Timer::QueryPerformanceCounter();
-	const auto frequency = Common::Timer::QueryPerformanceFrequency();
-	frame_num++;
-	fps_frames++;
-	if (now - fps_start >= frequency) {
-		current_fps = static_cast<double>(fps_frames) * static_cast<double>(frequency) /
-		              static_cast<double>(now - fps_start);
-		fps_start   = now;
-		fps_frames  = 0;
-	}
-
 	const auto* device_name = graphic_ctx.GetPhysicalDeviceProperties().deviceName.data();
 	auto text = fmt::format(
 	    "[{} | {}] {}{}{}{}{}{}[{}] [{}], frame: {}, fps: {:.0f}", KYTY_BUILD_LABEL, build_type,
@@ -954,16 +956,7 @@ void WindowContext::UpdateTitle() {
 	    (has_title_id ? ", " : ""), (has_app_ver ? app_ver : ""), (has_app_ver ? " " : ""),
 	    device_name, processor_name, frame_num, current_fps);
 
-	struct TitleUpdate {
-		SDL_Window*  window;
-		std::string* text;
-	} update {window, &text};
-	EXIT_IF(!SDL_RunOnMainThread(
-	    [](void* data) {
-		    auto& title = *static_cast<TitleUpdate*>(data);
-		    SDL_SetWindowTitle(title.window, title.text->c_str());
-	    },
-	    &update, true));
+	SDL_SetWindowTitle(window, text.c_str());
 }
 
 } // namespace Libs::Graphics

src/graphics/presentation/window/windowInternal.h

@@ -27,6 +27,7 @@ struct WindowLoopState {
 	SDL_Event        event {};
 	bool             need_exit = false;
 	std::atomic_bool paused    = false;
+	std::atomic<uint64_t> presented_frames {0};
 };
 
 struct WindowContext {
@@ -42,7 +43,7 @@ struct WindowContext {
 	void                                                    RecreateSurface();
 	void                                                    RefreshSurfaceCapabilities();
 	void                                                    UpdateIcon();
-	void                                                    UpdateTitle();
+	void UpdateTitle(uint64_t frame_num, double current_fps);
 	void                                                    Resize(int width, int height);
 	void ProcessWindowEvent(const SDL_WindowEvent& event);
 	void ProcessDisplayEvent(const SDL_DisplayEvent& event);

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

@@ -109,6 +109,7 @@ constexpr MemoryOpcodeInfo DS_OPCODE_LIST[] = {
     {0x39u, Opcode::DS_READ_I8, 1, 8, true},     {0x3au, Opcode::DS_READ_U8, 1, 8},
     {0x3bu, Opcode::DS_READ_I16, 1, 16, true},   {0x3cu, Opcode::DS_READ_U16, 1, 16},
     {0x3du, Opcode::DS_CONSUME, 1, 32},          {0x3eu, Opcode::DS_APPEND, 1, 32},
+    {0x3fu, Opcode::DS_ORDERED_COUNT, 1, 32},
     {0x40u, Opcode::DS_ADD_U64, 2, 32},         {0x4au, Opcode::DS_OR_B64, 2, 32},
     {0x4du, Opcode::DS_WRITE_B64, 2, 32},        {0x4eu, Opcode::DS_WRITE2_B64, 4, 32},
     {0x4fu, Opcode::DS_WRITE2ST64_B64, 4, 32},   {0x76u, Opcode::DS_READ_B64, 2, 32},
@@ -214,6 +215,7 @@ uint32_t DsSourceCount(Opcode opcode) {
 		case Opcode::DS_READ_ADDTID_B32:
 		case Opcode::DS_CONSUME:
 		case Opcode::DS_APPEND: return 0u;
+		case Opcode::DS_ORDERED_COUNT: return 1u;
 		default: return IsDsWriteOpcode(opcode) || IsDsAtomicOpcode(opcode) ? 2u : 1u;
 	}
 }
@@ -412,6 +414,14 @@ void DecodeDs(uint32_t pc, std::span<const uint32_t> code, uint32_t word_index,
 	if (inst.opcode == Opcode::UNSUPPORTED) {
 		SetUnsupported(inst, Family::DS, opcode, "DS opcode is not implemented");
 	}
+	if (inst.opcode == Opcode::DS_ORDERED_COUNT) {
+		// Keep OFFSET1's wave type as well as its add/swap selector for GDS addressing.
+		inst.offset           = offset0 & 0xfcu;
+		inst.secondary_offset = offset1;
+		if (!inst.gds || ((offset1 >> 4u) & 3u) > 1u || ((offset1 >> 6u) & 3u) != 0u) {
+			SetUnsupported(inst, Family::DS, opcode, "DS ordered count variant is not implemented");
+		}
+	}
 	if (inst.opcode == Opcode::DS_SWIZZLE_B32 && inst.offset >= 0xe000u) {
 		SetUnsupported(inst, Family::DS, opcode, "DS swizzle FFT mode is not implemented");
 	}

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

@@ -382,6 +382,7 @@ enum class Opcode {
 	V_CMPX_LG_F32,
 	V_CMPX_GE_F32,
 	V_CMPX_O_F32,
+	V_CMPX_U_F32,
 	V_CMPX_NGE_F32,
 	V_CMPX_NLG_F32,
 	V_CMPX_NGT_F32,
@@ -579,6 +580,7 @@ enum class Opcode {
 	DS_BPERMUTE_B32,
 	DS_CONSUME,
 	DS_APPEND,
+	DS_ORDERED_COUNT,
 	DS_READ_I8,
 	DS_READ_U8,
 	DS_READ_I16,

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

@@ -232,7 +232,7 @@ constexpr VopcOpcodeInfo VOPC_OPCODE_LIST[] = {
     {0x11u, Opcode::V_CMPX_LT_F32},        {0x12u, Opcode::V_CMPX_EQ_F32},
     {0x13u, Opcode::V_CMPX_LE_F32},        {0x14u, Opcode::V_CMPX_GT_F32},
     {0x15u, Opcode::V_CMPX_LG_F32},        {0x16u, Opcode::V_CMPX_GE_F32},
-    {0x17u, Opcode::V_CMPX_O_F32},
+    {0x17u, Opcode::V_CMPX_O_F32},         {0x18u, Opcode::V_CMPX_U_F32},
     {0x19u, Opcode::V_CMPX_NGE_F32},       {0x1au, Opcode::V_CMPX_NLG_F32},
     {0x1bu, Opcode::V_CMPX_NGT_F32},       {0x1cu, Opcode::V_CMPX_NLE_F32},
     {0x1du, Opcode::V_CMPX_NEQ_F32},       {0x1eu, Opcode::V_CMPX_NLT_F32},
@@ -892,6 +892,7 @@ bool IsVopcFloatCompareOpcode(Opcode opcode) {
 		case Opcode::V_CMPX_LG_F32:
 		case Opcode::V_CMPX_GE_F32:
 		case Opcode::V_CMPX_O_F32:
+		case Opcode::V_CMPX_U_F32:
 		case Opcode::V_CMPX_NGE_F32:
 		case Opcode::V_CMPX_NLG_F32:
 		case Opcode::V_CMPX_NGT_F32:
@@ -1449,6 +1450,9 @@ bool SupportsNativeVop3ResultModifiers(Opcode opcode) {
 		case Opcode::V_MAD_F32:
 		case Opcode::V_FMA_F32:
 		case Opcode::V_FMA_F16:
+		case Opcode::V_MIN3_F16:
+		case Opcode::V_MAX3_F16:
+		case Opcode::V_MED3_F16:
 		case Opcode::V_CUBEID_F32:
 		case Opcode::V_CVT_PKRTZ_F16_F32:
 		case Opcode::V_LDEXP_F32:
@@ -1479,7 +1483,8 @@ bool HasUnsupportedNativeVop3Modifiers(Opcode opcode, bool permlane, bool mad_mi
 		return clamp != 0u || omod != 0u;
 	}
 	if (IsNativeVop3F16TernaryOpcode(opcode)) {
-		return opcode != Opcode::V_FMA_F16 && (clamp != 0u || omod != 0u);
+		// Clamp is applied by WriteF16 for every f16 ternary; omod only for V_FMA_F16.
+		return opcode != Opcode::V_FMA_F16 && omod != 0u;
 	}
 	if (IsNativeVop3I16TernaryOpcode(opcode)) {
 		return abs != 0u || (clamp != 0u && !clamp_modifier) || omod != 0u || neg != 0u;
@@ -1529,6 +1534,7 @@ bool IsVopcCompareExec(Opcode opcode) {
 		case Opcode::V_CMPX_LG_F32:
 		case Opcode::V_CMPX_GE_F32:
 		case Opcode::V_CMPX_O_F32:
+		case Opcode::V_CMPX_U_F32:
 		case Opcode::V_CMPX_NGE_F32:
 		case Opcode::V_CMPX_NLG_F32:
 		case Opcode::V_CMPX_NGT_F32:

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

@@ -883,6 +883,50 @@ void Translator::DS_SWIZZLE_B32(const Decoder::Instruction& inst) {
 	                                ReadU32(MemorySourceAt(inst, 1)), ir.GetExec()}));
 }
 
+void Translator::DS_ORDERED_COUNT(const Decoder::Instruction& inst) {
+	// Approximate ordered counting with one GDS atomic per wave. Wave-launch ordering and
+	// release/done synchronization are not emulated.
+	const auto memory = MemoryInfoFromDecoded(inst);
+	const auto m0 = ir.GetM0();
+	auto address = ir.BitwiseAnd(ir.ShiftRightLogical(m0, IR::U32(IR::Value(16u))),
+	                             IR::U32(IR::Value(0xfffcu)));
+	if (((inst.secondary_offset >> 2u) & 3u) == 1u) {
+		const auto packer = ir.BitwiseAnd(m0, IR::U32(IR::Value(0xfu)));
+		address = ir.IAdd(address, ir.ShiftLeftLogical(packer, IR::U32(IR::Value(2u))));
+	}
+	const auto exec   = ir.GetExec();
+	const auto ballot = ir.Emit(IR::ValueOpcode::Ballot, {exec});
+	const auto low = ir.CompositeExtract(ballot, 0u);
+	const auto high = ir.CompositeExtract(ballot, 1u);
+	const auto active = ir.INotEqual(ir.BitwiseOr(low, high), IR::U32(IR::Value(0u)));
+	const auto first = ir.Select(
+	    ir.INotEqual(low, IR::U32(IR::Value(0u))),
+	    IR::U32(ir.Emit(IR::ValueOpcode::FindILsb32, {low})),
+	    ir.IAdd(IR::U32(IR::Value(32u)), IR::U32(ir.Emit(IR::ValueOpcode::FindILsb32, {high}))));
+	// An empty EXEC still returns the counter. Elect lane zero and perform an add of zero.
+	const auto selected = ir.Select(active, first, IR::U32(IR::Value(0u)));
+	const auto lane = IR::U32(ir.Emit(IR::ValueOpcode::LaneId));
+	const auto is_first = ir.IEqual(lane, selected);
+	const auto source = IR::U32(ir.Emit(IR::ValueOpcode::ReadFirstLane,
+	                                   {ReadU32(inst.src0), is_first}));
+	const auto value = ir.Select(active, source, IR::U32(IR::Value(0u)));
+	const auto flags = AddMemoryInfo(memory, inst.pc);
+	IR::U32 result;
+	if (((inst.secondary_offset >> 4u) & 3u) == 1u) {
+		const auto swapped = IR::U32(ir.Emit(IR::ValueOpcode::SharedAtomicSwap32,
+		    {address, value, ir.LogicalAnd(is_first, active)}, flags));
+		const auto unchanged = IR::U32(ir.Emit(IR::ValueOpcode::SharedAtomicIAdd32,
+		    {address, IR::Value(0u), ir.LogicalAnd(is_first, ir.LogicalNot(active))}, flags));
+		result = ir.Select(active, swapped, unchanged);
+	} else {
+		result = IR::U32(ir.Emit(IR::ValueOpcode::SharedAtomicIAdd32,
+		                        {address, value, is_first}, flags));
+	}
+	// DS_ORDERED_COUNT writes every destination lane, regardless of EXEC.
+	ir.SetVectorReg(static_cast<IR::VectorReg>(inst.dst.reg),
+	    IR::U32(ir.Emit(IR::ValueOpcode::ReadFirstLane, {result, is_first})));
+}
+
 void Translator::DS_PERMUTE(const Decoder::Instruction& inst, bool backward) {
 	const auto address = ir.IAdd(ReadU32(inst.src0), IR::U32(IR::Value(inst.offset)));
 	WriteOperand(inst.dst, ir.Emit(backward ? IR::ValueOpcode::BpermuteU32
@@ -1114,6 +1158,7 @@ void Translator::EmitMemory(const Decoder::Instruction& inst) {
 		case Decoder::Opcode::DS_SWIZZLE_B32: return DS_SWIZZLE_B32(inst);
 		case Decoder::Opcode::DS_PERMUTE_B32: return DS_PERMUTE(inst, false);
 		case Decoder::Opcode::DS_BPERMUTE_B32: return DS_PERMUTE(inst, true);
+		case Decoder::Opcode::DS_ORDERED_COUNT: return DS_ORDERED_COUNT(inst);
 		case Decoder::Opcode::DS_CONSUME:
 			return DS_APPEND_CONSUME(inst, IR::ValueOpcode::DataConsume);
 		case Decoder::Opcode::DS_APPEND:

src/graphics/shader/recompiler/frontend/translate/Translator.h

@@ -109,6 +109,7 @@ private:
 	void DS_ADDTID(const Decoder::Instruction& inst, bool write);
 	void DS_SWIZZLE_B32(const Decoder::Instruction& inst);
 	void DS_PERMUTE(const Decoder::Instruction& inst, bool backward);
+	void DS_ORDERED_COUNT(const Decoder::Instruction& inst);
 
 	IR::F32 SelectF32(IR::U1 condition, IR::F32 true_value, IR::F32 false_value);
 	IR::U32 ConvertF32ToU32Saturated(IR::F32 value, float upper_bound, float safe_upper,

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

@@ -344,6 +344,7 @@ void Translator::EmitVector(const Decoder::Instruction& inst) {
 		case O::V_CMP_O_F32: EmitFloatOrderedCompare(inst, true, false); return;
 		case O::V_CMP_U_F32: EmitFloatOrderedCompare(inst, false, false); return;
 		case O::V_CMPX_O_F32: EmitFloatOrderedCompare(inst, true, true); return;
+		case O::V_CMPX_U_F32: EmitFloatOrderedCompare(inst, false, true); return;
 		case O::V_CMP_CLASS_F32: EmitFloatClassCompare(inst, false); return;
 		case O::V_CMPX_CLASS_F32: EmitFloatClassCompare(inst, true); return;
 		case O::V_CMPX_CLASS_F16: EmitFloatClassCompare(inst, true, true); return;

src/kernel/pthread.cpp

@@ -3200,6 +3200,15 @@ static void* RunThread(void* arg) {
 	os_thread_id = GetHostThreadId();
 #endif
 	thread->host_thread_id = os_thread_id;
+#if KYTY_PLATFORM == KYTY_PLATFORM_LINUX
+	if (!thread->name.empty()) {
+#if defined(__APPLE__)
+		pthread_setname_np(thread->name.substr(0, 63).c_str());
+#else
+		pthread_setname_np(pthread_self(), thread->name.substr(0, 15).c_str());
+#endif
+	}
+#endif
 
 	LOGF("\tPthread run begin: %s, id = %d, os_thread_id = %" PRIu64 ", entry = 0x%016" PRIx64
 	     ", arg = 0x%016" PRIx64 ", stack_addr = 0x%016" PRIx64 ", stack_size = %" PRIu64 "\n",

src/launcher/forms/configuration_edit_dialog.ui

@@ -6,7 +6,7 @@
    <rect>
     <x>0</x>
     <y>0</y>
-    <width>840</width>
+    <width>900</width>
     <height>620</height>
    </rect>
   </property>
@@ -400,6 +400,13 @@ To use your own sound, replace the emulator's assets/sounds/trophy-unlock.wav wi
        <property name="toolTip"><string>Restore default settings</string></property>
       </widget>
      </item>
+     <item>
+      <widget class="QPushButton" name="settings_button">
+       <property name="minimumSize"><size><width>80</width><height>25</height></size></property>
+       <property name="text"><string>Game Configs</string></property>
+       <property name="toolTip"><string>Import or export game configs</string></property>
+      </widget>
+     </item>
      <item>
       <spacer name="buttonSpacer">
        <property name="orientation"><enum>Qt::Horizontal</enum></property>

src/launcher/forms/configuration_list_widget.ui

@@ -125,7 +125,7 @@
             </size>
            </property>
            <property name="toolTip">
-            <string>&lt;html&gt;&lt;head/&gt;&lt;body&gt;&lt;p&gt;Edit configuration&lt;/p&gt;&lt;/body&gt;&lt;/html&gt;</string>
+            <string>Edit game config</string>
            </property>
            <property name="text">
             <string/>
@@ -150,7 +150,7 @@
             </size>
            </property>
            <property name="toolTip">
-            <string>&lt;html&gt;&lt;head/&gt;&lt;body&gt;&lt;p&gt;Delete configuration&lt;/p&gt;&lt;/body&gt;&lt;/html&gt;</string>
+            <string>Clear game config</string>
            </property>
            <property name="styleSheet">
             <string notr="true"/>

src/launcher/include/configuration.h

@@ -6,6 +6,7 @@
 #include <QByteArray>
 #include <QChar>
 #include <QColor>
+#include <QJsonObject>
 #include <QMetaEnum>
 #include <QMetaType>
 #include <QObject>
@@ -14,9 +15,6 @@
 #include <QStringList>
 #include <QVariant>
 
-#define KYTY_CFG_SET(n) s->setValue(#n, QVariant::fromValue(n).toString());
-#define KYTY_CFG_GET(n) n = s->value(#n).value<decltype(n)>();
-
 template <class T>
 inline QStringList EnumToList() {
 	QStringList ret;
@@ -126,7 +124,7 @@ public:
 	int                    vblank_frequency            = 60;
 	int                    console_language            = DEFAULT_CONSOLE_LANGUAGE;
 	bool                   vulkan_validation_enabled   = false;
-	bool                   shader_validation_enabled   = true;
+	bool                   shader_validation_enabled   = false;
 	ShaderOptimizationType shader_optimization_type    = ShaderOptimizationType::Performance;
 	LogDirection           shader_log_direction        = LogDirection::Silent;
 	QString                shader_log_folder           = "_Shaders";
@@ -188,90 +186,16 @@ public:
 		game_comment    = other.game_comment;
 	}
 
-	void WriteSettings(QSettings* s) const {
-		KYTY_CFG_SET(name);
-		KYTY_CFG_SET(basedir);
-		KYTY_CFG_SET(game_path);
-		KYTY_CFG_SET(custom_settings);
-		KYTY_CFG_SET(screen_resolution);
-		KYTY_CFG_SET(user_name);
-		KYTY_CFG_SET(user_id);
-		KYTY_CFG_SET(audio_input_device);
-		KYTY_CFG_SET(present_mode);
-		KYTY_CFG_SET(gpu_index);
-		KYTY_CFG_SET(fullscreen_enabled);
-		KYTY_CFG_SET(hide_cursor_enabled);
-		KYTY_CFG_SET(readback_linear_images);
-		KYTY_CFG_SET(tessellation_enabled);
-		KYTY_CFG_SET(trophy_enabled);
-		KYTY_CFG_SET(vblank_frequency);
-		KYTY_CFG_SET(console_language);
-		KYTY_CFG_SET(vulkan_validation_enabled);
-		KYTY_CFG_SET(shader_validation_enabled);
-		KYTY_CFG_SET(shader_optimization_type);
-		KYTY_CFG_SET(shader_log_direction);
-		KYTY_CFG_SET(shader_log_folder);
-		KYTY_CFG_SET(command_buffer_dump_enabled);
-		KYTY_CFG_SET(command_buffer_dump_folder);
-		KYTY_CFG_SET(printf_direction);
-		KYTY_CFG_SET(printf_output_file);
-		KYTY_CFG_SET(profiler_enabled);
-		KYTY_CFG_SET(renderdoc_enabled);
-		KYTY_CFG_SET(amd_cpu_enabled);
-#if defined(_WIN32)
-		KYTY_CFG_SET(red_zone_protection_enabled);
-#endif
-		s->setValue("host_input_mapping", host_input_mapping);
-		KYTY_CFG_SET(elf);
-	}
+	void WriteSettings(QSettings* s) const;
+	void ReadSettings(QSettings* s);
 
-	void ReadSettings(QSettings* s) {
-		KYTY_CFG_GET(name);
-		KYTY_CFG_GET(basedir);
-		KYTY_CFG_GET(game_path);
-		KYTY_CFG_GET(custom_settings);
-		KYTY_CFG_GET(screen_resolution);
-		user_name          = s->value("user_name", user_name).toString();
-		bool user_id_ok    = false;
-		auto saved_user_id = s->value("user_id", user_id).toInt(&user_id_ok);
-		user_id            = user_id_ok && Config::IsConfiguredUserIdValid(saved_user_id)
-		                         ? saved_user_id
-		                         : Config::DEFAULT_USER_ID;
-		audio_input_device = s->value("audio_input_device", audio_input_device).toString();
-		KYTY_CFG_GET(present_mode);
-		gpu_index = s->value("gpu_index", -1).toInt();
-		if (EnumToText(present_mode).isEmpty()) {
-			present_mode = PresentMode::Mailbox;
-		}
-		KYTY_CFG_GET(fullscreen_enabled);
-		KYTY_CFG_GET(hide_cursor_enabled);
-		KYTY_CFG_GET(readback_linear_images);
-		KYTY_CFG_GET(tessellation_enabled);
-		trophy_enabled = s->value("trophy_enabled", trophy_enabled).toBool();
-		vblank_frequency = s->value("vblank_frequency", vblank_frequency).toInt();
-		console_language = s->value("console_language", console_language).toInt();
-		if (console_language < 0 || console_language > MAX_CONSOLE_LANGUAGE) {
-			console_language = DEFAULT_CONSOLE_LANGUAGE;
-		}
-		KYTY_CFG_GET(vulkan_validation_enabled);
-		KYTY_CFG_GET(shader_validation_enabled);
-		KYTY_CFG_GET(shader_optimization_type);
-		KYTY_CFG_GET(shader_log_direction);
-		KYTY_CFG_GET(shader_log_folder);
-		KYTY_CFG_GET(command_buffer_dump_enabled);
-		KYTY_CFG_GET(command_buffer_dump_folder);
-		KYTY_CFG_GET(printf_direction);
-		KYTY_CFG_GET(printf_output_file);
-		KYTY_CFG_GET(profiler_enabled);
-		KYTY_CFG_GET(renderdoc_enabled);
-		amd_cpu_enabled = s->value("amd_cpu_enabled", false).toBool();
-#if defined(_WIN32)
-		red_zone_protection_enabled =
-		    s->value("red_zone_protection_enabled", red_zone_protection_enabled).toBool();
-#endif
-		host_input_mapping = s->value("host_input_mapping", host_input_mapping).toStringList();
-		elf                = s->value("elf", elf).toString();
-	}
+	[[nodiscard]] QVariantMap GameSettings() const;
+	// Call on a temporary configuration: validation can fail after loading values.
+	bool SetGameSettings(const QJsonObject& settings, QString& error);
+
+private:
+	template <class Settings>
+	void ReadGameSettingsValues(const Settings& s);
 };
 
 Q_DECLARE_METATYPE(Configuration*)

src/launcher/include/configurationEditDialog.h

@@ -33,6 +33,8 @@ public:
 
 signals:
 	void PreviewControllerColor(const QString& color);
+	void ImportGameSettings();
+	void ExportGameSettings();
 
 private:
 	Ui::ConfigurationEditDialog* m_ui = nullptr;

src/launcher/include/configurationListWidget.h

@@ -74,6 +74,8 @@ private:
 	void               SelectItem(QTreeWidgetItem* witem);
 	void               ApplyCompatibility();
 	void               UpdateToolbarIcons();
+	void               ImportGameSettings(QWidget* parent);
+	void               ExportGameSettings(QWidget* parent) const;
 	[[nodiscard]] bool HasValidGameDirectory() const;
 
 	ConfigurationItem*            m_selected_item = nullptr;

src/launcher/src/configuration.cpp

@@ -0,0 +1,145 @@
+#include "configuration.h"
+
+QVariantMap Configuration::GameSettings() const {
+	QVariantMap values;
+#define KYTY_CFG_SET(n) values.insert(QStringLiteral(#n), QVariant::fromValue(n).toString())
+	KYTY_CFG_SET(screen_resolution);
+	KYTY_CFG_SET(user_name);
+	KYTY_CFG_SET(user_id);
+	KYTY_CFG_SET(audio_input_device);
+	KYTY_CFG_SET(present_mode);
+	KYTY_CFG_SET(gpu_index);
+	KYTY_CFG_SET(fullscreen_enabled);
+	KYTY_CFG_SET(hide_cursor_enabled);
+	KYTY_CFG_SET(readback_linear_images);
+	KYTY_CFG_SET(tessellation_enabled);
+	KYTY_CFG_SET(trophy_enabled);
+	KYTY_CFG_SET(vblank_frequency);
+	KYTY_CFG_SET(console_language);
+	KYTY_CFG_SET(vulkan_validation_enabled);
+	KYTY_CFG_SET(shader_validation_enabled);
+	KYTY_CFG_SET(shader_optimization_type);
+	KYTY_CFG_SET(shader_log_direction);
+	KYTY_CFG_SET(shader_log_folder);
+	KYTY_CFG_SET(command_buffer_dump_enabled);
+	KYTY_CFG_SET(command_buffer_dump_folder);
+	KYTY_CFG_SET(printf_direction);
+	KYTY_CFG_SET(printf_output_file);
+	KYTY_CFG_SET(profiler_enabled);
+	KYTY_CFG_SET(renderdoc_enabled);
+	KYTY_CFG_SET(amd_cpu_enabled);
+#if defined(_WIN32)
+	KYTY_CFG_SET(red_zone_protection_enabled);
+#endif
+	KYTY_CFG_SET(elf);
+#undef KYTY_CFG_SET
+	return values;
+}
+
+template <class Settings>
+void Configuration::ReadGameSettingsValues(const Settings& s) {
+#define KYTY_CFG_GET(n) n = s.value(QStringLiteral(#n)).template value<decltype(n)>()
+	KYTY_CFG_GET(screen_resolution);
+	user_name          = s.value("user_name", user_name).toString();
+	bool user_id_ok    = false;
+	auto saved_user_id = s.value("user_id", user_id).toInt(&user_id_ok);
+	user_id            = user_id_ok && Config::IsConfiguredUserIdValid(saved_user_id)
+	                         ? saved_user_id
+	                         : Config::DEFAULT_USER_ID;
+	audio_input_device = s.value("audio_input_device", audio_input_device).toString();
+	KYTY_CFG_GET(present_mode);
+	gpu_index = s.value("gpu_index", -1).toInt();
+	if (EnumToText(present_mode).isEmpty()) {
+		present_mode = PresentMode::Mailbox;
+	}
+	KYTY_CFG_GET(fullscreen_enabled);
+	KYTY_CFG_GET(hide_cursor_enabled);
+	KYTY_CFG_GET(readback_linear_images);
+	KYTY_CFG_GET(tessellation_enabled);
+	trophy_enabled   = s.value("trophy_enabled", trophy_enabled).toBool();
+	vblank_frequency = s.value("vblank_frequency", vblank_frequency).toInt();
+	console_language = s.value("console_language", console_language).toInt();
+	if (console_language < 0 || console_language > MAX_CONSOLE_LANGUAGE) {
+		console_language = DEFAULT_CONSOLE_LANGUAGE;
+	}
+	KYTY_CFG_GET(vulkan_validation_enabled);
+	KYTY_CFG_GET(shader_validation_enabled);
+	KYTY_CFG_GET(shader_optimization_type);
+	KYTY_CFG_GET(shader_log_direction);
+	KYTY_CFG_GET(shader_log_folder);
+	KYTY_CFG_GET(command_buffer_dump_enabled);
+	KYTY_CFG_GET(command_buffer_dump_folder);
+	KYTY_CFG_GET(printf_direction);
+	KYTY_CFG_GET(printf_output_file);
+	KYTY_CFG_GET(profiler_enabled);
+	KYTY_CFG_GET(renderdoc_enabled);
+	amd_cpu_enabled = s.value("amd_cpu_enabled", false).toBool();
+#if defined(_WIN32)
+	red_zone_protection_enabled =
+	    s.value("red_zone_protection_enabled", red_zone_protection_enabled).toBool();
+#endif
+	elf = s.value("elf", elf).toString();
+#undef KYTY_CFG_GET
+}
+
+void Configuration::WriteSettings(QSettings* s) const {
+	const auto values = GameSettings();
+	for (auto it = values.constBegin(); it != values.constEnd(); ++it) {
+		s->setValue(it.key(), it.value());
+	}
+	s->setValue("name", name);
+	s->setValue("basedir", basedir);
+	s->setValue("game_path", game_path);
+	s->setValue("custom_settings", QVariant::fromValue(custom_settings).toString());
+	s->setValue("host_input_mapping", host_input_mapping);
+}
+
+void Configuration::ReadSettings(QSettings* s) {
+	ReadGameSettingsValues(*s);
+	name               = s->value("name").toString();
+	basedir            = s->value("basedir").toString();
+	game_path          = s->value("game_path").toString();
+	custom_settings    = s->value("custom_settings").toBool();
+	host_input_mapping = s->value("host_input_mapping", host_input_mapping).toStringList();
+}
+
+bool Configuration::SetGameSettings(const QJsonObject& settings, QString& error) {
+	const auto fail = [&error](const QString& message) {
+		error = message;
+		return false;
+	};
+	auto values = GameSettings();
+	for (auto it = settings.constBegin(); it != settings.constEnd(); ++it) {
+		if (!values.contains(it.key())) {
+			return fail(tr("Unknown game setting: %1").arg(it.key()));
+		}
+		if (!it.value().isString()) {
+			return fail(tr("Game setting %1 must be a string.").arg(it.key()));
+		}
+		values.insert(it.key(), it.value().toString());
+	}
+
+	ReadGameSettingsValues(values);
+	const auto canonical = GameSettings();
+	for (auto it = settings.constBegin(); it != settings.constEnd(); ++it) {
+		if (canonical.value(it.key()) != values.value(it.key())) {
+			return fail(tr("Invalid value for game setting: %1").arg(it.key()));
+		}
+	}
+	if (user_name.trimmed().isEmpty() || user_name.toUtf8().size() > Config::MAX_USER_NAME_LENGTH) {
+		return fail(tr("User name must contain 1-16 UTF-8 bytes."));
+	}
+	if (gpu_index < -1) {
+		return fail(tr("GPU index must be -1 or greater."));
+	}
+	if (vblank_frequency < 30 || vblank_frequency > 360) {
+		return fail(tr("Vblank frequency must be between 30 and 360."));
+	}
+	if ((shader_log_direction == LogDirection::File && shader_log_folder.isEmpty()) ||
+	    (printf_direction == LogDirection::File && printf_output_file.isEmpty()) ||
+	    (command_buffer_dump_enabled && command_buffer_dump_folder.isEmpty())) {
+		return fail(tr("Enabled logging and command buffer dumps require an output path."));
+	}
+
+	return true;
+}

src/launcher/src/configurationEditDialog.cpp

@@ -19,6 +19,7 @@
 #include <QLineEdit>
 #include <QListView>
 #include <QListWidget>
+#include <QMenu>
 #include <QMessageBox>
 #include <QPushButton>
 #include <QResizeEvent>
@@ -126,6 +127,7 @@ ConfigurationEditDialog::ConfigurationEditDialog(Configuration& info, QWidget* p
 	setMinimumWidth(width());
 	InitGameDirectories();
 	m_ui->controller_group->setVisible(false);
+	m_ui->settings_button->setVisible(false);
 
 	connect(m_ui->ok_button, &QPushButton::clicked, this, &ConfigurationEditDialog::save);
 	connect(m_ui->cancel_button, &QPushButton::clicked, this, &QDialog::reject);
@@ -354,6 +356,11 @@ void ConfigurationEditDialog::InitGameDirectories() {
 
 void ConfigurationEditDialog::SetGlobalSettings(const QStringList& dirs) {
 	m_global_settings = true;
+	auto* menu        = new QMenu(m_ui->settings_button);
+	menu->addAction(tr("Import..."), this, &ConfigurationEditDialog::ImportGameSettings);
+	menu->addAction(tr("Export..."), this, &ConfigurationEditDialog::ExportGameSettings);
+	m_ui->settings_button->setMenu(menu);
+	m_ui->settings_button->setVisible(true);
 	m_game_dirs_list->clear();
 	m_game_dirs_group->setMinimumWidth(GLOBAL_SETTINGS_GAME_DIRS_MIN_WIDTH);
 

src/launcher/src/configurationListWidget.cpp

@@ -24,9 +24,11 @@
 #include <QDialog>
 #include <QDir>
 #include <QFile>
+#include <QFileDialog>
 #include <QFileInfo>
 #include <QHeaderView>
 #include <QIcon>
+#include <QInputDialog>
 #include <QJsonDocument>
 #include <QJsonObject>
 #include <QJsonParseError>
@@ -40,6 +42,7 @@
 #include <QProcess>
 #include <QPushButton>
 #include <QRegularExpression>
+#include <QSaveFile>
 #include <QSet>
 #include <QSettings>
 #include <QSignalBlocker>
@@ -765,7 +768,7 @@ void ConfigurationListWidget::edit_configuration() {
 
 	auto                    info = CreateConfiguration(*item);
 	ConfigurationEditDialog dlg(*info, this);
-	dlg.setWindowTitle(tr("Edit game settings"));
+	dlg.setWindowTitle(tr("Edit game config"));
 	connect(&dlg, &ConfigurationEditDialog::PreviewControllerColor, this,
 	        &ConfigurationListWidget::PreviewControllerColor);
 
@@ -789,8 +792,8 @@ void ConfigurationListWidget::delete_configuartion() {
 		return;
 	}
 
-	if (QMessageBox::Yes == QMessageBox::question(this, tr("Clear custom settings"),
-	                                              tr("Do you want to clear custom settings?"))) {
+	if (QMessageBox::Yes == QMessageBox::question(this, tr("Clear game config"),
+	                                              tr("Clear this game's config and use global settings?"))) {
 		delete m_custom_infos.take(item->GetInfo().game_path);
 		item->GetInfo().custom_settings = false;
 		WriteSettings();
@@ -810,6 +813,10 @@ void ConfigurationListWidget::edit_global_settings() {
 	dlg.SetGlobalSettings(m_game_dirs);
 	connect(&dlg, &ConfigurationEditDialog::PreviewControllerColor, this,
 	        &ConfigurationListWidget::PreviewControllerColor);
+	connect(&dlg, &ConfigurationEditDialog::ImportGameSettings, this,
+	        [this, &dlg]() { ImportGameSettings(&dlg); });
+	connect(&dlg, &ConfigurationEditDialog::ExportGameSettings, this,
+	        [this, &dlg]() { ExportGameSettings(&dlg); });
 
 	if (dlg.exec() == QDialog::Accepted) {
 		m_global_info.CopyEmulatorSettingsFrom(info);
@@ -822,7 +829,133 @@ void ConfigurationListWidget::edit_global_settings() {
 			ScanGameDirectory();
 		}
 	}
-	emit Select();
+	SelectItem(m_ui->cfgs_list->currentItem());
+}
+
+static bool IsGameSettingsTitleId(const QString& title_id) {
+	static const QRegularExpression pattern(QStringLiteral("\\APPSA[0-9]{5}\\z"));
+	return pattern.match(title_id).hasMatch();
+}
+
+void ConfigurationListWidget::ImportGameSettings(QWidget* parent) {
+	const auto path = QFileDialog::getOpenFileName(parent, tr("Import game configs"), {},
+	                                               tr("JSON files (*.json)"));
+	if (path.isEmpty()) {
+		return;
+	}
+	QFile file(path);
+	if (!file.open(QIODevice::ReadOnly)) {
+		QMessageBox::warning(parent, tr("Import failed"), file.errorString());
+		return;
+	}
+	const auto document = QJsonDocument::fromJson(file.readAll());
+	if (!document.isObject() || document.object().isEmpty()) {
+		QMessageBox::warning(
+		    parent, tr("Import failed"),
+		    tr("Expected a JSON object containing game configs keyed by PPSA code."));
+		return;
+	}
+	const auto settings = document.object();
+	for (auto it = settings.constBegin(); it != settings.constEnd(); ++it) {
+		Configuration info;
+		QString       error;
+		if (!IsGameSettingsTitleId(it.key()) || !it.value().isObject() ||
+		    it.value().toObject().isEmpty() ||
+		    !info.SetGameSettings(it.value().toObject(), error)) {
+			QMessageBox::warning(parent, tr("Import failed"),
+			                     tr("Invalid game config for %1. %2").arg(it.key(), error));
+			return;
+		}
+	}
+	struct ImportedGame {
+		ConfigurationItem*             item;
+		std::unique_ptr<Configuration> info;
+	};
+	std::vector<ImportedGame> imported;
+	auto                      unmatched = settings.keys();
+	for (int index = 0; index < m_ui->cfgs_list->topLevelItemCount(); ++index) {
+		auto*      item     = static_cast<ConfigurationItem*>(m_ui->cfgs_list->topLevelItem(index));
+		const auto title_id = item->GetInfo().title_id.trimmed().toUpper();
+		if (!settings.contains(title_id)) {
+			continue;
+		}
+		auto    info = CreateConfiguration(*item);
+		QString error;
+		if (!info->SetGameSettings(settings.value(title_id).toObject(), error)) {
+			QMessageBox::warning(parent, tr("Import failed"),
+			                     tr("Invalid game config for %1. %2").arg(title_id, error));
+			return;
+		}
+		info->custom_settings = true;
+		imported.push_back({item, std::move(info)});
+		unmatched.removeAll(title_id);
+	}
+	auto message = imported.empty() ? tr("No matching games found in the library.")
+	                                : tr("Configs will be updated for the following games:\n");
+	for (const auto& game: imported) {
+		message += tr("\n%1 — %2").arg(game.info->title_id.trimmed().toUpper(), game.info->name);
+	}
+	if (!unmatched.isEmpty()) {
+		message += tr("\nGames not found: %1").arg(unmatched.join(QStringLiteral(", ")));
+	}
+	if (imported.empty()) {
+		QMessageBox::information(parent, tr("Import game configs"), message);
+		return;
+	}
+	QMessageBox confirmation(QMessageBox::Question, tr("Import game configs"), message,
+	                         QMessageBox::Apply | QMessageBox::Cancel, parent);
+	confirmation.setTextFormat(Qt::PlainText);
+	confirmation.setDefaultButton(QMessageBox::Cancel);
+	if (confirmation.exec() != QMessageBox::Apply) {
+		return;
+	}
+	for (auto& [item, info]: imported) {
+		const auto game_path = info->game_path;
+		delete m_custom_infos.take(game_path);
+		m_custom_infos.insert(game_path, info.release());
+		item->GetInfo().custom_settings = true;
+		item->Update();
+	}
+	WriteSettings();
+}
+
+void ConfigurationListWidget::ExportGameSettings(QWidget* parent) const {
+	QMap<QString, const ConfigurationItem*> games;
+	for (int index = 0; index < m_ui->cfgs_list->topLevelItemCount(); ++index) {
+		const auto* item = static_cast<ConfigurationItem*>(m_ui->cfgs_list->topLevelItem(index));
+		const auto& game = item->GetInfo();
+		const auto  title_id = game.title_id.trimmed().toUpper();
+		if (m_custom_infos.contains(game.game_path) && IsGameSettingsTitleId(title_id)) {
+			games.insert(tr("%1 — %2 (%3)").arg(title_id, game.name, game.game_path), item);
+		}
+	}
+	if (games.isEmpty()) {
+		QMessageBox::information(parent, tr("Export game config"),
+		                         tr("No game configs with a PPSA code were found."));
+		return;
+	}
+	bool       accepted = false;
+	const auto selected = QInputDialog::getItem(parent, tr("Export game config"), tr("Game:"),
+	                                            games.keys(), 0, false, &accepted);
+	if (!accepted) {
+		return;
+	}
+	const auto& game     = games.value(selected)->GetInfo();
+	const auto  title_id = game.title_id.trimmed().toUpper();
+	const auto  path =
+	    QFileDialog::getSaveFileName(parent, tr("Export game config"),
+	                                 title_id + QStringLiteral(".json"), tr("JSON files (*.json)"));
+	if (path.isEmpty()) {
+		return;
+	}
+	const auto settings =
+	    QJsonObject::fromVariantMap(m_custom_infos.value(game.game_path)->GameSettings());
+	const auto data =
+	    QJsonDocument(QJsonObject {{title_id, settings}}).toJson(QJsonDocument::Indented);
+	QSaveFile file(path);
+	if (!file.open(QIODevice::WriteOnly) || file.write(data) != data.size() || !file.commit()) {
+		QMessageBox::warning(parent, tr("Export failed"), file.errorString());
+	}
 }
 
 void ConfigurationListWidget::edit_input_mapping() {
@@ -1040,11 +1173,11 @@ void ConfigurationListWidget::show_context_menu(const QPoint& pos) {
 	                   tr("Remove save data..."), this, SLOT(remove_save_data()));
 	menu.addSeparator();
 	QAction* action_edit =
-	    menu.addAction(style()->standardIcon(QStyle::SP_FileIcon), tr("Edit game settings..."),
+	    menu.addAction(style()->standardIcon(QStyle::SP_FileIcon), tr("Edit game config..."),
 	                   this, SLOT(edit_configuration()));
 	QAction* action_delete =
 	    menu.addAction(style()->standardIcon(QStyle::SP_DialogDiscardButton),
-	                   tr("Clear custom settings"), this, SLOT(delete_configuartion()));
+	                   tr("Clear game config"), this, SLOT(delete_configuartion()));
 
 	if (item == nullptr) {
 		menu.addSeparator();

src/libs/libKernel.cpp

@@ -2831,6 +2831,8 @@ constexpr int32_t KERNEL_AIO_STATE_ABORTED    = 4;
 constexpr int32_t KERNEL_AIO_STATE_NOTIFIED   = 0x10000;
 constexpr int32_t KERNEL_AIO_MAX_QUEUE        = 512;
 constexpr int32_t KERNEL_AIO_MAX_REQUESTS     = 128;
+constexpr uint32_t KERNEL_AIO_WAIT_AND        = 1;
+constexpr uint32_t KERNEL_AIO_WAIT_OR         = 2;
 
 struct KernelAioResult {
 	int64_t  return_value;
@@ -2933,26 +2935,28 @@ int KYTY_SYSV_ABI KernelAioSubmitWriteCommands(KernelAioRwRequest* req, int32_t
 	return OK;
 }
 
+static int32_t kernel_aio_get_state(int32_t id, bool notify) {
+	if (!kernel_aio_is_valid_id(id)) {
+		return LibKernel::KERNEL_ERROR_ESRCH;
+	}
+
+	auto current = g_kernel_aio_state[id].load(std::memory_order_acquire);
+	while (notify && (current == KERNEL_AIO_STATE_COMPLETED || current == KERNEL_AIO_STATE_ABORTED)) {
+		if (g_kernel_aio_state[id].compare_exchange_weak(current, current | KERNEL_AIO_STATE_NOTIFIED,
+		                                                std::memory_order_acq_rel)) {
+			break;
+		}
+	}
+	return current == 0 ? LibKernel::KERNEL_ERROR_ESRCH : current;
+}
+
 int KYTY_SYSV_ABI KernelAioPollRequest(int32_t id, int32_t* state) {
 	PRINT_NAME();
 
 	if (state == nullptr) {
 		return LibKernel::KERNEL_ERROR_EFAULT;
 	}
-
-	if (!kernel_aio_is_valid_id(id)) {
-		*state = LibKernel::KERNEL_ERROR_ESRCH;
-		return OK;
-	}
-
-	auto current = g_kernel_aio_state[id].load(std::memory_order_acquire);
-	while (current == KERNEL_AIO_STATE_COMPLETED || current == KERNEL_AIO_STATE_ABORTED) {
-		if (g_kernel_aio_state[id].compare_exchange_weak(current, current | KERNEL_AIO_STATE_NOTIFIED,
-		                                                std::memory_order_acq_rel)) {
-			break;
-		}
-	}
-	*state = current == 0 ? LibKernel::KERNEL_ERROR_ESRCH : current;
+	*state = kernel_aio_get_state(id, true);
 	return OK;
 }
 
@@ -2969,31 +2973,53 @@ int KYTY_SYSV_ABI KernelAioPollRequests(int32_t* ids, int32_t num, int32_t* stat
 	return OK;
 }
 
-int KYTY_SYSV_ABI KernelAioWaitRequest(int32_t id, int32_t* state, uint32_t* usec) {
+int KYTY_SYSV_ABI KernelAioWaitRequests(int32_t* ids, int32_t num, int32_t* states, uint32_t mode,
+                                        uint32_t* usec) {
 	PRINT_NAME();
 
-	if (state == nullptr) {
+	if (ids == nullptr || states == nullptr) {
 		return LibKernel::KERNEL_ERROR_EFAULT;
 	}
-
-	if (!kernel_aio_is_valid_id(id)) {
-		return KernelAioPollRequest(id, state);
+	if (num <= 0 || num > KERNEL_AIO_MAX_REQUESTS ||
+	    (num != 1 && mode != KERNEL_AIO_WAIT_AND && mode != KERNEL_AIO_WAIT_OR)) {
+		return LibKernel::KERNEL_ERROR_EINVAL;
 	}
 
-	uint32_t waited  = 0;
-	auto     current = g_kernel_aio_state[id].load(std::memory_order_acquire);
-	while (current == KERNEL_AIO_STATE_PROCESSING) {
-		if (usec != nullptr && *usec != 0 && waited >= *usec) {
-			*state = current;
-			return LibKernel::KERNEL_ERROR_ETIMEDOUT;
+	const auto collect_states = [&](bool notify) {
+		int32_t finished = 0;
+		bool newly_complete = false;
+		for (int32_t i = 0; i < num; i++) {
+			const auto current = kernel_aio_get_state(ids[i], notify);
+			states[i] = current;
+			finished += current != KERNEL_AIO_STATE_PROCESSING;
+			newly_complete |= current == KERNEL_AIO_STATE_COMPLETED || current == KERNEL_AIO_STATE_ABORTED;
+		}
+		return finished == num || (mode == KERNEL_AIO_WAIT_OR && newly_complete);
+	};
+	const auto start = std::chrono::steady_clock::now();
+	for (;;) {
+		const bool complete = collect_states(false);
+		const auto elapsed = std::chrono::duration_cast<std::chrono::microseconds>(
+		                         std::chrono::steady_clock::now() - start).count();
+		const bool timed_out = usec != nullptr && elapsed >= *usec;
+		if (complete || timed_out) {
+			// A competing waiter can claim the newly completed request after the first scan.
+			if (collect_states(true)) {
+				if (usec != nullptr) {
+					*usec = elapsed < *usec ? *usec - static_cast<uint32_t>(elapsed) : 0;
+				}
+				return OK;
+			}
+			if (timed_out) {
+				return LibKernel::KERNEL_ERROR_ETIMEDOUT;
+			}
 		}
-
 		Common::Thread::SleepMicro(10);
-		waited += 10;
-		current = g_kernel_aio_state[id].load(std::memory_order_acquire);
 	}
+}
 
-	return KernelAioPollRequest(id, state);
+int KYTY_SYSV_ABI KernelAioWaitRequest(int32_t id, int32_t* state, uint32_t* usec) {
+	return KernelAioWaitRequests(&id, 1, state, KERNEL_AIO_WAIT_AND, usec);
 }
 
 int KYTY_SYSV_ABI KernelAioDeleteRequest(int32_t id, int32_t* ret) {
@@ -3376,6 +3402,7 @@ LIB_DEFINE(InitLibKernel_1) {
 	LIB_FUNC("2pOuoWoCxdk", KernelAioPollRequest);
 	LIB_FUNC("o7O4z3jwKzo", KernelAioPollRequests);
 	LIB_FUNC("KOF-oJbQVvc", KernelAioWaitRequest);
+	LIB_FUNC("lgK+oIWkJyA", KernelAioWaitRequests);
 	LIB_FUNC("XQ8C8y+de+E", KernelAioSubmitWriteCommands);
 	LIB_FUNC("nu4a0-arQis", KernelAioInitializeParam);
 	LIB_FUNC("il03nluKfMk", LibKernel::KernelRaiseException);

tests/KernelFileSystemTests.cpp

@@ -169,12 +169,14 @@ void TestAioBatches() {
   using Batch = int (KYTY_SYSV_ABI *)(int32_t *, int32_t, int32_t *);
   using Single = int (KYTY_SYSV_ABI *)(int32_t, int32_t *);
   using Wait = int (KYTY_SYSV_ABI *)(int32_t, int32_t *, uint32_t *);
+  using WaitBatch = int (KYTY_SYSV_ABI *)(int32_t *, int32_t, int32_t *, uint32_t, uint32_t *);
   const auto submit = reinterpret_cast<Submit>(find("HgX7+AORI58"));
   const auto poll = reinterpret_cast<Batch>(find("o7O4z3jwKzo"));
   const auto erase = reinterpret_cast<Batch>(find("Ft3EtsZzAoY"));
   const auto poll_one = reinterpret_cast<Single>(find("2pOuoWoCxdk"));
   const auto erase_one = reinterpret_cast<Single>(find("5TgME6AYty4"));
   const auto wait = reinterpret_cast<Wait>(find("KOF-oJbQVvc"));
+  const auto wait_batch = reinterpret_cast<WaitBatch>(find("lgK+oIWkJyA"));
   constexpr char Payload[] = "AIO payload";
   const int fd = FileSystem::KernelOpen("/savedata0/aio.dat", 0x602, 0777);
   Check(fd >= 3 && FileSystem::KernelWrite(fd, Payload, sizeof(Payload)) == sizeof(Payload),
@@ -191,14 +193,22 @@ void TestAioBatches() {
   }
   ids[2] = -1;
   std::array<int32_t, 3> states {-1, -1, -1};
-  Check(poll(ids.data(), ids.size(), states.data()) == OK &&
+  uint32_t timeout = 0;
+  Check(wait_batch(ids.data(), ids.size(), states.data(), 2, &timeout) == OK &&
             states[0] == 3 && states[1] == 3 && states[2] == Kernel::KERNEL_ERROR_ESRCH,
-        "batch poll writes every state and per-request invalid-ID error");
+        "OR wait observes completed reads and invalid-ID errors even with a zero timeout");
   Check(poll_one(ids[0], &states[0]) == OK && states[0] == (3 | 0x10000) &&
             poll(ids.data(), 2, states.data()) == OK &&
             states[0] == (3 | 0x10000) && states[1] == (3 | 0x10000) &&
             wait(ids[0], &states[0], nullptr) == OK && states[0] == (3 | 0x10000),
         "single and batch polls and waits share completion notification state");
+  timeout = 1000000;
+  Check(wait_batch(ids.data(), ids.size(), states.data(), 1, &timeout) == OK &&
+            states[0] == (3 | 0x10000) && states[1] == (3 | 0x10000) &&
+            states[2] == Kernel::KERNEL_ERROR_ESRCH && timeout <= 1000000 &&
+            wait_batch(ids.data(), ids.size(), states.data(), 2, nullptr) == OK &&
+            wait_batch(ids.data(), 1, states.data(), 0, nullptr) == OK,
+        "AND and OR waits return when all IDs are terminal and one-ID waits ignore mode");
   Check(erase(ids.data(), ids.size(), states.data()) == OK &&
             states[0] == OK && states[1] == OK && states[2] == Kernel::KERNEL_ERROR_ESRCH,
         "batch deletion writes per-request results");
@@ -221,6 +231,22 @@ void TestAioBatches() {
               }),
           "maximum-sized batch returns all invalid-ID errors");
   }
+  states.fill(42);
+  Check(wait_batch(nullptr, 1, states.data(), 1, nullptr) == Kernel::KERNEL_ERROR_EFAULT &&
+            wait_batch(ids.data(), 1, nullptr, 1, nullptr) == Kernel::KERNEL_ERROR_EFAULT &&
+            wait_batch(ids.data(), 0, states.data(), 1, nullptr) == Kernel::KERNEL_ERROR_EINVAL &&
+            wait_batch(ids.data(), 129, states.data(), 1, nullptr) == Kernel::KERNEL_ERROR_EINVAL &&
+            wait_batch(ids.data(), 2, states.data(), 0, nullptr) == Kernel::KERNEL_ERROR_EINVAL &&
+            wait_batch(ids.data(), 2, states.data(), 3, nullptr) == Kernel::KERNEL_ERROR_EINVAL &&
+            states == std::array<int32_t, 3> {42, 42, 42},
+        "invalid wait arguments do not modify outputs");
+  std::array<int32_t, 128> invalid_ids {};
+  std::array<int32_t, 128> errors {};
+  Check(wait_batch(invalid_ids.data(), invalid_ids.size(), errors.data(), 1, nullptr) == OK &&
+            std::all_of(errors.begin(), errors.end(), [](int32_t error) {
+              return error == Kernel::KERNEL_ERROR_ESRCH;
+            }),
+        "maximum-sized wait returns every invalid-ID error");
   Check(FileSystem::KernelClose(fd) == OK, "close AIO read fixture");
 }
 

tests/ShaderRecompilerComputeTests.cpp

@@ -32243,6 +32243,66 @@ TestCase DsAppendUsesEncodedGdsSelector() {
   return test;
 }
 
+TestCase DsOrderedCountAddressAndExec(u32 wave_size, bool pixel_counter) {
+  using O = ShaderOpcode;
+  TestCase test;
+  test.name = pixel_counter
+                  ? (wave_size == 64 ? "DsOrderedCountPixelAddressWave64"
+                                     : "DsOrderedCountPixelAddressWave32")
+                  : (wave_size == 64 ? "DsOrderedCountAddressAndExecWave64"
+                                     : "DsOrderedCountAddressAndExecWave32");
+  auto &code = test.code;
+  // Both the base and OFFSET0 are aligned independently; M0's low bits are
+  // a wave index, or a pixel packer ID, rather than a byte address or size.
+  AppendSMovLiteral(&code, 124, 0x01030003u);
+  code.push_back(EncodeSop1(0x04, 10, 126)); // Save the full EXEC mask.
+  code.push_back(EncodeVop2(0x25, 2, InlineU32(1), 0)); // ADDR value = lane + 1.
+  AppendVMovU32(&code, 9, 99); // DATA0 must not supply the count.
+  const u32 selected_lane = wave_size == 64 ? 47u : 7u;
+  const std::array<uint64_t, 3> masks = {
+      wave_size == 64 ? UINT64_MAX : uint64_t{UINT32_MAX},
+      uint64_t{1} << selected_lane, 0};
+  u32 counter = 1000;
+  for (bool exchange : {false, true}) {
+    for (size_t mode = 0; mode < masks.size(); ++mode) {
+      AppendVMovU32(&code, 3, 0xdeadbeefu);
+      AppendSMovLiteral(&code, 126, static_cast<u32>(masks[mode]));
+      AppendSMovLiteral(&code, 127, static_cast<u32>(masks[mode] >> 32u));
+      const u32 release_done = exchange && mode == 2 ? 3u : 0u;
+      const u32 offset1 = release_done | (pixel_counter ? 4u : 0u) |
+                          (exchange ? 0x10u : 0u);
+      code.push_back(EncodeDs0(0x3f, (offset1 << 8u) | 7u, true));
+      code.push_back(EncodeDs1(3, 9, 2));
+      code.push_back(EncodeSopp(0x0c, 0)); // Wait for the returned GDS value.
+      code.push_back(EncodeSop1(0x04, 126, 10)); // Restore EXEC before reading all lanes.
+      AppendStoreVgprAtLaneDwordOffset(&code, 3, 0,
+                                      static_cast<u32>(test.expected.size()));
+      test.expected.insert(test.expected.end(), wave_size, counter);
+      if (mode != 2) {
+        const u32 value = mode == 0 ? 1u : selected_lane + 1u;
+        counter = exchange ? value : counter + value;
+      }
+    }
+  }
+  AppendEnd(&code);
+  test.initial.assign(test.expected.size(), 0xdeadbeefu);
+  const u32 counter_address = pixel_counter ? 0x110u : 0x104u;
+  test.gds_initial.assign(0x120u / sizeof(u32), 100);
+  test.gds_initial[counter_address / sizeof(u32)] = 1000;
+  test.expected_gds = test.gds_initial;
+  test.expected_gds[counter_address / sizeof(u32)] = counter;
+  test.compute_info.threads_num[0] = wave_size;
+  test.compute_info.threads_num[1] = 1;
+  test.compute_info.threads_num[2] = 1;
+  test.compute_info.thread_ids_num = 1;
+  test.compute_info.wave_size = wave_size;
+  test.has_compute_info = true;
+  test.opcodes = {O::S_MOV_B32, O::S_MOV_B64, O::V_ADD_NC_U32, O::V_MOV_B32,
+                  O::DS_ORDERED_COUNT, O::S_WAITCNT, O::V_LSHLREV_B32,
+                  O::BUFFER_STORE_DWORD, O::S_ENDPGM};
+  return test;
+}
+
 TestCase DsAppendConsumeGdsRegionBounds() {
   using O = ShaderOpcode;
   struct Access {
@@ -35538,6 +35598,10 @@ std::vector<TestCase> MakeCases() {
   cases.push_back(DsAppendAllocatesAcrossWaves(32));
   cases.push_back(DsAppendAllocatesAcrossWaves(64));
   AddCase(DsAppendUsesEncodedGdsSelector);
+  for (u32 wave_size : {32, 64}) {
+    cases.push_back(DsOrderedCountAddressAndExec(wave_size, false));
+    cases.push_back(DsOrderedCountAddressAndExec(wave_size, true));
+  }
   AddCase(DsAppendConsumeGdsRegionBounds);
   AddCase(DsGdsSubdwordAndAtomicWrites);
   AddCase(DsReadWrite2Variants);
@@ -41059,6 +41123,14 @@ int main(int argc, char **argv) {
     }
     return 0;
   }
+  if (argc == 2 && std::strcmp(argv[1], "--ds-ordered-count-only") == 0) {
+    VulkanHarness vulkan;
+    for (u32 wave_size : {32, 64}) {
+      RunCase(&vulkan, DsOrderedCountAddressAndExec(wave_size, false));
+      RunCase(&vulkan, DsOrderedCountAddressAndExec(wave_size, true));
+    }
+    return 0;
+  }
   if (argc == 2 && std::strcmp(argv[1], "--ds-atomics-only") == 0) {
     VulkanHarness vulkan;
     for (const auto &test : MakeCases()) {